Open Collections

UBC Theses and Dissertations

UBC Theses Logo

UBC Theses and Dissertations

Long-run energy resource economics : reconciling uncertain carbon signals for integrated assessments… Ritchie, William Justin 2018

Your browser doesn't seem to have a PDF viewer, please download the PDF to view this item.

Item Metadata

Download

Media
24-ubc_2018_may_ritchie_william.pdf [ 95.78MB ]
Metadata
JSON: 24-1.0364580.json
JSON-LD: 24-1.0364580-ld.json
RDF/XML (Pretty): 24-1.0364580-rdf.xml
RDF/JSON: 24-1.0364580-rdf.json
Turtle: 24-1.0364580-turtle.txt
N-Triples: 24-1.0364580-rdf-ntriples.txt
Original Record: 24-1.0364580-source.json
Full Text
24-1.0364580-fulltext.txt
Citation
24-1.0364580.ris

Full Text

         B.Sc., Physics, B.Eng., Electrical Engineering, The University of North Carolina-Charlotte, 2008 M.A.Sc, Materials Engineering, The University of British Columbia, 2013  A THESIS SUBMITTED IN PARTIAL FULFILLMENT OF THE REQUIREMENTS FOR THE DEGREE OF  Doctor of Philosophy in THE FACULTY OF GRADUATE AND POSTDOCTORAL STUDIES (Resources, Environment and Sustainability)  THE UNIVERSITY OF BRITISH COLUMBIA (Vancouver)       Long-run energy resource economics: reconciling uncertain carbon signals for integrated assessments of global environmental changebyWilliam Justin Ritchie© William Justin Ritchie, 2018 March 2018 ii Abstract Studies of global environmental change require a long-term perspective that must contend with uncertain future human and Earth system processes. In this context, the scientific community frames possibilities for energy resource use with integrated assessment models (IAMs). IAMs combine various threads of scientific knowledge to allow systematic studies of hypothetical socioeconomic and technological developments.  Engineering-focused IAMs maintain economic concepts of energy use initiated by studies which responded to the 1970s energy crises by anticipating that growing demand for energy could rely on a coal backstop supply. Thus, many scenarios of vast coal combustion were produced to illustrate this outlook, where humanity had no choice but to become “the intelligent mole”. Such coal backstop scenarios played an important role in early climate model development because they provided a strong carbon signal.  Initial economic models of climate policy costs were based on assuming that the high-carbon backstop would always be cheaper than the low-carbon backstop. These ideas anchored expectations for future climate change as the IPCC assessment process was established, and continue to shape the uncertainty range considered by today’s studies. This thesis examines modeling concepts used to structure uncertain energy resource developments for long-term studies of global environmental change with a special focus on coal. The concept of a vast coal backstop energy supply is evaluated and these findings are applied to develop empirical constraints for an IAM coal supply curve. In the example considered by this thesis, an empirically consistent coal backstop scenario produces climate policy costs for a 1.5°C target equivalent to those for a 2°C goal that must overcome a vast coal backstop supply: the default configuration of many IAMs. An energy system phase space method is developed to map whether these long-run scenarios provide sufficient coverage of future uncertainties. It is found that IAM scenarios are needlessly constrained to produce outlooks for transitions toward a global energy supply with increasing carbon intensity. When these energy system scenarios are combined with socioeconomic projections for global per-capita income convergence, they serve to reproduce a style of reasoning that links aspirational equity goals with worst-case environmental consequences.          iii Lay Summary Studies of global environmental change involve processes that require a multi-decade perspective on developments in future society. To frame uncertain future possibilities, the scientific community uses models that combine hypothetical concepts for long-run socioeconomics and energy technologies. These models have their conceptual origin in the aftermath of the 1970s energy crises, where securing oil and gas supplies seemed precarious. At that time, coal was thought to be the most reliable energy source for ever growing demand, and many future scenarios were created to illustrate how coal was the best option to substitute for oil and gas. These outlooks became the template for the worst-case climate change scenarios still used by the scientific community today. This thesis studies the way energy resources are conceptualized for research on climate change, with a focus on calibrating the range of future energy system developments that could result from an updated understanding of coal’s potential role.    iv Preface The research program underlying this thesis addresses the context of global change research which extends over the course of many decades. Subsequent chapters contribute concepts to aid in studies of the future global energy system by using methods that can be broadly classified under the field of energy economics and data science.  I am primarily responsible for writing and conducting this work. My PhD supervisor Hadi Dowlatabadi was very helpful in shaping these concepts, verifying my analysis, tolerating my incoherent first drafts and drawing attention to the conceptual shortfalls they contained.  A working paper edition of the material included in Chapter 2 was published as Ritchie, J., & Dowlatabadi, H. (2017). Evaluating the Learning-by-Doing Theory of Long-Run Oil, Gas, and Coal Economics (No. 17-14). Resources for the Future, and it is currently undergoing peer review.  A version of Chapter 3 was published as Ritchie, J., & Dowlatabadi, H. (2017). The 1,000 GtC coal question: Are cases of vastly expanded future coal combustion still plausible? Energy Economics, 65, 16–31. http://doi.org/10.1016/j.eneco.2017.04.015. Much of the content in Chapter 5 was published as Ritchie, J., & Dowlatabadi, H. (2017). Why do climate change scenarios return to coal? Energy, 140, 1276–1291. http://doi.org/10.1016/j.energy.2017.08.083 A version of Chapter 6 was published as Ritchie, J., Dowlatabadi, H., 2018. Defining climate change scenario characteristics with a phase space of cumulative primary energy and carbon intensity. Environ. Res. Lett. 13, 024012. http://doi.org/10.1088/1748-9326/aaa494 Chapters 2, 4 and 7 are in various stages of reviews and revision.   v Table of Contents  Abstract  ................................................................................................................................................. ii Lay Summary ............................................................................................................................................. iii Preface  ................................................................................................................................................ iv Table of Contents ....................................................................................................................................... v List of Tables ............................................................................................................................................ viii List of Figures ............................................................................................................................................ ix List of Key Abbreviations ........................................................................................................................ xiv Acknowledgements ................................................................................................................................ xvii Dedication  ............................................................................................................................................. xviii Chapter 1: Introduction to studies of energy system uncertainties relevant for global environmental change research ..................................................................................................................... 1 1.1 How to study energy-economy models? A world in the model that can be examined in stages ......................................................................................... 1 1.2 Structuring deep uncertainty in future energy scenarios: two theories ....... 2 1.3 Quantitative studies of global change: integrated assessment models ...... 8 1.4 Summary ....................................................................................................... 24 Chapter 2: Evaluating the learning-by-doing theory of long-run fossil energy economics .................. 27 2.1 Empirical trends in upstream oil and gas (1978-2008): production costs, market dynamics and reserves .................................................................... 30 2.2 Nordhaus (2009) on the perils of a learning model: extended to energy resources ....................................................................................................... 44 2.3 Implications of chosen learning rates for long-run energy economics and climate change mitigation cost projections ................................................. 52 2.4 Summary ....................................................................................................... 61 Chapter 3: The 1,000 GtC coal question: Are scenarios of vastly expanded future coal combustion still plausible? ........................................................................................................................ 63 3.1 Reference case trends in coal reserves, resources, production, and prices ....................................................................................................................... 66 3.2 Interpreting the information dynamics of global coal reserves: why are modern reserve assessments smaller than legacy outlooks? ................... 73 3.3 Implications of modern coal reserve assessments for the conceptual basis of future energy scenarios ............................................................................ 78 3.4 Case study: coal resources in climate change scenarios .......................... 83  vi 3.5 Summary - recommendations for a recalibrated 21st-century coal reference case .............................................................................................. 91 Chapter 4: An integrated climate change assessment with empirically constrained coal supply and transportation demand ......................................................................................................... 94 4.1 Empirically constrained coal: a prototype IAM reference case ................. 95 4.2 The GCAM reference energy system (GCAM-ref) and its high-carbon backstop ........................................................................................................ 99 4.3 Developing alternate GCAM reference cases based on empirically constrained supply and demand factors ................................................... 110 4.4 Climate policy case I - solving for a moderate mitigation target: RCP4.5 ..................................................................................................................... 129 4.5 Climate policy case II – solving for a strong mitigation target of well below 2˚ .................................................................................................................. 147 4.6 Summary ..................................................................................................... 155 Chapter 5: Why do climate change scenarios return to coal?............................................................. 157 5.1 Energy system reference cases in the SSP-RCP framework: a brief meta-analysis illustrates the return to coal hypothesis applied by IAMs .......... 161 5.2 Enough coal for the end of time: the learning-by-extracting theory of fossil energy resource supply .............................................................................. 170 5.3 Energy scenarios that return to coal: looking at the future with one eye on perfect foresight and no hindsight ............................................................. 176 5.4 The return to coal hypothesis and SSP5-RCP8.5 .................................... 187 5.5 Why climate change scenarios return to coal: a strong carbon signal to overwhelm uncertainty in climate models ................................................. 188 5.6 Summary ..................................................................................................... 194 Chapter 6: A general approach to energy system uncertainty in global change scenarios – structure neutral phase methods for graphical and analytical solutions ......................................... 195 6.1 Developing an energy system phase space for climate change scenarios ..................................................................................................................... 197 6.2 IPCC AR5 energy system reference cases: a k-means cluster analysis of IAM no policy scenarios ............................................................................. 203 6.3 AR5 reference case clusters: described as characteristics of oil, gas, and coal combustion .......................................................................................... 210 6.4 Comparing IPCC AR5 energy system outlooks to IEA and SSP scenarios in the RCP phase space............................................................................. 216 6.5 Summary ..................................................................................................... 223 Chapter 7: Maintaining Malthus (1798): a global change scenario archetype continues .................. 226  vii 7.1 Malthus (1798): a global change archetype established ......................... 227 7.2 Maintaining Malthus (1798) in today’s global change scenarios ............. 229 7.3 Linking narratives of equity with ecological disaster: an explicit contrast illustrated by SSP5 vs. SSP4 storylines .................................................... 247 7.4 Model structures that produce the high-coal transportation future: a case study of GCAM-ref ...................................................................................... 252 7.5 Summary ..................................................................................................... 256 Chapter 8: Conclusion ............................................................................................................................ 259 8.1 The story of the models produce a scientific meta-narrative ................... 260 8.2 Journal submissions associated with this thesis ...................................... 262 8.3 Research contributions of this thesis ......................................................... 264 Chapter 9: Summary of implications for the ongoing climate policy discourse .................................. 266 Bibliography ............................................................................................................................................ 266 Appendices ............................................................................................................................................. 290 Appendix A - Aggregation of individual nations to the five region levels used in the AR5 database ...................................................................................................... 290 Appendix B - AR5 population scenarios ........................................................................... 292 Appendix C - Excerpts of peer reviews collected from journal submissions related to material in this thesis .................................................................................. 297    viii List of Tables Table 1-1 -  Oil, Gas and Coal: Production, Reserves and Resources as applied in EMF 27 ................... 22 Table 2-1 -  Alternate GCAM reference cases for conventional oil learning rates – effect on cost of oil (% change from unmodified GCAM-ref) ........................................................................................... 55 Table 2-2 -  Alternate GCAM reference cases for conventional oil learning rates – effect on oil production phase out and backstop deployment .......................................................................................... 57 Table 3-1 -  Global fossil energy resource base for two EMF studies (EMF 1995; McCollum et al. 2014) ....................................................................................................................................................... 64 Table 3-2 -  Mohr et al. (2015) estimate of cumulative production by rank and remaining ultimately recoverable resource (URR) – in exajoules (EJ) ....................................................................... 81 Table 4-1 -  Summary of four alternate GCAM reference case properties: climate, primary energy, coal emissions, and liquid fuels ........................................................................................................ 129 Table 4-2 -  Summary of four alternate GCAM reference case mitigation economics for a RCP4.5 consistent scenario .................................................................................................................... 147 Table 4-3 -  Summary of four alternate GCAM reference case mitigation economics 1.5˚C consistent scenarios .................................................................................................................................... 154 Table 5-1 -  IAM marker models and scenarios for climate change pathways .......................................... 159 Table 6-1 -  Conventional fossil reserve and resource estimates based on BGR (2015) ......................... 215 Table 6-2 -  Summary of fossil energy system reference case characteristics applied by AR5 WGIII scenarios to describe RCP ranges ........................................................................................... 216 Table 6-3 -  Cumulative oil, gas and coal resource use (2016-2100) - IEA, AR5 and SSP reference case scenarios .................................................................................................................................... 220 Table 7-1 -  AR5 WGIII scenarios: models and variables considered ........................................................ 231   ix List of Figures Figure 1.1  Series of cascading uncertainties characteristic of global change research: focus on climate change ............................................................................................................................................ 5 Figure 1.2  IIASA Energy in a Finite World – two global energy supply scenarios and historical ............... 9 Figure 1.3   McKelvey (1972) system for resource classification ................................................................. 21 Figure 2.1  Influence of learning rates on calculations of future oil supply costs ....................................... 29 Figure 2.2a  Upstream productivity trends in oil & gas (1978-2008).............................................................. 32 Figure 2.2b  Trends in direct lifting costs per barrel of oil equivalent for EIA FRS Companies (1980-2009) ....................................................................................................................................................... 32 Figure 2.3a Market price and upstream productivity trends multi-plot (1978-2008) .................................... 37 Figure 2.3b  Marginal upstream productivity rate per dollar of change in market price (1978-2008) ......... 38 Figure 2.4a  Development proportion of upstream costs (1978-2009) .......................................................... 42 Figure 2.4b  Two distinct cycles in reserves: quantity of proved reserves-to-Brent prices for oil (1955-2015) ............................................................................................................................................. 43 Figure 2.4c Range of estimates for ratio of proved reserves to oil price ..................................................... 44 Figure 2.5a  Relationship of productivity rates in Equation 2.8 ...................................................................... 50 Figure 2.5b  Possible values for productivity gains from each component for a measured learning rate of 1% per year .................................................................................................................................. 51 Figure 2.6  Cumulative discounted cost of 21st-century oil supply ............................................................. 53 Figure 2.7a 21st-century oil cost trajectories from GCAM alternate learning reference cases .................. 55 Figure 2.7b  Cumulative conventional oil supply curves (cost-quantity) produced by GCAM alternate learning scenarios ........................................................................................................................ 56 Figure 2.8a Conventional oil production outlooks for scenarios of alternate 21st-century learning rates . 58 Figure 2.8b Unconventional oil production outlooks for scenarios of alternate 21st-century learning rates ....................................................................................................................................................... 58 Figure 2.8c Cumulative oil and coal production outlooks for alternate learning .......................................... 59 Figure 2.9a Total 21st-century carbon emissions from conventional oil learning rates .............................. 60 Figure 2.9b Year-2100 discounted carbon price across alternate scenarios .............................................. 60 Figure 3.1a  Trends in global coal market benchmark prices ........................................................................ 68 Figure 3.1b  Annual hard coal production as reported by IEA Coal Information Reports, WEC and BGR 69 Figure 3.1c  Coal reserves in mass units from successive WEC and BGR reports indexed to WEC (2001) ....................................................................................................................................................... 69 Figure 3.1d  Reserve-to-production ratio for global coal (mass-basis) .......................................................... 70 Figure 3.2a  Coal reserves as a fraction of recent coal resource assessments ........................................... 72 Figure 3.2b  Salient assessments of the total coal resource base ................................................................ 72 Figure 3.3a  USGS (2009) coal recoverability study detail by basin ............................................................. 75 Figure 3.3b  Process for calculating economically recoverable coal adapted from USGS (2009) .............. 75 Figure 3.4   Geologic carbon supply curve for Earth’s fossil occurrences (Rogner 1997) mapped to coal reported at that time ..................................................................................................................... 83 Figure 3.5 Cumulative 21st-century primary energy from coal in IPCC AR high-emission scenario baseline cases .............................................................................................................................. 86 Figure 3.6a  Cumulative 21st-century primary energy from coal in original RCP baselines ........................ 87 Figure 3.6b  Marchetti (1977) curves of world primary energy substitution for coal, oil and gas ................ 88 Figure 3.6c 21st-century cumulative CO2 emissions from fossil fuels in final RCP scenarios .................. 88 Figure 3.7 Cumulative primary energy from coal in the 223 AR5 WGIII baseline runs ............................ 90 Figure 4.1a Supply curves for coal adopted by IAMs .................................................................................... 97 Figure 4.1b Original RCP8.5 total primary energy supply ............................................................................. 98  x Figure 4.1c A prototype long-run energy supply outlook based on modern coal assessments ................. 99 Figure 4.2a  GCAM-ref energy system (1990-2100) .................................................................................... 103 Figure 4.2b  GCAM-ref refinery inputs by technology (1990-2100) ............................................................. 103 Figure 4.2c  GCAM-ref refinery inputs per-capita (1971-2100) ................................................................... 104 Figure 4.2d  GCAM-ref primary energy resource costs (1990-2100) .......................................................... 104 Figure 4.3a Case A: reference case of carbon supply curves with geologically present carbon supply . 107 Figure 4.3b Case B: carbon and low-carbon supply curves with empirically constrained carbon supply107 Figure 4.4a  CO2 Emissions in GCAM-ref by sector .................................................................................... 109 Figure 4.4b  Radiative forcing in GCAM-ref by GHG component ................................................................ 109 Figure 4.5a  Supply curve for coal empirically constrained from I-GCAM-ref to II-GCAM-ECCo ............. 112 Figure 4.5b  Coal supply costs in I-GCAM-ref and II-GCAM-ECCo ............................................................ 113 Figure 4.5c  Change in primary energy by resource: II-GCAM-ECCo from I-GCAM-ref ........................... 114 Figure 4.5d  Change in refined fuels: II-GCAM-ECCo from I-GCAM-ref..................................................... 114 Figure 4.5e  Change in CO2 emissions by sector: II-GCAM-ECCo from I-GCAM-ref ............................... 115 Figure 4.6a  Transportation service demand income elasticities for the original GCAM reference case . 117 Figure 4.6b  Transportation service demand income elasticities for the alternate III-GCAM-DECOR reference case ............................................................................................................................ 118 Figure 4.6c  Global transportation service demand by mode in I-GCAM-ref .............................................. 119 Figure 4.6d  Global transportation service demand by mode in III-GCAM-DECOR .................................. 120 Figure 4.6e  Global transportation service demand change by major modes in III-GCAM-DECOR  from I-GCAM-ref .................................................................................................................................... 121 Figure 4.6f  Change in primary energy by resource: III-GCAM-DECOR from I-GCAM-ref ...................... 121 Figure 4.6g Change in refined fuels: III-GCAM-DECOR from I-GCAM-ref ................................................ 122 Figure 4.6h  Change in CO2 emissions by sector: II-GCAM-ECCo from I-GCAM-ref ............................... 122 Figure 4.7a  Change in primary energy by resource: IV-GCAM-STOiC from I-GCAM-ref ........................ 123 Figure 4.7b Change in refined fuels: IV-GCAM-STOiC from I-GCAM-ref .................................................. 124 Figure 4.7c  Change in CO2 emissions by sector: IV-GCAM-STOiC from I-GCAM-ref ............................ 124 Figure 4.8a  Change in radiative forcing by component: II-GCAM-ECCo from I-GCAM-ref...................... 125 Figure 4.8b  Change in radiative forcing by component: III-GCAM-DECOR from I-GCAM-ref ................. 126 Figure 4.8c  Change in radiative forcing by component: IV-GCAM-STOiC from I-GCAM-ref ................... 126 Figure 4.8d  Change in global mean surface temperature: II-GCAM-ECCo ............................................... 127 Figure 4.8e  Change in global mean surface temperature: III-GCAM-DECOR .......................................... 127 Figure 4.8f  Change in global mean surface temperature: IV-GCAM-STOiC ............................................ 128 Figure 4.9a  I-GCAM-ref-RCP4.5 energy system (1990-2100) ................................................................... 131 Figure 4.9b Change in primary energy by resource: I-GCAM-ref-RCP4.5 from I-GCAM-ref ................... 131 Figure 4.9c Change in primary energy by resource: II-ECCo-RCP4.5 from I-ref-RCP4.5 ....................... 132 Figure 4.9d  Change in primary energy by resource: III-DECOR-RCP4.5 from I-ref-RCP4.5 ................... 132 Figure 4.9e  Change in primary energy by resource: IV-STOiC-RCP4.5 from I-ref-RCP4.5..................... 133 Figure 4.10a I-GCAM-ref-RCP4.5 refinery inputs by technology (1990-2100) ........................................... 134 Figure 4.10b Change in refined fuels: I-GCAM-ref-RCP4.5 from I-GCAM-ref ............................................ 135 Figure 4.10c Change in refined fuels: II-ECCo-RCP4.5 from I-ref-RCP4.5 ................................................ 135 Figure 4.10d Change in refined fuels: III-DECOR-RCP4.5 from I-ref-RC4.5 .............................................. 136 Figure 4.10e Change in refined fuels: IV-STOiC-RCP4.5 from I-ref-RC4.5 ................................................ 137 Figure 4.11a CO2 Emissions in I-GCAM-ref-RCP4.5 by sector .................................................................. 138 Figure 4.11b Change in CO2 emissions by sector: II-ECCo-RCP4.5 from I-ref-RCP4.5 .......................... 139 Figure 4.11c Change in CO2 emissions by sector: III-DECOR-RCP4.5 from I-ref-RCP4.5 ...................... 139 Figure 4.11d Change in CO2 emissions by sector: IV-STOiC-RCP4.5 from I-ref-RCP4.5 ........................ 140  xi Figure 4.12a Change in total cumulative primary energy with carbon capture and sequestration: [II], [III] and [IV] from I-ref-RCP4.5 ........................................................................................................ 141 Figure 4.12b Change in global mean surface temperature: GCAM RCP4.5 scenarios ............................. 141 Figure 4.12c Radiative forcing in I-GCAM-ref-RCP4.5 by GHG component .............................................. 142 Figure 4.13a Cumulative undiscounted climate policy (2020-2100) costs to achieve RCP4.5 for four alternate GCAM reference cases [I-IV] .................................................................................... 143 Figure 4.13b Cumulative discounted climate policy costs to achieve RCP4.5 for four alternate GCAM reference cases (2020-2100) ................................................................................................... 144 Figure 4.13c Undiscounted carbon prices to achieve RCP4.5 for four alternate GCAM reference cases (2020-2100) ............................................................................................................................... 145 Figure 4.13d Discounted carbon prices to achieve RCP4.5 for four alternate GCAM reference cases     (2020-2100) ............................................................................................................................... 