Unlocking the Chemical Space for Rechargeable Batteries with a Generative Solvent Design System

Electrolyte discovery for rechargeable batteries today relies on heuristic trial-and-error or high-throughput screening of existing molecules. Here, we introduce a Generative Solvent Design System (GSDS) that integrates a graph-based deep molecular generator with machine learning (ML) property predictors to design rechargeable battery solvents de novo. We enable this by constructing a battery-specific prior data set (Batt-SLM, 115,756 molecules) and fine-tuning a graph-based molecular generator using physics-informed ML surrogates for redox potential, viscosity, melting point, donor number, and dielectric constant. We validated the performance of GSDS on the rediscovery of both fluorinated and phosphorus-containing compounds not seen during training. This allows us to propose application-specific candidates (top 0.2 ‰) for alkali metal batteries─fluorinated diluents and nonfluorinated weakly solvating electrolytes─that pass a posterior verification funnel including property evaluation, synthetic accessibility, candidate prioritization, and literature checks. We conclude that GSDS establishes a tractable solvent design layer of a broader electrolyte-design framework and can be expanded toward salt-aware, interface-informed, and mixture-included optimization for next-generation rechargeable batteries.