Large eddy simulation of particle-laden turbulent channel ow

Qunzhen Wanga) and Kyle D. Squires

Department of Mechanical Engineering, 209 Votey Building, University of Vermont, Burlington, Vermont 05405

Received 25 July 1995; accepted 26 January 1996 Particle transport in fully-developed turbulent channel ow has been investigated using large eddy simulation LES of the incompressible NavierStokes equations. Calculations were performed at channel ow Reynolds numbers, Re , of 180 and 644 based on friction velocity and channel half width ; subgrid-scale stresses were parametrized using the Lagrangian dynamic eddy viscosity model. Particle motion was governed by both drag and gravitational forces and the volume fraction of the dispersed phase was small enough such that particle collisions were negligible and properties of the carrier ow were not modied. Material properties of the particles used in the simulations were identical to those in the DNS calculations of Rouson and Eaton Proceedings of the 7th Workshop on Two-Phase Flow Predictions 1994 and experimental measurements of Kulick et al. J. Fluid Mech. 277, 109 1994 . Statistical properties of the dispersed phase in the channel ow at Re 180 are in good agreement with the DNS; reasonable agreement is obtained between the LES at Re 644 and experimental measurements. It is shown that the LES correctly predicts the greater streamwise particle uctuation level relative to the uid and increasing anisotropy of velocity uctuations in the dispersed phase with increasing values of the particle time constant. Analysis of particle uctuation levels demonstrates the importance of production by mean gradients in the particle velocity as well as the uid-particle velocity correlation. Preferential concentration of particles by turbulence is also investigated. Visualizations of the particle number density eld near the wall and along the channel centerline are similar to those observed in DNS and the experiments of Fessler et al. Phys. Fluids 6, 3742 1994 . Quantitative measures of preferential concentration are also in good agreement with Fessler et al. Phys. Fluids 6, 3742 1994 . 1996 American Institute of Physics. S1070-6631 96 00905-4

I. INTRODUCTION

Gas-phase turbulent ow elds laden with small heavy particles occur in a wide range of engineering and scientic disciplines. Examples are as diverse as pollutant dispersion in the atmosphere and contaminant transport in industrial applications. In these as well as many other instances the primary interactions are from uid to particles only, i.e., in the dilute regime in which particle collisions as well as the effect of particles on uid mass and momentum transport is negligible. Even with the restriction to dilute ows, accurate prediction of particle-laden turbulence is important in order to gain a better understanding of particle transport by turbulence as well as ultimately improve the design of the engineering devices in which two-phase ows occur. The principal difculty with the prediction of particleladen turbulent ows is that traditional approaches model particle transport using gradient transport hypotheses and do not accurately account for the complex interactions between particles and turbulence. Traditional methods are usually based on the Reynolds-averaged NavierStokes RANS equations in which the entire spectrum of velocity uctuations is represented indirectly using various parametrizations, e.g., K- models. A primary shortcoming of RANS methods for the prediction of particle-laden turbulent ows is related to deciencies associated with the model used to predict
a

Phone: 802 656-1940; FAX: 802 656-1929; electronic mail: qwang@emba.uvm.edu, squires@emba.uvm.edu

properties of the Eulerian turbulence eld. Accurate prediction of particle transport is strongly dependent upon providing a realistic description of the time-dependent, threedimensional velocity eld encountered along particle trajectories e.g., see Berlemont et al.,1 Simonin et al.2 . Deciencies in the prediction of turbulence properties in RANS calculations will in turn adversely impact prediction of dispersed phase transport. Thus, predictive techniques which are generally applicable to a wide class of turbulent twophase ows and accurately represent the underlying structure of turbulence are needed. The most sophisticated approach to representing the underlying structure of turbulence and, hence, particle transport in turbulent ows is direct numerical simulation DNS . In DNS the NavierStokes equations are solved without approximation other than those associated with the numerical method . DNS has been successfully employed in a number of studies which have increased our fundamental understanding of many of the mechanisms governing particle interactions with turbulence. Much of this work has been performed in isotropic turbulence and has shown, for example, that particles with time constants of the order of the Kolmogorov timescale are preferentially concentrated into regions of low vorticity and high strain rate e.g., see Squires and Eaton,3 Wang and Maxey4 . Preferential concentration results from an inertial bias in the particle trajectory and also leads to substantial increases in particle settling velocities in homogeneous turbulence Wang and Maxey4 . DNS has also been used to examine particle transport in wall-bounded shear
1996 American Institute of Physics 1207

Phys. Fluids 8 (5), May 1996

1070-6631/96/8(5)/1207/17/$10.00

Downloaded19Jul2004to195.83.232.39.RedistributionsubjecttoAIPlicenseorcopyright,seehttp://pof.aip.org/pof/copyright.jsp

ows e.g., see Rashidi et al.,5 Pedinotti et al.,6 Rouson and Eaton7 . Rashidi et al.5 and Pedinotti et al.6 have shown that particles can become preferentially concentrated in the lowspeed streaks which characterize the near-wall region and are subsequently resuspended by ejections from the wall. Rouson and Eaton7 found that particles possessing similar material properties to those used in the experiments of Kulick et al.8 see also Fessler et al.9 demonstrated effects of preferential concentration. Application of DNS to the study of particle deposition in turbulent boundary layers by McLaughlin,10 Brooke et al.,11,12 and Chen et al.13 has clearly shown that particles accumulate in the near wall region. Aside from the interactions between particles and turbulence resulting in preferential concentration in wall-bounded shear ows, mean gradients in the uid and particle velocities have complex effects on particle velocity uctuations. Both DNS and experiments have demonstrated that particle uctuations consistently lead those of the uid near the wall and, contrary to measurements in isotropic turbulence, that the streamwise velocity variance increases with increasing particle response time e.g., see Rogers and Eaton,14 Kulick et al.,8 Rouson and Eaton7 . Analyses by Liljegren15 and Reeks16 are consistent with these ndings and have demonstrated the important effect of mean shear in both the uid and particle velocity eld. Some Eulerian-based models have also recognized the importance of accounting for the effect of mean shear on particle uctuation levels e.g., see Kataoka and Serizawa,17 Abou-Arab and Roco,18 Reeks,19 Hwang and Shen,20 Simonin et al.2 . The effects of preferential concentration on particle transport as well as other features such as the effect of mean velocity gradients on particle uctuation levels are very difcult to represent using conventional predictive methods. DNS, while an extremely powerful tool for supplying information which cannot be obtained from experiment, is not practical for use as a predictive tool because it remains restricted to relatively low Reynolds number turbulent ows. An approach which is not as severely restricted in the range of Reynolds numbers as DNS is large eddy simulation LES . In LES the large, energy containing scales of motion are calculated directly while only the effect of the smallest subgrid scales of motion are modeled. Thus, LES predictions are less sensitive to modeling errors than in RANS calculations and, since the subgrid scales are more universal than the large scales, it is also possible to represent the effect of the subgrid scales using relatively simple models. A signicant advantage of LES over RANS methods is that it permits a much more accurate accounting of particleturbulence interactions. The primary drawback to the application of LES for the prediction of complex turbulent ows has traditionally been much the same as that which currently limits RANS methods, i.e., an inability of the subgrid-scale SGS model to account for changes in the spectral content of the turbulence under a variety of conditions, e.g., changes in the Reynolds number, type of ow, etc., without ad hoc tuning. The development of dynamic SGS modeling Germano et al.21 , however, has considerably improved the viability of LES as
1208 Phys. Fluids, Vol. 8, No. 5, May 1996

