A third of what the theory promised

Euler’s column is the best-behaved thing in structural mechanics. Its critical load is exact, a column built with an initial bow of a thousandth of its length reaches 95% of it, and a century of testing has confirmed the formula to within the scatter of the steel.

A cylindrical shell in axial compression has a critical stress that is just as exact, just as easy to derive, and it is wrong by a factor of three or more on every specimen ever tested. The theory has not failed. What differs is something neither critical load calculation contains.

The critical load is the same number in all three. What decides the strength of the real structure is the slope of the equilibrium path immediately past that load, and that slope is invisible to every eigenvalue calculation ever performed.

Three shapes, and only one of them is forgiving #

The three cases have names and they are worth attaching to objects.

Stable symmetric. The path rises after the critical load, so the structure can carry more than its critical load once it has bent — it just cannot do so without deflecting. A pin-ended column is this case, and so is a plate, which is why a plate’s post-buckling strength is a design resource rather than a hazard: a slender plate carries several times its critical load by shedding stress to its edges.

Unstable symmetric. The path falls after the critical load and does so symmetrically, so it does not matter which way the structure goes. A cylindrical shell in axial compression is the standard example, and a spherical shell under pressure is worse.

Asymmetric. The path falls one way and climbs the other. A portal frame buckling in sway is the structural example — the same sway that second-order analysis chases — and the consequence is peculiar: the strength depends on the sign of the imperfection, so two nominally identical frames built to opposite tolerances have different capacities.

The stable case is worth stating on its own, because it is the one the whole subject was built on. A column that was never straight deflects from the first increment of load and heads for a critical load it never reaches; there is no maximum on its curve at all. It does not fail by buckling — it fails when its stresses reach yield, and the critical load is an asymptote rather than an event. That is why a bow of a thousandth costs a column almost nothing, and it is exactly what the two falling cases do not do.

The cleanest way to see how much of that is the path rather than the imperfection is to redraw all three at a twentieth of the crookedness and read the peaks again.

Two things are worth taking from the pair. The first is that even at a thousandth of a radian the asymmetric system is already down five per cent, so the sensitivity is not something that appears only in badly built structures. The second is the ratio: for that system, cutting the imperfection by twenty recovered eighteen points, and the next factor of twenty would recover under four — the deficit falls as roughly the square root of the imperfection, so a twentieth of it leaves 23 per cent of the deficit rather than 5. The return on straightness is falling from the very first increment, which is the practical face of the exponent the next section measures.

The exponent is the whole subject #

Take the unstable case and vary the imperfection. The maximum load falls, and the interesting question is how.

Those two numbers are two thirds and one half, and they are Koiter’s, from a doctoral thesis written in occupied Holland in 1945 and unread outside it for fifteen years. What is worth noticing here is how they were obtained: by computing the maximum of an exactly solved equilibrium path at twenty-four imperfection amplitudes and fitting a straight line through the logarithms. The exponents are measured on this site rather than quoted, which is the only arrangement in which agreement with the theory is evidence of anything.

The practical content of an exponent below one is the infinite slope at the origin. For the unstable case,

$$1 - \lambda_{\max} \approx 1.45\,\varepsilon^{2/3}$$

so an imperfection of a ten-thousandth costs 0.3%, a thousandth costs 1.5%, and a hundredth costs 7%. Each factor of ten in the imperfection costs only a factor of 4.6 in the deficit, which is why the last increment of straightness is the one that buys nothing and why tolerance specifications on shells reach a point of diminishing returns quickly.

Why a cylinder is the worst case there is #

The shell is worse than the model above, and the reason is a genuinely different phenomenon: a cylinder has an enormous number of buckling modes at almost exactly the same critical load.

For an axially compressed cylinder, every combination of axial half-waves and circumferential waves satisfying one relation gives the same critical stress. Dozens of modes are coincident, they interact, and the interaction of several unstable-symmetric modes is far more sensitive than any of them alone. That is why measured strengths cluster at a fifth to a third of theory rather than at the 70% the single-mode calculation would give.

The shape of that curve is the part that carries meaning, and it is the same statement as the exponent. The imperfection that matters is measured against the thickness — a dent of half a millimetre is nothing on a 20 mm plate and is a sixth of a 3 mm one — so a thinner shell is a rougher shell in the only units that count, and the knockdown falls accordingly.

The claim that measured strengths cluster at a fifth of theory is not rhetorical, and the same curve read further along says where the fifth is.

The curve has another end, and reading it is what stops the knockdown factor being treated as a property of shells rather than of thin ones. Walk back along the same line to a shell an inspector would call heavy — 10 mm on a metre radius — and the fraction recovers most of the way.

