Fluid Mechanics: Numerical Methods. In J.-P. Fran coise, G.L. Naber and S.T. Tsou (eds.

)
Encyclopedia of Mathematical Physics vol. 2, pp. 365374 (2006). Oxford: Elsevier. (ISBN 978-0-1251-
2666-3)
Jean-Luc Guermond, LIMSI, CNRS, University Paris
Sud, Orsay, France and Texas A&M University, Col-
lege Station, Texas
The objective of this article is to give an overview
of some advanced numerical methods commonly
used in uid mechanics. The focus is set primarily
on nite elements methods and nite volume meth-
ods.
1. Fluid Mechanics Models
Let be a domain in R
d
(d = 2, 3) with bound-
ary and outer unit normal n. is assumed to
be occupied by a uid. The basic equations govern-
ing uid ows are derived from three conservation
principles: conservation of mass, momentum, and
energy. Denoting the density by , the velocity by
u , and the mass specic internal energy by e
i
, these
equations are

t
+(u) = 0 (1)

t
(u) +(u u) = + f, (2)

t
(e
i
) +(ue
i
) = : + q
T
j
T
, (3)
where is the stress tensor, =
1
2
(u + u)
T
is
the strain tensor, f is a body force per unit mass
(gravity is a typical example), q
T
is a volume source
(it may model chemical reactions, Joule eects, ra-
dioactive decay, etc.), and j
T
is the heat ux. In ad-
dition to the above three fundamental conservation
equations, one may also have to add L equations
that accounts for the conservation of other quan-
tities, say

, 1 L. These quantities may

be the concentration of constituents in an alloy, the
turbulent kinetic energy, the mass fractions of vari-
ous chemical species by unit volume, etc. All these
conservation equations take the following form
(4)
t
(

)+(u

) = q

, 1 L.
Henceforth the index is dropped to alleviate the
notation.
The above set of equations must be supple-
mented with initial and boundary conditions. Typ-
ical initial conditions are
|t=0
=
0
, u
|t=0
= u
0
,
and
|t=0
=
0
. Boundary conditions are usu-
ally classied into two types: the essential bound-
ary conditions and the natural boundary conditions.
Natural conditions impose uxes at the boundary.
Typical examples are
(n +Ru)
|
= a
u
,
(j
T
n + r
T
e
i
)
|
= a
T
and
(j

n + r

)
|
= a

.
The quantities R, r
T
, r

, a
u
, a
T
, a

are given.
Essential boundary conditions consist of enforcing
boundary values on the dependent variables. One
typical example is the so-called no-slip boundary
condition: u
|
= 0.
The above system of conservation laws is closed
by adding three constitutive equations whose pur-
pose is to relate each eld , j
T
, j

to the elds ,
u, and . They account for microscopic properties
of the uid and thus must be frame independent.
366 Fluid Mechanics: Numerical Methods
Depending on the constitutive equations and ade-
quate hypotheses on time and space scales, various
models are obtained. An important class of uid
model is one for which the stress tensor is a linear
function of the strain tensor, yielding the so-called
Newtonian uid model:
(5) = (p + u)I + 2.
Here p is the pressure, I is the identity matrix, and
and are viscosity coecients. Still assuming
linearity, common models for heat and solute uxes
consist of assuming
(6) j
T
= T, j

= D,
where T is the temperature. These are the so-called
Fouriers law and Ficks law, respectively.
Having introduced two new quantities, namely
the pressure p and the temperature T, two new
scalar relations are needed to close the system.
These are the state equations. One admissible
assumption consists of setting = (p, T). An-
other usual additional hypothesis consists of as-
suming that the variations on the internal energy
are proportional to those on the temperature, i.e.,
e
i
= c
P
T.
Let us now simplify the above models by assum-
ing that is constant. Then, mass conservation im-
plies that the ow is incompressible, i.e., u = 0.
Let us further assume that neither , , nor p de-
pend on e
i
. Then, upon abusing the notation and
still denoting by p the ratio p/, the above set
of assumptions yields the so-called incompressible
NavierStokes equations:
u = 0, (7)

t
u +uu u +p = f. (8)
As a result, the mass and momentum conservation
equations are independent of that of the energy and
those of the solutes:
c
P
(
t
T +uT) (T) = 2: + q
T
, (9)

