Accelerating Numerical Simulations of CO2 Geological Storage in Deep Saline Aquifers via Machine-Learning-Driven Grid Block Classification

Abstract

The accurate prediction of pressure and saturation distribution during the simulation of CO2 injection into saline aquifers is essential for the successful implementation of carbon sequestration projects. Traditional numerical simulations, while reliable, are computationally expensive. Machine learning (ML) has emerged as a promising tool to accelerate these simulations; however, challenges remain in effectively capturing complex reservoir dynamics, particularly in regions experiencing rapid changes in pressure and saturation. This article addresses the challenges by introducing a fully automated, data-driven ML classifier that distinguishes between regions of fast and slow variation within the reservoir. Firstly, we demonstrate the variability in pressure across different reservoir grid blocks using a simple brine injection and production scenario, highlighting the limitations of conventional acceleration approaches. Subsequently, the proposed methodology leverages ML proxies to rapidly and accurately predict the behavior of slow-varying regions in CO2 injection simulations, while traditional iterative methods are reserved for fast-varying areas. The results show that this hybrid approach significantly reduces the computational load without compromising on accuracy. This provides a more efficient and scalable solution for modeling CO2 storage in saline aquifers.