Insulation Prediction and Descriptor Selection Based on Random Forest Algorithm

To find more efficient and environmentally friendly substitutes for sulfur hexafluoride (SF₆), it is still necessary to screen more potential insulating gases. However, the commonly used experimental methods are inefficient. In this article, random forest (RF) algorithm are chosen to predict gas insulation strength with gas molecular descriptors. The origin of descriptors includes those used in existing journals (such as electric dipole moment µ (IEEE Trans Dielectr Electr Insul 20(3):Jun [1]), positive surface area, and surface area (Neurocomputing 216 [2])A_s) (J Comput Chem 38(10):721–729 [3]), and the highly weighted ones in our pre-experiment with RF (such as Hall-Kier alpha connected with charge distribution and Kappa to describe the degree of molecular curvature). The descriptors sorted by RF reveals more correlations between the molecular properties and insulation, and the trained model provides important insights to design and to screen other SF₆ substitutes.