A design rule that is an empirical envelope under sixty years of test scatter is not a theory that failed. It is an admission that the quantity governing the strength is the geometry of the real shell, which is not on the drawing, and that the cheapest way to bound it is to measure a great many shells.

Where the sensitivity comes from #

It is worth being precise about the mechanism, because “imperfections reduce strength” is true of everything and explains nothing.

In a stable-symmetric system the imperfect structure deflects early and then finds more stiffness as it goes, so the deflection stabilises and the load keeps climbing. In an unstable one the imperfect structure deflects early and finds less stiffness, so the deflection accelerates and the load reaches a maximum while the structure is still moving in the direction it was pushed.

The second is the same event as a snap-through: a maximum on an equilibrium path, at which the structure has nowhere nearby to go.

So the family is one family. A limit point, an unstable bifurcation and an imperfection-sensitive shell are three descriptions of a structure whose stiffness is falling faster than its load can be taken away, and the only question in each case is what put it on the falling branch.

The theory that does work, and what it is for #

None of this makes the classical calculation useless, and it is worth stating what it is good for.

The column curve is what imperfection sensitivity looks like in the stable case, and the comparison is worth making explicitly. It is the envelope of squashing and Euler buckling with the real columns falling below both, and the whole of the departure is the rounding of the corner where the two failures compete — a few tens of per cent at worst, and negligible at either extreme. The shell’s curve is not a rounded corner. It is a factor of three, at every slenderness at which buckling governs at all, and it does not go away as the shell gets thinner. Same phenomenon, same word, two orders of consequence apart.

The critical load fixes the scale of the answer. A knockdown factor is a fraction of it, so an error in the classical stress — in the effective length, say, which the end conditions decide — is an error in the design stress in exact proportion, and the empirical curve is only usable because the theory it multiplies is exact. An empirical factor applied to an exact theory is a different thing from an empirical formula, and the difference is that the first generalises to a shell nobody has tested.

The critical load also identifies the mode, which decides where a stiffener has to go. And it tells the designer which regime the structure is in: below the crossover the shell squashes, above it the shell buckles and the knockdown applies. A calculation that is wrong by a factor of three about the strength can still be exactly right about which failure is coming.

Measuring the imperfection instead of assuming it #

The modern alternative to the empirical envelope is to stop treating the imperfection as unknown. Measure the actual geometry of the shell, put it into a nonlinear analysis, and compute the maximum of its own equilibrium path.

That is now practical and it changes the character of the problem. The shell’s strength becomes a computed quantity with a measured input rather than a fraction of an ideal, and the sensitivity curve becomes a way of deciding how accurately the geometry needs to be surveyed. It also produces an uncomfortable finding — one this collection meets whenever it looks at what a section can absorb before it fails — which is that the shape of the imperfection matters as much as its size: an imperfection resembling the buckling mode is far more damaging than a larger one that does not, so a shell with a 5 mm dent in the wrong pattern can be weaker than one with a 15 mm dent in the right place.

The Southwell plot is the older version of the same instinct — extracting a critical load from measurements taken well below it.

The same shape of argument, in another field #

This site has met a structure whose strength is decided by the largest flaw in it once already, and the parallel is close enough to be worth drawing precisely.

A brittle plate’s strength is not a material property in the way a yield stress is. It is $K_c/(Y\sqrt{\pi a})$, with $a$ the size of the worst crack, so the strength of the object is set by a defect nobody put on the drawing and the theory that predicts it is exact about a plate whose flaw size is known. Test a hundred nominally identical specimens and the results scatter over a factor of two, because the flaw sizes do.

The parallel holds in four places at once. Both are governed by a defect rather than by the nominal geometry. Both have a sublinear dependence on the defect size — a square root there, a two-thirds power here — so the first small defect costs disproportionately. Both produce large scatter in tests of identical specimens. And in both the engineering response has been the same: bound the defect by specification and inspection, then design to a lower envelope of the data.

Where they differ is instructive. A crack is local and can be found by looking; an imperfection in a shell is distributed, and the damaging ones are the components of it that resemble a buckling mode, which no inspection procedure is naturally shaped to find. And a crack grows under repeated load, so fatigue makes the defect worse with time, whereas a shell’s imperfections are fixed the day it is rolled — which is the one respect in which the stability problem is the easier of the two.

The general statement they share is worth having in plain words. Some structures are governed by their nominal properties and some by their departures from them, and no calculation of the first kind will ever indicate which sort a structure is. What indicates it is the shape of the path past the limit, and finding that costs an analysis nobody does unless they have been told to.

