Heat Transfer · Design background

Designing a Solar Flat-Plate Heat Exchanger

Before any equation gets typed into a solver, an engineer has to decide which physics matters, how to make the problem dimensionless, and which mathematical machinery actually applies. Here's that whole chain, end to end.

▶ Play the lab📖 Design background

1The physical setup

Picture a flat-plate solar collector: an incompressible fluid flows steadily through a thin rectangular channel of dimensions $x_0, y_0, z_0$, absorbing heat from the sun-warmed plate and losing some of it back to the surroundings by convection (coefficient $h$, fluid conductivity $k$). The question that matters for design: how hot does the fluid get by the time it exits? Answering that requires solving the energy balance for the temperature field inside the channel.

2Nondimensionalize first

Raw physical variables — flow rate $\dot m$, specific heat $C_p$, conductivity $k$, channel size — are a mess to carry through a derivation. Group them into two dimensionless numbers instead:

$$ \text{Graetz number } \phi = \frac{\dot m\, C_p\, y_0}{x_0 z_0 k} \qquad \text{Nusselt-like number } Nu = \frac{h\,y_0}{k} $$

$\phi$ tracks how far along the channel you are, in a sense scaled by the flow's ability to carry heat away; $Nu$ is a Biot-style ratio of convective strength to conductive strength. Every plot in the lab is a curve of dimensionless exit temperature $\psi$ against $\phi$, for a chosen $Nu$ — once you're in these units, the result applies to any channel size or flow rate that gives the same two numbers.

3Separate variables — with a twist at the boundary

The energy balance is a PDE. Separating variables, the usual move from the heat equation, splits it into an ODE in the cross-channel direction. Where this problem gets interesting is the boundary condition: heat doesn't just stop at the wall, it convects away, so the condition mixes the temperature and its derivative (a Robin boundary condition) instead of fixing either one outright. Substituting the ODE's general solution into that mixed condition is what produces the eigenvalue equation:

$$ \beta_n \tan\beta_n = Nu $$

There is no algebraic way to solve for $\beta_n$ — tan and a linear term can't be isolated from each other. This is the signature of a convective boundary condition; you'll see the exact same transcendental form in fin analysis, lumped-capacitance validity charts, and any other 1-D diffusion problem with convection at the edge.

4Bracket and solve for the eigenvalues numerically

Plot $f(x)=x\tan x - Nu$ and a pattern appears: on each interval $(n\pi,\, n\pi+\pi/2)$, $f$ starts at $-Nu$ (negative) and shoots to $+\infty$ as $x\to(n\pi+\pi/2)^-$ — guaranteeing exactly one root per interval. That's a perfect bracket for bisection or Newton's method, the same techniques from the root-finding lab. The lab computes the first 21 eigenvalues live this way, and they match the original paper's table to 4–5 significant figures.

▶ In the lab

Scroll the eigenvalue list as you change Nu — watch β₀ (the fundamental) shift noticeably, while the higher βₙ barely move (they're pinned close to $n\pi+\pi/2$ regardless of $Nu$). Open the lab →

5Assemble the eigenfunction series

Just like a Fourier series sums sine/cosine modes weighted by coefficients, the full ("distributed") solution sums an infinite family of modes built from these eigenvalues:

$$ \psi_{\text{distributed}} = 1 - 2\sum_{n} \frac{Nu^2\,e^{-\beta_n^2/\phi}}{\beta_n^2\big(Nu(Nu+1)+\beta_n^2\big)} $$

The $e^{-\beta_n^2/\phi}$ factor decays faster for larger $\beta_n$, so far from the inlet (large $\phi$) only the first mode or two actually contribute — the same convergence behavior as a smooth Fourier series, just dressed up in different eigenvalues.

6The lumped shortcut, and when it's allowed

If you're willing to assume the temperature is uniform across the channel cross-section at every point along its length — exactly the same simplifying assumption as lumped capacitance in transient conduction — the whole series collapses to one exponential:

$$ \psi_{\text{lumped}} = 1 - e^{-Nu/\phi} $$

Lumped capacitance is valid when the Biot number is small (no significant internal gradient); here the analogous criterion is $Nu$ small. The lab's comparison shows this directly: for $Nu<1$, the lumped curve tracks the full distributed series closely enough to use in practice — matching the original paper's conclusion — but for $Nu\ge1$ the spatial profile genuinely matters and you need the full series.

Key takeaways

EngineeringCandy · Learn · the design thinking behind the playground