Vous êtes sur la page 1sur 24

On the permeability approach

Chahid K. Ghaddar

of unidirectional

fibrous media: A parallel computational

Computer Aided Process Engineering, Inc., Suite 220, 888 WorcesterStreet, Wellesley,Massachusetts02181 (Received 6 April 1995; accepted 7 July 1995) The problems of viscous and inertial flows through unidirectional fibrous porous media are addressed using an entirely parallel computational approach. The pertinent partial differential equations, derived from homogenization theory, are solved by a parallel finite element method in conjunction with Monte Carlo techniques to predict the statistical permeability coefficient. A nip-element method, which mitigates the frequent geometry-induced numerical difficulties, while providing both accurate approximations for the permeability coefficient, and rigorous error estimates, is also presented. The seepage permeability coefficient is determined for a wide range of fiber concentration. It is shown to deviate markedly at low porosities from the behavior predicted by earlier cell models, while exhibiting generally good agreement at high, and moderate porosities with the cell models, and with the limited available experimental and analytical results. Limited but illustrative inertial flow results at moderate Reynolds numbers are also presented for both regular and random arrays. For regular arrays, the flow is found to be unsteady for Reynolds numbers greater than approximately 150 at which traveling waves characterized by distinct periods and amplitudes are observed. Some modest discrepancy is found in comparison with available data which is attributed to the unsteady effects and other numerical issues. For random arrays, several configuration permeability values are calculated and compared satisfactorily against the Ergun correlation. 0 I995 American Institute of Physics.

1. INTRODUCTION Transport phenomena in fibrous porous media play a critical role in many man-made and natural processes. Examples include: aerosols filtration, fluidization3 convecY2 tive heat exchangers,4 and cell mechanics5P6 just to name a few. Needless to say, a detailed microscopic analysis of the phenomena is plagued by many difficulties such as the statistical nature of the microstructure, the presence of a large range of disparate length scales, and nonlinearity among many others issues. From an engineering view point, however, it usually suffices to predict the macroscopic permeable character of the media rather than the intrinsically complicated details of the phenomena. In this paper we consider the permeability behavior of one class of fibrous porous media in which the flow is strictly transverse to the axes of the fibers. Such media which can be accurateIy modeled by two-dimensional (2D) random arrays of cylinders, appear to have accumulated fewer dedicated efforts to understand their permeable character as opposed to random beds of spheres for example. In general, the efforts (reviewed below) have either involved artificial assumptions to simplify the problem to the extent an analytical solution was possible, or have lacked the flexibility to perform a large-scale numerical studies and address nonlinear effects. As a result, rather limited theoretical results are available and are typically short on statistical analysis. The permeability behavior of 2D random arrays of cylinders is addressed in this work using an entirely parallel computational approach, based on first principles and coupled with proper statistical analysis. In particular, the present strategy departs from earlier efforts in several aspects: First, by exploiting the well-known benefits associated
Phys. Fluids 7 (il), November 1995

with parallel processing, it is possible to perform better statistical treatment and address more ambitious problems while keeping the intensive computational requirements within reason. Second, me geometry-induced numerical difficulties which are typical problems facing numerical treatment, have been treated, rather than avoided, with a nip-element method which enjoys rigorous approximation properties. Third, nonlinear effects due to moderate Reynolds-number flows which are not amenable to analytical analysis have been considered. In the following we briefly review previous studies for both creeping and inertial flows and then describe the organization of the current paper.

A. Creeping flow regime The permeability in the viscous flow regime, is governed by the well-established7b Darcy law s

where (u) is the superficial fluid velocity in the direction of the imposed external pressure drop Ap across the thickness L of the medium, K is the permeability coefficient (in general tensorial) which depends only on the bore geometry for sufficiently small Reynolds numbers, and ,u is the fluid viscosity. Darcy equation has been the point of departure for s many earlier capillary models aimed at correlating the pressure drop-flow rate relation, of which the well-known Carman-Kozeny equation has often been used. The latter correlation - (referred to hereafter simply as the Kozeny equation) in its simplest form provides the following expression for K:
0 1995 American Institute of Physics 2563

1070-6631/95/7(11)/2563/24/$6.00

Downloaded 29 Dec 2010 to 210.212.58.168. Redistribution subject to AIP license or copyright; see http://pof.aip.org/about/rights_and_permissions

K=

-, k

(2)

where k is the Kozeny constant supposedly independent of the porosity E, and rh is the mean hydraulic radius defined as the ratio of free volume to the wetted area, which is, for a fibrous medium of fiber diameter d, equal to (3) There is a large volume of data for beds consisting of a variety of spherical and nonspherical particles which indicate that a value of k-5.0 independent of shape and porosity in the range 0.26<~sO.8 yields a good agreement of the The same value of k has also been found Kozeny model. t2*13 satisfactory for three-dimensional fibrous mediat4-I6 in the porosity range 0.6~ ~~0.9. (The author is not aware of any experimental data on fibrous media for porosities less than 0.6.) Theoretical attempts to predict the Kozeny constant (or the permeability behavior) are, in general, plagued by the essential difficulty of characterizing the microstructure. This has Led mainly to two complementary approaches: the method of bounds; the structure-dependent analyses. In the former method, variational principles are used to construct upper and lower bounds for the permeability coefficient using limited microstructure information, like the component concentration. The bounds constructed by this approach are generally difficult to evaluate, or not tight for a precise estimate. (See Torquator7 for a comprehensive review on these techniques.) In the structure-dependent analyses, certain assumptions are made concerning the microstructure. Early efforts include the simplified cell models of Happel, Kuwabara,tg and * Spielman and Goren. The models consider an isolated fiber enclosed within an imaginary cell with the assumption that the influence of the surrounding solids upon the isolated fiber can be lumped into an appropriate boundary condition on the cell boundaries,18~or into a simple damping term added to g For the Stokes equations.20,21 example, in the models of Happel and Kuwabara the cell is chosen, for mathematical convenience, to be a concentric circle enclosing the tiber, with the assumption of, respectively, no shear or no vorticity at the outer boundary. For flow perpendicular to the fiber, the model of Happel predicts a Kozeny constant of k-5 in the porosity range 0.26~ ~~0.7, while the Kuwabara model predicts a higher value of approximately 7. More recent approaches consider supercells with multiple fibers and periodic boundary conditions. Homogenization theory, by means of multiple scale analysis, allows for the rigorous deduction of the supercell flow equations. Analytical solutions have been possible for regular (square or hexagonal) periodic arrays. Semi-analytical solutions 3 based on superposition of fundamental solutions followed by numerical treatment of truncated sums have been successful in treating random supercells with limited statistical averages. Cruz and PateraZ4presented a parallel computational strategy combined with proper statistics to predict the effective properties of multicomponent media and obtained lim2564 Phys. Fluids, Vol. 7, No. 11, November 1995

ited creeping-flow permeability results. The numerical treatment however, was frequently plagued by the induced geometrical stiffness due to the random supercell configurations. B. Inertial flow regime There are relatively few theoretical results for inertial porous-media flows, except for a few studies of structured arrays. Eidsath et al3 and Edwards et aZ.26performed numerical calculations for various regular arrays. Dybbs and EdwardsZ7 performed flow visualization studies through media consisted of Plexiglas spheres in hexagonal packing, and Plexiglas rods arranged in a random three dimensional geometry. Their study shed light on the details of the flow regime as a function of Reynolds number, and on the onset of unsteadiness. The author is not aware of any published theoretical studies on inertial flows through random arrays. However, many empirical correlations describing the flow-rate/ pressure-drop relation which are typically presented in the form V>=a(z4)+b(u)m, (4)

extend classical Darcy law by introducing a second nonlins ear term to account for inertial losses. The coefficients a and b are constants that depend on the porosity and pore characteristics, and m is typically 2. Many attempts have been made to accurately define the parameters in (4) to fit the body of experimental data as compiled by Scheideggerg and Davidson et a2.The most widely accepted relation for practical engineering applications is the Ergun relation2* given in terms of a modified friction factor, f , and modified Reynolds number, Re I as follows: 150 f =Re + 1.75, where (5)

flVP
Re =Re

&,
1 l--E

(6) (7)

and Vp is a nondimensional pressure drop defined as

v,= .A!zT
C. Organization

L/d p(u)

(8)

where p is the fluid density. of paper

This paper is organized into four sections as follows: in the next section we present the mathematical formulations of both the Stokes and inertial flow problems which are founded upon the periodic supercell model and homogenization theory. A nip-element method designed to mitigate the subsequent numerical treatment is also presented. In Section III we discuss the pa.ra.llelnumerical treatment of the formulations. Results are presented in Section IV which inChahid K. Ghaddar

Downloaded 29 Dec 2010 to 210.212.58.168. Redistribution subject to AIP license or copyright; see http://pof.aip.org/about/rights_and_permissions

dUi -z=O, .i

in a,

Yl

where R is the region of the supercell occupied by the fluid, and Sij is the Kronecker delta. The velocity must vanish on all obstacles boundaries and both the velocity and pressure must be X-periodic in both y t and y2. It is further required for uniqueness that Japdy= 0. It is important to note that although the Stokes equations do not permit a solution for flow, past an isolated cylinder in an infinite domam3t our supercell, however, thanks to the finite domain and periodic boundary conditions, constitutes a well-posed problem except possibly in the limit of diminishing concentration values which are not the focus of this work. It is advantageous to work with the variational formulation of (9)-(10) since it is required for the ensuing finite element analysis. This is based on the well-known variational principles2 u=arg m=hEzJfdv>, (11) (related to the excess dissipa-

l33. 1. Periodic supercell [0,X] containing N cylindrical inclusions of unity diameter d, with centers located at yi , i = l,.. . ,N.

where the functional Jo(v) tion) takes the form

(12) eludes three subsections: In Section IV A we present two demonstrative examples using the nip-element method which also serve as a benchmark for the implementation. In Section IV B we present creeping flow permeability results and compare against the predictions of earlier cell models as well as available experimental measurements and analytical results. Finally, in Section IV C, we present inertial porous-media flow results for regular and random arrays for selected concentrations and Reynolds numbers along with a comparison with the limited available data. II. PROBLEM FORMULATION We consider a fluid-fiber structure composed of a periodically replicated supercell of size X, y E [ O,h] X [0,X] ClR2, consisting of two phases: an incompressible fluid phase of constant viscosity p= 1, and density p; a dispersed phase of N immovable co-oriented cylindrical obstacles of diameter d = 1, centers yi E {Y}~= (yi , . . . , ylv), and volumetric concentration, c= 1 - E, as depicted in Fig. 1. A fixed imposed global pressure gradient, Vp = - 1 jt , [where it is a unit vector in the direction of the y, axis (see Fig. l)], is responsible for a net transverse fluid motion parallel to the y1 axis. [Note the choice (d= l,@= l,Vp= - 1 y^i) renders p the only variable parameter which may be, equally well, interpreted as a nondimensional control parameter.] A. Viscous flows The fluid motion governing equations in the supercell are deduced from periodic homogenization theory.2990 Here the velocity vector u(y)=(u1(yi,y2),u2(y1,y2)) and the pressure, p = p (y), satisfy the steady Stokes equations, -p+dp=*,i, Here Z={(ui .u2). E (H~,(fl),H~,(R))~diuv=O}, and H,&(n) is the space of all X-doubly periodic functions which vanish on 65l, and for which both the function and derivative are square-integrable over 0. In order to arrive at the standard velocity-pressure weak form, the constrained maximization (11) is transformed into an unconstrained saddle problem by enlarging the velocity include all functions space to (UlPU2) in and introducing a Lagrange multiplier W&#W>~~,W)), - the pressure, p - to impose the incompressibility constraint. Taking the tirst variation of the resulting Lagrangian, weak obtain the form: Find we CU1442,P)
E (H~#(n>,H~,(n>,L~,o(~)) such that

dy-

I,$pdy=

lau,dy,

(13)
V(Ul a21 E e&w2,

q - dy=O, I Cl dYi

dUj

vq -%,oWL

where Lip(a) is the space of all h-doubly periodic functions q(y) which are square-integrable over 0 (note that candidate pressures need not be continuous), and for which Sdy= 0. transverse conjiguration permeability, The K= ~({y}~,c,h), is defined through Darcy relation (1) and s takes the form

