Generating pseudo well logs for a part of the upper Bakken using recurrent neural networks

To develop a reservoir, we need to understand the distribution of key reservoir properties. Those formation properties are mostly derived from well log data. However, obtaining well log data is so expensive that the data we acquire is never sufficient. Therefore, we need to create pseudo data from real measured data. One approach is to generate synthetic well logs using recurrent neural networks (RNNs). Though relatively nascent, it is proven that RNNs allow for the generation of well logs with reasonable accuracy at a fraction of the price. Currently, gamma ray is the most cultivated well log in the E&P process today (owing to its decisive role during drilling operations and affordable price). All the same, a more extensive log suite is required to comprehend the producibility and frackability of hydrocarbon reservoirs like shale reservoirs. In this study, an upper Bakken database of 97 wells which contain gamma ray, bulk density, photoelectric index, density porosity, neutron porosity, true resistivity, compressional and shear slowness was used to develop deep learning models apt to creating synthetic logs for another set of 40 wells in the upper Bakken. The latter wells only had measured gamma ray. Using petroleum engineering concepts and RNNs, data-driven models capable of generating the remaining logs were created. The data-driven models were validated using wells with logs that have been removed from the database to serve as blind validation wells. The performance of the models (5–10 Mean Average Percentage Error and R-squared of 0.6–0.8) suggests the workflow described in this study is a viable way to maximize the value of a gamma ray log. With the ability to generate logs of similar quality for all wells in a shale asset (starting only with gamma ray), superior reservoir evaluation and thus improved reservoir management can be achieved.