3.1. Overview of flow through an array of cylinders

In this section, an overview of 3-D flow through an array of cylinders is presented for the four wake regimes observed in He et al. (Reference He, Draper, Ghisalberti, An, Branson, Cheng and Ren2022, Reference He, Ghisalberti, An, Draper, Branson, Ren and Cheng2024b). In presenting this, comparison with results from 2-D simulations is made, which provides insights into the applicability of 2-D DNS in modelling flow through an array of cylinders.

The four wake regimes are: (i) the coupled individual wake (CI), in which the array shear layers are stabilised, and the array wake behind the array is dominated by the element-scale wakes of individual cylinders (figure 5a,e); (ii) the Kelvin–Helmholtz instability wake (KH), where the array shear layers are susceptible to Kelvin–Helmholtz instability and form two rows of KH vortices (figure 5b,f); (iii) the ‘steady + shedding’ wake (SS), in which the two array shear layers remain independent in the steady region, before interacting to result in a von Kármán vortex street downstream (figure 5c,g); and (iv) the vortex street wake (VS), where a von Kármán vortex street forms immediately behind the array (figure 5d,h). As the flow transitions from regime CI through to VS, the formation of element-scale vortices in the wake of individual cylinders is progressively suppressed within the array, coinciding with the increasingly dominant array-scale vortices developed behind the array.

Figure 5.

Flow characteristics for four regimes of the array-scale wake. (a–d) Cross-sectional instantaneous flow field (at z/D = 0) from 3-D DNS, visualised by spanwise vorticity ${\omega _z}d/{U_\infty } = [ - 0.4,0.4]$ for (a) coupled individual wake (CI), (b) Kelvin–Helmholtz instability wake (KH), (c) steady + shedding wake (SS) and (d) vortex street wake (VS). Isosurfaces of ${\omega _x}d/{U_\infty } ={\pm} 0.5$ (coloured by ${\omega _x}$) for (e) CI, (f) KH, (g) SS and (h) VS. The four cases in regimes CI, KH, SS and VS have G/d = 8, 4.5, 2.5, 1.2 and N = 31, 109, 31, 31, respectively. As the flow transitions from regime CI through to VS, the vortex structures behind the array progressively evolve from the element scale to the array scale, and the flow exhibits 3-D features largely behind the array rather than within the array.

Figure 5 shows that the flow exhibits 3-D features and that 3-D flow structures can arise behind individual cylinders or the array depending on the wake regime. For instance, in regime CI (see figure 5a,e), it is seen that 3-D vortices develop in the wakes of cylinders at either side of the array. These vortices will merge and be dissipated rapidly in the near wake of the array as they are convected downstream. As the flow transitions into regime KH (see figure 5b,f), whilst there is no 3-D flow structures forming within the array, the array-scale streamwise vortices form when the two rows of KH vortices, associated with each array shear layer, interact with each other to form a vortex street further downstream. In comparison to regimes CI and KH, the flow in regime SS exhibits relatively weak 3-D flow features (figure 5c,g), occurring where the two array shear layers interact to form vortex shedding. Finally, in regime VS, the porous array behaves similarly to an isolated solid cylinder in terms of the formation of a 3-D von Kármán vortex street immediately behind the array (figure 5d,h).

The distribution of three-dimensionality in the system can be further illustrated by examining the time-mean spanwise kinetic energy, which is defined as

(3.1)\begin{equation}{E_z} = \frac{1}{T}\int_0^T {\frac{1}{2}{w^2}/U_\infty ^2\,\textrm{d}t} ,\end{equation}

where T (5000 time units) is the sample duration associated with the fully developed flow and ${w^2}/U_\infty ^2$ is post-processed on a spanwise-averaged flow field. This quantity has been used previously to check the three-dimensionality in the cylinder wake (e.g. Tong et al. Reference Tong, Cheng, Zhao, Zhou and Chen2014; Jiang et al. Reference Jiang, Cheng, Draper, An and Tong2016).

Whilst in regime CI, non-zero ${E_z}$ values are distributed behind the array and around a few cylinders in the rear of the array (figure 6a). Alternatively, in regimes KH, SS and VS non-zero ${E_z}$ values are only observed behind the array (figure 6b–d). Clearly, the flow exhibits 3-D features largely behind rather than within the array. This is consistent with figure 5 where the three-dimensionality is visualised by isosurfaces of streamwise vorticity.

Figure 6.

Distribution of time-mean spanwise kinetic energy ${E_z}$ (defined in (3.1)) for an array of 31 cylinders. The four cases are the same as in figure 5: (a) CI, (b) KH, (c) SS and (d) VS. Non-zero ${E_z}$ values are mostly distributed behind rather than within the array, indicating the development of three-dimensionality predominantly behind the array.

Although 3-D structures form within the array in regime CI (figures 6a and 5a), the three-dimensionality of the element-scale wakes is relatively weak at $R{e_d} = 200$, and this is expected given that this Reynolds number is close to the critical Reynolds number (194) for onset of three dimensionality for an isolated cylinder (Williamson Reference Williamson1996; Jiang et al. Reference Jiang, Cheng, Draper, An and Tong2016). Furthermore, at $R{e_d} = 200$ the maximum difference in the average drag coefficient $\overline {{C_d}} $ for two tandem cylinders across the range of G/d = 1.5–8 is only 1 % (Koda & Lien Reference Koda and Lien2013), which suggests limited influence of element-scale three-dimensionality on the element drag. More importantly, studies have found that around 90 % of the total mean drag is associated with the shear layers attached to the cylinder surface, with less dependence on the flow structure in the far wake (e.g. Fiabane, Gohlke & Cadot Reference Fiabane, Gohlke and Cadot2011). This suggests that the array-scale three-dimensionality behind the array, far from most of element wakes, has limited influence on the element drag.

The influence of three-dimensionality on the drag of a single cylinder near a wall is limited when ${E_z}$ has magnitude less than or equal to 10⁻² (Jiang et al. Reference Jiang, Cheng, Draper and An2017). In the present study, each cylinder is confined by surrounding cylinders. Figure 6 shows that the magnitude of ${E_z}$ within the array is between 10⁻³ and 10⁻² for all the regimes. With reference to Jiang et al. (Reference Jiang, Cheng, Draper and An2017), the three-dimensionality in the system should therefore have limited influence on the total drag force of the array.

