You are on page 1of 27

IMA Journal of Mathematics Applied in Business and Industry (2000) 11, 229-255

Application of an exactly incompressible lattice Bhatnagar-Gross-Krook method to internal pressure driven and depth-averaged flows
D. M. WHITE

Materials Research Institute, Sheffield Hallam University, Pond Street, Sheffield, SI 1WB, UK
I. HALLIDAY

Downloaded from imaman.oxfordjournals.org at Universitatea Transilvania, Braşov on February 8, 2011

Division of Applied Physics, School of Science and Mathematics, Sheffield Hallam University, Pond Street, Sheffield, SI 1WB, UK
C. M. CARE

Materials Research Institute, Sheffield Hallam University, Pond Street, Sheffield, SI 1WB, UK
AND C. YOUNG AND A. STEVENS

Rolls Royce & Associates, PO Box 31, Derby, DE24 8BJ, UK


Modifications to the lattice Bhatnagar-Gross-Krook (BGK) method eliminate compressibility errors from the simulation of macroscopic fluids governed by the steady-state Navier-Stokes equation. We describe the means by which this modified scheme makes it possible to compute both pressure drops and flow fields, bringing internal, pressure driven, isothermal flows within the scope of the method. The scheme is applied to the simulation of flow of a viscous incompressible fluid past a sudden expansion in an infinite aspect ratio duct (back-facing step geometry). Further, we devise the means by which an exactly incompressible (El) LBGK scheme may be extended and calibrated to calculate depthaveraged flow fields in ducts of uniform depth. By analytical calculation and by simulation we demonstrate the validity of our approach. Keywords: Simulation; incompressible hydrodyamics; depth-averaged

1. Introduction Commonly, to avoid the computational expense of full three-dimensional (3D) calculation, internal flows in ducts of constant depth are computed in two dimensions, without explicitly modelling the (shallow) depth of the duct (Fig. 1). In computational fluid dynamics (CFD) the results of such calculations are generally assumed to be depthaveraged quantities, and the influence of the 'unmodelled' dimension is accounted for by using different momentum sink terms in the flow momentum equations. In Section 2 we explore the basis of these assumptions and in Section 3 we proceed to examine the domain of applicability of two-dimensional (2D) calculations to laminar flow. In engineering calculations of duct flow, the principal agency generating the motion of the fluid (liquid) is a pressure difference. It is therefore essential properly to adapt any
The Inititute of Mathematics and its Application] 2001

230

D. M. WHITE ETAL

Downloaded from imaman.oxfordjournals.org at Universitatea Transilvania, Braşov on February 8, 2011

FlG. 1. Geometry of depth average flow calculations. In the D2Q9 EILBGK simulations the xy plane was explicitly modelled, with the z-depth entering through parameter k (equation (42)). For the finite-difference calculations of Section 3, the mesh occupied the yz plane.

LBGK scheme focused on depth-averaged quantities to accommodate boundary conditions which are given principally in terms of pressure. Of all Lattice Boltzmann equation (LBE) approaches to computation, the eponymous lattice Bharnagar-Gross-Krook (LBGK) scheme (Bhatnagar et al., 1954; Qian et al., 1992) is the simplest and most recent variant to attract interest. The standard approach to LBGK hydrodynamics recovers the incompressible Navier-Stokes equations only if spatial gradients in lattice density may be neglected. Because the latter maps onto simulated fluid pressure, the effect of pressure gradients is least problematic to LBGK simulation when modelled by impressing a spatially uniform body force over the lattice fluid. Whilst often accurate, this device becomes unacceptable where the pressure gradient is expected to vary throughout the system or where pressure varies over its boundaries. Moreover, the well-known association between lattice density and fluid pressure couples pressure and velocity fields and a need to restrict the attending compressibility errors further deters application of the standard LBGK method in systems containing significant nonuniform pressure gradients. In consequence, whilst standard LBGK achieves, with good accuracy, Reynolds numbers (Re) of around one thousand (Miller, 1995; Hou etal, 1995) in systems with fairly uniform pressure, there have been few applications to pressure-driven flows, and currently LBGK simulation in this area is dwarfed by traditional CFD. Hou et al. (1995) quantify compressibility effects arising in LBGK simulation of 2D lid-driven cavity flow, where the error introduced thereby is not great. For out application, so limiting (in terms of attainable Re; Section 4) are compressibility effects, that an attempt to assess them is of little value and we concentrate instead upon the detail of implementing a modified LBGK scheme. After a relaxation parameter co, the LBGK scheme contains only its equilibrium distribution function through which the governing macrodynamical equations might be manipulated (Qian et al., 1992; Zou et al, 1995). The modified LBGK scheme due to Zou et aL (1995) and independently Lin et al. (1996), recently shown exactly to recover momentum equations identical to the time-independent Navier-Stokes equations, entails a reinterpretation of the usual LBGK lattice velocity and density, but essentially adjusts the

APPLICATION OF LBGK METHOD TO INTERNAL FLOW

231

equilibrium distribution function in a simple way to eliminate compressibility error. These ideas we discuss further in Section 4. In Section 5, by appropriate adjustment of the evolution equation, we devise a derivative EILBGK scheme for direct calculation of flow quantities which have been averaged over the small spatial dimension of, for example, a confining duct (that is depth-averaged (z-averaged) velocity and pressure fields). Centrally, the scheme we devise is amenable to pressure boundary conditions. As part of the process, the need properly to calibrate the means by which friction is introduced in the flow through noslip conditions is highlighted. We also demonstrate analytically and from our simulation results how properly to calibrate the lattice evolution to obtain the correct influence from the unmodelled (z) dimension when the scheme is applied to depth-averaged flow. We present our conclusions in Section 7. 2. Incompressible depth-averaged flow Throughout Section 2 Greek subscripts refer only to coordinates x and y in the explicitly modelled plane (Fig. 1), d denotes physical fluid density, and all other symbols have their usual meaning. Taking isothermal, laminar flow in the shallow duct of constant depth depicted in Fig. 1 to be governed by the continuity and incompressible NS equations and writing the inertial derivatives in the latter in conservation form, one may easily perform a partial (z)-integration to yield: vzva
z

Downloaded from imaman.oxfordjournals.org at Universitatea Transilvania, Braşov on February 8, 2011

+ ^-(.Blva\l=Lz-alva\l=Lo)
3 -I\Ll

(i)

