2.1 Numerical method

Direct numerical simulations of single-phase flows are performed for an incompressible Newtonian fluid. A pseudospectral method is employed in the periodic directions (streamwise, $x$ , and spanwise, $z$ ) and second-order finite differences are used for spatial discretization in wall-normal $y$ direction. The solution is advanced in time by a third-order Runge–Kutta scheme. Incompressibility is achieved by correcting the pressure contribution, which is a solution of a Poisson equation. The fluid velocity and pressure fields are a solution of the continuity equation (2.1) and momentum balance equations (2.2) and (2.3):

(2.1) $$\begin{eqnarray}\displaystyle & \displaystyle \frac{\unicode[STIX]{x2202}u_{j}}{\unicode[STIX]{x2202}x_{j}}=0, & \displaystyle\end{eqnarray}$$

(2.2) $$\begin{eqnarray}\displaystyle & \displaystyle \frac{\unicode[STIX]{x2202}u_{i}}{\unicode[STIX]{x2202}t}+u_{j}\frac{\unicode[STIX]{x2202}u_{i}}{\unicode[STIX]{x2202}x_{j}}=-\frac{1}{\unicode[STIX]{x1D70C}_{f}}\frac{\unicode[STIX]{x2202}p}{\unicode[STIX]{x2202}x_{i}}+\unicode[STIX]{x1D708}\frac{\unicode[STIX]{x2202}u_{i}}{\unicode[STIX]{x2202}x_{j}\unicode[STIX]{x2202}x_{j}}+\frac{1}{\unicode[STIX]{x1D70C}_{f}}F_{i}, & \displaystyle\end{eqnarray}$$

(2.3) $$\begin{eqnarray}\displaystyle & \displaystyle F_{i}^{k}=-\mathop{\sum }_{k_{s}=1}^{8}\mathop{\sum }_{n=1}^{N_{k_{s}}}\frac{w_{k_{s}}^{n}}{\triangle V_{k_{s}}}f_{i}^{n}. & \displaystyle\end{eqnarray}$$

Using a projection technique, the body force in the momentum balance equation (2.3) accounts for the particle momentum contribution to node $k$ of the fluid based on particles in all of the eight computational volumes ( $k_{s}$ ) which share this node. Parameter $w_{k_{s}}^{n}$ is the linear geometric weight for each particle $n$ based on its distance from node $k$ , and the inner summation is over all $N_{k_{s}}$ particles in the volume $k_{s}$ .

Numerical simulations of particle trajectories and suspension flow dynamics are based on the standard Lagrangian point-particle approximation where the particle-to-fluid density ratio $r\equiv \unicode[STIX]{x1D70C}_{p}/\unicode[STIX]{x1D70C}_{f}\gg 1$ and the particle size is smaller than the smallest viscous dissipation scales of the turbulence. As a consequence of this and the low volume concentrations (a maximum volume fraction of $\overline{\unicode[STIX]{x1D6F7}_{V}}$ less than $1\times 10^{-3}$ is used in this study), only the Stokes drag force and two-way coupling have been incorporated (see Balachandar & Eaton Reference Balachandar and Eaton2010). The velocity of particle $n$ is governed in (2.4) and particle trajectories are then obtained from numerical integration of the equation of motion (2.5):

(2.4) $$\begin{eqnarray}\displaystyle & \displaystyle \frac{\text{d}u_{p,i}^{n}}{\text{d}t}=\frac{f_{i}^{n}}{m_{p}}=\frac{1}{\unicode[STIX]{x1D70F}_{p}}[1+0.15(Re_{p}^{n})^{0.687}](u_{f,i}^{n}-u_{p,i}^{n}), & \displaystyle\end{eqnarray}$$

(2.5) $$\begin{eqnarray}\displaystyle & \displaystyle \frac{\text{d}x_{i}^{n}}{\text{d}t}=u_{p,i}^{n}, & \displaystyle\end{eqnarray}$$

where $\unicode[STIX]{x1D70F}_{p}=\unicode[STIX]{x1D70C}_{p}{d_{p}}^{2}/18\unicode[STIX]{x1D707}$ is the Stokes relaxation time of the particle, $m_{p}$ is the particle mass and the particle Reynolds number $Re_{p}^{n}=|u_{f,i}^{n}-u_{p,i}^{n}|d_{p}^{n}/\unicode[STIX]{x1D708}$ is based on the magnitude of the particle slip velocity $(u_{f,i}^{n}-u_{p,i}^{n})$ and particle diameter $d_{p}^{n}$ . In this work, the average $Re_{p}^{n}$ is between $0.125$ ( $r=80$ , two-way) and $3.25$ ( $r=8000$ , one-way), which is far smaller than the suggested maximum $Re_{p}\approx 800$ for the Stokes drag correction in (2.4) (Schiller Reference Schiller1933). Thus the $Re_{p}$ correction to the Stokes drag is minimal in this study. In order to highlight the effect of particle response time, we do not consider gravitational settling. Other terms in the particle momentum equation (see Maxey & Riley Reference Maxey and Riley1983) are neglected since they remain small compared with drag when the density ratio $r\gg 1$ . These neglected terms are also found to have less effect on the stability analysis in relatively low-Reynolds-number turbulence as demonstrated by Klinkenberg, de Lange & Brandt (Reference Klinkenberg, de Lange and Brandt2014). Particle–particle collisions are not taken into consideration due to the low volume concentrations, and we exert a purely elastic collision between particles and the upper/lower walls. This purely elastic wall collision is commonly used in gas–solid turbulence (Li et al. Reference Li, McLaughlin, Kontomaris and Portela2001; Sardina et al. Reference Sardina, Schlatter, Brandt, Picano and Casciola2012; Zhao et al. Reference Zhao, Andersson and Gillissen2013). However, we have tested the restitution coefficient $|u_{p,init}^{n}/u_{p,final}^{n}|$ between $0.5$ and $1$ and do not observe significant changes to particle distributions or two-way coupling, consistent with Li et al. (Reference Li, McLaughlin, Kontomaris and Portela2001).

