Soaring Chemical Space of Battery Electrolytes Using Universal Machine Learning Potential

Li-ion batteries, widely used in electronic devices, electric vehicles, and aviation, demand high energy density, fast charging capabilities, and broad operating temperature ranges. Computational approaches combined with experimental design have gained increasing attention for electrolyte development. However, the inherent complexity of electrolytes—arising from their diverse compositions and varying proportions—poses a significant challenge. Classical molecular dynamics often fails due to inaccuracies in force field parameters, while ab initio calculations are limited by prohibitive computational costs. Machine learning molecular dynamics (MLMD) offers a promising alternative, combining efficiency with ab initio accuracy. However, its potential has been hindered by the transferability limitations of machine learning potentials (MLPs). In this work, we develop a universal machine learning potential (uMLP) for electrolytes by generating randomly composed electrolytes and employing an iterative training approach to collect representative datasets, effectively overcoming these limitations. The uMLP enables accurate computation of key properties, including density, solvation structure, viscosity, ionic conductivity, and operating temperature range, for a broad range of electrolytes through MLMD simulations. Furthermore, coordination dynamics analysis of Li+ , by quantifying the coordination lifetime (¯τ), provides a direct, quantitative measure of solvation strength. Shorter ¯τ, indicative of weak solvation, correlates with faster ion transport and higher ionic conductivity. The uMLP for electrolytes facilitates the prediction and optimization of electrolyte properties, offering a powerful tool for electrolyte design.