1 K (Ul)= T;Znu,dy. I
By equating the first variation of the functional (12) to zero and using (1 l), K can be expressed as
Chahid K. Ghaddar

d2Ui djjdyj

$Vi

in a,

for i=1,2,

(9)

Phys. Fluids, Vol. 7, No. 11, November 1995

2565

Downloaded 29 Dec 2010 to 210.212.58.168. Redistribution subject to AIP license or copyright; see http://pof.aip.org/about/rights_and_permissions

1
K=-T A-

(16)

which is related to the Helmholtz stationary dissipation principle.33-35 It is worthy to note that although we consider the isotropic permeability, the extremizing property (16) extends to the full permeability tensor following the development in Nir et a1.36 The permeability, of course, depends on the particular inclusions centers {y}N. By presuming the latter are random variables, {Y}N, prescribed according to a joint probability density function (JPDF), ftyJN( { Y}~, c , A), renders the permeability a random variable whose average, K,(c,A) (independent of {YIN) is then given by

K({Y}N,C,h)f~y)N({Y}N,c,A)dY. K,(C,A) =I (Co,AlxCoAIP

(17)

The effective permeability, K, , is representative of a random periodic fibrous media whose microstructure can be adequately predicted by f{vI,. Nevertheless, it is plausible to believe that as X--+00, the periodic media will resemble a truly random media (provided the macroscale L%-A) and, hence, their ,corresponding statistical properties are applicable to each other. More precisely, the limiting process limb,, K,(C,A), or realistically, K,(c, A> A) for some h(c), assumes a well-defined value which, for all practical purposes, representative of true random media. This has been demonstrated numerically for an analogous problem of determining the effective conductivity in thermal composites24 where A(c) is defined such that for XaR(c), (i) K,(c,A) no longer changes appreciably, and (ii) the standard deviation of the random variable K=K({Y}~,c,A), uK(c,A), is sufficiently small. In order to compute (17) one needs to know the inclusion JPDFffvlN which is typically unknown. Here the inclusion JPDF is assumed to be isotropic and homogeneous based on a random acceptance-rejection sequential addition process738 which is believed to be relevant to many 24 physical systems. [Note that a particular configuration, in general, may be anisotropic, but, on the average, the properties are expected to be isotropic (scalar).] B. Inertial flows In the event of important convective inertial effects, the fluid motion in the periodic supercell is governed by the (possibly unsteady) Navier-Stokes equations, p[~+Uj~]-~+~=fili, in f1, for i=l,:;8)

rigorously derived using periodic homogenization theory as has been possible for the Stokes problem. Here the nonlinear convective term precludes complete decoupling between the tensorial components of the homogenized, and supercell subproblems [save for the special case of weak inertial effects3 (~4 1 )] which poses a series of difficulties for the subsequent solution. The adopted model, widely accepted in the literature,25V26 justified by a control-volume type arguis ment in which the obvious contribution of the homogenized problem to the supercell problem is the linearized pressure appearing locally as p = - I yt . Furthermore, the interest here is not in the details of the tlow field of the macroscale problem but rather the bulk permeability relating the pressure drop to the how rate, which is precisely what we calculate from the supercell problem The latter should be, due to spatial averaging, less sensitive to the flow field details. The weak form of equations (18) and (19) takes the following form. Find (ut ,uZ,p) E (H~#(n),H~,(n),~~,(n)). such that

tidYi

di pdy=

nu ,dy,

v(~,

,u2j E (H&W>)2T

(20)
-

IL aYi

dUi q dy=O,

vq E L;,,(sz).

Turning to the permeability expression, we adopt the basic definition provided by Darcy equation (1) which ins volves no modeling assumptions. The transverse configuration permeability, equation (15), remains valid, although temporal averaging is required in unsteady situations,
K=;

o*{U,(t))dt= J

;~o j-~U~(t)dydt.

(22)

where T is the period of the unsteady flow. The permeability, K, however, will also depend on the flow Reynolds number,
Re,

(23) where the last equality follows from the choice (d= I,pu= 1), and equation (15). Last, we show that the permeability can only decrease for nonzero Reynolds numbers. In fact, we have from (20) P
(19)

subject to the no-slip condition at the obstacles and the A-periodicity of both the velocity and both yt and y2 directions. It should be pointed supercell equations are actually rationalized,

boundaries, pressure in out that the rather than

E uldy, R J
where we have used the incompressibility condition (19). Due to A-periodicity of the velocity, the convective term, SnUiUj (duildyj) dy, vanishes. Indeed
Chahid K. Ghaddar

2566

Phys. Fluids, Vol. 7, No. 11, November 1995

Downloaded 29 Dec 2010 to 210.212.58.168. Redistribution subject to AIP license or copyright; see http://pof.aip.org/about/rights_and_permissions

P)
FIG. 2. Porous-media nip-element strategies : (aj nip-region enlargement (upper bound); (b) nip-region blockage (lower bound).

FIG. 3. The excised nip region, 22 Upper bound (left); lower bound (right).

uillj-

azci

aYj

dy=i

1
s
an

UjU~ .njdS=

0 ,

05) where ds is a differential element of &I defined by the outer unit normal n. Here we again use (19) followed by application of the divergence theorem. Using (25), averaging (24) over one period, T, and recognizing that the period average ldt) dy vanishes, we obtain of Sn$(duf
u*dydt.

(26)

From (22) and (26) we conclude that (27) which is identical to the temporally-averaged (16) despite the presence of convective contributions. Now, given that u is not the Stokes solution as defined by (13) and (14), and hence the maximizing property (16) is lost, it follows that Jo(u) and hence, K can only decrease relative to the Stokes permeability. C. A nip-element method

shown in Fig. 2(b), preventing flow across the gap. These models will not change the weak variational statement of the original problem as given by equations (13) and (14), except for replacing the original domain s1 by the new modified domain, defined as g=n\@for fi, the lower bound, and as c= $1 U L8 for the upper bound, where 58is the excised nip region as defined by Fig. 3 for the two different nips. The nip region is defined by two geometrical parameters: the interparticle spacing, a; the nip height, /?, above the center line connecting the two forming inclusions. Thus, any two inclusions with a center-to-center distance less than 1 + (Y, are considered forming a nip, where LY, is a specified critical value. The proofs for the claimed hierarchy of the bounds, KLB< KC KUB, are based on variational embedding arguments, in which a maximizing solution on a smaller domain is extended (by zero, due to the no slip boundary condition) to construct an admissible but non-maximizing candidate on a larger domain. To translate this statement mathematically let un be the correct Stokes solution for the bare (no nip) problem defined over the domain 61; uq be the correct Stokes solution for the nip-augmented problem defined over 5Yand Olg be the zero field extending over the region .LV. For the upper-bound permeability, K~, we can write [using equations (11) and (16)]:
K-A

Jn(Ua)

The computation of the configuration permeability is frequently plagued by the presence of excessively close pairs of inclusions forming very small gaps-nip regions. This, at best, results in stringent mesh generation, ill-conditioned systems and poor load balancing. At worst, numerical treatment may simply not be possible and the particular realizations must be discarded. The nip-element method is a computational artifice proposed in order to significantly mitigate these difficulties at the expense of introducing small bounding errors. The method is discussed below. 1. viscolls flows The (Stokes) permeability, K, enjoys the variational principle (16) which facilitates rigorous construction of nipregion-based lower and upper bounds. To generate an upper bound, K[JB , we simply enlarge the nip region as shown in Fig. 2(a), thus allowing for a larger flux of fluid across the gap than would actually pass. To obtain a lower bound, KLB, we simply replace the nip region with a blockage as
Phys. Fluids, Vol. 7, No. 11, November 1995

1 =pn(un>+Jdo)

=KUBT

08)

where the last inequality follows because the constructed field un U 018, although admissible, is not the correct Stokes solution over the domain 8 and thus will not maximize the functional Jdv) . The proof for the lower-bound permeability, KI.B , is analogous and is presented in Ghaddar.39 We remark that sharper bounds can also be constructed in an analogous approach to the effective conductivity problem presented in Cruz, Ghaddar and Patera. For example, to obtain a sharper lower bound, we can assume a one dimensional Poiseuille flow in the nip gap region in which the
Chahid K, Ghaddar
2567

Downloaded 29 Dec 2010 to 210.212.58.168. Redistribution subject to AIP license or copyright; see http://pof.aip.org/about/rights_and_permissions

candidate velocity profiles are reduced to parallel parabolic ones. .The lower bound will obtain physically because the model implies an infinite viscosity in the direction normal to the flow, and mathematically because of the additional restriction on the functional space in which we search for a solution. The mathematical treatment is, of course, more complicated with the presence of the Lagrange multiplier, the pressure. The Poiseuille flow assumption, in fact, is very close to reality in the nip regions. Therefore, we expect our crude models to be rather sharp. This is because of the second order dependence of the velocity, and hence, the permeability on the inter-particle spacing. This has been observed numerically and presented in Sec. IV A. 2. Extension to inertial flows original domain

Unfortunately, the Stokes bounds are no longer rigorous for the Navier-Stokes problem due to the loss of the functional-maximizing property of the permeability, equation (16). However, it is expected that, the bounds will still work since the tlow is expected to be effectively inertia-free in the narrow. nip regions even when the global Reynolds number may be large. Sensitivity derivatives, could be used (e.g., with respect to say, the interparticle spacing) to assess the merits of this naive approach. III. PARALLEL, NUMERICAL TREATMENT The numerical solution of the supercell problem encompasses four main tasks: domain partitioning; mesh generation; finite element solution; parallel implementation. These tasks are discussed below. A. Domain partitioning A tensor-product partitioning strategy is employed to divide the original supercell domain into P smaller nonoverlapping Subdomains or subcells, where P is the target number of processors. The procedure, presented in detail in Ghaddar,39 is accomplished by first, overlaying an initial structured grid upon the square supercell domain. The grid lines are then individually repositioned to eliminate any existing geometrical degeneracies. Next, the necessary geometrical and topological information associated with the individual subcells (identified with the grid bricks) are obtained. Last, the bare supercell is augmented with the desired type of nip elements, and the required modifications are performed. The tensor-product partitioning procedure is illustrated in Fig. 4 for an original realization containing 50 inclusions at a concentration of 0.5, which is partitioned into P= 16 subcells. Generally, the resulting subcells will comprise smaller disconnected domains. This fact, however, will not sacrifice the structured topology associated with the initial grid which has a significant impact on reducing the communication overhead in the parallel computing environment and hence. boosting the performance. B. Mesh generation Mesh generation is performed in parallel and is based on triangular finite elements which provide greater geom&ric
2568

nip-augmented partitioned domain


FIG. 4. Example of a tensor-productpartition of a realization containing 50 inclusions at 0.5 concentrationinto lower-bound nip-augmented16 subcells.

flexibility for decomposition of complex domains as opposed to quadrilateral elements. In this work we employed a Voronoi-based triangulator, MSHPTG,4* for its demonstrated robustness, minimal input information, and satisfactory result.24 MSHPTG requires on input a boundary grid and the corresponding topological description. It honors the boundary grid density in the generation of interior elements, which makes it suitable for achieving a desired mesh density distribution in a fully automated algorithm by simply controlling the boundary grid. The main problem is then reduced to how to discretize the boundary curves. The boundary resolution which is propagated smoothly to the interior, must therefore respect regions of potentially high gradients if discretization errors are not to dominate or even destroy the numerical solution. These high gradients are expected to occur at regions of clustered inclusions. The procedure used for gridding the boundaries is presented in detail in Ghaddar.39 It makes use of the so-called minimal distance as a guide in the node distribution
Chahid K. Ghaddar

Phys. Fluids, Vol. 7, No. 11, November 1995

Downloaded 29 Dec 2010 to 210.212.58.168. Redistribution subject to AIP license or copyright; see http://pof.aip.org/about/rights_and_permissions

simplicity, parallel implications associated with the iterative solution algorithms are deferred until Sec. Ill D. 1. The Stokes problem
Discretization. The space-restricted statement of the weak variational equations (13) and (14) takes the following form: Finduh E [V,(fL)] andph E P,(a) suchthat

uiUh,V)-b(Ph,V)=(cSil,V), ~v4V,W)12,
-biq,ufJ=O, where a(.,.) vqe Phi n),

(31) (32)

is the standard continuous Laplacian operator,

dub ,v)= I n V.vV-u,dCl;


b (. , .) is the bilinear form,
FIG. 5. Finite element mesh generated in parallel for a realization of 17 inclusions at 0.55 concentration,partitioned into four subcells. Here 11 nips are present.The mesh is composed of 4243 triangles.

