Selection of Optimal Vector-Valued Intensity Measures for Seismic Fragility Analysis in Shield Tunnels Based on LSTM Neural Networks

This research introduces a novel approach for seismic fragility assessment by employing a long short-term memory (LSTM) neural network to identify the most effective scalar and vector intensity measures (IMs). This approach enables the rapid and accurate plotting of vector fragility surfaces for shield tunnels embedded in layered soils and subjected to seismic actions. First, an extensive suite of two-dimensional, fully nonlinear soil–structure interaction analyses was executed to generate ground–motion–structure response pairs. These records were subsequently leveraged to train the LSTM network, which received free-field acceleration time histories and directly output critical engineering demand parameters along the tunnel lining. The developed framework significantly mitigates computational expenses while maintaining an acceptable level of fidelity relative to the reference finite element results. Consequently, it serves as an alternative to traditional time history evaluation techniques. Second, we conducted an IM screening process using the results of the LSTM predictions. On the basis of criteria such as relevance, efficiency, practicality, and professionalism, we benchmarked 17 scalar IM and 3 vector IM candidate schemes. The findings indicate that the peak ground velocity (PGV) serves as the most effective scalar IM, whereas the combination of peak ground acceleration (PGA) and PGV forms the optimal vector IM. Finally, probabilistic demand and capacity models are integrated within a fully analytical fragility formulation to derive both scalar and vector fragility estimates. Comparative evaluation reveals that vector IM based fragility surfaces markedly reduce epistemic uncertainty and furnish refined probabilistic descriptions of damage states (DSs) across the seismic demand space.