A Bi-Directional Long Short-Term Memory Neural Network for Geologic-Facies Classification

Abstract

Facies interpretation from well logs is a common, oftentimes ambiguous, and laborious task during reservoir characterization. Although machine-learning methods for facies interpretation have been used for more than 30 years, only recently have computer-assisted methods approached human-level performance owing to advances in processing power and the availability of large, high-quality petrophysical datasets. In many prior studies, independent training (and testing) samples at specified depths in individual or multiple wells did not account for cyclic facies alterations. In this deep-learning approach, cyclic stratal information is accounted for, and cyclic stratal correlations are identified and stored using a bi-directional long short-term memory (BLSTM) network. To overcome over-fitting, AdaGrad, early stopping, and k-fold cross-validation were implemented. This new approach was tested on well data from the Hugoton and Panoma fields of western Kansas and Oklahoma that were used in the 2016 machine-learning competition sponsored by the Society of Exploration Geophysicists and Enthought Inc. After hyperparameter tuning, the median accuracy from 100 realizations predicting facies in two blind wells was 0.700. This value eclipsed the contest winner’s value of 0.639, which was the most accurate of 300 submitted solutions. The success of this BLSTM approach highlights the robustness of a well-designed deep-learning method and the importance of integrating domain knowledge into machine learning techniques.

AAPG Datapages/Search and Discovery Article #90350 © 2019 AAPG Annual Convention and Exhibition, San Antonio, Texas, May 19-22, 2019