Flow Lab

Syllabus / Module 11: Fluid Dynamics: Navier–Stokes and Beyond

Lesson 11.9

Turbulence: An Introduction

Transition, Reynolds decomposition, energy cascade, Kolmogorov scales, RANS vs LES vs DNS, and what that means for simulators.

Written byClaude Opus

🌊
Why this matters for simulation: Understanding why a real wind-tunnel simulator needs turbulence modeling.

By the end of this lesson you can…

  • Describe transition to turbulence and split a velocity into mean and fluctuation (Reynolds decomposition) with the averaging rules.
  • Average Navier–Stokes to find the Reynolds stress $-\rho\overline{u'v'}$, explain its sign, and explain the closure problem.
  • Explain the energy cascade, derive Kolmogorov's $k^{-5/3}$ spectrum and the scales $\eta$, $\tau_\eta$, $u_\eta$ by dimensional analysis, and evaluate them for the light wing's layer.
  • Derive the $\mathrm{Re}^{9/4}$ cost of direct simulation, evaluate it for the wing, and place RANS (k–ε) and LES (Smagorinsky) as models.
  • Compute an energy spectrum with the FFT and fit its slope over the right range.
On this page
·
  1. Transition: why a laminar layer breaks down
  2. Reynolds decomposition and averaging
  3. Reynolds stresses and the closure problem
  4. The energy cascade
  5. Kolmogorov’s −5/3 and the smallest scales
  6. Worked example: the light wing’s turbulent layer
  7. The cost of resolving everything: DNS, LES, RANS
  8. Computing it
  9. Concept checks
  10. Problems
  11. Key Takeaways

Last lesson’s laminar layer had a short life. On the light wing (a 1.51.5 m chord at 5050 m/s, ReL=5.135×106\mathrm{Re}_L = 5.135\times10^6) the local Reynolds number Rex=Ux/ν\mathrm{Re}_x = Ux/\nu reaches 5×1055\times10^5 at x=0.1461x = 0.1461 m, where a smooth plate’s layer starts to break down (from memory). Behind that point the Blasius profile is gone: the layer becomes turbulent, a churning sheet that mixes fast outer air down to the wall, rubs harder, and holds on longer against an adverse pressure gradient. This lesson asks why it forms, why averaging the equations leaves an unknown, where its energy goes, and why no simulator you will write can resolve it.

Tools: 11.8 Boundary Layers, 11.6 Vorticity Dynamics & Circulation, 11.5 Non-dimensionalization & the Reynolds Number and 7.10 Chaos & the Lorenz System.

