3.1. Two-way coupling simulation

The DNS of the particle-laden turbulence is explained below. The volume-averaged equations (2.10) and (2.11) are also regarded as the basic equations of a general two-way coupling simulation. For clarity about the averaging volume size, the notations $V_s$, $\alpha _s$ and $\langle\, \cdot\, \rangle _s$ are used for the DNS instead of $V$, $\alpha _c$ and $\langle\, \cdot\, \rangle _c$, respectively. The volume $V_s$ is defined as the sphere of radius $R_s$. In the DNS, the characteristic length of the averaging volume is comparable to the grid spacing ($\Delta x$) and the minimum scale of the background turbulence. Therefore, the fluid residual stress $\boldsymbol {\tau }_c$ is negligible and omitted from (2.11), whereas the model of the interaction force $\boldsymbol {F}$ is necessary. To treat the finite-sized particle comparable to the Kolmogorov length scale, we use the models for the fluid force on the particle surface developed by Fukada et al. (Reference Fukada, Takeuchi and Kajishima2016, Reference Fukada, Fornari, Brandt, Takeuchi and Kajishima2018, Reference Fukada, Takeuchi and Kajishima2019, Reference Fukada, Takeuchi and Kajishima2020), which are explained in the following.

The particles are tracked individually by the equations

(3.1)$$\begin{gather} \frac{{\rm d}\kern0.06em \boldsymbol{x}_p}{{\rm d}t}=\boldsymbol{w}_p, \end{gather}$$

(3.2)$$\begin{gather}\frac{{\rm d}\boldsymbol{w}_p}{{\rm d}t}=\frac{\boldsymbol{f}}{m_p}, \end{gather}$$

(3.3)$$\begin{gather}\frac{{\rm d}\boldsymbol{\varOmega}_p}{{\rm d}t}=\frac{{\rm \pi}\rho_c\nu d_p^{3}}{I_p} \left(\frac{1}{2}\boldsymbol{\nabla}\times\boldsymbol{U}_{ud}-\boldsymbol{\varOmega}_p\right), \end{gather}$$

where $\boldsymbol {x}_p$ is the particle centre, $m_p$ is the particle mass, $\boldsymbol {f}$ is the fluid force on the individual particle, $\boldsymbol {\varOmega }_p$ is the particle angular velocity, $I_p$ is the moment of inertia, $d_p$ is the particle diameter, and $\boldsymbol {U}_{ud}$ is the estimation of the undisturbed fluid velocity (Fukada et al. Reference Fukada, Fornari, Brandt, Takeuchi and Kajishima2018). According to Fukada et al. (Reference Fukada, Fornari, Brandt, Takeuchi and Kajishima2018, Reference Fukada, Takeuchi and Kajishima2020), $\boldsymbol {f}$ is modelled as

(3.4)\begin{equation} \boldsymbol{f}=f_F({Re}_s,Q)\,\boldsymbol{m} -\frac{{\rm \pi} d_p^{3}}{4}\,\boldsymbol{\nabla} P_{ud} -{\rho_c}\,\frac{{\rm \pi} d_p^{3}}{12}\,\frac{{\rm d}\boldsymbol{w}_p}{{\rm d}t}, \end{equation}

where

(3.5)\begin{equation} {Re}_s=\frac{\alpha_s\left|\left\langle\boldsymbol{u}\right\rangle_s(\boldsymbol{x}_p)-\boldsymbol{w}_p\right|d_p}{\nu} \end{equation}

is the particle Reynolds number based on the volume-averaged velocity relative to the particle,

(3.6)\begin{equation} \boldsymbol{m}=\frac{\left\langle\boldsymbol{u}\right\rangle_s (\boldsymbol{x}_p)-\boldsymbol{w}_p}{\left|\left\langle\boldsymbol{u}\right\rangle_s(\boldsymbol{x}_p)-\boldsymbol{w}_p\right|} \end{equation}

is the unit vector in the direction of the relative velocity, $P_{ud}$ is the estimation of the undisturbed fluid pressure, and $Q=2R_s/d_p$ is the radius ratio between $V_s$ and the particle. The first term on the right-hand side of (3.4) indicates the viscous contribution, and the other terms are the effect of the acceleration. The viscous contribution is modelled based on the disturbed velocity $\langle \boldsymbol {u}\rangle _s$ instead of the undisturbed velocity, and this estimation reflects partially the history effect (Fukada et al. Reference Fukada, Fornari, Brandt, Takeuchi and Kajishima2018), which is an advantage of the present model as other undisturbed-flow-based models require a specific history model. As the value of $\langle \boldsymbol {u}\rangle _s$ depends on $R_s$, the viscous contribution $f_F$ includes the effect of $R_s$ as the non-dimensional parameter $Q$. The function $f_F$ is modelled based on the particle-resolved simulation around a single particle (Fukada et al. Reference Fukada, Takeuchi and Kajishima2019, Reference Fukada, Takeuchi and Kajishima2020):

(3.7)\begin{equation} f_F=3{\rm \pi}\rho_c\nu^{2}\,A_F(Q)\,{Re}_s\{1+0.15\,{Re}_s^{B_F(Q)}\}, \end{equation}

where $A_F$ and $B_F$ are fitting functions. As $Q$ becomes larger, $A_F$ and $B_F$ approach $1$ and $0.687$, respectively, and $f_F$ coincides with the Schiller–Naumann correlation (Clift, Grace & Weber Reference Clift, Grace and Weber1978).

The reaction force on the fluid is expressed as

(3.8)\begin{equation} \boldsymbol{F}=\frac{1}{V}\sum_{particles} \boldsymbol{F}_F, \end{equation}

where the function $\boldsymbol {F}_F$ is part of the reaction force from one particle modelled as a function of the relative direction $\boldsymbol {m}\boldsymbol {\cdot}(\boldsymbol {x}-\boldsymbol {x}_p)$ and the distance $|\boldsymbol {x}-\boldsymbol {x}_p|$ as well as the fluid force $\boldsymbol {f}$, to reflect the surface stress distribution on the particle (Fukada et al. Reference Fukada, Takeuchi and Kajishima2020). The model $\boldsymbol {F}_F$ is constructed to satisfy the momentum conservation

(3.9)\begin{equation} \int_{{whole\ space}}\boldsymbol{F}\,{\rm d}V={-}\sum_{particles}\boldsymbol{f}. \end{equation}

In contrast to conventional two-way coupling simulations, the consistency between $\boldsymbol {F}$ in (2.11) and $\boldsymbol {f}$ computed with $\langle \boldsymbol {u}\rangle _s$ was established, as the common averaging volume $V_s$ is applied. For more detail, see Fukada et al. (Reference Fukada, Takeuchi and Kajishima2016, Reference Fukada, Fornari, Brandt, Takeuchi and Kajishima2018, Reference Fukada, Takeuchi and Kajishima2019, Reference Fukada, Takeuchi and Kajishima2020). By determining $R_s=(\sqrt {3}/2)\,\Delta x$, (2.10), (2.11), (3.1), (3.2) and (3.3) are solved numerically with the uniform grid. For the fluid phase, the variables $(\langle \boldsymbol {u}\rangle _s,\langle p\rangle _s)$ are defined at staggered grid points, and the second-order central difference scheme is used for the spatial derivatives. The volume fraction $\alpha _s$ is computed directly based on the relative position $|\boldsymbol {x}-\boldsymbol {x}_p|$ and the sizes of the averaging volume $R_s$ and the particle $d_p$. The second-order Runge–Kutta method is applied for the convective and viscous terms in (2.11) and for (3.1)–(3.3). The pressure is obtained by solving the Poisson equation constructed by substituting the intermediate velocity from (2.11) into (2.10) as the term $\partial \alpha _s/\partial t$ is determined explicitly (here, subscript $c$ is replaced with $s$). This procedure is an extension of the fractional step method (Kim & Moin Reference Kim and Moin1985) to the multiphase flow. The detail of the treatment of the pressure is found in Fukada et al. (Reference Fukada, Takeuchi and Kajishima2016). The validation of the numerical method is shown in Appendix A.

3.2. Numerical condition

For the initial condition, the particles are located regularly at cubic grid points. Initially, the fluid velocity is zero, and the particle translational and angular velocities are also zero.

The forced turbulence is considered in the cubic periodic box of length $L_{cube}$. The forcing method is according to Eswaran & Pope (Reference Eswaran and Pope1988). The external force term in (2.11) is

(3.10)\begin{equation} \alpha_c\boldsymbol{g}_c=\sum_{0<|\boldsymbol{k}|\leqslant\sqrt{2}k_0}\boldsymbol{a}_{\boldsymbol{k}} \exp({\rm i}\boldsymbol{k}\boldsymbol{\cdot}\boldsymbol{x}), \end{equation}

where $\boldsymbol {k}=(k_x,k_y,k_z)$ is the wavenumber vector, $k_0=2{\rm \pi} /L_{cube}$ is the minimum wavenumber, and $\boldsymbol {a}_{\boldsymbol {k}}$ is the complex vector to be given in the simulation; for the isotropic turbulence simulation, $\boldsymbol {a}_{\boldsymbol {k}}$ is given randomly, and the forcing parameters are the acceleration variance ($\sigma ^{2}$) and the time scale ($T_L$), which will be detailed in Appendix B.

