Simulation-based machine learning for real-time reliability analysis of different tunnel shapes under quasi-static loading

Circular and rectangular tunnels are commonly studied under seismic loading, while sub-rectangular tunnels have recently been adopted in practice. This study uses the Finite Difference Method (FDM) to analyse six tunnel shapes under seismic conditions by quasi-static analysis displacement-based methods, considering both soil and tunnel parameters. The resulting data on internal force distribution in the linings form a valuable dataset for machine learning applications. An ensemble machine learning (ML) framework, Extreme Gradient Boosting (XGBoost), is proposed as an efficient alternative to the computationally demanding FDM for tunnel design optimisation. XGBoost achieves rapid predictions with strong agreement to FDM results. Feature importance analysis identifies key design variables affecting performance. Additionally, full-field predictions of internal forces in six tunnel shapes demonstrate the model’s high efficiency. Considering uncertainties in seismic and soil characteristics, a reliability-based framework using Monte Carlo Simulation estimates tunnel lining reliability. Finally, a user-friendly application has been developed to predict internal force extremes and distributions and perform real-time reliability analysis, offering a practical tool for preliminary tunnel design under seismic loading.