Using neural networks and the Markov Chain approach for facies analysis and prediction from well logs in the Precipice Sandstone and Evergreen Formation, Surat Basin, Australia

Abstract

Abstract Facies analysis is crucial for reservoir evaluation because the distribution of facies has significant impact on reservoir properties. Artificial Neural Networks (ANN) are a powerful way to use facies interpretations from core to determine equivalent facies from wireline logs in uncored wells. However, ANN do not incorporate information that relates to facies successions. This has limited the ability to effectively use facies information derived from logs alone in reservoir modelling, especially at the regional scale where data is often sparse, clustered, or incomplete. In this study, based on observations of 8 cored wells with a total thickness of ∼2000 m, 20 core facies were defined that range from 0.22 m to 11.56 m thick. Facies were based on grain size, sedimentary structures, and ichnological characteristics; Each facies corresponds to a distinct depositional sub-environment within the broader context of a large nearshore to shallow marine system. It was essential that these facies were incorporated into reservoir models to accurately map the distribution of reservoir and seal geobodies for CO2 storage assessment in the Surat Basin, Australia. However, core data are few and far between in the Surat Basin. To use core-defined facies in the absence of core, six wireline log parameters – gamma ray, density, sonic, neutron, photoelectric factor, and deep resistivity were plotted in multidimensional space and examined using Linear Discriminator Analysis. Combined with model recognition and Fisher Canonical Discriminance, the 20 core facies were simplified into 10 representative wireline log facies (WLF) with unique petrophysical parameters. We then used the Markov Chains Approach (MCA) to determine the significance of vertical facies transitions, which supported the interpretation that facies group into 5 distinct associations: (1) channel-levee complex; (2) lower delta plain; (3) subaqueous delta; (4) shoreface and; (5) tidal flats and channels. Based on the facies analysis and statistical classification, Multilayer Perceptron Classifier, a type of neural network method was applied using a training set of three cored wells that had all 6 wireline log data and using the facies successions determined from the MCA. Results show that the accuracy of WLF prediction ranges from 66% to 99% (ca. 83%). The accuracy of facies recognition decreased step wise with a decreasing number of logs as input data, such that when only gamma ray, density, deep resistivity, and sonic were used to train neural networks the accuracy dropped to between 45 and 98% (ca. 67%), depending upon the facies. This was considered the lowest acceptable threshold of accuracy for facies determination for input into reservoir models for carbon capture and storage. The results of this study show that sedimentary facies can be accurately predicted for uncored intervals in the Precipice Sandstone and Evergreen Formation to improve facies mapping and static reservoir modelling. Additionally, wireline log facies are helpful for interpreting Lower Jurassic stratigraphy, depositional setting, and basin evolution in the Mesozoic of Eastern Australia.