Anda di halaman 1dari 175

Approximating functions - scalar problems 91

7 and simpler parts and then, based on these results, it is established for the entire
region.
Choice of approximating The first route (a ) does not present an attractive general approach since the
region differs from proble m to prob lem. Therefore, this approach is geared towards
functions for the FE the solution of specific proble ms having in principle the same geometry. What we
are looking for is a method applicable to general geometries and this objective is
method - scalar problems achieved by the second alternative, route (b), described abo ve.
If we split the body, i.e. the region of interest, into smaller parts, i.e. finite elements,
for which we can establish an approximation for the unknown temperature T, then
we are able to analyze bodies with arbitrary geometries. As already touched upon
in Chapter I, it is a characteristic feature of the FE method that the approximation
of the unknown function is carried out ' elementwise'. Actually, the FE method owes
its name to this characteristic feature.
We have emphasized tha t the FE method is based on the weak formula tion of the In this chapter we shall treat the app roximation of a scalar function like the
problem. As a pro totype of such a weak formulation we may consider (6.46) for temperature over one-, two- and three-dimensional elements and we shall show how
two-dimensiona l heat flow: the approximation of the scalar function over the entire region of the body of interest
can be established based on this elementwise approximation.
(Vv)TtDVTdA = - f Vht d.5!' - f vq,td.5!' + f vQtdA (7.1) In orde r to achieve such an approximation we have to make a choice for the
fA .:4. ~ A
type of approximation. For instance, should trigonometric or exponential functions
In this equation, terms involving the unknown temperature T and the arbitrary be used in the approximation procedure? It turns out that the use of polynomials is
weight function v appear. Therefore, in order to achieve an F E formulation from especiallyadvantageous and convenient for establishing an elementwise approximation
(7.1), we need to make certain choices for the approximation of the temperature T of the unknown scalar, and in order to understand this importan t point, it is
as well as for the weight function v. appropriate to evaluate the criteria for convergence.
We can therefore summarize the basic steps in the FE formulation as follows:

Basic steps in the F E formulation 7.1 Gene ral requirements


I. Establish the strong formulat ion of the problem.
2. Obtain the weak form of the problem . 7.1.1 Convergence criteria
3. Choose approximations for the unknown function - in the present
case the unknown temperat ure T. The FE method pro vides an approximate solution to the problem at hand and it is
4. Choose the weight function v. obvious that the more elements we use (i.e. the more computer power we utilize) the
more accurate the approximate solution. In the limit, when the elements are infinitely
small, we require that our approximate solution is infinitely close to the exact
solution. This is the convergence requirement.
Considerable emphasis has already been given to steps 1 and 2. In the present chapter Let us consider the inhomogeneous temperature field over the two-dimensional
we shall treat the manner in which approximations are made for the unknown body shown in Figure 7.1. The boundary along AC is insulated and the temperature
function T, i.e. step 3. In the following chapter we shall finally make an evaluation is prescribed along the boundary ABOC. In Figure 7.1 (b), a very coarse element
of a suitable choice of the weight function v, i.e, step 4. After these discussions we mesh consisting of only two triangular elements is used, whereas many triangular
will eventually be in a position to present the FE method. elements are used in the fine element mesh shown in Figure 7.l(c).
In ord er to establish an approximation for the unknown scalar - the temperature Within each element the temperature varies drastically in the coarse mesh and
T - we may in principle adopt two basically different routes : (a) the unknown
less dramatically in the fine mesh. In the limit, for infinitely small elements, we then
function is approximated directly over the entire region of interest; (b ) the region is
have within each element either a consta nt temperature gradient or a constant
divided into smaller parts and the approximation is first carried out over these smaller
temperature. If the approximation of the temperature over each element is to
90
92 Introducti on to the fin i te el ement me thod
Ap p r ox ima ting functions - scala r problems 93

aJ T ~ ' OO' C B
ArL.L....LL.1--L..l....J-+~ T= 100'C

A ""--- - _...1
Insulated I - - r - Lineor variat ion
Figure 7.2 Two neighbouring elements

along the dotted line. We a re then led to the so-called compatibility or conf orming
T~ 50' C requirement :

b) cJ
Compatibility or conforming requirement
The approximation of the tempe rature over element boundaries must
be continuous.

The convergence requ irement, stating that for infinitely small elements the
Figure 7.1 ( a) Two-dimensional heat problem ; (b) coarse element mesh ; (c) fine element mesh appro ximatio n approaches the exact temp erature field, is clearly fulfilled if both the
completeness and compatibility requirements are fulfilled, i.e. we obtain the important
statement
approach the exact temperature for decrea sing element size, it follows that the
approximation must fulfil the so-called completeness requirements : convergence criterion = completeness + compatibility requirements
Here only qualitative argu ments have been presented for the fulfilment of
Comp leteness convergence, but a more stringent evalu ati on is presented by Strang and Fix ( 1973),
and 10 Chapter 9 we shall rederiv e the compatibility requirement using a different
The approximation must be able to rep resent an arbitrary constant
approach. However, it turns out that it may even be allowable to relax the
temperature gradient.
compa tibility (i.e. the conforming) requirement and still be able to fulfil the
The approximation must be able to repre sent an arbitrary constant
convergence criterion. Indeed, even if a discontinuity exists across an element
temperature .
boundary, this discontinuity ma y approach zero for infinitely small elemen ts. Fini te
elements fulfilling the compatibility (i.e. the confo rming ) requ irement are called
conf orming elements , otherwise they are termed non-conforming elements. Most finite
Th e completeness requi rements therefore imply that the approximation of the
elemen ts are conforming, but even non-conforming elements are used, especially in
tem perature at least must include terms corresponding to an arbitrary linear
plate analysis as discussed in Chapter 18.
polynomial. Th at is, in three dimensions we must have the following approximation:
As indicated above, the fulfilmen t of both completeness and compatibility is
sufficient for convergence. However, whereas the requirement of compatibility may
T'" "" + "', x + "', y+ IX,' + possibly other terms (7.2)
be relaxed, the fulfilment of completeness is clearly a necessary condition for
convergence that must be satisfied.
Consider now two neighbouring elements as shown in Figure 7.2, where
As the completeness requirement must be met, one element is always able to
triangular elements are used for convenience. The temperature va riatio n along the
rep roduce exactly an arbitrary linear temperature variation. Clearly, if a n arbitrary
common element boundary AB is considered. In each element an approximation of
assemblage of conforming elements is considered, these elements a re also able to
the temperature is made a nd if the temperature variation in element 1 along the
reproduce exactly an arbitrary linear temperature variation over the assemblage. Thi s
dashed line is evaluated, it may differ fro m the temperature variation in element 2
94 Introduction to the tinite elemen t m ethod Approximating functi ons - s calar p roblems 95

observation lead s to the so-called patch test for no n-conforming element s. To check written as
that non-conforming elements fulfil the con vergence requiremen t they must pass this
test. The essence of the patch test is to check that an arbitrary assemb lage (i.e. patch) T(x, y) = To + ( aT) (x _ x o) + ( aT) (y - Yo)
of elements is able to reproduce exactly an arbitr ary linear temperature variation. ox 0 oy 0
Apart from some plate elements discu ssed in Chapter 18, in th is introductory text
we shall only con sider con forming elements and we refer to Hughes (1987), Strang + H~:~)o (x - xo)' + (o~~Jo (x - xoHy - Yo) (7.3)
and Fix (1973) and Zienkiewicz an d Taylor (1989 ) for a detailed treatment of the

+ ~ (~:n 0 (y
patch test and the behaviour of non-c on forming elements.
Let us return to the completeness requirements which mus t be fulfilled in order - Yo)' + 0(r 3 )
to achieve convergence. Referring to (7.2) we note that an approx imation which at
least involves an arbitrary linear polyno mial is necessary in order to fulfil the where the term 0(r 3 ), which is read as 'of order 3', is defined in general as
completeness criterion . With th is result at hand it will come as no surprise to see O(r") = Cr" for r --+ 0 (7.4)
later that the elementwise approximations always cons ist of polynomials possibly
containing even higher-order terms than linear ones. In this expression, C is a constant and r is the distance between the current point
(x , y) and the fixed point (x o, Yo) (cf. Figure 7.3). Moreover, in (7.3) To is the exact
temperature at point (x o, Yo) and, likewise, (oT / ox )o, (oT/ oy)o, (o2T/ox 2)o,
(o'T/ ox oY )o and (OZT/ oy2)o are the exact derivatives evaluated at the point (x o, Yo)'
It appears that (7.3) may be written as
7.1.2 Rate of convergence - Pascal's triangle
T(x, y) = a , + a,x + a 3y + a,x' + a,xy + a6y2 + Cr 3 (7.5)
Having established that the approximation of the un kn own funct ion (temperature) where at . .. a6 are constants.
is obtained using a pol ynomial, which at least includes all linear terms , we need In a region which contains the point (xo, Yo) we now assume that the
information on which higher-order terms it may be advantageous to include. F or approximation to the true temperature field is taken as
this purpose we shall discuss the conc ept of rate oj convergence, which provides
information on how fast the approximate solution converges towards the exact (7.6)
solution when the element size is decreased. where the no tation TOPP has been used in order to emphasize that (7.6) represents
Consider the two-dimensional body shown in Figure 7.3. Assume th at the exact the approximate temperature field. Subtracting (7.6) from (7.5) yields
temperature distribution T is kn own and let us make a Taylor expansion of this
exact solution about the point (xo , Yo). It follows that the true solutio n T can be T(x, y) - T"PP(x, y) = Cr' (7 .7)
and it appears that T - T"PP is the error in the approximation process. Let us now
consider a finite element (i.e. a region) which contains the point (x o, Yo) and let h
denote the maximum distance from the point (x o, Yo) to the element boundaries.
y Within this finite element the maximum error is written as IT - TOPP lmn and from
(7.7) we find that
IT - T"PPlm., = Ch'
We may consider the distance h as representative of the element size and to emphasize
this issue, the result above is written as
IT - T:PP lm., = cv (7.8)
where T:PP refers to the approximation for an element of size h.
'- x
Let us assume now that instead of an element of size h we use another,
geometrically similar element of size ah. In analogy to (7.8) it follows that
Figure 7.3 Taylor expansion of true solution about point (x o. Yo) IT - T:P lm., = Ca'h' (7.9)
96 Intro du ction to the finite el emen t method Approximating [unctions - scala r p roblems 97
Combining (7.8) and (7.9) results in more accurate than (7.12), owing to the term a,xy, and this is reflected by the fact
IT - T:~P lm81 = (Xl I T - r: PP lm.1l (7.10)
that , in general, C, $ C I (cf. (7.13) and (7.17. However, the important point is that
the rate of convergence is unchanged. Terms in the approximation which do not
If, for example, ex is chosen as t, i.e. the new element size is half its original size, we improve the rate of convergence are called parasitic terms.
obtain Obviously, unless other specific properties of the approximating function are
IT - T:r:l'I=, = I IT - T:PPlmu (7.11) required, one would try to avoid parasitic terms and instead use polynomials which
conta in terms of lower order before higher-order terms are invoked. A polynomial
This implies that with an element size of h/2, the maximum error in the which contains all terms of a specific order as well as all lower-order term s is called
approximation is one-eighth of the maximum error using an element size of h. The a complete polynomial (not to be confused with the completeness requirements,
term ~, present in (7.10) gives the rate of convergence, which in the present case is discussed previously ). In one dimens ion , the approximations a l + a, x and
of order 3. It follows that the rate of convergence provides information on how fast at + a2x +a 3 x 2 are complete, whereas the approximation a t + a2x + a3x 3 is
the approximate solution converges towards the true solution when the element size incomplete because it lacks the quadratic term .
is decreased. Observe tha t as the exact solution T is unknown in practice, it is no t In order to keep track of the terms in a polynomial used for two-dimensional
possible to calculate the value of the error IT - PpPlmu . approximation it is convenient to make use of Pascal's triangle, which lists the terms
In (7.6) we considered all terms up to and including the quadratic terms. Suppose in a systematic manner :
now that only the linear terms are considered in the approximation, i.e.
PPP(x, y) = a l + a,x + a,y (7.12) Pascal's triangle
As, for instance , the term (8'T/ 8x')o(x - xo)' IS of order O(r') we obtain by
subtracting (7.12) from (7.3) that
x y
T(x, y) - T'PP(x, y) = Clr' (7.13) x' xy y'
where C I is a constant. As in the derivation of (7.10), we obtain from (7.13) that Xl x 2 y xy2 y3

x 4 x3y X2y2 xy3 y4


(7.14)
.'
i.e. the rate of convergence is of order 2 for the approximation given by (7.12). Ifthe
element size is halved, i.e. ~ = t, (7.14) yields

