Explainable machine learning for multiscale thermal conductivity modeling in polymer nanocomposites with uncertainty quantification

Graphene-based polymer nanocomposites show great potential for thermal management, but accurately predicting their thermal conductivity remains challenging due to multiscale structural complexity and parameter uncertainty. We propose an innovative approach integrating interpretable stochastic machine learning with multiscale analysis to predict the macroscopic thermal conductivity of graphene-based polymer nanocomposites. Our bottom-up framework addresses uncertainties in meso- and macro-scale input parameters. Using Representative Volume Elements (RVEs) and Finite Element Modeling (FEM), we compute effective thermal conductivity through homogenization. Predictive modeling is powered by the XGBoost regression tree-based algorithm. To elucidate the influence of input parameters on predictions, we employ SHapley Additive exPlanations (SHAP) and Local Interpretable Model-agnostic Explanations (LIME), providing insights into feature interactions and interpretability. Sensitivity analyses further quantify the impact of design parameters on material properties. This integrated method enhances prediction accuracy, reduces computational costs, and bridges data-driven and physical modeling, offering a scalable solution for designing advanced composite materials for thermal management applications. • A multi-scale framework is developed to predict the thermal conductivity of graphene- based polymer nanocomposites, integrating meso- and macro-scale analyses for comprehensive material characterization. • SHapley Additive exPlanations (SHAP) and Local Interpretable Model-agnostic Explanations (LIME) provide actionable insights into the impact of design parameters on material performance through sensitivity analysis. • The Explainable machine learning approach bridges data-driven and physics-based modeling, minimizing reliance on complex analytical simulations in material design. • The proposed methodology significantly enhances predictive accuracy while reducing computational costs compared to traditional stochastic multi-scale modeling techniques.