Natural Convection Fundamentals

Walk into a room on a winter evening and stand near a large glazed facade. Even if the thermostat reads a comfortable 21 C, the space next to the glass can feel chilly. What you feel is not simply the cold surface radiating at you. A sheet of room air has been cooled by contact with the glass, has grown denser than the air around it, and is now sliding downward along the pane, spilling across the floor toward your ankles. That is the cold-window downdraft, and it is pure natural convection: motion caused by nothing more than temperature-induced density differences, with no fan anywhere in the picture.

The same mechanism runs in reverse above every radiator. Water at 60 or 70 C heats the metal panels, the panels heat a thin skin of adjacent air, and that air becomes lighter than its surroundings and rises. As the plume climbs it drags surrounding room air along with it, a process called entrainment, so the plume widens and slows with height while a broad, gentle circulation fills the room: air rises at the radiator, travels across the ceiling, sinks along the far walls, and returns along the floor. Baseboard heaters, convectors, and heated floors all drive variants of this loop. A heated floor is a particularly interesting case because warm air sits beneath cooler air, which is an unstable arrangement, so the flow organizes itself into rising plumes spaced across the slab rather than one tidy circulation cell.

What physically drives all of this is the pairing of gravity with thermal expansion. Warm a parcel of air and, at roughly constant pressure, it expands and its density falls. Gravity then pulls the heavier surrounding fluid down more strongly than the lighter parcel, and the net effect is an upward buoyant force on the warm parcel. There is no imposed velocity scale, which is the deep difference from forced convection: in a duct you know the mean velocity because a fan set it, but in a room the air decides its own speed by balancing buoyancy against viscous friction. That self-determination is exactly what makes natural convection both fascinating and harder to calculate. The temperature field drives the flow, the flow reshapes the temperature field, and the two must be solved together.

The consequences reach well beyond comfort. Natural convection sets the heat loss from hot water pipes in a plant room, the cooling rate of a thermal mass wall overnight, the stratification in an atrium or a tall stairwell (warm air pooling at the top is why top-floor corridors overheat), the performance of a solar chimney, and the frost pattern on a poorly insulated duct. In equipment it governs how fast a transformer or a power electronics cabinet cools when its fan fails, and it is the mechanism behind the stack effect that moves air through tall buildings in winter. Anywhere a surface and a fluid sit at different temperatures with no fan blowing, natural convection is working, whether the designer accounted for it or not.

This course builds the working knowledge you need to account for it deliberately. Session 2 writes down the governing equations and introduces the Boussinesq approximation, the simplification that makes the mathematics tractable. Session 3 converts those equations into the four dimensionless groups that organize the whole subject: Grashof, Rayleigh, Prandtl, and Nusselt numbers. Session 4 looks at the boundary layers where almost all the action happens, and Session 5 gives you the empirical correlations, the compact formulas distilled from decades of experiments, that turn a geometry and two temperatures into a heat transfer coefficient. Session 6 works a complete numerical example end to end, and Session 7 collects the mistakes that trap even experienced engineers and points toward what changes when you move the problem into CFD.

This relation, Newton's law of cooling, is the accounting framework for the entire course. Everything that follows is really about one thing: finding a trustworthy value of h for a surface sitting in still air, so that this simple equation gives a useful answer.

Key takeaways

• Natural convection is fluid motion driven by density differences from temperature gradients, with gravity supplying the force and no fan involved.
• The cold-window downdraft and the radiator plume are the two canonical building examples; both are buoyancy-driven boundary-layer flows.
• Unlike forced convection, there is no imposed velocity: the flow speed emerges from the balance of buoyancy and viscous drag.
• Natural convection affects comfort, equipment cooling, stratification, and heat loss throughout buildings, so ignoring it is a design decision with consequences.
• Newton's law of cooling frames the whole course: the central task is estimating the heat transfer coefficient h.

