A framework of the transport model for high-order eddy viscosity tensor in two-dimensional turbulent flow

Xingguang Zhou, Dalin Zhang, Wenxi Tian, G.h. Su, Suizheng QiuarXiv 2023 · DOI · arXiv

This academic paper presents a new theoretical framework for turbulence modeling, specifically focusing on a transport model for a high-order eddy viscosity tensor. By moving beyond the traditional scalar approach, the authors aim to better capture the directional dependencies (anisotropy) of turbulent flows.

TL;DR

Researchers have developed a transport model for a fourth-order eddy viscosity tensor ($\zeta_{ijkl}$) derived directly from second-order moment equations. Unlike standard models that assume turbulence is the same in all directions, this framework successfully predicts complex phenomena like flow separation in diffusers and aligns closely with Direct Numerical Simulation (DNS) data.

What Problem Does This Paper Address?

Most industrial Computational Fluid Dynamics (CFD) relies on the Boussinesq hypothesis, which relates Reynolds stress to the mean velocity gradient using a single scalar value: the eddy viscosity ($\nu_t$). While computationally efficient, this "isotropic" assumption is a major flaw. It assumes turbulence acts like a simple fluid where friction is the same in every direction.

In reality, turbulence is highly anisotropic—it has "memory" (history effects) and behaves differently depending on the direction of the flow, especially near walls or in regions with pressure changes. Existing "Non-Linear Eddy Viscosity Models" (NLEVM) try to fix this with complex mathematical expansions, but these are often empirical and can become numerically unstable. The authors seek a more fundamental, theoretically grounded transport equation for the eddy viscosity tensor itself.

Methods

The authors employ tensorial analysis to derive a transport model for a fourth-order eddy viscosity tensor in a 2D-3C (two-dimensional flow, three-component velocity) structure.

  1. Derivation: They started with the second-order moment transport equation (specifically the LRR-IP model) and used the derivative of the tensor function to extract the transport characteristics of the eddy viscosity tensor.
  2. The Equation: The resulting transport equation for $\zeta_{ijkl}$ includes terms for:
  3. Transient/Convection: How the tensor moves and changes over time.
  4. Production (Strain & Vorticity): How the tensor is generated by flow deformation and rotation.
  5. Diffusion: How turbulence intensity spreads.
  6. Dissipation: How turbulent energy is lost to heat.
  7. Numerical Implementation: The model was implemented in OpenFOAM using the SIMPLE algorithm. To maintain numerical stability, they solved a total of 17 equations (including $k$, $\epsilon$, and 12 independent components of the $\zeta$ tensor).

Datasets

The model was validated against two classic benchmarks in fluid mechanics:

  • 2D Straight Channel Flow: Compared against the landmark DNS dataset by Kim et al. (1987) at $Re = 40,000$.
  • Asymmetric Planar Diffuser: Compared against the experimental measurements of Obi et al. (1993). This case is particularly difficult because it involves an adverse pressure gradient that causes the flow to detach from the wall (separation).

Key Insights and Ideas

  • The Lumley Triangle Validation: The authors used the "Lumley triangle"—a visual tool to define the possible states of turbulence anisotropy. Their model’s trajectory in the triangle matched DNS data almost perfectly, whereas the standard $k-\epsilon$ model stays stuck at the origin (assuming perfect isotropy).
  • Capturing the "Separation Bubble": Standard $k-\epsilon$ models famously fail to predict flow separation in diffusers. This new model successfully "saw" the separation bubble, proving that capturing Reynolds stress anisotropy is the key to predicting complex flow structures.
  • Physical Meaning of Negative Components: The paper discusses how components of the tensor can be negative. This corresponds to Converse Gradient Transport (CGT), where momentum moves from low-region to high-region—a phenomenon that standard scalar models physically cannot represent.

Key Results

  • Accuracy Boost: In the diffuser case, the prediction accuracy for streamwise mean velocity increased by nearly 300% compared to the standard $k-\epsilon$ model and 200% compared to the LRR (Reynolds Stress Model).
  • Separation Points: The model predicted flow detachment at $6.16H$ and reattachment at $27H$. While the detachment point was slightly earlier than the experimental $11H$, it was a massive improvement over standard models that predicted no separation at all.
  • Dominant Components: Numerical analysis revealed that $\zeta_{1111}$ and $\zeta_{2222}$ are the primary drivers of the Reynolds viscous stress in these flows.

Limitations and Future Work

  • Computational Cost: Solving 17 coupled non-linear equations is significantly more expensive than the 2 equations required for $k-\epsilon$. The authors admit the model is currently more of a theoretical tool than a practical engineering one.
  • Numerical Stability: To get the simulation to converge, the authors had to use an approximation for the production term ($P_k^{apx}$). They suggest that a more robust numerical algorithm is needed to solve the "strict" form of the equations.
  • Machine Learning Integration: The authors propose using machine learning or neural networks in the future to calibrate a "damping function" (similar to the one used in the Spalart-Allmaras model) to make the tensor model more stable and easier to use in industrial applications.

Summary updated

Abstract

Motivated by the concept of eddy viscosity tensor in improved Boussinesq hypothesis, a transport model of high-order eddy viscosity tensor in 2D-3C turbulence structure is derived from the second-order moment model by tensorial analysis.
BibTeX
@misc{zhou2023framework,
  author = {Xingguang Zhou and Dalin Zhang and Wenxi Tian and G.H. Su and Suizheng Qiu},
  title = {{A framework of the transport model for high-order eddy viscosity tensor in two-dimensional turbulent flow}},
  year = {2023},
  eprint = {2302.05626},
  archivePrefix = {arXiv},
  url = {https://arxiv.org/abs/2302.05626},
}
Figure