Ground-Motion Flow (GMFlow): Large-Scale 3D Ground-Motion Synthesis with Physics-Inspired Latent Operator Flow Matching

This dataset contains processed 300 synthetic earthquake events for the San Francisco Bay Area, used to evaluate the GMFlow model. The raw simulatoin (5300 events in .h5 format) are provided via https://huggingface.co/datasets/Yaozhong/GMFlow_raw_simulation. Code is released via https://github.com/yzshi5/GMFlow

Dataset Summary

The collection includes three magnitude classes $M_{w}M\_w$ with 100 events each: $M_{w}M\_w$ 4.4, 6.0 ,7.0. Each event is a 4D npy file: [4, 256, 128, 96]([C, H, W, T], spatial grid of 256x128, time of 24 seconds with a sampling frequency of 4 Hz).

• Channels 0–2: Three-component velocity wavefields.
• constant Channel 3: ${\log }_{10}(\sigma )\backslash log\_\{10\}(\backslash sigma)$ (where $\sigma \backslash sigma$ is the standard deviation of original event used for normalization).
• Recover the original 3C wavefield data data[:3] * np.pow(10, np.mean(data[3]))
• Hypocenter and Magnitude file rupture_hypo_test.npy

Weights for Pretrained model

• Super Resolution Operator: under folder SNO
• Auotencoding Operator under folder AENO
• Flow Matching under folder FM

Reference

If you find this repository useful for your research, please consider citing our work

@article{shi_large-scale_2026,
    title = {Large-{Scale} {3D} {Ground}-{Motion} {Synthesis} with {Physics}-{Inspired} {Latent} {Operator} {Flow} {Matching}},
    author = {Shi, Yaozhong and Lavrentiadis, Grigorios and Tsalouchidis, Konstantinos and Ross, Zachary E. and McCallen, David and Zou, Caifeng and Azizzadenesheli, Kamyar and Asimaki, Domniki},
  journal= {arXiv:2603.17403 [cs]},
  year={2026}
}