AuTopology—End to End Force Field Parametrization Workflow for Polymer Electrolytes Using Machine Learning

Solid polymer electrolytes (SPEs) are seen as promising alternatives to conventional liquid electrolytes in lithium battery systems due to their low density, mechanical compliance, and low flammability but are challenged by lower ionic conductivity. Molecular dynamics (MD) simulations can be used to guide the design of novel SPEs by allowing quantitative determination of separable anion and cation diffusions as well as local solvation environments and correlated ion motions. Classical potential MD simulations update molecular conformations by the net force on each atom due to covalent and nonbonded interactions. However, these classical potentials are often not well defined for novel systems as they require materials- and local environment-specific parameters such as unique bond stiffnesses which are either meticulously hand-tuned across decades or unchangeable due to proprietary licenses.<br/><br/>In this work, we explore the effects of anharmonic bonded interactions on ionic solvation and conductivity mechanisms in polymer and liquid electrolyte systems. An in-house, machine learning-based workflow, named AuTopology, was used to autonomously learn the interatomic potential parameters for prototypical solid and liquid electrolyte systems at experimentally-relevant concentrations with both harmonic OPLS-based and anharmonic PCFF+-based models from DFT forces as training data. These simulations of tens-of-thousands of atoms were allowed to run for hundreds of nanoseconds to determine the individual anion and cation diffusivities and resulting conductivities. Using this framework and an in-house database of molecular conformations, we have been able to reproduce wB97XD3-level DFT forces from trained OPLS force fields to within 5.5 kcal/mol-A. Correlation-based conductivities between the OPLS and PCFF+ models for polyethylene oxide (PEO) and liquid carbonate systems with different lithium salts were found to match legacy parameterizations to the same order of magnitude while matching or improving conductivity predictions relative to experiments.