Du Pan, Parikh Meet Hemant, Fan Xiantao, Liu Xin-Yang, Wang Jian-Xun
Aerospace and Mechanical Engineering, University of Notre Dame, Notre Dame, IN, USA.
Mechanical and Aerospace Engineering, Cornell University, Ithaca, NY, USA.
Nat Commun. 2024 Nov 29;15(1):10416. doi: 10.1038/s41467-024-54712-1.
Eddy-resolving turbulence simulations are essential for understanding and controlling complex unsteady fluid dynamics, with significant implications for engineering and scientific applications. Traditional numerical methods, such as direct numerical simulations (DNS) and large eddy simulations (LES), provide high accuracy but face severe computational limitations, restricting their use in high-Reynolds number or real-time scenarios. Recent advances in deep learning-based surrogate models offer a promising alternative by providing efficient, data-driven approximations. However, these models often rely on deterministic frameworks, which struggle to capture the chaotic and stochastic nature of turbulence, especially under varying physical conditions and complex, irregular geometries. Here, we introduce the Conditional Neural Field Latent Diffusion (CoNFiLD) model, a generative learning framework for efficient high-fidelity stochastic generation of spatiotemporal turbulent flows in complex, three-dimensional domains. CoNFiLD synergistically integrates conditional neural field encoding with latent diffusion processes, enabling memory-efficient and robust generation of turbulence under diverse conditions. Leveraging Bayesian conditional sampling, CoNFiLD flexibly adapts to various turbulence generation scenarios without retraining. This capability supports applications such as zero-shot full-field flow reconstruction from sparse sensor data, super-resolution generation, and spatiotemporal data restoration. Extensive numerical experiments demonstrate CoNFiLD's capability to accurately generate inhomogeneous, anisotropic turbulent flows within complex domains. These findings underscore CoNFiLD's potential as a versatile, computationally efficient tool for real-time unsteady turbulence simulation, paving the way for advancements in digital twin technology for fluid dynamics. By enabling rapid, adaptive high-fidelity simulations, CoNFiLD can bridge the gap between physical and virtual systems, allowing real-time monitoring, predictive analysis, and optimization of complex fluid processes.
涡旋分辨湍流模拟对于理解和控制复杂的非定常流体动力学至关重要，对工程和科学应用具有重要意义。传统的数值方法，如直接数值模拟（DNS）和大涡模拟（LES），虽然提供了高精度，但面临着严重的计算限制，限制了它们在高雷诺数或实时场景中的应用。基于深度学习的替代模型的最新进展通过提供高效的数据驱动近似提供了一种有前途的替代方案。然而，这些模型通常依赖于确定性框架，难以捕捉湍流的混沌和随机性质，特别是在不同的物理条件和复杂、不规则几何形状下。在这里，我们介绍了条件神经场潜扩散（CoNFiLD）模型，这是一种生成学习框架，用于在复杂的三维域中高效、高保真地随机生成时空湍流。CoNFiLD将条件神经场编码与潜扩散过程协同集成，能够在不同条件下高效且稳健地生成湍流。利用贝叶斯条件采样，CoNFiLD无需重新训练即可灵活适应各种湍流生成场景。这种能力支持诸如从稀疏传感器数据进行零样本全场流重建、超分辨率生成和时空数据恢复等应用。大量数值实验证明了CoNFiLD在复杂域中准确生成非均匀、各向异性湍流的能力。这些发现强调了CoNFiLD作为一种通用、计算高效的实时非定常湍流模拟工具的潜力，为流体动力学数字孪生技术的进步铺平了道路。通过实现快速、自适应的高保真模拟，CoNFiLD可以弥合物理系统和虚拟系统之间的差距，实现对复杂流体过程的实时监测、预测分析和优化。
