Ensemble learning predictive model for thermal conductivity integrating priori physical knowledge and interpretive analysis

Accurate determination of constitutive relationships for hydrothermal parameters is crucial for multi-field coupling studies. However, the presence of multi-scale influencing factors and highly nonlinear response behaviors makes it difficult for existing models to accurately capture coupling effects, heat transfer pathways, and transport mechanisms. Consequently, the development of reliable and robust parameter models remains a significant challenge. To address these issues, a physics-informed ensemble learning (PIEL) model was introduced, combining a priori physical knowledge with interpretive analysis. Using soil thermal conductivity (λ) as a case study, the PIEL model’s accuracy, robustness, and physical consistency were comprehensively evaluated. Sensitivities, response patterns, and coupling effects in the decision-making process were visualized using Shapley additive explanations (SHAP) and partial dependence plots (PDPs). The results demonstrate that the proposed ensemble learning framework effectively captures the complex nonlinear coupling behavior governing thermal conductivity, improving prediction accuracy by a factor of 3 to 6 compared to traditional models. By incorporating a priori heat-transfer knowledge, the PIEL model effectively prevents physically implausible predictions-a common limitation of purely data-driven approaches-thereby substantially enhancing the physical consistency of model outputs. Among the evaluated methods, the physics-informed extreme gradient boosting (PXGBoost) model optimized via the sparrow search algorithm exhibited the highest predictive accuracy and robustness. Sensitivity analyses and response patterns revealed by SHAP and PDPs are consistent with established heat-transfer theory, further validating the physical interpretability of the PIEL-based decision-making process. Furthermore, the optimal predictors identified through cumulative SHAP values significantly outperform traditional parameter analysis methods, enabling the development of more computationally efficient and accurate simplified models. The PIEL model, integrating physical knowledge, represents a powerful tool for geotechnical parameter prediction and paves the way for advancing AI-enabled hydrothermal simulations.