Chapter6

EIGENVALUEAND
TIME-DEPENDENTPROBLEMS
I
I,',
6.1 EIGENVALUEPROBLEMS
, I
6.1.1 Introduction
An eigenvalue problem is defined tobeonein which weseek thevalues of the parameter A
,I: such that theequation
'I
, ,
A(u) =AB(u) (6.1.1 )
is satisfied for nontrivial values of u. Here A and B denote either matrix operators or
differential operators, and values ofAfor which Eq. (6.1.1) issatisfied are called eigenvalues.
Foreach value of Athere is a vector u, called aneigenvector or eigenfunction. Forexample,
theequation
d
2
with A=-- B=]
X
2'
d
which arises inconnection with natural axial vibrations of a baror the transverse vibration
of a cable, constitutes an eigenvalue problem. Here Adenotes the square of the frequency
of vibration, to.
Ingeneral, thedetermination of theeigenvalues is of engineering as well as mathemat-
ical importance. In structural problems, the eigenvalues denote either natural frequencies
or buckling loads. In fluid mechanics and heat transfer, eigenvalue problems arise in con-
nection with the determination of the homogeneous parts of the transient solution, as will
be shown shortly. In these cases, eigenvalues often denote amplitudes of the Fourier com-
ponents making up the solution. Eigenvalues are also useful in determining the stability
characteristics of temporal schemes, as discussed in Section 6.2.
In this section, we develop finite element models of eigenvalue problems described by
differential equations. In view of the close similarity between the differential equations
governing eigenvalue and boundary value problems, thesteps involved intheconstruction
of their finite element models areentirely analogous. Differential eigenvalue problems are
reduced to algebraic eigenvalue problems (i.e., [A](X} = A[B](X}) by means of the finite
element approximation. Themethods of solution of algebraic eigenvalue problems arethen
used to solve for theeigenvalues and eigenvectors.
. ,
292 ANINTRODUCTION TOTHEFINITEELEMENT METHOD
6.1.2 Formulation of Eigenvalue Problems
Parabolic Equation
Consider the partial differential equation
peAau _ !- (kA au)=q(x,t)
(6.1.2)
at ax ax
which arises in connection with transient heat transfer in one-dimensional systems (e.g.,
a plane wall or a fin). Here u denotes the temperature, k the thermal conductivity, p the
density, Athe cross-sectional area, e the specific heat, and qthe heat generation per unit
length. Equations involving thefirst-order time derivative are called parabolicequations.
The homogenous solution (i.e., the solution when q= 0) of Eq. (6.] .2) is often sought
intheform of a product of a function of xand a function of t(i.e., through the separation-
of-variablestechnique):
uli(x,t)= V(x)T(t) (6.1.3)
Substitution of this assumed form of solution into the homogeneous form of (6.1.2) gives
dT d ( dV)
peAV--- kA- T=O
dt dx dx
Separating variables of t and x (assuming that peA and kA are functions of x only), we
arrive at
1 dT 1] d ( dV)
Tdt= peAVdx kAdx (6.1.4)
Note thattheleft-hand side of this equation is a function of t only while theright-hand side
is a function of x only. For two functions of two independent variables to be equal for all
val ues of theindependent variables, both functions must beequal tothesame constant, say
-'A ('A>O):
] dT
Tdt=
lId ( dV)
peAVdx kAdx =-'A
or
dr
-=-'AT
dt
(6.1.5)
d ( dV)
dx dx
-- kA- - 'ApeAV =0 (6.1.6)
The negative sign of the constant 'A is based on the physical requirement that the solution
V (x) be harmonic in x while T(t)decay exponentially with increasing t.
The solution of (6.1.5) is
Ke-
Ai
T(t)= (6.1.7)
where K is a constant of integration. The values of 'A are determined by solving (6.].6),
which also gives Vex). With T(t) and Vex) known, we have the complete homogeneous
solution (6.1.3) of Eq. (6.1.2). Theproblem of solving (6.1.6) for 'A and V(x) is termed an
eigenvalueproblem, and 'A is called the eigenvalue and V(x) the eigenfunction. When k,
CHAPTER 6; EIGENVALUE ANDTIME-DEPENDENT PROBLEMS 293
A, p, and c areconstants, thesolution of Eq. (6.1.6) is
2 pc
U(x) = C sinax +D cosax, a =-A (6.1.8)
k
where C and D areconstants of integration. Boundary conditions of the problem areused
to find algebraic relations among C and D. The algebraic relations can be expressed in
matrix form as ([A] - a[BDIV} = 10}, where [A] and [B] depend, ingeneral, on k, c, and
p, and {V} is the vector of integration constants C and D. Since C and D both cannot be
zero (otherwise, we obtain tile trivial solution), werequire the determinant of the coefficient
matrix [A] - arB] tobezero. This results inan algebraic eigenvalue problemwhose solution
yields a (hence, A) and IV} = (C, D).
To fix the ideas, consider Eq. (6.1.6) subject tothe boundary conditions (e.g., afin with
specified temperature at x =0 and insulated at x. =L)
[
dU]
U(O) =0, kA- =0 (6.1.9)
dx x=t
We note that nonhomogeneous boundary conditions can be converted to homogeneous
boundary conditions by a change of variables. Using the above boundary conditions in
(6.1.8), we obtain
o= C . 0 +D . I, 0 = a (C cos aL - D si na L)
or
(6.1.l 0)
For nontrivial solution (i.e., not both C and D areequal tozero), weset the determinant of
the coefficient matrix in (6.1.10) to zero and obtain (since a cannot be zero)
(211 - l)n
cosaL =0 allL = 2
Hence, the homogeneous solution becomes [note that the constant K of Eq. (6.1.7) is
absorbed into C
n
]
11 -t. I . 2 ( k ) (211 - l)n
u (x, t) = LJ Ci, smc;, All = all - , all = --- (6.1.11) n
11=1 pc 2L
Theconstants C; aredetermined using theinitial condition of the problem, u(x, 0) = uo(x):
00
i(x,O)= LCn sin all x =uo(x)
11=]
Multiplying both sides with sino.,, integrating over the interval (0, L), and making use
of the orthogonality condition
t 1, ifm #11
sin anx sin alii x dx = L if _
(6.1.12)
i
0
o 2' I m-11
weobtain
CII = - uo(x) sin all x dx (6.1.J 3)
L 0
21
L
294 AN INTRODUCTION TOTHE FINITE ELEMENT METHOD
The complete solution of Eq. (6.1.2) is given by the sum of the homogeneous solution and
the particular solution u(x, l) = uh(x, t) +uP(x, t).
The example discussed above provides the need for determining eigenvalues (an) in
thecontext of finding the transient response of a parabolic equation. Next, weconsider the
transient response of a second-order equation in time, known as a hyperbolic equation.
Hyperbolic Equation
The axial motion of a bar, for example, is described by the equation [see Reddy (2002),
pp. 185-187]
a
2u
a ( au)
pA- - - EA- = f(x, t) (6. 1.l4)
al
2
ax ax
where u denotes the axial displacement, E the modulus of elasticity, A the cross-sectional
area, p thedensity, and f the axial force per unit length. The solution of (6.1.14) consists
of two parts: homogeneous solution u" (i.e., when f = 0) and particular solution up. The
homogeneous part is determined by the separation-of-variables technique, as we discussed
for the parabolic equation.
The homogenous solution of Eq. (6.1.14) is also assumed to be of the form in (6.1.3).
Substitution of (6.1.3) intothe homogeneous form of (6.1.14) gives
2T
d d ( dU)
pAU--- EA- T=O
dt? dx dx .
Assuming that pA and EA are functions of x only, we arrive at
1 d
2
T 1 1 d ( dU) 2
T dt 2 = pA Udx kA dx =-a
or
d
2T
-2 +a
2
T = 0 (6.1.l5)
dl
--
d (
EA-
dU)
-a
2
pAU=O (6.1.16)
dx dx
The negative sign of the constant a is based on the physical requirement that the solution
u(x, l) be harmonic in x and t,
The solution of (6.1.15) is
Ke-
i a l
T(l) = = K[ cosor +K
2
sinal (6.1.l7)
where K 1 and K
2
areconstants of integration. Thesolution of (6.1.16), when E, A, and p,
areconstants, is
-2 P 2
U(x) = C sinex +D cosex, a =-a (6.1.l8)
E
where again C and D areconstants ofintegration. In theprocess ofdetermining theconstants
C and D using theboundary conditions of theproblem, onceagain wearerequired to solve
aneigenvalue problem (the steps areanalogous tothose described for a parabolic equation).
Unlike in the case of heat conduction problem, the quantities an in the case of bars have
direct physical meaning, namely, they represent natural frequencies of the system. Thus, in
CHAPTER 6: EIGENVALUE ANDTIMEDEPENDENT PROBLEMS 295
structural mechanics we may be interested in determining only the natural frequencies of
thesystem but not thetransient response.
When weare interested only innatural frequencies ofthesystem, theeigenvalue problem
may be formulated directly from the equation of motion (6. I.14) by assuming a solution
form that is periodic intime t
U(x, t) = U(x)e-
i w1
(6.1.19)
where to denotes the frequency of natural vibration, i equals P, and U(x) denotes the
configuration of the structure at that frequency, called the mode shape (i.e., for each value
of t, there is an associated mode shape). Substitution of (6.1.19) into the homogeneous
form of (6. I.14) gives
[
- PAW
2
U _ (EA dU)] e-
i 1ul
=0
dx dx
or
d ( dU) 2
-- EA- -w pAU=O
dx dx
which isidentical toEq. (6.1. I 6)with a =to. Similar equations can befonnuJated fornatural
vibration of beams. Another eigenvalue problem that arises directly from the governing
equilibrium equation is that of buckling of beam-columns. We shall return to this topic
shortly.
In summary, eigenvalue problems associated with parabolic equations areobtained from
thecorresponding equations of motion by assuming solution of the form
u(x, t) =U(x)e-
l1l
, A=a (6.1.20a)
whereas those associated with hyperbolic equations areobtained by assuming solution of
the form
(6. I.20b)
where Adenotes theeigenvalue. Finite element formulations of both problems are presented
next.
6.1.3 Finite Element Formulation
Comparison of Eqs. (6.1.6) and (6.1.16) with the model equation (3.2.1) reveals that the
equations governing eigenvalue problems arespecial cases of the model equations studied
in Chapters 3 and 5. Here we summarize the steps in the finite element formulation of
eigenvalue problems for the sake of completeness and ready reference. We will consider
eigenvalue problems described by (a) a single equation in a single unknown (e.g., heat
transfer, bar, and Euler-Bernoulli beam problems), and (b) a pair of equations in two
variables (e.g., Timoshenko beam theory).
Heat Transfer and Bar-Like Problems
Consider theproblem of solving theequation
d [ dU]
-- a(x)- +c(x)U(x) =ACo(x)U(x) (6.1.21)
dx dx
\
I
,.
:
I'
f
"
!
I.
296 ANINTRODUCTION TOTHEFINITE ELEMENT METHOD
for Aand V(x). Here a, c, and Co areknown quantities that depend onthe physical problem
(i.e., data), Ais the eigenvalue, and V is the eigenfunction. Special cases of Eq. (6.1.21) are
given below.
Heat transfer: a = kA, C == Pj3, Co = pcA (6.1.22)
Bars: a = r,A, c=O, co=pA (6.1.23)
Over a typical element Qe, we seek a finite element approximation of V in the form
11
=Luj1/Jj(x)
(6.1.24)
j=1
The weak form of (6.1.2 I) is
0= t
b
(a
dW
dV +cwV(x) -ACOWV) dx - - (6.1.25)
J,
'(1
dx dx
where IV is the weight function, and and are the secondary variables at node 1 and
node 11, respectively (assume that Qf =0 when i # I and i 1'= n)
, (6.1.26)
dx X(I dx Xb
Substitution of thefinite element approximation into theweak form gives thefinite element
model of the eigenvalue equation (6.1.21):
(6.1.27a)
where
r [ ] t
b
d1/Je
K
0
=JX(I a(x) d; +c(x)1/Jt1/Jj dx, M
0
=JX(I co(x)1/Jr1/Jj dx (6.1.27b)
Equation (6.1.27a) contains the finite element models of the eigenvalue equations (6.1.6)
and (6.1.16) as special cases.
The assembly of element equations and imposition of boundary conditions on the as-
sembled equations remain thesame asinstatic problems ofChapter 3. However, thesolution
of the condensed equations for the unknown primary nodal variables is reduced to an alge-
braic eigenvalue problem in which the determinant of the coefficient matrix is set to zero
to determine the values of Aand subsequently the nodal values of the eigenfunction Vex).
These ideas are illustrated through examples.
Example 6.1.1 __ -,-__
Consider a plane wall, initially at a uniform temperature To, which hasboth surfaces suddenly
exposed toafluid attemperature Tco. The governing differential equation is (forconstant data)
a
2r
et
k-=pco- (6.1.28a)
ax
2
at
and theinitial condition is
T(x, 0) =To (6.1.28b) j

:1
CI1APTER 6: EIGENVALUE ANDTIME-DEPENDENT PROBLEMS 297
, ,
wherekis thethermal conductivity,pthedensity, andCo thespecificheatat constantpressure.
Equation(6.1.28a)isalsoknownasthediffusionequationwithdiffusioncoefficienta= k/pci:
We consider two sets ofboundary conditions, each being representative ofadifferent
scenarioforx= L.It amounts tosolving Eq. (6.J.28a) and (6.l.28b) fortwo different setsof
boundaryconditions.
Set 1. If theheattransfercoefficientatthesurfaces ofthewall isassumed tobeinfinite,
theboundaryconditionscanbeexpressedas
T(O, t)= Too, T(L, t)= 1'00 fort > (6.1.290)
Set2. If weassumethat thewall atx= Lis subjected to ambient temperature, wehave
T(O, t)= Too,
[
kaT +(3(T - Too)] I =0 (6.1.29b)
ax x=L
Equation(6.1.28a)canbenormalizedtomaketheboundaryconditionshomogeneous.Let
k T- Too
rv-_
/./=--- u-
To - Too PCo
Thedifferentialequation(6.1.280),initialcondition(6.l.28b),andboundaryconditions(6.1.290)
and (6.1.29b) become
(6.1.300)
U(O,t)=o, u(I,t)=O u(x,O)=1 (6.1.30b)
{3L
U(O, t)= 0,

