Systematic error patterns in dynamic modulus predictive models of asphalt concrete

Owing to a reasonable correlation with field performance, dynamic modulus (vertical bar E*vertical bar) has been used as an input in pavement design specifications. With experimental determination of vertical bar E*vertical bar issues, researchers have resorted to the use of predictive models. This study evaluates the underlying patterns in prediction error in these models using vertical bar E*vertical bar database developed during NCHRP 1-40D study. The significant difference between the global and mixture-wise dataset observed in the correlation analysis was substantiated by the T-test. Thus, predictive models were recalibrated, and statistical indicators indicated improvement with mixture-wise calibration compared to global calibration. Q-Q and cumulative distribution plots constructed using 'difference parameter (DP)' introduced in this study indicated highest and lowest error with AI-Khateeb and Original Witczak models, respectively. Finally, the entire range of measured vertical bar E*vertical bar was divided to analyze the error patterns in detail. In general, lower prediction error was observed in the middle range where all the predictive models showed comparable performance. However, they showed significant prediction error at the extreme values of vertical bar E*vertical bar, and the DP showed skewed and flatter distribution. In this region, Hirsch and Al-Khateeb models performed poorly whereas Original Witczak and South Korean models performed well. Sensitivity analysis identified binder properties as highest sensitive.