Physics-guided deep learning for generative design of large-diameter tunnels under existing metro lines

Abstract

The overlapping construction of large-diameter tunnels is inevitable, but the construction control faces great challenges due to the complexity of underground environments. A generative design method for large-diameter tunnels under existing metro lines based on physic-guided deep learning is proposed, aiming at optimizing tunnel layouts from a physical perspective to ensure effective construction control. A topology-optimized model dataset considering soil uncertainties is fed into a physics-guided Wasserstein generative adversarial network (PGWGAN) for training, producing numerous physically consistent schemes. The optimal scheme is selected using the multiple-attribute decision-making (MADM) method under multi-working conditions. A case study on large-diameter tunnel construction demonstrates the method's feasibility, showing that it meets the safety requirements across various conditions and achieves significant improvement. This paper contributes a physics-guided generative design method for large-diameter tunnel overlapping construction. It accounts for multiple working conditions and includes an evaluation module that integrates parametric finite element analysis (FEA) with multi-attribute evaluation.