Data-driven geotechnical characterization and prediction of mechanical properties using machine learning framework

Geotechnical characterization is limited by soil variability, data imbalance, and poor spatial transferability. This study proposes an integrated ASMOTE–PINO–SABO framework to predict soil cohesion and internal friction angle reliably. The framework combines Adaptive Synthetic Minority Oversampling Technique (ASMOTE) for balanced class representation, Physics-Informed Neural Operator (PINO) to embed geotechnical constraints, and Self-Adaptive Bayesian Optimization (SABO) for network tuning. A leakage-free workflow utilizing GroupKFold spatial validation evaluates model performance across unseen borehole locations. Using a balanced dataset of 3,250 samples, the framework ensures physical consistency through physics-guided learning. The optimized model achieves (R2) values of 0.894 for cohesion and 0.918 for friction angle, with RMSE values of 4.12 kPa and 1.21°, respectively. The PINO component reduces physics residual errors by 91.05%. Spatial validation confirms stable transferability at an 87.1 km distance with (R2) values of 0.854 and 0.882. Finite element validation via ABAQUS demonstrates engineering utility, maintaining settlement prediction errors below 2.1%. This integrated framework establishes a robust, physically consistent pathway for data-driven geotechnical characterization.