Electrolytomics: A Unified Big Data Approach for Electrolyte Design and Discovery

Abstract

Electrolyte discovery is the bottleneck for developing next-generation batteries. For example, lithium metal batteries (LMBs) promise to double the energy density of current Li-ion batteries (LIBs), while next-generation LIBs are desired for operations at extreme temperature conditions and with high voltage cathodes. However, there are no suitable electrolytes to support these battery chemistries. Electrolyte requirements are complex (conductivity, stability, safety), and the chemical design space (salts, solvents, additives, concentration) is practically infinite; hence, discovery is primarily guided through trial and error, which slows the deployment of such next-generation battery chemistries. Inspired by artificial intelligence (AI)-enabled drug discovery, we adapt these machine learning (ML) approaches to electrolyte discovery. We assemble the largest small molecule experimental liquid electrolyte ionic conductivity data set and build highly accurate ML and deep learning models to predict ionic conductivity across a wide range of electrolyte classes. The developed models yield results similar to those of molecular dynamics (MD) simulations and are interpretable without explicit encoding of ionic solvation. While most ML-based approaches target a single property, we build additional models of oxidative stability and Coulombic efficiency and develop a metric called the electrolyte score ( eScore ) to unify the predicted disparate electrolyte properties. Deploying these models on large unlabeled data sets, we discover distinct electrolyte solvents, experimentally validate that the electrolyte is conductive (>1 mS cm –1 ), stable up to 6 V, supports efficient anode-free LMB, and even LIB cycling at extreme temperatures. Our work marks a significant step toward efficient electrolyte design, accelerating the development and deployment of next-generation battery technologies.