Artificial Neural Network Modeling of Theoretical Maximum Specific Gravity for Asphalt Concrete Mix

The maximum specific gravity of asphalt concrete (AC) mix (G_mm) is an important parameter without which asphalt mix design cannot be realized. But the experimental procedure for measuring the G_mm requires time, consumes electric energy, and generates wastewater. Huge amount of experimental data that can enable the virtualization of the AC mix design process exists. But to date, all standardized AC mix-design procedures are mainly experimental. In this study, non-linear regression analysis and multi-layer Artificial Neural Network (ANN) were utilized to develop prediction models for the G_mm of AC mixes. The study utilized 4158 superpave mix-design data points from the Long-Term Pavement Performance (LTPP) information management system (IMS) database. The input variables are asphalt specific gravity G_b, asphalt binder content P_b, and combined bulk specific gravity of aggregates G_sb. The ANN-model (R=0.9843,MSE=0.00016) performed better than the regression model (R=0.9241,MSE=0.00076). A standalone user-friendly MATLAB-based app was developed for the trained ANN-model. The ANN-model is capable of predicting G_mm within AASHTO and ASTM standard single-operator precision requirements (± 0.011) 85.9% of the time. The model can predict G_mm within a margin of ± 0.021 with a 95% success rate. The resulting air voids which were estimated using the predicted G_mm met air-void precision tolerance of ± 0.5 and ± 1.0% in 85.6 and 96.3% of the tests, respectively. The proposed model could minimize the time, energy, and material resources needed during the mix-design process of AC. Standards for AC mix-design should be revised to accommodate more use of prediction models so as to make the design process more sustainable.