Extrapolation performance of Generative Waveform Models in data-scarce regimes

Determine how well Generative Waveform Models for seismic ground motions extrapolate beyond their training parameter ranges—including earthquake magnitude, hypocentral distance, site condition V_S30, and faulting type—and quantify their performance at the data-scarce edges of these ranges.

Background

The paper introduces a conditional generative latent denoising diffusion model trained on Japanese strong-motion records, conditioned on earthquake magnitude, hypocentral distance, V_S30, and faulting type. As with all strong-motion modeling, short-distance recordings of large-magnitude events are scarce, which challenges synthesis performance in precisely the regimes most relevant to hazard.

The authors note that while GWMs could augment datasets, the capability of such models to extrapolate beyond their trained ranges and to perform reliably at the sparse edges of the parameter space remains unresolved. Establishing this would guide practical deployment of GWMs in regions or scenarios with limited data.