Pressure stabilises the shell, and emptying it is the dangerous act #

There is a design lever on imperfection sensitivity that costs nothing and is usually already present, and it is the reason the failures happen when they do.

An internal pressure stretches a cylinder circumferentially. A stretched membrane resists being pushed out of plane — the same reason a drum head is stiffer than a slack one — so the inward lobes an axial buckle needs are opposed by the hoop tension, and the imperfections that would have grown into them are held flat.

The effect is large. A cylinder with a knockdown factor of 0.2 when empty can recover much of its classical capacity under a modest internal pressure, and the recovery is steepest at low pressures — the first increment is worth the most, exactly as the first imperfection cost the most.

Which inverts the intuitive reading of when a tank is at risk. A full pressurised vessel is carrying its contents’ weight and is stabilised against buckling by the very pressure that is stressing its walls. An empty one is unstressed, unstabilised and at its most imperfection-sensitive, and it is carrying its own weight, its roof, any snow and any wind. Silo and tank failures cluster around emptying, and the mechanism is on this page rather than in the loading.

The same shell has two instabilities and only one of them is savage #

The factor of three is a statement about axial compression, and it does not generalise even to the shell it was measured on.

Put the same cylinder under external pressure — a vacuum in a tank, a pipeline below water, a silo being emptied too fast through a sealed roof — and it buckles circumferentially, into a small number of long lobes running the length of the shell. The classical load is

$$p_{cr} \approx \frac{2E}{1-\nu^2}\left(\frac{t}{R}\right)^3$$

and the knockdown factor for it is around 0.75 rather than 0.2. The same shell, the same imperfections, and a sensitivity three or four times milder.

The reason is the one the mode-spacing argument gives. Axial compression of a cylinder has dozens of modes crowded within a few per cent of each other — different axial half-wavelengths and different circumferential wave numbers all reaching criticality together — so any imperfection whatever has a large component near a critical mode. External pressure has a handful of well-separated modes: two lobes, three, four, at loads far apart. There is nothing to cluster, and a shell with well-separated modes behaves like a column.

So “shells are imperfection-sensitive” is too coarse a statement to design with. The sensitivity belongs to a mode, not to a shell, and the same vessel can be treated with confidence in one loading and with a knockdown factor of five in another. Which of the two a particular structure is in is decided by how it is being loaded on the day — and, for a tank, that changes with the valve.

The practical residue is a short list of questions to ask of any thin shell, in order. What is the mode? How close together are its neighbours? Is anything present that separates them — a pressure, a stiffener, a ring, a filling? And is that thing present in every condition the shell will be in, including the ones nobody drew: empty, part-built, being cleaned, with a vent blocked. The knockdown factor is a number about the worst of those, and the worst is usually a state the structure spends very little time in.

What the picture cannot show #

The three models are rigid bars on springs. They reproduce the exponents and the shapes exactly, which is what they are for, and they contain no material, no stress and no real geometry. A real shell’s behaviour is a superposition of interacting modes and the single-degree-of-freedom picture is a caricature — an accurate one about the property being illustrated, and a caricature.

The knockdown curve is a lower bound and not a prediction. It is drawn under the data, so a real shell will usually be stronger than it says, by an amount nobody can compute in advance. Designing to a lower bound means accepting that the material is being used at an efficiency that is unknown and probably poor.

The imperfection is a single number. A real structure’s departure from its drawing is a field — a function over the whole surface — and reducing it to one amplitude is the step that makes the arithmetic possible and the prediction approximate. Two shells with the same measured maximum deviation can differ by 40% in strength, and everything above would report them identically.

Nothing here treats plasticity. The stockiest shell drawn above, at R/t = 100, has a classical stress of 1,271 MPa and therefore fails by yielding long before it buckles; the transition between the two regimes is a region in which both matter — the same interaction the column curve rounds off and for the same reason.

Where the ladder goes #

The first rung is the interaction of modes, which is where the shell’s factor of three actually comes from and which needs more than one degree of freedom to show. Two coincident unstable modes produce a sensitivity exponent lower than either alone, and a structure optimised so that two modes coincide — which is what optimisation naturally does — has been optimised into the most sensitive configuration available.

The second is the practical business of imperfection tolerances: how much out-of-roundness a silo may have, how it is measured, and what the specification is worth given that the shape matters more than the size.

The third is the connection this essay has been circling. A structure’s sensitivity to its own construction is a computable property, and it is not the factor of safety, not the stiffness and not the redundancy. It belongs beside them as a thing a designer should know about a structure, and it is the one that decides whether the drawing describes what was built. It also decides how much an inspection is worth, which is the question every combined-failure check eventually reduces to: which of the two things being combined is actually known.