iw

b(q,v) =

sn

qV -vdR;

(34)

(.,.) is the inner product,


(q.u) = /gpvdO:

process. The minimal distance between a point y and a surface S, denoted as d&y,S) is defined by dHi9.S)=m4y-xl,
XES

(35)

and the subspaces V,, and Ph are defined as

(29)

where [y-xl is the usual Euclidean distance between two points. This distance is used in conjunction with the following distribution functior? for meshing of arc and line segments making up the boundaries of the subcells, 1 h(y) =
mldH(y,S) + l/h,,,

I -1
.r

(30)

where h is the mesh spacing at point y; h,,, is a nominal mesh spacing parameter which is defaulted to, for sufficiently large dH ; nz is a refinement parameter that controls the near-cylinder density; and r is a global control refinement parameter typically set to unity. In effect equation (30) honors the relative closeness of neighboring inclusions when distributing the nodes along the inclusion surface. The resolution will be denser in clustered regions and lighter in wellconditioned regions. Figure 5 shows the finite element mesh for a realization of 17 inclusions at 0.55 concentration partitioned into four subcells. C. Finite element treatment There are two basic ingredients in the finite element analysis: discretization of the weak variational form; solution of the resulting discrete system of equations. For the first task we use isoparametric second order triangular finite elements for accurate boundary representation and greater meshing flexibility. For the second task, we adopt iterative solution algorithms since they require both dramatically less memory and significantly fewer operations than direct methods, and they are more readily and efficiently parallelized. In the following we discuss the Stokes and the NavierStokes problems, respectively. We note that, for purpose of
Phys. Fluids, Vol. 7, No. 11, November 1995

where lok reads: restricted to finite element k. The choice of the approximation spaces for both the velocity and pressure is made such that the inf-sup condition4 is satisfied. In particular, the choice is designed on the use of the Taylor-Hood 43 PI-P1 quadratic and linear triangular elements pair that are known to honor the inf-sup condition. Proceeding with the discretization we arrive at the following linear system:

[Q,

-It*

I;](

;;}=(By,

(38)

where A is the discrete Laplacian operator,

A = ij

(39)

Dr=D,,i is the derivative operator corresponding to the bilinear form b, Dmni z ji,,~ don;

i40)

- I

where 4, and (6, are the nodal basis associated with the linear and quadratic meshes, respectively; B = Bij is the mass matrix defined by (41) and (1) is a vector of unity entries. It should be pointed out that the system matrix in (38) as constructed, does not incorChahid K. Ghaddar 2569

Downloaded 29 Dec 2010 to 210.212.58.168. Redistribution subject to AIP license or copyright; see http://pof.aip.org/about/rights_and_permissions

porate the associated boundary conditions of X-periodicity, no slip inclusions walls and the zero average of the pressure vector. In the context of iterative solver, these boundary conditions are imposed as additional steps to the iterative algorithm (described below). The dissipation-related expression for Kh corresponding to (16) is given by 1 Kh=--T x(42)

the old values. The uniqueness condition for the pressure is effected by imposing the zero average requirement on ph. Finally we remark that no special treatment is required in the presence of nip elements other than enforcing the no slip condition for the velocity along the nip edges. 2. The Navier-Stokes problem Discretization. The spatial discretization of the variational, unsteady, incompressible, Navier-Stokes equations (20) and (21). takes the form (for clarity we drop the subscript h)
-+C(U)Ui - DiUi=O ,

Due to the functional dependence of Kh on the Laplacian operator A, Kh iS expected to inherit the convergence properties of A (see Wang and Fix44), that is, third order convergence with second order finite elements. This is verified in Section IV Al. Solution strategy. The solution is based on the iterative Uzawa algorithm,45~46which decouples the pressure and velocity (via application of block Gaussian elimination to the system matrix) into two equations, 2 sphEtz, -D,IA- BaiI 9 (43) (44)

=-Aui+DTp+BSil,

(47)

for i= 1,2, where C(u) = C,(U) is the skew-symmetric discrete convection operator: (48) Temporal discretization is carried out by treating the viscous term implicitly using first order backward Euler integration scheme, and the nonlinear convection term explicitly by the third order Adams-Bashforth. This leads to the following semi-implicit time-stepping scheme:

AUi,h= D,Tph+ BSi,,

where S is defined by 2
S=C

i=l

DiA- D

I .

(45)

Equation (43) is first solved for the pressure by a nested conjugate gradient iteration48,49 followed by two elliptic solutions for the velocity components. (Unfortunately, no preconditioning was used.) The A-periodicity of the velocity and pressure is discretely imposed on the iterate by identifying appropriate pairs of nodes as the same node along periodic edges,50.51 that is, summing the contributions and writing the sum over

mte. B u;+ -D;pfpn+ A+ At


n

=&Buy-pi

q=fJ

a,C(U-q)u~-q+s~i*I

(49)

-Di,;+l=O,

(5oj

where At is the time step at iteration ~1. The coefficients LYE defined by an algebraic functions of the three last time are steps,

~2At-1(Atn-1+Atn-2)+6Atn(2Atn-*fAt-2)+4Atn2

Atn-1(Atn- +Atn-2)

(51)

*=-E

1 6At(At- +At-*)+4Atn2 I At-1&-2 6APAt- +4Atn2 (At,,- I+ Atn-2)AtTL-2

1 9

(52)

63)

For the case of a constant At, the coefficients aq reduce to the classical constants documented with third order AdamsBashforth, namely LYE s , a1 = 2 and ~ys= 2. = The merit of this temporal scheme mixing is twofold: first by treating the nonlinear term explicitly, the equations become Iinear at every time step which would significantly simplify the solution; and second, the implicit treatment of the diffusion operator alleviates the stringent time step re2570

quirement associated with the stability of the diffusion operator. The time step is then governed solely by the maximum eigenvalue of the convection operator which must lie inside the stability region of the explicit scheme to honor the well-known Courant condition.39 A value of 0.5 for the latter has proved a stable limit for the moderate range of Reynolds numbers considered. The temporally-averaged numerical permeability follows from the dissipation related expression, (42):

(54) or, equivalently, from the Darcy definition: s


Chahid K. Ghaddar

Phys. Fluids, Vol. 7, No. 11, November 1995

Downloaded 29 Dec 2010 to 210.212.58.168. Redistribution subject to AIP license or copyright; see http://pof.aip.org/about/rights_and_permissions

1?T Khoc I

Buldt,

(59

where T is the period of flow oscillations. The sum in (55) is over the entire length of the global vector Bul , where B is the system mass matrix [equation (41)]. The two expressions should be equivalent which provides a useful check for the implementation. The temporal averaging in unsteady situations is effected by averaging the permeability over a full time period. Solution strategy. One possible solution strategy would be to apply the Uzawa algorithm to (49)-(50) at each time step. While this is an accurate method, it suffers from the very ill-conditioned pressure-solve operator; in addition, the expensive nested Helmholtz inverse solution makes this approach very inefficient. Instead a recent operator splitting method is used which breaks the original system into a sequence of three decoupled subproblems integrated in a new intermediate variable5253 as follows: 2 fw+ =
D;p-l-&, Buy-px q=o

&b.d vertex

a,C(un-q)u;-q+

Bail,

0 (56)

.. .. ,oo::

E(P n+l_p)=-&

D&f,

(57)

,&;+l+; Ill rri

&nB-lD;(p+ -p),

(58)
0 r I

intermediate where ti is the velocity field, H = [A + (p/A tj B] is the Helmholtz operator, and E is the consistent Poisson operator given by 2
E= C DiB- DT.

FIG. 6. Definitions of essential parallel constructs.

I=1

(59)

It is important to note that this first order split formulation incurs additional temporal errors of order At. However, if steady time-independent solutions are reached, these errors vanish completely. Furthermore, this splitting scheme is performed on the well-posed discrete original equations based upon a consistent approximation spaces and thus requires no artificial or special treatment of boundary conditions, which plagues the more classical fractional splitting method based on the continuous original equations.34 The first and third steps are solved by the standard conjugate gradient iterations. On the other hand, the pressure solution requires a nested conjugate gradient iteration similar to the Uzawa used for the Stokes problem. The mass matrix, however, is very well-conditioned and when pre-conditioned by the diagonal lumped mass matrix, the number of inner conjugate gradient iterations is reduced to order unity, which significantly mitigates the otherwise serious drawback of the algorithm. To circumvent this inner solve completely we replace the mass matrix, B, by the diagonal lumped mass matrix ji defined elementally as %=~ace(BkjlfikI
Phys. Fluids, Vol. 7, No. 11, November 1995

where Iok] represents the area of element ank. Note that this substitution also eliminates the iterations requirement for the last step (58). The error incurred due to this approximation vanishes completely if a steady state time-independent solution is obtained, as is evident from equation (46). For the case of an unsteady solution, a small damping effect was observed when using the lumped mass matrix (see Section IV c 1). D. Parallel considerations In this section we describe briefly some of the aspects of the parallel implementation on a distributed-memory, MIMD, message-passing architecture, here the Intel iPSC/ 860 32-node hypercube. In particular, we do not intend to discuss the broader context of parallel processing,46,55,56 but rather to outline briefly the parallel constructs necessary for performing the communication-involved tasks, as well as the parallel performance measures of the implementation. 1. Parallel constructs The parallel constructs are best described by referring to the schematic shown in Fig. 6 in which a supercell is partitioned into, here, four subcells. The relevant information associated with each subcell is then loaded onto the assigned processor. This information includes the geometrical and toChahid K. Ghaddar 2571

(60)

Downloaded 29 Dec 2010 to 210.212.58.168. Redistribution subject to AIP license or copyright; see http://pof.aip.org/about/rights_and_permissions

pological descriptions which are needed for the subsequent meshing; pointers that identify common boundaries with other processors along with other associated information for the required exchange of data. The subcells are subsequently triangulated followed by a global checking to ensure the integrity of the entire supercell mesh. The data structures are based on the now standard technique of elemental evaluation of operators in which no assembling of the system operator is required.6*47 Instead, the quadratures are performed only to assemble elemental matrices. All the iterates are stored locally based on the individual grids of the subcells. For example, in Fig. 6, the solution vectors are split into four locally defined data structures. At the end of the solution stage, each subcell will hold the correct information based on its mesh, and hence the solution is completely known over the entire supercell. The back bone of the Uzawa and the splitting scheme is the iterative conjugate gradient algorithm. The computational complexity of the latter involves two main operations: matrix-vector product and vector-vector inner product. The latter is performed locally on each processor followed by a global log2P gather-scatter operation which sums from all the processors and redistributes the total sum back. The former is effected using the elemental evaluation technique by carrying out the multiplications locally on each processor. This is followed by a direct stiffness summation, DSS, procedure which sums the local values at the elemental local nodes associated with a global node and redistributes the total sum back to the elemental local nodes (see Fig. 6 for definitions of local and global nodes). By virtue of the partitioning procedure and also due to periodicity, certain global nodes will appear on more than one (and as many as four) processors. This will require in addition to the local DSS on each processor parallel communication (exchange of data) among processors either to attain the correct global values locally, or to impose periodic boundary conditions. It be-. comes critical for parallel efficiency to minimize the number of messages (sends/receives) and to maximize their lengths. For this purpose, boundary nodes are classified and grouped into either vertices if shared between more than two processors, or edges if shared between strictly two processors. The Parallel part of the DSS (in addition to the serial part performed locally on each processor), is then to sum the contributions from all local nodes associated with edges and vertices and redistribute the correct sum back to the contributing local nodes. (For details on the implementation see Ghaddar?9)

(62) the load balance, A,, , Ab= mqsp Ndof- fidof p +1= fid0-f maxpEP Ndof p fidof

