Optimized machine learning approaches for the prediction of viscoelastic behavior of modified asphalt binders

Abstract

Additives are commonly used in pavement engineering to improve the original bitumen's rheological and mechanical characteristics to meet severe loading and climatic condition requirements. To select the optimum dosage of an additive for modifying the original bitumen, it is essential to predict the viscoelastic behavior of modified bitumens, which can be performed by implementing the constitutive viscoelastic parameters of the complex shear modulus (G*) and phase angle (δ). In this work, a comprehensive experimental database consisting of the results of the frequency sweep mode of a dynamic shear rheometer (DSR) at seven test temperatures (−22 ~ 22 °C) was used. Prediction models for the viscoelastic behavior of bitumen modified with different dosages of crumb rubber, styrene–butadienestyrene (SBS), and polyphosphoric acid (PPA) were developed by optimizing and applying different machine learning approaches, including Artificial Neural Networks (ANN), Robust Linear Regression, Linear Support Vector Regression, Decision Tree Regression, Gaussian Process Regression (GPR), and Ensemble Regression, on the data. By comparing the various studied model outputs in terms of performance measurements, such as the correlation of coefficients, relative root mean square error, scatter index, relative error, and Nash-Sutcliffe efficiency coefficient, it was determined that the Ensemble Regression method has the highest performance in predictions.