\Ll

where the overbar denotes a depth average


_ i ,Lt

va(xyz)dz.
Lz JO

(3)

Invoking no-slip boundary conditions on velocity in equations (1) and (2) a pseudo 2D system is formed, in which, controlling the depth-averaged quantities, are the incompressible continuity equation exactly and a momentum equation after the incompressible NavierStokes augmented by a momentum sink. Importantly, in the latter, the inertial derivatives take a modified form:

v^=
where
Sa = V-

~^M 2

TSa
(5)

dva 3z

v-

a? z=0

232

D. M. WHITE ETAL.

Defining Cap -T^Vp(4) may be written v^Vp = da ( - j dpCap +-?-. (7)


Downloaded from imaman.oxfordjournals.org at Universitatea Transilvania, Braşov on February 8, 2011

Va~Vp,

(6)

Evidently the depth-averaged velocity and pressure fields are governed by the equations of 2D fluid flow with an effective body force of the particular form dpCap + j (8)

in which expression (8) the second stress term represents the influence of the unmodelled surfaces, z = constant, the first cumulant term an inertial discrepancy. By considering the case of unidirectional flow parallel to the ^-direction (Fig. 1), characterized by a peak (central) velocity MO. it is possible to obtain order of magnitude estimates for the two terms in (8) 1 1 0
v

Sa ~ V = T --

Ly

from which a comparison of the stress and cumulant terms follows


Sg/Lz VUO Ly _ 1 A, (9)

where Re = Lzuo/v is calculated from the unmodelled depth, and A = Ly/Lz is the aspect ratio of the duct cross-section. Supported by the observations of the next section, it becomes clear from (9) that, for ducts of sufficiently large aspect ratio, the effective body force impressed on the notionally 2D fluid may be approximated by the stress term Sa of expression (8) alone. Because this observation is central to the subsequent development of our EILBGK simulation, it is further investigated in the next section. 3. Calculations on steady flow in a duct To develop the assumptions of Section 2 we consider here numerical and analytical calculations of steady-state incompressible flow in a wide, shallow duct. In unidirectional, steady flow along the *-direction of a long duct with geometry of Fig. 1, general arguments allow the velocity and pressure fields to be written vx = vx(y, z) and p(x) po Mx. Consequently, the momentum equation reduces to a Poisson equation: M = , (10)

APPLICATION OF LBGK METHOD TO INTERNAL FLOW

233

1400

1200

1000 <t 800

Downloaded from imaman.oxfordjournals.org at Universitatea Transilvania, Bra&#351;ov on February 8, 2011

600

200

260

FIG. 2. Dimeasionless ratio ft (,Sx/L{)/(dpCzp) of stress term to cumulant term obtained by finite-difference calculations of z-averaged flow in the duct of Fig. 1.

the continuity equation being identically satisfied. Approximating the second derivatives in (10) with central differences evaluated on a rectangular yz-mesh and applying the usual no-slip condition on vx, numerical solutions of (10) were obtained for a range of A and Re values, the latter being controlled through the size of the pressure gradient M inducing the flow. Depth-averaged quantities were then calculated from trapezium rule integration, and derivatives were evaluated from mesh differences. In this way, values for the cumulant and stress terms, Sa/Lz and dpCap in (8) were obtained. Figure 2 shows the y-averaged value of their ratio, R, estimated in (9) plotted as a function of A/Re, with a clear linear trend emerging for the range of 80 s A < 1400 and 7 < Re ^ 35 000 used. These % results support the view that, for large values of A/Re, the artifice of a fictitious body force is sufficient for flow to be treated in two dimensions, the influence of the formally unmodelled surfaces manifesting itself in an effective body force impressed throughout the 2D fluid and represented by stress term Sa of expression (8) alone. Whilst, for ducts in which A is sufficiently large, we aim to simulate in two dimensions, we consider the parent 3D flow, sampled across the unmodelled z-direction and take it to exhibit that parabolic velocity profile which is a solution of the incompressible NavierStokes equation for flow parallel to two parallel-plane walls:
4UQ

= -JYZ(,LZ
L

-Z),

(11)

where, of course, MO depends upon the applied pressure gradient (Landau & Lifshitz, 1966). One may now obtain form (11) expressions for the derivatives in definition (5): dvx
i

4vuo
z=0

(12)

dz z=Lz

234

D. M. WHITE ETAL

In equation (12) o may be eliminated in favour of z-averaged velocity, obtained from (11), whereupon we obtain from (4) and (5) the equations governing pseudo 2D flow in a shallow duct of large aspect ratio at steady-state, 3pvp = 0 (13)

Downloaded from imaman.oxfordjournals.org at Universitatea Transilvania, Bra&#351;ov on February 8, 2011

where, we remind the reader, an overbar denotes averaging over the shallow z direction, Greek subscripts refer only to x, y (Fig. 1) and the de facto velocity dependent body force in the right-hand side has its origin in the stresses from the unmodelled surfaces. The latter, we note, might be modified to account for the influence of turbulence in the flow, but for the present we confine our attention to the laminar regime and proceed, in the next sections, to devise appropriate modifications to an EELBGK scheme which move its governing macroscopic equations towards equations (13) and (14).

4. Modified LBGK simulation Correspondence between the behaviour of the standard LBGK macroscopic observables (15) )ci (16)

and those of viscous incompressible hydrodynamics is well publicized (see, for example, Qian et al., 1992) and the fuller account in the appendix of Hou et al. (1995)) and arises as a consequence of imposing on /j (x, t) a process of relaxation towards a local equilibrium. The local equilibrium density is designed to recover isotropy, mass conservation and Galilean invariance in the equations governing the macroscopic quantities defined in (15) and (16), which are (to O( 3 )): d,p + dapua = 0 p dtPUa + Updppiia = da + v(co)dp(dppua + dapup), where 1/2 (17) (18)

and all symbols have their usual meaning (Qian et al, 1992). To develop an isomorphism between equations (17) and (18) and the continuity (incompressible Navier-Stokes) equation, it is usual to neglect variation of p in all but

APPLICATION OF LBGK METHOD TO INTERNAL FLOW

235

the pressure term of (19) and to map d-*p, p-+\p, ->. (20) (21) (22) (23)
Downloaded from imaman.oxfordjournals.org at Universitatea Transilvania, Bra&#351;ov on February 8, 2011

