2.1 Governing equations and particle-phase methods

Consider a dilute monodisperse cloud of particles suspended in an incompressible Newtonian gas. The carrier phase has a density $\unicode[STIX]{x1D70C}_{f}$ and viscosity $\unicode[STIX]{x1D707}_{f}$ and satisfies the Navier–Stokes equations

(2.1) $$\begin{eqnarray}\displaystyle & \displaystyle \unicode[STIX]{x1D735}\boldsymbol{\cdot }\boldsymbol{U}=0, & \displaystyle\end{eqnarray}$$

(2.2) $$\begin{eqnarray}\displaystyle & \displaystyle \unicode[STIX]{x1D70C}_{f}\frac{\unicode[STIX]{x2202}\boldsymbol{U}}{\unicode[STIX]{x2202}t}+\unicode[STIX]{x1D70C}_{f}\boldsymbol{U}\boldsymbol{\cdot }\unicode[STIX]{x1D735}\boldsymbol{U}=-\unicode[STIX]{x1D735}p+\unicode[STIX]{x1D707}_{f}\unicode[STIX]{x1D6FB}^{2}\boldsymbol{U}+\unicode[STIX]{x1D70C}_{f}\boldsymbol{g}-\boldsymbol{F}, & \displaystyle\end{eqnarray}$$

where $\boldsymbol{U}$ is the fluid velocity, $p$ is the pressure, $\boldsymbol{g}$ is the gravitational acceleration and $\boldsymbol{F}$ is the momentum coupling force with the dispersed phase. For dilute solid particles of density $\unicode[STIX]{x1D70C}_{p}$ , diameter $d_{p}$ , velocity $\boldsymbol{V}$ , relaxation time $\unicode[STIX]{x1D70F}_{p}=\unicode[STIX]{x1D70C}_{p}d_{p}^{2}/(18\unicode[STIX]{x1D707}_{f})$ , with diameters small enough that the particle Reynolds number $Re_{p}=|v-u|d_{p}/\unicode[STIX]{x1D708}<1$ (Maxey & Riley Reference Maxey and Riley1983), this coupling is captured by the Stokes drag

(2.3) $$\begin{eqnarray}\boldsymbol{F}=\unicode[STIX]{x1D70C}_{p}\unicode[STIX]{x1D719}\frac{\boldsymbol{U}-\boldsymbol{V}}{\unicode[STIX]{x1D70F}_{p}},\end{eqnarray}$$

where $\unicode[STIX]{x1D719}$ is the local volume fraction. The remaining hydrodynamic forces are neglected. Note that the gravitational force acting on the particles does not directly force the fluid motion. Instead, the gravitational force alters the particle velocity in a way that leads to a particle–fluid drag force (2.3) that resembles gravitational loading when the difference of velocities is nearly relaxed to the terminal velocity at low Stokes numbers.

It is important to note that semi-dilute suspensions for which the average volume fraction $\langle \unicode[STIX]{x1D719}\rangle$ is small can still have a strong coupling between the two phases. Owing to the large density ratio $\unicode[STIX]{x1D70C}_{p}/\unicode[STIX]{x1D70C}_{f}$ of solid particles in gaseous flows, the mass loading $M=\langle \unicode[STIX]{x1D719}\rangle \unicode[STIX]{x1D70C}_{p}/\unicode[STIX]{x1D70C}_{f}$ , a measure of the coupling between the phases, can be of order unity. This means ensembles of particles can collectively modulate the flow on a level on par with the mean fluid velocity.

When treating homogeneous shear flows, it is customary to do a decomposition that isolates the perturbations from the mean homogeneous shear motion

(2.4) $$\begin{eqnarray}\displaystyle & \displaystyle \boldsymbol{U}=\unicode[STIX]{x1D6E4}y\boldsymbol{e}_{x}+\boldsymbol{u}, & \displaystyle\end{eqnarray}$$

(2.5) $$\begin{eqnarray}\displaystyle & \displaystyle \boldsymbol{V}=\unicode[STIX]{x1D6E4}y\boldsymbol{e}_{x}+\boldsymbol{v}. & \displaystyle\end{eqnarray}$$

Here $\boldsymbol{u}$ and $\boldsymbol{v}$ are the fluctuating fluid and particle velocities, $\unicode[STIX]{x1D6E4}$ is the shear rate, $\boldsymbol{e}_{x}$ is the streamwise direction and $y$ is the coordinate along the cross-direction $\boldsymbol{e}_{y}=-\boldsymbol{g}/g$ which is also parallel to gravity. The fluid fluctuations are given by the following governing equations:

(2.6) $$\begin{eqnarray}\displaystyle & \displaystyle \unicode[STIX]{x1D735}\boldsymbol{\cdot }\boldsymbol{u}=0, & \displaystyle\end{eqnarray}$$

(2.7) $$\begin{eqnarray}\displaystyle & \displaystyle \unicode[STIX]{x1D70C}_{f}\frac{\unicode[STIX]{x2202}\boldsymbol{u}}{\unicode[STIX]{x2202}t}+\unicode[STIX]{x1D70C}_{f}\boldsymbol{u}\boldsymbol{\cdot }\unicode[STIX]{x1D735}\boldsymbol{u}=-\unicode[STIX]{x1D735}p+\unicode[STIX]{x1D707}_{f}\unicode[STIX]{x1D6FB}^{2}\boldsymbol{U}+\unicode[STIX]{x1D70C}_{f}\boldsymbol{g}-\boldsymbol{F}-\unicode[STIX]{x1D70C}_{f}\unicode[STIX]{x1D6E4}u_{y}\boldsymbol{e}_{x}-\unicode[STIX]{x1D70C}_{f}\unicode[STIX]{x1D6E4}y\frac{\unicode[STIX]{x2202}\boldsymbol{u}}{\unicode[STIX]{x2202}x}. & \displaystyle\end{eqnarray}$$

In the absence of particles, the homogeneous shear flow contains only two intrinsic scales: the shear characteristic time $\unicode[STIX]{x1D70F}_{f}=\unicode[STIX]{x1D6E4}^{-1}$ , and the shear viscous dissipation length scale $L_{\unicode[STIX]{x1D707}}=\sqrt{\unicode[STIX]{x1D707}_{f}/(\unicode[STIX]{x1D70C}_{f}\unicode[STIX]{x1D6E4})}$ . The unboundedness of the flow removes any other scale. New scales arise in particle-laden shear subject to gravity. As explained in Kasbaoui et al. (Reference Kasbaoui, Koch, Subramanian and Desjardins2015), the unperturbed sheared particle-phase sediments at the speed of the particle settling velocity $V_{g}=\unicode[STIX]{x1D70F}_{p}g$ . Based on this characteristic velocity and the characteristic time $\unicode[STIX]{x1D70F}_{f}=\unicode[STIX]{x1D6E4}^{-1}$ , non-dimensionalization of the two-fluid equations reveals a characteristic length equal to the settling distance in one shear time, $L_{g}=V_{g}/\unicode[STIX]{x1D6E4}$ . This is a scaling that focuses on the collective dynamics of the particles in shear, rather than the particle-scale dynamics. It follows that a total of four independent dimensionless numbers control this flow: the mass loading $M$ , average volume fraction $\langle \unicode[STIX]{x1D719}\rangle$ , Stokes number $St=\unicode[STIX]{x1D6E4}\unicode[STIX]{x1D70F}_{p}$ and Reynolds number $Re=\unicode[STIX]{x1D70C}_{f}V_{g}L_{g}/\unicode[STIX]{x1D707}_{f}=(L_{g}/L_{\unicode[STIX]{x1D707}})^{2}$ . In this context, the Reynolds number is a measure of the relative strength of inertial accelerations at the scale set by the particles’ sedimentation to the viscous dissipation. It also shows that fluid perturbations induced by slow settling particles are suppressed by viscosity. It is noteworthy that, while shear reinforces preferential concentration (Kasbaoui et al. Reference Kasbaoui, Koch, Subramanian and Desjardins2015), it can also lead to reduced flow modulation by strengthening the dissipation compared to the gravitational effects ( $Re=(L_{g}/L_{\unicode[STIX]{x1D707}})^{2}\propto 1/\unicode[STIX]{x1D6E4}$ ).