To further confirm the applicability of 2-D simulations, a comparison of velocity profiles is made between 3-D and 2-D simulations in figure 7. In the streamwise direction (figure 7a–d), for all the regimes, the streamwise velocity decreases with x/D in the upstream of the array and then continuously decreases throughout the array though some velocity fluctuations are observed due to the flow heterogeneity induced by individual cylinders; downstream of the array, the velocity shows complex variation but eventually increases towards the upstream velocity ${U_\infty }$. As the flow transitions from regime CI to VS, the transverse profile indicates a larger wake deficit in the array wake (at x/D = 1) (figure 7e–h).

Figure 7.

Profiles of time-mean streamwise velocity along $y = 0$ in 3-D simulations compared with 2-D results: (a) CI, (b) KH, (c) SS and (d) VS. Profiles of time-mean streamwise velocity along $x/D = 1$: (e) CI, (f) KH, (g) SS and (h) VS. Whilst there is good agreement of velocity profiles upstream of and within the array, there are differences in the downstream velocity profiles due to the generation of 3-D flow structures.

Upstream of and within the array $(x/D \le 0.5)$, there is a good agreement between the streamwise velocity profiles (figure 7a–d), with $\textrm{RMS}{\textrm{E}_\varOmega }$ of $\bar{u}/{U_\infty }$ between 3-D and 2-D simulations of 0.2 %, 0.3 %, 0.8 % and 3.3 % for the four regimes, respectively. This suggests that the flow upstream of and within the array modelled in the 2-D simulation is representative of that in the 3-D simulation.

Downstream of the array (x/D > 0.5), the level of agreement between 3-D and 2-D results depends on the wake regime. In regime CI, the profiles in 3-D and 2-D simulations are in excellent agreement for both streamwise and transverse profiles (see figure 7a,e), respectively with $\textrm{RMS}{\textrm{E}_\varOmega } = 1.8\,\%$ and 1.0 %. In regimes KH and SS, 3-D and 2-D transverse profiles agree reasonably well (see figure 7f,g), both with $\textrm{RMS}{\textrm{E}_\varOmega }$ less than 3.0 %. However, the 3-D and 2-D streamwise profiles show a similar trend but start to deviate in the KH and SS regimes, with higher $\textrm{RMS}{\textrm{E}_\varOmega } = 12.7\,\%$ and 31.3 %, respectively (figure 7b,c). In regime VS (figure 7d,h), both the 3-D and 2-D streamwise and transverse profiles start to deviate, with $\textrm{RMS}{\textrm{E}_\varOmega } = 33.4\,\%$ and 22.4 %, respectively.

The differences in values of $\bar{u}/{U_\infty }$ downstream of the array are generated due to the following reasons. Firstly, due to missing turbulent diffusion in the 2-D simulation, the 2-D vortex structures in the array wake persist in strength as they travel downstream (not shown) and hence cause stronger wake entrainment. The streamwise velocity is, in turn, recovered in a faster rate relative to that in the 3-D simulation. This faster rate becomes obvious when large-scale 3-D vortex structures are generated at the wake centreline (comparing figures 7b–d and 5b–d). Furthermore, without vortex stretching 2-D array shear layers are much stronger and interact to form vortex shedding closer to the array (not shown), which hence induces wake entrainment earlier than in the 3-D simulation. This is consistent with earlier recovery of $\bar{u}/{U_\infty }$ in the streamwise profile (see figure 7c,d).

The focus of the present study is arrangement effects on the bulk flow through and bulk drag on the array, which are associated with the variation of element-scale flow and drag characteristics with arrangement (demonstrated later). Therefore, no further attempt was made to demonstrate the differences in the array wake between 2-D and 3-D simulations.

The discussions based on figures 5–7 as a combination suggest that 2-D simulations may be a reasonable tool that helps us increase the resolution of the parameter space in investigating the arrangement effects on the drag on the array elements and the bulk velocity through the array, though caution is needed in interpretation of 2-D array-scale wake.

3.2. Arrangement effects: gap ratio and array-to-element diameter ratio

To interpret the arrangement effects at the element scale, a first step is to understand the element-scale flow and drag characteristics within a finite circular array. Since the array is a combination of multiple lines of cylinders, the scenario of flow past a line of cylinders is studied first.

The key findings of flow past a single line of cylinders from existing literature (Hosseini et al. Reference Hosseini, Griffith and Leontini2020; Zhu et al. Reference Zhu, Zhong and Zhou2021; Eizadi et al. Reference Eizadi, An, Zhou, Zhu and Cheng2022) can be summarised as: (i) the flow evolution along the line is chiefly controlled by G/d when $R{e_d} \le 200$ and $N \ge 3$; (ii) the flow transitions from shear layer reattachment (SLR) and two-row structure (TRS) occurring at $G/d \approx 3.7$; and (iii) SLR and TRS can cause significant drag reduction to the elements covered by these characteristic flow structures.

Two typical cases at G/d = 2 and 4.5 with N = 11 from 3-D simulations are used to illustrate the two characteristic flow structures and drag reduction mechanisms associated with them (see cross-sectional flow fields in figure 8). The flow for G/d = 2 remains 2-D, while the flow is 3-D for G/d = 4.5 (details are not shown). For G/d = 2 (see figure 8a) the two (positive, negative) separated shear layers from the upstream cylinder reattach on the downstream cylinder from C2 to C3 and SLR develops downstream into regular vortex shedding; these vortices develop a unique pattern downstream where the positive and negative vortices are separated into two parallel rows spanning from C8 to C11, i.e. TRS. For G/d = 4.5 (figure 8b), von Kármán vortices develop from C1; these vortices develop TRS downstream spanning from C2 to C4. Further downstream the flow becomes chaotic.

Figure 8.

Flow and drag characteristics of a single line of 11 cylinders for θ = 0° in 3-D DNS. Instantaneous cross-sectional (z/D = 0) flow fields for G/d = 2 (a) and G/d = 4.5 (b). (c) Distribution of drag coefficient along the line. Cylinders with TRS or SLR are marked in dark grey in (a–c).

In the wake of a single cylinder, the primary von Kármán vortices were found to develop downstream into TRS when L_y/L_x > 0.365, where L_y is the cross-wake spacing between centres of positive and negative wake vortices and L_x is the distance between centres of adjacent vortices of the same sign (vortex centre is defined by the maximum of spanwise vorticity ${\omega _z}d/{U_\infty }$). This threshold has been theoretically identified based on inviscid theory and validated through both experiments (Durgin & Karlsson Reference Durgin and Karlsson1971; Karasudani & Funakoshi Reference Karasudani and Funakoshi1994) and numerical simulations (Thompson et al. Reference Thompson, Radi, Rao, Sheridan and Hourigan2014). In the present study, TRS is similarly developed from primary von Kármán vortices from C1 in a line. Therefore, this threshold L_y/L_x > 0.365 is adopted here to quantitatively classify TRS, with the two length scales L_y, L_x illustrated in figure 8(a).