In addition to the point force approximation as implemented in (2.3), there are other techniques that can be used to account for two-way coupling in particle-laden flows. For example Pan & Banerjee (Reference Pan and Banerjee1996) use the pressure gradient and stress tensor from an analytic velocity disturbance field to replace the point force, Capecelatro & Desjardins (Reference Capecelatro and Desjardins2013) use a volume filtering operator to replace the point sources by smoother, locally filtered fields and Gualtieri et al. (Reference Gualtieri, Picano, Sardina and Casciola2015) present an exact regularized point particle method which includes the far-field disturbance by applying a multipole expansion to obtain the Stokeslet. In order to validate our current method, we compared both with the results of Zhao et al. (Reference Zhao, Andersson and Gillissen2013) at $Re_{\unicode[STIX]{x03C4}}=180$ with $\unicode[STIX]{x03C4}_{p}^{+}=30$ as well as with results from Capecelatro & Desjardins (Reference Capecelatro and Desjardins2013) at $Re_{\unicode[STIX]{x03C4}}=235$ with $\unicode[STIX]{x03C4}_{p}^{+}=65$ . The mean velocity profile, concentration profile and root mean square (RMS) velocity profile agree well with each other (not shown here), which suggests that the current particle in cell method for two-way coupling is suitable for our study, and is not sensitive to the two-way coupling scheme. Additional details and validation of the numerical scheme can be found in previous work for both one-way coupling and two-way coupling (Richter & Sullivan Reference Richter and Sullivan2013; Sweet, Richter & Thain Reference Sweet, Richter and Thain2018).

2.2 Single-phase PCF

Figure 2.

Comparison between current study and published data from Kawahara & Kida (Reference Kawahara and Kida2001) and Wang, Abbas & Climent (Reference Wang, Abbas and Climent2017) in single-phase flow at $Re_{b}=400$ : (a) mean velocity profiles; (b) fluctuation velocity RMS in three directions.

Simulation of PCF turbulence in the miniunit configuration is used in the rest of this paper, at a relatively low Reynolds number (just above the onset of transition to turbulence). The flow is driven by two walls moving at equal and opposite velocities, and by imposing a no-slip condition on each. Boundary conditions in the lateral, wall-parallel directions are periodic. The Reynolds number is defined using half of the relative wall velocity and half of the gap between the two walls. The turbulent state is obtained by inducing fluctuations into a laminar flow at a high Reynolds number and once a turbulent state is developed, the Reynolds number is decreased gradually to a desired Reynolds number. The grid used in this work is $N_{x}\times N_{y}\times N_{z}=32\times 64\times 32$ (grid convergence tests (not shown here) have ensured that results are independent of grid resolution) corresponding to a resolution in wall units of $\unicode[STIX]{x0394}x^{+}\times \unicode[STIX]{x0394}y^{+}$ (wall, centre) $\times \unicode[STIX]{x0394}z^{+}\approx 6.8\times (1,1.5)\times 4.7$ . Figure 2 shows the mean velocity profile and velocity fluctuations in all three directions, in comparison with Wang et al. (Reference Wang, Abbas and Climent2017) who used a second-order centred finite volume method and Kawahara & Kida (Reference Kawahara and Kida2001) who used a pseudospectral discretization in all three directions. We can see that the RMS velocity fluctuations from the present simulations agree well with Kawahara & Kida (Reference Kawahara and Kida2001), whereas there is a small discrepancy between the current work and Wang et al. (Reference Wang, Abbas and Climent2017) in the streamwise and spanwise directions near the centre region. This might be due to the slight difference of domain size in the spanwise direction which confines the shape of two counter-rotating rolls in miniunit.

2.3 Suspension flow configurations

Table 1.

Parameters of baseline numerical simulations. The Reynolds number $Re_{b}\equiv U_{w}h/\unicode[STIX]{x1D708}$ , where $U_{w}$ is half of the relative velocity and $h=L_{y}/2$ is half of the gap between two walls. The friction Reynolds number is $Re_{\unicode[STIX]{x1D70F}}\equiv u_{\unicode[STIX]{x03C4}}h/\unicode[STIX]{x1D708}$ and the particle Stokes relaxation time is $\unicode[STIX]{x03C4}_{p}\equiv \unicode[STIX]{x1D70C}_{p}d^{2}/(18\unicode[STIX]{x1D70C}_{f}\unicode[STIX]{x1D708})$ , where $d$ represents the particle diameter. The superscript ‘ $+$ ’ is the dimensionless number based on viscous scale, where $\unicode[STIX]{x1D6FF}_{\unicode[STIX]{x1D708}}$ , $u_{\unicode[STIX]{x03C4}}$ and $\unicode[STIX]{x1D708}/u_{\unicode[STIX]{x03C4}}^{2}$ correspond to the viscous length scale, velocity scale and time scale, respectively. FCM (force coupling method) is a low-order multipole expansion for capturing finite-size effects used by Wang et al. (Reference Wang, Abbas and Climent2017).

Table 1 contains several selected parameters for the statistical and modal analysis in this study (i.e. the cases used for all analyses but not including the transition tests of § 3.1). The case number indicates the bulk Reynolds number, density ratio and particle volumetric concentration used in that particular simulation. Turbulence–laminar transition occurs near case 380–900–4. Hamilton & Abernathy (Reference Hamilton and Abernathy1994) found a minimum threshold of streamwise vorticity circulation of the LSVs (below which the transition happens), and therefore we use this case to check if this threshold is modified by the presence of inertial particles.

