隨著深度學習技術在計算流體力學（CFD）領域的應用日益興起，構建高保真且具備物理感知能力的神經代理模型已成為提升模擬效率的關鍵。本研究旨在探討生成式擴散模型在流場預測中的物理準確性，並系統性比較四種建模策略的方法論差異與適用限制。
研究初期採用 Stable Diffusion 1.5 進行流場圖像生成，實驗結果顯示，儘管該模型能生成與目標幾何高度相似的視覺結果，但其本質僅限於像素層級的紋理模仿，缺乏對流體力學基本規律的理解，導致預測結果呈現物理失真。為建立數值比較基準，本研究開發一套輕量化卷積神經網路（CNN），導入雷諾數條件通道與空間座標資訊，並配合納維—斯托克斯方程殘差作為物理損失函數進行約束訓練。實驗顯示，CNN 在訓練工況下具備合理的速度場重建能力，然而在物理損失引入後的訓練中期，模型易收斂至對稱性平凡解，揭示了純判別式架構缺乏全局物理先驗的根本限制。
本研究進一步將物理損失約束嵌入 Stable Diffusion 1.5 的潛空間架構中。對比實驗揭露了深層的機制性衝突：去噪損失梯度與物理損失梯度在擴散模型參數空間中存在方向競爭，導致物理殘差雖大幅降低，速度場重建精度卻同步惡化，卡門渦街特徵在整個訓練過程中始終無法有效重建。本研究將此現象定義為數值拮抗（Numerical Antagonism），並將其識別為以數值損失約束生成式擴散模型的方法論根本限制。
針對數值拮抗問題，本研究提出 Physics-Aware Prompt 策略，將物理知識從損失函數層次提升至語意條件層次，透過思維鏈六階段流場演化描述引導去噪 U-Net 的生成方向，從根本上消除梯度競爭。定性評估結果顯示，生成流場呈現清晰的交替渦核結構、合理的渦脫落相位多樣性，以及比 CNN 更豐富的尾流非線性細節，驗證了語意層次物理條件引導的可行性。

This study investigates physically consistent flow field generation using conditional latent diffusion models, with cylinder wake flow as the primary benchmark. A systematic com-parison is conducted across four modeling strategies: a baseline Stable Diffusion 1.5 mod-el, a lightweight CNN with physics-informed constraints, a physics-informed diffusion model (SD1.5+PINN), and a Physics-Aware Prompt strategy based on Chain-of-Thought semantic design.
The CNN baseline achieves strong quantitative reconstruction accuracy but exhibits con-vergence slowdown in the physics loss stagnation region when physics constraints are in-troduced, revealing the fundamental limitation of deterministic mapping without global physical priors. Incorporating PINN constraints into the diffusion model reduces physics residuals substantially; however, gradient competition between the denoising loss and physics loss in the diffusion model parameter space leads to a counterintuitive outcome where physical residuals decrease while velocity field accuracy degrades—a phenomenon this study terms numerical antagonism—and the Kármán vortex street structure remains unrecoverable throughout training.
The Physics-Aware Prompt strategy resolves this conflict by encoding physical knowledge at the semantic level through structured natural-language descriptions partitioned into six Chain-of-Thought flow evolution stages. Conditioning the denoising process via cross-attention layers replaces loss-level physical penalties with semantic-level generative guid-ance, eliminating gradient competition at its source. Qualitative evaluation demonstrates clear reconstruction of alternating vortex cores, reasonable phase diversity across generat-ed samples, and richer nonlinear wake details compared to both baselines.

[1] R. Rombach, et al., "High-resolution image synthesis with latent diffusion models." Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR). p. 10684–10695. 2022.
[2] J. Ho, A. Jain, and P. Abbeel, "Denoising diffusion probabilistic models." Advances in Neural Information Processing Systems (NeurIPS). 33: p. 6840–6851. 2020.
[3] A. Vaswani, et al., "Attention is all you need." Advances in Neural Information Processing Systems (NeurIPS). 30: p. 5998–6008. 2017.
