Professional Documents
Culture Documents
4 (1992) 247-267
C0ntinuumMechanics
T h and
erm0uynamlcs
]
9 Springer-Verlag 1992
In this paper a new approach to computational fluid dynamics is introduced in which the fluid is regarded as a stochastic dynamical system. The
velocity of the fluid is related to a stochastic process governed by an appropriate master equation acting in a discrete phase space. The method is
explained by means of (1 + 1)-dimensional flow phenomena. It is shown
that the stochastic approach naturally leads to transparent numerical
algorithms for stochastic simulations of fluid motion. By simulating plane
Poiseuille flow it is demonstrated that the probabilistic approach yields a
correct description of laminar fluid motion. Furthermore, soliton-like and
shock wave solutions of Burgers' equation are generated by stochastic
simulations of the underlying stochastic process.
1 Introduction
The basic equations describing the dynamics of fluid motion, i.e. the NavierStokes equation, have been known for a long time. Nevertheless, the integration of these partial differential equations remains a challenging problem in
spite of the increasing computational capabilities available today. For this
reason, we suggest in this paper a new approach to fluid dynamics which treats
the fluid as a stochastic dynamical system governed by a discrete master equation. Thereby this approach avoids a mathematical formulation in terms of
deterministic partial differential equations for macroscopic fields.
One of the central ideas of our approach is to relate the velocity field of
the fluid to an appropriately chosen discrete stochastic process. Once a definition of the phase space, i.e. the space of states of the fluid, has been given,
a master equation governing the stochastic process can be formulated. On the
other hand, the velocity which appears as a field variable in the Navier-Stokes
equation is then interpreted as the expectation value of the stochastic process
defined by the master equation. This transition from random variables to expectation values provides the connection between the stochastic and the
macroscopic description and applies also to variables other than the velocity.
248
~" -
2 Hydrostochastics
In this section we first define the space of states of the fluid. Once the accessible states of the fluid are known it is possible to construct a master equation
governing the probabilistic time evolution of the system. Furthermore, we
demonstrate how the deterministic equations o f motion emerge from the
master equation as the equations governing the time evolution of the expectation value of the random velocity.
249
(1)
u(x z, t) = ~u(N~(t)).
The above equation provides the connection between the macroscopic and the
stochastic description of fluid motion. Within the stochastic description the
velocity is a time-dependent random variable N~, i.e. a stochastic process
governed by a master equation to be defined below. On the other hand, the
velocity on a macroscopic scale, that is on the scale accessible to standard experimental observation, is given by an expectation value, and therefore obeys
the deterministic Navier-Stokes equation. The velocity unit Ou has been introduced in order to obtain a discrete stochastic process N~. Thus it
represents the size of the smallest possible changes of the state of the fluid
in the discretized phase space. This means that within our description a
positive value Ou. N~ of the random velocity in the cell 2 can be interpreted
as the presence of N~ velocity particles each carrying the velocity Ou. Correspondingly, a negative value Ou. N~ is to be interpreted as the presence of
INk] antiparticles of velocity. Defining the positive and negative part of N~ as
N2u+::
[Nu~,
N~>0
O,
N~<O
)~
Nu-:=
f 0,
N~>0
~-N~,
N~<O
(2)
we write:
N~ = N~+ + N~_.
(3)
Thus we see that within our description the mesoscopic state of the fluid may
be viewed as a many velocity particle state and is completely fixed by specifying the number N~ of velocity particles in each cell 2. Formally, the resulting
phase space may be written as
~
r = {(N ~)~z
JN~ ~ z}.
(4)
250
(5)
Op
-
Ot
(6)
Ox
O,
(7)
Ox
neither u nor Op/Ox can depend on the spatial coordinate. Therefore, assuming
that Op/Ox does not depend on time equation (6) describes a uniform linear
growth of the velocity u with time.
In order to model this deterministic behaviour within our mesoscopic picture we regard u(t) as the expectation value of an integer valued Poisson random process [4] Nu(t), i.e. we define
u( t) = fiu(Nu( t) ).
