File size: 3,401 Bytes
fc761fe 324e0b3 fc761fe 324e0b3 787882c 019a444 787882c 324e0b3 |
1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60 61 62 63 64 65 66 67 68 69 70 71 72 73 74 75 76 |
---
license: cc-by-nc-4.0
language:
- en
datasets:
- HEMEWS-3D
metrics:
- mae
- mse
- frequency_bias
- Goodness-Of-Fit
tags:
- neural operator
- scientific machine learning
- earthquake
- seismology
- ground motion
---
# MIFNO
Train a Multiple-Input Fourier Neural Operator (MIFNO) to predict the solution of 3D source-dependent Partial Differential Equations (PDEs). The MIFNO is described in the article [Multiple-Input Fourier Neural Operator (MIFNO) for source-dependent 3D elastodynamics](https://doi.org/10.1016/j.jcp.2025.113813). It extends the 3D Factorized Fourier Neural Operator (F-FNO) to PDEs with a source term. As such, the MIFNO contains a dedicated *source branch* that takes as input a vector of source parameters.
<img src="graphical_abstract.png" alt="MIFNO" width="1000" height="1000">
## Data
The MIFNO is trained on the [HEMEW<sup>S</sup>-3D database](https://doi.org/10.57745/LAI6YU) that contains 30,000 simulations of the 3D elastic wave equation in heterogeneous media with different sources. The folder `HEMEWS3D_S32_Z32_T320_fmax5_rot0_test` contains 20 test samples that have already been preprocessed and are readily usable for machine learning tasks. To get the entire train dataset, you need to download the data from the [HEMEW^S-3D repository](https://entrepot.recherche.data.gouv.fr/dataset.xhtml?persistentId=doi:10.57745/LAI6YU) and follow the [preprocessing pipeline](https://github.com/lehmannfa/HEMEW3D).
## Training and Evaluation
The [Github repository](https://github.com/lehmannfa/MIFNO) contains all the code to train and evaluate the MIFNO. A model checkpoint obtained after 300 epochs is provided in this repository.
## Visualization
You can use the [Jupyter notebook](https://huggingface.co/lehmannfa/MIFNO/blob/main/Ground%20motion%20predictions%20with%20MIFNO.ipynb) in this repository to run inference predictions with the MIFNO and plot ground motion predictions.
## References
If you use this code, please cite
```
@article{lehmannMultipleinputFourierNeural2025,
title = {Multiple-Input {{Fourier Neural Operator}} ({{MIFNO}}) for Source-Dependent {{3D}} Elastodynamics},
author = {Lehmann, Fanny and Gatti, Filippo and Clouteau, Didier},
year = {2025},
journal = {Journal of Computational Physics},
volume = {527},
pages = {113813},
issn = {00219991},
doi = {10.1016/j.jcp.2025.113813},
}
```
If you use the HEMEW<sup>S</sup>-3D database, please cite
```
@misc{lehmannPhysicsbasedSimulations3D2023a,
title = {Physics-Based {{Simulations}} of {{3D Wave Propagation}} with {{Source Variability}}: \${{HEMEW}}{\textasciicircum}{{S-3D}}\$},
shorttitle = {Physics-Based {{Simulations}} of {{3D Wave Propagation}} with {{Source Variability}}},
author = {Lehmann, Fanny},
year = {2023},
publisher = {[object Object]},
doi = {10.57745/LAI6YU},
url = {https://doi.org/10.57745/LAI6YU}
}
```
and
```
@article{lehmannSyntheticGroundMotions2024,
title = {Synthetic Ground Motions in Heterogeneous Geologies from Various Sources: The {{HEMEW}}{\textbackslash}textsuperscript\{\vphantom\}{{S}}\vphantom\{\}-{{3D}} Database},
author = {Lehmann, F. and Gatti, F. and Bertin, M. and Clouteau, D.},
year = {2024},
journal = {Earth System Science Data},
volume = {16},
number = {9},
pages = {3949--3972},
doi = {10.5194/essd-16-3949-2024},
url = {https://essd.copernicus.org/articles/16/3949/2024/}
}
``` |