Prediction of tunnelling-induced longitudinal deformation based on knowledge-guided neural network
Transportation Geotechnics, 61:102103, 2026.
Summary
This study develops a physics-informed approach for predicting longitudinal ground deformation caused by shield tunnelling when monitoring observations are sparse. The model incorporates governing equations for elastic ground deformation and the associated boundary conditions into neural network training. Automatic differentiation evaluates the physical constraints. To represent layered ground with contrasting properties, a multidomain formulation divides the problem into separate regions and imposes continuity conditions at their interfaces. This treatment targets the gradient discontinuities that can undermine a single-network approximation in composite strata. Comparisons with FLAC3D simulations test displacement predictions for both homogeneous and layered configurations. The results support the use of physical constraints to supplement limited observations and show that the multidomain approach improves behavior near material interfaces. The framework combines observed data and mechanical assumptions, with its scope determined by the adopted elastic ground model.