Non-Linear Support Vector Machine Prediction of the Mechanical Properties of Asphalt Binders Subjected to Varying Temperatures and Frequencies Based on SARA

Abstract

Listen

Translate article

This study investigates the effects of chemical fractions on the mechanical properties of asphalt binders and predicts the mechanical properties of asphalt binders based on the chemical fractions. Initially, four fractions—saturate, aromatic, resin, and asphaltene (SARA)—were isolated from 36 asphalt binders using a thin-layer chromatography with flame ionization detection (TLC-FID) analyzer. Subsequently, the complex modulus and phase angle of the asphalt binders were determined for a range of frequencies and temperatures. The relationships between SARA content, heavy components, colloidal instability index, and the complex modulus and phase angle were analyzed. Advanced models, including quadratic polynomial and non-linear support vector machine (SVM) with sigmoid and RBF (Gaussian) kernels, were employed to predict the complex modulus and phase angle of asphalt binders based on the SARA data, and the reliability of these prediction models was critically assessed. The findings indicate that the contents of asphaltenes, resins, aromatics, and saturates significantly influence the rheological properties at different frequencies, though a clear correlation between SARA contents and both the complex modulus and phase angle was not established. Alternative methods should be considered for studying the mechanical properties of asphalt derived from SARA. The RBF kernel demonstrated superior performance compared to the quadratic polynomial and non-linear SVM with the Sigmoid kernel. While the non-linear SVM with the RBF kernel accurately predicts the complex modulus, it fails to predict the phase angle at low frequencies. The validation of this model confirmed its efficacy in capturing the relationship between input (SARA) and output (complex modulus and phase angle) vectors for each asphalt binder. The predicted complex modulus master curves closely match the experimental results, yet the model only approximates the trend of phase angle variation with frequency.