146 Figure 4.14a Change in global mean surface temperature: GCAM RCP3.1 scenarios ............................. 148 Figure 4.14b Change in global mean surface temperature: GCAM 1.5˚ - overshoot (OS) scenarios ...... 149 Figure 4.14c Radiative forcing in GCAM RCP3.1 by GHG component ...................................................... 149 Figure 4.14d Radiative forcing in GCAM 1.5˚ overshoot (OS) by GHG component .................................. 150 Figure 4.15a Discounted cumulative climate policy costs to achieve 1.5˚C of warming without overshoot: RCP3.1 scenarios for baselines [I] and [IV] ............................................................................ 151 Figure 4.15b Discounted cumulative climate policy costs to achieve 1.5˚C of warming with overshoot: scenarios for baselines [I] and [IV] ........................................................................................... 152 Figure 4.15c Discounted carbon prices to achieve 1.5˚C of warming without overshoot: RCP3.1 scenarios    for baselines [I] and [IV] ............................................................................................................ 153 Figure 4.15d Discounted carbon prices to achieve 1.5˚C of warming with overshoot: scenarios for baselines [I] and [IV] ................................................................................................................. 154 Figure 5.1   The four representative concentration pathways (RCPs) ....................................................... 158 Figure 5.2 IAM uncertainty ranges for 21st-century fossil resource production outlooks from IPCC AR5 with SSP marker scenarios and RCP reference cases ........................................................... 162 Figure 5.3 Historical trends in per-capita energy resource use compared to the IPCC AR5 range with SSP marker scenarios and RCP reference case average ...................................................... 167 Figure 5.4 Examples of learning-by-extracting (LBE) oil, gas and coal supply curves used in IAMs to structure energy supply projections for 21st-century climate change scenarios .................... 174 Figure 5.5a  Reserves-to-production (R-P) ratio for oil (black line) and coal (red-line) from 1965-2015 .. 179 Figure 5.5b  Horizon of information for 21st-century energy demand projections (green line) compared with data vintages of coal (red line) and oil (black line) ........................................................... 180 Figure 5.5c  Ratio of coal substitution in AR5 scenarios .............................................................................. 181 Figure 5.6  Multi-decade projections of primary energy from IIASA Energy in a Finite World and a range of US energy policy studies ....................................................................................................... 183 Figure 5.7a  CMIP5 Global Mean Temperature (tas) projections from GCMs and ESMs with historical observation from NASA GISS ................................................................................................... 192 Figure 5.7b  CMIP5 Global Mean Temperature (tas) projections from GCMs and ESMs – RCP scenarios ..................................................................................................................................................... 193 Figure 5.7c CMIP5 Global Mean Temperature (tas) projections from GCMs and ESMs with historical observation from NASA GISS ................................................................................................... 193 Figure 6.1a  RCP phase space diagram ....................................................................................................... 200 Figure 6.1b  Points of reference within the phase space ............................................................................. 201 Figure 6.1c  IAM scenarios in the phase space ............................................................................................ 202 Figure 6.2a  AR5 reference cases plotted as time-series – primary energy  ........................................... 204 Figure 6.2b  AR5 reference cases plotted as time-series - oil ................................................................... 205 Figure 6.2c  AR5 reference cases plotted as time-series - gas ................................................................ 205  xii Figure 6.2d  AR5 reference cases plotted as time-series  - coal ............................................................... 206 Figure 6.3  RCP phase space populated with IPCC AR5 WGIII scenarios from IAMs.......................... 207 Figure 6.4a  k-means cluster analysis of AR5 WGIII database reference cases for k = 12 .................... 208 Figure 6.4b  k-means cluster analysis elbow plot of WGIII database reference cases for 2 ≤ k ≤ 12 .... 209 Figure 6.5  RCP phase space populated with cluster centers IPCC AR5 WGIII scenarios .................. 210 Figure 6.6a  Cumulative oil, gas and coal combustion in AR5 reference cases shaded by cluster grouping (2016-2100) - oil [O] ................................................................................................. 211 Figure 6.6b  Cumulative oil, gas and coal combustion in AR5 reference cases shaded by cluster grouping (2016-2100) - gas [G] .............................................................................................. 212 Figure 6.6c  Cumulative oil, gas and coal combustion in AR5 reference cases shaded by cluster grouping (2016-2100) - coal [C] .............................................................................................. 213 Figure 6.7  Box plots of resource combustion by cluster ......................................................................... 214 Figure 6.8  Time-series plots of SSP IAM scenarios and IEA reference case global energy system .. 217 Figure 6.9a  Energy system phase space - IEA, SSP and IPCC AR5 scenarios (2016-2100) ............... 219 Figure 6.9b  Energy system phase space for IEA, SSP and IPCC AR5 scenarios (2016-2100) ............ 219 Figure 6.10a Climate policy trajectories for 2˚ per SSP marker scenario (RCP3.4 - 2020-2100) ........... 222 Figure 6.10b Climate policy trajectories for 2˚ - SSP alternate scenarios (RCP3.4 - 2050-2100) .......... 223 Figure 7.1a  IPCC AR5 WGIII database scenarios of global GDP per capita .......................................... 232 Figure 7.1b  IPCC AR5 WGIII database scenarios of transportation service demand ............................ 233 Figure 7.1c  IPCC AR5 WGIII database scenarios of final energy for transportation services ............... 233 Figure 7.2a  IPCC AR5 WGIII database scenarios of regional GDP per-capita: OECD ......................... 235 Figure 7.2b  IPCC AR5 WGIII database scenarios of regional GDP per-capita: Economies in transition .................................................................................................................................................. 236 Figure 7.2c  IPCC AR5 WGIII database scenarios of regional GDP per-capita: Asia ............................. 237 Figure 7.2d  IPCC AR5 WGIII database scenarios of regional GDP per-capita: Middle East and Africa .................................................................................................................................................. 238 Figure 7.2e IPCC AR5 WGIII database scenarios of regional GDP per-capita: Latin America ............. 239 Figure 7.3a  Global liquid fuel production: AR5 WGIII scenarios .............................................................. 240 Figure 7.2b  Coal-to-liquids production: AR5 WGIII scenarios .................................................................. 241 Figure 7.2c  Proportion of liquid fuel supply from coal backstop supply: AR5 WGIII ............................... 242 Figure 7.3a  IPCC AR5 WGIII database scenarios of regional primary energy resource use per-capita: OECD ....................................................................................................................................... 243 Figure 7.3b IPCC AR5 WGIII database scenarios of regional primary energy resource use per-capita: Economies in Transition .......................................................................................................... 244 Figure 7.3c IPCC AR5 WGIII database scenarios of regional primary energy resource use per-capita: Asia ........................................................................................................................................... 245 Figure 7.3d IPCC AR5 WGIII database scenarios of regional primary energy resource use per-capita: Middle East and Africa ............................................................................................................ 246 Figure 7.3e IPCC AR5 WGIII database scenarios of regional primary energy resource use per-capita: Latin America ........................................................................................................................... 247 Figure 7.4a SSP5 and SSP4 scenarios of regional GDP per-capita convergence ................................. 250 Figure 7.4b SSP5 and SSP4 scenarios of regional GDP primary energy convergence ........................ 251 Figure 7.4c SSP5 and SSP4 scenarios of regional transportation final energy per-capita convergence .................................................................................................................................................. 252 Figure 7.5a GCAM-ref and SSP2 socioeconomic scenarios .................................................................... 253 Figure 7.5b GCAM-ref: global passenger travel demand by mode .......................................................... 254 Figure 7.5c Cost of transportation services in GCAM-ref by mode.......................................................... 254 Figure 7.5d GCAM-ref refinery inputs ........................................................................................................ 255  xiii Figure 8.1 Year 2100 forcing by component in RCP scenarios, IPCC (1990) BAU and their consistent hypotheses of the global energy system ................................................................................ 262 Figure 8.2 Timeline of journal submissions associated with this thesis ................................................. 263 Figure B.1 Population scenarios AR5 WGIII scenarios: World ............................................................... 292 Figure B.2 Population scenarios AR5 WGIII scenarios: OECD .............................................................. 293 Figure B.3 Population scenarios AR5 WGIII scenarios: Economies in transition.................................. 294 Figure B.4 Population scenarios AR5 WGIII scenarios: Asia ................................................................. 294 Figure B.5 Population scenarios AR5 WGIII scenarios: Middle East and Africa ................................... 295 Figure B.6  Population scenarios AR5 WGIII scenarios: Latin America .................................................. 296         xiv List of Key Abbreviations Alphabetical  Proved reserves ................................................................................................................................................ 1P Proved and possible reserves ......................................................................................................................... 2P Intergovernmental Panel on Climate Change Fifth Assessment Report .................................................... AR5 Business-as-usual ......................................................................................................................................... BAU Billion cubic meters ......................................................................................................................................... bcm Bundesanstalt für Geowissenschaften und Rohstoffe (German Ministry of Natural Resources) ............ BGR Barrel of oil equivalent ..................................................................................................................................... boe Biomass energy carbon capture sequestration ...................................................................................... BECCS Biomass-to-liquids .......................................................................................................................................... BTL Compound annual growth rate .................................................................................................................. CAGR Capital expenditures .................................................................................................................................... capex Carbon capture sequestration ...................................................................................................................... CCS Coupled model intercomparison project ..................................................................................................... CMIP Carbon dioxide emissions .............................................................................................................................. CO2 Coal-to-liquids ................................................................................................................................................. CTL Energy Information Administration ................................................................................................................. EIA Economies in Transition .................................................................................................................................. EIT Exajoule .............................................................................................................................................................. EJ Energy modeling forum ................................................................................................................................. EMF Earth systems models ................................................................................................................................. ESMs Fischer-Tropsch ................................................................................................................................................ FT Fossil fuels and industry ................................................................................................................................ FF&I General circulation models ........................................................................................................................ GCMs Gross domestic product ................................................................................................................................ GDP Greenhouse gas emissions ........................................................................................................................GHGs Gigajoule ........................................................................................................................................................... GJ Global mean surface temperature increase above pre-industrial levels .............................................. ΔGMST Gigatons carbon emissions ............................................................................................................................ GtC Gas-to-liquids .................................................................................................................................................. GTL Gigatons of oil equivalent .............................................................................................................................. Gtoe  xv Integrated assessment models .................................................................................................................... IAMs International Energy Agency ........................................................................................................................... IEA International Institute for Applied Systems Analysis ................................................................................. IIASA Intergovernmental Panel on Climate Change ............................................................................................ IPCC Kilowatt-hours ................................................................................................................................................ kWh Latin America ................................................................................................................................................. LAM Learning-by-extracting .................................................................................................................................... LBE Middle East and Africa .................................................................................................................................. MAF Millions of barrels per day (of oil).................................................................................................. mbd or mbdoe Monte Carlo simulation.................................................................................................................................... MC Market exchange rates .................................................................................................................................. MER Million tons of oil equivalent .......................................................................................................................... Mtoe Natural gas liquids ....................................................................................................................................... NGLs Organisation for Economic Co-operation and Development ................................................................... OECD Organization of the Petroleum Exporting Countries ................................................................................. OPEC Operational expenditures .............................................................................................................................. opex Per annum (annually) ...................................................................................................................................... p.a. Probability density functions ........................................................................................................................ PDFs Passenger kilometers traveled ...................................................................................................................... PKT Purchasing power parity ................................................................................................................................ PPP Representative concentration pathways .................................................................................................... RCPs Reforming economies ................................................................................................................................... REF Radiative forcing ............................................................................................................................................... RF Reserves-to-production ratio ..........................................................................................................................R-P Shared socioeconomic pathways ................................................................................................................ SSPs Total primary energy supply ....................................................................................................................... TPES Underground coal gasification ...................................................................................................................... UCG United Nations Framework Convention on Climate Change .............................................................. UNFCCC Ultimately recoverable resource ................................................................................................................... URR United States Dollar ....................................................................................................................................... USD United States Geological Survey ............................................................................................................... USGS World Energy Council ................................................................................................................................... WEC Working Group III ......................................................................................................................................... WGIII  xvi Zettajoule ............................................................................................................................................................ ZJ           xvii Acknowledgements  I’m forever grateful to my supervisor Hadi Dowlatabadi for his patience and advice in seeking to make this work ever more internally coherent, and that he somehow believed I could write a thesis after years of working in applied engineering, where eloquent writing skills were not emphasized, leaving my early drafts with a very long way to go. My committee members Margaret Schabas and Richard Howarth have been excellent guides by directing my focus and orienting this thesis to properly engage with concepts in economic thought. The support of the Canadian Social Sciences and Humanities Research Council through a Vanier Scholarship was crucial, because it provided the space and time to address this topic in a deep way. Further, I thank the US National Science Foundation (SES-1463493) which allowed me to present portions of this work to the exceptional research team at the Carnegie Mellon Center for Climate and Energy Decision Making (CEDM). This thesis was also supported by the Pacific Institute for Climate Solutions (Transition to a low GHG economy, 36170–50280). Further, I appreciate the dozens of anonymous peer reviewers who confirmed the technical accuracy and value of these ideas with an array of comments that only encouraged me to press onward. I thank my parents, Bill and Trish Ritchie for instilling the value of education from an early age, and for their continual encouragement as I’ve pursued that path. My wife Jane has been an incredible partner in this work, and so I dedicate this thesis to her.    xviii Dedication             For Jane.   1 Chapter 1: Introduction to studies of energy system uncertainties relevant for global environmental change research   Studies of global environmental change investigate human interactions with the Earth system on time scales of many decades (Camill, 2010; Munn, 2002; Schellnhuber et al., 1997). This context extends beyond the frontiers of today’s knowledge, calling for a consideration of possible developments in future society and technology.  Such a long-range approach inherently enters domains of deep uncertainty, defined by Lempert et al. (2003) as situations where scientists, analysts, and decision-makers do not know or cannot agree on: (i) the appropriate conceptual and mathematical models for understanding dynamics of the major driving forces that will shape the future, (ii) the probability distributions of key variables or parameters, or (iii) the way to value desirability of alternative outcomes.  These three characteristics aptly describe research on the economics of energy resources in the long-run. Processes of fossil resource extraction and combustion are recognized as a key driver of global environmental change (Alcamo et al., 1996; Dincer, 1999; Grubler, 2012; IPCC, 2014; Quéré et al., 2016; Vitousek et al., 1997). Therefore, it is important to understand possibilities for energy demand and supply given hypotheses of future development.  The field of global change research frames uncertainties relevant for the future global energy system by drawing from a tradition of integrated assessment models (IAMs). Accordingly, these models and the theories they apply to structure their economic understanding of energy resources are naturally a key focus of this thesis. Therefore, the following chapters examine energy-economy modeling within the world of IAMs. 1.1  How to study energy-economy models? A world in the model that can be examined in stages  This study draws motivation and inspiration from work on economic and energy models conducted by Mary Morgan (2012) and Martin Greenberger (1983).  Through providing quantitative frameworks that mediate both social and physical processes, IAMs are distinct from the models developed by economists. 1 Yet, they are used in analogous ways to the autonomous epistemic genre of economic modelling investigated by Morgan (2012). Morgan's work eloquently describes how economists craft their models as research tools to study the context of the world, but also as a means of enquiring into a world within the model. IAMs find equivalent treatment in global environmental change research by making disciplinary intuitions explicit. Therefore, any study of IAMs can draw insights from Morgan's approach to economists and their use of models.  Further to the specific context of energy-economy modeling, Greenberger (1983) studies the research program that arose from combined efforts of engineers and economists during the 1970s.                                                      1  This thesis does not provide a detailed comparison of economic models and the economic world of IAMs. Though IAMs intend to capture world dynamics, so they include at least a stylized representation of some economic theory at a global scale. Some IAMs are certainly economic models, such as the Nordhaus DICE model (Nordhaus, 2016; 1992; 1977; Nordhaus and Sztorc, 2013). Many IAMs simply have exogenous GDP growth that calibrates energy demand based on economic theories of productivity and production functions, elasticities of substitution, learning-by-doing, and equilibrium solutions (partial or general).   2 In doing so, he identifies three distinct phases of formal modelling efforts that provide a useful structure for positioning ideas analyzed and developed in the following thesis:  I. Ideation: The initial ideation phase is primarily informal, where ideas are generated to structure and define the focus of resulting work. Ideation proceeds with exploratory efforts, shaping ideas, establishing a hierarchy for proposing questions and outlining terms of reference. For Greenberger, ideation is the critical phase of modelling: an art shaped by wisdom, oriented by an intention to be, "approximately right, rather than precisely wrong." It is important to note that disciplinary intuition plays a major role in the ideation phase.  II. Classification: The second phase of classification, establishes an accounting scheme, allowing formal models to describe the relationships at hand. This specification allows for precise numerical work, and an, "effective quantitative structure for exercising judgment and setting priorities."  III. Codification: The final phase of codification forms explicit rules for interactions between system elements. These rules record hypothesized relationships between variables and parameters, maintaining internal consistency between assumptions, data, and results. Greenberger sees codification as the stage which facilitates communication beyond the horizon of the modeler's cognition and intuition.  This introductory chapter proceeds by recounting the prevailing theories applied to structure uncertainty in studies of global environmental change. Then, we proceed by briefly understanding the history and context of the energy-economy models applied to this task.  1.2   Structuring deep uncertainty in future energy scenarios: two theories  Outlooks for the global energy future are influenced by two contemporary schools of thought on how to address deep uncertainty in scenarios of global environmental change: (i) an approach using alternative storylines (de Vries, 2006; Robinson, 1990; 1982; Schwartz, 1991; Vergragt and Quist, 2011), and (ii) a focus on fully probabilistic scenarios (Casman et al., 1999; Dowlatabadi, 2002; Dowlatabadi and G. Morgan, 1993; Sokolov et al., 2009; Webster et al., 2003).  1.2.1 Alternative scenario method: uncertainties structured by storyline narratives The alternative scenario method acknowledges that possible future developments can result from a wide range of unknown and unknowable outcomes (Grübler and Nakićenović, 2001). In this sense, assigning subjective distributions using probability concepts drawn from natural sciences is interpreted as a distortion on further analysis.  This view is based on an understanding that verification of any distribution would require repeated experiments and measured outcomes that are impossible in the context of long-range future social development. van Vuuren et al. (2008) point out that energy models may only address a limited scope of complex relationships, so adopting a narrative approach can provide consistency for the aspects that cannot be captured within the world of the model.   Davidson (1991) outlines an economic perspective for this theory building from the understanding Keynes (1937; 1921) applies to the situations where probability distributions are not helpful in understanding real world uncertainty. For Davidson there are three distinct environments for decision-making under uncertainty: (i) the objective probability environment where the past is  3 considered a statistically reliable guide to the future; (ii) the subjective probability environment in which prospects for future outcomes can be weighted at the moment of choice; and (iii) a true uncertainty environment that exists when a decision-maker believes unforeseeable changes will occur between choices made today and future outcomes, even if there is significant evidence to support the existence of objective frequencies and/or subjective probabilities.  