The distribution of drag coefficient along the line for the two gap ratios are plotted in figure 8(c). It is seen that the cylinders within either SLR (C2, C3, C8–C11 marked in dark grey in figure 8a) or TRS (C2–C4 in figure 8b) have very small drag coefficients relative to other cylinders without these typical flow structures.

Figure 9 illustrates how the characteristic element-scale wake varies with increasing G/d (and hence D/d) for a circular array with a fixed total number of cylinders (N = 109) in 3-D DNS. The element-scale wake structure within a multiple-line array exhibits three flow states: (i) SLR develops for a number of cylinders within an array; (ii) the element-scale wake in the array is characterised by TRS, which covers at least two cylinders within a line; and (iii) over the entire array no downstream cylinders are within a region of TRS or SLR (termed ‘non-covered’, or NOC, here). Two cases with SLR are presented in figure 9(a,b). For (G/d, D/d) = (2, 22.2) (figure 9a), SLR fully occupies the three lines of cylinders in the middle of the array (lines 1⁻, 0, 1⁺). The SLR is progressively broken down as the array edges are approached (figure 9a). The resultant wakes of individual cylinders remain steady but form a staggered pattern. The formation of the staggered pattern is attributed to flow diversion towards either side of the array, as indicated by diverging streamlines (figure 9a). At larger G/d = 2.5, SLR can develop downstream into vortex shedding (lines 2⁻, 1⁻, 0, 1⁺, 2⁺ in figure 9b). With further increases in G/d, it is seen that TRS develops for a number of cylinders (figure 9c). The TRS covers from C3 to C5 in four lines (the filled circles in lines 2⁻, 1⁻, 1⁺, 2⁺) and up to C6 in the middle line (line 0). Finally, NOC is observed for the array with the largest G/d = 8 (figure 9d), where TRS is confined between C2 and C3 in the three lines (0, 1⁻, 2⁻) without covering any downstream cylinder in each line. With increasing G/d, SLR or TRS covers a smaller number of cylinders within a line (see figure 9a–d).

Figure 9.

Cross-sectional instantaneous flow field (at z/D = 0) for arrays of 109 cylinders with various gap ratios in 3-D DNS. Cylinder lines are numbered, and the cylinders covered by TRS are marked in dark grey in (c,d), which is classified based on the criterion L_y/L_x > 0.365. With increasing G/d, the extent of elements covered by SLR or TRS over the array is reduced.

Flow transition processes similar to those of a single line of cylinders are observed in multiple-line arrays. A transition from SLR to vortex shedding is seen in both a single line and an array for small gap ratio (figures 8a and 9b); for large gap ratio, a downstream transition from primary vortex to secondary vortex and eventually to a chaotic wake is observed in both the single line and the array. Comparisons of spectra of lift coefficient combined with the instantaneous flow field for the array (G/d = 8, N = 109) with those of the counterpart single-line case are used to demonstrate this (see figure 10). The flow and lift spectra for different lines within the array show similar evolution, such that only lines 0 and 4⁻ are presented for demonstration. The primary frequency dominates the first two cylinders in each line within the array (figure 10a ₁,a ₂,c ₁,c ₂). This shows that the von Kármán vortex shed from C1 governs the frequency of the TRS occurring behind C2 (see figure 10b). The secondary frequency emerges on the spectrum of C2 and then becomes dominant over C3 to C6 (figure 10a ₃–a ₆,c ₃–c ₆). Further downstream, no dominant frequency is observed (figure 10a ₈,c ₈), which characterises the chaotic flow (see figure 10b,d). Such downstream evolution of lift spectrum and vortex structures is similar to that for the single line (see figure 10e ₁–e ₈,f), demonstrating the similarity of flow evolution within the array to that for a single line of cylinders.

Figure 10.

Similarity of downstream evolution of the lift coefficient spectrum and instantaneous flow between an array of 109 cylinders and a single line of 11 cylinders (both at G/d = 8) in 3-D DNS. Here ‘P’ and ‘S’ represent primary and secondary frequencies, which are indicative of the element-scale primary vortex and secondary vortex, respectively.

Similarity is also seen in the distribution of drag coefficients of cylinders along the line. This is demonstrated in figure 11 for an array of cylinders with N = 109, G/d = 4.5, for which the instantaneous flow field is shown in figure 9(c). It is seen that the drag distributions along lines of cylinders are very similar to that of the single line especially in the middle of the array (comparing figure 11 with figure 8c), with a sharp decrease of $\overline {{C_{d,i}}} $ through the first three cylinders, followed by an increase and a decrease over downstream cylinders.

Figure 11.

Distribution of drag coefficient of cylinders within an array of cylinders with N = 109, G/d = 4.5 in both 3-D and 2-D DNS. The cross-sectional instantaneous flow field for this case is shown in figure 9(c). The distributions of $\overline {{C_{d,i}}} $ along the line of cylinders within the array in both 3-D and 2-D simulations are similar to that of the single line of cylinders shown in figure 8(c), especially in the middle of the array, suggesting the applicability of using 2-D DNS in characterising the element drag.

The variation of $\overline {{C_{d,i}}} $ along each row of cylinders within the array in 2-D simulations follows the trend of that in 3-D simulations. Values of $\overline {{C_{d,i}}} $ for some cylinders in 3-D simulations are slightly lower than those in 2-D simulations. For all the cases in the present study, the $\textrm{RMS}{\textrm{E}_\varOmega }$ for $\overline{C_{d, i}}$ is less than 8 %. This level of discrepancy of $\overline {{C_{d,i}}} $ between 3-D and 2-D simulations is comparable to that of flow past two tandem cylinders (e.g. Reference Koda and LienKoda & Lien 2013). This demonstrates the minor effect of using the 2-D assumption to simulate the flow within the present scope.

Despite the similarity, the presence of the adjacent lines of cylinders in proximity in arrays makes the flow evolution along a line within the array more complicated than that along the single line. The adjacent lines have two effects on the flow: (i) adjacent lines behave similarly to bounding walls (Zdravkovich Reference Zdravkovich1997; Sahin & Owens Reference Sahin and Owens2004) to confine the flow motions in the cross-flow dimension (referred to herein as a confinement effect), and (ii) vortices shed from adjacent lines interact with each other (vortex interaction), destroying the regular vortex structures. The confinement effect is illustrated by comparing the instantaneous flow field between the single line and the lines within the array for the same G/d = 2. For the single-line case (figure 8a), the flow develops downstream from SLR to vortex shedding. In contrast, the flows along three lines within the array (1⁻, 0, 1⁺ in figure 9a) are stabilised with SLR persisting throughout the array due to the confinement of adjacent lines of cylinders.

