Performance characterisation of machine learning models for geotechnical axial pile load capacity estimation: an enhanced GPR-based approach
Ibrahim Haruna Umar, Mahir Sukairaj Salga, Hang Lin, Jubril Izge Hassan, Abdulaziz Ahmad, Awaisu Shafiu Ibrahim, Benjamin Kuma Jechira
Abstract
Traditional methods for predicting axial pile capacity often rely on simplified assumptions, leading to potential inaccuracies in diverse geotechnical conditions. This study investigates the key parameters influencing axial pile capacity through comprehensive statistical and machine learning analyses in Kano, Nigeria. Six models–Bayesian additive regression trees (BART), explainable boosting machine (EBM), gradient boosting machine (GBM), Gaussian process regression (GPR), improved GPR, and multivariate adaptive regression splines (MARS)–were evaluated using a dataset of 100 training and 25 testing samples from reinforced concrete piles. Statistical analysis revealed strong correlations between pile length and embedment depth (r=0.94) and moderate correlations between axial capacity and length (r=0.78). The improved GPR model demonstrated superior performance with the highest R2 (0.9873) and lowest MARE (0.0678) during training, maintaining robust performance during testing (R2=0.9812). Feature importance analysis identified the uncorrected number of blows at the pile tip (N) as the most influential parameter (35.2–42.3%), followed by embedment depth (17.97–32.8%). Partial dependence analysis revealed non-linear relationships between design parameters and axial capacity, with diminishing returns observed beyond certain thresholds. These findings provide valuable insights for optimising pile foundation design in geotechnical engineering applications.