▶ Video A laminar layer breaking into turbulence, a velocity split into mean and fluctuation, eddies stretching each other down to the Kolmogorov scale, and why a grid that holds them all is out of reachcaptions only
Transcript
  1. Last lesson ended on a promise: the laminar layer is only a lower bound. At Reynolds x of 5 times ten to the fifth, just 0.1461 metres back, it breaks down. The wing's Reynolds number is 5.135 times ten to the sixth, so the chord runs far past that point. Small waves grow, turn three-dimensional, and burst into turbulent spots that merge into one churning sheet.
  2. Turbulence is chaotic: two flows that start a hair apart soon differ completely. So we predict statistics, not instants. Split every quantity into a mean and a fluctuation, u equals u bar plus u prime. By construction the average of u prime is zero. But a product of fluctuations need not average out: u prime v prime is not zero in general.
  3. Average the momentum equation and the quadratic term leaves a stress behind: minus rho u prime v prime, momentum carried across the layer by eddies. It is positive, because eddies carry fast air down toward the wall. So a layer turbulent from the leading edge rubs 15.46 newtons per metre, against the laminar 2.693. But the new unknown has its own equation, with triple products, and those have fourth-order ones, forever. That is the closure problem.
  4. Where does the energy go? The mean shear feeds the largest eddies, the size of the layer. They stretch smaller eddies, by 11.6's vortex stretching, and those stretch smaller ones: Richardson's cascade. Epsilon, the energy per unit mass per second, passes through every scale at the same rate, about u cubed over ell, and ends as heat only where the eddies' own Reynolds number is about one. In two dimensions nothing stretches vorticity.
  5. Kolmogorov, in 1941: small eddies forget how they were made. In the inertial range the spectrum depends only on epsilon and k, and dimensional analysis forces E of k equals C epsilon to the two thirds, k to the minus five thirds. The synthetic field's FFT fits minus 1.676, against minus 1.667. Honest status: a hypothesis plus dimensional analysis, not a theorem.
  6. Now the wing's turbulent layer. Delta is 25.25 millimetres. The friction coefficient, 2.693 times ten to the minus three, gives u tau of 1.835 metres per second, and epsilon equals u tau cubed over delta: 244.6 watts per kilogram. The Kolmogorov length is 59.7 micrometres, and delta over eta is 423.
  7. To simulate directly, the grid must reach eta, so the point count is Reynolds number to the nine fourths. For the wing that is 1.25 times ten to the fifteen. Counting only the turbulent layer's volume still gives 3.55 times ten to the eleven points, 2.12 times ten to the fourth times a 256 cubed grid. Backwards, a 128 grid reaches only Reynolds number 645.
  8. So we model. RANS solves only the mean flow and models the Reynolds stress with an eddy viscosity. The k-epsilon model builds it from k squared over epsilon, with a constant, 0.09, fitted to simple flows. LES resolves the big eddies and models the small: Smagorinsky's viscosity. A filter at k c equals 16 keeps 79.1 percent of the energy but 10.4 percent of the dissipation. Both are tuned models, poor where wings get interesting.
  9. Your own solver models turbulence too, by accident. Semi-Lagrangian advection interpolates, which averages, and that adds a numerical viscosity, u delta x times one minus C over two. A sine wave at N equals 64 decays at 2.4834 measured, 2.4674 predicted. It is an uncontrolled, implicit subgrid model.
  10. No grid we can afford holds the wing's eddies. So the solver must at least get right what it cannot fudge: keeping the air incompressible on a grid. Next, 11.10: split the step, then enforce divergence zero with a pressure projection.

Transition: why a laminar layer breaks down

Blasius is an exact solution, but an exact solution can be unstable. That is 7.7 Nonlinear Systems & Linearization applied to a PDE: linearize about the laminar profile, look for modes eσte^{\sigma t}, and check the sign of the real part of σ\sigma. On a flat plate, wavy disturbances (Tollmien–Schlichting waves, from memory) grow once Rex\mathrm{Re}_x passes a critical value, turn three-dimensional, and break into turbulent spots that spread and merge. 11.8’s Rex≈5×105\mathrm{Re}_x \approx 5\times10^5 is typical in a quiet stream; roughness, noise and pressure gradients move it far. Accelerating flow delays it, which is how natural-laminar-flow airfoils keep long laminar runs (from memory).

The turbulent state is chaotic (7.10 Chaos & the Lorenz System): deterministic, but two flows that start a hair apart soon differ completely. Lorenz had three variables; by the count below, the wing’s layer has hundreds of billions. The instantaneous velocity is unpredictable; its statistics (means, variances, spectra) are reproducible, and engineering needs nothing else.

Reynolds decomposition and averaging

Split every quantity into a mean and a fluctuation:

u=uˉ+u′,v=vˉ+v′,p=pˉ+p′.u = \bar u + u', \qquad v = \bar v + v', \qquad p = \bar p + p' .

The bar is an average over a time long compared with the eddies but short compared with changes in the mean flow. Four rules follow:

  • u′‾=0\overline{u'} = 0, by construction.
  • uˉ v′‾=uˉ v′‾=0\overline{\bar u\,v'} = \bar u\,\overline{v'} = 0.
  • ∂u/∂x‾=∂uˉ/∂x\overline{\partial u/\partial x} = \partial\bar u/\partial x: averaging commutes with derivatives.
  • u′v′‾≠0\overline{u'v'} \ne 0 in general: a product of fluctuations need not average out.

So averaging a linear term puts a bar on it, while a product gives

uv‾=(uˉ+u′)(vˉ+v′)‾=uˉvˉ+u′v′‾.\overline{uv} = \overline{(\bar u + u')(\bar v + v')} = \bar u\bar v + \overline{u'v'} .

