GeoFWI3D Dataset

GeoFWI3D is a large-scale open-source benchmark dataset of realistic 3D subsurface models designed to accelerate deep learning (DL) assisted seismic imaging and full waveform inversion (FWI). While modern seismic AI research is rapidly evolving toward three-dimensional architectures, DL models trained exclusively on 2D datasets often exhibit limited generalization capability when applied to realistic 3D seismic imaging problems. GeoFWI3D addresses the scarcity of large-scale, realistic training datasets by providing models of increasing geological complexity. The complete dataset comprises 10,000 unique, realistic 3D subsurface volumes. Each model has spatial dimensions of $96\times96\times96$. To provide a comprehensive benchmark and prevent class collapse during training, the dataset is uniformly divided into four structural complexity classes. Class 1 Pure stratigraphy (2,500 volumes) Class 2 Faulted networks (2,500 volumes) Class 3 Salt diapirism (2,500 volumes) Class 4 Complex coupled deformations with intersecting salt bodies and tectonic faulting (2,500 volumes) Data Records & Modalities: vp3d.bin: The ground-truth 3D compressional P-wave velocity model, with cell values stored explicitly in meters per second ($m/s$). fault3d.bin: A discrete 3D semantic mask tracking tectonic deformation, where a value of 0 indicates intact stratigraphy and non-zero integer values map specific fault plane indices. rgt3d.bin: The continuous 3D Relative Geologic Time field, establishing a monotonically increasing coordinate space tied directly to depositional chronology. image3d.bin: The synthetic 3D seismic volume, representing zero-offset acoustic p-reflectivity. Note: Binary masks for salt bodies can be generated on-the-fly by identifying voxels with zero values in the RGT volumes, as the salt bodies physically overwrite and terminate the stratigraphic RGT field. Format and Storage Architecture:To preserve total numerical integrity and prevent compression-induced interpolation artifacts, all physical property volumes and structural labels are saved as raw, uncompressed binary arrays using a uniform 32-bit floating-point precision (float32) format. To optimize network transfer bandwidth, the 10,000 models are batched and compressed into 10 individual tarball archives, each containing a sequential block of 1,000 distinct models. Further quick start guides and baseline benchmarking scripts can be accessed through the project repository at: https://github.com/aaspip/geofwi3d