A complete description of particle-laden flows requires governing equations for the particle phase. Below we examine three approaches used in the present study.

Euler–Lagrange method (EL): Based on the work of Maxey & Riley (Reference Maxey and Riley1983), the equation of motion of particle ‘ $i$ ’ in homogeneous shear is given by

(2.8) $$\begin{eqnarray}\displaystyle & \displaystyle \frac{\text{d}\boldsymbol{x}^{i}}{\text{d}t}=\boldsymbol{v}^{i}+\unicode[STIX]{x1D6E4}y^{i}\boldsymbol{e}_{x}, & \displaystyle\end{eqnarray}$$

(2.9) $$\begin{eqnarray}\displaystyle & \displaystyle \frac{\text{d}\boldsymbol{v}^{i}}{\text{d}t}=\frac{\boldsymbol{u}(\boldsymbol{x}^{i},t)-\boldsymbol{v}^{i}}{\unicode[STIX]{x1D70F}_{p}}+\boldsymbol{g}, & \displaystyle\end{eqnarray}$$

where $\boldsymbol{x}^{i}$ and $\boldsymbol{v}^{i}$ are respectively the position and fluctuating velocity of this particle. Only Stokes drag is retained here for the same reasons as in (2.3).

Euler–Lagrange approaches rely on solving these equations for all $N$ discrete particles in the flow. The Eulerian particle velocity at the location $\boldsymbol{x}$ is computed from the Lagrangian velocities in the following way:

(2.10) $$\begin{eqnarray}\boldsymbol{v}(\boldsymbol{x},t)=\frac{1}{N}\mathop{\sum }_{i}^{N}v^{i}(t)q(\Vert \boldsymbol{x}-\boldsymbol{x}^{i}\Vert ),\end{eqnarray}$$

where $q$ is a filter kernel with width $\unicode[STIX]{x1D6FF}_{f}$ comparable to the grid spacing $\unicode[STIX]{x0394}x$ . The particle volume fraction $\unicode[STIX]{x1D719}$ is obtained in a similar way. The construction of these Eulerian quantities allow the evaluation of the forcing term (2.3) exerted by the particles on the gas.

Two-fluid Euler–Euler method (TF): Euler–Euler formulations derive from the observation that on scales significantly larger than the particle diameter but smaller than the confining apparatus, the particle phase bears a coherent, organized, fluid-like motion. In these approaches the Eulerian particle velocity $\boldsymbol{v}$ becomes the subject of models and equations to be solved. The discrete particulate view is lost in favour of a continuum description in terms of number density $n=\unicode[STIX]{x1D719}\unicode[STIX]{x03C0}d_{p}^{3}/6$ and particle velocity field $\boldsymbol{v}$ .

In two-fluid methods for semi-dilute suspensions, the population balance is dictated by pure advection by the particle velocity field. The latter is found by taking the derivative in (2.9) along the particle trajectory $\text{d}/\text{d}t=\unicode[STIX]{x2202}/\unicode[STIX]{x2202}t+\boldsymbol{v}\boldsymbol{\cdot }\unicode[STIX]{x1D735}$ . The governing equations for the dispersed phase have the following conservative form:

(2.11) $$\begin{eqnarray}\displaystyle & \displaystyle \frac{\unicode[STIX]{x2202}(\unicode[STIX]{x1D70C}_{p}n)}{\unicode[STIX]{x2202}t}+\unicode[STIX]{x1D735}\boldsymbol{\cdot }(\unicode[STIX]{x1D70C}_{p}n\boldsymbol{v})=-\unicode[STIX]{x1D6E4}y\frac{\unicode[STIX]{x2202}(\unicode[STIX]{x1D70C}_{p}n)}{\unicode[STIX]{x2202}x}, & \displaystyle\end{eqnarray}$$

(2.12) $$\begin{eqnarray}\displaystyle & \displaystyle \frac{\unicode[STIX]{x2202}(\unicode[STIX]{x1D70C}_{p}n\boldsymbol{v})}{\unicode[STIX]{x2202}t}+\unicode[STIX]{x1D735}\boldsymbol{\cdot }(\unicode[STIX]{x1D70C}_{p}n\boldsymbol{v}\boldsymbol{v})=\unicode[STIX]{x1D70C}_{p}n\boldsymbol{g}+\unicode[STIX]{x1D70C}_{p}n\frac{\boldsymbol{u}-\boldsymbol{v}}{\unicode[STIX]{x1D70F}_{p}}-\unicode[STIX]{x1D70C}_{p}n\unicode[STIX]{x1D6E4}v_{y}\boldsymbol{e}_{x}-\unicode[STIX]{x1D6E4}y\frac{\unicode[STIX]{x2202}(\unicode[STIX]{x1D70C}_{p}n\boldsymbol{v})}{\unicode[STIX]{x2202}x}. & \displaystyle\end{eqnarray}$$

The shear decomposition is assumed in the equations above. Together with the fluid’s mass and momentum conservation equations, equations (2.11), (2.12), (2.1) and (2.2) provide a complete set of equations to describe semi-dilute gas–solid flows.

Anisotropic-Gaussian Euler–Euler method (AG): Equation (2.12) relies on the fundamental hypothesis that the particle velocity field is single valued. This is not always true for gas–solid flows. In fact, a feature of inertial particulate flows is the ability of trajectories of particles with different histories to cross leading to multiple particle velocities at a single point. Kinetic based methods recognize this as a limitation and borrow from the rarefied-flow community to handle the crossing characteristics. The idea is to allow the particle velocity field to take multiple values, by solving a probabilistic equation governing the number density probability density function (pdf) $f(t,\boldsymbol{x},\boldsymbol{v})$ . Following Williams (Reference Williams1958), the equation in phase space for the pdf of a dilute system of collisionless particles interacting only through the mean fluid velocity is given by

(2.13) $$\begin{eqnarray}\frac{\unicode[STIX]{x2202}f}{\unicode[STIX]{x2202}t}+\boldsymbol{c}\boldsymbol{\cdot }\unicode[STIX]{x1D735}_{x}\,f+\unicode[STIX]{x1D735}_{c}\boldsymbol{\cdot }\left(f\left(\frac{\boldsymbol{u}-\boldsymbol{v}}{\unicode[STIX]{x1D70F}_{p}}+\boldsymbol{g}\right)\right)=0.\end{eqnarray}$$

Because of the larger space of variables, equation (2.13) is vastly more complicated to solve. To make this method computationally tractable, many methods rely on solving for a finite number of moments instead of the full pdf $f$ . Since the transport of every moment relies on the next-order moment, the difficulty resides in finding a closure to the high-order ones. One approach is to formulate closures in the form of constitutive relations linking high-order moments to low-order ones given in physical space. These are usually applicable for a limited range of Stokes number. See Simonin et al. (Reference Simonin, Fevrier and Laviville2002), Fvrier, Simonin & Squires (Reference Fevrier, Simonin and Squires2005) and Kaufmann et al. (Reference Kaufmann, Moreau, Simonin and Helie2008) for small Stokes number closures, and Masi & Simonin (Reference Masi and Simonin2014) for large Stokes number ones. Another approach is based on postulating a form of the number density pdf itself. With a choice of pdf using only low-order moments, the high-order moments can be found from direct integration of the pdf, hence, providing closure to the moment transport equations (Desjardins et al. Reference Desjardins, Fox and Villedieu2008; Fox Reference Fox2008). This is the method we follow in this paper.

