Prediction of fluid oil and gas volumes of shales with a deep learning model and its application to the Bakken and Marcellus shales

Abstract

The fluid oil and gas volumes (S1) retained within the shales are one of the most important parameter of producible fluid oil and gas saturations of shales together with total organic carbon content. The S1 volumes can directly be obtained by Rock-Eval pyrolysis analysis. However, it is time consuming and not practical to obtain samples from all intervals of all wells in any shale play. S1 volumes prediction with a deep learning (DL) model have increasingly became important with the booming exploration and development of shale oil and gas resources. S1 volumes of shales are controlled by organic matter richness, type and maturity together with reservoir quality and adsorption capacity which are mainly effected by age, depth, organic content, maturity and mineralogy. A dataset consisting of 331 samples from 19 wells of various locations of the world-class organic-rich shales of the Niobrara, Eagle Ford, Barnett, Haynesville, Woodford, Vaca Muerta and Dadaş has been used to determination of a DL model for S1 volumes prediction using Python 3 programing environment with Tensorflow and Keras open-source libraries. The DL model that contains 5 dense layers and, 1024, 512, 256, 128 and 128 neurons has been predicted S1 volumes of shales as high as R2 = 0.97 from the standard petroleum E&P activities. The DL model has also successfully been applied to S1 volumes prediction of the Bakken and Marcellus shales of the North America. The prediction of the S1 volumes show that the shales have lower to higher reservoir quality and, oil and gas production rate that are well-matches with former studies.