ABSTRACT Machine learning (ML) and artificial intelligence (AI) are rapidly advancing approaches with wide‐ranging applications in petroleum geosciences and energy resource exploration. Artificial neural networks (ANNs) are now increasingly used to address complex geoscientific problems with high accuracy. The key is to design and apply the appropriate ML algorithm to achieve the target result. The focus must be on the user's experience and the domain knowledge; that is, the ANN should be trained and guided by human intelligence. Among many algorithms, feed‐backward neural networks (FBNNs) have been proven to yield accurate results and accomplish complex tasks while optimizing time and cost. The Kohonen self‐organizing maps (SOMs) are among the most successful tools for pattern recognition and feature mapping and can resolve many issues related to complex geological features. The descriptive nature of the sedimentological data acquired from borehole images (BHIs) or whole‐rock cores made them very expensive and rarely available in all field wells. Therefore, a practical approach is required to get the learning from the available data in a few wells and predict it in the rest of the field wells. The FBNN is the best candidate for designing this workflow because of its adaptive, iterative nature. This article reviews the theory behind ANN algorithms and presents an example from the NEAG 2 Field in the North East Abu Gharadig Basin in the context of ANN applications for classifying and predicting reservoir facies. The facies associations described from the BHI available in three field wells, have a specific texture and composition that cannot be differentiated by conventional models, as there is no clear difference in their log responses to be indexed. To overcome these challenges, the 27 BHI‐described facies are grouped into four definitive facies groups.The most indicative petrophysical logs are selected to be used in the ANN model based on the Principal component analysis (PCA) which is then validated by star plots and a scatter plot matrix (SPLOM). Supervised by the BHI‐described facies, the ANN gets its training and learning to classify the facies and assign a log response combination for each facies type. A blind well test validated the facies grouping geologic concept and the accuracy of the ANN model. The facies classification task generated electrofacies curves for the three indexed wells, which were then used to predict the curves for the other wells in the field, as a compensation for the data gaps. These results allowed for the distribution of the different reservoir facies within a 3D multi‐facies geological model and provided a practical solution to the common geologic challenges.
Rashad et al. (Tue,) studied this question.