GlobalTomo

The first global dataset for physics-ML seismic wavefield modeling and full-waveform inversion

Project structure

GlobalTomo/
│
├── conf/
│   ├── config.yaml      # Configuration file
│
├── images/              
│   └── overview.jpg
│
├── scripts/             
│   ├── eq.py            # Defination of PDE constraints
│   ├── load_data.py     # Dataloader from .h5 files
│   ├── meta_info.py     # Meta information about the data
│   ├── misc.py          # Custmized training tools
│   ├── model.py         # Custmized ML models
│   └── plot.py          # Visualization tools
│
├── .gitignore           # Contents ignored by git
├── infer.py             # Evaulation functions
├── inverse.py           # Inversion functions
├── LICENSE              # License
├── README.md            # An instruction for use
├── requirements.txt     # Dependencies
└── train.py             # Training script

Getting started

conda create -n globaltomo python==3.8
conda activate globaltomo
pip install -r requirements.txt

Training forward models

Define your training configuration files within the conf directory. We provide example YAML files in the repository for an easy start. Typically, you should specify the data tier, model type, and desired output in these files. Ensure that the custom.name in your configuration matches the file name.

To initiate training, execute the following command:

python train.py --config-name='acoustic_wfslice_MLP'

Once training is complete, the checkpoints will be automatically stored in the outputs directory. Note that you can safely comment the self._record_constraints() in anaconda/env/globaltomo/lib/python3.8/site-packages/modulus/sym/trainer.py L607 to save time and avoid errors in training.

Evaluating forward models

Before evaluating the trained models, specify the wave_type, output_type, and model_name in infer.py.

Then, run the following command to start evaluation.

python infer.py

If the model path exists, the code will automatically load the configurations and the model parameters for evaluation.

Inversion

1. Optimization-based inversion

To perform an inversion using the trained forward models, specify the model path in inverse.py. Then, execute the following command to begin optimizing the structure from a random starting point:

python inverse.py

2. Direct inverse mapping

An alternative strategy is to train a neural network that directly maps the observed seismogram to the velocity structure. To do this, create a config file with custom.velocity set to true and define other parameters as shown in the example file conf/acoustic_velocity_MLP.yaml. Use a command similar to the one used for training forward models:

python train.py --config-name='acoustic_velocity_MLP'

Download the data

To access the data, please click this link and fill the request form.