Integrating experimental data and machine learning to predict collapse potential of loessic and lacustrine-alluvial soils using suction and SWCC parameters.

This study investigates the influence of geological characteristics and Soil-Water Characteristic Curve (SWCC) parameters on the collapse potential (CP) of unsaturated soils from loessic and lacustrine-alluvial deposits in Iran. Laboratory experiments-including SWCC determination, suction-controlled modified consolidation tests, and microstructural analyses (SEM and XRF)-were conducted on samples reconstructed to match natural moisture and in-situ density. The results show that matric suction significantly affects compressibility and collapse behavior, with higher suction and Pressure at wettings increasing CP. SEM observations revealed microstructural rearrangements and denser particle packing consistent with hydraulic and mechanical responses, reflecting the influence of depositional environment and mineralogy. To enhance predictive capability, machine learning models were trained on the integrated dataset, with Gradient Boosting achieving the highest predictive performance (R² = 0.859, RMSE = 1.273). Permutation importance analysis identified Pressure at wetting, matric suction, and saturated volumetric water content (θs) as the most influential mechanistic predictors. These findings demonstrate a strong quantitative association between suction-controlled SWCC parameters and collapse potential, indicating that the incorporation of hydraulic variables significantly improves predictive accuracy compared with conventional index-based approaches. The proposed integrated experimental-hydraulic-machine learning framework provides a reliable preliminary tool for collapse assessment and strengthens the mechanistic basis for geotechnical design in unsaturated collapsible soils.