Hand calculations hide the governing equations inside correlations, but every correlation in Session 5 and every CFD simulation in Session 7 is an answer to the same three equations written below. It is worth meeting them directly, because once you see where buoyancy enters the mathematics, the dimensionless groups of Session 3 stop looking like alphabet soup and start looking inevitable.

We work with the standard engineering simplifications: steady flow, a Newtonian fluid, constant viscosity, thermal conductivity and specific heat, and speeds low enough that compressibility and viscous heating are negligible. The one property we refuse to hold constant is density where it matters most, in the gravity term, for reasons that will become clear. The first equation is conservation of mass. For these low-speed flows it takes the incompressible form, which says that the velocity field has no divergence: what flows into a small control volume flows out.

The momentum equation is Newton's second law applied to a fluid particle: mass times acceleration equals the sum of forces. On the left sits the inertia of the fluid, the acceleration of a particle as it is carried along its path. On the right sit the pressure force, the viscous friction force, and the weight of the fluid, the body force ρg. Written for steady flow with the assumptions above:

The last term is the entire subject of this course in one expression. Here is how it arises. Start with the full weight term ρg and split the density into a constant reference value plus a small variation, ρ = ρ_∞ + (ρ - ρ_∞). The constant part combines with the pressure gradient to form the reduced pressure p*, which carries all the boring hydrostatic weight of the fluid column and drives no motion. What remains is the density variation multiplied by gravity: the buoyancy force. Warm fluid has ρ below ρ_∞, so this term points upward, opposite to g, and cold fluid is pulled down. Nothing else in the momentum equation knows about temperature, so this single term is the coupling that makes the flow "natural."

The Boussinesq approximation is the decision to treat density as constant everywhere except in that one gravity term, where it is linearized with temperature. The derivation idea is a Taylor expansion: expand ρ(T) about the reference temperature T_∞ and keep only the first-order term, which introduces the volume expansion coefficient β = -(1/ρ)(∂ρ/∂T)_p, a tabulated property that measures how strongly density falls as temperature rises. Higher-order terms are dropped on the grounds that the temperature differences are modest. The result is the linear relation used above:

This is a superb approximation for room-air problems: with ΔT of 20 K in air near 300 K, βΔT is about 0.07, so density varies by only a few percent while the buoyancy force, which depends on the difference, is captured correctly. The third equation, conservation of energy, then tracks how temperature is carried by the flow and spread by conduction. With no heat sources and negligible viscous dissipation:

On the left is advection, temperature swept along by the velocity field; on the right is diffusion, heat conducted through the fluid. Their ratio, set by the flow the momentum equation produces, is what the Prandtl number will quantify in Session 3.

Now the limits, which matter because the approximation is so convenient that people stretch it too far. It breaks down when β(T_s - T_∞) is not small compared to 1, because then the neglected higher-order density terms and the variation of the other properties across the thermal boundary layer are no longer negligible. For air at room temperature, β is about 1/300 K^-1, so a 30 K difference gives 0.1, acceptable, but a 300 K difference, a furnace wall, a fire plume, a hot industrial process, gives order 1 and the approximation fails; there you need variable-property or compressible formulations. It also fails where the ideal-gas relation β = 1/T does not apply and you forgot to look up the real β for a liquid, where speeds approach compressible regimes, and where large property variations across the layer make evaluating everything at a single film temperature a poor representative choice. For the building and HVAC problems in this course, with temperature differences of tens of degrees, you are on safe ground.

Key takeaways

• Continuity, momentum with a buoyancy body force, and the advection-diffusion energy equation govern every natural-convection flow.
• Buoyancy appears when the hydrostatic pressure field is subtracted out, leaving the density variation times gravity as the only temperature-coupled term.
• The Boussinesq approximation holds density constant except in the gravity term, where ρ = ρ_∞[1 - β(T - T_∞)].
• β measures thermal expansibility; for ideal gases β = 1/T in kelvin.
• The approximation is valid when β(T_s - T_∞) is small compared to 1, which holds for typical building ΔT values but fails for furnaces, fires, and other large-ΔT flows.

