Designing Asphalt Concrete Mixtures Based on Balanced Mix Design Requirements Utilizing a Two-Step Approach

Abstract

The Illinois Flexibility Index Test (I-FIT) and the Hamburg Wheel Tracking Test (HWTT) have been used for asphalt concrete (AC) balanced mix design in Illinois. Automating the selection of initial asphalt binder content and aggregate gradation, based on target mix type, I-FIT flexibility index (FI), and HWTT rutting depth (RD), could reduce the effort, time, and material resources in mix design. This study proposes a two-step framework combining machine learning (ML) and an optimization method to predict asphalt mix constituents. Initially, six ML models—support vector regression, a deep neural network, an autoencoder, adaptive boosting, a random forest (RF), and extreme gradient boosting (XGB)—were trained using an extensive database. The Step 1 model utilized a subset of samples, which included the AC mix properties, I-FIT FI, and HWTT RD, to provide preliminary predictions of asphalt binder content and aggregate gradation after data preprocessing. The Step 2 model applied a genetic algorithm to a broad I-FIT and HWTT database to refine Step 1 predictions, aligning them with predefined FI and RD targets. The Step1-XGB model accurately predicted binder content and aggregate gradation, both with and without long-term aging (LTA) considerations (with LTA: $R^2$  = 0.959; without LTA: $R^2$  = 0.956). The subsequent Step2-RF model, enhanced with a genetic algorithm, optimized AC mixture design to meet specified FI and RD requirements. The proposed two-step approach was validated through other AC mix designs and uncertainty analysis. An interactive tool was developed to implement this framework and enable practical, first-trial mix design solutions.