Turbulence

The Number That Predicts the Transition

In 1883, Osborne Reynolds set up a transparent glass pipe in his Manchester laboratory and injected a thin filament of dye into water flowing along it. At low speeds the dye stayed a single straight thread across the entire pipe length. Speed it up, and at a specific threshold the thread began to wobble, broke into wavy filaments, and finally dispersed into a turbid cloud. Reynolds realized that the threshold depended not on any single quantity but on a combination: the product of fluid density, average speed, and pipe diameter, divided by the fluid’s viscosity. The combination is dimensionless. We now call it the Reynolds number, and a century and a half later it is still the first thing anyone computes about a flow.

For pipe flow the threshold sits at a Reynolds number near 2,300. Below that the flow stays laminar. Above it, a large enough disturbance at the inlet sets off turbulence that keeps itself going, and ordinary pipes always have such disturbances. The 2,300 threshold is specific to circular pipes; boundary layers, jets, and shear flows each have their own critical numbers. But the general story is universal: low Reynolds means viscosity wins, the flow stays orderly. High Reynolds means inertia wins, and order collapses into chaos.

The Reynolds numbers of the real world span an absurd range. A bacterium swimming through water lives at a Reynolds number of about a ten-thousandth, where viscosity is so dominant that the bacterium feels water as we would feel honey. Blood in the finest vessels: about a thousandth. Blood leaving the heart at the peak of each beat: several thousand, enough for brief bursts of turbulence. A garden hose: about 10,000. A river: about ten million. An airliner’s wing in cruise: about a hundred million. The Sun’s churning outer layers, which generate its magnetic field: roughly 10¹³. Jupiter’s Great Red Spot: a hundred times lower, around 10¹¹. Flow in a pipe can stay smooth far above 2,300 only with extreme care over the water coming in; the most careful experiments have held it smooth to about a hundred thousand.

The Energy Cascade

Andrei Kolmogorov, in three short papers in 1941, proposed the deepest organizing idea anyone has had about turbulence. Forget the details of how the flow is stirred, he said. Once turbulence is well developed, energy flows downhill through scales like water through nested sieves. You inject energy at some large scale – the propeller stirring the swimming pool, the wing slicing through air, the Sun heating the convection zone. The big eddies are unstable and break apart into smaller eddies. Those smaller eddies break apart into yet smaller eddies. The process repeats over many orders of magnitude until finally the eddies are so tiny that viscosity overwhelms their motion and turns it into warmth, the jiggle of the fluid’s molecules.

i Energy runs down to ever smaller eddies, then becomes warmth ×

Between the input scale and the dissipation scale, Kolmogorov argued, there is a wide range of sizes where neither stirring nor viscosity matter directly. Eddies there only hand their energy on to smaller ones; physicists call this middle stretch the inertial range. In it the statistics depend only on the rate at which energy flows down the cascade and on the size of the eddy you are asking about. From that alone, by matching units, he predicted how the energy is shared among sizes: an eddy twice as large holds about 1.6 times as much, the two-thirds power of two. Written as a spectrum, the way physicists usually plot it, the same rule is the minus five-thirds law. The minus five-thirds law is one of the most universally confirmed scaling relations in all of physics. It shows up in tidal channels, in jet engines, in supernova remnants, in atmospheric soundings – anywhere with enough Reynolds number to develop a long inertial range.

The smallest scale of the cascade has a name too. The Kolmogorov dissipation scale is the size at which viscous diffusion takes over. For an atmospheric eddy a kilometer across, the dissipation scale is roughly a millimeter. So a gust 100 meters wide can cascade down five orders of magnitude before viscosity finally erases it. That ratio is why direct numerical simulation of any realistic flow is so brutal: the grid has to span both scales at once, and the total computing cost grows roughly as Reynolds number cubed.

Kolmogorov’s theory is also famously incomplete. Real turbulence is not statistically uniform. Energy dissipation is intermittent, concentrated in thin filaments and sheets that occupy a tiny fraction of the volume. Higher-order statistics of velocity fluctuations deviate measurably from the smooth Kolmogorov predictions. Kolmogorov himself patched this in 1962 with a correction for the patchiness. The patch helps but does not close the issue. Intermittency remains one of the deep unsolved features of the cascade.