KH-

where d is the density, v the kinematic viscosity, p the pressure and v the velocity of a physical liquid fluid flowing in equivalent geometry. These isomorphisms entail an equation of state p = p/3 for the lattice fluid, whence lattice fluid speed of sound c, = (dp/dp)1/2 - 1/V3 is seen to be no greater than O(l). Now, taking the incompressible limit in hydrodynamics decouples liquid density d from pressure p by approximating 3d/dp = 0: valid for cs = (dp/dd)i large, Mach number M = V/cs small. With cs large the significant pressure changes occuring in faster duct flows may be reconciled with negligible density changes, and traditional CFD successfully computes pressure-driven athermal flows up to intermediate Re, in the incompressible limit. The same range of Re is not accessible to standard LBGK applied to the problem considered in this paper, however. Compressibility effects are introduced into the lattice velocity field through the steadystate form of its continuity equation (17):

from which we see that the lattice velocity field is divergence-free only as long as may be assumed small. cs is C(l) and p = l/3p, so this requirement entails confining the pressure gradient, leading to low velocity and low Re. On lattices of approximately 40 x 1000 (geometry of Fig. 3) compressibility error was observable (as diverging stream function contours well upwind of the duct expansion) by Re = 25 or so (certainly by Re = 73; Fig. 4c below); for pressure-driven flow the problem is essentially that a pressure gradient must be modelled, in standard LBGK, by a gradient in p, the magnitude of which, when restricted to limit compressibility effects, inhibits the Re obtainable with a given size of lattice. We proceed now to describe the variant in which Zou et al. (1995) and Lin et al. (1996) adapt the LBGK method, making it more strongly incompressible, by identifying ), (24) (25) which new definition of velocity is considered in a recent book by Rothman & Zaleski (1997). For the present work the //(*, t) collide and propagate on the square two-speed D2Q9 lattice (Qian et al, 1992) (Table 1) and are governed by a minimal relaxation form

236

D. M. WHITE ETAL. 600

30 /

Downloaded from imaman.oxfordjournals.org at Universitatea Transilvania, Bra&#351;ov on February 8, 2011

FIG. 3. The backward-facing step geometry simulated. The step height H is one third of the channel width W. Parabolic profiles were imposed using equilibrium forcing (Section 6.2) at inlet and outlet. On all other boundaries v = 0 was imposed using wall collision steps calibrated to give the correct friction factor / * , using flow in an infinite aspect ratio duct (Section 6.3). As indicated in the diagram, the lattice used to resolve the geometry was 600 x 30 sites.

of evolution (26) (27) in which co (equivalent to 1/r of other authors) is a scalar, and fi(v) is the equilibrium density, considered below. That governing macrodynamical equations for (24) and (25) be isotropic, Galilean invariant, and result from update rules on the fi(x_, t) in which local mass and momentum conservation and relaxation are implicit, prove sufficient to close a modified LBGK scheme onto steady-state hydrodynamics if an equilibrium density fi(v_, t): = fi(0)(p + 3v c,- \{v(28)

be used (Zou et ai, 1995; Lin et al., 1996; White et al., 1999). In (28), as in the standard LBGK (Qian et al., 1992), we have (relative to Table 1): \ /HO) = 5 fi(0) = ^ for for for =9 i even / odd (rest), (speed 1), (speed 2).

Modified scheme macrodynamics at steady state is Zou et al. (1995); Lin et al. (1996): (29) 3ava = 0, (30)

where v(w) is given by (19) again. Note that both the continuity equation and the advective term in the lattice momentum equation (29) now become exactly incompressible, whilst the form for the diffusive term is preserved, but at the cost of the time derivative (omitted from (29), the correct (hydrodynamic) form of which is lost. The modified scheme is thus

APPLICATION OF LBGK METHOD TO INTERNAL FLOW 32400

237

Downloaded from imaman.oxfordjournals.org at Universitatea Transilvania, Bra&#351;ov on February 8, 2011

40000

100

200

300 400 Position along channel x

500

600

FIG. 4. (a) Total LBGK mass M(t) = J^^fi{x_,i) as a function of time for the geometry of Fig. 1; see Section 6 for stimulation detail, (b) Lattice density p plotted as a function of distance, x, along an infinite aspect ratio duct, forced with an open boundary condition outlet; solid (dotted) line represents p in the modified (usual) LBGK scheme.

238
0.154 (c) 0.152 0.150 0.148 0.146 0.144 0.142 0.140

D. M. WHITE ETAL.

Old Scheme (RE - 78) New Scheme (RE = 73)

Downloaded from imaman.oxfordjournals.org at Universitatea Transilvania, Bra&#351;ov on February 8, 2011

__.

I t

0.138 0.136 0.134 0.132 0.130 0.128 0.126 0


1 1

. , '

.**
,.'

100

200

300 400 Position along channel x

500

600

FIG. 4. (c) Average LBGK velocity v(x) m f0 v(y, x) d> plotted as a function of distance, x, along the channel. Solid (dotted) line represents, v in the modified (usual) LBGK scheme. (Note that no origin in vertical direction is shown).

TABLE 1 Square two speed D2Q9 lattice i 1 2 3 4 5 6 7 8 9 (rest)


Cix CiV
c

-1 0 1 1 1 0 -1 -1 NA

1 1 1 0 -1 -1 -1 0 NA

i 2 1 2 1 2 1 2 1 0

APPLICATION OF LBGK METHOD TO INTERNAL FLOW

239

valid only at steady state (Zou et ai, 1995). However, within the steady state, it is now necessary to make no approximations to obtain viscous incompressible hydrodynamics: LBGK p maps onto hydrodynamic p/d without any implied association between p and d (real fluid density), full mappings being 7 - P, a V.^v, (31) (32)
Downloaded from imaman.oxfordjournals.org at Universitatea Transilvania, Bra&#351;ov on February 8, 2011

and (23) again. It is appropriate to emphasize that moment (25) yields flow velocity and that quantity is now exactly divergence free (note 30). From (25) and Stokes' theorem / fs v dS = 0, whence the flux of velocity is zero. If, over the surface of some volume, pressure (LBGK p) varies then / fs pvdS_^ 0. Thus, in pressure-driven flow, product pv will vary along even the straight sections of any duct (Fig. 4). Standard LBGK of course conserves momentum in this situation by progressively increasing upwind velocity (Fig. 1) as the pressure (LBGK density) decays, producing, as we have already remarked, compressibility errors even at Re 25. We return to this point in Section 6.2. We proceed in Section 5 to adapt this exactly incompressible scheme to the computation of depth-averaged flow. 5. EILBGK scheme for simulating internal depth-averaged flows To modify this 2D scheme in a manner calculated to generate additional body force terms in the macrodynamic momentum equation (17), isomorphic with the additional terms in (14), we write fi(x+ci,t+l) Centrally Sfi is taken to be = fi(x,t) + tli+8fi. (33)