(8)
The Poisson process underlying the partial differential equation (6) is defined
by the following master equation governing the time evolution of the probability distribution P ( N u, t)
OP
- - = k ( P ( N u - 1, t) - P ( N u, t)).
Ot
(9)
(10)
The transition rate k has to be chosen such that this expectation value shows
the correct time dependence, i.e.
k =
aOp
6u Ox"
01)
251
The important fact to be noticed here is that the Poisson process has, in contradistinction to other stochastic processes, a deterministic limit. In fact, in the
limit
f u --+ 0
and
fuN, = const.,
(12)
we obtain
P(Nu't)~f(Nufu-kfut)=f(
Nufu+ Op
"
(13)
(14)
P = P(N~; t)
reads
OP
- E kZ(E~-I - 1 ) P ,
Ot
2
(15)
where, again, k ~ denotes the space dependent negative pressure gradient divided by flu. In the above equation we introduced the shift operators Ea
defined by
E~F( . . . . N~ . . . . ) = F( . . . . N~ + 1. . . . )
E ; 1 F ( . . . . N~ . . . . ) = F ( . . . . N~ - 1 , . . . ) ,
(16)
OP
at - E k~ (Ex--1 - 1 ) P + ~
kx>0
- k X ( E a - 1 ) P = : dpP,
(17)
kx<0
252
H. E Breuer, F. Petruccione
(18)
u ( x z, t) = ~u(N~u(t))
Ot
d = const..
(20)
It is easy to see from the above equation that in the limit of infinitesimal small
random steps,
~bz(t) -~ 6(x = Ol. 2, t),
(21)
D : = lira (t~l2 d) .
~l~o
(22)
The formal analogy of the diffusion equation (22) with the one dimensional
Navier-Stokes equation (18) is obvious. However, there is a fundamental difference between these two equations. The diffusion equation (22) describes the
time evolution of a probability distribution which is normalized and positive
definite. In contrast, the one dimensional Navier-Stokes equation (18) is an
equation for an expectation value which might be negative and not normalizable. Thus, the stochastic process underlying (18) can not be described
by a one particle random walk.
Therefore, one has to leave the one particle picture. To this end, we consider a collection of independent velocity particles each of which is governed
by Eq. (20). The state of the resulting collective system [4] is characterized by
the set of numbers N~ of velocity particles in all cells ft. Thus, in this manyparticle picture one disregards the identity of the individual particles and is
merely interested in the occupation numbers N~ of particles in each cell.
(Note, that in the one-particle picture the state of the system is completely
specified by giving the location )~ of the particle.)
In the framework of the many particle picture the one-particle density
4z(t) is replaced by the many particle probability distribution P(N~; t). The
253
Ot -
1
ROl 2
~ [(E;11E~ -
1) N~ + (E;+11 E2 - 1) Null P.
(23)
where fil denotes the cell size. Since the expectation value of N~ is a statistical
guess of the one-particle density at the point 2 the velocity field (19) obeys
the diffusion equation (18) as required.
Until now we implicitly restricted our attention to the collective random
walk of positive velocity particles. Obviously, to be consistent with our general
picture of particles and antiparticles of velocity we have to generalize the
master equation describing viscous fluids in order to allow for the correct
treatment of antiparticles of velocity. Since a one-particle jump to the right
is equivalent to a one-antiparticle jump to the left and vice versa, the master
equation (23) has to be modified in the following way: In order to describe
the diffusion of antiparticles the numbers N~ = N~_ are replaced by their absolute values tN~[ = -N~_ and the shift operators EZ.11Ea by E2 1EF 1. Consequently, if both particles and antiparticles of velocity are present the master
equation can be written:
OP
1
--
Ot
[ (EJ_IE
-1 2 - 1) Nu~+ + (EZ+11E~ - 1) N~u+]P
RSI 2 ~
1
Rill 2 E
Z
=:
ddP,
(24)
at
+ u
Ox
R Oxz"
(25)
Burgers' equation has been used as a simple model of turbulent motion. Furthermore, it admits solutions representing nonlinear wave motion [6]. Note that
254
1 ~
(E~-+11Ez- l) 2 Nu(Nu - 1)P.
