A boiling point prediction method based on machine learning for potential insulating gases

[Display omitted] •Gradient Boosting Regression with RDKit (GBR-RDKit) model was designed to predict boiling points.•A boiling point database composed six elements for potential insulating gases was constructed.•Characteristics related to boiling point properties were identified by SHAP analysis.•RDKit-GBR method had superior performance in different levels of model complexity. The boiling point is a crucial indicator for assessing the suitability of insulating gases. Its theoretical prediction has consistently garnered significant attention from the scientific community. In this study, a boiling point database composed of hexa-element (C, H, O, N, F, S) for potential insulating gases was constructed. The model of Gradient Boosting Regression with RDKit descriptors (RDKit-GBR) achieved superior predictive ability on the test set with a coefficient of determination of 0.97, a mean absolute error of 17.74 °C, and a root-mean-squared error of 27.83 °C. The SHapley Additive exPlanations analysis showed that the “Ipc” feature in RDKit, which represents the spatial relationship and interaction between pairs of atoms within molecules, plays a central role in predicting the boiling points for insulation gases. Furthermore, the applicability of RDKit-GBR method was further validated across several elemental combinations. Eventually, compared with the previously reported models, the hexa-element model achieves excellent accuracy.