(7.15) Any polynomial which contains all the terms above a certain horizontal line in
Pascal's triangle is a complete polynomial. For one-dimensional problems Pascal's
A comparison with (7.11) illustrates that the more higher-order terms we include in
triangle degenerates to I, x, x', x' ,... and even for three-dimensional problems it is
the approximation, the faster is the rate of convergence.
rather straightforward to generalize the results above .
However, this conclusion should be used with caution as the following example
With these results in mind, we can summarize our discussion for the elementwi se
will show. Consid er the approximation given by
approximation procedure:
PPP(x, y) = a l + a,x + a,y + a, xy (7.16)
Even though we include one quadratic term, the true solution (7.3) contains other
quadratic terms which are of the order O(r'). That is, subtracting (7.16) from (7.3)
Elementwise approximation
yields A polynomial is used.
This polynomi al must include at least an arb itrary linear polynomial -
T(x, y) - T'PP(x, y) = C,r' (7.17) then the completeness requ irements are fulfilled.
where C, is a constant. A comparison of (7.13) and (7.17) shows that the two Choose the polynomial so that cont inuity across element boundaries is
expressions are similar in structure even though C I i' C,. This means that achieved - then the compatibility or conforming requirement is fulfilled.
approximation (7.16) also leads to (7.14) just like approximation (7.12), i.e. we have If possible, avoid parasitic term s.
not improved the rate of convergence. The approximation given by (7.16) is in general
p

98 Introdu ction to the tinite element m ethod Approximating functions - scalar problems 99
Having established these general guidelines, we shall now discuss the elementwise
approximation of a scalar function like the temperature for one-, two- and
temperatures and
Figure 7.4 we obtain
7; ~ at the nodal points i and j, respectively. From (7.18) and

three-dimensional problems, and we will show how the approximation of the scalar

[~J = [: ::J[::J
function over the entire region of the body of interest can be established based on ~ = eel + Cl2X j

these elementwise approximations. For convenience the scalar function is denoted 7j = Cil + (J.2 X j or
by T, as it may represent the temperature of a body. which can be written as
(7.21)

7.2 One-dimensional elements where

7.2.1 Simplest possible one-dimensional element a' [7;J; C [I xiJ


=
~
=
1 Xj
(7.22)

Referring to (7.2), it appears that the simplest possible one-dimensional approxima- It appears that the column matrix a' contains the temperature at the nodal points
tion, which fulfils the completeness requirements, is given by of the element. From (7.21) it follows that
(7.18) (7.23)
where Cil and (X2 are constant parameters for each element. In a moment we will and insertion into (7.19) yields
express these parameters in terms of the temperature at the nodal points of the
T=NC -la' (7.24)
element. As we have two parameters, the element contains two nodal points.
Moreover, it is advantageous to locate the nodal points at the ends of the element, This expression can be written as
i.e. we obtain the element with the length L as shown in Figure 7.4.
Let us rewrite (7.18) in the following matrix form (7.25)
T = NIX (7.19)
where
where
N' = [N f N j] = Nc- ' (7.26)

N = [I x]; IX = [:: J (7.20) From (2.46) it is easy to determine the inverse matrix C - l and we obtain

It turns out to be advantageous to express the parameters 1X , and 1X2 in terms of the -X'J
1 whereL =x.) -x ,.
I (7.27)

Using this expression in (7.26) it follows that


T
I N' = [N~ Nj] (7.28)

where

1
N~
, = - -L (x - x.)
)
(7.29)
L--:e::::======:}-_ x
Xi
N~) = ~L (x - x .)
'

Figure 7.4 Simplest possible one-dimensional element The functions Nf and N j are called element shape functions or element interpolation
100 Introdu ction to the hnite element method App roximating functions - scalar problems 101

T T

"
''':'N '~ T.,
-.
'''.
Figure 7.5 Variation of element shape functions with position x
T
f unctions and the matrix N' is termed the element shape function matrix. It will tum
out later that different meth ods exist to establ ish the elemen t shape functions and
the specific method we ha ve adopted here is termed the C-matrix method (cf. (7.21).
It is of interest to evaluate how the elemen t shape functions Nf( x) and Nj(x)
va ry with position x, and Figure 7.5 sho ws thi s dependence. F rom (7.29) it is obvious
that N f( x ) a nd N j(x) var y linearly with position x. More over, from (7.29) and Figure
7.5 it follows that the element shape functions possess the following important ~====:=3- x
properties:
Figure 7.6 Approx imation of temperature in linear element
N~ ( x,) = I ; N f(x j) = 0
(7.30)
N j(x ,) = 0; N j(xj) = I
derived directly from (7.31). As an example, for x = x, we have T = T; , i.e.
These properties will also turn out to be valid for the more complex elements to be
[Nf(x,) - IJ T; + Nj( x ,)7j = 0
discussed later.
We note that (7.25) can be written as As this expression holds for arbitrary T;- and Trval ues, we conclude that Nf( x ,) = I
and Nj(x,) = 0, in accordance with (7.30). Likewise, if x = x j is used in (7.31) we
T( x ) = N ~(x)T; + N j( x)7j (7.31) conclude that Nf(x j) = 0 and N j(x j ) = 1, also in accordance with (7.30).
Referring to the weak formulation (4.36) of the one-dimensional heat flow
With expressions (7.25) and (7.31) we have obtained a formulation which separates problem, it is important to determine the gradient of the temperature T. F rom (7.25)
the influence of geom etry via the element sha pe functions from the influence of physics or (7.31) it follows that
via the temperatu resat the nodal points. Thi s a ppealing sepa ration is not present in
dT dN' dN' dN'
(7.18), where the parameters a t and a, depend on geometry as well as on nodal point _ = _ a t = _ ' 'f, + - j T. (7.32)
temperatures. It follows that once the geometry of the element is known, the element dx dx dx dx '
sha pe functions can be established directly. Thi s important feature is common for Let us define the matrix B' by
all types of finite elements and it grea tly facilitates implementation on a computer.
Expression (7.31) shows clearly that the approximate temperature is expres sed
as a suita ble interpolation between the temperatures at the nodal points. Thi s I B,=dN'
dx
I (7.33)
interpola tion is given by the element sha pe functi ons, which in the present case are
linear functio ns. Moreover, the approximate temperature T depends linearly on the Then (7.32) can be written as
temperatures T; and 7j at the nodal points, i.e. the appro ximation of the temperat ure
ca n be illustrated as shown in Figure 7.6.
~ (7.34)
When the fundamental properties (7.30) for the element shape funct ions were
establi shed , use was made of the expression s given by (7.29). However, (7.30) can be
LE::J
100 In trod uction to the finite element method
App roximatin g fu nc ti ons - scalar problems 101
N~ N!
" ) T T

T,
"
'";''N ''T.,
-,
'''.
Figure 7.5 Variation of element shape functions with position x
T
fun ctions and the matrix N' is termed the element shape function matrix. It will turn
out later that different meth ods exist to establish the element shape functions and T;
the specific method we have ad opted here is termed the C-matrix method (cr. (7.21 .
It is of interest to eva luate how the element shape functions NHx ) and Nj(x)
va ry with position x, and Fig ure 7.5 sho ws th is dependence. From (7.29) it is o bvious
that NH x) and N j(x) vary line arly with position x. More over, from (7.29) and Figure
7.5 it follows that th e elem ent shape function s possess the following impo rtant ~==~- x
properties:
Figure 7.6 Approximation of temperature in linear element
NH x ,) = I ; NH x J) = 0
(7.30)
N j(x,) = 0; N j(x J) = I
deri ved directl y from (7.31). As an example, for x = x, we have T = 7;, i.e.
Th ese properties will also tu rn o ut to be valid fo r the more complex elements to be
discussed later.
[N~(x, ) - 1]7; + N j( x,) 7j = 0
We note that (7.25) can be written as As this expression holds for arbitrary 7;- and Tjvalues, we conclude that N Hx ,) = 1
and Nj(x,) = 0, in accordance with (7.30 ). Likewise, if x = x J is used in (7.31 ) we
T(x ) = N~( x)7; + N j(x) 7j (7.31) conclude that N~(xJ) = 0 and N j( x J) = I, also in accordance with (7.30 ).
Referring to the weak formulation (4.36) of the one-dimensional heat flo w
With expressions (7.25) and (7.3 1) we ha ve obtained a formulation which separa tes problem, it is important to determine the gradient of the temperature T. Fr om (7.25)
the influence of geometry via th e elemen t shape function s from the influen ce of ph ysics or (7.31) it follow s that
via the temperatures at the nodal points. Th is appea ling sepa ra tio n is not pre sent in
(7. 18), where the param ete rs a t and a, depend on geom etry as well as on no dal point
d T dN'
- =_ at = _
dN'
1 1i + dN~
_ J 7j (7.32)
tempera tures. It follo ws tha t once the geometry of the element is known, the element dx dx dx dx
sha pe function s can be esta blished directly. This impo rta nt feature is co mmon for Let us defin e the matrix B' by
all types of finite elements a nd it greatly facilitates implementation o n a computer.
Expression (7.31) shows clearly that the approximate temperature is expressed
as a suita ble interpola tio n between the temperatures at the nodal points. Th is
interp olat ion is given by th e elem ent shape functi ons, which in the present case are
I B' =dN"
dx (7.33 )

linear functions. Moreover, the approximate temperature T depends linearly on the


Then (7.32) can be written as
temperatures 7; and 7j at the nodal points, i.e. the approximation of the temperature
can be illustrated as shown in Figure 7.6.
When th e fundamental properties (7.30) for the element shape functions were
established , use was made of the expressions given by (7.29 ). However, (7.30 ) can be
I dT = B' a'
dx
I (7.34)
102 Introduction to the finite element m ethod Ap proxime tinq functions - scalar problems 103

Therefore the matrix B' relates the temperature gradient to the temperatures at the the element. We note that, due to (7.26) and (7.20), B' may be written as
nodal points of the element. From (7.33) and (7.29) we obtain dN' dN 1
B' = - = - C =[0 I]C ' (7.36)
dx dx
B' = [dNi dNiJ = [- ~ ~J (7.35) where C - 1 is given by (7.27).
dx dx L L
We have already mentioned that the approximation chosen fulfils the complete-
ness requirements. Let us now check that the element formulation also fulfils the
As B' here is a constant matrix, the gradient dTldx is constant within the element
compatibility or conforming requirement stating that the temperature approximation
in accordance with the fact that the temperature is assumed to vary linearly within
must vary continuously across common element boundaries. For this purpose,
consider the body consisting of three elements shown in Figure 7.7, where, for
instance, NJ is the element shape function of nodal point j for element 1, whereas
T Ny is the element shape function of nodal point j for element 2. For each element,
Tj we have the approximation shown in Figure 7.6 and as we interpolate between the
in e lement 1 nodal points, and as, for instance, nodal point j is common for two elements, we
T; clearly have a continuous variation of the temperature over common element
dN'T
, / ' AI J J boundaries as shown in Figure 7.7. This means that the element is conforming.
N j Ti / " However, we note that the temperature gradient, in general, varies discontinuously
x over common element boundaries.
CD k m Figure 7.7 shows how the approximate temperature variation over the entire
T body is established from the approximation over each element. We shall now
Tj in element 2 formulate the approximate temperature variation over the entire body in a slightly

