UBC Theses and Dissertations
Exploring heterogeneous ice nucleation by molecular simulations and machine learning Soni, Abhishek
Heterogeneous ice nucleation (HIN) has applications in climate science, nanotechnology, and cryopreservation. Pure water does not freeze homogeneously above ∼ −39◦C. Ice nucleation on the Earth’s surface or in the atmosphere usually occurs heterogeneously involving foreign substrates (ice nuclei). Foreign substrates can be mineral dust, soot, or bacteria. Experiments identify good ice nuclei but lack sufficient microscopic resolution to answer the basic question: What makes a good ice nucleus? We employ molecular dynamics (MD) simulations to address this question and investigate microscopic mechanisms of HIN. We consider substrates that display a wide range of ice-nucleating abilities. These are α-alumina, kaolinite, gibbsite, hematite, mica, β-AgI, and K-feldspar. The α-alumina (0001) surface nucleates the (0001) basal plane of hexagonal ice (Ih), whereas the kaolinite (001) plane stabilizes the (10¯10) prism face. Mica (001) surfaces nucleate ice by an unusual mechanism, the ice nucleus consists of ordered layers of hexagonal and cubic ice. The primary prism face (10¯10) of β-AgI nucleates the prism face of Ih, which significantly enhances current understanding of HIN by AgI. All substrates mentioned above are good ice nuclei in experiments. HIN is not observed for gibbsite, hematite, and boehmite, again consistent with their experimental classification as poor ice nuclei. Experimentally, K-feldspar is an excellent ice nucleus, but our simulations show no icelike order for its (001), (010), and (100) planes. This suggests that some surface rearrangement or defect is responsible for HIN by K-feldspar. Simulations of multiple surfaces allow us to identify important features of a good ice nucleus. We develop a two-dimensional (2D) lattice description of water ordering in the surface layer. Basal and prism faces ice bilayers decompose into two, 2D lattices (triangular for basal, rectangular for prism), and a substrate must stabilize both to promote HIN. We demonstrate how lattice match, hydrogen bonding, and atomistic surface morphology crucially influence HIN efficiency. We employ machine learning to predict the likelihood of HIN using surface features, together with liquid water properties local to the surface, which can be obtained from short simulations. This achieves 89% accuracy in the classification of surfaces as good or bad ice nuclei, with water properties the dominant factor.
Item Citations and Data
Attribution-NonCommercial-NoDerivatives 4.0 International