Prediction of pile settlement due to parallel twin tunnel excavation using machine learning

In urban tunnel excavation, parallel twin tunnel excavation can reduce construction time under environmental constraints. This study analyses ground and structural behaviour during simultaneous excavation of parallel twin tunnels, comparing it with staggered excavation. A laboratory model test was conducted, assuming tunnel excavation as volume loss, with pile length, vertical pile-tunnel offset, and pile row count as variables. LVDTs measured settlement at the surface above the pile and tunnel. A two-dimensional finite element analysis was performed for back-analysis to estimate soil properties and evaluate settlements and vertical ground displacement. Various geometric and ground conditions were considered, generating 1296 cases for automated numerical analysis of both excavation methods. Data were analysed using Random Forest, XGBoost, and LightGBM through supervised learning, and model performance was compared to determine the most suitable prediction model. Predictions were validated against laboratory model test and back-analysis results. Finally, key variables influencing pile settlement were quantified, enabling a comparative assessment of staggered and simultaneous excavation.