The Smagorinsky constant for the single-phase LES is influenced by the forcing condition (Germano et al. Reference Germano, Piomelli, Moin and Cabot1991). To compare the behaviour of the particle stress models, the anisotropic forcing condition is also applied; the unidirectional and constant complex vector $\boldsymbol {a}_{\boldsymbol {k}}$ is given as

(3.11)\begin{equation} \boldsymbol{a}_{\boldsymbol{k}}=(\sigma,0,0) \quad {\rm for}\ \boldsymbol{k}=(0,\pm k_0,0), \end{equation}

and the external force term becomes $\alpha _c\boldsymbol {g}_c=(2\sigma \cos (2{\rm \pi} y/L_{cube}),0,0)$.

Table 1 shows the computational conditions and the results of the single-phase turbulence as a reference for the particle-laden turbulence. We adopt two cases for the number of computational cells $N_{cell}=256^{3}$ and $384^{3}$, and distinguish the case names by appending I (i.e. isotropic) or U (i.e. unidirectional) to $N_{cell}^{1/3}$. The time increment is $\Delta t=1.7\times 10^{-6}(\nu k_0^{2})^{-1}$ for the cases I256 and U256, and $\Delta t=8.5\times 10^{-7}(\nu k_0^{2})^{-1}$ for the other cases. The numerical results in the present study are averaged from $2\times 10^{5}$ to $4\times 10^{5}$ time steps unless noted otherwise. Although the mean velocity over the entire domain is not always zero, this value does not influence the energy spectra and the other statistical quantities (e.g. dissipation rate and particle stress) in this study.

Table 1.

Numerical condition and results of single-phase turbulence. Here, $\sigma$ and $T_L$ are the intensity of acceleration and the time scale, respectively. For more detail, refer to Appendix B.

The Kolmogorov length scale is

(3.12)\begin{equation} \eta=\left(\frac{\nu^{3}}{\epsilon}\right)^{1/4}, \end{equation}

and the Reynolds number based on the Taylor length scale is

(3.13)\begin{equation} {Re}_{\lambda}=\frac{u_{rms}\lambda}{\nu}, \end{equation}

where $\epsilon$ is the dissipation rate, $u_{rms}$ is the r.m.s. value of each component of velocity, and $\lambda$ is the Taylor length scale computed by

(3.14)\begin{equation} \lambda=\sqrt{\frac{15\nu u_{rms}^{2}}{\epsilon}}. \end{equation}

Although (3.12)–(3.14) are used conventionally for isotropic turbulence, the same definitions are applied for the unidirectional forcing cases (i.e. U256 and U384). Figure 1 shows some snapshots of the velocity field on the $x$–$y$ and $y$–$z$ planes. Three-dimensional turbulent flows are confirmed for both the isotropic and unidirectional forcing cases. According to Yeung, Sreenivasan & Pope (Reference Yeung, Sreenivasan and Pope2018), the resolution $\eta /\Delta x=0.5$ is adequate for low-order statistics. Although the second-order centred finite difference is used in the present study, the values of $\eta /\Delta x$ are larger than 0.5 for all the cases. The effect of the grid resolution is assessed in Appendix A.

Figure 1.

Velocity vector for (a,b) case I256, and (c,d) case U256 on (a,c) the $x$–$y$ plane, and (b,d) the $y$–$z$ plane.

Table 2 shows the numerical conditions for particle-laden turbulence. To study the effect of the particle inertia, particles with three kinds of diameters (distinguished as D1, D2 and D3) are added to the base settings in table 1. The particle density is $\rho _p=1000\rho _c$, and the dispersed phase volume fraction in the whole domain is $1.0\times 10^{-4}$ for all cases; for example, the numbers of particles in I256D1 and I256D3 are 85 921 and 3182, respectively. The Stokes number ${St}_K$ is defined by

(3.15)\begin{equation} {St}_K=\frac{\tau_p}{\tau_K}, \end{equation}

where $\tau _p$ is the particle relaxation time scale,

(3.16)\begin{equation} \tau_p=\frac{\rho_p d_p^{2}}{18\rho_c\nu}, \end{equation}

and $\tau _K$ is the Kolmogorov time scale,

(3.17)\begin{equation} {\tau_K}=\sqrt{\frac{\nu}{\epsilon}}. \end{equation}

The variables $\eta$ and ${\tau _K}$ are based on the corresponding results for the single-phase turbulence, and ${St}_K$ is related directly to the particle size; a case with a larger diameter shows a larger ${St}_K$ value (see table 2). The particle diameter is comparable to the Kolmogorov length scale. The corresponding range of ${St}_K$ is up to $O(10^{2})$, which is larger than in the previous study (Moreau et al. Reference Moreau, Simonin and Bédat2010).

Table 2.

Numerical condition for particle-laden turbulence. The base setting indicates the number of grid cells and the forcing condition in table 1.

5.1. Models of particle stress

To extract the effect of the direction of the particle stress tensor $\boldsymbol {\tau }_{d}$, the deviatoric and isotropic parts are compared individually with their corresponding models. The isotropic and deviatoric parts of the tensors are denoted by the subscripts ‘iso’ and ‘dev’, respectively; for example, the particle stress $\boldsymbol {\tau }_{d}$ obtained by DNS is decomposed into

(5.1)$$\begin{gather} \boldsymbol{\tau}_{d,{iso}}=\tfrac{1}{3}\left({\rm tr}\,\boldsymbol{\tau}_d\right){\boldsymbol{\mathsf{I}}}, \end{gather}$$

(5.2)$$\begin{gather}\boldsymbol{\tau}_{d,{dev}}=\boldsymbol{\tau}_{d}-\boldsymbol{\tau}_{d,{iso}}, \end{gather}$$

where ${\boldsymbol{\mathsf{I}}}$ is the identity tensor.

The fluid residual stress is regarded as the particle stress model as the particles receive the fluid force in the direction of decreasing relative velocities to the fluid. The particle Smagorinsky model and the particle scale-similarity model are also introduced as models that represent implicitly the effect of the fluid motion. The fluid residual stress model for the dispersed phase $\boldsymbol {\tau }_d^{f}$ is given by $\boldsymbol {\tau }_c$ of (2.12) as

(5.3)\begin{equation} \boldsymbol{\tau}_d^{f}=\boldsymbol{\tau}_c; \end{equation}

the particle Smagorinsky model is only the deviatoric component as

(5.4)\begin{equation} \boldsymbol{\tau}_{d}^{pS}=\boldsymbol{\tau}_{d,{dev}}^{pS} ={-}(2R)^{2}|{\boldsymbol{\mathsf{S}}}_p|{\boldsymbol{\mathsf{S}}}_p, \end{equation}

where $R$ is the radius of the averaging volume $V$, with

(5.5)$$\begin{gather} {\boldsymbol{\mathsf{S}}}_p=\boldsymbol{\nabla}\langle\boldsymbol{w}\rangle_d+ \left(\boldsymbol{\nabla}\langle\boldsymbol{w}\rangle_d\right)^{\rm T}- \tfrac{2}{3}\left(\boldsymbol{\nabla}\boldsymbol{\cdot}\langle\boldsymbol{w} \rangle_d\right){\boldsymbol{\mathsf{I}}}, \end{gather}$$

(5.6)$$\begin{gather}|{\boldsymbol{\mathsf{S}}}_p|=\sqrt{\tfrac{1}{2}{\boldsymbol{\mathsf{S}}}_p\colon{\boldsymbol{\mathsf{S}}}_p}; \end{gather}$$

and the particle scale-similarity model is

(5.7)\begin{equation} \boldsymbol{\tau}_d^{pB}= \left( \left\langle\left\langle \boldsymbol{w}\right\rangle_d\left\langle\boldsymbol{w}\right\rangle_d\right\rangle_d -\left\langle\left\langle \boldsymbol{w}\right\rangle_d\right\rangle_d\left\langle\left\langle \boldsymbol{w}\right\rangle_d\right\rangle_d \right). \end{equation}

The models $\boldsymbol {\tau }_d^{pS}$ and $\boldsymbol {\tau }_d^{pB}$ are the counterparts of the models for the single-phase turbulence (Smagorinsky Reference Smagorinsky1963; Bardina, Ferziqer & Reynolds Reference Bardina, Ferziqer and Reynolds1983). For the isotropic part, we use the model introduced by Moreau et al. (Reference Moreau, Simonin and Bédat2010):

(5.8)\begin{equation} \boldsymbol{\tau}_{d}^{pY}=\boldsymbol{\tau}_{d,{iso}}^{pY}= (2R)^{2}|{\boldsymbol{\mathsf{S}}}_p|^{2}{\boldsymbol{\mathsf{I}}}. \end{equation}

