Machine Learning Prediction of Transport Properties in Amorphous Polymer Electrolytes Using Chemically Informed Structural Descriptors

Amorphous polymer electrolytes are promising candidates for next-generation lithium batteries. They offer improved safety and mechanical stability compared to conventional liquid electrolytes. However, predicting their transport performance remains challenging. Ionic conductivity and diffusivity depend on complex, coupled interactions between polymer structure, chemistry, and ion dynamics. This work leverages machine learning to predict four key transport properties in amorphous polymer electrolytes using a molecular dynamics data set of more than 6,000 polymer electrolyte systems. These are ionic conductivity, bis-(trifluoromethanesulfonyl)-imide (TFSI) diffusivity, lithium-ion diffusivity, and polymer diffusivity. To enhance predictive capability and preserve mechanistic interpretability, we engineered chemically and structurally meaningful features from Simplified Molecular-Input Line-Entry System (SMILES) representations. Selected features include elemental composition, functional group densities, polarity descriptors, and polymer flexibility metrics. Three modeling frameworks were trained and evaluated under 5-fold cross-validation, including random forest, gradient boosting, and neural networks. The neural network achieved the highest predictive performance across all four targets and exceeded prior benchmarks on three of them. It also exhibited substantially greater robustness to data partitioning than the tree-based baselines. Feature importance analysis was performed using both impurity-based rankings and Shapley additive explanations. Polymer backbone flexibility, polarity, and ionic environment were found to dominate transport behavior. Directional attribution further identified primary and secondary amines and carboxylate groups as specific negative contributors to lithium-ion mobility. The results of this work demonstrate that chemically informed structural descriptors can effectively predict electrolyte performance while yielding mechanistic insight. This enables accelerated screening and design of high-efficiency polymer electrolytes for future energy storage applications.