+HU) I =0,
H=- (6.1.30c)
ax x=1
k
wherethebarsoverxand1 areomitted in theinterestof brevity.
Solution of(6.1.300) and (6.1.30b) byseparation-of-variables technique (orsubstitute
A
u= Ue- ( ) leads to thesolutionof theeigenvalueproblem
d
2U
-- - AU=0, U(O) =0, U(l)=0 (6.1.31a)
dx
2
whereassolution of(6.1.300) and (6.1.30c) results in
(
dU+HU)1 =0 (6.1.3Ib)
dx .<=1
This differential equation is aspecial caseof Eq. (6.1.21) with 0 = I,C= 0,and Co = I.
Foralinearelement,theclement equations(6.1.27a)havetheexplicitform [seeEq.(3.2.34)
for elementmatrices]
(6.1.320)
[- - ]- .] }= {
Foraquadraticelement, we have [seeEq, (3.2.370) for elementmatrices]
I [ 7
- -8 (6.1.32b)
- [
(
3he I
-8 7 30 -I 2
I
298 ANINTRODUCTIONTOTHEFINITEELEMENT METHOD
Solution for Set 1. For a mesh of two linear elements (the minimum number needed for
Set I boundary conditions), with III=11
2
=0.5, the assembled equations are
( [
I -1 0] I[2 0]) lVI 11 Q: I
2 -I 2 -1 -A- I 4
I
1 V
2
= Qi+Qi
o -1 I 12 0 I 2 V
3
Q2
2
The boundary conditions V(0) =0, Qi+ Qi = 0, and V(I) =0 require VI = V
3
= O. Hence,
the eigenvalue problem reduces to thesingleequation
or AI=12.0, V2=F0
The modeshapeis given (within anarbitrary constant; take V
2
= I) by
O::::x::::O.5
V
Zep2(X)
= V(x)= = (21z - x)/h=2(1 - x) 0.5::::x:::: 1.0
For a mesh of one quadratic element, we have (11 = 1.0)
The corresponding eigenfunction is
The exact eigenvalues are All =(mr)2 andA1 =Jr2 =9.8696. The exacteigenfunctions for
Set 1boundary conditions are VII(x)=sinI1Jr x and VI(x)=sinnx. Clearly, one quadratic
element gives more accurate solution thantwolinearelements.
The eigenvalues and eigenfunctions can be used to construct the solution of the transient
problem. For example, thesolution of (6.1 JOa) and(6.IJOb) is
00 00
2
u(x, 1)=LC
IIVII
(x )e- 1.
1I
1 =LCII sinnx.c e-(n
Jr
J, (6.1.33)
11=1 11=1
where CII are constants to be determined using the initial condition of the problem [see
Eq. (6.1.13)]. The finite element solution of the sameproblem, when one quadratic element is
used, is given by
which is very closeto the one-term solution in (6.1.33).
Solution for Set 2. For a mesh of two linear elements, the Set 2boundary conditions
translate into VI =0 and + HV
3
=O. The condensed equations are
or
(6.1034)
h
j
CHAPTER 6: EIGENVALUE AND TIMEDEPENDENT PROBLEMS 299
- _._-- - - - ----- ..----------
:1
Fornontrivial solution (Le., at least oneoftheU, is nonzero), thedeterminant ofthecoefficient
matrix should be zero:
1
4 - 4X -(2 + X) I
-(2 + X) 2 + H - 2X = 0,
_ A
A=-
12
or
7F - 4(5+H)X +4(1 +H) =
The above equation is known as the characteristic equation of the eigenvalue problem. The
two roots of this quadratic equation for H = I are
_ 12J(12)2-7x8
A12=
, 7.
--+ XI =0.374167, X
2=3.0544
and theeigenvalues are (A = 12X)
Al = 4.49, A2 = 36.6529
Theeigenvectors associated with each eigenvalue can becomputed from equations (6.1.34)
[

-(2 + Aj)

3 - 2Aj U
3
0' i = 1,2
Forexample, for XI = 0.374167 (A} = 4.4900), wehave
2.5033ui
l
) - 2.3742Ui
1
) =0
Taking uil) = 1,we obtain
{
u (l}} = {Ui
l
} }
ujl}
= {1.0000 }
1.0544
On theother hand, if wenormalize thecomponents, (Uil})2 + (Ujl)2 = 1, weobtain
{
U(I )} = {Uil) } =
uil}
{0.6881 }
0.7256
In other words, weobtain amode shape whose amplitude isnotunique. Hence, theeigenfunction
corresponding to Al =4.4900 (h = 0.5) is
U(l)(x)
.
= {0.6881 x] h
0.6881 (1 - x)1h +0.7256(2x - 1)/2h
for 0 x 0.5
for 0.5 x 1.0
Similarly, theeigenvector components corresponding to A2 = 36.6529 is (notnormalized)
{
U(2)} = {UY} } = { 1.0000}
UP) -1.6258
Fora mesh of onequadratic element, thecondensed equations are
and thecharacteristic equation is
15F - 4(13 +3H)X+12(1 +H) =0, X=
10
300 ANINTRODUCTION TOTHEFINITEELEMENT METHOD
Thetworoots of this quadraticequationforH =1are
_ 32 /(32)2 - 15x24
A - ~ Xl =0.41545, X
2
=3.85121
1,2- 15-
and theeigenvaluesare(A =l2X)
Thecorrespondingeigenvectorcomponentsare(notnormalized)
(
U(I) = {ViI)} = {1.aoOO}
- V ~ J 1.0591
{
V(2)= {Vi
2
) } = {- 1.0000}
- V?) -2.9052
Theexacteigenfunctionsfor Set2boundaryconditionsare
VII (r) = sinAx (6.1.35a)
and theeigenvaluesAll arecomputedfrom the equation
HsinA+AcoSA=O (6.1.35b)
Thefirst two roots of thetranscendental equation (6.1.35b) are(for H = 1)
j):;=2.0288 ~ Al =4.1160; A=4.9132 ~ A2=24.1393-
A comparisonofthe eigenvalues obtained using meshes oflinearand quadraticelements
withtheexactvaluesispresentedinTable6.1.1.Notethatthenumberofeigenvaluesobtainedin
Table 6.1.1 Eigenvaluesoftheheatconductionequations(6.1.31a) and(6.l.31b) fortwo sets
of boundaryconditions.
Mesh )'1 A2 A) A4 AS A6 A7
2L 12.0000
(4.4900)* (36.6529)
4L 10.3866 48.0000 126.756
(4.2054) _(27.3318) (85.7864) (177.604)
8L 9.9971 41.5466 99.4855 192.000 328.291 507.025 686.5J2
(4.1380) .(24.9088) (69.1036) (143.530) (257.580) (417.701) (607.022)
lQ 10.000
(4.1545) (38.5121)
2Q 9.9439 40.000 128.723
(4.lJ96) (24.8995) (81.4446) (207.653)
4Q 9.8747 39.7754 91.7847 J60.000 308.253 514.891 794.794
(4.1161) - (24.2040) (64.7704) (129.261) (240.540) (405.254) (658.133)
Exact 9_8696 39.4784 88.8264 157.9137 246.740 355.306 483.611
(4.1160) (24.1393) (63.6597) (122.889) (201.851) (300.550) (4J8.987)
Thesecond linefor eachmesh correspondsto set 2boundary conditions.
____J
-- - - - - - - - - - - - ~ . - - ~ - - - ~ - - _ . ."".------ - ~ - - - - - - - - - - -----
,
CIIAPTER 6:EIGENVALUE ANDTIMEDEPENDENTPROBLEMS 301
--_.- .. ---
i
2.00
I
l..50
1.00
,.-..

0.50
:s-
ui
Q)
o:l
0..
0.00
.J::
en

0
-0.50
;;E
--81inear
-1.00
elements
-1.50
- 4quadratic
elements
-2. 00
Coordinate,x
j
I
Figure 6.1.1 Thefirstthreemodeshapesaspredictedbyamesh ofeightlinearelementsanda
I
I
mesh of four quadratic elements for the heat transfer problem of Example 6.1.1
(Set 1boundary conditions).
I
Coordinate,x
Figure 6.1.2 Thefirstthreemodeshapesof theheattransferproblemof Example6.1.1 (Set2
boundaryconditions).
.
, 0.4
0.0 0.2 0.6 0.8 1.0
2.00
1.50
1.00
r<:
>:
:5'
vi
0.50

fJ'
..c:
III

0.00
'0
0

-0.50
-1.00
-1.50
0.0 0.2 0.4 0.6 1.0
j
I
I
I
I
I
i
--
302 AN INTRODUCTION TOTHEFINITE ELEMENT METHOD
the finite element method isalways equal to the number ofunknown nodal values. As the mesh
isrefined, not only do we increase the number ofeigenvalues but we also improve the accuracy
of the preceding eigenvalues. Note also that the convergence of the numerical eigenvalues to
the exact ones is from the above, i.e., the finite element solution provides an upper bound to
the exact eigenvalues. For structural systems, this can be interpreted as follows: According
tothe principle of minimum total potential energy, any approximate displacement field would
overestimate the total potential energy of the system. This is equivalent to approximating the
stiffness ofthe system with a larger value than the actual one. Astiffer system will have larger
eigenvalues (or frequencies). The first three mode shapes ofthe system are shown inFigs. 6.].1
and6.1.2.
We note that the eigenvalue equations (6.1.31a) and (6.1.31b) can also be interpreted
as those arising in connection with the axial vibrations of a constant cross sectional mem-
ber. In that case, U denotes the axial displacement [u(x It) = U(x)e-
iwt
] and A= (j} p/ E,
w being the frequency of natural vibration. For example, the boundary conditions in
(6.1.31 b) can be interpreted as those of a bar fixed at the left end and the right end
being connected to a linear elastic spring (see Fig. 6.1.3). The constant H is equal to
k/ EA, k being the spring constant. Thus, the results presented in Table 6.1.1 can be in-
terpreted as w
2
p/ E of an uniform bar for the two different boundary conditions shown in
Fig. 6.1.3. .
Natural Vibration of Beams
Euler-Bernoulli Beam Theory
For the Euler-Bernoulli beam theory, the equation of motion is of the form [see Reddy
(2002), pp. 193-196]
(6.1.36)
wherep denotes themass density perunit length, A theareaof cross section, E themodulus,
and J the second moment of area (see Chapter 5). The expression involving pI is called
/E,A
x
L=l
E,A
k
I
J
L=l
-I
x J
Set I: U(O) = 0, U(L)=O
Set 2: U(O) = 0, [ c::: + flU] =0
. ;<=/.
Figure 6.1.3 Uniform bar with two sets of boundary conditions.
CHAPTER 6: EIGENVALUE AND TIME-DEPENDENT PROBLEMS 303
rotary inertia term. Equation (6.1.36) can be formulated as an eigenvalue problem in the
interest of finding the frequency of natural vibration by assuming periodic motion
w(x, 1) =W(x)e-
iwt
(6.1.37)
where to is the frequency of natural transverse motion and W(x) is the mode shape of the
transverse motion. Substitution of Eq. (6.1.37) into (6.1.36) yields
Z
d (dZW) ( dZW)
(6.1.38)
- EI- -'A pAW-pl-'.- =0
dx
z
dx
z
dx
z
z
where 'A =w .
Theweak form of Eq. (6.1.38) is given by
l
XI, ( dv dW) dZv dZW
0= EI- -z - 'ApAvW - 'Apl.--- dx
z
dx dx dx dx
n
Z
Z
X
I
[
d ( d W) d IV] I [( d) d
(6.].39)
XI> W ] XI>
+ V _. EI- +'Apl-'.- + EI-'-
dx dx
z
dx, dx dx
z
,