Turbulence Is Not Noise

The biggest surprise of twentieth-century turbulence research is that what looks chaotic from a distance contains organized, repeatable structures up close. They are not stable patterns the way a planetary orbit is stable. They flicker into existence, persist for many turnover times, then dissolve. But while they last they carry a disproportionate share of the energy and a disproportionate share of the dissipation. Three of them have become canonical.

i Vortex tubes, thin filaments of concentrated rotation, carry most of the dissipation ×

Vortex tubes are the most striking. They appear in direct simulations as long, thin, intense filaments of swirl – sometimes called “worms” in the early papers. Each filament is roughly as wide as the Kolmogorov dissipation scale but can be hundreds of times longer. They occupy less than a few percent of the total volume yet carry most of the flow’s spin, counted as the square of how fast the fluid turns at each point. They are the geometric face of intermittency.

Hairpin vortices are the dominant structure of turbulent boundary layers, where flow meets a wall. Theodorsen hypothesized them in 1952 from indirect measurements; modern particle-image velocimetry has confirmed them as U-shaped or lambda-shaped vortex loops arching out of the wall in self-generating packets. Walk through any turbulent boundary layer – the air over a wing, water over a ship hull – and you walk through a forest of hairpins.

Large-scale streamwise streaks appear close to walls as alternating high-speed and low-speed bands of fluid, roughly a thousand viscous units long. They are responsible for most of the wall’s drag on the flow above it. The takeaway from these three: turbulence is not pure noise. It is a fierce balance between coherent organization at certain scales and chaotic mixing between them, which is part of why it is so hard to model with simple statistics.

Why Planes Fly – and What the Textbooks Get Wrong

The usual story says the air over a wing’s curved top has farther to go, so it must hurry to meet the air from underneath at the back edge, and fast air has low pressure. The flaw is the meeting. Nothing makes the two streams rejoin: smoke let into a wind tunnel shows the air over the top reaching the back edge well before the air beneath. Air sped up only enough to keep that date would hold up a small fraction of a real plane.

i A wing turns a mass of air downward; the equal-and-opposite reaction lifts the plane ×

A wing meets the air tilted up a little, and it turns the passing air downward. Turning the air takes a push, and the air pushes back on the wing just as hard: that push back is lift. The low pressure over the top and the air sent down are one event counted at its two ends, and they give the same number. The air does not carry the push away. Each layer of sinking air presses on the one below, until the plane’s weight rests on the ground as a slight rise in pressure, spread too wide to feel. How much lift a wing gets is settled at its sharp back edge, where the air from above and below must flow off smoothly, side by side, with none curling round the edge. That sets how strongly the air circles the wing, and the lift on each meter of wing is the air’s density times the speed times that circling. Tilt the wing further and the circling, and the lift, grow.

Boundary Layers and the Drag Crisis

Fluid touching a surface clings to it, so right at the surface its speed is zero. The thin layer in which the speed climbs from zero to that of the flow outside is called the boundary layer, identified by Prandtl in 1904. It is one of the most consequential concepts in classical physics. Every friction calculation, every heat-transfer estimate, every prediction of where a flow will separate from a surface, rides on what the boundary layer is doing.

In 1912 Gustave Eiffel – the same Eiffel – discovered something counterintuitive. The drag on a smooth sphere does not rise steadily with speed. Around a Reynolds number of 300,000 it suddenly falls, even as the speed goes up. Measured against the full push of the oncoming air on the ball’s cross-section, the drag drops from about half of it to about a fifth, and lower still on a very smooth ball. The reason: at high enough speed, the boundary layer wrapping the sphere transitions from laminar to turbulent before it separates from the rear. A turbulent boundary layer stirs fast air down close to the surface, so it can push on against the rising pressure toward the back for longer. Separation moves rearward, the wake shrinks, and total drag plummets. This is the drag crisis.