The models $\boldsymbol {\tau }_{d,{iso}}^{pB}=(1/3)({\rm tr}\,\boldsymbol {\tau }_d^{pB}){\boldsymbol{\mathsf{I}}}$ and $\boldsymbol {\tau }_{d,{iso}}^{pY}$ for the isotropic part of the particle stress are similar as both represent the fluctuation intensity of the particle velocity at the scale larger than $R$. Note that $\boldsymbol {\nabla }\langle \boldsymbol {w}\rangle _d$ is determined uniquely as long as $\alpha _d>0$ because the derivatives of $\alpha _d$ and $\alpha _d\left \langle \boldsymbol {w}\right \rangle _d$ are well-defined (Fukada et al. Reference Fukada, Fornari, Brandt, Takeuchi and Kajishima2018). However, $\left \langle \boldsymbol {w}\right \rangle _d$ is the $C^{1}$ function, and the evaluation of $\boldsymbol {\nabla }\langle \boldsymbol {w}\rangle _d$ requires a spatial resolution that is finer than the averaging volume, which is not adequate practically. Therefore, the gradient $\boldsymbol {\nabla }$ in (5.5) is replaced with a discretisation operator $\tilde {\boldsymbol {\nabla }}$, and the particle Smagorinsky model relates the smoothed velocity fluctuation at the grid scale and the particle stress. The discretisation operator $\tilde {\boldsymbol {\nabla }}=(\tilde {\partial }_x,\tilde {\partial }_y,\tilde {\partial }_z)$ is defined as

(5.9)\begin{equation} \tilde{\partial}_i \langle\boldsymbol{w}\rangle_d=\frac{-\left\langle\boldsymbol{w}\right\rangle_d (\boldsymbol{x}-(1/2)\,\Delta X\, \boldsymbol{e}_i)+\left\langle\boldsymbol{w}\right\rangle_d (\boldsymbol{x}+(1/2)\,\Delta X\,\boldsymbol{e}_i)}{\Delta X}, \end{equation}

where $i$ is $x$, $y$ or $z$, and $\boldsymbol {e}_i$ the unit vector in the $i$ direction. The average represented by the outer brackets in the particle scale-similarity model (5.7) is computed with the values at the seven points $\boldsymbol {x}\pm (1/2)\,\Delta X\,\boldsymbol {e}_x$, $\boldsymbol {x}\pm (1/2)\,\Delta X\,\boldsymbol {e}_y$, $\boldsymbol {x}\pm (1/2)\,\Delta X\,\boldsymbol {e}_z$ and $\boldsymbol {x}$. The value of $\left \langle \boldsymbol {w}\right \rangle _d$ in (5.7) and (5.9) is computed directly from the numerical results.

The particle stress models are used by multiplying model coefficients $C$ with the corresponding subscript and superscript:

(5.10a–c)$$\begin{gather} C^{pS}_{dev}\boldsymbol{\tau}_{d,{dev}}^{pS}, \quad C^{pB}_{dev}\boldsymbol{\tau}_{d,{dev}}^{pB}, \quad C^{f}_{dev}\boldsymbol{\tau}_{d,{dev}}^{f}, \end{gather}$$

(5.11a–c)$$\begin{gather}C^{pY}_{iso}\boldsymbol{\tau}_{d,{iso}}^{pY}, \quad C^{pB}_{iso}\boldsymbol{\tau}_{d,{iso}}^{pB}, \quad C^{f}_{iso}\boldsymbol{\tau}_{d,{iso}}^{f}. \end{gather}$$

As the model coefficient $C^{pS}_{dev}$ is affected by the particle relaxation time and the time scale of the flow, $C^{pS}_{dev}$ is not a constant, in contrast to the Smagorinsky model for the single-phase turbulence. Although practical simulations require modelling of $\boldsymbol {\tau }_c$, the exact value of $\boldsymbol {\tau }_c$ obtained from the two-way coupling DNS result is used in the present study to focus on the relation between the particle and fluid unresolved motions. To compute the particle stress models in the two-way coupling DNS framework, the radius of the averaging volume $R$ and the virtual grid spacing $\Delta X$ need to be determined. For a large $R$ case, even though ${St}_K$ is larger than 1, the motion of the particles depends on the flow at the averaging scale. To investigate the effect of the flow at a scale larger than $\eta$, relatively large averaging volumes of $R/\eta =O(10)$ are considered, although this scale is not sufficient for a practical LES. The following $R$ values are employed: $k_0 R=0.75$, $1.13$ and $1.50$ (correspondingly, $R/\eta =44$, $66$ and $87$ for the case I256; see also $k_0\eta$ values in table 1). For each $k_0 R$ in each simulation condition, we compare the fully resolved particle stresses and their models at $20\,000$ different combinations of $\boldsymbol {x}$ and $t$.

Figure 5 shows the influence of $\Delta X$ on $\tilde {\partial }_x \langle w_x\rangle _d$ and $\tilde {\partial }_x \langle w_y\rangle _d$ along a line in the $x$ direction for the case of the smallest number of particles (I256D3) and the smallest averaging volume ($k_0R=0.75$). The intensity of the fluctuation is remarkably large for the smallest $\Delta X/R$ (${=}\sqrt {3}/6$), while the intensity is suppressed and similar trends are obtained for the other cases ($\Delta X/R=2\sqrt {3}/3$ and $\sqrt {3}$). In the present study, $\Delta X/R$ is set to be $2\sqrt {3}/3$ to reflect the collective motion of the particles that is not sensitive to small random disturbances.

Figure 5.

Components of discretised gradient (a) $\tilde {\partial }_x\langle w_x\rangle _d$ and (b) $\tilde {\partial }_x\langle w_y\rangle _d$, along a line parallel to the $x$-axis, for the case I256D3 with $k_0 R=0.75$.

To compare the roles of $\boldsymbol {\tau }_{d,{iso}}$ and $\boldsymbol {\tau }_{d,{dev}}$, the energy transfer of the dispersed phase between the grid scale and the subgrid scale is studied. By multiplying (2.6) by $\left \langle \boldsymbol {w}\right \rangle _d$, we obtain the equation corresponding to the energy transport of the dispersed phase at the grid scale:

(5.12)\begin{align} &\frac{\partial}{\partial t}\left(\frac{1}{2}\,\alpha_d\left\langle\boldsymbol{w}\right\rangle_d \boldsymbol{\cdot}\left\langle\boldsymbol{w}\right\rangle_d\right) +\boldsymbol{\nabla}\boldsymbol{\cdot}\left\{\frac{1}{2}\,\alpha_d\left\langle \boldsymbol{w}\right\rangle_d\left(\left\langle\boldsymbol{w}\right\rangle_d \boldsymbol{\cdot}\left\langle\boldsymbol{w}\right\rangle_d\right)\right\}\nonumber\\ &\quad ={-}\boldsymbol{\nabla}\boldsymbol{\cdot}\left(\alpha_d \boldsymbol{\tau}_d\boldsymbol{\cdot}\left\langle\boldsymbol{w}\right\rangle_d\right) +\alpha_d\boldsymbol{\tau}_d\colon\boldsymbol{\nabla}\langle\boldsymbol{w}\rangle_d +\boldsymbol{F}\boldsymbol{\cdot}\left\langle\boldsymbol{w}\right\rangle_d+ \alpha_d\boldsymbol{g}_d\boldsymbol{\cdot}\left\langle\boldsymbol{w}\right\rangle_d, \end{align}

where the second term of the right-hand side is the energy transfer between the scales owing to the particle stress. This energy transfer term is decomposed as

(5.13)\begin{equation} \boldsymbol{\tau}_d\colon\tilde{\boldsymbol{\nabla}}\langle\boldsymbol{w}\rangle_d= \boldsymbol{\tau}_{d,{iso}}\colon\tilde{\boldsymbol{\nabla}}\langle\boldsymbol{w}\rangle_d +\boldsymbol{\tau}_{d,{dev}}\colon\tilde{\boldsymbol{\nabla}}\langle\boldsymbol{w}\rangle_d, \end{equation}

where the gradient is replaced by $\tilde {\boldsymbol {\nabla }}=(\tilde {\partial }_x,\tilde {\partial }_y,\tilde {\partial }_z)$ defined by (5.9), and $\alpha _d$ is omitted. Figure 6 shows the p.d.f.s of $\boldsymbol {\tau }_d\colon \tilde {\boldsymbol {\nabla }}\left \langle \boldsymbol {w}\right \rangle _d$, $\boldsymbol {\tau }_{d,{iso}}\colon \tilde {\boldsymbol {\nabla }}\left \langle \boldsymbol {w}\right \rangle _d$ and $\boldsymbol {\tau }_{d,{dev}}\colon \tilde {\boldsymbol {\nabla }}\left \langle \boldsymbol {w}\right \rangle _d$ for the case I256D2 and $k_0 R=1.13$. As the magnitudes of the isotropic and deviatoric contributions are comparable, the particle stress models of both parts are important. The energy transfer corresponding to the isotropic part $\boldsymbol {\tau }_{d,{iso}}\colon \tilde {\boldsymbol {\nabla }}\left \langle \boldsymbol {w}\right \rangle _d$ is related to the divergence $\tilde {\boldsymbol {\nabla }}\boldsymbol {\cdot }\left \langle \boldsymbol {w}\right \rangle _d$; the p.d.f. of $\boldsymbol {\tau }_{d,{iso}}\colon \tilde {\boldsymbol {\nabla }}\left \langle \boldsymbol {w}\right \rangle _d/\epsilon$ takes a symmetric distribution centred at 0, although this is caused by a mechanism that is different from the divergence-free condition of the incompressible flow. The contribution of the deviatoric part $\boldsymbol {\tau }_{d,{dev}}\colon \tilde {\boldsymbol {\nabla }}\left \langle \boldsymbol {w}\right \rangle _d$ tends to have a negative value, and the sum of both contributions has a higher probability of being a negative value, meaning that the energy is transferred from the grid scale to the subgrid scale. Therefore, the particle stress model for the deviatoric part is important for the prediction of the net energy transfer.

