Development of Dynamic Modulus Prediction Model Using Artificial Neural Networks for Colombian Mixtures

Abstract

Dynamic modulus is a key material property to predict the performance of flexible pavements. Several prediction models have been developed worldwide to predict the dynamic modulus from aggregate, binder, and mixture properties with corresponding frequency and temperature. Artificial neural network (ANN)-based prediction models have shown better accuracy than regression-based models. In this study, two ANN-based prediction models were developed for the Colombian hot-mix asphalt (HMA) mixtures. One model (W-ANN) has similar input variables as the Witczak model, and the other model (H-ANN) has similar input variables as the Hirsch model. The ANN-based dynamic modulus prediction models were trained, validated, and tested using 972 data points. The coefficient of determination (R2), RMS error (RMSE), absolute average error (AAE), and Se/Sy indicated that the two ANN-based models performed better than the previous models. The W-ANN had slightly better performance than the H-ANN model. The parameters for both ANN-based models are reported to reproduce dynamic modulus values for future use and testing with high accuracy. A standalone closed-form equation was extracted from each ANN model, which makes the developed models easier for and more accessible to practitioners. Sensitivity analysis showed that both models are sensitive to binder and mixture properties. These models can be utilized in Colombia for the existing and future development of pavement design packages, and will reduce the necessity of extensive testing in the future.