Bayesian Optimization for Efficient Parameterization of LiX Forcefields for Molecular Dynamics

Hayden Scheiber and Grenfell Patey

University of British Columbia

In many ways, the accurate molecular dynamics simulation of lithium halides is a deceptively difficult problem. At first glance, the binary lithium salts appear to be about as simple a chemical system as one can conjure up. However, recent theoretical investigations of the landscape of lithium halide crystal structures paint a more complex picture. Due to the relative ion sizes of lithium vs its counter-anion, these salts have multiple low-energy crystal structures that compete to complicate their structural chemistry. A recent investigation by our group revealed that the relative energies of these crystal structures are highly dependent on the details of the attractive and repulsive interactions between ions, particularly the strength of London dispersion. Therefore, modeling the interaction potentials of lithium halides proves to be a particularly difficult problem compared with other binary salts. Many of the classical forcefields currently available for lithium halides get the lowest-energy crystal structure wrong, particularly for LiBr and LiI. Furthermore, even those forcefields that get the correct low-energy structure right tend to yield incorrect relative energies for the competing low-lying metastable structures. Furthermore, the melting points for most available lithium halide forcefields are far from experiment, indicating a recurring issue with the high-temperature thermodynamic behaviour of the solid phase, liquid phase, or both.

In the work presented here, we use multiple experimental structural, energetic, and thermochemical data points combined with a set of theoretical high-accuracy lithium halide metastable crystal energies and geometries as targets for a computationally efficient Bayesian Optimization scheme to optimize pairwise classical forcefield parameters for lithium halide molecular dynamics simulations. We validate the resulting forcefields in finite temperature molecular dynamics simulations to test their energetics and most stable crystal structure at high temperature. We also calculate model melting points and liquid properties, comparing our results with the available experimental data.

The figure below is an outline for the procedure used in this work. LiX Bayesian Optimization Overview

Back to List of Abstracts