Probabilistic Stability Analysis of a Tunnel Placed in a Spatially Varying Rock Mass Using Deep Learning Meta-model

This study aims to evaluate the impact of spatial variability on tunnels situated at varying embedment depths (H). A comprehensive array of rock strength metrics, including GSI, mi, and σci, are employed in the analysis of tunnel behavior within a rock mass adhering to the Hoek-Brown failure criterion. The uniaxial compressive strength of the intact rock (oci) is presumed to exhibit spatial variability, with random fields created using the Markovian autocorrelation function and Cholesky decomposition. Utilizing the strength, geometrical, and statistical spatial parameters, lower bound finite element limit analysis (LB-FELA) and random field theory (RFT) are applied through Monte Carlo simulations (MCS) to optimize the maximum surcharge (qs) the tunnel can endure. The result is presented in the form of a stability number N = qs / σci. Design charts are provided for computed stability numbers (N), its mean, and coefficient of variation. Additionally, a deep learning convolutional neural network (CNN) is developed as a meta-model for the lower bound random finite element limit analysis (LB-RFELA) initially incorporated in this study. Furthermore, the potential of CNN to learn complex features embedded in the prepared random fields through MCS is examined. The random fields are pre-processed accordingly to enhance the zone of influence and are further fed to the prepared network of CNN. Lastly, the autocorrelated random fields of (σci) with embedded tunnels are classified as stable or unstable, and the accuracy of the prepared model is presented.