where v is the weight function. Note that the rotary inertia term contributes to the shear
force term, giving rise to an effective shear force that must beknown at a boundary point
when thedeflection is unknown at the point.
To obtain thefinite element model of Eq. (6. J.39), assume finite element approximation
of theform
4
W(X) =L jcPj(X)
(6.1.40)
j=l
where j are the Hermite cubic polynomials [see Eqs. (5.2.1]) and (5.2.12)], and obtain
thefinite element model
(6.1.41 a)
where
XI, dZC dZcP dcP dC)
C
lXi' (
C
= El-'-i -,_.I d.x = +pI _i _J dx
lJ d Z d Z " lJ 'v, 'f'J d d '
X .x .x X
n
,xx
l
a
(6.J.41b)
Z
Z
== d
d ( dW )
+'ApI
T
dHI] I
'
( dW) I
[
El
dT
El
dT ,,x ,x ,x X .x X
n a
Z
2
d ( d W) dW] I ( d W) I (6.1.41 c)
== - - EI- +),pJ-. ' Q4==- EI--
[ 2
2
dx d.x dx r dx \'
, I> '/'
For constant values of EI and pA, the stiffness matrix [K
C
] and mass matrix [M
C
]
are
-3h
c
-6
-311, ]
2h
2
3h
c
IK'] = 2E,I, [-
'c
. 3h
c
6 3h
c
-3h
c
3h
c
2h
C
Z
304 ANINTRODUCTION TOTHEFINITEELEMENTMETHOD
[Me] = peAehe _
. 420 54 -13h
e
156 22h
e
13h -3h; 22h
e
4h;
36 -3he -36 -3he]
pet, -3h 4h; 3h
e
-h;
(6.1.42)
+30h
e
-36 3h
e
36 3h
e
[
-3h
e
-h; 3h
e
4h;
When rotary inertia is neglected, weomit the second part of the mass matrix in (6.1.42).
Timoshenko BeamTheory
The equations of motion of the Timoshenko beam theory are [see Reddy (2002), pp. 193-
196]
a
2w
a[ (aw )]
pA-- - - GAK
s
- +\}J = 0 (6.1.43a)
at
2
ax ax
2\}J
a a (a\}J) (ow)
pf - - - 1- +GAK
s
- +\}J = 0 (6.1.43b)
at
2
ax ax ax
where G is the shear modulus (G = E/[2(l +v)])and K.\. is the shear correction factor
(K, = 5/6). NotethatEq. (6.1.43b) contains therotary inertia term. Onceagain, weassume
periodic motion and write
W(x,t)= W(x)e-
i w1
, \}J(x, t)= S(x)e-
i w1
(6.1.44)
and obtain the eigenvalue problem from Eqs. (6.l.43a)and (6.1.43b)
_5 [GAKs (d.W +S)] - (ipAW = 0 (6.1.45a)
dx dx
(/S)+GAK
s
(dW +S) - uipiS= 0 (6.1.45b)
dx dx dx .
FOI" equal interpolation of W(x)and. S(x)
/I /I
W(x)= L WJ1fij(x), S(x)=L Sj1fi.j(x) (6.1.46)
j=l j=]
where V/j are the (11 - I) order Lagrange polynomials, thefinite element model is given by
(6.1.47)
1
where [K
e]
is the stiffness matrix and [Me] is the mass matrix
.1" d,lre d1/te
K
11
= GAK _'f'_i _Jdx
lJ.
s
dx dx
A"
--------------
CHAPTER 6: EIGENVALUE ANDTIME-DEPENDENT PROBLEMS 305
(6.1.48a)
== _ [GAKs (S + I ' == - (/S) I
dx ,. dx .r
An a
== [GAKs (S +d":)] I, == (/S) I (6.1.48b)
dx ,. dx X
.ll.b b
For the choice of linear interpolation functions, Eq. (6.1.47) has the explicit form (after
rearranging the nodal variables)
-3h
e
-6
+6A
e
) 3h
e
3h
e
6
h;(I.5 - 6A
e
) 3h
e
r
o]
e
(6.1.490)
o '
Lr,
(6.1.49b)
For Hermite cubic interpolation of W(x) and related quadratic approximation of S(x)
[i.e., interdependent interpolation element (IJE)], the resulting mass matrix is cumbersome
and depends on the stiffness parameters ( and G). We will notconsider it here [see Reddy
(I 999b, 2000)].
Example 6.1.2 _
Consider a uniform beam of rectangular cross section (B x H), fixed atx =andfreeat x =L
(i.e., a cantilever beam). We wish to determine the first four flexural frequencies of thebeam.
Theboundary conditions of thestructure arc
W(O) =0, =0, =0, =0 (6.1.50a)
dx .x X""/- dx ,l"o=L
in theEuler-Bernoulli beam theory and
W(O) '=0, 8(0)=0, [E/
S]
=0, [GAKs (:: +s)] _=0
(6.1.50b)
dx >:=1. x-I.
intheTimoshenko beam theory. Thefirst twoterms ineachcasearetheessential (orgeometric)
boundary conditions andthelast two arethenatural (or force) boundary.conditions.
...... .._--

306 ANINTRODUCTION TOTHE FINITE ELEMENT MEnloD

Thenumber ofeigenvalueswe wish todetermine dictates theminimumnumber ofelements
tobeused toanalyze the structure. First, consider theEuler-Bernoulli beam element. Since
two primary degrees offreedom arespecified (VI=V
2
=0),aone-element mesh will have
only twounknowndegreesoffreedom (U3 and V
4
) .Hence,wecan determineonlytwo natural
frequencies. Thus,a minimum of twoEuler-Bernoullibeamelementsareneeded todetermine
fournatural frequencies. If weusethereduced integration Timoshenko beam element, a mesh
oftwo linear elements canonly represent thefirst two mode shapes ofthecantilever beam
(seeFig. 6.1.4) because there areonly two independent dellection degrees offreedom that
are unspecified. In orderto representthefirstfourmodeshapes using thereducedintegration
element(RIE),wemust useatleast fourlinear elements or two quadratic elements. However,
thefour computedfrequencies may not be thelowestfour.
ForiJlustrative purpose, first weconsider theone-element mesh oftheEuler-Bernoulli
beamelement. We have
13L ]
-3L
2
22L
4L
2
W1=0,0
1
=0
W
2,02
W
1
=0,SI=0 W
2
,S2 W
3
,S3
x x
] I 2 3
11'(.1) =W11 (x)+0
1
2(x) ( W
e
't'
e
(x)+W
e
't'
e
(x)
}
I I 2 2 Perelement
+W
2
3(x)+O
2
4(x)
Sex) = +S;o/;(x)
W(x) - - - - Euler-Bernoullibeam element
-- Timoshenkobeam elements
Mode shape 2
(
I ) (1) /'Mode shape2
If =W-/ ""
2 I ""
W2(21
Mode shape I./"
/V(I)= W(2)
2 \ I
Modeshape J
Figure 6.1.4 Twopossiblenonzeromodeshapesof acantileverbeam that can be represented
byamesh oftwo linear Timoshenko beam elements (oroneEuler-Bernoulli
beamelement).
------------
r
CHAPTER 6: EIGENVALUE AND TIME-DEPENDENT PROBLEMS 307
The specified displacement and force degrees of freedom are WI = 0, 8
1
= 0, Q3 = 0, and
Q4 = 0. Hence, the condensed equations are
2E/ [ 6 3L]_u}PAL[156 3L]){W
2}={0}
(
V 3L 2L
2
420 22L 4L
2
30L 3L 4L
2
8 2
Setting the determinant of the coefficient matrix in the above equation to zero, we obtain a
quadratic characteristic equation inw
2
.
First, consider thecasein which therotary inertia is neglected. Thecharacteristic equation
for this caseis
2 pAl} 2
15120-1224A+A =0, A=--W
EI
whose roots are
=3.5327 J1, "'2 =34.8069 j1
A( = 12.48 A2 = 1211.52, or (VI
U pA
U/
U pA
The exact frequencies are WI = 3.5160 and W2 = 22.0345, where Wi = wi(L2JpAlE/). The
eigenvector components arecomputed from thefirst of thecondensed equations
W(i) = -L (1260 - ltA
i
) 8(i) i = 1, 2
2 420_ 13Ai 2 '
Hence, theeigenvectors are
W2 Ill) _ {0.7259L } W2 }(2) = {0.1312L }
{
8
2
- -1.0000 '
{
8
2
-1.0000
Forthecasein which rotary is not neglected, wewrite pI = pA (H
2
/ 12) and take
HI L = 0.01. Thecharacteristic equation is
whose roots are
3.5327/!f1 34.7984/!f1
Al = 12.4797, A2 = 1210.926 or WI = --2- -., W2 = 2 -
L pA L pA
Theexact frequencies areWI = 3.5158 and W2 = 22.0226, where Wi = wi(L
2
JpAlE/).
Next, weconsider theone-element mesh of theTimoshenko beam element (RIE). We have
2EI
/LoV -6 3L 6 3L 6 1
(
-3L L
2
82 3L L
2
8 ,
= H
2/12,
and where K, = 5/6, I' = II A
EI 1+v H
2
3
J=I.5+61\,
3
2=1.5-6A,
A= =---- /La = l2l\.
GAKsU 5 U'
The specified displacement and force degrees of freedom are WI = 0, 51 = 0, Q3 = 0, and
Q4 = 0; and thecondensed equations are
2EI [6 3L ] 2
PAL
[2 0]) {W2} {O}
(
f.Lo V 3L L
2(1.5
+6/\.) - W 6 21' S2 =
308 AN INTRODUCTION TOTHEFINITEELEMENT METHOD
Settingthedeterminant of thecoefficientmatrixto zero, weobtai nthecharacteristicequation
2
F -3 (I +3s
2
+ l2s
2
1\)X+108s
2
= 0, X= P ~ 1\w
wheresisthelength-to-heightratio,s = LI H.
When therotary inertia is neglected, wehave
- 121\
A=--
1+4A
Tofurthersimplify the expression, weassume1J= 0.25 andobtain
.1 12 E1
w-= .--
(l +H2 IV) pAL
4
and
'=10: w= 3.4469JEI
H V pA
The exact frequency isW= 3.5158 for LIH = 100 and W= 3.5092 for LI H = 10, where
Wi = to, (L
2
JpAlEI). If therotary inertiaisincluded, weobtain
.= 100: w= 3.4639JEl . .~ 10: w= 3.4413JEI
H L2 pA ' H V pA
The frequencies obtained bytheEuler-Bernoulli beam elements and Timoshenko beam
elements (RIE) areshown in Table6.1.2for twodifferent values of thelength-to-heightratio
LI H. Thevalue LI H == 100makes theeffect of sheardeformationnegligible, andweobtain
essentially thesameresult asintheEuler-Bernoulli beam theory. Since thefrequencies are
normalized,W= L
2
J(pA I El), itisnecessary only toselectthevalues of1Jand LIH. For
Table 6.1.2 Natural frequencies ofacantilever beam according toTimoshenko beam theory
(TBT)andEuler-Bernoullibeam theory(EBT) [w=wL
2(pAIEI)I/2].
L/H =100 L/H = 10
Mesh WI l1J2 5)3 l1J4 wI l1J2 l1J3 l1J4
4L 3.5406 25.6726 98.3953 417.1330 3.5137 24.1345 80.2244 189.9288
8L 3.5223 22.8851 68.8937 151.8431 3.4956 21.7004 60.6297 119.2798
16L 3.5174 22.2350 63.3413 127.5434 3.4908 21.1257 56.4714 104.6799
2Q 3.5214 23.3226 78.3115 328.3250 3.4947 22.0762 67.0884 181.0682
4Q 3.5161 22.1054 63.3271 133.9828 3.4895 21.0103 56.4572 108.6060
8Q 3.5158 22.0280 61.7325 121.4458 3.4892 20.4421 55.2405 100.7496
EBTT 3.5158 22.0226 61.6179 120.6152 3.4892 20.9374 55.1530 lOO.2116
TBTT 3.5158 22.03]5 61.6774 120.8300 3.5092 21.7425 59.8013 114.2898
EBTl 3.5160 22.0345 61.6972 120.9019 3.5160 22.0345 61.6972 120.9019
t Rotaryinertia.is included(Exact).
t Rotaryinertiais neglected;the resultsare independent of theratio1-1H (Exact),
CHAPTER6:EIGENVALUE ANDTIME-DEPENDENTPROIlLEMS 309
- _._-------- ~ _ _
computational purposes, wetake II = 0.25; also, note that
E 5 4EI l/2pA
GAK= BH-=- pl=--
2(1 +v) 6 [f2' 12
where BandHare thewidth and height, respectively,of thebeam.
We notefrom Table 6.1.2 thatthefinite elementresults converge with h-refinement (i.e.,
whenmore of thesame kind ofelements areused) and also withp-refinement (i.e.,whenhigher-
order elementsareused). Thep-refinementshows more rapidconvergence of thefundamental
(i.e., first) frequency. Note that theeffect ofshear deformation istoreduce themagnitude
ofnatural frequency when compared totheEBT. Inother words, the assumed infinite shear
rigiditymakes theEBT over-predict themagnitude of frequencies. Thefirst four mode shapes
ofthecantileverbeam, asobtained using the16-elcmentmesh oflinearelements, areshown
inFig. 6.1.5.
2.00
1.50
1.00
~
~
0.50

