Answers are grounded in this paper and cite their sources.
The optimization of well placements and injection rates for maximizing CO 2 storage remains a critical challenge for scaling up geological carbon storage (GCS) considering minimizing leakage outside of leased boundary and potential geomechanical risks. A deep reinforcement-learning (DRL) framework is developed in this paper, where a convolutional-neural-network (CNN) policy is trained to map surrogate-predicted CO 2 saturation and pressure fields (states) to injection-well locations and dynamic injection rates (actions). Policy parameters are updated by Proximal Policy Optimization (PPO) with a multi-objective reward that increases with stored CO 2 mass while imposing continuous penalties for pressure buildup and boundary leakage. Accurate surrogate models trained on coupled flow-geomechanics simulations significantly reduce computational cost and enable the agent and environment interactions required to train the DRL policy, supporting high-resolution evaluation across numerous geological realizations. The effectiveness of the proposed DRL framework is evaluated using a modified conditional geostatistical model based on the Aquistore CO 2 storage site in Saskatchewan, Canada. Model-parameter uncertainties are handled using a domain randomization scheme, where multidimensional scaling and cluster analysis compress 1,000 realizations into 20 clusters to streamline agent training. Performance is benchmarked against both fixed-rate reference configurations and the covariance-matrix adaptation evolution strategy (CMA-ES). Across all validation samples, the PPO-based policy consistently delivers superior well configurations and dynamic injection rates, achieving higher reward as injection volume while maintaining operational safety. On a representative validation sample, the PPO policy achieved 43.3 Mt of cumulative CO 2 storage, significantly outperforming the 37.1 Mt achieved by the CMA-ES benchmark
No takes yet. Share an insight, caveat, or question.
Erten et al. (2026) studied this question.
Synapse has enriched 10 related papers. Consider them for comparative context: