Large-Scale 3D Ground-Motion Synthesis with Physics-Inspired Latent Operator Flow Matching

GMFlow Paper

Simulation pipeline

GMFlow Raw Simulation is a large-scale open dataset of high-fidelity three-dimensional earthquake ground-motion simulations generated for the GMFlow project. It contains both point-source and finite-rupture earthquake scenarios for San Francisco Bay Area, covering magnitudes Mw 4.4, Mw 6, and Mw 7 . Please download the dataset (.h5 files) from https://huggingface.co/datasets/Yaozhong/GMFlow_raw_simulation

The simulations are performed on the supercomputer Perlmutter at the National Energy Research Scientific Computing Center (NERSC), more than 5300 events with a total size of 3.27 TB, this dataset is intended to support research in earthquake engineering, seismology, scientific machine learning, operator learning, generative modeling, and uncertainty-aware surrogate modeling etc.

Detailed preprocessing files are provided in simulation_process folder, please run the preprocess.py, post_process.py and final_norm.py sequentially for each magnitude

Model architecture

Results

Setup and quick start (no need to download the raw simulation dataset)

First download the processed test dataset from https://huggingface.co/datasets/Yaozhong/GMFlow, place the 300 test events under dataset folder. (100 events for each magnitude). To set up the environment, create a conda environment

# clone project
git clone https://github.com/yzshi5/GMFlow.git
cd gmflow

# create conda environment
conda env create -f environment.yml

# Activate the `gmflow` environment
conda activate gmflow

Download the pretrained weights for Super-resolution operator, Autoencoding operator and Flow Matching via the same link, place them under the checkpoints folder. Then to generate a new synthetic event, the event_id ranges (0, 300), where by default 0-100 Mw6, 100-200 Mw7, 200-300 Mw4.4.

# standard 
python evaluations/quick_test.py end2end_test event_id

# spectral calibarated with the mean of residual
python evaluations/quick_test_unbias.py end2end_test event_id

Training and Inference

run the training file under training_scripts to train SNO, AENO and SNO (128M or 16M). For inference, check the paper and folder evaluations, which contains files for residual plot, magnitude scaling, scenario comparison, zero-shot super-resolution, spectral calibaration etc. please first run files with_calc__ to save the calculated results, then run the corresponding file with__plot__ for visualization.

Video

one Mw4.4 scenario

one Mw7 scenario

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}
}