[4] A. Radford, et al., "Learning transferable visual models from natural language supervision." Proceedings of the International Conference on Machine Learning (ICML). 139: p. 8748–8763. 2021.
[5] O. Ronneberger, P. Fischer, and T. Brox, "U-Net: Convolutional networks for biomedical image segmentation." Proceedings of the International Conference on Medical Image Computing and Computer-Assisted Intervention (MICCAI). p. 234–241. 2015.
[6] M. Raissi, P. Perdikaris, and G.E. Karniadakis, "Physics-informed neural networks: A deep learning framework for solving forward and inverse problems involving nonlinear partial differential equations." Journal of Computational Physics. 378: p. 686–707. 2019.
[7] S. Wang, X. Yu, and P. Perdikaris, "When and why PINNs fail to train: A neural tangent kernel perspective." Journal of Computational Physics. 449: p. 110768. 2022.
[8] D. Shu, Z. Li, and A. Barati Farimani, "A physics-informed diffusion model for high-fidelity flow field reconstruction." Journal of Computational Physics. 478: p. 111972. 2023.
[9] N. Thuerey, et al., "Deep learning methods for Reynolds-averaged Navier–Stokes simulations of airfoil aerodynamics." AIAA Journal. 58(1): p. 25–36. 2020.
[10] H. Gao, L. Sun, and J.X. Wang, "PhyGeoNet: Physics-informed geometry-adaptive convolutional neural networks for solving parametric PDEs on arbitrary domains." Journal of Computational Physics. 428: p. 110079. 2021.
[11] C.H.K. Williamson, "Vortex dynamics in the cylinder wake." Annual Review of Fluid Mechanics. 28(1): p. 477–539. 1996.
[12] C.-G. Li, T. Makoto, and K. Onishi, "Feasibility investigation of compressible direct numerical simulation with a preconditioning method at extremely low Mach numbers." International Journal of Computational Fluid Dynamics. 28(6-10): p. 411–419. 2014.
[13] C.G. Li, et al., "Compressible direct numerical simulation with a hybrid boundary condition of transitional phenomena in natural convection." International Journal of Heat and Mass Transfer. 90: p. 654–664. 2015.
[14] C.G. Li, et al., "A sharp interface immersed boundary method for thin-walled geometries in viscous compressible flows." International Journal of Mechanical Sciences. 253. 2023.
[15] K. Komatsu, et al., "Parallel processing of the Building-Cube Method on a GPU platform." Computers & Fluids. 45(1): p. 122–128. 2011.
[16] M.D. de Tullio, et al., "An immersed boundary method for compressible flows using local grid refinement." Journal of Computational Physics. 225(2): p. 2098–2117. 2007.
[17] R. Ghias, R. Mittal, and H. Dong, "A sharp interface immersed boundary method for compressible viscous flows." Journal of Computational Physics. 225(1): p. 528–553. 2007.
[18] J.-H. Bastek, W. Sun, and D.M. Kochmann, "Physics-informed diffusion models." Proceedings of the International Conference on Learning Representations (ICLR 2025). 2025.
[19] Z. Dong, Z. Lu, and Y. Yang, "Fine-tuning a large language model for automating computational fluid dynamics simulations." Theoretical and Applied Mechanics Letters. 15: p. 100594. 2025.
[20] R. Li, et al., "Self-guided diffusion model for accelerating computational fluid dynamics." Proceedings of the International Conference on Machine Learning (ICML). 2025.
[21] J. Hu, Z. Lu, and Y. Yang, "Generative prediction of flow fields around an obstacle using the diffusion model." Physical Review Fluids. 10(9): p. 094903. 2025.
[22] J. Qiu, et al., "Pi-fusion: Physics-informed diffusion model for learning fluid dynamics." arXiv preprint arXiv:2406.03711. 2024.
[23] A. Zhou, et al., "Text2PDE: Latent diffusion models for accessible physics simulation." arXiv preprint arXiv:2410.01153. 2024.