In this work, we focus on the anisotropic-Gaussian closure, originally proposed for rarefied gases (Levermore & Morokoff Reference Levermore and Morokoff1998) and later extended to particle-laden flows (Vi et al. Reference Vi, Doisneau and Massot2015). In this method, the number density pdf is approximated by

(2.14) $$\begin{eqnarray}f(\boldsymbol{c})=\frac{\unicode[STIX]{x1D70C}_{p}n}{(2\unicode[STIX]{x03C0}|\boldsymbol{P}_{p}|)^{2/3}}\exp \left(-\frac{1}{2}(\boldsymbol{c}-\boldsymbol{v})\boldsymbol{\cdot }\boldsymbol{P}_{p}^{-1}(\boldsymbol{c}-\boldsymbol{v})\right),\end{eqnarray}$$

where $\boldsymbol{v}=(\unicode[STIX]{x1D70C}_{p}n)^{-1}\int \boldsymbol{c}f\,\text{d}\boldsymbol{c}$ is the Eulerian particle velocity (first-order moment) and $\boldsymbol{P}_{p}$ is the particle pressure tensor derived from the energy tensor (second-order moment) $\boldsymbol{E}=\int \boldsymbol{c}\boldsymbol{c}f\,\text{d}\boldsymbol{c}=\unicode[STIX]{x1D70C}_{p}n(\boldsymbol{P}_{p}+\boldsymbol{v}\boldsymbol{v})$ . It is clear that moments up to the second order need to be solved in order to build the presumed AG pdf. Using the shear decomposition, these read

(2.15) $$\begin{eqnarray}\displaystyle & \displaystyle \frac{\unicode[STIX]{x2202}\unicode[STIX]{x1D70C}_{p}n}{\unicode[STIX]{x2202}t}+\unicode[STIX]{x1D735}\boldsymbol{\cdot }(\unicode[STIX]{x1D70C}_{p}n\boldsymbol{v})=-\unicode[STIX]{x1D6E4}y\frac{\unicode[STIX]{x2202}\unicode[STIX]{x1D70C}_{p}n}{\unicode[STIX]{x2202}x}, & \displaystyle\end{eqnarray}$$

(2.16) $$\begin{eqnarray}\displaystyle & \displaystyle \frac{\unicode[STIX]{x2202}(\unicode[STIX]{x1D70C}_{p}n\boldsymbol{v})}{\unicode[STIX]{x2202}t}+\unicode[STIX]{x1D735}\boldsymbol{\cdot }\boldsymbol{E}=\unicode[STIX]{x1D70C}_{p}n\boldsymbol{g}+\frac{\unicode[STIX]{x1D70C}_{p}n\boldsymbol{u}-(\unicode[STIX]{x1D70C}_{p}n\boldsymbol{v})}{\unicode[STIX]{x1D70F}_{p}}-\unicode[STIX]{x1D70C}_{p}n\unicode[STIX]{x1D6E4}v_{y}\boldsymbol{e}_{x}-\unicode[STIX]{x1D6E4}y\frac{\unicode[STIX]{x2202}(\unicode[STIX]{x1D70C}_{p}n\boldsymbol{v})}{\unicode[STIX]{x2202}x}, & \displaystyle\end{eqnarray}$$

(2.17) $$\begin{eqnarray}\displaystyle \frac{\unicode[STIX]{x2202}\boldsymbol{E}}{\unicode[STIX]{x2202}t}+\unicode[STIX]{x1D735}\boldsymbol{\cdot }\boldsymbol{Q} & = & \displaystyle \boldsymbol{g}(\unicode[STIX]{x1D70C}_{p}n\boldsymbol{v})+(\unicode[STIX]{x1D70C}_{p}n\boldsymbol{v})\boldsymbol{g}+\frac{(\unicode[STIX]{x1D70C}_{p}n\boldsymbol{v})\boldsymbol{u}+\boldsymbol{u}(\unicode[STIX]{x1D70C}_{p}n\boldsymbol{v})-2\boldsymbol{E}}{\unicode[STIX]{x1D70F}_{p}}\nonumber\\ \displaystyle & & \displaystyle -\,\unicode[STIX]{x1D6E4}(\boldsymbol{E}\boldsymbol{\cdot }\boldsymbol{e}_{\boldsymbol{y}})\boldsymbol{e}_{x}-\unicode[STIX]{x1D6E4}\boldsymbol{e}_{x}(\boldsymbol{e}_{y}\boldsymbol{\cdot }\boldsymbol{E})-\unicode[STIX]{x1D6E4}y\frac{\unicode[STIX]{x2202}\boldsymbol{E}}{\unicode[STIX]{x2202}x},\end{eqnarray}$$

where $\boldsymbol{Q}$ is the particle heat flux third-order tensor, a third-order moment that requires closure. The reconstruction of the AG distribution provides closure by allowing the direct computation of this moment from the presumed pdf $\boldsymbol{Q}=\int \boldsymbol{c}\boldsymbol{c}\boldsymbol{c}f\,\text{d}\boldsymbol{c}$ .

Sabat et al. (Reference Sabat, Vié, Larat and Massot2016) show that the AG method reproduces the same level of clustering found in Euler–Lagrange simulations of one-way coupled decaying homogeneous isotropic turbulence. The agreement holds across the whole range of Stokes number in contrast to the TF method which leads to excessive clustering for moderate and large Stokes numbers. The success of this approach stems from its ability to capture the particle dispersion due to particle-trajectory crossing. It is also argued to be the most likely distribution, in the sense of entropy maximization, given the moments of $f$ up to the second order (Vi et al. Reference Vi, Doisneau and Massot2015).

Note that the AG method can be considered as an extension of the TF method. In the limit of infinitely small Stokes number, the particle pressure tensor vanishes and the AG distribution collapses onto the mono-kinetic distribution $f(t,\boldsymbol{x},\boldsymbol{c})=n(t,\boldsymbol{x})\unicode[STIX]{x1D6FF}(\boldsymbol{c}-\boldsymbol{v}(t,\boldsymbol{x}))$ . Integrating the moments of the population balance equation (2.13) with this distribution yields the TF equations (2.11) and (2.12).

To conclude this section, we shall note that many (Druzhinin Reference Druzhinin1995; Ferry & Balachandar Reference Ferry and Balachandar2001; Vi et al. Reference Vi, Doisneau and Massot2015) have argued, based on one-way coupling studies, that particle trajectory crossing for flows with small volume fraction and small Stokes number is negligible. However, this has not been verified in the two-way coupled semi-dilute regime. In the following, the granular temperature $\unicode[STIX]{x1D6E9}=\text{Tr}(\boldsymbol{P}_{p})/3$ from the AG and EL methods will serve as a measure of the presence of trajectory crossing as this quantity is identically zero in the TF method.