The Kolmogorov scale is difficult to determine and somewhat ambiguous at the low turbulence levels in these simulations, although we can estimate that the minimum dissipation scale is about $1.5$ times the viscous length scale at the wall, which itself is about $2.0$ times the viscous length scale in the centre (Pope Reference Pope2000). The ratio of the particle to fluid time scale defines the Stokes number $St_{turb}\equiv \unicode[STIX]{x03C4}_{p}/\unicode[STIX]{x03C4}_{f}$ , where $\unicode[STIX]{x03C4}_{f}\equiv L_{y}/\max (v^{\prime }|w^{\prime })$ is related to the turnover time scale of the LSVs (Massot Reference Massot2007; Wang et al. Reference Wang, Abbas and Climent2017). The estimation of $\unicode[STIX]{x03C4}_{f}$ is about $25$ time units (time unit defined as $h/U_{w}$ ) from figure 7(a). In the context of high-Reynolds-number atmospheric boundary layer flows, the particle Stokes number $\unicode[STIX]{x03C4}_{p}^{+}$ based on the viscous scale corresponds to a diameter of $d\approx 30~\unicode[STIX]{x03BC}\text{m}$ to $d\approx 90~\unicode[STIX]{x03BC}\text{m}$ at $Re_{\unicode[STIX]{x03C4}}\approx 10^{4}$ within $1$  m of the surface when the particle density is of the order of $1000~\text{kg}~\text{m}^{-3}$ . In order to gain insight into the finite-size effects on particle dispersion and turbulence modulation, we specifically select one group of parameters $\unicode[STIX]{x03C4}_{p}^{+}$ and $St_{turb}$ to be nearly the same as the finite-size particles used in Wang et al. (Reference Wang, Abbas and Climent2017). There are two cases, 500–5000–4 and 500–8000–4, with extremely high density ratios that are intended to help build a complete understanding of the effect of Stokes number on particle dispersion in the LSVs in our model flow. These high density ratios are simulated with one-way coupling only since they fully laminarize the flow at $Re_{b}=500$ .

3.1 Effects of particle response time and mass loading on transition

Figure 3.

Effect of inertial particles on the turbulence-to-laminar transition as indicated by the temporal evolution of $C_{f}$ . This is demonstrated by gradually decreasing $Re_{b}$ , density ratio ( $r$ ) and bulk volumetric concentration ( $\overline{\unicode[STIX]{x1D6F7}_{V}}$ ) starting from a fully turbulent simulation at $Re_{b}=500$ . (a) Case 500–80–4, decreasing $Re_{b}=500$ to $Re_{b}=290$ ; (b) case 500–900–4, decreasing $Re_{b}=500$ to $Re_{b}=400$ ; (c) fixed $Re_{b}=500$ and increasing density ratio $r=80$ to $r=3000$ ; (d) fixed $Re_{b}=500$ and two density ratios $r=1800$ and $r=3000$ , $\unicode[STIX]{x1D6F7}_{b}$ decreasing from $4\times 10^{-4}$ to $1\times 10^{-4}$ .

While linear stability analysis is a key tool in understanding the effects of a dispersed phase on transition, it often must assume that the flow is uniformly (or with some other specified distribution) laden with particles, and that the flow is two-dimensional (Saffman Reference Saffman1962; Michael Reference Michael1964; Dimas & Kiger Reference Dimas and Kiger1998; Rudyak, Isakov & Bord Reference Rudyak, Isakov and Bord1998). Clearly, this type of analysis is fundamentally limited in uncovering nonlinear effects such as the regeneration cycle or inertial accumulation of particles (Squires & Eaton Reference Squires and Eaton1990; Eaton & Fessler Reference Eaton and Fessler1994), and theoretical predictions are difficult to develop. The interactions of inertial particles with coherent structures (e.g. LSSs and LSVs in PCF; see figure 6 a) complicate matters and necessitate a fully nonlinear analysis. For this reason we examine the turbulent-to-laminar transition threshold from an empirical point of view, by simulating a fully developed turbulent PCF experiencing (i) successive reduction of the Reynolds number for several density ratios (figure 3 a,b), (ii) successive increase of the particle density at a fixed Reynolds number (figure 3 c) and (iii) successive increase of the particle mass loading for several density ratios at a fixed Reynolds number (figure 3 d). Identifying a rigorous turbulent–laminar transition threshold would require a large number of simulations to form converged relaminarization statistics. Instead of pinpointing this value, we are instead interested in classifying the particle influence as either stabilizing or destabilizing and examining the impact of the particles on flow features, and therefore do not attempt to specify the true critical Reynolds numbers.

Transition of single-phase flow is observed at $Re_{c}\sim 320$ which is nearly the same as in Wang et al. (Reference Wang, Abbas and Climent2018) in the miniunit domain. The wall friction coefficient $C_{f}=2\overline{\unicode[STIX]{x03C4}}_{w}/(\unicode[STIX]{x1D70C}U_{w}^{2})$ (summed on both walls) is used as an indicator of current flow regime. The same initial turbulent flow configurations were chosen from the single-phase flow simulations at $Re_{b}=500$ for all tests. The particles were then randomly seeded in the simulation domain. The two-phase flow simulations were integrated for at least $1200$ time units ( $h/U_{w}$ ) before the Reynolds number was abruptly decreased, in order to accurately evaluate the transition threshold. The evolution in time of the wall friction coefficient is shown in figure 3 for the various tests. Figure 3(a,b) shows the successive reductions of the Reynolds number down to turbulent-to-laminar transition for $r=80$ ( $Re_{c}\sim 290$ ) and $r=900$ ( $Re_{c}\sim 400$ ) with the same volumetric concentration of $\overline{\unicode[STIX]{x1D6F7}_{V}}=4\times 10^{-4}$ . We also perform similar tests (not shown in this figure) for $r=200$ ( $Re_{c}\sim 240$ ), $r=300$ ( $Re_{c}\sim 260$ ) and $r=500$ ( $Re_{c}\sim 310$ ). Figure 3(c) shows the successive increase of the particle density up to the turbulence-to-laminar transition for $Re_{b}=500$ (transition occurs between $r=900$ and $r=1800$ ) with the same volumetric concentration of $\overline{\unicode[STIX]{x1D6F7}_{V}}=4\times 10^{-4}$ . Finally, figure 3(d) shows the successive increase of the particle volume/mass loading for two density ratios $r=1800$ (transition occurs by $\overline{\unicode[STIX]{x1D6F7}_{V}}=4\times 10^{-4}$ ) and $r=3000$ (transition occurs by $\overline{\unicode[STIX]{x1D6F7}_{V}}=2\times 10^{-4}$ ).