+
T,
~
N2 T >
" /----...;.. N2 T
different manner. Previously, we determined the element shape functions (cf. (7.29)
and Figure 7.7) from which it appears that one element shape function is related to
j j ,/,
Y " nodal points i and m, whereas two element shape functions are related both to nodal
,/ "- , point j and nodal point k. Let us now define the global shape function s in such a
x
manner that one global shape function is related to each nodal point. The following
CD k m
definitions are adopted:
T
N! for x in element I
N= {
in elemen t 3
I 0 otherwise
T,
+ Tm

CD
N:T'~_N3T
k
/"

m
mm
X
n, =
r'
0
for x in element I
NI. for x in element 2
otherwise
(7.37)

r t for x in element 2
T
Tj N, = ~t for x in element 3
T, over en tire body otherwise
T;
for x in element 3
Tm Nm = {N;'
0 otherwise
x
These global shape functions as well as definitions (7.37) are illustrated in Figure 7.8.
CD k m
In this figure, it appears that the global shape function for a specific nodal point
Figure 7.7 Approximation of the temperature over three elements only differs from zero in those elements which contain this nodal point. This feature
104 Introdu ction to the finite element method Approximating func tions - scalar p robl ems lOS
ELEMENT SHAPE FUNCTI ONS GLOBAL SHAPE FUNCTION S T

iCDj
o
k CD
O--- x
m
t~. i CD j
o
k CD
0---- x
m
+
"L , <


N; T;

j
k m
x

t e>. ~.
Tj
/', ,
+ /
/- N J' T'J ~' <,
I
CD k CDm x
k m

IL ~
N,
T

CD
N,2 _______

-- '-. ,
/'

O--- x
/' ............... ~

k CD m
/N~
'L
Nm
CD
~~
CD m
xk
+
Tk
/' ""' -,
....... / -NkTk--',
x
k m
N'm ......",. /'

'L /
T

L
,/


CD
k
t

m
X

CD


k 0
~ x

m
+
NmTm--...... #

Figure 7.8 Definition of global shape functions from element shape functions --
k
<

m
x

will turn out to be of fundamental importance in the later FE formulation and it is T


characteristic of all types of finite elements.
Tj
With these definitions of the global shape functions, it is obvious that the Tk
approximation of the temperature over the entire body can be written as = Ti
Tm
T = N ;T, + N jT; + N,r. + NmTm (7.38)
as illustrated in Figure 7.9. The reader is encouraged to compare Figures 7.7 and 7.9 k m
and check that (7.38) holds. Figure 7.9 Approximation of the temperature over entire body using global shape functions
The temperatures at all nodal points of the body are collected into the column
matrix a given by Then (7.38) can be written as

._[~]
T= Na (7.41)
(7.39)
It follows that the gradient within the body is given by

Moreover, the global shape j unction matrix N is defined by I dTdx = Ua I (7.4 2)


N = [N; n, N, N m] (7.40) .
Approx imating f unc tions - scalar problems 107
106 Introdu ction to the fin ite el em en t method

where B is defined as Let us rewrite (7.45 ) according to


(7.46 )
T =~"
~ (7.43) where
~
i.e.
~ = [I X x
2
]; , =[::] (7.47)

B= [dN' dN J dN, dNmJ (7.44)


dx dx dx dx
In order to express the parameters " I' "2 and o: in terms of the temperatures at the
As these results are generalized directly to bodies comprising more elements, we have
nodal points, we expres s that (7.45) for x = x, provides T = 1;, etc. This yields
obtained a complete description of the approximate temperature over an arbitrary
one-dimensional body. Note that the elements may be of different lengths. 1; = (Xl + (X2X I + (X3 xl
1) = a l + a 2 x j + r1.3 Xf
7.2.2 Quadratic one-dimensional element ~= Ct': l + IX2X", + Cl.3 X;

which can be written as


Instead of the simple linear one-dimensional element presented above we may easily
(7.48)
construct higher-order elements where the approximation of the temperature contains
terms of higher order than the linear ones. As an example consider the quadratic
where
approximation given by
(7.45) x, X~]
Again it is advantageous to express the three parameters "" "2
and "3by the
temperature at the nodal points. That is, there is room for three nodal points as
a,=[H c=[: Xj

X I;
xl

xl
(7.49)

shown in Figure 7.10. In order to obtain continuity in the temperature across As before , the column matrix a' contains the temperatures at the nodal points of the
neighbouring elements, nodal points are located at each end of the element and the element. From (7.48) it follows that
third nodal point may be located arbitrarily. In practice, however, the third nodal
( 7.50)
point is located at the middle of the element as shown in Figure 7.10.
and insertion into (7.46) yields
T= ~C -la' (7.51)
T
This expression can be written as

T = N'a' I (7.52)

where
j k N' = [ N: Nj N:J = sc: (7.53)
0

XI Xj X, Let us now determine the inverse matrix C - 1. First, we determine the determinant
L/2 L/2 of C. From (2.40) and choosing j = I we obtain
L det C = I(x)x: - x,xI> - l (x /x: - x, x; ) + I (x ,x; - xJx;)

Figure 7./0 Quadratic one-dimensional element = xJx, (x, - xJ) - x ,x, (x , - x .) + x /x)(x) - x .)
108 Introduction to the finite element method Approximating functions - 'ica l~ r problems 109

Reference to Figure 7.10 gives T T T

det C
L
= - (xjx, - 2x,x, + x,x j)
L
= - [x,(x j - x,) + x,(x j - x,)]
.I
2 2 N:T k __/
+
= ~ (x.~- X~) = L' (x,- x .) = L'
/
2 2 '2 4 '4
o - ~

Using (2.46), after a little algebra we then obtain j --- k i \:~; ' J k

XjX' - 2x,x. X,X j ] I


T
col = ~, -(Xj + x.) 2(x, + x.) - (x, + x) (7.54 )
[ Tk
I -2 I
Inserting (7.47) and (7.54) into (7.53) provides - Tj
T;
N' = [N, Nj N:J (7.55)

where
Figure 7.12 Approximation of temperature in quadratic element

N, = 2.
L'
(X - x.)(x - x.)
J

4 the following fundamental property


Nj = - L 2 (x - x,)(X - x.) (7.56)

N~
,
= {I
0
at nodal pointi
at all other noda l points
(7.57)

As before N" Nj , N1 are called element shape functions, whereas N' is termed the in complete analogy to (7.30).
element shape function matrix. Just as for the linear element, the specific manner by Expression (7.52) may be written as
which we have determined the element shape functions is termed the C-matrix method
(cf. (7.48)). It appears that the element shape functions are quadratic functions of x, T = N,7; + N j1j + Np, I (7.58)
and Figure 7.11 shows their variation with position x.
From (7.56) and Figure 7.11 it follows that the element shape functions possess which shows that the temperature T depends linearly on the temperatures 7;, 1j and
I; at the nodal points ; that is, the approximation of the temperature can be illustrated
as shown in Figure 7.12. . . ..
Just as for the linear element, the fundamental property (7.57) for the element
shape functions may be derived directly from (7,.5~). Let us assume that x c= ;,f~r
which T = 7;, i.e. . .
[Ni(x,) - 1]7; + Nj(x,)1j + Nl(x,) I; = 0
As this expression holds for arbitrary 7;-, 1j- and I;-values, we conclude (pat
Ni(x,) = 1, N j(x,) = N1(x') = 0 in accordance with (7.57). The remaining properties
j '-~ j -- k given by (7.57) are obtained in a similar way by setting x = x j and x = x.. .
Figure 7.ll Shape functions for quadratic element In the weak formulation (4.36), the temperature gradient appears. In order to
110 Introd uction to the finite elem ent method Ap proximating functions - scalar problems III

determine this gradient we use (7.52) and (7.58) to obtain T

d_ T = dN' dN'
__ a" = __I ~ + dN'
_ _J 1j + -dN-'' 1k (7.59) in element 1
dx dx dx dx dx
As with the linear element, the matrix B' is defined by

I B,=dN' I (7.60) m n
dx T o
i.e. we have
+
I dT = B'a'
dx
I (7.61)

n
where the matrix B' relates the tempe rature gradient to the temperatures at the nodal
points of the element. With the element shapefunctions given by (7.56) it follows that

ov er e nti re body
BC= dN'
__, dN'
__J
dNkJ ~ x. - x,
[ dx dx dx
=
e [2x - J
- 2(2x - Xi - x,)

(7.62)
I I j k m n
It is no surprise that B' and thus the temperature gradient within the element depends
linearly on x since the temperat ure was assumed to vary quadratically with x (cr. CD 0
Figure 7.13 Approximation of temperature over two quadratic elements
(7.45)). We also observe that, due to (7.53) and (7.47), B' may be written in the
following alternative manner:
the global shape functions in such a manner that one global shape function is related
e dN' d~ -1 to each nodal point. The following definitions are made :
B = -= - C = [0 I 2x]C - 1 (7.63)
dx dx N' for x in element I
Nj = { 0 I

otherwise
where the constant matrix C - 1 is given by (7.54).
To illustrate how the temperature variation over the entire body is obtained
from the elementwise approximation in a manner ana logous to the approach for the
linear element, we consider the body shown in Figure 7.13, which is assumed to
Nj =
t'
0 j
for x in element I
otherwise
for x in element I

e'
comprise two elements.
The notation is as before, implying that Nt is the element shape function of N, = :; for x in element 2 (7.64)
nodal point k for element I, whereas N ] is the element shape function of nod al point
otherwise
k for element 2. Since we interpolate between the temperatures of the nodal points
and as, for instance, nodal point k is common to two elements, we clearly have a for x in element 2
N = { N;'
continuous variation in the temperature over common element boundaries. That is, m 0 othe rwise
the element is conform ing or compatible. However, the temperature grad ient varies,
in general, in a discontinuous manner over common element boundaries.
Just as for the linear element and with reference to Figure 7.13, we now define
N =
" t'
0
II
for x in element 2
otherwise

l
112 Introd ucti on to the finite element method Approximating functions - scalar p r oblems 113
ELEMENT SHAPE FUNCTI ON S GLOBAL SHAPE FUNCTIONS element shape function Ni(x) acco rdi ng to
N;

lL' \:"N ;I
o
j <:"' k
0 0

m
o-- x
n
l L~o
j ""' k
0

m
0------ x
n
N( x)
I
= {N,( x)
0
if element a co ntains the glo bal nod al point i
otherwise
(7.66)

<D 0 N <D 0 The reader is invited to check that (7.64) is contained in th is definition and th at even

'L:>.
(7.37), applicable for the linear elemen t, is co ntained in the form given by (7.65) a nd

'L/~-'\N:
(7.66). In fact , we shall adopt the definitio ns (7.65) and (7.66) for all types of finite
elements. These expre ssions imply that a global shape funct ion related to a specific
o.---x
N
o 0 0---)( nodal point differs from zero only in those elements which contain the nodal po int
j k m n j k m n in quest ion .
<D <D 0 It appears that the approxi mat ion over the enti re bod y given in Figure 7.13 can

lL
k
now be written as

~'L o/;>o
1 A N2
N k __ '- k
-/ V-, (7.67)
o---x
N o--X
(

0 0
,. "'- .;:: j k m
<:
n j k m n as illustr~ted in Figure 7.15. Th e tem peratures at all nod al point s of the bod y are
<D 0 m <D 0 collected into the column mat rix a given by

'L 0

j
<D
2
Nm -


k
(
,
(
/ ..... - ...... -,

0
0

m
-,
,
' Q----. X
n
'LN
n
0

<D
j
~
km
o---- x0

cD
n
a=
1;
7j
T, (7.68)
Tm

'L 0

j
<D

k""---
N'
n~

0
m
,-

0
/'
0--)(
n
'L <D
0

j
0

k <=::::='" m
0
/ o----x
n
T"
Moreover, the glob al shape functi on matrix N is defined by
N = [N ; Nj N, Nm N.] (7.69)
Figure 7./4 Definition of global shape functions from element shape functions Then (7.67) can be written as

T = Na (7.70)

For the bod y shown in Figure 7.13, these global sha pe functions take, as well as T he gradient within the bo dy becomes
definitions (7.64), the form sho wn in Figu re 7.14.
In order to write expression (7.64) in a more compact fashion, let us denote a
global nodal point by i. For a given position x, we may identify the associated element
, dT
dx
= Ba I (7.71)
with number a which contains this position; that is,
where B is defined as

~
position x => identification of associated element a (7.65)
(7.72)
With this observation, the glob al shape funct ion N;(x ) can be defined from the ~
11 4 Introd uction to the tinite elemen t method Approximating tunctions - scalar problems 115

i.e.

