TPE-Optimized Neural Network Framework for Predicting Settlement of Nodular Pile Foundations

This chapter explores the application of a hybrid machine learning framework that combines Artificial Neural Networks (ANN) with Tree-Structured Parzen Estimator (TPE) to predict the settlement of nodular piles under static loading conditions. The study begins with a descriptive analysis of input and output variables, including pile geometry, applied load, and soil resistance measured from the Standard Penetration Test (SPT). The methodology involves Bayesian optimization of ANN architecture using TPE, which is shown to outperform other machine learning models such as Random Forest, XGBoost, CatBoost, and LightBoost. The evaluation of the TPE-ANN model demonstrates high prediction accuracy and consistent residual behavior across data splits. The results indicate that the TPE-ANN model achieves a test R² of 0.939 and a RMSE of 3.093 mm, confirming its strong predictive ability. The chapter concludes by highlighting the reliability and robustness of the ANN-TPE model in predicting pile settlement, supporting more informed design decisions for nodular piles under various loading and soil conditions.