Peter Hansbo

Applied Partial Dierential

Equations for Computational
Science and Engineering
Part II: Nonlinear models in continuum
mechanics
March 10, 2011
Contents
1. Material Models in Small Deformation Solid Mechanics . . 1
1.1 Viscous stressstrain relations . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1
1.1.1 Maxwell materials . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1
1.1.2 Kelvin materials . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 2
1.1.3 Generalization to several dimensions . . . . . . . . . . . . . . . . 3
1.2 Elastoplastic stressstrain relations . . . . . . . . . . . . . . . . . . . . . . . 4
1.2.1 An elastoviscoplastic model . . . . . . . . . . . . . . . . . . . . . . . 9
2. General Principles in Continuum Mechanics . . . . . . . . . . . . . . 11
2.1 Velocity of a continuum . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 11
2.2 Flux through a surface . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 12
2.3 Conservation of mass . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 13
2.4 Reynolds transport theorem . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 14
2.5 Newtons equations of motion in a continuum setting . . . . . . . 17
3. Large Deformation Solid Mechanics. . . . . . . . . . . . . . . . . . . . . . . 19
3.1 Strain measures . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 19
3.1.1 The deformation gradient and a general approach to
strain . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 21
3.1.2 The boundary value problem on the deformed domain 23
3.1.3 Stressstrain relations in large deformation theory . . . . 23
4. Thermodynamics . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 27
4.1 Energy balance and the rst law of thermodynamics . . . . . . . . 27
4.2 Entropy and the second law of thermodynamics . . . . . . . . . . . . 29
5. Fluids . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 33
5.1 Inviscid compressible uids and conservation laws . . . . . . . . . . 33
5.2 Nonlinear conservation laws and shocks . . . . . . . . . . . . . . . . . . . 37
5.3 Linearization of the Euler equations and acoustics . . . . . . . . . . 39
5.4 Inviscid incompressible uids . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 40
5.5 Viscous incompressible uids . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 43
5.6 Turbulence . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 45
4 Contents
6. Stability problems . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 47
6.1 A simple example . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 47
6.2 The beam . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 50
6.3 Instability and the minimum of potential energy . . . . . . . . . . . . 54
1. Material Models in Small Deformation Solid
Mechanics
In this Chapter we extend the simple Hookes law to linear viscous models
and nonlinear elastoplastic modes.
1.1 Viscous stressstrain relations
We recall the simple Hookes law in one dimension:
= E = E
du
dx
where is the stress (force per unit area) and is the strain (increase in length
divided by initial length), and u is local displacement. We can view this law
as a generalization of an elastic spring and visualize the model as in Fig.
1.1 (left). In a viscous material, like a Newtonian uid mentioned in Applied
Partial Dierential Equations for Computational Science and Engineering,
Part I: Linear models in continuum mechanics (henceforth called Part I ), it
is instead the strain velocity that is proportional to the stress, so that
d
dt
=
1

, or =
1

, or
d u
dx
=
1

,
where (in solid mechanics) is known as the modulus of viscosity. This ma-
terial can be represented as a damper, Fig 1.1 (right). These two models can
now be combined in various ways to create ever more complex material mod-
els in order to try to predict realistic material behavior. We shall only give
the two simplest possible models.
1.1.1 Maxwell materials
The Maxwell material can be viewed as a damper and a spring coupled in
series, cf. Fig 1.2 (left). The the stress in the spring part,
s
, and the damper
part
d
, must be equal due to equilibrium,
=
s
=
d
, (1.1)
and the total strain is given by
2 1. Material Models in Small Deformation Solid Mechanics
Fig. 1.1. Spring model (left) end damper model (right).
=
s
+
d
. (1.2)
The constitutive relations are

s
= E
s
, (1.3)
and

d
=
1

d
. (1.4)
Thus,
=
1
E
+
1

. (1.5)
In this material, a loadingunloading cycle leaves a residual deformation (in
the damper part) that does not go away, only the deformation in the spring
part returns to its initial state.
1.1.2 Kelvin materials
Many viscous materials return to their original state after deformation (albeit
slowly). In order to model such materials, one can couple a spring and a
damper in parallel, see Fig. 1.2 (right). This is the Kelvin model, in which
the strain in the damper and the spring are equal,
=
s
=
d
, (1.6)
and the stress is the sum of stresses in the spring and in the damper,
=
s
+
d
. (1.7)
The constitutive relations are the same as for the Maxwell material, and thus
+E = . (1.8)
In this model, a loadingunloading cycle leaves a residual stress in the spring
which slowly takes the deformation back to its initial state.
1.1 Viscous stressstrain relations 3
E
E
Fig. 1.2. Maxwell model (left) and Kelvin model (right).
1.1.3 Generalization to several dimensions
The generalization of Kelvins onedimensional model to threedimensions is
straightforward. Hookes law for an isotropic material states that (cf. Part I )
= 2(u) + uI = 2(u) +tr (u) I,
where we recall that the trace of the strain matrix, tr , means the sum of
the diagonal elements of . A generalization to a Kelvin model could thus be
(t) = ( u) + 2(u) +tr (u) I. (1.9)
This constitutive relation must then be combined with the equilibrium equa-
tions
(t) = f(t),
and the problem is that of an initialboundarvalue problem in u. The cor-
responding extension of a Maxwell material to get an initialboundaryvalue
problem for u is more involved because of the presence of in the constitutive
law. We then rst need to invert Hookes law to obtain
(u) =
1
2


2(3 + 2)
tr , (1.10)
and a Maxwell constitutive model could be
( u) =
1

+
1
2


2(3 + 2)
tr . (1.11)
Viscoelastic models are thus linear, but time dependent. We next turn to
a model which is nonlinear.
4 1. Material Models in Small Deformation Solid Mechanics
1.2 Elastoplastic stressstrain relations
Elastic materials exhibit a linear relation between stress and strain. There is,
however, a limit to the load a material can carry before it breaks. In metals,
the elastic regime ends when atomic planes starts sliding along each other,
and irreversible strains occur. In Fig. 1.3 the principle is illustrated. Once
the stressstrain curve passes point 2, irreversible strains will occur. Upon
unloading, the stress-strain curve will return to = 0 along a straight line,
leaving residual strains. The actual graph of the function = () is complex
and diers from material to material, and there is a need to make simplifying
assumptions. A simple model for plasticity is to postulate that the material
Fig. 1.3. Stressstrain relations in metals.
can only carry a stress below a xed value
y
, the yield stress, and assume the
idealized stressstrain relation shown in Fig. 1.4. The extension of this simple
model to three dimensions can be done in various ways. Experimentally, one
nds that the material does not yield (the atomic planes do not slip) in
case of pure compression (pure volume change). It is therefore sometimes
assumed that the shear forces alone are responsible for plastic yield. It turns
out however, that a mathematically more pleasing theory can be developed
by splitting the stress matrix into the part responsible for volumetric change
and the part responsible for change of form. By Hookes law we have
= 2(u) + uI,
1.2 Elastoplastic stressstrain relations 5
Fig. 1.4. Idealized plasticity model in one dimension.
and we know that u can be seen as the volume change (cf. Part I ). Taking
the trace of the left hand side and of the right hand side, followed by division
by 3, we obtain the mean (or volumetric) stress

V
:=
1
3
(
11
+
22
+
33
) =
2
3
u + u := K u
where K = + 2/3 is called the bulk modulus. What remains, the stress
deviator

D
=
V
I (= 2
D
from Hookes law),
which has zero trace, is thus responsible for change of shape. Standard plas-
ticity theory thus typically posits that there is a limit to the maximum size
of the stress deviator. Richard von Mises suggested that we take as criterion
|
D
| =

D
:
D

y
_
2
3
,
where
y
is, as above, the yield stress in uniaxial tension. Note that the
left-hand side is an invariant and does not depend on the chosen coordinate
system.
It is possible to visualize von Mises yield criterion if we use the principal
stresses (the eigenvalues of the stress matrix) as coordinate axes. We are thus
looking at the case when the physical coordinates are chosen so that
=
_
_

1
0 0
0
2
0
0 0
3
_
_
.
If all stresses are equal,
1
=
2
=
3
=: , then the volumetric stress
V
=
and, since
6 1. Material Models in Small Deformation Solid Mechanics

D
=
_
_
0 0
0 0
0 0
_
_
I =
_
_
0 0 0
0 0 0
0 0 0
_
_
,
there is no yield whatever the size of . Otherwise we nd

D
=
_
_

1
0 0
0
2
0
0 0
3
_
_

1
3
_
_

1
+
2
+
3
0 0
0
1
+
2
+
3
0
0 0
1
+
2
+
3
_
_
or

D
=
1
3
_
_
2
1

3
0 0
0 2
2

3
0
0 0 2
3

2
_
_
and thus

D
:
D
=
1
9
_
(2
1

3
)
2
+ (2
2

3
)
2
+ (2
3

2
)
2
_
.
A closer inspection shows that the equation for the yield surface,
1
9
_
(2
1

3
)
2
+ (2
2

3
)
2
+ (2
3

2
)
2
_
=
2
2
y
3
,
describes a slanted cylinder in the space of principal stresses shown in Fig.
1.5. Thus, in von Mises model, the principal stresses are conned to stay
within this cylinder.
In many texts on plasticity, one employs the yield function, in our case
dened as
() :=

