Dimensional analysis and similarity

Full-scale testing is expensive, slow, and sometimes simply impossible — nobody builds a spare bridge to find out whether it will vibrate itself apart. What makes a scaled model useful instead of merely decorative is not intuition about what "roughly the same" ought to mean. It is a small piece of mathematics that tells you exactly which numbers have to match between the model and the real thing, and exactly which are free to differ. That mathematics starts from a rule so basic it barely feels like one.

Dimensional homogeneity

Every term you add, subtract, or set equal to another in a physically meaningful equation has to carry the same dimensions. Not the same units — metres and feet both measure length, and converting between them is bookkeeping — but the same combination of mass, length and time. Take Bernoulli's equation, written as a head:

A pressure, a kinetic energy density and a potential energy density look like three unrelated quantities until you break each one down. All three reduce to \(ML^{-1}T^{-2}\), and that is not a coincidence to be waved past — it is the precondition for the equation being allowed to exist in this form at all. If a derivation ever produces a sum where the terms do not match, the derivation has an error in it somewhere, and this is the cheapest way there is to catch it: check dimensions before you check arithmetic.

The rule cuts both ways. It disqualifies wrong equations immediately, and it also disqualifies certain forms from ever being considered in the first place — nobody writes \(F = V + D\) for a drag force, because a velocity and a length cannot be added no matter what coefficient you attach to fix it up. Every physically sound equation can also be divided through by one of its own terms to become dimensionless, since dividing a quantity by another of the same dimension leaves a pure number. That single observation — homogeneity guarantees an equation can be made dimensionless — is the seed the next section grows into a method.

It is worth being precise about what is and is not being claimed. Homogeneity does not tell you the constant of proportionality in an equation, or even the exact functional form. A dimensionally consistent guess can still be wrong. What it gives is a filter: any proposed relationship that fails the check is certainly wrong, and one that passes has at least earned the right to be checked against data.

The Buckingham Pi theorem

Suppose you do not know the equation at all — only which physical quantities are likely to matter. Drag on a sphere in a steady stream plausibly depends on the sphere's diameter, the fluid's density and viscosity, and the approach velocity, along with the drag force itself:

Five quantities, and every one of them is built from only three primary dimensions — mass, length, time. The Buckingham Pi theorem says a relationship among \(n\) dimensional variables, describable with \(j\) primary dimensions, can always be rewritten as a relationship among \(n-j\) dimensionless groups instead:

Here \(n=5\) and \(j=3\), so the five-variable problem is secretly a two-variable one. Finding the groups themselves is the method of repeating variables: choose \(j\) of the original variables that between them carry every primary dimension and do not themselves form a dimensionless group — \(\rho\), \(V\), \(D\) serve well here — then combine each repeating variable with every remaining one in turn, solving for the exponents that cancel every dimension.

Carried through, the five variables collapse to two dimensionless groups:

and the theorem's promise is that the original relationship is equivalent to a relationship between the two groups alone, \(\Pi_1 = f(\Pi_2)\) — a drag coefficient as a function of what turns out to be the Reynolds number. The theorem does not hand you that function; wind-tunnel data or a numerical solution does. What it hands you is the enormous reduction in the size of the experiment needed to find it. Testing five independent variables one at a time is a large programme. Measuring one curve, drag coefficient against Reynolds number, is an afternoon.

The one genuinely difficult step is deciding which variables belong on the list in the first place, and no theorem can do that for you. Leave out a relevant variable and the derived groups will not correlate cleanly against data; include an irrelevant one and the analysis produces an extra group that turns out not to matter, which is a wasted effort but at least an honest one. Physical judgement about the problem has to come before the algebra starts.

The common dimensionless groups

A handful of Pi groups recur so often across fluid mechanics that it is worth knowing them by name and, more importantly, by what ratio of forces each one actually represents.

