Geological multiphysics · Open data & benchmarks
A shared subsurface.
Four physical views.
Aligned 3D geology, physical properties, and geophysical
observations.
A common ground for learning across gravity, magnetics, MT, and
seismic.
Full-release counts from the public dataset card .
01 / The idea
A subsurface structure can look very different through gravity, magnetic, electrical, and seismic measurements. Studying these views together calls for data that share the same geological context.
GEMS-3D pairs three-dimensional property volumes with observations from all four modalities. Spatially aligned properties, anomaly labels, and acquisition metadata connect physical structure to its observed signatures.
Use these paired representations to study supervised forward surrogates, joint property reconstruction, anomaly recognition, and robustness to incomplete observations.
Igneous intrusions · Gas & hydrates · Brine faults · Sulfide bodies · Salt domes · Basement interfaces · Serpentinized zones
02 / Research figures
Explore original figures from the research manuscript. Enlarge a figure to read its labels and compare the details.
Enlarge figure
Representative samples from the manuscript. Columns show nine anomaly families. Rows show velocity, density, log-resistivity, susceptibility, MT, seismic, and gravity–magnetic responses. Colors are specific to each panel.
Enlarge figure
Model-side diagnostics from the manuscript: anomaly-family counts, row-normalized property signatures, coverage across background fault-structure classes, and family-wise property distributions.
Enlarge figure
The manuscript’s dataset pipeline connects geological conditions to aligned velocity, density, resistivity, and susceptibility volumes, then to gravity, magnetic, magnetotelluric, and seismic observations in a unified data package for downstream benchmarks.
03 / Inside a sample
Choose a modality to explore its property volume, observation, and fields in a prepared sample.
Connect the distribution of subsurface mass to its gravity signature. Pair a density volume with a spatial observation grid for supervised operator learning.
rho_modelgravity_dataStudy how magnetic susceptibility contrasts appear in surface observations. Shared geological context lets magnetic signatures be compared with the other physical views.
chi_modelmagnetic_dataExplore the electrical view through frequency-dependent apparent resistivity and phase. Frequency metadata and polarization channels accompany the prepared observations.
res_modelmt_app_res mt_phasemt_freqs_hzRelate velocity structure to time-dependent seismic observations. Source and receiver geometry retain the acquisition context for studying individual shots or joint predictions.
vp_modelseismic_data04 / Research directions
Use a common dataset to ask how physical information is learned, combined, and recovered.
Predict geophysical observations from their corresponding property volumes, or learn all four mappings jointly.
Physical properties Observations
Forward surrogate benchmarksReconstruct aligned property volumes from multiple observations, or recognize the underlying anomaly family.
Observations Subsurface properties
Joint inversion baselineEvaluate missing modalities and shuffled cross-modal pairings to examine robustness and the contribution of aligned information.
Complete views Incomplete views
Robustness evaluationBenchmark validation groups samples by background model to avoid overlap between training and validation backgrounds.
05 / Open datasets
Three release sizes, the same paired structure. Each example links a model bundle with an observation bundle.
Inspect the format and build your first data-loading workflow.
A subset for developing experiments before scaling to the full release.
The full paired release for larger-scale multiphysics experiments.
Counts and sizes follow the linked release cards. Review each card’s metadata, manifests, and license before use. Even Tiny contains approximately 20 GiB of data.
06 / Get started
Download the data, inspect a bundle, and build on the open benchmark loaders and models.
import numpy as np
# Use a trusted, downloaded dataset sample.
with np.load("path/to/model_bundle.npz",
allow_pickle=True) as sample:
vp = sample["vp_model"]
rho = sample["rho_model"]
res = sample["res_model"]
chi = sample["chi_model"]
print(vp.shape, rho.shape)Aligned volumes. A shared sample. Four physical properties.
The public benchmark release uses prepared data. Geological model generation and the numerical forward solvers are maintained separately.
07 / Reference
Reference the public dataset release when using the data. This resource entry points to the dataset card; consult the card for bibliographic updates.
Dataset card & release metadata@misc{gems3d_dataset,
title = {GEMS-3D: Geological Multiphysics
Earth Simulation in 3D},
howpublished = {Hugging Face dataset},
url = {https://huggingface.co/datasets/anomy4pub/GEMS-3D}
}