3.1 The effect of the number of stripes

In this subsection we present results of a series of simulations in which we varied the number of stripes while keeping $\ell _{C}=1/2$ . Four sets of cases were run for $Ra=10^{7}$ , $10^{8}$ and $10^{9}$ , and $\mathit{Pr}=1$ and $10$ .

Figure 3. (a) Nusselt number $Nu$ and (b) compensated Nusselt number $NuRa^{-0.3}$ against $Ra$ for various $f$ and $Pr=1$ . The markers show the actual results while the dotted lines indicate the trend between these simulations. For all simulations $\ell _{C}$ was fixed to $\ell _{C}=0.5$ .

First we show in figure 3(a) the curves of $Nu$ ( $Ra$ ) for the fully conducting case and three striped cases with $f=$ 1, 10 and 90. The markers show the actual results from the simulations and the dotted lines indicate the trend between the measurements. From figure 3(a) we see that the scaling does not differ significantly between all cases and all curves are almost parallel. We do see a strong dependence on the wavenumber of the pattern. For $f=1$ , we computed that $Nu$ is approximately two-thirds of the fully conducting case. Increasing $f$ results in a larger $Nu$ and the values almost converge with the fully conducting case. Both the $f=90$ and $f=10$ cases are relatively closer to the fully conducting case at $Ra=10^{7}$ than at $Ra=3\times 10^{9}$ . Our interpretation is that, due to the lower $Ra$ , the thermal boundary layer is thicker, which results in larger horizontal conduction of heat. As all curves have approximately the same scaling, we have compensated the data using $Ra^{-0.3}$ , which is shown in figure 3(b). Here it is clearly visible that, for low $Ra$ , the $f=90$ case is almost as efficient as the fully conducting case, but the differences increase with increasing $Ra$ . The two extreme cases, $f=1$ and the fully conducting case, follow a similar trend; however, the $f=1$ case is shifted to a lower, less efficient level. Both panels of figure 3 clearly show that an increase in wavenumber of the pattern results in an increase in  $Nu$ .

Figure 4. (a) Nusselt number $Nu$ for various $Ra$ and $\mathit{Pr}$ versus pattern wavenumber $k_{x}$ . The shaded area shows the data points for which the stripe width is smaller than the thermal boundary layer. (b) The same data as in panel (a) but normalized by $Nu_{fc}$ . The black points are the same as in the shaded area in panel (a). For both panels, the error bars shows the statistical convergence error.

To make this point even more clear, we plotted $Nu$ ( $k_{x}$ ) in figure 4(a). All datasets show a clear $k_{x}$ dependence: $Nu$ increases quite strongly with $k_{x}$ . The grey area indicates the data points for which the width of the stripes $L_{p}$ is smaller than the thermal boundary layer thickness $\unicode[STIX]{x1D706}_{T}$ , which can be estimated as $\unicode[STIX]{x1D706}_{T}=H/(2Nu)$ . Ripesi et al. (Reference Ripesi, Bigerale, Sbragaglia and Wirth2014) showed that the heat transfer increases monotonically with increasing $k_{x}$ until $L_{p}$ is comparable to $\unicode[STIX]{x1D706}_{T}$ . A similar trend is visible for $Ra=1\times 10^{7}$ and $Ra=1\times 10^{8}$ even if the $Ra=1\times 10^{7}$ case shows minor increases in heat transfer when $L_{p}$ is further decreased beyond $\unicode[STIX]{x1D706}_{T}$ . The change in $\mathit{Pr}$ from $\mathit{Pr}=1$ to $\mathit{Pr}=10$ does not have a significant effect on the heat transfer, at least not for $Ra=10^{8}$ , for which the two datasets are practically overlapping. This behaviour is similar to the standard RB case, in which the $\mathit{Pr}$ dependence of $Nu$ is also weak (Ahlers et al. Reference Ahlers, Grossmann and Lohse2009).