""-
vi'
G.l
0.
0.00
'"
,J::
CIl
G.l
'"
0
-0.50
~
-1.00
-1.50
-2.00
0.0 0.2 0.4 0.6 0.8 1.0
Coordinate,x
Figure6.1.5 Firstfour natural mode shapes ofacantilever beam, aspredicted using a 16-
elementmesh oflinearTimoshenkobeam elements(Lj H = 10).
----- .._- -----
Inclosing thediscussion onnatural vibration, it isnoted that when thesymmetry of the
system isused tomodel theproblem, only symmetric modes arepredicted.It isnecessary
tomodel thewholesystem inordertoobtain all themodes ofvibration.
Stability (Buckling) of Beams
ERT
Thestudyof buckling of beam-columnsalsoleadstoaneigenvalueproblem. Forexample,
theequation governing onset ofbuckling ofacolumn subjected toanaxial compressive
310 ANINTRODUCTIONTOTHEFiNITEELEMENTMETHOD
I I
1 1
L L
(a) (b) (c:)
Figure 6.1.6 Columns with different boundary conditions and subjected to axial compressive load,
NO. (a) Hinged-hinged. (b) Hinged-clamped. (c) Clamped-clamped. (d) Clamped-free,
force NO (seeFig, 6.1.6), according to EBTis [see Reddy (l999b)]
2 2 2
d ( d W ) d W
- 1- +N -=0 (6.1.51)
dx? dx? dx?
which describes aneigenvalue problem with A= NO. Thesmallest value of NO iscalledthe
critical buckling load.
The finite element model of Eq, (6.1.51) is
(6.1.52a)
where {L),ej and (Qej are the columns of generalized displacement and force degrees of
freedom at the two ends of the Euler-Bernoulli beam element:
e
[
d (d
2W)
odW]
- 1- +N-
e dx dx? dx 1
E1-
(PW. )
(
dx?
I
(6.1.52b)
2
[
d ( d W ) dW]
-- EI-. -N-
dx dx
2
. dx 2
d2W)
-EI-
(
dx
2
2
where the subscripts land 2 refer to element nodes I and 2 (at x =X
a
and x =Xb, respec-
tively). The coefficients of the stiffness matrix IKe] and the stability matrix Ice] are
Ab d2q/ d ~ jAb de de
K{; =
j
E I ~ ---1dx, Gfj = _J dx (6.1.52c) _I
"
Aa
dx dx
~
dx dx
CHAPTER 6: EIGENVALUE ANDTIME-DEPENDENT PROBLEMS 311
where cPf are the Hermite cubic interpolation functions. The explicit form of [G
e
] is [see
thesecond part of themass matrix inEq. (6. I.42)]
36 -3171' -36
-3171']
_17
2
. ["G e] = _1_ -3111' 417; 3171' 'I'
(6.J .52d)
, 30171' - 36 3171' 36 3171'
[
-3171' -17; 3171' 417
2
I'
TET
ForTBT, theequations governing buckling of beams are [see Reddy (1999b)]
2W
d [ (dW )] od
-- G K ~ ,-+5 +N -=0 (6.1.530)
d.x . dx dx
2
d ( d5) (dW)
- - 1- +GAK
s
-, +5 = 0 (6.J .53b)
dx dx dx
Here W(x) and 5(x) denote thetransverse deflection androtation, respectively, at theonset
of buckling.
Thefinite element model ofEqs. (6.1.530) and (6.1.53/)), with equal interpolation of W
and 5, is
(6.J.540)
where
J
XIJ cHedVf
e
K,I,I = GAK _v_" _J d.x
IJ r
s
dx dx
'n
1
.1'" d1/te
K.
12
= GAK
s
_,_i J ~ dx
IJ X dx.l
n
\It == Q ~
[
= -GAK
s
(
dW
-
dx
)
+5
dW]
+NO-
dx "
'1/
I'
\III ==
I'
Q
3
=
.
[ '(dW
GAK
s
-
d.x
)
+5 -
OdW]
N -
dx "
. b
,
e
M
1
==
I'
Q
2
=
( d5)
-1-
dx r
'1/
I'
Mil ==
e
Q
4
=
( d5)
-1-
dx ,_
~
(6.J.54b)
,
i
i
i
312 ANINTRODUCTIONTOT/-IEFINITEELEMENTMET/-IOD
Forthelinear RIE, thestiffness matrix [K
e
] isgiven inEq. (6.].49a). Thestability
matrix isgiven by
-] 0]
000
o I
000
(6.].55)
Example6.1.3
Considerauniform column (L, A, I, E,andv) fixed at oneendand pinned at theother, as
shown in Fig. 6.1.7. We wish to determine thecritical buckling load.
Let usconsider a mesh of two Euler-Bernoulli beam elements. The assembled system of
finite elementequationsis(h= L/2)
6 -3h -6 -3/1 o 0
-3h 2h
2
3h h
2
o 0
2EI
-6 3h 6+6 3h- 3h -6 -3h
h
3
-3h h
2
3h- 3h 2h
2
+2h
2
3h h
2
o o -6 3h 6 3h
o o -3h h
2
3h 2h
2
36 -3h -36 -3h o o
Ql
-3h 4h
2
3/1
_h
2
o o
QJ
-36 3h 36+36 3h- 3h -36 -3h
~ Qt
-3/1 _h
2
3h- 3h
4h
2
+4h
2
311
_h
2
~ + ~
o o -36 3/1 36 3h
~
o o -3h -h
2
3h 4h
2
~
whereVI denotethegeneralized displacements of theglobalnodes.
w=o 3
CD
-2
dw
W=.- =0
dx
Figure6.1.7 Acolumnfixedatx=0 andhingedatx= L, and subjectedtoaxial compressive
load NO.
L- .. _ ..
CHAPTER 6:EIGENVALUEANDTIME-DEPENDENTPROBLEMS 313
1 -- --- ----.-- - ~ . .._-.-.."..
;
i Thegeometric supports and equilibrium conditions ofthe column require that (note that
;
theaxial compressive force NO enters the equilibrium equation and theboundary condition
i
I
I
through Q ~
!
!
Q ~ +Q1 =0, Q ~ +Q ~ =0, Q ~ = O
I
Hence, thecondensed equationsbecome
0
2El [12 0 -311; NO [72 8'12
- 0 4h
2
/z2 - - 0
(
./z3 -3h 11
2
2h
2
3011 -311 _11
2
whichdefines theeigenvalueproblemfor thedetermination ofNO andUi.
Setting the determinant of thecoefficient matrix to zero, weobtain a cubic equation for i
NO. Thesmallest valueofNO isthecritical bucklingload, which is(obtained usingFEMID)
El
N
cri
=20.7088U
whereas the"exact" solution is 20.187El]L
2
[seeReddy (1999b), pp. 155-166].
Next, consideratwo-elementmesh ofthe RIE. Theassembledequations are
6 -3h -6 -3h
11
28}
-311
311 /z28
2
2El -6 3/z 12 0
h
22
2h
28,
-3h
2
o
/lO/z3
o 0 -6 3h
o 0 -3/z 11
28
2
o
o
-6
3/z
6
311
0
0
-311
h
28
2
3h
h
28,
I 0 -1 0 0 0
o 0 0 0 0 0
-] 0 2 0 -I 0
o 0 0 0 0 0
o 0 -1 0 1 0
o 0 0 0 0 0
where K, =5/6,and
Thecondensedequationsare
=
QJ
Qi
Q ~ +Qt
Ql + ~
Q ~
Q ~
Note that because ofthezero diagonal element inthe stability matrix, weobtain alinear
algebraicequationforNO (v =0.25and /\=H
2
/ L2)
N
O
= 180+144/\ El
2.25+54/\+36/\2U
314 ANINTRODUCTION TOTHEFINITEELEMENT METHOD
which is almost four times the correct value! Amesh of eight linear RIBs yields 21.348 EllL
2
and four quadratic elements yields 20.267 Ell L
2
for LIH = 100. Clearly, the RIB has slow
convergence rate for buckling problems.
We close this section with a note that eigenvalue problems (for natural vibration and
buckling analysis) offramestructures canbeformulated using theideas discussed inSection
5.4. The transformed element equations areof the form
_ ={Qc}, _ = [QC}
[K
e]
=[Te]T[ke][ye], [Me] = [Te]T[MeUTil [G
e
] = [ye]Trce][ye]
(6.1.56)
6.2 TIME-DEPENDENT PROBLEMS
6.2.1 Introduction
In this section, we develop the finite element models of one-dimensional time-dependent
problems and describe time approximation schemes to convert ordinary differential equa-
tions intimetoalgebraic equations. We consider finite element models ofthetime-dependent
version of the differential equations studied inChapters 3 and 5. Theseinclude the second-
order(inspace) parabolic (i.e., first timederivative) andhyperbolic (i.e., second time deriva-
tive) equations and fourth-order hyperbolic equations arising inconnection with thebending
of beams [seeEqs. (6.1.2), (6.1.14), (6.1.36), (6.1.43a), and (6.1.43b)].
Finite element models of time-dependent problems canbe developed intwoalternative
ways: (a) coupled formulation in which the time t is treated as an additional coordinate
along with the spatial coordinate x and (b) decoupled formulation where time and spatial
variations are assumed to be separable. Thus, the approximation in the two formulations
takes the form
II
u(x, t) uf,(x, t) = LJ,j1frj(x, t) (coupled formulation) (6.2.1a)
j=l
II
U (x, t) uf, (x, t) = L uj (t)1frj (x) (decoupled formulation) (6.2.1b)
j=l
where 1frj (x, t) are time-space (two-dimensional) interpolation functions and ft j are the
nodal values that are independent ofx and t, where 1frj(x) are the usual one-dimensional
interpolation functions inspatial coordinate xonlyandthenodal values ui (t) arefunctions
of timet only. Derivation of theinterpolation functions intwocoordinates will bediscussed
inChapters 8and9 inconnection with thefinite element analysis of two-dimensional prob-
lems. Space-time coupled finite element formulations are not common, and they have not
been adequately studied. In this section, weconsider thespace-time decoupled formulation
only.
CHAPTER 6: EIGENvALUE ANDTIME-DEPENDENT PROBLEMS 315
Deflectionattime's
Figure 6.2.1 Deflectionsof acableat different times.
Thespace-time decoupled finite elementformulation oftime-dependent problems in-
volves two steps:
1. Spatialapproximation, where the solution 11 ofthe equation under consideration isap-
proximated by expressionsoftheform (6.2.1b), and thespatial finite element model of
theequation is developed using theprocedures of static or steady-state problems while
carrying alltime-dependent terms inthe formulation. This step results inasetofor-
dinary differential equations (i.e., asemidiscrete system ofequations) intime forthe
nodal variables uj(t) oftheelement. Equation (6.2.1b) represents thespatial approxi-
mation ofuforany time t. When thesolution isseparable into functions oftime only
andspaceonly,11(x, 1)=T(t)X (x), theapproxiIllation(6.2.1b) isclearlyjustified.Even
when thesolution isnot separable, (6.2.1b) can representagood approximation ofthe
actual solution provided asufficiently small timestep isused.
2. Temporalapproximation, wherethesystemof ordinarydifferentialequationsarefurther
approximated in time, often using finite difference formulae forthetime derivatives.
This step allows conversion of thesystem of ordinary differential equations into a set of
algebraicequationsamong uj attimet
s
+I = (s +1)txt,where ~ t isthetimeincrement
,
and sisanonnegativeinteger.
i,
I
I:
All timeapproximation schemesseektofind 11 j attimets+1 using theknown valuesof
.'
~ i
.,
Uj from previous times:
"
compute {UJs+I using { u } ~ '{U}s-I, ...
"I .
Thus, atthe end ofthetwo-stage approximation, we have acontinuous spatial solution at
discreteintervalsoftime:
u(x, t
s
) ufr(x, t
s
)=L
11
uj(ts)Vlj(x) (s = 0,1, ...) (6.2.2)
j=1
Note that theapproximate solution (6.2.2) has thesame form asthatintheseparation-of-
variables technique used tosolve boundary value and initial value problems. Bytaking
nodal values tobefunctions oftime, weseethat the spatialpoints inan element take on
different valuesfordifferent times (seeFig. 6.2.1).
We study thedetails ofthetwo steps by consideringamodel differential equation that
containsboth second- and fourth-order spatial derivatives and first- and second-ordertime
,.
I
-I
I
316 AN INTRODUCTION TOTHEFINITEELEMENTMETHOD
derivatives:
(6.2.3a)
Theabove equation is subject toappropriate boundary and initial conditions. Theboundary
conditions areof theform
2u)
specify u(x, t) or _ a au + (b a
ax ax ax
2
2u
au a
specify or b- (6.2.3b)
ax ax
2
at x = 0, L, and the initial conditions involve specifying
(6.2.3c)
where it == au/at. Equation (6.2.3a) describes, for example, the following physical
problems:
(a) Heat transfer and fluid flow: C2 =and b =
(b) Transverse motion of a cable: a =T, Co =0, b =0, Cl =p, C2 =
(c) The longitudinal motion of a rod: a = EA, b = 0; if damping is not considered,
Cl =0, C2 = pA
(d) The transverse motion of an Euler-Bernoulli beam: a =0, b =EI, Co =k, c\ =0,
c2=pA
We will consider these special cases through examples.
6.2.2 Semidiscrete Finite Element Models
Thesernidiscrete formulation involves approximation of the spatial variation of the depen-
dent variable. The formulation follows essentially the same steps as described in Section
3.2. Thefirst step involves theconstruction of theweak form of the equation over a typical
element In thesecond step, we develop thefinite element model byseeking approximation
of the form in (6.2.lb).
Following the three-step procedure of constructing the weak form of a differential
equation, we can develop theweak form of (6.2.3a) over anelement Integration byparts is
used once onthefirst term and twice onthesecond term todistribute thespatial derivatives
equally between theweight function wand thedependent variable u:
au) a ( a
2
u)] aIV ( a
2
u)]Xb
+ W -a- +- b- +- -b-
[ [(
ax ax ax
2
ax ax
2
x"
CHAPTER 6: EIGENVALUE AND TIME-DEPENDENT PROBLEMS 317
i
"
2w 2 2u Xb
( aw au a a u au a )
= a--+b-
2
+cowU+CjW-+C2W -
2
-wf dx ,i
2-
X ax ax ax ax at at
a
where
2u)]
A [
2
au a ( a
QI = -a-+- b-
ax ax ax X
n
(6.2.4b)
2u)]
A [au a ( a
Q3 = - -a-+- b-
2
'
ax ax ax
XI,
Next, we assume that u is interpolated by an expression of the form (6.2.1b). Equa-
tion (6.2.lb) implies that, at any arbitrarily fixed timet > 0, the function u can be approxi-
mated by a linearcombination of the 1/JJ and u)(1), with u)(1) being the value of u at time
t at the jth node of theelement Qe. In otherwords, the time and spatial variations of u are
separable. This assumption is not valid, in general, because it may not be possible to write
. the solution u(x, t) as the product of a function of timeonly and a function of space only.
However, with sufficiently small time steps, it is possible to obtain accurate solutions to
even those problems for which the solution is not separable in time and space. The finite
element solution that we obtain at theendof the analysis is continuous in space but not in
time. We only obtain the finite element solution in the form
n n
u(x, ts) = L uj(1s)1/Jj(x) = L(uj)ev'j(x) (s = 1, 2, ...) (6.2.5)
J=:I J=:I
where (uiYis the value of u(x, t) at timet = t
s
and node j of theelement Qe.
Substituting w =1/Ji (x) (to obtain the i th equation of the system) and (6.2.1 b) into
(6.2.4a), we obtain
2
11 [
I 2 I
d
.uJ 2
d
UJ
]
=" (K .. +K..)u).+M..-+M..-.-. -F (6.2.6)
I) I) I) dt I) dt2
"
Jn matrix form, we have
[K]{u} +[M1J{tl +[M
2
]{ii } ={F} (6.2.7a)
318 AN INTRODUCTION TOTHE FINITEELEMENTMETHOD
where
(6.2.7b)
j
Xb
= Co 1/J(t/Jj dx,
X"
l
(6.2.7c)
M0 =
x b
C21/Ji1/Jj dx,
l
X(j
Xb Xb 2 2
2 d 1/Jid 1/J j
Kij = b----2 dx,
2
F, = jx" 1/Jddx+eli
x dx dx
Q
Equation (6.2.7a) is a hyperbolic equation, and it contains the parabolic equation as a
special case (set [M
2
]= [0]). The time approximation of (6.2.7a) for these two cases will
be considered separately. We begin with the parabolic equation.
6.2.3 Parabolic Equations
Time Approximation
Thetimeapproximation isdiscussed with thehelp ofasingle first-order differential equation.
Suppose that we wish to determine u(t) for t> 0 such that u(t) satisfies
du
a- +bu= f(t), 0 < t < T and u(O) =Uo (6.2.8)
dt
where a#- 0, b,andUo areconstants, and f is a function of timet.Theexact solution of the
problem consists of twoparts: thehomogeneous andparticular solutions. Thehomogeneous
solution is
The particular solution is given by
uP(t)= (1
1
e
kr
f(r)dr)
The complete solution is given by
(6.2.9)
In thefinite difference solution of (6.2.8), wereplace thederivatives with theirfinite dif-
ference approximation. Themost commonly used scheme for solving (6.2.8) is thea-family
ofapproximation in which a weighted average of the time derivatives at two consecutive
time steps is approximated by linear interpolation of the values of the variable at the two
steps (see Fig. 6.2.2):
. . . U.l+1 - U.l
(1 - a)u.l+aU.l+l = for 0 :::: a:::: I (6.2.lOa)
L".t.l+l
where Us denotes the value of u(t) at time t=t.l = L:=l!1ti, and L".t
s
=t
s
- t.l-lis the sth
time step. If the total time [0, T]is divided into equal time steps, then t
s
= s L".t. Equation
(6.2.1 Oa) can be expressed as
U.l+l = Us + L".t us+a
u
s+a=(1-a)us+aus+1
forO::::a::::1 (6.2. lOb)
CIIAPTER 6: EIGENVALUE AND TIME-DEPENDENT PROIlLEMS 319
,i "" us+1 - us, forward difference
s /0,./
s
/
{
Us "" Us backward difference
, s-I
u(l) lis
s=O s=2 s-l s 05+,1
1+:-+1 '
s=l
/o,./s
Figure 6.2.2 Approximation of thederivative of a function.
When a =0, Eq. (6.2.10a) gives
which is nothing but the slope of the function u(t) at ti me I = Is based on the values of the
function at time Is and 1
st
I. Since the value of the function from a step in front is used, it is
termed [orward difference approximation. When a =1, we obtain
. Ustl - Us . Us-Us-l
Ustl =
---* Us = ---