Clearly, two limiting cases exist: at small density ratios (thus small Stokes numbers since $d_{p}$ is held constant) the flow experiences destabilization and the critical Reynolds number is lowered, and at high density ratios (high Stokes numbers) the particles tend to stabilize the turbulence and increase the critical Reynolds number. This is consistent with the linear stability prediction given by Saffman (Reference Saffman1962). For particles with high density ratios, the damping effect on turbulence varies monotonically with mass loading (again consistent with previous studies (Saffman Reference Saffman1962; Dimas & Kiger Reference Dimas and Kiger1998; DeSpirito & Wang Reference DeSpirito and Wang2001)), at least in the dilute regimes considered here. The turbulence attenuation effect is enhanced with an increased particle mass loading as seen in figure 3(d). In pressure-driven flow, Klinkenberg et al. (Reference Klinkenberg, De Lange and Brandt2011) proposed a modified Reynolds number ( $Re_{m}=(1+\overline{\unicode[STIX]{x1D6F7}_{m}})Re_{b}$ ) for heavy particles via analysis of the standard Orr–Sommerfeld–Squire system. This effective Reynolds number can only predict turbulence damping (and an increase of the critical Reynolds number) and has a proportional increase with the particle mass loading. Comparing this estimate based on $Re_{m}$ to the present simulations, we obtain

$Re_{c}\sim 400{-}450$ for $r=900$ and $\overline{\unicode[STIX]{x1D6F7}_{V}}=4\times 10^{-4}$ (predicted $Re_{m}=435$ );

$Re_{c}\sim 500$ for $r=900$ to $1800$ and $\overline{\unicode[STIX]{x1D6F7}_{V}}=4\times 10^{-4}$ (predicted $Re_{m}=435{-}550$ );

$Re_{c}\sim 500$ for $r=1800$ and $\overline{\unicode[STIX]{x1D6F7}_{V}}=2\times 10^{-4}$ to $4\times 10^{-4}$ (predicted $Re_{m}=435{-}550$ );

$Re_{c}\sim 500$ for $r=3000$ and $\overline{\unicode[STIX]{x1D6F7}_{V}}=1\times 10^{-4}$ to $2\times 10^{-4}$ (predicted $Re_{m}=416{-}512$ ).

We therefore find that the transition Reynolds number found in the PCF laden with high-inertia particles ( $St_{turb}>0.5$ , where turbulence is attenuated) is in the range of the estimates using the modified Reynolds number predicted by Klinkenberg et al. (Reference Klinkenberg, De Lange and Brandt2011).

3.2 Conditional test of two-way coupling

For pointwise particles, two-way coupling is realized by applying a corrected three-dimensional Stokes drag with Navier–Stokes equation as in (2.4). The extra dissipation (the loss of kinetic energy due to fluid–particle interactions; see Zhao et al. (Reference Zhao, Andersson and Gillissen2013)) plays the key role in providing a stabilizing effect. In order to further investigate the key coupling components that lead to stabilization and turbulence attenuation, we perform a ‘conditional’ test in this section.

Figure 4.

Turbulence–laminar transition threshold indicated by the temporal evolution of $C_{f}$ . A ‘conditional’ test of sensitivity due to inertial particle drag force applied in different spatial regions in case 500–3000–4.

As shown in figure 4, we begin with the case 500–3000–4 (initialized from an unladen turbulent flow field at $Re_{b}=500$ ) with randomly distributed particles. The flow decays into laminar flow with normal, full two-way coupling. By artificially removing the interphase coupling force in either the streamwise $x$ direction or both the $y$ and $z$ directions throughout the whole domain, we first find that streamwise coupling is the primary mechanism that attenuates the turbulence (i.e. without $y$ and $z$ coupling the flow still transitions from turbulent to laminar). From here, the streamwise coupling is only removed (i) from the spatial region where the regeneration cycle is active ( $0.2<y/h<0.8$ ) or (ii) from the inactive region ( $y/h<0.2$ and $y/h<0.8$ ). Turbulence is sustained by removing streamwise coupling in the region associated with the regeneration cycle, suggesting that it is streamwise coupling in this region that is responsible for laminarization. Further, we remove the streamwise coupling in different radial positions relative to the LSVs (strong streamwise vorticity $|\unicode[STIX]{x1D714}_{x}|>0.1$ is in the central region of the LSVs, $|\unicode[STIX]{x1D714}_{x}|>0.015$ is a larger region containing $|\unicode[STIX]{x1D714}_{x}|>0.1$ and $0.015<|\unicode[STIX]{x1D714}_{x}|<0.1$ is an annulus formed by subtracting the area $|\unicode[STIX]{x1D714}_{x}|>0.1$ from $|\unicode[STIX]{x1D714}_{x}|>0.015$ ). This analysis, shown in figure 4 as red diamonds, demonstrates that streamwise coupling, specifically in the region associated with the annulus given by $0.015<|\unicode[STIX]{x1D714}_{x}|<0.1$ and in the region associated with the regeneration cycle, should be considered as the most sensitive region which might be altered by the inertial particles, leading to turbulence attenuation.

Figure 5.

(a,b) Contour of temporal average of streamwise vortex stretching term in one single cross-section ( $y,z$ plane) within $50$ time units and isolines of temporal and streamwise average of streamwise vorticity within $50$ time units. (c,d) Contour (logarithmic scale) of temporal and streamwise average of normalized concentration by the bulk value and isolines of temporal and streamwise average of streamwise vorticity within $50$ time units. (a,c) Case 500–500–4; (b,d): case 500–900–4.

To visualize this stabilization process that occurs in the annulus of region $0.015<|\unicode[STIX]{x1D714}_{x}|<0.1$ , figure 5(a,b) shows the contour of streamwise vorticity stretching $\unicode[STIX]{x1D714}_{x}\unicode[STIX]{x2202}u/\unicode[STIX]{x2202}x$ . This is a streamwise-dependent quantity, so we time-average (50 time units) at a single cross-section at $x=L_{x}/2$ . Figure 5(c,d) shows the contour of the normalized particle concentration with respect to the bulk concentration at the same location and averaged over the same time. The isolines in all panels are the streamwise-averaged streamwise vorticity ( $\overline{\unicode[STIX]{x1D714}_{x}}(y,z)$ ). Figure 5(a,c) and figure 5(b,d) show cases 500–500–4 and 500–900–4, respectively. It is clear that strong vorticity stretching happens in the range of $0.015<|\unicode[STIX]{x1D714}_{x}|<0.1$ , which is a key component of the self-sustaining regeneration cycle. The intensity of this stretching is reduced in case 500–900–4 (contour in figure 5 b) compared to that of case 500–500–4 (contour in figure 5 a). Simultaneously, we find that there are more particles present in the range of $0.015<|\unicode[STIX]{x1D714}_{x}|<0.1$ in case 500–900–4 (figure 5 d) than in case 500–500–4 (figure 5 c).

