SubsurfaceBench

A large-scale benchmark dataset of field-scale 3D subsurface velocity models inspired by real geological surveys, paired with 2D velocity slices and their corresponding acoustic wavefields and shot gathers. The dataset spans multiple geological settings and contains thousands of paired samples, providing a comprehensive benchmark for machine learning tasks in computational geophysics.

While currently tested with Fourier Neural Operators (FNOs) for predicting seismic wave propagation, SubsurfaceBench is designed to support a broad range of ML architectures and research directions in subsurface modeling.

Dataset Summary

Metric	Value
Total index rows	12828
Total HDF5 files	12828
3D velocity models	0
2D velocity slices	4276
Wavefields	4276
Shot gathers	4276
Train samples	4096
Test in-distribution	100
Test out-of-distribution	80
Model types	f3, fault, gom, penobscot, salt_canopy, seam
Frequency bands	3-6Hz

Dataset Structure

Parquet Index

The data/all.parquet file is a sidecar index that catalogs every HDF5 file in the dataset. Each row represents one HDF5 file. The slice_id column links related files together --- a velocity slice, its wavefield(s), and its shot gather(s) all share the same slice_id.

Browse the index using the Dataset Viewer above to explore all 25 columns interactively.

Data Types

data_type	Description	HDF5 Key	Count
model	3D SOS-smoothed velocity volume	velocity	0
slice	2D velocity slice (training input x)	velocity	4276
wavefield	2D acoustic wavefield (training target y)	wavefield	4276
gather	Shot gather (alternative target y)	shot_gather	4276

Splits

Split	Description
train	Training data (all model types)
test_in_dist	In-distribution test (same model types as train, different slices)
test_out_dist	Out-of-distribution test (held-out geology)

Propagation Time Constraint

• Wavefields exist at 5s propagation time only
• Shot gathers exist at both 5s and 8s propagation time

Schema (25 columns)

#	Column	Type	Description
1	slice_id	string	Links related files (e.g., f3_042_il0123). Null for 3D models
2	model_id	string	Source 3D model (e.g., f3_042)
3	data_type	string	model | slice | wavefield | gather
4	model_type	string	Geological category (e.g., f3, gom, fault)
5	split	string	train | test_in_dist | test_out_dist
6	file_path	string	Relative path to HDF5 file
7	orientation	string	inline or crossline
8	slice_index	int32	Index in original 3D volume
9	slice_location_m	float32	Physical position in meters
10	slice_axis	int32	1 (inline) or 2 (crossline)
11	volume_shape	string	Source volume dims (e.g., 960x1000x1000)
12	depth_samples	int32	nz of this array (varies by model type)
13	width_samples	int32	nx of this array
14	propagation_time_s	float32	Simulation time: 5.0 or 8.0
15	frequency_band	string	e.g., 3-25Hz
16	f_min_hz	float32	Band minimum frequency
17	f_max_hz	float32	Band maximum frequency
18	velocity_min_m_per_s	float32	Min velocity (m/s)
19	velocity_max_m_per_s	float32	Max velocity (m/s)
20	velocity_mean_m_per_s	float32	Mean velocity (m/s)
21	velocity_std_m_per_s	float32	Std velocity (m/s)
22	source_x_km	float32	Source X position (km)
23	source_z_km	float32	Source Z depth (km)
24	source_x_idx	int32	Source X grid index
25	source_z_idx	int32	Source Z grid index

Directory Structure

dataset_root/
├── data/
│   └── all.parquet                      # Sidecar index (25 columns)
├── models/                              # 3D SOS velocity volumes (flat)
├── slices/                              # 2D velocity slices (individual HDF5)
├── wavefields/
│   └── {split}/5s/{freq_band}/       # Nested: 5s propagation ONLY
└── shot_gathers/
    └── {split}/{5s|8s}/{freq_band}/ # Nested: 5s AND 8s propagation

Usage

Browse the Index

import pandas as pd

# Load the full index
df = pd.read_parquet("data/all.parquet")

# Filter by split and data type
train_slices = df[(df.split == "train") & (df.data_type == "slice")]
train_wavefields = df[
    (df.split == "train") & (df.data_type == "wavefield")
    & (df.frequency_band == "3-6Hz")
]

# Pair velocity slices with wavefields for training
pairs = train_slices.merge(train_wavefields, on="slice_id", suffixes=("_vel", "_wf"))
print(f"Training pairs: {len(pairs)}")

Load Individual HDF5 Files

import h5py

# Velocity slice
with h5py.File("slices/slice_f3_042_il_0123.h5", "r") as f:
    velocity = f["velocity"][:]           # (nz, nx) float32
    model_type = f["velocity"].attrs["model_type"]
    orientation = f["velocity"].attrs["orientation"]

# Wavefield
with h5py.File("wavefields/train/5s/3-6Hz/wavefield_f3_042_il_0123.h5", "r") as f:
    wavefield = f["wavefield"][:]         # (nt, nx, nz) float32
    freq_band = f["wavefield"].attrs["frequency_band"]

# Shot gather
with h5py.File("shot_gathers/train/5s/3-6Hz/shot_gather_f3_042_il_0123.h5", "r") as f:
    gather = f["shot_gather"][:]          # (nt, nx) float32
    prop_time = f["shot_gather"].attrs["propagation_time_s"]

Dataset Creation

Source Data

Field-scale 3D velocity models inspired by publicly available subsurface surveys:

• F3 (Netherlands North Sea)
• GOM (Gulf of Mexico)
• Fault (synthetic fault models)
• Salt Canopy (synthetic salt body models)
• SEAM (SEG Advanced Modeling)
• Penobscot (offshore Canada, held out for OOD testing)

Models are processed through SOS (Sum of Squares) structural smoothing to create smooth background velocity fields suitable for acoustic wave propagation.

Wavefield Generation

2D acoustic wavefields are generated by solving the acoustic wave equation on each velocity slice using finite-difference time-domain (FDTD) simulation, with a Ricker wavelet source at the surface center.

Citation

@dataset{subsurfacebench,
  title={SubsurfaceBench},
  author={Stanford Exploration Project},
  year={2026},
  url={https://huggingface.co/datasets/jdstitt/subsurfacebench},
}

License

This dataset is released under CC BY 4.0.

Contact

Stanford Exploration Project (SEP), Stanford University.