2.2 An algebraic instability

The instability of particle-laden shear is the result of two concomitant effects: preferential concentration and an inertial acceleration or gravity. It is possible to show that preferential concentration is directly related to the extreme compressibility of the particle phase (Maxey Reference Maxey1987). For particles with small inertia, the rate of volumetric expansion and contraction is given by

(2.18) $$\begin{eqnarray}\unicode[STIX]{x1D735}\boldsymbol{\cdot }\boldsymbol{v}\simeq \unicode[STIX]{x1D70F}_{p}(S-R)+O(\unicode[STIX]{x1D70F}_{p}^{2}),\end{eqnarray}$$

where $S=1/2(\unicode[STIX]{x1D735}\boldsymbol{u}+\unicode[STIX]{x1D735}\boldsymbol{u}^{\text{T}}):1/2(\unicode[STIX]{x1D735}\boldsymbol{u}+\unicode[STIX]{x1D735}\boldsymbol{u}^{\text{T}})$ and $R=1/2(\unicode[STIX]{x1D735}\boldsymbol{u}-\unicode[STIX]{x1D735}\boldsymbol{u}^{\text{T}}):1/2(\unicode[STIX]{x1D735}\boldsymbol{u}-\unicode[STIX]{x1D735}\boldsymbol{u}^{\text{T}})$ denote the second invariants of the fluid strain and rotation tensors (Squires & Eaton Reference Squires and Eaton1991). When the local strain exceeds the local rotation particles accumulate. In the opposite scenario of local rotation dominating over strain, the particles are expelled. In pure particle-laden homogeneous shear, strain and rotation perfectly balance each other $S=R=\unicode[STIX]{x1D6E4}/2$ leading to no preferential concentration. Additionally, a cloud of homogeneously distributed particles exerts a uniform gravitational loading on the fluid, the effect of which is an increased hydrostatic pressure. This state is possible but not tenable as minute perturbations quickly alter these balances.

Figure 1 illustrates the way sinusoidal perturbations grow. First, a perturbation to the fluid velocity disturbs the balance of strain and rotation. The perturbation creates alternating regions of dominant strain and dominant rotation. Preferential concentration acts to expel particles from the rotational regions to the straining ones. The migration of the particles results in reinforced gravitational loading exerted on the fluid in the regions with excess particles, and lower loading in the depleted ones. The turning of the wave in the shear flow and the translation of the particles relative to the fluid due to sedimentation allow this inhomogeneous forcing to have components in phase with the initial fluid velocity perturbation. As a result the inhomogeneous gravitational forcing strengthens the initial fluid velocity perturbation.

Figure 1.

A sinusoidal velocity perturbation alters the balance of strain and rotation in homogeneous shear. Regions with excess straining accumulate particles by the preferential concentration mechanism while regions with excess rotation expel particles. As the particle volume fraction turns in the shear flow and translates due to sedimentation, the resulting particle distribution perturbation feeds back to the gas due to the stronger gravitational forcing exerted by the heavier loaded regions of the perturbation.

Most hydrodynamic instabilities, such as Kelvin–Helmholtz and Rayleigh–Taylor instabilities, are said to be normal and exhibit an exponential unbounded growth. On the contrary, in the homogeneous shear instability, perturbations initially grow algebraically with time and reach a finite amplitude at long times. It broadly falls under the category of non-modal instabilities (Schmid Reference Schmid2007). The homogeneous shear pumps energy into the perturbation helping it grow but also has a stabilizing effect over sinusoidal perturbations. A wave of any initial arbitrary orientation, i.e. wavevector $\boldsymbol{k}=(k_{x},k_{y})$ , is rotated by the shear in a few shear times $\unicode[STIX]{x1D70F}_{f}=\unicode[STIX]{x1D6E4}^{-1}$ such that the final orientation is parallel to the shear gradient as illustrated in figure 2. Formally, this is described by

(2.19) $$\begin{eqnarray}\frac{\text{d}\boldsymbol{k}}{\text{d}t}=\unicode[STIX]{x1D735}(\unicode[STIX]{x1D6E4}y\boldsymbol{e}_{x})\boldsymbol{\cdot }\boldsymbol{k}\rightarrow k_{y}(t)=k_{y,0}-k_{x}\unicode[STIX]{x1D6E4}t.\end{eqnarray}$$

When the velocity gradient direction is parallel to gravity, the behaviour depicted in figure 1 and explained above might or might not be sampled depending on the initial orientation of the perturbations. Waves that are initially oriented upstream ( $k_{x}k_{y}>0$ ) sample the normal orientation ( $k_{x}k_{y}=0$ ) for a brief time before the shear further turn them downstream ( $k_{x}k_{y}<0$ ). The normal orientation is a configuration of maximum coupling between the two phases and strongest growth. Downstream waves do not sample this configuration and hence do not grow substantially. Once the wavevector is nearly parallel to both shear gradient and gravity, the result is nearly vertically stacked layers of higher and lower number density. The particle phase cannot drive fluid velocity perturbations anymore and growth ceases completely in the linear regime. The activation of nonlinear states will depend on the magnitude of the initial perturbation.

Figure 2.

The homogeneous shear causes the turning of wave perturbations and the contraction of length scales. The sketch shows an initial upstream sinusoidal perturbation in the particle distribution strained at $\unicode[STIX]{x1D6E4}t=0,1,2$ and $4$ (left to right).

The growth and saturation of the number density has been studied with a linear stability analysis by Kasbaoui et al. (Reference Kasbaoui, Koch, Subramanian and Desjardins2015) for small heavy inertial solid particles in a gas. To make the derivation analytically tractable, an asymptotic solution to (2.12) was used in the LSA. Proposed by Ferry & Balachandar (Reference Ferry and Balachandar2001), the expansion is valid in the limit of small Stokes numbers $(St\ll 1)$ and leads to a relationship between the particle velocity and the local fluid velocity given by

(2.20) $$\begin{eqnarray}\boldsymbol{v}=\boldsymbol{u}+\boldsymbol{V}_{g}-\unicode[STIX]{x1D70F}_{p}\left(\frac{\unicode[STIX]{x2202}\boldsymbol{u}}{\unicode[STIX]{x2202}t}+(\boldsymbol{u}+\boldsymbol{V}_{g})\boldsymbol{\cdot }\unicode[STIX]{x1D735}\boldsymbol{u}+\unicode[STIX]{x1D6E4}y\frac{\unicode[STIX]{x2202}\boldsymbol{u}}{\unicode[STIX]{x2202}x}+\unicode[STIX]{x1D6E4}u_{y}\boldsymbol{e}_{x}\right)+O(\unicode[STIX]{x1D70F}_{p}^{2}).\end{eqnarray}$$

This expansion was introduced to make the analytical treatment of the linear stability analysis tractable. One needs only to solve the number density transport equation (2.11) with the evaluated particle velocity to completely describe the flow. However, there is no a priori way of determining how small the Stokes number needs to be for this expansion to be applicable.

In the LSA, the algebraic, non-modal growth of the instability was demonstrated with sinusoidal disturbances of the form

(2.21) $$\begin{eqnarray}\displaystyle & \displaystyle n(\boldsymbol{x},t)=\langle n\rangle +St\hat{n}(t)\exp (\text{i}(k_{x}x+k_{y}(t)y)), & \displaystyle\end{eqnarray}$$

(2.22) $$\begin{eqnarray}\displaystyle & \displaystyle \boldsymbol{u}(\boldsymbol{x},t)=\hat{\boldsymbol{u}}(t)\exp (\text{i}(k_{x}x+k_{y}(t)y)), & \displaystyle\end{eqnarray}$$

where the wavevector obeys (2.19), $\langle n\rangle =\langle \unicode[STIX]{x1D719}\rangle \unicode[STIX]{x03C0}d_{p}^{3}/6$ is the mean number density and $\hat{n}$ and $\hat{\boldsymbol{u}}$ are the wave amplitudes. The linearized evolution of the latter quantities is given by the coupled ordinary differential equations

(2.23) $$\begin{eqnarray}\frac{\text{d}\hat{n}}{\text{d}t}=-\text{i}\boldsymbol{V}_{g}\boldsymbol{\cdot }\boldsymbol{k}\hat{n}+\frac{\langle n\rangle }{\unicode[STIX]{x1D6E4}}\text{i}2\unicode[STIX]{x1D735}\boldsymbol{u}_{b}:\boldsymbol{k}\hat{\boldsymbol{u}},\end{eqnarray}$$