i Dimples trip the boundary layer turbulent, the wake shrinks, and drag drops ×

Golf ball dimples are there to cheat the drag crisis. A skilled golfer cannot hit a smooth ball fast enough to push it past the crisis on its own: the ball would have to leave the tee at over 300 kilometers per hour. So manufacturers add dimples that trip the boundary layer turbulent at a much lower speed, and the dimpled ball is past the crisis from the moment it is struck. Its drag is about half a smooth ball’s, and it carries roughly twice as far. Halving the drag alone accounts for about 1.6 times the carry. The rest is lift: a ball struck with backspin is lifted by the spin, and dimples make that lift steady, where a spinning smooth ball gets it far less reliably, at some speeds even pushed the wrong way. The same trick of tripping the air turns up in the textured sleeves of racing cyclists’ suits and in the small fins on airliner wings, called vortex generators, that keep the air from peeling off. It is one of the few cases where deliberately roughening a surface reduces drag, and it works only because turbulence is what saves you.

When the Fluid Conducts Electricity

Liquid metals, ionized gases, and astrophysical plasmas obey Navier-Stokes coupled to Maxwell’s equations. The combined framework is magnetohydrodynamics, or MHD. The magnetic field exerts a Lorentz force on flowing charges; the moving charges in turn alter the field. At high electrical conductivity the field lines become nearly frozen into the fluid – you can move them around the way you move a thread through cloth, but you cannot pull one out of the cloth without ripping something.

MHD is the language for the Sun, the solar wind, the cores of planets, every fusion reactor humans have built, and most of the interstellar medium. Inside the Sun, differential rotation winds the magnetic field up like a clock spring at the boundary between the evenly turning interior and the convecting outer envelope, generating the 11-year solar cycle. Sunspots are concentrated bundles of magnetic flux pushing through the photosphere. Solar flares are catastrophic releases when the wound-up field finally snaps and reconnects, sending bursts of plasma into space at thousands of kilometers per second. Earth’s outer core runs essentially the same physics at lower temperature and produces the geomagnetic field. Tokamaks confining fusion plasmas are entirely MHD stability problems. Get the math wrong and the plasma touches the wall and the experiment ends.

MHD turbulence is its own subspecialty because Alfvén waves – transverse oscillations of field plus matter – carry their own cascade alongside the ordinary hydrodynamic one. The two cascades couple, and the predicted spectral slope deviates from Kolmogorov’s minus five-thirds depending on which version of the theory you trust. The Iroshnikov-Kraichnan model predicts minus three-halves; Goldreich-Sridhar predicts a modified anisotropic minus five-thirds. Solar wind observations sit somewhere between. Plasma turbulence is the engine room of much of astrophysics, and it remains under active debate.

Computers Hit a Wall, AI Goes Around It

Direct numerical simulation of turbulence resolves every scale from the largest eddies to the Kolmogorov dissipation scale. No models, no fudge factors, just Navier-Stokes on a fine grid. The trouble is that the number of grid points grows roughly as the Reynolds number to the nine-fourths power, and the number of time steps as its three-fourths power. Total cost: about the Reynolds number cubed in the idealized estimate, closer to its fourth power in practice. Doubling the Reynolds number means sixteen times more computing. The current world record for a single homogeneous-isotropic-turbulence snapshot is a 32,768-cubed grid stored at the Johns Hopkins Turbulence Database, occupying half a petabyte for one instant in time. That snapshot reaches a Reynolds number of about 2,550, counted on the size of its middle eddies, the usual measure for such boxes – still astronomically far below a planetary atmosphere like Jupiter’s.

The practical workaround is Large Eddy Simulation, where you resolve the big eddies and use a model – called a sub-grid closure – to represent the effect of the small ones you cannot afford to compute. The closure is the central modeling choice and the central error source. Decades of careful tuning have produced workable closures for many applications, but no universal one.