t
+u
1

(D) =
1

. (10)
Another model allowing for a weak dependency
of on the temperature, while still enforcing incom-
pressibility, consists of setting =
0
(1(T T
0
)).
If buoyancy eects induced by gravity are impor-
tant, it is then possible to account for those by set-
ting f =
0
g(1 (T T
0
)), where g is the grav-
ity acceleration, yielding the so-called Boussinesq
model.
Variations on these themes are numerous and a
wide range of uids can be modeled by using nonlin-
ear constitutive laws and nonlinear state laws. For
the purpose of numerical simulations, however, it is
important to focus on simplied models.
2. The Building Blocks
From the above considerations we now extract a
small set of elementary problems which constitute
the building blocks of most numerical methods in
uid mechanics.
2.1. Elliptic equations. By taking the divergence
of the momentum equation (8) and assuming u to
be known and renaming p to , one obtains the
Poisson equation
(11) = f,
where f is a given source term. This equation plays
a key role in the computation of the pressure when
solving the NavierStokes equations; see (54b). As-
suming adequate boundary conditions are enforced,
this model equation is the prototype for the class
of the so-called elliptic equations. A simple gen-
eralization of the Poisson equation consists of the
advection-diusion equation
(12) u () = f,
where > 0. Admissible boundary conditions are
(
n
+ r)
|
= a, r 0, or
|
= a. This
type of equation is obtained by neglecting the time
derivative in the heat equation (9) or in the solute
conservation equation (10). Mathematically speak-
ing (12) is also elliptic since its properties (in par-
ticular, the way the boundary conditions must be
enforced) are controlled by the second-order deriva-
tives. For the sake of simplicity, assume that u = 0
in the above equation and that the boundary con-
dition is
|
= 0, then it is possible to show that
solves (12) if and only if minimizes the following
functional
J() =
_

(||
2
f) dx,
where | | is the Euclidean norm and spans
(13) H = {;
_

||
2
dx < ;
|
= 0}
Writing the rst-order optimality condition for this
optimization problem yields
_

=
_

f,
Fluid Mechanics: Numerical Methods 367
for all H. This is the so-called variational for-
mulation of (12). When u is not zero, no varia-
tional principle holds but a similar way to refor-
mulate (12) consists of multiplying the equation by
arbitrary functions in H and integrating by parts
the second-order term to give
(14)
_

(u) + =
_

f, H.
This is the so-called weak formulation of (12). Weak
and variational formulations are the starting point
for nite element approximations.
2.2. Stokes equations. Another elementary
building block is deduced from (8) by assuming
that the time derivative and the nonlinear term
are both small. The corresponding model is the
so-called Stokes equations
u +p = f, (15)
u = 0. (16)
Assume for the sake of simplicity that the no-slip
boundary condition is enforced: u
|
= 0. Intro-
duce the Lagrangian functional
L(v, q) =
_

(u:v qv fv) dx.

Set
X = {v;
_

|v|
2
dx < ; v
|
= 0}
M = {q;
_

q
2
dx < }.
Then, the pair (u, p) XM solves the Stokes
equations if and only if it is a saddle-point of L,
i.e.,
(17) L(u, q) L(u, p) L(v, p), (v, q)XM.
In other words, the pressure p is the Lagrange mul-
tiplier of the incompressibility constraint u = 0.
Realizing this fact helps to understand the nature
of the Stokes equations, specially when it comes to
construct discrete approximations. A variational
formulation of the Stokes equations is obtained by
writing the rst-order optimality condition, namely:
_

(u:v pv fv) dx = 0, v X,
_

qudx = 0, q M.
When the nonlinear term is not zero in the momen-
tum equation, or when this term is linearized, there
is no saddle-point, but a weak formulation is ob-
tained by multiplying the momentum equation by
arbitrary functions v in X and integrating by parts
the Laplacian, and by multiplying the mass equa-
tion by arbitrary functions q in M:
_

((uu)v+u:vpv) dx=
_

fv, (18)
_

qu = 0. (19)
2.3. Parabolic equations. The class of elliptic
equations generalizes to that of the parabolic equa-
tions when time is accounted for:
(20)
t
+u () = f,
|t=0
=
0
.
Fundamentally, this equation has many similarities
with the elliptic equation
(21) +u () = f,
where > 0. In particular, the set of bound-
ary conditions that are admissible for (20) and
(21) are identical, i.e., it is legitimate to enforce
(
n
+ r)
|
= a, r 0, or
|
= a. Moreover,
solving (21) is always a building block of any algo-
rithm solving (20). The important fact to remember
here is that if a good approximation technique for
solving (21) is at hand, then extending it to solve
(20) is usually straightforward.
2.4. Hyperbolic equations. When

