Practical machine learning model selection and interpretation for organophosphorus flame retardancy in Epoxy resin

The traditional trial-and-error method for developing organophosphorus flame retardants is time-consuming and expensive. This work constructed machine learning models for limiting oxygen index (LOI), peak heat release rate (PHRR), and UL-94 for epoxy resin on the basis of the collected multifactor database, including the structure of organophosphorus flame retardants, addition amounts, matrix combustion performance, and flux. The training and test sets were divided on the basis of molecular groups to avoid data leakage within the same molecule group, in contrast to conventional random splitting. The best results for LOI and PHRR prediction were achieved via the XGBoost algorithm and ECFP4 fingerprints, with mean absolute errors of 1.61% and 125.5 kW/m² on the test set, respectively. For UL-94 classification, the MACCS with XGBoost achieved 79% accuracy. The Shapley additive explanation for the model indicated that the addition amount and matrix combustion performance data were the two most important features. This work could help develop a more accurate and reliable prediction model for flame retardancy.