Machine Learning workflow to create pseudo-3D from 2D seismic We offer a new service to create a pseudo-3D volume from a 2D seismic dataset. Pseudo-3D cubes are interesting beca...
Machine Learning workflow to create pseudo-3D from 2D seismic
We offer a new service to create a pseudo-3D volume from a 2D seismic dataset.
Pseudo-3D cubes are interesting because they enable the application of 3D seismic interpretation workflows and 3D visualization techniques to 2D seismic datasets.
We developed two workflows in OpendTect’s Machine Learning platform to do this: a direct, fast approach and an approach that requires interpretation input to flatten / unflatten the data. Both workflows are applicable in settings with regular 2D grids that tie in to a 3D seismic volume.
The slider compares the interpolation results of both methods on two blind test lines in the Penobscot survey, offshore Nova Scotia. The input volume from which we extract 3D input cubelets is derived from the 3D volume. We pass seismic data every 1250 x 1250 m and we blank all other traces in between. The 3D bin-size is 25 x 25 m.
In the direct approach, we train a 3D Unet with shape 128x128x128 samples to infill the missing data. In the flattening / unflattening approach, we flatten both the input volume and the target volume before extracting examples.
The flattening (Wheeler transformation) is done with a model-driven HorizonCube that we construct from 6 interpreted horizons. In this case we train a 3D Unet with shape 64x64x128 samples.
Both interpolation results are encouraging. The direct result is easy because it does not require any interpretation. However, the interpolated reflection patterns are less continuous than those obtained with the flattening / unflattening approach. Interested? Contact us at info@dgbes.com.
#machinelearning #OpendTect