63)

where NFf is the number of degrees of freedom, residing on processor p and Ndcf is the average number of degrees of freedom which is equal to the total number of degrees of 6-eedom divided by the number of processors. Note Ab& 1, and Ab= 1 is the ideal value. This definition leads directly to the parallel efficiency in the limit of no communication overhead, 1 rl,=z 7 (64 where ;ip refers to the no-communication-overhead case which is, by definition, larger than rip [relation (64) follows by recognizing that, in the limit of no communication and no nonlinear hardware effects, t,& P) scales with maxp,$$ and t,,[,( 1) scales with the total number of def grees of freedom]; and last, the overall MFLOPS which can be estimated by multiplying the typical MFLOPS capability of a processor by the number of processqrs and the parallel efficiency. These defined measures arc generally problemdependent, and may vary significantly from one case to another. For a typical calculation with O(50,OOO) total degrees of freedom and good load balance, an estimated Q(80) MFLOPS and a parallel efficiency of ~90% was achieved on 16 processors of the Intel i860.39 IV. RESULTS AND DISCUSSION A. Nip-element examples The nip-element method offers potentially significant computational benefits particularly in conjunction with iterative solvers. This is a consequence of the elimination of otherwise element-dense nip regions which is directly reflected in improving the problem conditioning, the reduction of degrees of freedom, and the improvement of load balancing. These combined effects result in considerable CPU savings relative to the original nip-free problem.40 Here we present two examples: the first one is intended to demonstrate the rigor of the hierarchy of the Stokes bounds for an interesting single-nip configuration, whereas the second is intended to test the method in the presence of important convective effects for a random realization. 1. The Stokes permeability bounds We begin by verifying the third order convergence of the permeability expected with the second-order isoparametric discretization. This also serves as a benchmark for the implementation as well as provides a guide for the expected discretization error as a function of mesh density. In Fig. 7 we plot the absolute error, E, = K~ - K,,,,~ versus the representative mesh spacing ff = h nomlr [see equation (30)] using both subparametric and isoparametric implementations. The tests were carried out for a hexagonal array at concentration
Chahid K. Ghaddar

2. Parallel performance Turning to the parallel performance, it is generally characterized by the following four interrelated quantities: the speed up, S, ,

Ldc( Sp=r 1) talc(P)


where tcalc(n) refers to the calculation time on p2processors; the parallel efficiency, rip, 2572
Phys. Fluids, Vol. 7, No. 11, November 1995

Downloaded 29 Dec 2010 to 210.212.58.168. Redistribution subject to AIP license or copyright; see http://pof.aip.org/about/rights_and_permissions

FIG. 9. Special realization containing two inclusions at 0.1 concentration, (h=3.963).

FIG. 7. Convergenceof the permeability with mesh spacing for a hexagonal array at concentration 0.6. The third order convergenceis observed with isoparametricfinite elements.

0.6. (shown in Fig. 8) for which ~,,,,~=0.001452 (Sangani and Acrivos ). The calculations were performed on four pro cessors of the Intel i860 hypercube and consumed 30 to 1000 seconds of computation time. (Total degrees of freedom ranged from 1400 to 12000.) We observe the third order convergence of the permeability with isoparametric elements. Note that the skin effects due to subparametric representation of the boundary obscure the third order convergence (see Strang and FLK,~ pp. 105-116 and pp. 192204).

to the flow direction. The motivation behind this example stems from the argument that the flow blockage strategy incurs counter effects of reduced drag which may obscure the expected lower bound permeability. This is because the blockage strategy reduces the surface area exposed to the fluid and eliminates full dissipative regions, particularly in this symmetric flow situation in which the nip region has a very minor flow-passing role. The rigor of the boundhierarchy proof is verified by the numerical results shown in Table I in which the following quantities are presented: the relative error E, (dependent on the exact solution K),

where

K*

is either
K),

~~~

or

~~~

the relative error E, (i&e-

Next we consider the realization shown in Fig. 9 which has N=2 cylindrical inclusions at a concentration of 0.1, in which the centerline connecting the two inclusions is parallel

pendent of

x 100, where
#T+ KUB+KLB

(66)

2.

the degrees-of-freedom per velocity component, Ndof u . Indeed, these counter effects only render the bounds very sharp. We remark that these results were obtained for the case CZ=0.075,/I= 0.1 (recall LYis the inter cylinder spacing; j3 is the nip height above the center line) using a relatively very fine grid and stringent incomplete iteration control. [The

TABLE I. Bound results for the Stokes permeability of the symmetric realization shown in Figure 9. K* or i?
K KLB

E, *.* 4.OE-5 1.5E-4 ...

-6 ... *.a I** 3.4E-5

Ndof u 13697 13214 13357 ***

FIG. 8. A particular finite element mesh comprising 1965 finite triangular elements for a hexagonal array at concentration0.6.
Phys. Fluids, Vol. 7, No. 11, November 1995

KUB E

0.7439011 0.7439008 0.7439019 0.7439013

Chahid K. Ghaddar

2573

Downloaded 29 Dec 2010 to 210.212.58.168. Redistribution subject to AIP license or copyright; see http://pof.aip.org/about/rights_and_permissions

PIG. 11. Speed contours for the Navier-Stokes calculation at Reynolds number Re = 68.7 in the realization of Fig. 10(a).

PIG. 10. Random realization containing 14 inclusions at 0.3 concentration (A = 6.054) used in the inertial nip-element example of Section IV A 2: (a) original geometry; (b) lower-bound geometry; (c) upper-boundgeometry.

discretization error is expected to be 0( 10m7) based on the convergence test of Fig. 7. This error is smaller than the reported bound gap.] 2. Inertial porous-media flow example

As discussed in Section II, the bound hierarchy is still expected to obtain at finite Reynolds numbers. This is demonstrated using a random realization containing 14 inclusions at a concentration of 0.3 shown in Fig. 10(a). Here, p, is chosen to be 10000, at which a moderate flow Reynolds number, Re, is expected to result. The geometries of the lower and upper bound calculations are shown respectively in Figs. 10(b) and 10(c) in which four nips are present. Table II presents the bound results for the permeability calculations. In addition to the quantities defined in Table I, is also presented: GS, defined as the ratio of the maximum edgelength of all finite elements present to the minimum edgelength of all finite elements present; CPU(p) the (wall clock) processing time in seconds on p processors; Ab the load imbalance defined in (63).

The hierarchy of the bounds is obtained despite the relatively important convective contributions, [ Re = 0( 701, see Table II]. (Of course, this does not rule out the possibility of counter examples.) Note that the nip gaps form percolation ialves; the flow is effectively inertia-free through these passages, even though the global Reynolds number is large. This is made evident in Fig. 11 which contours the speed over the nip-free domain. The clustered inclusions in the middle almost block the flow entirely, forming a stagnation region. Finally we remark that, given the relatively moderate geometrical stiffness of the original problem (see Table II), the savings for this particular realization are somewhat moderate. (We only obtain a factor of 3 reduction in CPU time for both the lower bound and upper bound calculations compared to the original problem.) More typically, for worse geometrical distortion situations, the conditioning of the consistent Poisson operator [equation (59)], associated with the pressure solve deteriorates rapidly, resulting in dramatic observed savings. 6. Creeping flow permeability results

TABLE II:Bound results for the inertial-flow permeability of the random array realization shown in Figure 10 at p= 10000. K* or i?
K KLB

E, ..+ 0.9 2.3 ...

,f?, 1.. ... ... 0.7

Re 68.7 68.1 70.2 69.2

Ndof, 15508 13743 13783 -*-

CPU(16)
872510 278324 245500 523824

GS
84.9 18.7 14.7 ...

Ala 2.1
1.4 1.4 ...

KC/B i

0.006873 0.006811 0.007029 0.006920

The supercell permeability, K,(c,X), is calculated at a particular h, X0%7, for 27 concentration values of c=O.O5,0.75 and cj=0.1+j0.02,j=0 ,..., 24. The choice of A is conjectured to be reasonably largea to reduce spatial averaging effects and hence, K~(c,&,) is expected to be a good .approximation to the true random media. The nip-element procedure has been used rather extensively to overcome the dramatic negative effects of the geometrical stiffness on the computation time.40 In order to minimize the bound gap error, however, a stringent cr,, c~~=O.05, has been defined. The worst obtained nondimensionaI relative bound gap error [see equation (66)] was for the highest concentration considered, c = 0.58, and amounted to 0.0520. Turning to the Monte Carlo components of the algorithm, for any particular concentration, c, we approximate the integral (17) as average of the lower and upper bound the sampIe means,
Chahid K. Ghaddar

2574

Phys. Fluids, Vol. 7, No. II, November 1995

Downloaded 29 Dec 2010 to 210.212.58.168. Redistribution subject to AIP license or copyright; see http://pof.aip.org/about/rights_and_permissions

10

.............. c.. ............. ::::::::!:::::::::::::::::::::.::::::::::::~::~:::::..: ................................ I.. ............... ............... ^. .............................. .............................. ...............

. ................ . .......................... : ...............

. .. . . z.ii;,:L ::::::: ;z;:;::k::::> >:::::..:: * ..... ............ ... . ..................... .................................... ..A 4 ................................................................. ................ I ............... : ................. . ................................ :. . ................ ..............

::.

TABLE III. Mean values and associated varianCesfor the creeping-flow transversepermeability of 2D random arrays of cylinders at the listed inclusion concentrations.

................................

. ............

.;.

............

. ...............

IO~~iiii~ ................................ C I? ...........ii::::~rili!~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~:~~~~~~~~~~~~~:~~~~~~~~~~~~~~:~~~~~ _................................3.. ............. ........... :ii:iiiii:i_r--j~iiiiiii~iiiiiiiiiiiiiiiiiiiEi iiiiiiii~iiiiiiiiii~iiiiiiii~iiiiiiiiiiiiiii~ 0.05 9.721E-01 :................~ .g.. ......j.. ..............................i.. .............f.. .............. .. 0.675 5.848E-01 10-1.:............. ... - . ............ii:,:.:. ...... ............_..... ...... ........ .............. ...... ..:: 0.10 ..................... 3.622E-41 ............................................................. ................................ ........... .....I......I.. ......I.. ............I......>I ........................................ ............... ., .................................. ...................... /................................ ................. ................ .... ............. .................. ............... 0.12 3.337E-01 ............................... ... .....l.li!:l~iii::::~:::::::::::::::i:::::::::::: ...) . ................*. ........................ 0.14 1.980E-01 ................+............................ .....................+................ 0.16 1.4388-01 lii!iiiiiii:ii:::: :::::.:::::~~::~~~~~:~~:::~ .:$: i..llIiiii:lji~lii!~~~~~~~~~~ :j ~ ::::::::::::::aiiiiiiliijiiiiiiiiiittiiiriiilieiiiil::::::.,,S:Fpiiii,ii,iiiiiiiiiiiiiiii ..............*. ..............................;.. .........*.:: 0.18 1.257E-01 -::.::::::::::T:::::::::::::::i:::::::::::::::~:::~:::~:::::::~::,.~ ............. y&c.:~:I..:;:;::~ 0.20 8.753E-02 ..............i.. ................................ ................ ............

(5.2
1.0738-03 1.8328-02 1.2638-02 1.784E-02 1.321E-03 9.891E-04 1.981E-03 7.708E-04 8.792E-04 2.476E-04 1.1358-04 7.885E-05 8.609E-05 2.142E-05 3.847E-05 1.069E-05 7.553E-96 3.512E-06 3.117E-06 2.911E-06 5.249E-07 1.312E-06 4.03OE-07 4.947E-07 2.04OE-07 1.908E-07 7.614E-08

0.22 6.690E-02 10 .:::..:..::::::~:::::::::::::::::::::::::::~:~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~;:~ :.;:ii 0.24 5.259E-02 ,o~:-:-::~~~~l ii~ 0.26 3.715E-02
0 0.1 0.2

0.3 C

0.4

0.5

0.6

FIG. 12. The behavior of permeability as a function of fiber concentrationin 2D random arrays. The enor bars representone mean-normalizedstandard deviation plus 5% bound gap error.

