A Dynamic and Constrained Machine Learning Framework for CO ₂ Saturation Prediction in Geological Carbon Storage Reservoirs

Abstract Traditional machine learning (ML) methods often underutilize physical principles, which can lead to uncertain or unrealistic predictions, particularly when models are trained on limited data sets. We introduced a new ML framework designed to forecast CO 2 migration patterns in geological carbon storage (GCS) reservoirs. Using the Illinois Basin–Decatur Project (IBDP) GCS site as a case study, the framework incorporated physically meaningful inputs such as permeability, porosity, and capillary pressure. Effective permeability and capillary pressure were treated as dynamic inputs to account for feedback mechanisms from multiphase flow behavior in porous media. These two properties were updated based on predicted CO 2 saturation at each time step, explicitly incorporating the physics of multiphase flow. We also introduced a constrained loss function to enforce CO 2 saturation limits, as determined by porous media wettability, to improve model reliability. Our framework demonstrated the efficiency in predicting CO