Reservoir Facies Classification using Convolutional Neural Networks

Abstract

Distribution of Reservoir facies classes in a hydrocarbon reservoir is an important aspect for geologists to decide the location to drill a production well. Facies classes are based on rock characteristics that can indicate the locations of good quality sand, shale and hence the presence of hydrocarbon. However, due to the heterogeneous nature of earth subsurface, gathering facies information is challenging. Seismic data when properly correlated with the facies classes at the well locations can be suitable for effectively generating the facies information across the reservoir. Machine learning models had a great contribution in this field to successfully understand the reservoir characteristics from seismic data. However, advanced modelling techniques like deep learning still have not been explored in depth. In this paper, we put emphasis to understand the relation of seismic and facies classes using Convolution Neural Network (CNN) models, a popular deep learning model that takes into consideration the spatial relation present in the input data. CNN variations are adapted to improve the mapping relation of seismic to facies classes. When the models are compared with the traditional Machine Learning models, the CNN provided improved mapping relation as compared to the conventional models. More specifically, ResNet, an adaptation of CNN architecture with residual connections, outperformed all the other models with a Precision, Sensitivity and Specificity of 0.61, 0.69 and 0.84 respectively.