Figure 6.

P.d.f.s of the energy transfer of the dispersed phase $\boldsymbol {\tau }_d\colon \tilde {\boldsymbol {\nabla }}\left \langle \boldsymbol {w}\right \rangle _d$ and the components $\boldsymbol {\tau }_{d,{iso}}\colon \tilde {\boldsymbol {\nabla }}\left \langle \boldsymbol {w}\right \rangle _d$ and $\boldsymbol {\tau }_{d,{dev}}\colon \tilde {\boldsymbol {\nabla }}\left \langle \boldsymbol {w}\right \rangle _d$ for the case I256D2 with $k_0 R=1.13$.

5.2. Principal axes of the particle stress

The agreement of the principal axes (APA) of the particle stress models with those of the fully resolved particle stress is important to predict the sign of the energy transfer term. The eigenvalues of the tensor $\boldsymbol {\tau }$ are denoted as $\lambda _{\alpha }(\boldsymbol {\tau })$, $\lambda _{\beta }(\boldsymbol {\tau })$ and $\lambda _{\gamma }(\boldsymbol {\tau })$, and the corresponding eigenvectors (principal axes) are $\boldsymbol {e}_{\alpha }(\boldsymbol {\tau })$, $\boldsymbol {e}_{\beta }(\boldsymbol {\tau })$ and $\boldsymbol {e}_{\gamma }(\boldsymbol {\tau })$, respectively. The eigenvalues and eigenvectors are distinguished by $\lambda _{\alpha } > \lambda _{\beta } > \lambda _{\gamma }$; a case of two identical eigenvalues does not occur in this study. A degree of APA is given by the following direction cosines:

(5.14)$$\begin{gather} \cos\theta_{\alpha}=\left|\boldsymbol{e}_{\alpha}(\boldsymbol{\tau}_{d,{ dev}})\boldsymbol{\cdot}\boldsymbol{e}_{\alpha}(\boldsymbol{\tau}_{d,{dev}}^{model})\right|, \end{gather}$$

(5.15)$$\begin{gather}\cos\theta_{\beta}=\left|\boldsymbol{e}_{\beta}(\boldsymbol{\tau}_{d,{ dev}})\boldsymbol{\cdot}\boldsymbol{e}_{\beta}(\boldsymbol{\tau}_{d,{dev}}^{ model})\right|, \end{gather}$$

(5.16)$$\begin{gather}\cos\theta_{\gamma}=\left|\boldsymbol{e}_{\gamma}(\boldsymbol{\tau}_{d,{ dev}})\boldsymbol{\cdot}\boldsymbol{e}_{\gamma}(\boldsymbol{\tau}_{d,{dev}}^{ model})\right|, \end{gather}$$

where $\boldsymbol {\tau }_d$ is the particle stress based on the two-way coupling DNS results, and $\boldsymbol {\tau }_{d,{dev}}^{model}$ corresponds to $\boldsymbol {\tau }_{d,{dev}}^{f}$, $\boldsymbol {\tau }_{d,{dev}}^{pS}$ and $\boldsymbol {\tau }_{d,{dev}}^{pB}$. Although the eigenvectors may have opposite directions (e.g. $-\boldsymbol {e}_{\alpha }$ instead of $\boldsymbol {e}_{\alpha }$), these two directions have no substantial difference. Therefore, the absolute values are employed in (5.14)–(5.16).

Another degree of APA is assessed by using the quaternion. By a rotation of the coordinates, three eigenvectors $\boldsymbol {e}_{\alpha }(\boldsymbol {\tau }_{d,{dev}}^{model})$, $\boldsymbol {e}_{\beta }(\boldsymbol {\tau }_{d,{dev}}^{model})$ and $\boldsymbol {e}_{\gamma }(\boldsymbol {\tau }_{d,{dev}}^{model})$ can be transformed to $\boldsymbol {e}_{\alpha }(\boldsymbol {\tau }_{d,{dev}})$, $\boldsymbol {e}_{\beta }(\boldsymbol {\tau }_{d,{dev}})$ and $\boldsymbol {e}_{\gamma }(\boldsymbol {\tau }_{d,{dev}})$, respectively. The rotation is quantified by the quaternion

(5.17)\begin{equation} q(\boldsymbol{\tau}_{d,{dev}},\boldsymbol{\tau}_{d,{dev}}^{ model})=\cos\frac{\theta_q}{2}+ia_1\sin\frac{\theta_q}{2}+ ja_2\sin\frac{\theta_q}{2}+ka_3\sin\frac{\theta_q}{2}, \end{equation}

where $\boldsymbol {a}=(a_1,a_2,a_3)$ is the rotation axis, $\theta _q$ is the rotation angle, and $i$, $j$ and $k$ are the imaginary units of the algebra of quaternions (Farebrother, Groß & Troschke Reference Farebrother, Groß and Troschke2003). By determining $q(\boldsymbol {\tau }_{d,{dev}},\boldsymbol {\tau }_{d,{dev}}^{model})$, $\cos \theta _q$ can represent a degree of APA between $\boldsymbol {\tau }_{d,{dev}}$ and $\boldsymbol {\tau }_{d,{ dev}}^{model}$; a value close to 1 indicates a good agreement of the predicted particle stress field. As $q(\boldsymbol {\tau }_{d,{ dev}},\boldsymbol {\tau }_{d,{dev}}^{model})$ and corresponding $\theta _q$ are not unique (e.g. $q$ and $-q$ indicate the same rotation, and the corresponding $\theta _q$ are different), we take the smallest positive $\theta _q$ among all possible values.

Figures 7–9 show the profiles of the p.d.f.s of $\cos \theta _{\alpha }$, $\cos \theta _{\beta }$ and $\cos \theta _{\gamma }$ for the three stress models. The results for the cases I256D1, U256D1, I256D3 and U256D3 are shown to compare the effects of the isotropy of the external forces and ${St}_K$. For the eigenvector $\boldsymbol {e}_{\alpha }$ (figure 7), the high probabilities are observed in the range $\cos \theta _{\alpha }>0.9$ for $\boldsymbol {\tau }_{d,{dev}}^{pB}$ and $\boldsymbol {\tau }_{d,{ dev}}^{f}$ of all four cases of isotropic/anisotropic forcing with small/large particle diameters. As the probabilities of $\cos \theta _{\alpha }>0.9$ for $\boldsymbol {\tau }_{d,{dev}}^{pB}$ are larger than those for $\boldsymbol {\tau }_{d,{dev}}^{f}$, the scale-similarity assumption is effective to predict the principal axis of the largest eigenvalue $\lambda _{\alpha }$. For the particle Smagorinsky model $\boldsymbol {\tau }_{d,{dev}}^{pS}$, the largest probability takes place at approximately $\cos \theta _{\alpha }=0.8$ for the unidirectional forcing cases (figures 7b,d), unlike the other two particle stress models. For the isotropic forcing cases (figures 7a,c), the probabilities of $\cos \theta _{\alpha }>0.9$ for $\boldsymbol {\tau }_{d,{dev}}^{pB}$ and $\boldsymbol {\tau }_{d,{dev}}^{f}$ tend to decrease with ${St}_K$ (i.e. with larger particle diameters), whereas the probabilities for $\boldsymbol {\tau }_{d,{dev}}^{pS}$ remain at approximately the same level regardless of ${St}_K$, indicating the difference in the characteristics of the models with ${St}_K$.

Figure 7.

P.d.f.s of $\cos \theta _{\alpha }$: (a) I256D1, (b) U256D1, (c) I256D3, (d) U256D3. The particle stress models are indicated by different colours. The sizes of the averaging volumes are indicated by different symbols.

Figure 8.

P.d.f.s of $\cos \theta _{\beta }$: (a) I256D1, (b) U256D1, (c) I256D3, (d) U256D3. The particle stress models are indicated by different colours. The sizes of the averaging volumes are indicated by different symbols.

Figure 9.

