Abstract The accurate prediction of soil bearing capacity remains a critical challenge in geotechnical engineering, particularly given the complex non-linear relationships between soil properties and foundation performance. Traditional analytical methods often struggle to capture these complexities, leading to potential overestimation or underestimation of bearing capacity across different footing types. This study investigates the application of machine learning techniques for predicting soil bearing capacity across different footing types. The research utilized 200 datasets, comprising 175 institutional sources and 25 laboratory Direct Shear Test experiments, with an 80-20 split ratio for model development and validation. A Hybrid Tree-Based Ensemble Learning (HTBEL) methodology was developed and compared against conventional models (M5P, CatBoost, AdaBoost, SVR, and Decision Tree) and Terzaghi analytical equation. The HTBEL model demonstrated superior predictive accuracy with R² values exceeding 0.96 across all footing types, maintaining errors below 5% throughout the sample range. Square footings showed the highest bearing capacity (median ~3,400 kN/m²) due to favorable area-to-depth ratio, followed by circular footings (~3,200 kN/m²) benefiting from symmetrical stress transmission, while strip footings (~2,000 kN/m²) showed lower performance due to concentrated stress distribution along their length. Clustering analysis identified optimal configurations at 3 clusters (Silhouette Score: 0.5236) and 10 clusters (0.5315). This research establishes HTBEL as a robust methodology for bearing capacity prediction in geotechnical engineering applications.
Read full abstract