Channel flow DNS
Channel flow DNS: Theoretical Foundations
Overview
Professor, does DNS of channel flow hold a special position in the CFD world?
It's the most fundamental benchmark in turbulence research. The DNS at $Re_\tau = 180$ by Kim, Moin & Moser (1987) was groundbreaking and has since become the gold standard for turbulence model validation.
What is $Re_\tau$?
It's the friction Reynolds number, defined by the wall friction velocity $u_\tau$ and the channel half-width $\delta$.
Here, $\tau_w$ is the wall shear stress. $Re_\tau$ is a dimensionless number representing the ratio of the inner scale to the outer scale of wall turbulence.
Wall Law
Is the wall law also related?
The most important theoretical achievement for turbulent channel flow is the wall law. The distance from the wall is non-dimensionalized using inner variables.
In the viscous sublayer ($y^+ < 5$), there is a linear relationship:
In the logarithmic region ($y^+ > 30$), the log-law applies:
Here, $\kappa \approx 0.41$ (von Karman constant) and $B \approx 5.2$.
So DNS data is used to validate this law, right?
The DNS at $Re_\tau = 590$ by Moser, Kim & Mansour (1999) (commonly known as the MKM dataset) is an important dataset that confirmed the universality of the wall law with high precision. Currently, data up to $Re_\tau = 5200$ (Lee & Moser, 2015) exists.
Major DNS Databases
| Researchers | $Re_\tau$ | Grid Points | Year |
|---|---|---|---|
| Kim, Moin & Moser | 180 | $192 \times 129 \times 160$ | 1987 |
| Moser, Kim & Mansour | 180, 395, 590 | Up to $384 \times 257 \times 384$ | 1999 |
| Hoyas & Jimenez | 2003 | $6144 \times 633 \times 4608$ | 2006 |
| Lee & Moser | 5200 | $10240 \times 1536 \times 7680$ | 2015 |
10 billion grid points at $Re_\tau = 5200$? That's an incredible computational scale.
The story of how the world's first turbulent DNS in 1987 took "several weeks"
When Kim, Moin & Moser published their DNS (Direct Numerical Simulation) of channel flow in 1987, the computation took several weeks using the highest-performance supercomputer of the time. The Reynolds number was only 180 (based on friction velocity), with about 4 million grid points. Today, this calculation could be completed in a few days on a laboratory workstation. What made this paper groundbreaking was that it "numerically visualized the internal structure of turbulence for the first time." The moment streak structures and streamwise vortices near the wall appeared as predicted by theory, fluid dynamics researchers worldwide were reportedly thrilled. That paper is still cited thousands of times per year.
Computational Methods for Channel flow DNS
DNS Method
What kind of algorithm is used to solve DNS?
DNS (Direct Numerical Simulation) resolves all scales of turbulence. It solves the Navier-Stokes equations directly without models. For channel flow DNS, the pseudo-spectral method is mainstream.
Pseudo-Spectral Method
For the streamwise ($x$) and spanwise ($z$) directions, periodic boundary conditions allow Fourier series expansion; for the wall-normal direction ($y$), Chebyshev polynomial expansion is used.
The nonlinear term (convection term) is calculated in physical space and transformed to wavenumber space via FFT. This is why it's called the "pseudo" spectral method.
So the nonlinear term isn't calculated directly in wavenumber space.
Calculating the convection term in wavenumber space results in a convolution, increasing computational cost to $O(N^2)$. Calculating in physical space and using FFT keeps it at $O(N \log N)$. However, aliasing errors occur, so the 3/2 rule (de-aliasing) is applied.
Grid Resolution Requirements
DNS requires grids smaller than the Kolmogorov scale $\eta$. For channel flow, this is expressed in inner variables.
| Direction | Recommended Resolution | Remarks |
|---|---|---|
| Streamwise $\Delta x^+$ | 5~10 | Resolves streak structures |
| Spanwise $\Delta z^+$ | 3~5 | Streak structure width $\lambda_z^+ \approx 100$ |
| Wall-normal $\Delta y^+_{wall}$ | < 1 | Resolves viscous sublayer |
| Wall-normal $\Delta y^+_{center}$ | 5~10 | Channel center |
Time Integration
What about time integration?
A hybrid approach is standard: implicit method (Crank-Nicolson) for the viscous term and explicit method (3rd-order Adams-Bashforth or 3-stage Runge-Kutta) for the nonlinear term. The CFL condition is determined by the explicit part, typically $\Delta t^+ \approx 0.1$~$0.5$.
How long do you need to integrate to obtain statistics?
A wash-out time of $T u_\tau / \delta > 10$ is needed to reach statistical stationarity, followed by sampling requiring $T u_\tau / \delta > 20$~$50$. Convergence of statistics is slower for higher-order moments.
Why the spectral method became the "standard" for channel DNS
The spectral method (Fourier expansion + Chebyshev polynomials) is widely used in channel flow DNS because it can "achieve the same accuracy at lower cost" compared to finite difference methods. Resolving all scales of turbulence requires extremely high accuracy, and finite difference methods would require an extreme increase in grid points to keep up. On the other hand, the spectral method can solve smooth, periodic flows with "mathematically maximum efficiency." However, it also has drawbacks, such as difficulty with complex geometries. That's why benchmark DNS calculations tend to concentrate on simple shapes like channels and pipes, which in turn has led to the enrichment of databases.
Channel flow DNS in Practice
Practical Guide
How can I use DNS data to validate RANS models?
Performing DNS itself requires large-scale computation, but in practice, it's important to utilize publicly available data to validate the accuracy of RANS or LES.
Related Topics
Experience the theory firsthand with the interactive simulator for this field
All Simulators