P.d.f.s of $\cos \theta _{\gamma }$: (a) I256D1, (b) U256D1, (c) I256D3, (d) U256D3. The particle stress models are indicated by different colours. The sizes of the averaging volumes are indicated by different symbols.

The trends of $\cos \theta _{\beta }$ and $\cos \theta _{\gamma }$ (figures 8 and 9) are similar to each other in some aspects, as explained below. The probabilities of $\cos \theta _{\beta }>0.9$ and $\cos \theta _{\gamma }>0.9$ for $\boldsymbol {\tau }_{d,{dev}}^{f}$ are larger than those for $\boldsymbol {\tau }_{d,{dev}}^{pB}$ at the small ${St}_K$ cases (figures 8a,b and 9a,b), and the advantage of the scale-similarity model observed in figure 7 no longer exists. For the large ${St}_K$ case under the isotropic forcing condition (figures 8c and 9c), the probabilities for $\boldsymbol {\tau }_{d,{dev}}^{pS}$ are higher than those of $\boldsymbol {\tau }_{d,{dev}}^{pB}$ and $\boldsymbol {\tau }_{d,{dev}}^{f}$. In addition, the difference in the trends of $\cos \theta _{\beta }$ and $\cos \theta _{\gamma }$ for $\boldsymbol {\tau }_{d,{dev}}^{pS}$ is confirmed for the unidirectional forcing cases (figures 8b,d and 9b,d); the maximum values of the probabilities are at around $\cos \theta _{\beta }=1$ and $\cos \theta _{\gamma }=0.7$.

The intensity of the fluid velocity fluctuation inside the averaging volume $V$ is related to the radius $R$. The probabilities of $\cos \theta _{\alpha }>0.9$, $\cos \theta _{\beta }>0.9$ and $\cos \theta _{\gamma }>0.9$ for $\boldsymbol {\tau }_{d,{dev}}^{f}$ increase with $k_0R$ for all the cases (figures 7–9). Therefore, although the particle motions are different from the background flow around the individual particles, the motions are affected by the flow of scales close to $R$, and this effect enhances with the volume size.

Considering that the collective particle motion is influenced by the flow inside the averaging volume, and that the turbulence intensity varies in space, it is important to evaluate the models based on locally averaged information. The Stokes number reflecting the effect of local information in the averaging volume is defined as

(5.18)\begin{equation} {St}_R=\tau_p\,\frac{\epsilon}{u_R^{2}}, \end{equation}

where $u_R^{2}$ is the fluctuation intensity of the fluid velocity inside the volume obtained by

(5.19)\begin{equation} u_R^{2}=\tfrac{1}{3}\,{\rm tr}\,\boldsymbol{\tau}_c, \end{equation}

and the dissipation rate is calculated as

(5.20)\begin{equation} \epsilon=\int |\boldsymbol{k}|^{2}\,E(|\boldsymbol{k}|)\,{\rm d}|\boldsymbol{k}| + \frac{1}{L_{cube}^{3}}\sum_{particles}\boldsymbol{f}\boldsymbol{\cdot} \alpha_s(\left\langle\boldsymbol{u}\right\rangle_s(\boldsymbol{x}_p)-\boldsymbol{w}_p). \end{equation}

The second term on the right-hand side of (5.20) indicates the unresolved dissipation in the vicinity of the particle surface. By analogy with the scale estimation of turbulence (Tennekes & Lumley Reference Tennekes and Lumley1972), the value $u_R^{2}/\epsilon$ is considered as the time scale for dissipating the eddy of velocity $u_R$. Note that $\epsilon$ is the averaged value and that the effect of the averaging volume size is reflected in (5.18) through $u_R^{2}$ (e.g. $u_R^{2}\to 0$ for $R\to 0$). As an indicator similar to ${St}_R$, the locally defined time scale based on the strain rate and the eddy viscosity of LES was introduced as the model input for the particle acceleration in the Euler–Lagrangian framework (Gorokhovski & Zamansky Reference Gorokhovski and Zamansky2018). Using (5.18), the degree of APA based on the quaternion (5.17) is evaluated by the conditional probability $A_x({St}_R)$ of being $\cos \theta _q\geqslant x$ for a fixed ${St}_R$. A large value of $A_x$ indicates that the model describes accurately the particle stress field in terms of the degree of APA.

Figure 10 shows the conditional probability $A_{0.8}$ for all the cases of the two-way coupling simulations. The values of $A_{0.8}$ for $\boldsymbol {\tau }_{d,{dev}}^{pS}$, $\boldsymbol {\tau }_{d,{dev}}^{pB}$ and $\boldsymbol {\tau }_{d,{ dev}}^{f}$ are regarded as the functions of ${St}_R$, and the different trends for the models are observed clearly. In particular, $\boldsymbol {\tau }_{d,{dev}}^{f}$ is predominant for ${St}_R<1$ for all the cases, which corresponds to the result for the small ${St}_K$ cases shown in figures 8(a), 8(b), 9(a) and 9(b). Bragg, Ireland & Collins (Reference Bragg, Ireland and Collins2015) showed theoretically that the particle motion is related to the fluid velocity gradient at scale $r$ if the Stokes number based on $r$ is smaller than 1; otherwise, the effect of the larger flow structure is dominant. This theoretical result coincides with the predominance of $\boldsymbol {\tau }_{d,{dev}}^{f}$ as the particle motion is affected by the flow structure at the scale $R$ for ${St}_R<1$. Note that although $R$ is much smaller than the upper limit of effective scales (e.g. the scale $O(10^{3})$ times $\eta$ for ${St}_K=0.1$; (Tom & Bragg Reference Tom and Bragg2019)) and the particles do not follow the subgrid fluid flow, the flow at the scale $R$ is still effective for the particle stress.

Figure 10.

Degree of APA for the models to those for the fully resolved particle stress based on the quaternion $\cos \theta _q$ against ${St}_R$ defined by (5.18): (a) I256D$X$, (b) U256D$X$, (c) I384D$X$, (d) U384D$X$, where D$X$ is D1, D2 or D3. The particle stress models are indicated by different colours: red, blue and black represent $\boldsymbol {\tau }_{d,{dev}}^{pS}$, $\boldsymbol {\tau }_{d,{dev}}^{pB}$ and $\boldsymbol {\tau }_{d,{dev}}^{f}$, respectively. The sizes of the averaging volume are indicated by different line types. The particle conditions (D$X$) are indicated by different symbols.

The conditional probability $A_{0.8}$ for $\boldsymbol {\tau }_{d,{dev}}^{pB}$ is larger than that for $\boldsymbol {\tau }_{d,{dev}}^{f}$ for ${St}_R>5$ as $\boldsymbol {\tau }_{d,{dev}}^{pB}$ (related to the spatial variation of $\left \langle \boldsymbol {w}\right \rangle _d$) is related to the flow structure larger than $R$. For the isotropic forcing cases (figures 10a,c), $A_{0.8}$ for $\boldsymbol {\tau }_{d,{dev}}^{pB}$ decreases with increasing ${St}_R$. Considering the case that the $xx$-component of the particle stress is dominant owing to the positive and negative $w_{px}$ of the individual particles inside the averaging volume, the decreasing trend of $A_{0.8}$ for $\boldsymbol {\tau }_{d,{dev}}^{pB}$ is understood as $\left \langle w_x\right \rangle _d$, and its spatial variation can be almost zero regardless of the flow structure at the scale $R$ for large ${St}_R$.

For the unidirectional forcing cases, as already shown in figure 3, $E_y< E_z\approx E_x$ for the low-wavenumber region and $E_y\approx E_z< E_x$ for the high-wavenumber region were observed, in contrast to the isotropic forcing cases where an equal distribution of energy was observed. This uneven distribution of $E_i$ suggests that the $zz$-component (related to $E_z$) of $\boldsymbol {\tau }_{d,{dev}}^{pB}$ estimated on the large scale tends to be large, while the predominance of this component is reduced for the particle subgrid stress $\boldsymbol {\tau }_{d,{dev}}$ for small ${St}_R$, indicating that $A_{0.8}$ for $\boldsymbol {\tau }_{d,{ dev}}^{pB}$ is suppressed. Therefore, $A_{0.8}$ for $\boldsymbol {\tau }_{d,{dev}}^{pB}$ does not exhibit a clear decreasing trend with increasing ${St}_R$ in the unidirectional forcing cases (figures 10b,d) in contrast to the isotropic forcing cases (figures 10a,c).

To examine the validity of using the averaged $\epsilon$ for ${St}_R$ defined by (5.18), the above result for $A_{0.8}$ is compared with that defined with the local dissipation $\epsilon _{loc}$ inside $V$. The probability $A_{0.8}$ based on $\epsilon _{loc}$ in figure 11 shows similar but diffused profiles in comparison with that based on $\epsilon$ in figure 10, indicating that ${St}_R$ defined with $\epsilon$ is preferable to extract the particle stress behaviour. The energy loss of the flow inside $V$ is caused by the convective transport as well as the viscous dissipation, and the rate of the energy loss is different from $\epsilon _{loc}$. In the case that the energy is supplied from the outside to maintain the flow structure inside $V$, the lifetime of the flow structure becomes longer than the estimated value $u_R^{2}/\epsilon _{loc}$. Therefore, the locally defined $\epsilon _{loc}$ is not always adequate, and the mean value $\epsilon$ is employed for ${St}_R$ in this study.

