PHYSICAL REVIEW A VOLUME 43, NUMBER 12 Generalized hydrodynamic transport in lattice-gas automata Li-Shi Luo Center for Nonlinear Studies, Los Alamos National Laboratory, Los Alamos, New Mexico 87545 and School of Physics. Georgia Institue of Technology, Atlanta, Georgia 30332-0430 Hudong Chen Center for Nonlinear Studies, Los Alamos National Laboratory, Los Alamos, New Mexico 87545 ‘and Department of Physics, Dartmouth College, Hanover, New Hampshire 03755 Shiyi Chen Center for Nonlinear Studies, Los Alamos National Laboratory, Los Alamos, New Mexico 87545 ‘and Bartol Research Institute, University of Delaware, Newark, Delaware 19716 Gary D. Doolen and Yee-Chun Lee Center for Nonlinear Studies, Los Alamos National Laboratory, Los Alamos, New Mexico 87545 (Received 14 December 1990) ‘The generalized hydrodynamics of two-dimensional latice-gas automata is solved analytically in the linearized Boltzmann approximation. The dependence of the transport coeficents (kine ‘matic viscosity, bulk viscosity, and sound speed) upon wave number K is abtained analytically. Anisotropy of these coefficients due to the lattice symmetry is studied for the entire range of wave ‘number, k. Boundary effects duc to a finite mean free path (Knudsen layer) are analyzed, and accurate comparisons are made with lattice-gas simulations. 15 JUNE 1991 Since it was shown theoretically that lattice-gas auto- mata (LGA) can simulate the Navier-Stokes equa- tions,'? lattice-gas automata have been extensively used to study many physical problems in the ficld of hydro- dynamics. Major advantages of the lattice-gas method include its parallel nature and the capability of han- dling complicated boundary geometries. One significant achievement of the LGA method is,the accurate simula~ tion of flow through porous media.‘ Here, we calculate analytically the generalized hydrodynamic LGA transport coefficients using the linearized lattice Boltzmann equa- tion The usual hydrodynamics requires significant modifi- cation when the characteristic length scale is the same or- der as the mean free path. A generalization is required in order to include the wave-number dependence of the transport coefficients in such a way that the constitutive relation is preserved.® This generalization produces a nonlocal hydrodynamic response to fluctuations, General- ized hydrodynamics has been studied previously in the context of a hard-sphere fluid.” ‘The significance of studying generalized hydrodynamics in the context of LGA models is justified by the fact that the LGA method has been used to solve difficult physical problems such as flow through porous media," a situation where generalized hydrodynamics ean be important, The purpose of this pa- per is to present analytic results for the wave-number Gependence of transport coefficients in LGA models for the entire range of wave number k and to examine bound ary effects due to a finite mean free path. Previously re- ported transport coefficients results are valid only in the hydrodynamic limit, k—0.* Here, the linearized lattice Boltzmann equation, which is written as an eigenvalue a equation valid for allk, is derived and solved analytically. Before presenting our results, we now briefly describe the lattice-gas model. This many-body latice-gas system resides on a two-dimensional (2D) hexagonal lattice space fx with unit lattice constant and with discrete time 16 (0,1,2,...]. Each particle has unit mass and unit speed.” The six possible velocity directions of particles along the links of the hexagonal lattice are €,=cosl(a—1)2/3]8+sinl(a~1)8/319, ae f2,...,6 (Greek indices a, f, ... always run from 1 to 6 unless specified otherwise.)' There can be either | or 0 particles, with the velocity é, at the same lattice site, x, and time # (exclusion principle). The evolution of the system consists, of two steps: collision and streaming. At each time step (a1=1), particle interactions occur through local col- lisions, i.e., each collision only involves particles at the same site, ‘The collision rules are designed to conserve the