Evaluation of Machine Learning Applications for the Complex Near-Critical Phase Behavior Modelling of CO2–Hydrocarbon Systems

The objective of this study was to evaluate the capability of machine learning models to accurately predict complex near-critical phase behavior in CO2–hydrocarbon systems, which are crucial for enhanced oil recovery and carbon storage applications. We compared the physical Peng–Robinson equation of state model to machine learning algorithms under varying temperatures, pressures, and composition, including challenging near-critical scenarios. We used a direct neural network model and two hybrid model approaches to capture physical behavior in comprehensive compositional space. While all the models showed great performance during training and validation, the Direct Model exhibited unphysical behavior in compositional space, such as fluctuations in equilibrium constants and tie-line crossing. Hybrid Model 1, integrating a single Rachford–Rice iteration for physical constraints, showed an improved consistency in phase predictions. Hybrid Model 2, utilizing logarithmic transformations to better handle nonlinearities in equilibrium constants, further enhanced the accuracy and provided smoother predictions, particularly in the near-critical region. Overall, the hybrid models demonstrated a superior ability to balance computational efficiency and physical accuracy, closely aligning with the reference of the Peng–Robinson equation of state. This study highlights the importance of incorporating physical constraints into machine learning models for reliable phase behavior predictions, especially under near-critical conditions.