Figure 11.

Degree of APA for the models to those for the fully resolved particle stress based on the quaternion $\cos \theta _q$ against ${St}_R$ based on $\epsilon _{loc}$: (a) I256D$X$, (b) U256D$X$, (c) I384D$X$, (d) U384D$X$, where D$X$ is D1, D2 or D3. The particle stress models are indicated by different colours: red, blue and black represent $\boldsymbol {\tau }_{d,{dev}}^{pS}$, $\boldsymbol {\tau }_{d,{dev}}^{pB}$ and $\boldsymbol {\tau }_{d,{dev}}^{f}$, respectively. The sizes of the averaging volume are indicated by different line types. The particle conditions (D$X$) are indicated by different symbols.

5.3. Correlation coefficient for the deviatoric part

To investigate the characteristics of the magnitudes (related to the eigenvectors) as well as the principal axes of the particle stress models, the correlation coefficient of the deviatoric part is defined as

(5.21)\begin{equation} D_{dev}({St}_R)=\frac{\displaystyle\sum\nolimits^{St_R} \delta\boldsymbol{\tau}_{d,{dev}}\colon\delta\boldsymbol{\tau}_{d,{dev}}^{model}} {\displaystyle\sqrt{\sum\nolimits^{St_R}\delta\boldsymbol{\tau}_{d,{dev}}\colon\delta\boldsymbol{\tau}_{d,{ dev}}} \sqrt{ \displaystyle\sum\nolimits^{St_R}\delta\boldsymbol{\tau}_{d,{ dev}}^{model}\colon\delta\boldsymbol{\tau}_{d,{dev}}^{ model}}}, \end{equation}

where $\sum ^{{St}_R}$ indicates the summation at a fixed ${St}_R$ (hereafter, the conditional summation), and $\delta \boldsymbol {\tau }_{d,{dev}}$ and $\delta \boldsymbol {\tau }_{d,{dev}}^{model}$ are

(5.22a,b)\begin{equation} \delta\boldsymbol{\tau}_{d,{dev}}=\boldsymbol{\tau}_{d,{dev}}- \frac{\displaystyle\sum\nolimits^{{St}_R} \boldsymbol{\tau}_{d,{ dev}}}{\displaystyle\sum\nolimits^{{St}_R} 1},\quad \delta\boldsymbol{\tau}_{d,{ dev}}^{model}=\boldsymbol{\tau}_{d,{dev}}^{ model}-\frac{\displaystyle\sum\nolimits^{{St}_R}\boldsymbol{\tau}_{d,{ dev}}^{model}}{\displaystyle\sum\nolimits^{{St}_R}1}, \end{equation}

respectively. Figure 12 compares the dependence of $D_{dev}$ on ${St}_R$ for each model ($\boldsymbol {\tau }_{d,{dev}}^{pS}$, $\boldsymbol {\tau }_{d,{dev}}^{pB}$ and $\boldsymbol {\tau }_{d,{dev}}^{f}$). The trend of $D_{dev}$ is similar to the degree of APA (figure 10) such that the models $\boldsymbol {\tau }_{d,{dev}}^{f}$ and $\boldsymbol {\tau }_{d,{ dev}}^{pB}$ show the highest and the second highest $D_{dev}$ for ${St}_R<1$, which also confirms the effectiveness of ${St}_R$ for extracting the characteristics of the particle stress in terms of the correlation coefficient. In general, by superposing a single vortex of the scale $R$ and disturbances, $\boldsymbol {\tau }_{d,{dev}}^{f}$ and $\boldsymbol {\tau }_{d,{iso}}^{f}$ reflect the structures of the scale $R$ and smaller scales, respectively. Therefore, the correlation coefficient $D_{dev}$ for the model $\boldsymbol {\tau }_{d,{dev}}^{f}$ is regarded as the contribution of the fluid motion of the scale $R$ for ${St}_R<1$. The differences between figures 10 and 12 are also observed; as $D_{dev}$ for $\boldsymbol {\tau }_{d,{dev}}^{pB}$ in the case U256D$X$ ($X=1,2,3$; figure 12b) shows a slightly decreasing trend with increasing ${St}_R$ in contrast to the increasing trend of $A_{0.8}$ (figure 10b), the prediction of the magnitude of the particle stress by $\boldsymbol {\tau }_{d,{dev}}^{pB}$ is preferable for small ${St}_R$. For ${St}_R>5$, the $D_{dev}$ values for $\boldsymbol {\tau }_{d,{dev}}^{pS}$ are close to those for $\boldsymbol {\tau }_{d,{dev}}^{pB}$ for all the cases as $\boldsymbol {\tau }_{d,{dev}}^{pS}$ and $\boldsymbol {\tau }_{d,{ dev}}^{pB}$ are related to the similar fluctuation intensities of $\left \langle \boldsymbol {w}\right \rangle _d$ at the same scale larger than $R$.

Figure 12.

Correlation coefficients of the particle stress models for the deviatoric part $D_{dev}$: (a) I256D$X$, (b) U256D$X$, (c) I384D$X$, (d) U384D$X$, where D$X$ is D1, D2 or D3. The particle stress models are indicated by different colours: red, blue and black represent $\boldsymbol {\tau }_{d,{dev}}^{pS}$, $\boldsymbol {\tau }_{d,{dev}}^{pB}$ and $\boldsymbol {\tau }_{d,{dev}}^{f}$, respectively. The sizes of the averaging volume are indicated by different line types. The particle conditions (D$X$) are indicated by different symbols.

In addition to $D_{dev}$ as the direct evaluation of the particle stress model, the correlation coefficient of the energy transfer rate $\boldsymbol {\tau }_{d,{dev}}\colon \tilde {\boldsymbol {\nabla }}\langle \boldsymbol {w}\rangle _d$ is also defined as

(5.23)\begin{equation} H_{dev}({St}_R)=\frac{ \displaystyle\sum\nolimits^{{St}_R}\delta h_{ dev}\,\delta h_{dev}^{model}} {\sqrt{ \displaystyle\sum\nolimits^{{St}_R} \delta h_{dev}\,\delta h_{dev}} \sqrt{ \displaystyle\sum\nolimits^{{St}_R} \delta h_{dev}^{model}\,\delta h_{dev}^{model}}}, \end{equation}

where

(5.24)$$\begin{gather} \delta h_{dev}=\boldsymbol{\tau}_{d,{dev}}\colon\tilde{\boldsymbol{\nabla}} \langle\boldsymbol{w}\rangle_d- \left.\sum\nolimits^{{St}_R} \boldsymbol{\tau}_{d,{dev}}\colon\tilde{\boldsymbol{\nabla}}\langle \boldsymbol{w}\rangle_d\right/\sum\nolimits^{{St}_R}1, \end{gather}$$

(5.25)$$\begin{gather}\delta h_{dev}^{model}=\boldsymbol{\tau}_{d,{dev}}^{model}\colon\tilde{\boldsymbol{\nabla}}\langle\boldsymbol{w}\rangle_d- \left.\sum\nolimits^{{St}_R}\boldsymbol{\tau}_{d,{dev}}^{ model}\colon\tilde{\boldsymbol{\nabla}}\langle\boldsymbol{w} \rangle_d\right/\sum\nolimits^{{St}_R}1. \end{gather}$$

Figure 13 shows the dependence of $H_{dev}$ on ${St}_R$ for each model. For the models $\boldsymbol {\tau }_{d,{dev}}^{pS}$ and $\boldsymbol {\tau }_{d,{dev}}^{pB}$ that are related to the fluctuation intensity of the particle velocity $\left \langle \boldsymbol {w}\right \rangle _d$, the values of $H_{dev}$ tend to be higher than those of $D_{dev}$ (figure 12). This tendency for $H_{dev}$ is consistent with the result of Moreau et al. (Reference Moreau, Simonin and Bédat2010). As a result, the models based on the particle velocity $\left \langle \boldsymbol {w}\right \rangle _d$ are preferable in the sense of the energy transfer rate ($H_{dev}$) for the range ${St}_R>1$. For ${St}_R<1$, the model $\boldsymbol {\tau }_{d,{dev}}^{f}$ is predominant as well as the model $\boldsymbol {\tau }_{d,{dev}}^{pB}$ for the energy transfer rate.

Figure 13.

Correlation coefficients of the energy transfer rate by the particle stress models for the deviatoric part $H_{dev}$: (a) I256D$X$, (b) U256D$X$, (c) I384D$X$, (d) U384D$X$, where D$X$ is D1, D2 or D3. The particle stress models are indicated by different colours: red, blue and black represent $\boldsymbol {\tau }_{d,{dev}}^{pS}$, $\boldsymbol {\tau }_{d,{dev}}^{pB}$ and $\boldsymbol {\tau }_{d,{dev}}^{f}$, respectively. The sizes of the averaging volume are indicated by different line types. The particle conditions (D$X$) are indicated by different symbols.

