Natural Convection in Enclosures

The differentially heated square cavity is the canonical problem of natural convection in enclosures, and it deserves its place at the start of this course because almost every practical enclosure, from a wall cavity to a sealed electronics box, can be understood as a variation on it. The geometry is deliberately simple: a square of side L. The left vertical wall is maintained at a uniform hot temperature T_h, the right vertical wall at a uniform cold temperature T_c, and the horizontal top and bottom walls are adiabatic, meaning no heat passes through them. The cavity is filled with a Newtonian fluid, most often air, and gravity acts downward. Nothing moves the fluid except buoyancy.

Before any flow is considered, the driving force needs a scale. In the Boussinesq approximation, which is the standard treatment for this problem, density variations are neglected everywhere except in the buoyancy term, where density is linearized with temperature through the thermal expansion coefficient beta. The result is a single dimensionless group that measures the strength of buoyancy relative to viscous and thermal damping: the Rayleigh number based on the cavity side L. When the Rayleigh number is small, the buoyancy force cannot overcome viscous resistance and the fluid barely moves. When it is large, buoyancy dominates and a vigorous circulation develops. The Prandtl number, the ratio of momentum diffusivity to thermal diffusivity, plays a secondary role: it sets the relative thickness of the velocity and thermal boundary layers but does not decide whether convection happens.

Heat transfer across the cavity is reported with the average Nusselt number on the hot wall: the actual heat flux divided by the heat flux that pure conduction would produce across a stagnant fluid layer of the same thickness. A Nusselt number of 1 means the fluid might as well be a solid slab conducting heat; anything larger measures how much the flow enhances the transfer.

The flow structure changes character as Ra_L increases, and the sequence is worth memorizing because it recurs in every enclosure you will meet. At very low Rayleigh numbers (on the order of 10³ and below), the flow is a slow, single, weak circulation and the temperature field is nearly the linear conduction profile; Nu stays close to 1. As Ra_L rises through roughly 10³ to 10⁴, convection takes over: fluid rises along the hot wall and sinks along the cold wall in a single primary cell, the isotherms tilt, and thin boundary layers begin to form on the vertical walls. In the boundary-layer regime, roughly from 10⁴ upward, the vertical boundary layers carry most of the heat, the cavity core becomes stably stratified (warm fluid sitting over cold fluid, which resists vertical motion), and secondary recirculation zones can appear in the corners where the boundary layers turn. As Ra_L climbs toward 10⁶ and beyond, the circulation grows increasingly vigorous, the boundary layers thin further, and at very high Rayleigh numbers (on the order of 10⁹) the flow transitions to turbulence, first unsteadily in the boundary layers and then throughout the cavity.

The stratified core deserves a closer look because it explains several counterintuitive behaviors of enclosed convection. In the boundary-layer regime the core is not well mixed, as it would be in forced convection. Instead, temperature increases smoothly with height and the core fluid is nearly stagnant, with most of the motion confined to thin sheets along the walls. This means the heat transfer is set almost entirely by the boundary layers, which is exactly why correlations for enclosures take the boundary-layer form Nu proportional to Ra^1/4 in the laminar regime and Ra^1/3 when the boundary layers themselves become turbulent. It also means the top and bottom walls matter more than their adiabatic label suggests: the boundary layers turn at these walls and dump or collect heat there, so the corner regions and the exact thermal condition of the horizontal walls influence the Nusselt number measurably.

Key takeaways

• The Rayleigh number Ra_L = gβΔTL³/(να) is the single governing parameter for the onset and vigor of convection in the cavity; Pr is secondary.
• Nu_L = 1 means pure conduction; the sequence with rising Ra_L is conduction-dominated flow, a single-cell circulation, a boundary-layer regime with a stratified core, and eventually turbulence near Ra_L of order 10⁹.
• In the boundary-layer regime the vertical wall boundary layers carry almost all the heat while the core sits stably stratified and nearly stagnant.
• Always evaluate properties at the mean temperature and use the wall-to-wall spacing as the length scale in Ra_L.

