The application of machine learning to a parametric analysis of fluid migration during geological carbon sequestration in a sandstone reservoir

Abstract:

During carbon capture and sequestration (CCS), it is critical to monitor the fluid flow migration and fluid pressure propagation for CO₂ storage efficiency and safety. As a result, physics-based, multi-phase fluid flow models are applied to study the site-scale behavior of CCS projects, while, the physics-based simulations are computationally expensive and time consuming. To understand the simultaneous effect of uncertainties in capillary pressure and relative permeability to fluid flow migration and pressure perturbation, artificial neural networks (ANN), one of widely used machine learning methods, is applied to capture features of CO₂ saturation and fluid pressure propagation based on 460 physics-based simulations describing different combinations of parameters from capillary pressure and relative permeability models. In addition, the CO₂ saturation and pressure migration under unknown capillary pressure and relative permeability conditions are predicted. It demonstrates that the trained ANN model accelerates the CO₂saturation and fluid pressure distribution prediction by at least 25,000%.

Results are analyzed on the basis of response surface maps to understand systematic variations of fluid migration across a complete 3-D parameter space (Po, λ, and Sgr), and it shows that entry pressure (Po) is the dominant parameter controlling CO₂ plume geometry and pressure accumulation. Small Po encourages fluid migration and large Po inhibits fluid migration. Meanwhile, the effect of λ and Sgr increase with increasing Po: (1) large λ and small S_gr encourages the migration of CO₂ saturation; and (2) high S_gr encourages the migration of pressure to a large extent comparing to λ.