Leveraging Machine Learning To Predict the Atmospheric Lifetime and the Global Warming Potential of SF6 Replacement Gases.

Abstract

The global warming potential (GWP) is a relative measure of the capability of a molecule to trap the Earth's infrared radiation as heat. The measurement or prediction of the GWP of a molecule is based on two factors: the radiative efficiency and atmospheric lifetime of a molecule. While the calculation of the radiative efficiency of a molecule using the computational chemistry approach, such as density functional theory (DFT), is well-established and robust, the development of a computational approach to estimate the atmospheric lifetime remains challenging and limited to date. In this contribution, we developed a machine learning (ML) approach to estimate a molecule's atmospheric lifetime and GWP100 based on electronic and geometrical features. We benchmarked the state-of-the-art computational workflow with the developed ML model in estimating the atmospheric lifetime and GWP100. The developed ML model outperforms the existing approach with the mean absolute error values of 0.234 (ML-predicted atmospheric lifetime) and 0.249 (direct ML model for GWP100) compared with 0.535 (Atkinson's method) and 0.773 (Kazakov et al.) from previous works. The developed models were used to screen >7000 molecules in PubChem and bigQM7 data sets in a search for SF6 replacement gas for the electric industry and identified 84 potential candidates.