Our focus is to develop numerical methods for physics simulations. A particular emphasis lies on simulating fluid flow and Navier-Stokes problems with deep learning techniques. We are interested in combining all kinds of neural networks, be it convolutional, recurrent, or adversarial, with physical knowledge and priors. Such methods have significant potential for efficient and robust simulations, e.g., in the context of visual effects and digital games, but also in engineering or medical areas.