Dielectric strength (E_r) is a critical factor in screening and evaluating SF₆ replacement gas. The traditional experimental methods of measuring E_r are not only extremely time-consuming but also very costly. In this work, an E_r prediction model for SF₆ replacement gases is constructed using machine learning methods. First, an exhaustive literature survey is performed to collect 88 high-quality experimental E_r values. Second, a total of 32 insightful microscopic descriptors are accurately calculated for each compound using 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 through five-fold cross-validation and hyperparameter optimization, are used to train and test the 88 experimental E_r data and their relevant microscopic descriptors. Finally, the result shows that the Ada Boost regression model exhibits 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 E_r. The results indicate that polarizability is the predominant factor significantly affecting E_r, 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 E_r. There is an approximately positive correlation between the molecular weight and the E_r of gases. To verify the reliability of the Ada Boost regression model for E_r prediction, the E_r of SF₆ and six known environmentally friendly replacement gases are tested within an absolute error of 0.02–0.33. This study provides a feasible method for accelerating the search for SF₆ replacement gases.