Improved prediction of bimodal soil-water characteristic curves for expansive soils from mercury intrusion porosimetry using machine learning

The soil-water characteristic curve (SWCC) occupies a central position in unsaturated soil mechanics, yet for expansive soils its direct experimental characterization is challenging by the concurrent demands of resolving a bimodal pore material and tracking suction-induced volumetric deformation across a measurement process that routinely exceeds several months. To avoid these constraints, the present study develops a microstructural-machine learning (ML) framework for predicting bimodal SWCCs over the entire suction range, including low to high swelling potential expansive soils. Mercury intrusion porosimetry (MIP) was conducted on compacted specimens, and the recovered pore-throat size distributions were converted to equivalent matric suction profiles through the Young-Laplace equation; systematic corrections (for suction, s , and degree of saturation, S ) were then applied to reconcile MIP-inferred retention with laboratory-measured values by incorporating suction scaling factors and the volumetric shrinkage history recorded along the full drying path. Fourteen regression algorithms were benchmarked under identical cross-validation conditions; XGBoost was selected as the most defensible architecture on the basis of its balance of predictive accuracy and resistance to overfitting rather than the single highest cross-validated score, achieving R² = 0.997 on the held-out test partition, a level of accuracy comparable to several other tree-based ensembles evaluated, but combined with materially greater generalisation stability. Shapley additive explanations analysis decomposed this prediction into contributions from individual inputs and identified five key physical descriptors. Those five descriptors were subsequently adopted as the terminal alphabet for gene expression programming, which produced two transparent, computationally inexpensive closed-form analytical models whose predictive accuracy was preserved against nine independent experimental datasets drawn from the open literature, the primary practical contribution of this work. Furthermore, the model serves as a vital tool in unsaturated soil mechanics to ensure the safe and cost-effective design of infrastructure founded on these problematic soils.