Data-driven modeling for evaluating deformation of a deep excavation near existing tunnels

Abstract

This study explores an integrated framework combining in-situ test-based numerical and data-driven modeling to assess the performance of a deep excavation-tunnel system. To achieve the goal, a case history of deep excavations adjacent to existing tunnels in silt/sand-dominated sediments is introduced to establish a base three-dimensional finite element (3D-FE) model. In-situ tests such as cone penetration test (CPT/CPTU) and seismic dilatometer test (DMT/SDMT), as an alternative to laboratory testing, are used to determine a set of advanced constitutive model parameters. The established excavation-tunnel numerical model is then validated against filed monitoring data. A dataset from numerical simulation is created for training and testing four machine learning models (i.e., artificial neural network (ANN), support vector machines (SVM), random forest (RF), and light gradient boosting machine (LightGBM)), which predict the maximum wall deflection, ground surface settlement, horizontal and vertical displacements of the tunnel. Results show that the ANN model outperforms other models in prediction capacity. Its generalization ability in practice is further enhanced by comparing field measurement data and empirical equations. The findings suggest that, with the integrated in-situ tests, FE and ANN modeling could be used to predict deformation responses of deep excavations close to existing tunnels in soft soil. The present study is useful and valuable for practical risk assessment and mitigation decisions.