Data-driven modeling of ultimate load capacity of closed- and open-ended piles using machine learning

Abstract

Field pile load tests are fairly expensive experiments that can be applied to certain pile types required to be installed in full scale. Hence, it is neither practical nor efficient to perform a load test for every installed pile. While there exist many empirical relations for predicting pile capacities, such methods typically suffer from accuracy and generality. Therefore, current geotechnical practice still looks for methods to accommodate full-scale pile load testing to serve as accurate and practical tools. In this study, load bearing capacities of closed- and open-ended piles in cohesive and cohesionless soils are predicted using machine learning. Nine such methods are utilized in the analyses where Cone Penetration Test (CPT) and pile data are considered as the learning features necessary to teach those methods the database gathered via a comprehensive search. Then, machine learning models are developed, and the databases are separated into five-folds according to the cross-validation-principle, which are used for both training and testing of the machine learning methods. Model predictions are validated with classical CPT-based equations. Results indicate that Relevance Vector Regression and the Random Forest methods typically generate considerably better predictions than the other methods and empirical equations. Thus, machine learning methods are found as reliable tools to predict the pile load capacities of both open-ended and closed-ended pile provided that there is a large enough database and that an appropriate method is used.