As a result, we associate the turbulence attenuation with the streamwise coupling of high-inertia particles (e.g. $St_{turb}=0.625$ ), and their presence in the range of $0.015<|\unicode[STIX]{x1D714}_{x}|<0.1$ is of key significance since this is the region which has the strongest streamwise vortex stretching effect. Below, we argue that the streamwise vortex stretching and the LSVs are the key sub-steps in the regeneration cycle, and coupled to each other. The preferential presence of particles in this streamwise vortex stretching region is an important phenomenon which alters the regeneration cycle and thus the transition from turbulent to laminar flow, and has been observed at higher Reynolds numbers (Lee & Lee Reference Lee and Lee2015).

3.3 Particle distribution and velocity profile

For inertial pointwise particles, their spatial distribution is determined through the drag force exerted from large-scale turbulent structures to the particles. Due to inertia, heavy particles tend to be expelled from the vortex centre to the low-vorticity but high-strain-rate regions (Druzhinin & Elghobashi Reference Druzhinin and Elghobashi1998). As argued above, this process plays a key role in turbulence modulation since particles tend to accumulate in regions associated with streamwise vortex stretching. In a similar study focused on neutrally buoyant finite-sized particles, Wang et al. (Reference Wang, Abbas and Climent2018) found that particles accumulating in the streaks tend to enhance the turbulence whereas particles accumulating in the large-scale rolls hardly modify turbulence levels.

Figure 6.

(a) Contours of the magnitude of the streamwise flow velocity (colourbar in upper left panel) and instantaneous particle positions projected onto the $(y,z)$ cross-section when the LSS is strongest before breakdown, coloured according to magnitude of particle streamwise velocity (colourbar in upper right panel). The particle size shown in the panels is magnified four times for better visualization. Top-left: case 500–80–4; top-right: case 500–500–4; bottom-left: case 500–900–4; bottom-right: case 500–8000–4 (one-way coupling). (b) Mean volume concentration normalized by the bulk concentration. (c) Mean streamwise velocity scaled by $U_{w}$ .

Figure 7.

RMS velocity fluctuations in cases with density ratios. (a) Fluid phase in three directions. (b–d) Particulate phase in three directions: (b) streamwise direction fluctuation; (c) wall-normal direction fluctuation; (d) spanwise direction fluctuation. In (b–d), fluid phase in single phase is plotted as a reference.

Figure 6(a) shows an instantaneous particle distribution over a $(y,z)$ cross-sectional plane with different density ratios ( $r=80{-}8000$ ) at the same Reynolds number ( $Re_{b}=500$ ) and volumetric concentration ( $\overline{\unicode[STIX]{x1D6F7}_{v}}=4\times 10^{-4}$ ) at a point in time corresponding to the strongest LSSs during the regeneration cycle. From left to right, top to bottom, it is clear that particles with low to moderate inertia tend to accumulate in the LSSs whereas particles with a higher inertia distribute more homogeneously and spread throughout the Couette gap, a behaviour seen in other particle-laden, wall-bounded turbulent flows. Along the same lines, particles with density ratio from $r=300$ to $500$ are mostly trapped inside the LSSs, even during the streak breakdown process (not shown). Either lower (e.g. $r=80$ ) or higher (e.g. $r=900$ ) than this density ratio, the particles can stray from the LSS regions, consistent with known behaviour of inertial particles (Maxey Reference Maxey1987; Druzhinin & Elghobashi Reference Druzhinin and Elghobashi1998). In a Taylor–Green vortex (TGV) set-up, Massot (Reference Massot2007) proposed a threshold of the ratio between particle response time and the TGV turnover time scale. Below the threshold, particles will stay inside this cell whereas particles tend to move to the other cells above the value. We numerically obtained the single-particle trajectory in TGV flow with the same particle time scale corresponding to particles in the present simulated turbulent LSVs. The particle behaviour in TGV flow is analogous to particle distribution in turbulent PCF with the same particle time scale (data are not shown here). Case 500–80–4 has a low $St_{turb}\approx 0.056$ causing the particles to behave more as tracers and stay in one LSV, whereas case 500–8000–4(1way) with a high $St_{turb}\approx 5.56$ leads to particles which cannot follow the streamlines and cross the LSVs after being ejected by ejection events. Particles with intermediate Stokes numbers, as in case 500–500–4 with $St_{turb}\approx 0.35$ , are expelled from the LSVs and become trapped inside the LSSs.

This non-monotonic change in particle distribution can be seen in figure 6(b), where the normalized particle volume concentration is shown. With increasing density ratio from 500–80–4 to 500–300–4, the concentration decreases in the centre whereas it increases in the near-wall region. We observe an opposite tendency when further increasing the density ratio from 500–300–4 to 500–1200–4. At higher density ratios with two-way coupling, transition to laminar flow occurs in the present simulations. At higher flow Reynolds number ( $Re_{b}=2025$ ), the increase of particle concentrations in the centre with increasing particle response time ( $\unicode[STIX]{x03C4}_{p}^{+}\approx 90$ corresponding to $St_{turb}\approx 1.06$ ) is also observed by Richter & Sullivan (Reference Richter and Sullivan2013). For the sake of highlighting the long particle response time effect on the particle distribution, we also show cases 500–3000–4 and 500–8000–4 with one-way coupling, where we can find a nearly homogeneous concentration profile because particles cannot follow the streamlines.