UL
0,
where U is the reference velocity scale and L is the
reference length scale, (20) degenerates into the so-
called transport equation
(22)
t
+u = f.
This is the prototypical example for the class of
hyperbolic equations. For this equation to be well-
posed it is necessary to enforce an initial condition

|t=0
=
0
and an inow boundary condition, i.e.,

|
= a, where

= {x ; (un)(x) < 0}
is the so-called inow boundary of the domain. To
better understand the nature of this equation, in-
troduce the characteristic lines X(x, s; t) of u(x, t)
dened as follows:
(23)
_
d
t
X(x, s; t) = u(X(x, s; t), t),
X(x, s; s) = x.
If u is continuous with respect to t and Lipschitz
with respect to x, this ODE has a unique solution.
368 Fluid Mechanics: Numerical Methods
Furthermore, (22) becomes
(24) d
t
_
(X(x, s; t), t)

= f(X(x, s; t), t).

Then
(x, t) =
0
(X(x, t; 0)) +
_
t
0
f(X(x, t; ), ) d
provided X(x, t; ) for all [0, t]. This shows
that the concept of characteristic curves is impor-
tant to construct an approximation to (22).
3. Meshes
The starting point of every approximation tech-
nique for solving any of the above model problems
consists of dening a mesh of on which the ap-
proximate solution is dened. To avoid having to
account for curved boundaries, let us assume the do-
main is a two-dimensional polygon (resp. three-
dimensional polyhedron). A mesh of , say T
h
, is a
partition of into small cells hereafter assumed to
be simple convex polygons in two dimensions (resp.
polyhedrons in three dimensions), say triangles or
quadrangles (resp. tetrahedrons or cuboids). More-
over, this partition is usually assumed to be such
that if two dierent cells have a nonempty intersec-
tion, then the intersection is a vertex, or an entire
edge, or an entire face. The left panel of Figure 1
shows a mesh satisfying the above requirement. The
mesh in the right panel is not admissible.
Fig 1: Admissible (left) and nonadmissible (right)
meshes.
4. Finite Elements: Interpolation
The nite element method is foremost an inter-
polation technique. The goal of this section is to
illustrate this idea by giving examples.
Let T
h
= {K
m
}
1mN
el
be a mesh composed of
N
el
simplices, i.e., triangles in two dimensions or
tetrahedrons in three dimensions. Consider the fol-
lowing vector spaces of functions
(25)
V
h
= {v
h
C
0
(); v
h|Km
P
k
, 1 m N
el
},
where P
k
denotes the space of polynomials of global
degree at most k. V
h
is called a nite element ap-
proximation space. We now construct a basis for
V
h
.
Given a simplex K
m
in R
d
, let v
n
be a vertex
of K
m
, let F
n
be the face of K
m
opposite to v
n
,
and dene n
n
to be the outward normal to F
n
,
1 n d + 1. Dene the barycentric coordinates
(26)
n
(x) = 1
(x v
n
) n
n
(v
l
v
n
) n
n
, 1 n d +1,
where v
l
is an arbitrary vertex in F
n
(the denition
of
n
is clearly independent of v
l
provided v
l
be-
longs to F
n
). The barycentric coordinate
n
is an
ane function; it is equal to 1 at v
n
and vanishes
on F
n
; its level-sets are hyperplanes parallel to F
n
.
The barycenter of K
m
has barycentric coordinates
(
1
d + 1
, . . . ,
1
d + 1
).
The barycentric coordinates satisfy the following
properties: For all x K
m
, 0
n
(x) 1, and
for all x R
d
,
d+1

n=1

n
(x) = 1, and
d+1

n=1

n
(x)(x v
n
) = 0.
Consider the set of nodes {a
n,m
}
1nn
sh
of K
m
with barycentric coordinates
_
i
0
k
, . . . ,
i
d
k
_
, 0 i
0
, . . . , i
d
k, i
0
+. . .+i
d
= k.
These points are called the Lagrange nodes of K
m
.
It is clear that there are n
sh
=
1
2
(k + 1)(k + 2) of
these points in two dimensions and n
sh
=
1
6
(k +
1)(k + 2)(k + 3) in three dimensions. It is remark-
able that n
sh
= dimP
k
.
Let {b
1
, . . . , b
N
} =

