Machine Learning-Based Prediction of Allowable Bearing Capacity for Deep Mat Foundations: A Comprehensive ۳D Numerical Study

The accurate prediction of the allowable bearing capacity (qall) for deep mat foundations, governed by a settlement criterion, is a critical yet complex challenge in geotechnical engineering. This study presents a robust predictive framework by integrating high-fidelity numerical modeling with machine learning. An extensive database of ۵۰۰۰ samples was generated using three-dimensional FLAC۳D simulations employing the Hardening Soil model. Three machine learning algorithms—Random Forest, Support Vector Regression, and Gradient Boosting (GB)—were developed and compared. The GB model demonstrated exceptional predictive accuracy (R۲=۰.۹۹۵۴), significantly outperforming a hierarchy of three traditional elasticity-based analytical methods. A subsequent sensitivity analysis using SHapley Additive exPlanations (SHAP) revealed that the model autonomously learned complex and physically meaningful geotechnical principles, including the inverse relationship between foundation width and the settlement-governed allowable pressure. This research validates the synergy between numerical simulations and machine learning, offering a highly accurate and interpretable tool that surpasses conventional methods for the detailed design of deep mat foundations.

Keywords:

Bearing Capacity , Deep Foundations , Machine Learning , Numerical modeling , SHapley Additive Explanations