Prediction of the char formation of polybenzoxazines: The effect of heterogeneities in the crosslinked network to the prediction accuracy in quantitative structure-properties relationship (QSPR) model

Molecular Operating Environment (MOE) software has great potential when combined with the Quantitative Structure-Property Relationship (QSPR) approach, and was proven to be useful to make good prediction models for series of polybenzoxazines [1-3]. However, the effect of heterogeneities in the crosslinked network to the prediction accuracy is yet to be tested. It was found that polybenzoxazines with polymerisable functional group (e.g. acetylene-based benzoxazines) form up to 40% higher char yield compared to their analogue polybenzoxazines due to the contribution of the polymerisable functional group (e.g. ethynyl triple bond) in the cross-linked network. In order to investigate the effect of the inconsistent cross-linking network, a data set consisting of thirty-three benzoxazines containing various structures of benzoxazines was subdivided into two smaller data sets based on their functional group, either benzoxazines with polymerisable functional group (acetylene-based benzoxazines set (Ace-M)) or non-polymerisable functional group (aniline-based benzoxazines (Ani-M)). Char yield predictions for the polybenzoxazines for these data sets (Ace-M and Ani-M) were compared with the larger thirty-three polybenzoxazines data set (GM) to investigate the effect of the inconsistency in crosslink network on the quality of prediction afforded by the model. Prediction performed by Ace-M and Ani-M were found to be more accurate when compared with the GM with total prediction error of 3.15% from both models compared to the GM (4.81%). Ace-M and Ani-M are each better at predicting the char yields of similar polybenzoxazines (i.e. one model is specific for a polymerisable functional group; the other for non-polymerisable functional group), but GM is more practical as it has greater 'general' utility and is applicable to numerous structures. The error shown by GM is considerably small and therefore it is still a good option for prediction and should not be underestimated.