GeoFWI3D

Large-scale 3D velocity models for deep-learning full waveform inversion (FWI) and other seismic processing workflows.

Overview

GeoFWI3D provides many synthetic 3D Earth models, each stored as raw float32 (little-endian) binaries. Per model you get:

File	Contents
vp3d.bin	P-wave velocity
image3d.bin	Synthetic p-reflectivity
rgt3d.bin	Relative geologic time (RGT)
fault3d.bin	Fault index mask

Each volume has shape 96 × 96 × 96 in (X, Y, Z) order. On disk the layout is contiguous C-order over those dimensions.

Download

Compressed archives are hosted on Box:

Download GeoFWI3D data

Each models_batch_*.tar.gz contains 1000 models. Place the archive files in the same directory as this repository, then extract.

Extract

From the repository root:

./extract_models.sh

This creates allmodels/ and extracts every models_batch_*.tar.gz into it.

Directory layout

After extraction, models look like:

allmodels/
├── model_0000/
│   ├── image3d.bin
│   ├── vp3d.bin
│   ├── rgt3d.bin
│   └── fault3d.bin
├── model_0001/
│   └── ...
└── ...

Folder names use four-digit zero padding: model_0000, model_0001, …

Quick start

1. Install Python dependencies used by the example notebook: NumPy, Matplotlib, and scikit-image (for marching_cubes).
2. Run Jupyter with working directory quick_start/ so from plotting import plot3d works. Open read_data.ipynb: it defines read_models and plot_all_models, sets data_root / model_folders / shape, and walks through loading and plotting.

Model categories from parameters.csv

The repository includes parameters.csv, with one row per model and a key sample_index column matching folder names like model_{sample_index:04d}.

Useful label columns for categorization:

• yn_fault: 1 if model contains faults
• yn_salt: 1 if model contains a salt body

Example (from quick_start/read_data.ipynb) to build category index lists and select models:

import csv
from pathlib import Path

with Path("../parameters.csv").open(newline="") as f:
    rows = list(csv.DictReader(f))

for r in rows:
    r["sample_index"] = int(r["sample_index"])
    r["yn_fault"] = int(float(r["yn_fault"]))
    r["yn_salt"] = int(float(r["yn_salt"]))

def select_indices(rows, **flag_equals):
    return [
        r["sample_index"]
        for r in rows
        if all(int(r[k]) == int(v) for k, v in flag_equals.items())
    ]

layered_indices = select_indices(rows, yn_fault=0, yn_salt=0)
fault_indices = select_indices(rows, yn_fault=1, yn_salt=0)
salt_indices = select_indices(rows, yn_salt=1, yn_fault=0)
fault_and_salt_indices = select_indices(rows, yn_fault=1, yn_salt=1)

idx = salt_indices[0]
image, vp, rgt, fault, salt = read_models(data_root, model_folders, idx)

Salt + Fault model

plot_all_models(
    data_root, 
    model_folders, 
    9177, 
    "Salt model with faults",
    save_path="../gallery/salt_fault_model.png"
)

Fault mask

fault3d.bin stores a fault index per voxel. To plot a single fault (here index 5):

image, vp, rgt, fault, salt = read_models(data_root, model_folders, 9177)
fault_mask = fault.T == 5
plot3d(
    fault_mask.astype(np.float32),
    cmap='Reds',
    frames=[45, 45, 46],
    ifnewfig=True,  
    showf=False,     
    close=False,     
    ifinside=False,
    figname="../gallery/fault_mask_5.png"
)
plt.show()

Salt mask

Salt bodies have RGT = 0 in rgt3d.bin. Mask and plot with:

salt_mask = rgt == 0
salt_mask = salt_mask.astype(np.float32)
plot3d(
    salt_mask.T,
    cmap='Reds',
    frames=[45, 45, 46],
    ifnewfig=True,   
    figname="../gallery/salt_mask.png"
)
plt.show()

---

Plotting helpers in quick_start/plotting.py are adapted from pyseistr (https://github.com/aaspip/pyseistr) utilities.