Apply generative earthquake-load models to time-variant reliability

Establish how generative machine-learning models for earthquake ground motions can be integrated into systematic time-variant structural reliability analysis, particularly when recorded data are sparse, non-independent, and representative of nonstationary random processes.

Background

The review discusses generative adversarial networks and related methods for producing ensembles of earthquake ground motions. It emphasizes unresolved concerns involving data scarcity, violation of the independent-and-identically-distributed assumption, time-frequency nonstationarity, and the physical fidelity of generated records.

Although generative models may provide flexible ensembles, their role in reliability analysis—especially first-passage and other time-variant reliability calculations—has not been established in the reviewed literature.