A consequence of adjacent lines is a more chaotic flow within the array, in comparison with that along the single line of cylinders. For a single line of cylinders, a distinct low frequency $(fd/{U_\infty } = 0.035)$ is observed from C4 to C8 (see figure 10e ₄–e ₈). This frequency corresponds to the large-scale tertiary vortex reported in Eizadi et al. (Reference Eizadi, An, Zhou, Zhu and Cheng2022). This is, however, absent for the array (figure 10a ₄–a ₈,c ₄–c ₈) where the flow exhibits a chaotic shedding feature (see figure 10b,d).

In addition to the influence of adjacent lines, flow through an array is more complex still than that of a single line due to the diversion of flow around the array. With this diverted flow, element wakes (particularly near array edges) are directed laterally towards the edge of the array, as shown by streamlines in figure 9(a). Hence typical flow structures such as SLR and TRS, formed due to the alignment of element wakes with the incident flow, progressively break down as array edges are approached (figure 9a–d). This is the reason why the drag distribution along a line of cylinders in an array increasingly deviates from that of a single line of cylinders when the line is closer to the edge of the array (comparing figure 11 with figure 8c).

The identified characteristic flow structures, SLR and TRS, play important roles in determining $\overline {{C_d}} $. A plot of $\overline {{C_d}} $ for arrays with N = 31 (figure 12a), combined with the cross-sectional instantaneous flow fields in 3-D simulations for critical G/d (figure 12b–d), is used to demonstrate this. It is seen that with increasing G/d, $\overline {{C_d}} $ generally increases towards an asymptotic value of roughly 0.8 (figure 12a). The variation of $\overline {{C_d}} $ with G/d in 3-D simulations is close to that in 2-D simulations. Therefore, 2-D simulations are employed to increase the resolution with G/d and identify critical values (≈3.3, 7.8) for the transition between different regimes. Despite the general increasing trend, the variation of $\overline {{C_d}} $ with G/d depends on the element-scale flow structure. SLR remains dominant in the range G/d = 1.8–2.7 (figure 12a), with a typical instantaneous flow field shown in figure 12(b). There is therefore little variation of $\overline {{C_d}} $ in this range (figure 12a). Beyond G/d = 2.7, SLR starts to break down. With diminished drag reduction in SLR, the $\overline {{C_d}} $ value increases sharply above G/d = 2.7. With further increase in G/d, TRS begins to dominate the element wake. The extent of cylinders covered by TRS is virtually unchanged over the range G/d = 3.4–4.7, in which the instantaneous flow field is similar to that shown in figure 12(c) and hence the $\overline {{C_d}} $ value stabilises around 0.55. There is a second sharp increase of $\overline {{C_d}} $ from G/d = 4.7 to 8 where with increasing G/d the extent of cylinders covered by TRS in a line is progressively reduced upstream and eventually to only C2 for lines in the middle of the array (see figure 12d). The element wake is in NOC for G/d > 8, causing the near-constant value of $\overline {{C_d}} $ for G/d > 8 (figure 12a). The above interpretation reveals the underlying link between the characteristic flow structure (and its variation with G/d) and the averaged element drag coefficient in finite arrays.

Figure 12.

Demonstration of the critical role of element-scale flow structures in determining the average element drag coefficient in both 3-D and 2-D DNS. (a) Variation of $\overline {{C_d}} $ with G/d for arrays with N = 31. Dashed lines represent the range of G/d where SLR, TRS or NOC become dominant. (b–d) Cross-sectional instantaneous flow fields of N = 31 for critical gap ratios (G/d = 2.7, 4.5, 8) at z/D = 0 in 3-D simulations marked in (a). In (c,d) cylinders covered by TRS are marked in dark grey. Despite the general increasing trend of $\overline {{C_d}} $, values of $\overline {{C_d}} $ are controlled by the characteristic flow structures and its variation with G/d.

More arrays with N = 7, 19, 31, 55, 109 were simulated to demonstrate these three states (SLR, TRS, NOC) in the (G/d, D/d) parameter space (figure 13). This figure includes 10 3-D cases, and 73 complementary 2-D cases due to the computational cost. It is seen from figure 13 that the element-scale wake structures in 3-D and 2-D simulations are in the same flow state for the same value of G/d and D/d. The element-scale wake structure is primarily set by G/d in most cases. For small gap ratio, SLR dominates the element wake within an array at $G/d\lesssim 3$, whose threshold is less than that for the single line $(G/d\lesssim 3.7)$. For intermediate gap ratio $(3\lesssim G/d\lesssim 8)$, the element-scale wake is characterised by TRS. Cases with large gap ratio $(G/d\gtrsim 8)$ display NOC. With increasing G/d, the transition from SLR into TRS coincides with the decrease in the extent of cylinders within an array covered by SLR, as shown in figure 9(a–c). A similar decreasing trend is observed for the extent of TRS as the flow transitions from TRS to NOC (figure 9c,d). The reduction in the extent of these two characteristic flow structures in an array with increasing G/d is similar to that observed for the single-line scenarios (e.g. Hosseini et al. Reference Hosseini, Griffith and Leontini2020).

Figure 13.

The element-scale wake structure and average drag coefficient $\overline {{C_d}} $ of an array of cylinders mapped out in the parameter space of G/d and D/d for θ = 0° for both 3-D and 2-D DNS. Cases along each dashed line have the same total number of cylinders in the array (N, marked on the right). Cases with the same flow state cluster together, demonstrating dependence of characteristic element-scale flow feature on G/d and D/d. The value of $\overline {{C_d}} $ varies significantly with G/d and D/d from 0.23 to 0.90, highlighting the significant impact of cylinder arrangement.

For a single line (e.g. Eizadi et al. Reference Eizadi, An, Zhou, Zhu and Cheng2022) and an infinite array (da Silva et al. Reference da Silva, Luciano, Utzig and Meier2019) of cylinders, the formation of characteristic element-scale flow features is controlled by the gap ratio when the cylinder Reynolds number is fixed. It appears that this is also the case for a finite circular array when D/d > 10, as shown by the separation of three characteristic element-scale wake states at two virtually invariant G/d values (3, 8) for $N\gtrsim 31$ in figure 13. This suggests that the array-scale (D/d) has secondary influence on the formation of characteristic element-scale wake if the array has a sufficiently large number of cylinders $(N\gtrsim 31)$ and large size $(D/d\gtrsim 10)$.