To perform a better comparison between the different $Ra$ , we normalize the resulting $Nu$ using $Nu_{fc}$ , the Nusselt number of the fully conducting case. The normalized $Nu$ for different stripe configurations are shown in figure 4(b). We see the same trend for all $Ra$ and $\mathit{Pr}$ . At the lowest $k_{x}$ , in which we only have a single conducting and single insulating stripe, the effective $Nu$ is approximately two-thirds of the fully conducting case. When the number of stripes, i.e. $k_{x}$ , is increased, we see that, for all tested $Ra$ and $\mathit{Pr}$ , the Nusselt number slowly converges to almost the fully conducting case. So, remarkably, even if only half of the plate is conducting, it can be almost as effective as if the plate is fully conducting. In this compensated plot it is also clearly visible that, for the largest $k_{x}$ of $Ra=1\times 10^{7}$ , for which $L_{p}$ goes below the size of $\unicode[STIX]{x1D706}_{T}$ , the heat transfer is still increasing.

The differences between the different $Ra$ are not so clear at the lowest $k_{x}$ ; however, for a slight increase of $k_{x}$ we see that the curves order themselves. The $Ra=10^{9}$ case increases slightly slower with $k_{s}$ when compared to the $Ra=10^{8}$ case. The $Ra=10^{7}$ case increases slightly faster than the $Ra=10^{8}$ case and it ends at 99.7 % of the fully conducting case for the largest number of stripes used. The explanation for this trend is the difference in boundary layer thickness, which decreases when we increase $Ra$ . The boundary layer controls the heat transport from the bulk to the conducting region. Below the insulating region, this heat transport must also go in the horizontal direction. By increasing $Ra$ and thus decreasing the thickness of the boundary layer, the same amount of heat needs to be conducted through a smaller ‘channel’. This explains the decreased effectiveness with increasing  $Ra$ .

In these sets of simulations, we also compare $\mathit{Pr}=10$ with $\mathit{Pr}=1$ for $Ra=10^{8}$ . The results for both $\mathit{Pr}$ are quite similar. At lower $k_{x}$ , we see that the $\mathit{Pr}=10$ case has only a marginally larger $Nu$ and these differences become smaller with increasing $k_{x}$ and are within the uncertainty of the simulation.

Figure 5. (a) Temperature of the bulk fluid $\unicode[STIX]{x1D703}_{bu}$ for various Ra and $\mathit{Pr}$ versus pattern wavenumber $k_{x}$ . (b) Temperature below the insulating stripes $\unicode[STIX]{x1D703}_{in}$ , averaged over the entire insulating area, for different $Ra$ and $\mathit{Pr}$ against pattern wavenumber $k_{x}$ .

Figure 5(a) shows the average bulk temperature of the fluid plotted against the wavenumber. This average was computed over the horizontal plane at mid-height of the system. The effect is similar to what we see for $Nu/Nu_{fc}$ in figure 4(b). At the lowest $k_{x}$ the average bulk temperature has increased to about $2/3$ . In that case the top plate is split in half and only one half is contributing to heat transfer. While ignoring the adiabatic area, we can divide the conducting areas into three equally sized parts, with two parts on the bottom plate and one on the top plate. Using the same reasoning as that used for the symmetric case, we obtain the following equality for the average bulk temperature: $\unicode[STIX]{x1D703}_{bulk}=(2/3)\unicode[STIX]{x1D703}_{bottom}+(1/3)\unicode[STIX]{x1D703}_{top}=2/3$ . As with $Nu/Nu_{fc}$ , we see that the average bulk temperature approaches the fully conducting case of $\unicode[STIX]{x1D703}_{bulk}=1/2$ for increasing $k_{x}$ . When comparing the various curves, they all appear similar, with a maximum difference of about 0.02–0.03 in the average bulk temperature. Unfortunately, Wang et al. (Reference Wang, Huang and Xia2017) do not report their average temperature and a comparison to their calculations with large-wavelength imperfect BCs is impossible.

