Machine learning-based prediction of global warming and ozone depletion potentials of refrigerants

Abstract

Global warming potential (GWP) and ozone depletion potential (ODP) are pivotal environmental metrics of refrigerants, whereas accurate predictive methods for predicting GWP and ODP are still scarce. By collecting the state-of-the-art molecular GWP and ODP database, this study introduces QSPR models considering diverse machine learning algorithms to perform the demand-driven tasks of GWP regression, GWP classification, and ODP classification. Through a rigorous validation scheme, the best models are identified for each task, which achieve an R² of 0.8790, a classification accuracy of 0.6890, and a classification accuracy of 0.9971 on the test sets, respectively. Following that, SHAP analysis is performed for the optimal models to gain deeper insights into the contributions of individual features. Finally, a high-throughput screening with the final models is conducted to identify potential green refrigerants. This work provides a reliable tool for predicting the environmental properties of refrigerants and supporting the design of refrigerant systems.