KmT
h
{a
1,m
, . . . , a
n
sh
,m
}
be the set of all the Lagrange nodes in the mesh.
For K
m
T
h
and n {1, . . . , n
sh
}, let j(n, m)
{1, . . . , N} be the integer such that a
n,m
= b
j(n,m)
;
j(n, m) is the global index of the Lagrange node
a
n,m
. Let {
1
, . . . ,
N
} be the set of functions in
V
h
dened by
i
(b
j
) =
ij
, then it can be shown
that
(27) {
1
, . . . ,
N
} is a basis for V
h
.
The functions
i
are called global shape functions.
An important property of global shape functions is
that their supports are small sets of cells. More pre-
cisely, let i {1, . . . N} and let V
i
= {m; n; i =
j(n, m)} be the set of cell indices to which the node
Fluid Mechanics: Numerical Methods 369
b
i
belongs, then the support of
i
is
mVi
K
m
. For
k = 1, it is clear that
i|Km
=
n
for all m V
i
and
all n such that i = j(n, m), and
i|Km
= 0 other-
wise. The graph of such a shape function in two di-
mensions is shown in the left panel of Figure 2. For
k = 2, enumerate from 1 to d+1 the vertices of K
m
,
and enumerate from d+2 to n
sh
the Lagrange nodes
located at the midedges. For a midedge node of in-
dex d + 2 n n
sh
, let b(n), e(n) {1, . . . , d + 1}
be the two indices of the two Lagrange nodes at
the extremities of the edge in question. Then, the
restriction to K
m
of a P
2
shape function
i
is
(28)
i|Km
=
_

n
(2
n
1), if 1 n d + 1,
4
b(n)

e(n)
, if d + 2 n n
sh
,
Figure 2 shows the graph of two P
2
shape functions
in two dimensions.
Fig 2: Two-dimensional Lagrange shape functions:
Piecewise P1 (left) and Piecewise P2 (center and right).
Once the space V
h
is introduced, it is natural to
dene the interpolation operator
(29)
h
: C
0
() v
N

i=1
v(b
i
)
i
V
h
.
This operator is such that for all continuous func-
tion v, the restriction of
h
(v) to each mesh cell is
a polynomial in P
k
and
h
(v) takes the same val-
ues as v at the Lagrange nodes. Moreover, setting
h = max
KmT
h
diam(K
m
), and dening
r
L
p = (
_

|r|
p
dx)
1
p
for 1 p < ,
the following approximation result holds
(30) v
h
(v)
L
p + h(v
h
(v))
L
p
ch
k+1
v
C
k+1
()
,
where c is a constant that depends on the quality of
the mesh. More precisely, for K
m
T
h
, let
Km
be
the diameter of the largest ball that can be inscribed
into K
m
and let h
Km
be the diameter of K
m
. Then,
c depends on = max
KmT
h
h
Km
/
Km
. Hence, for
the mesh to have good interpolation properties, it is
recommended that the cells be not too at. Fami-
lies of meshes for which is bounded uniformly with
respect to h as h 0 are said to be shape-regular
families.
The above example of nite element approxima-
tion space generalizes easily to meshes composed
of quadrangles or cuboids. In this case, the shape
functions are piecewise polynomial of partial degree
at most k. These spaces are usually referred to as
Q
k
approximation spaces.
5. Finite Elements: Approximation
We show in this section how nite element ap-
proximation spaces can be used to approximate
some model problems exhibited in 2.
5.1. Advection-diusion. Consider the model
problem (21) supplemented with the boundary con-
dition (
n
+ r)
|
= g. Assume > 0,
+
1
2
u 0, and r 0. Dene
a(, ) =
_

(( + ) + ) dx
+
_

r ds.
Then, the weak formulation of (21) is: Seek H
(H dened in (13)) such that for all H
(31) a(, ) =
_

f dx +
_

g ds.
Using the approximation space V
h
dened in
(25) together with the basis dened in (27), we
seek an approximate solution to the above prob-
lem in the form
h
=

N
i=1
U
i

i
V
h
. Then, a
simple way of approximating (31) consists of seek-
ing U = (U
1
, . . . , U
N
)
T
R
N
such that for all
1 i N
(32) a(
h
,
i
) =
_

f
i
dx +
_

g
i
ds.
This problem nally amounts to solving the follow-
ing linear system
(33) AU = F,
where A
ij
= a(
j
,
i
) and F
i
=
_

f
i
dx +
_

g
i
ds. The above approximation technique is
usually referred to as the Galerkin method. The fol-
lowing error estimate can be proved
(34)
h

L
p + h(
h
)
L
p
ch
k+1

C
k+1
()
,
370 Fluid Mechanics: Numerical Methods
where in addition to depending on the shape-
regularity of the mesh, the constant c also depends
on , , and .
5.2. Stokes equations. The line of thought devel-
oped above can be used to approximate the Navier
Stokes problem (15)(16). Let us assume that the
nonlinear termuu is linearized in the formvu,
where v is known. Let T
h
be a mesh of , and as-
sume that nite element approximation spaces have
been constructed to approximate the velocity and
the pressure, say X
h
and M
h
. Assume for the sake
of simplicity that X
h
X and M
h
M. As-
sume that bases for X
h
and M
h
are at hand, say
{
1
, . . . ,
Nu
} and {
1
, . . . ,
Np
}, respectively. Set
a(u, ) =
_

