Evaluation of Automated Lithology Classification Architectures using Highly-Sampled Wireline Logs for Coal Exploration

Abstract

Wireline logs are a supplemental data source to conventional core logging. The recent explosion of machine learning algorithms has provided researchers with ample opportunity to develop automated statistical tools for classifying lithology from wireline logs, which geologists can use to produce first-pass interpretations or to validate existing interpretations. Such automated interpretations can be particularly valuable information in the case of missing or damaged core samples. There exists, however, a need to evaluate said machine learning algorithms in the case where available wireline logs contain a wide range of different logs which are highly-sampled. This paper explores different machine learning algorithms and architectures for lithology classification using wireline data from project area Jundah East, 60 km north-west of Wandoan, Queensland, which is well known for coal mineralisation. We used seven well logs each containing 19 wireline logs sampled at 1 cm-1, available through the Queensland Digital Exploration (QDEX) data system. Three popular supervised machine learners, namely the Naïve Bayes classifier, Support Vector Machine, and Multilayer Perceptron (an artificial neural network), are tested under two architectures: committee (one classifier per well log) and singular (one classifier for all well logs). The results show the Naïve Bayes classifier, although computationally simple, achieves good results in general when training using a committee architecture on a large data set. For coal classification in particular, it achieved the sensitivity score of 0.79 and the specificity score of 0.97. While the committee and singular architectures generated similar results, the committee architecture provides the benefits of faster computation time as well as a flexible platform for the training of additional well logs.