A multi-fidelity framework with adaptive transfer learning for risk assessment of structural systems subject to natural hazards

Recently, low-fidelity models (e.g., data-driven emulators) have gained extensive acceptance for accelerating the evaluation of nonlinear dynamic systems, thereby enhancing the computational efficiency in estimating small failure probabilities against extreme events, such as natural hazards.However, directly employing low-fidelity models for uncertainty propagation can potentially lead to biased predictions.To address this challenge, this work develops a multi-fidelity scheme with adaptive transfer learning for efficiently estimating small failure probabilities.In this approach, the probability space is partitioned into several strata, within which deep learning-based emulators are developed by adaptive transfer learning.These emulators serve as strata-wise low-fidelity models.By effectively combining the strata-wise high-and low-fidelity evaluations, this method allows for efficient estimation of small probabilities of failure associated with rare events.Through an illustration on an full-scale steel building subject to seismic excitation, the proposed scheme is demonstrated to enable accurate estimation of small failure probabilities, while enabling remarkable speedup compared to stratified sampling schemes relying entirely on high-fidelity model evaluations.