((vu)+ u:dx
and
b(v, ) =
_

v dx.
Then, we seek an approximate velocity u
h
=

Nu
i=1
U
i

i
and an approximate pressure p
h
=

Np
k=1
P
k

k
such that for all i {1, . . . , N
u
} and
all k {1, . . . , N
p
} the following holds
a(u
h
,
i
) + b(
i
, p
h
) =
_

f
i
dx, (35)
b(u
h
,
k
) = 0. (36)
Dene the matrix A R
Nu,Nu
such that A
ij
=
a(
j
,
i
). Dene the matrix B R
Np,Nu
such that
B
ki
= b(
i
,
k
). Then, the above problem can be
recast into the following partitioned linear system
(37)
_
A B
T
B 0
_ _
U
P
_
=
_
F
0
_
,
where the vector F R
Nu
is such that F
i
=
_

f
i
.
An important aspect of the above approximation
technique is that for the linear system to be invert-
ible, the matrix B
T
must have full row rank (i.e., B
has full column rank). This amounts to
(38)
h
> 0, inf
q
h
M
h
sup
v
h
X
h
_

q
h
v
h
dx
v
h

X
q
h

M

h
,
where
v
h

2
X
=
_

|v
h
|
2
dx, q
h

2
M
=
_

q
2
h
dx.
This nontrivial condition is called the
LadyzenskajaBabuskaBrezzi condition (LBB) in
the literature. For instance, if P
1
nite elements
are used to approximate both the velocity and the
pressure, the above condition does not hold, since
there are nonzero pressure elds q
h
in M
h
such
that
_

q
h
v
h
dx = 0 for all v
h
in X
h
. Such elds
are called spurious pressure modes. An example is
shown in Figure 3. The spurious function alterna-
tively takes the values 1, 0, and +1 at the vertices
of the mesh so that its mean value on each cell is
zero.
0
0
0 0
0
0
0
0
0 0
0 0
1
1
1
1
1
1
1
1
1
1
1
1
+1 +1
+1
+1
+1 +1
+1
+1
+1 +1
+1 +1
Fig 3: The P1/P1 nite element: the mesh (left); one
pressure spurious mode (right).
Couples of nite element spaces satisfying the
LBB condition are numerous. For instance, assum-
ing k 2, using P
k
nite elements to approximate
the velocity and P
k1
nite elements to approx-
imate the pressure is acceptable. Likewise using
Q
k
elements for the velocity and Q
k1
elements for
the pressure on meshes composed of quadrangles or
cuboids is admissible.
Approximation techniques for which the pressure
and the velocity degrees of freedom are not associ-
ated with the same nodes are usually called stag-
gered approximations. Staggering pressure and ve-
locity unknowns is common in solution methods for
the incompressible Stokes and NavierStokes equa-
tions; see also 7.3.
6. Finite Volumes: Principles
The nite volume method is an approximation
technique whose primary goal is to approximate
conservation equations, whether time-dependent or
not. Given a mesh, say T
h
= {K
m
}
1mN
el
, and a
conservation equation
(39)
t
+F(, , x, t) = f,
( = 0 if the problem is time-independent and
= 1 otherwise), the main idea underlying every
nite volume method is to represent the approx-
imate solution by its mean values over the mesh
Fluid Mechanics: Numerical Methods 371
cells (
K1
, . . . ,
KN
el
)
T
R
N
el
and to test the con-
servation equation by the characteristic functions
of the mesh cells {1
K1
, . . . , 1
KN
el
}. For each cell
K
m
T
h
, denote by n
Km
the outward unit nor-
mal vector and denote by F
m
the set of the faces
of K
m
. The nite volume approximation to (39)
consists of seeking (
K1
, . . . ,
KN
el
)
T
R
N
el
such
that the function
h
=

