GSA Annual Meeting in Phoenix, Arizona, USA - 2019

Paper No. 286-3
Presentation Time: 2:10 PM


AUCHTER, Neal1, FALIVENE, Oriol1, KLEIPOOL, Luuk2 and ESPEJO, Irene1, (1)Interpretation Technology, Shell International Exploration and Production Inc., 3333 Highway 6 South, Houston, TX 77082, (2)Interpretation Technology, Shell Global Solutions International, Amsterdam

Cored intervals of subsurface stratigraphy offer a high resolution, empirical, 1D representation of the subsurface. As such, core is a fundamental component for many subsurface studies because it provides detailed information about depositional processes, stratigraphic stacking, and is used to calibrate derivative subsurface data such as petrophysical logs and seismic surveys. Core description, and more importantly interpretation, is the process of converting a section of rock acquired from the subsurface into useable subsurface data. In industry, this process is commonly performed by a specialist geoscientist, which tends to make it expensive, time-consuming, and biased by individual experience. These challenges, along with the repetitive nature of the task, make it an appealing option for automation using machine learning. Herein, we test the efficacy of deep learning convolutional neural networks (CNNs) for generating first-pass core descriptions from core photos.

We present initial results characterizing deep-water turbidite deposits from two distinct basins. We assembled a database of core images and lithofacies descriptions from multiple wells totaling >300 meters for each basin. These data, together with several data augmentation strategies, were used to train the CNNs. To improve predictive performance, we used multiple CNNs organized hierarchically, each one optimized with transfer learning. Each CNN identifies a specific characteristic of the core, such as lithology or structure, as well as intervals in which the core is not suitable for description. To test the predictive capabilities of the CNNs with new well data, wells from each dataset not used for training were systematically validated. Predictive capabilities vary with the amount and distinctiveness of training data for each lithofacies category. However, accuracies on the order of 80% or more were achieved for all lithofacies, and these can likely be improved by increasing the quality and amount of training data. These results corroborate the idea that successfully trained CNNs can be used to expedite description of new cores and digitize legacy core descriptions, thereby alleviating specialist geoscientist resourcing. Standardization of core description databases will facilitate more objective stratigraphic analyses at basin and global scales.