An Interpretable Machine Learning Approach for Predicting Bearing Capacity of Driven Piles Using SHAP
نویسندگان
1 Assistant Professor, Faculty of Civil, Water and Environmental Engineering, Shahid Beheshti University, Tehran, Iran
2 BSc, Faculty of Civil, Water and Environmental Engineering, Shahid Beheshti University, Tehran, Iran
3 BSc, Faculty of Civil, Water and Environmental Engineering, Shahid Beheshti University, Tehran, Iran
doi
10.61882/NMCE.2510.1102چکیده
Determination of the ultimate pile bearing capacity is still one of the major concerns in geotechnical engineering due to the complex interaction between soil and structure. This study employs interpretable machine learning models to provide precise predictions of pile capacities while identifying the role of the key design variables. A detailed data set of 100 steel and concrete piles is evaluated by including eight important design variables: effective pile length, cross-sectional area, Flap number, drained cohesion, drained soil friction angle, effective unit weight of soil, pile–soil friction angle, and pile material. Prediction models for the pile capacities are established using the Random Forest, XGBoost, CatBoost, and Extra Trees algorithms, which are validated through a strict 5-fold cross-validation. The results show that the Extra Trees algorithm is the most stable and has the highest predictive capability, with a coefficient of determination (R2) of 0.95 ± 0.03 and RMSE of 1806 ± 999. Furthermore, SHapley Additive exPlanations (SHAP) analysis is performed to calculate the importance of the design parameters, indicating that effective pile length, cross-sectional area, and Flap number are the major contributing factors. This reveals that the proposed unique combination of Flap number with cutting-edge machine learning analysis is an accurate, clear, and viable process for pile capacities.