(26)
In Sec. II.C we demonstrate how Burgers' equation emerges from the master
equation formulation as an equation for the expectation values. It can be
shown that the above form of the master equation for the convection term
leads to a discretization of the differential operator uO/Ox which is of order
Jl. In order to obtain a discretization of this differential operator which is of
order Jl 2 the following symmetrized form of the master equation will be used:
OP
Ot
_ 1 Ju
2 Jl
(E~-+11E~- 1) 2 N~u(N~ - 1) P
p~
q- E (E~-I Ez-1 -- 1) ~- N u ( N u - 1)
)
).
= : ~r
(27)
where we abbreviated the effect of the right hand side by defining the "convection operator" 5r c.
Constructing Eq. (27) we assumed that N~u _>_0. The general master equation describing both the presence of velocity particles and antiparticles can
now easily be derived. Let us assume, that cell 2 contains ]NuZl =-NZu_ antiparticles of velocity. The elementary jumps of the antiparticles can be obtained
from those of the particles of velocity by simply reversing the directions of
the jumps, i.e. all antiparticles jump to the left. However, since the process
255
Having discussed how to interpret the various terms occuring in the equations
of fluid dynamics as stochastic processes let us now combine these processes
to obtain the unified master equation formulation of Hydrostochastics. Obviously, the combined effect of different terms can be obtained by just adding
the corresponding operators. For example, the master equation describing a
pressure driven viscous fluid in one space dimension reads
OP
ot = I @ +
e,
(28)
where the pressure and diffusion operators @ and J d have been defined in
Eqs. (15) and (24). As another example let us construct the master equation
corresponding to Burgers' equation (25); we obtain from Eqs. (24) and (27):
OP
Ot
1
-- Rt~l 2 E
Rgl 2 E
+2
+ E
(E;11E
- 1) 2 IN I(IN I- 1)P
1 [N~I (IN.~I - 1) PI .
(E;1E;~-I - 1) 2
(29)
),
Although the above equation may look rather complicated it allows for a
transparent physical interpretation within the many-particle picture. Recall,
that the system is characterized by giving the numbers N~ of velocity particles
in each cell of the discretized space. The above master equation defines the
possible transitions between the states in the phase space F and the corresponding transition rates. These transitions solely consist of one particle jumps to
neighbouring cells each particle carrying the amount gu of velocity. The latter
fact turns out to be of fundamental importance as it makes possible the construction of efficient numerical schemes which simulate the stochastic process
underlying the master equation (see Sec. 2.5). Furthermore, the nonlinear convection term in the master equation is responsible for the fact that in the
many-velocity particle system a nonlinear self-interaction is present,
256
Let us now demonstrate in some detail how the macroscopic equations may
be derived from the corresponding master equations. As an example we consider the master equation (29) corresponding to Burgers' equation (25).
Generally, the time derivative of the expectation value ua = (OuN~) may be
written as the expectation value of a commutator
(30)
Ot
From the properties of the shift operators Ez it follows that (recall that
Nu =Nu+ +N~_)
t~bt ,,h12+1 __2Nu~+ +Nu~+l 4_ jVz+1 - 2 N ~ _ +N~_ )
([6uN~u, del ) = R 6l 2 v, u+
---. u-
t~U
,~N2 + 1
- R M 2 ,~.,
-2Nu+Nu
2-1
).
(31)
we obtain neglecting
~u 2
([6uNZu, d c l ) -
((N.u2+l) 2 - (Nu~-l)2).
(32)
4dl
(33)
Ot
1 U2+l -+
U]_ 1
-
2(5l
(34)
Obviously, this equation is nothing but the discretized version of the Burgers
equation (25) which emerges in the continuum limit ~l-~ 0.
As has been mentioned in Sec. 2.2.3 the form (27) of the convection term
leads - in the equation for the expectation value (Su(N~) - to an approximation of the differential operator uO/Ox which is of order (~/)2.
