Data-driven approach augmented by attention mechanism in critical and boiling thermophysical properties prediction of fluorine/chlorine-based refrigerants

Refrigerants are ubiquitous in modern society, indispensable for various applications ranging from air conditioners and refrigerators to spacecraft and hospitals. Improving predictive methodologies for designing and implementing novel fluids is therefore essential. This paper presents an integrated prediction model augmented by attention mechanism for the molecular weight, boiling point, critical temperature, and critical pressure of fluorine/chlorine - based refrigerants. The model employs molecular groups as molecular structure descriptors and utilizes a machine learning algorithm with attention mechanisms. The prediction accuracy is good (all R2 > 0.9), far ahead of other methods, such as Lydersen, Joback, Klincewicz-Reid, Constantinou-Gani, and multilayer perceptron methods. Furthermore, to gain insights into the influence of group features on property prediction, an interpretable method known as Shapley Additive Explanation (SHAP) was employed. This method computes the contributions of group features, thus providing a deeper understanding of the relationship between molecular groups and thermophysical properties. The results verified that the proposed models effectively establish this relationship.