Dynamic fracture of a bicontinuously nanostructured copolymer: A deep-learning analysis of big-data-generating experiment

Abstract

Here, we report measurements of detailed dynamic cohesive properties (DCPs) beyond the dynamic fracture toughness of a bicontinuously nanostructured copolymer, polyurea, under an extremely loading rate, from deep-learning analyses of a dynamic big-data-generating experiment. We first describe a new Dynamic Line-Image Shearing Interferometer (DL-ISI), which uses a streak camera to record optical fringes of displacement-gradient vs time profile along a line on sample's rear surface. This system enables us to detect crack initiation and growth processes in plate-impact experiments. Then, we present a convolutional neural network (CNN) based deep-learning framework, trained by extensive finite-element simulations, that inversely determines the accurate DCPs from the DL-ISI fringe images. For the measurements, plate-impact experiments were performed on a set of samples with a mid-plane crack. A Conditional Generative Adversarial Networks (cGAN) was employed first to reconstruct missing DL-ISI fringes with recorded partial DL-ISI fringes. Then, the CNN and a correlation method were applied to the fully reconstructed fringes to get the dynamic fracture toughness, 12.1kJ/m^2, cohesive strength, 302 MPa, and maximum cohesive separation, 80.5 um, within 0.4%, 2.7%, and 2.2% differences, respectively. For the first time, the DCPs of polyurea have been successfully obtained by the DL-ISI with the pre-trained CNN and correlation analyses of cGAN-reconstructed data sets. The dynamic cohesive strength is found to be nearly three times higher than the dynamic-failure-initiation strength. The high dynamic fracture toughness is found to stem from both high dynamic cohesive strength and high ductility of the dynamic cohesive separation.