Reynolds stresses and the closure problem

With continuity, the advection term is uux+vuy=(u2)x+(uv)yuu_x + vu_y = (u^2)_x + (uv)_y, so 2D xx-momentum reads

(u2)x+(uv)y=−pxρ+ν(uxx+uyy).(u^2)_x + (uv)_y = -\frac{p_x}{\rho} + \nu(u_{xx} + u_{yy}) .

Averaging with the rules above:

(uˉ2)x+(uˉvˉ)y=−pˉxρ+ν(uˉxx+uˉyy)−(u′2‾)x−(u′v′‾)y.(\bar u^2)_x + (\bar u\bar v)_y = -\frac{\bar p_x}{\rho} + \nu(\bar u_{xx} + \bar u_{yy}) - \big(\overline{u'^2}\big)_x - \big(\overline{u'v'}\big)_y .

In a thin layer 11.8’s scaling keeps only the yy-derivative of the extra terms, and the mean boundary-layer equation becomes

uˉuˉx+vˉuˉy=UeUe′+1ρ∂τ∂y,τ=μuˉy−ρ u′v′‾.\bar u\bar u_x + \bar v\bar u_y = U_eU_e' + \frac{1}{\rho}\frac{\partial\tau}{\partial y}, \qquad \tau = \mu\bar u_y - \rho\,\overline{u'v'} .

The fluctuations act on the mean flow like an extra shear stress, the Reynolds stress −ρu′v′‾-\rho\overline{u'v'}: momentum carried across the layer by eddies.

Why it is positive, and why the wall rubs harder. In the layer uˉ\bar u grows with yy. A parcel moved down (v′<0v' \lt 0) comes from faster air, so it arrives faster than its new neighbours (u′>0u' \gt 0). A parcel moving up (v′>0v' \gt 0) carries slow air into a fast region (u′<0u' \lt 0). Either way u′v′<0u'v' \lt 0, so −ρu′v′‾>0-\rho\overline{u'v'} \gt 0: momentum flows toward the wall far faster than viscosity alone carries it. The mean profile fills out and the wall gradient steepens, and so does τw\tau_w. A from-memory turbulent-plate correlation, CD=0.074 ReL−1/5C_D = 0.074\,\mathrm{Re}_L^{-1/5}, puts the light wing’s two-face friction at 15.4615.46 N/m if turbulent from the leading edge, against 11.8’s laminar 2.6932.693 N/m (the cost panel prints both); the real wing, laminar for its first 10%10\%, sits between, near the top. The same fast near-wall air is why the layer holds on longer: it has more momentum to spend climbing an adverse gradient (11.8’s golf-ball argument).

The closure problem. Unknowns: uˉ\bar u, vˉ\bar v, pˉ\bar p and the new u′v′‾\overline{u'v'} (in 3D, six Reynolds stresses). Equations: continuity and momentum, one short (six in 3D). An equation for u′v′‾\overline{u'v'}, derived from Navier–Stokes, contains triple products such as u′u′v′‾\overline{u'u'v'}, whose equations contain fourth-order products, forever. The cause fits in one line: average x˙=−x2\dot x = -x^2 over noisy runs,

dxˉdt=−x2‾=−xˉ2−x′2‾,\frac{d\bar x}{dt} = -\overline{x^2} = -\bar x^2 - \overline{x'^2} ,

and the mean’s equation needs the variance. Every turbulence model is a choice of where to cut this chain and what to put there; none follows from Navier–Stokes alone.

The energy cascade

The mean shear feeds the largest eddies, the size of the layer. Their Reynolds number uℓℓ/νu_\ell\ell/\nu is large, so viscosity barely touches them. Instead they strain smaller eddies, by 11.6’s mechanism: a vortex tube stretched along its axis gets thinner and, with its circulation conserved (Kelvin), spins faster. Energy moves into smaller, faster eddies, which stretch smaller ones still: Richardson’s 1922 picture (from memory) of big whirls feeding little whirls. It turns into heat only at scales whose own Reynolds number is about 11.

Call the energy fed in per unit mass per second ε\varepsilon (W/kg == m2^2/s3^3). In a steady state the same ε\varepsilon passes through every scale and is dissipated at the bottom, and it is set at the top:

ε∼uℓ3ℓ,\varepsilon \sim \frac{u_\ell^3}{\ell} ,

(an eddy hands on its energy uℓ2u_\ell^2 in one turnover time ℓ/uℓ\ell/u_\ell). Lower the viscosity and the dissipation does not drop; the cascade just runs to smaller scales. That is an experimental fact, not a theorem.