MYljCja,
j

(34)

which, on appeal to the cia values for D2Q9 (Table 1) may be shown to have moments
Vi

(36)
Vi

Vi

Higher moments of Sfi (up to the fourth) are zero. With these properties, our model's macrodynamics can generate, on using the usual multiscale (Chapman-Enskog) expansion methods at O(e) in steady-state, a modified Euler equation: = -da- + -va. (38)

240

D. M. WHITE ETAL

With the Sfi terms contributing no further at C(e 2 ) scales, the derivation of the final macroequations governing our particular EILBGK scheme in the steady state is straightforward, resulting in a form wherein the isomorphism with (14) is obvious:
Vf,dpva

= - 3 O ! -p 1 + v(io)(3p8pva + dadpvp) + -va,


daVa = 0,

(39)
(40)
Downloaded from imaman.oxfordjournals.org at Universitatea Transilvania, Bra&#351;ov on February 8, 2011

in which v(a>) is given by (19) again. This EILBGKs continuity equation and its momentum equations' advective term are now seen to be exact, whilst the form for the diffusive term in the latter is preserved (by appropriate choice of the equilibrium distribution function (equation (28)) but, we stress again, at the cost of the time derivatives, the correct (hydrodynamic) form of which is lost (39). Of more immediate importance, its macrodynamics are (overall) seen to be isomorphic with depth-averaged flow, with the implied association (41) whence our depth parameter k introduced into (33) through (34) is seen to embody the means by which the unmodelled depth of the duct is input. Using the EILBGKs v we obtain, in lattice units (42) Used in conjunction with an exactly incompressible scheme, depth parameter k may be shown to stand in correct relation to lattice pressure (density) variations, as we shall show in subsequent sections.

6. Simulation using first an infinite aspect ratio, then depth-averaged EILBGK schemes Here we describe the forcing and calibration employed in our simulations of the EILBGK scheme introduced in Section 4 and present results. The forcing applied over what we shall term 'open boundaries' allows the simulator to gain fuller control of the fluid pressure field and avoids the use of the more usual device of a uniformly impressed body force (White et al., 1999). We adapt the idea to produce for LBGK simulation a means of bounding the velocity field analagous to what traditional CFD terms 'upwinding'. The particular means by which these boundary conditions are implemented, we believe, are applicable to EILBGK simulation in general (see White et al, 1999) and clearly bring it closer to CFD practice by allowing for calculation of the flow's pressure drops, that is to say, the approach is especially productive when the principal boundary conditions (describing the agency generating flow) are given in terms of p alone.

APPLICATION OF LBGK METHOD TO INTERNAL FLOW

241

6.1 Reynolds number One can normalize steady-state equation (29) to characteristic lattice velocity U* and length L* (lattice units): Vfi8fiVa = -dag(p) + dpdpVc He in which all velocities (lengths) are now dimensionless by U*(L*) and U*L*
R

(43)

Downloaded from imaman.oxfordjournals.org at Universitatea Transilvania, Bra&#351;ov on February 8, 2011

3~1?-

(45)

Thus in LBGK simulations, the relative importance of the advective and diffusive terms is measured Re defined in a manner precisely analagous to viscous incompressible hydrodynamics. In an LBGK simulation, lattice size (and possibly time-step) determine a value of L*, boundary conditions and forcing (see below) influence U* and the value of co in use sets v(co), to give from (44), our characterizing lattice fluid Re. 6.2 Modelled walls In this section we describe the means by which explicitly modelled walls are represented in our simulations. A decoupled lattice density allows a simulator freedom to force target boundary node velocities v through the corresponding equilibrium densities fi(v), without undermining momentum conservation considerations. Hence, certain boundary pressures may be allowed freely to develop into values consistent with the other boundary conditions (e.g. no-slip Dirichlet conditions on t>: see below) and the elliptic nature of the system of equations effectively under solution. For example, in the back-facing step geometry used (Fig. 3), having chosen an inlet pressure, inlet and outlet velocity profiles, and to apply Dirichlet conditions on all other channel walls, our calculation of the flow avoids specification of outlet pressure. Such boundaries we designate 'open'. They are easily represented in the modified scheme where lattice site density influences pressure and not lattice momentum. We thus 'force' target boundary velocities, v (consistent with actual mass conservation), by continually resetting link densities at boundary position * with / (u), i.e. to a value close to that consistent with the desired boundary velocity. Whilst it is the equilibrium densities /-(u). not the fi(x_, t) terms, which are used to force in this way, to a given v, this mode of forcing flow drove our system towards a parabolic profile based upon the fi terms with considerably improved spatial efficiency. Notwithstanding, it is the case that a parabolic profile is not set on the LBGK lattice extremities in the same sense as Dirichlet conditions are set on velocities in CFD. We consider a back-facing step of height H in a channel of height W. Parabolic flow at the inlet (height (W - H)) and outlet (height W) of the expanded duct of Fig. 3 was forced after the manner described above by overwriting, at every step, link densities at duct inlet sites (on the lattice edge) to values consistent with a parabolic profile with peak

242

D. M. WHITE ETAL

Downloaded from imaman.oxfordjournals.org at Universitatea Transilvania, Bra&#351;ov on February 8, 2011

FIG 5. The infinite aspect ratio duct used to calibrate boundary conditions and pressure changes along the duct

(characteristic) lattice velocity


Vr
=

W2-2HW + H2

(y - W)(y - H)

(46)

with node density p = 1-8, and to the matched profile:


