Multifidelity operator learning for predicting displacement fields of tunnel linings under external loads

Abstract

In practical tunnel projects, the deformation of tunnel linings affects the service performance and structural reliability of tunnels. However, currently, several monitoring points must be installed to represent the specific deformation patterns of tunnel linings accurately, incurring high labor and financial costs. In recent years, the rapid development of artificial intelligence has provided potential solutions to this issue. To solve the aforementioned problem, this study proposed a multifidelity DeepONet framework, which comprised two neural networks. The first low-fidelity network was trained with data provided by a macro-level numerical model validated by experimental campaigns to learn the physical deformation patterns of tunnel linings. The second network was trained with limited high-fidelity monitoring data to learn the correlations between observations and numerical models. Even with very limited monitoring data, the proposed framework could still predict the mechanical behavior of tunnel linings under different loading scenarios. In this study, data collected from noncircular tunnel projects were used as case studies. The results demonstrated that the final output conformed to the deformation pattern obtained with the numerical simulations and was consistent with the actual measurements, achieving seamless fusion of the experimental campaigns and numerical models.