New Diffusion Model Generates Physics-Informed Data Efficiently

Diffusion models are powerful for generating spatiotemporal signals, such as fluid dynamics images, but often lack the ability to incorporate underlying physical laws. This limitation can lead to visually plausible but physically inaccurate generated samples. A new study introduces a simple yet effective physics-informed approach based on diffusion guidance with self-generated data augmentation. The proposed method learns the data distribution conditioned on the degree of deviation from physically correct dynamics. It then generates samples by explicitly setting this deviation condition to zero, ensuring adherence to physical laws. A key innovation is that the evaluation of governing equations is decoupled from the diffusion model's training and sampling processes. This decoupling eliminates the need to solve computationally expensive governing equations at every denoising iteration, leading to significantly faster sample generation. Experimental results demonstrate that this model not only substantially reduces physical deviations compared to standard diffusion models but also achieves further improvements when combined with existing physics-constrained diffusion methods.