D
:
D

y
_
2
3
,
and then the stress must reside in the set E of admissible stresses dened by
E = { : () 0},
i.e., inside the cylinder of Fig. 1.5. We note that this is a convex set and thus
tools of convex analysis can be used.
The simplest model for plasticity is the Hencky (or total deformation)
model. In standard engineering notation this model may be formulated as
follows: for a given body load f (and given boundary conditions), nd the
displacement eld u such that
= f (equilibrium)
= 2
e
+tr
e
I (constitutive equation for stress)
()
ij
=
1
2
_
u
i
x
j
+
u
j
x
i
_
(small strain)
=
e
+
p
(additive decomposition, cf. (1.2))

p
= g() (constitutive equation for plastic strain)
1.2 Elastoplastic stressstrain relations 7
Fig. 1.5. Yield surface in the space of principal stresses.
where
e
is the elastic strain,
p
is the plastic strain, and g is the over-
stress function (to be dened). The Hencky model is suitable for proportional
loading only; for general loading/unloading, these equations have to be time
dependent, replacing u by u, strains by strain rates, etc.
Now, for a convex set E, there must hold, for E, where E is the
surface of E,
( ) : q 0 E
if q is an outward-pointing normal to E, cf. Fig 1.6. This simple fact
of convex sets has been given a mechanical interpretation as the Maximum
Plastic Dissipation (MPD) property: the actual stress in a plastic state,
E, is the one, among all admissible stresses E, leading to maximum
plastic dissipation D, where
D :
p
0.
Clearly, D 0 is implied by thermodynamics since it physically means a
dissipation of energy (the stresses must perform work when rupturing the
material and this work cannot be regained). The maximization of D, however,
is a physical assumption that may or may not hold for a given material. We
may write MPD as
:
p
:
p
E
or
8 1. Material Models in Small Deformation Solid Mechanics
( ) :
p
0 E,
and thus we can identify q =
p
; in other word:
p
must by necessity be
normal to the yield surface. This means that

p
=

()|
=0
(or g() =

()|
=0
),
where is called the plastic multiplier. (The reason that is called a mul-
tiplier lies in the fact that we have a minimization problem (of potential
energy) with a side condition () 0, which can be solved using Lagrange
multiplier techniques).
Fig. 1.6. Convex set of admissible stresses.
A very simple way of approaching the problem of obtaining a constitutive
law for the plasticity problem, without the use of multipliers, is to introduce
the projection onto the yield surface, i.e.,
(
D
) =
_

D
if () 0,

D
|
D
|

y
_
2
3
if () > 0.
This corresponds to a projection () of as () = (
D
) +
V
I which
we note is always in E. Alternatively, we can view the projection as an elasto-
plastic constitutive relation and formulate our original plasticity problem as
that of nding the displacement eld u such that
1.2 Elastoplastic stressstrain relations 9
= f (equilibrium)
= (2
D
(u)) +
3 + 2
3
tr (u) I (constitutive equation for stress)
()
ij
=
1
2
_
u
i
x
j
+
u
j
x
i
_
(small strain)
and need not concern ourselves with plastic strains at all. All that needs to
be done is to project the deviatoric part of the elastic stress onto the yield
surface. This approach is known as the radial return method (we scale the
elastic stress radially back to the yield surface) and can only be used for
plasticity models where the yield surface is convex and where MPD holds.
We remark that plasticity models with non-convex yield surfaces as well
as models violating MPD are in use, in particular for friction materials such
as sand. The basic idea of projecting back the elastic stress to a yield surface
is often used in these cases as well, albeit in a more complicated fashion.
The elastic strains can be recovered from (1.10) and the plastic strains
from the dierence between total strain (computed fromu) and elastic strain.
1.2.1 An elastoviscoplastic model
The projection operator onto the yield surface can also be used in a penalty
method. Starting from a time-dependent setting and using the additive split
( u) =
e
+
vp
, where
vp
is a viscoplastic strain, we can argue as in the
case of a Maxwell material and dene
( u) =
1
2


2(3 + 2)
tr +
1

( ()) .
This is known as the DuvautLions model of elasto-viscoplasticity. Note that

vp
=
1

( ())
_
=
1

D
(
D
)
_
, the mean stress cancels out
_
which takes the place of the viscous strain rate / in Maxwells model,
forces onto the yield surface when 0 in order to stay nite. In this
sense, the viscoplastic model is close to the von Mises plasticity model. The
constitutive parameter is known as the relaxation time.
The main dierence between this viscoplastic model and the plasticity
model from which it is derived lies in the fact that in the former we have an
approximation of the gradient of the yield function (in the time dependent
plasticity model) so that

p
=

( ()) / =
vp
.
This approximation makes the problem more regular than the plasticity prob-
lem, and thus easier to solve numerically.
2. General Principles in Continuum Mechanics
In this Chapter we introduce the general conservation and balance laws that
hold in all continuum mechanical theories. They thus form the basis for uid
and solid mechanics in general.
2.1 Velocity of a continuum
Two types of description of motion are mainly used: referential description
and spatial description. The referential description uses a coordinate system
of material coordinates X referring typically to the conguration occupied
by the continuum in question at a reference time, often t = 0. A referential
coordinate of a particle thus remains constant during motion through space.
If we use the coordinates of space itself, through which the object moves,
then we talk of a spatial description with coordinates x. So we can say that
x is the place occupied by the particle which was originally at X, and we
may write
x = x(X, t).
Note that the material coordinates X do not depend on time! X and x may
or may not be measured with respect to the same coordinate axes.
The spatial description is very important in uid mechanics, where one
usually looks at xed volumes in space through which the uid ows. Often
the velocity in a uid is viewed a a primal variable, and as a spatial quantity,
v = v(x, t),
but if we want to dene velocity, we must look at particles and their velocities.
Thus the velocity is dened as a function of material coordinates,
v = v(X, t) :=
x(X, t)
t
.
Clearly
v(X, t) v(x(X, t), t),
so by the chain rule
12 2. General Principles in Continuum Mechanics
v(X, t)
t
=
v(x, t)
t
+
x(X, t)
t
v(x, t) =
v
t
+v v =:
Dv
Dt
,
here we reintroduced v = v(x, t) and introduced the notation D/Dt for
the material time derivative obtained when following the motion (the time
derivative obtained when keeping the reference coordinate xed).
2.2 Flux through a surface
Fig. 2.1. Flux through an innitesimal surface element dS during an innitesimal
time interval dt.
In Fig. 2.1 we show a small element dS of a ctitious surface S, xed in
space, with a uid medium owing through it. A normal to the surface is
denoted n, and the uid velocity is denoted v. By denition of velocity, we
see that during a time interval dt, particles initially on dS will move out to
form a slanted cylinder with side length |v|dt. The height of the cylinder is
immediately found as v ndt, since v n is the length of v in the direction n.
Hence the volume of material owing through dS in time dt equals v ndt dS.
The amount of volume per unit time owing through dS, v ndS, in known
as the ux of volume through dS. The total ux of volume through S is thus
given by
Volume ux =
_
S
v ndS.
Similarly, we can compute the total mass per unit time owing through the
surface. Since mass equals density times volume, we have
2.3 Conservation of mass 13
Mass ux =
_
S
v ndS,
where is the density of the uid. In general, any quantity carried by the
ow (energy, heat, momentum, etc.), has a corresponding ux, so that
Flux of =
_
S
v ndS.
Here is not necessarily a scalar, momentum for example is a vectorial
quantity. Typically we are interested in quantities per unit mass, and then
the ux is instead dened as
Flux of =
_
S
v ndS.
If the surface is closed and the normal points outwards from S, then the
integral denotes the total net outow of x through S. In Part I we addressed
heat conduction. There we had two uxes, diusive ux and convective ux.
The convective ux was of the type discussed here, i.e., heat carried by uid
ow. The diusive ux, on the other hand, is a nonmechanical ux, which
of course also can occur; another such example is electric current ow. Thus,
nonmechanical uxes do not require motion of the medium in order to occur.
In some respects they can still be treated as if they were uid ows; indeed,
heat was originally perceived as a uid, caloric.
2.3 Conservation of mass
Consider an arbitrary volume V , xed in space, with surface S. The total
mass M contained in the volume is given by
M =
_
V
dV.
If the density in the volume changes with time, then so does the mass, and
we have
dM
dt
=
_
V

t
dV.
If mass is not created or destroyed inside V , then the rate of change of mass
must also equal the total ux of mass through S. If there is an increase of
mass, so that

t
is a positive quantity, then the rate of inow of mass must
equal the rate of change of mass, so

_
S
v ndS =
_
V

t
dV,
or, by use of Gauss divergence theorem,
14 2. General Principles in Continuum Mechanics