which is termed, for obvious reason, the backward difference approximation.
Returning to Eq. (6.2.8), we note that it is valid for all times I > O. In particular, itis
valid at times 1 =Is and I =I
st
l . Hence, from Eq. (6.2.8) we have
. I . I l
Us = UI - bus) I Ustl =- (fstl - 7Ustl)
a a
Substituting the above expressions into Eq. (6.2. lOa), we arrive at
, (Ustl - us)
(l-a)(fs-bus)+a(.fltl-bUstl)=a A
Dlstl
Solving for Us+l, we obtain
[a +a L\lstl b]U
s+l
= 10 - (l - a) L\ls+1 b]Us +Lilstl [a fltl +(I - a).fl]
(6.2.1Ia)
or
a - (1- a) L\tstl b [a fstl +(l - a)fs]
Ustl = Us + Litstl (6.2.11b)
a+aL\tstlb
I'
Thus, Eq. (6.2.l1) can beused repeatedly to march in time and obtain the solution at times
t = Istl' I
st
2, ... , IN, where Ntime is the number of time steps required to reach the final
:I
. 320 ANINTRODUCTION TOTHEFINITEELEMENT METHOD
time T (or thesolution reaches a steady state). At the very beginning, i.e., S =0, thesolution
U1 is calculated using the initial value Uo:
a - (l - a) 8.1I b [afI +(l - a) h]
U\= uo+8.11 (6.2.12)
a + a 8.11 b a + a 8.1J b
We may alsodevelop a time approximation scheme using thefinite element method. To
this end, weconsider the problem in (6.2.8). We wish to determine Us+I in terms of Us' The
weighted-integral form of (6.2.8) over the time interval ; l
H
I ) is
0= t' V(tJ(a -v bu - f)dt
(6.2.13)
where v is the weight function. Assuming a solution of the form
II
LUj1f;j(t)
j==1
where 1f;j (t) areinterpolation functions of order(n - 1).TheGalerkin finite element model
is obtained by substituting the above approximation for U and v =1f;i into (6.2.13). We
obtain
[AHu) = (F) (6.2.14)
where
1S 1
I
ls 1
l
+ ( d1f;j )' +
Aij = Is 1f;i(l) ad! +b1f;j dt, F, = I, 1f;i(t)I(l) dt
(6.2.15)
Equation (6.2.14) is valid with thetime interval v.. i.; J), and it represents a relationship be-
tween thevalues UI, Uz, ... , UII, which arethevalues ofu at time 1.
1
, l
S
+ ,:1 I/ (II- I) , l
S
+ Z,:1 I/ (II- 1) ,
... , l
s
+l , respectively. This would yield a multistep approximation scheme.
To obtain a single-step approximation scheme, i.e., write UHI in terms of Us only, we
assume linearapproximation (i.e., n = 2)
where 1f;1 (t) = and1f;z(l) = .In thesamefashion, 1(1) canbe represented interms
of its values at 1,\' and t.; J:
For this choice of approximation, Eq. (6.2.15) becomes
(
a [-I I] b8.1 [2 1]) IUs I 8.1 I Is I
(6.2.16)
2 -1 I +6 1 2 Us+I = 6 2!1+1
Assuming that Us is known, we solve for UHI from the second equation in (6.2.16)
2b8.1) (b8.1) (Is 2
I s
+
l
)
(
a + -3- Us+I = a - -3- Us + 8.1 :3 + -3-
(6.2.17) ,
Comparing Eq, (6.2.17) with Eq. (6.2.11a), we find that the Galerkin scheme is a special
\
I,
,I',
case of thea-family of approximation, with a = 2/3.
1\
CHAPTER 6: EIGENVALUE AND TIME-DEPENDENT PROBLEMS 321
Stable and Conditionally Stable Schemes
Equation (6.2.11 b) can be written in the form
a - (l - a) i11.1+ 1 b
Us+1 =A(u
s
) +Fr.s+l, A =
a +a i11s+1b
F i1 [a fr+1 +(1 - a)fs] (6218)
<s,s+J = ts+1 + A I ..
a a u1
s+1
')
Theoperator A is known astheamplification. operator. SinceUs is anapproximate solution,
the error E, =ua(ts) - Us at time t
s
(where u., is the exact solution) will influence the
solution at ls+l. The error will grow (i.e., E, will be amplified) as we march in time if
the magnitude of the operator is greater than 1, IAI > 1. When the error grows without
bound, the computational scheme (6.2.1 1b) becomes unstable (i.e., solution U
s
+l becomes
unbounded withtime). Therefore, in orderfor the scheme to be stable, it is necessary that
IAI:s 1:
a - (1- a) i11s+1b I <I
(6.2.19)
a+ai1t
s
+l b -
I
Theabove equation places a restriction onthe magnitude of thetimestep for certain values
of a. When theerrorremains bounded for any time step[i.e., condition (6.2.19) is trivially
satisfied for any value of M], thescheme is stable. If theerrorremains bounded onlywhen
the time stepremains below certain value [inorder to satisfy (6.2.19)], the scheme is said
to be conditionally stable.
Fordifferent values of a, thetimeapproximation scheme in (6.2.11b) yields a different
scheme. The following well-known time-approximation schemes along with theirorderof
accuracy and stability should be noted:
0, the forward difference (or Euler) scheme (conditionally
stable); order of accuracy = O(i1t)
a=
~ the Crank-Nicolson scheme (stable); O(i1t)2
~ the Galerkin method (stable); O(i1t)2
1, the backward difference scheme (stable); O(i1t) (6.2.20)
Of these, the Crank-Nicolson method is themost commonly used scheme.
Fully Discretized Finite Element Equations
We now have thetools necessary toconvert thesetof ordinary differential equations (6.2.7a)
to a set of algebraic equations in much the same way we converted a single differential
equation (6.2.8) to an algebraic equation (6.2.1 1b). Here, we work with matrix equation
[i.e., [M
I
] = [M] and 1M
2
] = [0] in Eq. (6.2.7a)]
[M]!u} +[K]!u} = IF} (6.2.2Ia)
subject to the initial conditions
(u10 =(uo}
(6.2.21b)
322 ANINTRODUCTION TOTHEFINITEELEMENTMETHOD
where {u}o denotes the vector of nodal values of u at time l = 0, whereas {uo} denotes the
column of nodal values ujO. As applied to a vector of ti me derivatives of the nodal values,
thea-family of approximation reads as
6l
s+1
[(1 - a){ll L+a{u }s+Il = {Uls+1 - {uL for 0::: a::: I (6.2.22a)
or
{U}s+1 = {u}s +6l{U}s+a
(6.2.22b)
{U}S+a = (l - a){u}s +a!u}s+1
for 0 ::: a::: 1. Equation (6.2.22a) can be used to reduce the ordinary differential equations
(6.2.21 a) to algebraic equations among theUj at time t.; I. Since (6.2.2] a) is valid for any
l > 0, we can write it for times l =t, and l =ls+l:
[M]{uL +[KJs{uL = {F}s (6.2.23a)
[M]{uL'+J +[K]s+duL+I ={FL+I
(6.2.23b)
where it is assumed that the matrix [M] is independent of time. Premultiplying both sides
of (6.2.22a) with [M] we obtain
6l
s+la[M]{u
L+l +6l
s+1
(l - a)[M]{itL = [M]({uL+1 - {u}s)
Substituting for [M]{u L+I and [M]{u L from (6.2.23a) and (6.2.23b), respectively, we
arrive at
Solving for the vector !U}s+l, we obtain
[KL+l!U}s+l = [KL{uL +{F}s,s+l
(6.2.24a)
where
[kls+
l
= [M] +al [K]s+I, [Rt = [M] - a [ l ~
{F}s,S+1 =6l
s+1
[a{FL+l +(I - a){Fls] , (6.2.24b)
al =a6l
s+l,
a2=(l-a)6l
s+1
Note that, in deriving (6.2.24a) and (6.2.24b), it has been assumed that [M] is independent
of time and that thetime step is nonuniform.
Equations (6.2.24a) and (6.2.24b) arevalid for a typical finite element whose sernidis-
cretized equations are of the form (6.2.21a). In other words, Eqs. (6.2.24a) and (6.2.24b)
hold for any problem, independent of thedimension and method of spatial approximation,
as long as theendresult isEq. (6.2.21a). Theassembly, imposition of boundary conditions,
and solution of the assembled equations are the same as described before for steady-state
problems. Calculation of [k] and {F} at time l = 0 requires knowledge of the initial condi-
tions !u}o and thetime variation of {F}. Note that fora =0(theforward difference scheme),
we obtain [k] = [M]. If matrix [M] is diagonal, (6.2.24a) becomes
s+1
= u --
1 (/l
"k!u
S
+Fs,s+l
)
(no sum on i)
(6.2.25)
1M" L IJ J I '
(II) j=1
.!
CIl/\JYrER 6: EIGENVALUE AND TIME-DEPENDENT PROBLEMS 323
Thus, noinversion ofthecoefficient matrix isrequired insolving for {uy+1 .Such ascheme is
called explicit. Ascheme issaid tobeimplicit when it is not explicit (i.e., animplicit scheme
requires the inversion of a coefficient matrix). Thus, explicit schemes arecomputationally
less expensive compared to implicit schemes; implicit schemes are more accurate and have
larger critical time steps. Inreal-world problems, thecost of computation precludes theuse
of implicit schemes.
Inconventional finite element formulations, I.M] isseldom diagonal. Therefore, explicit
schemes in finite element analysis can exist only if the time-approximation scheme is such
that [k] =[M] and [M] is diagonal. Thematrix [M] computed according tothedefinition
(6.2.7c) iscalled the consistent (mass) matrix, and itis notdiagonal unless 1/Ii are orthogonal
functions over theelement domain. There are several ways to diagonalize mass matrices;
these will be discussed in Section 6.2.5.
Consistency, Accuracy, 'and Stability
Since (6.2.22a) represents an approximation, which is used to derive (6.2.240), error is
introduced into the solution {u }s+ 1 at each time step. In addition to the truncation error
introduced in approximating thederivative, round-off errors can be introduced because of
the f nite arithmetic used in the computations. Since the solution at time t
s
+I depends on
the solution at time t., the error can grow with time. As discussed earlier, ifthe error is
bounded, the solution scheme is said to be stable. The numerical scheme (6.2.240) is said
to be consistent with thecontinuous problem (6.2.7a) if the round-off and truncation errors
go to zero as f..t ---+ O. Accuracy of a numerical scheme is a measure of the closeness be-
tween the approximate solution and the exact solution whereas stability of a solution is a
measure of the boundedness of the approximate solution with time. As wemight expect,
the size of the time step can influence both accuracy and stability. When we construct an
approximate solution, we like it to converge to the true solution when the number of el-
ements or the degree of approximation is increased and the time step f..t is decreased. A
time-approximation scheme is said tobeconvergent if,for fixed t
s
, thenumerical value {u}s
converges to its true value {u(t.()} as f..t ---+ O. Accuracy is measured in terms of therateat
which the approximate solution converges. If a numerical scheme is stable and consistent,
it is also convergent.
For all numerical schemes in which a ~ the a-family of approximations is
stable only if the time step satisfies the following (stability) condition [follows from
Eq. (6.2.19)]:
, 2
f..t < f..tcri == --- (6.2.26)
(I - 2a)A
where Ais the largest eigenvalue associated with the problem and f..tcri is (see below)
(1M] - A[K]) {u} = {Q} (6.2.27)
Notethatthesame mesh as that used for the transient analysis must beused tocalculate the
eigenvalues.
324 AN INTIWDUCTIONTOTHE FINITEELEMENTMETHOD
6.2.4 Hyperbolic Equations
Time Approximation
Consider matrix equations of theform [set [M
I
] =[C) and [M
2
] = [M] in Eq. (6.2.7a)]
[K](u} +[C](U} +[M](u} = (F} (6.2.28a)
subjected to initial conditions
(u(O)} = (uo}, (u(O)} = (va} (6.2.28b)
Such equations arise in structural dynamics, where [M] denotes mass matrix, [C) the
damping matrix, and [K] thestiffness matrix. Thedamping matrix [C) is often taken to be
a linear combination of the mass and stiffness matrices, [C) = f3J [M] +f32[KJ, where f31
and f32 aredetermined from physical experiments. Inthepresent study, wewill notconsider
damping (i.e., [C) =(0)) inthenumerical examples, although thetheoretical developments
will account for it. Transient analysis of both bars and beams lead to equations of the type
given in (6.2.28a) and (6.2.28b). The mass and stiffness matrices for these problems are
discussed below.
Axial Motion of Bars
The equation of motion is given by Eq. (6.1.14). The semidiscretization results in Eq.
(6.2.28a), with [Me] and [XC] given by Eq. (6.1.27b) and [C) = [0).
Transverse Motion of Euler-Bernoulli Beams
Theequation of motion is given byEq. (6.1.36). Thesemidiscretized equation is the same
as Eq. (6.2.28a), with [Me] and [K
C
] given by Eq. (6.1.41b) and [C) =[0).
Transverse Motion of Timoshenko Beams
Theequations of motion aregiven byEqs. (6.1.43a) and (6.1.43b). The semidiscretization
results in Eq. (6.2.28a), with [Me] and [K
e]
given by Eq. (6.1.48a). Note that [Me] and
[K
e
] are thematrices given inEq. (6.1.47).
There are several numerical methods available to approximate the second-order time
derivatives and convert differential equations to algebraic equations. Among these, the
Newmark family of time-approximation schemes is widely used in structural dynamics
[Newmark (1959)]. Other methods, such as theWilson method [Bathe andWilson (1973)]
and theHoubolt method [Houbolt (1950)), can also be used to develop thealgebraic equa-
tions from thesecond-order differential equations (6.2.28a). Here we consider theNewmark
family of approximation schemes.
Newmark's Scheme
Inthe Newmark method, thefunction and itsfirst time derivative areapproximated according
to
1
(Uls+l = (uls + L1t(u}s + 2:(L1t)2(u}s+y (6.2.29a)
(U}s+1 = (uls +(u}s+a
L1 t (6.2.29b)
t
I
CHAPTER6 EIGENVALUEANDTIME-DEPENDENT 325
where
{idsH= (1- e){ids +elii }s+l
(6.2.29c)
anda andy (= 2{3) areparametersthatdeterminethestabilityandaccuracyof thescheme.
Equations (6.2.290) and(6.2.29b) canbeviewed asTaylor's series expansions ofu and u.
Thefollowing schemes are special cases of(6.2.290) and(6.2.29b):
a = y =2{3 =! Constant-averageaccelerationmethod (stable)
a = y =2{3 = Linearacceleration method(conditionallystable)
a = y =2{3 =0 Central differencemethod (conditionally stable) (6.2.30)
a = y =2{3 = Galerkinmethod (stable)
a = y =2{3 =2 Backward differencemethod (stable)
Forallschemesinwhich y < a and a ::: thestabilityrequirementis
(6.2.31)
where(})max isthemaximum natural frequency ofthesystem(6.2.28a) without [C]:
(6.2.32)
Fully Discretized Finite Element Equations
Eliminating (it}s+l from Eqs. (6.2.29a) and(6.2.29b) andwriting theresult for(it}s+l, we
obtain
(6.2.33)
where
(6.2.34)
a6 ={3:t' a7 = - 1, as = - 1) 6t
Now premultiplying Eq. (6.2.29a) with [M]s+! andsubstituting for[M]s+dii}s+l from
Eq.(6.2.28a), weobtain
([M]s+1 +{3(6t)2[K]s+l) IU}s+l =[M]s+1 {b}s +{3(6t)2IF}s+!
-{3(6t)2[C]s+llu}s+1
(6.2.35)
where
1
IbIs = lu}s +6tlu}s +2: (1 - y) (6t)2Iii}s
(6.2.36)
Now, multiplying throughoutwith 03 =1/[{3(6tf], wearriveat
(a3[M]s+1 +[K]s+l) IU}s+1 =a3[Ml,+db}s +{FIs+l - [C]s+dU}s+l
(6.2.37)
326 ANINTRODUCTION TOTHEFINITEELEMENTMETHOD
UsingEq. (6.2.33)for[UL+l inEq. (6.2.37) and collecting terms, weobtain thefinal result
(6.2.38)
where
[KL'+1 =[K]s+1 +a3[M]s+1 +a6[C]s+J
[FL,s+1 =[F}s+l +[M]s+1 [A}s +[C]s+1[B}s
[A}s =a3[b}s =a3[uls +a4[ul
s
+as[il}s
(6.2.39)
{Bls =a6{u}s +a7[u}s +as{il}s
and
(6.2.40)
Note that thecalculation of[K] and [F} requires knowledge oftheinitial conditions
{u}o, [u}o, and [il}o. Inpractice, wedonot know [il}o. Asanapproximation, itcanbe
calculatedfrom (6.2.28a) (we often assume that theapplied force iszero att = 0):
[U}o = [Mr
l
([F}o - [K][ulo - [C][ulo) (6.2.41)
At the endofeach time step, the new velocity vector [itls+1 and acceleration vector {il}S+l
arecomputedusing [from Eqs. (6.2.29a) and (6.2.29b)]
[il}s+J =a3({u}s+1 - {u}s) - a4[lds - Qs[il}s
{itls+l ={U}s +a2{iils +al [U}s+l
(6.2.42)
al =af..t, a: = (l - a)f..t
Theremaining procedurestays the same asinstatic problems.
Thefully discretized model in(6.2.38) isbased ontheassumption that y #- O. Ob-
viously, forcentered difference scheme (y =0),theformulation must bemodified (see
Problem6.26).
6.2.5 MassLumping
Recallfromthetimeapproximationof parabolicequationsthatuseof theforwarddifference
scheme (i.e., a =0)results inthe following time marching scheme [see (6.2.24a) and
(6.2.24b)]:
(6.2.43)
Themass matrix [Me] derived from the weighted-integral formulations ofthegoverning
equation iscalled theconsistent mass matrix, and itissymmetric positive-definite and
nondiagonal. Solution oftheglobal equations associated with (6.2.43) requires inversion
of theassembledmass matrix.If themass matrixis diagonal,then theassembledequations
canbesolveddirectly (i.e., withoutinvertingamatrix)
(UI )Hl = Mi/ [ Mil CUd, t1t KIJ (UJ)'+ t1t(Fd'] (6.2.44)
CHAPTER 6: EIGENVALUE ANDTIME-DEPENDENT PROBLEMS 327
and thus saving computational time. The explicit nature of (6.2.44) has motivated analysts
,
to find rational ways of diagonalizing the mass matrix.
There areseveral ways of constructing diagonal mass matrices, also known as lumped
mass matrices. The row-sum and proportional lumping techniques arediscussed here.
Row-Sum Lumping
The sum of the elements of each row of the consistent mass matrix is used as the diagonal
element:
MI;
(6.2.45)
= t, f pV/Nj dx = f PV/;' d.x
where the property VI] = I of theinterpolation functions isused. When p is constant,
(6.2.45) gives
l'M
e]
= pile [J
(linear element)
. 2