Real enclosures are rarely square. A cavity in a framed wall might be 2.4 m tall and 90 mm deep; the air gap behind a rainscreen panel might be 40 mm deep and meters tall; a double-skin facade can run the full height of a building. The aspect ratio captures this stretching in one number:

A = 1 is the square cavity of Session 1. A much larger than 1 is a tall slot; A much smaller than 1 is a shallow layer. The aspect ratio does not just rescale the square-cavity answer; it changes the flow topology, the threshold at which convection starts, and which correlation you should use.

Consider the tall cavity first, since it is the building industry's everyday case. Stretch a square cavity vertically and the single primary cell stretches with it. The hot-wall boundary layer now has a long run, and as it rises it thickens and can become unstable, shedding secondary cells that stack vertically in the core. Whether the flow stays unicellular or breaks into multiple stacked cells depends on both the aspect ratio and the Rayleigh number: at a fixed Ra_L, increasing A eventually triggers the multicellular pattern, and at a fixed A, increasing Ra_L does the same. The practical consequence is that a tall cavity can transfer noticeably less heat than a naive correlation would predict, because the stacked secondary cells are weaker than one tall cell would be, and because tall narrow slots suppress the onset of convection to higher Rayleigh numbers than the square-cavity value near 10³. This is good news for wall cavities and rainscreen gaps: the tall, thin geometry is inherently self-limiting. It is also why the correlations in Session 3 carry an explicit (H/L) factor with a negative exponent: for the same gap spacing and temperature difference, a taller cavity has a lower Nusselt number.

The shallow cavity, A below 1, behaves differently. For a side-heated shallow cavity the single roll stretches horizontally, the top and bottom walls play a larger role because the flow spends more of its circuit in contact with them, and the heat transfer per unit area of the isothermal walls falls as the walls move apart in the horizontal direction. The scaling with Ra_L remains boundary-layer-like, but the prefactor drops with increasing width. Note that in building physics the more common shallow case is a horizontal or slightly tilted layer heated from below, such as a roof build-up or a floor void; that configuration is covered in Session 4 on inclined enclosures.

Three building details make this session concrete. First, the cavity inside a framed or masonry wall: tall (A of 10 to 30), narrow (50 to 150 mm), side-heated by the indoor-outdoor temperature difference. Convection here is weak and often multicellular, which is one reason filling the cavity with insulation is so effective: the insulation kills the small residual convection outright. Second, the ventilated gap behind cladding or a rainscreen: geometrically a tall narrow slot, but open at top and bottom, so it behaves partly like the open channels of Session 3 (Elenbaas) and partly like a tall enclosed cavity; the stack effect in the gap is what dries the wall, and its heat transfer is a secondary concern. Third, the double-skin facade: a very tall slot (tens of meters) that may be naturally ventilated, mechanically ventilated, or sealed depending on the design. In sealed or weakly ventilated operation the cavity stratifies strongly, with warm air pooling at the top, and the heat transfer to the inner skin is set by the local boundary layer at each height rather than by any single average. Treating a 20 m tall facade with a correlation validated to A = 12 is an extrapolation, and a responsible design either keeps the extrapolation explicit or breaks the cavity into ventilated segments.

A useful mental picture for tall cavities is the local boundary layer view: at any height, the hot-wall boundary layer behaves like a natural-convection boundary layer on a vertical plate, growing with height, while the core provides a stratified background temperature that the boundary layer sees as its "free stream." This picture explains both the negative aspect-ratio exponent in the correlations and the multicellular breakdown: when the boundary layer grows too thick relative to the gap, or the stratification becomes strong enough, the single cell can no longer span the height and stacked cells take over. It also tells you where CFD meshes need resolution in tall cavities: along the full height of both isothermal walls, not just near the top and bottom.

