Predictive modeling of pore pressure build-up in vibratory pile driving through machine learning

Abstract

In the context of large-scale infrastructure projects, such as major bridges and docks on shorelines, understanding the behavior of deep piles in saturated sandy soils is crucial. This study employs three-dimensional numerical modeling of vibratory pile driving using Midas GTS NX finite element software and the UBCSAND constitutive model, challenging several common simplifying assumptions found in prior research. The efficacy of the numerical model in predicting pile driving processes and potential liquefaction was rigorously evaluated and validated against experimental data from previous studies. Sensitivity analyses were performed to investigate how pore pressure and liquefaction potential are influenced by various factors, including vibratory pulse counts, pile length-to-diameter ratios, and soil properties. The results from these analyses were utilized to train artificial neural networks and symbolic regression models. The performance of these models was assessed using a range of performance metrics and ROC curves. To enhance interpretability, symbolic regression provided a clear mathematical expression capturing the relationship between key features and soil liquefaction. Furthermore, SHapley Additive exPlanations were employed to offer detailed insights into feature importance and the model’s decision-making process. Design charts were developed based on these models to offer practical guidance for practitioners. Overall, this study underscores the effectiveness of integrating advanced numerical simulations with machine learning techniques, demonstrating significant improvements in understanding and predicting pile driving behavior and liquefaction potential in saturated sandy soils.