TabTransformer model for facies classification

Abstract Geoscientists now have access to modern technology and a wide range of open-source tools for reservoir characterization, including facies classification. Successful oil and gas field exploration and development depend heavily on accurate facies classification. This study focuses on facies classification using data from the Council Grove gas reservoir, located in the shallow marine environment of the Hugoton and Panoma Fields, Southwest Kansas. While machine learning (ML) algorithms such as Multi-Layer Perceptron (MLP), Gradient Boosted Decision Trees (GBDT), and Support Vector Machines (SVM) are commonly used for facies classification, they often face challenges such as noise sensitivity and suboptimal performance in semi-supervised learning. This study proposes the use of a transformer model, specifically the TabTransformer, which has shown promise in natural language processing and classification tasks. The TabTransformer applies self-attention mechanisms and uses parametric embeddings to transform categorical and numerical data into a contextualized form for classification. The TabTransformer achieved an accuracy of 83% in classifying facies, surpassing IBM Watson AutoAI models (Accuracy: 82% for Extra Trees Classifier, 79% for XGBoost). Additional accuracy measures, such as the area under the curve (AUC) and the F1 score, showed an AUC above 86% and an F1 score of 81%. However, the model’s generalization on blind data was 63%, which is attributed to limited training data. The TabTransformer model demonstrates significant potential for facies classification, particularly in shallow marine environments where rock coring is challenging. With sufficient data and the right optimization, the model can achieve optimal results for predicting reservoir rocks.