The singularity a layer makes for itself

When the flow lets go and the three essays after it are about where a boundary layer separates. This one is about what happens to the equations there, and the answer is that they stop having a solution.

The march, and what it starts from #

Howarth’s flow is the standard test: an outer stream decelerating linearly, $U_e = U_0(1 - \beta x)$, which is the simplest adverse pressure gradient there is.

The computation is an implicit finite-difference march of Prandtl’s equations — implicit in $y$, with the convective coefficients lagged in $x$ and refreshed a few times per station — on four hundred points across the layer with a step of $5\times10^{-6}$.

It starts from Blasius. The starting profile’s wall shear is 7.4042 against the closed form’s 7.4028, and its shape factor is 2.5953 against 2.59, which is the check that the machinery is solving the right equations before it is asked anything interesting.

Why a parabolic problem can fail at all #

Before the numbers, the structure, because “the solution cannot be continued” is a strange thing to say about a marching problem.

A parabolic problem marches, and marching is usually robust: each station is computed from the one before, and there is no eigenvalue problem, no matrix to be singular, nothing to fail. That is exactly what makes Prandtl’s equations attractive and what makes their failure surprising.

The mechanism is the convective coefficient. The momentum equation carries $u\,\partial u/\partial x$, so the streamwise derivative is multiplied by $u$ — and near the wall $u$ is small and getting smaller. As the shear falls the coefficient of the marching direction falls with it, and the problem loses its parabolic character in a thin region next to the wall.

That is why the failure appears at the wall rather than in the middle of the layer, and why it appears as a divergence in a wall quantity rather than as a blow-up everywhere. The equations are still parabolic in the bulk; they are degenerate in the sublayer, and the degeneracy is exactly co-located with the vanishing shear.

What happens #

The wall shear falls, its slope steepens, and at

$$x_s = 0.119831$$

it reaches zero. Howarth’s value is 0.1198, so the agreement is 0.026 per cent, and the shape factor at the last computed station is 3.82 — the value a separating laminar layer has.

Past that station there is nothing. Not a solution with reversed flow, not a solution that continues badly — the march cannot be continued, and the reason is not the scheme.

The profile sequence is worth reading for what it says about the mechanism.

Early on the layer is Blasius-like: full, with a healthy wall gradient. As the deceleration acts, the fluid nearest the wall — which has the least momentum and feels the same pressure rise as everything else — is slowed the most, and the profile develops an inflection and then a nearly vertical section at the wall.

That is the sequence how much uphill a layer can take quantifies from the integral side, and the gradient that does both reads from the other end: the same adverse gradient that separates the layer is the one that puts an inflection point in it and makes it unstable. Here both happen and neither is what the essay is about; what is about to happen is that the equations run out.

Goldstein’s result #

What Goldstein showed in 1948 is that the failure has a definite analytic form. Near the separation station,

$$\tau_w \propto (x_s - x)^{1/2}, \qquad \frac{d\delta^*}{dx} \propto (x_s - x)^{-1/2}.$$

So the wall shear reaches zero with an infinite slope, and the displacement thickness’s slope diverges. The solution cannot be continued because continuing it would require complex values.

Measured here, over a window one decade wide and one decade out from separation, the exponent is 0.601. That is not one half, and the discrepancy has a cause worth being honest about: the scheme is first order in $x$, and the measured exponent inherits that. At a step of $4\times10^{-5}$ it is 0.7375, at $10^{-5}$ it is 0.6354, and at $5\times10^{-6}$ it is 0.6011 — converging on a half, from above, and slowly.

The evidence that it is real #

The two behaviours together are what say the singularity belongs to the equations.

Over the same factor of eight in the step, the separation station moves by 0.108 per cent while the exponent moves by 0.14. A numerical artefact would do the opposite: its location would move with the grid and its apparent form would be whatever the truncation error had.

A real singularity being resolved slowly moves the exponent and not the station, and that is what the sweep shows.

The window dependence is worth a note because it is the honest face of the measurement. Very close to $x_s$ the last few stations are a step apart and the structure being resolved is smaller than a step, so the local exponent there is the discretisation’s and it rises towards one. One decade out it is 0.60 and falling with refinement, and further out still the asymptotic form has not taken over.

There is a window in which the measurement means something, it is a decade wide, and its position is stated on every figure that uses it. That is the ordinary situation for measuring an asymptotic exponent on a finite computation, and it is the same discipline the range a real Reynolds number does not have applies to a spectrum.

And it is not Howarth’s flow’s fault #

The last check is that the phenomenon is generic.

Halve the deceleration and separation moves twice as far downstream: over a factor of eight in $\beta$, the product $x_s\beta$ is constant to 0.25 per cent. The layer separates after the same fractional loss of free-stream speed whatever the rate, and the exponent does not move at all.

So the singularity is a property of Prandtl’s equations with a prescribed adverse pressure gradient, not of Howarth’s particular choice.

What the displacement thickness does #

The other half of Goldstein’s result is about the displacement thickness, and it is the half that explains the resolution.

$d\delta^*/dx \propto (x_s - x)^{-1/2}$: the layer’s effective thickness is growing with an infinite slope as separation is approached. The third thickness is the collection’s account of what the various thicknesses measure, and the displacement one is the one the outer flow sees.

So at the very station where the equations are failing, the quantity that would feed back to the outer flow is diverging. That is not a coincidence: the two are the same fact. The layer is trying to displace the outer flow at a rate that would change the pressure completely, and the prescribed-pressure formulation is refusing to let it.

The body the outer flow sees sets up the coupling and records that it “cannot reach a trailing edge” — which is this failure, met from the other side.

Whose fault it is, which is the boundary condition’s #

Here is the resolution, and it is the reason this essay sits next to the length the limit invents.

Prescribing the pressure is a limit: it is what the Reynolds number going to infinity does to the coupling between the layer and the outer flow, taken in the outer flow first. In that limit the outer flow is computed as though the body were solid, its pressure is handed down, and the layer has to live with it.

A real layer does not have to. As it thickens it displaces the outer flow, the outer flow’s pressure changes, and the change relieves the gradient the layer is struggling against. The one-way coupling forbids that relief, and the singularity is the layer being asked to survive a pressure rise that a real flow would not have imposed on it.

The triple deck restores the coupling over a region of length $Re^{-3/8}$, with the pressure as an unknown. In it, the flow goes through separation with no singularity anywhere.

Same equations. Different thing prescribed. Different answer.

The inverse method, which is the cheap version of the same insight #

There is a practical demonstration of the same point that needs no asymptotics.

Run the boundary-layer equations inverse: prescribe the displacement thickness $\delta^*(x)$ as a smooth function that continues past separation, and solve for the pressure gradient that produces it. The equations are the same; the unknown has changed.

Marched that way, the computation goes straight through separation. The wall shear passes through zero and becomes negative, a reversed-flow region appears, and nothing diverges. What was a singularity is a smooth zero crossing.

That is as direct a demonstration as could be asked for that the singularity is in the boundary condition. The interactive and semi-inverse methods used in industrial codes are all versions of it: let the pressure be an unknown, iterate, and the difficulty disappears.

A short history, because it took a long time #

The chronology is worth having because it shows how hard the diagnosis was.

Prandtl wrote the equations in 1904. Within a decade it was known that computations of decelerating flows broke down, and for forty years the breakdown was widely taken to be a numerical difficulty or a consequence of the approximations rather than a property of the equations.

Goldstein settled it in 1948 by finding the local form: a square-root approach, an infinite slope, and a proof that the solution cannot be continued. That established that the failure was real and did not explain it.

The explanation came in 1969, in three independent papers — Stewartson, Messiter and Neiland — with the triple deck. Twenty-one years between knowing the singularity existed and knowing what it was a consequence of, and the answer was a structure with three layers and a new power of the Reynolds number in it.

That is a long time for a question with a two-sentence answer, and the reason is that the answer is about a limit’s ordering rather than about anything in the flow. Nothing in a laboratory would have suggested it.

What “separation” then means #

It is worth revisiting the word in the light of all this, because it has been used loosely in three different senses.

Zero wall shear is the classical criterion and it is what when the flow lets go computes. In a steady two-dimensional flow it is the right one.

The Goldstein singularity is what the prescribed-pressure equations do at that station, and it is a property of the formulation rather than of the flow.

And the physical separation — the point where the streamline leaves the surface, where the layer’s scaling breaks down, and where the interaction region begins — happens over a distance $Re^{-3/8}$ around the same place, and asking for a single station for it is asking a question the flow does not answer sharply.

