Virtual Screening of Novel Eco-Friendly Gaseous Dielectrics through Dimensionless Bond Decomposition and Machine Learning Algorithm.

Abstract

Identification of environmentally friendly gaseous dielectrics to replace the most potent greenhouse gas SF₆ is urgently desired in the worldwide high-voltage electrical industry. However, the great challenge for SF₆-free technology remains because of numerous contradictory requirements it has to meet simultaneously: high dielectric strength, low boiling points, low global warming potential, high arc quenching capability, low acute/subchronic inhalation toxicity, and low flammability. Herein, the chemical bonds are revealed to be the universal, unique, and unified descriptors to develop the predictive models for efficient virtual screening of novel gaseous dielectrics. By means of the automatic bond decomposition mechanism toward the dimensionless SMILES formula, excellent correlations between experiments and theory have been obtained successfully for eight types of key properties of the insulation gases using the optimized artificial neural networks. The bond-based machine-learning algorithm is stable to both training and test sets within the leveraging applicability domains. Mechanistic interpretations of the inherent coupling effect have been carried out by the normalized importance of weights of the bond descriptors. The bond-based networks were applied to a total of 3727 C, H, N, O, S, and F-containing compounds as curated from PubChem. Properties of each species were predicted, and the overall performance was ordered by scoring with respect to SF₆. Although no gas could be identified to be superior to SF₆ in all aspects, a shortlist of promising replacement gases with well-balanced dielectric performance has been found by virtual screening and might stimulate experimental synthesis and tests for practical use. Moreover, the present work provides guidelines for the rational design of structural characteristics of novel compounds influential for gaseous dielectrics.