Near-Instantaneous Physics-Based Ground-Motion Maps Using Sparse-to-Dense Deep Learning

Rapid and accurate estimates of ground-motion intensity measures are critical for seismic hazard assessment and disaster response. Empirical ground-motion models provide fast predictions, but suffer from large uncertainties, especially in regions with sparse observations. Physics-based simulations offer physically consistent shaking intensity estimates but remain computationally prohibitive for real-time applications and large-scale scenario analyses. We present a machine-learning framework that predicts high-resolution ground-motion intensity maps conditioned on earthquake source parameters, combining physics-consistent predictions with near-instantaneous inference. The framework predicts a suite of intensity measures widely used in seismic hazard and earthquake-engineering studies -- including peak ground velocity (PGV), peak ground acceleration (PGA), and response spectra -- for arbitrary double-couple sources embedded in a realistic 3D medium, inherently capturing complex geological and topographic effects.

Our approach leverages two complementary training datasets obtained from waveform simulations: spatially sparse shaking intensity maps generated via reciprocity methods and spatially dense intensity maps from forward simulations. A conditioned U-Net is first pretrained on abundant spatially sparse maps to learn global spatial features, subsequently fine-tuned using a limited set of spatially dense maps. This staged training strategy significantly reduces training data requirements while maintaining high predictive accuracy. Applied to the San Francisco Bay Area and Wellington, New Zealand, the framework produces physics-consistent intensity maps with speedups of 6–7 orders of magnitude compared to traditional wave-propagation simulations. This enables scalable, near-instantaneous hazard assessment for both rapid disaster response and comprehensive scenario-based analyses.