L 0
I [1
0
[Meh = p;e 4 (quadratic element) (6.2.46)

0
Compare these lumped mass matrices with the consistent mass matrices
}
[MeJc = P:e (linear element)
1
l'
[MeJc = pile [ (quadratic element) (6.2.47)
j'

30 -1 2 4 1
II
t
Here subscripts Land C refer to lumped and consistent mass matrices, respectively.

J
'I
." ,I
Proportional Lumping
Here the diagonal elements of the Jumped mass matrix arecomputed to beproportional to
the diagonal elements of the consistent mass matrix while conserving the total mass of the
element:
(6.2.48)
a= f pdx/ t f p1frNI dx
Forconstant p, proportional lumping gives thesame lumped mass matrices as those obtained
in the row-sum technique for the Lagrange linear and quadratic elements.
The use of a lumped mass matrix in transient analyses can save computational time in
two ways. First, for forward difference schemes, lumped mass matrices result in explicit
algebraic equations not requiring matrix inversions. Second, the critical time step required
for conditionally stable schemes is larger, and hence less computational time is required
when lumped mass matrices areused. To seethis, consider the stability criterion in (6.2.31)
for the case a = ! and f3 = O. For a one linear element model of a uniform barof stiffness
328 AN INTRODUCTION TOTHEFINITEELEMENTMETHOD
A and mass pA, fixed at the left end, the eigenvalue problem with aconsistent mass matrix
IS
(6.2.49)
Since U
J
=0 and Qi =0, we have
2 EA /PAh 3
W =- --=-
h 3 ph
2
Substituting this into the critical time step relation (6.2.31), wehave,
(6t
cr
ik = 2/ W
max
= h(4p/3) 1/2
If we use the lumped matrix, to is given by
W =(2/ p) 1/2/ h
and thecritical time step is
(6.2.50)
Thus, explicit schemes require larger time steps than implicit schemes.
6.2.6 Applications
Here weconsider two examples ofapplications of f nite element models ofone-dimensional
problems. Example 6.2.1 is taken from transient heat transfer (parabolic equation), and
Example 6.2.2 is taken from solid mechanics (hyperbolic equation).
Example 6.2.1 _
Consider thetransient heat conduction problem described bythe differential equation
au
-
at
-
a
2
u'
-
ax
2
= for < x < I
(6.2'sla)
with boundary conditions
au
u(O, t) =0, ax (1, t) =0 (6.2'slb)
and initial condition
u(x,0)=1.0 (6.2.51c)
where u isthenondimensionalized temperature. Theproblem athandisaspecial caseof(6.2.3a)
with a = 1, b=O, Co =0, c, = 1, C2 =0, and f =0 [or Eq. (6.1.2) with kA = 1, pcA = 1, and
q =0]. The finite element model is
[Me]{u} + [Ke]{u} == {Qe}
-------
CHAPTER 6:EIGENVALUEANDTIME-DEPENDENTPROBLEMS 329
,--- ---_.-- -.
where
l
Xb

K
e
= -' -}dx
IJ dx dx .
.to
For the choice of linearinterpolation functions, the semi discrete equations of atypical
elementare
he[2 1] }+ [ I-1]{U1 }={Q! }
6 1 2 U
2
he -1 1 U
2
Q
2
where heisthelength oftheclement. Useofthea-familyofapproximation (6.2.100) results
intheequation[see(6.2.240)]
where t'J.t isthetimestep.
First, considera one-elementmesh. We have(h== 1)
I t'J.t
-h+a- -h 1 - a- V [ 1 (1 - a)- 1 t'J.t ] -h- t'J.t
-h+(1-a)-
t'J.t ]
IV 1
1 t'J.t 1 M V - 1 t'J.t
1 !J.t
3 h 6 h 1 _ 3 h 6 h J
[ 2
V
2
-h-a- -li-va-r- -h+(I-.)- -h-(1-a)- I1",
6 h 3 h 6 h 3 h
HI 1
where Qi =a(QJ)5+1 +(l - a)(Q!)s'Theboundaryconditionsoftheproblemrequire
(V1)s == 0, :::0 foral1 s>0(i.e., t>0)
However, the initial condition(6.2.51c) requires
, Sincetheinitialcondition should beconsistentwith theboundary conditions, wetake(U1)0 == O.
Thenitfollows that(V
2
)o:::1.
Using the boundary conditions, wecanwrite for theone-element model (h:::J.0)
(6.2.52)
which canbesolvedrepeatedly forV2 atdifferent times,S :::0,1, ....
Repeated useof(6.2.52) cancausethetemporal approximation errortogrow withtime,
depending onthevalue ofa. Asnoted earlier, theforward difference scheme (a== 0) isa
conditionallystablescheme.Todeterminethe criticaltimestepfor theone-elementmesh, we
first calculatethemaximumeigenvalueoftheassociated system
The eigenvalueis (thereis onlyone)
J _I 2
-';{AhV
2+h
V
2==0.
or )..,=3/h =3
.I
330 ANINTRODUCTION TOTHEFJNITE ELEMENT METHOD
Hence, from Eq. (6.2.26) wehave D.teri =2/), =0.66667. Thus, in orderfor theforward dif-
ference scheme(6.2.52)
tobestable, thetime stepshould besmallerthan 6.t
e
ri = 0.6667; otherwise, thesolution will
beunstable,asshown inFig. 6.2.3.
Foratwo-element mesh, wehave (11 I = h
2
= h = 0.5);thecondensed equations ofthe
time-marchingschemearegiven by
2 6.1
-h+2a-
-h-a-IVj -h-2(1-a)-
1
I 6.1]
3 h
[2 6.t
-h+(1-a)-.
6.1 ]
IV j
6 h 2 _ 3 h
6 h 2
[
I 6.1 I 6.1 V
3
- I 6.1
I 6.t V
3
-li - a-
-h+a- ,1+1 -h+(1-a)- -11 - (1- a)-
6 It 3 h 6 h
3 It
,I
with (V
2
)0and(U3)0. Theforward differenceschemeyields
66.1
[4 1] {V2 } = [4-2J.L 1+J.L] {V
2
} ,
J.L = h2
6 1 2 V3 .<+1 6 1+J.L 2 - J.L V3 .\
-1.50 I
/.
I
-2.00
2linearelements(a=0.0, =0.065)
Time,t
Figure6.2.3 Stability of theforward difference (0: = 0.0) and Crank-Nicolson (a = 0.5)
schemes asapplied toaparabolicequation.
2linearelements
I
I
I
I
I
\
I
I
I
2.00
1.50
1.00
0;::-
n
0.50 -
'S"
i
0.00

