One diaphragm, every wave

Take a long tube, put a diaphragm across the middle, fill one half with gas at ten atmospheres and the other with the same gas at one, and burst the diaphragm.

The initial condition has no waves in it — two uniform states at rest, with a discontinuity between them and nothing else. What comes out is every kind of wave the equations of compressible flow can carry, all three at once, moving in different directions at different speeds and obeying different laws.

Three waves, three kinds #

The Euler equations in one dimension have three characteristic families, with speeds $u - a$, $u$ and $u + a$, and the Riemann problem produces one wave of each.

The shock, running into the low-pressure gas. Everything jumps across it — pressure, density, temperature, velocity, entropy — and the jumps satisfy the Rankine–Hugoniot conditions rather than any assumption about how the discontinuity is made.

The expansion fan, running back into the high-pressure gas. Nothing jumps: the pressure falls smoothly through it, the entropy does not change at all, and it widens with time because its leading edge travels faster than its trailing edge. That spreading is the difference between a compression and an expansion, and is why an expansion cannot become a shock.

The contact surface, between them. This one is the odd member of the family and is often overlooked: the pressure and the velocity are continuous across it and only the density jumps. It is not a wave in the sense of carrying a signal; it is a material boundary, made of the two gases that were on either side of the diaphragm, drifting at the fluid’s own speed.

What is actually solved #

The remarkable thing about the Riemann problem is how little of it is numerical.

The two sides must agree in the middle: the velocity behind the shock, computed from the right-hand state, must equal the velocity behind the fan, computed from the left-hand state, and the pressure must be common. Each side gives a formula for the velocity as a function of the unknown middle pressure — a shock branch if that pressure is above the far one, an isentropic fan branch if below — and the two branches meet with matching value and slope where they cross over.

That gives one scalar equation in one unknown, and it is monotone. Everything else follows in closed form: the two middle densities, the shock speed, the fan’s head and tail, and the state at any similarity coordinate $x/t$.

The solve is a bisection rather than a Newton iteration, and deliberately so: a monotone bisection cannot converge to the wrong root, and this problem has exactly one root that matters.

What the solver computed, and how it was checked #

Each wave is checked against its own law, and the laws are different — which is what makes the check meaningful rather than circular.

Across the shock: mass, momentum and energy, in the shock’s own frame, with residuals of $3\times10^{-15}$, $2\times10^{-15}$ and $3\times10^{-16}$. None of those three was used to find the shock speed, which came from the pressure match.

Across the fan: the entropy $p/\rho^\gamma$ is unchanged to a part in $10^9$, and the Riemann invariant $u + 2a/(\gamma-1)$ is unchanged to the same precision. Neither was imposed.

Across the contact: the pressure and velocity must be continuous and the density must not be. The last of those is a check with teeth: a solution whose contact has been flattened away is refused, because a two-wave answer to a three-wave problem is precisely the failure a careless solver produces.

And the shock must be supersonic relative to the gas ahead of it — Mach 1.656 for the classic case — because a discontinuity that is not is not a shock.

Reading the solution as three questions #

There is a way to hold the structure that makes the algebra unnecessary, and it is how the problem is actually set up.

Ask three questions in order. What pressure do both sides agree on? — one root find. How fast is the material between them moving? — read off either branch, since they now agree. What is the density on each side of that material boundary? — different on each, because the two gases arrived at the common pressure by different routes, one through a shock and one through an isentropic expansion.

That last point is the whole reason a contact exists. A shock raises entropy and an expansion does not, so gas that reached 30 kPa by being shocked is at a different density from gas that reached 30 kPa by expanding — even though both are at the same pressure and moving at the same speed. The contact surface is the entropy difference made visible, and if both waves happened to be isentropic there would be no contact at all.

Why this problem is the standard #

Every compressible flow solver ever written is tested on this, and the reason is not tradition.

It has an exact answer. Not a series, not an asymptotic expansion — an answer, from one root find and some algebra, against which a numerical result can be compared point by point.

It contains every difficulty at once. A scheme must capture a discontinuity without oscillating, resolve a smooth expansion without smearing it, and hold a contact surface sharp without any mechanism forcing it to. Those three requirements pull in different directions, and a scheme tuned for one usually fails another.

