Physically Grounded Descriptors for Machine-Learning-Guided Diluent Design in Lithium Metal Batteries

Abstract

Localized high-concentration electrolytes require diluents that avoid direct Li+ coordination and preserve Li+-anion aggregates while remaining miscible with the primary solvent. Rational discovery has remained difficult because these competing requirements arise from chemically distinct interactions that have lacked computable molecular design coordinates. Fluorination has therefore been widely used to suppress electron donation, but the environmental persistence and synthetic burden of fluorinated molecules motivate alternative strategies for controlling diluent functionality. Here, we resolve the key interaction requirements of diluents using two complementary descriptors: BF3 affinity for specific electron-pair donation and the molecular polarity index (MPI) for distributed nonspecific electrostatic interactions. These descriptors reproduce established donor-number and Kamlet–Taft π* trends and, more importantly, distinguish noncoordinating yet miscible diluents from solvating and immiscible solvents. Coupling this interaction-based framework with uncertainty-aware machine learning enabled screening of 1.32 million nonfluorinated molecules and identified 368 candidates while recovering false negatives and excluding false positives near the decision boundaries. Newly selected aromatic hydrocarbons, including 1,2,4-trimethylbenzene and 2-ethyltoluene, sustained stable cycling in Li-metal full cells for more than 200 cycles, whereas a descriptor-excluded analogue underwent rapid capacity decay. These results establish specific coordination and distributed electrostatics as predictive design coordinates for diluent discovery.