Heuristic Principles to Predict the Effect of Crumb Rubber Gradation on Asphalt Binder Rutting Performance

The objective of this study was to employ an artificial neural network (ANN) to predict asphalt-rubber (AR) rutting performance characteristics using binder properties, crumb rubber (CR) gradations, and mechanical test parameters. The scope included advanced asphalt binder rheological characterization using a dynamic shear rheometer (DSR), encompassing preparation of a total of 18 laboratory-blended AR binders with two base binders and nine CR gradations, totaling over 2,200 data points. Principles of ANNs were used to predict the three AR binder performance parameters: $\eta$, $G^{*}/\sin \delta$, and $\tan \delta$. Eight input parameters constituting test temperature and frequency, five CR gradation components, and base binder viscosity were employed to develop the ANN model. A back-propagation learning algorithm with scaled conjugate gradient (SCG) as the training algorithm in a feed-forward, two-hidden-layer neural network with seven and three neurons, respectively, was chosen as the best ANN architecture. The statistical goodness of fit measures $R$ for the total data set were, respectively, 0.994, 0.997, and 0.977 for $\eta ,G^{*}/\sin \delta$, and $\tan \delta$. ANN modeling conceptualized as part of the study indicated that rubber inclusions in asphalt binders would aid in the improvement of the materials’ rutting resistance. The magnitudes of weights and biases provided in this study for the eight chosen AR binder material input parameters could be well utilized in predicting the three binder material performance parameters. Overall, it is envisaged that the algorithm developed in this research pertinent to asphalt binders’ advanced rheological characterization would further the state of the art in designing rut-resistant rubber modified asphalts.