Since 2022 something unexpected has been happening. Machine-learning weather models trained on decades of reanalysis data have started outperforming traditional partial-differential-equation solvers on certain forecast horizons. Google DeepMind’s GraphCast in late 2023 beat the European Centre for Medium-Range Weather Forecasts’ flagship deterministic model on more than 90% of verification targets. By 2026, ECMWF runs daily forecasts from its own AIFS model alongside GraphCast, Aurora from Microsoft, and Pangu-Weather from Huawei. Several of these models are now demonstrating skillful forecasts somewhat past the traditional two-week predictability limit. Whether this represents a genuine extension of the chaos-imposed horizon or just a better extraction of large-scale flow patterns is an open debate – and, tellingly, these data-driven models still tend to underpredict record-breaking extremes. In parallel, a string of 2024 and 2025 papers used Bayesian sparse regression and physics-informed neural networks to discover compact analytic closures for sub-grid turbulence. The age of hand-crafted closures may be ending.

The Million-Dollar Question

In 2000, the Clay Mathematics Institute listed seven outstanding mathematical problems and attached a million-dollar prize to each. One of those was the question of whether the three-dimensional incompressible Navier-Stokes equations always have smooth solutions for all time given smooth starting conditions, or whether the equation can spontaneously produce a singularity – a point where velocity goes infinite in finite time. The equation works brilliantly in practice. Aircraft are designed with it, hurricanes are forecast with it, blood flow in the brain is modeled with it. What we do not have is a proof that the equation will not blow up.

The problem is still open. Only one of the seven Clay problems has ever been settled: Grigori Perelman proved the Poincaré conjecture in the early 2000s, and then declined the prize money when it was offered in 2010. Claimed solutions to Navier-Stokes arrive regularly, lately including machine-assisted ones, and none has survived scrutiny. The Clay Institute requires two years of community vetting after peer-reviewed publication before it will consider a payout. The prize remains uncollected.

The cleanest result in the field over the past decade was Terence Tao’s 2014 construction of an averaged Navier-Stokes equation that has the same energy identity, same scaling, and same harmonic-analysis estimates as the real one – and that he then proved blows up in finite time. The implication is sobering. Any proof of global smoothness for the real Navier-Stokes equation cannot rely on energy arguments and scaling alone. It has to exploit some finer feature of the actual nonlinearity, some structural property the averaged equation lacks. This rules out a whole family of proof strategies. We do not know which family does work.

There is also a converse research line. Buckmaster, Vicol, and collaborators have used a technique called convex integration to construct multiple distinct “weak” solutions of the equation from the same initial data – technically violating uniqueness in the broader function classes where the equation is solvable. For the inviscid cousin of Navier-Stokes, the Euler equation, Onsager’s 1949 conjecture about energy dissipation at the threshold Hölder regularity exponent of one third was fully resolved by Isett in 2018. The work is technical, beautiful, and discouraging for anyone hoping the Navier-Stokes prize will fall soon.

Knowing the Law Is Not Solving It

i Jupiter’s atmosphere, turbulence at Reynolds numbers around 10¹¹ ×

Turbulence is where chaos, statistical mechanics, and emergence meet most violently. The deterministic Navier-Stokes equation produces unpredictable trajectories with a finite memory horizon, the central object of chaos theory. The cascade produces universal statistical laws that depend only on energy flux, the central premise of emergence: simple rules at the bottom, surprising large-scale order at the top, with the bridge between them too tangled to traverse analytically. The flux through scales is a non-equilibrium steady state, the cleanest example physics has of a system that constantly transfers a conserved quantity through its degrees of freedom rather than approaching detailed balance. The same equations describe the smoke curling off a blown-out candle, the wake behind a swimming bacterium, the Great Red Spot at Jupiter, and the convecting interior of the Sun.

The lasting lesson is that knowing the law is not the same as solving it. Newton wrote down a single line of mathematics that runs the orbits of all the planets, and within a generation we could predict eclipses thousands of years in advance. Navier-Stokes is the same depth of insight applied to fluids, and almost two centuries on, we still cannot predict where the next eddy will form in a glass of milk. The honest summary is that turbulence has resisted every approach humans have brought to bear on it, and the consolation is that universe runs on the answer regardless of whether we ever find it.