[3H1-OS-9a] OS-9

10:00 AM - 10:15 AM JST(1:00 AM - 1:15 AM UTC)

[3H1-OS-9a-05] Learning Diffusion Models with Physics-Informed Classifier-Free Guidance

〇Akira Osaka¹, Naoya Takeishi¹, Takehisa Yairi¹ (1. The University of Tokyo)

Keywords:

Diffusion Model,Classifier-Free Guidance,Physics-Informed Machine Learning

Diffusion models have demonstrated strong performance in generating physical phenomena, such as images of fluid dynamics. However, a fundamental limitation of standard diffusion models is the lack of guarantees that the generated samples satisfy underlying physical laws. A previous study has attempted to mitigate deviations from the governing equations by incorporating residual terms into the training loss. Nevertheless, this approach presents notable drawbacks that physical constraints are not explicitly enforced during the sampling phase, and ODE/PDE solvers must be propagated at every training iteration. In this study, we propose a novel physics-aware generation framework that integrates classifier-free diffusion guidance conditioned on zero residuals. We evaluate the proposed method on fluid image generation tasks. Experimental results demonstrate that our approach effectively reduces deviations from the governing physical laws.