1: %
2: % version of 07/11/01
3: %
4:
5: \documentclass[figures]{epl}
6:
7: \title{Multi-particle-collision dynamics:
8: Flow around a circular and a square cylinder}
9: \shorttitle{Cell particle dynamics: Flow around a cylinder}
10: \author{A. Lamura\inst{1}, G. Gompper\inst{1}, T. Ihle\inst{2},
11: \And D. M. Kroll\inst{2} }
12: \shortauthor{A. Lamura \etal}
13: \institute{
14: \inst{1} Institut f\"{u}r Festk\"{o}rperforschung,
15: Forschungszentrum J\"{u}lich,\\
16: D-52425 J\"{u}lich, Germany\\
17: \inst{2} Supercomputing Institute, University of Minnesota,\\
18: 1200 Washington Ave. S., Minneapolis, MN 55415, USA
19: }
20: \pacs{02.70.Ns}{Molecular dynamics and particle methods}
21: \pacs{47.11.+j}{Computational methods in fluid dynamics}
22: \pacs{82.20.Wt}{Computational modelling; simulation}
23:
24: \begin{document}
25:
26: \maketitle
27:
28: \begin{abstract}
29: A particle-based model for mesoscopic fluid dynamics is used to simulate
30: steady and unsteady flows around a circular and a square cylinder in a
31: two-dimensional channel for a range of Reynolds number between 10 and
32: 130. Numerical results for the recirculation length, the drag
33: coefficient, and the Strouhal number are reported and
34: compared with previous experimental measurements and computational fluid
35: dynamics data. The good agreement demonstrates the potential of this
36: method for the investigation of complex flows.
37: \end{abstract}
38:
39: \section{Introduction}
40: The simulation of the hydrodynamic behavior of complex fluids has attracted
41: considerable attention in recent years. Several computational fluid dynamics
42: approaches, such as lattice gas automata~\cite{fris86}, lattice-Boltzmann
43: methods~\cite{mcna88,higu89}, and dissipative-particle dynamics~\cite{pago98},
44: have been proposed and developed in order to describe the dynamical behavior of
45: these systems on mesoscopic length scales.
46: The first two methods, despite of their conceptual simplicity, suffer
47: mainly from the lack of Galilean invariance. Moreover, since these are
48: lattice based methods, practical applications involving irregular geometries
49: often require the use of adapted computational meshes. Dissipative particle
50: dynamics is an off-lattice, particle-based method which does not have these
51: problems; however, it is often complex and difficult to analyze analytically.
52:
53: In this Letter, we investigate in detail another particle-based
54: simulation method which is a modification of Bird's Direct Simulation Monte
55: Carlo algorithm \cite{bird76}. The fluid is modeled by ``particles'' whose
56: positions and velocities are continuous variables, while time is discretized.
57: The system is coarse-grained into the cells of a regular lattice with no
58: restriction on the number of particles per cell.
59: The evolution of the system occurs in two steps: propagation and collision.
60: Each particle is first streamed by its displacement during the time
61: interval $\Delta t$.
62: A novel algorithm for the collision step was introduced recently by
63: Malevanets and Kapral~\cite{male99}.
64: The cells are the collision volumes for many-particle collisions
65: --- which is why we refer to this method as multi-particle collision
66: dynamics (MPCD) ---, which conserve mass, momentum and energy.
67: In Ref.~\cite{male99} it was shown that an H-theorem holds for this
68: dynamics and that the correct hydrodynamic equations for an ideal gas are
69: recovered. This model has been carefully studied in Ref.~\cite{ihle01},
70: where the Galilean invariance is critically discussed and it is shown
71: that the original algorithm has to be modified at low temperatures to
72: guarantee this symmetry.
73:
74: We present the results of a quantitative analysis of the
75: MPCD method in order to determine whether it can provide a convenient
76: alternative to other computational fluid dynamics approaches. In particular,
77: we study the problem of two-dimensional incompressible
78: flow around a circular and a square cylinder. The available
79: literature on this classical flow problem allows us to test and
80: validate this new model.
81:
82: After describing the method, we present computational results for a wide
83: range of Reynolds number, covering both steady and unsteady periodic flows.
84: The results are compared with previous experimental and numerical studies.
