double-blind-FN's picture
Tweak README wording (drop 'publication-style')
b67761b verified
---
license: cc-by-4.0
task_categories:
- image-to-image
tags:
- benchmark
- geoscience
- seismology
- geophysics
- subsurface
- velocity-model
- acoustic-wavefield
- wave-propagation
- scientific-computing
- HDF5
- preview-sample
pretty_name: Field-Scale Dataset (Preview Sample)
size_categories:
- n<1K
configs:
- config_name: default
data_files:
- split: all
path: "data/all.parquet"
---
# Field-Scale Dataset — Preview Sample
A curated **single-family preview** of the
[`subsurfacegen/field-scale-dataset`](https://huggingface.co/datasets/subsurfacegen/field-scale-dataset)
full release. Designed for reviewers to download in **under 5 minutes** and
inspect the dataset's content + structure without pulling the full 12 TB.
> ## 📊 Quick visual tour
>
> **All five figures from this preview are pre-rendered below**
> (and again, fully interactive, in [`view_preview.ipynb`](view_preview.ipynb)).
> The notebook is **rendered inline by HuggingFace with embedded plot outputs**,
> so reviewers can see every figure on this page without downloading or
> running anything.
![Shot-gather panels — the manuscript intro figure](figs/shot_gather_panels.png)
*Above: shot-gather panels. Velocity model on the left shows the three
color-coded source positions (Left/Middle/Right = source idx 8/32/56);
the three dark panels on the right are the corresponding 8 s gathers
recorded by the 1,000-receiver streamer. Generated from the included
`shot_gathers/8s/3-25Hz/...gom_151_il_0507.h5`.*
This preview ships exactly the four HDF5 files that produce the manuscript's
intro figure, so reviewers can reproduce that figure 1-for-1. Contents:
| File | Size | What it is |
|---|---:|---|
| `models/gom_d619/gom_151_sos.h5` | 2.19 GB | 3D SOS-smoothed velocity volume — Gulf of Mexico realization 151. Shape `(619, 1000, 1000)` on a 10 m grid. HDF5 key `velocity`. |
| `slices/slice_gom_151_il_0507.h5` | 1.1 MB | 2D inline-507 velocity slice extracted from the volume above. Shape `(619, 1000)`. HDF5 key `velocity`. |
| `wavefields/5s/3-6Hz/wavefield_gom_151_il_0507_srchorizontal4.958km.h5` | 508 MB | 5 s acoustic wavefield, 3-6 Hz band, single source at horizontal x = 4.958 km. Shape `(358, 1000, 619)` = `(nt, nx, nz)`. HDF5 key `wavefield`. |
| `shot_gathers/8s/3-25Hz/shot_gather_cube_gom_151_il_0507.h5` | 100 MB | 8 s shot-gather cube, 3-25 Hz band, **64 sources stacked**. Shape `(64, 572, 1000)` = `(n_src, nt, n_rec)`. HDF5 key `shot_gather_cube`. |
| **Total** | **~2.80 GB** | |
Plus a slimmed `data/all.parquet` (4 rows) covering only these four files and
the same 25-column schema as the full repo.
## Visual tour (all 5 figures)
The five panels below are PNGs rendered by `view_preview.py` from the four
HDF5 files included here. The same plots appear interactively in
[`view_preview.ipynb`](view_preview.ipynb).
### 1. 3D cutaway of the SOS-smoothed velocity volume
![3D cube cutaway of gom_151_sos.h5](figs/model_3d_cube_cutaway.png)
*Three exposed faces (top, front, left) of the 619 × 1000 × 1000 volume
on a 10 m grid, diverging seismic colormap centered at 3100 m/s.*
### 2. Three orthogonal slices through the volume
![Orthogonal slices through gom_151_sos.h5](figs/model_3d_orthogonal_slices.png)
*The 2D slice we ship (`slice_gom_151_il_0507.h5`) is the inline-507 panel on
the left. Red dashed lines mark where each panel's slice was taken in the
other two views.*
### 3. 2D velocity slice with acquisition geometry
![2D velocity slice with streamer + source overlay](figs/slice_2d_velocity.png)
*`slice_gom_151_il_0507.h5` overlaid with the streamer (every 25th of 1000
receivers @ 10 m depth, 0.6–9.4 km) and the gold star marking the wavefield
source x = 4.958 km.*
### 4. Wavefield time progression
![Wavefield time slices, 3-6 Hz, source @ x=4.958 km](figs/wavefield_time_progression.png)
*Six time snapshots (0.20 → 5.00 s) from the included 3-6 Hz wavefield.
Velocity model in grayscale (alpha 0.4) sits behind the wavefield amplitude
(seismic-alpha cmap, transparent near zero).*
### 5. Shot-gather panels (manuscript intro figure)
![Shot-gather panels](figs/shot_gather_panels.png)
*1×4 layout: velocity model with the three color-coded source positions on
the left, then the three corresponding shot-gather panels (Left/Middle/Right
= source idx 8/32/56) recorded over the 8 s simulation.*
## Quick start
```bash
# 1. Download the preview repo (replace ./preview_data with any local path)
pip install huggingface_hub h5py hdf5plugin numpy matplotlib
huggingface-cli download subsurfacegen/field-scale-dataset-preview \
--repo-type=dataset --local-dir=./preview_data
cd ./preview_data
# 2. Render all 5 figures (3D cutaway, orthogonal slices, 2D slice w/ geometry,
# wavefield time progression, shot-gather panels) to ./figs/
python view_preview.py --plot all --output-dir ./figs
```
The `view_preview.ipynb` notebook in this repo also embeds all five figures
inline — HuggingFace renders the saved outputs directly on the dataset page,
so you can read through the visual narrative without downloading anything.