Four dimensionless groups carry almost the entire conceptual load of natural convection. Once you can read them the way a structural engineer reads a load diagram, correlations stop being mysterious formulas and become statements about which physical effect dominates. We take them in the order you meet them when reasoning about a problem: how good is the heat transfer, what fluid is it, how strong is the driving force, and what regime results.

The Nusselt number answers the first question. It is the ratio of the actual convective heat transfer to the heat transfer that would occur by conduction alone through a stagnant fluid layer of the same thickness. Equivalently, in its working form, it is the dimensionless heat transfer coefficient:

A Nusselt number of 1 means the fluid might as well be a solid slab conducting heat; there is no convection benefit. A Nusselt number of 100 means convection carries a hundred times what conduction alone would. In room air, natural-convection Nusselt numbers for meter-scale surfaces routinely land in the tens to hundreds, which tells you immediately that the air motion is doing most of the work. Note the subscript L: the Nusselt number is meaningless until you say which length it is based on, and choosing that length correctly is one of the most error-prone steps in hand calculations, as Session 7 will emphasize.

The Prandtl number is a pure fluid property, the ratio of momentum diffusivity to thermal diffusivity:

It tells you which boundary layer is thicker. In air, Pr is about 0.7, so heat diffuses a little faster than momentum and the thermal boundary layer is slightly thicker than the velocity layer. In water Pr is around 7, and in oils it can be in the hundreds or thousands, meaning the velocity field spreads far beyond the thin skin where the temperature changes. For HVAC work you live mostly at Pr near 0.7, but the number matters because every serious correlation includes it: a correlation tuned for air will mislead you if you apply it to a thermal oil without the Prandtl correction.

The Grashof number is the natural-convection analogue of the Reynolds number. Where Reynolds compares inertia to viscous forces using an imposed velocity, Grashof compares the buoyancy force to the viscous force, building its own velocity scale out of g, β, and ΔT:

Large Grashof means buoyancy overwhelms viscosity and the flow is vigorous; small Grashof means viscosity damps everything into a creeping, conduction-dominated regime. The L³ dependence is worth pausing on: double the size of a surface and the Grashof number grows eightfold. Natural convection is strongly size-dependent, which is why a correlation validated on a 10 cm laboratory plate cannot be blindly trusted on a 3 m atrium wall without checking the Rayleigh range.

The Rayleigh number multiplies Grashof by Prandtl, folding the fluid's thermal behavior into the driving-force measure:

This is the master parameter of the subject. It decides whether convection happens at all, whether it is laminar or turbulent, and which correlation branch applies. Low Rayleigh, order 10³ and below in confined gaps, means conduction rules and the fluid sits nearly still. Moderate Rayleigh gives smooth laminar boundary layers. Push past roughly 10⁹ on a vertical plate and the boundary layer transitions to turbulence. Every correlation in Session 5 is really a curve of Nu versus Ra, with Pr as a secondary correction.

The practical skill is the thirty-second order-of-magnitude estimate. Take a 1 m tall interior wall, surface 20 K warmer than the room air, properties of air near 300 K (ν ≈ 15.9 x 10^-6 m²/s, α ≈ 22.5 x 10^-6 m²/s, β ≈ 1/300 K^-1). The numerator is 9.81 x (1/300) x 20 x 1³ ≈ 0.65; the denominator is ν x α ≈ 3.6 x 10^-10; the ratio is roughly 1.8 x 10⁹. That single number tells you three things before any correlation is opened: convection utterly dominates conduction here, the boundary layer is at or past transition to turbulence, and you need a correlation valid up to Ra of order 10⁹. Shrink the same problem to a 1 cm electronics component and Ra drops by a factor of a million (L³ again), landing near 10³, where you would question whether convection does anything at all. That is the power of the estimate: it tells you which physics regime you are in, which is the prerequisite for every calculation that follows.

