Accelerating core characterization and interpretation through deep learning with an application to legacy data sets

Abstract

Core descriptions provide an essential element of ground truth to subsurface reservoir characterizations, but adequate core characterization and interpretation require a significant time commitment from geoscientists because they are usually done manually. Moreover, core characterizations and interpretations often are inconsistent because they are made at different times by different people. We have used deep-learning segmentation networks, trained using worldwide data sets labeled by specialists, to infer lithofacies and structural features on core images. These networks are not intended to provide perfect core descriptions, but rather to provide high-level fully digital characterizations that can be quickly checked for quality and manually fixed if needed. This ensures that all available and relevant data can be used and relieves geoscientists from the most repetitive characterization tasks to allow them to focus on more complex and value-adding integration and interpretation. We test the previously trained networks in legacy public data from a UK field as an example. There are more than 1300 m of released core images in the field, together with multiple vintages of fault interpretations, that significantly differ from one another. Integrating core and seismic-scale structural characterizations is critical because the best reservoir facies are the most prone to deform via permeability-reducing processes. Detailed structural characterization and identification of individual structural features achieved through deep learning provide an unprecedented amount of data to highlight areas where the seismic-scale fault interpretations and core-scale structural interpretations are inconsistent, yielding valuable insights to validate fault interpretation.