N
el
m=1

Km
1
K1m
satises
the following: For all 1 m N
el
(40) |K
m
|d
t

Km
(t) +

Fm
F
m,
h
(
h
,
h

h
, t) =
_
K
f dx,
where
|K
m
| =
_
K
dx,

h
is an approximation of , and F
m,
h
is an
approximation of
_

F(, , x, t)n
Km
d.
The precise denition of the so-called approximate
ux F
m,
h
depends on the the nature of the problem
(e.g., elliptic, parabolic, hyperbolic, saddle point)
and the desired accuracy. In general the approxi-
mate uxes are required to satisfy the following two
important properties:
1. Conservativity: For K
m
, K
l
T
h
such that
= K
m
K
l
, F
m,
h
= F
l,
h
.
2. Consistency: Let be the solution to (39), and
set
h
=
1K
1
|K1|
_
K1
dx+. . .+
1K
N
el
|KN
el
|
_
KN
el
dx, then
F
m,
h
(
h
,
h

h
, t)
h0

F(, , x, t)nd.
The quantity
|F
m,
h
(
h
,
h

h
, t)
_

F(, , x, t)nd|
is called the consistency error.
Note that (40) is a system of Ordinary Dier-
ential Equations (ODEs). This system is usually
discretized in time by using standard time-marching
techniques such as explicit Euler, Runge-Kutta, etc.
The discretization technique above described is
sometimes referred to as cell-centered nite volume
method. Another method, called vertex-centered
nite volume, consists of using the characteristic
functions associated with the vertices of the mesh
instead of those associated with the cells.
7. Finite Volumes: Examples
In this section we illustrate the ideas introduced
above. Three examples are developed: the Poisson
equation, the transport equation, and the Stokes
equations.
7.1. Poisson problem. Consider the Poisson
equation (11) equipped with the boundary condi-
tion
n

|
= a. To avoid technical details, assume
that =]0, 1[
d
. Let K
h
be a mesh of composed
of rectangles (or cuboids in three dimensions).
The ux function is F(, , x) = ; hence,
F
m,
h
must be a consistent conservative approxima-
tion of
_

n
Km
d. Let be an interior face of
the mesh and let K
m
, K
l
be the two cells such that
= K
m
K
l
. Let x
Km
, x
K
l
be the barycenters
of K
m
and K
l
, respectively. Then, an admissible
formula for the approximate ux is
(41) F
m,
h
=
||
|x
Km
x
K
l
|
(
K
l

Km
),
where || =
_

d. The consistency error is O(h)

in general, and is O(h
2
) if the mesh is composed of
identical cuboids. The conservativity is evident. If
is part of , an admissible formula for the ap-
proximate ux is F
m,
h
=
_

a d. Then, upon
dening F
i
Km
= F
Km
\ and F

Km
= F
Km
,
the nite volume approximation of the Poisson
problem is: Seek
h
R
N
el
such that for all
1 m N
el
(42)

F
i
Km
F
m,
h
=
_
Km
f dx +

Km
_

a d.
7.2. Transport equation. Consider the transport
equation

t
+(u) = f, (43)

|t=0
=
0
,
|
= a, (44)
where u(x, t) is a given eld in C
1
( [0, T]). Let
T
h
be a mesh of . For the sake of simplicity, let
us use the explicit Euler time-stepping to approxi-
mate (40). Let N be positive integer, set t =
T
N
,
set t
n
= nt for 0 n N, and partition [0, T] as
follows
[0, T] =
N1
_
n=0
[t
n
, t
n+1
].
372 Fluid Mechanics: Numerical Methods
Denote by
n
h
R
N
el
the nite volume approxi-
mation of
h
(t
n
). Then, (40) is approximated as
follows
|Km|
t
(
n+1
Km

n
Km
) +

Fm
F
m,
h
(
h
,
h

h
, t
n
) =
_
K
f(x, t
n
) dx, (45)
where
0
Km
=
_
Km

0
dx. The approximate ux
F
m,
h
must be a consistent conservative approxi-
mation of
_

(un
Km
)d. Let be a face of the
mesh and let K
m
, K
l
be the two cells such that
= K
m
K
l
(note that if is on , belongs to
one cell only and we set K
m
= K
l
). If is on

,
set
(46) F
m,
h
=
_

(un
Km
)a d.
If is not on

, set u
n
m,
=
_

(un
Km
) d and
dene
(47) F
m,
h
=
_

n
Km
u
n
m,
if u
n
m,
0,

n
K
l
u
n
m,
if u
n
m,
< 0.
The above choice for the approximate ux is usu-
ally called the upwind ux. It is consistent with the
analysis that has been done for (22), i.e., informa-
tion ows along the characteristic lines of the eld
u; see (24). In other words, the updating of
n+1
km
must be done by using the approximate values
n
h
coming from the cells that are upstream the ow
eld.
An important feature of the above approxima-
tion technique is that it is L