True uncertainty in the Keynesian sense is often referred to as irreducible uncertainty (O'Donnell, 2013; 1989), or as Knightian uncertainty in reference to the distinction made by Knight (1921):  But Uncertainty must be taken in a sense radically distinct from the familiar notion of Risk, from which it has never been properly separated. The term "risk," as loosely used in everyday speech and in economic discussion, really covers two things which, functionally at least, in their causal relations to the phenomena of economic organization, are categorically different... The essential fact is that "risk" means in some cases a quantity susceptible of measurement, while at other times it is something distinctly not of this character; and there are far-reaching and crucial differences in the bearings of the phenomenon depending on which of the two is really present and operating. There are other ambiguities in the term "risk" as well, which will be pointed out; but this is the most important. It will appear that a measurable uncertainty, or "risk" proper, as we shall use the term, is so far different from an immeasurable one that it is not in effect an uncertainty at all. We shall accordingly restrict the term "uncertainty" to cases of the non-quantitative type. It is this "true" uncertainty, and not risk... which forms the basis of a valid theory of profit accounts for the divergence between actual and theoretical competition. (p.20)  Davidson (1996) further argues that economic modeling based on equilibrium microfoundations often proceeds by assuming or implying an ergodic reality - one that consists of essentially predetermined and immutable outcomes that result from hypothetical centers of gravity that stem from tendencies like ‘equilibrium’. Thus, he concludes that sampling from past data to understand true uncertainty in economic futures is equivalent to an implausible "sampling from the future". Davidson contrasts this view with a nonergodic interpretation of the future economy, where decision makers recognize that they are dealing in-part with an uncertain but creative reality that may not occupy positions consistent with past states.    Grubler and Nakicenovic (2001) employ a similar ‘creative reality’ concept by arguing that since each future outcome is path-dependent due to a number of conditional ‘what-if, then’ assumptions, narratives are the way to determine an internally consistent structure for uncertainty. Schweizer and Kriegler (2012) describe internal consistency as a, “scenario’s ability to represent dynamics consistent with current knowledge regarding plausible trends.” In this sense, formulating long-run scenarios by starting with narratives aims to avoid arbitrary and implausible combinations of social conditions, such as prospective societies where ‘high infant mortality’ leads to ‘low fertility rates’ (Grübler and Nakićenović, 2001) or 'high levels of wealth' occurs alongside 'low educational attainment' (Schweizer and Kriegler, 2012). Such an uncertainty range of outcomes produced by storylines allow stylized facts to structure the relationships and underlying logic among variables for further quantitative exploration (van Vuuren et al., 2008).   4 1.2.2 Fully probabilistic scenarios: projecting energy system futures from distributions  The probabilistic approach to long-run global change understands inputs as distributions of possibilities. This partitions resulting scenarios into probabilistic intervals which denote median expected results and extreme outliers (Capellán-Pérez et al., 2016; Dowlatabadi and G. Morgan, 1993; Keppo et al., 2007; Schneider and Mastrandrea, 2005; Sokolov et al., 2009; Webster et al., 2003).   Schneider (2002; 2001) argues that subjectivity is fundamentally inherent in all future projections, regardless of discipline or research program, and so it should be addressed explicitly and with quantitative rigor. Probabilistic studies of the global energy system are important to Schneider because relevant uncertainties feed into other downstream aspects of global environmental change research, driving an ever-larger cascading range of uncertain impacts that complicate risk assessments.  This perspective from Schneider (2002) is illustrated in Figure 1.1 for uncertainties relevant to studies of climate change, segmented by each stage of the research process. Though the public and decision makers may be most interested in stage VI, the range of possible impacts are contingent on uncertainties within each individual stage from I-V. Any uncertainties in the initial input conditions (Stage I) become amplified throughout this chain. Thus, a comprehensive mapping of uncertainties relevant for the initial stages is viewed as an important undertaking.   Without any attempt at subjective probability assessments, Schneider (2001) argues there is a 'probability vacuum' which leads to arbitrary interpretations. Even if such use is unintended, some outcomes are perceived as implicitly more likely than others. Lempert et al. (2006) extend this argument, proposing that understanding this sequence of interdependent uncertainties with equally likely storylines compromises decision-making. In the view of these authors, developing scenarios of future outcomes with quantitative probabilities will aid in determining priorities for future investments since resources are not unlimited.  Gillingham et al. (2015) understand subjective probabilities as 'degrees of belief' regarding future uncertainties. This follows from Ramsey (1926) which recognized that while it is impossible to obtain objective measurements of psychological variables there is still an important distinction to be made:  I do not see how we can sharply divide beliefs into those which have a position in the numerical scale and those which have not. But I think beliefs do differ in measurability in the following two ways. First, some beliefs can be measured more accurately than others; and, secondly, the measurement of beliefs is almost certainly an ambiguous process leading to a variable answer depending on how exactly the measurement is conducted. This interpretation of subjective probability is employed by Gillingham et al. (2015) as, "the odds informed scientists would take when wagering on the outcome of an uncertain event,” where, “the 'wager' is understood to bound the calculation of probability.”  5  Figure 1.1  Series of cascading uncertainties characteristic of global change research: focus on climate science research outputs – each initial stage (I-V) contains relevant uncertainties amplified by each latter stage; adapted from Schneider (2002)  Morgan and Henrion (1990) offer three arguments in support of explicitly representing uncertainty in studies relevant to global change because: (1) it aids in identifying the most or least important factors, the source of disagreement, and in anticipating the unexpected; (2) it clarifies the opinion of experts about how much they think they know and whether they disagree; (3) when the uncertainties of the past have been carefully articulated we can improve confidence in appropriate use of earlier work.  Though Morgan and Henrion provide these arguments in the context of policy analysis and decision making, their perspectives also apply to research in the physical and social sciences: why should any global change researcher focus significant research time on future outcomes beyond the range of possibilities? 1.2.3 Structuring uncertainty in global change research: juxtaposing these two theoretical approaches  An ongoing debate has addressed the meaning and adequacy of uncertainty in each of these contrasting theories, framing their ability to serve robustly guide broader scientific research on what to expect from the global energy future (Allen et al., 2001; Cooke, 2013; Dessai and Hulme, 2004; Grübler and Nakićenović, 2001; Grübler et al., 2006; Kandlikar et al., 2005; G. Morgan and Keith, 2008; Parson, 2008; Parson et al., 2007; Patt and Dessai, 2005; Pittock et al., 2001; Reilly, 2001; Trutnevyte et al., 2016; Webster et al., 2002).  van Vuuren et al. (2008) position these two traditions by relating them to disciplinary approaches in the ideation phase of study. These authors argue the fully probabilistic approach can represent a ‘positivist control systems engineering’ paradigm (where the system is well-known enough to make meaningful estimates of probabilities). The alternative storyline framework for future scenarios is  6 understood as within the ‘constructivist social science tradition’ (where alternative visions are created without assigning likelihood) (de Vries, 2006; Dreborg, 1996; Robinson, 2003; 1990; 1982).  Narratives and quantitative probabilities are seen complimentary efforts by van Vuuren et al. (2008) because they can address multiple dimensions of uncertainty in long-term studies which include: ontic uncertainty (natural randomness), epistemic uncertainty (incomplete knowledge), disagreement among experts (pluralism or disciplinary), and human reflexivity (unknowns in response to change).  Dessai and Hulme (2004) interpret this bifurcation as representing biophysical and social approaches to risk. They highlight that studies of biophysical impacts tend to be more comfortable with probabilistic studies and often use a long-term perspective to examine physical changes to ecosystems and geologic features. Social approaches are more focused on the shorter-term horizon and the human element of change, which is much harder to anticipate and thus tend to resist rigid probabilistic uncertainty analysis.   Walker et al. (2003a) synthesize both definitions, characterizing uncertainty as, "any departure from the unachievable ideal of completely deterministic knowledge." This is a broader definition of uncertainty than one focused on situations defined by inadequate information because it recognizes that phenomena under study may be non-deterministic.  Several studies have developed conditional probability estimates of population and other factors of global change scenarios by embedding probability distribution functions (PDFs) within narrative storylines (O'Neill, 2005; 2004; van Vuuren et al., 2008). However, the broader research community has tended to follow from the philosophical position outlined by Grübler et al. (2006) which argues there are simply too many uncertainties relevant to important questions for global change research without narratives. Thus, storylines are developed that provide a structure which captures, "how future societies will operate, how fast the population will grow, and how technological progress will change things." In this view, simply focusing a few 'best guess' scenarios ignores the important lessons of history regarding technological change, i.e. simply relying on subjective probability assessments of experts from the past would mean nuclear power is still expected to be too cheap to meter. Therefore, equally probable storylines have provided the most widely applied method for addressing uncertainty in global change research.   Two prominent examples of energy scenarios developed with this approach include the storyline based IPCC Special Report on Emission Scenarios (SRES) (Nakicenovic et al., 2000) and Shared Socioeconomic Pathways (van Vuuren et al., 2017). Much of the dialogue between the two schools of uncertainty took place in response to the SRES (Grübler and Nakićenović, 2001; Schneider, 2001).  Parson et al. (2007) address the controversy that arose from the six SRES storyline scenarios when the narrative developers argued that all storylines of future global change were ‘equally sound’. The rationale for treating each storyline scenario as equally likely stems from arguments that: (i) the multivariate possibilities which vaguely define boundaries of the outcome space leave no coherent way to distinguish whether a probability is attributed to the interval between scenarios or if it addresses the probabilities within a scenario, (ii) quantitative probabilities may induce reflexivity that influences the behaviors driving the scenarios, so that the storylines become ‘self-fulfilling’ or ‘self- 7 denying’, (iii) it is not the role of scenario developers to make judgements of likelihood but the users, especially when scenarios are used to inform high-stakes decisions and (iv) that the joint distributions of underlying drivers are simply too complex on the global scale to convey meaningful information in a way that isn’t ‘spurious’ or ‘arbitrary’. Parson et al. further note that while the SRES scenarios began their life as ‘equally likely’ storylines, they finished their use more like probabilistic scenarios, as many quantitative likelihood assessments followed in their wake.  The more recent Shared Socioeconomic Pathways (SSPs) eschew any focus on assessing the likelihood of storylines, drawing mainly from the arguments used to support the initial release of the SRES. A common concern is that focusing on a few ‘spuriously’ selected ‘most likely’ scenarios could lead to disregard for low-probability high-impact scenarios, which may be the most relevant for approaches to risk management (Riahi et al., 2017; Rozenberg et al., 2014). For studies of climate change policies, inclusion of low probability scenarios is considered essential (Rozenberg et al., 2014; Wagner and Weitzman, 2015). Both the SRES and SSPs structure uncertainty with narrative-based approaches, and expand these concepts using extensive quantitative modeling efforts. These narrative and quantitative methods play a major role in constructing long-run scenarios of the global energy system. However, questions of uncertainty in global change take on additional dimensions within the world of a quantitative model.  1.2.4 Uncertainty in global change codified by the world of the model Models abstract from the world to give ideas an explicit form, allowing further inquiry to proceed within a structure and style for reasoning (M. S. Morgan, 2012). A model can be conceptual (a line and box diagram) or a mathematical formulation, structured by equations and coded into a computer (Walker et al., 2003).  Modelling is said to become 'formalized' as it moves from vague ideas about the world toward more explicit and exact rule bound interactions between system components (M. S. Morgan, 2012). Moving from the ideation phase of a formal modeling effort, into stages of classification and codification, introduces additional dimensions of uncertainty.  Explicit to the context of global change modelling, van Vuuren et al. (2008) identify uncertainties in three forms: (a) conceptual theories, (b) structures, and (c) parameterizations. These three issues are analogous to each stage of the model development process identified by Greenberger (1983) for uncertainty in stages of ideation, classification and codification.  Walker et al. (2003) provide a parallel logic, defining a three-dimensional taxonomy for uncertainty in a model based on axes of:  • Location: Location refers to the position of uncertainty within a modeling exercise. This refers to uncertainty in model (i) boundaries, (ii) structural components, (iii) technical implementation, (iv) inputs (v) parameters and (vi) outcomes (prediction error).    • Level: This dimension of uncertainty captures the spectrum between knowledge and ignorance, i.e. the difference between myopia and perfect foresight. This is the dimension of uncertainty addressed by work that follows from Knight (1921) and Keynes (1921).   8 • Nature: The nature of uncertainty distinguishes between epistemic and natural variability. Walker et al. (2003) highlight that natural variability can result from the inherent randomness of nature, human behavior (micro-scale), society (macro scale), and technological change (new developments, breakthroughs or side-effects) – e.g. ‘ontic uncertainty’ as termed by van Vuuren et al.  Chapter 2 of this thesis focuses on a single key parameter used by IAMs to represent the global energy future, so it is useful to address parametric uncertainty with more detail in this section. Parameters define constant relationships within a model. They are contingent on the chosen context and scenario (Walker et al., 2003). Parameters can be exact (universal constants), fixed (well defined by previous investigations), a priori (difficult to identify and are chosen to be invariant based on experience), and calibrated (unknown values determined by comparison of model outcomes for historical data series). Uncertainty in a priori and calibrated parameters can be difficult to constrain, e.g know the plausible scope, such as upper or lower boundaries of their possible values.  Large ensembles of global change models are applied to address uncertainty in outcomes through multi-model comparisons (e.g. Clarke et al., 2014; O'Neill et al., 2016a). Studies on the macroeconomics of the long-run global energy system are commonly conducted through multi-model comparisons of IAMs (Kriegler et al., 2014a). These intercomparison exercises intend to provide uncertainty ranges for outcomes by running similar scenarios across a range of different models, producing an ensemble based on measuring model uncertainty. This is equivalent to asking the same question to a range of different oracles, with their responses forming the range of expected outcomes.  With the theories applied to structure uncertainty in global change research defined and structured, we can now focus on the specific context of IAMs.  1.3   Quantitative studies of global change: integrated assessment models Today's IAMs build from a discipline of global systems modelling efforts that started in 1960s, following from the World 3 model used by the 1972 Limits to Growth (LTG) study (de Vries, 2006; Forrester, 1971; 1969; Martens and Rotmans, 1999; Donella Meadows et al., 1972; 1982). The World 3 model faced criticisms that it did not sufficiently address regional dynamics, the process of resource discovery, behaviors in response to price (substitution, elasticities) or technological change (Donella H Meadows and Dennis Meadows, 2007; Nordhaus, 1973; Nørgaard et al., 2010; Vieille Blanchard, 2010). Further critiques focused on data calibration and whether the trajectory of human society presented in the LTG study was overly constrained by poor modelling choices (Cole, 1973).   This discourse inspired economists, engineers, physicists, geologists, mathematicians, psychologists and other disciplines to seek an integrated approach toward a new suite of global change models (Costanza et al., 2006; de Vries, 2006). The International Institute for Applied Systems Analysis (IIASA) served as a nexus for many these initial steps toward a discipline of integrated assessment modelling of the world system (Häfele, 1976; Häfele and Buerk, 1976; Häfele et al., 1974; Donella Meadows et al., 1982).    9 1.3.1 The IIASA Approach establishes foundations for the discipline of world energy modelling   IIASA was founded in late 1972 by the Soviet Union, the United States, and ten other countries to cultivate scientific research on global challenges across the divide of the Cold War. Against the backdrop of energy challenges during the 1970s, namely the 1973 OPEC oil embargo, world modelling of energy systems became an important research thrust for IIASA. The First IIASA Energy Program (1973-1979) aimed to study possible developments in global long-range energy systems through the year 2030.  The IIASA Approach to integrated global energy modelling combined a set of models to examine energy use and supply in seven world regions. Population and economic growth were specified exogenously for the MEDEE model of final energy demand, after which the MESSAGE supply model determined the optimal supply-cost combination of environmental, technological and resource constraints. The IMPACT model evaluated economic impacts of the MESSAGE supply strategies, and the MACRO model illustrated consequent investment and consumption. Each region was evaluated iteratively until an internally consistent set of results emerged to produce two scenarios of global energy use from 1975 to 2030: IIASA-High and IIASA-Low (Figure 1.2).  Results from the IIASA energy models led to the publication of the IIASA Energy in a Finite World (EFW) study (Häfele, 1981; 1980a; 1980b; Häfele et al., 1981). Work conducted at IIASA in preparation for EFW convened scientists from around the world, helping to establish important conventions in integrated modelling of long-range developments in world energy economics. These efforts motivated the next generation of energy studies conducted with integrated models of the global economy (Edmonds and Reilly, 1985; Nordhaus, 1979). 2   Figure 1.2  IIASA Energy in a Finite World – two global energy supply scenarios and historical – the MESSAGE energy model led to the (a) IIASA-low and (b) IIASA-high scenarios, compared to (c) historical developments in global primary energy from BP (2017); historical data 2016-2030 trend based on extrapolated decadal growth rate per source; oil (red), gas (blue), coal (yellow), nuclear (green), hydro (purple), solar (grey), other (pink)                                                         2  There is a parallel strand of national and regional energy-economy modelling with roots in the World 3 and IIASA approach (Greenberger, 1983; Donella Meadows et al., 1982), however this section is focused on global energy models and so these are left unaddressed. Chapter 5 of this thesis touches on some examples from the United States.   10 1.3.2 World energy models focus on climate change: integrated assessment models grow in scope and scale Analogous approaches to energy modelling formed the basis for a new class of IAMs focused on global climate change (Dowlatabadi, 1995; Edmonds and Reilly, 1983a; 1983b; Lashof and Tirpak, 1990; Mintzer, 1987; Nordhaus, 1993; 1992; Rotmans, 1990). Early methods of integrating climate change simply extended the energy models of the 1970s and 1980s with fuel-specific emission coefficients to produce scenarios of greenhouse gas (GHG) emissions from energy system projections (Häfele, 1980b).  IAMs study climate change by linking developments in macroeconomics, the global energy system, demographics, land-use and their influence on the climate. Integrating so many complex factors leads IAMs to include reduced form representations of each underlying system as a way to maintain numerical tractability (Casman et al., 1999; Parson and Fisher-Vanden, 1997; van Vuuren et al., 2009). IAMs inform assessments of economics and policy relevant for climate change research by connecting socioeconomic and energy system projections with simple climate models that emulate the physical system response of full climate models (Clarke et al., 2014; van Vuuren et al., 2009).  More detailed studies of physical climate processes use general circulation models (GCMs) and earth system models (ESMs). GCMs divide the atmosphere into vertical layers of grid-cells coupled with ocean layers, allowing for calculations of the mass-energy balance within this hierarchy of tiers (Flato et al., 2013; IPCC, 2013; Wilby and Wigley, 2016). ESMs add further comprehensive detail by providing explicit representations of atmospheric exchange with the global carbon cycle, including ocean and plant ecology, land use and biogeochemistry (Heavens et al., 2013). Studies conducted with GCMs and ESMs draw from IAMs for relevant detail on their input scenarios of future GHG emissions (Nakicenovic et al., 2000; O'Neill et al., 2016b; van Vuuren et al., 2011).  Throughout the 1990s a growing number of IAMs were developed to address the challenge of climate change (Dowlatabadi and G. Morgan, 1993; Edmonds et al., 1997; Hope et al., 1993; Janssen, 1998; Martens and Rotmans, 1999; Nordhaus 1993, Rotmans and Dowlatabadi, 1998; Schneider, 1997; Tol, 1995; 1997). Surveying this generation of IAMs Weyant et al. (1996) describe two types of IA model: (i) policy optimization models that produce projections of climate change in the context of optimal policy for emission control rates or carbon taxes based on goals that could include maximizing welfare or minimizing costs and (ii) policy evaluation models which assess environmental, economic and social consequences of specific policies, or inform strategies on how to achieve them.  Rotmans and van Asselt (1999) and Dowlatabadi and Rotmans (1998) classify IAMs along a spectrum of: (i) macroeconomic oriented models – simple parameterized decision-analytic formulations of complex problems, and (ii) biosphere-oriented models – process-oriented descriptions of complex problems with a responsive environment.  The macro and microeconomic world in many IAMs draw from a neo-classical ideation (Rotmans and van Asselt, 1999; Weintraub, 2005) which informs a partial or general equilibrium structure (classification) with conventional approaches to optimization for utility, capital accumulation, savings, investment, productivity and economic growth (Manne et al., 1995; Nordhaus, 1992). Major engineering-economic IAMs include the Global Change Assessment Model (GCAM) (Edmonds et  11 al., 1997), the MIT Integrated Global Systems Model (IGSM) (Prinn et al., 1999), the IIASA MESSAGE (Messner and Schrattenholzer, 2000) and the Asia-Pacific Integrated Modeling (AIM)  group (Fujimori et al., 2012). These employ neoclassical concepts in a more general way, and are less identifiable as pure economic models, representing hybrids of various degrees between the macroeconomic and biosphere oriented IAM approaches.  The diversity of IAMs available to the global environmental change research community has grown in recent years. Many of these are either direct or tangential descendants of the models developed during the 1990s. Twenty-nine IAMs took part in the scenario intercomparison exercises which provided data to the Intergovernmental Panel on Climate Change (IPCC) Working Group III (WGIII) database (IPCC WGIII, 2014). These included the Energy Modelling Forum (EMF) 27 study which included 18 energy-economy IAMs originating from the European Union (IMAGE,MESSAGE,POLES,REMIND,WITCH), the United States (GCAM,MERGE,FARM,Phoenix), Canada (EC-IAM,TIAM-WORLD), Japan (AIM, BET, DNE, GRAPE), India (GCAM-IIM) and the OECD (ENV-Linkages) (Kriegler et al., 2014b).  IAMs are developed with distinctly heterogeneous approaches and adapted for studies of various scopes and scale. Therefore, it is important to draw clear boundaries for analysis and to avoid lumping all models together. Accordingly, this thesis focuses on evaluating concepts used in the engineering-economy models traditionally applied to develop outlooks for energy macroeconomics in Energy Modelling Forum (EMF) studies and for Intergovernmental Panel on Climate Change (IPCC) assessments (Blanford et al., 2014; Clarke et al., 2009; Krey et al., 2014; Kriegler et al., 2014a).  1.3.3 Understanding the meaning of the world in the model: evaluating the performance of IAMs and large global models more generally   Today’s global energy-economy modeling efforts benefit from much greater overall technical and conceptual sophistication than the world models developed during the 1970s. However, processes of global change are complex and human knowledge is inherently limited. Even if we could fully capture and understand the world as it is today with perfect information, such a snapshot would be bounded by time and human comprehension. 3 Though separate modeling sciences have distinct approaches, the process of modelling leads to endemic challenges in assessing the meaning of the world in the model, and how its epistemological resonance can be interpreted.  Simon (1996) suggests many complex structures are redundant, so we can use this repetition to simplify our description of planetary phenomena. This process of simplification in the development of models sharpens their focus during the ideation phase of development. However, ideation necessarily leads to subjective assumptions regarding structure and interactions between system components.                                                       3  If we imagine a complete implementation of a big data fantasy, wide-scale diffusion of quantum computing with server farms would capture and process data while 8 billion distinct agents represent human behavior analyzed by an artificial intelligence. Such a research program would still be dominated by uncertainties that result from theoretical interpretation of data over long time-frames, and unknown environmental responses in an open system and the depth of the human psyche. Reducing these uncertainties would therefore proceed in the form of making the human world more like the model, rather than vice versa.     12 Macroeconomist Robert Lucas describes this process of ideation in Spencer and Macpherson (2014) as one of enabling focus on the inquiry at hand: The construction of theoretical models is our way to bring order to the way we think about the world, but the process necessarily involves ignoring some evidence or alternative theories – setting them aside. That can be hard to do – facts are facts – and sometimes my unconscious mind carries out the abstraction for me: I simply fail to see some of the data or some alternative theory. This failing can be costly and embarrassing… but I don’t think it has any effect on the advance of knowledge. Others will see the blind spot… keep what is good, and correct what is not.  In other words, modelling sciences proceed by requesting simplified abstract representations of the world, which require subjective evaluations to determine where phenomena are condensed or expounded.  IAMs intend to capture a far grander scope than macroeconomic modelling. And with this challenge, there arises the question of how to evaluate and interpret theories, concepts, parameters, variables and their technical implementation in IAMs given 'what gets left out' and 'what gets put in' (Risbey et al., 1996). 1.3.3.1 Framing archetypes of model evaluation Computer-based mathematical models have long been influenced by Newtonian concepts of mechanistic, deterministic, reductionist and equilibrium-based explanations of the world (Mirowski, 2002; 1989) After World War II, conventional economic concepts fused with this modeling paradigm during a period of stable growth, favorable demographics for a consumer economy, and technical engineering successes at ever larger scales (Erickson et al., 2013; Janssen, 2002; M. S. Morgan, 2012).  The archetype for evaluating models consistent with this philosophy draws from creating simulations of reality that cannot be rejected (Janssen, 2002). Therefore, a 'successful' model represents a physical (or human) process that leads to some form of accurate predictive power, validated by the ex post testing and observation of its subcomponents. For example, validation of an econometric time-series model can proceed by looking for similar autocorrelation functions and residuals that are random or 'close to zero' (Chatfield, 2005; Hamilton, 1994; Pindyck and Rubinfield, 1991), even if many parameters have no clear physical representation. In early econometrics, data could be rejected but not theory: if results were suspect, it was interpreted as a problem in the data because the theory could not be wrong (M. S. Morgan, 1990). However studies of global environmental change involve non-equilibrium states of ecosystems and biological processes, and so many traditional disciplinary model validation techniques do not apply to such a context (Janssen, 2002; Risbey et al., 1996). Models developed within this archetype must face a system structure that is constantly evolving and adapting to change. Accordingly, Oreskes (1998; 1994) argues that verification and validation of mathematical models of natural systems is impossible because, "natural systems are never closed and model results are always non-unique."  