High-Throughput Screening and Molecular Property Prediction for Environmentally Friendly Insulation Gases

The rapid expansion of the power industry and growing concerns over the greenhouse effect have intensified the demand for environmentally friendly insulation gases as alternatives to SF<sub xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">6</sub> in power equipment. Accelerating the development of eco-friendly insulation gases requires an efficient and accurate assessment of key molecular properties, including high dielectric strength, low boiling point, low global warming potential, and minimal toxicity. To address this challenge, a molecular property prediction framework based on a universal pre-trained model with high predictive accuracy is proposed, replacing traditional trial-and-error methods and enabling high-throughput screening of potential insulation gases. This approach facilitates rapid identification of molecules meeting physical and chemical criteria without relying on complex first-principles calculations. Molecular datasets are collected and used to fine-tune the pre-trained model, achieving high predictive accuracy with mean absolute errors of 0.21 for dielectric strength, 10.90 K for boiling point, and 0.56 for the logarithm of global warming potential. For toxicity classification, the model demonstrates strong performance with an area under the receiver operating characteristic curve of 0.81 on the test set. High-throughput screening of a two-million-substructure database yields five promising candidate molecules for SF<sub xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">6</sub> replacement. Furthermore, atomic-level analysis of structure-property relationships provides valuable insights for future molecular design.