The more complex and elaborate the annotations become, or the less frequently certain classes occur in a dataset, the more costly the implementation of an ML evaluation becomes, and the more attractive the generation and use of synthetic training data becomes. There are several approaches to generating synthetic data in the literature, not only in materials science, but perhaps even more so in other disciplines. A common approach is data-driven methods such as generative adversarial networks. However, for rare classes, they suffer from the same problem of insufficient training data and are therefore not practically applicable to real-world use cases where little data is available. Model-based approaches, on the other hand, may be too difficult to implement for microstructural images.

This work adapts a texture synthesis approach from graphics design for application to microstructural images. This approach belongs to the group of non-parametric example-based image generation algorithms and allows the generation of new images by remixing single or multiple examples, thus working in low-data regimes.

The applicability of this approach is evaluated by generating macroscale defect structures for a classification task and microscopic microstructural images for a semantic segmentation task. Firstly, the quality of the synthetic data is determined by expert round robin tests and the application of appropriate image metrics. Secondly, it is discussed how the targets needed for texture synthesis, which require user input, can be as automatic as possible, but also unbiased and representative of the microstructures. Finally, it is shown that the synthetic training data can improve the performance of the trained ML models.