Knowledge-based machine learning for mineral classification in a complex tectonic regime of Yingxiu-Beichuan fault zone, Sichuan basin

Abstract

The present study aims to better understand the mineralogy and thermal structure of the Yingxiu-Beichuan fault zone (YBFZ), Sichuan basin, China, which was lacking previously. The previous research on the Wenchuan earthquake Fault Scientific Drilling (WFSD) project was focused on mineral classification, fault analysis, and geochemical research utilizing original logs using WFSD-1 well and at a shallow depth of 700 meters. No investigations are conducted at a depth of around 1550 meters utilizing multiple dimensionality reduced well-logs. Thereafter, we sought to categorize the minerals along the YBFZ using Machine Learning (ML) technologies and concentration-number (C–N) modeling using multiple WFSD-1 and WFSD-2 wells. In the categorization of rocks, three classifiers are discussed: Support Vector Machines (SVM), Feed Forward Back Propagation (BPNN), and Radial Basis Function Neural Networks (RBFN). Linear Discriminant Analysis (LDA) and Principal Component Analysis (PCA) are used to normalize the log data in a geologically complicated region. By using the BPNN classifier to PCA, LDA, and all 17 logs, we achieved respective accuracy rates of 75.54%, 86.32%, and 72.54%, which are higher than the accuracy rates of RBFN (76.03%, 77.01%, and 65.3%), and SVM (74.45%, 78.03%, and 70.06%). These results suggest that BPNN shows improved accuracy rates for mineral classification in a complex tectonic regime. In addition, concentration-number (C–N) fractal model technique and log-log plots are also used to characterize geothermal features and the results of C–N modeling supports the results of ML models. High gamma ray (GR) ranges of 349.7API, heat production (HP) ranges of 5.5 μW/m3, and low thermal conductivity (TC) ranges of 0.28 W/km show that the fault region is associated with comparatively strong radiogenic activity compared to its surroundings.