The shell model below is a toy: one complex number unu_n per octave of wavenumbers kn=2nk_n = 2^n, coupled to its neighbours by quadratic terms that conserve energy as Navier–Stokes’ advection term does (the Sabra model; its form is from memory). Energy is forced into shell 00 and removed by −νkn2un-\nu k_n^2u_n. The right panel plots the time-averaged flux through shell nn (the rate nonlinearity carries energy past it) over the dissipation rate. In test runs at ν=10−6\nu = 10^{-6} it sits at 1.001.00 over the first ten shells, then drops to zero where viscosity takes over, at the printed kηk_\eta. That flat line is the cascade. From ν=10−4\nu = 10^{-4} to 10−710^{-7} the flat stretch lengthens while the measured ε\varepsilon stays between about 0.0770.077 and 0.0950.095.

◆ Interactive Energy cascade in a shell model: energy enters shell 0 and flows at a constant rate through the inertial range to the viscous end; slide the viscosity

Loading simulation…

Kolmogorov’s −5/3 and the smallest scales

Kolmogorov proposed in 1941 (from memory) that eddies much smaller than ℓ\ell forget how they were made: their statistics depend only on ε\varepsilon and ν\nu, and in the inertial range, well above the viscous scales, on ε\varepsilon alone.

The spectrum by dimensional analysis. E(k)E(k) is energy per unit mass per unit wavenumber (∫0∞E dk=12∣u′∣2‾\int_0^\infty E\,dk = \tfrac12\overline{|\mathbf u'|^2}), units m3^3/s2^2. In the inertial range it depends only on ε\varepsilon (m2^2/s3^3) and kk (1/m). Try E=CεakbE = C\varepsilon^ak^b, the 11.5 method:

m3 s−2=m2a s−3a m−b:s: −2=−3a⇒a=23,m: 3=43−b⇒b=−53.\mathrm{m}^3\,\mathrm{s}^{-2} = \mathrm{m}^{2a}\,\mathrm{s}^{-3a}\,\mathrm{m}^{-b}: \qquad \text{s: } -2 = -3a \Rightarrow a = \tfrac23, \qquad \text{m: } 3 = \tfrac43 - b \Rightarrow b = -\tfrac53 . E(k)=C ε2/3k−5/3,E(k) = C\,\varepsilon^{2/3}k^{-5/3} ,

with C≈1.5C \approx 1.5 measured (from memory). Its status: a hypothesis plus dimensional analysis, not a theorem from Navier–Stokes. It agrees well with wind-tunnel, jet and atmospheric data; small departures (intermittency, because small-scale activity is patchy) show up in higher statistics. The shell model shows one: its fitted slope of ∣un∣2‾\overline{|u_n|^2} lands between about −0.74-0.74 and −0.80-0.80 in test runs, against Kolmogorov’s −2/3-2/3 (each shell holds a band of width about knk_n, so E≈∣un∣2‾/knE \approx \overline{|u_n|^2}/k_n and −2/3-2/3 means −5/3-5/3). The flux plateau is the robust result; the exponent is a little off, as in real flows.

The Kolmogorov scales. At the bottom, ε\varepsilon and ν\nu (m2^2/s) fix one length, one time and one speed:

η=(ν3ε)1/4,τη=(νε)1/2,uη=(νε)1/4.\eta = \Big(\frac{\nu^3}{\varepsilon}\Big)^{1/4}, \qquad \tau_\eta = \Big(\frac{\nu}{\varepsilon}\Big)^{1/2}, \qquad u_\eta = (\nu\varepsilon)^{1/4} .

Their Reynolds number is uηη/ν=1u_\eta\eta/\nu = 1: viscosity can just smooth them. Combine with ε∼uℓ3/ℓ\varepsilon \sim u_\ell^3/\ell:

ℓη=ℓ(uℓ3ℓ ν3)1/4=(uℓℓν)3/4=Reℓ3/4.\frac{\ell}{\eta} = \ell\Big(\frac{u_\ell^3}{\ell\,\nu^3}\Big)^{1/4} = \Big(\frac{u_\ell\ell}{\nu}\Big)^{3/4} = \mathrm{Re}_\ell^{3/4} .