Key takeaways

• Nu_L = hL/k measures how much convection enhances heat transfer over pure conduction; Nu = 1 means conduction only.
• Pr = ν/α is a fluid property describing whether momentum or heat diffuses faster; air sits near 0.7.
• Gr_L is the buoyancy-to-viscous force ratio, the natural-convection counterpart of the Reynolds number, and it scales with L³.
• Ra_L = Gr_L x Pr is the master parameter: it sets the regime (conduction, laminar, turbulent) and selects the correlation.
• A quick Ra estimate, computable by hand in under a minute, tells you whether natural convection matters and which correlation range you need.

Almost everything interesting in natural convection happens in a thin region hugging the surface. Outside it, the fluid barely knows the wall is there; inside it, velocity rises from zero at the wall to a maximum and falls back toward zero at the layer edge, while temperature falls from the surface value to the ambient value. Understanding this boundary layer is what turns correlations from black boxes into readable statements, and it is also what determines mesh requirements when you eventually move to CFD.

Consider a vertical plate hotter than the surrounding air. At the bottom leading edge the boundary layer has zero thickness and the flow starts from rest. Buoyancy accelerates fluid upward inside the layer, viscous drag at the wall holds it back, and the layer thickens as it climbs because it keeps entraining slow ambient fluid at its outer edge. The velocity profile is distinctive: zero at the wall by the no-slip condition, climbing to a peak somewhere inside the layer, then decaying to zero far out, since the ambient air is still. The temperature profile is simpler, maximum at the wall and decaying outward, but its thickness relative to the velocity layer depends on Prandtl number: for air the thermal layer is a little thicker than the velocity layer, while for high-Prandtl oils the temperature drops across a thin inner skin of a much broader velocity layer.

Scaling analysis extracts the essential quantitative behavior without solving anything exactly. Balancing buoyancy against viscous friction and advection against conduction in the boundary layer equations yields the laminar result that the layer thickness grows as the inverse fourth root of Rayleigh number:

The consequences are immediate. Because the heat transfer coefficient scales as k/δ, the Nusselt number inherits the fourth-root dependence, Nu ~ Ra^1/4, which is exactly the exponent you see in the laminar branches of the Session 5 correlations. It also tells you that boundary layers in room air are thin: at Ra_L = 10⁹ the fourth root is about 178, so the layer is roughly half a percent of the plate height, a few millimeters on a meter-scale wall. That thinness is why natural convection can move so much heat despite air's miserable thermal conductivity, and it is why a CFD mesh needs fine near-wall resolution, a point Session 7 returns to.

As the layer climbs and thickens, it eventually loses its laminar composure. Disturbances that viscosity could damp in the thin, fast lower region grow in the thicker, slower upper region, and the flow transitions to turbulence. For a vertical plate in air the commonly cited transition sits at Ra_L around 10⁹, though in practice it is a range rather than a line, sensitive to surface roughness, vibrations, and upstream disturbances. Beyond transition the layer mixes vigorously, the heat transfer coefficient stops declining with height as quickly, and the average Nusselt number shifts to a steeper Ra^1/3 dependence. The one-third power carries a neat physical message: since h ~ (k/L)Ra^1/3 and Ra itself contains L³, the length cancels out, so in the turbulent regime the heat transfer coefficient becomes approximately independent of plate height. A tall turbulent wall and a short turbulent wall, at the same ΔT, transfer heat at nearly the same rate per unit area.