Key takeaways

• Aspect ratio A = H/L changes the flow topology, not just the numbers: tall cavities develop stacked multicellular flow and suppress convection onset relative to the square cavity.
• The (H/L) factor in enclosure correlations has a negative exponent: taller gaps transfer less heat for the same spacing and temperature difference.
• Wall cavities, rainscreen gaps, and double-skin facades are all tall narrow slots; their weak, often multicellular convection is why cavity insulation works so well.
• Always check the validated aspect-ratio range of a correlation before applying it to a real detail, and flag extrapolations explicitly.

Correlations are where enclosure theory meets billable engineering. Nobody sizes a wall cavity or a heat sink fin array by running a research-grade simulation each time; you reach for a validated correlation, check that your geometry and operating point fall inside its validated range, and compute. This session covers the two correlation families you will use most: enclosed vertical cavities (MacGregor and Emery, 1981) and open parallel-plate channels (Elenbaas, 1942). The critical skill is not memorizing the formulas but knowing which family your problem belongs to and which length goes into the Rayleigh number.

For a vertical enclosure heated from one side and cooled from the other, the widely used MacGregor and Emery (1981) correlations give the average Nusselt number in two Rayleigh branches. In both branches L is the gap spacing between the hot and cold walls (the short dimension, the one heat must cross), and H is the cavity height. The aspect ratio H/L enters explicitly with a negative exponent, exactly as Session 2 anticipated: taller cavities transfer less.

Several things are worth noting. Below Ra_L of about 10³ the Nusselt number is essentially 1 (conduction dominates), and the range from 10³ to 10⁴ is a transition zone where convection is establishing itself. The 1/4 exponent in the lower branch is the laminar boundary-layer scaling; the 1/3 exponent in the upper branch is the turbulent boundary-layer scaling, where the heat transfer becomes independent of the cavity height. Notice that the aspect-ratio factor disappears in the upper branch: once the boundary layers are turbulent, the cavity height no longer matters. Also note the Prandtl ranges: the lower branch was validated for a very wide Prandtl range (1 to 2000, covering oils), while the upper branch covers 1 to 20. Air at Pr near 0.7 sits just below the stated lower limit of 1; in practice the correlations are still commonly applied to air with that caveat noted, since the Prandtl dependence is weak.

The second family is the open channel: two parallel plates, heated, with the top and bottom (or the ends) open to a large fluid reservoir so that fluid can freely enter and leave. The classic application is a naturally cooled heat sink: an array of vertical parallel fins, each channel between two fins drawing cool air in at the bottom and exhausting warm air at the top. This is not an enclosure at all, so enclosure correlations do not apply. Elenbaas (1942) correlated the average Nusselt number for this case as follows, with S the channel spacing (the short dimension between plates) and L the channel height (the vertical length of the plates):

The group (S/L) * Ra_S is sometimes called the modified or channel Rayleigh number. Physically it measures the buoyancy driving against the viscous resistance of the tall narrow channel: a tall channel (large L) throttles the flow more, hence S/L appears. The correlation's upper limit of 10³ on this group corresponds to the onset of the boundary-layer regime inside the channel; beyond it the flow develops distinct boundary layers and different treatments apply.

Choosing between the two families comes down to one question: can fluid enter and leave the gap freely? If the gap is sealed or nearly sealed (a glazing unit, a closed wall cavity, a sealed electronics enclosure), use an enclosure correlation with Ra based on the gap spacing L. If the gap is open to a reservoir at both ends (heat sink fins, a ventilated rainscreen gap, an open Trombe-wall channel), use the Elenbaas channel correlation with Ra based on the channel spacing S. The length in the Rayleigh number is always the short dimension across which heat flows and across which the temperature difference is imposed... in the enclosure it is the wall-to-wall gap L; in the channel it is the plate-to-plate spacing S. Getting this wrong, for example putting the 2.5 m cavity height into Ra_L for a wall cavity, inflates the Rayleigh number by the cube of the aspect ratio and produces nonsense.

