AI to Improve the Reliability and Reproducibility of Descriptive Data: A Case Study Using Convolutional Neural Networks to Recognize Carbonate Facies in Cores
Geology has traditionally been a descriptive science with a significant portion of the data coming from observations of features at a range of scales. Modern practices in the oil industry still rely in a large part on this legacy of observational data, for instance when rock facies are used to derive regional stratigraphic trends from core data, or as a building block for petrophysical classifications. However, a recent study has shown that even experienced carbonate sedimentologists will often classify the same facies using different textural names. This problem is compounded in industry by large teams often collaborating on a project, resulting in a heterogeneous attribution of facies to similar rocks despite the use of a common classification scheme. This problem reduces the reliability of descriptive data.Here, we present the first results of machine learning applied to automatic identification of carbonate facies using the Dunham classification scheme. We used high-resolution core images from the Integrated Ocean Discovery Program (IODP) Leg 194. Core images were used to train a model written in the Python programming language using the TensorFlow machine learning library. Specifically, we used Google’s Inception V3 network as a pre-trained Convolutional Neural Network (CNNs), and applied a method called ‘transfer learning’ to train Inception V3 to recognize carbonate core images. Results show that our CNN can achieve up to 90% accuracy for identification of Mudstone to Rudstone and Crystalline Dolomite. The main misclassifications were between matrix and grain supported facies, and fine and coarse-grained facies, textures also commonly misclassified by control tests with geologists. Interestingly, the bias observed in core description by the algorithm is very similar to human biases: a tendency to give a greater weight to grains as they stand out from the matrix, called ‘saliency’. But the CNNs were able to identify facies 60 times faster than humans, and with a much greater consistency. The results of our study demonstrate the potential of artificial neural networks to reliably interpret and quantify descriptive data for the oil and gas industry, in a fast, automated, high-resolution manner. Current and future work will focus on acquiring a larger dataset of core and thin section images, improving the training of the neural network, and coupling image recognition with logging and petrophysical data estimation.
AAPG Datapages/Search and Discovery Article #90350 © 2019 AAPG Annual Convention and Exhibition, San Antonio, Texas, May 19-22, 2019