In figure 5(b) we show the horizontally and time-averaged temperature below the insulating area of the top plate as a function of the wavenumber. At the lowest wavenumber (top plate split into equal conducting and insulating regions), we see that the averaged temperature is almost equal for all $Ra$ and $\mathit{Pr}$ . When increasing $k_{x}$ , a dependence on $Ra$ emerges. After just a few additional divisions in stripes, all curves order themselves according to $Ra$ , with the lower $Ra$ values approaching the lower bounds faster than the larger ones. At the largest $k_{x}$ , the temperature difference between $Ra=10^{9}$ and $Ra=10^{7}$ is 15 %. As the temperature below the insulating area is slightly higher than for the area below the conducting area because of lack of cooling, we can conclude that, for larger $Ra$ , the whole top layer is, on average, hotter than for the lower  $Ra$ .

These results suggest that it could be possible to account for the $Nu(k_{x})$ relationship by using corrected non-dimensional variables in the spirit of Verzicco (Reference Verzicco2004). However, changing the (effective) thermal conductivity ignores the heterogeneities of the plate, which is the main source of the observed behaviour. In figure 6, we show the corrected Nusselt number $Nu^{\ast }=Nu/\unicode[STIX]{x1D703}^{\ast }$ against the corrected Rayleigh number $Ra^{\ast }=Ra\unicode[STIX]{x1D703}^{\ast }$ , where $\unicode[STIX]{x1D703}^{\ast }$ is the non-dimensional average temperature difference given by $\unicode[STIX]{x1D703}^{\ast }=L_{p2}\unicode[STIX]{x1D703}_{in}/L_{p}$ . No logical ordering can be seen, and, as expected, an attempt to describe the results with some global effective thermal conductivity fails.

Figure 6. (a) Corrected Nusselt number against compensated Rayleigh number for selected pattern wavelengths. (b) Compensated and corrected Nusselt number against corrected Rayleigh number. No collapse or natural ordering of the curves can be seen.

For horizontal slices of the instantaneous temperature field, a discrete 2D Fourier transform can be applied, which is defined as

(3.1) $$\begin{eqnarray}\unicode[STIX]{x1D703}(\unicode[STIX]{x1D6FE})=\mathop{\left|\mathop{\left\langle \mathop{\sum }_{y=0}^{n_{y}-1}\mathop{\sum }_{x=0}^{n_{x}-1}\exp \{-\text{i}[(2\unicode[STIX]{x03C0}/n_{x})j_{x}x+(2\unicode[STIX]{x03C0}/n_{y})j_{y}y]\}\unicode[STIX]{x1D703}(x,y,t)\right\rangle }\nolimits_{t}\right|}\nolimits_{j_{x}=\unicode[STIX]{x1D6FE},j_{y}=0},\end{eqnarray}$$

where $\langle \cdot \rangle _{t}$ is the time average and the $j_{y}$ mode was set to zero, leaving a single wavenumber parameter $\unicode[STIX]{x1D6FE}\equiv j_{x}$ . Using the described method on the horizontal slice just below the top boundary, we can identify the imprint the stripe-structured BCs leaves on the flow.

Figure 7. Premultiplied 2D Fourier transform of a horizontal slice at various distances from the top boundary wall, averaged in time. Except for (d), each panel is compiled using a single dataset. The colours help to identify the different modes that are present in each single dataset and are identified by the odd or even value. (a) At the closest grid point ( $\hat{\boldsymbol{z}}=0.998$ ), the two different modes are clearly visible. (b) One grid point further ( $\hat{\boldsymbol{z}}=0.996$ ), the distinction slowly fades. (c) While still inside the boundary layer at $\hat{\boldsymbol{z}}=0.992$ , both modes are practically overlapping. (d) Just outside the boundary layer, it is impossible to distinguish two different modes at all. As a reference, we also show the spectrum at the centre of the system ( $\hat{\boldsymbol{z}}=0.5$ ).