Key takeaways

• Sealed gap: use an enclosure correlation (MacGregor and Emery) with Ra based on the gap spacing L. Open channel: use Elenbaas with Ra based on the channel spacing S.
• The MacGregor and Emery lower branch (Ra^1/4, with aspect-ratio and weak Prandtl factors) covers 10⁴ to 10⁷; the upper branch (Ra^1/3, no aspect-ratio factor) covers 10⁷ to 10⁹.
• Elenbaas keys on the modified group (S/L) x Ra_S between 10 and 10³; beyond that the channel enters the boundary-layer regime.
• Always verify all three validated ranges (Ra, aspect ratio, Pr) and flag any extrapolation explicitly.

Tilt the square cavity of Session 1 and the physics changes in a way that matters enormously for buildings. A roof build-up, an attic air space under a pitched roof, or a flat-plate solar collector is an inclined air layer, heated from below by the warm ceiling or absorber plate and cooled from above by the roof deck or glazing. Unlike the side-heated cavity, where any temperature difference drives some circulation, a layer heated from below is either stable or unstable depending on the Rayleigh number: below a critical value the stratification (warm below, cold above) is top-heavy and any disturbance grows into convection cells; the layer conducts heat as if the fluid were solid. Tilt adds a second effect: the component of gravity along the layer drives a shear flow even at low Rayleigh numbers, while the component normal to the layer drives the familiar cellular instability. The competition between these two mechanisms is what makes the angle dependence non-monotonic and interesting.

The standard engineering treatment is the correlation of Hollands, Unny, and Raithby (1976) for an inclined air layer heated from below. Here L is the layer thickness (the short dimension, normal to the plates), Ra_L is based on that thickness, and theta is the tilt angle measured from the horizontal, so theta = 0 is a horizontal layer heated from below and theta = 90 degrees would be the vertical side-heated cavity. The correlation is recommended for theta from 0 to 75 degrees; beyond that the flow is essentially the vertical-cavity regime of Session 3.

The three terms after the leading 1 tell a physical story. The leading 1 is pure conduction, always present. The 1.44 term represents the cellular convection driven by the top-heavy stratification; it switches on when Ra_L * cos(theta) exceeds 1708, which is the classical critical Rayleigh number for a horizontal layer heated from below, reduced by the tilt through the cos(theta) factor. The second bracket inside that term, containing sin(1.8 * theta), accounts for the along-slope shear flow that appears once the layer is tilted: it modifies the cellular pattern and becomes important at larger angles. The final term, with the 1/3 exponent, is the turbulent high-Rayleigh contribution, switching on when Ra_L * cos(theta) exceeds 5830. The positive-bracket convention is the elegant part of this correlation: instead of stitching separate formulas for separate regimes, each mechanism contributes only when its own threshold is crossed, so a single expression covers conduction, cellular convection, and turbulent convection.

The critical-angle behavior deserves emphasis because it is the design lever. At theta = 0 (horizontal, heated from below), convection is suppressed until Ra_L reaches 1708; this is the strongest suppression available, which is why horizontal air gaps in floors and ceilings perform close to their conduction value at small thicknesses. As theta increases, the critical value 1708/cos(theta) rises slowly at first, but the along-slope component of gravity simultaneously starts driving a unicellular circulation that does not need any threshold at all. The net result is that the Nusselt number first stays near 1 (suppressed), then rises as tilt increases, with the correlation capturing the transition through the sin(1.8 * theta) term. Around 70 to 75 degrees the layer behaves nearly like a vertical cavity, and the recommended practice is to switch to the vertical-enclosure correlations of Session 3 rather than push this correlation past its validated range.

