Accelerating Electrolyte Discovery Using Generative Models and Molecular Dynamics Simulations

Abstract

Operating lithium-ion batteries (LIBs) at high voltages and low temperatures demands electrolyte solvents with highly tailored electronic properties. Because traditional discovery is limited by historical solvent libraries and human intuition, expanding the accessible chemical space requires advanced computational methods. Here, we deploy the generative machine learning model G-SchNet, trained on the QM9-GCDQE database, to invert the design process for three designed targeted electrolyte environments: general-purpose, high-voltage, and low-temperature. We filtered the generated molecular space using a multistage framework based on validity, uniqueness, and novelty, followed by DFT and molecular dynamics (MD) validation. We identified a computationally promising candidate, trifluoro((fluoromethoxy)methoxy)methane, which is a fluorinated ether that exhibits computed transport properties enhanced relative to a conventional EC/DEC electrolyte under identical simulation conditions. This workflow illustrates the capacity of generative models to identify computationally promising electrolyte candidates that warrant further computational and experimental investigation.