B = [d N; dN J dN, dN m d N'J (7.73)


dx dx dx dx dx
It is obvious that this concept generalizes directly to bodies comprising more
elements.

T -,
J
+ I
/ ' __ N r
\ J J
II '\
\ 7.2.3 Cubi c and quartic one-dimensiona l elements - Lagrange
x
0

k
0

m n
interpolation
T
T, , Previously we established the linear and quadratic elements and we may proceed to
I \ consider the cubic element given by
+ f \ __ Nk Tk
,-,
I (7.74)
I ,
0 0 x In order to express the parameter s a . . . a. in terms of the temperatures at the nodal
,., ~
k m ....- - " n
point, it is possible again to use the C-matrix concept given by (7.21) and (7.48),
T
which requires that the inverse col is established. For the linear element, this inverse
is easily obtained, but for the quadra tic element the calculations are rather
+ cumbersome. This obstacle will be much more prono unced when an expression like

--
Tm NmTm __ ... -- ....
, ,, (7.74) is considered, but it turn s out tha t a general proced ure exists by which the
0 0 x element shape functions can be written down directly.
k m n To establish this general procedure, we shall first ado pt a suitable number ing of
T the (local) nodal points belonging to an element. It is emphasized that this local
numbering has no effect on the numbe rs of the global nodal points. With the nodal

'"L
+ numbers of the linear element given in Figure 7.16(a), (7.29) provides the element
Nn Tn---. ,"" shape functions
. s-: o----J(
0 0
k
-_ ....
0
m n N,
,I
= - r: (x -x 2 ) ;
I
Ni=r:( x-x ,) (7.75)

T Likewise, with the nodal number s of the quadratic element given in Figure 7.16(b),
~Tj 'N, T,'N mTm 'NnTn

01 bl
x, x, x, x, x,
k m n
x 0 ~ x
2 2 3
L L/2 L/2
Figure 7. J5 Approximation of the temperature over entire body using globa l shape functions
L

Figu re 7. /6 (a) Linear element; (b) quadratic element


116 Introduction to the finit e elem ent method Approxima ting functions - scalar prob lems 117
x, x, x, x,
0 0 Q o 0 0 0 0
2 3 4 1 2 3 4 5
LI3 I LI3 I LI3 ~~~~
L L

Figure 7.17 Cubic element Figure 7.19 Quartic element

x, x,
: 0 x
2
- ", - ",
Figure 7.20 Linear element applicable for axially loaded bar

'-----..---;tL--""'----+--+-----'1r------7"-----"~--x directly by setting


x, Xn_l x n

Figure 7.18 Lagrange 's interpolation function


'-- ---'1 k = I, 2, . . . , n (7.79)

To illus trate the use of Lagrange's interpolation function, we shall re-establish


(7.56) resu lts in the element shape functions N~ and N~ for the linea r element shown in Figure 7.16(a).
As the element comprises two nodal po ints , we have n = 2 and (7.79) then provides
N~ =, L
2
(x - x,)(x - X3); N~ = - ,
4
L
(x - x,)(x - X3);
(7.76)
N, = Ii; k = 1, 2

2 F rom (7.78) we get


N'3 = , (x - x,)(x - x,)
11 _ x- X2 . _ X - X1
L 1 - ,
1
12 -
Xl - Xl X2 - Xl
Consider now the cubic element in F igure 7.17. It is required that the element
shape functions allow us to write and as L = X 2 - X I ' we have rederived (7.75). In a similar way, the reader may verify
that for the q uadratic element, i.e. " = 3, use of (7.79) yields the results given by
(7.77) (7.76). Therefore, the elemen t shape functions of the cubic element shown in F igure
i.e. that the temperature is interpolated between its values at the nodal points, but 7.17 are obtained directly from (7.79) by putting" = 4, whereas the element shape
in order to do so the element shape functions must possess the fundamental property functions of the quartic element given by Figure 7.19 emerge for n = 5. This process
given by (7.57). We also know that the element shape functions must be a polynomial may be continued for arbitrary n-values. As the element shape functi ons for all the
and a polynomial fulfilling (7.57) is given directly by Lagrange's interpolationJormula. one-dimensional elements con sidered may be derived using Lagrange's interpolation
In fact, for n given points this formula provides a polynomial of order n - I given by formula, these elements are termed Lagrange elements.
Having discussed one-dimensional elements for heat flow problems in great
, _, (x - x,)(x - X2) ... (x - xk_,)(x - Xk+,) .. . (x - x,) .
Ik (x)= , detail, we finally mention that these elements obviously apply to axially loaded bars
(x, - x ,)(x k - X2) ... (x, - Xk _,)(Xk - Xk+ l) ... (Xk - x,) as well. In principle, however, one aspect needs a little attention. Whereas the
k = 1,2, . .. , n (7.78) temperature T is a quantity without any direction, the axial displacement u has
direction. When the differential equation (4.18) was derived for an axially loaded
where it should be noted that the term (x, - xk ) is not present in the denominator elastic bar, the axial displacement u was considered as positive when it was in the
and, likewise, the term (x - Xk) is not present in the numerator. It appears that direction of the x-axis, Taking account of these remarks, the simple linear element
I, -'(xk) = I and I, -I(X,) = 0 for k " i, as illustrated in Figure 7.18. M oreover, these applicable for the axial loading of a bar is illust rated in Figure 7.20, where u, and
properties are in accordance with the fundamental property for element shape U2 are the displacements of nodal points I and 2, respectively. A similar approach
functions given by (7.57). T herefore, the element shape functions can be constructed holds for high er-o rde r elements.
118 Introduction to the hnite element meth od Approximating functions - scalar problems 119

7.3 Two-dimensional elements T


Tj
I
7.3.1 Simple triangular element I
I
I
After this detailed discussion of one-dimensional elements, it is timely to consider I
two-dimensional elements. Referring to (7.2), the simplest possible two-dimensional T =al + Q.2x+ a.3 Y
element which fulfils the completeness requirements is given by I
I
(7.80) I Y
I
I
As we have three parameters at , a, and a" which are eventually to be expressed in
terms of the temperatures of the noda l points , the element must possess three nodal k-:!-r."T77n "W' j
points. These nodal points are located so that they define the element geometry , i.e.
we obtain the tri angular element shown in Figure 7.21. x
The interpolation given by (7.80) is illustrated in Figure 7.22. As this interpola tion
is linea r in x and y the element is termed the linear triangular element. It also follows
Figure 7.22 Variation of temperature over linear triangular element
that con tour curves within the element, i.e. curves for which T = constant, are straight
lines. values at the nodal point s, i.e.
In ord er to determ ine the element shape functions corresponding to (7.80) we
again use the C-matrix meth od. Therefore, let us write (7.80) as T = NiT, + Nj~ + Np; (7.83)
T = NIX (7.81)
where, as before, Ni(x, y) is the element shape function at nodal point i and T" ~ ,
where 1; are the temperatures at the nod al points . Therefore, inserting the coordinates for
the nodal point s, we obtain from (7.80) the following system of equations in the same
manner as for one-dimensi onal elements:
N =[I x y] ; IX =[::] (7.82)
(7.84)
where
Just as for one-dimens ional element s, our objective is to reformulate (7.80) so that
it clearly appears that the temperature within the element is inte rpolated between its
(7.85)

Y
Moreover, determ ination of IX by means of (7.84) and insertion into (7.81) yields
k
I x k Yk} (7.86)

crec =A
where
I Xj .Yj} N' = NC - ' (7.87)
and

L-- - - - - - N' = [Ni Nj Nn (7.88)


X

Figure 7.21 Simplest possible two-dimensional element It can be observed that (7.86) is in accordance with (7.83).
120 Introdu ction to the finite elemen t m ethod App roximating functions - scalar problems 121

To determine the inverse matrix C - l


, the determinant of C is calculated first. We observe that the element shape functions are linear functions of x and Y in
From (2.40) and choosingj = 1 we find accordance with (7.80)and their variation over the element is shown in Figure 7.23.
By inspection it is easily shown that (7.57) again holds, i.e.
det C = (x jYk - x,Y) - (x,Yk - x,y,) + (x'Yj - x jY,) (7.89)
Moreover, as the coordinates of the nodal points are known, trivial geometrical I at nodal point i
arguments show that the area A of the triangle in Figure 7.21 is given by N~ = { 0 at all other noda l points (7.93)

2A = det C (7.90)
Indeed this is a general property for all element functions and it follows directly from
One point needs some attention. In the matrix a' given by (7.85) we have listed the the form (7.83) in a similar manner as in the discussion related to (7.57).
nodal points in the sequence i,j and k. Referring to Figure 7.21 , it appea rs that this
Apart from the fundamental property given by (7.93) we also observe from Figure
sequence corresponds to a counter-clockwise direction. If a clockwise direction was
7.23 that, for instance, the element shape function Nk(x, y) belonging to the nodal
chosen, as given by the sequence i, k and j, for instance, this implies that the last two point k is zero along the element boundary given by i, j . To prove this formally we
rows of C, as given by (7.85), are interchanged. According to property (5) :or
recall that Nk(x" y,) = N :(x j, Yj) = 0, and as Nk(x, y) varies linearly it follows that
determinants (see page 18) this would change the sign of the determinant, i.e , N: = 0 along the element boundary i, j.
det C = - 2A. For convenience, we want (7.90) to apply always, t.e, we adopt the
Let us return to the interpolation formula (7.83). To appreciate fully how it
convention that the nodal points should always be taken in the counter-clockwise
works, the illustration shown in Figure 7.24 is essential.
direction, as expressed through the sequence i, j and k. However, we may equally
In the weak formulation of the two-dimensional heat flow problem (cf. (7.1)J,
well take the sequences j , k, i or k, i, j which are obtained from the sequence i, j, k
the temperature gradient V T enters the expression. Therefore, from (7.86) we derive
by a so-called cyclic permutation. The reason is that each cyclic permutation does
not change det C. To see this, consider the sequencej, k, i. From (7.85) the following
is implied: firstly that the first and second rows are interchanged, which ch~nges the
sign of det C; secondly that the second and third rows are interchanged, which agam VT = N' a']
aaxT]= [O
ax
changes the sign of det C so that it returns to its original value. [aT aN' a'
We conclude that for any counter-clockwise direction i, j and k where nodal oY oY
point i is arbitrary, the relation (7.90) holds.
The inverse matrix C- l can be established by means of (2.46) and some algebra which can be written as
will show that
VT= VN' a' (7.94)
x,Y, - x ,Yk x'Yj - X1Y']
Yk - Yi Yi - Yj (7.91) where VN' is defined according to
Xi - Xk Xj - Xi

The element shape functions are obtaine d from (7.87) and (7.88), which with VN' =
OaxN'] (7.95)
(7.82) and (7.91) yield [aN'
oy
Let us define the matrix B' by

(7.92)
B' = VN' =
OaxN'] (7.96)
[aN'
oy
122 Introduction to the finite elemen t method Approximating functions - sca lar problems 123
Tj

I
I
I
I Tk
I
I

) - - - - - -- - - - y
+ I
I
+
I
I

I
j
k

N'
J

x Figure 7.24 Approximation of temperature in triangular element

which implies that (7.94) takes the form

VT = B'a' (7.97)

