Interpretable Machine Learning for Thermomechanical Property Prediction of Geopolymer-Solidified Soils under Subzero Curing Conditions

Geopolymers have emerged as a promising eco-friendly alternative to conventional cement for soil stabilization in cold regions, offering both reduced carbon emissions and enhanced engineering performance. However, geopolymer-solidified soils (GSSs) exhibit significant degradation in thermal conductivity and mechanical strength subjected to subzero curing environment, posing challenges to practical applications. While laboratory testing remains essential for evaluating these properties, traditional experimental approaches face limitations such as prolonged durations, high resource demands, and difficulties in isolating multifactorial influences. To address these constraints, this study leveraged machine learning techniques, developing eight predictive models based on 1,440 experimental data sets to accurately estimate the thermal and mechanical behavior of GSS under subzero temperature curing. Utilizing the shapley additive explanations (SHAP) interpretability method, the research further identified and quantified the relative contributions of key input parameters governing GSS performance. Among the tested algorithms, the eXtreme Gradient Boosting (XGBoost) model demonstrated exceptional predictive accuracy (average R2 = 0.95) for thermal conductivity, unconfined compressive strength (UCS), shear strength, cohesion, and internal friction angle, with cohesion predictions showing the highest error margins, while thermal conductivity estimates were most precise. SHAP analysis revealed curing age as the dominant factor for thermal conductivity, cohesion, and internal friction angle (feature importance ≥0.39), whereas curing temperature (importance = 0.48) and normal stress (importance = 0.61) were critical for UCS and shear strength, respectively. These findings provide a robust data-driven framework for optimizing GSS mix designs in cold environments, ultimately supporting more sustainable and efficient geotechnical construction practices in permafrost and seasonal freeze–thaw regions.