In the view of Oreskes, models of earth systems can be evaluated for relative performance but that validation cannot proceed in absolute terms. Her perspective is that models can corroborate  13 hypotheses, or elucidate discrepancies in other models, but that their most important contribution is heuristic - as a guide for areas of further study. This is comparable to a work of fiction that tells a story which is a combination of objective, subjective and imaginative experience. She sees models as useful when they challenge existing formulations, rather than as tools verify or validate.  1.3.3.2 Subjective evaluation and testing of hypotheses in global change modelling Reflecting on the first major decade of global change modelling, Meadows, Richardson and Bruckmann (1982) recount a survey of large-scale multi-decade modeling teams conducted before the Sixth IIASA Symposium on Global Modelling, conveying insights on how the issue of evaluation is approached.  Meadows responds about her own experience developing the LTG model stating that the, "purpose of a model is not to dictate the truth, but to put forward a hypothesis for discussion and for attempts at disproof." In her view, the purpose of modeling is to simplify the system until it can be understood. She sees validation as a subjective concept and that it is, "best to admit that validity is a purely subjective concept and that each modelling school has its own unique approach to establishing confidence in its models.” Morgan and Henrion (1990) draw attention to this subjective model validation concept in Mar (1974) which suggested that modelers consider a model as validated when, "all variables they feel are important are included and none of the relationships between variables are incorrect by the modeler's standards." This subjective interpretation can inform a "uniform wince" criterion, where a critical look at one part of a model with its application in mind, one should not wince any more than looking critically at any other part (Howard and Matheson, 1984).  The Latin American World Model (LAWM) team from Candido Mendes University in Brazil provided a detailed response to the IIASA Symposium Global Model survey (D Meadows et al., 1982), sharing their understanding of model evaluation from the perspective of one of the first major global modelling teams from the 1970,:  [A] global model is a structured discourse... this means that it should never be understood as a 'reflection', or a 'synthesis' of reality. Rigorously speaking, a global model does not necessarily discourse about reality.  The LAWM team argue that every formal model (codification) is a discourse about a theoretical model (ideation) and that a homology is simply, "assumed to exist between the theoretical model and reality." In other words, the theoretical model is thought to 'resonate' with reality. When the formal model is built, another homology is assumed to exist with the theoretical model. Therefore, work conducted in a formal model is interpreted by the LAWM group as a discourse about the world in the model, but that may not include the theoretical model or representations of reality.  They highlight this distinction because structural resonance is 'passed through' with a priori and tacit assumptions, so that variables in the model are interpreted as faithful representations of the real historical process by default. Therefore, the lack of epistemological rigor in communication and presentation leads to confused interpretations of mathematical simulations as forecasts of the future.  14 They conclude that the descriptive dimension of a model only allows us to study what would happen if certain hypotheses are valid.  Despite many improvements in today’s generation of IAMs, there are intractable challenges in producing multi-decade outlooks with large-scale models. Thus, it is helpful to reflect on perspectives from Ayres (1984) who candidly assessed the difficulty of evaluating and interpreting the meaning of long-range models of human society.  1.3.3.3 Evaluating large-scale societal models – insights from Ayres (1984)  Ayres (1984) argues that the epistemology of long-range general equilibrium and systems dynamics models draw from a Leibnitzian philosophy of ‘trust derived from formal structure’. This is distinctly different from short-range econometric model evaluations based on empirical content. Therefore, he focuses on the formal structure of large-scale long-range models, identifying four key problems:  1. Applicability of economic theories for characterizing long-range historical development: Long-run models are often oriented around concepts of GDP, aggregate production functions and productivity that extend recent expectations for stylized relationships. For example, technological change is exogenous in neoclassical theory, leading modelers to specify constant rates of productivity increase. Ayres argues such formulations are likely to be misleading, e.g. external constant productivity increases compound over decades and lose analytical meaning. Though this critique is most applicable to the macroeconomic modelling school of IAMs, many IAMs are still influenced by neoclassical and conventional economic concepts by using constant productivity formulations. Here, Ayres echoes the concern of the ‘humbug production function’ which fit a Cobb-Douglas production function to data points that spelled out the word ‘HUMBUG’ (Shaikh, 1974).  2. Limitations of the statistical methods used in time-series analysis: Ayres suggests that best-fit methods from econometric time-series analysis only identify ‘strong’ relationships. This automatically limits models to a short-term perspective because more important ‘weak’ relationships may only appear as ‘noise’, even though their cumulative effect may be important. His examples include links between: energy prices and long-run productivity, resource scarcity that induces technological innovation, and demographics.  3. Ubiquitous nonlinearity of global processes: Since linear models do not apply to long-range phenomena, Ayres emphasizes the importance of nonlinear models. However, nonlinear models are extremely sensitive to parametric assumptions, and can be adapted to show essentially any dynamic behavior. Thus, he suggests their form is less important than their parameters.  4. Preoccupations with determinism (the Selden Paradox): Ayres explains the Selden Paradox (drawn from Isaac Asimov’s Foundation novel) as the belief that ‘human behavior on the macro-scale is absolutely predictable, while on the other hand, this course of history can be altered by a single intervention.’ Thus, the Selden Paradox forms an irreducible uncertainty for energy modeling, exemplified by the 1970s: historical GDP rates were projected by modelers to the end of the century, and the ratio of GDP to oil use was considered roughly constant, however these outlooks were completely decimated by the 1973 oil embargo which were contingent on choices of individual actors in OPEC.   Given these issues, Ayres concludes long-run global change models are most useful at communicating the extreme sensitivity of outcomes to small changes in the choice of control  15 variables. He sees them as useful accounting meta-frameworks, where the final uncertainty range of possibilities are their most important contribution.  1.3.3.4 Defining an evaluation concept for today’s generation of IAMs Work conducted with IAMs face a persistent existential dilemma that results from combining models of physical and social processes in ways that create a challenging epistemic discourse. Ackerman et al. (2009) contest that IAMs should be recognized as fundamentally different from physical science models because they combine descriptive analysis of physical and engineering systems with value judgments stemming from conventions in economics. These economic ideas can be opaque because they are couched within the same technical framework, or are simply exogenous in the form of productivity and GDP (Millner and McDermott, 2016). Therefore, unwise use of IAMs could imbue normative economic inputs or values with undue physical meaning.  Though models are useful tools for formalizing and communicating one's disciplinary intuition, the longevity of large-scale global models introduce a unique problem. IIASA's MESSAGE model has been used in studies since the 1970s, and despite ongoing code revisions and updates, many aspects of the structure and its objective functions for cost minimization have been maintained (according to the available documentation). Modeling teams may develop and work on a single IAM for many years, allowing the model to gradually supplant one's intuition on energy-economy interactions. Pindyck (2015) takes a more combative tone, arguing IAMs may do little more than provide a technical means of legitimizing subjective opinion. He sees economic models as means to organize thinking in logically consistent way, improving understanding of relationships among variables. However, from his view, developing elaborate economic and physical models, "might let us think that we are approaching the climate problem more scientifically, but like the Wizard of Oz, we would just be drawing a curtain around our lack of understanding."  While Pindyck's critique certainly applies to explicitly 'macroeconomic' IAMs, DICE is a model of this class which takes great effort to maintain simplicity and provide transparency. For example, Nordhaus has long provided code for DICE in the readily accessible format of Excel alongside detailed manuals to enable relatively seamless testing of alternative assumptions, even if end-users may be inclined to interpret results with limited sophistication. Pindyck’s arguments become more concrete when models like DICE are applied as key policy inputs for national and regional regulation as in the case of the social cost of carbon (G. Morgan et al., 2017; National Academies of Sciences, Engineering, and Medicine, 2016) .  Although engineering and biosphere oriented models of the global energy system face their own distinct set of challenges, developers generally provide enough documentation to understand the operational meaning of parameters and variables (Bridgman, 1927). That is, every term in a model has an operational meaning that develops from a description of how it would be measured, given sufficient means and license (Cooke, 2013). Even if comprehensive measurement is impossible, or imbued with deep uncertainty, an internally consistent operational meaning provides a cohesive  16 means of evaluation for IAMs because it takes a model at its word, aiding in an ability to distinguish the physical from the metaphysical. 4  1.3.3.5 Why use large-scale IAMs? Evaluating an appropriate use with a perspective from Morgan and Henrion (1990)  The discussion developed in this introductory chapter has highlighted the importance of avoiding the “run once for the answer mode” and the need to understand broader philosophical implications of modeling that stem from the integration of physical concepts with those of human systems (G. Morgan and Henrion, 1990). This warning is strengthened when large complex models can have their results replicated by simple pen, paper and calculator techniques (Keepin, 1984; Keepin and Wynne, 1984). However, healthy skepticism does not have to fuel pure cynicism (e.g. Pindyck, 2013). As emphasized by the developers of early large-scale long-range models, today’s IAMs are analogous in their ability to provide a common basis for structuring and organizing strands of scientific knowledge in ways that allow for testing hypotheses of the future (Rotmans and van Asselt Marjolein, 2001; van Asselt Marjolein and Rotmans, 2002). The array of large-scale energy system models which have been built and maintained with considerable effort can be applied as a testing bed for ideas and a valuable tool for characterizing uncertainty as suggested throughout this section by Ayres, Oreskes and Meadows. An insight in this vein is provided by Morgan and Henrion (1990) who suggest that once large research models are built and trusted, simplified analytical models can be developed to analyze specific domains of interest. Therefore, it is possible to distill them into surface response functions and phase space concepts to construct arguments for boundary conditions, scaling and other forms of more focused inquiry. This can allow for iterative development between a well-suited specific model which feeds back into the other realms of the world within the larger model. Chapter 6 revisits this idea with an example from IAM energy system outlooks in the development of a more general presentation of uncertainty in global change studies.     However, Morgan and Henrion (1990) emphasize that without thorough and systematic modeling and analysis of the uncertainty of the problem at hand, “we can not be sure that the results of a model, especially a very large and complex one, mean anything at all.” Therefore, these authors emphasize the importance of attention to model verification, input uncertainty, and clear definition of the modeling project objectives. These are all relevant considerations for studies of technological change in the future energy system.                                                       4  It is important to note that in defining the concept of operational meaning Bridgman (1927) does not intend to claim that physical knowledge obtained via measurement is inherently more valuable than metaphysical knowledge: “If the concept is physical, as of length, the operations are actual physical operations, namely, those by which length is measured; or if the concept is mental, as of mathematical continuity, the operations are mental operations, namely those by which we determine whether a given aggregate of magnitudes is continuous. It is not intended to imply that there is a hard and fast division between physical and mental concepts, or that one kind of concept does not always contain an element of the other.”    17 1.3.4 Formal studies of energy system uncertainty with IAMs Uncertainties in future energy system developments are generally addressed with multi-model comparisons of deterministic scenarios, i.e. model runs of a single evaluation of a given state of the world. These comparison projects apply common scenarios to produce an ensemble uncertainty range for key energy system outcomes such as total primary energy supply (TPES) or carbon dioxide emissions (CO2) based on the span generated by a range of models (Blanford et al., 2014; Calvin et al., 2012; 2013; Chen et al., 2013; Clarke et al., 2009; De Cian et al., 2013; Edenhofer et al., 2010; Energy Modeling Forum, 1995; Krey et al., 2014; Kriegler et al., 2014a; 2015; Luderer et al., 2013; 2011; Riahi et al., 2015).  Formal uncertainty analysis of the long-run global energy system is generally conducted with techniques for uncertainty propagation in IAMs – where probability distributions of input parameters are subject to Monte Carlo sampling, with each parameter combination run through a deterministic model (Manne and Richels, 1994; Reilly et al., 1987; Scott et al., 1999). Today’s IAMs used in the research community largely preserve the deterministic structure of earlier energy-economy models, and recent studies have employed these to conduct increasingly sophisticated formal uncertainty analyses (Bistline and Weyant, 2013; Bruckner et al., 1999; Capellán-Pérez et al., 2016; Gillingham et al., 2015; Kann and Weyant, 2000; Lemoine and McJeon, 2013; McJeon et al., 2011; Nordhaus, 2016; Rozenberg et al., 2014; Sokolov et al., 2009; van Vuuren et al., 2008; Webster et al., 2008; 2012). We now focus on a few of these papers to highlight their methods and key findings.  Nordhaus (2016) applies the DICE IAM to address parametric uncertainty in future climate change by developing probability density functions (PDFs) for equilibrium climate sensitivity, productivity growth, economic damages, the carbon cycle and the rate of decarbonization. These five PDFs are discretized into quintiles, creating 55 combinations (3,125) of model parameters. This structure is applied to estimate a mean SCC of 36 $/tCO2 (ton CO2) for the year 2015, a value that increases the mean estimate from a fully deterministic analysis with DICE. Therefore, Nordhaus suggests that optimal policy does not suggest waiting for uncertainties to be resolved.  Nordhaus emphasizes that uncertainty about physical parameters is ‘level’ (uncertainties remain roughly constant over time) but that economic variables are a ‘growth’ uncertainty (which tends to increase over time). Therefore, he concludes that outlooks for the year 2100 contend with greater uncertainty from economic variables than physical.  Webster et al. (2008) conduct a Monte Carlo (MC) simulation of 100 probabilistic parameters with the MIT Emissions Prediction and Policy Analysis (EPPA) model – the ‘human’ side of the IGSM. Relative contributions to uncertainty in baseline emissions are estimated for each parameter, and energy supply emerges as the dominant factor. 5 The authors note a surprising, and ‘puzzling’ result regarding the most important parameter that emerges: the elasticity of coal supply with respect to price, which explains nearly 25% of the variance in baseline emissions. They highlight that since coal is considered abundant, supply side factors were not traditionally thought to be important. However, this formal uncertainty analysis highlights varied rates of autonomous energy efficiency improvement (AEEI), and interfuel substitution that lead to higher-carbon emitting economies induce                                                      5  Variance explained by energy supply sums the effects of all parameters related to: technology costs, penetration rates, fossil resource availability and supply price elasticities.  18 a ‘call on coal’ that becomes a major influence on baseline emissions. In this study 2050 coal use ranges from 220-480 exajoules (EJ) with a median estimate of 320 EJ. Annual global emissions in the year 2100 reach a median level of 22 gigatonnes carbon (GtC) with an upper bound of 36.9 GtC and lower bound of 14.5 GtC. While baseline emissions are dominated by uncertainties related to energy supply (especially from coal), this study finds that uncertainties in carbon price trajectories are dominated by demand-side factors.  Bistline and Weyant (2014) provide a sequential decision-making model of electricity generation capacity technologies in the United States using multi-stage stochastic programming techniques with the MARKAL IAM. Sequential decision-making frameworks consider policy decisions at multiple time-periods, allowing for explicit discussion of hedging strategies in the face of uncertainty. 6 To develop cost projections for a hedging scenario, Bistline and Weyant apply discrete random variables for technological parameters in the context of policy decisions made during the period 2000-2025 under uncertainty regarding the strength of a CO2 cap in 2025 (no emissions cap, moderate cap, tight cap). Their hedging strategy suggests installations of nuclear before 2025 provide a strong hedge against low CCS availability and stringent climate policy targets. 7 Though this stochastic programing approach provides simulations of policy decision-making with somewhat more realism regarding timing, the authors highlight that numerical tractability constrains the number of scenarios that can be considered.  van Vuuren (2008) conduct a ‘conditional probabilistic’ analysis with the IMAGE/TIMER energy model using the Special Report on Emission Scenario (SRES) storylines to inform sampled ranges of parameter values PDFs for key CO2 emission driving forces. For example, in the A2 storyline of regionalization and low technology diffusion, the median AEEI is 50% lower than in a B1 storyline of higher sustainability, i.e. the A2 storyline has lower energy efficiency. Around the median range of each storyline’s contingent PDF a default of +- 15% is sampled. This method effectively develops +- 40% ‘error bars’ for the main SRES marker scenarios by bounding potential 21st-century outcomes, subject to the interpretation of each narrative’s influence on the global energy system. These authors find an overlap range for 21st-century GHG emissions between 1,400 and 1,600 GtC across storylines, smaller than the range of fully probabilistic studies of 1,100 to 1,700 GtC.  Gillingham et al. (2015) apply six IAMs (DICE, FUND, GCAM, MERGE, IGSM and WITCH) to study the influence of parametric uncertainty in population, total factor productivity and equilibrium climate sensitivity on estimates of CO2 emissions, climate change and the social cost of carbon. The authors develop a ‘two-track’ MC method based on: (Stage 1) model runs are conducted over a set of grid points to produce surface response functions that emulate model runs, then (Stage 2) input parameter PDFs are developed and sampled using MC simulations for each parameter. The final surface response functions for each model indicate the degree that outcomes deviate from a set of parameters which are different from the model’s baseline values. Based on this study, the authors estimate the relative contributions of parametric uncertainty vs. structural uncertainty to the final MC-                                                     6  This multi-stage approach rather than policies imposed ‘all at once’ as with approaches to uncertainty that conduct MC simulation for parameters with IAMs. 7  The electricity generation technology portfolio in this study focuses on coal with CCS, unabated coal, natural gas, nuclear and wind. Regarding solar, the authors state, “Solar technologies are rarely deployed due to the high assumed investment costs in the database and lower conventional technology costs, which make the marginal abatement costs for solar higher than competing low-carbon technologies. The lowest-cost solar technology has investment costs that increase over time to $2.33/W by 2050.”   19 simulation results. Analysis of variance for each parameter indicates that parametric uncertainties overwhelm structural model uncertainties. The authors conclude this result is a ‘sobering’ indication that the technique of multi-model ensembles to develop uncertainty ranges is highly deficient, and underestimates overall uncertainty by a significant amount. 8 Since this thesis is focused on the long-run macroeconomics of energy resources, we now direct our focus toward IAM studies on this topic.    1.3.4.1 Studying hypotheses of energy resources in IAMs: alternate dimensions of uncertainty  Multi-decade studies of the global energy system must anticipate how our knowledge of oil, gas and coal resources will change over time. As covered earlier in this chapter, a major criticism leveled at the World 3 model used in the LTG study was that it considered a fixed resource base which did not evolve with technological change (Cole, 1973; Nordhaus, 1973), therefore any scenario with an outlook for 'running out of resources' was classified from the start.  Since then, studies with IAMs have adopted a more dynamic approach to energy resources by considering the full extent of their geologic availability and by developing hypotheses of technological change which could lead to their eventual economic production (Bauer et al., 2016; e.g Gregory and Rogner, 1998; Rogner et al., 2012; Rogner, 1997). However, deposits of energy (and other mineral) resources lie beneath the surface of the Earth, and are challenging to characterize in a way that provides accurate knowledge of their quality and extent (McKelvey, 1972). Therefore, studies sensitive to future energy resource production face key uncertainties along three dimensions:  1. How much is geologically available? (reserves v. resources vs. total abundance in the Earth's crust); 2. What determines economic production? (demand, production technology, technological change, fuel conversion processes, end-use potentials, substitution); and 3. Is this information reliable? (classification schemes, data quality, economic vs. physical accounting, incentives for knowledge production, exploration techniques, regulatory regimes, political factors).  Research focused only on the first dimension can lead to outlooks that are overly pessimistic or optimistic: oil, gas and coal accounted as 'reserves' are but a fraction of the total amount of geologic resources; yet the total amount of resources are unlikely to ever be producible because of dimensions (2) and (3). Projections of end-use energy demand based solely on technologies (electric cars, hydrogen fuel cells, internal combustion engine) may extend beyond feasibility if they are not coupled with considerations of dimensions (1) and (3). Energy resources are commonly characterized with degrees of uncertainty such as proved, probable, possible; or measured, indicated, inferred. 9 Yet the application of these terms is often inconsistent across national boundaries, data vintages and resources (3), leading to confused outlooks, unclear future                                                      8  However, model structure was found to lead to vastly different SCC estimates – more than 80% of the difference between SCC values were explained by the model uncertainty, compared to 0.052 (CO2 concentrations), 0.062 (temperature), economic output (0.016), radiative forcing (0.020) and population (0.109).   9  So far in this literature review, we have touched on the concept of 'subjective probability distributions', these categories are excellent examples of widely-used subjective probabilities.   20 possibilities and overconfidence in decision-making. Thus, any study on long-run energy supply and demand must be conducted within an explicit framework that accounts for the influence of these three dimensions.  1.3.4.2 Classifying energy resources: reserves, resources and occurrences A widely used system for classifying the economic and geologic factors inherent in assessments of energy resources draws from McKelvey (1972). McKelvey develops a horizontal axis for geologic factors and a vertical axis for degrees of economic recovery (Figure 1.3) based on general definitions (Rogner et al. 2012) of:   • Reserves (dark blue) are the quantities indicated by geologic and engineering information that are likely to be recovered with reasonable certainty in the future from known reservoirs under existing economic and operating conditions.  • Resources (lighter blue) are the detected quantities that cannot be profitably recovered with current technology, but may be recoverable in the future, and quantities that are geologically possible but undiscovered.  • Occurrences (light blue) are the total amount of a mineral contained in the Earth’s crust in some recognizable form.  Further muddling this picture, the lowest box of Figure 1.3 includes unconventional resources. The distinction between conventional and unconventional resources is often confusing, based on various combinations of geological characteristics and production technologies. Authors tend to apply preferred definitions which can change depending on industry conventions and developments in technology.  Conventional generally refers to oil and gas produced from reservoirs using traditional drilling, pumping and compression techniques. Unconventional oil and gas is present in different geologic formations, requiring advanced production methods. Unconventional oils include shale oil (hydraulic fracturing), oil sands (bitumen), oil shale (kerogen), or extra heavy oils – these deposits may involve in situ heating or treatment before they can be extracted (McGlade, 2012). Unconventional gas refers to shale gas, tight gas or coal bed methane (McGlade et al., 2013).   21  Figure 1.3   McKelvey (1972) system for resource classification – with additional categories as noted by Rogner et al. (2012) – the vertical axis indicates various degrees of economic recovery (right axis) while the horizontal axis denotes magnitudes of geologic certainty (lower axis)   1.3.4.3 Energy resources in IAMs: assessing plausible oil, gas and coal production outlooks in the world of the model  Multi-decade studies conducted with IAMs understand the potential for future energy resource production by looking at the full extent of this McKelvey diagram. An example of the fossil energy potentials used by IAMs in the EMF 27 study is included in Table 1.1 (IEA, 2016; McCollum et al., 2014; Rogner et al., 2012).  As noted in Table 1.1, during 2014 production of conventional oil was 170 EJ/year. With reserves of 4,900-7,600 EJ, the reserve-to-production (R-P) ratio for oil is 30-45 years. However, some of the oil produced today was formerly classified as 'resources' meaning that a long-term study must consider the additional 4,200-6,200 EJ in oil resources, establishing a resources+reserves-to-production (R+R/P) ratio of 50-80 years. 10  Many engineering-economic IAMs apply a dynamic resource concept to characterize the boundary between today's reserves and the resources that could be re-assessed as reserves in the future. IAMs translate this concept into supply curves that include reserves and resources, structured by their potential extraction cost and the influence of technological change on improving their economic                                                      10  Geologists commonly frame production potentials through accounting for a ultimately recoverable resource (URR) which is a theoretical approach to determining the amount that has been extracted already and the ultimate amount that can ever be extracted. There is a great volume of literature of suitability and applicability of URR which is outside the scope of this introductory chapter.   22 production (Bauer et al., 2016; Rogner, 1997). Chapters 2 and 3 of this thesis examines this concept in detail.  Table 1-1 - Oil, Gas and Coal: Production, Reserves and Resources as applied in EMF 27   Type Production* (2014) Reserves  (EJ) Reserves-to-Production Resources  (EJ) R+R-to-Production† Oil Conventional Crude oil + NGL 170 EJ 4,900 - 7,600 30 - 45 4,200 - 6,200 50 - 80 Unconventional 16 EJ 3,800 - 5,600 250 - 350 11,300 - 14,800 950 - 1,300 Natural Gas Conventional 110 EJ 5,000 - 7,100 45 - 65 7,200 - 8,900 110 - 150 Unconventional 27 EJ 20,100 - 67,100 755 - 2,520 40,200 - 121,900 2,260 - 7,100 Coal  170 EJ 17,300 - 21,000 100 - 125 291,000 - 435,000 1,850 - 2,740 * IEA World Energy Outlook 2016; All other data from McCollum et. al (2014); † R+R = Resources and Reserves Studies on uncertainties in these reserve and resource outlooks conducted with IAMs examine varied outlooks for oil, gas and coal by modifying the cost or quantity as presented in their supply-availability curves (Calvin et al., 2013; Capellán-Pérez et al., 2014; McCollum et al., 2014; 2016; McJeon et al., 2014; van Ruijven and van Vuuren, 2009). We can now revisit a few key insights from these studies.  