Despite the slip velocity between the particulate phase and fluid phase, the shape of the mean particle velocity profile is mainly determined by the carrier phase. In particular, the mean particle streamwise velocity will reflect the local mean fluid velocity profile of the structure it is contained within (e.g. LSSs or LSVs). Hamilton et al. (Reference Hamilton, Kim and Waleffe1995) have shown that for the miniunit configuration, the characteristic ‘S’ shape of the mean velocity profile in PCF is governed primarily by the LSVs in single-phase flow. Figure 6(c) compares the mean fluid velocity (solid black line) to the mean particle velocities $\overline{u_{p}}/U_{w}$ of particles with varying density. There is a clear distinction between the nearly linear (respectively ‘S’) shape of the mean particle velocity curve for 500–300–4 (respectively 500–3000–4(1way)) if the particles are trapped in the LSSs (respectively across the whole domain). Again, this effect is non-monotonic with particle inertia. For cases 500–3000–4(1way) and 500–8000–4(1way), the qualitative shape of the mean particle velocity is similar to that of mean fluid velocity in single-phase flow; this discrepancy is due to the difference between the particle response time scale and the turnover time of the LSV.

3.4 Turbulence intensity

The RMS velocity fluctuations for the various cases are plotted in figure 7. Figure 7(a) shows the RMS velocities in all three directions for the fluid phase, where it is apparent that $u_{f\,rms}^{\prime }$ is nearly unchanged whereas in the core region $v_{f\,rms}^{\prime }$ and $w_{f\,rms}^{\prime }$ decrease slightly with increased particle inertia. The particulate RMS velocity fluctuations are shown in figure 7(b–d); the single-phase velocity fluctuations are also plotted as a reference (solid black lines). As noted by Yu, Vinkovic & Buffat (Reference Yu, Vinkovic and Buffat2016), particles moving away from the wall are associated mainly with ejections while particles moving towards the wall are associated with sweeps. As seen above, inertial particles with low to moderate Stokes numbers tend to remain in the high-strain-rate region (see figure 6) – regions associated with high $u_{f\,rms}^{\prime }$ . This results in high values of $u_{p\,rms}^{\prime }$ across the whole Couette flow gap which can be seen in figure 7(b). Thus the increase of particulate streamwise turbulent kinetic energy compared with single-phase flow is due to the accumulation of particles in the LSS which contains high streamwise turbulent kinetic energy.

The effect is opposite for $v_{p\,rms}^{\prime }$ and $w_{p\,rms}^{\prime }$ . The fluid velocity fluctuations $v_{f\,rms}^{\prime }$ and $w_{f\,rms}^{\prime }$ are high in the outer regions of the LSVs (not shown here). Therefore when the inertial particles collect in high-strain-rate regions (intermediate Stokes numbers), their spanwise and wall-normal velocity fluctuations are much smaller than the fluid average. At low Stokes numbers, long residence times in the LSVs allow particles to gain wall-normal and spanwise kinetic energy (figure 7 c,d). At high Stokes number, particles tend to again distribute homogeneously throughout the whole domain. Particles move across LSVs, but their inability to quickly adjust their velocity results in suppressed values of $v_{p\,rms}^{\prime }$ and $w_{p\,rms}^{\prime }$ .

As a final note, we have compared the particle distribution, mean velocity and RMS velocity fluctuations (not shown here) between case 500–80–4 and a simulation with finite-size particles from Wang et al. (Reference Wang, Abbas and Climent2017); both have the same particle response time. We find that even for the same Stokes number, finite-size particles collect more in the central region and that their mean velocity and fluctuations are similar to the fluid phase. This difference is due directly to finite-size effects which should be considered for physical systems where particle diameters are larger than the dissipation scales of the flow.

4.1 Intensity of the characteristic terms of regeneration cycle

Figure 8.

Modal decomposition of the various sub-steps composing the regeneration cycle as shown in figure 1 and outlined in § 4. Four cases are shown: 500–0–0 (unladen), 500–80 (500–80–4 and 500–80–10 with $St_{turb}=0.056$ , turbulence enhancement) and 500–900–4 ( $St_{turb}=0.625$ , turbulence attenuation). (a) ${\mathcal{M}}(n\unicode[STIX]{x1D6FC},m\unicode[STIX]{x1D6FD})$ as in (4.1) representing turbulent kinetic energy contained in the streaks; (b) ${\mathcal{C}}(0,\unicode[STIX]{x1D6FD})$ as in (4.2) representing for the intensity of the LSV; (c) ${\mathcal{S}}(0,\unicode[STIX]{x1D6FD})$ as in (4.5) representing the vorticity stretching effect to LSV; (d)  ${\mathcal{L}}(0,\unicode[STIX]{x1D6FD})$ representing the lift-up effect induced by LSV to enhance LSS as in (4.3). The initial $1200$ time units ( $h/U_{w}$ ) are shown.

Table 2.

The average and period of five characteristic terms of the regeneration cycle. The error of the period due to the resolution of the signals is in the range of $\pm 2.5$ time units.

Figure 8 shows the temporal evolution of the quantities defined in the previous section: the wall-normal integrated amplitude of LSSs ( ${\mathcal{M}}(0,\unicode[STIX]{x1D6FD})$ ; figure 8 a), the strength of the meandering streak (represented by mode ${\mathcal{M}}(\unicode[STIX]{x1D6FC},0)$ ; figure 8 a), LSV strength (represented by ${\mathcal{C}}(0,\unicode[STIX]{x1D6FD})$ ; figure 8 b), the vortex stretching term (represented by ${\mathcal{S}}(0,\unicode[STIX]{x1D6FD})$ which is the summation of all 12 pairs of $p=\pm 1$ and $r=\mp 1$ in combination with $q=-2$ to $3$ and $s=3$ to $-2$ ; figure 8 c) and the lift-up term (represented by ${\mathcal{L}}(0,\unicode[STIX]{x1D6FD})$ mode; figure 8 d). Clearly, all of the signals are fluctuating in time with a similar period corresponding to the three regeneration steps as shown in figure 1. In single-phase flow, Hamilton et al. (Reference Hamilton, Kim and Waleffe1995) estimated this period is slightly less than $100$ time units ( $h/U_{w}$ ) at $Re_{b}=500$ . In flow influenced by low-inertia finite-size particles, Wang et al. (Reference Wang, Abbas and Climent2017) also observed a similar period.

