A Deep Learning Framework for Predicting Global Warming Potential of Refrigerants for Sustainable Chemical Design

The rapid identification of environmentally sustainable refrigerants is essential to meet global climate targets and comply with international mandates such as the Kigali Amendment. This study presents a deep learning framework to predict the 100-year Global Warming Potential (GWP100) of single-component refrigerants using molecular descriptors and dimensionality reduction. We used descriptor sets from RDKit, Mordred, and alvaDesc, combined with principal component analysis and quantile transformation, to train ensemble neural networks. The RDKit-based model performed best, achieving a root-mean-square error (RMSE) of 481.9 and a coefficient of determination (R2) of 0.918 on the test set. Factor analysis showed that molecular weight, lipophilicity, and functional groups such as nitriles and allylic oxides contribute most to GWP. To promote accessibility, we developed a web tool that accepts RDKit-calculated descriptors as input and returns GWP predictions. This framework enables efficient virtual screening of refrigerant candidates and provides a foundation for future integration with other sustainability metrics.