Ensuring Reliability in CO2 Leakage Risk Assessment through AI-Driven Uncertainty Quantification Across Scales

Uncertainty Quantification plays a crucial role in the sustainable management of CCS technologies and in assessing the risk of CO2 leakage, which can be related to geological faults and fractures or mineralogical changes due to the medium acidification. However, subsurface uncertainties, arising from missing geological features at the reservoir scale, are challenging to identify. Therefore, we leverage insights gained from pore-scale and laboratory experiments to investigate the propagation of uncertainties across scales.

We present an efficient AI-driven data assimilation framework, based on Bayesian Physics-Informed Neural Networks, that incorporates robust uncertainty quantification to address multi-task and multi-scale Bayesian inference problems. This allows the combination of experimental data with physics-based regularization derived from prescribed models, while ensuring adaptive and automatic weighting of the different tasks’ uncertainties.

We showcase its application in pore-scale reactive inverse problems, using dynamical X-ray microtomography, to quantify morphological uncertainties in the porosity field and reliable ranges for the model reactive parameters. This allows to effectively address uncertainty quantification at the pore scale and propagate uncertainties for leakage risk due to carbonate dissolution in acidic fluids. Additionally, we address fault-related leakage by quantifying uncertainties in fractures conductivity, identifying relevant hydraulic apertures through the combination of data-driven and physics-based formulations.