a tool for prediction of complex ows since the eddy viscosity is calculated during the course of the computation and in turn reects local properties of the ow e.g., see Squires and Piomelli22 . SGS models which reect local properties of the turbulence are especially attractive for the computation of particle-laden ows since predictions of particle transport should be expected to be signicantly improved by more accurate SGS models. Thus, the principal objective of this work is application of large eddy simulation to computation of a well-dened turbulent shear ow, fully-developed channel ow, for which DNS results and experimental measurements exist for comparison and evaluation of LES predictions. As discussed above, particle interactions with turbulence in wall-bounded shear ows result in both complex statistical behavior of the particle velocity eld as well as complicated structural features, i.e., preferential concentration. Therefore, a primary interest of this study is to determine the utility of LES for prediction of these effects. Contained in Sec. II is an overview of the simulations. Comparison of LES predictions to DNS results as well as experimental data is presented in Secs. III and IV. In addition to statistical properties such as mean and uctuating particle velocities, instantaneous particle distributions near the channel wall and centerline are analyzed in detail, both qualitatively and quantitatively, to investigate the degree to which LES represents preferential concentration. A summary of the work may be found in Sec. V.
II. SIMULATION OVERVIEW A. LES of turbulent channel ow

The turbulent ow between plane channels driven by a uniform pressure gradient was calculated using LES of the incompressible NavierStokes equations. The equations governing transport of the large eddies obtained by ltering the NavierStokes equations are i u xi i u t 0, i j uu p xi 1 Re
2 ui ij i1 ,

1 2

xj

xj xj

xj

where u i is the uid velocity, p is the pressure, i j is the Kronecker delta, and i 1,2,3 refers to the x streamwise , y normal , and z spanwise directions, respectively the usual summation notation applies and an overbar, , denotes application of the ltering operation . The governing equations 1 and 2 have been made dimensionless using the channel half-width and friction velocity u , yielding a mean pressure gradient of 1 corresponding to i1 in Eq. 2 . The Reynolds number in 2 is Re u / , where is the kinematic viscosity. For fully-developed channel ow periodic boundary conditions for the dependent variables are applied in the streamwise and spanwise directions, whereas the no-slip condition is applied on the channel walls. The effect of the subgrid scales on the resolved eddies in 2 is represented by the SGS stress, i j u i u j i j . In this uu work i j is parametrized using an eddy viscosity hypothesis,
Q. Wang and K. D. Squires

Downloaded19Jul2004to195.83.232.39.RedistributionsubjecttoAIPlicenseorcopyright,seehttp://pof.aip.org/pof/copyright.jsp

ij

1 3

i j kk

TS i j

where the eddy viscosity is

T

C 2 , S

the resolved-scale strain rate tensor is dened as S ij 1 2 i u xj j u , xi 5

S and S 2S i ji j is the magnitude of i j . The lter width S is dened as ( ) 1/3 where , , and are 1 2 3 1 2 3 the grid spacings in the x, y, and z directions, respectively. The model coefcient C in 4 requires specication in order to close the system 1 and 2 . The model coefcient C is determined dynamically using the resolved velocity eld. Following Germano et al.,21 a second lter, the test lter denoted using , is introduced to derive an expression for C. Germano23 showed that the resolved turbulent stress, L i j i j i j , uu uu 6

the SGS stress i j , and the subtest-scale stress T i j u i u j i j , are related by uu L i j T i j j . i 7 Assuming that a closure similar to 3 may be used to model the test-eld stress T i j , it is possible to use 7 to derive an expression for C. The model coefcient C was calculated following the approach developed by Meneveau et al.24 in which the error in 7 is minimized along uid particle trajectories, resulting in an expression for the model coefcient of the form C x,t I LM . IMM 8

ther discussion . The test lter width in the streamwise and spanwise directions was twice the grid lter width also see Germano et al.21 . Test ltering in the streamwise and spanwise directions was carried out in Fourier space using a sharp cut-off lter. The governing equations 1 and 2 were solved numerically using the fractional step method on a staggered grid e.g., see Kim and Moin,25 Perot,26 Wu et al.27 . Second-order AdamsBashforth was used for advancement of the convective terms and part of the SGS term while the CrankNicholson method was applied for an update of the viscous terms and a portion of the SGS stress. The Poisson equation for pressure was solved using Fourier series expansions in the streamwise and spanwise directions together with tridiagonal matrix inversion. Large eddy simulations were performed at Reynolds numbers based on friction velocity and channel half-width of 180 and 644 corresponding to Reynolds numbers of 3,200 and 13,800 based on centerline velocity and channel halfwidth . At both Reynolds numbers the ow was resolved using 64 65 64 grid points in the x, y, and z directions, respectively. The channel domain for the calculation at 2 4 /3 and 5 /2 2 /2 at Re 180 was 4 Re 644. Previous computations of turbulent channel ow have shown these domain sizes are large enough to avoid contamination of the ow by periodic boundary conditions Piomelli28 . The grid spacing in wall coordinates in the x and z directions was x 35 and z 12 at Re 180 79 and z 16 at Re 644. A stretched grid and x was used in the wall-normal direction and for both Reynolds 1. numbers the rst grid point was at y
B. Calculation of particle trajectories

In principle, I LM and I M M are obtained from the solution of separate transport equations. However, to reduce the computational cost associated with calculation of C, the numerator and denominator of 8 are obtained using a simple time discretization, resulting in In 1 x LM
1 In M x M

The particle equation of motion used in the simulations describes the motion of particles with densities substantially larger than that of the surrounding uid and diameters small compared to the Kolmogorov scale: dvi dt
f p

3 CD v u vi ui 4 d

i1 ,

13

L nj 1 M nj i i
1

x 1

I n x n t , ui LM 9 I n M x n t , ui M 10

M nj 1 M nj i i

where H x max(x,0) is the ramp function, the timescale in 1/4 10 is dened as T 2 I LM , and t/T n . 1 t/T n 11

where v i is the velocity of the particle, u i is the velocity of the uid at the particle position, and g is the acceleration of gravity. The body force acts along the positive streamwise direction corresponding to ow in a vertical channel. The uid and particle densities in 13 are denoted f and p , respectively, and d is the particle diameter. Previous computations of particle-laden turbulent channel ow have shown that the particle Reynolds number, Rep v u d/ , does not necessarily remain small e.g., see Rouson and Eaton,7 Wang and Squires29 . Therefore, an empirical relation for C D from Clift et al.30 valid for particle Reynolds numbers up to about 40 was employed, CD 24 1 0.15Re0.687 . p Re p 14

The quantity M i j is dependent upon the closure approximation and for an eddy viscosity model is Mij 2 SS
ij 2

. S Sij I n M (x M

12

i t) and i t) are obtained The values of u u through linear interpolation see Meneveau et al.24 for a furPhys. Fluids, Vol. 8, No. 5, May 1996

I n (x LM

It should also be noted that 13 is appropriate for describing the motion of smooth rigid spheres and neglects the inuence of virtual mass, buoyancy, and the Basset history force on
Q. Wang and K. D. Squires 1209

Downloaded19Jul2004to195.83.232.39.RedistributionsubjecttoAIPlicenseorcopyright,seehttp://pof.aip.org/pof/copyright.jsp