McCollum et al. (2014) examine fossil energy resource production outlooks developed by 12 IAMs for the EMF 27 study. Though models employ varied cost-availability curves that lead to notable differences in reference scenarios, these collectively result in projections of future energy systems dominated by use of fossil fuels through year-2100 (58-80 ZJ - 3.1-4.3x historic use). Even though coal is massively deployed by several models (10-14x today's production rate by 2100), it is noted that the total resource base is extremely large and so these scenarios are considered plausible. Cumulative fossil fuel consumption across the range of model produced scenarios are well within the bounds of estimated reserves and resources, and so the authors interpret all of these outlooks as likely.  McJeon et al. (2014) analyze trajectories of GHG emissions under two scenarios: conventional gas (11,000 EJ of natural gas) and one of abundant gas (40,000 EJ) through the year 2050 in five IAMs (GCAM, MESSAGE, REMIND, WITCH and BAEGEM). The abundant gas scenario reduces the cost and increases the availability of natural gas, which then largely substitutes for coal, but also for lower carbon energy from nuclear and renewables. Natural gas releases about half as much carbon (56 kgCO2/GJ) per unit of energy than coal (96 kgCO2/GJ) so substitution of gas for coal is often thought to aid in decarbonizing the energy system. However, this multi-model study found that the abundant gas scenario had a limited impact on lowering CO2 emissions because gas substitutes for both high and lower carbon energy. Further, cheaper gas reduces the cost of energy, expanding total global energy use.   23 McCollum et al. (2016) conduct a similar study for oil, examining the sensitivity of CO2 emission trajectories to oil prices, creating a high price ($120/barrel) and a low price ($40-50/barrel) scenario for the MESSAGE IAM. These two scenarios are assessed in the context of biomass energy, biofuels, fossil synfuels, couplings between oil and natural gas prices, alternative vehicle technologies (electric, natural gas, hydrogen) and climate policy. The difference in CO2 emissions in the two scenarios is as much as 7 GtCO2 per year by 2050, or about 10% of the difference from the initial reference case. In the high oil price case, more carbon intensive alternatives are used, such as coal-to-liquid based synthetic fuels (synfuels), unconventional oils and biomass. In the low-oil price case, demand is higher and fewer high carbon-intensity alternatives are deployed. In summary, it is unclear from this study whether oil prices have a significant influence on trajectories of GHG emissions in MESSAGE, or if outlooks for CO2 emissions in MESSAGE are simply insensitive to energy prices.  van Ruijven and van Vuuren (2009) develop three price scenarios for the TIMER (IMAGE) energy model based on oil prices at low ($5/GJ), medium ($10/GJ) and high ($20/GJ) levels which are coupled to similar trajectories in natural gas markets. They introduce an exogenous price parameter for calculations of production costs to dynamics generally not presented in IAM supply curves. 11 Across all three of these scenarios, projections for primary energy supply in 2050 remains consistent, and coal provides the residual for lower or higher levels of oil or gas use. The high cost scenario in van Ruijven and van Vuuren (2009) leads to substitution away from oil over five decades, producing an outlook for transportation with biofuels and hydrogen. As most of the hydrogen is produced from electrolysis with coal-based electricity, a faster rate of coal production is projected. In the low-cost scenario, oil and natural gas is produced at a faster rate, and depletion raises the cost these fuels. Therefore, energy prices converge among all the scenarios. GHG emission projections for each of the scenarios also show little sensitivity, varying only as much as 15% after 2040.  Calvin et al. (2013) produce various combinations of high, medium and low scenarios for oil, gas and coal with the GCAM IAM. The high and medium scenarios assume a supply of 52 ZJ for oil, 56 ZJ for gas and 104 ZJ for coal. The low case assumes 30 ZJ of oil, 38 ZJ of gas and 39 ZJ of coal. These authors find that in scenarios with low oil availability, coal-to-liquids provides a significant portion of liquid fuels. In the high coal scenario, low cost coal substitutes for oil, leading to wide deployment of coal-to-liquids. The high fossil fuel scenario (high for all hydrocarbons) leads to the lowest deployment of coal-to-liquids by 2095 as abundant unconventional oil at 70 million barrels per day (mbd) reduces the 'call on coal' to only 23 mbd. Coal-to-liquids provides as much as 90 mbd of liquid fuels across all the scenarios. As with the other studies covered in this section, there is little change in CO2  emissions across the GCAM scenarios by 2050. The baseline, high fossil and high coal scenarios follow the exact same emission trajectories through mid-century, while the low fossil scenario leads to an outlook for emissions 15% lower than these projections.  Capellan Perez et al. (2016) develop a probabilistic assessment of fossil resource outlooks, applied to the GCAM IAM. These authors propose use of a vast array of ultimately recoverable resource (URR) estimates to account for uncertainties in the information conveyed by resource estimates from                                                      11  These include (1) strategic behaviours of oil producing countries, (2) potential underinvestment in production and refining capacity, (3) market uncertainties in response to political uncertainties, (4) rapid increases in demand, (5) limitations in the production rate of low-cost fields, (6) energy security policies in oil consuming countries, and (7) limitations in the rate at which new oil resources can be brought into production.  24 different agencies. URR intends to estimate the sum of past and future production in the context of physical factors related to reserve growth, resource discoveries, field sizes, and depletion rates. The authors synthesize a range of URR estimates from agencies (USGS, WEC, BGR, IEA, IIASA) and industry (ESSO, Shell) (Dale, 2012). From this database of URR estimates, they develop a distribution with uniform probability for each fuel source. The reserve and resource estimates applied in EMF 27 map to the top range of this database. These URR estimates are randomly assigned cost-availability curves shaped as either logistic, exponential or inverse. Then, 1,000 MC iterations are applied to produce a range of scenarios in GCAM, producing cumulative distribution functions (CDF) for outputs. This process achieves an interquartile estimate for 21st-century emissions of 970-1,470 GtC which is lower than the 1,370-1,700 GtC range produced by the EMF27 study. The authors find that a large portion of variance in cumulative CO2 emissions (0.73) stems from the URR of global coal.  This section has summarized recent research on uncertainties in energy cost and availability with IAMs. Issues noted by each study are: (i) the relative convergence of carbon emissions and energy prices across scenarios within models, and (ii) the significant role of coal in dominating uncertainty for future CO2 emission reference cases. Coal as a key uncertainty in GHG emission projections was also noted by Webster et al. (2012; 2008). Therefore, a key focus of this thesis is on outlooks for energy resource production in the world of IAMs, with special attention to coal.  1.4   Summary  This introductory chapter has reviewed key literature relevant to a systematic inquiry on energy resource uncertainties in global change research. Accordingly, we can summarize that:  • Studies of global environmental change must address long-time horizons beyond the boundaries of today's knowledge, where the global energy system is a key source of uncertainty.  • Integrated assessment models (IAMs) are applied in global change research to understand possibilities for future fossil energy production in this context.  • IAMs are models of the global energy-economy system where it is useful to distinguish between the 'model in the world' and a 'world in the model' created through ideation, classification and codification (Section 1.1). IAMs are used in ways that are analogous to the process by which economists use and create models: they give form to disciplinary intuitions and communicate these perceptions to others.   • Studies of global change structure uncertainty with two distinct approaches: one focused on scenarios constructed by narratives (Section 1.2.1) and the other with input PDFs (objective, subjective) (Section 1.2.2). Both concepts draw on IAMs for quantitative rigor (Section 1.2.3), which leads to specific forms and concepts of uncertainty relevant for modelling sciences (Section 1.2.4).   • IAMs provide quantitative meta-accounting frameworks for global change research (Section 1.3). Today's IAMs continue a tradition of global energy-economy models developed from collaborations between economists, engineers, environmental scientists and other disciplines (Section 1.3.1). Through the 1990s, this discipline of energy-economy modelling was increasingly adapted to study climate change (Section 1.3.2).  • Research conducted with IAMs faces challenges endemic to modelling sciences (Section 1.3.3), namely that by creating a simplified abstraction of the world, complex processes are condensed in a way that requires epistemological rigor for evaluation and communication of model results.   25 • Because IAMs define boundaries for analysis within open systems, it is not useful to understand them with an absolutist concept of 'validation', but to 'evaluate' them based on their ability to corroborate hypotheses and provide useful heuristic tools for studies of global change (Section 1.3.3.1).        • Early large-scale global model developers viewed the issue of model evaluation as one that is deeply subjective, where they were understood as a structured discourse about the world of the model, rather than 'reality' (Section 1.3.3.2).  • Confidence in these models draws from formal structure, rather than empirical content. However, this leads to challenges that stem from structural uncertainties inherent in the: (i) applicability of economic theory over the long-run, (ii) limitations of time-series calibration methods to detect important relationships, (iii) ubiquitous nonlinearity of global change, and (iv) deterministic nature of models outlooks that create paradoxes for interventions to redirect their trajectories (Section 1.3.3.3). • IAMs face a persistent existential dilemma that results from imbuing concepts from social science with explicit physical relevance. While this can lead to skepticism and cynicism regarding their usefulness, they are accompanied with operational definitions that provide a meaningful context for evaluations (Section 1.3.3.4).  • With careful evaluation and understanding of the world of the model, IAMs can be useful means of structuring and organizing strands of scientific knowledge in ways that test and communicate hypotheses of the future. However, large-scale IAMs are so complex, specific domains of interest are best explored with simple analytical models well suited for focused inquiry. This can take place with simple surface response functions that allow for bounding, scaling or calibration arguments that feed back into the larger model (Section 1.3.3.5).   • Studies of uncertainties in the future global energy system are conducted with IAMs through multi-model comparisons of deterministic scenarios, or through uncertainty propagation techniques that draw from input probability distributions (Section 1.3.4). Gillingham et al. (2015) conduct a multi-model study that combines both approaches, finding that parametric uncertainties overwhelm structural uncertainties. These authors indicate that IAMs can be redundant in their description of energy-economy systems, e.g. for some studies, running one large-scale IAM is equivalent to effectively running all models.  • Research on energy resources across the long-run must contend with uncertainty in dimensions that address supply, demand and information quality (Section 1.3.4.1). These must look beyond today's assessed reserves, to anticipate factors that could influence the recoverability of identified resources through conventional or unconventional techniques.   • IAMs approach such questions by varying their underlying oil, gas and coal input supply curves to analyze the influence of different assumptions on outlooks for primary energy production, key sectors of end-use energy and the influence on long-run CO2 emissions (Section 1.3.5.1). Two key insights can be distilled from these studies regarding trajectories of CO2  emissions produced by IAMs: (i) they are largely insensitive to significant changes in oil and gas costs or availability and (ii) are dominated by uncertainties regarding coal supply-side factors.  With this foundation established, the following thesis proceeds by evaluating the concepts applied to structure uncertainties relevant for fossil energy resource input supply curves in IAMs (Chapters 2-4). In Chapter 5, these ideas are synthesized to understand how they influence outlooks for scenarios used by the global environmental change research community. Chapter 6 proposes a method to address long-run outlooks for energy resources in a structure-neutral fashion, with the intent of contributing toward a more general discourse on uncertainty in studies of global change. Chapter 7 understands the socioeconomics  26 of these scenarios in the context of historical analogues. Chapter 8 concludes by recounting the key contributions of this work within a concise summary.  The work leading to this thesis started with a simple research question: How did historical concepts of  energy resources and technology in economic thought shape today’s models of how to achieve a future  low carbon transition? Initially, there was no intention that coal would have much to do with this line of  inquiry, since I wanted to focus on decarbonization rather than re-carbonization. Yet, since this fuel  source was considered the ultimate backstop technology when today’s climate models were in their early  stages of development, it was inevitable that coal would eventually play a leading role in the following  chapters.    27 Chapter 2:  Evaluating the learning-by-doing theory of long-run fossil energy economics   As addressed in Chapter 1, long-term studies of global change inherently extend beyond the scope of today’s knowledge, requiring a dynamic approach to future technological possibilities and the frontiers of currently available information. Understanding fossil energy resources in this context commonly begins with assessments of total geologic oil, gas and coal occurrences. After data limitations are acknowledged, hypotheses can be applied to anticipate future developments in the production technologies that could enable economic access to the full extent of these deposits.  Rogner (1997) addresses these questions of inherent long-run uncertainty with an innovative methodology by providing an approach to the ideation phase of energy modelling which grounds the total geologic presence of fossil energy resources in a theory of learning-by-doing. This seminal assessment combines diverse reports from governments and international agencies to formulate internally consistent cumulative availability curves for each hydrocarbon fuel (further referred to as the H-H-R Supply Curve).  Cumulative resource supply curves were originally proposed by Tilton and Skinner (1987) as a means of showing how the future availability of energy resources could vary according to price, and these build on the conceptual foundation established by Hotelling (1931). Aguilera (2014) discusses cumulative supply curves for global oil and gas, highlighting that they differ from the supply curves in microeconomics which are a static snapshot in time. Production cost-recoverable quantity (cost-quantity) curves, such as those covered in this chapter, intend to depict the general availability of energy resources over the long-run.  Rogner’s hypothesis is that continuous production will induce a learning curve effect independent of market prices, reducing the cost of accessing future resources. As more of the cost-quantity curve is produced, technological improvements accumulate as a compounding learning effect that leads to significant productivity gains in conventional and unconventional oil, gas and coal extraction technologies. In this theory, today’s reserves are understood as a "flow" continually replenished by the "stock" of total geologic occurrences, with a dynamic boundary characterized by learning that accumulates from increasing knowledge. The specific methodology applied by Rogner to structure resource cost-quantity curves is addressed with more detail in Chapter 5, as this chapter focuses primarily on the learning-by-doing element of their ideation.  Learning curves draw from a long history of studies on manufacturing, and in macroeconomics through endogenous modeling of technical change (Anzanello and Fogliatto, 2011; Arrow, 1962; Yelle, 1979). Since Wright (1936) observed productivity gains that resulted from repetitive tasks on airplane assembly lines, learning curves have provided effective and accurate mathematical accounts of performance improvements in continuing manufacturing processes when used in a relevant context (Yeh and Rubin, 2012).  Economic models of learning-by-doing anticipate productivity improvements that result from ongoing use of tools and techniques by workers, which lead to shortcuts and process optimizations that reduce the time, cost and materials involved in executing a specific task. Macroeconomic concepts of learning-by-doing have drawn from these strong microeconomic foundations, as in the work by  28 Arrow (1962) and Lucas (1988) which consider learning effects for an endogenous model of technical change in neoclassical growth theory. 12 However, the remainder of this chapter focuses on the conceptualization of learning-by-doing for cumulative resource supply curves, and not on its implementation in growth theory and other areas of economic thought.  Rogner (1997) adapts the concept of learning-by-doing to cumulative resource supply curves, creating an elegant foundation for energy models by condensing the complex factors shaping hydrocarbon economics into a numerically tractable solution. This learning-by-extracting (LBE) theory calculates future fossil energy supply potentials with a non-price induced learning rate (𝜌𝜌) – an expected outcome of ongoing production. The resulting cost-quantity curve for future supply is then simplified to focus on two dimensions: assumptions varying the rate of future learning and the total geologic stock of the resource.  IAMs regularly apply the LBE theory to develop fossil resource input supply curves (Clarke et al., 2014; IPCC, 2000; IPCC WGIII, 2014; Joint Global Change Research Institute, 2016; Luderer et al., 2013; Masui et al., 2011; McCollum et al., 2014; Riahi et al., 2011; van Vuuren, 2007). Detailed and publicly accessible global oil, gas and coal productivity data is often outside the budget of public and academic researchers, making the original H-H-R Supply Curve one of the very few available to the research community for more than a decade after its publication. Rogner et al. (2012) provide an updated outlook for energy resources within this conceptual framework of learning-by-doing.  Though each IAM applies unique variations of Rogner’s initial concept, the basic theoretical approach has remained consistent for decades. McCollum et al. (2014) review the details of learning driven fossil energy resource supply costs in a range of IAMs. Recent efforts by Bauer et al. (2016a) place this method within a framework that scales fossil availability curves based on scenario assumptions for trajectories of future socioeconomic development. However, this geologic learning model of productivity has yet to be empirically assessed for consistency with its operational definition by comparing against oil, gas and coal industry data (Bauer et al. 2016a). This gap in the literature leaves studies reliant on the LBE theory with an untested concept of technological change inherently sensitive to its key parameter: the chosen learning rate.  To illustrate the influence a selected learning rate can have on cost projections of future energy supply, Figure 2.1 reproduces the original H-H-R Supply Curve for oil with annual rates of learning driven productivity gains (𝜌𝜌) that vary from +1.0% to -1.0%. Each cost-quantity curve for oil intersects with its equivalent amount of carbon dioxide emissions (top x-axis) at a common backstop price for low emission oil alternatives of roughly $120/barrel of oil equivalent (boe). Calculating the next century of oil economics with this 2% total variation in learning rates results in an 1,800 GtCO2 span of uncertainty across the supply curve, e.g. roughly a half-century of current annual total CO2 emissions from all fossil fuels. The total price effect contributed by learning (𝚸𝚸) in this case estimates an aggregate productivity improvement for oil supply costs across the century as high as +170%, or as low as -60%. 13                                                       12  Lucas (1988) articulated a case for learning-by-doing in macroeconomics where each good has a different potential for learning induced productivity gains.  13  Note: Each curve starts at a different place on the y-axis as it reflects the learning effect throughout the entire century.  29 Throughout this chapter the annual learning-by-extracting effect is denoted as ρ: the rate of learning driven productivity gain, or dollar value of upstream cost reduction. The cumulative productivity gain induced by learning over the duration of the projection is 𝚸𝚸.   Figure 2.1  Influence of learning rates on calculations of future oil supply costs – Rogner (1997) oil supply curve (bottom x-axis in gigatons oil equivalent) modeled with varied assumptions on the rate of productivity gains from learning (-1.0% ≤ 𝜌𝜌 ≤ +1.0%). The equivalent amount of emissions from carbon dioxide (GtCO2) are shown on the top x-axis; the range of GtCO2 spanning the variation in each learning driven supply curve is shown on the top bar (green)  Nordhaus (2009) argues that learning curve models of future energy technologies produce estimates of long-run productivity with a consistent upward bias. He suggests this is “dangerous” because costs for any energy supply strategy calculated with such a technique is highly sensitive to a chosen learning rate that is difficult to evaluate (and possibly indistinguishable from model or data artifacts, or normative preferences).  The following chapter considers LBE assessments of fossil energy resources as a specific illustration of Nordhaus’ argument. This is highlighted by extending Nordhaus' theoretical formulation to account for the specific physical and geological factors related to oil, gas and coal recovery. Production of fossil energy resources occur under conditions distinct from other manufacturing processes, adding further complexity to the issue Nordhaus raises for robust determination of an appropriate learning rate.  This chapter evaluates the suitability of the Rogner (1997) hypothesis for long-run fossil resource economics in the context of contemporary data. As proposed in Chapter 1, this resource concept in IAMs can be evaluated based on the operational definition described by Rogner - namely that, "hydrocarbon resource exploration, development and production is subject to a compounded productivity gain of 1% per year. This 1% productivity growth rate approximates the average long-term historically observed rates in the hydrocarbon upstream sectors." (Rogner 1997, p.251) Though  30 IAM scenarios can apply different learning rates, this chapter evaluates the essential ideation of a long-run stable learning rate independent of price.   To conduct this evaluation, Section 2.1 revisits productivity trends in upstream oil and gas since the LBE model was developed in the mid-1990s, providing a first step in linking the theory proposed by Rogner (1997) to empirical evaluation. Section 2.2 extends the theoretical case of Nordhaus (2009) to the specific context of energy resources. With these issues articulated, Section 2.3 understands the conceptual influence of various empirically consistent learning rates on outlooks for oil production, economics and backstop resources with a simple conceptual model, and with the GCAM IAM. Section 2.4 concludes this chapter with a short summary, which establishes a background for evaluating the dynamic reserve-resource concept for coal (Chapter 3).  2.1   Empirical trends in upstream oil and gas (1978-2008): production costs, market dynamics and reserves Rogner (1997) develops a cumulative fossil resource availability curve to understand the economics of oil, gas and coal through the 21st-century. We have now experienced more than 15% of this period, allowing us to revisit the basic tenets of the LBE theory for empirical evidence of its core assumptions regarding: (i) autonomous compounding upstream productivity driven by learning, (ii) long-term stable upstream costs independent of market price effects, and (iii) the relevance of a reserve-to-production (R-P) equilibrium range for framing the economics of future production.  2.1.1 Data on oil and gas upstream costs: evidence of compounding productivity driven by learning?  The US Energy Information Administration (EIA) conducted a regular survey of major US energy companies with its Financial Reporting System (FRS) through 2011. Subsequent FRS reports analyzed data on the financial performance of domestic and worldwide operations for companies that included ExxonMobil, Shell, ConocoPhillips, Chevron and BP. The latest EIA FRS publication provides an internally consistent time-series from 1977 through 2009 that allows for examination of aggregate industry productivity data (Bawks et al., 2011). Figure 2.2(a-b) displays these reported FRS company cost trends for two key elements of oil and gas production that map to the operational definition of Rogner (1997): total upstream expenditures per barrel of oil equivalent (boe) for oil and gas (Figure 2.2a) and production costs less royalties (Figure 2.2b).  Figure 2.2a overlays the upstream cost trends for each decade of available FRS data, calculated by compound annual growth rate (CAGR) with a 3-year moving average (top axis). Decadal trends in upstream costs indicated by these FRS data are: +0.9% (1978-1988), +0.5% (1988-1998) and +9.6% (1998-2008). Assuming a stable declining trend for total upstream costs would be inconsistent with the FRS data since the calculated productivity rate is negative in each ten-year period (𝜌𝜌 < 0). 14                                                       14   The resolution of trend analysis is of importance to note: while the 1978-1988 trend shows a slight increase in costs, a smoothed compound annual growth rate would miscalculate the costs in nearly every year during the decade, missing the extreme cost increase from 1979-1983 and decline from 1984-1987. This reflects price volatility in the market for a global commodity. Alternative supply strategies that compete with oil through demand for manufactured products (such as wind turbines or solar panels) may have price trends that more directly relate to the learning-by-doing model for manufacturing.     31 Across the full three decades in the FRS data, total upstream costs increased at a rate of +3.6% per year. The right-axis (dark purple) of Figure 2.2a overlays the Brent market price in constant dollars ($2009). Exploration costs show the highest stability, fluctuating between 10% and 30% of Brent crude. Production costs dominate throughout the early portion of the time-series (1978-1996) until development expenditures become the highest proportion of upstream spend from 1997 onward. Notably, the three-year moving average of upstream FRS expenditures exceed Brent market price for much of the period during 1997-2002 - signaling market prices that reached unsustainable levels for the industry long-term. Development costs increasingly dominated upstream costs from the late 1990s, indicating a growth trend in industry capital expenditures that contributed to the supply-side conditions for the following decade’s oil bull market. 15  Industry trends for production, development and exploration costs reported in the FRS (Figure 2.2a) align with the initial formulation of the LBE model in the period leading up to its original publication. From 1988-1996 total upstream costs per boe of oil and gas output experienced an average -1.1% annual cost decline (𝜌𝜌 ≈ +1%). Though these productivity gains did not translate to subsequent decades, this portion of the time-series shows gradual improvements in oil and gas production costs as a learning effect would expect for a homogenous product. Cost ranges reported in the H-H-R Supply Curve show correspondence with industry metrics reported as direct lifting costs. 16 Lifting costs account for the expenditures required to extract developed reserves after they are found and acquired. EIA FRS data on direct lifting costs provided in Figure 2.2b indicate the technical cost of extracting oil and gas rose +0.7% p.a. from 1980 to 2009.  The three decadal cost trends range from negligible (1980-1990), to a sharp compound annual decline of -5.5% (1990-2000) and a rapid increase (+9.0%) in the first decade of the 21st-century. From 1980-1992 the trend indicates an annual cost decline from improving productivity at a rate of -1.0% per year. This sub-period appears to show the effect of learning from continuous production with conventional technologies in well-characterized geographic regions. This shaped the initial formulation of the LBE concept which anticipated these productivity gains would continue for all geologic oil and gas resources.                                                      15  Note: The FRS data report aggregate oil and gas production, so these values are not directly indicative of actual producer marginal cost, or useful for calculating profit margin. While providing an internally consistent data set for upstream costs and production, the aggregation of oil and gas data makes disaggregation dependent on a series of complex assumptions.  