85:
86: \section{The model}
87: The system we study consists of $N$ particles of unit mass with continuous
88: positions $\vect{r}_i(t)$ and velocities $\vect{v}_i(t)$ in two dimensions,
89: $i=1, 2, ..., N$. The evolution is given by repeated streaming and collision
90: steps. Assuming a unit time interval, $\Delta t = 1$, the particle positions
91: change according to
92: \begin{equation}
93: \vect{r}_i(t+1) = \vect{r}_i(t) + \vect{v}_i(t)
94: \label{streaming}
95: \end{equation}
96: during the streaming step.
97: For the collision step, the system is divided into the cells of a regular
98: lattice of mesh size $a_0$, with $a_0=1$. Each of these cells is the collision
99: volume for a multi-particle collision defined by
100: \begin{equation}
101: \vect{v}_i(t+1) = \vect{u}(t) + \mathsf{\Omega} \; [\vect{v}_i(t) - \vect{u}(t)] ,
102: \label{collision}
103: \end{equation}
104: where $\vect{u}$ is the macroscopic velocity, defined as the average velocity
105: of the colliding particles, which we assume to have coordinates of the center
106: of the cell. $\mathsf{\Omega}$ denotes a stochastic rotation matrix which
107: rotates by an angle of either $+\pi/2$ or $-\pi/2$ with probability $1/2$.
108: The collisions are simultaneously performed on all the particles in a cell
109: with the same rotation $\mathsf{\Omega}$, but $\mathsf{\Omega}$ may differ
110: from cell to cell. The local momentum and kinetic energy
111: do not change under this dynamics.
112:
113: The division of space into a fixed cell lattice for the collision step
114: breaks Galilean invariance. This is a minor effect as long as the
115: mean-free path is comparable or larger than the cell size $a_0$, because
116: subsequent
117: collisions of a particle typically occur in different cells. On the
118: other hand, when the mean-free path is small compared to the cell size $a_0$,
119: particles remain within the same cell for many collisions, and become
120: correlated. These correlations
121: depend on the presence of flow, and Galilean invariance is broken.
122: It was shown in Ref.~\cite{ihle01} that this problem can be cured by
123: performing the collision operation in a cell grid which is
124: shifted each time step by a random vector with components in the interval
125: $[-a_0/2,a_0/2]$.
126: The collision environment is then independent of the macroscopic velocity,
127: and Galilean invariance is exactly restored.
128:
129: \section{No-slip boundary conditions}
130: For planar walls which coincide with the boundaries of the
131: collision cells, no-slip boundary conditions are conveniently simulated
132: in the case of a fixed cell lattice
133: by employing a bounce-back rule, i.e. the velocities of particles which
134: hit the wall are inverted after the collision. However, the walls will
135: generally not coincide with, or even be parallel to, the cell boundaries.
136: Furthermore, for small mean-free paths, when a shift operation of
137: the cell lattice is required to guarantee Galilean invariance,
138: partially occupied boundary cells are unavoidable, even in the simplest
139: flow geometries.
140:
141: \begin{figure}[ht]
142: %\twoimages[scale=0.35]{poiseuille_shift.eps}{poiseuille_fixed.eps}
143: \caption{Poiseuille flow through a channel of size $H=30a_0$ and
144: $L=50a_0$, for $n_{av}=35$.
145: (a) $k_BT=0.01275$, corresponding to a mean-free path
146: $\ell = \Delta t (k_BT/m)^{1/2} = 0.11$ (with $\Delta t=m=1$).
147: The simulation is
148: carried out with the shift operation of the collision cells. Results are
149: shown for simple bounce-back boundary conditions (open circles) and for
150: the boundary condition (\ref{newcoll}) (full circles). The dashed and solid
151: lines are fits to a parabolic flow profile.
152: (b) $k_BT=0.4$, corresponding to a mean-free path $\ell = 0.63$.
153: The simulation is
154: carried out for a fixed cell lattice, which is displaced with respect to
155: the walls by a quarter of the lattice constant, so that the cell lattice
156: fills the channel asymmetrically.