Worked example: the light wing’s turbulent layer

Assumptions: sea-level air (ν=1.4607×10−5\nu = 1.4607\times10^{-5} m2^2/s); a layer turbulent from the leading edge, with the from-memory correlations δ=0.37 L ReL−1/5\delta = 0.37\,L\,\mathrm{Re}_L^{-1/5} and Cf=0.0592 ReL−1/5C_f = 0.0592\,\mathrm{Re}_L^{-1/5}; largest eddies of size δ\delta moving at the friction velocity uτ=UCf/2u_\tau = U\sqrt{C_f/2} (layer fluctuations are a few uτu_\tau, from memory). An order-of-magnitude estimate; lab block A prints:

  1. ReL1/5=21.98\mathrm{Re}_L^{1/5} = 21.98, so at the trailing edge δ=0.37⋅1.5/21.98=25.25\delta = 0.37\cdot1.5/21.98 = 25.25 mm, nearly eight times the 3.253.25 mm laminar layer of 11.8.
  2. Cf=2.693×10−3C_f = 2.693\times10^{-3}, uτ=501.3465×10−3=1.835u_\tau = 50\sqrt{1.3465\times10^{-3}} = 1.835 m/s.
  3. ε=uτ3/δ=6.176/0.02525=244.6\varepsilon = u_\tau^3/\delta = 6.176/0.02525 = 244.6 W/kg.
  4. η=59.7 μ\eta = 59.7\ \mum, τη=2.44×10−4\tau_\eta = 2.44\times10^{-4} s, uη=0.2445u_\eta = 0.2445 m/s.
  5. Check: δ/η=423=(uτδ/ν)3/4=31713/4\delta/\eta = 423 = (u_\tau\delta/\nu)^{3/4} = 3171^{3/4}.

The smallest eddies are a twentieth of a millimetre across and live a quarter of a millisecond; nearer the wall they are smaller still (the wall unit ν/uτ\nu/u_\tau is 7.96 μ7.96\ \mum).

◆ Interactive A synthetic turbulent field with a prescribed spectrum: its FFT spectrum, the Reynolds number the grid can hold, and an LES filter that keeps the energy but loses the dissipation

Loading simulation…

The field is synthetic: random phases with E∝ksE \propto k^{s} from k=3k = 3 to a cut-off kd=3 Re3/4k_d = 3\,\mathrm{Re}^{3/4}; it has the spectrum but no coherent vortices, a test signal rather than turbulence. Its FFT spectrum fits −1.676-1.676 at Re=100\mathrm{Re} = 100 (prescribed −1.667-1.667). Past Re=149\mathrm{Re} = 149, kdk_d leaves the 256×256256\times256 grid.

The cost of resolving everything: DNS, LES, RANS

Direct numerical simulation (DNS) uses no model, so the grid spacing must reach η\eta while the box holds the largest eddies, LL: per direction L/η=Re3/4L/\eta = \mathrm{Re}^{3/4} points; in 3D,

Npoints∼(Re3/4)3=Re9/4.N_{\text{points}} \sim \big(\mathrm{Re}^{3/4}\big)^3 = \mathrm{Re}^{9/4} .

A step must not carry an η\eta-sized eddy past its own size, Δt≲η/U\Delta t \lesssim \eta/U, over a run of a few L/UL/U: Re3/4\mathrm{Re}^{3/4} steps, Re3\mathrm{Re}^3 work. Doubling the speed multiplies the grid by 29/4=4.762^{9/4} = 4.76 and the work by 88.