Using the Fourier transform we can identify the imprint of the BCs just below the top boundary in the flow itself. Using this distinct signature we can find out how long this pattern is still present once one moves away from the boundary wall. The distance from the top boundary wall is indicated using $\hat{\boldsymbol{z}}$ . In figure 7, we see the compensated spectra for $f=1$ at five different planes for decreasing $\hat{\boldsymbol{z}}$ . The colours are used to identify the different modes and except for figure 7(d) are compiled using a single dataset (odd or even value). When moving away from the wall, we see that the two distinct modes approach each other and just outside the boundary layer, at $\hat{\boldsymbol{z}}=0.970$ , it is hard to distinguish the two different curves at all. Within the boundary layer the signature of the pattern almost completely fades away and in the bulk flow is not visible at all. The difference between the Fourier transform just outside the boundary layer ( $\hat{\boldsymbol{z}}=0.970$ ) and at mid-height of the system ( $\hat{\boldsymbol{z}}=0.5$ ) is marginal. These findings hold for the complete range of  $f$ .

It is quite remarkable that, even for the most extreme case at $f=1$ , the pattern is not visible in the bulk region. This means that in the boundary layer, in which conduction dominates, the temperature differences of the top plate are averaged such that an effective, slightly higher cold plate is seen by the bulk flow.

Figure 8. (a) Comparison of the normalized Nusselt number $Nu/Nu_{fc}$ for the single- and double-sided cases with $Ra=10^{8}$ , $\mathit{Pr}=1$ and $\ell _{C}=0.5$ . For the lowest $k_{s}$ we see that the system is about two-thirds and one-half of the fully conducting case for the single- and double-sided cases, respectively. (b) Average temperature above or below the insulating stripes for the single- and double-sided cases with $Ra=10^{8}$ , $\mathit{Pr}=1$ and $\ell _{C}=0.5$ . The dashed line shows the average bulk temperature, calculated at mid-height for the double-sided system.

3.2 Patterns on both plates

Until now we have only applied the insulating and conducting patches to the top boundary. This resulted in a normalized heat transfer of approximately two-thirds for the lowest $k_{x}$ and almost the fully conducting case at the highest $k_{x}$ . Using a simple argument we could indeed rationalize the value of two-thirds for the normalized heat transfer for the lowest wavenumber. If we now apply the same pattern also on the bottom plate, can we still get to the same efficiency as if we had only applied the pattern to the top boundary wall?

The comparison of the normalized Nusselt number $Nu/Nu_{fc}$ between the single- and double-sided cases is shown in figure 8(a). For the lowest $k_{x}$ we see that the double-sided case conducts heat at approximately half the rate of the fully conducting case. This is in line with the expectations, as we only have half the effective area on both boundary walls. What is also visible is that the curve for the double-sided case is steeper than the curve of the single-sided case, therefore reducing the difference between both cases with increasing $k_{x}$ . For the largest $k_{x}$ the heat transfer of this system again is almost as efficient as if it were fully conducting. There is only half of the area available for heat entering the system and only half the area for heat leaving the system. Still, the same amount of heat transfer as if the system were fully conducting is achieved. For the single-sided case, as for the double-sided one, for the lowest $k_{x}$ the efficiency is not exactly $2/3$ and $1/2$ but slightly above these values. The geometry of the double-sided case can be decomposed into a regular RB cell and a neutral domain, both with identical dimensions, positioned next to each other. The top and bottom boundaries from this neutral domain are both insulating and heat can only enter and exit from the sides. As we have periodic BCs in the horizontal direction, both sides of the regular RB area are connected to this neutral domain, which acts as a buffer for heat. This extra buffer is the only difference and thus the cause for the small difference.

