Physics-informed deep learning method for predicting tunnelling-induced ground deformations

Tunnelling-induced ground deformations inevitably affect the safety of adjacent infrastructures. Accurate prediction of tunnelling-induced deformations is of great importance to engineering construction, which has historically been dependent on numerical simulations or field measurements. Recently, some surrogate models originating from machine learning methods have been developed, showing satisfactory prediction performance with high computational efficiency. However, these purely data-driven models show weak robustness in the absence of sufficient training data. This study proposed a hybrid deep learning model integrating both data-driven and physics-based strategies to decrease calculation costs and eliminate the dependence on large numbers of training data. The underlying physical mechanism of ground deformation due to tunnel excavation is coupled into the deep learning framework to form a physics-informed neural network (PINN) model. The performance of the hybrid model is first assessed by comparing it with the classical Verruijt-Booker solution and a conventional purely data-driven model. The potential of the proposed PINN model for engineering applications is then illustrated. It is found that the proposed PINN model can reasonably reproduce ground deformation fields obtained numerically with only a small amount of training data. This paper provides a new paradigm for incorporating hybrid deep learning frameworks and field monitoring systems to predict ground deformation fields in real time.

Automated summary

This study proposes a hybrid deep learning model that combines data-driven and physics-based strategies to reduce calculation costs and reliance on large amounts of training data. The model is able to reasonably reproduce ground deformation fields with a small amount of training data, showing potential for engineering applications.