Two applications anchor this. An attic under a pitched roof: the air layer between the ceiling insulation and the roof deck is heated from below in winter, when the ceiling is warm relative to the cold roof deck, at the roof pitch angle, typically 20 to 40 degrees. Whether that attic air convects or sits still determines how much heat bypasses the insulation through the air space, and the Hollands correlation with the actual pitch angle gives the answer directly. A flat-plate solar collector: the air gap between the absorber plate and the glazing is heated from below at the collector tilt (often set near the latitude angle). Here convection in the gap is a loss mechanism, degrading collector efficiency, and the correlation tells the designer how the loss grows with gap thickness and tilt. In both cases the design instinct is the same: keep the layer thin enough, or the temperature difference small enough, that Ra_L * cos(theta) stays near or below the cellular threshold, and the gap performs close to conduction.

Key takeaways

• Tilt introduces two gravity components: the normal component drives cellular instability above a critical Ra_L * cos(theta) of 1708, while the along-slope component drives a shear circulation.
• The Hollands, Unny, and Raithby correlation uses positive brackets so each mechanism (conduction, cellular, turbulent) switches on at its own threshold; it is valid for theta from 0 to 75 degrees.
• Convection suppression is strongest near the horizontal; near 75 degrees the layer behaves like a vertical cavity and Session 3 correlations take over.
• The correlation is strictly for heating from below; heating from above gives a stably stratified layer with Nu near 1.

A glazing unit is a natural-convection enclosure with commercial consequences: the air or gas gap between panes is typically 6 to 20 mm thick and a meter or more tall, heated on one side by room air and cooled on the other by the outdoors. Three heat transfer paths operate in parallel across that gap. Conduction through the gas is always present. Natural convection adds a parallel path whose strength depends on the gap Rayleigh number, exactly as in Sessions 1 through 3. Thermal radiation between the two glass surfaces adds a third path, and for standard uncoated glass with an emissivity near 0.84 on each side, the radiative path is comparable in magnitude to the conductive-convective path. Any serious window calculation must handle all three; a hand calculation that treats the gap as conduction only is a lower bound, not a rating.

The convection in a glazing gap is usually weak but not negligible. With a 12 mm air gap and a 15 K temperature difference, the gap Rayleigh number lands in the low thousands (Session 6 works this out in full): above the idealized conduction threshold, but far below the boundary-layer regime, and further weakened by the enormous aspect ratio of a full-height pane, which suppresses convection as Session 2 described. The practical upshot is a Nusselt number only modestly above 1 for typical double glazing. This is why widening the gap has a limit: up to roughly the low teens of millimeters, widening the gap reduces the conduction (heat flux proportional to k/L) faster than convection grows, so performance improves; beyond that, the Rayleigh number, which grows with L cubed, pushes the gap into stronger convection, and the total transfer starts rising again. The optimum is a trade-off, and it shifts with the fill gas: argon and krypton have lower conductivity than air, so the conduction term drops and the optimum moves, which is part of why gas-filled units outperform air-filled ones at the same geometry.

So far this describes the center of the glass, far from edges. Real windows have spacers, frames, and edge seals, and each degrades the local performance. The spacer bar that separates the panes at the perimeter is a thermal bridge: aluminum spacers conduct heat around the edge of the gap, creating a cold strip along the glass edge that shows up in winter as condensation risk. Warm-edge spacers made of stainless steel, polymers, or composites reduce this bridging substantially. Frames add their own two- and three-dimensional conduction paths, plus their own small enclosed air cavities that convect weakly. The result is that the whole-window U-value is always worse than the center-of-glass U-value, and the difference is set by edge and frame detailing rather than by the gap physics. When a client asks why a window rated at some U-value still condenses at the edges, the answer is almost always the spacer and the frame, not the gap.

Partitions take the aspect-ratio lesson of Session 2 and turn it into a product concept. If a tall gap convects, divide it: inserting horizontal partitions (or a honeycomb insert) breaks one tall enclosure into a stack of short cells, each with an aspect ratio near 1 and each seeing only a fraction of the total temperature difference. The Rayleigh number of each sub-cell, based on its small height and small local temperature difference, drops back toward the conduction regime, and the stack as a whole transfers heat at close to the conduction rate. The partitions must be thin, low-conductivity, and reasonably sealed against bypass flow to work; any leakage path that lets fluid circulate around the partition defeats the purpose. Honeycomb and slat inserts have been studied for exactly this reason in tall glazing cavities, Trombe walls, and solar collectors: they are a geometric way to buy back the conduction regime without evacuating the gap or filling it with a solid.