(47)

with unspecified density at the outlet boundary sites. Whilst continuity requires that inlet and outlet profile maxima are related through vo = \v'Q, outlet sites' pressures were allowed to evolve towards a steady value by setting their fi(v) using the average value of p propagated onto the outlet boundary sites at the previous time-step. Thus an 'upwinded' steady-state outlet density (pressure) develops to a measurable value determined by inlet conditions and the other boundaries in the system. In this way, after appropriate calibration (see Section 6.3), the pressurefieldresulting from a particular set of inlet driving conditions was obtained. Following our comments at the end of Section 4 we note that, with an open boundary condition, p decays along the simulation (consistent with the expected reduction in hydrostatic pressure), and the simulation mass must be allowed to 'equilibrate' to an asymptotic steady-state value (Fig. 4a). Figure 4b shows a typical variation of p along a uniform channel of infinite aspect ratio (Fig. 5), for both the new scheme and, for comparison, a similar variation obtained from a forced boundary interpretation of the usual scheme at similar Re. In Fig. 4c, we see the associated variations, along the channel, of average velocity v. For the new scheme the latter is constant, whereas in the standard scheme continuity couples density changes to changes in velocity, leading to compressibility effects which, again, enter into simulations even at low Re, making compressibility effects the primary determinant of accessible Re values in LBGK hydrodynamics applied in the usual manner to the present problem. To impose set pressure drops on the system at open boundaries we have also implemented what we term free profile conditions. Again, at the inlet, both a pressure and velocity profile are specified. However, with free profile conditions, a uniform pressure (lattice density) is set at the outlet, by the user, with the velocity v from the node downstream now used to set all boundary densities to equilibrium value / , (u, p). Allowing outlet velocity freely to develop, after the 'upwinding' technique of traditional CFD,

APPLICATION OF LBGK METHOD TO INTERNAL FLOW

243

applies we believe 'fully-developed flow' Von-Neumann type boundary conditions to the LBGK lattice extremities (White et at., 1999). Having dealt with open boundaries, we proceed to relate the means by which the noslip boundary conditions required for the ducts' static, resolved wall were calibrated with reference to flow friction factor. Indeed, working within an exactly incompressible scheme, the flow friction factor also provides a useful verification of the influence of the unmodelled walls, i.e. of depth parameter it.
Downloaded from imaman.oxfordjournals.org at Universitatea Transilvania, Bra&#351;ov on February 8, 2011

6.3

Friction factor

In the case of actual fluids, the friction factor / relates average flow profile velocity V to the pressure Ap developed along a duct and the usual fluid transport parameters. One can write for an infinite aspect ratio uniform duct of height Wo (hydraulic diameter 2 Wo) and length Lo (Fig. 5), for which V = \ Vo, (48)

a which, for infinite aspect ratio, reduces to

L.Q

/=,
with Re now based upon hydraulic diameter 2 Wo, i.e.

(49)

(50)

The LBGK friction factor / * provides a means by which 'no-slip' Dirichlet boundary conditions on velocity were calibrated. / * was measured by driving flow in an infinite aspect ratio straight duct, using forcing to matched parabolic velocity profiles at the inlet and outlet (Section 6.2). No-slip or Dirichlet conditions, applied along the sides of the duct and step ('wall sites') were originally imposed in the manner most common in the literature, using a boundary-specific collision step in which all densities on links with a component perpendicular to the line of the boundary are 'bounced-back' (have their propagation direction rotated through n rads), thereby introducing friction into the layer of fluid in contact with the wall. Symmetry alone implies a pseudoparabolic flow profile in the present geometry (Fig. 3) and a more stringent test of the boundary conditions lies with the stresses communicated to the wall by the fluid, measured by the friction factor, which is determined from the pressure gradient in the duct. Initially this took a value different from that suggested by equation (49). However, the precise value of / * may be improved by diluting the amount of bounce-back applied to densities on wall nodes with specular reflection, to obtain the correct pressure drop along the duct. Adapting equations (48) and (49) to modified LBGK simulations, in which Ap/d is determined by j Ap, (the total density drop along a full channel) we obtain a lattice version

244

D. M. WHITE ETAL

of equation (48):

in which v is the height-averaged lattice fluid velocity (see equation (25)) and LQ and Wo are now measured in lattice units. Note that the (modelled) width of the duct, Wo, is not exactly the geometrical width of the duct but is subject to correction, because such rules as we describe here for applying boundaries do not, it has been shown (Cornubert et al., 1991) and later (Ginzbourg & Adler, 1994)), locate a zero velocity on the node in which the wall rules act, rather at a finite distance outside that node. For current purposes, the geometrical width of calibration channels was by trial and error made sufficiently large that any error incurred in Wo was not measurable using the techniques we apply here. For our particular lattice orientation, short (evenly subscripted) lattice links lay either perpendicular or parallel to the simulation walls (see Figs 3 and 5) and were accordingly reflected or left unadjusted. For out revised wall collision step, however, densities on long (odd) links ,- (with both components perpendicular and parallel to the wall nonzero) were coded partly to reflect specularly (deflect through 7r/2 radians) and partly to reverse (reflect through n radians) in a variable ratio b. A finite fraction of bounce-back, b, is, in principle, sufficient to introduce frictional stresses and hence a rest layer parallel to wall sites (static boundaries). The correct value of b was, however, only determined after several sets of measurements of / * (obtained for given b using (51) for different Re. Research initiated by d'Humieres and co-workers (Cornubert et al, 1991) shows that this general means of implementing no-slip boundary conditions (bouncing-back) will not locate the wall at the boundary node, rather a fraction of a lattice vector off-lattice. Therefore care was exercized to ensure that all ducts simulated were of sufficient width to render this indeterminacy in width of negligible significance to our results. One expects a plot of / * versus 1 /Re* to have a gradient controlled by bounce-back fraction b, and by plotting this gradient, designated m(b), over a range of b (Fig. 6) it is possible to determine that value of b necessary to obtain a correct gradient m(b) of 96. Note that the value of collision scalar w in use influences the value of b required to give the correct / * . Whilst to obtain a consistent pressure drop along the system requires the use of a particular bounce-back fraction, the velocity field was observed to be less sensitive to b. Considering the pattern of flow generated in the backward-facing step (Section 6.4), varying b between zero and unity produced, in qualitative terms, little effect on the stream function \j/{z, y),

Downloaded from imaman.oxfordjournals.org at Universitatea Transilvania, Bra&#351;ov on February 8, 2011

iK*.y)= f yvx(x,y')dy', Jy=o

(53)

and attempts to measure its influence on the reattachment length also showed b to have little effect. Variations in the value of parameter b were observed to have a marked effect upon values of the total pressure drop along a given section of the system. The pressure