This expression for the temperature gradient is analogous to the formulations given
for one-dimensional elements (cf. (7.34) and (7.61)). The matrix B' relates the
:>....;,....--c- ---- y temperature gradient in the element to the temperatures at the nodal points of the
element.
With N' given by (7.88) it follows that

aN1 aN'
x __J

ax ON:]
ax
B' = ax (7.98)
[ aNi aN'
__J aN:
Figure 7.23 Element shape functions for linear triangular element oy oy ay
124 Introducti on to the fin ite element method Approximating functions - scalar problems 125
y

L- x Figure 7.26 Global shape function Ni(x, y) for nodal point i for linear triangular elements

Figure 7.25 Two neighbouring triangular elements


over the entire body can be written as

The element shape functions are given by (7.92), i.e. T=Na (7.102)

The column matrix a contains the temperatures at all nodal points of the body, i.e.
(7.99)
T,
T,
which shows that B' is a constant matrix. Therefore, in accordance with the linear a= (7.103)
expression (7.80), the temperature gradient VT is constant within the element.
Having established the elementwise approximation, our attention is now directed
towards the temperature approximation over the entire body. For this purpose we
T"
consider two neighbouring triangular elements (Figure 7.25). The element shape where n is the total number of nodal points in the body. The global shape function
functions Nt and N~ both relate to nodal point k, but Nt refers to element I and matrix N is then given by
N~ refers to element 2. From (7.93) it follows that Nt(x k, Yk) = N~(Xk' y,) = I and N=[N, N, ... N,] (7.104)
Nt (x j' y) = N~(xj' y) = 0, and as both Nt are N~ are linear functions, we conclude
that Nt = N~ holds along the common element boundary j, k. The temperature gradient within the body becomes
Considering a specific nodal point, let us construct a global shape function related
to that point in the same manner as for one-dimensional elements (cf. (7.65) and VT= Ba (7.105)
(7.66)). Therefore, for a given position (x, y), we may identify the associated element
a which contains this position, i.c, where B is defined by

aN,

['"'
position (x, y) =0> identification of associated element a (7.100) ...

For a given nodal point i the global shape function N, is then defined from the
B = VN = ax
aN,
ax
aN, ...
'en,
N"j
ax
(7.106)

element shape functions Nf according to ay ay ay

Nf (X, y) if element a contains global nodal point i The approximation of the temperature given by (7.80) clearly fulfils the
N,(~y)= . (7.101) completeness requirements. Let us now check whether the simple triangular element
{o otherwise considered is conforming, i.e. whether the approximate temperature distribution
varies continuously over common element boundaries. For this purpose consider
With these definitions the variation of a global shape function is shown in Figure 7.26. the common element boundary given by j, k in Figure 7.25. As the temperatures
From the global shape functions defined above, the temperature approximation at nodal points j and k are the same for the two elements and as the temperature
126 Introduction to the tinite elem en t m eth od Approximating functions - scalar problems 127

varies linearly between nod al points j and k, it follows that it varies in a continuous between the temperatures at the noda l points using the element shape functions , i.e.
manner over the common element boundary j, k. This means that the element is
conform ing. (7.108)
The linear triangular element dealt with so far is a very simple element. It was
suggested by Tu rner et 01. (\ 956), and even though in many respects it marked the where
beginning of the FE meth od, it is still widely used. It is essentia l that the element T1
can take the form of an arbitrary trian gle, which implies that the geometry of any
irregular two-dimensional body can be approximated as closely as requi red. This T,
Nt = [ N~ Ni Nj N~ ] ; at = (7.109)
ability to model arb itrary geometries is one of the essential ad vantages of the FE T,
meth od.
T.
and where the element shape functions fulfil the usual requirement given by
7.3 .2 Four-node rectangular element - the Lagrange element (7.93). This form may be obtained by expressing the parame ters a , .. . a. in terms of
the temperature s at the nodal points in the trad itional manner using the C-matrix
Let us now consider an element which con tains four nodal points, i.e. nodes. These method (cf. (7.84)). However, another method is possible using Lagrange's
nodal points are chosen so that they define the geometry of the element and we are interpolation formula (7.78).
then led to the rectangular element shown in Figure 7.27. For reasons that will be We first observe that in the x-direction, i.e. when y = constant, (7.107) implies
revealed later, the sides of the element are parallel to the coord inate axes. As we have that T varies linearl y with x. Likewise, in the j-direction, i.e. when x = constant,
four nodal points, (7.2) and Pascal's triangle result in the approximation (7.107) shows that T varies linearly with y. Let us consider next, for instance, noda l
point 1 of Figure 7.27. In the x-direction we adopt Lagrange's interpolation form ula
(7.107)
for n = 2 given by
The element defined by Figure 7.27 and (7.107) is occasionally referred to as the
Melosh element (Melosh, 1963), and it involves the parasitic term xy. Instead of this !l( x) = x - x,
quadratic term xy, one may in principle choose the term x ' or y' , but the 'choice xy Xl - x2
is in general preferable because it implies that the dependence on x and y is similar, which fulfils ll(x,) = I and ll( x,) = O. In the y-direction we choose
i.e. the approximation is of the same type in the x- and y-directions. Despite the
presence of the quadratic term x y, it is also observed that (7.107) for y = constant !l (y) = y - Y.
implies a linear variation with x and for x = constant a linear variation with y. For y , - Y.
this reason, (7.107) is often referred to as a bilinear expression.
i.e, !l(Y l) = I and !l(y. ) = O. The element shape function N; is then given as
As before, we want to reformulate (7.1 07) so that the temperatu re is interpolated
x - x, y - Y. I
N ;( x , y) = l: (x )!l (y ) = = - (x - x ,)( y - Y. )
x, - x, Yl - Y. 4ab

y y

1>---'------'----'-----l> 2
.......- - - - -- - - - - -- - - _ _ x
'---- - - -- -- - -x
Figure 7.27 Fou r-node rectangular element Figure 7.28 Element shape function N~ for four-node rectangular element
128 Introdu ction to the finite el em ent method Approximating [unctions - scalar p roblems 129

where 2a = x, - x, = x, - x. and 2b = Y. - y, = y, - y, (cf. Figure 7.27). T he where


reader may verify that Nl(x" y,) = I and Nj(x" y,) = N;(x" y,) = Nj(x., Y.) = 0
in accordance with the general requirement given by (7.93). Using the same approach ON ' ] [ OW oN , oN~ ON~]
for the other nodal points, we find that B' = VN' = ox = ox' ox ox ox (7.112)
oN ' oN ; oN 1 oN~ oN~
[- ------ - -
oy oy oy oy oy
N; = _1_ (x - x,)(y _ Y.)
4ab T he element shape functions are given by (7.110), i.e.

N1 = - _1_ (x - x,)(y - y,)


4ab
B' - I [y - Y. Y, - Y Y- y, y, - YJ (7.113 )
(7.110) 4ab x- X2 Xl - X X - X4 X3 - X

N~ = _1_ (x - x.)(y - y,)


4ab With the global shape funct ions defined in the usual manner (cf. (7.100) and
(7.101)) we obtain similarly to (7.102) -( 7.106)
N~ = - _1_ (x - x,)(y - y,)
4ab T =Na (7.114)

By inserting (7.110) in (7.108) it can easily be checked that the form given by (7.107) and
is obtained. Since the element shape functions may be derived using Lagrange's
interpolation formula, the four -node rectangular element is an example of a Lagrange VT= Ba (7.115)
element .
The variation of one element shape function over the element is illustrated in where
Figure 7.28. The appearance of the other element shape functions is quite similar.
Whereas the element shape functions vary linearly for lines parallel to the coordinate
axes, the presence of the term xy in (7.107) implies that they vary non-linearly along
= nN =
v
o~']
B [0en,;' 0;, ... e, (7.116)
all other lines. We observe that an element shape function related to a specific nodal oN ,
point is zero along element boundaries which do not contain the nodal point in oy oy oy
question. This property was also found for the triangular linear element and in order
T he approximation (7.107) fulfils the completeness requirement (cf. (7.2)). To
to prove this property formally for the four-node element, consider, for instance, the
check that the element is conforming, i.e. that the temperature varies continuously
element shape function N~ shown in Figure 7.28 along the element boundary 2,3.
across common element boundaries, consider the two elements shown in Figure 7.29.
As x = constant along element boundary 2,3, N~ varies linearly with y along this
Along the common boundary 2,5 we have x = constant, i.e. according to (7.107)
boundary. Moreover, as N~(x" y,) = N4(x" y,) = 0, we conclude that N~ = 0 along
element boundary 2,3. The same type of argument implies that if a nodal point is
shared by two neighbouring elements, then the two element shape functions belonging y
to that nodal point are identical along the common element boundary.
t; 5 6
As before, the temperature gradient is of interest and similar to (7.94)-(7.98) we

I I I
obtain from (7.108)

VT = VN'a' CD ~

which can be written as 2 3

x
I VT = B'a' (7.111)
Figure 7.29 Conforming elements
130 Introdu cti on to the fi nite el em en t method App rox imating fu nctions - scalar p r oblem s 131
y be shown that the elements in Figure 7.31 are conforming, i.e, cont inuit y across
common element boundaries exists.
s 6
The restriction that the four-node element must be rectangul ar with its element

I I
boundaries parallel to the coordinate axes has been emphasized. However, we shall
see in Chapter 19 that this restriction can be removed by means of the so-called
isoparametric fo rmulation of finite elements.
2 3
'------- - -- -- - -- - - -- _ x 7.3.3 More complicated rectangular and triangular
elements - serendip ity elements
Figure 7.30 No n-conforming elements
It is obvious tha t the manner in which Lagrange's interpolation formula was used
y to construct the element shape functions for the four-node rectangular element can
be generalized. Withou t going into detail, consider the rectangular element shown
in Figure 7.32. The specific feature of this six-node element is that it contains three

'---
I x
nodes in the x-direction and only two nodal points in the y-direction. Using Pascal's
triangle the temperature approximation becomes
T = IXI + 1X 2X + Cl 3Y +.(X4X 2 + cx~ xy + (X6X 2Y
which implies a quadratic variation in the x-direction and a linear variat ion in the
y-direction, in accordance with Figure 7.32. The element shape functions may be
(7.117)

Figure 7.31 Conforming elements constructed directly in a manner similar to that leading to (7.110), i.e. any element
shape function is obt ained as a suitable multiplication of a second-order Lagran ge
polynomial in x and a first-order Lagrange polynomial in y. The element of Figure
the temperature varies linearly with y for both elements. As the temperatures at nodal 7.32 is therefore another example of a Lagrange element and may be of interest for
points 2 an d 5 are identical for the two elements we conclude that continuity exists problems where it is known, a prio ri, that the temperature variation is more
across the common element boundary; tha t is, the element is conforming. pronounce d in the x-direction than in the y-direction,
However, it is import ant that this continuity is present only when the element Another Lagrange element that is used more often, and which exhibits
bou ndaries are parallel to the coo rdinate axes. To show this consider Figure 7.30. geometrical isotropy, is the rectan gular nine-node element shown in Figure 7.33. As
The common element boundary 2,5 is given by the equation y = ax + b. Insertion we have a quadratic variation in the x- and y-directions and nine nodal points, use
of this expression into (7.107) implies that of Pascal's triangle yields the following approximation:
T = ~l + ~ 3 b + (~2 + ~ 3a + ~4b) x + ~4 ax2 T= :XI + a: 2x + 1X3Y + 1X 4X 2 + IX sxy + C'X 6Y2 + IX, X 2 Y + :Xsxy 2 +
Ct: 9 x
2 2
y
i.c, T varies quadrat ically with x along boundary 2,5. This quadratic function is (7.118)
defined uniquely by three temperat ures, but as only the two temperatures T2 and T, y
are known, the quadratic expression above is not uniquely determined. This means
that the temperature variation along the common boundary in element 1 differs from
that of element 2. Therefore, continuity across common element boundaries is
violated implying that the element is non-conforming. In general, we would like to
avoid this situation, which means that the four-node element should be used in a
form where the element boundaries are parallel to the coordina te axes.
It appears that the four-node rectangular element in itself is unable to model
ar bitrary geometries of the body. However, if it is combined with the simple triangular L
x
element, arbitrary geometries can be modelled, and an example of such a combination
is shown in Figure 7.31. Using arguments similar to the above discussion it can easily Figure 7.32 Six-node rectangular Lagrange element
132 Introd uction to the fin i te ele ment method Approxima ting functio ns - scalar p rob l ems 133
y y

