Suitable Tests and Machine Learning Approach to Predict Moisture Susceptibility of Hot-Mix Asphalt

Abstract

The objectives of this study were to determine a suitable set of tests to use with a moisture-conditioning process and to develop a machine learning model to predict the moisture susceptibility of hot mix asphalt. Laboratory-compacted samples of 17 plant-produced mixes with known field performance were subjected to mechanical tests before and after moisture conditioning with the moisture-induced stress tester (MiST). Statistical analysis showed that seismic modulus and indirect tensile strength were effective in distinguishing the poor-performing mixes from the well-performing mixes. Principal component analysis was conducted on the test data, and a reduced set of dimensions that were capable of explaining much of the variance in the data was identified. The significant test properties were used to develop machine learning models with two supervised classification approaches. The k-nearest neighbor model was found to be very accurate in differentiating the mixes. The use of MiST conditioning, the specified physical tests, and machine learning methods is recommended for the identification of moisture-susceptible hot-mix asphalt.