Accelerated Screening of Electrolyte Solvents for High Safety Batteries Using Machine Learning

We present a high-throughput screening approach to identifying safer nonaqueous solvents to replace or modify the flammable, carbonate-based solvents currently used in commercial batteries. We trained graph neural networks using the DIPPR 801 database and used them to predict flash point, melting point, dielectric constant, and liquid viscosity directly from molecular structure. By screening over one billion molecules, aggregated from the Mcule catalog and GDB-13, we identify thousands of promising candidates that are dynamically stable and easily synthesizable. Among the candidates is one known battery additive and many substances previously overlooked for battery applications but used, for example, as food additives and fragrances. Our two-pronged approach, screening databases covering both catalogs of existing molecules (Mcule) and chemical enumeration of hypothetical molecules (GDB-13), yielded a diverse range of readily available and de novo candidates.