In deriving the above macroscopic equation for the expectation value ux
we neglected, of course, all higher moments of the stochastic process N~. As
is well-known the above master equation, defining a nonlinear stochastic process, leads to an infinite dimensional system of coupled differential equations
for the moments. Thus, in order to derive more rigorously from our master
equation the macroscopic equation and to investigate the dynamics and influence of fluctuations one has to employ a more systematic method. Such
a method is provided by the well-known O-expansion [4]. Applying this expansion to the master equations of Hydrostochastics reveals that, in fact, the
257
macroscopic equations are equivalent to the equations of fluid dynamics. Furthermore, it can be shown that within the linear noise approximation the fluctuations superimposed on the macroscopic dynamics can be identified with
those fluctuations derived from the theory of fluctuating hydrodynamics [7].
Thus, the stochastic dynamics implied by our master equation formulation
can, indeed, be given a clear physical interpretation. However, these considerations are beyond the scope of this paper and will be presented elsewhere [8].
In this work, we confine ourselves to the treatment of (1 + 1)-dimensional
fluid dynamical problems. The generalization of the master equation to two
and three space dimensions can be performed within the same framework of
the theory presented above.
2.4 Boundary and initial conditions
In the preceeding sections we have derived the fundamental master equations
of our theory. As one might expect the correct implementation of boundary
conditions is decisive for the description of fluid motion within the context
of Hydrostochastics as it is within the formulation of fluid dynamics with the
help of partial differential equations. In Hydrostochastics the correct boundary
conditions have to be imposed on the master equation, e.g. eq. (29), describing
fluid motion.
In this section we want to describe how to take care of a boundary
separating a viscous fluid and a solid surface. Typically, there exist two types
of boundary conditions. Firstly, at fixed solid surfaces the no-slip boundary
conditions require that the velocity vanishes. Secondly, for moving boundaries
the no-slip condition requires the velocity of the fluid at the solid surface to
be equal to that of the boundary.
The boundary condition at fixed solid surfaces
/~[ Boundary = 0
(35)
results from the fact that adhesion forces create an infinitesimal fluid layer of
zero velocity along the solid surface. As we already described in detail one of
the fundamental steps in the formulation of Hydrostochastics is the parametrization of the fluid bulk in cells 2. The infinitesimal zero velocity fluid layer
can thus be modelled by introducing a corresponding layer of cells. The cells
within this layer act as sinks of diffusing velocity particles. In practice, this
means that velocity particles jumping into the layer cells disappear. On the
other hand, this implies that no velocity particle can leave the layer cells since
the velocity of the boundary layer is zero.
In the case moving solid surfaces the boundary condition reads
b~j Boundary = /Q,
(36)
where 0 is the velocity of the moving solid surface. As in the case of the solid
surface at rest we again introduce a layer of boundary cells. Now, the layer
cells are supposed to move with the constant velocity U.. In other words, this
means that in each layer cell the number of velocity particles is kept constant.
If for example the solid surface moves with constant velocity U in the x direc-
258
tion, the number of velocity particles Nu in each layer cells has to be chosen
such that U = 6u(Nu)= 0uN,. There is thus a fundamental difference to the
case of a fixed boundary: Now velocity particles may leave the layer cells.
However, in the latter case the velocity of the layer ceils or equivalently the
number of velocity particles in these cells is kept constant. As for the fixed
boundary, particles diffusing from the neighboring cells into the layer ceils
disappear.
For the investigation of the properties of the fluid bulk it might be
necessary to use periodic boundary conditions. Also the implementation of
these boundary conditions does not cause any problem within our approach,
since all transitions have been reduced to one-step processes (see Sec. 3).
As far as initial conditions are concerned, let us assume that the initial
velocity in cell X is u0. This initial condition can be realized by fixing in an
appropriate way the number of velocity particles in the cell under consideration. In our example it is sufficient to put at the beginning of the simulation
N O = Uo/OU velocity particles in cell )~.
dt-
W(N~(to) ) lnr/.
