Diagenetic Facies Classification in the Arbuckle Formation Using Deep Neural Networks

• Published in: Mathematical geosciences Vol. 53; no. 7; pp. 1491 - 1512
• Main Authors: Deng, Tianqi, Xu, Chicheng, Lang, Xiaozheng, Doveton, John
• Format: Journal Article
• Language: English
• Published: Berlin/Heidelberg Springer Berlin Heidelberg 01.10.2021; Springer Nature B.V
• Subjects: Accuracy; Artificial neural networks; Cambrian; Carbonates; Chemistry and Earth Sciences; Classification; Computer Science; Diagenesis; Dissolution; Dissolving; Earth and Environmental Science; Earth Sciences; Gamma rays; Geotechnical Engineering & Applied Earth Sciences; Hydrogeology; Identification; Image classification; Labels; Learning algorithms; Logging; Long short-term memory; Machine learning; Magnetic resonance; Neural networks; NMR; Nuclear magnetic resonance; Ordovician; Permeability; Photoelectricity; Physics; Porosity; Predictions; Reservoirs; Sedimentary facies; Statistics for Engineering; Well logs; Diagenetic facies classification; Neural networks; Machine learning
• ISSN: 1874-8961; 1874-8953
• DOI: 10.1007/s11004-021-09918-0

Dissolution is a common diagenetic effect in carbonate formations. Vugs caused by dissolution significantly impact carbonate reservoir quality by affecting the porosity and permeability of the reservoir. However, without core and image logs, the identification and classification of vugs using wireline logs only is challenging, because logging tool responses reflect a mixed effect of changing mineral/fluid composition and diagenetic features. This paper presents a data-driven approach using neural networks to identify vugs and classify vug facies based on vug size. The purpose is to predict wells/intervals with limited measurements by machine learning models trained with core data from key wells. The input features for vug identification are conventional well logs (i.e., gamma ray, resistivity, neutron/density porosity, photoelectric factor, and acoustic slowness) from the Cambrian-Ordovician Arbuckle formation, Kansas. Two classification labels are used as the prediction target for the neural networks: (1) a binary vuggy index derived from nuclear magnetic resonance (NMR) measurements using a cutoff on T2 distribution, which presents the proportion of large pores over the total porosity, and (2) vug size labels from depth-by-depth core visual descriptions. A one-hidden-layer shallow neural network is compared against deep neural networks, including structures such as one-dimensional convolutional layers (1-D CNN) and long short-term memory (LSTM) layers. Results suggest that using a combination of multi-mineral analysis results and original well logs will increase the prediction accuracy of vug facies. Shallow and deep neural networks show a similar ability to identify vugs, with average accuracy of around 80%. However, to predict vug-size-based facies labels, deep neural networks outperform shallow neural networks, with overall accuracy improved by as much as 10%. The proposed method shows that deep neural networks (1-D CNN and LSTM) are reliable tools for vug facies prediction.