Turbulence and Advanced Topics (Overview)

TL;DR

You'll see why turbulence turns a clean PDE into one of the hardest unsolved problems in physics. You'll learn the Reynolds number's role as the tipping point, the cascade of energy from big eddies to small ones, and why we average equations instead of solving them exactly. By the end you'll know what RANS, LES, and DNS actually mean and when each is used.

1. The Mental Model

Turbulence isn't a different equation — it's the same Navier-Stokes equation behaving chaotically once inertia dominates over viscosity. Smooth flow becomes a tangle of swirling eddies at every scale, from the size of the pipe down to the size where friction finally eats the motion. You can't track every eddy, so you either average the mess, simulate only the big parts, or brute-force the whole thing on a massive computer. Turbulence is order breaking into structured chaos — and taming it means choosing what detail you're willing to sacrifice.

2. The Core Material

2.1 Why Flows Become Turbulent: The Reynolds Number

The Navier-Stokes momentum equation for incompressible flow is:

$$\rho\left(\frac{\partial \mathbf{u}}{\partial t} + \mathbf{u}\cdot abla\mathbf{u}\right) = - abla p + \mu abla^2\mathbf{u}$$

Two forces fight here: the inertial term $\mathbf{u}\cdot
abla\mathbf{u}$ (momentum shoving fluid around) and the viscous term $\mu
abla^2\mathbf{u}$ (friction damping it out). The Reynolds number measures which one wins:

$$Re = \frac{\rho U L}{\mu} = \frac{UL}{ u}$$

where $U$ is a characteristic velocity, $L$ a characteristic length, and $
u = \mu/\rho$ the kinematic viscosity. Low $Re$: viscosity smooths out any disturbance before it grows — you get smooth, layered (laminar) flow. High $Re$: viscosity is too weak to damp small perturbations, so they amplify. Tiny wobbles cascade into large-scale chaos. For pipe flow, the classic transition happens around $Re \approx 2300$; above that, turbulence appears in patches and eventually takes over the whole flow.

This is why Navier-Stokes is deceptively simple-looking but practically unsolvable in general — the nonlinear term $\mathbf{u}\cdot
abla\mathbf{u}$ is precisely what generates this instability, and there's no closed-form way to predict exactly when or how it triggers turbulence for arbitrary geometry.

2.2 The Energy Cascade and Kolmogorov Scales

Turbulence isn't uniform chaos — it has structure across scales. Energy enters the flow at large scales (think: the size of the object stirring the fluid), then cascades down through progressively smaller eddies until it reaches a scale small enough that viscosity finally dissipates it as heat. Richardson described this poetically: "big whirls have little whirls that feed on their velocity, and little whirls have lesser whirls, and so on to viscosity."

Kolmogorov quantified this in 1941. He argued that at small enough scales, turbulence forgets how it was generated and depends only on two things: the energy dissipation rate $\varepsilon$ (energy lost to heat per unit mass per unit time) and viscosity $
u$. Dimensional analysis gives the Kolmogorov length scale:

$$\eta = \left(\frac{ u^3}{\varepsilon}\right)^{1/4}$$

Below $\eta$, eddies are smooth and get smoothed out by viscosity — no more cascading. This scale can be absurdly small: in atmospheric turbulence, $\eta$ is often around a millimeter, while the large eddies (storms) span kilometers. That's a ratio of scales spanning $10^6$ to $10^9$ in 3D — which is exactly why direct simulation is so expensive: you need a grid fine enough to resolve millimeter eddies across a domain kilometers wide.

2.3 Three Ways to Actually Compute Turbulence

Since resolving every scale exactly is often computationally impossible, there are three main strategies, trading accuracy for cost.

DNS (Direct Numerical Simulation): Solve Navier-Stokes directly, resolving every scale down to $\eta$. No modeling, no approximation — just brute force. The grid points needed scale roughly as $Re^{9/4}$ (3 spatial dimensions, each needing resolution proportional to $Re^{3/4}$). Double the Reynolds number and your computational cost explodes by roughly $2^{2.25} \approx 4.8\times$. This is why DNS is limited to modest Reynolds numbers and academic problems — you simply can't DNS a full aircraft wing.

LES (Large Eddy Simulation): Resolve the large, energy-containing eddies directly, but model the small ones you can't afford to resolve. This works because Kolmogorov's theory says small eddies are roughly universal — so a generic "subgrid-scale model" (like the Smagorinsky model) can stand in for them without needing exact detail. Much cheaper than DNS, still expensive.