-stable in the sense

that
max
0nN,1mN
el
|
n
Km
| c(u
0
, f)
if the two mesh parameters t and h satisfy the
so-called CourantFriedrichsLevy (CFL) condition
u
L
t/h c(), where c() is a constant that
depends on the mesh regularity parameter =
max
KmT
h
h
Km
/
Km
. In one dimension c() = 1.
7.3. Stokes equations. To nish this short review
of nite volume methods, we turn our attention to
the Stokes problem (15)(16) equipped with the ho-
mogeneous Dirichlet boundary condition u
|
= 0.
Let T
h
be a mesh of composed of triangles (or
tetrahedrons). All the angles in the triangulation
are assumed to be acute so that for all K T
h
, the
intersection of the orthogonal bisectors of the sides
of K, say x
K
, is in K. We propose a nite volume
approximation for the velocity and a nite element
approximation for the pressure. Let {e
1
, . . . , e
d
} be
a Cartesian basis for R
d
. Set 1
k
Km
= 1
Km
e
k
for all
1 m N
el
and 1 k d, then dene
X
h
= span{1
1
K1
, . . . , 1
d
K1
, . . . , 1
1
KN
el
, . . . , 1
d
KN
el
}.
Let {b
1
, . . . , b
Nv
} be the vertices of the mesh, and
let {
1
, . . . ,
Nv
} be the associated piecewise linear
global shape functions. Then, set
N
h
= span{
1
, . . . ,
Nv
},
M
h
= {q N
h
;
_

q dx = 0};
see 4. The approximate problem consists of seek-
ing (u
K1
, . . . , u
KN
el
) R
dN
el
and p
h
M
h
such
that for all 1 m N
el
, 1 k d, and all
1 i N
v
,

Fm
1
k
Km
F
m,
h
+ c(1
k
Km
, p
h
) =
_
Km
1
k
Km
f dx, (48)
c(u
Km
,
i
) = 0, (49)
where
c(v
Km
, p
h
) =
_
Km
v
Km
p
h
dx.
Moreover,
F
m,
h
=
_

_
||
|x
m
x
l
|
(u
Km
u
K
l
) if = K
m
K
l
,
||
d(x
m
, )
u
Km
if = K
m
,
where d(x
Km
, ) is the Euclidean distance between
x
Km
and . This formulation yields a linear sys-
tem with the same structure as in (37). Note in
particular that
(50) sup
v
h
X
h
c(v
h
, p
h
)
v
h

= p
h

L
1,
Since the mean-value of p
h
is zero, p
h

L
1 is a
norm on M
h
. As a result, an inequality similar to
(38) holds. This inequality is a key step to proving
that the linear system is wellposed and the approx-
imate solution converges to the exact solution of
(15)(16).
Fluid Mechanics: Numerical Methods 373
8. Projection methods for NavierStokes
In this section we focus on the time approxima-
tion of the NavierStokes problem:

t
u u +uu +p = f, (51a)
u = 0, (51b)
u
|
= 0, (51c)
u
|t=0
= u
0
. (51d)
where f is a body force and u
0
is a solenoidal ve-
locity eld. There are numerous ways to discretize
this problem in time, but, undoubtedly, one of the
most popular strategies is to use projection meth-
ods, sometimes also referred to as ChorinTemam
methods.
A projection method is a fractional-step time-
marching technique. It is a predictorcorrector
strategy aiming at uncoupling viscous diusion and
incompressibility eects. One time step is com-
posed of three substeps: in the rst substep, the
pressure is made explicit and a provisional veloc-
ity eld is computed using the momentum equa-
tion; in the second substep, the provisional velocity
eld is projected onto the space of incompressible
(solenoidal) vector elds; in the third substep, the
pressure is updated.
Let q > 0 be an integer and approximate the time
derivative of u using a backward dierence formula
of order q. To this end, introduce a positive integer
N, set t =
T
N
, set t
n
= nt for 0 n N, and
consider a partitioning of the time interval in the
form
[0, T] =
N1
_
n=0
[t
n
, t
n+1
].
For all sequences v
t
= (v
0
, v
1
, . . . , v
N
), set
(52) D
(q)
v
n+1
=
q
v
n+1

q1

j=0

j
v
nj
,
where q 1 n N 1. The coecients
j
are
such that
1
t
(
q
u(t
n+1
)
q1

j=0

j
u(t
nj
))
is a qth-order backward dierence formula approx-
imating
t
u(t
n+1
). For instance,
D
(1)
v
n+1
= v
n+1
v
n
,
D
(2)
v
n+1
=
3
2
v
n+1
2v
n
+
1
2
v
n1
.
Furthermore, for all sequences
t
=
(
0
,
1
, . . . ,
N
), dene
(53)
,n+1
=
q1

