Hybrid machine learning framework for predicting soil-caisson interaction under seismic loading using XGBoost and xLSTM

Abstract

The soil-caisson interaction under seismic loading is an important consideration in design and safety assessment of deep foundation systems especially in seismic prone areas. Traditional finite element method (FEM) and boundary element method (BEM) cannot accurately simulate the time-dependent and nonlinear characteristics of soil-caisson systems under dynamic loading, primarily due to the need for making assumptions and using high calculation accuracy when taking into account the fully nonlinear characteristics of soil. To solve these problems, a three-step approach is developed in this study: (1) Establishment of the finite element model of soil-caisson interaction in ABAQUS; (2) validation and verification of the finite element model using shaking table test data, and (3) training and optimization of a hybrid machine learning framework, combining Extreme Gradient Boosting (XGBoost) with eXtreme Long Short-Term Memory (xLSTM) networks with Particle Swarm Optimization (PSO). This framework integrates the effectiveness of ensemble learning and deep learning for temporal modeling to forecast seismic responses like acceleration and stress distribution. Both the hybrid models were found to be highly successful in reproducing the dynamic soil-caisson behavior with both visual and statistical comparisons. The proposed framework was found to have better prediction accuracy and computational efficiency than the FEM and BEM approaches. The results provide insights into the promise of combining numerical modeling with experimental testing and advanced machine learning approaches to improve seismic design, risk assessment, and infrastructure resilience in geotechnical earthquake engineering.