This repository contains both the dataset and the codebase for ViT-EFWI, a project aimed at replicating the findings detailed in our submitted manuscript titled "Improved elastic full-waveform inversion with Vision Transformer reparameterization and Recurrent Neural Network reformulation".
Fig1. The schematic architecture of the proposed ViT parameterization shown in red box integrated within RNN-based EFWI framework shown in blue box.
Fig2. The inverted elastic models for Marmousi2 model using clean data, (a) Vp, (b) Vs, and (c) density obtained by Deepwave EFWI; (d) Vp, (e) Vs, and (f) density obtained by the proposed ViT-EFWI. The MSEs are marked on each plot for quantitative comparison.
The list of files:
.
├── README.md
├── data
│ ├── generate_model_polt.ipynb
│ ├── model
│ ├── observed
│ └── raw_model
├── fig
│ ├── marmousi2_vit_deepwave.png
│ └── vitefwi_flow.png
├── paper_result
│ ├── bp
│ ├── marmousi2
│ └── toy
├── result
│ └── marmousi2
├── vit-efwi.ipynb
└── vit-efwi.yml- All the models and associated seismic data are generated by a Jupyter Notebook ./data/generate_model_polt.ipynb
- All trained models used in the paper are save in ./paper_results/
- All codes to generate results in the paper are gathered into a sigle Jupyter Notebook ./vit-efwi.ipynb
You can then set up a conda environment with all dependencies like so:
conda env create -f vit-efwi.yaml
conda activate vit-efwi