Combining deep neural networks and Gaussian processes for asphalt rheological insights

Abstract

Asphalt binders are critical for asphalt pavement performance, and understanding their rheological behavior is essential for designing durable roadways. The complex shear modulus (G⁎) and phase angle (δ) are primary parameters characterizing binder rheology. This study introduces a novel hybrid machine learning model combining deep neural networks (DNN) and Gaussian process regression (GPR) to predict G⁎ and δ for bituminous binders and binder-filler systems (mastics). DNN excel at capturing complex, nonlinear relationships among eleven binder and thirteen mastic input parameters, including aging conditions, chemical and physical properties, and test parameters. However, standalone DNN struggle with small datasets, common at the binder scale, and lack inherent uncertainty quantification, limiting reliability in engineering applications. GPR improves DNN by refining predictions through probabilistic modeling, while providing uncertainty estimates, and enhancing accuracy with limited or noisy data. The hybrid model leverages DNN's feature extraction capabilities and GPR's ability to smooth predictions, significantly improving performance over standalone DNN. The hybrid model achieves high prediction accuracy, with R2 values of 0.997 for G⁎ and 0.947 for δ for binders, and 0.993 for G⁎ and 0.972 for δ for mastics, reducing G⁎ prediction error from 22.7% to 0.031% for fresh asphalt binder compared to standalone DNN. Feature importance analysis using random forest and SHAP techniques identifies test temperature, aging conditions, and penetration as key influencers of G⁎ and δ. This hybrid approach enhances the characterization of complex asphalt materials, offering pavement engineers a robust, reliable tool for predicting material behavior under diverse conditions.