_
V
( v) dV =
_
V

t
dV,
and, since V is completely arbitrary, we obtain the continuity equation

t
+ ( v) = 0. (2.1)
If we carry out the spatial dierentiation,

x
i
(v
i
) =

x
i
v
i
+
v
i
x
i
,
this can be written

t
+v + v = 0. (2.2)
Recalling the material time derivative,
D
Dt
:=

t
+v ,
we can write the continuity equation as
D
Dt
+ v = 0. (2.3)
If the uid is incompressible, so that the density of an innitesimal volume
is constant as it moves, then the continuity equation takes the simpler form
v = 0. (2.4)
This is also known as the incompressibility condition.
2.4 Reynolds transport theorem
In solid mechanics, the material is moving through space, so an important
question is how to evaluate the material time derivative of integrals over
moving volumes V (t), volumes that contain a given mass system (think of
V (t) as the volume of an object). For example, by conservation of mass
D
Dt
_
V (t)
dV = 0,
since the total mass of a body does not change as we follow its motion.
In Reynolds theorem we think of a particular time t when V (t) is sur-
rounded by a surface S. If this S is xed in time, then S is the boundary of
V (t) only at the particular time t when the volume of mass passes though it.
In that case we have, for a quantity per unit mass (scalar, vector, tensor)
2.4 Reynolds transport theorem 15
_
V (t)

t
( ) dV =
D
Dt
_
V (t)
dV
_
S
v ndS (2.5)
or: the rate of increase of inside the xed surface S equals the rate of
increase of at this moment inside S minus the ux of through S.
The derivation of this theorem diers between uid and solid mechanics.
In uid mechanics, one rst denes a control volume, xed in space, through
which the material ows. In solid mechanics, this control volume is simply
the object (or a part of an object) at a particular time t, in which case the
surface S is simply the boundary of V (t), S = V (t). Here, we will give the
standard uid mechanics explanation for the transport theorem.
Fig. 2.2. A set of material points, a system, passing through a control volume
(between the dotted lines) at time t and at time t +t.
In Figure 2.2 we have a system occupying a volume V (t) going through
a control volume V
CV
. The total amount of in the system is denoted by
B
sys
, so that
16 2. General Principles in Continuum Mechanics
B
sys
(t) :=
_
V (t)
dV,
and the amount of within the control volume is denoted by
B
CV
(t) :=
_
VCV
dV.
Note that B
CV
is also a time dependent quantity since the amount of inside
the control volume may change in time. The objective is to compute
D
Dt
B
sys
(t) = lim
t0
B
sys
(t +t) B
sys
(t)
t
.
Now, obviously be have that
B
sys
(t) = B
CV
(t) +B
in
,
and
B
sys
(t +t) = B
CV
(t +t) +B
out
,
where B
out
is the amount of that is leaving the control volume and B
in
is the amount of that going in to the control volume. Thus
D
Dt
B
sys
(t) = lim
t0
B
CV
(t +t) +B
out
B
CV
(t) B
in
t
or
D
Dt
B
sys
(t) = lim
t0
_
B
CV
(t +t) B
CV
(t)
t
+
B
out
t

B
in
t
_
(2.6)
Now, since the control volume is xed in space, we have
lim
t0
B
CV
(t +t) B
CV
(t)
t
=
_
VCV

t
( ) dV,
the local rate of change of inside the control volume. We recall from Chapter
2.2 that the amount of that ows through a surface S over time t can be
found as
B = t
_
S
v ndS
The dierent signs on B
in
and B
out
in (2.6) reects the dierent signs on
v n, and we arrive nally at
D
Dt
_
V (t)
dV =
_
VCV

t
( ) dV +
_
S
v ndS
which, if we consider the instant in time when V
CV
= V (t), gives Reynolds
transport theorem (2.5).
2.5 Newtons equations of motion in a continuum setting 17
By Gauss theorem, we can rewrite (2.5) as
D
Dt
_
V (t)
dV =
_
V (t)
_

t
( ) + ( v)
_
dV, (2.7)
where we interpret
( v) :=

x
j
( v
j
).
We can exemplify by setting = 1, in which case
0 =
D
Dt
_
V (t)
dV =
_
V (t)
_

t
+ (v)
_
dV,
which again gives the continuity equation (2.1).
By rearranging terms, we nd that the integral on the right in (2.7) can
be written
_
V (t)
_

t
+v
_
+
_

t
+ (v)
__
dV.
By the continuity equation, the second expression equals zero, and we obtain
Reynolds transport theorem in its simplest form
D
Dt
_
V (t)
dV =
_
V (t)

D
Dt
dV, (2.8)
which is the most important form of Reynolds transport theorem in solid
mechanics.
2.5 Newtons equations of motion in a continuum
setting
Newtons law of motion states that the (time) rate of change of linear momen-
tum must equal external forces. For a body occupying V (t) moving through
space, we have the rate of linear momentum as
D
Dt
_
V (t)
v dV,
where v is the velocity of the material points in the body. Newtons law of
motion states that the time rate of change of linear momentum must equal
the applied force, so we have that
D
Dt
_
V (t)
v dV =
_
V (t)
f dV +
_
S
t dS,
18 2. General Principles in Continuum Mechanics
where f is a volume force per unit mass and t is a surface traction. In Part I
we derived the relation between the surface traction and the stress tensor as
t = n.
By Gauss theorem and the transport theorem we thus have
_
V (t)
f dV =
_
V (t)

Dv
Dt
dV
_
S
ndS =
_
V (t)

Dv
Dt
dV
_
V (t)
dV.
Since the volume is arbitrary, we nd the equations of motion for a continuum
to be

Dv
Dt
= f, (2.9)
or

v
t
+ v v = f, (2.10)
which must be supplied with initial and boundary conditions. The equations
(2.10) are valid in the spatial coordinate system, i.e., a coordinate system
xed with respect to the motion. Next, we shall look at some models for
computing the stress matrix in large deformations.
3. Large Deformation Solid Mechanics
In this Chapter we derive the basic relations in large deformation solid me-
chanics. We extend the notion of strain from small to large straining and
discuss some basic stressstrain relations in the large deformation setting.
3.1 Strain measures
In small deformation theory we have, in the two dimensional case, denes the
strain by

xx
=
u
x
x
,
yy
=
u
y
y
,
xy
=
1
2
_
u
y
x
+
u
x
y
_
.
These denitions require the deformations to be small, so that there is no
practical dierence between initial and deformed state. Even if only the dis-
placements are large, and the material remains stress free, this denition fails.
Consider the case of a wheel undergoing a 90

rotation around its median

axis. The corresponding displacements from P to P

is given by
u
x
= Y X, u
y
= X Y.
and if we extend the small deformation strain by siumply replacing derivatives
with respect to (x, y) with derivatives with respect to the initial coordinates
(X, Y ) we obtain

xx
=
yy
= 1,
xy
= 0.
Thus in small strain theory, a rigid body rotation will give rise to straining
which is clearly not desirable. We need a better denition of strain so that
rigid body rotations do not give rise to straining. Consider an innitesimal
segment dX that undergoes deformation to length ds, see Fig. 3.2. The length
after deformation can be evaluated as
ds
2
=
_
dX +
u
x
X
dX
_
2
+
_
u
y
X
dX
_
2
.
Green introduced the following denition of large deformation strain:
20 3. Large Deformation Solid Mechanics
P
P' (X,Y)
(-Y,X)
Fig. 3.1. Counterclockwise rotation 90

from (X, Y ) to (Y, X).

E
xx
=
1
2
ds
2
dX
2
dX
2
=
1
2
_
_
1 +
u
x
X
_
2
+
_
u
y
X
_
2
1
_
=
u
x
X
+
1
2
_
_
u
x
X
_
2
+
_
u
y
X
_
2
_
,
(3.1)
and similarly
E
yy
=
u
y
Y
+
1
2
_
_
u
x
Y
_
2
+
_
u
y
Y
_
2
_
,
and (less straightforwardly)
E
xy
=
1
2
_
u
x
Y
+
u
y
X
_
+
1
2
_
u
x
X
u
x
Y
+
u
y
X
u
y
Y
_
.
Taking again the case of the rigid body motion discussed earlier, we nd
E
xx
= 1 +
1
2
(1 + 1) = 0, E
yy
= 1 +
1
2
(1 + 1) = 0,
E
xy
=
1
2
(1 + 1) +
1
2
(1 1) = 0,
which is as expected. We note that the quadratic terms can be neglected
if the displacements are small, which justies the small deformation strain
matrix. Next, we shall take a more general approach to arrive at the Green
stain.
3.1 Strain measures 21
X,x
Y,y
dX
ds
u(X,Y)
u(X+dX,Y)
(u
y
/X)dX
(1+u
x
/X)dX
Fig. 3.2. Large deformations in two dimensions.
3.1.1 The deformation gradient and a general approach to strain
The most important quantity in large deformation analysis is the deformation
gradient F. Consider a material point P located initially at X and connected
to a point Q by the vector dX. After deformation during a time interval t,
P has moved to p at x(X, t) and Q has moved to q at x(X + dX, t). The
vector connecting p and q is denoted dx, and can be found by
22 3. Large Deformation Solid Mechanics
dx = FdX, where F =
x
X
=
_