For rated performance, the industry does not use hand correlations. ISO 15099 is the established international standard method for the detailed thermal calculation of windows, doors, and shading devices: it defines how to combine the gap convection and radiation models, the frame and spacer conduction, and the boundary conditions into a consistent whole-window calculation, and it is the basis on which simulation tools used for ratings are built. This course teaches the physics so you can sanity-check those tools, not so you can replace them: if a rating tool reports a gap Nusselt number far from what the correlations in Sessions 3 and 4 suggest for that geometry, that is a flag to investigate the inputs, not to override the standard.

Key takeaways

• Heat crosses a glazing gap by conduction, weak natural convection, and radiation in parallel; for uncoated glass the radiative path is significant and a conduction-only hand calc is a lower bound.
• Typical double-glazing gaps sit at low-thousands Rayleigh numbers with very tall aspect ratios, so convection is weak and Nu is only modestly above 1.
• Widening the gap helps only up to a point; beyond it the L-cubed growth of Ra drives stronger convection and the total transfer rises again.
• Spacers and frames create edge and frame losses that separate whole-window performance from center-of-glass performance.
• Partitions and honeycomb inserts suppress convection by dividing a tall gap into short, low-Ra cells; ISO 15099 is the standard method for the full detailed window calculation.

This session puts the whole course to work on one realistic detail: a vertical double-glazed unit with a 12 mm air gap, glass height 1.5 m, warm inner pane at 15 C and cold outer pane at 0 C. The goal is the steady heat flux across the gap by conduction and natural convection, with a clear statement of what is included and what is left out.

Step 1: reference temperature and properties. Film and gap properties are evaluated at the mean temperature, (15 + 0)/2 = 7.5 C = 280.65 K. Tabulated air properties are given at 280 K (ν = 14.11 x 10^-6 m²/s, α = 19.9 x 10^-6 m²/s, k = 0.02439 W/(m K), Pr = 0.713) and at 300 K (ν = 15.89 x 10^-6, α = 22.5 x 10^-6, k = 0.0263, Pr = 0.707). The interpolation fraction is (280.65 - 280)/(300 - 280) = 0.0325, so the interpolated values are essentially the 280 K values: ν = 14.11 x 10^-6 + 0.0325 x (15.89 - 14.11) x 10^-6 = 14.17 x 10^-6 m²/s; α = 19.9 x 10^-6 + 0.0325 x (22.5 - 19.9) x 10^-6 = 19.98 x 10^-6 m²/s; k = 0.02439 + 0.0325 x (0.0263 - 0.02439) = 0.02445 W/(m K); Pr = 0.713 + 0.0325 x (0.707 - 0.713) = 0.7128. The thermal expansion coefficient for air as an ideal gas is β = 1/280.65 = 0.003563 K^-1.

Step 2: Rayleigh number. The length scale is the gap spacing L = 0.012 m, the dimension across which heat flows and across which the 15 K difference is imposed. Ra_L = g * β * ΔT * L³ / (ν * α) = 9.81 x 0.003563 x 15 x (0.012)³ / (14.17 x 10^-6 x 19.98 x 10^-6). The numerator is 9.81 x 0.003563 x 15 x 1.728 x 10^-6 = 9.06 x 10^-7; the denominator is 2.83 x 10^-10. Ra_L = 3.2 x 10³.

