Deep Learning–Based Prediction of Tunnel Face Stability in Layered Soils Using Images of Random Fields

Abstract

The stability analysis of tunnel faces in multilayered soils presents challenges due to the inherent variability in natural soils. Although the random field finite-element methods offer a reliable approach to address such variability, their heavy computational demands have been a significant drawback. To overcome this limitation, this study presents a novel deep learning–based method for efficient tunnel face stability analysis in layered soils with spatial variability. By combining the merits of convolutional neural networks (CNNs) and U-Net, the proposed method trains surrogate models using a small but sufficient number of random field images to effectively learn high-level features that encompass spatial variabilities, which significantly enhances computational efficiency. In particular, U-Net generates precise displacement field images based on random field images, enabling the discrimination of tunnel face collapse failure modes. To validate the effectiveness of this proposal, a comprehensive case study involving layered soils with spatial variabilities is conducted. The remarkable agreement between the outputs of CNNs and U-Net and the predictions of finite-element simulations underscores the promising potential of using deep-learning models as a surrogate for analyzing the stability of tunnel faces in spatially variable layered soils. Last but not least, the key innovation of this work lies in the pioneering application of U-Net for geotechnical reliability analysis.