RANS (Reynolds-Averaged Navier-Stokes): Don't resolve turbulent fluctuations at all — average them out entirely and solve for the mean flow only. Split velocity into mean and fluctuation:

$$u_i = \bar{u}_i + u_i'$$

Plug this into Navier-Stokes and average. Nonlinear terms produce an extra term called the Reynolds stress, $-\rho\overline{u_i'u_j'}$, representing momentum transport by turbulent eddies. This term is unknown — it depends on details you just averaged away. This is the closure problem: averaging creates more unknowns than equations, and you need a turbulence model (like $k$-$\varepsilon$ or $k$-$\omega$) to approximate the Reynolds stress in terms of mean-flow quantities. RANS is the cheapest and most widely used in engineering — cars, planes, pipes — because it's fast enough to iterate on designs.

flowchart TD
    A["Navier-Stokes at high Re"] --> B{"How much detail can you afford?"}
    B -->|"Full detail, huge cost"| C["DNS: resolve all scales down to eta"]
    B -->|"Resolve large eddies, model small ones"| D["LES: Smagorinsky subgrid model"]
    B -->|"Average out all fluctuations"| E["RANS: solve mean flow + closure model"]
    C --> F["Cost scales ~Re^2.25 — research use only"]
    D --> G["Moderate cost — used for complex unsteady flows"]
    E --> H["Cheapest — everyday engineering design"]

2.4 The Millennium Problem Connection

The Clay Institute's million-dollar problem asks: do smooth, physically reasonable solutions to the 3D Navier-Stokes equations exist for all time, or can they blow up (velocity going to infinity) in finite time? Turbulence is the physical manifestation of this mathematical uncertainty — the equations permit velocity gradients to sharpen without bound as energy cascades to ever-smaller scales, and nobody has proven whether this process stays finite or truly diverges. Kolmogorov's theory assumes a statistical steady cascade, but rigorously proving smoothness (or finding a blow-up) remains open. This is why turbulence isn't just "hard to simulate" — its mathematical foundations aren't fully settled either.

3. Worked Example

Let's estimate the DNS cost for two real flows and see why engineers reach for RANS or LES instead.

Flow 1: Flow around a small laboratory cylinder.
Say $U = 1$ m/s, $L = 0.02$ m (2 cm diameter), and water with $
u = 1\times10^{-6}$ m²/s.

$$Re = \frac{UL}{ u} = \frac{(1)(0.02)}{1\times10^{-6}} = 20{,}000$$

Grid points needed for DNS scale as $N \sim Re^{9/4}$:

$$N \sim (20{,}000)^{2.25} \approx 4.3 \times 10^{10}$$

That's about 43 billion grid points just spatially — before even counting the many time steps needed (time step also shrinks as resolution increases). This is at the ragged edge of what modern supercomputers can do for a research DNS study, and this is just a cylinder, not a real vehicle.

Flow 2: Airflow over a commercial aircraft wing.
Say $U = 250$ m/s (cruise speed), $L = 3$ m (chord length), air with $
u \approx 1.5\times10^{-5}$ m²/s.

$$Re = \frac{(250)(3)}{1.5\times10^{-5}} = 5 \times 10^{7}$$

$$N \sim (5\times10^7)^{2.25} \approx 3.6\times10^{18}$$

That's roughly 3.6 quintillion grid points. No computer on Earth today can handle this — not even close. This single number is the reason why aircraft designers use RANS models (with empirical turbulence closures tuned from wind-tunnel data) for daily design work, and reserve LES or hybrid RANS-LES only for small, critical regions like the trailing edge of a wing or a flap gap, where the mean-flow model isn't trustworthy enough.

The takeaway from the numbers themselves: going from $Re = 2\times10^4$ to $Re = 5\times10^7$ (a 2500× increase) inflates the DNS grid requirement by a factor of about $2500^{2.25} \approx 8.4\times10^7$. This superlinear blow-up in cost, driven directly by the $Re^{9/4}$ scaling from Kolmogorov's cascade argument, is the single most important practical fact in computational fluid dynamics: it's why almost no real-world engineering flow is ever solved by brute force.

4. Key Takeaways

4.1 Most Important Concepts

• Reynolds number is the switch: it's the ratio of inertial to viscous forces, and crossing a critical value flips flow from laminar to turbulent.
• The nonlinear term causes it all: $\mathbf{u}\cdot
  abla\mathbf{u}$ is the source of the instabilities that grow into turbulence — without it, Navier-Stokes would be linear and fully solvable.
• Energy cascades downward: large eddies transfer energy to smaller ones until viscosity dissipates it at the Kolmogorov scale $\eta$.
• DNS cost scales as $Re^{9/4}$: this single scaling law explains why turbulence simulation is so expensive and why full brute-force solutions are rare.
• The closure problem is unavoidable in averaging: RANS requires modeling the Reynolds stress term because averaging destroys information.
• LES is a middle ground: resolve what you can, model what's too expensive, relying on the (approximate) universality of small-scale turbulence.
• Navier-Stokes smoothness is still unproven in 3D: turbulence's fine-scale behavior connects directly to the unsolved Clay Millennium Problem.

4.2 Common Mis