Prediction of Load Capacities of Closed-Ended Piles Using Boosting Machine Learning Methods

In this study, a novel data-driven model is developed using boosting-type machine learning algorithms with the aim of predicting the ultimate load-bearing capacities of closed-ended piles. A comprehensive database is gathered using the full-scale load test data with four features. Special boosting type machine learning methods are trained and tested with the database. Once predictions are made, a newly developed machine learning algorithm called Shapley method is utilized to decide the effectiveness of the selected features in predicting pile capacities. Results indicate that the pile cross-section area and length features are sufficient to achieve accurate predictions covering the parameters on the pile side and the CPT-based tip resistance is the only parameter needed on the soil side. While different boosting methods result in different levels of accuracy in predicting the load bearing capacities of closed-ended piles, it is generally possible to determine the minimum number of features necessary to satisfy a high goodness of fit. In the end, optimum number of features are determined in the prediction process using the Shapley method through the boosting algorithms giving us a valuable prediction tool for estimating the bearing capacity of closed-ended piles.