In today’s post we share an elegant workflow proposed by Roar Heggland of Equinor ASA to compute Fault Dip and Fault Azimuth for Machine Learning predicted Fault Likelihood (ML-FL).
ML-FL produces Fault Likelihood in a very fast and easy way. It is a one-click workflow in OpendTect’s Machine Learning solution. To compute ML-FL you simply select the pre-trained Unet of shape 128x128x128 from the library of pre-trained models and apply this to your 3D seismic volume. ML-FL is often cleaner than Fault Likelihood computed with the original algorithm. However, whereas Fault Likelihood also outputs Fault Dip and Fault Azimuth, ML-FL only outputs Fault Likelihood itself. This creates a problem for OpendTect’s Fault Thinning algorithm and for Automatic Fault plane Extraction. Both algorithms in the Faults & Fractures plugin need all three components as input.
To compute the missing components, Roar simply applies the Fault Likelihood algorithm to a new volume: “1 minus ML-FL”. The idea behind this is that ML-FL separates faults and none faults much better than a seismic volume does. Therefore, we expect better hits in the Fault Likelihood scanning process with ML-FL input than with seismic input. We input “1 minus ML-FL” because faults in seismic data correspond to low semblance values, not high values as in ML-FL.
The slider shows an example of ML-FL, Thinned ML-FL and Automatic Fault planes Extracted from ML-FL. The missing Dip and Azimuth components for ML-FL were computed with the workflow described above.