Abstract
To assess the environmental performance of SF(6) substitute gases, it is essential to develop a predictive model for the global warming potential. In this study, 165 molecules are first selected to construct machine learning models using group contribution method. The predictive performance of various models is analyzed, including Artificial Neural Network, Random Forest, Gradient Boosted Decision Trees, and Support Vector Machines. Then 58 chemical activity descriptors are calculated using the M06-2X method and def2-TZVP basis, and the key descriptors are identified through Pearson correlation coefficient. These descriptors are used to build several machine learning models. The performance of these models constructed by the two approaches is compared. The result indicates that the descriptor-based models outperform the group-based models, with the descriptor-based Random Forest model achieving the best performance. The R (2) of test set reached 0.82, with an MSE of 0.015, an RMSE of 0.024, and an MAE of 0.09. Moreover, the descriptor-based model demonstrated higher stability and robustness across 1000 training iterations.