KLB

-+- KVB

lT=
where

2 -.
Ki ,
(69)

0.28 0.30 0.32 0.34 0.36 0.38 0.40 0.42 0.44 0.46 0.48 0.50 0.52 0.54 0.56 0.58

3.161E-02 2.424E-02. 1.767E-02 1.516E-02 l.l89E-02 9.943E-03 7.179E-03 6.372E-03 4.964E-03 3.363E-03 2.972E-03 1.969E-03 1.720E-03 l.O99E-03 8.101E-04 6.428E-04

k, = j$

r=

and the sample size, here N, is K,= ~~({Y)~,c,h~),i= l,...,N,, is either the lower- or upper-bound configuration permeability. Likewise, the variance associated with 2, G is approximated by the average , of the lower and upper bound sample variances,
;i'2, dB t &B 2 9

(70)

where

cr,(ch)- +--$Ii=, (Ki-k*j2. N_

(70

prominent of which is the increased drag due to increased fluid-solid interaction and the likelihood of percolation-type flow mechanism. This sharp decrease in the permeability can be also interpreted in terms of simple channel flow mechanism. In the latter, the average velocity is proportional to the square of the plate separation; in porous medium, the average nearest neighbor interparticle spacing, 0 (c- ), plays the role of the channel separation. Predicting the form of dependence on the concentration is not a trivial task and has been carried out analytically only in the limit of dilute concentrations.57 It is, interesting, however, to note that the plot suggests an exponential decay with increased concentration of the form

The sample size N, is set to 20 which is expected to reasonably control the noise contribution. (For details on the sequential acceptance-rejection sampling procedure see Ghaddar3 For N,=20 realizations, the most time consum) ing calculation (~ 0.58) requires 42 hours of CPU time, and costs roughly $600 on 16 processors at relatively good and 80 MFLOPS). This padIe performance ( VI6 -90% calculation is simply prohibitive on a workstation. Figure 12 shows the behavior of the permeability data points of the 2D random arrays as a function of the inclusion concentrations, along with error bars representing one meannormalized standard deviation plus 5% as a conservative account for the bound gap error. (The data points are listed in Table III.) We first note that the sharp diminishing of the permeability with increased inclusion concentration is a well-known fact on account of many factors. The most
Phys. Fluids, Vol. 7, No. 11, November 1995

K,(C)=cq)e-PO=,

(72)

where the constants a0 and /Ia are determined from a weighted least square fit to the data based on the meanstandard deviation as cya= 1.171 and normalized PO= 12.736. Clearly this simple model is only applicable in the indicated range of concentrations; in the limit c--+c,,,=O.82 (Berrymat? the model will predict a ), not zero value of the permeability which is not consistent with the physics of the problem. (No attempt made to consoliis date the exponential behavior with any analytical analysis.) Next we evaluate the applicability of the Kozeny model for the 2D random arrays over the wide range of concentrations. In Fig. 13 we plot calculated values of the Kozeny constant, k [equation (2)] based on the simulation permeability data points versus the porosity points 1 - cj . We also plot
Chahid K. Ghaddar 2575

Downloaded 29 Dec 2010 to 210.212.58.168. Redistribution subject to AIP license or copyright; see http://pof.aip.org/about/rights_and_permissions

IO3

..f . . 1.............. L . . L . ..t....... -.L . I . . . :...: . . . . . . . . . . . . > .,.........., < . . . . . ../.............. % . . . . . . . . . . L. . . . ::: . . . . . . . . . . . _. . . . . . . . . . . . . . . . . . . . . . . . . . . ..,,............ z . . . . . . . . . . . . . . ._,.. _. . . . . . . . ,............ i.. ...................... . . .. * . .. ..r........; . ............................. ............................ i.. j............................. L.. ....... ............... current ... . .......................... <. .......................... ..... . ............. . .I;. ........ ........... ... ........................................ . . <.......... .. . ...1.-. pw&+.[rg59] ........................................... ..i..

........................................ ,~appel$g5g~. ~ . . ................ .

j(f ,::::::::::::i :::::::::::::i ::::::: :::::::: :::: j ,~~~.~:~~:nd ~~~~~ ........................... ........................ I...... ........................... ................................. ............ I ...... ............. .............. I ...... ....... ......... i.. +..:&&qgq::::::;::::::::::::: i ,..... ........ t
k
..................... . .............

. .

. ...

. ....

. . ..

..............

.._

....... I.. ..................... L ............. . ....... .._ ... . ..............

_-__-.-_---_---....................................... ............. I.. ........... . ..............

. ..............

..................................... ...f .............I..............,...,........ i.. .......I.. 1 I I lOOI 4 9 I I


0.3 0.4 0.5 0.6 e 0.7 0.8 0.0 I

t .,.......

. ..i ..I..........

{ . . . . . . . . . . . . . ... . . . . . . . . . . . . .. . . . . . . . . . . . i . . . . . . . . . . . . i . . . . . . . . . . . 1

FIG. 13. Calculated values of the Kozeny constantbasedon the current 2D random array calculations, cell models, and some related experimentaImeasurements.(The error bars representone mean-normalizedstandarddeviation plus 5% bound gap error.)

the error bars which represent one mean-normalized standard deviation plus 5% to account for the bound gap error. [It is worth noting that the way we calculate k based on the sample mean permeability value does not give us directly information on the standard deviation associated with k. However, due to the inversely linear relation of the permeability and the Kozeny constant (for a fixed porosity), it is easily shown that the mean-normalized standard deviations are, in fact, equivalent (A k/k = AZ?/&. ] On the same figure we also plot the Kozeny constant predicted form the cell models of Happel, l8 Kuwabaralg and Spielman and Goren* for flow perpendicular to cylinders, as well as few experimental measurements performed on predominantly transverse fibrous media. We make several remarks on the figure, o The noise contribution to the data is certainly high, and disappointing. Similar high noise levels have also been reported by Cruz and Patera, but it is not readily clear why the noise is significantly pronounced in the creeping flow problem relative to an analogous effective conductivity calculation comprising identical number of samples.58 Furthermore, the higher concentrations possess a larger number of inclusions for the specified X0 and thus, due to spatial averaging effects, onewould expect smaller associated variances, however this is not observed here. It thus appears that the permeability is very sensitive to the inclusion distribution which may be blamed on the high sensitivity of the local pressure to the configuration details which consequently affects the mean flow and hence the permeability. Further analysis, however, is required to shed light on this observation. o The present data points indicate that, in the porosity range E E l-0.56,0.88], k can be modeled as a constant with an average value of 10.025 which is noticeably larger than the -5 values associated with 3D beds of spheres as well as
2576
Phys. Fluids, Vol. 7, No. II, November 1995

3D fibrous media (media with random orientation of fibers axes) and granular beds. The higher value, - 10, is blamed on a larger surface area in contact with non-stagnant fluid. In other words, the likelihood of forming stagnation (trapqingj regions is smaller, leading to more energetic fluid in contact with the fibers relative to a 3D random bed at the same porosity. 0 The author is not aware of any experimental data on the transverse permeability of strictly unidirectional fibrous media with a microstructure that could beadequately modeled by a random .sequential addition process (the method employed to generate the current realizations). On the other hand, there are measurements156 in the porosity range 0.7g~GO.9 on predominantly transverse fibrous media using air as the working fluid which predict a Kozeny constant of approximately 6-7 (see Fig. 13). Unfortunately, the lack of the microstructure details associated with these measurements, in addition to possible compressibility effects preclude rigorous comparison. Nevertheless, it is clearly observed that most of these measurements lie, in fact, within the confidence levels of our data points, which indicates, at least, reasonable agreement. 0 The models of Happel and Kuwabara appear to form lower and upper bounds, respectively, to our data points in the porosity range [ 0.76,0.925], whereas the Spielman and Goren model exhibits excellent agreement with our data in s this range. At lower porosities, the permeability behavior of the 2D arrays, which is seen to increase sharply as the maximum packing is approached,5g deviates markedly from the behavior in 3D beds which agrees with the curve of Happel with approximately a constant value of k = 5. It is important to note that this behavior in the 2D arrays is fundamentally different from the behavior of the Spielman and Goren model which breaks down at lower porosities. s The latter which is based on the Brinkman model, breaks s down due to physically unrealistic assumptions, since the effect of many solid boundaries in the immediate proximity of a fiber can not be lumped into a simple Darcy damping term. The behavior of the 2D random arrays is attributed to pure geometrical effects and is best elucidated by considering regular arrays: as the maximum packing is approached, a percolation mechanism for the flow prevails until the flow is totally blocked and the Kozeny constant tends to infinity. This can be proven rigorously based on the available asymptotic expression for the drag in regular arrays near the maximum packing. 226o This mechanism is also expected in random arrays although the author is not aware of any data or asymptotic expressions near the maximum packing that support this claim. (Unfortunately, our random sequential addition Monte Carlo sampling procedure is prohibitively inefficient for concentrations greater than a jamming threshold of 0.6 which precludes considerations of lower porosities than present.) On the other hand, in 3D beds, a path for the how is always guaranteed even at the maximum packing and the drag will not increase indefinitely. Here it is important to note that the agreement of the 2D cell models of Happel and Kuwabara (which are inconsistent with our 2D data) with the behavior of true 3D beds is probably due to the no shear or no vorticity assumptions on the cell boundary.
Chahid K. Ghaddar

Downloaded 29 Dec 2010 to 210.212.58.168. Redistribution subject to AIP license or copyright; see http://pof.aip.org/about/rights_and_permissions

TABLE IV. A comparison of the creeping-f-lowtransversepermeability of 2D random arrays. Here N is the number of cylinders, N, is the sample size, X is the supercell size, and the % diff. is with respect to the present data. c N N, x Kccr % diff. N N, x i?G % diff. N N, h kk(T %diff. Presentdata 6 20 6.865 0.3622 t- 0.1124 ... 17 20 6.671 0.02424 +- Q.28E-3 .. . 27 20 6.512 0.001969 t 6.35E-4 ... Sangani and Yao 16 5-8 11.209 0.41750 + 0.03 15.3 16 5-x 6.472 0.02550 t 2.OE-3 5.2 9 5-8 3.760 0.00235 + 4.24E-4 19.3 Howell? 1.. ... ... 0.4475 23.6 ... ... ... 0.02425 0.004 1.. ... ... 0.001925 2.2

TABLE V. Temporally-averaged permeability results for the square array at

the indicated concentrationsand Reynolds numbers. c


P K Re

Steady/Unsteady S S S S S S S u u S S S S
U S

0.1

0.2

0.3

0.3

0.5

0.4

0 At the higher porosities, E+ I, the permeability is known to vary as l/c in 3D beds (Batchelo? Using this ). fact in the Kozeny equation, it is easily shown that the Kozeny constant varies also as l/c and the model is not valid in that limit. The cell models of Happel and Kuwabara are consistent with this fact as can be seen in Fig. 13. The behavior of the 2D random arrays appears to be in agreement with the cell models although the dependence on c is likely to be different. Unfortunately, we do not have sufficient data in that limit to fully demonstrate this behavior. We conclude the discussion on Stokes permeability with a comparison to the few semi-analytical results reported by Sangani and YaoZ3 and the theoretical predictions of Howells*i based on a correction to the Brinkman model. s Referring to the comparison presented in Table IV we comment the following.

0.5

0.6

0 295.8 635.0 900.0 1632.2 2025.0 2500.0 2809.0 3271.8 0 894.0 1953.6 5112.2 10261.7 0 2540.2 5610.0 14932.8 30870.5 0 7956.6 18198.0 51256.9 118956.0 0 32869.7 74911.7 212152.4 491120.6

7.62OE-2 6.602E-2 6.141E-2 6.0558-2 5.752E-2 5.644E-2 5.532E-2 5.346E-2 5.1648-2 2.544E-2 2.221E-2 2.042E-2 1.8958-2 1.6568-2 9.0098-3 7.963E-3 7.221E-3 6.404E-3 5.5878-3 2.949E-3 2.617E-3 ?.315E-3 2.055E-3 1:750E-3 7.424E-4 6.742E-4 5.885E-4 4.8678-4 3.878E-4

0 19.5 40.0 54.5 93.9 114.3 138.3 150.2 168.4 0 19.9 39.9 96.9 169.9
0

20.2 40.5 95.6 172.5


0

S S S
U S

20.8 42.1 105.3 208.2 0 22.2 44.1 103.2 190.4

S S S IJ S S S S u

0.075 -

0.08 !----0.07 -

I-----I
0 current x Edwards et al

0.065 -

x 0

x 0

0
6 0
0 0 3:

20

40

60

80

100

-2-l
120 140 160 180

Re PIG. 14. Permeability values of the squarearray at c = 0.2 for several Reynolds numbers.
Phys. Fluids, Vol. 7, No. 11, November 1995

(Note that in the two articles the radius is assumed unity as opposed to the diameter in the present work. Therefore, their data is renormalized such that the radius is half.) In general the comparison should be regarded as good even at the largest difference since the confidence intervals overlap. However, it appears that the present data point at the lowest concentration c = 0.1 is less accurate than the two others which is evident by the large associated standard deviation of the current data point and the good agreement of the other two. Now given that our sample size is over two times larger than the sample size used by Sangani and Yao, one suspects that the chosen supercell size of X-7 is perhaps not large enough to filter out the special effects at the lower concentrations. On the other hand, the present data point at the highest concentration, c=O.5, seems to be more accurate than the value reported by Sangani and Yao given our twotimes larger supercell size and sample size in addition to the good agreement with Howells. (In fact, Sangani and Yao report convergence difficulties at higher concentrations and larger number of cylinders.) It is also interesting to note that Howells correction to the Brinkman model by finding the s s resistance due to a second cylinder and then averaging over all its possible positions not only stabilizes the model, but also provides accurate results.