Figure 8 shows that the amplitudes of ${\mathcal{M}}(0,\unicode[STIX]{x1D6FD})$ and ${\mathcal{M}}(\unicode[STIX]{x1D6FC},0)$ are diminished by the inertial particles of case 500–900–4 (dotted black line), which is consistent with the behaviour observed by Wang et al. (Reference Wang, Abbas and Climent2017) for finite-size particles. However for low particle inertia, the amplitude of ${\mathcal{M}}(\unicode[STIX]{x1D6FC},0)$ is nearly unchanged due to low mass fraction (case 500–80–4) whereas it is suppressed with the presence of more particles of the same Stokes number (case 500–80–10). The presence of a large number of inertial particles (either when they enhance or attenuate the turbulence) tends to stabilize the amplitude of LSSs ( ${\mathcal{M}}(0,\unicode[STIX]{x1D6FD})$ ) but not significantly affect the time-averaged intensity of the streaks.

The time-averaged intensities of ${\mathcal{C}}(0,\unicode[STIX]{x1D6FD})$ , ${\mathcal{S}}(0,\unicode[STIX]{x1D6FD})$ and ${\mathcal{L}}(0,\unicode[STIX]{x1D6FD})$ are provided in table 2, where we see that these characteristic terms are all suppressed when turbulence attenuation occurs due to particles (case 500–900–4). However, for cases with turbulence enhancement (cases 500–80–10 and 500–80–4), the primary difference is the amplified intensity of the circulation ( ${\mathcal{C}}(0,\unicode[STIX]{x1D6FD})$ , also seen in figure 8 b), and this amplification increases with particle concentration. The strengthened LSVs are critical for sustaining the turbulence and thus are consistent with the observed lower transitional Reynolds number ( $Re_{c}\sim 290$ in case 500–80–4 versus $Re_{c}\sim 320$ for single-phase flow as shown in figure 3 a). As found by Hamilton & Abernathy (Reference Hamilton and Abernathy1994), there is a minimum threshold of circulation below which the turbulence collapses and the flow becomes laminar. At lower Reynolds numbers and/or higher particle concentrations, we indeed see that this is the case (not shown here), and it appears that the primary effect of the particles is to modify the LSV strength during both stabilization and destabilization.

4.2 Periodic behaviour and phase difference of the regeneration cycle

To better understand the effects of particles on the timing of the regeneration cycle, we calculate a temporal autocorrelation of the five signals in figure 8:

(4.6) $$\begin{eqnarray}R_{ss}(\unicode[STIX]{x0394}t)=\frac{\overline{s^{\prime }(t)s^{\prime }(t+\unicode[STIX]{x0394}t)}}{s_{rms}^{\prime ^{2}}},\end{eqnarray}$$

where $s^{\prime }$ is the fluctuation of a signal with respect to the time-averaged value.

Figure 9.

Temporal autocorrelation functions of the five signals shown in figure 8. (a) Single-phase flow, case 500–0–0. Suspension flow: (b) case 500–80–4; (c) case 500–900–4.

The temporal autocorrelation is calculated over $3600$ time units to ensure converged statistics. Figure 9(a) plots the five temporal autocorrelations in single-phase flow, case 500–80–4 is shown in figure 9(b) and case 500–900–4 is shown in figure 9(c). The time difference between the first two maxima of the correlation coefficient is taken as the period of the cycle and these are summarized in table 2. We can see that the period changes due to the presence of the particles, and that in cases of turbulence enhancement, the period is shortened (similar phenomenon is observed by Pan & Banerjee (Reference Pan and Banerjee1996), where particles enhance the turbulence with an increasing frequency of the sweep activity), while for turbulence suppression, the period is lengthened. Along these lines, in single-phase flow, Hamilton et al. (Reference Hamilton, Kim and Waleffe1995) observed that the period of the regeneration cycle is shortened at lower Reynolds numbers, e.g. 18 cycles in 1500 time units at $Re_{b}=400$ versus 16 cycles in 1500 time units at $Re_{b}=500$ . Therefore comparing to single-phase flow, the period of the regeneration cycle increases in case 500–900–4 (turbulence is attenuated, behaving as a lower Reynolds in single-phase flow) whereas the period of the regeneration cycle decreases in cases 500–80–4 and 500–80–10 (turbulence is enhanced, behaving as a higher Reynolds in single-phase flow).

The regeneration cycle consists of three sequential sub-processes sketched in figure 1: streak formation, streak breakdown and streamwise vortex regeneration. The key mechanisms are: the LSSs are generated by a linear lift-up process by the LSVs, the LSSs break down to meandering streaks by complex nonlinear interactions and wavy streaks ( $x$ -dependent, contributing to $\unicode[STIX]{x2202}u/\unicode[STIX]{x2202}x$ ) interact with streamwise vorticity ( $\unicode[STIX]{x1D714}_{x}$ ) to strengthen the LSVs, which is another nonlinear process. This cycle is recognized as a robust, spatial–temporal evolution and time-stable self-sustaining process. We noted above that the period of this cycle (e.g. signal ${\mathcal{M}}(0,\unicode[STIX]{x1D6FD})$ ) is around $104$ time units in single-phase flow, $94$ time units in case 500–80–4 and $121$ time units in case 500–900–4. It is instructive to further analyse the phase difference during the regeneration cycle, especially in order to quantify the inertial particle effect during turbulence modulation. As can be seen in figure 8, the LSSs ( ${\mathcal{M}}(0,\unicode[STIX]{x1D6FD})$ ) and the meandering streaks ( ${\mathcal{M}}(\unicode[STIX]{x1D6FC},0)$ ) nearly have opposite phase. Two subsequent sub-steps, the vortex stretching ( ${\mathcal{S}}(0,\unicode[STIX]{x1D6FD})$ ) and the lift-up ( ${\mathcal{L}}(0,\unicode[STIX]{x1D6FD})$ ) induced by LSVs, are closely synchronous with the temporal evolution of circulation ( ${\mathcal{C}}(0,\unicode[STIX]{x1D6FD})$ ). Vortex stretching, lift-up and circulation all seem to remain in phase with the meandering streaks. To verify this, we perform a temporal cross-correlation study of the five signals based on (4.7):

