Multifidelity well-control optimization for geological CO₂ sequestration with brine extraction using a coarse-grid network model

by Billal Aslam, Stein Krogstad, Olav Møyner, Knut-Andreas Lie, Bicheng Yan

Year: 2026 DOI: https://doi.org/10.1016/j.jcp.2026.115307

Extra Information

Journal of Computational Physics, Volume 567 (2026)

Abstract

In geological CO₂ sequestration, effective pressure management is crucial to prevent caprock failure and reduce the risk of leakage. Controlled brine extraction offers a practical mechanism to regulate pressure while increasing storage capacity. However, optimizing coupled injection and extraction strategies typically relies on repeated evaluations of high-fidelity reservoir simulators, leading to prohibitive computational costs.

We propose a sequential multifidelity optimization (SMFO) framework that combines a physics-based reduced-order model (CGNet) with a high-fidelity simulator to accelerate well-control optimization. The CGNet model is obtained through aggressive coarsening of the fine-scale reservoir model and retains the dominant flow connectivities of the original model. Implemented within a differentiable simulator (MRST), the CGNet model is calibrated using adjoint-based misfit minimization to reproduce the dynamic responses of the high-fidelity simulator. Within the SMFO framework, the low-fidelity model is used to efficiently explore the control space during early optimization stages, while the high-fidelity simulator is employed in later iterations to ensure the solution accuracy.

Application to the Johansen formation benchmark demonstrates that the reduced CGNet model achieves a sevenfold reduction in simulation time while preserving key dynamic features, including saturation evolution and well responses. Scaled mismatches remain below 10⁻², and plume similarity metrics indicate strong agreement with high-fidelity results (SSIM > 0.93, RMSE < 0.06). When embedded in the SMFO framework, the proposed approach attains the same optimal net present value as direct high-fidelity optimization while reducing total CPU time by more than 50%. These results indicate that differentiable physics-based surrogates, when combined with gradient-based optimization, provide an effective and scalable approach for computationally efficient reservoir management in large-scale CO₂ storage applications.