exact particle number, momentum, and kinetic energy. Alter collision, particles move in the direction of their ve- locities to adjacent sites. ‘The evolution of the Frisch-Hasslacher-Pomeau (FHP) lattice-gas system is completely determined by the micro scopic equation Nat Eat 41) = MAG +CAN), w where V denotes all N(x.) and the Boolean function CAN) = Ei (55 — 50D TIN#G.0U = Nx” ® 7097 ©1991 The American Physical Society7098 represents creation or annihilation of N,(x,t) due to col- lision for all possible incoming configuration s and outgo- ing configuration sata given site x. ‘The quantity is a random Boolean number in time and space, satisfying the normalization condition Dewe vs and the conservation conditions LE swheO, where hy is any linear combination of 1, (, ~1)s/3} and (,),=sinl(a— 1)x/3]. The average value of Sur, Au’ (Eue, defines a transition probability from state's tos, for any arbitrary s and s'. Moreover, Sw has rotational symmetry, ie. for any s and s', é is invariant if states s and s' are both subjected to simultaneous proper ‘or improper rotations. Also, different values for Aw pro- duce different values for the transport coefficients, but their qualitative behavior will not change. The ensemble average of Eq. (1) using the random- phase approximation yields the lattice Boltzmann equa- tion Sabot AV) =felX O46), @) where f-(x,1) is the single-particle distribution function, and f denotes all of fa(x,1). The operator M, is equal to Coif Ny and Ey are replaced by fp and Ay, respectively, in Eq. (2). ‘The particle-number density n(x,0) and momentum density p(x,t) are related to fa(x,t) by nx) =ZofelK0, and plat) =DekafalRit)y_ respec- tively. With the random-phase approximation, which ne~ glects the instantaneous correlations between particles in- volved in a collision process, one can show that the single- particle distribution function fa(x,1) obeys the Fermi- Dirac distribution.” Assume that fo(x,1) dll +e.(x0] with |gel <1, where d is the equilibrium density for zero mean vel We can linearize Eq. (3) to obtai Atel t Dm HOFE APG, — where the linearized collision operator {) is the 6x6 real circulant matrix: “aa, © where o,=Zo5q is the number of particles in configur- ation s. Note that 1°? only depends on the equilibrium density d for a given set of collision rules. The Fourier transform of Eq. (4) is, galt +1) = ENE uot 1D 0 plbu0) © ay Zi s.)oh— 5p) Ave ‘This can be written in vector form lok + D> =A Ged|o(k.0)), ” where the component of the fluctuation vector (al(k,1)) is elk), ‘The matrix H(k)=D()HO), where H(0) =1+0, is the evolution operator. ‘The diagonal matrix” D(k) =diag(exp(—sk-€),exp(—ite-€2), LUO, CHEN, CHEN, DOOLEN, AND LEE a ‘exp(—ik-€,)) is the displacement operator. In general, the eigenvalue problem for H(k) cannot be solved analytically except for special cases, whereas that for H(0) can be trivially solved because (0) is a circu- lant matrix. The matrix #1(0) has three unit eigenvalues ‘corresponding to the three hydrodynamic modes, i.c., the corresponding eigenvectors are associated with the con- served quantities of the system. (0) also has three non- unit eigenvalues corresponding to three kinetic modes. It can be rigorously showa that these modes can be analyti- cally continued to the k=0 region [as indicated by Eq. (24) later]. Although our procedure is general, we illus- trate it using the six-bit colision-saturated nondeterminis- tic FHP model* in what follows. In this case, HO) =cirelpupapyPaprp2) with prmimdG+3d), pr=d+4d)/2, ps=d/2, pam —d(+d), and d=d(1—d). The eigenvalues of HO) are Aize"l, Ass—1—3d(1+2d), and 261 6d". The eigenvectors of H(0) with unit eigenvalue are 1 2 ° 1 1 1 -1 1 Im=e]i]. bom se] =2} andlno=] o | 1 -1 = 1 1 = (8) which are the density, x-momentum, and y-momentum modes, respectively. Denoting @ as the angle between k and € (=8), uJ), and |u + (kc) as the generalized hy- drodynamie eigenvectors, the zeroth order (in k) hydro- dynamic eigenvectors of 1 (k) are |u,(0))=cos6|p,0) ~ sin@l ps0) =|p,) , 9) 1 lu + (0)) =—L-In) + (cos6 