CoNFiLD: Conditional Neural Field Latent Diffusion Model Generating Spatiotemporal Turbulence

P. Du, MH Parikh, X. Fan, XY Liu, J.-X. Wang. CoNFiLD: Conditional Neural Field Latent Diffusion Model Generating Spatiotemporal Turbulence. arXiv preprint arXiv:2403.05940. 2024 Mar 9. https://arxiv.org/abs/2403.05940

Generating equilibrium inflow turbulence of 3D channel flows

We demonstrate the CoNFiLD model on synthesizing sequences of instantaneous inlet velocity and pressure fields for 3D turbulent channel flows, highlighting its utility in generating accurate inflow turbulence boundary conditions, critical for eddy-resolving simulations.

Generating non-equilibrium turbulence of periodic hill

we further demonstrate CoNFiLD’s capability in generating spatiotemporal non-equilibrium turbulence flows through a classical periodic hill benchmark case, featuring a broad spectrum of complex flow behaviors including separation, recirculation, and reattachment.

Generating 3D wall-bounded turbulence with wall roughness

After showcasing CoNFiLD’s effectiveness in synthesizing cross-sectional spatiotemporal turbulence, we extend its application to a more challenging scenario: the spatiotemporal generation of sophisticated instantaneous wall-bounded turbulent flows within 3D domains featuring regular wall-roughness elements. Turbulent flows over a rough surface are ubiquitous in various naval systems due to manufacturing processes or service-induced erosion and biofouling. Different roughness conditions significantly affect near-wall turbulence structures and the transfer of scalar, momentum, and energy, impacting the safety, performance, and efficiency of marine systems.

Zero-shot full spatiotemporal flow reconstruction from sparse sensor data.

We first explore an application of significant practical importance: full-field spatiotemporal reconstruction of flow from sparse sensor data through zero-shot conditional generation, underpinned by Bayesian posterior sampling. This capability is essential across various engineering domains, where obtaining comprehensive full-field flow information is challenging due to complex setups, prohibitive computational costs, or the inherent sparsity and noise in direct measurements.

Zero-shot spatiotemporal super-resolution of low-fidelity data

Super-resolution techniques are rapidly being adopted across various computational and experimental communities to derive significant details from low-resolution (LR) images and data. Analytical, physics-based, and deep learning super-resolution techniques have shown promising results, from improving low-fidelity simulation results to enhancing under-resolved 4D flow MR imaging data. Motivated by these advancements, we present another capability of our proposed CoNFiLD model—creating highly detailed instantaneous flows from LR counterparts, showcasing significant potential for large-scale super-resolution challenges. Through the turbulence channel flow case, we demonstrate the zero-shot super-resolution capability of the trained CoNFiLD model, regardless of the quality of LR data.