시간 간격에 따른 인공지능 난류 예측 성능 비교: U-Net과 Diffusion Model의 비교 연구

초록

This study investigates the impact of time increments on the autoregressive prediction of two-dimensional turbulence slices extracted from three-dimensional direct numerical simulations, comparing a U-Net with a diffusion-based generative model. Experiments demonstrate that the diffusion model significantly outperforms the deterministic approach, reducing the rollout-averaged mean squared error by 28.3% and relative energy error by 61.9%. Notably, the generative model effectively mitigates the smoothing of fine-scale structures and error accumulation observed in deterministic methods, preserving physical statistics such as energy spectra even at large time steps. Although the diffusion-based approach incurs higher computational costs, these findings highlight its superior capability in maintaining physical consistency for long-term turbulence prediction under partial observation conditions

키워드