Turbulence and Advanced Topics (Overview)

TL;DR

Turbulence is what happens when the nonlinear term in Navier-Stokes overwhelms viscosity, creating chaotic, multi-scale motion that's still deterministic but practically unpredictable. You'll learn how the Reynolds number sets the stage, how energy cascades from big eddies to small ones, and why we still don't have a rigorous proof of what happens to smooth solutions in 3D. This note gives you the map before you go deeper into any one region of it.

1. The Mental Model

Turbulence is the fluid version of a chain reaction: one twist creates smaller twists, which create smaller ones still, all the way down to where viscosity finally eats them. The Navier-Stokes equations don't change — the same PDE describes a calm stream and a raging rapid. What changes is how sensitive the flow becomes to tiny disturbances. Turbulence is order hiding inside chaos, governed by one equation that we can write down but can't fully solve.

2. The Core Material

2.1 The Reynolds Number and the Onset of Chaos

Photo by Steve A Johnson on Pexels

Everything starts with one dimensionless ratio:

$$Re = \frac{UL}{\nu}$$

where $U$ is a characteristic velocity, $L$ a characteristic length, and $\nu$ the kinematic viscosity. This number compares inertial forces (the nonlinear term $(\mathbf{u}\cdot\nabla)\mathbf{u}$) to viscous forces ($\nu\nabla^2\mathbf{u}$) in the momentum equation:

$$\frac{\partial \mathbf{u}}{\partial t} + (\mathbf{u}\cdot\nabla)\mathbf{u} = -\frac{1}{\rho}\nabla p + \nu \nabla^2 \mathbf{u}$$

At low $Re$, viscosity wins. Disturbances get smoothed out and the flow stays laminar — think honey, or slow flow through a narrow pipe. At high $Re$, the nonlinear term dominates. Small perturbations get stretched and amplified rather than damped, and the flow transitions to turbulence. Pipe flow famously transitions somewhere around $Re \approx 2300$, though the exact number depends on how carefully you avoid disturbing the inlet.

Why does the nonlinear term cause chaos and not just "faster laminar flow"? Because $(\mathbf{u}\cdot\nabla)\mathbf{u}$ couples different length scales together. A disturbance at one scale can generate motion at a completely different scale. This coupling is what viscosity, being linear, cannot do — it only smooths existing scales, it never generates new ones. That's the seed of everything that follows.

2.2 The Energy Cascade and Kolmogorov's Picture

Photo by Dr. Mohammad Hoque on Pexels

Once flow is turbulent, energy doesn't sit still — it moves through scales. Large eddies (set by the geometry of the problem, size $L$) are unstable and break into smaller eddies, which break into smaller ones, in a process called the energy cascade. This continues until eddies become so small that viscosity can finally dissipate their kinetic energy into heat.

Kolmogorov's 1941 theory (K41) gives a beautifully simple picture of this cascade for high-$Re$ turbulence. He assumed that at scales small enough, the flow "forgets" the large-scale geometry and depends only on two things: the rate of energy dissipation $\varepsilon$ (energy per unit mass per unit time) and viscosity $\nu$. Dimensional analysis alone gives you the smallest scale in the flow, called the Kolmogorov scale:

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

Between the large scale $L$ and the tiny scale $\eta$ sits the inertial subrange, where neither the large-scale forcing nor viscosity directly matters — only the cascade rate $\varepsilon$. In this range, Kolmogorov predicted the energy spectrum (how kinetic energy is distributed across wavenumber $k$):

$$E(k) = C\, \varepsilon^{2/3} k^{-5/3}$$

This $-5/3$ power law is one of the most experimentally confirmed results in all of fluid mechanics — measured in wind tunnels, the atmosphere, oceans, even astrophysical plasmas. It tells you turbulence isn't structureless noise; it has a precise statistical scaling law hiding inside the chaos.