oj
....
n>
a
-0.50
n>
f-
-1.00
6.0 5.0 4.0 3.0 2.0 1.0
-2.50
0.0
CHAPTER 6: EIGENVALUE ANDTlMEDEPENDENT PROBLEMS 331
Theassociatedeigenvalueproblem is
.Thecharacteristicequation is
-2 - - Ah
2
7A -IOA+l=O, A=-
6
whoseroots are AI =2.5967 and A2 =31.6891. Hence, thecritical time stepbecomes 6.t
c
ri =
2/31.6891 =0.0631. Atime step of6.t =0.065 results in anunstable solution, asshown in
Fig. 6.2.3.
For(unconditionally)stableschemes(a ~ ~ , there is norestrictiononthetime step(e.g.,
Crank-Nicolsonmethod withtwolinearelementsand6.t =0.065yieldverysmoothandstable
solution, asshown inFig. 6.2.3). However, toobtain asufficiently accuratesolution, the time
stepmustbetakenasafraction of6.t
c
ri ' Ofcourse,theaccuracyofthesolutionalsodependson
themeshsizeh. Asthisisdecreased(i.e.,the number ofelements isincreased), M
cri
decreases.
Plotsofu(l, t) versus timefor a =0anda =0.5with 6.t =0.05areshowninFig. 6.2.4.
Solutions predicted bymeshes ofone, two, orfour linear (L) orquadratic (Q) elements are
compared. Theconvergenceofthesolution with increasingnumberofelementsisclear. The
finiteelementsolutionsobtainedwithdifferentmethods, timesteps,and meshesarecompared
with theexactsolution inTable6.2.J.
t"t = 0.05
Jquadraticelement(a = 0.5)
,\ Jquadraticelement(a = 0.0)
o 2linearelements(a =0.5)
6. 2linearelements(a = 0.0)
,\
~
r
0.50
0.0 0.5 1.0 1.5 2.0 2.5
Time, t
Figure 6.2.4 Transient solution of aparabolic equationaccordingtolinearandquadraticfinite
elements.
332 AN INTRODUCTION TOTHEFINITE ELEMENT METHOD
Table 6.2.1 Acomparison of thefinite element solutions obtained using various timeapprox-
imation schemes and meshes with the analytical solution of a parabolic equation
(f:-,t = 0.05).
a=O a = 1 a=0.5
lL IL IL 2L 4L 4L IQ 2Q 4Q Exact
0.00 1.0000 1.0000 1.0000 1.0000 1.0000 1.0000 1.0000 1.0000 1.0000 1.0000
0.05 0.8500 0.8696 0.8605 1.0359 0.9951 0.9933 1.0870 0.9942 0.9928 0.9969
0.10 0.7225 0.7561 0.7404 0.9279 0.9588 0.9554 0.9819 0.9550 0.9549 0.9493
0.15 0.6141 0.6575 0.6371 0.8169 0.8639 0.8707 0.8693 0.8831 0.8725 0.8642
0.20 0.5220 0.5718 0.5482 0.7176 0.7557 0.7694 0.7679 0.7633 0.7731 0.7723
0.25 0.4437 0.4972 0.47 I7 0.6300 0.6759 0.6824 0.6780 0.6933 0.6855 0.6854
0.30 0.3771 0.4323 0.4059 0.5533 0.5906 0.6037 0.5987 0.6006 0.6070 0.6068
0.35 0.3206 0.3759 0.3492 0.4858 0.5250 0.5325 0.5286 0.5394 0.5358 0.5367
0.40 0.2725 0.3269 0.3005 0.4266 0.4608 0.47 J 3 0.4668 0.4710 0.474 [ 0.4745
0.45 0.2316 0.2843 0.2586 0.3746 0.4083 0.4158 0.4121 0.4201 0.4188 0.4194
0.50 0.1969 0.2472 0.2225 0.3289 0.3592 0.3676 0.3639 0.3687 0.3701 0.3708
0.55 0.1673 0.2149 0.1914 0.2888 0.3176 0.3247 0.3213 0.3275 0.3273 0.3277
0.60 0.1422 0.1869 0.1647 0.2536 0.2798 0.2868 0.2837 0.2883 0.2890 0.2897
0.65 0.1209 0.1625 0.1418 0.2227 0.2472 0.2535 0.2505 0.2556 0.2556 0.2561
0.70 0.1028 0.1413 0.1220 0.1955 0.2180 0.2238 0.2212 0.2253 0.2258 0.2264
0.75 0.0874 0.1229 0.1050 0.1717 0.1924 0.1979 0.1953 0.1995 0.1996 0.2001
0.80 0.0743 0.1069 0.0903 0.1508 0.1697 0.1747 0.1725 0.1761 0.1764 0.1769
0.85 0.0631 0.0929 0.0777 0.1324 0.1498 0.1544 0.1523 0.1557 0.1559 0.1563
0.90 0.0536 0.0808 0.0669 0.1162 0.1322 0.1363 0.1345 0.1375 0.1378 0.1382
0.95 0.0456 0.0703 0.0575 0.1020 0.1166 0.1205 0.1187 0.1216 0.1218 0.1222
1.00 0.0388 0.0611 0.0495 0.0896 0.1029 0.1065 0.1048 0.1074 0.1076 0.1080
Example 6.2.2
We wish to determine thetransverse motion of a beam clamped at both endsand subjected to
initial deflection using EBTand TBT. Thegoverning equations are
2 4
aw aw
- +- =0 for O< l; < I (6.2.53a)
at
2
ax
4
.
a,v O\V
w(O,t)=O, -(O,t)=O, w(l,t)=O, -(l,t)=O (6.2.53b)
ax ax
ow
w(x, 0) = sinrrx -rrx(1- x),
at(x, 0) =0 (6.2.53c)
Note that the initial deflection of the beam is consistent with the boundary conditions. The
initial slope is given by
(
aw)
8(x,0)=- - =-ncosrrx+rr(l-2x) (6.2.53d)
ax ~
Because of symmetry about x = 0.5 (center of the beam), weconsider only a half span of
thebeam forfinite element modeling. Herewe usethefirst halfof thebeam, 0 :s x :s 0.5, as the
CHAPTER6:EIGENVALUEANDTIME-DEPENDENTPROBLEMS 333
.---
,
I
j
computational domain. Theboundary condition atx =0.5iseCO.5, t)=-caw /ax)(o.5, t) =0.
Thesemidiscretized finite element model of atypical element is
156 -2211
e
54
-2217 e
4h
2
-J3h
e
e
420 54 -1317
e
156
[
13lz
e
-317
'e
2
22h,
We begin with a one-element mesh with theEuler-Bernoulli beamelement. Thesernidis-
cretizedmodel is .
-2211 54 -311 -6
[156
1z2
-22h 417
2
-1317
-3h
2
q2 + -311
211
2
317
U
2
13h ]rI[6 -3h]rI
420 54 -13h 156 2217 U
3
17
3
-6 311 6 311 U
3
1311 -311
2
2211 417
2
U
4
-u. 11
2
311 211
2
U
4
I
Theboundaryconditionsfortheone-elementmodel translateinto
VI =0, U
2
=0, U
4
=0, =0 forall t >0
Theinitial conditionscanbecomputedfrom C6.2.53c) and(6.2.53d)
l!1=0, l!2=0, l!3=0.2146, l!4=0} for(=O
VI=O, U
2
= O, U
3
=0 U
4
=0
Thecondensedequation of thetime marchingschemefor this casetakes theform
(K33+Q3M33)(U3)s+l = =.M33(a3U3 +a4U3 +QSU3)." s =0, 1, ...
whereQ3, Q4, andasaredefinedin(6.2.40).Thesecondderivative[hfortimet=0(i.e.,when
s=0)iscomputedfrom theequationofmotion:
.. K33(U.,, )O (12 ) 420
(U3)O=- . =- - x0.2146 -=-110.932
M
33
h
3
156h
Fory <!' wemust computethecriticaltimestep Litcri, which depends onthesquareof
maximum natural frequency ofthe beam [see Eq. (6.2.31)].Forthepresentmodel, ill
max
is
computedfrom theeigenvalueproblem
-3/z
e
211
2
e
311
c
11
2
e
Hence; thecritical timestepfor a:::: 0.5andy = (i.c., thelinearaccelerationscheme) is
Litcri :::: J12/illmax=0.15236
334 ANINTRODUCTIONTOTHEFINITEELEMENTMETHOD
Although there isnorestriction ontime integration schemes with CL =0.5and y >0.5,the
critical time step providesan estimate of the time step to beused toobtain transient solution.
I.
Figure6.2.5 shows plotsofw(0.5, t) versustimefor theschemewith CL =0.5and y = t.
Threedifferent timesteps, St=0.175,0.150,and0.05,areusedtoillustrate theaccuracy. For
/').t =0.175> /').tcriJ thesolution is unstable, whereas for/').t < /').fcri,itisstablebutinaccurate.
Theperiodof the solutionis given by
T=2rr/w=0.27635
For two- and four-element meshesof Euler-Bernoullielements,thecriticaltimestepsare
(detailsarenot presentedhere)
(Mcrih=0.00897, (/').f
cr
i)4== 0.00135
where thesubscripts refer tothenumber ofelements inthemesh. Thetransverse deflection
obtained with the oneandtwo Euler-Bernoulli elements (/').t =0.005) inhalfbeam fora
completeperiod (0, 0.28)areshownin Fig. 6.2.6.
Theproblemcan also beanalyzedusing theTimoshenko beamelement(RIE).In writing
the governingequations [see(6.1.43a) and (6.1.43b)] of theTBTasthey apply to thepresent
problem, we first identify thecoefficients OAK,I' EI, pA,and pI consistentwith those inthe
differentialequation (6.2.53a). WehaveEI= 1.0and pA= 1.0.ThenOAKs canbecomputed
as
E EI 12 5 4
OAKs = . BHK
s
= - - = -EI (6.2.54)
2(1 +I)) 2(1 +v) H2 6 H2
1.00
0.50
.-..
...
v)
0
'-'
0.00
d
.8 I
1) I
(\)
q:: -0.50
\
,
(\)
I
I 0 M=0.175}
i:::l
-1.00
\ 0 M=0.15 (y= 1/3)
a=O.5
, == 0.05
\
-1.50
\ 0 M=0.175 (y=0.5)
I
\
I
-2.00
0.0 0.5 1.0 1.5 2.0
Time,t
Figure 6.2.5 Centerdeflection w(O.5, t) versustimet foraclampedbeam.
CHAPTER 6: EIGENvALUE ANDTIME-DEPENDENT PROBLEMS 335
0.20
0.10
q
If)
0
l'
d'
.2
0.00
t)
Cl)
p;::
Cl)
0
-0.10
'D.I =0.005
- - Ielement
-0.20 2elements
0.0 0.\ 0.2
Time, f
Figure6.2.6Center transverse deflection versus time for aclamped beam subjected toan
initial transversedeflection (M = 0.005, CY. = 0.5,and y = 0.5).
where B isthewidth and H the heightof thebeam, and I = BH
3
, V = 0.25, and K, = are
used inarrivingatthelastexpression. Similarly,
1 3 1 ?
pI=p-BH=-pAH- (6.2.55)
12 12
Thus, thegoverningequations ofthe TBT for theproblem at handare
OJ 0.4 0.5
(6.2.56a)
(6.2.56b)
Thevalues ofw(O.5, I)asobtained using theEuler-Bernoulli and Timoshenko elements
(both elements include therotary inertia) forvarious numbers ofelements arepresented in
Table 6.2.2. Thetime stepistaken tobe6.1 =0.005,whichissmaller than thecriticaltimestep
of thetwo-element mesh of theEuler-Bernoullibeamelementwhen y = Plotsofw(0.5, t)
obtained with twoand four linear Timoshenko beam elements, two quadratic Timoshenko
beam elements forLIH =100 (since L= l.0, wetake H =0.01; forLIH =100 theshear
deformation effect isnegligible) along with thetwo-element solution oftheEuler-Bernoulli
beamelement areshown inFig. 6.2.7. Thetwo linearelement mesh ofTBTbeam elements
predictstransientresponsethatdifferssignificantlyfromtheEBTsolution,andtheTBTsolution
converges totheEBT solution asthenumber ofelements orthedegree ofapproximation is
increased. Should we useconditionally stableschemes, it canbeshown that theTimoshenko
beam element requires largerMcri than theEuler-Bernoulli beam element. This is because, as
LI H is decreased, theliJmax predicted bytheTBTissmaller than that predicted bytheEBT.
336 ANINTRODUCTIONTOTHEFINITEELEMENTMETHOD
...... ... . ....._.. - _.. -,-
Table6.2.2 Effect of mesh on thetransient response of abeam clamped at bothends(tlt=
0.005).
EBTelements TBTelements'
a=0.5, y= 1/3 a=0.5, y=0.5 a=0.5,y=0.5
Time t 2 2 4 2L 4L 2Q
0.00 0.2146 0.2146 0.2146 0.2146 0.2146 0.2146 0.2146 0.2146
0.01 0.2091 0.2098 0.2091 0.2098 0.2098 0.2124 0.2099 0.2116
0.02 0.1928 0.1951 0.1928 0.1951 0.1951 0.1938 0.1936 0.1951
0.03 0.1666 0.1696 0.1667 0.1696 0.1698 0.1550 0.1686 0.1690
0.04 . 0.1319 0.1346 0.1320 0.1348 0.1350 0.1145 0.1374 0.1427
0.05. 0.0904 0.0930 0.0905 0.0931 0.0935 0.0695 0.0980 0.1067
0.06 0.0442 0.0481 0.0443 0.0482 0.0483 0.0093 0.0073 0.0657
0.07 -0.0043 0.0014 ':"'0,0041 0.0014 0.0018 -0.0467 -0.0403 0.0234
0.08 -0.0525 -0:0462 -0.0523 -0.0459 -0.0917 . -0.0844 -0.0267
0.09 -0.0980 .-0.0926 -0.0978 -0.0923 -0.0916 -0.1422 .-0.1254 -0.0706
0.10 -0.1385 -0.1345 -0.1383 -0.1342 -0,1336 -0.1833 -0.1588 -0.1100
0.11 -0.1719 . -0.1685 -0.1717 -0.1685 -0.1682 -0,2010 ":"0.1875 -0,1461
0,12 -0.1964 . -0.1933 -0.1963 -0.1933 -0.1932 -0.2154 -0.2069 -0.1717
0.13 -0.2108 . -0,2088 -0.2108 -0,2088 -0,2087 -0.2205 -0.2063 -0.1969
0.14 -0.2144 -0,2153 -0.2144 -0.2150 -0.2148 -0.1986 -0.2138 -0.2110
0,]5 -0.2070 -0.2113 -0.2071 -0.2112 -0.211I -0.1696 -0.2134 -0.2146
2L=two linearelements;4L=four linearelements;2Q=two quadraticelements,
0.40
OJO
0.20
0.10
Q
vf
0
'-'
0.00
:s
I=f
0
-0.10
'P
o
Ill.
-0.20
/+=l
III
c
-OJO
)
TBT (2linearelements)
i
-0.40
TBT(4linearelements)
I
+ TBT (2quadratic elements)
-0.50
o BBT(2elements)
Time,t
Figure6.2.7 Transientresponse ofabeam clamped atbothends, according totheTBTand
EBT(EI=1,pA=1,If =0.01, M=0.005,a=0.5,and y =0.5).
. .. - .._._.. _-------------- -J
0.5 0.4 OJ 0.2 0.1
-0.60
0.0
CHAPTER6:EIGENVALUEANDTIME-DEPENDENTPROBLEMS 337
6.3 SUMMARY
Inthischapter, finite element formulations ofeigenvalue problems andtime-dependent
problems aredeveloped. One-dimensional, second- and fourth-order equations (beams)
have been discussed. Theeigenvalue problems studied include problems ofheattransfer
(and thelike), bars, andbeams. Inthecaseofbars andbeams, theeigenvalue problems
arise inconnectionwithnaturalvibrations andbuckling of columns. Except for thesolution
procedure, thefinite elementformulation ofeigenvalue problems isentirely analogous to
boundary valueproblems.
Finiteelementmodelsoftime-dependentproblemsdescribed byparabolicandhyper-
bolic equations have alsobeen presented. Atwo-step procedure toderive finite element
models from differential equations has been described. Inthefirst step, weseekspatial
approximations ofthedependentvariables oftheproblem aslinearcombinationsofnodal
values thatarefunctionsof time andinterpolation functionsthatarefunctionsof space. This
procedure isentirely analogous tothefinite element formulation presented forboundary
valueproblemsin Chapters 3-5.Theendresultof thisstepis a set ofordinarydifferential
equations (intime) among thenodal values. Inthesecond step, theordinary differential
equations arefurther approximated using thefinite difference approximation ofthetime
derivatives. Theresulting algebraic equations canbesolved forrepeatedly, marching in
time.Examplesusingbothtransientheat transferand beam bendingare presented.
PROBLEMS
Section6.1
Mostproblemsareformulative innature.Foreigenvalueproblems,weneed towrite thefinal charac-
teristicequationsfor theeigenvalues.
6.1Determine thefirst twoeigenvalues associated with theheattransfer problem whose governing
equationsandboundaryconditionsaregiven by (Fig. P6.1)
_.!!-(aau)+b
au
+cu=O forO<x<L
ax ax at
u(O) =0, (a ::+ f3
u
) L" =0
where a, b, c,andf3 areconstants. Use(a) twolinear finite elements and(b) onequadratic
elementinthedomain tosolvetheproblem.
c= fJP a = kA
u(O) = 0 ~ i t l
Fin'slateral surfaceaswellas
x, u
theendx = L areexposed to
ambienttemperature
~ L ------.I
FigureP6.1 FigureP6.2
338 ANINTRODUCTION TOTHEFINITEELEMENTMETHOD
6.2 Determine the first two longitudinal frequencies of a rod (with Young's modulus E, area of
cross section A, and length L) that is fixed at one end (say, at x = 0) and supported axially at
the other end (at x =L) by a linear elastic spring (with spring constant k), as shown in Fig.
P6.1:
2 2
au au
-EA-+pA-=O for Uc x c L
ax
2
at
2
u(O) =0, (EAdU+ku)1 =0
dx x=L
Use (a) two linear finite elements and (b) one quadratic element in the domain to solve the
problem. Answer: (a) The characteristic equation is n.
2
- (10 +4c))., + (I + 2c) =0, c =
kL/2EA, ).,=(ph
2/6E)u}.
6.3 Determine thesmallest natural frequency of a beam with clamped endsandof constant cross-
sectional areaA, moment of inertia I, andlength L. Use thesymmetry andtwo Euler-Bernoulli
beam elements in thehalf beam.
6.4 Resolve theabove problem with two RTEs inthe half-beam.
6.5 Consider a beam (of Young's modulus E, shear modulus G, area of cross section A, second
moment areaabout the axis of bending I, and length L) with its left end (x = 0) clamped and
its right end (x =L) supported vertically by a linearelastic spring (see Fig. P6.5). Determine
the fundamental natural frequency using (a) one Euler-Bernoulli beam element and (b) one
Timoshenko beam element (IlE) (use thesame mass matrix in both elements).
Figure P6.5
6.6 Consider a simply supported beam (ofYoung's modulus E, mass density p, areaofcross section
A, second moment of areaabout theaxis of bending I, andlength L) with an elastic support at
thecenter of the beam (seeFig. P6.6). ~ t n n n thefundamental natural frequency using the
minimum number of Euler-Bernoulli beam elements. Answer: The characteristic polynomial
is 455)"
2
- 2(129 + c))., +3 + 2c = O.
R
A,E
Pinned here
L
-
2
R
Figure P6.6
6.7 Determine the critical buckling load of a cantilever beam (A, I, L, E) using (a) one Euler-
Bernoulli beam element and(b) oneTimoshenko beam element.
CHAPTER6:EIGENVALUEANDTIME-DEPENDENTPROBLEMS 339
6.8 Thenatural vibration ofabeam under applied axial compressive load N isgoverned bythe
differential equation
d
4
d
2
w w
EI- +N
O
_ =AW
dx" dx?
whereAdenotes nondimensional frequency ofnatural vibrationandEI istheflexural stiffness
ofthebeam. (0) Determinethefundamental (i.e.,smallest)naturalfrequency (J) of acantilever
beam (i.e., fixed at oneendand free at theother end) of length L with axial compressiveload
No using onebeam element. (b) What isthebuckling load ofthebeam? You arerequired to
give thefinal characteristicequation ineach case.
6.9 Determine thefundamental natural frequency of thetruss shown inFig. P6.9 (you arerequired
onlytoformulate the problem).
2
10ft.
E= 30msi
A, =3in.
2
2EI
lOft.
A
2
= 4in.
2
A
E =30X 10
6
Ib/in.
2
, II=OJ
A = 10
2
in.
2
, 1= 10
2
in."
EJ CD
FigureP6.10
6.10 Determine thefundamental natural frequency of thetruss shown inFig. P6. J0(you arerequired
only toformulate theproblem).
6.11 Determine thefirst two longitudinal natural frequencies ofarod(A, E, L, 111), fixed atone
endand with anattached mass 1112 attheother. Usetwo linear elements. Hint: Note that the
boundary conditions for theproblem areu (0) =0and (EA aulax +m2 a
2ulat 2
) IX=L =o.
6.12 Theequation governing torsional vibration of acircular rod is
a
2
a
2