TABLE I. Particle parameters in turbulent channel ow, Re 28 m Lycopodium

p p

180. 70 m copper 4.50 810 0.00388 0.70

50 m glass 0.65 117 0.00277 0.50

/( /u )

a/ a

0.048 9 0.00139 0.25

particle motion. For particles with material densities large compared to the uid these forces are negligible compared to the drag. Previous investigations have shown that the effect of the lift force, while relevant to problems of particle deposition, is less signicant to this work and therefore the effect of shear-induced lift in the equation of motion has been neglected see Wang and Squires29 . Finally, the volume fraction of particles is assumed small enough such that particleparticle interactions are negligible. From computation of an Eulerian velocity eld, 13 was integrated in time using second-order AdamsBashforth. Since it is only by chance that a particle is located at a grid point where the Eulerian velocity eld is available, fourthorder Lagrange polynomials were used to interpolate the uid velocity to the particle position see Wang et al.31 for further discussion . Particle displacements were also integrated using the second-order AdamsBashforth method. For particles that moved out of the channel in the streamwise or spanwise directions periodic boundary conditions were used to reintroduce it in the computational domain. The channel walls are perfectly smooth and a particle was assumed to contact the wall when their center was one radius from the wall. Elastic collisions were assumed for particles contacting the wall. Properties of the dispersed phase were obtained by following the trajectories of 250,000 particles. The trajectories of a large ensemble of particles are required in order to present statistics for the dispersed phase in the same manner as for the uid, i.e., by averaging over homogeneous planes. Numerical experiments demonstrated adequate statistical convergence was obtained using this sample size see the Appendix for further discussion . The particles used in the simulations possess material characteristics identical to those in the DNS study of Rouson and Eaton7 and experiments of Kulick et al.8 and Fessler et al.9: 28 m diameter Lycopodium particles, 25, 50, and 90 m diameter glass beads, and 70 m copper particles. The particle response time and radius a expressed in terms of both channel variables ( and u ) and in wall units for the simulations at Re 180 and Re 644 are shown in Tables I and II, respectively the

2 response time p p d /(18 f ) shown in the table is the value corresponding to Stokes ow around the particle . The particle parameters at Re 644 have been normalized by the experimental values of channel half-width 20 mm and friction velocity u 0.49 m/s see Kulick et al.8 . The particle parameters in the DNS at Re 180 were scaled by Rouson and Eaton7 to match the Stokes numbers dened in terms of the Kolmogorov timescale in the experimental measurements of Kulick et al..8 For both Reynolds numbers the particle radius is much smaller than the lter width except very near the wall where the particle radius can be comparable to the wall-normal grid spacing the rst grid point was 0.45 at Re 180 and at y 0.84 at Re 644). at y Thus, the effect of a nonuniform uid velocity eld on particle motion near the wall may be less accurately represented for the larger particles.

III. PARTICLE VELOCITY STATISTICS

From an arbitrary initial condition the Eulerian velocity eld was time advanced to a statistically stationary state. The particles were then assigned random locations throughout the channel. The initial particle velocity was assumed to be the same as the uid velocity at the particle location. Particles were then time advanced in the ow eld to a new equilibrium condition in which particle motion was independent of initial conditions. Similar to the uid ow, statistics of the particle velocity were averaged over the two homogeneous directions, both channel halves, and time. Fluid statistics were averaged for 3.5 /u at Re 180 and for 4 /u at Re 644. The development time, i.e., the time required for particles to become independent of their initial conditions was larger for increasing values of the particle response time, e.g., 0.5 /u for the Lycopodium particles and 6 /u for the copper particles at Re 180. After an equilibrium condition had been reached, particle statistics were accumulated for 6 /u at Re 180 and 4.5 /u at Re 644. The mean streamwise velocity prole obtained from the LES calculations is shown in Figure 1 for each particle type considered in the DNS Fig. 1 a and experiments Fig. 1 b . Figure 1 a shows that at Re 180 there is good agreement between LES and DNS. As expected, both the LES and DNS show that the Lycopodium particles closely 10) the Lycopodium track the uid ow. Near the wall (y velocity prole from the DNS slightly lags that of the uid while the mean velocity from the LES is nearly equal to the uid velocity. Rouson and Eaton7 attribute the lag in the Lycopodium prole to preferential concentration of Lycopodium particles in the low-speed streaks. The discrepancy between LES and DNS results may be related to the fact that
644. 25 m glass 0.12 77 0.000625 0.403 50 m glass 0.44 283 0.00125 0.805 90 m glass 1.54 992 0.00225 1.449 70 m copper 3.10 1996 0.00175 1.127

TABLE II. Particle parameters in turbulent channel ow, Re 7 m Lycopodium

p p

14 m Lycopodium 0.01 6 0.00035 0.225

28 m Lycopodium 0.04 26 0.0007 0.450

/( /u )

a/ a

0.0025 2 0.000175 0.113

1210

Phys. Fluids, Vol. 8, No. 5, May 1996

Q. Wang and K. D. Squires

Downloaded19Jul2004to195.83.232.39.RedistributionsubjecttoAIPlicenseorcopyright,seehttp://pof.aip.org/pof/copyright.jsp

FIG. 1. Mean streamwise velocity in turbulent channel ow. a Re 180; b Re 644. LES:, uid elements; ---, Lycopodium; , 50 m glass; , copper; Rouson and Eaton7 in a and Kulick et al.8 in b : ; Lycopodium; , 50 m glass; *, copper.

the near-wall structures are less well resolved in the LES and, consequently, preferential concentration of particles in low-speed regions is less signicant than in the DNS, resulting in an over-prediction of the mean velocity of the Lycopodium particles see Sec. IV A for further discussion . For the glass beads and copper particles the proles in Figure 1 a become increasingly larger than the uid velocity for increases in the Stokes number. LES predictions of the mean velocity of 50 m glass beads are in good agreement with the DNS; the copper velocity prole from the LES calculations is also in good agreement with DNS results for 10 but it may be observed from the gure that the DNS y results are larger near the wall. Note also that both LES and DNS proles of the copper particles exhibit a slight plateau near the wall. The plateau may result from the transport of high velocity particles from the outer region of the channel to the near-wall region and this feature appears to be somewhat more pronounced in the DNS. However, it is also important to note that the difference between LES and DNS is relatively small and conned to a very thin region near the wall. Particle mean velocity proles at Re 644 from the LES calculations are shown in Figure 1 b together with the experimental measurements for 50 m glass beads and copper particles from Kulick et al.8 Similar to the results at the lower Reynolds number, LES predictions show greater difPhys. Fluids, Vol. 8, No. 5, May 1996

ferences in the mean particle velocity relative to the uid with increasing Stokes number; the largest differences occurring near the wall. As also observed at Re 180, near the wall the Lycopodium particles slightly lead the uid and are then nearly identical to the uid velocity in the outer region. Comparison of the proles for the 50 m glass beads demonstrates relatively good agreement for y greater than about 20 the results in Figure 1 b show that 20. However, for y the experimental measurements increase towards the wall while LES predictions do not. It may also be observed that there is relatively poor agreement between the LES and Kulick et al.8 for the copper particles. For y greater than roughly 10, the mean prole of the copper particles in the experiment is nearly equal to that of the uid while in the LES the copper particles lead the uid throughout the channel. As also apparent in Figure 1 b , Kulick et al.8 found that the mean velocity of the copper particles increased substantially near the wall, an effect not observed in the LES at either Reynolds number. It should also be noted that the mass loading corresponding to the experimental measurements shown in the gure is 2%. Kulick et al.8 also measured particle velocity statistics at higher mass loading where the effect of particles on uid turbulence becomes signicant. Thus, it is clear there are substantial differences between the LES and experimental measurements of the copper velocities at Re 644. Kulick et al.8 attributed the increase in near-wall copper velocity to the possibility that high-speed particles rebounding from the wall maintain a signicant fraction of their streamwise momentum. The plateau in the copper velocity prole near the wall is consistent with notion of elastic collisions of high-speed particles contacting the wall and is also in agreement with recent examinations of particle deposition which have shown that particles are brought to the near wall region by events with normal velocity much larger than the local turbulence intensity Brooke et al.12 . Unlike the experimental measurements, however, the resulting copper mean velocity in the LES does not increase in the near wall region. Comparison of rms particle velocity uctuations from the LES to both DNS results and experimental measurements is shown in Figure 2 for Re 180) and Figure 3 for Re 644). As may be observed in Figure 2, there is in general good agreement between LES predictions and the DNS results of Rouson and Eaton.7 It is clear from the gure that near the wall the streamwise uctuation levels increase with increasing values of the Stokes number while the wallnormal and spanwise uctuations are reduced. Comparison of the LES and DNS results also shows that the peak values in the LES proles occur at slightly larger y than in the DNS. It is further interesting to note that the streamwise rms velocities of the Lycopodium and glass beads from the LES calculations are smaller than the DNS values while the wallnormal and spanwise rms velocities in the LES are slightly greater than the corresponding values in the DNS. Rms intensities from the channel at Re 644 are shown in Figure 3. Velocity proles for the Lycopodium particles are not available from the experiments; LES predictions are shown in Figure 3 a for comparison and are consistent with
Q. Wang and K. D. Squires 1211

Downloaded19Jul2004to195.83.232.39.RedistributionsubjecttoAIPlicenseorcopyright,seehttp://pof.aip.org/pof/copyright.jsp

FIG. 2. Root-mean-square velocity uctuations in turbulent channel ow, Re 180. a Lycopodium; b 50 m glass; c copper. LES uid: , i 1;, i 2; i 3; LES particle:, i 1; i 2; ---, i 3; particle uctuations from Rouson and Eaton:7 , i 1; * i 2; i 3.

FIG. 3. Root-mean-square velocity uctuations in turbulent channel ow, Re 644. a Lycopodium; b 50 m glass; c copper. LES uid: , i 1;, i 2; i 3; LES particle:,i 1; , i 2; ---, i 3; particle uctuations from Kulick et al.:8 , i 1; * , i 2.

the data at the lower Reynolds number, i.e., the Lycopodium uctuations are nearly equal to the uid values but lead in the streamwise direction while lagging in the wall-normal and spanwise directions. Experimental measurements of particle velocities are available for the 50 m glass beads and copper particles and it is evident in Figure 3 b that there is good agreement between LES predictions and the measured streamwise intensities of the glass beads for y 10. The wall-normal uctuations in the experiment are greater than
1212 Phys. Fluids, Vol. 8, No. 5, May 1996

the LES values but the location of the peak in the wallnormal uctuations is reasonably well predicted. The greatest discrepancy between LES and experimental measurements occurs for the copper particles. Figure 3 c shows that the streamwise intensities in the experiment are larger than the LES predictions, and unlike the LES and DNS values at Re 180, the streamwise intensities in the experiment peak at around y 12. Kulick et al.8 showed that the probability
Q. Wang and K. D. Squires

Downloaded19Jul2004to195.83.232.39.RedistributionsubjecttoAIPlicenseorcopyright,seehttp://pof.aip.org/pof/copyright.jsp

distribution function pdf of the streamwise copper velocity 12, demonstrating that the streamwise was bimodal at y copper intensities should not be interpreted as the width of a Gaussian velocity distribution. Pdfs of the copper velocities in the LES around y 12 were examined and not found to exhibit a similar bi-modal structure as in the experiment. Finally, comparison of the spanwise uctuation levels for all the particles show a reduction with increasing Stokes number and greater similarity to the uid for the smaller Stokes numbers. It is difcult to speculate as to the precise cause of the differences between LES predictions and both DNS results and experimental measurements. Errors in the SGS model used in LES contribute to the differences as well as other factors such as the different particle sample sizes and the interpolation scheme used to obtain uid velocities at particle positions in the computations. Another source of error in LES predictions of the particle velocity statistics is due to the neglect of particle transport by SGS velocities. In LES the smallest scales of motion are not resolved by the computational grid, only their effect on the large eddies is represented via the SGS model. Thus, only the large-scale velocity eld is directly available in a LES computation for determining particle motion; for the results presented above the effect of subgrid-scale velocity uctuations on particle transport were neglected. One measure of the error is the value of the particle relaxation time relative to the smallest S S resolved timescale in the LES, T /( ) 1/ . At Re 180 the timescale T increases from about 0.011 near the wall to roughly 0.11 near the channel centerline while at Re 644 increases from 0.002 to 0.08. Thus, some of the particle relaxation times considered in the LES are comparable to the smallest resolved timescale see Tables I and II . The effect of the SGS velocity eld on particle transport was investigated by adding SGS uctuations to the uid velocity used in the particle equation of motion 13 . Calculations were performed in the channel ow at Re 644 in which the uid velocity in the particle equation of motion was the resolved component i , directly available in the u LES, plus a subgrid contribution u i . The magnitude of the SGS uctuation u i was determined by solving a transport equation for SGS kinetic energy, q 2 . The transport equation used for determination of q 2 is that proposed by Schumann,32 i.e., q2 t j u q2 xj 2 2 S
2 2

FIG. 4. Rms velocity uctuations in turbulent channel ow, Re q 2 /3; ---, u 1,rms ; , u 2,rms ; , u 3,rms .

644.

solved wall-normal velocity near the wall, and the SGS energy peaks at about the same location as the resolved streamwise uctuations. In the simulations SGS intensities were obtained from q 2 and specied to have the same relative magnitudes as the resolved-scale intensities. The component uctuations u i were then scaled by random numbers sampled from a Gaussian distribution and added to i at the particle location. The u complete velocity, i.e., i u i , was subsequently used in u 13 to determine the particle velocity. Figure 5 shows the wall-normal uctuations for the 28 m Lycopodium particles at Re 644 both with and without SGS velocity uctuations included in 13 . As is evident from the gure there is a negligible effect of the SGS uid velocity on particle uctuations. The relative difference is greatest near the wall where SGS uctuations are large compared to the resolved components, but the change in the wall-normal particle velocities due to the addition of the SGS uid velocity is less than 1%. Although not shown here, the difference in particle uctuations in the other directions is similar and there is essentially no effect on the mean streamwise particle velocity. Figure 5 demonstrates the strong ltering effect of par-

xj

1 l 3

q 2 /2

q2 xj 15

xj xj where c 2 3k 0
3/2

1 q3 c , 2 l

,l

min ,c y ,

16

with the Kolmogorov constant k 0 1.6. Shown in Figure 4 is the prole of the SGS uctuation q 2 /3 along with the resolved components. SGS uctuations are larger than the rePhys. Fluids, Vol. 8, No. 5, May 1996