It is self-similar. There is no length in the initial condition — two uniform states and a discontinuity have no scale between them — so the solution depends on $x/t$ alone, so a single computation at one time is a computation at all times, and a numerical result can be checked for whether it has the right similarity structure as well as the right values.

That absence of a length is the same property the counting argument would have predicted: with no length and no time in the problem, the only dimensionless combination available is $x/t$ over a speed, and the solution has to be a function of it. It is why one computation covers all times, and why a numerical scheme can be tested for whether it has the similarity structure as well as the values.

The contact is the usual failure. A first-order scheme smears it over more and more cells as time goes on, because there is no physical mechanism keeping it sharp — unlike a shock, which is held together by the nonlinearity that made it — and a solver that reports a beautifully sharp shock and a contact spread over forty cells has passed the eye test and failed the problem.

The strength that cannot be bought #

A shock tube’s shock is driven by the pressure ratio across the diaphragm, and the return on that ratio is poor.

Two ceilings are visible in that figure and they have different causes.

The density ratio saturates at $(\gamma+1)/(\gamma-1)$, which is 6 for air, because the Rankine–Hugoniot relations say so however strong the shock. No pressure ratio whatever can compress a perfect gas by more than six times in a single shock, which is why re-entry heating is a temperature problem rather than a density one.

The shock Mach number climbs only logarithmically, because the driver gas has to expand through its own fan to push, and its expansion is limited by its own speed of sound. That is why real shock tubes use a light driver gas — helium or hydrogen, whose speed of sound is three or four times air’s — rather than merely a higher pressure. The escape velocity of the driver is what matters, not its pressure.

The tube as an instrument #

The exact solution is also a design tool, because everything an experimenter needs is in it.

The test time is the gap between the shock arriving at a station and the contact surface arriving there, which the two computed speeds give directly: at ten to one, the shock runs at 554 m/s and the contact at 293, so a station one metre from the diaphragm has about 1.6 milliseconds of steady flow behind the shock before the driver gas arrives and spoils it.

The conditions in that window come from the same solution: a known pressure, a known density, a known temperature and a known velocity, all uniform, all computed from the initial states. That is what makes a shock tube useful — for a millisecond it produces flow conditions no steady facility can sustain, and it produces them predictably.

And the reflected wave extends it. A shock arriving at the closed end of the tube reflects, and the gas behind the reflection is brought back to rest at a much higher temperature — which is the standard configuration for chemical kinetics measurements, where what is wanted is not a flow but a few hundred microseconds of very hot stationary gas.

The second Riemann problem, and the condition that removes it #

The reflected shock mentioned above raises a question the first solution does not answer, and answering it turns out to decide the design of every hypersonic shock tunnel ever built.

The reflection runs back up the tube through gas it has just brought to rest, and after a while it reaches the contact surface — the boundary between the driven gas and the driver. What happens there is another Riemann problem, with a shock arriving at a discontinuity in density instead of at a wall, and the same matching argument decides it: the pressure and velocity behind the transmitted wave must agree with the pressure and velocity behind whatever runs back the other way.

Generically something does run back the other way. If the driver gas is acoustically softer than the gas the shock came from, the interaction sends an expansion back towards the end wall; if it is stiffer, a shock. Either one arrives at the stagnant hot test gas a few hundred microseconds later and changes its pressure, which ends the useful measurement — so the test time is not set by the tube’s length but by how long it takes a wave nobody wanted to come back.

And there is one choice for which nothing comes back at all. Arrange the two gases so that the reflected shock passes through the contact and continues on its way, with the states on the two sides already agreeing, and there is no reflected wave to return — the interaction is transparent. That is the tailored-interface condition, identified in the late 1950s, and it is a single equation relating the driver gas’s properties to the driven gas’s and to the shock strength.

Meeting it is why a driver is chosen rather than merely made light. The advice above — helium or hydrogen, because the driver’s own expansion is what limits the shock — is right about the incident wave and incomplete about the facility. A driver that is too soft under-tailors and returns an expansion; too stiff and it over-tailors and returns a shock; so real tunnels run mixtures, or heat the driver, to sit on the condition rather than at the end of the range. The reward is large: a tailored tunnel holds its reservoir conditions for milliseconds where an untailored one holds them for a fraction of one.

