Drifting models can replace slow iterative diffusion for CFD surrogates, enabling real-time flow field generation that's orders of magnitude faster while matching diffusion model quality.
This paper speeds up CFD simulations by using a generative model called "drifting" instead of traditional diffusion models. The model learns to generate realistic fluid flow patterns in a single pass rather than iteratively, making it 100x faster while maintaining accuracy. It uses a learned latent space and label-aware masking to ensure generated flows match boundary conditions.