Exploring Nontoxic Green Refrigerants Using Positive-Unlabeled Learning and High-Fidelity Search

Identifying environmentally sustainable refrigerants remains a challenging task, as it requires the reconciliation of thermodynamic performance with stringent safety criteria, particularly those related to low toxicity. In this study, we present a data-driven framework that emphasizes nontoxic refrigerant candidates while also considering flammability, thermal stability, and low global warming potential (GWP). Our approach integrates graph neural networks (GNNs) with positive-unlabeled (PU) learning to accurately identify refrigerants exhibiting the desired physicochemical properties, even on incomplete and biased data sets. We applied the proposed framework to a chemical library of over 33,000 compounds and identified seven promising refrigerants that satisfy all physicochemical and environmental criteria. We report detailed property predictions and uncertainty estimates for each candidate. This work introduces PU learning as a powerful tool for safer molecular design and provides a scalable alternative to traditional experimental or computational screening approaches.