Optimal control of geological carbon storage using multi-physical monitoring and deep reinforcement learning Available to Purchase

We propose the use of multi-physical monitoring to train and utilize a deep reinforcement learning (DRL) agent as a sequential decision-making tool for the optimal control of a geological carbon storage (GCS) operation. The optimization problem aims to maximize the profit from the operation while minimizing the risk of induced seismicity by learning an optimal carbon dioxide (CO2) injection policy. Numerical studies using DRL agents coupled with time-lapse AVO, time-lapse gravity, and geostatistical reservoir simulators show that: 1) the trained DRL policy outperforms constant injection strategies, especially given uncertainty in the permeability model; 2) a multi-physical monitoring approach enhances this policy, particularly when noise is present in the geophysical data.