APPLICATION OF LBGK METHOD TO INTERNAL FLOW

245

I
3

Downloaded from imaman.oxfordjournals.org at Universitatea Transilvania, Bra&#351;ov on February 8, 2011

20

30

40 50 60 Percentage bounce-back (>)

70

S O

90

100

FIG. 6. Fraction b of link densities not reflecting specularly at wall sites and the corresponding gradient m (b) of the plot of lattice friction factor / * versus lattice Reynolds number Re*. The correct value of m(b) to correspond with hydrodynamics is 96. Data corresponding to collision parameters a> = 1-3, 1-5 and 1 -9 are shown

field corresponding to a simulation of flow past the expansion is shown in Fig. 3. The nonuniformity of the pressure field is clear, and its qualitative features were stable. As we have seen in Section 3, the profile flow sampled down the unmodelled z-direction is parabolic, so
1

' d

Lo

v2'

(54)

which, of course, derivatives directly from the definition of / for flow along an infinitely deep duct with width Lz and length Lo (White et al., 1999). Using standard results for this flow, (54) may be reduced to a well-known result.
J

Re

(55)

where Re is based upon the hydraulic diameter 2LZ, i.e. (56) Having used / * in this way for the collision rule in use on the explicitly modelled, y = constant (Fig. 1) walls, the value of b was retained for use in all simulations of depthaveraged flow. The friction factor may also be used as a check on the simulation's behaviour in the unmodelled direction (z-direction, Fig. 1) and as a check upon our interpretation of the role of parameter k.

246

D. M. WHITE ETAL

Consider the steady-state of a 2D lattice of length LQ in the x-direction, infinite in the y-direction, driven with uniform, matched inlet and outlet profiles VQ = VQX. Let the inlet density be p\n and consider the outlet sites' densities to have evolved to a steady-state density pout in the manner described in Section 6.2. On general arguments of symmetry va(x, y) = rty&ax, An Pout oxP = : (57) ,,-, (58)
Downloaded from imaman.oxfordjournals.org at Universitatea Transilvania, Bra&#351;ov on February 8, 2011

For this situation all link densities will be given by their equilibrium form (28) so, in modified evolution equation (33), we may set i?,- = 0 and, with (57) and (58) in mind, obtain MVQ, p(x + cix)) = Mvo, p(x)) + kfMcixVx(59)

Substituting, from (28) for /-(wo. PC* + cix)) and fi(vo, p(x)) we note that, because the velocity field is uniform (57) and in the EILBGK scheme, variations in the lattice density are not communicated into site velocities, all terms involving UQ cancel and after a Taylor expansion and some algebra (59) yields ciadap =kciava, which, for the present geometry, reduces to the exact result An Pout 4p , : = =kvo.
LQ LQ

(60)

,,,. (61)

Equations (61) and (42) may be substituted into (51) to obtain reassurance of the result required by hydrodynamics:

/-:
here, Re* = 2VQLZ/VW is based upon the unmodelled depth of the parent 3D duct. Using periodic boundary conditions, it is straightforward to simulate this semi-infinite lattice, and by measuring the gradient in lattice density p along the flow direction, to calculate / * from (51) with Lz obtained from the set value of k through (42). Simulations were performed for values of unmodelled depth parameter k in 00008 < k < 0 0 1 , each being driven with a uniform velocity profile ranging 0023 ^ VQ ^ 0-271 in lattice units. The average velocity u was confirmed to be of constant value throughout each simulation, 40000 time-steps being sufficient easily to ensure a final steady state. Across all tests performed, w was set to 1-9 and pressure drops of the order 3-7 x 10~' were recorded. For a particular k value the lattice friction factor, / * , and Re were calculated for each velocity profile and the values plotted against each other. Over the range of parameters considered, the gradient m{b) was measured to be very close to 96. Figure 7 shows the depth-averaged velocity sampled across a long, uniform duct calculated using modified EILBGK simulation (open symbols) compared with that calculated numerically after Section 3 (continuous line) for the system of Fig. 2. The close correspondence between these two sets of results further underlines the validity of the EILBGK model.

APPLICATION OF LBGK METHOD TO INTERNAL FLOW


Y(EILBGK) 20 i| 30 M'

247

Downloaded from imaman.oxfordjournals.org at Universitatea Transilvania, Bra&#351;ov on February 8, 2011

HLBQK

300 Y (numerical)

FIG. 7. The continuous line shows the normalized variation of the z-averaged velocity across the duct of Fig. 1 calculated from the finite difference scheme of Section 3. Symbols show the equivalent normalized quantity obtained from calculations in the modified E1LBGK scheme for equivalent Reynolds number and effective Lz (k parameter).

6.4

Comparison with experiment for backward-facing step geometry

A calibration to assess maximum accessible Re was effected by fixing relaxation parameter to and increasing v through parameter VQ (see Section 6.2) for a set system size. Compressibility effects were determined to be small from an inspection of the velocity field (in the form of a contour plot of stream function V (*, y)) in the outlet of the channel (well behind the step) (Fig. 8). In the new scheme, \ff (x, y) is a set of parallel lines in this region and thus it may be inferred that the velocity field shows no evidence of systematic increase in v (Fig 4c). Moreover, the pressure field (Fig. 8) is fully consistent with uniform decay in that quantity. This was observed to remain the case for all values of inlet profile parameter urj up to that which induced instability in the LBGK simulation (Section 6.2). A system of size 600 x 30 was driven with matched equilibrium forcing (Section 6.2) at inlet and outlet, using a value of co 1-9. The value of b in use was determined on the basis of the value of /*, measured for flow in an infinite aspect ratio duct (Fig. 5), again with a) = 1-9. For purposes of comparison with experiment, the length of the duct outlet down-stream of the step (Fig. 3) was determined by trial and error, to be sufficiently long such that it had no measurable influence upon the reattachment length L of the recirculating region behind the expansion (Fig. 3). L was measured in units of the step height, as the distance upwind of the step to an interpolated zero in vx. Figure 9 shows dimensionless L/H as a function of Re. These data were obtained from the final steady state of simulations using a) = 1-9, initialized uniformly to ft = poft(O) (see after equation (28)), with a node density of po = 1-8, equilibrated for 40000 time-steps (although flows were observed to have equilibrated by 25 000 steps). The stability of the scheme was checked in the manner

248

D. M. WHITE ETAL

(a)

(b)

Downloaded from imaman.oxfordjournals.org at Universitatea Transilvania, Bra&#351;ov on February 8, 2011

