Abstract
We propose a novel denoising diffusion generative model for predictingnonlinear fluid fields named FluidDiff. By performing a diffusion process, themodel is able to learn a complex representation of the high-dimensional dynamicsystem, and then Langevin sampling is used to generate predictions for the flowstate under specified initial conditions. The model is trained with finite,discrete fluid simulation data. We demonstrate that our model has the capacityto model the distribution of simulated training data and that it gives accuratepredictions on the test data. Without encoded prior knowledge of the underlyingphysical system, it shares competitive performance with other deep learningmodels for fluid prediction, which is promising for investigation on newcomputational fluid dynamics methods.