A Bayesian deep learning approach for rheological properties prediction of asphalt binders considering uncertainty of output

Abstract

Traditional methods for assessing rheological properties are often time-consuming, labor-intensive, and subject to operator bias. Moreover, conventional deep learning models may suffer from overconfidence in their predictions, limiting their applicability in real-world scenarios. A Bayesian deep learning framework that provides accurate predictions of asphalt binder rheological properties and quantifies the associated uncertainties is proposed to address these challenges. Atomic Force Microscopy images were employed as input values for the model, and the Dynamic Shear Rheometer test for the rheological property measurements was used. The proposed Bayesian deep learning model used Markov Chain Monte Carlo techniques to estimate model parameters and associated uncertainties. The effectiveness of the approach through an application involving multiple additives to validate the generality of the proposed framework was demonstrated. The results indicate that the Bayesian deep learning model offers improved prediction accuracy, reduced testing time, and operator-independent results compared to conventional methods. Additionally, quantifying uncertainties in predictions enable more informed decision-making and risk assessments in asphalt binder selection and pavement design. Overall, this study provides a novel and robust approach for predicting the rheological properties of asphalt binders, paving the way for more efficient and reliable pavement engineering practices.