The ratio of largest to smallest scales grows with Reynolds number:

$$\frac{L}{\eta} \sim Re^{3/4}$$

This is why turbulence simulations get so expensive. To resolve every scale in 3D, the number of grid points needed scales like $(L/\eta)^3 \sim Re^{9/4}$. Double the Reynolds number and your computational cost goes up by roughly a factor of 4.8. This single scaling law is why direct numerical simulation of turbulence at industrially relevant Reynolds numbers (millions) remains out of reach even on today's largest supercomputers.

flowchart TD
    A["Large eddies (scale L, energy injected)"] --> B["Eddy breakup: instability stretches and folds eddies"]
    B --> C["Medium eddies (inertial subrange, E(k) ~ k^-5/3)"]
    C --> D["Smaller eddies, cascade continues"]
    D --> E["Kolmogorov scale eta = (nu^3/epsilon)^1/4"]
    E --> F["Viscous dissipation: kinetic energy to heat"]

2.3 Turbulence Modelling: DNS, LES, and RANS

Photo by Lora Rikky on Pexels

Because fully resolving turbulence is so expensive, engineers and scientists use a hierarchy of approximations. Understanding where each sits tells you what you're trading away.

Direct Numerical Simulation (DNS) solves the Navier-Stokes equations exactly, resolving all scales down to $\eta$. It's the gold standard for accuracy but limited to modest Reynolds numbers (research-grade DNS today handles $Re$ up to maybe tens of thousands, far below a real aircraft wing's $Re \sim 10^7$).

Large Eddy Simulation (LES) resolves the large, energy-containing eddies directly but models the small ones using a "subgrid-scale model." The idea is that small eddies are more universal and easier to approximate statistically — this leans on the same universality Kolmogorov assumed in K41. A common closure is the Smagorinsky model, which adds an eddy viscosity:

$$\nu_t = (C_s \Delta)^2 |\bar{S}|$$

where $\Delta$ is the filter width (roughly the grid size), $\bar{S}$ is the resolved strain rate, and $C_s$ is a tunable constant. This is far cheaper than DNS but still resolves unsteady turbulent structures.

Reynolds-Averaged Navier-Stokes (RANS) goes further: it doesn't resolve any turbulent fluctuations directly. Instead you decompose velocity into a mean and fluctuation, $\mathbf{u} = \bar{\mathbf{u}} + \mathbf{u}'$, plug into Navier-Stokes, and average. This produces the RANS equations, but averaging the nonlinear term creates a new unknown — the Reynolds stress tensor $-\rho\overline{u_i' u_j'}$ — which represents the extra apparent stress caused by turbulent fluctuations. This is the origin of the closure problem: you now have more unknowns than equations, and you need a model (like $k$-$\epsilon$ or $k$-$\omega$) to close the system. RANS is by far the cheapest option and dominates industrial CFD, but it only gives you time-averaged flow behavior, not the instantaneous turbulent structures.

flowchart LR
    A["DNS: resolve everything, exact but Re-limited"] --> D["Decreasing cost, increasing modeling"]
    B["LES: resolve large eddies, model small scales"] --> D
    C["RANS: model all turbulence statistically via closure"] --> D

2.4 The Millennium Problem: Existence and Smoothness

Photo by Vitaly Gariev on Pexels

Zoom out from engineering to pure math, and there's a deeper mystery. The Clay Millennium Prize asks: given smooth, well-behaved initial data in 3D, does the Navier-Stokes equation have a smooth solution for all time — or can it "blow up" (develop infinite velocity or vorticity) in finite time?

Nobody knows. In 2D, the problem is solved — smooth solutions exist for all time, largely because vortex stretching (a key mechanism that can amplify vorticity without bound) is absent in 2D. In 3D, the vortex stretching term in the vorticity equation

$$\frac{D\boldsymbol{\omega}}{Dt} = (\boldsymbol{\omega}\cdot\nabla)\mathbf{u} + \nu\nabla^2\boldsymbol{\omega}$$

can in principle amplify vorticity