lehmannfa commited on
Commit
787882c
·
verified ·
1 Parent(s): 82a2e6a

Update README.md

Browse files
Files changed (1) hide show
  1. README.md +58 -1
README.md CHANGED
@@ -6,4 +6,61 @@ metrics:
6
  - mae
7
  - mse
8
  - code_eval
9
- ---
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
6
  - mae
7
  - mse
8
  - code_eval
9
+ ---
10
+
11
+ # MIFNO
12
+ 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.
13
+
14
+ ![MIFNO](https://github.com/user-attachments/assets/e08b86b4-2374-41ee-8e4f-7a4ab8019805)
15
+
16
+ ## Data
17
+ 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).
18
+
19
+ ## Training and Evaluation
20
+ 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.
21
+
22
+ ## Visualization
23
+ 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.
24
+
25
+
26
+ ## References
27
+ If you use this code, please cite
28
+ ```
29
+ @article{lehmannMultipleinputFourierNeural2025,
30
+ title = {Multiple-Input {{Fourier Neural Operator}} ({{MIFNO}}) for Source-Dependent {{3D}} Elastodynamics},
31
+ author = {Lehmann, Fanny and Gatti, Filippo and Clouteau, Didier},
32
+ year = {2025},
33
+ journal = {Journal of Computational Physics},
34
+ volume = {527},
35
+ pages = {113813},
36
+ issn = {00219991},
37
+ doi = {10.1016/j.jcp.2025.113813},
38
+ }
39
+ ```
40
+
41
+ If you use the HEMEW<sup>S</sup>-3D database, please cite
42
+ ```
43
+ @misc{lehmannPhysicsbasedSimulations3D2023a,
44
+ title = {Physics-Based {{Simulations}} of {{3D Wave Propagation}} with {{Source Variability}}: \${{HEMEW}}{\textasciicircum}{{S-3D}}\$},
45
+ shorttitle = {Physics-Based {{Simulations}} of {{3D Wave Propagation}} with {{Source Variability}}},
46
+ author = {Lehmann, Fanny},
47
+ year = {2023},
48
+ publisher = {[object Object]},
49
+ doi = {10.57745/LAI6YU},
50
+ url = {https://doi.org/10.57745/LAI6YU}
51
+ }
52
+ ```
53
+ and
54
+ ```
55
+ @article{lehmannSyntheticGroundMotions2024,
56
+ title = {Synthetic Ground Motions in Heterogeneous Geologies from Various Sources: The {{HEMEW}}{\textbackslash}textsuperscript\{\vphantom\}{{S}}\vphantom\{\}-{{3D}} Database},
57
+ author = {Lehmann, F. and Gatti, F. and Bertin, M. and Clouteau, D.},
58
+ year = {2024},
59
+ journal = {Earth System Science Data},
60
+ volume = {16},
61
+ number = {9},
62
+ pages = {3949--3972},
63
+ doi = {10.5194/essd-16-3949-2024},
64
+ url = {https://essd.copernicus.org/articles/16/3949/2024/}
65
+ }
66
+ ```