16  Though Rogner (1997) argues full upstream costs from exploration, development and production are captured in this model because of evidence suggested by development in the United States and North Sea production, the supply curves produced by the LBE approach more closely correspond to the direct lifting costs associated with production (e.g. operational expenditures). This case is argued throughout the following Section 2.1.2.   32   Figure 2.2a   Upstream productivity trends in oil & gas (1978-2008) – EIA FRS reporting on worldwide expenditures for exploration, development, and production of oil and gas output in barrel of oil equivalent (boe) with a 3-year moving average (MA); (right-axis) annual Brent crude benchmark price (purple line) in constant dollars ($2009); (top-axis) the decadal compound annual growth rate (CAGR) trend for upstream costs  Figure 2.2b  Trends in direct lifting costs per barrel of oil equivalent for EIA FRS Companies (1980-2009) – domestic trends (blue) and rest of world (red); (right axis) proportion of 2009 Brent benchmark price; (top axis) CAGR trends for each decade   33 While the EIA is only one source for industry productivity indicators, these data on upstream cost trends mirror the general features of other academic studies (Fantazzini et al., 2011; J. V. Mitchell and B. Mitchell, 2014) financial institution publications (Citi Research, 2013; Deloitte, 2015; Goldman Sachs, 2014; 2013; JP Morgan Asset Management, 2015; Lewis, 2014) and reports from oil industry consulting agencies (e.g. Kopits, 2014; Rystad Energy, 2015). This chapter focuses on the EIA FRS data because it is the highest quality dataset we could find in the public domain and available to readers for additional scrutiny. Further efforts can harmonize these data with upstream trends from the most recent decade. Admittedly, while including worldwide measures for Canada, Europe, the Former Soviet Union, Africa, the Middle East and other parts of the world, these data are biased toward US operations. Therefore, an immediate question arises about the application of these upstream trends to studies of global oil and gas, where OPEC producers play a major role. On that point, Watkins (2006) highlights that the deregulation of US oil prices in 1981 plugged the domestic market into the world, allowing information from the US to provide a window into reserve prices and costs in all regions open to new investment. As non-US companies develop and explore for oil in the US with operations around the world, the EIA FRS data series can be considered to generally represent the ‘shape’ of costs in many parts of the world. Global price trends have mirrored these upstream costs, suggesting they are generally representative of industry marginal cost and performance trends. 17  The LBE theory expected that compounding gains in performance would lead to ongoing upstream cost declines from accumulated learning. Yet, total upstream costs indicate an extended period of aggregate performance declines for total global oil supply. Despite specific performance increases in some regions and rapid diffusion of innovations in new upstream technologies (e.g. especially horizontal drilling post-2005), sustained long-run productivity trends from 1978-2009 break from those anticipated by the LBE theory. This discontinuity indicates a model of autonomous non-price induced learning for conventional oil and gas supply technologies does not capture relevant characteristics of the frontier between production technologies of the past and those of the future. A continuous learning effect applied to a heterogeneous resource base will thus face essential constraints in modeling the productivity of new technologies needed to access different types of resources in varied geologic formations. Cost-quantity curves can therefore present an inaccurate picture when combining multiple production technologies.  The analysis in this section suggests that specific manufacturing processes for future oil and gas production must be considered in models of long-run technological change to resolve contradictions between empirical trends and theoretical expectations for contributions from learning. The importance of introducing higher resolution modeling for extraction technologies is further illustrated by the context of capital expenditures.  2.1.2 The influence of market prices on productivity measures: distinct patterns for operational and capital expenditures The LBE theory expects that a learning effect independent of market price is a suitable explanation for productivity improvements in upstream energy resource extraction costs. However upstream                                                      17  Operations in the US are less subject to political instability than many regions, however, they may be more expensive due to concerns about litigation and social license.  34 costs are also contingent on a range of non-technical factors: taxes, royalties, land valuations, political intentions and business cycles.  Osmundsen and Roll (2016) explore evidence of industry cycles on upstream expenditures and provide evidence that bullish periods increase costs per unit of output, reducing measured productivity. In periods of rapid expansion, oil rigs and other oilfield service equipment experience a faster hike in wages and rig prices which reduce measured productivity, due to pressure from higher rates of capacity utilization. Conversely, in a market slump, equipment utilization rates decline, rig rates fall, and upstream productivity measures increase. Rogner (1997) equates long-term price in LBE supply curves to marginal costs (P = MC) determined by technology that improves with learning to formulate cost projections dominated by supply-side factors. The FRS upstream costs we analyze mirror patterns in market price, but are these fluctuations in productivity more clearly shaped by demand or ‘supply driven’ gains from learning? If learning-by-doing dominates upstream costs, an autonomous stable productivity trend is an appropriate model, since the costs of investing in supply expansion are largely independent of demand. However, Osmundsen and Roll point to one important way that demand-led prices shape marginal cost profiles, suggesting upstream productivity modeled independent of market conditions may not be applicable. Figure 2.3(a-b) examines the relationship between prices and productivity in the EIA FRS data. Production costs are summarized as operational expenditures (opex) and development plus exploration costs as capital expenditures (capex) (Aguilera, 2014). Figure 2.3a plots the year-to-year change in Brent price (top) alongside measured productivity gains for opex (middle) and capex (lower) from the FRS data in Section 2.1.1. Price declines visibly precede productivity gains through the early 1980s, suggesting much of the ‘learning effect’ measurable over this period resulted from industry consolidation. Parallel productivity gains in this series for opex are far less volatile than capex, and consistent with what a learning model would expect: 1991-1999 shows an eight-year stable improvement in operational productivity (𝜌𝜌 ≈ +3%). Because long-term models smooth trends to maintain numerical tractability, the histograms (right-side) superimpose empirically consistent normal distributions with mean values over the entire time-series for opex of 𝜌𝜌 = -4% (dashed yellow line) and 𝜌𝜌 = -5% for capex (dashed orange line). To test for the relevance of price effects, Figure 2.3b shows the influence of market fluctuations with year-to-year marginal changes in productivity measured per dollar of market price (𝑑𝑑𝑑𝑑𝑑𝑑$). Once again, the theoretical framework of LBE corresponds to opex trends: production expenditures experience little sensitivity to market price throughout the time-series, as Rogner (1997) originally assumes for total hydrocarbon energy supply.  A simple time-series average for opex indicates the productivity of operational expenditures fell by 0.08% for every dollar increase in market price (𝑑𝑑𝑑𝑑𝑑𝑑$ = -0.08), but an equilibrium value is close to zero. This supports further confidence in a non-price induced productivity model for opex. However, this assumption does not extend to capital expenditures where marginal productivity rates fluctuate significantly from 1979-2008.  35 The overall relationship between price and capex in this time-series is broadly negative (𝑑𝑑𝑑𝑑𝑑𝑑$ <  0), suggesting the industry tends to commit capital investments when market prices increase. The cyclical nature of this trend indicates that the industry adjusts expenditures based on what market outlooks allow over any multi-year period. Large positive values for capex in 1990, 1997 and 2003 may indicate points where the industry was temporarily starved for capital from underinvestment over the preceding period, and they are playing catch-up. Significant increases in amplitude during the latter half of the series correspond with the scale-up of capital investments needed to extend production into areas that required deepwater drilling and hydraulic fracturing, alongside boom times for the industry in the early 21st-century.  Since many oil and gas companies employ significant teams for forecasting and strategy, decisions to commit development costs are undoubtedly contingent on scenarios for market outlooks. This analysis supports the relevance of simulating market conditions in-step with projections for upstream productivity over the long-run. It seems difficult to harmonize an outlook for optimal investments that result in supply-led marginal costs determined by a 1% p.a. learning improvement with an industry that undertakes marginal capital investments under an expectation of higher market prices.  As visible in the FRS data from 1998-2008, development expenditures continue to accelerate in line with market prices (Figures 2.2a and 2.3a), indicating that the projects expanding marginal supply from the expensive end of the cost curve receive a green light under outlooks for continually increasing prices. If a 1% p.a. total upstream productivity improvement had occurred from 1988-2008, total upstream costs would have fallen from $24.50 per barrel to $20 per barrel by 2008, and expenditures on capex would have declined from $13 to $11 per barrel. Such a projection would have underestimated total upstream costs over these two decades by an average of 60% per year and capex by 100% per year. In this case, the LBE theory would have anticipated an equilibrium Brent market price of $26/bbl through the period from 2000-2008 over which Brent market prices averaged $60.18 Overall, it is unlikely that year-to-year average productivity measures for capital would maintain such distinct volatility across the industry. We therefore interpret these fluctuations of measured annual productivity in capex as indicating the dominance of essential business cycle elements over a measurable level of pure endogenous learning in this time-series. These are the factors originally discussed by Schumpeter (1934; 1939): during an upturn, wages increase and labor productivity decreases, during downturns the opposite occurs, as companies throttle expenditures for production capacity based on market outlooks. 19  FRS data illustrate important and relevant macro-scale aspects of the trends explored at the micro-scale by Osmundsen et al. (2016): short and medium-run constraints on production equipment during booms drive up costs because limited supplies of oilfield capital and labor may command higher prices. Accounting for such demand-led marginal costs in a long-run supply model is                                                      18   This projection of market prices maintains average mark-up per barrel in the FRS series of 30%.  19  These measures are further complicated because of the long-run outlooks required to develop new fields, i.e. market price outlooks for development expenditures must look beyond 3-year moving averages. However, this comparison is developed with 3-year MA market prices to make a one-to-one comparison with the original FRS data. In many regards, a year-to-year measure of capex productivity is limited but this is provided to match the annual point estimates in the LBE model.  36 necessary: socioeconomic conditions of many long-run policy models are predicated on a ‘long-boom’ of equilibrium growth in economic output (Clarke et al., 2014).  Total upstream costs per unit of production decline in a market bust, but resulting productivity measures are dominated by the expenditure reductions driven by responses to market conditions, and not the influence of learning. Capex productivity improvements in these data under such an economic environment seem to primarily reflect curtailed expansion of production to new areas and consolidation. Even though market pressures drive innovation, aggregate industry productivity data requires a careful analysis that accounts for explicit technological improvements alongside potential bear or bull market conditions - an insight particularly relevant for oil, gas and coal production data collected during the commodity bear market that started in 2014.  While short multi-year downturns merely constrain future output growth, extended periods of low capital investment will eventually lead to maturing production and well depletion, a 9% p.a. decline that sustained investment tends to reduce by 3% in aggregate. 20 Measured productivity gains due to a period of oversupply and falling oil prices do not inherently translate to increased long-run output potential because the production of oil resources is inherently different from manufacturing of a homogenous product: the production profile of specific wells and fields declines over time.  This analysis of FRS data indicates: (i) the LBE model accurately captures non-price induced secular trends for spending on operations and that (ii) the performance of energy sector capex is poorly represented in a homogenous formulation of marginal costs driven by the accumulation of learning.  Accordingly, some element of observable market price effects must inform a model of long-term industry productivity trends to overcome the bias introduced by aggregating operational and capital investment dynamics in the LBE theoretical approach to upstream cost. Plausible simulations of long-run oil and gas supply costs requires an explicit representation of the industry decision context for capital expenditures.                                                      20  The IEA World Energy Outlook 2016 highlights that depletion rates for global mature fields are around 9%, but sustained investment reduces decline of producing fields to 6% (IEA, 2016). Fustler et al. (2016) review the academic literature on decline rates, estimating a 6.2% p.a. rate post-peak.   37  Figure 2.3a Market price and upstream productivity trends multi-plot (1978-2008) – time-series (left column) and histogram with superimposed normal distribution (right column) for 3-yr MA change in Brent market price (purple line - upper), measured productivity improvement for operational expenditures (yellow line - middle), and capital expenditures (orange line – lower); the original Rogner (1997) estimation of 1% annual productivity gain is overlaid on the time-series (dark blue dotted line - upper) along with the average productivity improvement for opex (dashed yellow line - middle) and capex (dashed orange line – lower); each plot includes gray lines for the other two data series out of focus  38  Figure 2.3b  Marginal upstream productivity rate per dollar of change in market price (1978-2008) – (𝑑𝑑𝑑𝑑𝑑𝑑$) for operational expenditures (yellow line) and capital expenditures (orange line)  2.1.3 How relevant is an equilibrium reserve-to-production range for calibrating future upstream cost profiles?   Reserves-to-production (R-P) ratios for oil and gas have maintained a relatively consistent range over the 20th century (Adelman and Watkins, 2008; BP, 2017; Watkins, 2006; Wellmer, 2008). The LBE theory draws from this data to expect that R-P values maintain an equilibrium range in the future, framing recoverable reserves as a continual flow from the total stock of geologic resources. In this sense, resources are continually reclassified as reserves with production at costs subject to productivity improvements driven by learning. The equilibrium R-P intends to represent the behavioral dynamic of producers who otherwise have little incentive to invest in knowledge at lower production rates. However, the last few decades of data challenge the relevance of this assumption for projections of upstream costs drawn from cumulative resource availability curves.  Even as R-P ratios for oil and gas can remain relatively stable, the expenditures necessary to develop reserves into production have varied. A growing reserve base doesn’t inherently ensure that oil is getting cheaper to produce, and can often mean the opposite. The costs of converting proven reserves into a producing well are accounted as development expenditures. The LBE theory conceptualizes a dynamic resource boundary, and development expenditures represent the costs of ‘moving’ this boundary which differentiates the total geologic stock of an energy resource from its reserves, and their eventual production. Figure 2.4(a-c) plots several relationships between development costs, reserves and production for oil and gas.  39 Figure 2.4a depicts the proportion of exploration, production and development costs in the EIA FRS. Development as a fraction of upstream costs remained relatively stable from 1978-1991 but grew steadily from 1992-2005. Data from EIA FRS companies indicate that total expenditures on development grew from $7.50/boe in the mid-1990s to $36.50/boe in the mid-2000s. Over this period, development costs grew 4.8%/year from 1978-2008, outpacing growth in operating costs by 56%.  As an aspect of total industry marginal cost, development costs will be reflected in market price. Over the last few decades, development costs have mirrored trends in market prices much closer than trends in exploration or production costs (Section 2.1.2). Based on this observation, we can suggest that the LBE supply curves applied thus far in the literature reflect only one-third of the marginal cost of oil and gas production by focusing on technical operating costs of producing wells. This has neglected the cost and performance dynamics of the boundary between resources, reserves and production – especially as significant unconventional resources were being developed. While the equilibrium R-P concept is useful for capturing basic features of producer exploration behavior, the development costs required to realize an equilibrium reserve base are highly sensitive to aspects of technical difficulty introduced by geology or geography at the boundary between today’s reserves and those of the future. Adelman (1995) characterizes industry upstream behavior with a warehouse metaphor, where reserves are the dynamic inventory. This warehouse inventory is replenished from the resource base, depleted through production, and reserves are established by development expenditures. Adelman highlights that production capacity is likely to increase if development costs are below the equilibrium market price, but intensive periods of development raise the marginal cost per barrel of output, continually testing the equilibrium value. This interplay between supply and demand converts the marginal warehouse inventory of reserves into production as fast as the equilibrium market can rise. The dynamic described by Adelman implies a relationship between reserve development expenditures and demand that drives cycles of marginal cost and market price which effectively fluctuate around the base of proved reserves. In this case, an autonomous projection of an equilibrium R-P ratio provides little information on the availability of long-run supply if applied independently of development cost trends. 21  A ratio of proved reserves to market prices (R-to-Price) can illustrate a stylized version of these cycles in Adelman’s metaphor. 22 This plot of industry data (Figure 2.4b) suggest the realization of the reserve warehouse has fluctuated through two major cycles between 1955 and 2015. Each point on the curve in Figure 2.4b represents the size of the global oil reserve warehouse and the cost of converting it into production. A lower value indicates fewer reserves or higher costs (more expensive warehouse withdrawals) and vice-versa. In this series, peaks in the ratio of proved reserves-to-price                                                      21   The snapshot of a reserve base at a point in time will include a portfolio of projects with a range of necessary development costs to realize production consistent with today’s output. The view of Adelman (1995) is that the stock of geologic oil resources is irrelevant, and what matters is the development cost needed to provide a regular flow of oil production.  22  Market prices are assumed to reflect some aspect of medium-run marginal costs related to mobilizing reserves – i.e. reserves anticipated to be economically viable are developed at expected market prices.  40 occur in 1970 and approximately 1997-2001; troughs occur in the mid-1950s, 1980, and perhaps 2015. 23  Costs in this cycle test the maximum threshold of market demand in each period. Once market equilibrium no longer supports further growth in development costs, pressure eases on the need to sustain high production growth rates. At this point, upstream cost consolidates around ongoing viable production at a lower market price level (the 1980s and 2014-2016). Regardless of the exact mechanism generating these two cycles, they indicate the industry conditions that lead to increasing reserves at lower cost characterize only part of each cycle since 1955. This suggests the initial formulation of the LBE theory is based on a convention that projects dynamics consistent with the most favorable portion of this cycle for producing low-cost oil and gas (1984-1999). 24 Though the LBE supply model as applied in Rogner (1997) and subsequent studies allows the total size of the warehouse to grow, development expenditures that would govern the rate and costs of ‘warehouse withdrawals’ are noticeably missing from these cost-quantity curves - such as Adelman’s (1995) observation that development costs increase rapidly during periods of high capacity use. Therefore, we can argue the LBE concept for a dynamic oil and gas resource base is a useful description of factors that add to the reserve warehouse’s possible inventory, but a poor model of the inventory’s potential for production.  The costs of oil and gas supply estimated by the LBE theory depict conditions unmoderated by the development costs and market prices that could diminish reserve growth, lower demand or decrease production in a competitive marketplace with a diversity of energy supply strategies. Hence, the long-term fossil supply curves shaped by the LBE theory describe prices and availability consistent with an ideal outlook for permanently optimal investment, where supply is expanded at the lowest possible cost in perfect foresight.  The projection of a learning effect point-estimate from any single state in this reserve-price cycle will result in an over-abundant or overly-scarce depiction of oil and gas supply - each point in the historic time-series of Figure 2.4b is a valid representation of the supply-demand balance for the reserve base at a snapshot in time. Projecting a learning trend that smooths this cycle by starting with a selected baseline period is likely to considerably miscalculate the cost of mobilizing reserves in all future periods by establishing overly-bearish or bullish conditions from the outset.  To illustrate how this distortion is likely to occur, a Monte Carlo (MC) simulation with 200 runs randomly selects a base year R-$ value from a uniform distribution of the underlying time-series data from Figure 2.4b for calculation of compounded learning across the 21st-century. Projections of this reserve-to-price ratio are overlaid on Figure 2.4c for 600 Gtoe of oil (~140 years of supply at 2016 levels) from Rogner (1997), GCAM (Joint Global Change Research Institute, 2016) and 570 Gtoe from MESSAGE (Riahi et al., 2012). Historical trends across a five-year moving average for proved                                                      23  Data for oil prices and reserves were collected from the BP Statistical Review (2017) and for 1948-1980 from the Oil Economists’ Handbook (Jenkins, 2005). 24 One possible interpretation of these cycles could build from the pattern of structural oil demand adoption that generates alternating states of pressure and release on the reserve base. As the reserves in the present become cheaper (upward ascent of each cycle), development costs accelerate to keep production capacity growing in-step with the demand that absorbs increased availability, e.g. as the rate of adoption increased from the 1990s through the early 2000s, development costs increased as a proportion of upstream spend (Figure 2.4a).  41 reserves (P1) to constant 2014 USD (R/$) are illustrated by the solid black line which re-produces Figure 2.4b.  The mean year-2100 value from the results of our MC simulation is R-to-$ = 70.43 billion barrels of reserves per dollar, which represents a steady-state reserve-base condition 55% higher than data indicate the industry has ever experienced. By mid-century, the mean value of all runs illustrates a sustained level that surpasses previous peaks in the late-1990s and before 1973.  Though the projections of Rogner (1997), GCAM and MESSAGE may appear like conservative median estimates when plotted against the full range of simulations, they are consistent with the MC runs which have already reached the most bullish observed industry conditions by 2050. Representing these states as the baseline for industry operation will significantly underestimate the cost and overstate net economic benefits of future supply. Future research on long-term energy resources must strike a balance in recognizing the nature of industry operations which are defined by distinctly bullish and bearish long-term cycles that operate on time-scales of one to two decades – these are not short-term fluctuations that should be smoothed out to enhance model numerical tractability. The boundary between reserves and resources can move in directions that allow production of more reserves at lower cost, and vice-versa.  This section has considered empirical upstream cost dynamics and profiles to evaluate the suitability of the LBE theory’s conceptual pillars. Though none of these assumptions have provided a particularly strong guide for multi-decade studies of oil and gas economics, they have attempted to specify a model of industry productivity that captures essential elements of long-run trends. Therefore, the question arises: if higher resolution data were available, would it be possible to identify and fully distinguish macro-level productivity gains attributable to learning?   42   Figure 2.4a  Development proportion of upstream costs (1978-2009) – as reported by EIA FRS (2011), with top chart displaying normalized Brent price normalized to 2005 for comparison; (middle axis) displays year, left axis shows proportion of upstream cost and bottom axis indicates the ratio of development costs to total upstream cost   43  Figure 2.4b  Two distinct cycles in reserves: quantity of proved reserves-to-Brent prices for oil (1955-2015) – (right-axis) indicates regions of each cycle which lead to increasing pressure on development costs and declining pressure on development costs; (top-axis) notes duration of cycle states from trough-to-peak and peak-to-trough    44   Figure 2.4c Range of estimates for ratio of proved reserves to oil price – historic trend with 5-year moving average (solid black); a Monte Carlo simulation of 200 runs randomly selects from a base year R-$ value between 1950-2014 with uniform distribution and projects this value at 𝜌𝜌 = 1% p.a. (thin lines); values for 600 Gtoe of oil (>140 years of supply at 2014 levels) are provided for Rogner (1997) and GCAM (2012) and 570 Gtoe for MESSAGE (Riahi et al., 2012)   2.2   Nordhaus (2009) on the perils of a learning model: extended to energy resources The LBE theory for future hydrocarbon supply is conceptualized by Rogner (1997) with long-run market prices determined by marginal production costs (P = MC). Market prices are exogenous in this model, as they are driven by the learning that results from continually extracting the geologic resource base.  Though markets for energy commodities are global in scale, resources are locally produced under myriad conditions dictated by firm structure, international politics, royalty and tax accounting, technology, geology and access to markets. Surmising an aggregate estimate of macro-level productivity improvements to inform an appropriate learning rate is at best a speculative venture because of uncertainties in the data, as demonstrated by the volatile year-to-year productivity rates in Section 2.1.3. Rogner (1997) initially acknowledged this, noting: “Because data are consistently poor and have limited availability, estimating productivity gains over extended periods of time is a risky undertaking. Hence, there could be a wide margin for error around this productivity estimate. The projection of a long-term 1% per year growth rate may well prove too conservative (or too optimistic).” The challenging task of specifying robust learning rates for energy technologies is identified throughout the broader literature. McDonald and Schrattenholzer (2001) survey the literature,  45 producing a guide to learning rate selection for a range of energy technologies. Aguliera (2014) reviews several shapes of such cumulative supply curves, and notes that while technological change is generally expected to improve the economics of unconventional and lower grade oil and gas resources, significant upstream investments will be needed to realize these productivity gains. Aguilera and Ripple (2012), the Global Energy Assessment (Rogner et al., 2012), and Bauer et al. (2016) also note that such cost-quantity curves assume perfect implementation of capital investments needed to achieve them. The original formulation of the Rogner (1997) productivity model acknowledges that upstream investments are needed, but they are taken as a given, and so in essence are provided for ‘free’ when IAMs use these supply curves.   