j=0

nj
,
so that

q1
j=0

j
p(t
nj
) is a (q1)th-order extrapo-
lation of p(t
n+1
). For instance, p
,n+1
= 0 for q = 1,
p
,n+1
= p
n
for q = 2, and p
,n+1
= 2p
n
p
n1
for
q = 3. Finally, denote by (uu)
,n+1
a q-th order
extrapolation of (uu)(t
n+1
). For instance,
(uu)
,n+1
=
_
u
n
u
n
if q = 1,
2u
n
u
n
u
n1
u
n1
if q = 2.
A general projection algorithm is as follows. Set
u
0
= u
0
and
l
= 0 for 0 l q 1. If q > 1,
assume that u
1
, . . . , u
q1
, p
,q
and (uu)
,q
have
been initialized properly. For n q 1, seek u
n+1
such that u
n+1
|
= 0 and
(54a)
D
(q)
t
u
n+1
u
n+1
+
_
p
,n+1
+

q1
j=0
j
t

nj
_
= S
n+1
,
where S
n+1
= f(t
n+1
) (uu)
,n+1
. Then solve
(54b)
n+1
= u
n+1
,
n

n+1
|
= 0.
And nally update the pressure as follows:
(54c) p
n+1
=
q
t

n+1
+ p
,n+1
u
n+1
.
The algorithm (54a)(54b)(54c) is known in the
literature as the rotational form of the pressure-
correction method. Upon denoting u
t
=
(u(t
0
), . . . , u(t
N
)) and p
t
= (p(t
0
), . . . , p(t
N
)), the
above algorithm has been proved to yield the follow-
ing error estimates
u
t
u
t

2
(L
2
)
ct
2
,
(u
t
u
t
)

2
(L
2
)
+p
t
p
t

2
(L
2
)
ct
3
2
.
where
t

2
(L
2
)
= t

N
n=0
_

|
n
|
2
dx.
A simple strategy to initialize the algorithm con-
sists of using D
(1)
u
1
at the rst step in (54a), then
using D
(2)
u
2
at the second step, and proceeding
likewise until u
1
, . . . , u
q1
have all been computed.
At the present time, projection methods count
among the few methods that are capable of solving
the time-dependent incompressible NavierStokes
equations in three dimensions on ne meshes within
reasonable computation times. The reason for this
374 Fluid Mechanics: Numerical Methods
success is that the unsplit strategy, which consists
of solving
D
(q)
t
u
n+1
u
n+1
+p
n+1
= S
n+1
, (55a)
u
n+1
= 0; u
n+1
|
= 0, (55b)
yields a linear system similar to (37), which usu-
ally takes far more time to solve than sequentially
solving (54a) and (54b). It is commonly reported
in the literature that the ratio of the CPU time for
solving (55a)(55b) to that for solving (54a)(54b)
(54c) ranges between 10 to 30.
See Also
Fluid Dynamics. Incompressible Euler equa-
tions. Newtonian uids and thermohydraulics.
Navier-Stokes turbulence. Partial dierential equa-
tions in uid mechanics. Variational methods in
turbulence.
Further Reading
Doering, C.R.; Gibbon, J.D. (1995) Applied
analysis of the Navier-Stokes equations. Cambridge
Texts in Applied Mathematics. Cambridge: Cam-
bridge University Press.
Ern A. and Guermond J.-L. (2004) Theory and
Practice of Finite Elements. Springer Series in Ap-
plied Mathematical Sciences, Vol. 159. New-York:
Springer-Verlag.
Eymard R., Gallouet T., and Herbin, R. (2000)
Finite volume methods. In: Ciarlet PG and Lions
JL (eds.) Handbook of numerical analysis, Vol. VII
7131020. Amsterdam: North-Holland.
Karniadakis, G.E.; Sherwin, S.J. (1999)
Spectral/hp element methods for CFD. Numeri-
cal Mathematics and Scientic Computation. New
York: Oxford University Press.
Rappaz M., Bellet M. and Deville M. (2003) Nu-
merical Modeling in Material Science and Engineer-
ing. Springer Series in Computational Mathemat-
ics, Vol. 32. Berlin: Springer-Verlag.
Temam, R. (1984) Navier-Stokes equations.
Theory and numerical analysis. Studies in Math-
ematics and its Applications, Vol. 2. Amsterdam:
North-Holland.
Toro, E.F. (1997) Riemann solvers and numeri-
cal methods for uid dynamics. A practical intro-
duction. Berlin: Springer.
Wesseling, P. (2001) Principles of computational
uid dynamics. Springer Series in Computational
Mathematics, Vol. 29. Berlin: Springer.