Figure 13 also illustrates that the averaged element drag coefficient $\overline {{C_d}} $ (the filled colour in each symbol) increases with G/d but decreases with D/d, meaning that larger and denser arrays have lower average drag coefficients. The $\overline {{C_d}} $ value varies significantly with G/d and D/d by a factor of more than 4 (from 0.23 to 0.90) in the numerical runs of this study.

3.3. Arrangement effects: incident flow angle

Varying the incident flow angle can generate a wide range of element-scale wake structures, and hence drag distributions, over an array. This is demonstrated through 3-D DNS by varying the incident flow angle for an array of cylinders with N = 31 and G/d = 4.5 in figure 14. For θ = 0°, the flow along each line of cylinders develops from von Kármán vortices into TRS (figure 14a). Channelised flow with streamwise velocity higher than the ambient velocity is formed between adjacent lines (figure 14b). In contrast, the mean velocity in the gap between cylinders in each line is extremely low (figure 14b). The formation of TRS coincides with suppression of lateral mixing (Durgin & Karlsson Reference Durgin and Karlsson1971), which causes the persistence of channelised flow with high velocity throughout the array. The suppression of lateral mixing is indicated by very low Reynolds shear stress magnitude around covered cylinders (marked with dark grey circles in figure 14c). Accordingly, the low gap velocity leads to very low drag force on the cylinders covered by TRS (figure 14d).

Figure 14.

The variation with incident flow angle of the fields of cross-sectional instantaneous spanwise vorticity (first row) (at z/D = 0), time-averaged streamwise velocity (second row), Reynolds shear stress (third row) and time-mean force and drag coefficient (fourth row) for an array of 31 cylinders with (G/d, D/d) = (4.5, 24.8) in 3-D DNS. The fields of Reynolds stress and mean flow are post-processed on a time-mean spanwise-average flow field. The first (a–d), second (e–h) and third (i–l) columns represent θ = 0°, 10° and 30°, respectively. Cylinders within the TRS are marked in dark grey in (a,c,e,g). The arrow in the top-right corner of (d,h,l) represents a unit time-mean force of 1. With increasing θ, the TRS in (a) and channelised flow in (b) are progressively suppressed, which creates an increase in mean drag coefficient in (d,h,l).

For the case with θ = 10°, TRS is identifiable in the lower half of the array but completely disappears in the upper half of the array (figure 14e). In the lower half, the flow diversion (towards the lower side) helps to direct the local flow along the geometric channels resulting in the formation of channelised flow. The diverging flow, however, cuts across the channel direction in the upper half and supresses channelised flow, such that most cylinders in the upper half are no longer in the low-velocity wake regions of upstream cylinders (figure 14f). The forces on most of the cylinders are therefore increased and at a significant angle to the primary axis of the line (figure 14h).

For θ = 30°, TRS and channelised flow completely disappear within the array (figure 14i–k), which leads to a significant increase in mean drag coefficient (figure 14l). The flow progressively develops downstream into a chaotic state, as evidenced by the vorticity field in figure 14(i) and the absence of a dominant peak in spectra of lift coefficient (not shown). These chaotic vortices cause stronger lateral mixing in comparison with the other two flow angles (comparing Reynolds shear stress in figure 14c,g,k). Accordingly, the wake deficit of each cylinder is recovered more rapidly than at smaller θ (comparing figure 14b,f,j). Any downstream cylinder is in turn out of the low-velocity region behind its upstream cylinder (figure 14j) and hence has large drag (figure 14l). The resultant average drag coefficient $\overline {{C_d}} $ increases, more specifically by 42 %, compared with that at θ = 0° (see the legends in figure 14d,l).

The incident flow angle becomes less important with increasing array diameter for a fixed gap ratio. This is illustrated by two arrays with D/d = 24.8 (figure 14) and 48.6 (figure 15) but the same value of G/d (4.5). For $\theta=0^\circ$, the fraction of elements within TRS decreases from 45 % (figure 14a) to only 14 % when doubling the array diameter (figure 15a). Consistently, the channelised flow is eliminated near the back of the larger array (figure 15a), in contrast to that persisting throughout the smaller array (figure 14b). As indicated by the instantaneous and mean flow fields in figure 15, for the large array, whilst no channelised flow is formed at θ = 30°, a large number of cylinders form chaotic vortex shedding (~39 %, identified from lift coefficient spectra) and are non-channelised for $\theta=0^\circ $. Therefore, when increasing θ from 0° to 30°, the resultant breakdown of TRS and channelised flow impacts fewer cylinders within the larger array.

Figure 15.

Cross-sectional instantaneous flow field (upper half) (at z/D = 0) and mean field of streamwise velocity (bottom half) for (a) θ = 0° and (b) θ = 30° in an array of 109 cylinders (G/d, D/d) = (4.5, 48.6) in 3-D DNS. The mean flow is post-processed on a spanwise-average flow field. The colour bar (not shown) of spanwise vorticity is the same as that used in figure 14. Cylinders within TRS are marked in dark grey. With increasing θ from 0° to 30°, the breakdown of TRS and channelised flow impacts relatively fewer cylinders in a large array relative to that in the smaller array in figure 14.

With diminished dominance of TRS and channelised flow, the total drag on the larger array becomes less sensitive to θ. This is best indicated by a transverse-averaged drag coefficient ${\langle {C_d}\rangle _x}$, in which the angled bracket represents the operation of averaging drag coefficients $\overline{C_{d, i}}$ of elements with identical streamwise (x, denoted in the subscript) but different transverse (y) locations. It is seen that for the smaller array (D/d = 24.8), ${\langle {C_d}\rangle _x}$ for θ = 0° is largely lower than that for θ = 30° (figure 16a). A similar contrast in ${\langle {C_d}\rangle _x}$ is seen in the larger array in the same region of 0 < (x − x ₀)/d < 25 (figure 16b), with x ₀ the coordinate of the front cylinder in the array. Lower values of ${\langle {C_d}\rangle _x}$ for θ = 0° in both the small and large arrays are associated with the development of channelised flow in the region 0 < (x − x ₀)/d < 25 (see figures 14b and 15a). The flow channelisation leads to cylinders in each line largely behaving as a single bluff body and hence imposing smaller drag forces. This difference in ${\langle {C_d}\rangle _x}$ between θ = 0° and 30° is dramatically diminished (and reverses in sign) downstream of (x − x ₀)/d = 25 within the larger array (figure 16b). This is because the channelised flow is progressively eliminated beyond (x − x ₀)/d = 25 in the large array at θ = 0°, which coincides with cylinders in this region subject to higher local velocity (figure 15a). Near either side of the array where flow diversion is strong, the cylinders beyond (x − x ₀)/d = 25 at θ = 0° even experience higher velocity than at θ = 30° (comparing figure 15a,b). This leads to greater $\overline {{C_{d,i}}} $ values for cylinders at (x − x ₀)/d > 25 than θ = 30°. This reversal of ${\langle {C_d}\rangle _x}$ at (x − x ₀)/d = 25 explains why the averaged drag coefficient $\overline {{C_d}} $ for the larger array is much less dependent on incident flow angle (values provided in figure 16 legends).

