Integrating SHAP Interpretability and Ensemble Machine Learning To Optimize Conductivity in Flexible Electronic Polymers

Conductive polymers are essential materials for flexible electronics and energy storage, yet their electrical performance is highly sensitive to multivariate synthesis parameters that are difficult to optimize experimentally. Here, we develop a machine learning framework that accurately predicts polymer conductivity based on nine key synthetic variables. Among five tested algorithms, EXtreme Gradient Boosting (XGBoost) achieves the highest predictive accuracy (test R2 = 0.933), while Light Gradient Boosting Machine (LightGBM) shows strong performance with minimal overfitting. SHapley Additive exPlanations (SHAP) analysis reveals that dopant species is the dominant determinant of conductivity (59.85%), followed by oxidant species (17.5%) and polymer type (8.5%). The model further uncovers critical nonlinear relationships, including conductivity maximization at 30–60 °C, a sharp decline beyond a 2 mol/L dopant concentration, and negligible dependence on reaction duration after equilibrium. Notably, optimal conductivity arises under synergistic conditions of elevated reaction temperature (>80 °C) and low oxidant concentration (<0.2 mol/L), which promote molecular alignment and enhance charge carrier mobility. Our findings highlight that dopant-driven acid–base doping strength, rather than classical acid–base doping strength, serves as the primary factor controlling carrier transport in conductive polymers. This insight provides the design principles for organic electronic materials and demonstrates how machine learning can accelerate materials discovery and optimization.