For the light wing (lab block B):

  • Textbook: ReL9/4=1.25×1015\mathrm{Re}_L^{9/4} = 1.25\times10^{15} points, ReL3=1.35×1020\mathrm{Re}_L^{3} = 1.35\times10^{20} work.
  • Sharper: turbulent volume over η3\eta^3 (both faces, 25.2525.25 mm thick, one metre of span): 3.55×10113.55\times10^{11} points, 2.12×1042.12\times10^4 times a 2563256^3 grid, and 2.51×1042.51\times10^4 steps per chord crossing, still ignoring the finer wall region.
  • Backwards: NN cells per side reach η\eta only if Re3/4≤N\mathrm{Re}^{3/4} \le N, so Re≤N4/3\mathrm{Re} \le N^{4/3}: 645645 for Stam’s 128128, 16251625 for 256256. The wing is at 5.135×1065.135\times10^6.
◆ Interactive Kolmogorov scales and grid counts for a wing: slide speed and chord, compare with grids a laptop runs (log scale)

Loading simulation…

RANS (Reynolds-averaged Navier–Stokes) solves only the mean-flow equations and models u′v′‾\overline{u'v'}, usually with Boussinesq’s eddy viscosity −u′v′‾=νt uˉy-\overline{u'v'} = \nu_t\,\bar u_y, where νt\nu_t belongs to the flow, not the fluid. The k–ε model builds νt\nu_t from the turbulent kinetic energy k=12∣u′∣2‾k = \tfrac12\overline{|\mathbf u'|^2} (m2^2/s2^2) and ε\varepsilon (m2^2/s3^3); the only viscosity they make is

νt=Cμk2ε,Cμ=0.09 (from memory),\nu_t = C_\mu\frac{k^2}{\varepsilon}, \qquad C_\mu = 0.09 \ \text{(from memory)},

with kk and ε\varepsilon each obeying a transport PDE whose terms are modelled, constants fitted to simple flows. It is poor exactly where a wing gets interesting: strong adverse gradients and separation. Aerodynamic codes favour other, also tuned, models (Spalart–Allmaras, k–ω SST, from memory). RANS is cheap because the grid resolves only the smooth mean profile.

LES (large-eddy simulation) resolves the energy-containing eddies and models the small ones. In the field sim a filter at kc=16k_c = 16 keeps 79.1%79.1\% of the energy but 10.4%10.4\% of the dissipation ∑k2E\sum k^2E, so the model must remove about 90%90\% of the cascade’s flux at the grid scale. Smagorinsky’s eddy viscosity uses the grid spacing Δ\Delta and the resolved strain rate ∣Sˉ∣|\bar S| (1/s); the only viscosity they make is

νt=(CsΔ)2∣Sˉ∣,Cs≈0.1–0.2 (from memory).\nu_t = (C_s\Delta)^2|\bar S|, \qquad C_s \approx 0.1\text{–}0.2 \ \text{(from memory)} .

It too is a model: it drains energy wherever there is shear, laminar regions and walls included, and must be damped there (damping functions or a dynamic procedure, from memory). An outer-layer LES of the wing with cells of δ/10\delta/10 needs 4.71×1064.71\times10^6 points (lab block B), laptop-sized; the wall region’s cost grows almost like Re2\mathrm{Re}^2 (from memory).

Computing it

🐍 PythonLab: the wing's Kolmogorov scales (A), the cost of DNS (B), numerical viscosity of semi-Lagrangian advection (C)
⌘/Ctrl + Enter

Block C’s measured rate is 0.6%0.6\% above the prediction because νnumk2\nu_{\text{num}}k^2 is the leading term of a Taylor series in kΔx=0.39k\Delta x = 0.39.

Spectra from data. A hot-wire in the wing’s layer records u(t)u(t); frozen eddies carried past give x=Utx = Ut (Taylor’s hypothesis, from memory). For a sampled periodic signal uju_j, j=0,…,N−1j = 0, \dots, N-1, the discrete Fourier transform u^m=1N∑juje−2πimj/N\hat u_m = \frac1N\sum_j u_je^{-2\pi imj/N} (NumPy’s rfft divided by NN) is the amplitude at wavenumber mm. For real uu, u^−m\hat u_{-m} is the conjugate of u^m\hat u_m, so the energy at ±m\pm m together is Em=∣u^m∣2E_m = |\hat u_m|^2, and Parseval reads ∑m≥1Em=12u′2‾\sum_{m \ge 1}E_m = \tfrac12\overline{u'^2}. The exercise’s signal has amplitudes m−5/6m^{-5/6}, so Em=14m−5/3E_m = \tfrac14m^{-5/3} exactly up to m=256m = 256, then falls exponentially: a known answer.

