Labeled training datasets for Geologically-informed AI model

We develop a geological and geophysical forward modeling workflow from the perspective of stratigraphic forward modeling, adding fold structures, building attribute models, building seismic data. Specifically, we first use PyBadlands (Salles et al., 2018) to simulate numerous stratigraphic layers under diverse forcing conditions. Then we perform the interpolation process to obtain a stratigraphic volume and add folding structures (Wu et al., 2020). For the attribute models construction, we first manually build an initial porosity model (P0 ) based on the RGT model, assigning different porosity values at different RGT layers and introduce lateral variations to enhance its realism. Then we generate two weighting matrices (w 1 and w2 ) to incorporate variations in depositional environment and compaction effects, respectively. Finally, we combine these elements to generate the final realistic porosity model, which can effectively and reasonably reflect the influence of the variations in rock properties, depositional environments, and burial depths on subsurface sediments. After obtaining the realistic porosity model, we then employ the Biot-Gassmann theory (Biot, 1941; Gassmann, 1951; Berryman, 1999) to compute the corresponding velocity, density, and impedance models. To be consistent with the field seismic data, we perform the depth-to-time conversion for all attribute models based on the velocity model. Finally, we convolve the reflectivity model (converted from impedance model) with a Ricker wavelet and add real noise to generate the final synthetic seismic image. Simultaneously, we also automatically obtain the corresponding RGT labels. Finally, we automatically obtain 2000 pairs of synthetic seismic images (520[crossline] × 300[Time]) and corresponding RGT labels contained diverse geological clinothems patterns.