FIG. 5. Wall-normal velocity uctuations for 28 m Lycopodium particles, Re 644. , without SGS velocity in the particle equation of motion; ---, with SGS velocity in the particle equation of motion. Q. Wang and K. D. Squires 1213

Downloaded19Jul2004to195.83.232.39.RedistributionsubjecttoAIPlicenseorcopyright,seehttp://pof.aip.org/pof/copyright.jsp

ticle inertia on the uid velocity spectrum. For particles with material densities large compared to the carrier ow the response of particles to the frequency spectrum of the turbulence can be shown to be proportional to 1/( p ) 2 ( is the frequency . Thus, for increasing values of the relaxation time and/or frequency the ltering of high frequency motions by particle inertia is signicant, consistent with the results in Figure 5. Figures 2 and 3 show that at both Reynolds numbers the streamwise uctuations of the particle velocities are signicantly larger than the wall-normal or spanwise values and are also larger than the streamwise uctuations in the uid. Furthermore, the wall-normal and spanwise uctuations are reduced for increasing particle inertia while the streamwise values increase see also Rogers and Eaton,14 Kulick et al.8 . The relative strength of the streamwise intensities relative to the other components is clearly illustrated through examination of the diagonal components of b u, v f2 fifi

17

where f is either the uid or particle velocity uctuation. If the uctuations f i are isotropic each component of b should be 1/3 and the deviation from this value is a measure of the anisotropy in f i . Shown in Figures 6 and 7 is a comparison of the diagonal components of the anisotropy tensor for each particle type to those in the uid. It is evident that for all particles considered in the calculations the anisotropy of the particle uctuations is larger than the uid. The uctuation levels of the copper particles exhibit the greatest anisotropy, e.g., the wall-normal and spanwise components are never greater than 0.1. Perhaps more signicant than the increased anisotropy of the particle uctuations is that the streamwise intensities of the particles exhibit signicant increases relative to the uid in the near-wall region and the difference becomes larger with increases in the Stokes number. The enhanced uctuation levels of the particle intensities and larger anisotropy, while counter-intuitive, demonstrates the signicant effect of the mean-velocity gradient on particle uctuations. The effect of the mean velocity gradient has been considered in analyses which are predicated upon an accurate prescription of the Lagrangian correlation of uid elements measured along the trajectory of a particle e.g., see Liljegren,15 Reeks16 . The effect of mean shear on particle intensities may also be considered through direct examination of the particle equation of motion 13 . As shown by Simonin et al.,2 the equation governing the transport of the particle intensities can be written as D v v Dt where D v v Dt
1214

FIG. 6. Diagonal components of the anisotropy tensor, Re 180. a Lycopodium; b 50 m glass; c copper. , b u ; --- b u ; , b u ; , b v ; 11 22 33 11 *, b v ; , b v . 22 33

18

v is the particle velocity uctuation, and V m is the particle mean velocity no summation on repeated Greek indices . The rst term on the right-hand side represents production by the mean gradient of the particle velocity:

Vm

xm

v v ,

19

2 v vm

V . xm

20
Q. Wang and K. D. Squires

Phys. Fluids, Vol. 8, No. 5, May 1996

Downloaded19Jul2004to195.83.232.39.RedistributionsubjecttoAIPlicenseorcopyright,seehttp://pof.aip.org/pof/copyright.jsp

where is the volume fraction of the dispersed phase. The last term represents turbulent momentum transfer and is expressed as
f p

3 CD v u v 2 d

u v

22

As shown by Simonin et al.,2

d d p p

can be approximated as 23

and are destruction and production of parwhere ticle uctuations, respectively, and may be written as
d

2
A

v v ,

2
A

u v ,

24

where
A

is the mean particle relaxation time,

p f

4 d 3 CD

1 v u

25

FIG. 7. Diagonal components of the anisotropy tensor, Re 644. a Lycopodium; b 50 m glass; c copper. , b u ; ---, b u ;; , b u ; b v ; *, 11 22 33 11 bv ; ; bv . 22 33

The second term on the right-hand side of 18 represents transport by the particle velocity uctuations, 1 xm

v v vm ,

21

( C D denotes the drag coefcient averaged over the particles . In canonical wall-bounded ows such as a twodimensional boundary layer or turbulent channel ow the production term P is non-zero only in the streamwise direction. A representative prole of the terms in 18 is shown in Figure 8 for the 50 m glass beads at Re 180. As can be observed in Figure 8 a streamwise uctuations are produced through both interaction with mean gradients, P 11 , as well as through the uid-particle covariance represented by p 11 . It is important to note that the production by the uidparticle covariance is of the same order of magnitude as P 11 and provides a means by which greater anisotropy can occur in the particle intensities. In fact, though not shown p here, the production term 11 for the Lycopodium particles is substantially larger than that due to mean gradients. As can also be observed in each of Figure 8, production of streamwise intensities is balanced primarily by the contribution from d . It is interesting to note, however, that close to the 11 wall the attenuation of particle intensities is balanced mostly by the turbulent transport term D was calculated as the difference between the other terms in Eq. 18 . Similar behavior has also been observed by Simonin et al.2 Comparison of the gures also shows that the terms in 18 for the wall-normal and spanwise components are substantially smaller than for the streamwise component. Peaks in both p and d for 2 and 3 are also collocated with the peaks in the intensities shown in Figure 2. The results in Figure 8 clearly show that turbulent moacts as both a source and sink for mentum transfer particle intensities. In particular, the uid-particle covariance has a substantial effect on particle uctuation levels through p . Shown in Figure 9 are the non-zero terms of the uidparticle covariance at Re 180. As is evident in Figure 9 a the covariance u 1 v 1 is the largest for the Lycopodium particles and decreases with increasing Stokes number. Comparison of the gures also shows that streamwise component of the covariance is substantially larger than the other components. It should then be expected that the production term p will become increasingly important at smaller Stokes
Q. Wang and K. D. Squires 1215

Phys. Fluids, Vol. 8, No. 5, May 1996

Downloaded19Jul2004to195.83.232.39.RedistributionsubjecttoAIPlicenseorcopyright,seehttp://pof.aip.org/pof/copyright.jsp

see also Reeks,19 Liljegren33 . Simonin et al.2 derived second-order closures for these quantities and show that production of u 2 v 1 is proportional to the mean gradient of the particle velocity while u 1 v 2 is proportional to the mean gradient of the uid velocity and that the non-symmetrical form of the closure is related to the modeling of the pressure correlations, specically the rapid pressure term. The other diagonal components of the uid-particle covariance in Figure 9 show that the peak of the wall-normal component, u 2 v 2 , occurs at slightly larger y than the spanwise value, u 3 v 3 . This behavior correlates both with the location of p p the peak in 22 and 33 .
IV. PREFERENTIAL CONCENTRATION A. Near-wall region

FIG. 8. Terms in 18 at Re 180 for 50 2; c 3. , D ; ---, d ; ,

m glass beads. a ; , P .

