Data-Driven Surrogate Models for Rock-Fluid Interaction: A Case Study in CO2 Injection

Summary The design of geological CO2 storage requires complex engineering analyses, including predicting CO2 plume migration and rock dissolution. These predictions are traditionally performed using Computational Fluid Dynamics (CFD) simulations. However, the prohibitive computational expense of CFD restricts the ability to run the numerous scenarios required for comprehensive uncertainty quantification (UQ) and optimisation studies. To address this limitation, this work introduces a novel data-driven surrogate modelling framework designed to accelerate porous media fluid-rock interaction predictions. The model is trained on a dataset featuring diverse pore structures, ensuring it remains geometry-independent and can accurately predict flow in unseen environments. We propose a framework that is grid-scale-consistent, a property that maintains neural network accuracy across domains of different sizes. Therefore, it significantly reduces memory consumption by allowing for training on small domains while enabling inference over larger, unseen domains. We also incorporate and validate two other innovations. First, the UNet++ architecture, originally designed for image segmentation, is shown to outperform the UNet architecture in terms of prediction accuracy for these complex fluid dynamics problems. Second, employing a rollout training strategy yields better performance for long, iterative predictions. Furthermore, the framework for surrogate modelling developed has broad applicability to general fluid flow problems.