Fast evaluation of pressure and saturation predictions with a deep learning surrogate flow model

Numerical models for flow through porous media are essential for forecasting subsurface fluid flow response to support optimal decision-making to develop subsurface resources such as groundwater, geothermal, and oil and gas. For reservoir engineering, numerical flow simulation modeling is applied to support and maximize well ultimate recovery and recoverable reserves to maximize project economics and safety while minimizing environmental impacts. Subsurface models explore subsurface uncertainty based on integrating geological, geophysical, petrophysical, and reservoir engineering data and expert interpretations. To evaluate uncertainty, we rely on multiple geostatistical subsurface heterogeneity realizations paired with flow simulation forecasts to test the sensitivity of various reservoir development parameters and build an uncertainty model of reservoir performance. However, with large reservoir models, numerical flow simulation time increases, leading to a significant amount of professional time and computational effort that increases project costs, and increases cycle times resulting in delay or diminished decision quality. This issue motivates the utilization of surrogate, computationally efficient approximative flow models. Current methods focus on prediction accuracy and minimizing prediction error. When uncertainty is significant, prediction accuracy is insufficient, and we must consider the entire uncertainty distribution. We propose a new and general workflow to generate accurate and precise machine learning-based surrogate flow models to predict the relationship between the reservoir rock and fluids, development parameters, and the reservoir flow simulation responses. We train a deep convolutional neural network using the results from a threedimensional two-phase flow simulator. Next, we use the trained model to generate ensemble predictions from the flow surrogate to evaluate the uncertainty based on the development parameters and geological information. The machine learning-based surrogate model uses exhaustive subsurface predictor features, porosity, permeability, well position, and an engineered feature representing non-dimensional time as input to predict exhaustive subsurface response features, pressure, and saturation distribution over discrete time steps as the output. The proposed workflow integrates the spatiotemporal aspect of subsurface flow modeling. It allows the practitioner engineer to explore subsurface uncertainty without the entire computational cost of numerical flow simulation to support optimum, timely development decision-making.