Open Collections

Scheiber, Hayden Owen

University of British Columbia

The lithium halide salts may at first appear to be simple chemical systems. However, previous research indicates lithium halides have complex and poorly understood crystallization pathways from aqueous solutions. While lithium halides exist in the rocksalt crystal structure under ambient conditions, common lithium halide classical force fields more often predict wurtzite as the stable structure. This failure severely limits their application in molecular simulations of crystallization. Research in this thesis focuses on presenting new results and computational methodology to better understand the structural chemistry of both lithium halides and alkali halides in general. Density functional theory together with classical force fields are employed to examine the relative stability of candidate crystal structures for lithium halides and produce accurate theoretical reference data. Dispersion interactions are shown to play a key role in the stability of rocksalt over closely competing crystal structures. Classical models can be corrected in their structural predictions by scaling up the strength of dispersion, indicating a pathway towards better lithium halide force fields. Convolutional neural networks are used for the structural classification of simulated alkali halides. The neural networks are trained on a large data set generated from molecular dynamics simulations of alkali halides across a range of temperatures. Time convolution filters out short-lived structural fluctuations. The structure classifiers are shown to be accurate in bulk phase simulations, then demonstrated on crystallization of model alkali halide systems from the melt. The neural network classifiers are implemented in a melting point calculation algorithm for model binary salts. Finite size effects are characterized, then melting points of alkali halides using common rigid-ion interaction potentials are calculated and discussed. The methods developed throughout the research are employed for the optimization of pairwise lithium halide force fields, fitted to reference data using a reinforcement learning approach, Bayesian optimization. Limitations on the Coulomb Lennard-Jones potential form are uncovered, which do not appear to hold for the more flexible Coulomb Buckingham potential form. By introducing advanced computational methodology, this research reveals the inherent structural complexity of lithium halides and emphasizes the importance of considering structural landscapes during classical forcefields construction for molecular simulation.

Text; Moving Image

2022-12-22

Attribution-NonCommercial 4.0 International

http://hdl.handle.net/2429/83478

Chemistry

University of British Columbia

UBCV

http://creativecommons.org/licenses/by-nc/4.0/