Deep Neural Network Surrogate Flow Models for History Matching and Uncertainty Quantification

Abstract

History matching and uncertainty quantification in subsurface flow settings typically require large numbers of flow simulations. The use of surrogate models in place of high-fidelity simulations can lead to substantial reductions in computational cost. In this chapter, we discuss deep neural network surrogate models in detail. Two categories of deep-learning models, namely physics-informed and data-driven methods, are considered. One particular data-driven approach, the recurrent residual U-Net, is described and applied for an example involving oil-water flow in 3D channelized geomodels. The accuracy of this surrogate model for flow predictions with new realizations is demonstrated, and the method is then used in rejection-sampling-based history matching. The chapter concludes with a discussion of potential directions for future research on deep-learning surrogate models.