Toward balanced proton conductivity and water uptake: interpretable condition-aware machine learning for sulfonated polyimide membrane design

Balancing proton conductivity (σ) and water uptake (WU) in sulfonated polyimide (SPI) membranes requires distinguishing shared ionization effects from target-specific responses to chain structure and measurement conditions. Here, condition-resolved literature datasets and interpretable machine learning are used to examine these structure–property relationships across 1,216 σ and 863 WU records, each with an independently frozen 20% test set. Global repeat-unit descriptors (F1), local structural environments (F2), and ionic/compositional variables together with test conditions (F3) are compared to separate their incremental contributions. The selected models achieve frozen-test R² = 0.829, RMSE = 0.02796 S cm⁻¹, and MAE = 0.01376 S cm⁻¹ for σ, and R² = 0.760, RMSE = 12.86 percentage points, and MAE = 7.721 percentage points for WU. Adding fingerprints gives similar cross-validated R² gains for the two targets (0.046 for both), whereas the joint addition of physical and condition information produces larger gains (0.599 and 0.140). SHAP analysis identifies the ion exchange capacity (IEC)–relative humidity (RH) interaction as a shared leading contributor. Humidity and thermal conditions are more prominent for σ, whereas IEC, degree of sulfonation (DS), and chain composition are more prominent for WU. These patterns are consistent with hydration-enabled proton transport and with structural control of chain packing, free volume, and water sorption. They explain why increasing sulfonation alone cannot ensure balanced membrane performance and support condition-fixed, dual-property design of SPI membranes.