'---- -- - -- - - - - - _ x L -- x
Figure 7.33 Nine-node rectangular Lagrange element Figure 7.35 Six-node triangular element

y
0

Figure 7.34
1\
Conforming elements
I:
L - -- - -
~
x

Figure 7.36 Conforming elements


which is symmetric in x and y. It appears that the last three terms are parasitic.
Moreover , as all linear and quadratic terms are present, this nine-node element is Using arguments similar to those given previously, it can easily be shown that
able to reflect an arbitrary linear variation of the temperature gradient. The element any element mesh consisting of six-node triangular elements is conforming . As the
shape functions can be obtained as a suitable multiplication of Lagrange polynomials six-node element may form any triangle, we conclude that any two-dimensional
of second order in the x- and y-direction or using the C-matrix method . As the geometry can be approximated as closely as requested. As shown in Figure 7.36, the
temperature varies parabolically along any element boundary, a combination of six-node triangle can also be used in combination with the rectangular nine-node
rectangul ar nine-node elements is conforming. However, analogous to the discussion element to model bodies with skew bound aries and this combination is easily shown
of Figure 7.30, this conforming behaviour requires that the element bound aries are to be conforming. The six-node triangle is treated in detail by, for instance,
parallel with the coordinate axes. This requirement can be relaxed by use of the Zienkiewicz and Taylor (1989) and Gallagher (1975).
isoparametric FE formulat ion presented in Chapter 19. A conforming combination At this point, it is no surprise that a variety of elements exists and can be
of the three-node triangle and the four-, six- and nine-node rectangular Lagrange suggested, and we shall conclude the discussion of two-dimensional elements by
elements is shown in Figure 7.34. Further information on the nine-node Lagrange considering the often used eight-node rectangular element shown in Figure 7.37.
element may be found, for instance, in Becker et al. (1981) and Hughes (1987). The eight-nod e element looks a little like that of Figure 7.33, but it is not a
It is also possible to obtain triangular elements of higher order than the simple Lagrange element because we only have two nodal points along the line given by
three-node element treated previously. An example is the six-node triangle shown in nodal points 5 and 7. Therefore, Lagrange polynomials cannot be used directly to
Figure 7.35. Referring to Pascal's triangle, we can write the approximation as establish the element shape functions. Nevertheless, it is possible to construct these
T = a, + a,x + a,y + a,x' + a,xy + a6y' (7.119) element shape functions directly. To facilitate this process, we make a coordinate
transformation from the xy-system to the ~ry-system (cf. Figure 7.37). The
which forms a complete polynomial suggesting an efficient element. Moreover, any ~ry-coo rdin a te system is called a local coordinate system and it is devised so that
linearly varying temperature gradient can be modelled by (7.119). I~ I = 1 and Iry I = I hold along the element boundaries. Clearly, it is easier to establish
134 Introdu ction to the fin ite elemen t m ethod Approxima ting [unctions - scalar problems 135
y
3
l - 1]~O

1] ~ 1
6
4 7 I 3

8 6

5
Figure 7.39 Construction of element shape function N1 for corner node

L -- -- - - x
Next let us establish the element shape function for a typical corner node like
Figure 7.37 Eight-node serendipity element nodal point I. Referring to Figure 7.39, we require that N ~ is zero along the sides
2,6,3 and 4,7,3. Moreover, N ~ is also required to be zero at nodal points 5 and 8. It
7 3 turns out to be advantageous to assume that N~ is zero along the line defined by
l- 1]~ O
points 5 and 8. These observations imply that N~ = - t(l - ~ )( I - ~ )( l + ~ + ~),
where the factor - t is determined such that Nl( - I, - I) = 1. Following the same
8 6 procedure for the other corner nodes, we obtain

N~ = - to - ~) (I - ~)(I + ~ + ~)
1
I
5 2
I
N~ = - to + ~ ) ( I - ~ )(I - ~ + ~ )
I I (7.121)
I
1 +g~ 0
I
1 -g ~ 0
N; = - to + ~ )( I + ~ )( I- ~- ~ )
N~ = - to - ~ )( I + ~)(I + ~ - ~ )
s
Figure 7.38 Construction of element shape function N for mid-side node
Even though we have presented a systematic approach for the establishment of the
element shape functions, these functions were originally formulated by inspection
(Ergatoudis et al., 1968). The eight-node element is therefore called a serendipity
the element shape functions in the local ~~-coordinates, and from these results the element after the famous princes of Serendip noted for their chance discoveries
element shape functions may be expressed in terms of xy-coordinates, if requested. (Horace Walpole, 1754). The element shape functions for a typical mid-side and
Let us first consider a typical mid-size node like nodal point 5 in Figure 7.38. It corner node are illustrated in Figure 7.40.
is required that the element shape function N; is zero along the sides 2,6,3 and As before, we have that the temperature T is related to the temperature at the
4,7,3 and 1,8,4. F rom Figure 7.38 we concl ude tha t N; is given by N; = nodal points a' through T = N' a", and using (7.120) and (7.121) it follows that
t(l - ~ )(I + ;)(1 - ~), where the factor t is determin ed such that N;(O, - I ) = I.
By similar arguments for the other mid-side nodes we obtain (7.122)
We conclude that the eight-node serendipity element can model any linear variation
N; =!<J - ~2) (I _~) of the tempera ture grad ient and that (7.122) is symmetric in x and y. The last two
N~ = W + ;)(1 _ ~ 2 ) terms of (7.122) are pa rasitic of third order and a comparison with the nine-node
N,=!<J _ ~ 2 )( 1 + ~ )
(7.120) Lagrange element (F igure 7.33 and (7.11 8), shows that the internal node of the
Lagrange element contributes a parasitic term of fourth order. This suggests tha t for
N,= W - ~)(I _~2) the same number of nodal points (i.e. degrees of freedom) in a body, the eight-node
serendipity element is slightly more efficient than the nine-node Lagrange element.
136 Introduction to the finite element method Approximating [unctions - scalar problems 137
oj z
z
01 bJ
k
4

bJ

x x
I
I
Figure 7.41 (a) Four-node tetrahedral; (b) ten-node tetrahedral

Figure 7.40 Element shape functions for eight-node serendipity element : (a) mid-side node;
(b) corner node
""
As before, the eight-node serendipity element must have element sides parallel to the
coordinate axes in order to be conforming. This obstacle can be removed by using
the isoparametric formulation presented later. 8 nodes 20 nodes
We emphasize that instead of the derivation prescribed above, it is always
Figure 7.42 Eight- and twenty-node prism element
possible to establish the element shape functions by means of (7.122) in combination
with the C-matrix method.
shown in Figure 7.41 (b). The approximation becomes

i,
7.4 Thre e -dime nsional e lements
(7.124)
Three-dimensional or solid elements can be devised directly from the exposition above.
A detailed discussion of three-dimensional elements is given in Hughes (1987), which is a complete polynomial in accordance with the approximation for the
Gallagher (1975) and Zienkiewicz and Taylor (1989), but as no new conceptual six-node triangle (cf. (7.119)).
mailers are involved, we shall confine ourselves here to some rudimentary Three-dimensional elements in the form of prisms or bricks with element sides
facts. parallel to the coordinate axes are easily derived, for instance from the four-node
The simplest three-dimensional element, the tetrahedral shown in Figure element of Figure 7.27 and the eight-node element of Figure 7.37.These solid elements
7.41(a), has four nodal points corresponding to the approximation are illustrated in Figure 7.42 and the temperature approximation for the eight-node
prism becomes
(7.123)
The generalization of the six-node triangle results in the ten-node tetrahedral as (7.125)
138 Introd uction to the finite element meth od Approximating [unctions - scalar problems 139
whereas the approximation for the twenty-node isoparametric element is given by In addit ion, we found that an element shape function related to a specific nodal point
is zero along element boundaries which do not contain the nodal point in question.
T= (Xl + !X1X + C(3Y + lX4Z + C(s x + C(6 y 2 + 0:7Z2 + O:sxy + Ci. 9 X Z + (X IOYZ
2
Let us now prove another general property for element shape functions. All the
+ Clll x 2 y + C,( 11 X y 2 + ct13 X 2 Z + Ci14 X Z 2 + ctlS y 2z + C(16 y z 2 + Ct: 17 x y z elements considered fulfil the completeness requirement, i.e. the approximation is
always of the form given by (7.2). By putting all terms except a, equal to zero, it
+ a18x' yz + a'9xy'z + a,oxyz' (7 .126) appears that all the elements considered are able to describe a constant temperature
!f x, y or z is assumed to be constant, i.e. if one of the sides of the bricks is considered, within the element. !fthis temperature is called 1; we obtain from (7.127) and (7.128)
it follows that (7.125) and (7.126) reduce to (7.107) and (7.122), respectively. Using that
the isoparametric formulation presented later, these three-dimensional elements may
also be used in a conforming manner even when the element sides are not parallel
with the coordinate axes. i.e.

I ,~,N~ = I I (7.130)

7.5 Summary By inspection we may convince ourselves that this relation holds true for the element
shape functions derived previously.
We have seen that the elementwise approx imation technique offers a systematic and From (7.127) the temperature gradient within the element is given by
unified approach of obtaining an approximation for the unknown function - the
temperature - over the entire body. The possibility for treating bodies with arbitrary
geometries is a hallmark of the FE approximation technique. In addit ion, the freedom
I VT = B'a' (7.131)
to choose elements with different sizes as well as different types greatly enhances the For one-dimensional elements we have
possibilities offered by this approximation technique.
Irrespective of the particular element chosen, the approximation of the unkn own VT = dT B' dN~
= [dN; dN:'J (7.132)
function - the temperature - is carried out as an interpolation between the nodal dx ' dx dx dx
points of the element. This is expressed through
Two-dimensional elements imply that
(7.127) aN~

VTf]
...
where N' is the element shape function matrix and a' contains the temperature at
aT '
B' = [aN'
ax
aN;
ax
aN~
aN']
ax
e: (7.133)
the nodal points . That is, n,
ay ay ay ay
T,
whereas three-dimensional elements yield
T,
(7.128)
aT aN; aN~ aN~

1;"
ax ax ax ax
aT aN; aN~ aN~
where n, denotes the number of nodal points in the element. The element shape VT = - BC= (7.134)
functions, N~, depend on the coordinates. We recall the following fundam ental ay ay ay ay
property for the element shape functions: aT aN~ aN~ aN~
az az az az
I at nodal point i
(7.129)
N~ = { 0 at all other nodal points To formulate the approximation of the temperature over the entire body we
introduce the global shape functions similarly to (7.100) and (7.101). That is, first
140 Introduction to th e finite element meth od Ap p roximatin g func tions - scala r p roblems 14 1

we observe that any position (x, y, z) identifies the associated element a which Two-dimensional elements imply
contains this position. Therefore
oN, .. .
position (x, y, z) => identification of associated element a (7.135)
VT~ [~:JoT '
B= [aN'
ox ox
oN, oN,
aN.]
ox
oN.
(7.142)

For a given global nodal point i the global shape function N, is then defined from oy oy oy oy
the element shape functions N~ according to
whereas three-dimensional elements results in

( )_{Ni(X,
N i X,Y,Z -
o
y, z) if element a contains global nodal point i
.
otherwise
oT
ox
oN,
ox
oN,
ox
e,
ox
(7.136) VT= -oT
oy B=
oN,
oy
oN,
oy
e,
oy (7.143 )

It is of fundamental importance that these definitions imply that a global shape oT oN, oN, en,
function related to a specific nodal point differs from zero only in those elements
which contain the nodal point in question.
oz oz oz oz
Using these preliminary remarks the approximation of the temperature over the A variety of elements has been discussed and it seems natural to ask whether it
entire body is given by is possible to nominate the best element. Unfortunately, this is not possible in general
because the 'best' element often differs from problem to problem. However, we may
T= Na (7.137) state that, as the order of an element increases, the total number of unknowns in a
problem can in general be reduced for a given accuracy of the approximation. This
where N is the global shape function matrix and a contains the temperatures at the is especially true for three-dimensional elements. Use of higher-order elements will
nodal po ints of the entire body. That is, also reduce the amount of data preparation for the F E program, but this advantage

.-m
may be offset by the more involved numerical manipulations. In practice, it is rare
to use elements of higher order than those discussed previously.
Apart from these general remarks, we note that in regions where the gradient
N-[N, N, ... N.J (7.138) varies rapidly, more elements are clearly needed in order to capture this variation
than in regions where the gradient varies slowly . Finally, we recall that in order to
obtain a conforming behaviour of two-dimensional rectangular and three-dimensional
where n denotes the number of nodal points for the entire body. As expression (7.137) brick elements, the element sides must be parallel with the coordinate axes. This
is able to reflect a constant temperature over the entire body in an exact manner, essential drawback can be removed by using the isoparametric formulation presented
we derive in a way similar to (7.130) that in Chapter 19.