_
x
X
x
Y
x
Z
y
X
y
Y
y
Z
z
X
z
Y
z
Z
_

_
.
Note that F is in fact the transpose of J which we have used for mapping
reference elements to physical elements in the nite element method. This is
because J transforms gradients whereas F transforms innitesimals.
Objects transforming material quantities such as dX to spatial quantities
such as dx will de denoted using capital boldface letters (such as F), the
reverse transformation by lowercase bold letters.
Now, the length of ds a segment dx in the deformend conguration can
be found by ds
2
= dx dx, and similarly, its original length dS
2
= dX dX.
We also see that
dx dx = dX (F
T
FdX) = dX CdX,
where C = F
T
F is called the right CauchyGreen deformation tensor (be-
cause F is on the right). In dierential geometry, where ds is typically a
physical length and dS is the length in a parameter space, C is known as the
metric tensor, because it it related to the measuring of distances.
If we interchange the order of F and its transpose we obtain b = FF
T
,
consequently called the left CauchyGreen deformation tensor, and we have
dX dX = dx b
1
dx
The Green strain tensor E is dened, in analogy with (3.1), by the relation
1
2
(dx dx dX dX) = dX EdX, (3.2)
and thus
E =
1
2
(C I). (3.3)
Alternatively, we can express the strain in spatial coordinates through the
relation
1
2
(dx dx dX dX) = dx edx,
where
e =
1
2
(I b
1
)
is called the Almansi strain tensor. We have here used uppercase letters
to denote quantities associated with the reference coordinate system and
lowercase for quantities associated with the spatial coordinate system. The
deformation gradient is the link between the two coordinate systems since it
is used as a map F : X x.
3.1 Strain measures 23
3.1.2 The boundary value problem on the deformed domain
We also need to dene stress in large deformation theory. The good news
is that the stress tensor in the deformed conguration must fulll the same
equilibrium equations as in small deformation theory. The bad news is that
we then have to deal with a domain that keeps changing with time. If we want
to dierentiate on a xed domain, this has to be the material domain where
we use X as coordinates. The problem is that we then have to transform
back the physical (Cauchy) stress from x to X.
Starting with the equilibrium on the deforming domain (t), from Chap-
ter 2.5, we have that the external equilibrium, in the absence of acceleration,
requires that volume forces and surface tractions have to be in balance, thus
= f in (t). (3.4)
The gradient is here taken with respect to x, and we can also view the primary
unknown as x since the displacement u = x X with X known.
With natural boundary conditions n = t on
N
(t), we have the
nonlinear boundary value problem: Find and x such that
= f in (t),
n = t on
N
(t),
x = g on
D
(t),
where g represents a known motion of the
D
(t) part of the boundary. We
now need to relate to x, or rather F, which is the objective of the next
Section.
3.1.3 Stressstrain relations in large deformation theory
A common assumption in large deformation theory is that there exists a
scalar strain energy density W from which the stress tensor can be derived.
In small deformation elasticity, this strain energy density takes the form
W =
_

0
( ) : d =
_

0
(2 +tr I) : d =
1
2
_
2||
2
+(tr )
2
_
. (3.5)
The point of dening this quantity is that the stress can be directly derived
from it through
=
W

, which means
ij
=
W

ij
.
If the stress can be derived from such a strain energy density, we say that
the material is hyperelastic. In large deformations, the assumption is that W
is a function of the deformation gradient or quantities associated with it, so
that for example
24 3. Large Deformation Solid Mechanics
W = W
F
(F) = W
E
(E) = W
C
(C) = W
b
(b).
Clearly, the strain energy should be independent of the coordinates used,
which means that the only the invariants of the tensors upon which it depends
can come into play, i.e., the trace, determinant, and norm of the tensor. Note
that the norm and trace of strain are exactly what is dening strain energy
for small deformations in (3.5).
When extending the stress concept from small deformation theory to large
deformations, we have some design criteria. The rst is the one mentioned
above, that only invariants be used in the denition of W, but we also want
the stress to be zero for zero strain, and we want to be able to regain small
strain theory in the limit of vanishingly small strains. The simplest extension
from small deformation is the St. VenantKirchho stress dened from the
strain energy density
W
E
(E) =
1
2
_
2|E|
2
+(tr E)
2
_
so that
=
W
E
(E)
E
= 2E +tr EI
where we used the uppercase to indicate that this stress is dened in the
reference coordinate system. This clearly converges to the linear elasticity
case as the strains become small. The stress is called the (second) Piola
Kirchho stress, and needs to be transformed to the spatial coordinate system
if we are to use the equilibrium equations on the form (3.4). For this trans-
formation, it turns out that the Piola transform is the natural choice. The
Piola transform for a vector is dened by taking a vector q in the reference
system to a vector q in the spatial system via
q(x) =
1
J
F q(X), x = x(X),
where J = det F corresponds to local volume change under the change of
variables. The inclusion of 1/J has the following eect: we will momentarily
show that

x
q =
1
J

X
q, (3.6)
and since 1/J is countered by the volume change J in the transformed integral
we can, with this transform, substitute Gauss theorem on the spatial domain
with a corresponding theorem on the referential domain:
_

n q ds =
_

x
q dx =
_

1
J

X
q JdX =
_

n q d s.
This is the key idea behind the Piola transform. We will here restrict the
proof of (3.6) to the twodimensional case. What we want to show is that
3.1 Strain measures 25
J
x
q(x) =
X
q =
X
(JF
1
q(x(X))).
We have that
JF
1
=
_

_
y
Y

x
Y

y
X
x
X
_

_, JF
1
q =
_

_
y
Y
q
x

x
Y
q
y

y
X
q
x
+
x
X
q
y
_

_.
Thus,

X
(JF
1
q) =

X
_
y
Y
q
x

x
Y
q
y
_
+

Y
_

y
X
q
x
+
x
X
q
y
_
= q
x
_

2
y
XY


2
y
Y X
_
+q
y
_

2
x
Y X


2
x
XY
_
+
y
Y
q
x
X

x
Y
q
y
X

y
X
q
x
Y
+
x
X
q
y
Y
.
Since
q
x
X
=
q
x
x
x
X
+
q
x
y
y
X
,
etc., we nd that

X
(JF
1
q) =
y
Y
_
q
x
x
x
X
+
q
x
y
y
X
_

x
Y
_
q
y
x
x
X
+
q
y
y
y
X
_

y
X
_
q
x
x
x
Y
+
q
x
y
y
Y
_
+
x
X
_
q
y
x
x
Y
+
q
y
y
y
Y
_
=
_
q
x
x
+
q
y
y
__
y
Y
x
X

y
X
x
Y
_
= J
x
q.
The corresponding transform in the case of stress tensors is the rule
=
1
J
FF
T
=
1
J
F (2E +tr EI) F
T
,
which denes the Cauchy stress for the St. Venant-Kirchho model. Note
that a simple way to dene a tensor (called a dyadic tensor) is to take the
outer product of a vector by itself:
B = aa
T
,
where a is a column vector. If a = F a, then
26 3. Large Deformation Solid Mechanics
B = F a a
T
F
T
= F

BF
T
,
where we can think of

B as being dened in the reference coordinate system.
The inclusion of 1/J is again done to counter volume change between referen-
tial and spatial coordinates. This allows for a traction to be dened in relation
to the PiolaKirchho stress in the referential domain, and a Gaussian the-
orem for stresses to hold in the referential domain. A further examination of
these relations are however beyond the scope of this course.
To return to the stress denitions, it is generally agreed that the St.
VenantKirchho model is not very accurate outside the small deformation
theory; it works well with large rotations but not give a physically realistic
deformation under large strains. A better model is given by the NeoHooke
stress denition
=

J
(b I) +
ln J
J
I, (3.7)
which is here directly dened in the spatial coordinate system, but is actually
derived from the referential system, and a PiolaKirchho stress tensor
=
_
I C
1
_
+C
1
ln J.
We note that
u = x X, F =
u
X
+I
b = FF
T
=
u
X
u
X
T
+
u
X
+
u
X
T
+I,
so that, in small deformations, b 2(u)+I. Further (two-dimensional case,
but three dimensions is similar),
J =

u
x
X
+ 1
u
x
Y
u
y
X
u
y
Y
+ 1

=
u
x
X
u
x
Y
+
u
x
X
+
u
x
Y
+ 1
u
x
Y
u
x
X
so that J 1 + u 1 and ln J u, since ln(1 +x) x for x small.
Thus, the Neo-Hooke model is also a generalization of small deformation
elasticity.
4. Thermodynamics
In this Chapter we give some fundamental facts from thermodynamics. I par-
ticular, we derive the rst and second laws of thermodynamics in a continuum
mechanics setting.
4.1 Energy balance and the rst law of thermodynamics
Apart from mechanical work, which is a form of energy, we must also take
into account heat energy. In fact, straining a body generates heat, and heat
leads to strains and, in general, to stresses in a body. Energy transfer may
also involve other non-mechanical entities such as electricity, but these eects
will not be taken into account here.
The power input P
in
resulting from forces acting on a body occupying the
volume V (t) moving wiht velocity v is given by
P
in
=
_
V (t)
t v ds +
_
V (t)
f v dV,
where t is a surface traction and f is a volume force per unit mass. The
traction is connected to the Cauchy stress via t = n, so by Gauss theorem
and the symmetry of we can write
P
in
=
_
V (t)
(( +f) v + : (v)) dV
By Newtons law of motion (2.9), we can replace the rst two terms by density
times acceleration, and thus
P
in
=
_
V (t)
_

Dv
Dt
v + : (v)
_
dV.
Since
Dv
Dt
v =
1
2
D
Dt
|v|
2
,
we have, using Reynolds transport theorem on the form (2.8),
28 4. Thermodynamics
P
in
=
1
2
D
Dt
_
V (t)
|v|
2
dV +
_
V (t)
: (v) dV. (4.1)
The rst term is the material derivative of the kinetic energy of the volume,
and the second is called stress power. Note that the stress power is that part
of the external power input which does not contribute to change of kinetic
energy.
We must also consider heat input rate Q
in
into the volume in order to
get at a total energy balance. In a moving body, this input consists solely of
conducted heat through the surface V (t) (convective heat ux only occurs
through a xed control volume). Thus
Q
in
=
_
V (t)
q nds +
_
V (t)
r dV,
where q is the heat ux vector, q n is the outward heat ux, and r is a
heat source per unit mass inside the body, resulting, e.g., from radiation. It
is assumed here that Q
in
and P
in
are expressed in the same units, so that the
total input of power is the straight sum P
in
+Q
in
(which requires a conversion
constant if Q
in
is expressed in thermal units).
It is well known from experiments that if a system goes through a cycle
back to its initial state, then
_
P
in
dt = 0,
_
Q
in
dt = 0,
where
_
denotes the sum of contributions during the cycle. This means that
P
in
dt and Q
in
dt are not dierentials of any scalar function (in which case
the integrals would be zero). In thermodynamics, a scalar function belonging
to the system, such as density (an intensive quantity) or mass (an extensive
quantity), is called a state function. The quality of being a state function lies
in the fact that we may talk of the system content of it at any given moment;
this is thus not true for the power and heat input. On the other hand, it has
been experimentally veried that
_
(P
in
+Q
in
) dt = 0,
meaning that there is a scalar function such that (P
in
+Q
in
) dt is a dierential
thereof. This function is known as the total energy of the system, E
tot
, and
we have
D
Dt
E
tot
= P
in
+Q
in
,
_
D
Dt
E
tot
dt = 0, (4.2)
total energy is conserved in a closed cycle. This is known as the rst law of
thermodynamics. (Often this law is stated as the total energy of a closed
system is constant; in a closed system there is no work or heat input and
D
Dt
E
tot
= 0.)
4.2 Entropy and the second law of thermodynamics 29
Usually, the total energy of a system is formally divided into two parts,
its kinetic energy E
K
and its internal energy E
I
(which may actually also be
of kinetic nature locally, as heat motion), E
tot
= E
K
+E
I
, where
E
K
=
1
2
_
V (t)
|v|
2
dV.
Let us now dene the internal energy density per unit mass, u, such that
E
I
=
_
V (t)
udV
The rst law of thermodynamics (4.2) then gives
D
Dt
_
V (t)
_

2
|v|
2
+u
_
dV =
1
2
D
Dt
_
V (t)
|v|
2
dV +
_
V (t)
: (v) dV

_
V (t)
q nds +
_
V (t)
r dV.
Using Gauss divergence theorem on the heat ux term together with (2.8),
we nd that
_
V (t)
_
q +
Du
Dt
: (v) r
_
dV = 0.
This leads to the energy equation

Du
Dt
= : (v) +r q, (4.3)
alternatively, with e the total energy density,

De
Dt
=
1
2

D
Dt
|v|
2
+ : (v) +r q, (4.4)
which are the dierential equations expressing conservation of energy.
4.2 Entropy and the second law of thermodynamics
The main use of the second law of thermodynamics in continuum mechanics
is for setting limits to constitutive laws. It deals with the inevitable loss
of energy to heat dissipation, and it does so by introducing a new state
function called entropy. Entropy is one of those concepts that is easy in
principle to understand but very dicult to nail down. Indeed, Wikipedia has
a disambiguation page spanning twelve entries for thermodynamical entropy,
six for information theory, and three for computer science (other uses not
30 4. Thermodynamics
counted). The purpose here is to show what the entropy inequality looks like
in continuum mechanics.
In the early days of steam engines it was soon realized that in order
for such a machine to work there needed to be temperature dierences: you
cannot get work out of a steam engine if the surrounding temperature is
that of boiling water. Indeed one early form of the second law is simply that
without input of work, heat always ows from higher to lower temperature.
Consider an innitesimal heat input per unit mass q, where is used
instead of d to indicate that q is not an exact dierential. We recall that
Q
in
=
_
V (t)
qdV,
and that Q
in
dt is not a dierential of any scalar function. A dierential of
the form
df = A(x, y)dx +B(x, y)dy
is called exact if _
df = 0,
which means that there exists a function f(x, y) such that
df =
f
x
dx +
f
y
dy
so that if df is exact there must hold
A
x
=

2
f
xy
=
B
y
.
Our q is not exact and in general
_
q = 0.
However, in the study of idealized reversible processes, it was found theoret-
ically that
_
q

= 0,
where is the (absolute) temperature at which the heat input is delivered.
This means that there exists an exact dierential ds of a state variable s,
called the entropy per unit mass, such that in a reversible process
ds =
q

.
However, no process in the real world is completely reversible without power
input (which indeed is the second law stated loosely); there exist no perpetuum
4.2 Entropy and the second law of thermodynamics 31
mobile. In an irreversible cyclic process thermodynamics still postulates that
there exists a state function s such that
_
ds = 0,
but on the other hand it is invariably found that
_
q

0.
If we now consider an irreversible process from state a to state b and its
corresponding imaginary reversible process, then
s =
_
b
a
_
q

_
rev
,
(the entropy increase equals the entropy input in a reversible process) and
we have that for a full cycle (comprising one reversible and one irreversible
branch)
0
_ _
q

_
irrev
=
_
b
a
_
q

_
irrev
+
_
a
b
_
q

_
rev
=
_
b
a
_
q

_
irrev
s.
This means that the entropy increase in the irreversible process from a to b
is larger that the entropy input by heating,
s
_
b
a
_
q

_
irrev
.
We shall now translate this inequality into a continuum mechanical setting.
Recall that
Q
in
=
_
V (t)
q dV =
_
V (t)
r dV
_
V (t)
q nds,
so the entropy input rate equals
_
V (t)
r

dV
_
V (t)
q

nds
and we must have
D
Dt
_
V (t)
s dV
_
V (t)
r

dV
_
V (t)
q

nds
or: the rate of entropy increase is larger than the rate of entropy input. On
local form this becomes
32 4. Thermodynamics

Ds
Dt

r

_
q

_
, (4.5)
known as the ClausiusDuhem inequality. This inequality is rarely used for
computing anything in practice in continuum mechanics (except in isentropic
processes); its importance lies in the limits in puts on modeling material
laws. Being able to show that a material law of your invention obeys (4.5)
is generally viewed as an important step. If your material law on the other
hand violates (4.5), you probably had better modify it.
5. Fluids
This Chapter concerns the modeling of compressible and incompressible
ows.
5.1 Inviscid compressible uids and conservation laws
In a uid at rest, shear stresses cannot be sustained. This means that the
only stress can be pressure, so that
= p
0
I,
where p
0
is called the static pressure. The static pressure is assumed, from
thermodynamics, to be functionally related to the density and (absolute)
temperature of the uid, so that there exists an equation f(p
0
, , ) = 0. The
thermodynamic pressure p is dened as the quantity that fullls the same
relation also when the uid is moving, i.e.,
f(p, , ) = 0,
called the equation of state. A classical example of such an equation is the
perfect gas law which states
p = R,
where R is called the gas constant. In basic books on thermodynamics this
law is stated as
pV = nR
0
=
m
M
R
0

where V is the volume of gas and n is the amount of gas (in moles), m is
the mass of the gas, and M is its molar mass (the mass of a given amount of
substance). We can rewrite this as
p
V
m
=
R
0
M
= R,
giving the relation between R and R
0
.
There is also a caloric equation of state which relates the internal energy
to temperature and density,
34 5. Fluids
u = u(, ).
In an ideal gas, the internal energy is simply proportional to temperature, so
that
u = c
v
,
where c
v
is the specic heat at constant volume. If energy is added to a gas
while its volume is kept xed, then the energy change is related to tempera-
ture change by du = c
v
d. But if the gas expands and instead the pressure in
the gas is held xed, then some energy is used to expand the specic volume
v = 1/, so that
du +pdv = c
p
d, or d(u +pv) = c
p
d,
where c
p
is the specic heat at constant pressure. The quantity
h := u +pv = u +
p

is known as the enthalpy of the gas. We note that in an ideal gas

p

= R = c
p
c
v
,
so also c
p
is constant. The ration of specic heats is denoted by ,
:=
c
p
c
v
,
and can be computed theoretically for dierent gases. For air we have = 1.4.
For an ideal gas we can then deduce
u = c
v
=
c
v
R
p

=
c
v
c
p
c
v
p

=
p
( 1)
,
and we obtain the most common form of the caloric equation of state:
e =
p
1
+
1
2
|v|
2
(5.1)
where e is the total energy density of the gas.
Another important property of a gas is its speed of sound c. As we saw
in Part I, speed of waves in general can be found as
c =

stiness
density
.
In a gas, the stiness related to sound speed is dened as the bulk modulus
K when no heat is added or removed from the uid (an adiabatic process).
Then the bulk modulus has been theoretically found to be approximately
5.1 Inviscid compressible uids and conservation laws 35
K = p,
so
c =
_
p

