A hybrid finite element method–data-driven model for predicting deep foundation pit deformation

This paper addresses the following research question: How can a physical FEM model integrate a data-driven model to accurately and efficiently predict deep foundation pit deformation during excavation, while mitigating issues of data scarcity and slow optimization convergence? To answer this question, a hybrid, data-driven finite element method (FEM) was developed to predict the lateral deformation of deep foundation pits. The model comprises: (1) a Finite Element Model (FEM) and a synthetic dataset that integrates numerical simulation results with field monitoring data; (2) a hybrid surrogate model employing Structured Kernel Ridge Regression and Least Squares Boosting (SKRR-LSBOOST); and (3) an Improved Hippopotamus Optimization (IHO) algorithm for optimizing soil-layer parameters. A case study is employed to validate the effectiveness and feasibility of our model. The results demonstrate that the hybrid model achieves an R 2 value of 0.904 for deformation prediction, outperforming other candidate models. It is suggested that the newly developed model can help site managers identify potential risks in advance and ensure the safety of people and the surrounding environment. Additionally, it enhances decision-making effectiveness during the excavation of deep foundation pits. • An ensemble learning surrogate model is employed to fuse the strengths of machine learning models to improve parameter inversion accuracy and mitigate overfitting risks. • An improved Hippopotamus Optimization algorithm is developed to accelerate model convergence, support global optimum searches, and improve the deformation prediction accuracy. • A real deep foundation project in China is taken as a case study for illustration. • The proposed approach demonstrates strong performance in deformation prediction during deep foundation pit excavation.