Access to the material response in mechanical experiments can be provided by modern optical methods such as digital image correlation (DIC). Such experiments generate large amounts of data that need to be automatically analyzed using classical algorithms, or, more recently, machine learning (ML). However, the training of such machine learning models often requires labelled data, which can be expensive. Therefore, unsupervised machine learning and generative methods become more popular.

In this work, we use Cycle-Consistent Generative Adversarial Networks (Cycle-GANs) to generate DIC-like physical displacement fields of fracture mechanics problems. The models are trained with DIC data from real fatigue crack growth experiments and finite element simulations. We show how to generate fake DIC data from corresponding finite element simulations using cycle-consistent GANs. With a trained generator, we are able to sample simulation-consistent DIC displacement data without the need for real experiments. Finally, we show that these synthetic data sets can be used to train machine learning models for crack tip detection and can be applied to reliably detect crack tips in cases without labelled data.