(2.24) $$\begin{eqnarray}\displaystyle \frac{\text{d}\hat{\boldsymbol{u}}}{\text{d}t} & = & \displaystyle -\!\left(\text{i}\frac{M}{1+M}\boldsymbol{V}_{g}\boldsymbol{\cdot }\boldsymbol{k}+\left(\boldsymbol{I}-2\frac{\boldsymbol{k}\boldsymbol{k}}{k^{2}}\right)\boldsymbol{\cdot }\unicode[STIX]{x1D735}\boldsymbol{u}_{b}^{\text{T}}+\frac{\unicode[STIX]{x1D708}}{1+M}k^{2}\right)\hat{\boldsymbol{u}}\nonumber\\ \displaystyle & & \displaystyle +\,\frac{\unicode[STIX]{x1D6E4}}{\langle n\rangle }\left(\boldsymbol{I}-\frac{\boldsymbol{k}\boldsymbol{k}}{k^{2}}\right)\frac{M}{1+M}\boldsymbol{V}_{g}\hat{n}.\end{eqnarray}$$

Although (2.24) is given in vectorial form, it is sufficient to solve for the velocity component along the gradient direction $y$ only. The other component can be found from the continuity equation

(2.25) $$\begin{eqnarray}k_{x}\hat{u} _{x}+k_{y}\hat{u} _{y}=0.\end{eqnarray}$$

Solutions to (2.23)–(2.25) constitute the LSA result (LSA) and will be compared to the full numerical solutions of the Euler–Lagrange (EL), two-fluid (TF) and anisotropic-Gaussian (AG) simulations.

3.1 Shear periodicity and homogeneous shear treatment

The numerical study of the present flow requires a specially crafted algorithm to capture the homogeneous shear effects. Our approach is described in Kasbaoui et al. (Reference Kasbaoui, Patel, Koch and Desjardins2017). Based on the earlier work of Baron (Reference Baron1982) and Gerz, Schumann & Elghobashi (Reference Gerz, Schumann and Elghobashi1989), the equations are solved in physical space without the need to remesh as in the popular method of Rogallo (Reference Rogallo1981). The key aspect is to enforce the so-called shear-periodic boundary conditions (see figure 3)

(3.1) $$\begin{eqnarray}f(x,Ly,z)=f(x-\unicode[STIX]{x1D6E4}tL_{y},0,z),\end{eqnarray}$$

where $f$ is any quantity of interest like the velocity fluctuation $\boldsymbol{u}$ , and $L_{y}$ is the extent of the computational domain in the shear gradient direction $y$ . The distortion terms $\unicode[STIX]{x1D6E4}y(\unicode[STIX]{x2202}f)/(\unicode[STIX]{x2202}x)$ are added directly in physical space in a split step approach. In the reference frame of the laboratory, these terms are responsible for the turning of the waves. Kasbaoui et al. (Reference Kasbaoui, Patel, Koch and Desjardins2017) show that this method is able to capture the rotation of the Fourier modes by the shear while maintaining their normal structure, an essential capability for the present study.

Figure 3.

Shear-periodic boundary conditions.

3.2 Euler–Euler simulations

The appearance of clusters and void regions in the particle phase, call for special attention to the numerics used in Eulerian simulations. When clustering is significant, the volume fraction can drop significantly to near zero values in the neighbouring depleted regions. Positivity-preserving, also called realizable, Euler–Euler methods maintain positive volume fraction even when the solution is nearly devoid of particles ( $\unicode[STIX]{x1D719}\rightarrow 0$ ). Stability and mesh convergences of solvers for the dispersed phase are contingent on this property (Estivalezes & Villedieu Reference Estivalezes and Villedieu1996; Desjardins et al. Reference Desjardins, Fox and Villedieu2008).

The particle-phase solver follows the quadrature based method of moments (Fox Reference Fox2008; Vi et al. Reference Vi, Doisneau and Massot2015; Kong et al. Reference Kong, Fox, Feng, Capecelatro, Patel, Desjardins and Fox2017). This is a realizable, second-order spatial discretization on a collocated mesh. The numerical strategy relies extensively on Lie operator splitting to ensure that every term in the governing equations is added in a realizable way. The distortion terms are treated in the same manner as in Kasbaoui et al. (Reference Kasbaoui, Patel, Koch and Desjardins2017).

At the beginning of the time step, we start by computing the drag term from the previous time step:

(3.2) $$\begin{eqnarray}\boldsymbol{D}^{n}=\frac{\unicode[STIX]{x1D719}^{n}\boldsymbol{u}^{n}-(\unicode[STIX]{x1D719}\boldsymbol{v})^{n}}{\unicode[STIX]{x1D70F}_{p}},\end{eqnarray}$$

where $(\unicode[STIX]{x1D719}\boldsymbol{v})^{n}$ is the particle momentum at the previous step treated as a whole. Note that the fluid solver mesh is staggered whereas the particle mesh is collocated, making interpolations of quantities from one mesh to the other necessary. At this point, we also compute the second-order drag tensor if using AG. Next, the pure convection of particle volume fraction, momentum and energy is advanced with a third-order Runge–Kutta scheme. For the volume fraction field this reads

(3.3) $$\begin{eqnarray}\displaystyle & \displaystyle \widetilde{\unicode[STIX]{x1D719}}=\unicode[STIX]{x1D719}^{n}-\text{d}t\unicode[STIX]{x1D735}\boldsymbol{\cdot }(\unicode[STIX]{x1D719}\boldsymbol{v})^{n}, & \displaystyle\end{eqnarray}$$

(3.4) $$\begin{eqnarray}\displaystyle & \displaystyle \widetilde{\widetilde{\unicode[STIX]{x1D719}}}={\textstyle \frac{3}{4}}\unicode[STIX]{x1D719}^{n}+{\textstyle \frac{1}{4}}(\widetilde{\unicode[STIX]{x1D719}}-\text{d}t\unicode[STIX]{x1D735}\boldsymbol{\cdot }\widetilde{(\unicode[STIX]{x1D719}\boldsymbol{v})}), & \displaystyle\end{eqnarray}$$

(3.5) $$\begin{eqnarray}\displaystyle & \displaystyle \overline{\unicode[STIX]{x1D719}}={\textstyle \frac{1}{3}}\unicode[STIX]{x1D719}^{n}+{\textstyle \frac{2}{3}}(\widetilde{\widetilde{\unicode[STIX]{x1D719}}}-\text{d}t\unicode[STIX]{x1D735}\boldsymbol{\cdot }\widetilde{\widetilde{(\unicode[STIX]{x1D719}\boldsymbol{v})}}), & \displaystyle\end{eqnarray}$$

where the numerical fluxes are computed with a second-order kinetic flux (Vikas et al. Reference Vikas, Wang, Passalacqua and Fox2011). Combined with a minmod slope limiter, the numerical fluxes are both stable in the presence of shocks and realizable for first-order forward in time integrations. The three stage Runge–Kutta extends this property to a realizable third-order time accurate integration by using convex combinations of realizable states. The overall scheme preserves the volume fraction positivity and is shock capturing.

Further updates are needed to add the remaining forces. In the next step, drag, gravity and convection of homogeneous shear by the fluctuations are added. For the particle momentum this reads

(3.6) $$\begin{eqnarray}\displaystyle \widehat{\unicode[STIX]{x1D719}\boldsymbol{v}}=\overline{\unicode[STIX]{x1D719}\boldsymbol{v}}-\text{d}t(\boldsymbol{D}^{n}+\boldsymbol{g}\unicode[STIX]{x1D719}^{n}-\unicode[STIX]{x1D6E4}(\unicode[STIX]{x1D719}v_{y})^{n}\boldsymbol{e}_{x}). & & \displaystyle\end{eqnarray}$$

This operation is also positivity preserving. The last remaining terms are the distortion terms $\unicode[STIX]{x1D6E4}y\unicode[STIX]{x2202}/(\unicode[STIX]{x2202}x)$ . These are added exactly using the characteristics method as in Kasbaoui et al. (Reference Kasbaoui, Patel, Koch and Desjardins2017)

(3.7) $$\begin{eqnarray}\displaystyle (\unicode[STIX]{x1D719}\boldsymbol{v})^{n+1}(\boldsymbol{x})=\widehat{\unicode[STIX]{x1D719}\boldsymbol{v}}(\boldsymbol{x}-\unicode[STIX]{x0394}t\unicode[STIX]{x1D6E4}y\boldsymbol{e}_{x}) & & \displaystyle\end{eqnarray}$$

and the boundary conditions are applied with time $n+1$ .