157: }
158: \label{fig_poiseuille}
159: \end{figure}
160:
161: We found that the simple bounce-back rule fails to guarantee no-slip boundary
162: conditions in the case of partially filled cells. This can be demonstrated,
163: for example, for
164: a simple Poiseuille flow at low temperatures (small mean-free paths). The
165: velocity profile for a channel of height $H$ and length $L$ (measured in
166: units of the lattice constant) is shown in Fig.~\ref{fig_poiseuille}(a).
167: The velocity profile does not extrapolate to zero at the walls, and
168: a strong slip is clearly visible.
169: We therefore need a generalization of the bounce-back rule for partially
170: filled cells. Many different schemes are possible. We propose the
171: following algorithm, which we found both to be very efficient and to give
172: good results --- as discussed below.
173: For all the cells of the channel which are cut by walls and therefore
174: have a number of particles $n$ smaller than the average number $n_{av}$
175: of the bulk cells, we fill the
176: `wall' part of the cell with virtual particles in order to make the
177: effective density of real plus virtual particles equal the
178: average density. The velocities of the wall particles are drawn from a
179: Maxwell-Boltzmann distribution of zero average velocity and the same
180: temperature $T$ as the fluid. The collision step, Eq.~(\ref{collision}),
181: is then carried out with the average velocity of {\em all} particles
182: in the cell. Since the sum of random vectors drawn from a Gaussian
183: distribution is again Gaussian-distributed, the velocities of the
184: individual wall particles never have to be determined explicitly. Instead,
185: the average velocity $\vect{u}$ in Eq.~(\ref{collision}) can be written
186: as
187: \begin{equation}
188: \vect{u} = \frac{\sum_{i=1}^{n} \vect{v}_i + \vect{a}}{n_{av}}
189: \label{newcoll}
190: \end{equation}
191: where $\vect{a}$ is a vector whose components are numbers from a
192: Maxwell-Boltzmann distribution with zero average and variance
193: $(n_{av}-n) k_B T$ \cite{note}. Results for Poiseuille flow
194: with partially filled cells, both with cell-shifting and
195: a fixed cell lattice, are shown in Fig.~\ref{fig_poiseuille}(a)
196: and \ref{fig_poiseuille}(b), respectively. The results are in
197: very good agreement with the expected parabolic flow profile
198: for kinematic viscosities $\nu=0.079\pm 0.001$ and $\nu=0.087\pm 0.001$,
199: respectively,
200: which should be compared with the values of $\nu=0.083\pm 0.001$ and
201: $\nu=0.088\pm 0.001$ obtained from the (short-time) decay of the
202: velocity autocorrelation function in an equilibrium system \cite{ihle01}.
203: It is important to notice that the center-of-mass velocity of the virtual
204: wall particles is {\em non-zero} in general, which reflects the non-zero
205: temperature of the wall. Simulations of Poiseuille flow using the same
206: boundary condition but with resting wall particles, i.e.
207: $\vect{a}\equiv 0$ in Eq.~(\ref{newcoll}), again fails to reproduce the
208: desired flow profile. In this case, also the density of (real) particles
209: increases drastically near the walls.
210:
211:
212: \section{Results and discussion}
213:
214: \begin{figure}[ht]
215: %\threeimages[scale=1.50]{flow1.eps}{flow2.eps}{flow3.eps}
216: \caption{Velocity field at the final steady state for $Re=30$ for the circular
217: (left panel) and square (center panel) cylinder. Velocity field at $Re=100$
218: for the square cylinder (right panel). Only a fraction of the simulation box is
219: shown in each case.}
220: \label{fig_flow}
221: \end{figure}
222:
223: The flow around a circular and a square cylinder with diameter $D$ inside a
224: plane channel of height $H$ and length $L$ was investigated. The blockage
225: ratio $B=D/H$ is fixed at $0.125$ for both the cylinders. In order to reduce
226: the influence of inflow and outflow boundary conditions, the length $L$ is
227: set to $L/D=50$. The cylinder is centered inside the channel at a
228: distance $L/4$ from the inflow region. The average number of
229: particles per cell is $n_{av}=10$ and the temperature is fixed at
230: $k_B T=0.4$; the corresponding viscosity is $\nu=0.110 \pm 0.004$. The
231: flow is driven by assigning Maxwell-Boltzmann-distributed
232: velocities with parabolic profile $v_x(y) = 4 v_{max} (H-y)y/H^2$ to
233: particles in the region $0 < x \leq 10$. The maximum flow velocity, $v_{max}$,
234: is fixed by the condition that the Mach number --- i.e. the velocity
235: relative to the speed of sound $c$, with $c^2 = [C_p/C_v] dp/d\rho =
236: 2k_BT/m$ --- is approximately
237: $1/4$, in order to avoid significant compressibility effects.
