Physics-informed deep learning and analytical patterns for predicting deformations of existing tunnels induced by new tunnelling

Abstract

Investigating the response of soil and existing tunnels to new undercrossing tunnelling is important for ensuring the serviceability of existing structures. Existing methods, such as analytical models or data-driven approaches, struggle to accurately simulate tunnel mechanisms and predict soil deformations. Integrating these two methods provides a promising solution for practical applications. This study proposes an analytical-physics-informed neural network (API-NN) enhanced with transfer learning and uncertainty quantification for a real-time prediction of existing tunnel deformations induced by new tunnelling. The proposed approach leverages analytical patterns by exploiting physical perspectives into data-driven networks and merging them with transfer learning to solve the forward and inverse problems of soil-tunnel interactions, particularly with sparse datasets. Two scenarios, a practical project in Xiamen City and a 3D finite element method, were applied to validate the proposed approach. Results revealed that the API-NN approach successfully incorporates both physical patterns and data-driven networks, achieving real-time forecasting of existing tunnel deformation. The proposed approach maintains its computational precision even when dealing with noisy data. Compared to the existing physics-informed method, the proposed approach realized a 13.9% reduction in mean absolute error, demonstrating higher forecasting precision that is essential for tunnel applications.