Data-driven Prediction of Dynamic Complex Modulus for Non-conventional Asphalt Mixtures

The dynamic modulus (|E*|) of asphalt mixtures is one of the main material characteristics that governs the quality of asphalt pavements in their design and construction. While the |E*| values are traditionally obtained from intensive laboratory testing or statistical predictive equations, this research attempted to apply several machine learning (ML) techniques to predict |E*|, especially for non- conventional asphalt mixtures. The study used 3906 lab-measured |E*| data points from different types of non-standard asphalt mixtures, such as recycled asphalt pavement, recycled asphalt shingles, warm mix asphalt, asphalt rubber, air-blown asphalt, and polymer-modified asphalt. Mixture temperature, loading frequency, aggregate gradation, mixture volumetric, and asphalt binder information were included as variables used in the ML techniques. Relative comparisons were made to answer the following question: which ML technique would provide a more accurate prediction for |E*| when non-conventional asphalt mixtures are considered in the design and construction? It was found that, among the five ML techniques used in the study, decision trees and random forests showed the best prediction capability. Linear regression showed the least accurate prediction. It was also found that the |E*| measured for asphalt-rubber asphalt mixtures was best predicted by ML techniques.