Evaluation of ANN-Based Dynamic Modulus Models of Asphalt Mixtures

Artificial neural network (ANN)-based dynamic modulus |E*| models were evaluated on South Carolina's asphalt mixtures, the majority of which contained recycled asphalt pavement (RAP). These ANNs contained similar input variables as the NCHRP 1-40D and Hirsch regression models and were implemented in the neural network toolbox of MATLAB version R2018b. Two previously published ANN-based |E*| models were also evaluated on the same database. Most ANNs in the literature have been shown to predict |E*| with good success; however, they have not been validated outside of their original studies. The results showed that (1) ANN-based |E*| models performed significantly better than regression models; (2) ANNs with few input variables (either Va, Vbeff, and Gb* or VMA, VFA, and Gb*) highly predicted |E*| with R2>0.99 on testing; (3) ANNs can accurately predict |E*| of recycled asphalt mixtures; (4) the validation performance of the two published ANNs on South Carolina's asphalt mixtures was ranked fair; and (5) locally customized ANNs are more accurate in the estimation of |E*| than globally calibrated ANNs or regression models.

Automated summary

ANN-based dynamic modulus models outperform regression models on South Carolina's asphalt mixtures, especially when using a few input variables to highly accurately predict |E*|. The accuracy of predicting |E*| is higher with locally customized ANNs.