Post‐Event Ground Motion Estimation using Graph Neural Networks

Accurate ground motion (GM) estimates are essential for forensic analysis of structural damage following major earthquakes when direct recordings at the location(s) of interest are unavailable. Contemporary post‐event GM estimation methods often leverage nearby observations to constrain estimates of intensity measures (IMs); however, existing approaches rely on empirical ground‐motion models with well‐known limitations in capturing spatial dependencies. This study introduces a graph neural network (GNN) approach for estimating ground‐motion IMs, leveraging a graph‐based representation to naturally encode spatial dependencies and allow for different observation types. Applied to a New Zealand (NZ) case study, the GNN achieves performance in line with the established multivariate normal conditional IM method, while learning spatial correlations directly from the data. These case study results illustrate the viability of GNNs for post‐event GM estimation, while the data‐ and graph‐based approach offers inherent advantages, such as support for any IM type and straightforward extensibility to additional observation types, for example, macroseismic intensity. Continued improvements in model architecture and increased data availability are expected to further enhance GNN performance and applicability for this problem.