Enhance the Vertical Resolution of Conventional Well Logs Using Auto-Encoder

Abstract Conventional well logs typically have insufficient vertical resolution for thin-bed evaluation. In this study, we use an auto-encoder and one high-resolution log to enhance the vertical resolution of conventional logs within the same well for improved thin-bed evaluation. An auto-encoder is a generative artificial neural network capable of learning latent representations of the input data without supervision. First, we construct a suitable filter based on the tool response function of the low-resolution target log and apply this filter to the high-resolution source log to generate the corresponding low-resolution source log. We then use this low-resolution source log as input and the high-resolution source log as output to train the autoencoder. Finally, we apply the trained autoencoder, using the low-resolution target log as input to generate the high-resolution target log. Each low-resolution target log must be processed separately in the workflow due to different tool response functions. To validate the concept and workflow, we first used a synthetic dataset in which two high-resolution logs (conductivity and permittivity) were available. A specific filter was applied to the high-resolution permittivity log to create a low-resolution permittivity log, which served as the low-resolution target log. We then applied the workflow, feeding the low-resolution permittivity log into the trained auto-encoder to generate the predicted high-resolution target log. This was compared with the original high-resolution permittivity log, and they matched well. The workflow was subsequently applied to a thin-bed field dataset to enhance the vertical resolution of the conventional resistivity log using a high-resolution source log derived from borehole image measurements. Fluid saturations were then evaluated using the autoencoder-generated high-resolution resistivity and compared with those from the original resistivity log. The results show that the auto-encoder-generated log has higher vertical resolution than the original log and reads higher values in thin-bed sandstone zones, resulting in higher hydrocarbon saturation. By enhancing the resolution of conventional well logs with the autoencoder, we gain deeper insights into reservoir properties such as lithology, porosity, permeability, and fluid saturation. This enables better decision-making regarding field development and production optimization, potentially leading to increased hydrocarbon recovery and improved profitability. In contrast to the typical practice of feeding the same data into an auto-encoder as both input and output, in this study, we trained the auto-encoder with high-resolution well logs as output and corresponding low-resolution logs as input. This approach forces the model to learn the transformation logic from low resolution to high resolution. Enhancing the vertical resolution of conventional well logs through an auto-encoder provides a novel, effective, and efficient method for thin-bed analysis.