AI-based augmentation of prediction potential for asphalts

There is a lack of studies when dealing with the comparison between regression methods and machine learning (ML)-type methods in terms of their ability to interpret and describe how the components of a bituminous mixture affect mechanistic performance. At the same time, artificial intelligence (AI)-driven approaches are becoming more popular in analysing asphalt mixtures, yet there are limited comparisons of regression and machine learning (ML) models for mechanistic performance interpretation. Consequently, a comparison of AI and statistical approaches is presented in this study for predicting bituminous mixture properties such as stiffness, fatigue resistance, and tensile strength. Some of the important input features are bitumen content, crumb rubber content, and air void content. The research uses random forest model (RFM), linear regression model (LRM), and polynomial regression model (PRM). RFM and PRM achieved an R2 as high as 0.94, with mean absolute error (MAE) less than 2.5, and are, therefore, good predictive models. Interestingly, RFM works best in one-third of instances, particularly when dealing with outliers, whereas traditional statistical models work better in two-thirds of instances. The results highlight AI's value in bituminous mixture optimisation, where RFM showed good prediction accuracy. In 30% of the cases, AI models outperformed the conventional statistical approaches. At the same time, analyses show that model performance varies significantly with scenarios and that even if AI models capture complex nonlinear relationships, they must not override DOE principles.