(4.7) $$\begin{eqnarray}R_{s_{1}s_{2}}(\unicode[STIX]{x0394}t)=\frac{\overline{s_{1}^{\prime }(t+\unicode[STIX]{x0394}t)s_{2}^{\prime }(t)}}{s_{1\,rms}^{\prime }s_{2\,rms}^{\prime }},\end{eqnarray}$$

where $s_{1}^{\prime }$ and $s_{2}^{\prime }$ are the fluctuations of signals $s_{1}$ and $s_{2}$ with respect to the time-averaged value. For instance, $R({\mathcal{M}}(0,\unicode[STIX]{x1D6FD}),{\mathcal{M}}(\unicode[STIX]{x1D6FC},0))$ represents temporal cross-correlation between ${\mathcal{M}}(0,\unicode[STIX]{x1D6FD})$ (LSS) and ${\mathcal{M}}(\unicode[STIX]{x1D6FC},0)$ (meandering streaks), thereby providing information on temporal lag between the two processes. If $R({\mathcal{M}}(0,\unicode[STIX]{x1D6FD}),{\mathcal{M}}(\unicode[STIX]{x1D6FC},0))$ is positive, it reflects that ${\mathcal{M}}(0,\unicode[STIX]{x1D6FD})$ occurs prior in time to ${\mathcal{M}}(\unicode[STIX]{x1D6FC},0)$ during one regeneration cycle – the physical explanation for this specific case is that the LSS is formed, and subsequently breaks down into meandering streaks.

Figure 10.

Temporal cross-correlation functions of the five signals shown in figure 8. (a) Single-phase flow, case 500–0–0. (b) Suspension flow, case 500–80–4. (c) Suspension flow, case 500–900–4. More than $3000$ time units of phase shifts are calculated but only the range of $-100$ to $100$ time units is shown.

Table 3.

Time difference (time units) between connected sub-steps of the regeneration cycle. The error due to the resolution of the signals is in the range of $\pm 2.5$ time units.

Figure 10 illustrates the phase difference between ${\mathcal{M}}(0,\unicode[STIX]{x1D6FD})$ , ${\mathcal{M}}(\unicode[STIX]{x1D6FC},0)$ , ${\mathcal{C}}(0,\unicode[STIX]{x1D6FD})$ , ${\mathcal{S}}(0,\unicode[STIX]{x1D6FD})$ and ${\mathcal{L}}(0,\unicode[STIX]{x1D6FD})$ averaged over $3600$ time units. Starting from the LSS ( ${\mathcal{M}}(0,\unicode[STIX]{x1D6FD})$ ) stage, we have chosen seven correlation coefficients between every pair of connected sub-steps. Figure 10(a) shows these coefficients for single-phase flow, in figure 10(b) for case 500–80–4 and in figure 10(c) for case 500–900–4. We quantify the phase shifts in table 3. Although inertial particles in case 500–900–4 have been found to attenuate the turbulence activity, the inertial particles actually do not significantly alter the basic regeneration cycle or its timing. The shortened or lengthened period of the regeneration cycle is mainly due to the particle modulation of streak breakdown from  ${\mathcal{M}}(0,\unicode[STIX]{x1D6FD})$ (LSS) to ${\mathcal{M}}(\unicode[STIX]{x1D6FC},0)$ (meandering streaks) and streak formation from ${\mathcal{C}}(0,\unicode[STIX]{x1D6FD})$ (LSV) to ${\mathcal{M}}(0,\unicode[STIX]{x1D6FD})$ (LSS). Here we should point out that the vortex regeneration is not necessary for every cycle; it might be absent for some cycles (Hamilton et al. Reference Hamilton, Kim and Waleffe1995), but its minimum value has to be above a threshold to produce unstable streaks.

Figure 11.

Sketch of phase difference between connected sub-steps and the period of the regeneration cycle in minimal unit: (a) case 500–0–0 in comparison with (b) case 500–80–4 and (c) case 500–900–4. (d) Qualitative comparison of the averaged value and the amplitude of the signals with single-phase flow.

Finally, we can summarize the primary effect of inertial particles on turbulence regeneration by continuing to use case 500–80 (including 500–80–4 and 500–80–10 with $St_{turb}=0.056$ ) and case 500–900–4 ( $St_{turb}=0.625$ ) as representative examples of turbulence enhancement and turbulence attenuation, respectively. A temporal evolution of the regeneration cycle can be seen in the schematic sketch in figure 11(a–c). In this figure, we highlight the key turbulent structures and sub-steps by their corresponding metrics. The summation of the transition from ${\mathcal{M}}(0,\unicode[STIX]{x1D6FD})$ to ${\mathcal{M}}(\unicode[STIX]{x1D6FC},0)$ , from ${\mathcal{M}}(\unicode[STIX]{x1D6FC},0)$ to ${\mathcal{C}}(0,\unicode[STIX]{x1D6FD})$ and from ${\mathcal{C}}(0,\unicode[STIX]{x1D6FD})$ to ${\mathcal{M}}(0,\unicode[STIX]{x1D6FD})$ yields roughly $101$ time units in single-phase flow, $90$ time units in case 500–80–4 and $119$ time units in case 500–900–4 corresponding to the whole period $104$ time units, $94$ time units and $120$ time units, respectively. Taking into consideration the resolution of signals, it is reasonable to believe that the summed phase differences are consistent with the calculated period of the regeneration cycle.

The regeneration cycle is therefore a temporal sequence of processes which has a stable periodicity, even under the influence of inertial particles – phase differences between linked sub-steps are nearly unchanged by the presence of particles except streak formation and breakdown. Instead, during turbulence enhancement (case 500–80), as can be seen in figure 11(d), the particles simply strengthen the magnitude of the LSVs ( ${\mathcal{C}}(0,\unicode[STIX]{x1D6FD})$ ), whereas during turbulence attenuation (case 500–900–4), the particles substantially reduce the magnitude of key turbulent structures and their corresponding sub-steps. Once the magnitude of the LSVs ( ${\mathcal{C}}(0,\unicode[STIX]{x1D6FD})$ ) is below a certain threshold (as in case 380–900–4), particles act to abruptly shut off the regeneration, thereby delaying turbulent-to-laminar transition. It is worth noting that a similar upward shift of the transitional Reynolds owing to the vortex suppression is also observed during turbulence modulation laden with polymer additives – the modifications to the ‘exact coherent states’ are due to the suppression of the streamwise vortices due to the polymers exerting an opposite force to the fluid motion in the vortices and weakening the streamwise vortices (see Stone et al. Reference Stone, Roy, Larson, Waleffe and Graham2004).