We introduce a novel approach that combines interpretable quantitative stochastic machine learning with multiscale analysis to predict the macroscopic thermal conductivity of graphene-enhanced polymer nanocomposites. Our method effectively addresses uncertainties in input parameters across meso and macro scales within a bottom-up modeling framework. By integrating Representative Volume Elements (RVE) with traditional Finite Element Modeling (FEM), we calculate the effective thermal conductivity through homogenization. We further enhance predictive modeling by employing the XGBoost regression tree method. To clarify the influence of input variables on model outcomes, we incorporate SHapley Additive exPlanations (SHAP) and Local Interpretable Model-agnostic Explanations (LIME). Additionally, sensitivity analyses are conducted to assess the impact of design parameters on material properties. This comprehensive approach improves both global and local interpretability, clarifying feature interactions in data-driven and physical models. It reduces the reliance on extensive analytical modeling and simulations, enhancing prediction accuracy and significantly lowering computational costs. Our method holds significant promise for the design of new composite materials optimized for thermal management.