Drifting Models for Surrogate Flow Modeling

CFD simulations provide high-spatial resolution data about flow fields, which are helpful for the optimization of intricate indoor environments. Nevertheless, the process of expanding the boundaries of design and its rapid iterations is restricted by the high computational costs associated with CFD. Although state-of-the-art generative surrogate models provide a better approximation of distributions compared to deterministic networks, the slow inference time incurred by iterative sampling makes them impractical. This paper attempts to generate single-pass, high-quality flow fields for the first time by the application of a new generative drifting method to the fluid mechanics domain. The proposed method uses a conditional architecture to perform drifting in the latent space of

a learned Variational Autoencoder (VAE). Additionally, it utilizes label-aware masking to guide generated samples in accordance with defined boundaries. This label-conditioned model shows two orders of magnitude faster inference speed compared to iterative diffusion methods while maintaining comparable flow accuracy and consistency. A further variant of this model, modified with spatial conditioning, illustrates a potential solution to improve the generalization for previously unseen flow geometries. In summary, the drifting approach proposed here represents the first opportunity for surrogate CFD models to be operational in real time.

Message: Hi FE, I need help with “Drifting Models for Surrogate Flow Modeling” in “Biotechnology & Biomedical Engineering”