Uncertainty-aware Digital Twins for Monitoring and Optimizing of Geological Carbon Storage

Geological Carbon Storage (GCS) is the only truly scalable pathway to net-negative emissions on the path to net-zero, yet ensuring safe, cost-efficient, and verifiable GCS demands far more than static reservoir models. We introduce an uncertainty-aware Digital Twin framework that delivers continuous, probabilistic monitoring of subsurface CO₂ plume dynamics and enables risk-informed operational decisions in real time. At the core of our approach is a machine-learning-enabled data assimilation framework where Simulation-Based Inference (SBI) replaces linear ensemble Kalman filtering, enabling the framework to handle highly nonlinear CO₂ dynamics and seismic observations. This allows the Digital Twin to characterize the full posterior distribution of reservoir states—fluid saturations and pressures—conditioned on multimodal time-lapse observations from imaged surface seismic data and well measurements. The framework operates recursively: neural networks are trained offline on ensembles of simulated reservoir states and simulated observations, capturing complex nonlinear subsurface dynamics and observations. Once trained, the Digital Twin assimilates incoming field data in near real-time, continuously updating the probabilistic reservoir model as new measurements arrive. By maintaining a probabilistic rather than deterministic view of reservoir state, our Digital Twin enables decision-making under uncertainty—optimizing CO₂ injection schedules while managing the risk of exceeding critical fracture pressures, and identifying monitoring strategies that maximally reduce uncertainty. Validated on realistic numerical simulations of GCS scenarios, this framework offers a concrete path toward safer, more transparent, and verifiable Geological Carbon Storage operations.