(37)
Thus, having determined the transition time to +dt one chooses a specific
transition with a relative probability which is given by the ratio of the corresponding one-particle transition rate to the total rate W(N~(to)). In this
way a new state N~(to +dt) is found. Repeating this procedure yields a trajectory N~u(t), i.e. a realization of the stochastic process.
259
u(x~, t) = Ou(N~u(t))= ~ E
(N~(t))j.
(38)
j=l
3 Examples
in t w o s p a c e - t i m e
dimensions
In this section we are going to apply Hydrostochastics to some typical ( l + l ) dimensional flow situations. We compare the results of the simulations with
well-known analytical results of Hydrodynamics.
(39)
i.e.,
R -
U.L
(40)
We choose coordinates in such a way that the plates are located at x = 0 and
x = 1. The flow is directed along the positive y-axis and we denote the y-component of the velocity by u = u(x, t).
The hydrodynamic description of this flow is given by the following form
of the Navier-Stokes equation in two space-time dimensions:
Ou
3t
dp
dy
1 02u
R OX2"
+ -
(41)
u(x, 0) = 0,
(42)
(43)
This initial and boundary value problem can be solved by standard means and
yields for the time-dependent velocity field:
u(x, t) = 4x(I - x) -
Z
n = 1,3,5 ....
32
~
793n
sin(~znx) exp
( - l r 2Rn 2 t ) .
(44)
260
H. R Breuer, E Petruccione
(45)
1,25
1.00
0.75-
0,50 -
0.25 -
0
0
0.1
0.2
0.3
0.4-
0.5
0.6
X .------',--
0,7
0.8
0.9
1.0
261
This example demonstrates that it is indeed possible to incorporate the effects of the deterministic pressure gradient as well as that of the viscosity term
in a unified stochastic interpretation of the Navier-Stokes equation. Furthermore, we see that the zero boundary conditions can be implemented quite
simply by regarding the solid walls at x = 0 and x = 1 as sinks for the diffusing velocity particles. Finally, the agreement between the simulation and the
analytical curve is remarkably good during the whole time evolution. This
reflects the fact that we are simulating a stochastic process and not merely a
stationary or equilibrium state.
3.2 Nonlinear waves
Originally Burgers proposed the Eq. (25) as a simple one-dimensional model
of homogeneous turbulence. The main features of the Navier-Stokes equations
are retained in the above equation. The nonlinearity has the same structure
as in the Navier-Stokes equation, and the dissipative term is also of the same
type. Only the pressure term is missing, so that one has to expect a relaxation
of turbulence with time. The model also lacks an equation of continuity, so
that it describes in practice a one-dimensional compressible flow. The Burgers
equation is particularly appealing because the analytical solutions of the initial
value problem is known [6]. These solutions represent, for example, nonlinear
wave solutions like shock waves and "humps". Thus, Burgers' equation is interesting as it makes possible the study of the interplay of nonlinear propagation and viscous diffusion.
3.2.1 Shock waves
The first example we are going to treat is the shock wave solution of Burgers'
equation. The shock wave solution is obtained for the following initial condition
I1, x < 0
(46)
u(x,O)=uo(x)=
O, x > 0 "
The diffusion and the convection of this initial step are described by the
following time-dependent solution
u(x, t) =
(47)
erfc ( - x / ~ x )
erfc ( x/-R/ 4t (x - t ) )
(48)
The function erfc(x) is known as the conjugated error function, and is given
by
erfc(x) -
] dy e x p ( - y 2 ) .