Chahid K. Ghaddar

2577

Downloaded 29 Dec 2010 to 210.212.58.168. Redistribution subject to AIP license or copyright; see http://pof.aip.org/about/rights_and_permissions

x Edwards et al

x
0.018 0.016 0 I 200

50

100

150

Re
c= .5

Re

IO

0 0

6.5

2.5 K

0 X 0

2
X

0 0

1.5

100

200

Re

Re

FIG. 15. Permeability values of the squarearray at several Reynolds numbers and four different concentrations.The 0 s indicate current data; the Xs are data reported by Edwards et al.

C. Inertial porous-media 1. Regu/ar arrays

flow results

Square array results. For the sake of making a comparison with the limited available numerical data, we restrict our selected concentrations and Reynolds numbers mainly to the comprehensive set reported in Edwards et aLz6 These concentrations and Reynolds numbers are l Cj E {0.2,0.3,0.4,0.5,0.6}, *Re E {0,20,40,95,180). Since our solution strategy does not permit defining Re a priori, Re is calculated a posteriori from (23), where p is estimated based on the permeability results provided in Edwards et a1.26 such that to yield the desired Re.
2578 Phys. Fluids, Vol. 7, No. 11, November 1995

Figure 14 shows the current permeability results (marked with OS) for the concentration c=O.2 at approximately the Reynolds numbers listed above and four additional values. (The data points are listed in Table V.) Also plotted on the same figure the data reported by Edwards et a1.26 (marked with Xs). The comparison for the remaining set of concentrations 0.3, 0.4, 0.5 and 0.6 is provided in Fig. 15. Figures 14 and 15 show, generally, a reasonable, agreement between the two calculations. However, there are conspicuous discrepancies that deserve to be investigated to establish the possible sources of errors and, consequently, the relative reliability of the two calculations. Referring to Figs. 14 and 15 we first note the relatively large discrepancy at the Stokes solution (Re =O). ComparChahid K. Ghaddar

Downloaded 29 Dec 2010 to 210.212.58.168. Redistribution subject to AIP license or copyright; see http://pof.aip.org/about/rights_and_permissions

Re = 150.2

r
25x

I
0 current +

+ Bergelin et al
* Eidsath et al x Edwards et al
x

20-0.4; 5 IO 1.5 20 25 30 35 I 40 VP 0.402 I 15 I

Re = 168.4

10 t

f x

a1

O -02-0.4 -

a0

IO

15

20

25

30

35

40

0' 0

50

100

time unit

Re FIG. 18. Comparisonof dimensionlesspressuredrop, VP, versus Reynolds number, for the square array at concentration 0.5 between the experimental data of Bergelin et al., the numerical data of Eidsath et al., Edwards et al., and the current data.

FIG. 16. History plots of the y,-component of the velocity at Re = 150.2, and Re= 168.4, respectively, for the square array at concentration0.2.

ing the two calculations to the exact solutions we find that our Stokes permeability values are virtually exact, whereas Edwards et aZ. permeability values deviate appreciably by as high as 6.2% at c = 0.6. This deviation in the Stokes limit of Edwards et al. indicates a lack of resolution due possibly to large iteration or/and discretization errors. In Edwards et al. a penalty finite element method based on the steady formulation of the cell problem is used. No iteration error control is reported for the Newton iteration scheme. The computational domain is restricted to one fourth of the square cell for the Stokes problem, and half the cell for the finite Reynolds-

number runs, with symmetry assumptions on both the velocity and pressure. Edwards et al. report using typically 400 isoparametric biquadratic square finite elements in half the cell. In this work a mixed finite element method based on the unsteady formulation is used, along with a stringent error control imposed on both the velocity and pressure for the steady runs. (More precisely, the Lz-norm of the residual in both the velocity and pressure is required to be less than

Lumoed Mass Matrix

-4.5 ;o 2l.5 1

:
36.5 iI

f
3j.5

:
L3*

:
31.5

Re=O

Re = 19.9

Exact M W ;s Matrix

46

2i.5

zi

2i.5

38

36.5

;I

3i.5

3A.5

time unit
Re = 96.9 Re = 169.9

FIG. 17. Time history plots for the yt-component of the velocity at p= 3271.8 using both the lumped mass matrix (upper) and the exact mass matrix (lower) for the square array at concentration c = 0.2.
Phys. Fluids, Vol. 7, No. 11, November 1995

FIG. 19. Streamlines contours for tlow through a square array of cylinders at 0.3 concentration at the indicated Reynolds numbers.
Chahid K. Ghaddar 2579

Downloaded 29 Dec 2010 to 210.212.58.168. Redistribution subject to AIP license or copyright; see http://pof.aip.org/about/rights_and_permissions

c=

.3
I

0.028 r 0.028 0.024

I
s? 50

0 current x Edwards et al

x 10 (, IOiig9
8-

c = .4

K
0.018 0.016

7-

B540

s?
X

,,I , ( 2 , ox 1
0 100 150 200

3-

Re

Re c= .I3
1.5 1.40

x109
2

1.31.2.
IE

1.1 X IX

2.5 -

0.9 0.8 -

X 0

0.71 0 Re

50

100

150

200

Re

FIG. 20. Permeability values of the hexagonalarray at severalReynolds numbersand four different concentrations. 0 s indicate current data; the Xs are The data reportedby Edwards et al.

1 X low5 in virtually all the steady runs reported.) The complete square array cell was considered with periodic boundary conditions for both the Stokes and inertia1 runs, with a typical 1600 second order isoparametric triangular finite elements, which is twice the number of elements reported by Edwards et al. The lack of resolution could be the primarily reason for the discrepancy at the finite Reynolds numbers results as well. Another important factor is the unsteadiness observed at Re, approximately 150 and greater for which the flow exhibited steady periodic oscillations. The unsteadiness results in larger dissipation and thus lower permeability. This effect is clearly observed by the sudden drop in the temporally-averaged permeability values for the last two largest Reynolds numbers at c = 0.2. Note that Edwards et al.
2580 Phys. Fluids, Vol. 7, No. 11, November 1995

steady approach with imposed symmetry assumption on the velocity and pressure is inconsistent at the higher Reynolds numbers. Figure 16 plots the history curves of the yt-component of the velocity at a selected point in the square array at c=O.2 for the two unsteady calculations at Reynolds numbers Re= 150.2 and Re= 168.4 respectively. We first note that the onset of unsteadiness at about Re = 150 is consistent with limited previous visual studies in regular arrays of spheres.6LY27 Re = 138.3 no oscillations were observed up [At to 30 time units.) Second, we also observe the expected increase in the amplitude of oscillations at the higher Reynolds number which is also accompanied with a slight increase in the frequency. Third, the unsteadiness phenomenon in periodic arrays is in the form of traveling waves characterized by
Chahid K. Ghaddar

Downloaded 29 Dec 2010 to 210.212.58.168. Redistribution subject to AIP license or copyright; see http://pof.aip.org/about/rights_and_permissions

TABLE VI. Temporally-averagedpermeability results for the hexagonal array at the indicated concentrationsand Reynoids numbers. c
P 0 967.2 2342.5 9467.3 K
Rt?

Steady/Unsteady
S S S S

0.3

15550.1 0
0.4 23342.6 4900.0 26114.6 60172.1 0 5625.0 12723.8 60811.6

0.5

0.6

119577.6 0 15351.2 37869.2 126878.4 226861.7

2.703E-2 2.081E-2 1.748E-2 1.311E-2 1.102E-2 l.O56E-2 8.713E-3 7.481E-3 4.784E-3 3.442E-3 4.108E-3 3.653B-3 3.16OE-3 2.025E-3 1.525E-3 1.45133-3 1.354E-3 1.200E-3 9.0928-4 7.314E-4

20.1 40.9 124.1 171.2 0 20.4 36.6 124.9 207.1 0 20.5 40.2 123.1 182.3 0 20.8 45.4 115.3 165.9

U S S S S U S S S S U S S S S U

distinct periods and amplitudes, and is fundamentally different from the periodic wake (vortex shedding) phenomenon behind bluff cylinders. (We remark that further analysis of unsteadiness phenomena is beyond the scope of the current study.) The unsteadiness raises an important question conceming the effect of the lumped mass matrix employed in place of the actual mass matrix, for reasons of computational savings. [The lumped mass matrix has virtually no effect on steady time-independent solutions (see Section III C ).] To investigate this effect, we re-compute the flow for the case p=57.2 (yields Re = 168.4 with the lumped mass method) using the exact mass matrix. In Fig. 17 we plot the time history of the yt-component of the velocity resulting from both strategies at an identical selected history point. The results indicate that the lumped mass matrix has a small damping effect as we observe the small increase in the oscillations amplitude when using the exact mass matrix. The exact mass matrix method wiIl thus yield a lower permeability due to the higher amplitude oscillations and thus higher dissipation. The actual reduction in the permeability however, was found to be inappreciable (the relative change is less than 0.33%). This slight improvement in the measurement does not warrant the 0 (5 >- 0 ( 10) more expensive exact matrix method. We next compare our results against the available experimental data of Bergelin et al.,63 and the numerical data of Eidsath et al.= at c= 0.5. The comparison is shown in Fig. 18. We first remark that the data are presented in terms of a dimensionless pressure drop, 6, defined in (8) rather than permeability, K, used in the previous comparisons. The two are, of course, related by the following relation: VP=&.
1

Bergelin et a2. than the data reported by Edwards et al. and Eidsath et al. Our last two data points happen to be at higher Reynolds numbers than the rest of the data which precludes accurate comparison. We conclude the discussion on square arrays by describing the flow structure at different Reynolds numbers. In Fig. 19 we plot the streamlines for the square array at concentration 0.3 for several Reynolds numbers. At finite Reynolds number a pair of symmetric vortices is seen to form in the gaps between cylinders. As Reynolds number is increased, the vortices intensify with their centers pushed apart resulting in parallel streamlines through the free path. At the highest Reynolds number the flow is unsteady and the plot represents an instantaneous streamlines configuration. A deeper qualitative picture of the flow pattern with increasing Reynolds numbers can be described in conjunction with the visual studies of Dybbs and Edwards. The latter considered porous media consisting of Plexiglas spheres in hexagonal packing and Plexiglas rods arranged in a complex, fixed three dimensional geometry. Although the pore structure is different than our simple square array, there is still a close parallel in the flow structure between the different structures. The visualization studies showed that at Re = 1 boundary layers begin to develop near the solid boundaries of the pores. At Re> 10, the boundary layers become more pronounced and an inertial core appears. As the Reynolds number is increased, the core flows enlarge in size and their influence become more and more pronounced on the overall flow picture. The first evidence of unsteady tlow is observed at Rem 150, in the form of laminar oscillations in the pores. These oscillations take the form of traveling waves characterized by distinct periods and amplitudes. Hexagonal array results. The next type of regular arrays considered, is that of a hexagonal arrangement as shown in Fig. 8. Again we restrict our selected concentrations and Reynolds numbers mainly to the set reported in Edwards et al. to allow for constructive comparison. These concentrations and Reynolds numbers are
l l cj~{0.3,0.4,0.5,0.6},

Re ~{0.20,40,120,180}.

