The PM6 method within the Ampac10 software package and Codessa 3.3 software were utilized to calculate 951 descriptors for 65 insulating gas molecules. This approach is faster and provides a wider range of descriptors compared to the DFT method. After preprocessing and feature variable screening using a genetic function algorithm, 12 significant feature descriptors influencing electric strength were identified. A QSPR model for electric strength was established using a neural network with 85% of the data as the training set and 15% as the prediction set. The model demonstrated a high prediction capability with a correlation coefficient R2 of 0.9951. The residual plot indicated no systematic errors during model establishment, and the error probability distribution chart showed that the errors were essentially normally distributed. The model was validated using insulating gas molecules with bonds not included in the 65 studied molecules, indicating its predictive ability for new molecular structures and potential use in screening the electric strength of new SF6 alternative gases. In summary, the PM6 method combined with Codessa 3.3 software, along with genetic function approximation and neural network algorithms, offers a convenient and accurate method for predicting the electric strength of insulating gas molecules.