1; b

numbers. It is also interesting to note that the off-diagonal components peak at approximately the same y . For the Lycopodium particles the covariance tensor is nearly symmetric with u 2 v 1 u 1 v 2 . For the glass beads and copper particles, however, the asymmetry of the covariance becomes more apparent with u 2 v 1 becoming increasingly larger than u 1 v 2 . Thus, lower-level closure models for these quantities should necessarily reect this asymmetry
1216 Phys. Fluids, Vol. 8, No. 5, May 1996

Both experimental measurements and numerical simulations have shown that inertial bias in particle trajectories results in a preferential concentration of particles in regions of low vorticity or high strain rate see Eaton and Fessler34 for a general review . In wall-bounded shear ows it is reasonably well known that particle concentration elds near the wall are non-uniform with the largest number densities occurring in the low-speed streaks Pedinotti et al.,6 Rouson and Eaton7 . It is also important to point out, however, that recent investigations have shown that preferential concentration occurs throughout the channel Fessler et al.9 . Furthermore, Wang and Maxey4 have shown that preferential concentration obeys Kolmogorov scaling, i.e., particles with time constants and settling velocities close to the Kolmogorov scales will exhibit the largest effects of preferential concentration. Since the smallest scales of motion in LES are not resolved, it is of interest to examine the degree to which preferential concentration is reproduced in the present LES. Shown in Figure 10 a are streamwise velocity contours in an x-z plane at y 3.6 at Re 180. As evident in the gure, the streaky structure of the near-wall region is represented in the LES. The instantaneous particle distribution in the same plane and at the same time is also shown in Figure 10. A streaky structure in the number density of the Lycopodium particles similar to that observed in the velocity eld is apparent in Figure 10 b . It is also clear from the gure that the number density is less well organized for the 50 m glass beads and the copper particle distribution in Figure 10 d appears random. Similar behavior was also observed by Rouson and Eaton7 using DNS and demonstrates that at Re 180 the LES does represent, at least qualitatively, preferential concentration of particles by turbulence. The stream4.8 and wise velocity contours in an x-z plane at y Re 644 are shown in Figure 11 a . As can be seen from the gure, the streaky structure of the streamwise velocity is also evident at the higher Reynolds number. Number density distributions for the 28 m Lycopodium, 50 m glass beads, and 70 m copper particles are shown in Figures 11 b , 11 c , and 11 d , respectively, and are similar to those obtained at the lower Reynolds number. Lycopodium particles again exhibit a structure somewhat similar to the velocity eld and appear more organized than the glass beads and copper particles.
Q. Wang and K. D. Squires

Downloaded19Jul2004to195.83.232.39.RedistributionsubjecttoAIPlicenseorcopyright,seehttp://pof.aip.org/pof/copyright.jsp

FIG. 9. Fluid-particle covariance in turbulent channel ow, Re d copper., u 1 v 2 ; ---, u 2 v 1 ; , u 2 v 2 ; , u 3 v 3 .

180. a

u 1 v 1 . , Lycopodium; ---, glass; , copper. b Lycopodium; c 50

m glass;

A quantitative measure of preferential concentration in the near-wall region is the ratio of the streamwise uid velocity at the particle location to the average uid velocity in the same plane. This measure was rst used by Pedinotti et al.6 and should be unity for a random distribution of particles i.e., no preferential concentration and smaller than unity if particles are preferentially concentrated in low-speed regions. Figure 12 shows the pdf of this quantity at both Reynolds numbers. Consistent with the instantaneous distributions in Figures 10 and 11, the pdf of the Lycopodium particles exhibits the greatest difference as compared to a random distribution while the glass beads and copper particles possess distributions closer to random. In particular, the higher probability of small values of the streamwise velocity being measured in the vicinity of Lycopodium particles indicates a preferential concentration in low-speed regions.
Phys. Fluids, Vol. 8, No. 5, May 1996

B. Channel centerline

Fessler et al.9 have recently demonstrated that preferential concentration also occurs along the centerline of turbulent channel ow. In addition to the ve types of particles examined in the experiments two sets of Lycopodium particles with smaller diameters were used in the LES at Re 644 to further examine the effect of response time on preferential concentration see Table II . Both visualizations and quantitative measures were obtained in the experiments and therefore provide a suitable benchmark for comparison to LES predictions. The instantaneous particle distributions along the channel centerline from the LES calculation at Re 644 are shown in Figure 13. Similar to the behavior observed by Fessler et al.,9 the gure show that the copper particles are randomly distributed whereas the Lycopodium particles and
Q. Wang and K. D. Squires 1217

Downloaded19Jul2004to195.83.232.39.RedistributionsubjecttoAIPlicenseorcopyright,seehttp://pof.aip.org/pof/copyright.jsp

FIG. 10. Velocity contours and particle distributions from LES at t 6 /u , y m glass; d 70 m copper.

3.6, Re

180. a Streamwise velocity; b 28

m Lycopodium; c 50

glass beads exhibit varying degrees of preferential concentration. Of the three types of particles shown in Figure 13 a c , the preferential concentration of the 25 m glass beads appears more signicant than for the Lycopodium particles, which is in turn stronger than for the 50 m glass beads. Thus, results in the gures again demonstrate that the LES reects preferential concentration of particles by turbulence and exhibits the same qualitative features as in the experiments. One approach to quantifying preferential concentration
1218 Phys. Fluids, Vol. 8, No. 5, May 1996

is through calculation of the pdf of the particle number density. For a random distribution of particles the pdf is Poisson distributed, Fp k e k!
k

26

where is the average number of particles per cell. Shown in Figure 14 are pdfs of the particle number density at Re 644 together with the Poisson distribution. The pdf was
Q. Wang and K. D. Squires

Downloaded19Jul2004to195.83.232.39.RedistributionsubjecttoAIPlicenseorcopyright,seehttp://pof.aip.org/pof/copyright.jsp

FIG. 11. Velocity contours and particle distributions from LES at t 3 /u , y m glass; d 70 m copper.

4.8, Re

644. a Streamwise velocity; b 28

m Lycopodium; c 50

calculated by subdividing the region 0.975 y/ 1.025 into a network of cells of cross-sectional area 2.6 mm2 in the x-z plane . The pdf of copper particles is very similar to the Poisson distribution whereas the other particles differ signicantly from a random distribution; the greatest difference appearing to occur for the 25 m glass beads. Thus, consistent with Figure 13 as well as Fessler et al.,9 Figure 14 shows that preferential concentration is not a monotonic function of the Stokes number. Fessler et al.9 also dened as a measure of preferential concentration the deviation from a Poisson distribution,
Phys. Fluids, Vol. 8, No. 5, May 1996

27

where and p are the standard deviations for the measured particle distribution and the Poisson distribution, respectively. For particles exhibiting preferential concentration some cells have large number densities, whereas other cells substantially lower number densities relative to the mean, resulting in large positive values of D. Since number densities are obtained by dening a network of cells in an x-z plane, it is therefore important to consider the effect of the
Q. Wang and K. D. Squires 1219

Downloaded19Jul2004to195.83.232.39.RedistributionsubjecttoAIPlicenseorcopyright,seehttp://pof.aip.org/pof/copyright.jsp

FIG. 13. Particle distribution for the region 0.975 y/ 1.025 at t 2 /u , Re 644. a 28 m Lycopodium; b 25 m glass; c 50 m glass; d 70 m copper.

