Diagnostics for Eddy Viscosity Models of Turbulence Including Data-Driven/Neural Network Based Parameterizations

William J. Layton, Michael SchneierarXiv 2019 · arXiv

Introduction

In computational fluid dynamics (CFD), resolving every scale of motion in high-Reynolds-number flows is often computationally impossible. To address this, researchers use eddy viscosity models to simulate the energy-draining effects of unresolved small-scale turbulence. While these models are essential, they frequently act as "black boxes" that can suffer from over-diffusion, damping fluid motion to the point of physical inaccuracy.

With the recent surge in using Neural Networks (NNs) to parameterize these models, the need for theoretical safeguards has become urgent. Unlike classical physics-based models, NN-based parameterizations often lack stability guarantees. This paper provides a mathematical "health check" for these simulations, deriving computable conditions to determine when a model is necessary and detecting when it is failing.

The Taylor Microscale and Mesh Necessity

A fundamental question in CFD is determining when a mesh is fine enough to render an eddy viscosity model unnecessary. Traditionally, it was believed that a model is required unless the mesh resolves the tiny Kolmogorov scale ($h \approx Re^{-3/4}$), where viscous dissipation dominates.

However, this research establishes a more lenient benchmark: the Taylor microscale ($h \approx Re^{-1/2}$). The authors prove that if the mesh width $h$ resolves the Taylor microscale, the aggregate energy dissipation is handled correctly without additional modeling. Conversely, if $h$ is significantly larger than the Taylor microscale, the simulation will be under-diffused, causing energy to "pile up" at the smallest resolved scales. This finding provides a practical threshold for practitioners to evaluate their simulation setups.

A Posteriori Energy Diagnostics

To monitor model performance during a simulation, the authors utilize a posteriori energy estimates. Rather than predicting behavior a priori, they derive bounds for energy dissipation that can be calculated directly from active simulation data.

They decompose the model's performance into three monitorable, non-dimensional statistics:

  1. Average turbulent viscosity: $avg(\nu_{turb})/LU$
  2. Average model length scale: $avg(l)/L$
  3. Predicted turbulent intensity: $I_{model}$

By monitoring these quantities, researchers can identify "red flags" in real-time. If these values deviate significantly from $O(1)$, the model is likely over-dissipating energy.

Identifying the Source of Failure

The paper provides a rigorous mathematical framework to isolate why a model might be failing. Through a series of theorems, the authors prove that the time-averaged energy dissipation rate is bounded by the average of the chosen eddy viscosity parameterization:

$$\text{avg}(\nu_{turb}) / LU \leq \mu \cdot (\text{avg}(l)/L) \cdot \sqrt{I_{model}}$$

This inequality allows researchers to determine if over-diffusion is caused by an incorrect length-scale parameterization ($l$) or an inaccurate prediction of turbulent intensity ($I_{model}$). For those developing data-driven or NN-based models, these bounds offer a path toward physics-informed machine learning. These energy constraints can be integrated into the training process of a Neural Network to ensure the resulting model adheres to physical laws and avoids excessive damping.

Limitations and Future Directions

While this framework provides a robust diagnostic for the fluid bulk, the authors note two primary areas for future development:

  • Boundary Layers: The current analysis assumes velocity and body forces vanish at the boundaries. In engineering applications like aerodynamics, the complex shear flows in boundary layers require a specialized energy dissipation analysis.
  • Numerical Dissipation: Practical CFD codes often introduce "artificial" viscosity through their numerical schemes. Future work must integrate the effects of this discretization-induced dissipation with the model-induced dissipation discussed here.

By providing these diagnostic tools, the paper moves turbulence modeling away from "black box" guesswork and toward a more transparent, mathematically grounded methodology.

Summary updated

BibTeX
@misc{layton2019diagnostics,
  author = {William J. Layton and Michael Schneier},
  title = {{Diagnostics for Eddy Viscosity Models of Turbulence Including Data-Driven/Neural Network Based Parameterizations}},
  year = {2019},
  eprint = {1911.02491},
  archivePrefix = {arXiv},
  url = {https://arxiv.org/abs/1911.02491},
}
Figure