Shake Anywhere: a simulation-free AI-based earthquake ground motion generator for any source/any geology.

Predicting earthquake ground motion in complex seismological and geological settings remains an open challenge for earthquake engineers and seismologists. While 3D numerical simulation offers valuable insights into the effects of source rupture, wave propagation, scattering and local site effects, its high computational cost and time-to-result hinder its adoption in regional-scale seismic hazard assessments.

Recent advances in neural operators, like MIFNO [1], have enabled fast inference of elastodynamics solution. Despite the accuracy of the 3D numerical simulations employed for training such neural operators, their performance is affected by high-frequency spectral bias [2]. Inferred time histories display a spectral falloff, resulting from the learning bias of deep networks towards low frequency features, generalizing across data. Generating high-frequency content is not only prohibitive from a numerical standpoint (high computational and calibration costs), but also because deep neural networks slowly learn irregular local features.

Previous efforts to improve numerical simulations and MIFNO predictions station-wise, using a diffusion transformer, helped with spectral accuracy [3,4], but this solution did not offer any guarantee to maintain spatial consistency across the entire 3D wave field.

To address this, we use a generative diffusion model trained on a high-resolution seismic dataset (HEMEWS-3D, [5]) that captures a variety of ground-motion scenarios in heterogeneous media. A 3D diffuser [6] first learns the distribution of physically plausible 3D geologies. It then leverages pretrained MIFNO's reconstruction guidance [7] approximation to ensure consistency with known physics, while adding missing high-frequency components and preserving spatial coherence. The approach is validated with frequency-based accuracy metrics.

This framework enables the generation of broadband earthquake scenarios anywhere and for any source, and providing a scalable method for realistic, high-fidelity ground-motion predictions. Not only this solution paves the way towards real-time inference of new broadband earthquake scenarios, but it devotes high-fidelity simulations to specific sites of interest, for fine-tuning the MIFNO, offering a promising solution for earthquake risk assessment.

References

(1) Lehmann et al. 2025, 527, 113813. https://doi.org/10.1016/j.jcp.2025.113813.

(2) Rahaman et al. 2019; Vol. PMLR 97. https://proceedings.mlr.press/v97/rahaman19a.html.

(3) Gabrielidis et al. 2026, 109930. https://doi.org/10.1016/j.cpc.2025.109930.

(4) Perrone et al. 2025. https://doi.org/10.48550/arXiv.2504.00757.

(5) Lehmann et al. 2024, 16 (9), 3949–3972. https://doi.org/10.5194/essd-16-3949-2024.

(6) Molinaro et al. 2024. https://doi.org/10.48550/arXiv.2409.18359.

(7) Bergamin et al. 2024 Workshop on AI4Differential Equations In Science. https://openreview.net/forum?id=1avNKFEIOL.