Figure 16.

Evolution of the transverse-averaged drag coefficient ${\langle {C_d}\rangle _x}$ along the two arrays for incident flow angles θ = 0° and 30° in 3-D DNS. The two arrays shown in (a,b) have the same value of G/d (4.5) but different values of D/d: (a) D/d = 24.8, N = 31; (b) D/d = 48.6, N = 109. Error bars represent the standard error of the $\overline{C_{d,i}}$ values averaged in the transverse (y) direction. Both arrays have a dependence of ${\langle {C_d}\rangle _x}$ on θ in the region $0 < (x - {x_0})/d < 25$. The larger array has minimal dependence on incident flow angle beyond (x − x ₀)/d = 25.

The incident flow angle effects are sensitive to the gap ratio. This is demonstrated by comparing the case of G/d = 2.7 with that of G/d = 4.5 described above. At θ = 0° instantaneous flow within the array is dominated by SLR (figure 17a) and channelised flow develops throughout the array (figure 17b). All cylinders in turn have very low drag except for the most upstream cylinder in each line (figure 17c). Shear-layer reattachment breaks down into vortex shedding for some cylinders for θ = 10° (figure 17d), with incident flow redirected to pass through the gap between adjacent lines of cylinders (figure 17e). For θ = 30°, SLR is completely suppressed, and the flow is characterised by a staggered steady wake pattern (figure 17g). Clearly, with increasing θ, the channel flow is progressively broken down from the top of the array towards its bottom. The average drag coefficient is increased by 15 % with an increase of θ from 0° to 30°, which is smaller than 42 % for the case with G/d = 4.5 (comparing legends between figures 14d,l and 17c,i).

Figure 17.

The variation of cross-sectional instantaneous flow field (first row) (at z/D = 0), time-averaged field of streamwise velocity (second row) and the distribution of force and drag coefficient (third row) with incident flow angle for an array of 31 cylinders with (G/d, D/d) = (2.7, 15.3) in 3-D DNS. The first (a–c), second (d–f) and third (g–i) columns represent θ = 0°, 10° and 30°, respectively. The colour bar of spanwise vorticity omitted in (a,d,g) is the same as in figure 14. With increasing θ, SLR in (a) and channelised flow in (b) within the array are progressively suppressed, which coincides with an increase in drag on cylinders over the array in (c,f,i).

In combination, figures 14–17 indicate that the variation of $\overline {{C_d}} $ with θ is dependent on both the gap ratio and array-to-element diameter ratio. This strong, nonlinear dependence of $\overline {{C_d}} $ on θ is further demonstrated in figure 18 with a combination of 3-D and 2-D simulations. Generally, the influence of incident flow angle on $\overline {{C_d}} $ become less significant with decreasing G/d for a fixed total number of cylinders N. On varying θ from 0° to 30°, the averaged drag coefficient can vary in a wide range. For instance, the $\overline {{C_d}} $ value varies in the ranges of 0.90–1.30 and 0.55–0.78 for the arrays with (G/d, D/d) = (17, 35), (4.5, 24.8), respectively (denoted in order ‘⬠’, ‘□’ in figure 18). All arrays exhibit their lowest value of $\overline {{C_d}} $ at θ = 0° due to the formation of channelised flow at this angle (e.g. figures 14b, 15a and 17b). However, the highest value of $\overline {{C_d}} $ occurs at a range of angles and the averaged drag coefficient does not necessarily increase with θ monotonically. It is therefore insufficient to infer the variation of $\overline {{C_d}} $ with θ based purely on the positioning of cylinders relative to incident flow or the projected area of cylinders normal to the incident flow. The controlling mechanism for the dependence of $\overline {{C_d}} $ on θ is the complex variation of element-scale wake interaction with the cylinder arrangement.

Figure 18.

The strong, nonlinear variation of $\overline {{C_d}} $ with θ across the range of arrays, demonstrating that the $\overline {{C_d}} $ value is not only related to the array geometry but also depends on the orientation of the array. The symbols with red and black edges represent 3-D and 2-D cases, respectively.

3.4. Universal descriptor of bulk flow through a porous array

In §§ 3.2 and 3.3, we have shown that the element-scale flow and drag characteristics can vary widely with arrangement parameters G/d, D/d and θ. However, none of these parameters independently control flow through the porous array. From the perspective of momentum balance, the bulk flow through a porous array is controlled by the array resistance, which is defined by the effective flow blockage parameter ${\varGamma ^{\prime}_D}$ (Chang & Constantinescu Reference Chang and Constantinescu2015). It is therefore hypothesised that the effective flow blockage parameter provides a universal description of bulk bleeding flow, which is demonstrated in this section.

Figure 19 presents the variation of ${\bar{u}_p}/{U_\infty }$ with Γ_D and ${\varGamma ^{\prime}_D}$, incorporating results from both 3-D and 2-D simulations. The $\textrm{RMS}{\textrm{E}_\varOmega }$ of ${\bar{u}_p}/{U_\infty }$ for the same values of Γ_D between 3-D and 2-D simulations is 2 %. Given this small discrepancy, the following discussion is based on a combination of 3-D and 2-D simulations.

Figure 19.

The relationship between the bleeding velocity ${\bar{u}_p}/{U_\infty }$ and flow blockage parameters. Here $\overline {{C_d}} = 1$ and direct measurement of $\overline {{C_d}} $ are used in (a) Γ_D and (b) ${\varGamma ^{\prime}_D}$. The filled grey symbols represent the cases varying θ from 0° to 30° with an interval of 5° (see table 4 in Appendix B). The data points are scattered in (a) but collapse well in (b), demonstrating that ${\varGamma ^{\prime}_D}$ controls the bulk bleeding flow.

First, the variation of bulk bleeding velocity ${\bar{u}_p}/{U_\infty }$, defined in (2.7), with the geometric flow blockage parameter Γ_D (assuming $\overline {{C_d}} = 1$) is examined as it has been commonly used to describe the bleeding flow in previous studies (e.g. Rominger & Nepf Reference Rominger and Nepf2011; Zong & Nepf Reference Zong and Nepf2012). It is seen that the bleeding velocity generally decreases with increasing Γ_D. As per (2.13), this means that a denser (low G/d), larger (high D/d) array typically has a lower bleeding velocity.

Despite the general decreasing trend, the bleeding velocity is not entirely controlled by Γ_D. Data from the same arrays (and thus the same value of Γ_D) but with different arrangements are highly scattered, especially in the intermediate range of Γ_D. For instance, figure 19(a) incorporates ten typical arrays with different θ from 0° to 30° with a constant increment of 5°, spanning across the investigated Γ_D range (see grey filled symbols). The difference between the maximum and minimum of ${\bar{u}_p}/{U_\infty }$ in the range of θ = 0°–30° is only about 0.02 for a very small or large flow blockage parameter (e.g. Γ_D = 0.3, 24.2). In contrast, for a medium Γ_D = 1.8, this difference can be up to 0.15, which is one order of magnitude higher than that for very small or very large Γ_D. By varying G/d, D/d and θ (between 0°–30°) simultaneously but keeping Γ_D = 1.8, the dimensionless velocity difference between the maximal and minimal values can be even 0.28 and the relative difference is 50 % (cases 8 and 33 in table 4 in Appendix B). These demonstrate that the impact of the cylinder arrangement on bleeding flow is most critical in the intermediate range of Γ_D, particularly in the range $1\lesssim {\Gamma_D}\lesssim 3$ where the difference between the minimum and maximum of ${\bar{u}_p}$, associated with variation of θ from 0° to 30°, is greater than 10 % of U_∞. The high scattering of data demonstrates that the cylinder arrangement can impose a first-order influence on the bleeding flow. The incompleteness of Γ_D in defining bleeding flow is because this parameter only defines the geometric bulk blockage of an array without fully incorporating information of arrangement.

The prevalence of the characteristic flow structures appears to be the underlying physical mechanism for the most critical arrangement effects in the intermediate range of Γ_D (1–3). It is found that this range of Γ_D corresponds to the intermediate range of G/d and D/d where TRS and SLR is the dominant flow feature within an array (see figure 13). With small variation in θ, the extent of TRS or SLR within the array (and the associated reduction in drag) can be greatly enhanced or suppressed. Accordingly, $\overline {{C_d}} $ (and thus ${\bar{u}_p}/{U_\infty }$) can be particularly sensitive to θ. Outside this range, the flow within the array is characterised by either non-covered element-scale wake (NOC) (low Γ_D) or very low local velocity (hence low drag) (high Γ_D); the bleeding flow is therefore less dependent on the cylinder arrangement.

Contrastingly, there is excellent collapse of bleeding flow data with the effective flow blockage parameter ${\varGamma ^{\prime}_D}$ using direct measurement of $\overline {{C_d}} $ (figure 19b). In fact, the scattered data points in the critical range $(1\lesssim {\Gamma_D}\lesssim 3)$ all collapse in the medium range of $0.5\lesssim {\Gamma^{\prime}_D}\lesssim 1.5$ where ${\bar{u}_p}/{U_\infty }$ has higher gradient against ${\varGamma ^{\prime}_D}$ than at the two ends of ${\varGamma ^{\prime}_D}$. Any change to the cylinder arrangement in this range will alter ${\varGamma ^{\prime}_D}$ and hence dramatically change ${\bar{u}_p}/{U_\infty }$. The improved collapse of ${\bar{u}_p}/{U_\infty }$ with ${\varGamma ^{\prime}_D}$ demonstrates the importance of hydrodynamic response of $\overline {{C_d}} $ to the array arrangement in controlling the bulk bleeding flow, which is, however, missed in geometric Γ_D. The response is manifested by the variation of element-scale flow and drag characteristics with arrangement as discussed in §§ 3.2 and 3.3. By allowing the controlling influence of array arrangement on $\overline {{C_d}} $, the effective flow blockage parameter, i.e. ${\varGamma ^{\prime}_D}$, not only controls the amount of flow passing through a porous array but also determines when the arrangement effects are critical.

There is vast variability of element-scale flow characteristics in ${\varGamma ^{\prime}_D}$. For instance, the two 3-D cases 13* and 20* (marked in bold text in table 4) with ${\varGamma ^{\prime}_D} \approx 1.5$ have different element-scale flow structures but almost identical values of ${\bar{u}_p}/{U_\infty } \approx 0.2$. Vortex shedding is formed within one array (G/d = 2.5, D/d = 27.5, N = 109) at θ = 0° but not in the other array (G/d = 2.3, D/d = 13, N = 31) at θ = 30° (see figure 23a,b in Appendix B). The effective flow blockage parameter basically characterises the net bulk resistance of the array, which determines the bulk bleeding flow. The resistance can be broken down into discrete point drag forces, which are physically modelled by individual cylinders in a circular domain herein. Depending on the array and incident flow properties (defined by G/d, D/d, α, θ, Re_d), the array can have various element-scale flow and drag characteristics at fixed ${\varGamma ^{\prime}_D}$.

Figure 19(b) demonstrates that the bulk bleeding velocity is a function of ${\varGamma ^{\prime}_D}$. Furthermore, this paper demonstrates the complex variation of $\overline {{C_d}} $, a critical component of ${\varGamma ^{\prime}_D}$, with arrangement, by varying selected arrangement parameters (G/d, D/d, θ) in §§ 3.2 and 3.3. As a combination, this shows how the arrangement changes the element-scale flow and drag characteristics and eventually changes the bulk force on and the bulk flow through the array. The relevance of $\mathit{\Gamma_{D}^{\prime}}$ and coupling between element and array scales are further explored in He et al. (Reference He, Draper, Ghisalberti, An and Branson2024a).

