Comparative prediction of axial pipe pile capacity using SVR, XGBoost, and neural networks

Accurately forecasting axial pile capacity remains a cornerstone challenge in geotechnical design because the empirical and semi-empirical equations that dominate practice often deviate widely from full scale load tests. To address this gap, we compiled a database of 546 instrumented load tests and adopted the Davisson Capacity as ground truth. Four mechanics-based traditional procedures ( FHWA, USACE, API, Revised Lambda ) were first applied to establish a realistic baseline, after which three machine learning (ML) models were trained on identical folds: support vector regression ( SVR ) as a base learner, extreme gradient boosting ( XGBoost ) as a boosted tree ensemble, and a multilayer perceptron ( MLP ). The MLP ’s depth, width, activation family, optimizer, learning rate schedule, batch size, dropout, L 1 /L 2 regularization, normalization strategy, kernel initializer, loss function, and gradient clipping threshold were exhaustively tuned with Optuna ’s Bayesian sampler under five-fold cross validation. Despite this intensive search, the optimized MLP could not produce a tangible improvement over XGBoost , while both MLP and XGBoost outperforming SVR and all four traditional methods.