(49)
262
H. P. Breuer, F. Petruccione
1.25 -
c n ~
1.00
0.75 -
0.50 -
0.25 -
\
0
-1.00
- 01.75
-01.50
-01,25
0.25
o:5o
0'.75
1:oo
X:
Fig. 2. T h e s t o c h a s t i c s i m u l a t i o n o f t h e s h o c k wave s o l u t i o n o f B u r g e r s ' e q u a t i o n corr e s p o n d i n g to the initial condition Eq. (46) for the three different times tl = 0.1,
(50)
263
The propagation of the single hump is described by the following exact solution
[exp (RA/2) _ 1] exp ( _ R 2)
u(x, t) =
(51)
1 [exp(RA/2'-1]erfc(~tx
In the simulation the initial condition was realized by setting initially N~ = 0 for
all but one cell ,1. = 0 for which N o was chosen such that A = 2, i.e. we have at
t=0:
0(fi~u2~/)
for / l . 0,
=
(52)
N"x(0)
int
, for /l = 0,
where int (y) denotes the integer part of y. In Figure 3 we show the result of the
stochastic simulation together with the analytical result (51) (solid line) for three
different times t = 0.01, t = 0.05, and t = 0.10. The Reynolds number was chosen
to be R = 10. Keeping in mind that a singular initial condition is studied here, the
agreement between the stochastic simulation and the analytical result is striking
although only 201 cells have been used. Furthermore, it should be emphasized
that we did not employ any special technique to obtain Fig. 3, nor did we perform
any analytical pre-integration in order to smooth the singular initial condition,
nor did we use any analytical a priori guess of the solution.
17.5 -
15.0 -
12.5 -
I 10.0-
7.5-
5.0-
2.5-
0
-1.00
'0'.75
-0'.50
-o'.25
0.25
0.50
075
1.~0
Fig. 3. The stochastic simulation of the single hump solution of Burgers' equation corresponding to the initial condition Eq. (50). The velocity is plotted for the three different
times t 1 = 0.01, t2 = 0.05, and t3 = 0.1. Reynolds number: R = 10.
Smooth curves: The analytical result according to Eq. (51).
Symbols: The stochastic simulation using 10 realizations and 201 cells. 5u was chosen to
be 0.01
264
u ( x + 2, t) = u ( x , t ) ,
Ennexp[-4(t+
- ct .
t+l
1) ( x - c t - n )
R
4(t+l)
E nexp
2]
+ c,
(54)
2]
(x-ct-n)
where the sums are extended over the even integers n = 0, 2, 4 . . . . . This
solution represents a periodic array of shocks separated by smooth ramps moving to the right with mean velocity
2
1 ~ dr u ( x , t) = c = const..
(55)
20
From the examples shown above it shuld be clear how to realize the initial
conditions corresponding to the exact solution (54) at t = 0 in a stochastic
simulation. We therefore only explain how to impose the periodic boundary
1.5-
1.0-
0 . 5 "~
T
0-
-0.5
-1.0-
o'.2s
o'.so
o%
.'oo
X"
1 25
.'so
1.'75
2%
i.
Fig. 4. The stochastic simulation of a periodic wave solution of Burgers' equation with
boundary condition (53), mean velocity c = 0.5, and Reynolds number R = 100. The
velocity is plotted for the three different times t 1 = 0.0 (initial condition), t2 = 0.5,
and t3 = 1.0.
Smooth curves: The analytical result according to Eq. (54).
Symbols: The stochastic simulation employing 10 realizations and 100 cells; 5u = 12001
265
266
References
267
6. Whitham, G. B. : Linear and Nonlinear Waves. John Wiley & Sons, New York, 1974
7. Landau, L.D. ; Lifshitz, E.M. : Fluid Mechanics, Volume 6 of Course of Theoretical
Physics. Pergamon Press, London, 1959
8. Breuer, H.P.; Petruccione, E: Hydrostochastics and Hydrodynamic Fluctuations,
preprint THEP 11/92, University of Freiburg (1992)
9. Sabelfeld, K. K.: Monte Carlo Methods in Boundary Value Problems. Springer,
Berlin 1991
10. Marshall, G.: Computer Physics Communications 56 (1989) 51
11. Canuto, C. ; Hussaini, M. Y.; Quarteroni, A. ; Zeng, T. A. : Spectral Methods in
Fluid Dynamics. Springer-Verlag, New York 1987.
12. Smoller, J.: Shock Waves and Reaction-Diffusion Equations. Springer-Verlag
New York 1983
H. P. Breuer
E Petruccione
Alber t-Ludwigs-Universit~t
Fakult~it for Physik
Hermann-Herder-Str. 3
W-7800 Freiburg i. Br.
FRG
Received June 12, 1992