Dielectric strength (Er) is a critical factor in screening and evaluating for SF6 replacement gas. The conventional experimental methods for measuring Er are not only exceptionally time-consuming but also prohibitively expensive. This work constructed an Er prediction model for SF6 replacement gases based on machine learning methods. First, an exhaustive literature survey is performed to collect 88 high-quality experimental Er values. Second, a total of 32 insightful microscopic descriptors are accurately calculated for each compound based on density functional theory, including both global parameters and molecular electrostatic potential parameters. Furthermore, five state-of-the-art machine learning algorithms, which have been carefully modified based on five-fold cross-validation and hyperparameter optimization, are utilized to train and test the 88 experimental Er data and their relevant microscopic descriptors. Finally, the result reveals that Ada Boost regression model demonstrates superior predictive performance with a coefficient of determination of 0.90, a mean absolute error of 0.17, and a root mean square error of 0.18. Moreover, Shapley Additive exPlanations analysis is used to reveal the correlation between the microscopic descriptors and Er. The results indicate that polarizability emerges as the predominant factor significantly affecting Er, which accounts for as high as 17.3%, followed by the molecular weight (14.1%). Specifically, molecules with high α are more prone to deformation under the action of an electric field, and their electron clouds are more likely to be polarized, which has a positive correlation with Er. There is an approximately positive correlation between the molecular weight and the Er of gases. To confirm the reliability of Ada Boost regression model for Er prediction, the Er of SF6 and six known environmentally friendly replacement gases were tested within an absolute error of 0.02-0.33. This study provides a feasible pathway to accelerate the search for SF6 replacement gases.
下载: