Prediction and uncertainty analysis of thermal-hydraulic parameters in bentonite-sand mixtures using TabPFN

摘要: Thermal conductivity (l) and hydraulic conductivity (k) are jointly governed by the microstructure of compacted bentonite. However, existing studies predict these parameters independently, ignoring their coupling effects. This poses challenges to the predictive accuracy and uncertainty analysis. To address this limitation, a unified interpretable framework based on the Tabular Prior-Data Fitted Network (TabPFN) was developed for the simultaneous prediction of l and k in compacted bentonite-sand mixtures. Heterogeneous datasets of compacted bentonites were established, including a l-dataset (1,936 instances and 8 inputs) and a k-dataset (896 instances and 12 inputs). TabPFN exhibited superior prediction performance compared to BNN, achieved R² values of 0.970 for l and 0.957 for k. SHapley Additive exPlanations (SHAP) analysis identified the saturation (S_r) and compacted dry density (r_d) as the predominant predictors of l and k predictions, respectively. Monte Carlo simulations further revealed that l reached a maximum value of 1.79 W/mK (95% confidence interval of 1.65-1.81) at the rd of 1800 kg/m3, reflecting the combined effects of compaction and saturation, whereas k decreased with increasing rd. This study provides a robust and interpretable machine learning framework with uncertainty quantification capability for the joint prediction of thermo-hydraulic properties in bentonite barrier materials, offering a practical tool for the design and long-term performance assessment of bentonite barriers in geological repositories.