Data-driven estimation models of asphalt mixtures dynamic modulus using ANN, GP and combinatorial GMDH approaches

The objective of the present study is to develop and evaluate machine learning-based prediction models, employing the artificial neural networks (ANNs), Genetic Programming (GP), and the Combinatorial Group Method of Data Handling (GMDH-Combi), for dynamic modulus (E*) of hot mix asphalt estimation. To develop the models, an experimental database including 1320 data was employed in which dynamic shear modulus of binder, binder phase angle, cumulative percent retained on (3/8-in.) sieve, No. 4 sieve, percent passing No. 200 sieve by total aggregate weight, volumetric percent air voids in the mix, and volumetric effective binder content were considered as the effective contributing parameters on E*. Considered input parameters can be readily measured using simple laboratory tests. The proposed models' performances were evaluated using common error criteria and compared with that of existing models for estimating E* such as the Witczak, modified Witczak, revised Bari-Witczak, Hirsch, Yu and Shen, Al-Khateeb (I) and (II), global and the simplified global models. The results indicated that the ANN model with correlation coefficients of 0.9821 and 0.9839, respectively, for training and testing, is more accurate than the GMDH (0.9500 and 0.9503) and GP (0.9493 and 0.9495)-based developed models. Additionally, the accuracies of the proposed models outperform entire existing models. Moreover, a wide range of input data was considered in the model development, and the proposed models' generality and robustness were investigated through a parametric analysis over the covered ranges. The main benefit of using the GMDH and GP approaches was providing closed-form explicit mathematical expressions, which leads to simple and straightforward applications for general engineers and researchers in the field.