PGVnet: a machine learning framework for the generation of rapid, physics-consistent PGV maps

SUMMARY The rapid and accurate estimation of strong ground motion is essential for seismic hazard assessment and near-real-time disaster response. Although empirical ground motion models enable fast intensity predictions, they simplify the underlying physics and exhibit large uncertainties. Conversely, physics-based simulations—while capable of more accurately predicting ground shaking—are computationally expensive, making them impractical for large-scale hazard assessments and real-time event response. To overcome these limitations, we introduce a novel two-step machine learning framework that predicts peak ground velocity (PGV) for arbitrary double-couple sources positioned anywhere within a given medium, combining the accuracy of physics-based models with near-instantaneous inference. In the first step, an ensemble of XGBoost predictors, trained on a reciprocal Green’s function data base, generates a sparsely sampled PGV map for any input source. In the second step, we refine this map into a continuous spatial prediction. By leveraging Green’s function reciprocity, our approach reduces the required number of simulations in training, lowering both computational cost and storage demands. Our framework provides spatially continuous PGV predictions and inherently accounts for complex 3-D geological and topographic effects. It can deliver results within seconds while maintaining accuracy up to the highest frequency captured in the physics-based simulations. This makes PGVnet ideal for applications such as rapid earthquake disaster response, as well as large-scale probabilistic seismic hazard analyses and multi-hazard digital ecosystems. Validated in the geologically complex San Francisco Bay Area, our approach generates PGV maps consistent with physics-based simulations, offering an effective balance between computational speed and accuracy.