Yeh and Rubin (2012) highlight a number of issues with technology experience curves, namely the difficulty of estimating key parameters and the uncertainty of their eventual shape (i.e. evidence for ‘S-shaped’ curves that show rapid growth initially but plateau). Clarke et. al (2006) review literature on learning-by-doing to emphasize there is no single source of learning that dominates single parameter formulations of productivity gains incurred by technological change, i.e. simple experience curve formulations condense a range of factors. Ferioli et al. (2009) support the hypothesis that a chosen learning rate is a ‘surrogate’ for more complex factors by demonstrating a measured “learning curve” can result as an artifact of multiple underlying and simultaneous component processes. 25 Gillingham et al. (2008) review many representations of technological change used in energy models, noting these face common problems with comprehensive empirical data to calibrate parameters in a convincing way.  As highlighted in this literature, learning models of future energy technologies face common problems with empirical evaluation. Therefore, it is most relevant to view them as primarily theoretical in nature, representing hypotheses of technological change. Nordhaus (2009) develops such a theoretical case in order to argue that learning curves for future energy supply strategies are potentially dangerous when applied in policy models: they are highly sensitive to artificial learning rates that could be indistinguishable from measurement errors and simply represent normative choices.  2.2.1 Nordhaus (2009) generic theoretical case  To explore whether exogenous factors could result in cost declines mistakenly measured as productivity gains created by learning, Nordhaus develops a theoretical case where exogenous technical change is denoted as ℎ and true endogenous learning as 𝑟𝑟.  The price function 𝑝𝑝𝑡𝑡 for a generic industry is assumed to equal instantaneous marginal cost 𝑐𝑐𝑡𝑡 , where the rate of cost declines equals the decline in price in Equation 2.1, with 𝑔𝑔𝑡𝑡 representing a constant growth rate for industry output. 𝑝𝑝𝑡𝑡 = 𝑐𝑐𝑡𝑡 = ℎ + 𝑟𝑟𝑔𝑔𝑡𝑡  (2.1)  With a constant marginal cost, price is determined exogenously to current demand. Growth in output (i.e. demand) can be expressed as in Equation 2.2 with constant price elasticity (𝜖𝜖), an elasticity of                                                      25  Notably, Ferioli et al. (2009) state, “Our primary finding is that, even when the learning curve is evaluated over a wide range (i.e., three orders of magnitude of cumulated production) quite different fits of the same set of data are imaginable and at least equally justifiable. We point out that products can often be described as the sum of a learning component and one for which no cost reductions occur.   46 per capita demand with respect to total output (𝜆𝜆), a growth of aggregate per capita output (𝑤𝑤𝑡𝑡), and a constant population growth of 𝑛𝑛. 𝑔𝑔𝑡𝑡 = 𝜖𝜖𝑝𝑝𝑡𝑡 + 𝜆𝜆𝑤𝑤𝑡𝑡 + 𝑛𝑛 = 𝜖𝜖𝑝𝑝𝑡𝑡 + 𝑧𝑧                                           (2.2) Nordhaus summarizes the autonomous, non-price induced growth as 𝑧𝑧𝑡𝑡 = 𝜆𝜆𝑤𝑤𝑡𝑡 + 𝑛𝑛. Substituting Equation 2.1 into Equation 2.2, the total price decline (𝑝𝑝) results (Equation 2.3), and time subscripts are dropped since this is an example of constant growth. Output growth (𝑔𝑔) is then determined by Equation 2.4.                𝑝𝑝 = ℎ + 𝑟𝑟𝑔𝑔 = ℎ + 𝑟𝑟(𝜖𝜖𝑝𝑝 + 𝑧𝑧) = ℎ+𝑟𝑟𝑟𝑟1−𝑟𝑟/𝜖𝜖 (2.3) 𝑔𝑔 = 𝜖𝜖(ℎ + 𝑟𝑟𝑔𝑔) + 𝑧𝑧 = 𝜖𝜖ℎ+𝑟𝑟1−𝜖𝜖𝑟𝑟                                         (2.4) The contribution of learning to declining prices would be calculated as 𝜌𝜌, where the learning curve is the ratio of price to growth, 𝑝𝑝𝑔𝑔 illustrated in Equation 2.5.    𝜌𝜌 = 𝑝𝑝𝑔𝑔= ℎ+𝑟𝑟𝑟𝑟𝜖𝜖ℎ+𝑟𝑟                                                     (2.5) Nordhaus argues that in Equation 2.5, the exact contribution of learning is difficult to determine because so many coefficients are present. It is challenging to know how much of the price decline can be confidently attributed to learning in Equation 2.5 unless we know the exact values of exogenous technical change, the demand elasticity and the rate of autonomous demand growth. Fro example, in the case of oil and gas there could be another variable denoting political goals to capture the influence of cartels (or domestic budget targets of oil producers).  Nordhaus uses this formulation to develop a quantitative case that illustrates calculations of learning for the generic manufacturing industry. He assumes plausible values for price elasticity (𝜖𝜖 = 1), exogenous demand growth (𝑧𝑧 = 0.04) and a rate of exogenous technical change (ℎ = 0.01). When true learning is zero (𝑟𝑟 = 0), substituting these values into Equation 2.5 results in a calculation of ‘learning’ at 20% (𝜌𝜌 = 0.20) from Equation 2.6. 𝜌𝜌 = 0.01+0×0.041×0.01+.04 = 0.010.05 = 0.20                                   (2.6) Nordhaus then considers that if the true learning value were greater than zero (𝑟𝑟 = 0.25) the learning rate receives an even larger upward bias - calculated as 𝜌𝜌 = 0.4. This logic supports his conclusion that the learning curves applied in models of endogenous technical change will tend toward a consistent upward bias because of complications induced by interactions between demand, output growth and exogenous technical change. Therefore, disentangling the explicit effect of learning induced price declines - independent of all other exogenous factors - requires a highly detailed study of specific industry conditions.  In summary, Nordhaus suggests:   47 (a)  there is a fundamental statistical identification problem in separating an endogenous learning effect from exogenous productivity gains;  (b)  the subsequent estimated learning coefficient is generally biased upwards;  (c)  model parameters intended to represent learning effects are not robust to alternative explanations and specifications;  (d)  overestimates of learning coefficients will underestimate the total marginal cost of output for a technology; and  (e)  models that rely on learning curves are likely to simply choose technologies which incorrectly or arbitrarily specify a high learning coefficient, e.g. an upwardly biased long-run learning rate for any technology will allow it to ‘rise above the rest’.  Though Nordhaus (2009) agrees that productivity benefits follow as workers in firms gain experience with a production process, he expresses skepticism that embodied learning can be measured reliably for large global systems. The cumulative 'supply' of learning could be embodied in firm, a group of workers, an individual worker, or it could result from international or interindustry spillovers (Clarke et al., 2006 also emphasizes the learning effect of spillovers). Further, a measured learning improvement may not be durable (Yeh and Rubin, 2012 highlight such discontinuities in learning).  2.2.2 Drilling into factors of oil and gas productivity: a framework for components of potential learning?  As highlighted by Ferioli et al. (2009) and Nordhaus (2009), learning rates can be the result of multiple underlying component factors. The case developed by Nordhaus (2009) for exaggerated learning in a generic industry can be extended to examine the LBE theory by considering specific technical features of oil and gas extraction.  Hamilton (2012) analyzes the impact of technology and price on oil production over the last century in the United States (1859-2010) and across the world (1973-2010), extending far beyond the small window covered by EIA FRS data in Section 2.1 of this paper. Hamilton draws on these data to conclude that individual oil producing regions have not demonstrated a pattern of continuously increasing productivity from ongoing technological progress.  In Hamilton’s view, price incentives and technology have reversed declines in output resulting from geological or geographic factors, but only temporarily. Measured productivity gains in oil producing regions initially increase as new fields are developed, followed by productivity declines dominated by the natural depletion rate of wells, and mitigated through enhanced oil recovery techniques. Hamilton suggests the primary historical source of industry productivity gains and increasing global oil production during the 20th-century has been the exploitation of new geographical areas. While Hamilton is only focused on empirical aspects of the past and does not consider the potential long-run theoretical contribution of learning and unspecified technological breakthroughs to productivity, his analysis serves as a reminder of the engineering factors related to geology and geography which distinguish the oil and gas industry from other forms of manufacturing. 26 The                                                      26 Further, given the important role of oil in the economy, wholly political decisions have resulted in rapid growth of output and subsequent price declines, outside of what any would consider as free market equilibrium conditions. At  48 determination of a true learning rate for oil and gas could be further distorted by such complex industry conditions, adding another element to the issues raised by Nordhaus (2009). For energy resources it is unclear whether the role of learning and upstream technology improvements can always be fully distinguished from productivity gains resulting from specific geographic or geologic factors.  This dilemma mirrors echoes Adelman (1990): that the oil industry is an “endless tug-of-war between diminishing returns and increasing knowledge.” As currently formulated, the LBE model projects future supply costs as determined by a function of increasing knowledge, which pulls Adelman’s tug-of-war in a single direction.  The cycles of cost and reserves reviewed in Section 2.1 suggest that Adelman’s metaphor is apt - the reserve base does get pulled in both directions. Though applications LBE supply curve generally use lowest cost resources first, this weakly captures the effect depletion may have on costs of accessing the full geologic resource stock over the long-run, and misses an opportunity to understand the investments needed to offset declines. 27 The use of compounded learning as the prime determinant of projected future oil and gas supply costs develops Adelman’s industry metaphor in a way that confirms the concern of Nordhaus (2009). Hamilton’s analysis highlights the factors of oil production that counter the increasing returns from knowledge and learning, indicating a path toward integrating the insights of Nordhaus and Adelman. Global oil and gas production in the 21st-century will balance, benefit and suffer from both increasing knowledge and diminishing returns.   2.2.3 Measuring learning-by-extracting alongside geological and geographical factors Accordingly, the case of Nordhaus (2009) can include additional factors relevant to productivity in the oil and gas industry. The parameter 𝑜𝑜 is introduced to represent upstream productivity that results from geographic expansion and geological conditions. When the price, cost, output and growth assumptions of Nordhaus (2009) are adopted, the original equation for declines in price (𝑝𝑝) as a function of productivity gains results as Equation 2.7.  𝑝𝑝 = ℎ + 𝑜𝑜 + 𝑟𝑟(𝜖𝜖 + 𝑧𝑧)  (2.7) In this equation, following the notation from Section 2.2.1 the rate of true endogenous learning is denoted by 𝑟𝑟, exogenous technical change by ℎ, constant price elasticity by 𝜖𝜖, and 𝑧𝑧 is a function of autonomous, non-price induced growth. With this modification, the industry cost function is assumed to involve factors specific to engineering for oil and gas extraction (𝑜𝑜), which may also influence productivity independent of learning induced technical change.                                                       the moment these political concerns are left aside, as well as oligopoly or monopoly features that could create price declines to consolidate market share.  27  Though energy models commonly draw from such cost-quantity curves sequentially, depleting resources left-to-right, today’s oil supply of approximately 68 mbd conventional, 15 mbd natural gas liquids, and 8 mbd unconventional production (IEA, 2016) would draw from multiple points in the availability curve at any one time in order to meet the time constraints of demand (see Chapter 5).    49 In a case that considers production from an oil well in Texas, 𝑟𝑟 would capture endogenous learning that leads to productivity improvements for on-site extraction (e.g. the local crew gets better at operating the well). The specific location and geologic nature of the oil well would impose productivity considerations captured by 𝑜𝑜, such as favorable drilling conditions resulting from the initial pressure at the wellhead or the natural profile of production increases and declines indicative of a maturing oil well. A calculation of the learning coefficient 𝜌𝜌 from Equation 2.7 results in Equation 2.8, where exogenous and true endogenous learning are combined with productivity gains enabled by geology and geography. If we adopt the values of exogenous technical change, used by Nordhaus in Section 2.2.1, with a true learning rate of 𝑟𝑟 = 0.25 and consider that geology or geography may contribute at a 1% productivity gain (𝑜𝑜 = 0.01) then the learning coefficient would be measured at 𝜌𝜌 = 0.5, twice the true rate of learning (𝑟𝑟). 𝜌𝜌 = 𝑝𝑝𝑔𝑔= ℎ+𝑜𝑜+𝑟𝑟𝑟𝑟𝑟𝑟+𝜖𝜖ℎ+𝜖𝜖𝑜𝑜= 0.01+0.01+.25×0.040.04+0.01+0.01 = 0.5 (2.8) With these plausible values for exogenous technical change, autonomous growth and demand elasticity, the sensitivity of 𝜌𝜌 to 𝑜𝑜 and its relationship to 𝑟𝑟 can be further considered: with 𝑜𝑜 → 2𝑟𝑟 the marginal contribution of 𝑜𝑜 to 𝜌𝜌 rapidly declines as the calculated learning curve approaches unity. Following from this case, even if 𝑜𝑜 were twice the value of 𝑟𝑟, the calculated effect would remain roughly unchanged from when 𝑜𝑜 < 0.8𝑟𝑟. This association is plotted in Figure 2.5a where the measured learning rate (black line) follows an asymptotic trajectory, the true learning rate (blue line) remains static, and the geographic-geologic productivity gain ranges from values of 0 ≤ 𝑜𝑜 ≤ 2𝑟𝑟.             50     Figure 2.5a  Relationship of productivity rates in Equation 2.8 – using the values in the Nordhaus (2009) case, the ratio of geographic-geologic productivity gain (o) to true learning (r) (bottom axis) is plotted against the measured productivity gain (% per year) (left axis); the black line represents the measured productivity gain (ρ), the yellow line the geographic-geologic productivity gain (o), for a static true productivity gain (r = 0.25)     51  Figure 2.5b  Possible values for productivity gains from each component for a measured learning rate of 1% per year (𝜌𝜌 = 1%) – possible values for productivity components (o – left axis and r – bottom axis) that could result in a measured learning rate of 1% per year, plotted for a range of values that correspond to the capex productivity measured in Section 2.1.2 with parameter values from Nordhaus (2009) (orange line) and a specification more consistent with historical values for demand elasticity and growth  If we interpret the H-H-R (1997) assumption of 1% endogenous learning for global oil and gas as 𝜌𝜌 =0.01, using the structural form of Equation 2.8, we can solve for plausible values of geographic-geologic and true learning based productivity that could produce such a productivity measurement. Figure 2.5b plots this relationship for a range of 𝑟𝑟 values that corresponds to annual capex productivity measurements from Section 2.1.2, and the slope indicates that 𝑜𝑜 and 𝑟𝑟 are negatively correlated. Two specifications of Equation 2.8 are provided in this figure based on the Nordhaus (2009) parameter values used in this section (orange line), and a case with more historically consistent values for oil (yellow). 28 Based on the relationship depicted in Figure 2.5b (orange line), in a case where geological conditions result in a negative contribution to productivity such that 𝑜𝑜 = -1.25%, then the equivalent learning rate to maintain a +1.0% productivity gain on average would need to sustain 7% per year. Conversely, a high true learning rate of 25% could be masked by a contribution from 𝑜𝑜 at -2% per year, underestimating the influence of technological change. 29 It follows that with a measured learning rate of +1% that 𝑜𝑜 and 𝑟𝑟 could easily mask their relative contribution, i.e. a sustained high level of true learning is needed to compensate for a slightly negative contribution by 𝑜𝑜, or conversely, a high level of true learning could appear low with an order of magnitude smaller contribution from 𝑜𝑜.                                                      28 The parameter values for the historically consistent case are exogenous productivity at 1% (h = 0.01%), demand growth at 1.1% (z = 1.1%) and elasticity of demand at 0.1 (𝜖𝜖 = 0.1).  29  In a case where true learning is zero, exogenous productivity gains dominate.   52 In summary, a measurement of learning induced productivity that fails to capture the effect of 𝑜𝑜 could readily obtain a biased value for an extrapolation of the learning-by doing effect for future oil and gas production. These considerations illustrate that truly disentangling the contribution of endogenous learning from geology, geography or exogenous factors is extremely difficult without explicit studies of producing fields. Establishing the appropriate value of a learning parameter for long-term fossil energy supply is a complex process that needs a further robust modeling effort to remain relevance in future studies on climate and energy policy.  For the FRS data considered in Section 2.1, the mean value for decadal upstream productivity appears to be negative (𝜌𝜌 < 0), further complicating the picture. Was learning-by-doing negative or did geology, geography or exogenous factors dominate cost increases? Presuming a deterministic level of true learning over a long-time frame needs to overcome measurement issues such as these to become a plausible description of future hydrocarbon economics. When an observable productivity trend is the product of two unknowns, guessing the value of each without an empirically constrained distribution of plausible values is difficult to separate from a normative choice. In the next section we consider the relevance of these considerations for projections of long-run technology costs.  2.3   Implications of chosen learning rates for long-run energy economics and climate change mitigation cost projections Fossil energy supply curves constructed with the LBE theory generally indicate that the vast quantity of fossil occurrences in the Earth’s crust will readily dominate 21st-century choices for energy supply. Thus, policy goals for reducing carbon emissions to limit future climate change face stringent competition from the low-cost hydrocarbon deposits expected to result from compounded learning. The projected cost of any backstop technology that could readily substitute for these resources can also receive a bias from any selected learning rate. This section considers a simple example, and then applies various estimates of the LBE learning parameter in the GCAM IAM to test its effect.   2.3.1 Influence of learning rates on total 21st-century cost of oil supply and required low-carbon backstop prices If annual oil production growth continues across the 21st-century at 1.1% p.a. (21st-century trend), 660 Gtoe is withdrawn from the H-H-R Supply Curve. Varied rates of productivity from +1% to -1% applied to the oil cost ranges calculated by Rogner (1997) (Figure 2.6) adjusts the total discounted cost of supply by more than a factor of 7. As this case for oil demonstrates, an upwardly biased learning rate for 21st-century fossil energy supply can easily underestimate the cost of future oil supply by 1.6-to-7.4x per barrel.  Understating the cost of oil supply will also overstate the investment required to mitigate its GHG emissions with a low carbon alternative. If we equate the cost of future oil supply in the 𝜌𝜌 = + 1.0% case to a market price of $50/bbl average over the century, a $120/bbl zero carbon backstop oil substitute available today appears as a significant cost: a 60% reduction in the backstop cost is required for substitution with no deadweight loss. Yet, with an oil supply calculated at 𝜌𝜌 = -1.0%, this backstop is already 200% more cost effective than oil over the long-run – optimal energy policy in  53 this case calculates that short-run substitution should be incentivized because of negligible deadweight losses. 30  This simple case depicts how slight changes to the learning rate applied for an oil, gas and coal dependent policy baseline can frame a consistent series of mitigation steps as either net costs or net benefits - substantiating Nordhaus’ concern that a learning model for developing long-term energy strategies is potentially ‘dangerous’.    Figure 2.6  Cumulative discounted cost of 21st-century oil supply – growth in oil production at 21st-century consistent level (1.1%/year) with H-H-R price bands at learning rates of ρ = +1%, 0.5%, 0%, -0.5% and -1% for discount rate of 5%; (right bar) multiples of total 21st-century oil supply cost compared to the ρ = +1% case Ferioli et al. (2009) provide a similar example from the case of wind turbine learning curves from 1995 applied to estimate multi-decade deployment and investment costs (Neij, 1997). The cost of wind power capacity in 1995 was estimated at 1,333 $(1995)/kW with a total installed capacity of 5 GW worldwide and a learning rate of 4% per year (Neij, 1997). These led to projected installation costs for the year 2004 that were 5-25% above the actual reported installation, in less than a decade from the initial estimate (Ferioli et al., 2009). Further, cost reductions in wind power capacity were achieved at only one-fifth the cumulative investment estimated by the initial learning curve.  As demonstrated by the case of oil withdrawn from the original H-H-R Supply curve and the projected cost of wind power, uncertainties in learning rates can have a significant impact on the economics of multi-decade energy projections. While the simple examples in previous sections illustrate the basic influence of a chosen learning rate, they do not provide an integrated description of the significance fossil resource productivity gains could have on supply and demand in a long-run scenario. Therefore, we turn to an IAM.                                                       30   Learning calculated at 0% also considers that the $120 backstop is already more effective than oil.   54 2.3.2 Fossil productivity in the long-run: the LBE model in the GCAM IAM  The GCAM IAM is widely used to develop long-run energy scenarios, and uses an exact implementation of the LBE theory to structure its oil, gas and coal supply curves. Therefore, we can conduct a sensitivity analysis of learning rates for future fossil energy production in GCAM based on data analyzed in this chapter.  By default, the GCAM IAM applies a 0.75% per year productivity gain to every fossil resource in all 32 regions. However, it allows learning rates to be set per region and time period. 31 Six alternate reference cases were constructed in GCAM with different specific rates of learning for conventional oil production from 2005-2100. Each case uses a conventional oil learning rate consistent with the range in the EIA FRS data set based on 𝜌𝜌 = +1%, +0.5%, 0%, -0.5% and -1% and a value that reflects the 1978-2008 average trend of 𝜌𝜌 = -3.6% (which also approximates the opex time-series average from Section 2.2.2). No other aspects of these six scenarios were modified from the original GCAM reference case (GCAM-ref). Results from the default GCAM-ref scenario are plotted as a dark gray line to provide comparison (𝜌𝜌 = 0.75%).  2.3.2.1 Effect of learning rates on prices and cost-quantity curves Figure 2.7a shows the varied cost trajectories produced by each reference case. In the 𝜌𝜌 = +1% (red) and GCAM-ref (dark gray) scenarios, the cost of conventional oil remains at a consistently low level, leading to higher demand for this resource. In these two scenarios, total depletion of conventional oil occurs before end of the century, and then the cost curve explodes upward. The thirty-year decadal trend scenario (light blue) highlights the absurdity of using long-run averages to inform the chosen learning rate, as effective prices double their level from GCAM-ref before 2040 to more than $170 per barrel. The other rates of learning (𝜌𝜌 = +0.5%, 0%, -0.5% and -1%) indicate a range of stable oil price trajectories without severe discontinuities. Table 2.1 summarizes the prices for each alternate scenario at twenty-year intervals, emphasizing that even slight differences between learning rates can lead to dramatic discrepancies after many decades.  The underlying 21st-century cost-quantity curves for each alternate scenario are plotted in Figure 2.7b, framed by two reference points for conventional oil resources from the BGR (2015) assessment – a light grey line for total conventional oil reserves (7,140 EJ) and a dark black line for conventional oil reserves and resources (7,140 EJ + 6,815 EJ = 13,960 EJ). Learning rates greater than zero are consistent with supply curves for 30% more oil than BGR (2015) reserves and resources, while the zero-learning case uses all reserves and resources.                                                         31  As Chapter 4 is focused on GCAM in greater detail, a more comprehensive description of the model’s core structure and functionality in that section.  55 Table 2-1 - Alternate GCAM reference cases for conventional oil learning rates – effect on cost of oil (% change from unmodified GCAM-ref)   Change in cost of oil supply (% change from GCAM-ref) Annual learning rate (𝜌𝜌) (2005-2100) 2020 2040 2060 2080 2100 +1.0% -0.8% -2.9% -4.3% -5.2% +3396.3% +0.5% +0.9% +3.3% +5.0% +6.2% -33.4% 0% +3.0% +11.3% +17.3% +22.7% -20.9% -0.5% +5.6% +21.2% +33.8% +46.3% -0.8% -1% +8.8% +32.1% +55.4% +80.6% +31.3% -3.6% +38.4% +133.7% +326.4% +718.3% +856.3%   Figure 2.7a 21st-century oil cost trajectories from GCAM alternate learning reference cases – (left axis) oil cost in $2014 USD per GJ, and per boe (right axis)  56  Figure 2.7b  Cumulative conventional oil supply curves (cost-quantity) produced by GCAM alternate learning scenarios – (left axis) oil cost in $2014 USD per GJ, and per boe (right axis) with total supply in exajoules (bottom axis) and in gigatons oil equivalent (top axis); vertical lines mark two reference points from the BGR (2015) resource assessment for conventional oil reserves (grey line) and reserves + resources (black line) 2.3.2.2 Effect of learning rates on production profiles and resource substitution Twenty-first century conventional oil production profiles for the GCAM learning rate scenarios are plotted in Figure 2.8a with a reference line to mark the year-2014 rate. In 2014 conventional crude oil was produced at 67.2 mbd, natural gas liquids (NGLs) at 14.8 mbd, and unconventional oil at 7.6 mbd (IEA, 2016). 32 Conventional oil production in the learning rate scenarios between 𝜌𝜌 = -1% and 𝜌𝜌 = +1% reaches a maximum between 2020 and 2030 around 95-100 mbd - about 20% higher than recent levels. Positive learning rates lead to more gradual production declines after 2030, and a production resurgence after 2060 with learning above 𝜌𝜌 = 0.5% as production experience accumulates, so that conventional oil is more competitive in later decades. Key features of each production profile are detailed in Table 2-2.  Conventional oil declines in GCAM are offset by backstop resources. A backstop resource is a substitute energy source that is expected to come online when cheaper source are exhausted (Hartwick and Olewiler, 1986). The basic idea of a backstop resource is that sufficient incentive will exist to develop an alternative if the prices of today’s energy sources are high enough. However,                                                      32   For simplicity, the NGL and conventional oil production rates are combined in the reference line at 170 EJ (82 mbd), as GCAM does not appear to distinguish NGL production and consumption from conventional oil. NGLs are hydrocarbons like ethane, propane, butane, isobutene, and pentane. Ethane makes up the bulk of NGL production and is used for ethylene as a plastic feedstock. NGLs are commonly used as fuels for space heat or cooking, as a blend for vehicle fuel, as inputs for petrochemical plants, or as a diluent to increase viscosity of unconventional oil through pipelines.     57 until prices for the conventional technology reach a certain threshold level, the backstop technology does not receive enough investment or attention, so it awaits its introduction into the marketplace.  Building on Hotelling (1931), Nordhaus (1973) defines backstop technologies as a set of processes that are: (i) capable of meeting demand requirements and (ii) have a virtually infinite resource base. Though the backstop technology could be expensive compared to today’s technology in a long-run model, "if it exists, it assures that the planning problem at least has a feasible solution.” Nordhaus states that, “in some sense, the current stage of history is a transitory phase between dependence on cheap but scarce resources, and dependence on more costly but abundant resources.” The ideation of GCAM’s resource use applies this theory. The abundant backstop resources for liquid fuels in GCAM are unconventional oil and coal-based liquids. Chapter 7 revisits the context of coal-based liquids in more detail.  Unconventional oil and coal rise to the occasion of conventional oil depletion in GCAM scenarios. Bearish outlooks for conventional oil lead to increasingly rapid deployment of these backstop resources (Table 2-2). Unconventional oil production outlooks for each alternate learning scenario are plotted in Figure 2.8b. Curiously, although the historic average productivity rate scenario (light blue) is not consistent with oil price trajectories, GCAM provides highly skillful projections for year-2014 unconventional and conventional oil production rates.  Figure 2.8c plots the cumulative 21st-century production outlooks for conventional oil and its backstop resources as a function of learning rates.  At 𝜌𝜌 = -1.0% unconventional and conventional oil production converge at a level of 10,000 EJ. Although the production profiles for coal differ slightly among learning rates (Table 2-2) the total amount of coal extracted remains roughly equivalent at learning rates below 0%. The cumulative amount of 21st-century coal is much less sensitive to learning rates than unconventional production profiles, varying only around 10%.  