Physics-informed machine learning predicting CO2 capture performances of organic mixtures

CO₂ capture through amine absorption is an effective technology for combating global warming. The development of innovative solvents can enhance this process by increasing CO₂ solubility and reducing the size of required absorption columns. However, these solvents often involve both physical and chemical absorption mechanisms, necessitating extensive experimentation to characterise new solvent mixtures, which slows innovation. This study introduces a hybrid physics-informed model to predict CO₂ solubility in absorption mixtures. The model is designed to predict the behaviour of novel mixtures by characterising individual absorption mechanisms and incorporating physics-based insights to evaluate the contributions of each mechanism according to the mixture type. We benchmarked the hybrid model against a data-driven approach, training it on comprehensive literature data across diverse mixture types. The hybrid model demonstrated superior performance with an R² of 0.929 on the test set, outperforming the data-driven model with an R² of 0.611. It also exhibited lower bias across mixture categories, greater robustness in predictions and their physical adherence as highlighted by the executed SHAP analysis. By enabling accurate digital predictions of novel solvent mixtures, this hybrid model promotes process intensification, accelerating the development of more sustainable CO₂ capture technologies and contributing to a greener future.