End-To-End Learning of Classical Interatomic Potentials for Benchmarking Anion Polarization Effects in Lithium Polymer Electrolytes

Abstract

Solid polymer electrolytes are an exciting solution for safe and stable solid lithium electrode battery systems but are hindered by low ionic conductivity and low lithium transference. All-atom molecular dynamics simulation has become an invaluable tool to probe lithium diffusion mechanisms and accelerate the discovery of promising polymer chemistries. Because of their low computational cost and despite their approximate nature, only classical interatomic potentials can access the time and length scales for appropriate statistics of polymer kinetics. Machine learning (ML) potentials trained end-to-end on ab initio data have proven more accurate but cannot be scaled to the necessary time- and length- scales yet. Historical approaches to parametrize classical force fields have been incremental, reliant on a manual combination of top-down and bottom-up fitting, and are often paywalled and hard to reproduce. We introduce a computational learning workflow to predict classical interatomic potential parameters using quantum mechanical computations as training data that combines the automation and end-to-end fitting of ML with traditional class 1 and class 2 functional forms. The fitting strategy produced potentials whose simulations improved the accuracy of lithium coordination environments, diffusivities, and conductivities relative to experimental approaches when compared to both naive and hand-tuned parameters for liquid and solid organic electrolyte systems. We show that chemistry-informed regularization is necessary to constrain predicted parameters in order to reproduce experimental solvation and kinetic properties. Finally, we explore the limitations of nonpolarizable, fixed point-charge schemes in describing electrolyte anions and compare the effects of two alternative schemes to fit point-charge distributions. The two strategies result in distinct lithium coordination mechanisms and highlight that closest parity to DFT forces and energies does not correlate to correct trends with lithium salt concentration in kinetic and solvation properties for fixed-point-charge classical interatomic potentials.