Predicting Flash Point of Ester-based Insulating Oils and Unveiling Molecular Structure Effects via Graph Neural Networks

Flash point (FP) critically determines the fire safety of transformer oils, yet ester-based oils, the sustainable alternative to mineral oils, face a critical trade-off between FP enhancement and viscosity control. Consequently, molecular structure modification strategies to enhance the FP of esters emerge as a forefront research focus in insulating materials. However, current investigations remain insufficient in elucidating structure-property relationships, as the underlying mechanisms by which key molecular parameters affect FP are not systematically understood, and reliable quantitative FP prediction models for ester-based oils are underdeveloped. Thus, defining clear structure-FP correlations and developing accurate prediction methods are of critical engineering importance to guide the optimization of ester modifications. In this study, a dataset comprising 906 FP data points for ester molecules is compiled through the systematic extraction of data from published literature and chemical databases. Based on this dataset, the predictive performance of four Graph Neural Network (GNN) models is comparatively evaluated, among which AttentiveFP performs best, achieving a mean coefficient of determination (R²) of 0.97 and a mean root-mean-square error (RMSE) of 11.59°C. To further elucidate molecular interaction mechanisms, the Gradient-weighted Class Activation Mapping (Grad-CAM) technique is employed to visualize atom-level contributions. By modifying the molecular structures and analyzing FP changes, we find that FP increases with carbon chain elongation and an increase in the number of ester groups. Ester group positions show a negligible correlation with FP, except when ester groups are located at side-chain terminals, where FP decreases. While the position of C=C double bonds has minimal effect, increasing their number slightly reduces FP.