Step 3: regime check and Nusselt number. The idealized conduction threshold for an enclosure is Ra_L near 10³; our value of 3.2 x 10³ sits just above it, in the weak-convection transition zone and far below the boundary-layer regime that starts near 10⁴. Two further facts keep convection weak here: the aspect ratio is H/L = 1.5/0.012 = 125, and as Session 2 explained, very tall slots suppress convection onset and develop only weak multicellular motion; and at Ra of a few thousand the circulation, such as it is, barely perturbs the conduction temperature profile. The standard hand-calculation treatment is therefore Nu_L = 1, the conduction value, with the understanding that the true Nusselt number is only slightly above unity. This is also consistent with the MacGregor and Emery lower branch starting at Ra_L = 10⁴: we are below its validated range on the conduction side, so the conduction limit is the correct branch to use.

Step 4: heat flux. With Nu_L = 1:

Step 5: interpretation. The gap conductance is q''/ΔT = 30.6/15 = 2.04 W/(m² K), or equivalently a gap thermal resistance of 0.49 m² K/W. The center-of-glass U-value builds on this by adding the resistances in series: the interior surface film, the two glass panes (small resistance), the gap resistance computed here, and the exterior surface film; the U-value is the reciprocal of that total resistance, so it is lower than the gap conductance alone. Two caveats from Session 5 apply before anyone quotes this number: radiation between the panes is not included, and for uncoated glass it adds a parallel path of comparable size, so 30.6 W/m² is a lower bound on the true gap transfer; and edge and frame effects are not included, so the whole-window value differs further.

The calculation also demonstrates the gap-width trade-off directly. Suppose the gap were widened to 24 mm at the same 15 K difference. The conduction term k/L halves, which looks like a win, but Ra_L grows with L cubed: doubling L multiplies Ra_L by 8, taking it from 3.2 x 10³ to about 2.6 x 10⁴, squarely into the boundary-layer regime where Nu grows as Ra^1/4. The Nusselt number roughly doubles (2^3/4 is about 1.68, plus the regime change), which largely cancels the halved conduction term: the total transfer stops falling and starts rising. Widening the gap past the conduction regime does not keep reducing heat transfer; it invites the very convection you were trying to avoid. That is the physical reason double glazing settled near the low teens of millimeters for air, and it is the single most useful intuition this example can give you.

Key takeaways

• Evaluate properties at the mean temperature by interpolation; for a 12 mm air gap at 15 K difference, Ra_L = 3.2 x 10³, just above the idealized conduction threshold but far below the boundary-layer regime.
• With Nu_L approx 1, the gap heat flux is q'' = kΔT/L = 30.6 W/m², a lower bound because inter-pane radiation is excluded.
• The gap conductance (2.04 W/(m² K)) feeds the center-of-glass U-value in series with surface films and panes; edge and frame effects then separate whole-window from center-of-glass performance.
• Doubling the gap width multiplies Ra_L by 8, pushing the gap into the convecting regime and erasing the conduction gain: wider is not always better.

Every CFD code that claims to handle buoyancy should be able to reproduce the differentially heated square cavity of Session 1, and the accepted yardstick is the benchmark published by de Vahl Davis in 1983. The setup is exactly the Session 1 geometry: square cavity, hot left wall, cold right wall, adiabatic top and bottom, laminar flow, Boussinesq approximation, air-like Prandtl number of 0.71. De Vahl Davis computed highly accurate reference solutions for Rayleigh numbers from 10³ to 10⁶ and reported a compact set of quantities that any new simulation can be checked against. Because the benchmark is laminar, steady (up to 10⁶), and two-dimensional, it isolates the core numerics, the pressure-velocity coupling, the buoyancy source term, and the thermal boundary conditions, from the additional modeling choices that turbulence or three-dimensionality would introduce. If your solver cannot match this case, nothing you compute in a harder case can be trusted.

The headline comparison quantity is the average Nusselt number on the hot wall. The commonly reported benchmark values at Pr = 0.71 are:

Read these numbers as a story, not just a table. From 10³ to 10⁴ the Nusselt number doubles: convection is establishing itself. From 10⁴ to 10⁶ it grows roughly as Ra^1/4 (2.243 to 8.800 is a factor of 3.9 over two decades of Ra, and 100^1/4 is 3.16, close given the low-Ra transition): the boundary-layer regime. A simulation that reproduces the 10³ value but drifts at 10⁶ is usually under-resolved in the boundary layers; a simulation that misses even at 10³ has a more fundamental problem, often in the boundary conditions or the buoyancy term.

