Dimensional Analysis and CAE Simulation

Simply put, it's essential for sanity-checking simulation results. Just verifying that the output stress dimension is Pa or that the heat flux is W/m² can catch half of the bugs caused by input errors.

Very common. A famous example is NASA's Mars Climate Orbiter in 1999. Lockheed Martin output data in pound-force (lbf), which NASA read as newtons (N) for trajectory calculations, causing the probe to enter the Martian atmosphere and be lost. A $328 million loss caused by a single unit system error.

Moreover, dimensional analysis has another powerful use. Using Buckingham's π theorem to non-dimensionalize parameters allows deriving scaling laws for experiments. Techniques used daily in practice, like extrapolating results from a 1/10 scale wind tunnel model to the full-scale object, or reducing the computational domain in CFD to decrease mesh count.

It's a label representing the "type" of a physical quantity. In SI, there are seven fundamental dimensions. The following four are particularly important in CAE:

• Mass M (kg)
• Length L (m)
• Time T (s)
• Temperature Θ (K)

All physical quantities can be expressed as products of powers of these fundamental dimensions. For example, force is:

Stress is force divided by area, so:

And the principle of dimensional homogeneity: In physically meaningful equations, the dimensions on both sides of the equality must always match. If they don't, the equation is wrong.

Exactly. Let's verify. The dimension of stiffness is $[K] = \mathrm{M \, T^{-2}}$ (N/m = kg/s²), displacement is $[u] = \mathrm{L}$ (m). So the left side is $\mathrm{M \, L \, T^{-2}}$ = dimension of force. It matches correctly. If this were broken, it would be evidence of mixed unit systems in the input data.

In a nutshell, it's a theorem that tells you "the number of independent dimensionless parameters governing a physical problem".

This means the original physical problem can be completely described by just $p$ dimensionless numbers. Since the number of variables is reduced, the number of parameter studies drastically decreases.

Good question. Let's list the physical quantities related to pressure loss $\Delta p$ in pipe flow:

• $\Delta p$ (pressure loss): $\mathrm{M \, L^{-1} \, T^{-2}}$
• $V$ (flow velocity): $\mathrm{L \, T^{-1}}$
• $D$ (pipe diameter): $\mathrm{L}$
• $L$ (pipe length): $\mathrm{L}$
• $\rho$ (density): $\mathrm{M \, L^{-3}}$
• $\mu$ (viscosity): $\mathrm{M \, L^{-1} \, T^{-1}}$
• $\varepsilon$ (surface roughness): $\mathrm{L}$

$n = 7$ variables, fundamental dimensions $k = 3$ (M, L, T), so $p = 7 - 3 = 4$ π groups exist.

This means pipe pressure loss can be expressed in the form $f = \phi(Re, L/D, \varepsilon/D)$. A 7-variable problem has been organized into 4 dimensionless parameters. Instead of doing a full combination parameter study in CFD, you only need to vary Re, $L/D$, and $\varepsilon/D$.

Exactly. This is the essence of dimensional analysis—it simultaneously achieves dramatic computational cost reduction and identifies the essential physical parameters.

Let's look at three dimensionless numbers that frequently appear in CAE, along with their derivation processes.

Derivation: Non-dimensionalize the Navier-Stokes equations with characteristic length $L$ and characteristic velocity $V$. The scale of the convective term $\rho(\mathbf{v} \cdot \nabla)\mathbf{v}$ is $\rho V^2/L$, and the scale of the viscous term $\mu \nabla^2 \mathbf{v}$ is $\mu V/L^2$. Their ratio is:

In practice, for pipe flow, $Re < 2300$ is laminar, $Re > 4000$ is turbulent. This is the criterion for deciding whether to use a turbulence model ($k$-$\varepsilon$, SST $k$-$\omega$, etc.) in CFD.

Derivation: Non-dimensionalize the energy equation. Taking the ratio of the heat flux at the wall $q = h(T_w - T_\infty)$ to conduction within the fluid $q = k_f \Delta T / L$ gives $Nu = hL/k_f$. $Nu = 1$ means "convection is no different from conduction", $Nu = 100$ means "convection transports heat 100 times faster".

Exactly. Moreover, there are many empirical correlations for the Nu number. For example, for turbulent pipe flow, the Dittus-Boelert equation:

If CFD results deviate significantly from this correlation, you should suspect insufficient mesh or an incorrect turbulence model choice. This is the sanity check via dimensional analysis.

Here $L_c$ is the characteristic length (= volume/surface area), $k_s$ is the solid's thermal conductivity. If $Bi < 0.1$, the temperature distribution within the solid can be considered nearly uniform, so a 3D thermal analysis is unnecessary and a lumped parameter model (Newton's law of cooling) is sufficient.

For example, consider an aluminum electronic component ($k_s \approx 200$ W/(m·K)) cooled by natural convection. Component size 10mm, $h \approx 10$ W/(m²·K):

The Bi number is extremely small, so the temperature gradient within the component is negligible. That means there's no need to create a 3D mesh and perform thermal analysis. On the other hand, for a plastic enclosure ($k_s \approx 0.2$ W/(m·K)), $Bi = 0.5$, so a 3D analysis considering internal temperature distribution is necessary.

Yes. Similarity (similitude) refers to "the conditions for correctly scaling the results of a scaled-down model (mock-up) to the full-scale object". There are three similarity conditions:

1. Geometric Similarity: Shapes are similar (all dimensions match with scale ratio $\lambda_L$)
2. Kinematic Similarity: Velocity fields are similar (velocity ratio $\lambda_V$ matches at all points)
3. Dynamic Similarity: Ratios of acting forces match (all relevant dimensionless numbers match)

Sharp observation. In fact, generally it is impossible to satisfy all dimensionless numbers simultaneously. For example, in ship model testing, it's impossible to simultaneously match the Reynolds number (viscous dominance) and the Froude number (gravity wave dominance). Therefore, in practice, we use partial similarity, which means "matching only the dominant dimensionless numbers".

Let's use a wind tunnel test as an example. Testing a 1/5 scale model ($L_m = 1$ m) of a real car (overall length $L_p = 5$ m) in a wind tunnel. If we want to match the Reynolds number:

Using the same air ($\nu_m = \nu_p$):

If the real car travels at 100 km/h, the model would need a wind speed of 500 km/h (approx. Mach 0.4). Since compressibility effects start to appear, in reality, a pressurized wind tunnel is used to increase density, or we give up on perfect Reynolds number matching and apply corrections.

If the Reynolds number matches, the drag coefficient $C_D$ will be the same for the model and the real object:

The drag force on the prototype can then be calculated as:

Since $C_D$ is the same, and all the relationships follow from dimensional analysis, you can transfer the model's aerodynamic data directly to the full-scale car. This is the power of similarity laws.

In CFD, we often reduce the computational domain size using periodicity and symmetry. For example, to analyze heat transfer in a pin fin array, you only need to mesh a single fin unit cell and apply periodic boundary conditions on the sides. The scaling laws tell you exactly how to interpret the results for the full array without simulating it all.

Another example: when the internal Rayleigh number exceeds a critical value, natural convection begins in an enclosed cavity. Using similarity laws based on Ra number, you can predict when this transition occurs without brute-force parameter sweeps. This is the practical value of dimensional analysis in CAE.