Machine Learning–based Surrogate Modeling for Risk and Resilience Analysis

This study develops machine-learning-based surrogate models for risk and resilience analysis. Probabilistic resilience assessments are largely driven by Monte Carlo sampling. Among existing implementations, the risk and resilience framework in the Rtx software-continuously developed at the Center for Infrastructure Sustainability and Resilience Research at Sharif University of Technology-performs probabilistic such analyses via Monte Carlo sampling. For decision-making processes aimed at risk mitigation, these analyses must be executed repeatedly within optimization procedures; at an urban scale, such repetition imposes a significant computational burden that can render the process impractical. To address this limitation, low-cost surrogate models are employed. In this study, two Light Gradient Boosting Machine (LightGBM) models are developed: the first predicts spectral acceleration at a period of one second, and the second, conditioned on the output of the first model, estimates aggregate seismic loss. Training data are generated using Monte Carlo sampling. The proposed framework is evaluated on a virtual city subject to a scenario earthquake. Results indicate that the differences in the mean and standard deviation of the loss between the surrogate and the reference model are approximately ۲%, while the surrogate model reduces computational burden substantially and thereby enables regional seismic risk analysis.

Keywords:

Surrogate Model , Machine learning , Risk analysis , Community resilience analysis , Monte Carlo sampling