The algorithm then solves for the fluid phase using the method explained in Kasbaoui et al. (Reference Kasbaoui, Patel, Koch and Desjardins2017) and with the drag term $-\boldsymbol{D}^{n}$ interpolated to the fluid mesh as a source term.

3.3 Euler–Lagrange simulations

The Euler–Lagrange strategy follows the method of Capecelatro & Desjardins (Reference Capecelatro and Desjardins2013b ). The numerical method uses a volume-filtered approach that captures the particle modulation of the fluid without resolving the flow on the particle scale. The position and velocity of each Lagrangian particle (2.8) and (2.9) is advanced using a second-order Runge–Kutta scheme. Eulerian particle data such as the volume fraction field are computed from the Lagrangian data using a Gaussian kernel of width $\unicode[STIX]{x1D6FF}_{f}$ multiple of the particle diameter. The unperturbed fluid velocity at the particle location is computed using the method of Ireland & Desjardins (Reference Ireland and Desjardins2017). This quantity is required for the correct calculation of the drag force and leads to correct predictions of the particle settling velocity. The method is fully conservative and yields grid-independent solutions in two-way coupled problems and has been verified against experiments (Capecelatro & Desjardins Reference Capecelatro and Desjardins2013a ; Capecelatro, Pepiot & Desjardins Reference Capecelatro, Pepiot and Desjardins2014).

The imposed homogeneous shear necessitates some additions. The shear term in (2.8) generally poses no computational challenges. Particles leaving from the top and bottom of the domain ( $y=\pm L_{y}/2$ ) are reintroduced from the other side with translated $x$ position and velocity. This procedure is the well-known Lees–Edwards boundary conditions (Lees & Edwards Reference Lees and Edwards1972) and is the discrete analogue of the shear-periodic boundary conditions (3.1).

5.1 Nonlinearities in EL formulation

We start with the analysis of the nonlinearities in EL simulations. The case in table 1 is run with an initial perturbation strength $A=0.25$ . At this level, a strong deviation from the linear evolution is guaranteed, since according to the LSA, the sinusoidal volume fraction perturbation would grow so much that the depleted regions, the troughs of the sinusoid, would have a non-physical negative volume fraction. Hence, it is expected that large void regions form and that the sinusoid deforms nonlinearly.

Figure 8.

Snapshots of the volume fraction field in EL simulation for the large perturbation strength $A=0.25$ . A secondary transverse Rayleigh–Taylor instability breaks the sheet of particles formed by the primary preferential concentration. The final state displays sustained clustering.

Snapshots of the volume fraction field during $1\leqslant \unicode[STIX]{x1D6E4}t\leqslant 4$ (see figure 8) show the gradual evolution towards a state dominated by clusters similar to those seen in the grid turbulence experiments of Yang & Shy (Reference Yang and Shy2005) with copper beads ( $\langle \unicode[STIX]{x1D719}\rangle =5\times 10^{-5}$ , $M=0.5$ , $St=0.36$ ). First, the preferential concentration instability causes the extreme accumulation of particles along the oblique direction of the rotating perturbation during $0<\unicode[STIX]{x1D6E4}t<2$ . Nearly all of the particles are distributed on a sheet of small thickness. Despite the stronger nonlinear amplification, this first stage is qualitatively understood from the LSA and the simulations in the linear regime. Shortly after, a transverse instability manifests around $\unicode[STIX]{x1D6E4}t=2.1$ and is clearly visible on the snapshot at $\unicode[STIX]{x1D6E4}t=2.325$ . This secondary instability is of a Rayleigh–Taylor type as we explain below. It adds spatial dimensionality to the particle distribution by breaking the quasi-one-dimensional long band of particles formed by the preferential concentration instability. Notice that this secondary instability grows in a double helix shape: two modes in antiphase grow simultaneously. These modes grow on a time scale much smaller than the shear time $\unicode[STIX]{x1D6E4}^{-1}$ and lead to the formation of clusters seen in the snapshots at $\unicode[STIX]{x1D6E4}t=3$ , 3.5 and 4.

Figure 9.

Evolution of the perturbation in the EL simulation. (a) Shows the fluctuations peak around $\unicode[STIX]{x1D6E4}t\sim 2$ . (b) Shows significant growth of the unperturbed modes $k_{x}^{\star }=2$ and 3. The collapse of these modes coincides with the appearance of the transverse secondary instability. (c,d) Show the volume fraction and wave-normal particle velocity as a function of the wave-normal coordinate $x^{\prime }=\boldsymbol{k}\boldsymbol{\cdot }\boldsymbol{x}/k$ centred on the high concentration band of particles. The steepening of the sinusoidal wave leads to a volume fraction shaped like an impulse ( $\unicode[STIX]{x1D6E4}t=1.6$ and 1.9). Particle-trajectory crossing leads to the spreading of the profiles at $\unicode[STIX]{x1D6E4}t=2.1$ .

5.1.2 Caustics and particle-trajectory crossing

The two antiphase transverse modes forming the double helix (snapshot at $\unicode[STIX]{x1D6E4}t=2.325$ in figure 8) originate from two streams of particles that cross at the high concentration sheet. To understand how this takes place, we report the Eulerian wave-normal particle velocity $v_{x^{\prime }}=(\boldsymbol{k}\boldsymbol{\cdot }\boldsymbol{v})/k$ , where $\boldsymbol{v}$ is the Eulerian particle velocity computed from the Lagrangian data on the mesh. The figure shows clear steepening of the original sinusoidal wave. This nonlinear process occurs as particles migrate to the location of the high concentration sheet. Particles from the two sides move in two opposing streams causing the slope of the centred wave-normal velocity to steepen and the volume fraction impulse to sharpen. When the two streams meet, around $\unicode[STIX]{x1D6E4}t=2.1$ , particle-trajectory crossing takes place, i.e. the two streams inter-penetrate and continue moving in opposite directions. This has four consequences. First, the gradient of the wave-normal velocity $v_{x^{\prime }}$ drops, as seen in figure 9(d), due to the presence of particles with opposing velocities within one grid cell in the crossing zone. Second, the crossing of the streams leads to the spreading of the volume fraction impulse (see figure 9 c). Third, particle-trajectory crossing is known to manifest as caustics (Gustavsson et al. Reference Gustavsson, Meneguz, Reeks and Mehlig2012), long intersecting filaments of aligned particles. This can be seen in figure 10(a) in a zoomed-in view around the high concentration sheet. Fourth, the crossing of the particle streams along with transverse perturbations lead to particles within one grid cell having non-parallel velocity vectors (see figure 10 b).

Figure 10.

Snapshots of the Lagrangian particles in a zoomed-in view around the high concentration band of particles at $\unicode[STIX]{x1D6E4}t=2.05$ . Particle-trajectory crossing in EL simulations leads to: (a) the formation of caustics, long intersecting filaments of particles, and (b) particles within one grid cell with widely different velocity vectors. The arrows indicate the velocity directions of individual particles. The Lagrangian particles have been magnified three times for easier visualization.