I .t,N, = 1 I (7.139)

F ro m (7.137) the temperature gradient is given by

I VT =Ba (7.140)

F or one-dimensional elements we have

VT=dT
dx '
B= [dN, dN, dN.] (7.141)
dx dx dx
8 ----, Ch oice of weight function

method is ob tained by the Galerki n method [Galerkin, 1915) which is also used in
other form ulations than the FE ap proach. More generally, the Galerkin method is
143

Choice of weight function - an exampl e of a so-called weighted residual method, and in order to substantiate our
choice of the Ga lerkin method as an integrated part of the F E formulation, we will
weighted residual methods discuss different weighted residual methods in this chapter.
Gen eral theories for weighted residual methods may be found in Crandall ( 1956),
Finlayson and Scriven (1966), Finlayson (1972) and Zienkiewicz and Taylor (1989).
However, our inten tion is not to provide the reader with comprehensive knowledge
of different weighted residual methods, but simply to obtain sufficient motivation for
our choice of weight function in term s of the Galerkin method. With this limitation
in mind we consider the following one-dimen sional differential equati on a pplica ble
in the range a " x " b:

Let us return to our model problem, namely tha t of heat flow. The strong form of Lu + 9 = 0; a" x " b (8.2)
the problem was first establis hed and, based on this result, we derived the
corresponding weak form ulation. The advantages of the weak for m as compared where u(x) is the unk nown function and g( x) is a known function ; L denotes a
with the strong form have already been discussed and considering, for example, differentia l operator. When L is chosen , it specifies the actua l form of the differential
two-dimensio nal heat flow, the weak form was given by (6.46), which reads equa tion given by (8.2). As an example, L may be given by L = d /dx, which implies
that (8.2) takes the form du / dx + 9 = O. As another example, for L = d 2 /d x 2 + I,

fA
(Vv)TtDVT dA = -f ~
vht d!t' - f
~
vq, t d!t' + f
A
vQt dA (8.1)
(8.2) implies the differential equa tion d 2u / d x 2 + u + 9 = O. It appears that by using
the format (8.2), it is possible to write quite genera l one-dimensional differen tial
equations in a compact fashion. In order to simplify the discussion, the boundary
The FE method is an approximate met hod for solving arb itrary differential equations. conditions for (8.2) are assumed to be given by
Therefore, an appro ximat ion is required for the unknown function , which in the
present case is the temperature T. In the previous chapter, a detailed discussion was uta) = u.; u(b) = Ub (8.3)
devoted to the specific type of approximation for the tempera ture that is ad opt ed in where u. and Ub are known quantities. The specification of two bo undary conditions
the FE method. This approximation was found to be characterized as an elementwise implies that the differential equation (8.2) is of second order.
approximation over the entire body. Referring to (8.1), it appears that the only topic It is characteri stic that the differential equation (8.2) with the bounda ry
which remains to be discussed is the choice of the weight function v. We recall that conditions (8.3) can only be solved analytic ally in terms of an exact closed-form
in the weak formulation this weight function is an arbitrary function. At this point solution for certain simple expressions for the differential opera tor L and the function
we can therefore summ arize the basic steps in the FE formulation as follows: g. What we are looking for are meth ods which enable us to solve (8.2) subject to
(8.3) for arbitrary expressions for Land g. This gain in generality is paid for in the
sense that our solution takes the form of an approximate solution.
Basic steps in the FE formulation For this purpose, we fi rst make a slight reformulation of (8.2). We may multi ply
I. Establish the strong formula tion of the problem. (8.2) by an arbitrary function v(x), the weight function, to obtain
2. Obtain the weak form of the problem.
v(Lu + g) = 0
3. Make an elementwise approximation over the enti re body of the
unknown function, i.e. the temperature T. and we may even integrate this expression over the region of interest, i.e,
4. Choose the weight function v.

f V(LU + g) dx = 0 (8.4)
The previou s chapters have been devoted to steps 1-3 and in the present chapter
we shall establish the choice made for the weight function v, i.e. step 4. It is obvious that (8.2) implies (8.4) and as v denotes any function, it is easily sho wn
It turns out that a particularly suitable choice of the weight function in the FE tha t (8.4) implies (8.2), i.e. the two forms are equivalent. To show that (8.4) implies

142
144 In troduction to the titiite elem ent meth od Choice of w eight f unction 145

(8.2), choose the arbitrary weight function vas v = Lu + g, i.e. (8.4) becomes surprising that u' PP will not , in general, satisfy the equation exactly. That is, we obtain
Lu' PP + g = e (8.9)
f(LU+9)'dX=0
where e(x) is a measure for the error. This measure is termed the residual. With (8.9),
(8.8) can be written as
which can only be true if (8.2) is fulfilled. We have thus proved that (8.2) and (8.4)
are equivalent. It is apparent that this discussion has much in common with that of
the weak formulation for one-dimensional problems (see Chapter 4). However, it is (8.10)
important to realize that the formulation (8.4) is not a weak formulation. In order
to achieve a weak formulation an integration by parts is necessary, whereby the order
of differentiation of the unknown function is decreased at the expense of the weight Since the residual e depends through (8.9) and (8.6) on the unknown parameters
function being differentiated (cf. for instance (4.36 )). given by a, (8.10) is an expression which depends on the parameter a, .. . a. The
With the se remarks in mind, we return to the formulation given by (8.4) . As we o bjective is to choose the weight function v so that these parameters can be
seek a numerical solution, an approximation must be chosen for the unknown determined, which, in turn, implie s that the approximate solution (8.6) is known.
function u. Quite generally, we make the following assumption, which is assumed to Equation (8.10) may be interpreted so that the residual, e(x ), is given a certain
fulfil the boundary conditions (8.3): weight, v( x ), and the integral of this weighted residual, v(x)e(x), over the region of
interest is required to be zero. The various weighted residual methods differ in how
(8.5) the weight function v is chosen. Obviously, the specific choice of the weight function
Here a, ... a, are unknown parameters and 1jJ , 1jJ, are functions of x, which we influences the values of the parameters a, ... a" which are to be determined. Another
specify in advance. These functions are called trial or basis functions. Once the feature also related to (8.10) is that of orthogonality. The orthogonality condition for
parameters a 1 ... a, are known, our approximate solution is given by (8.5), i.e. the two column vectors a and b is expressed through a Tb = O. For functions, one can
task is to determine the parameters a, . .. a,. The trial functions may be chosen also speak of orthogonality and the function v(x) is said to be orthogonal to the
arbitrarily, but it is obvious that some knowledge of the physical problem facilitates function e(x) in the interval a 5: x 5: b if (8.10) holds.
the choice of suitable trial functions. With reference to Chapter 7, we may even take Equation (8.10) serves as a vehicle to determine the unknowns a, .. . a" i.e. a.
the approximation (8.5) as an elementwise approximation. In that case a 1 . . a, To see how this can be accomplished, we write the arbitrary weight function quite
would be the values of u at the nodal points and the trial functiens would become generally as
the global shape functions. However, (8.5) may be taken as any approximation and v = VICI + V,c, + ... + v"c, (8.11)
this is the reason why IjJ 1 .. . 1jJ, are termed trial functions.
The approximation (8.5) may be written as where Vi . .. t;. are known functions of X, which we specify in advance, and c 1 . .. en
are certain parameters. We note that the number of terms in (8.11) is the same as
the number of terms in (8.5). With the definitions
I u' PP = ljJa (8.6)

where
v -r ' ... Yo] <-f::] (8.12)

(8.7)
/
(8.11) can be written as

Iv = Vo I (8.13)
We may substitute u in (8.4) by u' PP and require that

f
As the weight function is arbitrar y and as V is known, we conclude that the
parameters C 1 . CIl , i.e. c, are arbitrary. Moreover, as v is one number and as the
v(L u' PP + g) dx = 0 (8.8)
transpose of a number is equal to the number itself, i.e. v = VT, we may write (8.13) as
Likewise in the differential equation (8.2), u may be substituted by u' PP and it is not v= CTV T (8.14)
146 Introductio n to th e finite el emen t method Choice of w eight fun cti on 147

Inserting (8.14) into (8.10) and o bserving that eT does not depend on the coordinate sho wing that L (l/J ) is a matrix with dimen sion I x n. Use of (8.17) in (8.15 ) imp lies
x, we obtain that
(fVTL(IjI)dx)a = - fVTgdX (8.19)
cTf VTe dx = 0
where a is considered to be independent of x. With the definiti ons
As this expression should hold for arbitrary cT-matrices, it is concluded that
K=fVTL(IjI)dX; f =-fVTgdX (8.20)
fVTedx=o (8.15)
wher e both K and f are independent of a, (8.19) can be written as the system of linear
equatio ns
It is important to realize that as the column matrix VT has the dimensions n x I (cr.
(8.12)), expression (8.15) comprises, in fact, n equations. That is, (8.15) may be written
as
I Ka = f I (8.21)

f V,edx = 0
As VT has the dimen sion n x I and L (IjI ) the dimension I x n (cf. (8.18 )), the
expressions for K and f may be written as

fv,edx=o I V,L( l/J , )dx f V, L(l/J, ) dx I V,L(IjI, )dx I V,g dx

"""(8.16)

K= f V,L(l/JI) dx I V,L(l/J,)dx f V,L(l/J,) dx


; f=-
I V,gdx

f v"e dx = 0
f v"L( l/J ,) dx f v"L ( IjI, ) dx f v"L ( l/J,) dx f v"g dx
The derivation of (8.16) is based on (8.10) and (8.13) and uses the fact that the
weight function v is arbitrary, i.e. c is also arbitrary. Alternatively, in (8.10) one may
(8.22)
directly choose n arbitrary weight functions v and first choose v = V" then v = V,
and so on . Clearly, the result of this approach is again given by (8.16). The approach showing that K is a sq uare matrix with dimension n x n. That is, (8.21) consists of
based on (8.13) is preferred here, as it facilitates a short notation, especially in later n linear equations from which the n unknowns a l . .. a" . i.e. a. can be determined.
FE formulations, but one should be aware of the equivalence of the two approaches When a has been obtained from (8.21), (8.6) provides the required approximate
both leading to (8.16). solutio n.
Returning to (8.16), the essential point is that the residual e depends on the n The procedure described above applies to all weighted residual methods, and a
unknowns a, .. . a, (cf. (8.9) and (8.6)). Therefore (8.16) serves as a system of variety of different specific weighted residual methods is obtained depending on our
equations to determine these n unknowns. To see this more clearly, let us insert (8.6) choice of the weight function v, i.e. our choice for V (cr. (8.13)). Some frequently used
into (8.9) to provide weighted residual methods will be described below.

e= L(IjIa) + 9
The a-matrix does not depend on the coordinate x, i.e. 8.1 Point collocation method
e= L(IjI)a + 9 (8.17) In thi s meth od, the weight function v is cho sen based on Dirac's della fun ction
Following Chapter 2, different iation of a matrix means that each component is b(x - x.). This fun ction is defined as
differentiated, i.e. we obtain from (8.7) that
(8.18)
'( _x-
u X
)_{oo if x = Xi (8.23 )
L(IjI) = [L(l/J.J L(l/J,) . .. L(l/J,)] I 0 otherwise
148 In trod uc tion to th e fin ite elemen t method Choice of w eig ht function 149

5 elxl

_ _ _ _.L.. ~ < p . X

Figure 8.2 Point collocation method


Figure 8.1 Dirac's delta function

and obtain from (8.22) in a manner anal ogou s to (8.28)

f~~ b(x - Xi) dx =1 (8.24) L ( !/J ,( x, )) L(!/J , (x ,) L(nx t )) g( x, )


L ( !/J ,(x,)) L(!/J,( x,) L(!/J.(x,)) g(x,)
where X, is a given fixed value. Alternatively, due to (8.23), we may write (8.24) as K= f= - (8.30)

r "'( X - Xi) dx = I (8. ~) L (!/J,(x .)) L(!/J, (x.))


