PREDICTION OF SURFACE SETTLEMENTS INDUCED BY SHIELD TUNNELLING USING PHYSICS-INFORMED NEURAL NETWORKS

Tunnelling-induced surface settlement is a significant concern due to the complexity of urban environment. For the prediction of surface settlements, traditional machine learning methods ignore the underlying physical mechanism, resulting in the requirement of excessive training samples. Based on the framework of a DNN, the classical Verruijt-Booker solution is modified to connect tunnelling-induced surface settlements with the position of a tunnel face. The modified solution is integrated into a paralleled DNN framework to build the data-driven and physics-informed PINN model, in which the physical mechanism can constrain the neural network. The results of illustrative examples show that: compared to the traditional DNN model, the generalization capability of a PINN model has apparently increased. The results of engineering applications show that: based on the monitoring data at the early stage of a shield tunnel project, the PINN model can accurately predict the surface settlement in the process subsequent to construction; in addition, it increases the intelligence of surface settlement control, which can provide early warnings for potential risks and construction suggestions.