This proposal is for the scientific studies within the group of Hossein Azizpour particularly focusing on the applications to fluid flow. This computation is requested for collaboration with Ricardo Vinuesa's group at KTH with whom Azizpour has several collaborative funded projects with PhD students and postdocs. In Azizpour's group we develop fundamental methods for deep learning and use them for impactful scientific applications. The main fundamental aspects pertinent to this computation project are generative models and differentiable top-k. The main application area for this computation request is for fluid dynamic. Even the fundamental research projects either involve purely empirical hypotheses (e.g., a new architecture) or requires empirical validations of the theories (e.g., a robust learning algorithm). Such empirical investigations for modern deep generative models are expensive as they involve heavy usage of large amounts of simulated fluid data, long training of deep generative models including differential equations solvers, and complex backbones such as large 3D U-Nets and/or visual transformers. Such computations can only be enabled with the help of a large modern GPU cluster such as Arrhenius. Since we are applying in a medium round and the cap on Arrhenius is 5000 GPU-hour, for this round we simultaneously apply for Dardel GH and Dardel AMD GPUs. The plan is to fully move to Arrhenius at the next large round.