Prediction of low-temperature fracture resistance curves of unmodified and crumb rubber modified hot mix asphalt mixtures using a machine learning approach

Abstract

Fracture resistance curves (R-curves) provide a robust tool for a comprehensive insight into the crack propagation regime in engineering materials. In this paper, an extensive research program is conducted to determine R-curves for hot mix asphalt (HMA) mixtures with varying properties. The experimental results are then used to develop R-curve prediction models following a machine learning approach. Three-point single-edge notched beam (SE(B)) experiments were conducted on HMA mixtures incorporating 0%, 5%, 10%, 15%, and 20% crumb rubber at low temperatures. The temperature ranged from + 5 °C to −20 °C while limestone and siliceous aggregate with two gradations were used in developing mixtures with two base bitumen having performance grades of PG58-22 and PG64-22. It was observed that as the temperature is declined to −20 °C, the stable crack growth region is significantly diminished in the R-curves, and the mixtures undergo a brittle fracture with abrupt failure of the specimen. A temperature of −15 °C could be determined where the transition from quasi-brittle to brittle fracture occurs. Mixtures fabricated incorporating 20% crumb rubber exhibited a progressively rising R-curve at the lowest test temperature (−20 °C) even in the unstable crack propagation phase, which is a desirable material characteristic. Two prediction models were developed for R-curves. Artificial neural networks (ANN) were used in the first model resulting in an R-square value of 0.965. Due to the black-box nature of the ANN, the multi-gene genetic programming approach was also applied in the prediction of the R-curves to derive a mathematical equation between the input data and the outputs. The R-square equaled 0.870 in this method. R-curves could successfully be predicted by both methods considering the negligible to fair errors.