## How this sample was created
We selected the `gom_151_il_0507` family because it is the exact sample family
shown in the manuscript's intro figure. The recipe is fully reproducible against
the full dataset:
```python
import pandas as pd
df = pd.read_parquet("data/all.parquet") # full dataset's parquet
family = df[df.slice_id == "gom_151_il_0507"]
# Returns 1 slice + 5 wavefields (one per band) + 5 gathers (one per band) = 11 rows.
# We kept the 3-6 Hz wavefield (matches the manuscript figure) and the 3-25 Hz
# shot-gather cube (richest source bandwidth). Total = 4 files.
```
Properties of this sample family in the full parquet index:
| Property | Value |
|---|---|
| `model_id` | `gom_151` |
| `model_type` | `gom` (Gulf of Mexico) |
| `slice_id` | `gom_151_il_0507` |
| `orientation` | `inline` |
| `slice_index` | `507` |
| `slice_location_m` | `5070` (= 507 × 10 m grid) |
| `split` (slice + wavefield + gather) | **`train`** |
| `split` (3D model row) | `null` (3D models are not split-assigned by design) |
| Wavefield source x | `4.958031 km` (uniformly random along the slice) |
| Wavefield band | `3-6 Hz` (1 of 5 bands; lowest-frequency option) |
| Gather band | `3-25 Hz` (1 of 5 bands; widest bandwidth) |
| Gather sources | 64 equally-spaced from `0.6` to `9.4 km` |
## Schema (matches the full dataset)
The slimmed `data/all.parquet` has the same 25 columns as the full repo. The
key columns for joining the four files are:
- `slice_id` — links a slice to its derived wavefields and shot-gathers
- `model_id` — links a slice to its source 3D volume
- `data_type` — one of `model | slice | wavefield | gather`
See the full repo's README for the complete column list and definitions.
## Numerical simulation parameters
| Parameter | Value |
|---|---|
| Solver | Devito FDTD (Louboutin et al., 2019) |
| Grid spacing | 10 m × 10 m |
| Time step (native) | 1.0 ms |
| Wavefield temporal subsample | factor 14 (effective dt = 14 ms) |
| FD stencil space order | 8 |
| Absorbing boundary | 60-cell sponge |
| Top boundary | Free surface |
| Receivers | 1,000 per slice, streamer at 10 m depth, 0.6-9.4 km |
| Source depth | 10 m |
| Wavefield source wavelet | band-limited Ricker, 3-6 Hz, peak f₀ = 4.5 Hz |
| Gather source wavelet | band-limited Ricker, 3-25 Hz, peak f₀ = 14.0 Hz |
| Random seed | 42 (reproducible source placement) |
## Loading the files
```python
import h5py
import hdf5plugin # registers the compression filter used by wavefields + gathers
import json
# 3D velocity volume — root attr "metadata" is a JSON string
with h5py.File("models/gom_d619/gom_151_sos.h5", "r") as f:
vol = f["velocity"][:] # (619, 1000, 1000) float32
meta = json.loads(f.attrs["metadata"])
# 2D slice — dataset attr "metadata_json"
with h5py.File("slices/slice_gom_151_il_0507.h5", "r") as f:
sl = f["velocity"][:] # (619, 1000) float32
sl_meta = json.loads(f["velocity"].attrs["metadata_json"])
# Wavefield — dataset attr "metadata_json"
with h5py.File("wavefields/5s/3-6Hz/wavefield_gom_151_il_0507_srchorizontal4.958km.h5", "r") as f:
wf = f["wavefield"][:] # (358, 1000, 619) = (nt, nx, nz)
wf_meta = json.loads(f["wavefield"].attrs["metadata_json"])
# Shot-gather cube — dataset attr "metadata_json"
with h5py.File("shot_gathers/8s/3-25Hz/shot_gather_cube_gom_151_il_0507.h5", "r") as f:
cube = f["shot_gather_cube"][:] # (64, 572, 1000) = (n_src, nt, n_rec)
sg_meta = json.loads(f["shot_gather_cube"].attrs["metadata_json"])
```
## License
This preview is released under [CC BY 4.0](https://creativecommons.org/licenses/by/4.0/),
matching the full dataset's license.
## Citation
```bibtex
@dataset{field_scale_dataset_preview,
title={Field-Scale Dataset (Preview): single-family sample of SOS-smoothed velocity volumes,
2D slices, wavefields, and shot-gather cubes},
author={Anonymous},
year={2026},
url={https://huggingface.co/datasets/subsurfacegen/field-scale-dataset-preview},
}
```
## Contact
Removed for anonymous review. Full dataset and contact information will be
restored after acceptance.