FIG. 8. (a) Pressure field produced from LBGK simulation of geometry shown in Fig. 2. The extent and nonuniformity of the lattice fluid pressure variations is readily apparent (b) Stream function ir(x, y) (see equation (53)) evaluated numerically (using a trapezium rule integration) for the velocity field vx(xy). It can be seen from these streamlines that PoiseuiUe flow redevelops downstream of the step, (c) Detail of recirculating region directly behind the step. Note that in (a) the end of the step is shown; in both (b) and (c) the stream function is shown immediately right of the step.

discussed at the beginning of Section 6. Figure 4a shows the time variation of the total lattice mass M(t): /( 0 I

(63)

The simulation results in Fig. 9 are overlaid with experimental data lifted from pseudo 2D experimental measurements made in the central plane of the apparatus represented in Denham & Patrick (1974) Fig. 2, and with comparable calculations made using standard CFD (Young, 1996). Whilst agreement with CFD is good, that with the experimental data points (admittedly obtained crudely from enlargements of Fig. 7a of Denham & Patrick (1974)) shows a correspondence which is clearly less close (certainly compared with the match achieved using the usual scheme) against theory on lid-driven cell geometries (Miller, 1995; Hou et ai, 1995). This disparity possibly originates from the comparatively large pressure gradients present in our simulation, which are less marked in confined and mechanically-driven lid-driven cell geometries. The latter essentially experience restricted pressure changes because of the confinement of the fluid. 6.5 Depth-averaged flow results

We choose as an illustrative example, flow near a bifurcation in a shallow duct (Fig. 10), concentrating on the qualitative dependance of the deflection of the principal flow upon driving pressure. With Dirichlet conditions on velocity at the resolved walls, values of inlet velocity and k were chosen so that A/Re was large. Figure 11 shows by means of the stream function (see below) the flow profile obtained in a bifurcating duct (see Fig. 10), and Fig. 12 shows the associated pressure field. Open boundary conditions were imposed on the pressure at

APPLICATION OF LBGK METHOD TO INTERNAL FLOW 10


.

249

9 1 . 1

r 8
3

I
+

7 -

I
+

Downloaded from imaman.oxfordjournals.org at Universitatea Transilvania, Bra&#351;ov on February 8, 2011

6 5 -

"1
\
LBGK data CFD results \ Denham and Patrick data
. 1 . . , . 1

I
4 v

| 50
75

1 .

100

125 150 Reynolds number (Re)

175

200

FIG. 9. Comparison of reattachment length data obtained from LBGK backward-facing step simulations with experimental data (Ginzbourg & Adler, 1994).

45

P2

PI

P3

45

230
FIG. 10. Diagram showing geometry of bifurcating duct

250

D. M. WHITE ETAL.

Downloaded from imaman.oxfordjournals.org at Universitatea Transilvania, Bra&#351;ov on February 8, 2011

(c)

FIG. 11. Stream lines obtained from EILBGK calculation for laminar flow in a shallow bifurcating duct, where AI Re 50. (a) Inlet pressure higher thanright-handexit, which is higher than top exit; majority of flow deflected towards top exit some allowed to flow throught to right exit, (b) On increasing right-hand pressure only very negligable flow through to right exit (c) On further increase in pressure atright-handside, flow now enters here as well as inlet

both outlets, with Dirichlet conditions on the velocities corresponding to uniform profiles. Also, simulations were performed using our free profile conditions, where uniform profiles were set at the inlet but developed naturally at both outlets. Both approaches to setting the boundary conditions yielded qualitatively similar results, but only those from using the latter boundary conditions are presented here. The length of both outlet channels was increased until the qualitative feature of the flow in the region of the bifurcation was seen to be insensitive to any further elongation. The parameters used in simulation were w = 1-9, k = 0-3 (therefore Lz & 10) and Ly = 45. The pressures were set to P I = 0-60 and P2 = 0-48, and P3 was gradually increased using the values 0-514,0-528 and 0-540, until the flow direction was reversed in the horizontal exit

APPLICATION OF LBGK METHOD TO INTERNAL FLOW

251

Downloaded from imaman.oxfordjournals.org at Universitatea Transilvania, Bra&#351;ov on February 8, 2011

FIG. 12 Pressure contours corresponding to Fig. 4. Note the uniformity of the pressure gradient in the inlet and outlet.

(i

The pressure field corresponding to the flow close to the bifurcation (Fig. 12) is nonuniform. The extent to which the pressure varies throughout the simulated system is clear, and whilst the qualitative features of this field are stable, variations in the value of wall collision parameter b were observed to have a marked effect upon numerical values, underlining the importance of calibrating the modelled wall's collision step. Analagous results to those presented in Fig. 11, based upon modification of Section 5 transplanted into the usual scheme, would show, for modest Re values, ^(x, y) contours upstream of the bifurcation progressively diverging into the modelled walls, as continuity requirements drive the lattice fluid's velocity to increases to compensate for the reduction in lattice density (pressure), underlining the need to use an exactly incompressible scheme in pressure-driven flow applications.

252

D. M. WHITE ETAL

For the purposes of comparison, a 2D model of the laminar flow within the T-junction was set up using version 1.6 of the PHEONICS flow modelling package. The flow domain was represented using a nonuniform 60 x 40 calculational grid, in which the solid regions adjacent to the vertical leg of the T-piece were 'blocked off' by defining them as regions of zero porosity. Pressure boundary conditions were defined at the inlet and exit planes of the T-junction, consistent with those assumed in the accompanying analysis. As a consequence of the 2D treatment used in the work, the principal flow resistances arising from the shear stresses at the unmodelled upper and lower faces of the flow domain could not be modelled directly. Therefore it was necessary to represent this effect through the use of additional momentum sinks applied to each of the velocity components distributed throughout the flow domain. The form of the fluid momentum sinks used in this work was based on the following empirical pressure loss definition:

Downloaded from imaman.oxfordjournals.org at Universitatea Transilvania, Bra&#351;ov on February 8, 2011

For the case of flow in a smooth walled wide aspect ratio duct it can be shown that, in the limit, the friction factor ( / ) may be expressed as

,-
Re =
p|u|D/l

(66)

