Prediction of asphalt low-temperature performance by FTIR spectra using comparative modelling strategy

In this paper, partial least square regression (PLSR) and artificial neural network (ANN) were used to predict three asphalt low-temperature index: low-temperature performance grade (LTPG), low-temperature grade loss (LTGL) and critical tip opening displacement (CTOD). The results indicate that the prediction accuracy of both models are relatively ideal when second derivative is used as the data processing technique. Combined with the variable importance projection plot for PLSR-based model, the key factors affecting LTPG, LTGL and CTOD are thermal oxygen aging, uncrystallised wax content and solid crystalline wax content. Compared to PLSR-based models, these ANNs have a far better prediction accuracy and can be further optimised by selecting appropriate activation function. By comparing both models through Bland–Altman analysis, ANN exhibits an obvious improvement in predictive accuracy for LTPG, which is limited in predicting LTGL. Regarding CTOD, ANN is unsuitable for predicting brittle asphalt the predicted value tends to be overestimated.