The three coincide to leading order and they are not the same statement, and this collection has used all three.

The bubble, which is what the resolution predicts #

The interaction region’s length is not just a repair to a theory; it predicts a structure everybody has seen.

A laminar boundary layer separating on a gently curved surface does not simply leave. It separates, forms a short region of reversed flow, and reattaches — a laminar separation bubble — and the bubble is short: a few per cent of a chord on an aerofoil at moderate Reynolds number.

Its length is set by the interaction region’s, which is $Re^{-3/8}$ of the body scale. That is a prediction of the triple deck rather than a fitted correlation, and it is one of the theory’s successes.

Where the straight line stops is the collection’s essay on the separation station itself, and this is what happens in the few per cent past it — a region whose existence the classical theory cannot express and whose length the corrected theory computes.

What a computation should actually do #

The practical advice falls out and it is worth stating for anybody running a boundary-layer code.

A direct march with a prescribed pressure is correct up to separation and cannot pass it. If a code reports a failure there it is behaving properly, and refining the grid is the wrong response.

The failure has a signature: the wall shear falling with a steepening slope, the displacement thickness rising with a steepening slope, and the iteration count at each station climbing. A code that instead diverges violently, or produces oscillations, or gives a separation station that moves with the grid, has a different problem.

And a computation that needs to go past separation must change what it prescribes. There is no version of the direct problem that works.

Why this is a limit essay #

The question these essays share is what a limit leaves behind, and this one leaves behind a singularity that belongs to the limit and not to the flow.

$Re \to \infty$ produces two things. It produces Prandtl’s equations, which are excellent, and it produces the decoupling — the statement that the outer flow can be computed first and the layer afterwards. The second is the part that fails, and it fails at exactly the place where the layer’s displacement stops being small compared with what the outer flow is doing.

The residue is a mathematical object with no physical referent: an infinite slope, a square-root approach, a station past which nothing exists, all of it produced by an ordering of two calculations that a real flow does not perform.

Ideal against real is the collection’s map of where the exact theory and the real flow part company, and this is the sharpest local instance of it — a place where the approximation does not merely become inaccurate but stops being defined.

Limits recorded rather than smoothed over #

The scheme is first order in $x$. That is why the measured exponent is 0.60 rather than 0.50, and the convergence towards a half is demonstrated rather than reached. A second-order scheme, or Keller’s box method, would get closer at the same cost; nothing here needed it, because the station is what the essay is about and the station is converged.

The inverse computation is described and not performed. What is computed here is the direct march and its failure. That the inverse formulation marches through is a standard result and is stated as such.

Only laminar flow. A turbulent layer separates too, at a shape factor nearer 2.4 than 4, and it does not have a Goldstein singularity in the same form — the turbulence models used introduce their own behaviour there, and the classical analysis does not apply.

And the exponent window matters. Very close to $x_s$ the measurement is the discretisation’s, because the last few stations are a step apart and the structure being measured is smaller than that. The window used is stated on the figure, and a different one gives a different number.

What the exponent is worth knowing for #

A last practical note on the one half, since the computation here reaches only 0.60.

The exponent matters because it decides how a numerical scheme fails. A quantity approaching zero as $(x_s - x)^{1/2}$ has a derivative going as $(x_s-x)^{-1/2}$, so a finite-difference march loses accuracy progressively over a region rather than hitting a wall — which is why the failure looks like slowly worsening convergence rather than a crash, and why it was so long mistaken for one.

It also decides how far back from separation a direct march can be trusted. With a square-root approach the wall shear is within a few per cent of its smooth behaviour until about one per cent of the separation distance from the station, so a computation stopped ten per cent short of separation is clean.

And it is the exponent the triple-deck theory has to match to. The interaction region’s inner solution must reduce to Goldstein’s form as it is left, which is one of the conditions that fixes the deck’s scalings — so the half is doing work in the theory that repairs it as well as in the theory that produces it.

The residue #

The limit is $Re \to \infty$, taken in the wrong order.

What survives it is a singularity: an infinite slope in a physical quantity, at a finite station, with an exponent of one half, in a solution of equations that describe a perfectly ordinary flow going round a perfectly ordinary body.

Nothing in the fluid does anything singular at $x = 0.1198$. The singularity is entirely the property of an approximation, and identifying it as such took thirty years and produced the triple deck.