5.4. Correlation coefficient for the isotropic part

By replacing the subscript ‘dev’ with ‘iso’ in (5.21) and (5.22a,b), the correlation coefficient of the isotropic part $D_{iso}$ is investigated. Figure 14 shows $D_{iso}$ for the models $\boldsymbol {\tau }_{d,{iso}}^{pY}$ and $\boldsymbol {\tau }_{d,{iso}}^{pB}$. As ${St}_R$ is fixed for the conditional summation, $\boldsymbol {\tau }_{d,{iso}}^{f}=(\epsilon \tau _p/{St}_R){\boldsymbol{\mathsf{I}}}$ takes a constant component in the diagonals, and therefore $D_{iso}$ for $\boldsymbol {\tau }_{d,{iso}}^{f}$ is not defined. The correlation coefficient depends largely on the averaging volume and ${St}_K$, and the differences of the $D_{iso}$ distributions between the models are not as clear as observed in figure 12. In the scale-similarity model, the unresolved particle stress is extrapolated from the resolved scale and the effect of scales that are much smaller than $R$ is not predicted accurately, while the effect of the small scales is relatively weak for the deviatoric part. Therefore, even for ${St}_R<1$, $D_{iso}$ for $\boldsymbol {\tau }_{d,{iso}}^{pB}$ is not as high as $D_{dev}$ (figure 12) for all the cases. The model $\boldsymbol {\tau }_{d,{iso}}^{pY}$ is more effective than $\boldsymbol {\tau }_{d,{iso}}^{pB}$ for ${St}_R<1$ in the sense of the correlation coefficient $D_{iso}$.

Figure 14.

Correlation coefficients of the particle stress models for the isotropic part $D_{iso}$: (a) I256D$X$, (b) U256D$X$, (c) I384D$X$, (d) U384D$X$, where D$X$ is D1, D2 or D3. The particle stress models are indicated by different colours: red and blue represent $\boldsymbol {\tau }_{d,{iso}}^{pY}$ and $\boldsymbol {\tau }_{d,{iso}}^{pB}$, respectively. The sizes of the averaging volume are indicated by different line types. The particle conditions (D$X$) are indicated by different symbols.

Figure 15 shows the correlation coefficient of the energy transfer rate $H_{iso}$ defined similarly by replacing the subscript ‘dev’ with ‘iso’ in (5.23)–(5.25). The correlation coefficient $H_{iso}$ is almost over 0.8. As the isotropic components of the particle stress ${\rm tr}\,\boldsymbol {\tau }_{d,{iso}}$ and the models ${\rm tr}\,\boldsymbol {\tau }_{d,{iso}}^{model}$ are always positive, the prediction of the sign of $\boldsymbol {\tau }_{d,{iso}}^{ model}\colon \tilde {\boldsymbol {\nabla }}\langle \boldsymbol {w}\rangle _d$ is always correct, indicating higher values of $H_{iso}$ than those of $D_{iso}$ (figure 14). Even though the correlation coefficients are different ($D_{iso}$ and $H_{iso}$), the advantages of the models are similar.

Figure 15.

Correlation coefficients of the energy transfer rate by the particle stress models for the isotropic part $H_{iso}$: (a) I256D$X$, (b) U256D$X$, (c) I384D$X$, (d) U384D$X$, where D$X$ is D1, D2 or D3. The particle stress models are indicated by different colours: red and blue represent $\boldsymbol {\tau }_{d,{iso}}^{pY}$ and $\boldsymbol {\tau }_{d,{iso}}^{pB}$, respectively. The sizes of the averaging volume are indicated by different line types. The particle conditions (D$X$) are indicated by different symbols.

5.5. Model coefficients

As the behaviour of the particle stress depends on ${St}_R$ particularly for the deviatoric part, we propose the model coefficients (5.10a–c) and (5.11a–c) as functions of ${St}_R$. The model coefficients $C$ are determined to minimise the following functions:

(5.26)$$\begin{gather} \sideset{}{{}^{{St}_R}}{\sum} (\boldsymbol{\tau}_{d,{dev}}-\boldsymbol{\tau}_{d,{dev}}^{model})\colon (\boldsymbol{\tau}_{d,{dev}}-\boldsymbol{\tau}_{d,{dev}}^{model}), \end{gather}$$

(5.27)$$\begin{gather}\sideset{}{{}^{{St}_R}}{\sum} (\boldsymbol{\tau}_{d,{iso}}-\boldsymbol{\tau}_{d,{iso}}^{model})\colon (\boldsymbol{\tau}_{d,{iso}}-\boldsymbol{\tau}_{d,{iso}}^{model}). \end{gather}$$

Figure 16 plots the model coefficients for the deviatoric parts with respect to ${St}_R$. The behaviours of the coefficients for the respective stress models may be understood by a comparison of the averaged magnitudes of the particle stress and its models (e.g. $|\boldsymbol {\tau }_{d,{dev}}|$ and $|\boldsymbol {\tau }_{d,{dev}}^{f}|$) in this subsection. According to the statistical results of the radial relative velocity $w_r$ of two particles, the averaged magnitude of $|w_r|$ increases with the distance $r$ (order of magnitude of $\eta$), and ${\rm d}|w_r|/{\rm d}r$ $({>}0)$ decreases with increasing ${St}_K$ for $0.7\leqslant {St}_K\leqslant 5$ (Ray & Collins Reference Ray and Collins2014). In other words, the magnitude of the difference of two particle velocities inside a small volume increases with increasing ${St}_K$. By analogy with the result for $w_r$, considering the effect of the flow at the scale $R$ instead of $\eta$, the fluctuation intensity of the particle velocity $\boldsymbol {w}_p$ inside the volume $V$ (i.e. $|\boldsymbol {\tau }_{d}|$) increases relatively with increasing ${St}_R$ (${\approx }1$) compared with the increases in $|\boldsymbol {\tau }_{d}^{pS}|$ and $|\boldsymbol {\tau }_{d}^{pB}|$ that are related to the spatial variation of $\left \langle \boldsymbol {w}\right \rangle _d$. Therefore, $C_{ dev}^{pS}$ and $C_{dev}^{pB}$ increase with increasing ${St}_R$ (figures 16a–d). For the fluid residual stress model (figures 16e, f), the increasing trend of $C_{dev}^{f}$ is observed as the increase of ${St}_R$ indicates a decrease in the magnitudes of $u_R^{2}$ and $|\boldsymbol {\tau }_{d,{dev}}^{f}|$.

Figure 16.

Model coefficients of the particle stress models for the deviatoric part $C_{dev}$: (a,c,e) isotropic forcing cases; (b,d, f) unidirectional forcing cases. Black and red colours represent low (I256D$X$, U256D$X$) and high (I384D$X$, U384D$X$) Reynolds number cases, respectively, where D$X$ is D1, D2 or D3. The sizes of the averaging volume are indicated by different line types. The particle conditions (D$X$) are indicated by different symbols.

Most plots of $C_{dev}^{pS}$ under the different conditions overlap with each other (figures 16a,b), while the coefficients of the other two models exhibit the effect of the Stokes number ${St}_K$, the radius of the averaging volume $k_0R$, and the Reynolds number ${Re}_{\lambda }$. For $C_{dev}^{f}$ (figures 16e, f), the lines of D1, D2 and D3 are independent of each other, indicating the effect of ${St}_K$. For a fixed ${St}_R\propto {St}_K/u_R^{2}$, the increase of ${St}_K$ implies the same proportional increase of the local fluctuation intensity of the fluid velocity $u_R^{2}$. However, the intensities of the flow structures of scales larger than $R$ do not increase in the same way as $u_R^{2}$, and the effect of the larger flow structures on $\boldsymbol {w}_p$ is suppressed relatively for the particles having large inertia. Therefore, with increasing ${St}_K$ through increasing the particle size from D1 to D3, the increase of $|\boldsymbol {\tau }_{d,{dev}}|$ is relatively small compared with the intensity of the fluid velocity (i.e. $|\boldsymbol {\tau }_{d,{ dev}}^{f}|$), and therefore $C_{dev}^{f}$, decreases (figures 16e, f).

The coefficient $C_{dev}^{pS}$ for the unidirectional forcing case (figure 16b) tends to be smaller than those for the isotropic case (figure 16a). Therefore, a larger gap between $|\boldsymbol {\tau }_{d,{dev}}^{pS}|$ and $|\boldsymbol {\tau }_{d,{dev}}|$ is indicated for the unidirectional forcing case. This tendency is explained by the energy spectra (figure 4) in which the decreasing rate of $E$ for the unidirectional forcing case is larger, and the disturbing effect of the small flow structure on $|\boldsymbol {\tau }_{d,{dev}}|$ is relatively low.