(73)

For the first two Reynolds numbers, the figure suggests that our data lie closer to the experimental measurements of
Phys. Fluids, Vol. 7, No. 11, November 1995

Unlike the square array, the permeability of the hexagonal array differs along any two orthogonal directions. Here, we compute the permeability along the y t-direction for the hexagonal configuration shown in Fig. 8, which we shall denote simply as K. (Our yt-direction matches y,-direction with respect to the Edwards et al. hexagonal cell.) It is important to note here that, due to the nonlinear Navier-Stokes equations, knowledge of the permeability for any two orthogonal directions can not be used (by simply adding vectors) to calculate the permeability along a different oblique direction. This statement is not true in the case of linear Stokes flow in which knowledge of the permeability tensor Kij will uniquely specify the permeability along any desired direction. We note that our computational domain is four times larger than that considered in Edwards et al. (Our hexagonal cell contains four inclusions whereas the Edwards et al. cell contains two inclusions but they consider only half of it with symmetry assumptions.) We have used typically O(3000)
Chahid K. Ghaddar

2581

Downloaded 29 Dec 2010 to 210.212.58.168. Redistribution subject to AIP license or copyright; see http://pof.aip.org/about/rights_and_permissions

0.0220.02 n 0.018 0.016 -

O x 0
0 X 0

Re
c= 4.5, x103 .5 I 16 x lOA < 4
X

Re
c== .6

I4 12lO-

x
X

3.5 tc
i 0 X

80
0 0

X 0 0

2.5 6ox
0

21.5 0

4Y , 200 2

100

300

50

100

150

200

Re

Re

FIG. 21. Comparison between square and hexagonal arrays permeability values at several Reynolds numbers and four different concentrations.The OS indicate square array data; the Xs indicate hexagonal array data.

tinite elements, that is, O(800) elements per quarter of the domain which is twice the number reported in Edwards et al. Figure 20 shows our permeability results (marked with 0 s) and those reported by Edwards et al. (marked with Xs) for the combination of concentrations and Reynolds numbers above. (-The data points are listed in Table VI.) Again we observe notable discrepancies particularly at c=O.4 and c=O.6, which are also believed due to lack of resolution in Edwards et al. results. This is evident by the consistent deviation of Edwards et al. Stokes permeability values from the exact values, as well as the nonsmooth change in Edwards et al. data at Rem40 (see Fig. 20). [Note that the Edwards ef al. value at c=O.4 at the (unsteady) Rem 180 case lies below our temporally-averaged value, contrary to what has been typically observed] On the other
2582 Phys. Fluids, Vol. 7, No. 11, November 1995

hand, a good agreement is observed for c=O.3 and c=OS. At c=OS, in fact, we find a very good agreement between the two calculations for all the reported Reynolds numbers including Re- 180, for which the flow is unsteady. This point was targeted using a finer mesh with virtually insignificant improvement in the reported value. This good agreement, despite Edwards et al. steady approach, is probably due to the fact that the permeability, which is a spatial average, is insensitive to the how details. It is interesting to compare the permeability results for the square and hexagonal arrays. Figure 21 provides this comparison, which indicates the existence of a crossover between the two arrays at all the concentrations shown, expect the highest c = 0.6, at which no crossover occurs in the range of Reynolds numbers considered. At the Stokes limit, the
Chahid K. Ghaddar

Downloaded 29 Dec 2010 to 210.212.58.168. Redistribution subject to AIP license or copyright; see http://pof.aip.org/about/rights_and_permissions

TABLE VIL Temporally-averagedupper-boundpermeability resuhs for random arrays at the indicated concentrations,Reynolds numbers, and the final integration times, fr . c
P ~(c,R@,h=7) Re

Steady/Unsteady S S S S S s
S

tf
... ... ... ... ... ...
...

0.2
Re=O Re = 20.1

0.3

0.4

0.5

0 289 625 0 900 2025 4900 10000 0 2500 5625 8100 18225 50625

9.5861E-2 6.2602E-2 4.9754E-2 2.7473E-2 1.7954E-2


1.3931E-2 l.O161E-2

0 18.1 31.1 0 16.2 28.2


Lo.0

7.0291E-3 1.3925E-2 9.5538E-3 7.7745E-3


1.8017E-3

70.3 0 23.9 43.7


14.6

U S S S
S

29.18 ... ... ...


...

1.2816E-3
l.l223E-3

23.4 56.8

S S

... ...

Re = 124.1

Re = 171.2

2. Random arrays
FIG. 22. Streamlines contours for flow through a hexagonal array of cylinders at 0.3 concentration at the indicated Reynolds numbers.

hexagonal arrays possess higher permeability than the square arrays. The Reynolds number at which a crossover occurs appears to increase with increasing concentrations. The crossover Reynolds numbers are approximately 12, 50, and 125 for the concentrations c=O.3, c = 0.4, and c= 0.5, respectively. It is very likely that a crossover will also occur for the concentration c= 0.6 at a higher Reynolds number value. The reasoning behind this behavior can be explained as follows: the hexagonal array possesses less surface area relative to a square array at the same concentration which leads to higher permeability for small Reynolds numbers. At higher Reynolds numbers, the tortuous character of the hexagonal array leads to a higher dissipation than does the straight through character of the square array due to the shielding effect, and hence, a crossover occurs. As the concentration is increased, the surface-related drag effects become stronger (the square array has larger surface area), which, in turn, could explain the delay in the crossover with increasing concentration. We conclude this discussion with a brief description of the flow structure in the hexagonal array at various Reynolds numbers. Figure 22 depicts the streamlines contours for the hexagonal array at concentration 0.3 for several Reynolds numbers. At Re = 0 the how is symmetric and no vortices are present (,or at least that could be detected numerically). As the Reynolds number is increased, a symmetric pair of vortices is seen to appear in the gaps between successive inclusions. These vortices grow with increasing Reynolds number with their centers being drawn farther apart forcing the streamlines to be parallel in a channel-like flow (see the flow at Re= 124.1). At Re= 171.2 the flow is unsteady and the picture represents an instantaneous configuration of the how pattern.
Phys. Fluids, Vol. 7, No. 11, November 1995

We present, here, conjgura~ion upper-bound permeability results which are calculated for randomly-generated realizations at selected concentrations and flow Reynolds numbers: (1) c=O.2, Re E { 18,31}, (2) c=o.3, Re E { 16,28,60,70},
(3) (4) c = 0.4, c=o.5, Re E { 24,44}, Re E ( 15,23,57}.

(We remark that the above Reynolds numbers have been rounded up to an integral value. The actual Reynolds numbers and the permeability values are listed in Table VII.) The size, h, of the supercells is chosen reasonably large, X0=7, to insure a higher likelihood that our configuration values do represent the stipercell (multiple realization ensemble) permeability values. [We note that, due to computational considerations, only upper bound nip elements are used (c~~=O.l,/3= 0.1). However, based on the results of Section IV A 2, we do not expect the one-sided bound error to be significant.] For each of the first three concentrations, we have only varied the Reynolds number using the same realization. The three upper bound random realizations are shown in Figs. 23(a), 23(b), and 23(c), respectively. For the concentration c = 0.5, three different random realizations [shown in Figs. 23(d), 23(ej and 23(f)] are used for the four Reynolds numbers listed. The number of elements used ranged from 7200 (c=O.2) to 11200 (c=O.5). It is believed, based on wiggle free history plots (see Fig. 24) and the moderate Reynolds numbers, that this resolution is adequate for accurate permeability prediction. The flow has been found to be steady for all the Reynolds numbers listed, except the case (c= 0.3,Re = 70). It is worthy to note that these calculations are computationally intensive. The higher Reynolds.number calculations require approximately 40 hours each to reach a steady state. (Had the upher bound nips not been used, this time would have significantly increased.) The author is not aware of any published studies on inertial flows through random arrays of cylinders, and hence
Chahid K. Ghaddar

2583

Downloaded 29 Dec 2010 to 210.212.58.168. Redistribution subject to AIP license or copyright; see http://pof.aip.org/about/rights_and_permissions

----L-JOO.@ 0 900 iiiiiz u


1 0.5 Ul 0 -0.5 0 2

c = .3, &e = 60.0

IO

12

14

c = .4, Re = 43.7

(4

Ul

I 3

c = .5, Re = 56.8 1

-0.2'

L 0 12 3 4 5 6 7 8 9

I lo

time unit

(b)

(4

id

0 8% o&g Oo 0 0 P m (f) -I

I;IG. 24. Time history plots of u, for selected steady Navier-Stokes calculations at the indicated Reynolds numbers and concentrations.

FIG. 23. Upper-bound random realizations used for the inertial permeability calculations: (a) c= 0.2 and various Reynolds numbers; (bj c=O.3 and various Reynolds numbers; (c) c=O.4 and various Reynolds numbers: (d) c=O.5 and Re= 14.6: (e) c=O.5 and Re=23.4; (f c=O.5 and Re=56.8. )

contact with energetic fluid. On the other hand, more free paths are available for the fluid in the 3D bed which lowers its resistance. These two combined effects (among possibly others) lead to the observed behavior. Third the discrepancy between our data points and the Ergun equation is not systematic in the porosity, or one would have expected say, a better agreement with data points at porosity 0.6 (given the behavior at porosity 0.5). Nevertheless, given the fact that these data points are configuration based, and the relatively high noise level observed in the creeping flow results, it is

no data is available for constructive validation and comparison. We shall, however, compare our results to the Ergun correlation for true 3D granular beds. In Fig. 25 we plot * our data points in terms of the modified friction factor, equation (6), along with the Ergun relation. This figure raises more questions than one can provide answers to based on the limited data. However, we make a few remarks. First, the agreement between the Ergun relation and our data points at porosity 0.7 is indeed remarkable although nonintuitive, and no plausible explanation could be provided for this behavior. Second, the data points at porosity 0.5 which fall in the porosity range 0.436 ~60.52 characteristic of the Ergun relation happen to fall above the curve. This suggests that a 2D random array possesses lower permeability than a comparable granular bed. (Note that our permeability values are already upper-bounded.) This is consistent with the observation made earlier on creeping flows as well (see Section IV B). The reasoning behind this follows the same argument made for the creeping flow results and is hlamed on pure geometrical effects: on one hand fewer trapping regions are expected to form in the 2D random array due to its more structured microstructure as opposed to a 3D granular bed. This, in turn, allows for a larger surface area in
2584
Phys. Fluids, Vol. 7, No. 11, November 1995

e=.8

0 e=.7

x e=.6 * t-=.5 -Ergun eq.

50

100

150

200 Re&

250

300

xl0

400

I 450

J 500

FIG. 25. Comparison of the modified friction factor [equation (6)] predicted by Fzgun equation versus values calculated for random 2D arrays of cylinders at selected concentrationsand Reynolds numbers.
Chahid K. Ghaddar