Horizontal plates tell a different story because orientation decides stability. A hot surface facing upward, or equivalently a cold surface facing downward, places light fluid beneath heavy fluid, which is unstable: the boundary layer cannot remain a tidy sheet and instead breaks into rising plumes, or falling sheets for the cold case, spaced across the surface. Heat transfer is vigorous and the correlations carry the turbulent one-third power even at moderate Ra. Flip the plate, hot facing downward or cold facing upward, and heavy fluid sits beneath light fluid, a stable stratification that suppresses motion. The boundary layer stays thin and laminar, heat creeps across largely by conduction with a weak lateral drift, and Nusselt numbers are markedly lower. This asymmetry is the single most important qualitative fact about horizontal-plate natural convection, and it is why Session 5 gives you different correlations for the two orientations: using the wrong one can easily misstate the heat transfer by a factor of two or more.

Key takeaways

• The boundary layer is where velocity and temperature change; outside it the fluid is nearly undisturbed.
• On a vertical plate the layer starts thin at the leading edge, thickens with height, and carries a velocity peak inside the layer.
• Laminar scaling gives δ/L ~ Ra_L^-1/4, which implies Nu ~ Ra^1/4; boundary layers in room air are millimeters thick on meter-scale surfaces.
• Transition to turbulence on vertical plates is commonly cited near Ra_L ≈ 10⁹; in the turbulent regime h becomes nearly independent of plate height.
• Horizontal plates are orientation-dominated: hot-facing-up is unstable with plume eruption and high heat transfer, hot-facing-down is stably stratified with low heat transfer.

Correlations are where theory meets the laboratory. The boundary-layer equations of Session 2 admit exact similarity solutions in a few idealized cases, but real surfaces have edges, real rooms have drafts, and real flows transition to turbulence, so the engineering community has distilled thousands of experiments into compact formulas of the form Nu = C x Raⁿ, sometimes with a Prandtl-number correction. Used within their validated ranges, with properties evaluated correctly, they predict average heat transfer coefficients within the scatter of the experiments themselves. Used outside those ranges, they are just confident-sounding guesses. This session gives you the two correlation families you will reach for most often in building work.

For a vertical plate, the workhorse is the Churchill and Chu (1975) correlation. Its form looks more elaborate than a simple power law because it is a blending function: it asymptotes to the laminar fourth-root behavior at low Ra and to the turbulent one-third behavior at high Ra, transitioning smoothly between them, which is why a single expression can span an enormous range:

The correlation is generally reported as valid over roughly Ra_L from 10^-5 to 10¹², which covers essentially every building-scale situation from a small warm sensor to an atrium wall. The characteristic length L is the plate height, the vertical dimension along which the boundary layer develops. Note that the correlation returns the surface-averaged Nusselt number, so the h you compute is the average over the whole plate, exactly what you need for total heat loss, but not the local value at a particular height.

Horizontal plates demand more care, because as Session 4 showed, orientation changes the physics. The widely used correlations in the style of Fujii and Imura (1972) split into two cases. First, the unstable case, a hot surface facing upward or a cold surface facing downward, where plumes erupt from the surface:

Second, the stable case, a hot surface facing downward or a cold surface facing upward, where stratification suppresses the flow and two Rayleigh ranges are distinguished:

For horizontal plates the characteristic length is not a side dimension but the area-to-perimeter ratio, L = A_s/P, where A_s is the surface area and P is the perimeter. For a rectangular plate of sides a and b this gives L = ab/[2(a+b)], which for a large plate approaches half the short side. The physical idea is that flow must travel inward from the edges, so the relevant scale is set by how far the edge is on average. This definition surprises newcomers and it is one of the most common sources of error in hand calculations.

Properties in every correlation above are evaluated at the film temperature, the arithmetic mean of surface and ambient temperatures:

Evaluating at T_f is a compromise: the boundary layer spans temperatures from T_s to T_∞, and the midpoint is the accepted representative value. For ideal gases the expansion coefficient follows directly, β = 1/T_f with T_f in kelvin. For liquids, look up β in property tables rather than using the ideal-gas relation.