238: Periodic boundary conditions are imposed in the $x$-direction, no-slip
239: boundary conditions on the channel and cylinder walls.
240:
241: Reynolds numbers, $Re\equiv v_{max} D/\nu$, in the range
242: $10 \leq Re \leq 130$ were investigated. Both steady flows, with a
243: closed steady
244: recirculation region consisting of two symmetric vortices behind
245: the body, and unsteady flows, with the well-known von Karman vortex street
246: with periodic vortex shedding from the cylinder, are observed to occur for
247: this range of Reynolds numbers. The critical Reynolds number above which
248: flow becomes unsteady, is approximately $49$ for the circular
249: cylinder~\cite{will96} and $60$ for the square cylinder~\cite{breu00}.
250: Figure~\ref{fig_flow} shows the macroscopic velocity field of the final
251: steady-state for $Re=30$, and of the periodic vortex shedding for $Re=100$.
252:
253: \begin{figure}[ht]
254: %\twofigures[scale=0.35]{recirc.eps}{drag.eps}
255: \caption{
256: The recirculation length as a function of the Reynolds number.
257: Square cylinder: ($\bullet$) this study with fixed cell lattice,
258: ($\circ$) this study with cell-shifting, ($-\!\!\!-\!\!\!-\!\!\!-$)
259: Breuer {\it et al.}~\cite{breu00};
260: Circular cylinder: ($\star$) this study with fixed cell lattice,
261: ($\triangle$) this study with cell-shifting, ($- - -$) Coutanceau and
262: Bouard~\cite{cout77}.}
263: \label{fig_recirc}
264: \caption{The drag coefficient $C_d = 2F_\parallel/(n_{av} v_{max}^2 D)$ ---
265: where $F_\parallel$ is the force exerted on the cylinder in the direction
266: parallel to the flow --- as a function of the Reynolds number.
267: Square cylinder: ($\bullet$) this study, ($-\!\!\!-\!\!\!-\!\!\!-$)
268: Breuer {\it et al.}~\cite{breu00};
269: Circular cylinder: ($\star$) this study, ($\circ$) Tritton~\cite{trit59},
270: ($\triangle$) He and Doolen~\cite{he97}.}
271: \label{fig_drag}
272: \end{figure}
273:
274: Such flow patterns are consistent with experiments ({\it cfr.} Fig.~4 of
275: Ref.~\cite{cout77} for the circular cylinder) and simulations ({\it cfr.}
276: Fig.~2 of Ref.~\cite{breu00} for the square cylinder). The length, $L_r$,
277: of the recirculation region, from the rear-most
278: point of the cylinder to the end of the wake, has been measured in units of
279: $D$. Results, with error bars, are shown in Fig.~\ref{fig_recirc} for both
280: the square and circular cylinders. The errors
281: are estimated to be $1/(2D)$, since the position
282: of the macroscopic velocity $\vect{u}$ in each cell is arbitrarily taken to
283: be the center of the cell. Our results for $L_r/D$ are consistent with a
284: linear increase with Reynolds number.
285: Our data for the circular cylinder compare very well with
286: experimental measurements for a similar value of the blockage ratio
287: ($B=0.12$)~\cite{cout77}. Indeed, $L_r/D$ is a sensitive function on $B$, and
288: increases with decreasing $B$~\cite{cout77}. No numerical data are
289: available for a similar
290: value of $B$. In the case of the square cylinder, our data can be compared
291: with the numerical results of Ref.~\cite{breu00}, obtained using
292: lattice-Boltzmann and finite-volume methods with the same blockage ratio
293: $B=0.125$ and same parabolic inflow velocity profile as in our study. A
294: linear fit to our data obtained both with cell-shifting
295: as well as the fixed-cell lattice data for $Re=10$ gives
296: \begin{equation}
297: L_r/D = - a + b \cdot Re
298: \end{equation}
299: with $a=0.148 \pm 0.069$ and $b=0.0525 \pm 0.0020$
300: while Breuer {\it et al.} find $a=0.065$ and $b=0.0554$. The slopes
301: are very similar, but our data are shifted to somewhat smaller values
302: of $L_r/D$. For the square cylinder, no experimental
303: data could be found in the literature.