🧪 Coding exerciseExercise: compute an energy spectrum by FFT and fit the inertial-range slope
⌘/Ctrl + Enter

Graded: E10=0.0053861E_{10} = 0.0053861, Parseval (∑E=0.52345\sum E = 0.52345), the slope −5/3-5/3 over 4≤k≤1284 \le k \le 128, and the warning: over 1≤k≤20471 \le k \le 2047 the same data give −5.40-5.40. Measured data do not mark their inertial range; you choose it, and the answer depends on the choice.

Concept checks

✅ Concept check

The light wing's turbulent layer dissipates about $\varepsilon = 244.6$ W/kg in sea-level air ($\nu = 1.4607\times10^{-5}$ m$^2$/s). What is the Kolmogorov length $\eta = (\nu^3/\varepsilon)^{1/4}$, in micrometres?

✅ Concept check

Why does averaging the Navier–Stokes equations leave more unknowns than equations?

✅ Concept check

DNS needs about $\mathrm{Re}^{9/4}$ grid points. By what factor does the grid grow when $\mathrm{Re}$ grows tenfold?

✅ Concept check

In the field sim, a filter at $k_c = 16$ keeps $79.1\%$ of the energy but $10.4\%$ of the dissipation. What does that imply for an LES on that grid?

✅ Concept check

What is the status of Kolmogorov's $E(k) \propto \varepsilon^{2/3}k^{-5/3}$?

Problems

✏️ Pen & paperKolmogorov's scales by dimensional analysis●○○

Using only ν\nu (m2^2/s) and ε\varepsilon (m2^2/s3^3), find the unique length, time and speed. Show that their Reynolds number is 11 and that ℓ/η=Reℓ3/4\ell/\eta = \mathrm{Re}_\ell^{3/4} when ε=uℓ3/ℓ\varepsilon = u_\ell^3/\ell. Then show that τη/(ℓ/uℓ)=Reℓ−1/2\tau_\eta/(\ell/u_\ell) = \mathrm{Re}_\ell^{-1/2}.

Show worked solution

Length νaεb\nu^a\varepsilon^b: m: 2a+2b=12a + 2b = 1, s: −a−3b=0-a - 3b = 0, so η=(ν3/ε)1/4\eta = (\nu^3/\varepsilon)^{1/4}. Time: 2a+2b=02a + 2b = 0, −a−3b=1-a - 3b = 1, so τη=(ν/ε)1/2\tau_\eta = (\nu/\varepsilon)^{1/2}. Speed: η/τη=(νε)1/4\eta/\tau_\eta = (\nu\varepsilon)^{1/4}. Then uηη/ν=ν1/4ε1/4ν3/4ε−1/4/ν=1u_\eta\eta/\nu = \nu^{1/4}\varepsilon^{1/4}\nu^{3/4}\varepsilon^{-1/4}/\nu = 1. With ε=uℓ3/ℓ\varepsilon = u_\ell^3/\ell: ℓ/η=ℓ uℓ3/4ℓ−1/4ν−3/4=(uℓℓ/ν)3/4\ell/\eta = \ell\,u_\ell^{3/4}\ell^{-1/4}\nu^{-3/4} = (u_\ell\ell/\nu)^{3/4}. And τη=(νℓ/uℓ3)1/2\tau_\eta = (\nu\ell/u_\ell^3)^{1/2}, so τηuℓ/ℓ=(ν/(uℓℓ))1/2=Reℓ−1/2\tau_\eta u_\ell/\ell = (\nu/(u_\ell\ell))^{1/2} = \mathrm{Re}_\ell^{-1/2}. For the wing’s layer, Reδ=3171\mathrm{Re}_\delta = 3171: the smallest eddies turn over 5656 times faster than the largest.

✏️ Pen & paperThe sign of a Reynolds stress●●○

At a point in a shear layer, suppose u=U0+asin⁡tu = U_0 + a\sin t and v=−bsin⁡tv = -b\sin t with a,b>0a, b \gt 0. Find uˉ\bar u, vˉ\bar v, uv‾\overline{uv} and u′v′‾\overline{u'v'} (average over a period), and say which way momentum is carried.