In figure 16(c), in the range ${St}_R<1$ for the isotropic forcing condition, $C_{dev}^{pB}$ for $k_0R=0.75$ (corresponding to $R/\eta =75$, based on table 1) in the high ${Re}_{\lambda }$ case (I384D1) is smaller than values for $k_0R=1.13$ and $1.50$ (corresponding to $R/\eta =66$ and $87$, respectively) in the low ${Re}_{\lambda }$ case (I256D1). This non-monotonic trend of $C_{dev}^{pB}$ with respect to $R/\eta$ is also observed for the unidirectional forcing condition for ${St}_R<1$ (figure 16d); the values of $R/\eta$ for $k_0R=0.75$ in the case U384D1 (at ${Re}_{\lambda }=183$), $k_0R=1.13$ in the case U256D1 (at ${Re}_{\lambda }=125$) and $k_0R=1.50$ in the case U256D1 are $58$, $51$ and $68$, respectively. Therefore, the result indicates that the above non-monotonic trend is influenced by ${Re}_{\lambda }$ at a fixed $R/\eta$. According to the result for the radial relative velocity at ${St}_K\approx 1$ (Ray & Collins Reference Ray and Collins2011), the magnitude of $|w_r|$ at a fixed $r/\eta$ is increased particularly for $r/\eta >1$ with increasing ${Re}_{\lambda }$. By replacing $\eta$ with $R$, the increase of $|\boldsymbol {\tau }_{d,{dev}}^{pB}|$ reflecting the spatial variation of $\left \langle \boldsymbol {w}\right \rangle _d$ at the scale larger than $R$ is suggested, indicating the decrease of $C_{dev}^{pB}$ with increasing ${Re}_{\lambda }$. A different trend is observed for $C_{dev}^{f}$. The above relation between $|w_r|$ and ${Re}_{\lambda }$ also indicates an increase of $|\boldsymbol {\tau }_{d,{dev}}|$ with increasing ${Re}_{\lambda }$. Therefore, for the isotropic forcing condition, the coefficients $C_{dev}^{f}$ for the high ${Re}_{\lambda }$ case (I384) are larger than those for the low ${Re}_{\lambda }$ case (I256) regardless of $R/\eta$ (figure 16e). For the unidirectional forcing condition (figure 16f), the increase of $C_{dev}^{f}$ at the high ${Re}_{\lambda }$ case (U384D1) with respect to its value at the low ${Re}_{\lambda }$ case (U256D1) is suppressed as $C_{dev}^{f}$ at $k_0R=0.75$ in the case U384D1 is smaller than values at $k_0R=1.13$ and $1.50$ in the case U256D1. This suppression indicates that the fluctuation intensity of $\boldsymbol {w}_p$ in $V$ is not reduced by decreasing ${Re}_{\lambda }$ as the fixed $R/\eta$ for lower ${Re}_{\lambda }$ corresponds to the larger fraction of the wavelength of the time-independent sinusoidal (unidirectional) forcing that induces the spatial distribution of $\boldsymbol {w}_p$.

By analogy with the result that ${\rm d}|w_r|/{\rm d}r$ (${>}0$) becomes larger for smaller $St_K$ (${\approx }1$) (Ray & Collins Reference Ray and Collins2014), it is understood reasonably that the increase of $|\boldsymbol {\tau }_{d,{dev}}|$ corresponds to the increase of $C_{dev}^{f}$ with increasing $k_0R$ for ${St}_R<1$, as observed in the cases I256D1 and I384D1 in figure 16(e) and the cases U256D1 and U384D1 in figure 16( f). The increase of $C_{dev}^{pB}$ with increasing $k_0R$ for ${St}_R<1$ is also observed (figures 16c,d). The almost independent trend of $C_{dev}^{pS}$ from $k_0R$ (figures 16a,b) indicates that the dependence of $|\boldsymbol {\tau }_{d,{dev}}^{pS}|$ (related to the spatial variation of $\left \langle \boldsymbol {w}\right \rangle _d$) on $k_0R$ is similar to that of $|\boldsymbol {\tau }_{d,{dev}}|$.

Figure 17 plots the model coefficients for the isotropic parts with respect to ${St}_R$, and the trends are similar to those observed in figure 16. In the study of Moreau et al. (Reference Moreau, Simonin and Bédat2010), the Stokes number ${St}_K=5.1$ and the averaging volume $R/\eta \leqslant 8$ were considered in the isotropic turbulence, whereas $7.6\leqslant {St}_K\leqslant 127.6$ and $R/\eta \geqslant 34$ are the focus in the present study for the large-scale multiphase flow. The model coefficients reported in Moreau et al. (Reference Moreau, Simonin and Bédat2010) were $C_{dev}^{pS}=0.025$, $C_{iso}^{pY}=0.051$ and $C^{pB}=2.2$ (without decomposing the isotropic and deviatoric parts), and these values have orders of magnitude similar to those of the values in this study.

Figure 17.

Model coefficients of the particle stress models for the isotropic part $C_{iso}$: (a,c,e) isotropic forcing cases; (b,d, f) unidirectional forcing cases. Black and red colours represent low (I256D$X$, U256D$X$) and high (I384D$X$, U384D$X$) Reynolds number cases, respectively, where D$X$ is D1, D2 or D3. The sizes of the averaging volume are indicated by different line types. The particle conditions (D$X$) are indicated by different symbols.

Although the scale-similarity model can represent the inverse cascade of energy from the subgrid scale to the grid scale, this model tends to make the computation unstable. To suppress the instability and cover the weakness of each model, the following linear combination models are also considered:

(5.28)$$\begin{gather} \boldsymbol{\tau}_{d,{dev}}^{lc}= C'^{pS}_{dev}\boldsymbol{\tau}_{d,{dev}}^{pS} +C'^{pB}_{ dev}\boldsymbol{\tau}_{d,{dev}}^{pB} +C'^{f}_{dev}\boldsymbol{\tau}_{d,{dev}}^{f}, \end{gather}$$

(5.29)$$\begin{gather}\boldsymbol{\tau}_{d,{iso}}^{lc}= C'^{pY}_{iso}\boldsymbol{\tau}_{d,{iso}}^{pY} +C'^{f}_{iso}\boldsymbol{\tau}_{d,{iso}}^{f}, \end{gather}$$

where $C'$ are the weight coefficients with the corresponding subscript and superscript. With the exception of the fluid residual stress term, similar models have been considered, and the weight coefficients that are smaller than those for the individual models were proposed (Moreau et al. Reference Moreau, Simonin and Bédat2010). As $C^{pY}_{iso}$ shows trends similar to $C^{pB}_{iso}$, and $\boldsymbol {\tau }_{d,{iso}}^{pY}$ tends to exhibit a higher correlation coefficient than $\boldsymbol {\tau }_{d,{iso}}^{pB}$ (figure 14), the model $\boldsymbol {\tau }_{d,{ iso}}^{pB}$ is excluded from (5.29). The term $C'^{f}_{ iso}\boldsymbol {\tau }_{d,{iso}}^{f}$ in (5.29) is treated as a constant for a fixed ${St}_R$. The weight coefficients in (5.28) and (5.29) are determined in order to minimise (5.26) and (5.27) by reading $\boldsymbol {\tau }_{d,{dev}}^{model}=\boldsymbol {\tau }_{d,{dev}}^{ lc}$. In figure 18, the weight coefficients for the deviatoric part show trends similar to those in figure 16, although the magnitudes are decreased. In figure 19, the reduction of $C'^{pY}_{iso}$ from $C^{pY}_{iso}$ (figures 17a,b) is significant, indicating the importance of the constant term $C'^{f}_{ iso}\boldsymbol {\tau }_{d,{iso}}^{f}$ as the model of the isotropic part requires a bias, which compensates for the difference between $\boldsymbol {\tau }_{d,{iso}}$ and $C'^{pY}_{iso}\boldsymbol {\tau }_{d,{ iso}}^{pY}$. For the deviatoric part, the constant term is not necessary as the averages of $\boldsymbol {\tau }_{d,{dev}}$ and the models should be zero.

Figure 18.

Weight coefficients of the linear combination model for the deviatoric part $C'_{dev}$: (a,c,e) isotropic forcing cases; (b,d, f) unidirectional forcing cases. Black and red colours represent low (I256D$X$, U256D$X$) and high (I384D$X$, U384D$X$) Reynolds number cases, respectively, where D$X$ is D1, D2 or D3. The sizes of the averaging volume are indicated by different line types. The particle conditions (D$X$) are indicated by different symbols.

Figure 19.

Weight coefficients of the linear combination model for the isotropic part $C'_{iso}$: (a,c) isotropic forcing cases; (b,d) unidirectional forcing cases. Black and red colours represent low (I256D$X$, U256D$X$) and high (I384D$X$, U384D$X$) Reynolds number cases, respectively, where D$X$ is D1, D2 or D3. The sizes of the averaging volume are indicated by different line types. The particle conditions (D$X$) are indicated by different symbols.