304:
305: Figure~\ref{fig_drag} shows the drag coefficient $C_d$ as a function of
306: the Reynolds number. For the square cylinder, agreement between our
307: data and the previous
308: numerical measurements is satisfactory, with a small but systematic
309: deviation to larger values compared to Ref.~\cite{breu00}
310: for $Re > 20$. In the case of the circular
311: cylinder, our results are compared with the experimental measurements of
312: Tritton~\cite{trit59}, and with numerical simulations by He and
313: Doolen~\cite{he97} performed with a lattice-Boltzmann method implemented
314: on a variable mesh. In contrast to the current simulations,
315: Refs.~\cite{trit59} and \cite{he97} both used a constant incoming
316: velocity profile and a very small blockage ratio.
317:
318: Our results fall below (about $5\%$) the literature data in this case.
319: This is mainly due to the different boundary conditions and blockage
320: ratios used in the various studies. Numerical investigations with the
321: lattice-Boltzmann method \cite{wagn94} showed
322: that $C_d$ depends on the profile of the incoming flow. In Ref.~\cite{he97},
323: lateral periodic boundaries and a constant inflow velocity profile
324: instead of no-slip boundary conditions with a parabolic inflow
325: velocity profile were used. It was
326: shown in Ref.~\cite{wagn94} that $C_d(no-slip)/C_d(periodic) = 0.82$ for
327: $B=0.1$ (no error bar given). It seems reasonable to expect a similar behavior
328: also for the present case. Indeed, the comparison of our results with those of
329: the lattice-Boltzmann simulation \cite{wagn94} gives a ratio of
330: $C_d(no-slip)/C_d(periodic) = 0.90\pm 0.03$.
331:
332: \begin{figure}[ht]
333: %\onefigure[scale=0.35]{strouhal.eps}
334: \caption{The Strouhal number as a function of the Reynolds number.
335: Square cylinder: ($\bullet$) this study, ($\diamond$)
336: Breuer {\it et al.}~\cite{breu00};
337: Circular cylinder: ($\star$) this study, ($\circ$) Tritton~\cite{trit59},
338: ($\triangle$) He and Doolen~\cite{he97}, ($\ast$) Williamson~\cite{will88}.
339: Both lines are guides to the eye.}
340: \label{fig_str}
341: \end{figure}
342:
343: In the unsteady flow regime we computed the lift coefficient
344: $C_l= 2F_\perp/(n_{av} v_{max}^2 D)$, where $F_\perp$ is the force excerted
345: on the cylinder in the direction perpendicular to the flow.
346: The characteristic frequency $f$ of the vortex shedding was determined by a
347: spectral analysis (fast Fourier transform) of the time
348: series of $C_l$ and used to determine the Strouhal number
349: $St\equiv fD/v_{max}$~\cite{note}.
350: Our data as a function of $Re$ are shown in Fig.~\ref{fig_str}.
351: They lie within the range of scatter of the published data.
352: We want to mention again that
353: for the circular cylinder, the existing literature were all
354: obtained using different inflow conditions than in our study.
355:
356: \section{Conclusions}
357: The multi-particle collision dynamics method has been successfully applied
358: to simulate two-dimensional flow around a circular and a square cylinder.
359: The results are found to be in good agreement with previous experimental
360: measurements and computational fluid dynamics studies.
361: The computational efficiency of the MPCD method can be estimated from
362: the time requirement of about 1.7 seconds (1.2 seconds) per time step for
363: a system of 2 million particles with (without) cell-shifting on a
364: Compaq XP1000 (667 MHz) workstation.
365:
366: The main difference between multi-particle collision dynamics and lattice
367: Boltzmann and finite-volume methods, which have been employed in most other
368: studies of this flow geometry, is the presence of thermal fluctuations.
