Abstract:Classical eddy viscosity models add a viscosity term with turbulent viscosity coefficient whose specification varies from model to model. Turbulent viscosity coefficient approximations of unknown accuracy are typically constructed by solving associated systems of nonlinear evolution equations or by data driven approaches such as deep neural networks. Often eddy viscosity models over-diffuse, so additional fixes are added. This process increases model complexity and decreases model comprehensibility, leading to the following two questions: Is an eddy viscosity model needed? Does the eddy viscosity model fail? This report derives a posteriori computable conditions that answer these two questions.
What problem does this paper attempt to address?
This paper attempts to solve the following two key problems:
1. **Is it necessary to use the eddy viscosity model?**
2. **Will the eddy viscosity model fail (i.e., over - dissipation)?**
### 1. Is it necessary to use the eddy viscosity model?
Traditional eddy viscosity models simulate turbulence by adding a viscous term with a turbulent viscosity coefficient \(\nu_{\text{turb}}\). However, these models usually lead to under - diffusion when the grid cannot resolve energetically significant vortices (for example, when the grid size \(\Delta x\approx\text{Re}^{-3/4}L\), i.e., the Kolmogorov micro - scale). Therefore, the classical view is that the eddy viscosity model needs to be used in this case.
However, this paper shows through theoretical analysis that if the grid can resolve the Taylor micro - scale (i.e., \(\Delta x\approx\sqrt{15}\text{Re}^{-1/2}L\)), the overall flow will not have an under - diffusion phenomenon. Specifically, Theorem 1 shows that when the grid size satisfies certain conditions, the overall dissipation rate of the flow will be close to zero, which means that the eddy viscosity model is not necessary in this case.
### 2. Will the eddy viscosity model fail?
The most common failure mode of the eddy viscosity model is over - damping, which will cause excessive dissipation of the solution energy, making the flow too smooth or even reaching a non - physical steady state. To evaluate this, the author introduced a computable statistic - the average model dissipation rate - and proved its relationship with model parameterization.
Specifically, Theorem 2 shows that the time - averaged total energy dissipation rate can be expressed as:
\[
\langle\varepsilon_0+\varepsilon_{\text{turb}}\rangle\leq\left(\frac{2}{2 - \beta}+\frac{2}{\beta(2 - \beta)\text{Re}^{-1}}+\frac{1}{\beta(2 - \beta)}\frac{\text{avg}(\nu_{\text{turb}})}{LU}\right)\frac{U^3}{L}
\]
where \(\text{avg}(\nu_{\text{turb}})\) is the average turbulent viscosity coefficient, \(\text{Re}\) is the Reynolds number, and \(U\) and \(L\) are the characteristic velocity and length scale, respectively.
Furthermore, Theorem 3 decomposes the average turbulent viscosity coefficient into the influence of the mixing length \(l\) selected by the model and the turbulence intensity \(k'\):
\[
\frac{\text{avg}(\nu_{\text{turb}})}{LU}\leq\mu\frac{\text{avg}(l)}{L}\sqrt{I_{\text{model}}(u)}
\]
where \(\mu\) is a calibration parameter and \(I_{\text{model}}(u)\) is the turbulence intensity predicted by the model.
By monitoring these computable statistics, it can be determined whether the eddy viscosity model is over - dissipated, and the model parameterization can be improved according to the spatial distribution information.
### Summary
This paper solves the problems of whether the eddy viscosity model is necessary and whether it will fail by introducing computable statistics. The research shows that when the grid can resolve the Taylor micro - scale, the eddy viscosity model may not be necessary; when using the eddy viscosity model, the performance of the model can be evaluated by monitoring statistics such as the average turbulent viscosity coefficient to avoid the over - dissipation problem.