.
An ideal frictionless uid is dened as a uid that cannot support shear
stresses even in motion. No such uid exists, but it is a convenient approxi-
mation for a large class of ows. Then
= pI,
where p is the thermodynamic pressure. These uids are said to obey the
Euler equations, which are characterized by the conservation of mass (2.1),

t
+ ( v) = 0.
energy (4.4),

De
Dt
=
1
2

D
Dt
|v|
2
+ : (v) +r q,
and linear momentum (2.9),

Dv
Dt
= f.
We want to write these on conservation form
u
t
+

i
q
i
(u)
x
i
= 0, (5.2)
where the q
i
are known as the ux vectors. The reason for this will be eluci-
dated below. We have for our ideal gas that
: (v) = pI : (v) = p v.
If we further assume no heat ow or heat source and no volume forces, we
have from transport of linear momentum
0 =
Dv
Dt
=
v
t
+ v v +p. (5.3)
From the conservation of mass we nd
(v)
t
=

t
v +
v
t
=
v
t
(v) v
so that (5.3) can be written
36 5. Fluids
0 =
(v)
t
+ (v) v + v v +p.
Let us consider rst the onedimensional case. Here we can write
(v)
x
v + v
v
x
=
(v
2
)
x
and thus the conservation of momentum takes the form
(v)
t
+

x
(v
2
+p) = 0
For the energy equation, where we similarly have
(e)
t
=
e
t
(v) e,
and, using also the momentum equation

Dv
Dt
= p,
we can write, in one dimension,
0 =
De
Dt

1
2

D
Dt
v
2
+p
dv
dx
=
e
t
+ v
e
x
v
Dv
Dt
+p
v
x
=
(e)
t
+

x
(v) e + v
e
x
+v
p
x
+p
v
x
=
(e)
t
+

x
(ve +vp).
Thus the Euler equations in one dimension is of the form (5.2) with
u =
_
_

v
e
_
_
, q
1
=
_
_
v
v
2
+p
v(e +p)
_
_
,
where
p = ( 1)
_
e
1
2
v
2
_
from (5.1). The variables , v and e are for this reason called conservation
variables. The Euler equations, expressing conservation of mass, momentum,
and energy, can also be written in nonconservative form by using other
variables, or simply by expanding the derivatives and writing
u
t
+A(u)
u
x
= 0, where A(u) :=
q
1
u
,
5.2 Nonlinear conservation laws and shocks 37
also called the quasilinear form.
Similarly in three dimensions, we can derive (5.2) with
u =
_

v
1
v
2
v
3
e
_

_
, q
i
= v
i
u +p
_

_
0

1i

2i

3i
v
i
_

_
,
where
ij
is the Kronecker delta. Again we must add the caloric equation of
state (5.1),
p = ( 1)
_
e
1
2
|v|
2
_
.
5.2 Nonlinear conservation laws and shocks
To illustrate the fundamental problem of nonlinear conservation laws, we
shall consider a scalar problem of the type
u
t
+
q(u)
x
= 0,
more specically Burgers equation where q(u) = u
2
/2. On quasilinear form,
this problem is
u
t
+u
u
x
= 0,
and we note that this equation says that the material derivative following the
particle paths
x
t
= u(x(t), t)
is zero. This has the consequence of not allowing for a unique function u as
a solution since if u varies, the characteristics for high u will overtake those
of low u giving several values of u for a given x, see Fig. 5.1. This cannot be
the physically correct solution to this problem, and the explanation to what
happens is that we have completely neglected dissipation. Consider instead
the viscous Burgers equation
u
t
+
1
2

x
(u
2
)

2
u
x
2
= 0,
with a very small viscous parameter. Now, however small is there are
instances when the eect of this term cannot be neglected. As u becomes
steeper the second derivatives will increase in size and counter the overtaking
characteristics. We will obtain an almost vertical front, a shock, traveling
with a characteristic speed s.
38 5. Fluids
Fig. 5.1. Physically impossible solutions to Burgers equation for smooth initial
data.
Using the conservation property of our equation, we have that for a solu-
tion that takes on the constant value u
l
on the left of the shock and u
r
on
the right of the shock, called a Riemann problem,
_
a
a
u
t
dx = q(u
l
) q(u
r
) =
1
2
u
2
l

1
2
u
2
r
=
1
2
(u
l
+u
r
)(u
l
u
r
)
for Burgersequation (assuming a is away from the shock). By denition of
s, we have
u(x, t) =
_
u
l
x < st
u
r
x > st,
and then
_
a
a
u(x, t) dx = (a +st)u
l
+ (a st)u
r
so that
_
a
a
u
t
dx = s(u
l
u
r
)
which gives the shock speed s = (u
l
+ u
r
)/2. More generally, this same
argument gives the RankineHugoniot jump condition
s =
q(u
l
) q(u
r
)
u
l
u
r
,
which for systems of conservation laws becones
q(u
l
) q(u
r
) = s (u
l
u
r
).
We note that if we have a linear system
q = Au,
5.3 Linearization of the Euler equations and acoustics 39
where A is a constant matrix, then A(u
l
u
r
) = s (u
l
u
r
), and s is an
eigenvalue to A.
Now, conservation laws are very sensitive to manipulations. Consider for
example multiplying Burgers equation by 2u; then we get, if the derivatives
exist,
2u
u
t
+ 2u
2
u
x
= 0 =

t
u
2
+
2
3

x
u
3
= 0.
For smooth solutions this gives the same solution, but if u is not dierentiable
the implication is not true and we note that the shock speed in the modied
equation is
s
mod
=
2
3
_
u
3
r
u
3
l
u
2
r
u
2
l
_
and it is a simple exercise to show that
s
mod
s =
1
6
(u
l
u
r
)
2
u
l
+u
r
,
i.e., the manipulation resulted in shock waves moving with the wrong speed.
This is the crucial fact concerning conservation laws: they have to be on
conservation form, and not only that, the right physical variables have to be
conserved.
5.3 Linearization of the Euler equations and acoustics
If we consider an on average still inviscid liquid of constant density
0
subject
to small oscillatory movements of its constituents particles, we can introduce
the uid displacement u = u(x, t) around the average state of rest. Then the
increase in volumetric stress (over the already existing ambient pressure in
the uid) fullls

v
= K u, or p = K u,
where K is the bulk modulus and p now denotes the deviation from the
ambient pressure, known as the acoustic pressure. Introducing the sound
speed
c
2
=
K

0
,
this can be written
p +
0
c
2
u = 0. (5.4)
The velocity of the particles is given by
v =
u
t
,
40 5. Fluids
and the conservation of linear momentum can be written

2
u
t
2
+p = 0. (5.5)
We can now eliminate the uid displacement by taking the divergence of (5.5)
and two time derivatives of (5.4), leading to

2
p
t
2
c
2

2
p = 0, (5.6)
the acoustic wave equation, or by taking the gradient of (5.4), leading to

2
u
t
2
c
2
( u) = 0. (5.7)
The formulation (5.6) is the most common for computations in acoustics
since it only involves the scalar pressure. It describes the propagation of an
acoustic pressure wave in the uid, moving with the speed of sound c, but its
derivation involves taking an additional spatial derivative of the pressure not
motivated by physical reasoning (to put it another way: the pressure in (5.6)
has more regularity than can be motivated from the original equations). In
the weak form of (5.6) this leads to natural boundary conditions involving
normal derivatives of the pressure whereas in a stress computation we have
the pressure itself in the natural boundary condition. This leads to diculties
if we want to couple acoustics in a room with the sound propagation in the
walls of the room (for example modeled as an elastic wave equation), since
the transmission conditions involve continuity of displacements and of stress,
not continuity of derivatives of stress.
The second form (5.7) does not suer from this increase in regularity
since the gradient of the pressure is already present in (5.5). Equations of the
type (5.7) are called graddiv equations and similar problems are common
in electrodynamical wave propagation problems (where articially increasing
regularity as in (5.6) typically does not work).
5.4 Inviscid incompressible uids
For compressible uids, we the density is an unknown and part of the problem.
In an incompressible uid on the other hand, the density does not change
following the motion we have v = 0 as derived in (2.4). For these types of
ows there can be no shocks since a shock wave requires compression. Another
assumption often used for incompressible ows is that the equation of state
relating pressure, density, and temperature is independent of temperature.
Then the equation of state takes the simple form = constant, and we can
drop the energy equation (unless we are interested in the transport of heat in
5.4 Inviscid incompressible uids 41
the uid). Note that the energy equation cannot be dropped for compressible
ows since we need the energy in order to compute the pressure, as in (5.1).
We still have the relation = pI, and the conservation of momentum
can then be written (and used) on nonconservative form as

v
t
+ v v +p = f. (5.8)
Together with the continuity equation v = 0 this gives a system of four
equations for the three components of velocity and for the pressure. These
are known as the incompressible Euler equations. Note that we do not here
need an extra equation to compute the pressure and that the pressure in fact
takes the role of a Lagrange multiplier enforcing divergence zero.
A further simplication of inviscid ows can be derived under the as-
sumption that the ow is irrotational, that is v = 0, where we recall
that
v =
_
v
z
y

v
y
z
,
v
x
z

v
z
x
,
v
y
x

v
x
y
_
is a measure of the rotation of the uid. This means that we can assume that
the velocity eld is the gradient of a potential function
v = (x, y, z),
since the curl of a gradient is always zero. Applying the continuity equation
to this velocity, we obtain

