Scalable reliability analysis of structures under earthquake via sparse graph data and state-space metamodel

Existing approaches for earthquake reliability analysis of large structures are often tailored for fixed structural layouts, making them inapplicable to different configurations, even those with minor modifications such as removing or adding elements. This study introduces a framework, named Sa3, in which a single model covers a set of structural configurations declared in advance, obtained from one training run and one hyperparameter setting. Furthermore, the proposed method offers several advantages, including fast inference time, high accuracy, and reliability results for various safety criteria. The framework’s core idea involves, first, designing a sparse graph representation for structural data, including geometric topology and material properties; second, utilizing a selective state-space-based learning component to transform the graph representation, ground motions, and historical responses into latent data that are beneficial to forecasting tasks. The viability and performance of Sa3 are demonstrated, via three 3D spatial structures of different complexities. The results show that Sa3 can provide reliability results with high accuracy for all investigated structures without requiring retraining or modifications of the framework architecture or parameters, while significantly reducing time complexity. Moreover, ablation and sensitivity studies provide a deeper understanding of the Sa3 framework’s operation.