Formulation of low temperature mixed mode crack propagation behavior of crumb rubber modified HMA using artificial...

Determining mixed mode fracture parameters asphalt concrete mixtures remains an engineering challenge due to non-homogeneity and inelasticity of the material. In this research, a study was conducted to determine the low-temperature R-curves of unmodified and crumb rubber modified Hot Mix Asphalt (HMA) under mode I and mixed-mode (I/II) loading conditions. Single edge notched beam (SE(B)) testing was employed to collect data, and three key fracture parameters-cohesive energy, energy rate, and fracture energy-were extracted to represent different stages of fracture and crack propagation. Within the scope of this study, it was observed that for the AC 85/100 paving grade bitumen, a temperature of - 20 °C serves as a critical temperature, shifting fracture from quasi-brittle to brittle. At this temperature, the stable crack growth region in the R-curves significantly shrinks, causing abrupt specimen failure. The incorporation of 20% crumb rubber demonstrated favorable material characteristics, with a progressively rising R-curve even during the unsfi crack propagation phase. The central goal of this research is to establish prediction models for the mixed-mode (I/II) crack propagation parameters G_b, G_f, and G_i. The features selected for modeling are G_b0, G_f0, and G_i0 (mode I), percentage of crumb rubber, type of aggregate, binder content, nominal maximum aggregate size, temperature, and normalized offset ratio. Two dataset configurations were used: dataset 1 contains all entries, while dataset 2 excludes G_b0, G_f0, and G_i0 (mode I). Five machine learning techniques, Regression, Multi-Gene Genetic Programming (MGGP), Support Vector Regression (SVR), Random Forest, and Artificial Neural Networks were employed to predict three key fracture parameters. Although slightly less accurate than SVR and Random Forest, MGGP offers the key advantage of yielding explicit mathematical expressions for crack propagation prediction. The R² index for the MGGP model in Dataset 1 was 0.93 for G_b, 0.94 for G_f, and 0.92 for G_i. For dataset 2, the indices were 0.89, 0.93, and 0.88, respectively.