The reliable workflow, then, is a short checklist: identify the geometry and orientation, choose the characteristic length (height for vertical plates, A_s/P for horizontal), compute the film temperature and look up or interpolate properties there, form Ra_L, verify it falls inside the correlation's stated range, compute Nu_L, convert to h = Nu_L x k / L, and finally q = h x A x ΔT. Session 6 runs this checklist end to end with real numbers.

Key takeaways

• Churchill and Chu (1975) gives the average Nusselt number for vertical plates over roughly Ra_L = 10^-5 to 10¹², with L as the plate height.
• Horizontal plates use L = A_s/P and orientation-specific correlations: unstable (hot up) transfers markedly more heat than stable (hot down).
• Always confirm Ra_L lies within the correlation's stated range before trusting the result.
• Evaluate all properties at the film temperature T_f = (T_s + T_∞)/2; for ideal gases β = 1/T_f with T_f in kelvin.
• The output of every correlation is a surface-averaged h, suitable for total heat rate q = hAΔT.

This session does the entire calculation for a realistic case, with every number shown, so you can follow along with a calculator and reproduce each step. The scenario: a vertical wall panel 1.2 m tall and 0.6 m wide, its surface held at 45 C, facing room air at 20 C. We want the natural-convection heat loss from the room-facing side of the panel. No radiation, no fans, just the buoyancy-driven boundary layer.

Now interpret the number rather than just filing it. First, the sanity checks all pass: h = 4.1 W/m²-K sits squarely in the 2 to 10 W/m²-K band typical of natural convection in room air, and a quick scaling estimate h ≈ (k/L)Ra^1/4 gives about 5.4 W/m²-K, the same order of magnitude, confirming no arithmetic blunder. Second, the magnitude is physically sensible: a 0.72 m² panel at 25 K above room temperature delivering roughly 73 W by convection alone means natural convection is a substantial but not dominant contributor next to radiation, which for a 45 C surface in a 20 C room would add a comparable or larger term. Anyone sizing this panel from convection alone would undersize it, and anyone ignoring convection would oversize the radiation estimate's share. Third, scale matters: had the panel been wall-mounted with its back side blocked, 73 W is the whole story; had it been free-standing in the room, the second face would contribute a similar amount, roughly doubling the total. Finally, remember this is a surface-averaged, steady result. It answers "how much heat leaves this panel," which is the right question for load calculations and equipment sizing, but it says nothing about where the warm air goes, which is a job for the plume thinking of Session 1 or the CFD of Session 7.

Key takeaways

• The workflow is fixed: film temperature, interpolated properties, Rayleigh number, range check, correlation, h, then Q = hAΔT.
• For the 1.2 m x 0.6 m panel at 45 C in 20 C room air: Ra_L = 3.61 x 10⁹, Nu_L ≈ 183, h ≈ 4.07 W/m²-K, Q ≈ 73 W from one face.
• βΔT = 0.082 confirms the Boussinesq basis of the correlation is valid for this ΔT.
• The result sits in the expected 2 to 10 W/m²-K band for room-air natural convection, and a scaling cross-check agrees to order of magnitude.
• Convection here is significant but partial: radiation adds a comparable term, and a free-standing panel would roughly double the total with its second face.

Natural-convection hand calculations are short, which makes them feel safe, but their brevity concentrates the risk: a single wrong choice early in the chain propagates through the L³ in the Rayleigh number and the fractional exponents of the correlations, and the final answer can be off by factors, not percent. The mistakes below are collected from real practice, and each has embarrassed experienced engineers. Read them as a pre-flight checklist.

The most damaging mistake is the wrong characteristic length. For a vertical plate L is the height; for a horizontal plate it is A_s/P; for a cylinder it is the diameter. Because Ra scales with L³, using the 0.6 m width instead of the 1.2 m height in Session 6's example would have cut Ra by a factor of eight and the resulting h by roughly a factor of 1.7. Mixing up the horizontal-plate definition is worse: using a side length where A_s/P belongs can shift Ra by orders of magnitude. Always write down which L you used and why, as if a colleague will audit it, because eventually one will.

