When the flow lets go
The rear half of any body is a problem. The flow has accelerated round the widest point, so it is moving fast and its pressure is low; by the time it reaches the back it has to have slowed down again and climbed back up to roughly the pressure it started with.
Air with plenty of momentum manages that. Air that has been rubbing against a wall for the whole length of the body sometimes does not.
When it does not, the flow separates — leaves the surface entirely — and almost everything interesting about real fluid mechanics follows from that one event.
The adverse gradient
The mechanism has a name and it is worth being precise about it.
Over the front of a body the flow speeds up and the pressure falls. That is a favourable pressure gradient: it pushes the fluid along, and the boundary layer stays thin and healthy.
Past the widest point the flow must slow down again, so the pressure rises. That is an adverse gradient, and it pushes backwards against fluid that is already moving forwards. Out in the free stream there is momentum to spare. Inside the boundary layer, next to the wall, there is not — that fluid has spent the whole journey losing energy to friction.
So the adverse gradient acts hardest on exactly the fluid least able to resist it. The velocity profile near the wall flattens, then develops a point of inflection, then reverses.
The instant the velocity gradient at the wall falls to zero, the flow has separated.
What separation costs
Two things happen at once, and both are expensive.
The pressure stops recovering. Downstream of the separation point the body sits in a broad region of slow, roughly constant, low-pressure fluid. The high pressure on the front is no longer balanced by high pressure on the back. That imbalance is pressure drag, and for a bluff body it is most of the total.
The effective shape changes. The outer flow no longer sees the body; it sees the body plus its wake, which is a much fatter and blunter object. So the pressure distribution over the entire front adjusts too.
This is the resolution of d’Alembert’s paradox in one paragraph. Ideal flow predicts zero drag because it assumes the pressure recovers perfectly. Separation is the flow refusing to recover it.
Where it happens, and when
The separation point moves, and what moves it is the Reynolds number.
At very low Reynolds number there is no separation at all — viscosity is strong enough to keep everything attached, and the flow is nearly symmetric front to back. As the Reynolds number rises the separation point moves forward from the rear, a recirculating bubble appears and grows, and by Reynolds number a few hundred the wake is wider than the body.
The counter-intuitive fix
Here is the part that seems backwards and is the most useful practical fact in the subject.
A turbulent boundary layer separates later than a laminar one.
Turbulence mixes. A turbulent layer is constantly carrying fast-moving fluid from the outer part of the layer down towards the wall, replenishing the momentum that friction has removed. So the near-wall fluid arrives at the adverse gradient with more to spend, and can climb further before giving up.
The cost is friction: a turbulent layer has a much steeper velocity gradient at the wall and therefore much more skin friction. For a streamlined body, where friction dominates, that is a bad trade. For a bluff body, where pressure drag dominates, it is an excellent one — delaying separation narrows the wake, and the pressure drag saved is far larger than the friction added.
Which is why a golf ball has dimples. The dimples trip the boundary layer into turbulence deliberately. The ball has more skin friction and a much narrower wake, and the net drag is roughly halved. A smooth golf ball would go about half as far.
The velocity profile, and the moment it turns over
The clearest way to see separation coming is to watch the shape of the velocity profile across the boundary layer as it moves along the surface.
Under a favourable gradient the profile is full: velocity rises steeply from zero at the wall and flattens quickly into the free stream. There is plenty of momentum near the surface.
Under an adverse gradient the profile thins out near the wall. The fluid closest to the surface is being decelerated hardest, so the steep initial rise slackens. A point of inflection appears part-way up — a sure sign of trouble, because an inflected profile is also hydrodynamically unstable.
Continue further and the gradient right at the wall reaches zero. That is the separation point, defined precisely: at .
Past it the near-wall velocity is negative — fluid is running forwards into the body — and the outer flow has left the surface entirely.
Every one of those stages is happening in a film thinner than the pencil line that would draw it, which is why separation is invisible in a photograph of a flow and obvious in a plot of the profile.
Why the wake is low-pressure
One step in the argument is often skipped, and it is the step that turns separation into drag.
The separated region behind a body is not still. It is a slow, recirculating mass of fluid, and it is connected to the outer flow along a shear layer. What sets its pressure is that outer flow: the separated region cannot support a pressure gradient of its own, so it takes roughly whatever pressure existed at the separation point.
At the separation point the outer flow is still moving fast. So the pressure there is low — and that low pressure is then imposed over the whole rear face of the body.
The body therefore has near-stagnation pressure pushing on its front and separation-point pressure — much lower — over its back. The difference, times the frontal area, is the pressure drag. Separating early means separating where the flow is faster, which means a lower wake pressure and more drag, which is precisely why delaying separation is worth so much.
The same mechanism is stall
A wing at increasing incidence is asking for an increasingly severe pressure recovery over its upper surface. The suction peak near the leading edge gets sharper, and the climb back to trailing-edge pressure gets steeper.
At some incidence the boundary layer cannot make the climb. It separates, and the separation point runs forward over the upper surface. The flow no longer leaves the trailing edge smoothly, so the Kutta condition has nothing to enforce, the circulation collapses, and the lift goes with it.
That is stall, and it explains several things at once: why it happens at an angle rather than a speed; why it is sudden for some sections and gentle for others, depending on whether separation starts at the leading edge or the trailing edge and how fast it spreads; and why the ideal theory’s lift curve goes on rising for ever, since it has no boundary layer to lose.
Bluff and streamlined, in numbers
The practical size of the effect is worth having, because “separation causes drag” is qualitative and the magnitude is what surprises.
A circular cylinder in a typical flow has a drag coefficient around 1.2. A streamlined strut of the same thickness — the same frontal area, but with a long tapering tail — comes in near 0.06.