2
= 0,
and we do not even need to invoke the conservation of momentum in order
to nd the velocity eld! The boundary conditions for this equation are of
Neumann type if the boundaries do not allow for ow going through, since
then
0 = n v = n ,
and in case we know that part of the domain is irrotational, we can couple
this equation to the full Euler equations via an interface by the coupling
condition
n = n v
E
where v
E
is the velocity in the rotational part of the ow.
In two space dimensions, we can get rid of the pressure for more general
ows. Introducing the vorticity through
:= v =
v
y
x

v
x
y
,
and subtracting the yderivative of the rst equation of (5.8) from the
xderivative of the second, we obtain (under the assumption that f is con-
servative, i.e., a gradient of a potential eld)
42 5. Fluids

t
_
v
y
x

v
x
y
_
+

x
(v v
y
)

y
(v v
x
) = 0
and we also have

x
(v v
y
)

y
(v v
x
) =
v
x
v
y
+v
v
y
x

v
y
v
x
v
v
x
y
= v
_
v
y
x

v
x
y
_
+
v
x
x
v
y
x
+
v
y
x
v
y
y

v
x
y
v
x
x

v
y
y
v
x
y
= v +
_
v
x
x
+
v
y
y
__
v
y
x

v
x
y
_
which, since v = 0, leads to

t
+v = 0,
and thus the vorticity is constant following the motion in an inviscid uid.
Next, the condition
v
x
x
+
v
y
y
= 0
means that there exists a stream function (x, y) such that
(v
x
, v
y
) =
_

y
,

x
_
.
Then
=
v
y
x

v
x
y
=

x
2


2

y
2
=
2
,
and we obtain the stream functionvorticity problem of nding (, ) such
that

t
+
_

y
,

x
_
= 0,

2
= ,
(5.9)
with suitable boundary conditions, in particular for . Consider the ux q
through a surface, i.e., line, connecting points P
0
and P
1
, with normal n. For
an innitesimal line segment ds we know that n = (dy/ds, dx/ds) and thus
q =
_
P1
P0
n v ds =
_
P1
P0
(v
x
n
x
+v
y
n
y
) ds =
_
P1
P0
(v
x
dy v
y
dx)
=
_
P1
P0
_

y
dy +

x
dx
_
=
_
P1
P0
d = (P
1
) (P
0
).
At a solid boundary, where n v = 0, must be constant, and on other
boundaries, where we know the ux, we know by how much must increase.
5.5 Viscous incompressible uids 43
5.5 Viscous incompressible uids
In Part I we gave the Stokes equations of creeping incompressible ow. We
shall now extend these equations to include the eect of the material deriva-
tive.
Stokes assumption was more generally that the dierence between the
stress in a static uid and a moving uid is a function of the symmetric
velocity gradient, so that such a uid carries stresses on the form
= p I +F((v)),
where p is the thermodynamic pressure. F is called the viscous stress, and if
F is linear the uid is called Newtonian. Such a linear relation can in general
be written, componentwise,

ij
= p
ij
+

l
c
ijkl

kl
(v),
where c
ijkl
are the components of a fourth order tensor. Now, it turns out
that if the uid is isotropic and the stress symmetric, it is possible to prove
mathematically that the general form of this relation must be on the form
= p I + 2(v) +(tr (v)) I, (5.10)
where and are material parameters. This is known as the NavierPoisson
law of a Newtonian uid.
Denoting, as above, the volumetric stress by
v
= (tr )/3, the deviatoric
stress by
D
=
v
I, and the deviatoric strain rate by
D
= (tr )I/3,
we can rewrite NavierPoisson as

D
= (
v
p)I +
_
+
2
3

_
tr (v) I + 2
D
(v).
Taking the trace of this equation gives us

v
p +
_
+
2
3

_
tr (v) = 0,
so NavierPoisson can be written as

D
= 2
D
(v) and
v
= p +K v = p
K

D
Dt
, (5.11)
where K is the bulk modulus, K = + 2/3. In the last step conserva-
tion of mass was used. This means that the negative volumetric stress equals
the thermodynamic pressure in two situations: if the ow is incompressible,
D/Dt = 0, or if Stokes hypothesis: K = 0 is fullled. Since typically only
44 5. Fluids
one material parameter, the viscosity, is considered, Stokes hypothesis is usu-
ally assumed to hold for compressible ows. Thus the NavierPoisson law is
usually written
= p I + 2
D
(v), (5.12)
valid both for compressible and incompressible ows. Introducing this consti-
tutive relation into the conservation of momentum equation (2.10), written
in spatial coordinates, we obtain the incompressible NavierStokes equations

v
t
+ v v 2
D
(v) +p = f,
v = 0.
(5.13)
As in Part I, we can rewrite these equations using the incompressibility con-
dition to obtain

v
t
+ v v
2
v +p = f,
v = 0.
(5.14)
Dividing by the constant density, we obtain the variant
v
t
+v v
2
v + p = f,
v = 0,
(5.15)
where = / is known as the kinematic viscosity and p = p/. As remarked
in Part I, the natural boundary conditions are dierent for (5.13) and (5.14)
in a weak formulation. The formulation (5.13) gives the traction as a natural
boundary condition which is physically correct and must be used in case of a
free boundary ow. On the other hand, (5.14) gives a good articial boundary
condition for pipe ows. If no natural boundary conditions are present, both
formulations give the same weak solution.
Also for the NavierStokes equations, we can eliminate the pressure in
two dimensional ows. Following the derivation of (5.9), we nd that we can
seek (, ) such that

t
+
_

y
,

x
_

2
= 0,

2
= ,
(5.16)
but now the boundary conditions for must be taken into account and
represent an additional diculty. The question of course is how the weak
form is derived. For stationary Stokes, where
2
= 0 and
2
= , the
functional that takes on a stationary value is
5.6 Turbulence 45
F(, ) =
_

_

1
2

2
_
dxdy,
Assuming is xed on this leads to
_

v dxdy = 0, v H
1
0
()
_

( w w)dxdy = 0, w H
1
(),
and we see that the second equation is in fact the equation for vorticity! Thus
no boundary condition for vorticity is usually assumed in the weak form of
(5.16).
5.6 Turbulence
Laminar ow is when there is (almost) no mixing between adjacent layers
of uid moving in parallel. In such situations, the viscosity can be measured
and the NavierPoisson law can be evaluated. It is then found to accurately
describe Newtonian uids.
In a turbulent ow the velocity, and hence stress, uctuates more or less
randomly in time at any given point. It is then not possible to experimentally
verify the NavierPoisson law, which nevertheless is still assumed to hold.
Vortices in the ow can then occur at all scales down to the a limit set by the
viscosity, where nally a vortex would dissipate. In order to analyze turbulent
ows, it is usually assumed that there exists an average ow v from which
the actual ow is perturbed, so that v = v+v

. Thus NavierStokes, without

force terms, takes the form

_

t
( v +v

) + ( v +v

) ( v +v

)
_
= ( p +p

) +
2
( v +v

),
and we can rewrite the nonlinear term as
( v
j
+v

j
)

x
j
( v
i
+v

i
) =

x
j
_
( v
i
+v

i
)( v
j
+v

j
)
_
( v
i
+v

i
)

x
j
( v
j
+v

j
)
where the last term vanishes when summed on j because of the divergence
zero condition. Thus we have component wise

_
_

t
( v
i
+v

i
) +

x
j
(( v
i
+v

i
)( v
j
+v

j
))
_
_
=

x
i
( p+p

)+
2
( v
i
+v

i
).
Finally, this equation is averaged. Under the rules of ensemble averaging, we
have
46 5. Fluids
a = a, a +b = a +

b, a

b = a

b.
Then all linear terms will lose their perturbations, because by denition these
average to zero, and
( v
i
+v

i
)( v
j
+v

j
) = v
i
v
j
+ v
i
v

j
+v

i
v
j
+v

i
v

j
= v
i
v
j
+v

i
v

j
.
The chain rule gives

x
j
( v
i
v
j
) = v
j
v
i
x
j
+ v
i
v
j
x
j
,
where the last term vanishes when summed on j. Thus we have the averaged
NavierStokes equations

_
_
v
i
t
+

j
v
j
v
i
x
j
_
_
=
p
x
i
+

x
j
_

v
i
x
j
v

i
v

j
_
,
or

_
v
t
+ v v
_
= p +
_
v v

_
.
The term v

which is a 3 3 matrix in a coordinate system can thus

be interpreted as a stress tensor, called Reynolds stress. The uctuations
obviously feed back into the equations for the average ow, and modeling the
resulting Reynolds stress is a fundamental problem, mainly in the numerical
solution of NavierStokes equations, since the vortices cannot be resolved
by mesh renement for most physically relevant ows. Boussinesq (1887)
proposed relating the turbulent stresses to the mean ow to close the system
of equations, in this setting as

turb
:= v

=
turb
v
2
3
k I,
where
turb
is a turbulent (eddy) viscosity, to be determined experimen-
tally, or by additional physical reasoning, and k is the average local kinetic
energy of the perturbations
k =

2

i
v
2
i
.
Of course this means that v

must be very precisely related to v, which

is a pretty strong assumption. In numerical simulations, however, this trick
is common in order to more or less roughly estimate the eect on the mean
ow by the uctuations. We still need to dene
turb
and get an equation for
k, and this is a (small) part of turbulence research, which is still the object
of intense eort (turbulence is considered to be one of the main unanswered
problems in continuum mechanics and applied mathematics).
6. Stability problems
In this Chapter we introduce the general conservation and balance laws that
hold in all continuum mechanical theories. They thus form the basis for uid
and solid mechanics in general.
6.1 A simple example
Consider the situation depicted in Figure 6.1, consisting of two rigid beams
connected by a rotation spring fullling the linear material law M = k,
where M is the rotating moment, kis a material constant, and is the angle
of rotation.
Fig. 6.1. A simple mechanical model.
The equilibrium, geometry, and constitutive law thus gives the system of
equations
_