The average temperatures just below or above the insulating boundaries are shown against $k_{x}$ in figure 8(b). The difference in temperature below the top insulating boundary between the single-sided case, shown in green circles, and the double-sided case, shown in orange squares, is only significant at the lowest $k_{x}$ . Only at the lower $k_{x}$ is the single-sided system hotter, just below the boundary. At larger $k_{x}$ , the asymmetry does not make a difference to the temperatures and both temperatures converge to the conducting plate temperature. In the same plot, we also show the temperature just above the insulating area of the bottom plate and the bulk temperature. For the double-sided case, for the lowest $k_{x}$ , the top, bottom and bulk temperatures are very similar. This indicates that for the lowest $k_{x}$ all the fluid in the neutral domain, the large area confined by the insulating bottom and top plates, has approximately the same temperature and hardly contributes to the heat transfer. This fully agrees with $Nu/Nu_{fc}\approx 0.5$ seen in figure 8(a). The bulk temperature of the double-sided case stays approximately 0.5 for the full range of $k_{x}$ , as must hold for a symmetric system. Temperatures above the bottom boundary and below the top boundary are also symmetric with respect to the bulk temperature, confirming the statistical convergence of our calculations. As for the single-sided case, figure 8(b) shows that for the largest $k_{x}$ the temperatures close to the insulating areas are very close to their conducting counterparts. The heat conduction in the boundary layer makes the bulk fluid see an almost perfect heat conductor.

3.3 Variation of the insulating fraction

All the systems that we have discussed until now had a conducting area with size equal to the insulating area, i.e. $\ell _{C}=1/2$ . We can now vary $\ell _{C}$ and look into its effect on the heat transfer. In the previous simulations we used $k_{x}$ as the wavenumber and this sets the number of insulating and conducting stripes in the 2D case. The width of the system, $L_{x}$ , was divided into $f$ equal pairs of these stripes. In the previous subsection, these divisions were of equal areas. Now we will change the ratio of areas to make the top plate less and less conducting. For these simulations we fixed $k_{x}=9$ , $Ra=10^{8}$ and $\mathit{Pr}=1$ . Then the ratio between the insulating and conducting areas was varied, namely we simulated $\ell _{C}=1.0,~0.875,~0.75,~0.625,~0.50$ and $0.25$ , thus gradually reducing the conducting area of the top plate from fully conducting to 25 % conducting.

Figure 9 shows the results of the simulations, as well as the data from the rectangular tank of the experiments in Wang et al. (Reference Wang, Huang and Xia2017) (extrapolated from their table 1). In the case where only 12.5 % of the area is insulating, we see that the difference with the fully conducting case is almost negligible and within the statistical error. On increasing the amount of insulating area to 25 %, the heat transfer is still more than 90 % of the fully conducting case. Even at 50 % conducting region we still get the effectiveness of 80 %. At the largest ratio of 75 % insulating fraction, we are still above 60 % of the heat transfer of the fully conducting case. In other words, the effective area of the top plate is only a quarter of the fully conducting case but still we get a system that is only 40 % less effective in conducting heat.

The rectangular tank of Wang et al. (Reference Wang, Huang and Xia2017) has patches with dimensions comparable to the system size. Two data points for the two different wavelengths are available for each $\ell _{C}$ . The data point with a smaller wavelength (‘ACA’ and ‘CAC’ patterns) corresponds to the higher values of $Nu$ . While the points at $\ell _{C}=1/3$ show a considerably lower $Nu$ than the DNS with a much smaller wavelength, a relatively good match between DNS and the experiment for $\ell _{C}=2/3$ provides some indication that at higher values of $\ell _{C}$ the saturation wavenumbers, for which $Nu\approx Nu_{fc}$ , are smaller.

Figure 9. (a) Normalized Nusselt number $Nu/Nu_{fc}$ for various fractions $\ell _{C}$ of the conducting plate, while keeping the pattern frequency $f=9$ , $Ra=10^{8}$ and $\mathit{Pr}=1$ (green circles), and the rectangular tank of Wang et al. (Reference Wang, Huang and Xia2017) (orange squares). (b) Average temperature just below the insulating area $\unicode[STIX]{x1D703}_{in}$ and average bulk temperature $\unicode[STIX]{x1D703}_{bu}$ , both plotted against $\ell _{C}$ . Other parameters are the same as for panel (a).

