Machine Learning–Finite Element Mesh Optimization-Based Modeling and Prediction of Excavation-Induced Shield Tunnel Ground Settlement

Ground settlement prediction for shield construction is highly important and challenging. This study introduces a machine learning algorithm combined with finite element numerical simulation, i.e., machine learning–finite element mesh optimization. For surface subsidence prediction, 16 combination models of ANN, KNN, RF and SVR were optimized by PSO, GA, BT and BO, involving raw data preprocessing, principal component analysis, hyperparameter selection and prediction accuracy evaluation. A subway shield tunneling project was analyzed, in which the meshes of finite element numerical models were discretized into different sizes from 1.0[Formula: see text]m to 2.0[Formula: see text]m. In total, 360 sets of data points were extracted from the simulation results, including stress, strain, shield jacking force, internal friction angle, cohesion force, and settlement, of which 252 data points were used as the input parameters of machine learning model. Analysis of average error rate of finite element–machine learning coupling models showed that the finite element model had the highest accuracy of settlement prediction when the mesh size of the finite element model was 1.4[Formula: see text]m, and the GA-SVR model had the highest accuracy and generalization ability in ground settlement prediction. This study highlights the uniqueness of machine learning–finite element mesh optimization model in application.