Downloaded 29 Dec 2010 to 210.212.58.168. Redistribution subject to AIP license or copyright; see http://pof.aip.org/about/rights_and_permissions

likely that a more systematic study (albeit expensive) will yield a more consistent behavior. V. CONCLUSIONS A parallel finite element procedure is developed and employed in conjunction with proper statistics to determine the statistical transverse permeability of two-dimensional random arrays of cylinders. Kesults are obtained for both low and moderate Reynolds number flows and over a wide range of area fractions. The main points of this work are summarized below. o The variational formulation of the Stokes and Navier-Stokes problems are presented, and it is mathematically shown that the permeability can only decrease for nonzero Reynolds numbers. 0 A nip-element method which provides rigorous and accurate upper and lower bounds for the configuration permeability is presented. The method is necessary to first, mitigate the geometry induced numerical difficulties, and second, to improve the statistical analysis by incorporating otherwise numerically intractable realizations. l Creeping flow permeability values are calculated in the concentration range [0.05,0.58] by averaging over 20 random realizations for each concentration value. The results are compared to the predictions of earlier cell models; 8-20 to few related experimental measurements; 56 to three analytical results reported by Sangani and YaoD at concentrations, c E {0.1,0.3,0.5}, and to the corresponding predictions of Howells. The permeability which exhibits an exponentiallike behavior is found to deviate markedly at low porosities from the cell models predictions while exhibiting good agreement at moderate and high porosities. A reasonable agreement is found with respect to the experimental measurements as well. A relative difference of approximately 15% and 19% is found respectively, at c= 0.1 and c = 0.5 with the data of Sangani and Yao while an excellent agreement is found at the intermediate value. The results are also found to be in excellent agreement with the predictions of Howells at c= 0.3 and c =0.5 while deviating by approximately 23% at c = 0.1. The discrepancy between the present and the analytical data is discussed based on the available facts. l Inertial flow permeability values are calculated for both regular and random arrays of cylinders up to a flow Reynolds number of approximately 200. ReguZar arraq Both square and hexagonal arranges. ments are considered. The flow is found to be unsteady for Reynolds numbers around or greater than 150. The unsteadiness is in the form of traveling waves characterized by distinct periods and amplitudes. The permeability values are compared to the results reported by Edwards et aZ.26 and some modest discrepancies are found. The discrepancies are attributed, on one hand, due to possibly lack of resolution in Edwards et al. calculations as suggested by the deviations from the known exact Stokes solutions, and on the other hand, due to their artificially imposed steady solution. Good agreement is found against the experimental measurements of Bergelin et ~1.~ It is also found that the lumped finite element mass matrix introduces a small damping effect when
Phys. Fluids, Vol. 7, No. 11, November 1995

used in place of the exact mass matrix for numerical efficiency reasons. The damping effect, however, has no appreciable effect on the permeability Random arrays. The configuration permeability is calculated for several random arrays at various concentrations and Reynolds number flow conditions. The results are compared against the Ergun correlation2 for general three-dimensional granular beds and the comparison is found to be, for the most part, satisfactory. ACKNOWLEDGMENTS This work was supported by the Advanced Research Projects Agency under Grant No. N00014-91-J-1889, by the Office of Naval Research under Grants No. NOOO14-90-J4124 and No. N00014-89-J-1610, and by the Air Force Office of Scientific Research under Grant No. F49620-94-10121. The work was performed in the Fluid Mechanics Laboratory of the Massachusetts Institute of Technology. The author is indebted to Professor Anthony T. Patera of the Massachusetts Institute of Technology for fruitful discussions. The author also wishes to acknowledge the assistance of CAPE, Inc. in meeting the publication costs of this manuscript.
K. Friedlander,Smoke Dust and Haze: Fundamentals ofAerosol BehavS. ior (Wiley, New York, 1977). Shapiro and H. Brenner, Dispersion/reaction model of aerosol filtraM. tion by porous filters, J. Aerosol Sci., 21, 97 (1990). 3J. Davidson, R. Clift, and D. Harrison, Fluidization, 2nd ed. (Academic Press,London, 1985). 4G. Hetsroni, Handbook of Multiphase System (McGraw-Hill, New York, 1982). 5F. E. Curry and C. C. Michel, A fiber matrix model of capillary permeability, Microvascular Res. 20, 96 (1980). 6J. R. Levick 1987, Flow through interstitium and other fibrous matrices, Q. J. Exp. Physiol. 72, 409 (1987). 7C. Y. Chen, Filtration of aerosolsby fibrous media, Chem. Rev. 55,595 (1955). *S. C. Stern, H. W. Zeller, and A. I. Schekman, The aerosol efficiency and pressuredrop of a fibrous filter at reduced pressures,J. Colloid Sci. 15, 546 (1960). 9A. E. Scheidegger,The Physics of Flow Through Porous Media, 3rd ed. (University of Toronto Press,Toronto, 1974). IoR F Probstein Physicochemical Hydrodynamics (Butterworths, Stoneh&MA, 198;). F. A. L. Dullien, Porous Media: Fluid Transport and Pore Structure, 2nd ed. (Academic Press, New York, 1992). Carman, Fluid flow through granular beds. Trans. Inst. Chem. Eng. *X. 15, 150 (1937). I3J. Happel and H. Brenner, Low Reynolds Number Hydrodynamics (Martinus Nijhoff, The Hague, 1983). 14C.P. Kyan, D. T. Wasan, and R. R. Kintner, Flow of single-phasefluids through fibrous beds, Ind. Eng. Chem. Fundam. 9, 596 (1970). t5E. Lord, Air flow through plugs of textile fibers. Part I: General flow relations, J. Text. Inst. 46, T191 (1955). i6J C Brown Determination of the exposedspecific surface of pulp fibers from air permeability measurements, Tappi 33, 130 (1950). 17S.Torquato S., Random heterogeneousmedia: Microstructure and improved bounds on effective properties,Appl. Mech. Rev. 44, 37 (1991). 5. Happel, Viscous flow relative to arrays of cylinders, Am. Inst. Chem. Eng. J. 5, 174 (1959). Kuwabara, The force experienced by randomly distributed parallel as: circular cylinders or spheres in a viscous flow of small Reynolds numbers, J. Phys. Sot. Jpn. 14, 527 (1959). L. Spielman and S. L. Goren, Model for predicting pressure drop and filtration efficiency in fibrous media, Env. Sci. Technol. 2, 279 (1968). *II. D. Howells, Drag due to the motion of a Newtonian fluid through a
Chahid K. Ghaddar

2585

Downloaded 29 Dec 2010 to 210.212.58.168. Redistribution subject to AIP license or copyright; see http://pof.aip.org/about/rights_and_permissions

sparserandom array of small fixed rigid objects, J. Fluid Mech. 64, 449 (1974). * S. Sanganiand A. Acrivos, Slow flow past periodic arrays of cylinders A. with application to heat transfer, Int. 5. Multiphase Flow 8, 193 (1982). a3A. S. Sangani and C. Yao, Transport processesin random arrays of cylinders II. Viscous Row, Phys. Fluids 31, 2435 (1988b). M. E. Cruz and A. T. Patera, A parallel Monte-Carlo tit&e-element procedure for the analysis of multicomponent random media. Int. J. Num. Methods Eng. 38, 1087 (1995). 15A. Eidsath, R. G. Carbonell, S. Whitaker, and L. R. Herrmann, Dispersion in pulsed systems-B& Chem Eng. Sci. 38, 1803 (1983). 26D. A. Edwards, M. Shapiro, I? Bar-Yoseph,and M. Shapira, The influence of Reynolds number upon the apparentpermeability of spatially periodic arrays of cylinders, Phys. Fluids A 2, 45 (1990). A. Dybbs and R. V. Edwards, A new look at porous media fluid mechanics-Darcy to Turbulent, in Fundumentals of Transpott Phenom~ in Porous Media, edited by J. Bear and M. Y. Corapcioglu (Martinus na Nijhoff, Boston, 1984), pp. 199-256. z8S. Ergun, Fluid flow through packed columns, Chem. Eng. Progr. 48, (1952). zgA. Bensoussan, J.-L. Lions, and G. C. Papanicolaou, Asymptotic Analysis for Periodic Structures (North-Holland, Amsterdam, 1978). 3oC.C. Mei and J.-L. Auriault, The effect of weak inertia on Bow through a porous medium, J. Fluid Mech. 222, 647 (1989). H. Lamb H., Hydrodynamics, 6th ed. (Cambridge University Press,Cambridge, 1932). 32V. Girault and P-A. Raviart, Finite Element Methods for Navier-Stokes Equations: Theory and Algorithms (Springer-Verlag,Berlin, 1986). s3G, K. Batchelor,An Introduction to Fluid Dynamics (Cambridge University Press,Cambridge, 1967). 34R.Hill and G. Power, Extremum principles for slow viscous flow and the approximatecalculation of drag, Q. J. Mech. Appl. Math. 9, 313 (1950). .S. Kim and S. J. Karrila, Microhydrodynamics: Principles and Selected Applications (Butterworths, Boston, MA, 1991). Nii, H. F. Weinberger,and A. Acrivos, Variational inequalities for a a6A body in a viscous shearing flow, J. Fluid Mech. 68, 739 (1975). 37B Widom Random sequentialaddition of hard spheresto a volume, J. &em. Phys. 44, 3888 (1966). Y. Rubinstein. Simulation and the Monte-Carlo Method (Wiley, New 8R. York, 1981). C. K. Ghaddar, Parallel computational methods for multicomponent media: Application to thermal composites and porous-media flows, Ph.D. thesis, Massachusetts Institute of Technology, 1995. 40M. E. Cruz, C. K. Ghaddar, and A. T. Patera, A variational-bound nipelement method for geometrically stiff problems; Application to thermal composites and porous media, Proc. R. Sot. London Ser. A 449, 93 (1995). F. Hecht and X. Sahel, Emc2: Editeur de maillages et de contours bidimensionnels,Manuel d utilisation, Rapport Technique No. 118, INRIA, 1990. 42M. D. Gunzburger, Finite Element Methods for viscous Incompressible Flows: A Guide to Theory, Practice, and Algorithms (Academic Press,San Diego, CA, 1989).

43C.Taylor and P. Hood, A numerical solution of the Navier-Stokes equations using the finite element method, Comput. Fluids 1, 73 (1973). @G. Strung and G. J. Fix, An Analysis of the Finite Element Method (Prentice-Hall, Englewood Cliffs, NJ, 1973). 45Y. Maday, D. Meiron, A. T. Patera, and E. M. Rdnquist, Analysis of iterative methodsfor the steadyand unsteadyStokesproblem: Application to spectralelementdiscretizations,SIAM J. Sci. Comput. 14, 310 (1993). %? F. Fischer and A. T. Patera, Parallel simulation of viscous incompressible flows, Annu. Rev. Fluid Mech. 26, 483 (1994). 47P.F. Fischer and A. T. Patera, Parallel spectral element solution of the Stokes problem, J. Comput. Phys. 95, 380 (1991). 48G H Golub and C. F. Van Loan, Mutti Computations (Johns Hopkins Umvkrsity Press,Baltimore, MD, 1983). 49G. Strung, Introduction to Applied Mathematics (Wellesley-Cambridge Press,Wellesley, MA, 1986). A. E. Saez and R. G. Carbonell, On the performance of quadrilateral finite elements in the solution to the Stokes equations in periodic structures, Int. Num. Methods Fluids 5, 685 (1985). 5 Borne, R. Chambon, and J. Auriault, Conforming finite element comL. putations applied to a spatially periodic, harmonic Navier-Stokes problem, Int. Num. Methods Fluids 5, 601 (1985). 52Y. Maday, A. T. Patera, and E. M. Rdnquist, A high-order splitting schemefor the unsteadyStokes and unsteady Navier-Stokes equations, in preparation. 53NEKTON V2.85 Manual, Fluent, Inc., Centerra ResourcePark, 10 Cavendish Court, Lebanon, NH 03766. 54G. E. Kamiadakis, M. Israeli, and S. A. Orzag, High-order splitting methods for the incompressible Navier-Stokes equations, J. Comput. Phys. 97, 414 (1990). K. Agarwal and J. C. Lewis, Computational fluid dynamics on parallel -(R processors,CFD Algorithms and Applications for Parallel Processors, FED-Vol. 156 (American Society of Mechanical Engineers, New York, 1993). s6G. Fox, M. Johnson, G. Lyzenga, S. Otto, J. Salmon and D. Walker,
Solving Problems on Concurrent Processors: General Techniques and Regular Problems (Prentice-Hall, Englewood Cliffs, NJ, 1988), Vol. I.

57G. K. Batchelor, Sedimentation in a dilute dispersion of spheres, J. Fluid Mech. 52, 245 (1972). S. Yepilyurt, C. Ghaddar,M. Cmz, and A. T. Patera, Bayesian-validated surrogates for noisy computer simulations SIAM J. Sci. Comput. (in pressj. 59J.G. Berryman, Random close packing of hard spheresand disks, Phys. Rev. A 27, 1053 (1983). mA. S. Sangani and A. Acrivos, Slow tlow through a periodic array of spheres,Int. J. Multiphase Flow 8, 343 (1982b). 6tT. H. Wenger,A. J. Karabelas,and T. J. Hanratty, Visual studies of flow in a regular array of spheres,Chem. Eng. Sci. 26, 59 (1971). A. Roshko, On the developmentof turbulent wakes from vortex streets, NACA Report 1191, 1954. e30. P. Bergelin, G. A. Brown, and S. C. Doberstein, Heat transfer and fluid friction during flow acrossbanks of tubes IV, Trans.ASME 74, 953 (1952).

2686

Phys. Fluids, Vol. 7, No. 11, November 1995

Chahid K. Ghaddar

Downloaded 29 Dec 2010 to 210.212.58.168. Redistribution subject to AIP license or copyright; see http://pof.aip.org/about/rights_and_permissions

Vous aimerez peut-être aussi