Application of GAN to Resolution Enhancement of LWD Real-Time Image Logs to Support Decision Making

Abstract

In the current scenario of project management, where the agility and optimization of operations have been prioritized, the practice of logging while drilling (LWD) has gained space compared to traditional wireline logging. In theory, acquiring quality petrophysical properties during drilling brings greater agility in decision making about completion and optimizes operation costs. However, regarding borehole image logs, due to limitations in transmission capacity, the actual available data in real time contain about 50% (for resistivity images) of the full azimuth information, being insufficient for the identification of critical geological structures capable of impacting the communication between production or injection zones or the quality of cementation, such as fractures, caves, and geomechanical collapse zones. The tool’s memory data with the full information may take a few days after the end of drilling to be delivered by the service company, which in some cases is not enough for fast decision making regarding completion. In this work, we tested models based on generative adversarial neural networks (GANs) to reconstruct the complete memory data based on real-time input. As in conventional GAN schemes, a generator is trained to receive a real-time input and create a “memory-like” image, while a discriminator is trained to tell real and fake images apart. To regularize the convergence of training, we used an architecture known in the literature as CycleGAN, where another generator-discriminator pair is trained simultaneously to do the reverse process, recreating the real-time data. Variations of the training process and data sets were used to generate different CycleGAN models. They were trained using logs of presalt reservoirs in Buzios Field, and performance was assessed on logging intervals not seen by the algorithms during training. The results achieved so far have been very promising, as in certain intervals, resultant models were able to capture the presence of fractures and caves. This methodology represents a way of circumventing telemetry limitations, where missing information is added indirectly to the real-time data as the artificial intelligence (AI) algorithm learns the main characteristics of a field/reservoir. Therefore, previous knowledge from the field can be used to continuously optimize future operations, efficiently incorporating the available database into the workflow of petrophysicists for the recognition of geological and geomechanical structures in time to support decision making in completion operations.