The use of machine learning to assess failure risks of shallow continuous tunnels subjected to active dip-slip faults

To date, multiple methods, including experimental and numerical approaches, have been proposed to address the behavior of a tunnel colliding with a dip-slip fault. However, the potential of artificial intelligence techniques remains largely untapped for evaluating tunnel behavior. In this study, three predictive models were applied to forecast fault displacement and tunnel fragility: multilayer perceptron (MLP), self-organizing map (SOM), and nondominated sorting genetic algorithm (NSGA). The MLP model excelled in predicting displacement associated with dip-slip faults (R2 = 0.99), suggesting an extremely high level of forecast accuracy. Additionally, SOM analysis revealed that the behavior of continuous tunnels in response to varying angles of normal faults is more complex than that observed with reverse faults. Notably, lowering the height-to-diameter ratio (h/D) and decreasing the fault angle below 70° dramatically increased tunnel fragility, adversely impacting its performance. Furthermore, the findings from NSGA indicate that dip-slip faults with angles 50–63° coupled with low overburden heights (3.5–6 m) pose a serious and substantial risk to the stability of thick, continuous tunnels.