It is a pleasing place for the essay’s own machinery to end up. The whole solution above rests on finding the one pressure at which two branches agree; the tailoring condition is that same agreement demanded a second time, at a surface which the first solution created and which exists only because a shock and an expansion reach the same pressure by different routes. The contact surface is made by an entropy difference, and tailoring is the arrangement in which that difference costs nothing.

What the picture cannot show #

No viscosity, no conduction, no thickness. The shock here is a mathematical discontinuity. A real one is a few mean free paths thick, and the contact surface diffuses into a mixing layer that grows as $\sqrt{\nu t}$ — which in a real tube is millimetres by the time the wave reaches the end.

Nothing here is one-dimensional in practice. The boundary layer on the tube wall is the growing kind rather than anything held in check, and it eventually swallows the uniform core behind the shock. A real diaphragm does not vanish; it ruptures, in pieces, over microseconds, and the wave takes some distance to become planar. The boundary layer on the tube wall grows behind the shock and eventually swallows the uniform core, which sets a limit on how long the test time lasts.

And the gas is calorically perfect. At the temperatures behind a strong shock, air’s specific heats change, its molecules dissociate, and $\gamma$ falls — which is precisely the regime where Newton’s impact theory becomes exact. The solution here would need real-gas thermodynamics to describe it, and would then no longer be closed-form.

The entropy is what forbids the mirror image #

One question the solution raises is why the shock runs into the low-pressure gas and the fan into the high — why not the other way round, with an expansion shock and a compression fan?

The equations permit it. A discontinuity from low pressure to high, running the other way, satisfies mass, momentum and energy across itself perfectly; it is the expansion shock, and this site has computed its residuals at exactly zero. What refuses it is the second law: it lowers entropy, and nothing else in the problem objects.

So the direction of every wave in the tube is fixed by thermodynamics rather than by mechanics, and the asymmetry of the answer — one side gets a shock, the other a fan — is that fact made geometric. A tube whose diaphragm separates two states is a small machine for demonstrating that the conservation laws have one more solution than the world does.

Where the model stops #

The Riemann problem is exact for the Euler equations, which are the compressible equations with viscosity and conduction deleted. That deletion is the whole reason the discontinuity is allowed: a genuine solution of the full equations has no discontinuities at all, only very steep gradients, and the shock is the limit of those gradients as the viscosity goes to zero.

The solution also assumes the two gases are the same, or at least that they have the same $\gamma$. Two different gases give a Riemann problem with a jump in $\gamma$ across the contact — solvable by the same method, and the basis of shock-tube experiments in which one gas drives another.

And it is a one-dimensional problem. In two or three dimensions the Riemann problem does not have a closed-form solution, which is why the modern solvers that use it — Godunov’s method and its descendants — apply the one-dimensional answer cell by cell, at every interface, along the normal direction. That is the sense in which this exact solution is the workhorse of computational compressible flow: it is not solved once, it is solved millions of times per timestep.

Who found it, and when #

Riemann formulated and solved the problem in 1860, in the paper that also introduced the idea that a compression wave must steepen into a discontinuity. He got the discontinuity right and the thermodynamics wrong — he assumed the flow through it was isentropic — and Rankine in 1870 and Hugoniot in 1887 fixed it by insisting on energy rather than entropy across the jump.

The shock tube as an instrument is Vieille’s, in 1899, and it became the standard tool of high-speed gas dynamics in the 1940s and 1950s: a way of producing, in a metre of steel tubing, flow conditions that no wind tunnel could sustain, for a millisecond at a time.

Godunov’s 1959 method put Riemann’s solution at the heart of numerical gas dynamics by using it as the building block of a scheme, and Sod’s 1978 comparison of methods on this exact problem is why the ten-to-one case is called Sod’s problem and why it appears in every paper that proposes a new scheme.

Where the ladder goes next #

This tube was choked by nothing: the waves ran freely and the only thing limiting them was the driver’s own expansion. Two other things can stop a compressible flow, and both of them stop it at exactly the same place — friction and heat, which are different physics with the same destination.