-GJ-+mJ-=O
ox
2
al
2
where is theangular displacement, J themoment of inertia, Gtheshear modulus, and III the
:i
density. Determine thefundamental torsional frequency of arod with disk(./1) attached at each
end. Usethesymmetryand (0) two linearelements and (b) onequadraticelement.
6.13 Theequations governing themotion of abeam according toTBTcan be reduced tothesingle
equation
FigureP6.9
where0
2
=EljmA and1/ =1I A.HereE istheYoung's modulus,Gistheshear modulus,III is
themassperunitlength,A istheareaof crosssection,and I isthemoment of inertia.Assuming
that(b
2
ml kG) < < I(i.e.,neglectthelast term inthegoverningequation),formulate thefinite
elementmodel ofthe(0) eigenvalueproblem for thedetermination of natural frequencies and
(b) fully discretized problem for thedetermination of thetransientresponse.
340 AN INTRODUCTION TOTHEFINITEELEMENTMETHOD
6.14 Use the finite element model of Problem 6.13 to determine the fundamental frequency of a
simply supported beam.
6.15 Findthecritical buckling load Peri bydetermining theeigenvalues of theequation
c,
d
4
w d
2
w
E1- +P
eri-
=0 for Il-cr < L
4 2
dx dx
w(O)=w(L)=O,
UseoneEuler-Bernoulli element in thehalf-beam. Answer: Peri = 9.9439E1/ L
2

Section 6.2
6.16 Consider the partial differential equation arising in connection with unsteady heat transfer in
an insulated rod:
au _ (a au) = f for< x< L
at ax ax, .
u(O, t) =0, u(x,O)=uo, [a au +f3(U-u
oo)+Q].1
=0
ax x=L
Following theprocedure outlined in Section 6.2, derive thesemidiscrete variational form, the
semidiscrete finite element model, andthefully discretized finite element equations fora typical
element.
'I
6.17 Using a two-element (linear) model and the semidiscrete finite element equations derived in
:
I Problem 6.16, determine the nodal temperatures as functions oftime for the case in which
r
a = 1, f =0, Uo = 1, and Q=0. Use the Laplace transform technique [see Reddy (2002)] to
solve theordinary differential equations in time.
6.18 Consider a uniform barof cross-sectional areaA, modulus of elasticity E, mass density m, and
length L. The axial displacement under the action of time-dependent axial forces is governed
bythe wave equation
Determine the transient response [i.e., find u(x, t)] of the bar when the end x = is fixed and
theendx = L issubjected toa force Po. Assume zero initial conditions. Useonelinear element
toapproximate the spatial variation of thesolution, andsolve theresulting ordinary differential
equation in time exactly to obtain
PoL x
U2(X, t) = --(l-cosor),
AE L
6.19 Resolve Problem 6.18 with a mesh of twolinear elements. UsetheLaplace transform method
to solve thetwo ordinary differential equations in time.
6.20 Solve Problem 6.18 when the right endis subjected to an axial force F
o
and supported by an
axial spring of stiffness k. Answer:
3F
o a ( kL)I/2
U2(t) = c(1 - cosf3t), c=---:- f = ~ 1 +-
mALf32 ' L EA
CHAPTER6:EIGENVALUEANDTIME-DEPENDENTPROBLEMS 341
6.21 Abaroflength L movingwith velocity va strikesaspringofstiffnessk. Determinethemotion
u(x, t) from theinstantwhen theend x =0 strikes thespring. Useonelinear element.
6.22 A uniform rodoflength L andmass m isfixed atx =0 and loaded withamass M atx =L.
Determine themotion u(x, t) ofthesystemwhen themass M issubjected toaforce Po. Use
onelinearelement.Answer:
PoL
a (3M )-1
uz(t) =c(l - cosAt), c=- A ~ -+m
AE' L AL
6.23Theflow ofliquid ina pipe, subjected toasurge-of-pressure wave (i.e., awater hammer),
experiences asurgepressure p, which is governed bytheequation
where m isthemass density andK thebulk modulus ofthefluid, D isthediameter andb
thethickness ofthepipe, andE isthemodulus ofelasticity ofthepipematerial. Determine
thepressure p(x, t) using onelinear finite element forthefol/owing boundary andinitial
conditions:
ap
p(O, t) =0, -(L,t)=O, p(x,O)=Po, jJ(x,O) =0
ax
6.24Consider theproblem of determining the temperature distribution of asolid cylinder, initially
atauniform temperature To andcooled inamedium ofzerotemperature (i.e., Too =0). The
governingequation of theproblemis
pc aT _ ~ (I' kaT) =0
at I' ar ar
Theboundary conditions are
Theinitial condition isT(r, t) =To. Determine thetemperaturedistribution 1'(1', t) using one
linearelement.TakeR =2.5cm,k =215W/(m .0C),f3 =525W/(mC), 1'0 =130C,p =2700
kg/m", andc=0.9kJ/(kgC). What is theheat loss at thesurface?Formulatetheproblem.
6.25Determine thenondimensional temperature0(I', t) intheregionbounded bytwolongcylindrical
surfacesofradii R
1
andR
z
. Thedimensionless heatconductionequation is
withboundaryandinitialconditions
ao
-(R1,t)=0, O(Rz,t)=I, 0(1',0)=0
ar
6.26 Show that (6.2.28a), (6.2.28b), (6.2.290), and (6.2.29b) canbereworked tomatch theform in
Eq. (6.2.38) as
anddefine [H] and(17).,+I.
342 ANINTRODUCTION TOTHEFINITE ELEMENTMETHOD
6.27 Auniformcantileverbeamoflength L, momentofinertia I, modulusofelasticity E, andmass
m beginstovibrate with initial displacement
andzero initial velocity. Find itsdisplacement atthefree endatany subsequent time. Use
oneEuler-Bernoulli beam element todetermine thesolution. Solve theresulting differential
equations intimeusing theLaplacetransform method.
6.28 ResolveProblem6.27 usingoneTimoshenko beamelement.
REFERENCESFORADDITIONALREADING
1. Argyris, J.H.andScharpf, O.w., "FiniteElements in TimeandSpace,"Aeronautical Journal of the Royal
Society, 73,1041-1044, 1969.
2. Bathe, K. J.,Finite Element Procedures in Engineering Analysis, Prentice-Hall,Englewood Cliffs, NJ,1982.
3. Bathe,K.J. andWilson,E. L., "StabilityandAccuracyAnalysis of DirectIntegration Methods,"lnternational
Journal of Earthquake Engineering andStructural Dynamics, 1, 283-291, 1973.
4. Belytschko, T., "AnOverview of Sernidiscretization andTimeIntegration Procedures," in Belytschko T. and
HughesT. J.R.(eds.),Computational MethodsforTransient Analysis, North-Holland, Amsterdam,pp.1-65,
1983.
5. Goudreau,G.L., andTaylor,R. L., "EvaluationofNumericalIntegrationMethodsinElastodynamics,"Journal
of Computer Methods in Applied Mechanics and Engineering, 2(1), 69-97,1973.
6. Hilber,H. M.,"AnalysisandDesignof NumericalIntegrationMethodsinStructuralDynamics,"EERCReport
no.77-29, EarthquakeEngineeringResearchCenter,Universityof California,Berkeley,CA,November 1976.
7. Houbolt, J.C;"ARecurrence Matrix Solution fortheDynamic Response ofElastic Aircraft," Journal of
Aeronautical Science, 17,540-550,1950.
8. Newmark, N.M.,"AMethod ofComputation forStructural Dynamics," Journal of Engineering Mechanics
Division, ASCE, 85,67-94,1959.
"
r 9. Nickell, R.E.,"OntheStabilityofApproximation Operators inProblemsofStructural Dynamics," Interna-
tional Journal of Solids andStructures, 7,301-319, 1971.
10. Reddy, J.N.,Theory and Analysis of Elastic Plates, TaylorandFrancis,Philadelphia, PA, 1999a.
II. Reddy, J. N., "OntheDynamic Behavior of theTimoshenko Beam FiniteElements,"Sadhana (Journal ofthe
Indian AcademyofSciences),24, Part 3,175-198,1999b.
12. Reddy, J. N., "On theDerivation of theSuperconvergent Timoshenko Beam FiniteElement," Int. 1.Comput.
Civil andStruct. Engng., 1(2),71-84, ~
13. Reddy,J.N.,Energy Principles andVariational Methods inApplied Mechanics, JohnWiley, NewYork,2002.
14. Wood,W. L., "Control of Crank-NicolsonNoise in theNumericalSolution oftheHeatConductionEquation,"
International Journal for Numerical Methods in Engineering, 11,1059-1065, 1977.