Table 2-2 - Alternate GCAM reference cases for conventional oil learning rates – effect on oil production phase out and backstop deployment      Unconventional Oil Coal Total Backstop Annual learning rate (𝜌𝜌) Maximum Year Post-Peak Minimum Year Annual Decline Rate* Resurgence Maximum Year Year 2050  Mbdoe/yr Year 2100  Mbdoe/yr  Year 2050  EJ/yr Year 2100  EJ/yr Year 2050  EJ/yr Year 2100 EJ/yr +1.0% 2030 2050 -0.89% 2080 13  147 370 600 400 910 +0.75% 2030 2060 -0.86% 2075 14 134 380 600 410 880 +0.5% 2025 2100 -0.7% --- 16 83 390 620 420 800 0% 2025 2100 -1.7% --- 21 100 400 650 440 860 -0.5% 2025 2100 -3.0% --- 27 120 420 660 480 910 -1% 2020 2100 -4.7% --- 35 130 430 650 500 920 -3.6% 2010 2060 -11.8% --- 64 140 430 640 560 930 * Year-to-year decline rate from maximum to post-peak minimum  58  Figure 2.8a Conventional oil production outlooks for scenarios of alternate 21st-century learning rates – (left axis) annual conventional oil production in exajoules (EJ/yr), and annual average million barrels per day (right axis – mbd/yr) with reference line for year-2014 conventional oil production (dotted gray line)    Figure 2.8b Unconventional oil production outlooks for scenarios of alternate 21st-century learning rates – (left axis) annual unconventional oil production in exajoules (EJ/yr), and in units of annual average million barrels per day (right axis – mbd/yr) with reference line for year-2014 unconventional oil production (dotted gray line)     59  Figure 2.8c Cumulative oil and coal production outlooks for alternate learning rates – (left axis) cumulative 21st-century primary energy (EJ), and annual average from 2005-2100 (right axis – EJ/yr) for conventional oil (black line), unconventional oil (orange line), and coal (yellow line) with reference line for GCAM-ref learning rate (dotted black line); scenario conventional oil learning rates (bottom axis) The alternate scenarios show a positive correlation between higher levels of backstop deployment and 21st-century cumulative carbon emissions. Intuitively, more combustion of high-carbon backstop resources that substitute for relatively lower-carbon conventional oil leads to increased carbon intensity of the global energy supply (Figure 2.9a). Backstop resources combine to constitute between 60-74% of total 21st-century carbon emissions in each scenario – ranging between 1,160 GtC and 1,510 GtC (the span between the RCP4.5 and RCP6.0 carbon supply curves – see Chapter 5). Coal combustion results in approximately 1,000 GtC of cumulative emissions for each learning rate scenario (975-1080 GtC). Alternate reference cases solved for a 2˚C climate policy goal 33 indicates that policy costs and carbon prices show the same relationship as the CO2 outcomes: higher levels of high-carbon backstop deployment lead to higher carbon prices by 10-30% (Figure 2.9b).  This result confirms the relevance of the sketchbook example from Section 2.3.1 and gives pause to consider whether high-carbon backstop resources are inevitable – as the chosen high-carbon backstop could significantly influence the outlook for 21st-century climate economics given that modest variations in conventional oil learning rates lead to considerable changes to GCAM’s carbon prices.                                                        33  A 2˚ climate policy goal in this thesis refers to steps that intend to limit end-of-century warming to 2˚C above pre-industrial levels.   60   Figure 2.9a Total 21st-century carbon emissions from conventional oil learning rates – (left axis) 21st-century cumulative carbon emissions (GtC) for all sources and backstop resources, with reference line for GCAM-ref learning rate (dotted black line); scenario conventional oil learning rates (bottom axis); proportion of cumulative carbon emissions from backstop resources (top axis)  Figure 2.9b Year-2100 discounted carbon price across alternate scenarios – (left axis) year-2100 carbon price discounted at 3% ($2014/ton carbon), and % difference from GCAM-ref (right axis) for varied scenario conventional oil learning rates (bottom axis); note y-axis break  61 2.4   Summary This chapter has considered the widely used ideation for long-term resource assessments in climate and energy policy studies – a learning-by-doing theory. Rogner (1997) articulates a theory of future hydrocarbon energy supply driven by annual productivity gains that result from learning, applied to the global oil, gas and coal resource base at a rate of 1% p.a. (𝜌𝜌 = +1.0%). When compounded across a century, this results in an aggregate productivity improvement that reduces the cost of fossil energy extraction by more than 170% (𝚸𝚸 = 170%/boe). While this approach is commonly used in integrated assessments of climate change and energy economics, there are no systematic studies available in the literature to calibrate learning rates per resource and industry.  EIA data on the financial metrics of US and worldwide oil and gas production indicate that decadal productivity trends are not stable and may significantly vary (Section 2.1).  From 1988-1996 they were broadly consistent with a +1.0% upstream productivity improvement as projected by Rogner (1997); overall, depending on the measure, oil and gas productivity declined by 3.0% (total upstream) to 5.2% (capex) p.a. from 1977-2009.   The volatility of year-to-year productivity changes in these data indicate the relevance of modeling stochastic processes of discovery, innovation and market conditions when calibrating long-term costs: projection of aggregate productivity gains in oil and gas as a stable secular trend over a century is likely to mislead any model of technology adoption patterns. Further, it appears likely that sustained periods of high demand (e.g. bull markets) reduce short-run and medium-run productivity, dominating any gains induced by learning during the same period (Section 2.1.2). Because conditions of sustained rapid fossil energy adoption are often projected in many long-term energy reference cases scenarios, these factors will be of explicit interest to the research community (Bauer et al., 2017; Clarke et al., 2014; Dellink et al., 2016; Riahi et al., 2017). Conversely, productivity estimates drawn from recent industry bear market conditions for cost reductions in US shale and unconventional technologies must be understood in the full context of labor, capital, markets and producer investment decisions. Recent upstream cost declines may not immediately translate to the next bull market.  LBE supply curves have accurately captured features of secular trends in oil and gas production operational expenditures, however these are one-third of the marginal costs reported by producers. A focus on learning driven productivity gains as a primary determinant of long-run energy resource costs has blurred the distinction between the specific processes, technologies and decisions on capital expenditures required to produce fossil energy commodities. Each fuel type, geography and geology is not a single manufacturing process, but requires a portfolio of technologies with their own learning rates and potentials. However, there is reason to expect that operational learning curves can extend well into the future to capture potential effects from automation, machine learning and information technology.  Across a century time-span, the production of fossil energy will require multiple manufacturing processes. The history of oil illustrates a dramatic span of technology between the horizontal and deepwater drilling of today and the original 19th-century wells of Drake, Pennsylvania. The work in this chapter suggests the present generation of fossil supply curves structured by the LBE theory have extended a learning model suitable for a single manufacturing process in a given facility  62 beyond its plausible boundaries. Recent expansion in supply has required new technologies with a different pattern of development costs, indicating the confines of this framework.    Though the LBE model poorly captures development costs, these expenditures play a significant role in driving marginal cost throughout cycles of demand on the ‘warehouse stock’ of reserve vintages (Section 2.1.3). EIA FRS data suggests development expenditures constitute more than half of oil and gas production in times of fast growing demand, a dynamic articulated by Adelman. While an equilibrium R-P ratio describes essential features of oil and gas producer behavior for knowledge of energy resources, this framework appears of limited value for determining future costs and availability. Varied costs of developing reserves into production has contributed to distinct cycles in the oil and gas reserve base over the last six decades (Section 2.1.3).  The learning-driven productivity modeled in LBE Supply Curves is explained by Rogner (1997) to result from optimal investment: this is only possible in an environment where prices may increase without a maximum threshold of demand on market price, e.g. where demand faces no constraints. This chapter shows that industry investment in capital expenditures is not optimal but faces boundary conditions at regular intervals. Relaxing the assumption of optimal investment for supply expansion opens an avenue for recalibrating long-term outlooks on fossil energy supply.   Nordhaus (2009) argues a learning model of productivity is dangerous for long-term energy studies: there are too many exogenous factors to isolate the contribution of learning-by-doing for a homogenous global energy technology (Section 2.2). Consequently, the assessed contribution of learning is generally biased upwards. Where Adelman (1990) sees the future of the oil and gas industry determined by a tug-of-war between diminishing returns and increasing knowledge. The argument was developed that the LBE theory of fossil supply curves has only pulled this rope in one direction.  Section 2.3 indicates that Nordhaus’ concerns apply to the LBE model of future fossil energy, which has the potential to distort the necessary policy costs for reducing future GHG emissions. The sketchbook example in Section 2.3.1 illustrates how estimates of energy resources based on autonomous learning can readily bias energy system reference cases for energy and climate policy: projected costs of mitigation and backstop technologies can easily be framed as net costs or benefits with slight changes to a chosen learning rate. Therefore, a rigorous justification is necessary for any selected learning rate in studies that apply Rogner's theory.  Section 2.3.2 confirms that slight changes to the learning rate for conventional oil in an IAM reference case (GCAM) has a significant influence on projected oil production profiles, and the deployment of high-carbon backstop resources: unconventional oil and coal. This establishes a positive correlation between scenarios with more bullish outlooks for backstop resources, higher levels of 21st-century carbon emissions, and increased costs for climate policy. This finding suggest that high-carbon backstop resource costs and availability could have a significant influence on the economics of climate change since they constitute as much as two-thirds of the total emissions in an IAM scenario. Though unconventional oil is the primary backstop resource for liquid fuel, coal also plays a major role, producing roughly 1,000 GtC of cumulative emissions regardless of the conventional oil learning rate. Accordingly, the next three chapters examine the concept of this coal supply curve, and the role of a coal backstop in IAM scenarios of global energy supply.    63 Chapter 3:   The 1,000 GtC coal question: Are scenarios of vastly expanded future coal combustion still plausible? As in the GCAM reference cases from Chapter 2, studies of global energy futures commonly derive scenarios by assuming continued growth in demand for primary energy resources. Long-run compounding in these projections lead to outlooks for a significant increase in fossil energy supply from today’s levels. When demand for transportation and industrial fuels exceeds available output from oil and gas, reference cases customarily illustrate vastly expanded coal production as the backstop (Clarke et al., 2014; Energy Modeling Forum, 1995; Grübler et al., 1998; Häfele, 1981; Nakicenovic et al., 2000; World Energy Council, 1993).  Through much of the 20th-century ratios of reserves-to-production (R-P) for coal were very high, providing a theoretical basis underlying such long-term energy scenarios: if existing coal reserves were depleted, producers could presumably have sufficient incentive to readily explore for more coal, reclassifying marginal geologic resources as recoverable reserves. Vintage reserve assessments indicated coal R-P ratios of well over 300 years. Global recoverable reserves of hard coal reported in the 1960s amounted to nearly 2,000 gigatons (Gt) - an R-P of more than 900 years (Flawn, 1966). This contributed to a common perception that the total occurrences of coal in our planet’s crust could provide the ultimate assurance of energy security, supporting ambitious growth in primary energy demand, or fully compensating for depletion of oil and gas.  However, efforts to determine the potentially recoverable portion of world coal resources have been fragmented, compromising time-series analyses with notoriously inconsistent and poor data (Gordon, 1987; Smil, 2003). Unreliable information has meant any determination of the plausible extent for future coal use amounted to a choice between Scylla, Charybdis, and the full scope of Hades.  These options can be characterized as three distinct modeling approaches:  • Method A - projecting a trend from any selected baseline window by assuming continued momentum in consumption growth, as common in medium-term outlooks. 34 • Method B - adopting reserve figures known to be inconsistent and incomplete, or  • Method C - expecting that coal follows the dynamics of oil and gas, and thus marginal geologic occurrences are reclassified as reserves, maintaining a range of equilibrium R-P values to replace depletion (Adelman and Watkins, 2008; Rogner, 1997; Watkins, 2006; Wellmer, 2008; Wellmer and Berner, 1997). Chapter 2 (Section 2.1.3) reviewed the equilibrium R-P concept for oil, and the following chapter analyzes this important energy model ideation in the context of coal.  Demand-focused projections using Method A attempt to infer likely growth rates for future consumption. However, estimates that extrapolate compounding demand can become abstracted from feasible trends in coal technology, production, and economics. Studies spanning longer timeframes tend to express the ultimate potential for realizable supply by integrating Methods B and C. Future energy scenarios derived from this approach interpret global coal reserve estimates as a                                                      34  Examples include the International Energy Agency’s World Energy Outlooks and the Energy Information Administration’s Annual Energy Outlooks.  64 conservative lower bound, since it is often considered that the vast resource base could be tapped into, replacing depletion over time while maintaining a constant range of R-P ratios.  Energy Modeling Forum (EMF) studies illustrate how this method is used to understand total geologic occurrences of coal in comparison to ultimately recoverable resource figures of oil and gas (Table 3.1). While oil and gas resources are characterized by total resource-to-production ratios that describe centuries, coal is portrayed on the order of several millennia. McCollum et al. (2014) draws on these numbers to explain that although baseline runs of many 21st-century scenarios depict cumulative coal production which greatly exceeds today’s reserve estimates, such outlooks should be considered plausible because they are well within use of the total geologic resource base.  Table 3-1 - Global fossil energy resource base for two EMF studies (EMF 1995; McCollum et al. 2014)   EMF 14 (1995)*  EMF 27 (2014)✝ Total resource✝ Quantity Exajoules (EJ) Resource-to-production 1995 production rate  Quantity  Exajoules (EJ) Resource-to-production 2014 production rate Oil (conventional) 14,200  100  13,800 80 Gas (conventional) 16,340 200  16,000 120 Coal 300,000 3,190  456,100 2,840 *EMF 14 (1995) figures for crude oil and natural gas include reserves alongside undiscovered resources, while the quantity reported for coal is listed as “ultimately recoverable resources”; the EMF 27 study provides clearer distinction between the conventional and unconventional quantities of oil and gas ✝Includes reserves and resources   The conceptual basis for this interpretation of total coal resource assessments applies an equilibrium R-P value of several centuries to the totality of geologic occurrences, depicting a vast potential for production growth. Yet, over the past three decades, the global R-P ratio did not maintain equilibrium and continually declined.  This chapter examines whether the use of total coal resource base figures have any analytical meaning for energy system reference cases, through seeking to understand the information conveyed by coal assessments, and the distinct context of terms such as ‘resources’, ‘reserves’ and ‘recoverable’ for descriptions of coal. In doing so, this process attempts to harmonize definitions suitable for recoverable coal with those of oil and gas for long-term studies of resource use which span the course of a century or more.   Coal reserves have continually dwindled through the early 21st-century despite theoretically favorable market conditions. Future projections for vastly expanded global coal production will need to revisit the heuristic of a vast coal backstop enabled by a dynamic R-P ratio maintained in equilibrium. Examining the relevance of implied market and technology trends can aid in this analysis.   65 Arguing that an equilibrium R-P for coal will continue to inspire confidence in vastly expanded supply for the coming century assumes a substantial fraction of the total resource base can readily be economically mined. This assertion must be supported by evidence on the direction and consequence of prices, technological change in the coal industry, and the process of determining the portion of coal deposits that are economically recoverable. As information on world coal has improved (e.g. learning), the economically recoverable portion of initially assessed deposits has been more accurately identified - always a much smaller quantity than the initial in situ amount recorded as reserves. Reserves are only a fraction of the total potential amount of geologic coal occurrences classified as resources (Chapter 1 provides a detailed McKelvey box diagram to clarify these definitions). Through this chapter, we can distinguish recoverable coal from the broader terms of 'reserves' and 'resources' once a recovery and other economically and socially relevant factors are applied. This definition allows for a contrast between coal in-place and the coal available as a potential source of fuel for economic use. With appropriate caveats about possible changes in demand, technology, economics, and uncertainties due to information quality, observation of the changes in global reserve assessments can indicate the broader dynamics that would empirically constrain the plausible quantities of recoverable coal in multi-decade scenarios.  To inform 21st-century coal reference cases, we can use the main sources compiling global coal assessments to trace and understand the history of a dynamic reserve base over the past three decades: the World Energy Council (WEC) and the German Federal Institute for Geosciences and Natural Resources (BGR). The most prominent trends over this period are rising production (especially in China) and higher prices. Both have more than doubled between 1990 and the early 21st-century. Despite these conditions, reported global coal reserves have continued to decline.  Since coal assessments tend to report physical quantities in mass-based measures, long-term scenarios of primary energy use have relied on secondary sources that convert these metrics. Rogner (1997) and Rogner et al. (2012) provide estimates of the primary energy available in hard coal reserves. These studies also indicate significant declines since 1990 – from as much as 47,000 EJ to around 18,000 EJ.  Using the methodology of these studies to harmonize their estimates of primary energy content with the mass-based units from more recent reports indicates a value on the order of 15,000 to 17,000 EJ is a reasonable upper bound for remaining recoverable coal in today’s long-run scenarios. The figure used in this chapter of 15,300 EJ is consistent with the range provided by recent studies and analyzed in Section 3.3.1 such as Mohr et al. 2015. This smaller value stands in contrast to the 440,000 EJ of primary energy theoretically available in the total geologic coal resources distributed throughout the Earth's crust (BGR 2015, 2014). However, pinpointing the precise quantity of primary energy reflected by hard coal reserve assessments is of subsidiary importance to understanding why these estimates have continued to decline. Long-run studies must capture the limitations of today’s knowledge by considering the potential for dynamic boundaries in definitions of reserves and resources. This requires identifying  66 the general influence of aggregate technological progress and increasing information on our collective understanding of hydrocarbon energy deposits.  Today, coal is the fuel used for more than 40% of the world’s electricity generation, and stands at a 29% share of global primary energy supply (IEA 2016). Coal is also a critical input to steelmaking and industrial processes. The coal that generates electric power is steam or thermal coal, while metallurgical or coking coal is used by industrial operations, and typically of higher quality. Coal is classified by a rank that measures its stage of geologic progression from lignite to anthracite. High ranked coals of anthracite and bituminous have a higher energy density and greater carbon content than lower ranks of sub-bituminous and lignite. This paper applies the BGR (2015) definition of hard coal: sub-bituminous, bituminous and anthracite with energy content greater than or equal 16,500 kJ/kg. To examine the appropriate context for coal in scenarios of the global energy future, this chapter is organized as follows: Section 3.1 addresses the trends in coal supply since 1990 suitable for inclusion in 21st-century reference cases and understands whether total coal resources have been a ‘stock’ that replenished the ‘flow’ of reserves as with oil and gas. Section 3.2 explains why modern coal assessments indicate fewer reserves than vintage reports and understands what these concepts mean for future coal availability curves. Section 3.3 provides a case study to illustrate how legacy assessments have influenced widely used future energy scenarios in the climate change research community. Section 3.4 concludes with a summary and recommendations for integrating future coal use in multi-decade energy system reference cases which provides a context for the work in Chapter 4. 3.1 Reference case trends in coal reserves, resources, production, and prices  Calibration of a relevant 21st-century coal reference case must focus on hard coal reserves, which contain approximately 90 percent of the energy in the world’s coal resource base (Mohr et al., 2015; Rogner et al., 2012; Rogner, 1997). A future coal backstop would rely heavily on global trade in hard coal, or its transformed products, since the economics of lignite encourage consumption close to the site of extraction in regional electricity generation because of the fuel’s low energy density and high water content (BGR, 2015). Therefore, this section focuses on trends in hard coal reserves.  Trends in hard coal reserves, production, and markets since the late 1980s are depicted in Figure 3.1. Reserve and production data in this figure draw from regular reports by the WEC and German Federal Institute for Geosciences and Natural Resources (Bundesanstalt für Geowissenschaften und Rohstoffe, BGR) (WEC 1989, 1992, 1995, 2001, 2004, 2007, 2013, 2016; BGR 2009, 2014, 2015). The International Energy Agency (IEA) annually relays select years of figures from both sources, along with regular updates on production (IEA 2001–2016).  Figure 3.1a plots the range of reported constant US dollar (USD) prices from 1989 to 2016 across 10 major coal market indices (BP, 2015; EIA, 2012; World Bank, 2016) shaded as a gray band. 35                                                      35  Detail on included coal market indices: (a) BP: Northwest Europe Marker Price, US Central Appalachian Spot Price Index, Japan Coking Coal Import, Japan Steam Coal, Asian Market Price; (b) EIA: Bituminous and Anthracite; (c) World Bank: Coal (Australia), thermal GAR, f.o.b. piers, Newcastle/Port Kembla from 2002 onward , 6,300 kcal/kg (11,340 btu/lb), less than 0.8 percent, sulfur 13 percent ash; previously 6,667 kcal/kg (12,000 btu/lb), less than 1.0 percent sulfur, 14 percent  ash, International Coal Report; Coal Week International; Coal Week; Bloomberg; IHS  67 Average benchmark coal prices (red) more than doubled between the early 1990s and the first decade of the 21st-century. These rising average prices were concurrent with a doubling of global hard coal production through 2014 (Figure 3.1b).  Given the heterogeneous domestic conditions for coal markets among regions, the impact of exchange rates is important to note, otherwise interpretations of price are overly conditioned by the United States perspective. Though USD denominated coal markets have declined from 2012-2016, local exchange rates in major exporting nations such as Russia, Colombia, South Africa and Australia held domestic prices steady over this four-year period (IEA 2016). Since many of the production costs in these countries are paid in rubles, pesos, rand and Australian dollars, average national coal market prices through 2016 remained flat or higher than 2012. This context of devaluation allows domestic producers to cover costs, holding their output steady, creating uncompetitive bear market conditions for the relatively expensive US coal industry.   All else equal, conventional economic theory would expect that higher sustained commodity prices (demand) reclassify marginal geologic deposits (supply) as economically recoverable reserves. Yet, since the doubling of coal prices and production in 2000, reserves declined by roughly 15 percent (Figure 3.1c). Reported reserves show a modest increase around 2000: as a decade-long expansion of coal production began, new mines were opened and initial supply contracts were signed, temporarily increasing reported reserves. Once the rate of mining continued to increase, however, total reported reserves declined despite rising market prices. Because long-run global coal reserve and price trends have not moved as expected from simple equilibrium supply-demand assumptions, the conceptual foundation of multi-decade coal resource economics are ripe for revision. Rogner (1997) and Rogner et al. (2012) convert mass-based assessments to energy units and provide important secondary references on coal supply for future energy projections. Rogner et al. (2012) report two-thirds less energy in the hard coal reserve base from the earlier assessment based on BGR (1989) and WEC (1992). This decline in available energy from coal marks a rapid decrease in the global coal R-P ratio from more than 300 to 100 years (Figure 3.1d).  Because of uncertainties in the energy content of recoverable coal, normalized values are provided to understand this decline. 36 WEC reports that the large number of assessed reserves from the late 1980s in Figure 3.1d results from an accidental reclassification of China’s reserves as “proved recoverable” from a previous definition as “proved amount in-place.”                                                        McCloskey Coal Report; (World Bank) Coal (South Africa), thermal NAR, f.o.b. Richards Bay 6,000 kcal/kg  from 2006 onward; during 2002–2005 6,200 kcal/kg (11,200 btu/lb), less than 1.0 percent, sulfur 16 percent ash; years 1990–2001 6390 kcal/kg (11,500 btu/lb) (International Coal Report; Coal Week International; Coal Week; World Bank), Coal (Colombia), thermal GAR, f.o.b. Bolivar, 6,450 kcal/kg, (11,200 btu/lb), less than 1.0 percent, sulfur 16 percent ash from August 2005 onward; during 2002–July 2005, 11,600 btu/lb, less than 0.8 percent sulfur, 9 percent ash, 180 days forward delivery (International Coal Report; Coal Week International; Coal Week; World Bank). 36  Throughout this paper, where physical units of global coal are converted to energy units, the Rogner (1997) methodology of applying average energy content per regional reserves is used to calculate an internally consistent value for harmonized comparison of assessment vintages. This method results in an estimate of 15,300 EJ of primary energy in BGR (2015) reported hard coal reserves. This is slightly lower than the value reported for hard coal reserves by BGR (2015) of 17,390 EJ and this is explained in Section 3.3.1. Note: The Rogner (1997) and Rogner et al. (2012) values in Figure 3.1c and 3.1d are converted to a mass-basis for comparison with IEA, WEC and BGR.   68 Declining reserves over time indicate that a stable R-P ratio for coal is unobserved, and does not provide a workable assumption to support long-run energy scenarios which tap into the larger assessed coal resource base.    Figure 3.1a  Trends in global coal market benchmark prices (BP, 2015; EIA, 2012; World Bank, 2016) – minimum and maximum values indicated by gray range, while red line follows the average of benchmark prices    69   Figure 3.1b  Annual hard coal production as reported by IEA Coal Information Reports, WEC and BGR (indexed to IEA reported values for 2001 in mass units); note y-axis break   Figure 3.1c  Coal reserves in mass units from successive WEC and BGR reports indexed to WEC (2001) – the WEC-BGR synthesis reported by Rogner (1997), and the updated Rogner et al. (2012) normalized to WEC (2001) using harmonized energy-to-mass units; note y-axis break   70  Figure 3.1d   Reserve-to-production ratio for global coal (mass-basis) – Rogner (1997) and Rogner (2012) illustrate the distinction between legacy and modern assessments; note y-axis break 3.1.1 Can we use a learning hypothesis to characterize the coal reserve-resource boundary?  Chapter 2 analyzed the learning-by-doing model for future fossil energy supply in the context of cumulative resource availability curves. A similar ideation is applied to structure coal resources for IAMs. As noted in Section 2.3.2, GCAM scenarios with a learning rate of zero or lower were empirically constrained to all known conventional oil resources.  An original aggregate assessment for global coal resources was established by the 1913 International Congress of Geologists (IGC) in Toronto. Since then, the World Energy Council (WEC), previously the World Power Council (WPC), and the German Federal Institute for Geosciences and Natural Resources (Bundesanstalt für Geowissenschaften und Rohstoffe, BGR) have maintained regular publications on total gl