Probabilistic reliability assessment of twin tunnels considering fluid–solid coupling with physics-guided machine learning

Accurate predictions of the internal force of a tunnel lining (IFTL) and maximum ground surface deformation (MGSD) are critical for avoiding unexpected accidents. This work proposes a physics-guided machine learning approach for probabilistic reliability assessment of IFTL and MGSD during the excavation of twin tunnels in a water-rich region. A physics-guided database including 1000 groups of input combinations is established with the help of the finite element method (FEM) incorporating fluid–solid coupling theory while a categorical boosting (CatBoost) method is employed to formulate the meta-models of the IFTL and MGSD. Comparisons between the results from the FEM model and the measured data present that the established FEM model can effectively capture the ground surface deformation. To further assess the performance of the developed meta-model, importance analyses of input features and sensitivity analyses of geometry parameter combinations are conducted. In the illustrative scenarios, the uncertainties of all input features and the meta-model are considered based on the Monte Carlo simulation and the prediction interval (PI) in the probabilistic reliability assessment. The promising potential of the CatBoost-based meta-model for reliability analysis during the excavation of twin tunnels in a water-rich region is presented in contrast to three popular machine learning approaches.