Leveraging data-driven machine learning techniques to enhance bearing capacity estimation in prebored and precast piles

The rapid pace of urbanization and the increasing demand for reliable deep foundation systems in metropolitan areas necessitate advanced methodologies for accurately predicting the bearing capacity of pile foundations. This study presents a novel application of machine learning (ML) techniques to estimate the pile bearing capacity (PBC) of prebored and precast piles (PPP) constructed with pretensioned prestressed high-strength concrete (PHC). Unlike traditional approaches that rely on empirical dynamic driving formulas, this research explores the potential of ML models to provide more accurate and practical solutions for construction quality control. Three ML algorithms—K-Nearest Neighbor Regression (KNR), Extreme Gradient Boosting (XGB), and Deep Neural Networks (DNN)—were employed to predict PBC at two critical stages of PPP installation: the End of Initial Driving (EOID) and Restrike phases. A dataset comprising 217 dynamic load tests was used for model development and validation. The results demonstrate that the ML models, particularly the DNN, significantly outperform traditional dynamic driving formulas such as the Hiley, Danish, and Gates methods in terms of both prediction accuracy and generalization capability. The DNN model achieved R 2 values of 0.960 for EOID and 0.915 for Restrike phase predictions, with minimal prediction errors across varying site conditions. Importantly, this study introduces a direct comparison between AI-driven models and conventional formulas, providing empirical evidence that ML-based predictions are less susceptible to the overestimations and underestimations often observed in traditional methods. Furthermore, interpretability was enhanced through SHapley Additive exPlanations (SHAP), which identified key influencing factors, such as pile diameter and ram weight, offering valuable insights for on-site decision-making and adaptive control during pile driving operations. By validating ML predictions against field-tested PDA results and widely used empirical formulas, this research establishes the practical feasibility of integrating AI into existing construction workflows. The proposed ML approach not only enhances prediction reliability but also streamlines quality control processes, reducing dependence on empirical judgments and contributing to more efficient and cost-effective project delivery.