peo) +sin6pyo)) lus 7 I 086 |p. IPy0) 1 = tlp), ao a ln where |p)) and |p)) denote the transverse and longitu- dinal_ momentum with respect to Kk. If we let $,(k,0) =(u(K)16(K,0) and 4K, Gre (1600), then the generalized hydrodynamic equations for the linearized lattice Boltzmann equation are (kt +1) 21h) 6 (kt) ee 9/(k,0), (11) balk +1) 922 es (kt) et 95 (k,0), (a2) ‘and the generalized hydrodynamic transport coefficients are defined in analogy to their definitions in hydrodynam- vik) = In(ei(k)) e $v 0k) + nk) = Retin 2 6))1 fon Imlin( ah eeea GENERALIZED HYDRODYNAMIC TRANSPORT IN LATTICE- where v, n, and c, are the kinematic viscosity, the bulk. viscosity, and the sound speed, respectively. By expanding, Dk) ink, Le., D(k) = En —ik)"P"CO)/nl with P(0) =diaglcos0,cos(0—x/3),... ,cos(+x/3)1, the coefficients of perturbation expansions for the eigen- values can be calculated. The coefiients in perturbation expansions for the transport coeficients are just the cumu- lants' of the corresponding eigenvalues, As an example, the first few coefficients in perturbation expansions for 2,(k) and the kinematic viscosity v(k) are given as fol- lows: ak) 12K OKA, where 2 (+A )/80-29), 2 = 2/48 —24) 1 -25)"1 12+ 2g Daa + Dhsdy —SAF+ WAG) = £05(60)(1 —As)L+SAs+ SAG HAAG, vik) =vo~ 2k. where voz, and vy—(2;)7/2—z/%. The zeroth or- der (in k) results of the transport coefficients obtained by this perturbative method are identical to the previous re- sults.?? In the case of @=0 or 02/6, the eigenvalues of H(k) are obtained analytically for arbitrary k. For @=0, H(k) can be decomposed into the direct sum of a 2x2 and a 4x4 matrix. For 0/6, H(k) becomes a direct sum of two 33 matrices. Therefore the eigenvalues are roots of ‘quadratic, quartic, or cubic algebraic polynomials. The ‘decomposition of H1(k) along the special directions 8=0 and 02/6 is consequence of the fact that the collision operator Mg is invariant under the complete lattice sym= metry group.’ This decomposition is also applicable to other FHP 2D models. For 0=0, 2k) = $ (1+2s)c08(+k) + SUF Poo —a. aa) The kinematic viseosity possesses an imaginary part be- tween the two branch points which satisfy (1++A3)? xc0s"(k/2)—423—0. As k increases from 0 to the first branch point, z)(k) and the eigenvalue of the kinetic mode coupled by the quadratic equation collide with each other. ‘Then both of them become complex conjugates until they collide again at the second branch point. After the second collision they separate along the real axis of the z plane. Numerical results indicate that a similar situation occurs when 0<0<2/6 for the transverse mode. However, when @=2/6 the kinematic viscosity has no imaginary part in the physical region (0 03] 00 on F 00 0.1 1.0) wave number i FIG. 1. Kinematic viseosity vk), bulk viseosity n(k), and sound speed es(k) vs k for density equal t0 0.2, and 00 (solid line), n/12 (dotted line), and 2/6 (dashed line). (a) v(k) vs Ki b) nf) vs (6) eC ¥5 viscosity v, the bulk viscosity n, and the sound speed c, for density d=0.2 and 60, x/12, and x/6 are shown in Figs. 1(a)-1(c). For 0=n/12, results are obtained by caleulat- ing the eigenvalues of H7(k) directly. We also caleulated vk) and c,(k) by perturbation expansion for 9=x/12. ‘The eighth order perturbative result of v(k) agrees with the exact result for k =0.5; for cs(k), results agree when k'<0.2. One can clearly see that these transport coef- ficients are highly anisotropic even for moderate values of K. This fact indicates that in LGA simulations, a large number of cells must be averaged over in order to over- come the anisotropy. In Fig. 2, the analytic result and LGA simulations of ey are compared for the direction @=7/6. Simulation results, confirm the prediction of the analytic result that cy in- creases as k increases along the direction 0/6. In Fig. 3, the momentum profile of forced flow between parallel plates (Poiseuille flow)" from our analysis and LGA simulations are compared. The geometric arrange- ment and forcing rules of the simulation are the same as 0.74 0.72 sound speed 0, 0.70) 