IMPROVED MODELS OF THE SUBSURFACE USING DEEP LEARNING FOR GEOLOGICAL CORE CHARACTERIZATION

The development of a geologically realistic subsurface model generally requires integrating low-resolution yet areally extensive 3D reflection seismic data with high-resolution effectively ‘point source’ core data. In general, the seismic data provide constraints on the larger scale (roughly 10 m +) elements of the model including geometries and location of faults, and the core data provide constraints on the smaller scale (roughly 1 meter and less) elements including lithology and structural features such as veins, stylolites, fractures, and deformation bands.

Despite its importance, core characterization is often under-utilized as it is time and labor intensive, requires specialist skills, and often produces inconsistent results when done by different people and/or at different times. To help counter these limitations, we trained deep learning networks by manually labelling between 100 and 300 m of core images in multiple geological settings. When applied to test datasets, the accuracies of the predictions from these networks generally exceeded 70% and were as high as 90%.

The application of this technique is not intended to produce perfect results, but rather to create a consistent high-level characterization of all available core very quickly in a model area, allowing for identification of lithological and structural trends, aiding core-seismic calibration, and identifying portions of the core that may require manual interpretation or validation by specialists. In short, it provides geoscientists more time to think about the broader geological ‘picture’ in the area of their subsurface model by helping to ensure that they take advantage of all available data.