The Nusselt number alone is not a complete validation, because two wrongs can make a right-looking average: an overpredicted boundary layer on one part of the wall can cancel an underpredicted one elsewhere. The benchmark therefore also reports the maximum horizontal velocity on the vertical mid-plane and the maximum vertical velocity on the horizontal mid-plane, together with the locations (the y-coordinate of the horizontal-velocity peak and the x-coordinate of the vertical-velocity peak). These quantities test whether your simulation puts the boundary layers in the right place with the right thickness: the vertical-velocity peak should sit very close to the hot and cold walls at high Ra, and its distance from the wall is a direct measure of boundary-layer resolution. The benchmark additionally reports the temperature profile along the mid-planes, which tests the core stratification: your simulation should reproduce the nearly linear, stably stratified core temperature at high Ra, not a well-mixed uniform core. Matching all of these together, velocities, locations, and profiles, is a far stricter test than matching Nu alone, and it is the standard you should hold an in-house validation to.

What does agreement tell you in practice? First, that the buoyancy source term is implemented and coupled correctly: the Boussinesq body force must use the same reference temperature as the boundary conditions, and the gravity vector must point the right way (a surprisingly common setup error). Second, that the mesh resolves the boundary layers: at Ra = 10⁶ the thermal boundary layer on the vertical walls is thin, and you need several cells across it, with the first cell well inside it, to capture the wall temperature gradient that sets the Nusselt number. A mesh that gives the right Nu at 10⁴ may be far too coarse at 10⁶; the correct workflow is a mesh-independence study at the highest Ra you intend to run, watching Nu and the peak velocities converge as the mesh is refined. Third, that the discretization is adequate: second-order spatial discretization is the usual expectation for a benchmark-quality match, and first-order upwinding typically smears the thin boundary layers enough to depress the Nusselt number visibly at high Ra. If your Nu comes in low at 10⁵ and 10⁶ while the flow pattern looks right, numerical diffusion from a low-order scheme or a coarse near-wall mesh is the prime suspect.

A practical validation procedure runs as follows. Build the square cavity with the exact Session 1 boundary conditions: isothermal vertical walls, adiabatic horizontals, no-slip everywhere. Run steady laminar cases at Ra = 10³, 10⁴, 10⁵, and 10⁶ with Pr = 0.71, refining the mesh until the hot-wall average Nusselt number changes by a small tolerance between successive meshes. Compare against the four benchmark values above, compare the peak mid-plane velocities and their locations, and compare the mid-plane temperature profiles. Check that the hot-wall and cold-wall Nusselt numbers agree with each other (energy conservation) and that residuals or monitors are truly flat, since buoyancy-driven flows can settle into slow limit cycles that look converged on residuals alone. Document the mesh counts, the scheme order, and the deviations from the benchmark: that document is your evidence, to yourself and to any client, that the solver is fit for enclosure work.

Key takeaways

• The 1983 de Vahl Davis benchmark is the standard validation case for natural-convection CFD: square cavity, hot left wall, cold right wall, adiabatic top and bottom, Pr = 0.71, Ra from 10³ to 10⁶.
• Benchmark average Nusselt numbers: approx 1.118 at 10³, 2.243 at 10⁴, 4.519 at 10⁵, 8.800 at 10⁶.
• Validate beyond Nu: peak mid-plane velocities and their locations test boundary-layer placement, and mid-plane temperature profiles test core stratification.
• Agreement confirms the buoyancy implementation, near-wall mesh resolution, and discretization order; mesh independence must be demonstrated at the highest Ra you plan to run.
• Watch for setup traps: wrong gravity direction, wrong reference temperature, non-adiabatic horizontals, turbulence models active in a laminar benchmark, and false convergence on residuals alone.