Acceleration of Superpave Mix Design: Solving Multi-Objective Optimization Problems Using Machine Learning and the Non-Dominated Sorting Genetic Algorithm-II

The traditional asphalt mix design requires the preparation of many samples to test, which consumes much time and labor. Moreover, selecting aggregate gradation and asphalt content based on individual experience until a mixture’s properties meet a specification is a trial-and-error procedure. To accelerate the current design process and broaden the search range of the design space, this study proposed a fast and cost-saving Superpave volumetric mix design framework to automatically determine the asphalt mix proportion using the prediction ability of machine learning (ML) and Pareto optimal solutions of multi-objective optimization (MOO) problems generated by the non-dominated sorting genetic algorithm-II. Two goals were pursued in the MOO problems, namely material cost and whether the air void (AV) is close to 4%. To transform the Superpave volumetric mix design process into MOO problems, maximum theoretical specific gravity (G mm ) and AV prediction models were developed. To do so, a comprehensive traditional hot mix asphalt design dataset was collected from the literature. With the inputs of binder performance grading, asphalt content, aggregate gradation and density, and compaction work, seven ML models were trained based on the dataset. The results show that gradient boosting and extreme gradient boosting performed the best at predicting the G mm and AV, with coefficients of determination of 0.93 and 0.86, respectively. The proposed design process was implemented in an actual project and quickly yielded the optimum designs for designing 12.5- and 19-mm nominal maximum aggregate size mixtures. Furthermore, the volumetric properties of these designs closely matched the laboratory-measured values.