Accelerating Reservoir Modeling Workflows With Neural Operators and GPU-Based Full-Physics Simulations on the Cloud

Abstract Large-scale applications in reservoir engineering, such as history matching, uncertainty quantification, well placement optimization, and reserve estimation, require numerous simulations of high-fidelity models. Despite rapid advances in high-performance computing, these simulations are computationally intensive and challenging. Integration of scientific machine learning (ML) techniques with high-fidelity full-physics computational models for developing high quality proxy/surrogate models offers a promising solution to overcome traditional challenges. This work presents a workflow for generating full-field proxy models for predicting the time-evolving state variables (pressure and saturations). This is achieved by integrating a state-of-the-art ML framework with a GPU-accelerated full-physics simulator, the latter to help generate large amounts of training data in the fastest possible time. For the ML part, we specifically utilize Fourier Neural Operators (FNOs) for better generalization. We demonstrate the effectiveness of our framework using two synthetic models and varying the positions of the wells, mimicking well placement optimization scenarios. Both the models considered are 2-phase oil-water systems with one injector well and four producer wells. The size of the grid is 64×64×64. For the first model, the permeability and porosity fields are assumed homogeneous while the second model has large variations in porosity and permeability inside the reservoir. Our results show that FNO-based ML approach holds promise for addressing full-field proxy generations for challenging models with complex variations in saturation front topologies. The total water injection rates and oil production rates were accurately captured by the current model. The study showed based on the complexity of the model, selecting the right number of Fourier modes is critical for ensuring accuracy. While using a lower number of modes may resulted in smoothing of the saturation front that may not always be desirable, a higher number of modes resulted in small and non-physical bounded (stable) variations along the saturation fronts.