A hybrid physics-machine learning model to predict density of mixtures of CO2 with impurities

Abstract

Carbon Capture, Utilization, and Storage (CCUS) plays a pivotal role in reducing greenhouse gas emissions, essential for limiting global warming to below 1.5 °C by 2100 and achieving carbon neutrality by 2050. Modeling carbon dioxide (CO₂) density is crucial for optimizing CO₂ transportation and storage systems. However, captured CO₂ streams from power sources often contain impurities such as Oxygen, Nitrogen, Carbon Monoxide, Argon, Sulfur Dioxide, Hydrogen, Methane, Water, and Hydrogen Sulfide. These impurities significantly impact transmission properties and challenge the predictive capabilities of equations of state (EoSs) thermodynamic models.

This study investigates the effects of impurities on CO₂ stream density using a comprehensive dataset of 134,204 density data points. Fourteen EoSs, including cubic, virial, physical, and multi-parameter equations, were evaluated to determine optimal modeling conditions. Moreover, machine learning models trained with experimental and synthetic data from equation of state (EoS) models were employed towards a high predictive capability model. This synthetic data was generated, within CCUS pipeline operating conditions, using the best-performing EoSs, primarily multiparameter equations with an Absolute Average Relative Deviation <3 %. Random Forest and Artificial Neural Networks provided robust density predictions, even in complex thermodynamic regions with a Coefficient of Determination >0.96.

This hybrid approach offers a novel pathway for improving density predictions of CO₂-rich systems, supporting more efficient and reliable transportation models. To the best of our knowledge, no previous study considered such a comprehensive dataset and EoSs for predicting the density of CO₂ rich mixtures using this hybrid approach.