Machine learning methods for uncertainty quantification in subsurface reservoirs

Abstract

We investigate current challenges in the reservoir engineering pipeline that can be addressed using recent machine learning techniques. Our emphasis is on improving the performance of uncertainty quantification tasks which are ubiquitous in subsurface reservoir simulations. In one work, we accelerate multiscale methods by embedding a neural network surrogate for the fast computation of the custom basis functions, replacing the need to solve the local elliptic problems normally required to obtain them. In a different work, we address current challenges in obtaining geological parametrizations that can capture complex geological structures. We adopt a neural network parametrization using a recent unsupervised learning technique, obtaining an effective parametrization that can reproduce high-order statistics of flow responses. In a follow-up work, we introduce a method for post-hoc conditioning of the neural network parametrization to generate conditional realizations by training a second neural network to sample from a Bayesian posterior and coupling it with the original network. In our final work, we introduce a framework for exemplar-based parametric synthesis of geological images based on a recent kernel method, obtaining a neural network parametrization of the geology using a single exemplar image.