369: In equilibrium, these fluctuations are responsible for a logarithmic
370: divergence of the viscosity with time \cite{erns70,ihle01}, a divergence
371: which is cut off by the finite system size at long times. A thermal
372: renormalization of the viscosity could indeed be responsible for the
373: small deviations of the values
374: for the recirculation length and the drag coefficient compared to the
375: lattice Boltzmann results.
376:
377: The present model provides a simple alternative scheme that can be used to
378: treat a wide class of physical and chemical problems. Many directions are
379: accessible to exploration. The results
380: presented here are relevant for studies of the
381: interactions among large colloidal particles in solution. The algorithm
382: can also be used for a mesoscopic model of the solvent dynamics which
383: can be coupled to a microscopic treatment of solute particles.
384:
385: \acknowledgments
386: A.L. thanks the Supercomputing Institute of the University of Minnesota for its
387: hospitality. Support from the National Science Foundation under Grant Nos.
388: DMR-9712134 and DMR-0083219, and the donors of the Petroleum
389: Research Fund, administered by the ACS, are gratefully acknowledged.
390:
391:
392: \begin{thebibliography}{99}
393:
394: \bibitem{fris86}
395: \Name{Frisch U., Hasslacher B. \And Pomeau Y.} \REVIEW{Phys. Rev. Lett.}{56}
396: {1986}{1505}.
397:
398: \bibitem{mcna88}
399: \Name{McNamara G. R. \And Zanetti G.} \REVIEW{Phys. Rev. Lett.}{61}{1988}{2332}.
400:
401: \bibitem{higu89}
402: \Name{Higuera F. J. \And Jimenez J.} \REVIEW{Europhys. Lett.}{9}{1989}{663}.
403:
404: \bibitem{pago98}
405: \Name{Pagonabarraga I., Hagen M. H. J. \And Frenkel D.} \REVIEW{Europhys. Lett.}
406: {42}{1998}{377}.
407:
408: \bibitem{bird76}
409: \Name{Bird G. A.} \Book{Molecular Gas Dynamics}\Publ{Clarendon Press, Oxford}
410: \Year{1976}.
411:
412: \bibitem{male99}
413: \Name{Malevanets A. \And Kapral R.} \REVIEW{J. Chem. Phys.}{110}{1999}{8505};
414: \REVIEW{ibid.}{112}{2000}{7269}.
415:
416: \bibitem{ihle01}
417: \Name{Ihle T. \And Kroll D. M.} \REVIEW{Phys. Rev. E}{63}{2001}{020201(R)}.
418:
419: \bibitem{note}
420: Details will appear elsewhere.
421:
422: \bibitem{will96}
423: \Name{Williamson C. H. K.} \REVIEW{Annu. Rev. Fluid Mech.}{23}{1996}{477}.
424:
425: \bibitem{breu00}
426: \Name{Breuer M., Bernsdorf J., Zeiser T. \And Durst F.}
427: \REVIEW{Int. J. Heat Fluid Flow}{21}{2000}{186}.
428:
429: \bibitem{cout77}
430: \Name{Coutanceau M. \And Bouard R.} \REVIEW{J. Fluid Mech.}{79}{1977}{231}.
431:
432: \bibitem{trit59}
433: \Name{Tritton D. J.} \REVIEW{J. Fluid Mech.}{6}{1959}{547}.
434:
435: \bibitem{he97}
436: \Name{He H. \And Doolen G.} \REVIEW{J. Comput. Phys.}{134}{1997}{306};
437: \REVIEW{Phys. Rev. E}{56}{1997}{434}.
438:
439: \bibitem{wagn94}
440: \Name{Wagner L.} \REVIEW{Phys. Rev. E}{49}{1994}{2115}.
441:
442: \bibitem{will88}
443: \Name{Williamson C. H. K.} \REVIEW{Phys. Fluids}{31}{1988}{2742}.
444:
445: \bibitem{erns70}
446: \Name{Ernst M. H., Hauge E. H., \And van Leeuwen J. M. J.}
447: \REVIEW{Phys. Rev. Lett.}{25}{1970}{1254}.
448:
449: \end{thebibliography}
450:
451: \end{document}
452:
453:
454: