Bayesian Variational Methods in Carbon Storage Monitoring

Abstract

This thesis explores the use of deep learning and variational inference in Carbon Capture and Storage (CCS) monitoring. It consists of an introductory part and three scientific papers. The first chapter introduces CCS and techniques for monitoring of such sites from both a marine and subsurface perspective. The next chapter introduces the basic concepts of deep learning. Further, an introduction to Bayesian neural networks and variational inference as tools for assessing uncertainty in deep models is presented. The last chapter of the introductory part presents how dropout in a neural network can approximate variational inference and different variational auto-encoder models. In part two, three papers are presented that use variational inference to obtain uncertainty estimates in different applications. The first paper presents an algorithm that probabilistically classifies time series from a marine environment to determine if it arise from a CCS leakage incident and uses a Bayesian decision rule to optimally decide whether to initiate costly actions. The second paper presents a novel variational auto-encoder model for probabilistic reconstruction of flow fields given sparse measurements. The last paper presents a variational auto-encoder for reservoir monitoring. The proposed variational auto-encoder can both reconstruct the above zone monitoring interval (AZMI) pressure and classify the flux rate of the CO2 leakage given only the sparse observations from the AZMI wells.