3.5. Variability of averaged drag coefficient

The scattering of data points in figure 19(a) is related to the variation of $\overline {{C_d}} $ for the same Γ_D (geometric flow blockage). By accounting for the $\overline {{C_d}} $ variation, figure 19(b) presents a relation between ${\bar{u}_p}/{U_\infty }$ and ${\varGamma ^{\prime}_D}$. However, prediction of ${\bar{u}_p}/{U_\infty }$ in reality still requires quantitative information of $\overline {{C_d}} $. Therefore, the variability of $\overline {{C_d}} $ in and with Γ_D is investigated in figure 20(a), which compiles data from this study and previous work (listed in figure 20b). It is seen that the present data agree with existing data of various arrangements and Re_d. The $\textrm{RMS}{\textrm{E}_\varOmega }$ of ${\bar{u}_p}/{U_\infty }$ for the same values of Γ_D between the present 3-D and 2-D simulations is 8 %.

Figure 20.

(a) An integration of data of $\overline {{C_d}} $ against Γ_D from previous studies and the present work. (b) List of references shown in (a). The symbols from the present study represent the same cases as shown in the legend in figure 19(a). Note: Exp, laboratory experiment; Num, numerical simulation. The plot in (a) demonstrates that there is clear variability of $\overline {{C_d}} $ with Γ_D, but also scatter across systems.

With increasing Γ_D the averaged drag coefficient generally decreases and variability of $\overline {{C_d}} $ in Γ_D reduces. The averaged drag coefficient varies from minimum to maximum by up to 50 % at low Γ_D = 0.5, but only by 17 % at Γ_D = 24 in the range θ = 0°–30°. For large Γ_D, the array can be either very densely packed or very large, which respectively reduces the freedom of altering arrangement and limits the impact of typical flow structures in changing arrangement. The overall drag force in turn becomes much less dependent on arrangement. Instead, it is more dependent on the extremely low bleeding velocity. This leads to convergence of $\overline {{C_d}} $ at high Γ_D.

Most data in figure 20(a) have $\overline {{C_d}} $ values much lower than unity, the value adopted in previous studies (e.g. Zong & Nepf Reference Zong and Nepf2012). While $\overline {{C_d}} = 1$ was assumed with reference to a single solid cylinder in steady flow, the averaged drag coefficient for a porous array is normally lower than that for a single cylinder due to arrangement effects. The complex variation of element-scale flow characteristics with independent variables (G/d, D/d, θ, α, Re_d) makes it difficult to collapse those independent parameters down to a single parameter to accurately characterise $\overline {{C_d}} $. Nevertheless, figure 20(a) enables a rough estimation of $\overline {{C_d}} $ for real systems, with uncertainty range for given Γ_D. When combined with figure 19(b), it provides predictive capacity for the bulk bleeding velocity.

3.6. Fluctuating force characteristics

In previous sections, force characteristics of individual elements have been discussed through time-mean quantities (e.g. $\overline {{C_d}} $). To show the whole picture of element-scale hydrodynamic features, in this section, the fluctuating components of lift and drag forces are explored. These two components are quantified by the average root mean squares of lift and drag coefficients of individual elements, as defined in (2.5). The $\textrm{RMS}{\textrm{E}_\varOmega }$ of $\overline {{C_{d,rms}}} $ and $\overline {{C_{l,rms}}} $ for the same values of ${\varGamma ^{\prime}_D}$ between the present 3-D and 2-D simulations are 4 % and 10 %, respectively.

The variations of $\overline {{C_{d,rms}}} $ and $\overline {{C_{l,rms}}} $ with ${\varGamma ^{\prime}_D}$ are associated with both element-scale and array-scale wake structures (figure 21). With an increase of ${\varGamma ^{\prime}_D}$ from 0, the $\overline {{C_{d,rms}}} $ and $\overline {{C_{l,rms}}} $ values increase and peak at ${\varGamma ^{\prime}_D} \approx 1$ and then decrease sharply to approach zero at ${\varGamma ^{\prime}_D} \approx 1.5$. This increasing trend is due to the transition of element-scale wake structure from NOC to TRS state with stronger flow fluctuations within the array. In contrast, beyond ${\varGamma ^{\prime}_D} \approx 1$, the number of cylinders with element-scale vortex shedding is reduced with the increase of ${\varGamma ^{\prime}_D}$ due to lower bleeding velocity and smaller gaps between individual elements. This is responsible for the decrease of $\overline {{C_{d,rms}}} $ and $\overline {{C_{l,rms}}} $ to 0.

Figure 21.

The variations of average root mean square lift and drag coefficients with the effective flow blockage parameter. The symbols represent the same cases as shown in the legend in figure 19(a). Note that the dashed ellipses mark the cases 12* and 19* (table 3 in Appendix B), which have different arrangements but the same value of Γ_D = 3. The $\overline {{C_{d,rms}}} $ and $\overline {{C_{l,rms}}} $ values show a different variation with ${\varGamma ^{\prime}_D}$ on either side of ${\varGamma ^{\prime}_D} \approx 1.5$.

In comparison with this decrease, the $\overline {{C_{d,rms}}} $ and $\overline {{C_{l,rms}}} $ values start to increase when ${\varGamma ^{\prime}_D}\gtrsim 1.5$. Whilst the element-scale vortex shedding is completely suppressed in this ${\varGamma ^{\prime}_D}$ range, the fluctuations of lift and drag forces are driven by flow oscillations induced by the array-scale vortex structures behind the array. The bleeding velocity is reduced with the increase of ${\varGamma ^{\prime}_D}$ such that the array shear layers become much stronger and the location of generating stronger large-scale vortex structures shifts towards the array. This causes stronger oscillations to the flow within the array by changing the pressure in the near wake of the array, which leads to larger fluctuations of lift and drag forces on individual cylinders.

Similar to $\overline {{C_d}} $ and ${\bar{u}_p}/{U_\infty }$, the $\overline {{C_{d,rms}}} $ and $\overline {{C_{l,rms}}} $ values can also vary widely with the cylinder arrangement, especially in the range of ${\varGamma ^{\prime}_D}\lesssim 1.5$. For instance, through changing the cylinder arrangement (G/d, D/d, θ) but keeping Γ_D = 3, both $\overline {{C_{d,rms}}} $ and $\overline {{C_{l,rms}}} $ values can vary by two orders of magnitude (see 3-D data points denoted by dashed ellipses in figure 21a,b). Even considering the same value of ${\varGamma ^{\prime}_D} = 1.2$, the value of $\overline {{C_{l,rms}}} $ can vary by a factor of two (from 0.3 up to 0.6). The effective flow blockage parameter ${\varGamma ^{\prime}_D}$ controls the time-mean bulk flow through the array ${\bar{u}_p}/{U_\infty }$ but not for the instantaneous flow within the array (hence $\overline {{C_{d,rms}}} $ and $\overline {{C_{l,rms}}} $).