Predicting Axial Capacity of Pipe Piles Using Artificial Neural Networks

Traditional methods for estimating the axial capacity of open-ended pipe piles often exhibit poor agreement with measured load test results, leading to potential under- or over-conservative designs. This study applies a multilayer perceptron (MLP) model to predict pile capacity from pile features and soil type only, using a database of 105 full-scale load tests spanning a wide range of diameters (0.25–2.5 m) and lengths (5–98 m). Rather than relying on a fixed architecture, model hyperparameters including network depth, width, activation functions, regularization, etc., were optimized using Optuna, a Bayesian hyperparameter optimization library to systematically search over model configurations. The final model demonstrated improved predictive performance compared to traditional design methods, highlighting the promise of machine learning for foundation design while underscoring the need for larger, more diverse datasets and physics-informed features to ensure generalization.