Generative Adversarial Networks (GANs) are gaining increasing popularity through their ability to learn data distributions from training samples and generate high quality samples of it. While this was shown extensively for 2D images, this thesis does investigate their use to reproduce three-dimensional data as done by other research. These learned samples do offer more diversity, since the generative model is able to produce a continuous representation between the training data samples and is stored much more efficiently. An interesting area to use these learned representations in is in the context of the performance evaluation of data visualization techniques. Therefore, networks were designed in consideration of ongoing research, which are able to be trained with a single configuration for a variety of data sets to produce scalar field data. The similarity of the generated distribution in comparison to the original data's distribution is measured through its execution performance when visualized by a volume rendering algorithm. By running large scale tests we were able to show that the created samples may serve as a plausible substitute for real data sets and displayed its ability to mimic features and the rendering performance of the training data. In addition, the model's outputs are compared to the original distribution through metrics in order to verify the results. Differences are inspected in detail to show causes of deviations from the model's performance and their statistics. Additionally, we discuss the observed properties of the model's output as well as impairments which may be introduced.