Interpretable Machine Learning-Based QSPR Modeling for Predicting Hydrogen Adsorption Performance of Ester Oil Impregnating Agents

Abstract

Understanding the relationship between molecular structures and hydrogen adsorption performance is essential for the rational screening of ester oil impregnating agents for high-voltage capacitor applications. Herein, an interpretable machine learning-based quantitative structure–property relationship (QSPR) framework was developed to predict hydrogen adsorption performance of ester oils. A data set of 56 ester molecules was established through experimental characterization, and robust molecular descriptors were identified from a high-dimensional descriptor space using a bootstrap-assisted least absolute shrinkage and selection operator (LASSO) stability selection strategy. Multiple machine learning models were evaluated, with the optimized XGBoost model achieving excellent predictive performance (R2 = 0.940, RMSE = 10.316, and MAE = 8.707). Explainable analyses based on SHapley Additive exPlanations (SHAP) and partial dependence plots revealed key molecular features and nonlinear structure–property relationships governing hydrogen adsorption behavior. This study provides an interpretable data-driven strategy for screening ester-based functional liquids and offers molecular insights for optimizing capacitor impregnating materials.