Figure 9(b) shows two different temperatures, namely, the average temperature just below the insulating region $\unicode[STIX]{x1D703}_{in}$ and the average bulk temperature $\unicode[STIX]{x1D703}_{bu}$ measured at mid-height of the system. For $\ell _{C}=0.85$ , $\unicode[STIX]{x1D703}_{in}$ is very close to zero, i.e. the top wall temperature. This is consistent with $Nu$ nearly having the value of the conducting case (figure 9 a). Also the bulk temperature is very close to the fully conducting case $\unicode[STIX]{x1D703}=0.5$ . When $\ell _{C}$ is decreased, making the top plate less and less conducting, $\unicode[STIX]{x1D703}_{in}$ increases gradually, reaching $\unicode[STIX]{x1D703}_{in}=0.5$ when $\ell _{C}=0.25$ . This equals the bulk temperature in a fully conducting system. However, when $\ell _{C}$ is decreased, also the bulk temperature gradually increases and nearly reaches 0.6 for $\ell _{C}=0.25$ . This rise in the bulk temperature is much slower than the rise in the temperature above the insulating area, meaning that the gradient between insulating regions at the plate and the bulk decreases and thus so does the heat transfer (see figure 9 a). Even though these simulations were conducted using only $f=9$ , one expects similar trends to apply for other values of $f$ . From the previous simulations we found that, when increasing $f$ , $Nu/Nu_{fc}$ will rise and $\unicode[STIX]{x1D703}_{ins}$ will decrease. The same response can be achieved by increasing $\ell _{C}$ as we increase the conducting area and approach the case of the fully conducting system.

3.4 Mixed insulating and conducting patterns in two dimensions

Until now, all the patterns that have been applied to the top and bottom boundary walls were one-dimensional (1D) stripe-like patterns. We only varied the width of the patches and the ratio between the insulating and conducting fractions. In this subsection we add an additional spatial dependence to these patterns to make them chequerboard-like:

(3.2) $$\begin{eqnarray}\left.\begin{array}{@{}c@{}}\unicode[STIX]{x1D703}(\hat{x},{\hat{y}},\hat{\boldsymbol{z}}=1)=0,\quad \hat{x}\in [iL_{px},L_{px1}+iL_{px}],~{\hat{y}}\in [jL_{py},L_{py1}+jL_{px}],~i,j\in \mathbb{Z},\\ \unicode[STIX]{x2202}_{z}\unicode[STIX]{x1D703}(\hat{x},{\hat{y}},\hat{\boldsymbol{z}}=1)=0,\quad \hat{x}\notin [iL_{px},L_{px1}+iL_{px}],~{\hat{y}}\notin [jL_{py},L_{py1}+jL_{px}],~i,j\in \mathbb{Z},\\ \unicode[STIX]{x1D703}(\hat{x},{\hat{y}},\hat{\boldsymbol{z}}=0)=1,\quad \forall ~\hat{x},{\hat{y}}.\end{array}\right\}\end{eqnarray}$$

A schematic of a set of four patches, two insulating and two conducting, is shown in figure 10. The dimensions of both types of patches were kept equal, i.e. $L_{px1}=L_{py1}=L_{px2}=L_{py2}$ , meaning $\ell _{c}=1/2$ .

Figure 10. Top view of the top plate BCs with applied chequerboard pattern ( $f=1$ ). Here $L_{px}$ and $L_{py}$ are the horizontal and vertical dimensions of a set of patches. The set itself is divided into two insulating (white) and two conducting (black) areas, all with equal dimensions: $L_{px1}=L_{py1}=L_{px2}=L_{py2}$ .

