Exploring deep-learning and equivariant machine-learning models for molecular dynamics of aqueous Li–Mn–Cl battery electrolytes

This annual report summarizes the work I have done so far during my tenure at ARL as a Distinguished Postdoctoral Fellow under Contract no. W911NF-19-2-0186 in the Battery Sciences Branch. My work so far has centered on learning the theory and tools of my new field and conducting preliminary computational studies of aqueous zinc battery electrolytes. I first present my initial efforts to develop machine-learning (ML) models to accelerate molecular dynamics (MD) simulations of aqueous Li₂ZnCl₄ · x H₂O and Li₂MnCl₄ · x H₂O electrolytes. These initial exploratory ML models predicted unphysical behavior and produced unstable MD simulations. I then discuss how I simplified my approach to focus on models of bulk water in order to better diagnose issues with the ML training procedures. I share initial results from my studies of bulk water and investigate additional families of ML models that lead to more stable MD simulations. My ultimate goal is to develop a reproducible protocol for training ML models that function as fast and generalizable interatomic potentials for electrolyte systems like Li₂MnCl₄ · x H₂O, or those with even more complex salt compositions.

Report