Twenty times less drag, from the same width. The strut has more surface area, and therefore more skin friction; it wins because it has almost no pressure drag. Its tail brings the pressure back up gently enough that the boundary layer can follow it nearly to the trailing edge, the wake is thin, and the fore-and-aft cancellation that ideal flow predicts is very nearly restored.
That reframes what streamlining is doing. It is not making a body slippery. It is arranging conditions under which the inviscid theory’s beautiful, wrong answer becomes nearly right.
A flat plate held edge-on has a drag coefficient near 0.001; held face-on, near 1.2. Same plate, same material, same surface finish. The factor of a thousand is entirely a question of where the flow separates.
What a designer can actually do about it
Four levers, and they are all attempts to give the near-wall fluid more momentum or ask less of it.
Shape the pressure recovery. Make the rear taper long and gentle. This is the cheapest and most effective, and it is why streamlined bodies look as they do.
Trip the layer turbulent. Accept more friction to get later separation. Golf ball dimples, turbulator strips on glider wings, the roughened leading edges of some racing bicycle wheels.
Add momentum directly. Vortex generators — small vanes that stir high-energy fluid down towards the surface — are visible on the upper wing surfaces of most airliners, usually just ahead of the control surfaces they are protecting.
Remove the tired fluid. Boundary layer suction, drawing the slowest air away through a porous surface. Effective, and expensive enough that it stays largely experimental.
Each is a response to the same diagnosis, and the diagnosis is always the pressure gradient rather than the friction.
What the solver computed
The recirculation in these figures is measured, not drawn.
lib/flow.js steps the incompressible equations in vorticity–streamfunction form with no-slip imposed
at the body. After the flow has settled, the streamwise velocity is sampled along the centreline
behind the cylinder and the extent of negative velocity — genuine reverse flow — is recorded.
At Reynolds number 1 and 10 there is none, and the caption says attached. At 40 and 100 there is, and the length is quoted. The site does not decide in advance which regime it is in; it measures.
Two limitations are stated rather than hidden. The computed bubble at Reynolds number 40 is about 0.85 diameters, against a laboratory value near 2.2 — the build-time grid is too coarse to resolve it properly, and the qualitative behaviour is right while the number is not. And above Reynolds number about 47 a real wake becomes unsteady and sheds a vortex street; this solver settles to a steady asymmetric state instead, so no figure here claims a street.
Reattachment, and bubbles that close
Separation is not always permanent, which matters for aircraft more than for cylinders.
A layer can separate, become turbulent in the shear region just outside the separated zone, and then — now carrying much more momentum — reattach to the surface further downstream. What is left is a closed separation bubble rather than an open wake.
Short bubbles of this kind sit on the upper surfaces of many aerofoils at moderate incidence and cost relatively little. The trouble is that they can burst: at slightly higher incidence the bubble fails to reattach, and the aerofoil goes from nearly normal lift to fully stalled with very little warning.
That is the leading-edge stall, and it is the reason some sections have a sharp, unforgiving stall and others give plenty of buffet first.
Separation is not always a fault
It is easy to read all this as a catalogue of failure, and some of the most useful devices in engineering separate the flow deliberately.
Spoilers on an airliner’s wing are exactly that: panels raised to force separation over part of the upper surface, destroying lift on demand during descent and after touchdown.
Bluff-body flow meters put a deliberately unstreamlined obstacle in a pipe and count the frequency of the vortices shed behind it, which is proportional to flow rate.
Vortex generators work by creating small controlled separations that energise the layer downstream and prevent a larger uncontrolled one.
And a sail or a delta wing at high incidence is operating with a large, stable separated vortex over its upper surface which is generating most of the lift. Separated flow is not automatically useless; it is uncontrolled separated flow that is.
Where the model stops
The grid is coarse. Separation is resolved; the internal structure of the layer is not.
No transition model. Whether a layer is laminar or turbulent at a given point is decided by history, roughness and disturbance level, none of which are computed here. The turbulent-separation argument above is stated, not simulated.
No vortex shedding. The unsteady wake is beyond this solver at present.
Two-dimensional. Real separation over a wing is strongly three-dimensional, and stall usually begins in a patch rather than along a line.
Who found it, and when
Prandtl’s 1904 paper contains the separation idea alongside the boundary layer itself, and includes photographs from a small water tunnel he built to show it.
The engineering consequences arrived over the following decades: Ludwig Prandtl and his students on aerofoil stall, and a long line of experimental work on bluff bodies. The golf ball, characteristically, was ahead of the theory — players had noticed in the nineteenth century that scuffed balls flew further than new smooth ones, and manufacturers were dimpling them decades before anybody could say why.
The one number that predicts it
There is no simple criterion for where a boundary layer separates, and the absence is worth noting because it is the kind of thing a reader expects a mature subject to have.
Separation depends on the pressure gradient’s history along the surface — not its value at a point, but the whole record of what the layer has been through since the leading edge. Two bodies with the same local gradient at a given station can separate at quite different places because one of them made its layer work harder earlier.
There are approximate criteria — Thwaites’ method for laminar layers, various integral methods for turbulent ones — and they are useful and not exact. Beyond them the honest answer is to solve the layer, which is what this site’s solver does, coarsely.
So separation is predictable in the sense that a computation will find it, and not in the sense that a formula will. That is a fair description of most of viscous fluid mechanics, and it is the reason the subject leans so heavily on dimensionless regimes — knowing which world a flow is in is often the most that can be said in advance.
The ladder from here
Nearby: the velocity profile through a separating layer, drawn; laminar against turbulent separation compared quantitatively; the drag crisis, where a sphere’s drag falls as it goes faster; and separation control — vortex generators, suction, blowing.
Then across to the thin layer that all of this happens in, and to the Reynolds number that decides which regime a body is in.