Deep learning-enhanced efficient seismic analysis of structures with multi-fidelity modeling strategies

Seismic assessment and design of structures nominally require large numbers of analyses with high-fidelity models—e.g., to obtain fragility curves–especially for regional scale tasks. In this study, a deep learning-enhanced multi-fidelity modeling approach is devised that can dramatically increase the computational efficiency of such analyses. This approach uses high- and low-fidelity numerical models for generating small and large sample responses first. Then, a deep learning-based projection model is trained with the limited high-fidelity data to learn the correlations within multi-fidelity results with the objective of having a trained model that can predict high-fidelity results from low-fidelity simulations. For validating this approach, a reinforced concrete frame and a high-rise shear-wall structure are used as validation and application examples, and the impacts of various key factors in training and model generation are examined. The results indicate that the proposed approach can effectively accelerate seismic analyses without compromising accuracy.