Geometric Deep Learning Predicts Electrostatic Potential for High-Throughput Electrolyte Design

Abstract

Developing advanced electrolytes is pivotal for the realization of next-generation high-energy lithium batteries. However, the combinatorial explosion within the chemical space of candidate molecules renders both traditional experimental and standard density functional theory (DFT)-based high-throughput screening practically and computationally prohibitive. To overcome this bottleneck, we developed an automated high-throughput virtual screening workflow. By systematically constructing a highly diverse library comprising over 129,026 electrolyte molecular variants and employing SchNet, a threedimensional geometry-aware deep learning architecture, we achieved rapid and accurate predictions of molecular surface electrostatic potential (ESP), including both ESPmax and ESPmin as key extrema descriptors for electrolyte screening. Furthermore, by leveraging interpretable machine learning and multi-dimensional statistical analysis, we explored the expansive chemical space to elucidate the competitive and synergistic effects of molecular skeletons, substitution chain lengths, fluorination, and etherification. Specifically, the skeleton defines the baseline ESP distribution, the substituent chain length regulates the spatial attenuation of polar effects, and fluorinated/etherified substitutions tune the electronic pushpull balance that governs solvation capability and oxidative stability. On this basis, electrolyte candidates can be selected according to state-specific ESP requirements: cationic frameworks should be biased toward high ESPmax to enhance oxidative stability and weaken excessive cation-solvent interactions; anionic frameworks should retain sufficiently deep ESPmin to promote Li + coordination and salt dissociation; and neutral solvents should maintain an intermediate ESP window that balances salt solubility, desolvation kinetics, and high-voltage stability. These rules further suggest that the fluorine/oxygen ratio should not be optimized by a universal standard, but should be adjusted according to the charge state and functional role of each electrolyte component. This work provides a data-driven framework for translating ESP extrema into actionable molecular design rules, thereby paving a rational pathway for future advanced battery electrolytes.