Closed-Loop CO₂ Storage Control With History-Based Reinforcement Learning and Latent Model-Based Adaptation

Closed-loop control of geologic CO2 storage must be robust to geological uncertainty while operating with limited monitoring, where dense subsurface states (e.g., pressure and saturation fields) are not available at deployment. We present a model-free reinforcement learning (RL) workflow that trains controllers on an ensemble of high-fidelity digital twins and evaluates them on a held-out reference simulator used strictly for inference. At each control step, the agent selects multi-well injection/production set-points and is rewarded for storage-oriented objectives while penalizing operational and pressure-related constraint violations, consistent with practical CCS control.

Please find the full abstract below.