An artificial intelligence method for improving upscaling in complex reservoirs

Abstract

The upscaling of geological properties is a fundamental requirement to construct a suitable simulation model since the geological model typically contains millions of cells, which makes it computationally difficult to simulate it on such a scale. The process consists of a scale transfer that adapts the petrophysical properties of a high-resolution grid to a coarser grid. Nevertheless, this is not a trivial operation, especially for absolute permeability, which is a non-additive property. One of the most reliable methods is to perform a flow simulation on fine-scale cells that correspond to the coarse block and derive the single permeability value that reflects the same flow value. Still, this is a very time-consuming method and depends on the imposed boundary conditions. This work aims to take advantage of recent advances in artificial intelligence to produce results with similar or better quality of flow-based methods, and more adaptive. It handles models with multiple geological realizations by employing machine learning to capture patterns from a subset of scenarios and use it to generalize for all others. The methodology is divided into two stages — local and global. In the first one, a deep neural network is trained with a fraction of geological realizations using the flow-based results as a reference. In the second stage, clustering analysis is employed along with a neural network optimization to learn an adjustment procedure for the coarse simulation model concerning the cumulative field production predicted by the fine model simulation. Afterward, the trained network performs the upscaling of remaining realizations, but more efficiently in terms of computational time and providing better output in terms of production and injection. The method was applied to a benchmark model and the experimental results demonstrate that the local technique was capable of reproducing very similar values to the flow-based upscaling using only one scenario in training and at the global stage the coarse model was improved even further matching the field oil production forecast of the fine model. Different scenarios were used for training and testing and the results were consistent showing no bias towards a specific configuration and capability of generalization to different scenarios. Ultimately, the proposed artificial intelligence approach performed an accurate upscaling, surpassing the reference approach with forecast production similar to the fine model, and also fast to compute when considering multiple geological realizations, since it reduces the required numerical simulations to a fraction of the total.