The automation potential of mechanical components is typically assessed manually. The NeuroCAD project from Fraunhofer IPA uses deep learning to decouple this assessment from the knowledge and experience of experts. The next step of NeuroCAD is optimizing the automation potential of components. This would enable the design of components more suitable for automated assembly processes. The thesis pioneers this step and aims at adapting the latent space manipulation method StyleCLIP for optimizing the grabability of mechanical components, as well as exploring the suitability of StyleCLIP for this task. For this, StyleGAN was adapted for the synthesis of mechanical components and CLIP was replaced with a comparator network for grabability. Given only limited data, the focus was set to optimizing screws and bolts. StyleCLIP was also extended to allow for the protection of specified structures of the component during the optimization. Two StyleCLIP methods, latent optimization and optimization using a latent mapper, were analyzed and compared. Most of the optimized components look almost realistic and the optimizations are by large reasonable according to the definition of grabability used. The optimization using the latent mapper yields superior results compared to the latent optimization. Investigating this, gradual and iterative optimization methods such as the latent optimization are unsuitable for the optimization of grabability. The optimization using the latent mapper seems suitable since its possible to resolve the identified issues and the results should improve given larger datasets.

Keywords: Assembly Automation, Automation Potential Optimization, NeuroCAD, StyleCLIP