Using this model described above, the resulting flow field was solved and the predicted streamlines are presented in Fig. 13 for comparison with the accompanying work in Fig. 11. 7. Discussion and conclusions The reinterpretation of velocity according to Zou et al. (1995); Rothman & Zaleski (1997); Lin etal. (1996) ~~ i(x,t)cia (67)

and the revision of equilibrium distribution function to /(") = M0)(p + 3v c. - \v2 + I(v c,-)2) (68)

in the modified LBGK scheme has been shown to suppress compressibility effects which restrict the range of Reynolds numbers accessible to the standard LBGK. When used with boundary conditions designed for a fuller control of the pressure field, we have been able to access Re R 250 (compared with Re 25 with standard LBGK). We have applied modified LBGK simulation to internal, pressure-driven flow in which nonuniform pressure gradients are to be expected, demonstrating with this approach a need to adjust the fraction of bounce-back, b, applied to link densities at no-slip boundaries of LBGK simulations to recover correct mechanical stresses at a static wall. Adjusting b in

APPLICATION OF LBGK METHOD TO INTERNAL FLOW

253

(a)

Downloaded from imaman.oxfordjournals.org at Universitatea Transilvania, Bra&#351;ov on February 8, 2011

(b)

(c)

FIG. 13. Streamlines obtained from CFD calculation of bifurcating problem, using same pressures as specified for LBGK calculation.

this way has allowed the LBGK method to recover a correct value of friction factor / * for flow in an infinite aspect ratio rectangular duct. This value of b significantly influences the measured pressure drops when applying modified LBGK simulation to internal, pressuredriven flow. Use of forced boundary conditions allows development of stable, parabolic boundary profiles in relatively small regions of lattice, and thus introduces an improvement in the efficiency of the method. Forced boundary conditions also have inherent flexibility (being capable of direct application where a boundary velocity profile and pressure are required) and bring LBGK simulation closer to CFD practice. For example, a solenoidal flow field is readily established using a forced boundary condition, but is tedious to apply using usual

254

D. M. WHITE ETAL

traditional LBGK methods. Moreover, we have seen that it is possible to simulate open boundary conditions in a modified scheme with ease and without recourse to continuity considerations, making for simple implementation, which is again more consistent with the CFD practice of upwinding. Comparisons of LBGK simulation with particular analytical solutions of the lid-driven cavity test bench test problem (Miller, 1995) deal with completely confined fluids so that the pressure fields encountered are more homogeneous. This possibly accounts for the closer agreement between theory and CFD that these simulations produce, over that which we report here with experiment. However, the data developed for comparison, from traditional numerical solution of the incompressible Navier-Stokes and continuity equations, is in better agreement with our LBGK results, and whilst for our application the former offers considerably greater efficiency, it appears not to produce significantly better agreement with experiment. The influence of the bounce-back fraction b might be understood in similar terms, for, as we have observed, the necessity to calibrate this quantity really emerges only when computing the pressure fieldoften the quantity of engineering significance. In the work presented here, we have analysed the Navier-Stokes equation and devised for ducts of uniform depth, a condition for the application of laminar, depth-averaged flow modelling over fully-resolved 3D calculation. Modifications to the EILBGK simulation method, mapping the latter onto the structure of a partially-integrated, steady-state Navier-Stokes equation (with new associations for the unmodelled depth of duct, which is controlled through our parameter k) have been presented and shown to promote correct hydrodynamics. It has further been demonstrated, both by analytical calculation and by simulation, that parameter k lies in correct relation with pressure jumps obtained by application of open boundaries to EILBGK scheme applied to internal, pressure-driven flows. We have increased the capacity of LBGK simulation to deal with flows in which the principal (flow inducing) boundary conditions are specified in terms of pressure. In doing so we have brought LB methods closer to CFD practice: adapting the EILBGK method has facilitated open boundaries, allowing us to avoid both the error due to compressibility effects (whilst actually calculating pressure jumps) and the artifice of a uniformly impressed body force to drive flow. Both these benefits acme from the way in which boundary conditions have been applied and not from any modification to the bulk scheme. They are, therefore, quite general and not confined in their utility to depthaveraged flows. In modifying the EILBGK scheme, the possibility of generating a sink term in the continuity equation has been deliberately avoided. Such a term may have proved useful in applications of the method to ducts of nonuniform depth and possibly as a means of modelling shallow water equations. Our model is still laminar. In traditional CFD the influence of turbulence is usually incorporated by some 'law of the wall'. Turbulence in the layer of fluid in contact with the no-slip boundaries in our model would undoubtedly modify the stress terms (5), but (provided the appropriate Reynolds number could be accessed) an appropriate modification is all that obstructs incorporation of some turbulent effects.

Downloaded from imaman.oxfordjournals.org at Universitatea Transilvania, Bra&#351;ov on February 8, 2011

APPLICATION OF LBGK METHOD TO INTERNAL FLOW

255

REFERENCES

BHATNAGAR, P., GROSS, E. P., & KROOK, M. K. 1954 Phys. Rev., 94, 511. CORNUBERT, R., D'HUMIERES, D., & LEVERMORE, D. 1991 Physica D, 47, 241.
DENHAM, M. K. & PATRICK, M. A. 1974 Trans. Inst. Chem. Eng., 52, 361.

GlNZBOURG, I. & ADLER, P. M. 1994/ Phys. II(France), 4, 191-214.


Hou, S., Zou, Q., CHEN, S., DOOLEN, G., & COGLEY, A. C. 1995 J. Stat. Phys., 118, 329.

LANDAU, L. D. & LIFSHITZ, E. M. 1966 Fluid Mechanics. Pergamon Press.


LIN, Z., FANG, H., & TAO, R. 1996 Phys. Rev. E, 54, 6323. MILLER, W. 1995 Phys. Rev. E, 51, 3659.

Downloaded from imaman.oxfordjournals.org at Universitatea Transilvania, Bra&#351;ov on February 8, 2011

QlAN, Y. H., D'HUMIERES, D., & LALLEMAND, P. 1992 Europhys. Lett., 17, 479-484. ROTHMAN, D. H. & ZALESKI, S. 1997 Lattice Gas Cellular Automata. Cambridge University Press.
W H I T E , D. M., HALLIDAY, I., CAVE, C. M., & STEVENS, A. 1999 Physica D, 129, 68-82.

YOUNG, C. 1996 PhD Thesis, Rolls Royce & Associates, Derby.


Zou, Q., Hou, S., CHEN, S., & DOOLEN, G. D. 1 9 9 5 / Stat. P/ryj.,81,319.

You might also like