Ultimate Bearing Capacity of Closed-Ended Piles Using Nonlinear Machine Learning Methods

In this study, axial load bearing capacities of closed-ended piles in cohesive and cohesionless soils are predicted through a data-driven study. Seven machine learning methods are used based on CPT and pile load test data identified as the learning features, which are necessary to teach those methods with the related databases. Nonlinear machine learning models are developed and the gathered databases are separated into five subsets according to the cross-validation-principle, which are subsequently used for both training and testing of the machine learning methods. Predictions of nonlinear machine learning methods are validated with classical empirical equations of load bearing capacity. Relevance Vector Regression and Random Forest methods are found to generate more accurate predictions than the other machine learning methods as well as empirical equations. Hence, nonlinear machine learning methods are determined to be reliable tools in predicting the pile load capacities of closed-ended piles provided that there is a large enough database.