GeoFWI3D: Large-scale 3D Velocity Model Dataset for Deep Learning-assisted Seismic Imaging

The 30-second take

• What: The authors publish a 10,000-model 3D subsurface benchmark spanning stratigraphy, faults, salt, and coupled fault–salt systems, with aligned labels and starter tasks for imaging and inversion.
• Why it matters: 3D full-waveform inversion is scarce and expensive; a shared, realistic training set is how learned imaging can become a default complement rather than a one-lab demo.
• Who should care: Seismic imaging and FWI groups, neural-operator researchers, and anyone who lacked large 3D geologic training volumes.

What the paper actually did

GeoFWI3D is an open benchmark of geologically plausible 3D subsurface models aimed at deep-learning-assisted seismic imaging and full waveform inversion (FWI). Traditional FWI is nonlinear, non-unique, ill-posed, and often trapped in local minima when the starting model is poor — and 3D FWI is described as prohibitively expensive. Learned mappings from shot gathers to subsurface properties need large realistic training sets, which the authors say have been the bottleneck. The release has 10,000 velocity models at 96×96×96, in four complexity classes: pure stratigraphy, faulted networks, salt diapirism, and coupled fault–salt systems. Each model has co-registered labels: compressional velocity (Vp), zero-offset seismic reflectivity, relative geologic time (RGT), and semantic fault/salt masks. Benchmark tasks include fault detection, joint salt-body segmentation and chronostratigraphy, FWI, wavefield and traveltime surrogates with neural operators, and generative modeling with a 3D diffusion model. Baseline results are provided. The dataset is CC BY 4.0.

What makes this disruptive

The scarce object is not another inversion algorithm — it is 3D geologic diversity at training scale. Ten thousand labeled 96³ volumes with four structural classes and multi-modal labels (Vp, reflectivity, RGT, masks) are a community lever. Bundling baselines from segmentation to neural-operator wavefields to 3D diffusion makes the release a scoreboard, not only a zip file. CC BY 4.0 is the abundance license. This does not solve FWI’s local-minima problem; it makes attacking that problem with learning more default.

Why it matters (outside the lab)

Abundance lens: accurate subsurface imaging is still expensive, expert, and capital-heavy — relevant to energy and earth-system monitoring. Shared 3D training data can pull that intelligence toward wider use. Horizon is mid: physics, compute, and real-field transfer remain gates. Near-term, use GeoFWI3D as a public baseline. Medium-term, whether 96³ synthetic geology transfers to field data decides if this becomes ordinary infrastructure.

Limitations & open questions

Models are synthetic and 96³ — resolution and geologic realism are designed, not field volumes. Four complexity classes omit many real-earth styles. Labels are co-registered by construction; real surveys do not hand you RGT and salt masks. Baseline tasks establish reference numbers, not production FWI quality. Learned FWI can still fail when shot gathers and physics differ from the generator. Open license does not remove compute cost of 3D training. Preprint plus dataset release.