The mechanical characterization of unsaturated problematic geomaterials remains a significant challenge in geotechnical engineering. This study introduces a comprehensive hybrid framework for unsaturated severely sandy dolomite (SSD) by unifying an advanced physics-based constitutive model (CM) with a data-driven artificial neural network (ANN) predictive tool. The constitutive component innovatively integrates subloading surface (SLS) theory with the Barcelona Basic Model (BBM), establishing a modified Cam-clay framework that explicitly accounts for the coupled effects of matric suction, strain-softening, and overconsolidation. The model is rigorously calibrated and validated against an extensive suite of laboratory experiments, including consolidated undrained triaxial tests on SSD specimens from the Central Yunnan Water Diversion Project under varying confining pressures (100–400 kPa) and saturation states. The results demonstrate that the SLS formulation effectively overcomes the limitations of traditional single-yield-surface models, achieving exceptional predictive accuracy ( R 2 > 0.92) for the complex stress–strain responses observed. Parameter sensitivity analysis identifies the critical state slope ( M ) and saturation compression index (λ(0)) as dominant factors. The model's practical utility is confirmed via a C++ implementation in FLAC3D, accurately replicating experimental stress paths with a mean error below 8%. Complementing this, a shallow feedforward ANN model (2-5-4 architecture) is developed, which predicts key shear strength parameters ( C cu , C ′, φ cu , φ′) from basic index properties (moisture content, saturation) with near-perfect fidelity ( R 2 > 0.998). This dual-pronged approach provides a robust and versatile toolkit for the analysis and design of critical infrastructure, such as tunnels in arid regions, founded on unsaturated geomaterials.
Wang et al. (Tue,) studied this question.