Development of Fast Machine Learning-based Proxy Models for Subsurface Carbon Storage: A Comprehensive Study of the Versatility of Standard Algorithms

Abstract:

Carbon capture, utilization, and storage (CCUS) is a promising technology for reducing atmospheric emissions of anthropogenic CO₂. To maximize the benefits of this technology for slowing climate change, operators are racing to balance the technical and economic viability issues attending large-scale deployment of CCUS. The economic burden imposed by large-scale implementation of CCUS over many decades into the future underpins the need for informed decisions, but current industry approaches are insufficient for rapid forecasting of subsurface storage and related impacts associated with CCUS. Machine learning techniques are a very attractive option capable of accelerating real-time decisions in subsurface storage of anthropogenic CO₂. In this contribution, we present a study of widely used off-the-shelf supervised machine learning algorithms to assess the limits of their applicability to aid forecasting of subsurface flow processes. Two-dimensional and three-dimensional synthetic data sets generated by numerical reservoir simulation and representing both homogeneous and heterogeneous reservoirs were used with CO₂ pressure plume, saturation plume, and water extraction rate as state variables to be predicted. The algorithms evaluated include multilayer perceptron neural network (MLP), convolutional neural network (CNN), long short-term memory (LSTM), and gated recurrent unit (GRU). The hyperparameters that define the architecture of these networks and how they learned patterns in the data were optimized and the incremental gain in training speed was investigated using high-performance GPUs. All the algorithms investigated predicted the data to near 100% accuracy, although the MLP outperformed the other algorithms in training speed. A key outcome of this study is the finding that limits can be placed on network design parameters to avoid over designing neural networks, with associated reduction in training and prediction times. This is very useful because large volumes of data may be generated in CCUS projects and over-design of neural network architectures imposes penalties that are counterproductive to the goal of near-real time forecasting.