Quantifying extreme failure scenarios in transportation systems with graph learning.

Statistical analysis of extreme events in complex engineering systems is essential for system design and reliability and resilience assessment. Due to the rarity of extreme events and the computational burden of system performance evaluation, estimating the probability of extreme failures is prohibitively expensive. Traditional methods, such as importance sampling, struggle with the high cost of deriving importance sampling densities for numerous components in large-scale systems. Here, we propose a graph learning approach, called importance sampling based on graph autoencoder (GAE-IS), to integrate a modified graph autoencoder model, termed a criticality assessor, with the cross-entropy-based importance sampling method. GAE-IS effectively decouples the criticality of components from their vulnerability to disastrous events in the workflow, demonstrating notable transferability and leading to significantly reduced computational costs of importance sampling in large-scale networks. The proposed methodology improves sampling efficiency by one to two orders of magnitude across several road networks and provides more accurate probability estimations.