Figure 11.

Granular temperature in the EL simulation. (a) Shows a peak of the domain-averaged granular temperature near the time caustics start appearing. (b) Shows the granular temperature peaks sharply at the location of the high concentration sheet of particles.

The spreading of the particle number density field by caustics effectively puts an end to the nonlinear amplification of the grid-resolved mean volume fraction variation due to preferential concentration. A quantitative measure of trajectory crossing is given by the granular temperature defined as the variance of the velocities of particles within a grid cell. Figure 11(a) shows that the domain-averaged granular temperature $\langle \unicode[STIX]{x1D6E9}\rangle$ reaches a peak around the same time $\unicode[STIX]{x1D719}_{rms}$ peaks. Furthermore, figure 11(b) shows this peak is spatially located near the region of particle crossing.

5.1.3 Rayleigh–Taylor instability

The nonlinear amplification of the preferential concentration instability triggers a transverse instability seen at $\unicode[STIX]{x1D6E4}t=2.325$ . The secondary instability happens on a time scale significantly shorter than that of the shear. As a result the one-dimensional wave resulting from the preferential concentration instability at $\unicode[STIX]{x1D6E4}t\sim 2$ can be considered a quasi-steady base state for the secondary instability. The shear forcing is negligible over the short duration of the development of the secondary instability and gravity is left as the sole driving force. To describe this secondary instability, we revert to a simpler formalism treating the particles and fluid as mixture. Using the expansion (2.20), one can show that a simplification of the governing equations of the particle-laden flow at a lower order in Stokes number yields the equations for a variable density fluid

(5.1) $$\begin{eqnarray}\displaystyle & \displaystyle \unicode[STIX]{x1D70C}_{m}\left(\frac{\unicode[STIX]{x2202}\boldsymbol{u}_{f}}{\unicode[STIX]{x2202}t}+\boldsymbol{u}_{f}\boldsymbol{\cdot }\unicode[STIX]{x1D735}\boldsymbol{u}_{f}\right)=-\unicode[STIX]{x1D735}p+\unicode[STIX]{x1D707}\unicode[STIX]{x1D6FB}^{2}\boldsymbol{u}_{f}+(\unicode[STIX]{x1D70C}_{m}-\unicode[STIX]{x1D70C}_{f})\boldsymbol{g}+O(\unicode[STIX]{x1D70F}_{p}), & \displaystyle\end{eqnarray}$$

(5.2) $$\begin{eqnarray}\displaystyle & \displaystyle \unicode[STIX]{x1D735}\boldsymbol{\cdot }\boldsymbol{u}_{f}=0, & \displaystyle\end{eqnarray}$$

(5.3) $$\begin{eqnarray}\displaystyle & \displaystyle \frac{\unicode[STIX]{x2202}\unicode[STIX]{x1D70C}_{m}}{\unicode[STIX]{x2202}t}+\boldsymbol{u}\boldsymbol{\cdot }\unicode[STIX]{x1D735}\unicode[STIX]{x1D70C}_{m}=O(\unicode[STIX]{x1D70F}_{p}), & \displaystyle\end{eqnarray}$$

where $\unicode[STIX]{x1D70C}_{m}=\unicode[STIX]{x1D70C}_{f}+\unicode[STIX]{x1D70C}_{p}\unicode[STIX]{x1D719}$ is the mixture density, and the preferential concentration terms are confined in the $O(\unicode[STIX]{x1D70F}_{p})$ terms. When completely neglecting the latter, the stability studies of unbounded variable density flows by Batchelor & Nitsche (Reference Batchelor and Nitsche1991) become of great relevance. Among the many base states they considered two of them are amenable to the configurations studied here: (i) sinusoidally varying density in the vertical direction, and (ii) long horizontal sheet of high density. Batchelor & Nitsche (Reference Batchelor and Nitsche1991) show that these configurations are subject to a Rayleigh–Taylor instability. Hence, to the extent that the quasi-steady flow and negligible preferential concentrations effects assumptions hold, the transverse secondary instability seen in figure 8 can be considered of a Rayleigh–Taylor type due to the mixture behaving as a variable density fluid.

Figure 12.

Evolution of mode $k_{x}^{\star }=1$ for the initial perturbation strengths $A=0.15$ , 0.20 and 0.25. The later decline of this mode for diminishing $A$ shows that the transverse perturbations are also delayed until a more favourable orientation and amplification are attained.

The results of Batchelor & Nitsche (Reference Batchelor and Nitsche1991) could be used to estimate the growth rate of the secondary modes. Case (i) might describe a transverse Rayleigh–Taylor instability growing at the early stages of the evolution of the mode $k_{x}^{\star }=1$ , i.e. while it remains sinusoidal. We focus on case (ii) since this is almost identical to the final stage of the nonlinear preferential concentration instability, around $\unicode[STIX]{x1D6E4}t=2$ , when the secondary Rayleigh–Taylor instability becomes evident. In the present case, the long band of particles is inclined rather than horizontal. Nevertheless, the results of the analysis of Batchelor & Nitsche (Reference Batchelor and Nitsche1991) can be used replacing the natural gravity by the effective wave-normal gravity $g_{n}=\Vert \boldsymbol{g}\boldsymbol{\cdot }\boldsymbol{k}\Vert /k=gk_{y}/k$ . To make use of the analysis, we will also assume a quasi-steady base state at $\unicode[STIX]{x1D6E4}t=2$ , meaning that the transverse perturbations grow on a time scale significantly smaller than $\unicode[STIX]{x1D6E4}^{-1}$ . Batchelor & Nitsche (Reference Batchelor and Nitsche1991) also assume that the transverse modes evolve in an inviscid fashion and have a wavelength much larger than the thickness of the band. Given these assumptions, a transverse mode with a wavenumber $\unicode[STIX]{x1D6FC}$ grows exponentially at the rate

(5.4) $$\begin{eqnarray}\unicode[STIX]{x1D70E}=\left(\frac{\unicode[STIX]{x03C0}}{2}\right)^{1/4}(\unicode[STIX]{x1D6FC}^{3}lg_{n}^{2})^{1/4}B^{1/2},\end{eqnarray}$$

where $l$ and $B$ represent the width and height of the mixture density impulse: $\unicode[STIX]{x1D70C}_{m}=\unicode[STIX]{x1D70C}_{f}(1+B\exp (-y^{2}/l^{2}))$ and $B\simeq \unicode[STIX]{x1D70C}_{p}/\unicode[STIX]{x1D70C}_{f}\unicode[STIX]{x1D719}_{max}-1$ . The ratio of time scales can be written in the following way:

(5.5) $$\begin{eqnarray}\frac{\unicode[STIX]{x1D70E}}{\unicode[STIX]{x1D6E4}}=\left(\frac{\unicode[STIX]{x03C0}}{2}\right)^{1/4}(\unicode[STIX]{x1D6FC}l)^{3/4}\sqrt{\frac{L_{g}}{l}\frac{k_{y}}{k}\frac{B}{St}}.\end{eqnarray}$$

The dependence on the inverse of the Stokes number shows that for small Stokes particles a quasi-steady assumption ( $\unicode[STIX]{x1D70E}/\unicode[STIX]{x1D6E4}\gg 1$ ) is valid. Expression (5.5) also shows that waves that achieve a more modest nonlinear growth by the preferential concentration instability, i.e. lower $B$ , will destabilize at an orientation closer to horizontal, i.e. larger $k_{y}/k$ . This trend is confirmed in figure 12 representing the evolution of the mode $k_{x}^{\star }=1$ for different initial strengths. The time at which this mode collapses indicates the time at which the transverse Rayleigh–Taylor instability appears. We see that for weaker initial strengths $A$ (and thus lower $B$ ) the secondary instability is delayed, which allows the wave to further turn towards the horizontal.

