In reservoir simulation, geological models are constructed using data such as well and seismic survey data, consisting of millions of grid blocks to capture small scale reservoir heterogeneity. This makes numerical reservoir simulations time-consuming. Therefore, upscaling, which merges multiple grid blocks into a fewer-blocks representing small scale heterogeneity, is essential. However, maintaining the original model's heterogeneity during upscaling is challenging. Over the past two decades, various methods such as renormalization, pressure-solver, and tensor permeability techniques have been proposed. These methods often involve complex computation, like solving partial differential equations (PDEs) or conducting pre-dynamic simulations, making upscaling very time-consuming and complicated.
To overcome these deficiencies, this study focused on applying deep learning, a powerful tool in the field of Artificial Intelligence (AI), to upscaling of both the single-phase property(absolute permeability) and the multi-phase property (relative permeability). First, a large dataset was generated for training AI, comprising fine grid reservoir models with two-phase fluids (water and oil) in 2- dimensions and various reservoir properties such as absolute and relative permeability. To ensure dataset diversity, absolute permeabilities were assigned following a normal distribution, while relative permeabilities were distributed to follow normal distribution patterns for the end point and curvature values.
Next, to ensure the consistency in predicted reservoir behavior between fine grid and coarse grid models, the absolute and relative permeabilities of coarse grid models were optimized using the Differential Evolution (DE) method, a numerical optimization technique. Subsequently, AI was trained to learn the relationships between fine grid properties and optimized coarse grid properties using Deep Neural Networks (DNN) and Convolutional Neural Networks (CNN). Various datasets were used to demonstrate the potential of applying deep learning to upscaling.
As a result, it was revealed that DNN and CNN effectively predicted absolute and relative permeabilities with high accuracy in a short time. This indicates that the deep learning could be a promising approach for automating the upscaling process, significantly reducing the time and complexity involved while maintaining the necessary heterogeneity characteristics expressed in the original geological model.
