Exploring the use of machine learning to predict metrics related to asphalt mixture performance

Agencies responsible for construction and maintenance of roadways often use some measure of performance to qualify asphalt mixtures before being used in construction. As of this writing, the state of Texas uses the Hamburg wheel tracking test and indirect tensile strength test to qualify a hot mix asphalt produced for roadway construction and maintenance. Optimizing the mixture design to produce mixtures with the desired performance criteria has been a topic of interest for asphalt researchers and industry personnel. This study explores the use of machine learning methods to estimate the rut depth from the Hamburg wheel tracking test and the indirect tensile strength from the mixture design and volumetric information. Support vector regression analysis and decision tree based ensemble methods, including bagging, random forests, extra-trees, and gradient boosting algorithms were trained with data collected by the Texas Department of Transportation for quality control and quality assurance purposes. Metrics related to mixture design including aggregate gradation and absorption, asphalt binder content and performance grade, use of warm mix asphalt, recycled materials, and laboratory-molded density as well as test information, such as number of wheel-passes applied in the Hamburg wheel tracking test, were used as input variables. The analysis showed that all of the machine learning algorithms adopted in this study were able to estimate the mixture performance criteria from the mixture design and volumetric properties when the models were trained with curated and sufficient data. While extra-trees provided the best performance in terms of the coefficient of determination, gradient boosting and support vector regression models were found to learn from the imbalanced data better than the other methods. This study offers opportunities for the development of data-driven performance-oriented mixture design optimization

Automated summary

This study explores using machine learning methods to estimate performance criteria of asphalt mixtures, finding that various algorithms performed well with data collected by the Texas Department of Transportation. Extra-trees showed the best performance in terms of coefficient of determination, while gradient boosting and support vector regression models were found to handle imbalanced data better.