FIG. 12. Probability distribution function of the difference between the uid velocity at the particle location and average uid velocity. a Re 180, pdf measured for 5.4 y 13.4; b Re 644, pdf measured for 5.7 y 13.7. ,random distribution; ---, Lycopodium; , 50 m glass; copper.

cell size when calculating D. For very small cell sizes, the dimension of a cell is smaller than the Kolmogorov lengthscale and the particle distribution will appear random, resulting in D being zero. For very large cells, regions of high and low particle number density will be contained within the same cell and the resulting value of D will also be close to zero. Between these two extremes there is a cell size where D is maximized see Fessler et al.9 for a further discussion . As shown in Fessler et al.,9 the cell size at which 27 is maximum is a function of the particle type. Therefore, pdfs of the number density eld along the channel centerline were calculated using several cell sizes. The maximum value from 27 , D max , in the LES is compared to Fessler et al.9 in Figure 15 a . Both the LES and the experiments indicate that D max exhibits a peak and is largest for the 25 m glass beads. Wang and Maxey4 used DNS of isotropic turbulence and showed that preferential concentration of heavy particles obeys Kolmogorov scaling. Fessler et al.9 estimated the Kolmogorov timescale and found that the ratio of the response time for 25 m glass beads to the Kolmogorov timescale is 2.2. However, the Stokes number dened in terms of the Kolmogorov timescale for the 28 m Lycopodium is 0.74 and thus the results in Figure 15 a seem to contradict Wang and Maxey.4 Fessler et al.9 attributed a possible cause of the discrepancy to the wider range of lengthscales in the experiment as compared to DNS. In this regard it is interesting to
1220 Phys. Fluids, Vol. 8, No. 5, May 1996

note that while the LES is performed at the same Reynolds number as the experiment, the range of scales in the computation is smaller since the subgrid-scale motions are not resolved. Thus, it is unlikely that the 25 m glass beads exhibit slightly stronger effects of preferential concentration because of an extended range of scales. A more likely cause of the discrepancy, which is discussed by Fessler et al.,9 is the difference between the cell size used for computation of the number density eld relative to the Kolmogorov lengthscale in the experiment and DNS. Fessler et al.9 found that for smaller cell sizes the Lycopodium particles exhibited slightly larger values of D as compared to that for the 25 m glass beads. Similar behavior was observed when analyzing the LES results. Wang and Maxey4 used a somewhat analogous measure as Fessler et al.9 in quantifying preferential concentration, Dk F k F p k 2 , 28

k 0

where F(k) and F p (k) are the pdfs for the actual and random distributions, respectively. LES results show that, similar to D, D k also exhibits a maximum which is dependent upon the cell size over which the number density eld is dened. The maximum value is shown in Figure 15 b and again shows a peak for the 25 m glass beads, conrming that particles with relaxation times close to the Kolmogorov timescale exhibit the strongest effects of preferential concentration.
V. SUMMARY

Large eddy simulations were carried out and particle transport was studied in turbulent channel ows at Reynolds 180 and Re 644, corresponding to numbers Re u /
Q. Wang and K. D. Squires

Downloaded19Jul2004to195.83.232.39.RedistributionsubjecttoAIPlicenseorcopyright,seehttp://pof.aip.org/pof/copyright.jsp

FIG. 14. Probability distribution function of particle number density in 0.975 y/ 1.025 at t 2 /u , Re mm. a 28 m Lycopodium; b 25 m glass; c 50 m glass; d 70 m copper.

644. Cell size used for calculation of pdf is 2.6

Rec U c / 3,200 and Rec 13,800, respectively. The Lagrangian dynamic eddy viscosity model of Meneveau et al.24 was used to parametrize subgrid-scale stresses in the calculations. Several statistical measures of the particle velocity and concentration eld together with instantaneous number density distributions were obtained from the calculations and compared to DNS results and experimental measurements. Both the particle mean velocity and rms uctuations from the LES are in good agreement with the DNS results of Rouson and Eaton7 at Re 180 and demonstrates that LES is nearly as accurate as the DNS. Reasonable agreement is obtained with the experimental measurements of Kulick et al.9 at the higher Reynolds number, except in the near-wall region. DNS results also do not agree particularly well with the experiments near the wall, indicating that the discrepancy in this region is probably related to the modeling assumptions used for the particle phase which is not treated exactly, even in the DNS . Consistent with previous analyses and measurePhys. Fluids, Vol. 8, No. 5, May 1996

ments, particle uctuation levels in the streamwise direction are greater than those in the uid and increase with increasing particle inertia. LES results show the importance of production by both the mean velocity gradient as well as the uid-particle covariance term. Preferential concentration was found to be reasonably well reproduced in the LES both near the wall and along the channel centerline. Visualizations and statistical measures are in good agreement with those obtained by Rouson and Eaton7 and Fessler et al.9 Discrepancies between the numerical simulations, whether LES or DNS, and experimental measurements may be due to factors such as electrostatic effects and particle collisions present in the experiment but currently not incorporated into either DNS and LES. Sommerfeld35 has shown that small changes in the particle-wall collision model can have a relatively large effect on statistics of the particle velocity, especially for particles with large time constants. Incorporation of electrostatic effects, different wall collisions
Q. Wang and K. D. Squires 1221

Downloaded19Jul2004to195.83.232.39.RedistributionsubjecttoAIPlicenseorcopyright,seehttp://pof.aip.org/pof/copyright.jsp

FIG. 15. Maximum values of D dened in 27 and D k dened in 28 for particles in the region 0.975 y/ 1.025, Re 644. a D max ; b D k,max . , LES; *, Fessler et al.9.

models, as well as representations of particle-particle collisions should be expected to shed light on the differences between the LES and DNS and experimental measurements. Accounting for particle collisions may be quite important since their effect is thought to reduce the anisotropy of particle uctuation levels Simonin et al.2 . Incorporation of a model for SGS velocity uctuations yielded a very little effect on particle velocity statistics. Given the relatively ac-

FIG. 17. Effect of sample size on the rms uctuating velocity in turbulent channel ow, Re 180. a streamwise; b wall-normal; c spanwise. , Eulerian; ---, 100 000 particles; , 250 000 particles; , 500 000 particles.

curate predictions of particle statistics at the moderate Reynolds numbers considered in this study, higher Reynolds number calculations i.e., coarser numerical resolution or statistics more sensitive to the small-scale velocity eld are needed to examine the effect of the SGS velocity on particle transport.
APPENDIX A: SAMPLE SIZE
FIG. 16. Effect of sample size on the mean velocity in turbulent channel ow, Re 180. , Eulerian; ---, 100 000 particles; , 250 000 particles; , 500 000 particles. 1222 Phys. Fluids, Vol. 8, No. 5, May 1996

Sample sizes necessary for obtaining a continuum representation of particle statistics were determined using velocity
Q. Wang and K. D. Squires

Downloaded19Jul2004to195.83.232.39.RedistributionsubjecttoAIPlicenseorcopyright,seehttp://pof.aip.org/pof/copyright.jsp

elds from LES of the channel ow at Re 180. Varying numbers of particles were randomly distributed throughout the channel. Statistics were then calculated by interpolating uid velocities to the particle position and averaging over x-z planes. Statistics obtained in this manner were then compared to the mean velocity and rms intensities obtained from the Eulerian grid. For a sufciently large sample size, the statistics of the velocity v i obtained from the particles should be the same as that obtained from the Eulerian eld i . u Shown in Figure 16 is a comparison of the Eulerian mean velocity to the mean proles obtained using 100 000, 250 000 and 500 000 particles. It is clear that the mean prole is adequately resolved using 100 000 particles. However, Figure 17 suggests that this is not the case for the root-meansquare rms uctuating velocity. The rms uctuations obtained using 100 000 particles differ from the value obtained from the grid while the proles obtained using 250 000 and 500 000 particles are quite close to the Eulerian values. Based upon the results in Figures 16 and 17, a sample size of 250 000 particles is adequate for resolution of the mean and rms proles and was used for the simulations reported in this work.
ACKNOWLEDGMENTS

