Lithofacies prediction from core images using Bayesian neural networks
Wei Xie, Jinyu Zhang, Kyle T Spikes · 2021
Machine learning techniques have gained much interest in reservoir characterization. However, the model prediction often lacks interpretability and reliability due to the ‘black-box’ nature of, for example, neural networks. In this paper, we present a method to identify core lithofacies and measure the uncertainty by U-net based Bayesian Neural Networks. This approach employs variational dropout and maximum a posterior (MAP) estimate, where we can quantify the uncertainties from both the data and model. We applied the method to predict the lithofacies of cores from the Wilcox Group, Gulf of Mexico. A prediction accuracy of 85% was obtained, and the associated uncertainties were evaluated. Although our current database is not complex and large enough to mimic the realworld scenario, it demonstrates the importance of understanding the uncertainties from data and model. Such uncertainties improve the data predictions and enhance the model reliability. They can be employed to quality control the training data and to improve the performance of the trained model.