Developing Machine Learning Models for Accurate Prediction of Radiative Efficiency of Greenhouse Gases

Background: Greenhouse gases (GHGs), particularly halocarbons, contribute significantly to the radiative forcing of climate change due to their long lifetime. Accurate prediction of GHG radiative efficiency (RE) is required to assess the relative impacts of GHGs on climate change. Methods: An RE dataset consisting of around 82 K molecules was generated using a narrow band model (NBM) and density functional theory (DFT) computed infrared spectra. This dataset, along with 568 experimentally available RE data, was utilized to develop machine learning (ML) models through a transfer learning approach. Significant findings: The accuracy of REs calculated with NBM and DFT was evaluated against experimental data. A moderate level of agreement was observed, with a tendency to overestimate the results. ML models were then trained using DFT data, and their performance was compared using various model architectures. The directed message-passing neural network (DMPNN) model outperformed other models, and when combined with quantum mechanical features (ml-QM-DMPNN), a slight improvement was observed. The DMPNN and ml-QMDMPNN models were refined by the transfer learning technique using the experimental dataset, which resulted in predicting experimental REs with root mean square errors (RMSEs) of 0.0582 and 0.0564 Wm ? 2ppb-1, respectively. This study offers critical insights into developing ML models that can predict REs of molecules, thereby facilitating the design of environmentally-friendly chemicals and materials.