Data Driven Screening of Electrolyte Matrix–Ion Pairs with Enhanced Seebeck Coefficients in Thermoelectrochemical Cells

Here we report a machine learning based framework to predict the seebeck coefficient of ionic thermoelectric materials composed of matrix–ion donor combinations for thermoelectrochemical applications. A dataset of 73 ionic thermoelectric samples was compiled from published literature, and molecular descriptors for both matrix materials and ion donors were extracted using the RDKit toolkit. Recursive feature elimination was employed to identify an optimal subset of 12 descriptors, enabling accurate prediction with minimized feature parameters. Interpretable SHAP and correlation analyses identified FractionCSP3 and molecular weight as key determinants of thermoelectric performance. Among fifteen evaluated models, Decision Tree–based methods Decision Tree, Extra Trees, Random Forest achieved the highest predictive accuracy and were combined in a weighted ensemble to robustly predict seebeck coefficients for 100 previously unseen matrix–ion donor systems. The ensemble results identified several high-performance candidates, particularly polyurethane-, cellulose-, and polyvinyl-alcohol-based systems, with predicted seebeck coefficients reaching up to 38.73 mV K⁻¹. This study demonstrates that an interpretable machine-learning model with reduced feature dimensionality can effectively accelerate the discovery of high-performance ionic thermoelectric materials for wearable thermoelectrochemical applications.