Prediction of SF₆ Replacement Gases Based on Machine Learning

Global Warming Potential (GWP) is a crucial criterion for assessing the impact of SF₆ replacement gases on climate change, which is determined by radiative efficiency (RE) and atmospheric lifetime (τ). This work leverages six modified machine learning (ML) methods to estimate GWP and its critical parameters (τ and RE), as well as explore their correlations via molecular descriptors, aiming at high-throughput screening and evaluation of potential SF₆ replacement gases. Based on the coefficient of determination, mean absolute error, and root-mean-square error, the optimal ML methods are Histogram Gradient Boosting Regression for GWP (0.95, 0.28, and 0.33), Gradient Boosting Regression for τ (0.92, 0.28, and 0.40), and Extreme Tree for RE (0.90, 0.04, and 0.05). Furthermore, molecular descriptors analysis reveals that τ and RE are mainly influenced by the energy of the highest occupied molecular orbitals and the gap between the highest occupied molecular orbitals and the lowest unoccupied molecular orbitals, respectively, which jointly affect GWP. Ultimately, the developed models are applied to screen the GWP₁₀₀ of 853 molecules from the QuanDB data set, yielding six SF₆ replacement gases with low boiling points, low GWP, and high electrical strength. This work not only predicts viable alternatives to SF₆ replacement gases but also provides insight into intrinsic correlations between GWP and its critical parameters (τ and RE).