Video Frame Interpolation Approach for Reservoir Simulation Visualization Utilizing Stable Diffusion Model

Abstract Reservoir simulation is a fundamental yet computationally intensive task in oil and gas exploration and production. It involves solving complex mathematical equations over time to predict the behavior of subsurface fluid flow under varying geological and operational conditions. This task is essential for evaluating production strategies, optimizing field development, and making informed operational decisions. To operate these complex systems, engineers leverage high-performance computers to execute such simulations at high spatial and temporal resolutions to handle the substantial computational power required. Moreover, this task also demands high storage capacity to store 3D reservoir models, which comprise a large number of grids representing the reservoir's geological structure, as well as the dynamic properties of the fluid, including water and oil saturation, pressure, and temperature, among others. In this work, we present a novel approach that utilizes the concept of video frame interpolation, a well-known problem in the computer vision field, to enhance reservoir simulation capabilities by increasing its resolution without affecting computational or storage capacity. The objective of this study is to improve simulation results with minimal computational processing. Our proposed approach employs a type of Generative AI Model, specifically a Diffusion Model, which utilizes a U-net architecture as its backbone to train the model. The results of our framework demonstrate remarkable alignment with outputs from full-scale simulation runs, achieving visually and numerically consistent results with only minimal deviations.