Re-evaluation of machine learning models for predicting ultimate bearing capacity of piles through SHAP and Joint Shapley methods

This study examines the prediction of the load-bearing capacity of closed and open-ended piles using machine learning (ML) methods. Full-scale load test results and CPT data are used to gather two comprehensive databases for such piles. ML models are developed employing input features associated with pile geometry and CPT resistances along with the ultimate bearing capacity being the only output feature. Following the training/testing sequences, the interpretability of ML predictions is examined through the Shapley and Joint Shapley value methods. Shapley values for multiple feature combinations allow ML models to decide the number of features necessary to make the most accurate predictions. Using updated input features, the models are rebuilt and predictions are repeated with the new input feature set; hence, the re-evaluation of ML models is focused on at this point. These features are twofold: One has two geometric features attributed to piles: cross-section area and length, and the other is a single feature attributed to soil, the average CPT tip resistance, which are overall sufficient in predicting the load capacity of both closed and open-ended piles. To the best of our knowledge, this study is one of the pioneers of its kind for pile foundations. The results show that the predictions of ML methods aided with strong interpretability techniques prove to be necessary in providing accurate results. As a result, Shapley is determined to be a useful tool for other geotechnical engineering applications as well.