ABSTRACT
This paper presents a probabilistic model for predicting the dynamic modulus |E*| of asphalt concrete (AC). A Bayesian Neural Network (BNN) trained on a substantial dataset collected from various states was employed. This approach accounts for the inherent stochasticity in the data generation process (i.e. aleatoric uncertainty) and addresses epistemic or model uncertainty as well. The model successfully predicted dynamic moduli for unseen testing datasets. In practice, predicted moduli could be as accurate and effective as measured values, considering the model uncertainty and |E*| test variability. However, an elevated epistemic uncertainty at extremely low and high |E*| ranges would be expected due to relatively low data points. To enhance interpretability, Shapley Additive Explanations (SHAP) analysis was utilised, demonstrating adherence to physical laws. The primary factors influencing predicted moduli, as identified by the analysis, are temperature, frequency, voids in mineral aggregates, reclaimed asphalt pavement content, and binder low-performance grade. The BNN model was deployed in a web application, making it accessible for application.
Acknowledgments
The authors acknowledge the Illinois Department of Transportation (IDOT) for funding parts of this study. The content of this paper reflects the views of the authors, who are responsible for the facts and accuracy of the data presented herein. The content does not necessarily reflect the official views or policies of Illinois Center for Transportation. This paper does not constitute a standard, specification, or regulation.
Disclosure statement
No potential conflict of interest was reported by the author(s).