The relation (5.5) can also be used to estimate the most likely transverse mode to grow. The scaling of the growth rate in $(\unicode[STIX]{x1D6FC}l)^{3/4}$ suggests that shorter wavelengths grow faster and the most likely mode to grow is such that $\unicode[STIX]{x1D6FC}l\sim 1$ . From figure 9(c), we estimate $l\simeq 0.05L_{g}$ , $B\simeq 17M\simeq 9.18$ and $k_{y}/k\simeq 0.67$ at $t=1.9$ . This gives a transverse wavelength $\unicode[STIX]{x1D706}_{RT}=2\unicode[STIX]{x03C0}/\unicode[STIX]{x1D6FC}\simeq 0.31L_{g}$ growing at the fast rate $\unicode[STIX]{x1D70E}\simeq 41\unicode[STIX]{x1D6E4}$ . The estimated wavelength is in agreement with the one seen in the snapshot at $\unicode[STIX]{x1D6E4}t=2.325$ in figure 8. The wavelength of the triggered mode in the transverse direction also depends on the strength of the initial perturbation driving the preferential concentration instability. Figure 13 shows that increasing the perturbation strength $A$ from 0.15 to 0.30 leads to the selection of larger wavelengths, in addition to triggering the instability earlier. A rigorous quantitative description of the most likely transverse perturbation to grow at finite Stokes number requires an analysis beyond the analogy made with the study of Batchelor & Nitsche (Reference Batchelor and Nitsche1991). In particular, the assumptions of quasi-steady base flow and infinitely small Stokes number behind expression (5.5) need to be relaxed.

Figure 13.

Effect of varying the initial perturbation strength $A$ on the orientation and wavelength of the transverse mode.

5.2 Nonlinearities in EE formulations

Having described in detail the nonlinear dynamics revealed in Euler–Lagrange (EL) simulations, we now turn to consider the fidelity with which Euler–Euler formulations capture these effects. In particular, we show results of the two-fluid equations (TF) which consider the particle velocity field to be single valued and the anisotropic-Gaussian (AG) model that allows for multiplicity of particle velocities in a single grid cell. The preferential concentration instability in these simulations is seeded with a fluid velocity perturbation of strength $A=0.25$ as in the EL simulations. Some small perturbation with variations normal to the imposed wave vector is required to trigger secondary instabilities. In the EL simulations, these perturbations were provided by the initial random distribution of particle positions. To provide a close comparison with this case we initialize the EE simulations with a volume fraction field computed based on grid cell averages of the initial discrete particle distribution used in EL simulations.

Figure 14.

Snapshots of the volume fraction field in TF simulation at $A=0.25$ .

Figure 15.

Snapshots of the volume fraction field in AG simulation at $A=0.25$ .

The nonlinear perturbation triggers the same qualitative dynamics in all three simulation approaches. The snapshots of volume fraction, shown in figures 14 and 15 for TF and AG simulations depict similar dynamics as shown in figure 8 for EL simulations: the initially sinusoidal wave of volume fraction sharpens into a highly concentrated particle sheet, which becomes unstable at $\unicode[STIX]{x1D6E4}t\sim 2.3$ to a transverse Rayleigh–Taylor instability leading eventually to a state of sustained clustering. The r.m.s. of the volume fraction over the entire cell grows nonlinearly and peaks at $\unicode[STIX]{x1D6E4}t\sim 2$ in all three methodologies (see figure 16).

Figure 16.

Evolution of the volume fraction fluctuations in the three simulation methods.

While AG and TF simulations lead to a chain of mechanisms similar to the one analysed for EL simulations in § 5.1, the particle sheet developed at $\unicode[STIX]{x1D6E4}t=2.1$ and the clusters developing by $\unicode[STIX]{x1D6E4}t=4$ are finer with higher local concentration in the TF simulations. This behaviour can be linked to the effects of particle-trajectory crossing. The TF method used here is based on a mono-kinetic assumption whereby all particle velocity vectors are equal within one grid cell. Thus it does not capture particle-trajectory crossing. Without the dispersive effect of particle-trajectory crossing, particle concentrations can become very large and particle velocity fields can become nearly discontinuous in space limited only by a slope limiting condition applied in the numerical evaluation of the TF model. Figure 17(b) shows the wave-normal velocity for the times $\unicode[STIX]{x1D6E4}t=1.6,$ 1.9 and 2.1 in TF and EL simulations. One can see that both methods collapse at the two earliest times during the wave steepening. However, at $\unicode[STIX]{x1D6E4}t=2.1$ the velocity profile broadens in the EL simulation as a result of trajectory crossing, while the TF velocity profile retains a large gradient. The shock-like wave normal velocity profile in the TF method leads to a sharp volume fraction impulse as seen in figure 17(a). In contrast, the wave-normal velocity profile broadens in the AG method by $\unicode[STIX]{x1D6E4}t=2.1$ in qualitative agreement with the EL simulations as seen in figure 17(d). This results in a spread out volume fraction profile in AG simulations at $\unicode[STIX]{x1D6E4}t=2.1$ (figure 17 c) as in EL simulations, although this profile appears smoother likely due to limited numerical resolution.

A key element in removing the discontinuity in AG is the incorporation of a poly-kinetic sub-grid model for the particle number density pdf controlled by the particle pressure tensor. The trace of the latter, i.e. the granular temperature, is seen to reach a high peak at the particle sheet location in figure 18(a). Despite differences in the magnitudes reached in AG and EL simulations, the trends are qualitatively similar. In particular, the time for the onset of significant granular temperature indicative of trajectory crossing is $\unicode[STIX]{x1D6E4}t\sim 2$ in both AG and EL simulations as seen in the plot of the domain-averaged granular temperature in figure 18(b).

Thus, the AG method reproduces all the major aspects of the nonlinear state seen in the EL simulations as outlined in § 5.1: the nonlinear growth and sharpening of the wave due to the preferential concentration, particle-trajectory crossing and the secondary Rayleigh–Taylor instability. The most striking difference between the AG and EL results seen by comparing figures 15 and 8 is the presence of smaller wavelength features in the EL simulations. This is seen both during the onset of the secondary instability at $\unicode[STIX]{x1D6E4}t\sim 2$ –2.3 and the final clustered state at $\unicode[STIX]{x1D6E4}t=4$ . The fine-scale features in the EL simulations can be attributed in part to initial sub-grid-scale particle volume fraction fluctuations captured by the randomly distributed particles in the EL formulation which eventually influence the grid-averaged structure. In addition, EE methods by their nature exhibit numerical diffusion of resolved small-scale volume fractions while the EL method is able to retain small-scale volume fractions through the initial growth period so that they can participate in the later time $\unicode[STIX]{x1D6E4}t\geqslant 2$ dynamics.

Figure 17.

Volume fraction and wave-normal velocity as function of the centred wave-normal coordinate in TF (a,b) and AG (c,d) simulations. The inability of the TF method to resolve particle-trajectory crossing leads to a higher volume fraction peak (a) and steeper wave-normal velocity (b) than in EL at $\unicode[STIX]{x1D6E4}t=2.1$ . The assumption of anisotropic-Gaussian particle number density pdf leads to a spread out volume fraction field (c) and lower gradient of the wave-normal velocity (d) at $\unicode[STIX]{x1D6E4}t=2.1$ than in TF.

Figure 18.

Granular temperature in AG simulations. Similarly to EL, graphs of the granular temperature in the wave-normal direction (a) display a sharp peak at the particle sheet location, although lower to the one observed in EL. The domain-averaged granular temperature (b) in AG and EL display qualitatively similar evolutions. The peak at $\unicode[STIX]{x1D6E4}t\sim 2$ due to the two particle streams crossing at the high concentration band is well reproduced in AG.