FuTE-FWI

Towards complex seismic layers: A vision Transformer approach to full waveform inversion.

This repository contains the source code for Fusion vision Transformer Enhanced network for Full Waveform Inversion (FuTE-FWI).

Enviroment Requirement

python = 3.10.14
pytorch = 2.3.0
pytorch_msssim = 1.0.0
lpips = 0.1.4
deepwave = 0.0.20
numpy
einops
skimage
matplotlib
tqdm

Suggest using anaconda to manage the environment:

conda create -n futefwi python=3.10

Data Preparation

OpenFWI

First, download model data from OpenFWI official repository. In our experiment, FlatVel-A, CurveVel-A, FlatFault-A and CurveFault-A dataset are required.

OpenFWI-140

To generate any dataset of OpenFWI-140, use gen_dataset.py. Here is an example for generating FlatVel-D dataset:

python gen_dataset.py -d FlatVel -v A

The dataset splitting strategy is hard-coded in dataset.py. Modify this file as needed. The default strategy is as follows:

Dataset	Train / test Split	Corresponding .npy files
Vel Family	24k / 6k	data(model)1-48.npy / data(model)49-60.npy
Fault Family	48k / 6k	data(model)1-96.npy / data(model)97-108.npy

Training

To train FuTE-FWI, VelocityGAN and InversionNet, use the training script respectively. For example, to train FuTE-FWI on FlatVel-D, run the following command:

python train_futefwi.py -d FlatVel -v A

To reproduce the ablation study, add the --ablation flag to specify the variant:

python train_futefwi.py -d FlatVel -v A --ablation sfe

Two options are available for --ablation: sfe (w/o SFE) and tm (w/o TM).

Refer to the help text if you want to specify the default parameters.

Testing

To evaluate the models, use test.py. The model and dataset name is required. For FuTE-FWI, you can add the --ablation option to evaluate the ablation variants.

python test.py -m FuteFWI -d FlatVel -v A; % Test FuTE-FWI on FlatVel-D.
python test.py -m VelocityGAN -d CurveVel -v A; % Test VelocityGAN on CurveVel-D.
python test.py -m InversionNet -d FlatFault -v A --draw; % Test InversionNet on FlatFault-D. Add --draw if you need visualization.
python test.py -m FuTE-FWI -d CurveFault -v A --sample-eval; % Test FuTE-FWI on CurveFault-D. Add --sample-eval to evaluate and visualize one sample.