As we took both horizontal dimensions to be equal, we define a single frequency $f$ in both directions. The plate is divided into $f$ sets of patches, which have dimensions $L_{px}=L_{py}=L_{p}$ . When $f=1$ the complete boundary consists of a single set of patches. By increasing $f$ to 2, the plate will consist of four sets of four patches each. To give a further impression how the temperature fields resulting from these BCs look, two visualizations of instantaneous temperature fields for $f=4$ and $f=20$ are shown in figure 11. The visualizations show respectively 16 and 400 sets of patches, which each consist of two insulating and two conducting areas.

Figure 12(a) shows the normalized Nusselt number $Nu/Nu_{fc}$ for the 1D (stripe) and 2D (chequerboard) patterns as a function of $k_{x}$ . For these cases, $Ra=10^{8}$ , $\mathit{Pr}=1$ and the error bars indicate the statistical convergence error. At first sight, the heat transfer is not that different when applying 1D or 2D patterning. At the lowest division, where we have divided the system into two or four areas for the 1D and 2D case, respectively, the difference is negligible. Also at the highest value of $k_{x}$ , the difference is inside statistical errors and seems insignificant.

Figure 11. Two 3D visualizations of the instantaneous temperature field with different pattern frequencies at the top boundary. Hot fluid is shown in red while the cool fluid has a blue colour. For both visualizations, $Ra=10^{8}$ and $\mathit{Pr}=1$ . (a) Visualization with $f=4$ , which results in 16 sets of patches each containing two insulating patches and two conducting patches. (b) Visualization with $f=20$ , which results in 400 sets of patches. Hot plumes rise from the bottom plate while cold plumes are ejected from the top boundary.

Figure 12. (a) Normalized Nusselt number $Nu/Nu_{fc}$ for 1D and 2D patterning versus $k_{x}$ for $Ra=10^{8}$ and $\mathit{Pr}=1$ . The error bars show the statistical convergence error. (b) Average temperature $\unicode[STIX]{x1D703}_{in}$ just below the insulating area and average bulk temperature $\unicode[STIX]{x1D703}_{bu}$ , both as a function of $k_{x}$ for the 1D and 2D patternings. The $Ra$ and $\mathit{Pr}$ values are the same as for panel (a).

At first glance, the 1D stripe pattern and the 2D chequerboard pattern may seem idealized representations. However, all other stripe and rectangular chequerboard patterns fall within these extreme cases, as the stripe pattern has $L_{px}/L_{py}\rightarrow \infty$ (or conversely $\rightarrow 0$ ), while for the chequerboard pattern $L_{px}/L_{py}=1$ . Our work shows that, for the relatively small wavelengths considered here, the shape of these patterns does not have a significant influence on the flow dynamics. While we do not expect $\ell _{C}$ to affect this statement considerably, larger wavelength patterns could show some dependence on their shape. In the experiments of Wang et al. (Reference Wang, Huang and Xia2017), the patches have wavelengths comparable to the system height and their distribution considerably affects the flow structure. Here, we consider wavelengths that are much smaller, affecting the flow only below the thermal boundary layer thickness, and reducing the Nusselt number. The largest wavelength considered in this work is still smaller than the smallest wavelength in the rectangular tank of Wang et al. (Reference Wang, Huang and Xia2017). There appear to be two clear regimes: the large patch regime, which shows a clear influence on the flow dynamics and whose effect is likely to be shape-dependent, and the small patch regime, which only affects the boundary layers and thus the heat transfer. While there must exist a cross-over regime between the two, it appears to be just outside the wavelengths we are considering.

When we look at $\unicode[STIX]{x1D703}_{in}$ , the temperature just below the insulating region in figure 12(b), we see a similar trend. The temperature below the insulating region is on average always lower for the 2D pattern. This is even the case for the higher and lower limits of $k_{x}$ . The heat can escape towards a conducting region in any direction, while for stripes it can only escape in one direction. Thus, the distance to a sink is smaller and thus the temperature is smaller. The bulk temperatures for both the 1D and 2D cases are almost identical, which means that the change in pattern has no effect on this quantity. From these results we can conclude that the impact of the two different patterns is very similar, and the quantitative differences between stripe and chequerboard patterns are at most small.