0.0 0.1 02 0.3 04 05 ‘wave number & FIG. 2. Sound speed cy(k) vs k for density d=0.2 and O=%/6, The analytic result is represented by the solid line and the LGA simulations by “+”. The relative differences between, the LGA simulation and the analytic result are less than 1%10 momentum p 1 6 im 16 channel width FIG. 3. Momentum profile of Poiseille low for density 40.2 and a channel width of 16 lattice sites. To obtain @ steady state, 10° time iterations were run hefore the average ‘The momentum p, is averaged over Ly and over 210 time iterations. The analytic result [Eq. (16)] is represented by the solid line and the LGA simulation by “0”. The graph is rescaled 50 that pax I. Note the agreement for the nonzero momen- tum at the walls due to the finite mean free path. The simul tion was run on a Cray-YMP computer, with Fo=L0641 x10" in Ref. 11: plates are arranged parallel to a velocity diree- tion, and periodic boundary conditions are applied. ‘The system size is LyX, 512*32, where Ly is twice the channel width. The forcing is a square-wave function be- tween plates: it is of uniform magnitude on sites |= 516 and of opposite uniform magnitude on sites 17 Sy 5322, ‘Assume Eq. (4) with a term of the square wave forcing along the & direction, which is parallel to the boundaries. Let k be along the 9 direction and assume a steady state, ie, lolk.r+1))=[6(k,0)), Then the discrete Fourier transform of the momentum profile is 6 a+—___6 ___ ( sin WV5(2k = DDa/ANI je Asinl2k=Dav/N] 92 coy 2NQk=De OSS? 1, as) Fo pay 'U. Frisch, B. Hasslacher, and Y. Pomeau, Phys. Rev. Lett. 6, 1505 (1986). 2U, Frisch eta, Complex Syst. 1, 649 (1987). 3S. Wolfram, J Stat. Phys. 45,471 (1986). 48. Chen eral, Physica D 47, 72 (1991); D. H. Rothman, Geo- physics 83, 509 (1988), 5G. R. McNamara and G. Zanet (1988). ©B.J. Alder and W. E. Alley, Phys. Today 37 (1), $6 (1984) YL M. de Schepper and E, G. D. Cohen, J. Stat, Phys. 27, 223, ‘982. 5M. Hénon, Complex Syst. 1, 763 (1987). Phys, Rev. Lett, 61, 2332 LUO, CHEN, CHEN, DOOLEN, AND LEE a where Fis the magnitude of forcing, N=L,/2, a=(2—3d~15d°)/13d°(2~3d ~6d")1 and b=3d(1+2d)/42~3d~-6d") In the thermodynamic limit (W—> ee, d~+0, and 2Nd—+do=const), the following parabolic profile for e(9) is obtained PG, tim pc? _[rolsiv0-p) +a), o=7<1, Fol4b—1)G-2) al, 15 <2, where @=1/(3d3), B=3do/8, and F=y/N. We can rewrite d=2/, where T's the mean free path in the limit. Therefore, the discontinuity of the momentum profiles, which presents the slip velocity at the walls," is propor- tional to the mean free path in the limit. Note that in Fig. 3 the discontinuity of the momentum profiles at the boun= aries (the slip momentum) is accurately predicted by the analysis. This phenomenon is a manifestation of existence of a Knudsen layer." In conclusion, we have obtained an analytical solution of the linearized lattice Boltzmann equation. We have quantitatively analyzed boundary effects due to a finite mean free path (Knudsen layer) and anisotropy effects due to the lattice symmetry, and find agreement with LGA simulations. a6) We are very grateful to Dr. H. A. Rose for his valuable suggestions. We thank Professor David Levermore for many helpful conversations and for showing us Ref, 13 be- fore publication, and we thank Dr. E. Y. Lob, Jr. for gui- dance in the use of the Connection Machine. One of the authors (S.C.) would like to acknowledge the support from the NASA Innovative Research Program under Grant No. NAGW-1648. This work is supported in part by the U.S. Department of Energy and by the Defense Advanced Research Project Agency Grant No, DPP88- 50. °P, J, Davis, Circulant Matrices (Wiley, New York, 1979). 9G. A. Korn and T. M. Korn, Mashematical Handbook for Sclemtists and Engineers, 2nd ed, (McGraw-Hill, New York, 1968). pp. 600-601 LP. Kadanoff, G. R, MeNamara, and G, Zanetti, Complex Syst. 1,791 (1987), "8D, K. Bhattacharya and G. C. Lie, Phys. Rev. Lett 62, 897 (1989); G, Mo and F. Rosenberger, Phys. Rev. A 42, 4688 (1990) "AR, Cornubert, D. d'Humiéres, and D. Levermore, Physica D 47,241 (1991).