Machine Learning Approach to Predict the Illite Weight Percent of Unconventional Reservoirs from Well-Log Data: An Example from Montney Formation, NE British Columbia, Canada

Abstract

Shale mineralogy is critical for the proper design and execution of hydraulic fracturing operations and for evaluating production potential. There has been relatively little research into using artificial intelligence for mineralogical prediction for the Montney Formation. This study aims to predict the Montney Formation illite wt.% using readily available conventional logs, where illite is one of the constituents of shale and can aid in analyzing the brittle and ductile zones within the shale formation. The wt.% of illite is often determined by examining core samples or cuttings using XRD or QEMSCAN; both techniques are time-consuming, costly, and cannot be performed without physical samples. Based on conventional log readings, this study uses artificial neural networks (ANNs) and principal component analysis (PCA) to construct an accurate prediction model for illite wt.%. The feed-forward neural network (FFNN) obtained good overall performance in illite wt.% prediction (R2 = 92%) utilizing the backpropagation algorithm and the B.R. technique for eight input parameters. The ANN model was tested by randomly selecting three wells from the same log dataset excluded from the core training and testing phases. Overall, R2 = 88.5% was found in the tests, which is encouraging. This work demonstrates the viability of employing the ANN in applications involving evaluating mineralogical components of a target source rock using traditional logs, especially when geochemical data are missing or inadequate.