_
M
PL
2
sin = 0,
= 2,
M = k,
leading to the relation
sin =
4k
PL
. (6.1)
This is a nonlinear equation since we are taking into account the change in
loading on the spring due to deformation.
48 6. Stability problems
Now, for positive angles , the solution to this equation depends on the
size of 4k/(PL), as shown in Figure 6.2. We see that for P 4k/L there
is only one solution = 0, but in the case P > 4k/L there is also another
solution =
1
> 0.
0 0.5 1 1.5 2 2.5 3 3.5
0
0.2
0.4
0.6
0.8
1
1.2
1.4


4k/PL > 1
4k/PL = 1
4k/PL < 1
Fig. 6.2. Solutions to (6.1).
We now pose the question: what happens if we disturb the equilibrium
by a small positive perturbation ? Then
M = M( +)
PL
2
sin( +)
= 2k( +)
PL
2
(sin cos + cos sin )
2k + 2k
PL
2
sin
PL
2
cos
=
_
2k
PL
2
cos
_
,
and if the disturbance is from the equilibrium point = 0 we see that
M =
_
2k
PL
2
_
< 0 if P >
4k
L
.
Since M is the dierence between the moment holding back deformation
and the moment increasing deformation, the negative sign of M means that
the spring is unable to hold back the deformation and we have an unstable
equilibrium.
The linearization could have been done directly in the equilibrium equa-
tion; if is small we have
6.1 A simple example 49
2k =
PL
2
sin
PL
2
=
_
4k
PL
1
_
= 0,
so we have two possibilities: either = 0 or we have a critical load
P
cr
=
4k
L
in which case is indeterminate (unlike the nonlinear case where we can
solve the nonlinear equation to get =
1
). This is the simplest form of an
eigenvalue problem. To see what happens in a slightly more complicated sit-
uation, consider the system in Figure 6.3, where the geometrical linearization
has been performed. We have
Fig. 6.3. A problem with two rotational springs.

B
= + = 2 ,
A
= ( ) = 2 .
from equailibrium: M
B
= PL, M
A
= PL, and from the constitutive law:
M
B
= k
B
, M
A
= k
A
. Together, this gives
2 =
PL
k
,
2 =
PL
k
,
or
__
2 1
1 2
_

_
1 0
0 1
___

_
=
_
0
0
_
so that := PL/k is an eigenvalue in a generalized eigenvalue problem. The
corresponding eigenvalues are
1
= 1,
2
= 3. For
1
we nd the eigenvector
by solving
_
1 1
1 1
__

_
=
_
0
0
_
giving = , a symmetric deformation mode, and for
2
we nd
50 6. Stability problems
_
1 1
1 1
__

_
=
_
0
0
_
giving = , an antisymmetric deformation mode. The lowest load
P
cr
= k/L is associated with the symmetric deformation. We realize that as
the number of springs increases, so do the number of possible modes and as-
sociated eigenvalues. This is a simplied version of a reduced elasticity model,
the beam. The beam is a long thin structure which makes possible some
simplifying assumptions in the elasticity equations.
6.2 The beam
Already in the 1490:s Leonardo da Vinci gured out (see Fig. 6.4) that a
beam elongates on one side and compresses on the other if it is bent. Thus
there is an axis, going through the centroid of the cross section of the beam,
that remains unstretched.
Fig. 6.4. Leonardo draws a beam.
If we put the beam in a coordinate system with the xaxis along the
centroid of the cross section, and bending is assumed to take place around
the yaxis, then the deection in the zdirection is w(x) := u
z
, see Fig. 6.5.
We can also see that u
x
= z, where = dw/dx is the (small) angle of
rotation. We thus have

xx
=
u
x
x
= z
d
2
w
dx
2
,
If we assume that the only stress is
xx
along the beam, Hookes law gives

xx
= zE
d
2
w
dx
2
.
Summing up the stresses on the cross section surface, Fig. 6.6, we nd that
6.2 The beam 51
Fig. 6.5. Deformation of a beam.
_
A

xx
z dydz = M,
so that
M =
_
A
E
d
2
w
dx
2
z
2
dydz = E
d
2
w
dx
2
_
A
z
2
dzdy = EI
d
2
w
dx
2
where I is known as the second moment of area.
Fig. 6.6. Forces at the cross section of a beam.
Inside the beam (assuming no axial forces) acts the moment M(x) and
the shear force T(x), which together balance the external distributed load q,
see Fig. 6.7. From equilibrium we have
T(x +x) T(x) = qx,
dT
dx
= q,
and
M(x +x) M(x) T(x +x)x +qxx/2 = 0,
dM
dx
= T.
52 6. Stability problems
Thus
d
2
M
dx
2
= q,
and
d
2
dx
2
_
EI
d
2
w
dx
2
_
= q. (6.2)
This is known as the beam equation and can (in special cases) be integrated
to give the deections of the beam. Boundary conditions for the beam include
w = 0, a support; dw/dx = 0, no rotation; d
2
w/dx
2
, no moment (a friction
free joint).
Fig. 6.7. Equilibrium in a beam.
We are now interested in the stability problem of the beam. In order to
nd this, we have to study equilibrium in a deformed setting, Fig. 6.8.
Fig. 6.8. Equilibrium in a beam including the eect of an axial force.
6.2 The beam 53
The axial force N acts along the centroid line, and w = w(x+x)w(x).
For the equilibrium of bending moments we now get
M(x +x) M(x) T(x +x)x N(x +x)w +qxx/2 = 0,
leading to (N is constant, since N(x +x) N(x) = 0)
d
2
M
dx
2
= q +N
d
2
w
dx
2
.
or
d
2
dx
2
_
EI
d
2
w
dx
2
_
N
d
2
w
dx
2
= q.
Now, if q = 0 and N = P, we have, with constant EI, that
d
4
w
dx
4
+
P
EI
d
2
w
dx
2
= 0.
This is a continuous eigenvalue problem with := P/EI, which can be solved
using the characteristic equation
r
4
+
P
EI
r
2
= 0 = r
1
= r
2
= 0, r
3
= i
_
P
EI
, r
4
= i
_
P
EI
.
and the exact solution is on the form
w(x) = Asin
_
P
EI
x +Bcos
_
P
EI
x +Cx +D.
The constants are decided from the boundary conditions. For example, a
simply supported beam has boundary conditions w(0) = w

(0) = w(L) =
w

(L) = 0, so
w(0) = 0 B +D = 0,
w(L) = 0 Asin
_
PL
2
EI
+Bcos
_
PL
2
EI
+CL +D = 0,
w

(0) = 0
P
EI
B = 0,
w

(L) = 0 A
P
EI
sin
_
PL
2
EI
B
P
EI
cos
_
PL
2
EI
= 0.
which means B = D = 0 and
Asin
_
PL
2
EI
= 0,
leading to C = 0. Thus, either w = 0 or
54 6. Stability problems
sin
_
PL
2
EI
= 0 =
_
PL
2
EI
= n.
The lowest nontrivial root to this equation is n = 1, and thus
P
cr
=

2
EI
L
2
.
This is known as the buckling load for which a catastrophic instability occurs.
Again, large deformation analysis may be used to predict the deection after
buckling, which is unknown in this linearized problem.
In 1991, a similar phenomenon led to the collapse of the Sleipner A oil
platform, evidently due to the analysts missing to take into account the sta-
bility problem; the cost of this little oversight: $700 million.
6.3 Instability and the minimum of potential energy
Returning to our original simple problem in Fig 6.1, we can set up the poten-
tial energy . It consists of the elastic energy stored in the spring, minus the
potential decrease of the external load. From Fig. 6.1 we see that the load
has moved the distance
= L
L
2
cos
L
2
cos = (1 cos )L.
Thus the potential energy of the system is
() =
1
2
k
2
PL(1 cos ) = 2k
2
PL(1 cos ).
To nd the rotation angle, we may use the principle of minimum potential
energy to nd
0 =
d
d
= 4k PLsin ,
which we already deduced. We may now also see if this is a true minimum of
the potential energy. Thus we compute
d
2

d
2
= 4k PLcos .
Now, since | cos | 1 we have d
2
/d
2
> 0 as long as 4k/L > P. On the
other hand, if P > 4k/L, d
2
/d
2
< 0 at = 0, and the equilibrium is
unstable. The situation is depicted in Fig. 6.9. Note that the local minimum
changes to a local maximum as PL/k > 4.
For a continuous problem, the energy is a functional, and we must use
variational derivatives to deduce the sign of the second derivative of potential
energy. The change of sign of second derivatives of potential energy is echoed
in the nite element approximation of large deformation problems in which
the linearized system becomes singular. (The system itself being the rst
derivative of potential energy.)
6.3 Instability and the minimum of potential energy 55
4 3 2 1 0 1 2 3 4
5
0
5
10
15
20
q
W


PL/k = 1
PL/k = 4
PL/k = 6
Fig. 6.9. Potential energy function ().