The Reynolds number is inertial force over viscous force. A large Re means inertia dominates and viscosity is a minor correction; a small Re means the opposite, and this single ratio is what laminar and turbulent flow turns on. The Froude number plays the same role for gravity:

inertial force over gravitational force, and it is the number that governs anything with a free surface — a ship's wake, a weir, a hydraulic jump — because gravity is what shapes a free surface in the first place. The Euler number is pressure force over inertial force,

and a pressure coefficient is simply \(2Eu\) under the more common convention of dividing by \(\tfrac{1}{2}\rho V^2\) instead — the factor of two is a bookkeeping choice, not a different physical idea. The Weber number closes the set, inertial force over surface tension:

It matters wherever a length scale is small enough that surface tension competes with inertia at all — a droplet breaking off a jet, a bubble, a thin film — and is safely ignored the moment \(L\) is measured in metres rather than millimetres. Every one of these is the same trick as the sphere's drag coefficient: a ratio of two force scales, arrived at by dimensional reasoning rather than by solving the governing equations directly.

Similarity

A model earns the right to stand in for a prototype in three stages, and each stage is a stricter requirement than the one before it. Geometric similarity is the weakest: the model is the prototype scaled by a single length ratio \(\lambda = L_{\text{model}}/L_{\text{prototype}}\) in every direction, so every angle is preserved and every length is in the same proportion. A model with a fatter fuselage or a shallower wing sweep than its prototype has already failed here, before flow enters the picture at all.

Kinematic similarity asks more: the velocity field itself must be a scaled copy, so that at every pair of corresponding points the ratio of velocities is the same constant, and streamline patterns look identical once redrawn to a common scale. A model can be geometrically perfect and still fail kinematic similarity, if the flow separates in a different place or the wake sits at a different angle.

Dynamic similarity is the strongest of the three, and the one that actually licenses a comparison of forces: every relevant dimensionless group has the same value on the model as on the prototype.

Match Reynolds number — or Froude, or whichever group the Pi theorem said governs this problem — and the dimensionless force coefficients, drag coefficient, lift coefficient, pressure coefficient, come out identical between model and prototype, because those coefficients are themselves the very Pi groups the theorem produced. Dynamic similarity requires kinematic; kinematic requires geometric. The chain only runs one way, and skipping a link invalidates everything built on top of it.

Scale models

Achieving dynamic similarity sounds like a design instruction: build the model at scale \(\lambda\), then run it at whatever speed makes the governing Pi group match. For a Reynolds-dominated problem with the same fluid on both sides, \(Re_m=Re_p\) forces \(V_m = V_p/\lambda\) — a smaller model has to run faster, not slower, and the arithmetic turns hostile quickly. A one-fifth-scale car model in air demands a test section moving at five times the road speed, which for anything above about 25 m/s starts running into compressibility effects the full-size car never sees. This is the real reason some automotive and aerospace tunnels are pressurised, or in the most demanding cases cryogenic: raising density or lowering temperature lowers the fluid's kinematic viscosity, which lets Re be matched at a test speed the tunnel can actually reach.

A free-surface problem is kinder, because it is Froude number that governs wave-making resistance rather than Reynolds number:

A 1:25 scale model of a ship doing 10 m/s is towed at \(10\sqrt{1/25}=2\) m/s — a speed any towing tank can hold steady for as long as the test needs. The honest complication is that matching Froude number and matching Reynolds number at the same scale ratio are mutually exclusive with a single fluid; a model towed slowly enough to match Fr is nowhere near matching Re. Naval architects do not pretend otherwise. The standard resolution splits the resistance into a wave-making part, scaled by Froude number from the model test, and a frictional part, estimated separately from a flat-plate correlation at the full-scale Reynolds number, and adds the two back together. It is not a single elegant scaling law. It is two honest approximations used where each one applies.

Once dynamic similarity — or its deliberate, documented partial version — holds, the payoff is that a dimensionless result measured on the model transfers unchanged. A drag coefficient read off a wind-tunnel model at the matched Reynolds number is the same drag coefficient the full-size vehicle will show, and the actual force on the prototype comes from putting that coefficient back together with the prototype's own density, velocity and area — never the model's.