Physics-Constrained Machine Learning on a Unified CO2–Brine Experimental Database: Predicting Relative Permeability & Capillary Pressure Curves

Carbon Capture Storage projects success hinges on reliable relative-permeability and capillary-pressure curves for CO2–brine displacement to predict plume migration and emplacement. Yet, core-flood tests under reservoir conditions are costly and scarce. We’ve harmonized 150 published experiments, extracting Brooks–Corey parameters, and trained machine learning models on easily measured inputs (porosity, permeability, salinity, pressure, temperature) to predict those curve parameters. The machine learning models were imposed with physics-informed monotonic constraints to ensure physically consistent trends, and the original database was subjected to adaptive data augmentation. The final cross-validated models deliver an average Mean Absolute Percentage Error of just 0.32 % for the predicted relative permeability and capillary pressure curve parameters, which outperforms the baseline model by two orders of magnitude. This workflow together with the frist-in-class CO2–brine relative permeability and capillary pressure database, makes possible to generate full drainage, imbibition and capillary pressure curves tied to uncertainty ranges, enabling rapid screening, sensitivity analyses, and real-time flow-model updates without sacrificing accuracy.