Grid-Based Surrogate Reservoir Modeling (SRM) for Fast Track Analysis of Numerical Reservoir Simulation Models at the Grid block Level

Abstract Developing proxy models has a long history in our industry. Proxy models provide fast approximated solutions that substitute large numerical simulation models. They serve specific useful purposes such as assisted history matching and production/injection optimization. Most common proxy models are either reduced models or response surfaces. While the former accomplishes the run-time speed by grossly approximating the problem the latter accomplishes it by grossly approximating the solution space. Nevertheless, they are routinely developed and used in order to generate fast solutions to changes in the input space. Regardless of the type of model simplifications that is used, these conventional proxy models can only provide, at best, responses at the well locations, i.e. pressure or rate profiles at the well. In this paper we present application of a new approach to building proxy models. This method has one major difference with the traditional proxy models. It has the capability of replicating the results of the numerical simulation models, away from the wellbores. The method is called Grid-Based Surrogate Reservoir Model (SRM) since it is has the unique capability of being able to replicate the pressure and saturation distribution throughout the reservoir at the grid block level, and at each time step, with reasonable accuracy. Grid-Based SRM performs this task at high speed, when compared with conventional numerical simulators such as those currently in use (commercial and in-house) in our industry. To demonstrate the capabilities of Grid-Based SRM, its application to three reservoir simulation models are presented. Fist is a giant oil field in the Middle East with a large number of producers, second, to a CO2 sequestration project in Australia, and finally to a numerical simulation study of potential carbon storage site in the United States. The numerical reservoir simulation models are developed using two of the most commonly used commercial simulators1. Two of the models presented in this manuscript are consisted of hundreds of thousands of grid blocks and one includes close to a million cells. The Grid-based SRM that learns and replicates the fluid flow through these reservoirs can open new doors in reservoir modeling by providing the means for extended study of reservoir behavior with minimal computational cost. Surrogate Reservoir Modeling (SRM) is classified as an AI-Based reservoir model (Mohaghegh, 2011) referring to a process that accomplishes the task of proxy modeling by learning the specific behavior of a numerical reservoir simulation model through training on a uniquely developed spatio-temporal dataset. The spatio-temporal dataset is developed for each model using only a handful of simulation runs.