Next is properties at the wrong temperature. The film temperature exists because the boundary layer spans T_s to T_∞, and evaluating at either endpoint biases every property in the same direction. For the modest ΔT values in buildings the error is small, a few percent, but it grows with ΔT, and in high-temperature problems it becomes the dominant error. Related is the β blunder from Session 2: β = 1/T_f needs kelvin, and liquids need tabulated β, not the ideal-gas relation at all. A spreadsheet that inverts a Celsius temperature produces a β off by an order of magnitude, and every downstream number inherits it silently because the units still look right.

Then come the correlation errors: applying a correlation outside its validated Ra range, and confusing the horizontal-plate orientations. Extrapolating a laminar correlation into the turbulent regime, or vice versa, discards the physics the correlation was fitted to; the range check in Session 6 takes ten seconds and is never optional. The orientation error is conceptual: a hot surface facing downward was assigned the hot-facing-upward correlation, which assumes plume eruption that cannot physically occur in the stable stratification, and the answer comes out roughly double the truth. Sketch the density arrangement first, decide stable versus unstable, then choose the formula.

Three quieter mistakes deserve mention. Forgetting radiation: a 45 C panel in a 20 C room radiates significantly, and a "natural convection heat loss" that omits it answers a question nobody asked. Counting the wrong area: one face versus two, or including edges that the correlation never covered. And treating the average h as a local value: correlations give surface-averaged coefficients, so using them to predict a local hot spot temperature is misuse, even if the total heat rate is right.

Be honest about uncertainty, too. Even a flawless hand calculation carries the scatter of the underlying experiments, typically on the order of 10 to 25 percent for these correlations, plus your uncertainty in the actual surface temperature, which in the field is often a guess. Hand calculations are for sizing, for checking, for deciding whether something matters, and for sanity-checking simulations. They are not guarantees.

When the problem outgrows hand methods, CFD removes the correlations entirely and solves the Session 2 equations directly on a mesh. That freedom trades one set of responsibilities for another. The mesh must resolve the boundary layer: Session 4's scaling, δ/L ~ Ra_L^-1/4, tells you the layer is millimeters thick on meter-scale walls, so the first cell off the wall must be a fraction of that, with a gentle growth ratio outward, or the computed h is fiction. The domain must be large enough that artificial boundaries do not choke the entrainment feeding the plume; a room modeled too tightly returns a different flow than the real room. Buoyant flows are often unsteady, with plumes that sway and shed, so a steady solver may never converge or may converge to a non-physical symmetric solution, and the turbulence model must handle buoyancy-influenced mixing rather than just shear. The Boussinesq approximation usually carries over into the CFD setup, subject to the same βΔT check from Session 2, and radiation often needs to be coupled in rather than added afterward. None of this makes CFD unreliable; it makes it a method whose errors live in different places than the hand calculation's, which is precisely why the two validate each other so well.

As a rule of thumb for choosing your tool: reach for the hand calculation when you need total heat rates, quick sizing, or a check on someone else's number; reach for CFD when you need the flow pattern itself, local temperatures, comfort fields, or interactions between multiple buoyant sources that no correlation covers. And whichever you chose, run the other one at least approximately: a hand calculation that agrees with the CFD to within 20 percent is the strongest evidence either of them is right.

Key takeaways

• The dominant hand-calculation errors are wrong characteristic length, properties at the wrong temperature, out-of-range correlations, and horizontal-plate orientation mix-ups.
• Because Ra scales with L³, length errors propagate brutally; always document which L was used and why.
• Average correlations give total heat rates, not local values, and they omit radiation unless you add it.
• In CFD there are no correlations, but the boundary layer must be meshed (δ/L ~ Ra_L^-1/4 sets the scale), the domain must admit entrainment, and unsteadiness plus turbulence modeling must be handled.
• Use hand calculations for sizing and checking, CFD for flow fields and complex interactions, and let each validate the other.