Show worked solution

uˉ=U0\bar u = U_0, vˉ=0\bar v = 0, u′=asin⁡tu' = a\sin t, v′=−bsin⁡tv' = -b\sin t. u′v′‾=−ab sin⁡2t‾=−ab/2\overline{u'v'} = -ab\,\overline{\sin^2t} = -ab/2, so uv‾=uˉvˉ+u′v′‾=−ab/2≠uˉvˉ=0\overline{uv} = \bar u\bar v + \overline{u'v'} = -ab/2 \ne \bar u\bar v = 0. Upward motion comes with slow air: the Reynolds stress −ρu′v′‾=ρab/2>0-\rho\overline{u'v'} = \rho ab/2 \gt 0 carries xx-momentum downward, toward a wall below: the fast-air-down mechanism in its simplest form.

✏️ Pen & paperWhy DNS scales as Re^3●●○

A cube of side LL has largest eddies of size LL and speed UU. Show that resolving η\eta takes Re9/4\mathrm{Re}^{9/4} points and, with Δt=η/U\Delta t = \eta/U over a time L/UL/U, Re3\mathrm{Re}^{3} point-updates. By what factors do both grow if the wing flies at 100100 m/s instead of 5050? What Re\mathrm{Re} can a 5123512^3 grid hold?

Show worked solution

L/η=Re3/4L/\eta = \mathrm{Re}^{3/4} per direction, cubed: Re9/4\mathrm{Re}^{9/4}. Steps: (L/U)/(η/U)=Re3/4(L/U)/(\eta/U) = \mathrm{Re}^{3/4}; product Re3\mathrm{Re}^3. Doubling UU doubles Re\mathrm{Re}: points ×29/4=4.76\times2^{9/4} = 4.76, work ×8\times8. A 5123512^3 grid holds Re≤5124/3=4096\mathrm{Re} \le 512^{4/3} = 4096, under a thousandth of the wing’s 5.135×1065.135\times10^6.

Key Takeaways

  • Past a critical Rex\mathrm{Re}_x small waves in a laminar layer grow, break into spots, and the layer turns turbulent: near x=0.146x = 0.146 m on the light wing (Rex=5×105\mathrm{Re}_x = 5\times10^5, from memory).
  • Turbulence is deterministic chaos; we predict statistics, and with u=uˉ+u′u = \bar u + u' averages of products of fluctuations survive.
  • Averaging Navier–Stokes adds the Reynolds stress −ρu′v′‾-\rho\overline{u'v'}, positive because eddies carry fast air down: more friction (15.4615.46 against 2.6932.693 N/m for the wing), later separation. Its equation needs higher moments: the closure problem.
  • Vortex stretching cascades energy to small scales at a rate ε∼uℓ3/ℓ\varepsilon \sim u_\ell^3/\ell independent of ν\nu; the shell model shows the constant flux. 2D flow has no forward energy cascade.
  • Kolmogorov’s E=Cε2/3k−5/3E = C\varepsilon^{2/3}k^{-5/3} is a hypothesis plus dimensional analysis, not a theorem. η=(ν3/ε)1/4\eta = (\nu^3/\varepsilon)^{1/4}, ℓ/η=Reℓ3/4\ell/\eta = \mathrm{Re}_\ell^{3/4}; in the wing’s layer η=59.7 μ\eta = 59.7\ \mum.
  • DNS needs Re9/4\mathrm{Re}^{9/4} points and Re3\mathrm{Re}^3 work: 3.55×10113.55\times10^{11} points for the wing’s layer, while a 128128 grid holds Re=645\mathrm{Re} = 645. RANS (k–ε) and LES (Smagorinsky) are tuned models.
  • Semi-Lagrangian interpolation is a numerical viscosity uΔx(1−C)/2u\Delta x(1-C)/2: an implicit, uncontrolled subgrid model.
  • Next, 11.10: no grid we can afford holds the wing’s eddies, so the solver we build must at least get right the one thing it cannot fudge, keeping the air incompressible on a grid. 11.10 splits Navier–Stokes into steps and enforces ∇⋅u=0\nabla\cdot\mathbf u = 0 with a pressure projection.
Finished this lesson?

Mark it complete to track your progress (saved in this browser). Reading every section does this for you.