Automating the determination of low-temperature fracture resistance curves of normal and rubberized asphalt concrete in single-edge notched beam tests using convolutional neural networks

As materials undergo large-scale yielding or exhibit large sizes of fracture process zone in the crack tip region, multi-parameter fracture concepts should be employed to describe the complex crack-tip stress-strain fields. Fracture resistance curves (R-curves) are an established tool in characterizing the entire fracture process of such materials. However, for complex materials such as bituminous mixtures, the development of these curves is subject to experimental and computational intricacies. In this research, a framework is developed to automate the construction of R-curves for normal and rubberized asphalt concrete (AC) mixtures. AC mixtures are produced using PG58–22 and PG58–28 binders. Limestone and siliceous aggregates are used, and three binder contents are considered for the mixtures. Single-edge notched beam (SE(B)) fracture testing is conducted on AC beams with two different notch patterns. A convolutional neural network (CNN) model is developed and trained over 1260 test images with varying temperatures, notch geometries, and setups. The CNN model was used to detect the growing crack on the beam surface and each crack-detected image was sent to an image processing framework to measure the crack length. Crack extension increments were calculated and synchronized with test time and magnitude of load, load-line displacement, and cumulative fracture energy, and the R-curve could be constructed. A training accuracy of 0.91 was obtained for the model and a loss of below 0.10 as a result of a hyperparameter tuning indicating reliable classifications by the CNN architecture. The R-curves showed desirable agreement for control mixtures at temperatures of 0 °C and −15 °C. As the mixtures are rubberized, the R-curves showed favorable agreement in the crack blunting phase, transition zone, as well as the unstable propagation phase at −20 °C. Cohesive energy magnitudes were compared for the two methods with a Pearson coefficient of 0.81 while fracture rate and fracture energy magnitudes showed favorably close magnitudes with coefficients of 0.89 and 0.98 respectively.