The authors are grateful to Professor John Eaton and Mr. Damian Rouson for supplying the DNS results and experimental measurements as well as providing helpful comments on the manuscript. This work is supported by the National Institute of Occupational Safety and Health Grant No. OH03052-02 . Computer time for the simulations was supplied by the Cornell Theory Center.
A. Berlemont, P. Desjonqueres, and G. Gouesbet, Particle Lagrangian simulation in turbulent ows, Int. J. Multiphase Flow 16, 19 1990 . 2 O. Simonin, E. Deutsch, and M. Boivin, Large eddy simulation and second-moment closure model of particle uctuating motion in two-phase turbulent shear ows, in Turbulent Shear Flow 9, edited by F. Durst, N. Kasagi, B. E. Launder, F. W. Schmidt, and J. H. Whitelaw SpringerVerlag, Heidelberg, 1995 , p. 85. 3 K. D. Squires and J. K. Eaton, Preferential concentration of particles by turbulence, Phys. Fluid A 3, 1169 1991 . 4 L.-P. Wang and M. R. Maxey, Settling velocity and concentration distribution of heavy particles in homogeneous isotropic turbulence, J. Fluid Mech. 256, 27 1993 . 5 M. Rashidi, G. Hetsroni, and S. Banerjee, Particle-turbulence interaction in a boundary layer, Int. J. Multiphase Flow 16, 935 1990 . 6 S. Pedinotti, G. Mariotti, and S. Banerjee, Direct numerical simulation of particle behaviour in the wall region of turbulent ows in horizontal channels, Int. J. Multiphase Flow 18, 927 1992 . 7 D. W. I. Rouson and J. K. Eaton, Direct numerical simulation of particles interacting with a turbulent channel ow, Proceedings of the 7th Workshop on Two-Phase Flow Predictions, edited by M. Sommerfeld Erlangen, Germany, 1994 . 8 J. D. Kulick, J. R. Fessler, and J. K. Eaton Particle response and turbulence modication in fully turbulent channel ow, J. Fluid Mech. 277, 109 1994 .
1

J. R. Fessler, J. D. Kulick, and J. K. Eaton, Preferential concentration of heavy particles in a turbulent channel ow, Phys. Fluids 6, 3742 1994 . 10 J. B. McLaughlin Aerosol particle deposition in numerically simulated channel ow, Phys. Fluid A 1, 1211 1989 . 11 J. W. Brooke, K. Kontomaris, T. J. Hanratty, and J. B. McLaughlin, Turbulent deposition and trapping of aerosols at a wall, Phys. Fluids A 4, 825 1992 . 12 J. W. Brooke, T. J. Hanratty, and J. B. McLaughlin, Free-ight mixing and deposition of aerosols Phys. Fluids 6, 3404 1994 . 13 M. Chen, K. Kontomaris, and J. B. McLaughlin, Dispersion, growth, and deposition of coalescing aerosols in a direct numerical simulation of turbulent channel ow, Gas-Particle Flows, ASME-FED, 1995, Vol. 228, p. 27. 14 C. B. Rogers and J. K. Eaton The behavior of solid particles in a vertical turbulent boundary layer in air, Int. J. Multiphase Flow 16, 819 1990 . 15 L. M. Liljegren, The effect of a mean uid velocity gradient on the streamwise velocity variance of a particle suspended in a turbulent ow, Int. J. Multiphase Flow 19, 471 1993 . 16 M. W. Reeks, On the constitutive relations for dispersed particles in nonuniform ows, I: Dispersion in a simple shear ow, Phys. Fluids A 5, 750 1993 . 17 I. Kataoka and A. Serizawa, Basic equations of turbulence in gas-liquid two-phase ow, Int. J. Multiphase Flow 15, 843 1989 . 18 T. W. Abou-Arab and M. C. Roco, Solid phase contribution in the twophase turbulence kinetic energy equation, J. Fluid Eng. 112, 351 1990 . 19 M. W. Reeks, On the continuum equations for dispersed particles in nonuniform ows, Phys. Fluids A 4, 1290 1992 20 G.-J. Hwang and H. H. Shen, Fluctuation energy equations for turbulent uid-solid ows, Int. J. Multiphase Flow 19, 887 1993 . 21 M. Germano, U. Piomelli, P. Moin, and W. H. Cabot, A dynamic subgrid-scale eddy viscosity model, Phys. Fluids A 3, 1760 1991 . 22 K. D. Squires and U. Piomelli, Dynamic modeling of rotating turbulence, Turbulent Shear Flows 9, edited by F. Durst, N. Kasagi, B. E. Launder, F. W. Schmidt, and J. H. Whitelaw Springer-Verlag, Berlin, 1995 , p. 71. 23 M. Germano, Turbulence: The ltering approach, J. Fluid Mech. 238, 325 1992 . 24 C. Meneveau, T. S. Lund, and W. Cabot, A Lagrangian dynamic subgrid-scale model of turbulence, Proceedings of the 1994 Summer Program, Center for Turbulence Research, Stanford University, 1994, p. 271. 25 J. Kim and P. Moin, Application of a fractional-step method to incompressible NavierStokes equations, J. Comput. Phys. 59, 308 1985 . 26 J. B. Perot, An analysis of the fractional step method, J. Comput. Phys. 108, 51 1993 . 27 X. Wu, K. D. Squires, and Q. Wang, On extension of the fractional step method to general curvilinear coordinate systems, Num. Heat Transfer, Part B: Fundam. 27, 175 1995 . 28 U. Piomelli, High Reynolds number calculations using the dynamic subgrid stress model, Phys. Fluids A 5, 1484 1993 . 29 Q. Wang and K. D. Squires, Large eddy simulation of particle deposition in a vertical turbulent channel ow, to appear Int. J. Multiphase Flow. 30 R. Clift, J. R. Grace, and M. E. Weber, Bubbles, Drops and Particles Academic Press, New York, 1978 . 31 Q. Wang, K. D. Squires, and X. Wu, Lagrangian statistics in turbulent channel ow, Atmos. Env. 29, 2417 1995 . 32 U. Schumann, Subgrid length-scales for large-eddy simulation of stratied turbulence, Theor. Comput. Fluid Dyn. 2, 279 1991 . 33 L. M. Liljegren, The inuence of a mean uid velocity gradient on the particle-uid velocity covariance, Int. J. Multiphase Flow 20, 969 1994 . 34 J. K. Eaton and J. R. Fessler, Preferential concentration of particles by turbulence, Int. J. Multiphase Flow 20, 169 1994 . 35 M. Sommerfeld, Modelling of particle-wall collisions in conned gasparticle ows, Int. J. Multiphase Flow 18, 905 1992 .

Phys. Fluids, Vol. 8, No. 5, May 1996

Q. Wang and K. D. Squires

1223

Downloaded19Jul2004to195.83.232.39.RedistributionsubjecttoAIPlicenseorcopyright,seehttp://pof.aip.org/pof/copyright.jsp