Machine learning prediction of lower and upper flammability limits of refrigerant mixtures toward safe refrigerant discovery

Given mixing different refrigerants as a typical approach for refrigerant design, reliable predictive models for evaluating the lower and upper flammability limits (LFL and UFL) of refrigerant mixtures are of great interest from the safe application point of view. In this study, based on a comprehensive refrigerant mixture database, six machine learning (ML) algorithms and two dataset split schemes (i.e., the point-based and the mixture-based) are thoroughly compared for accurate structure-property modeling of LFL and UFL. In addition, to enhance the modeling performance for refrigerant mixtures, the feasibility of data augmentation by incorporating a pure organic substance dataset are explored, with the test MAE of LFL reduced by 25% and that of UFL reduced by 14.1%. Following that, the SHAP method is performed for the interpretability of the final models. Finally, as a demonstration of model application, the LFL and UFL of hundreds of binary mixtures consisting of state-of-the-art pure refrigerants are predicted by the final models, providing valuable references for the discovery of safe refrigerants. • Machine learning on flammability limits of refrigerant mixtures is presented. • Both lower and upper flammability limits are covered for rigorous ML modeling. • Including pure molecular data effectively enhances the mixture property modeling. • LFL and UFL of hundreds of novel refrigerant mixtures are predicted by final model.