The differentiation indicated by the operator L, for instance L(!/J,(x,), should be
g( x.)

where x ~ and x : denote x-values slightly larger than and smaller than x, understood as follows: first the differentiation of the function !/J,(x) is carried out
respectiv~IY. Using'these definitions, Dirac's delta function is illustrated in Figure 8.1. and then the specific x-value, x = x 2 , is inserted.
In the point collocation method, the weight function v in (8.13) is chosen such that
v= [b(x - x , ) b(x - x,) .. . b( x - x.)] (8.26)
The fixed points x, . . . x, are chosen arbitrarily within the region a " x " b and are 8.2 Subdomain collocation me thod
called collocation points. In order to illustrate the consequence of this choice of weight
function, we evaluate (8.16), which becomes In the point collocation method n points were chosen. In the subdomain collocation

f V,e(x) dx = f b(x - x ,)e(x ) dx = 0; i = I, .. . , n


,
(8.27)
method , however, the region of interest is divided into n sub regions. Each of these
subregions is given by X i ~ X S X i + 1 where 1 ~ i :5: n and both X i and X i + 1 are located
in the region a"X" b. The mat rix V (cr. (8.12), which defines the weight function,
As Dirac's delta function is zero unless x = x, (see (8.23)) we have is chosen such that

bb(x - f X."b(x - f X."b(x -


f
V = {I if X I S x S Xi + 1 i = I, . . " n (8.31)
x,je(x) dx = x ,)e(x ) dx = e(x,j Xi ) dx = e(x i) 0 otherwise
41 X. Xi
(8.28) as illustrated in Figure 8.3.
where (8.25) has been used. From (8.27) and (8.28) we obtain Th is means that (8.16) becomes

f V, e(x) dx = e(x i) = 0; i = I, .. . , n (8.29)


f b

V,e(x ) dx = f"' x,
e(x) dx = 0; i = 1, .. . n (8.32)

That is, the point collocation method forces the residual e(x ) to be zero at the That is, the subd omain collocat ion method forces the average of the residual over
collocation points, as illustrated in Figure 8.2. each subdomain to be equal to zero, as illustrated in Figure 8.4.
In order to determine the specific form of the system of equations (8.21), we With (8.31) and using (8.22), the coefficient matrix K and the right-hand side f
150 In trod uction 10 the fin ite elemen t method Ch oi ce of we ight function 151

Vi lxl 8.3 Least-squares method


Su bdamain i
I I Due to (8.17) we ob serve that the residual e depends on both x and the parameters
a 1 a", i.e.

b
e = e(x, a t . . . all) (8.34)
a
Xi In the least-squares me/hod. a typical component of the matrix V given by (8.12) is
chosen as
Figure 8.3 Weight function in subdomain collocation method
oe
Vj = - ; i= 1 ... ,n (8.35)
oa j

which implies that (8. 16) takes the form


. l xl

Subdomoin i \ f
' oe e dx
Q OOj
= 0 ; i = I. . . n (8.36)

In order to evaluate this choice. consider the following quantity I :


a b

(8.37)

Figure 8.4 Subdomain collocation method


Since we integrate with respect to x, it follows that the quantity f depends on the
parameters a,. . . a, only. i.e. 1= I(a, . .. a,). From (8.37) we conclude that
of (8.21) take the form s
oedx;
!'!'-=2f' e 1= 1. .. n (8.38)

I X> L(l/! ,)dX


x,
oa, Q oaj

An evalu ation of (8.37) shows that I is the square of the error. i.e, the residual .
K= [ L(l/!,) dX integrated over the region of interest. A comparison of (8.36) and (8.38) reveals that
. . 01
- = 0; i = l, . .. ,n (8.39)
r 'L(l/! ,) dx r 'L(l/!z) dx r 'L(l/!,) dx oa,

(8.33) i.e. our choice of weight function given by (8.35) implies that I is stationary. As the
residual e. and thus also the quantity I . can be made arbitrarily large. we conclude
X>
I 9 dx
that the stationarity of I as expressed through (8.39) is a minimum. Consequently.
the choice of weight function given by (8.35) implies that the square of the error is

I
x~,
a minimum ; hence the terminology of the least-squares method.
9 dx To determine specifically the coefficient matrix K and the right-hand side f of
f= - x,
the system of equations (8.21) when the least-squares method is adopted. we first
insert (8.17) into (8.35) to obtain

v, = L(l/! ,); i = I . .. n (8.40)


152 Int rodu ction to the finite elemen t m eth od Ch oice of w eight fu n cti on 153
Th at is. (8.22) yields (8.22) to provide

f L (ljI,)L( ljI,) dx f L (ljI,)L( ljI, ) dx f L (ljI ,)L(ljI. ) dx


f : ljI,L( ljIl ) dX f: ljI,L (ljI,) dX f ljl, L (ljI. ) dX

K= f L ( ljI, )L( ljI,)dx f L ( ljI, )L( ljI, ) dx f L ( ljI, )L( ljI.) dx


K = f ljI, L( ljIl ) dx f ljI, L(ljI,) dx f ljI, L(ljI .) dx

f L(ljI.)L( ljI.) dx

(8.41)
(8.45)

fL( ljI, )9dX

f= -
f L ( ljI, )g dx
.\
f= -

f L ( ljI.)g dx

which shows that the coefficient matrix K is symmetric.

8.5 Example
8.4 The Galerkin method
To illustr ate the weight residual methods described above. we consider the differential
In the Galerkin method ( 1915) the components of the matrix V given by (8.12) are equation
chosen in accordance with d' u
-dxl + u + x = O 0 ,;; x ,;; I

(8.46)
II; = ljI, I j = I , .. .. n (8.42)
In the notation of (8.2) it follows that
i.e. the components II; are equ al to the trial functions ljI,. Loosely speak ing we express
d'
this by saying that L = - 2 +I g= x (8.47)
dx '

weight function s = trial functions (8.43) The bounda ry conditions are given as

With (8.42). (8.16) yields u(O) = u(J ) = 0 (8.48)

f ljI,edx = 0; i = I, . . . n (8.44)
As easily proved by inspection, the exact solution to (8.46) and (8.48) is

u= -
sinx
- - x (8.49)
Referring to the discussion of (8.10). it follows that in the Galerkin method the trial sin 1
function s are orthogonal to the residual. To determine explicitly the coefficient ma trix In order to obtain an expre ssion for the approximation to u in the form of (8.5).
K and the right-hand side f of the system of equat ions (8.21), (8.42) is inserted into we may explore a num ber of possibilities. One possibility is given by the trigonometric
\
154 In troducti on to the finite elem ent m eth od Cho i ce of w eigh t func ti on 155

series Next, consider the [east-squares method, where K and f are given by (8.41) which
implies that
u' PP = bo + (a, sin ex + b, cos cx) + (a, sin 2cx + b, cos 2cx)
+ ... +(a. sin ncx + b. cos ncx) (8.50)
K = [{ L(!/J ,)L(!/J ,) dx J = [{ (_n 2 sin nx + sin nx)2 dx J = [~ (l - n 2)2 J
This appro ximation is required to fulfil the boundary conditions (8.48) and it appears
that u'''(O) = 0 is fulfilled if
(8.5 1)
2
f= - [ { L (!/J , )g dX = _ [ { (_ n sin nx+ sin nX)XdxJ = -[~ (l - n')J
bo = b , = b, = ... = b. = 0
Mo reover, the con dition u''' ( 1) = 0 is fulfilled if (8.60)

(8.52) and where it was used that


c=n
With (8.51) and (8.52), (8.50) red uces to
uapp
= at sin nx + a2 sin 2nx + ... + an sin nnx
\
(8.53) f . 2
sm x dx
X
= - -
2
1
-
4
sin Zx:
'
f x sin x dx = sin x - x cos x

As a very simple approximation, we take just one term in this series, i.e. From (8.21) it follows that
u3 PP
= at sin nx (8.54) 1 I 2
- (1 - n' f a , = - - (1 -n') i.e. a, = '" 0.0718 (8.61)
and the pr ob lem is no w to determine the parameter a, . With reference to (8.6) we have 2 n n(n' - 1)
t/1 = [t/1,] = [ sin nx] (8.55)
Finally, from (8.45) the Galerkin method implies that
Consider first the point collocation metho d. From (8.47) and (8.30) we obtain

K = [L( t/1,(x,))] = [ _n 2 sin nx, + sin nx ,]


(8.56)
K = [{!/J ,L(!/J,) dx = [{Sin nx( - n' sin nx + sin nx) dx J = G (1 - n 2)J
f= - [g(x , )] = -[x,]
f= - [ { !/J , g dXJ = - [ { x sin nx dxJ= -GJ
Choose the collocation point, for instance, as the midpoint of the interval, i.e. x, = t.
F ro m (8.21) it then follows that

1
(-n z + 1)a,=- -1 .i.e. a, , '" 0.0564 (8.57) ul xl
2 2(n - 1)
0.10 Subdomoin coflocution
Fo r the subdomain collocation method, K and f are given by (8.33). In the present
'<, EXQct
example, only one sub domain is involved and this is convenientl y chosen as the ent ire
region of inte rest. This results in
0.0 8 y -"
/
/ ~. ...-' ' .
-, Gal erk in and l east squares
0.0 6

0.04 / /;/-----<>"
. ,," "', .
(8.58) 11/~ ,,~~
/~/ Point colloca tion ,,~.
0.02
a '~
Use of these exp ressions in (8.21) results in "-------<-- --+-- - -+-----4-_x
o 025 0.5 075 1.0
2 2 1 n
(l - n ) - a, = - - i.e. a, = 2 '" 0.0885 (8.59)
n 2 4( n - 1) Figure 8.5 Comparison of different weighted residual methods
156 In troduction to the fi ni te elem en t m ethod

That is, (8.21) yields 9


1 2
- (I - n 2
2
)a! =1
- -
n
I.e. a! = 2
n(n - 1)
'" 0.0718 (8.62) FE formulation of
It appears that the Galerkin and least-squares methods in this example provide
the same results. The results of the four weighted residual methods are compared
one-dimensional heat flow
with the exact solution in Figure 8.5 and the superiority of the Galerkin and
least-squares methods is obvious for the example considered.

\ 8.6 Concluding remarks


In the previous chapters we have discussed strong and weak forms of the heat flow
The point collocation method requires that the residual is equal to zero at the
problem, the elementwise approximation technique used in the FE formulation for
collocation points. This condition seems intuitively less exacting than that used in
the unknown function and various choices for the weight function. It was emphasized
the subdomain method where the average of the residual over each subdomain is
that the FE approach is based on the choice of weight function in accordance with
required to be zero. Thus one may expect the subdomain method to be more accurate
the Galerkin method . The basic steps in the FE method can therefore be summarized
than the point collocation method. The least-squares method and the Galerkin
as follows:
method turn out to be very efficient. The least-squares method always results in a
symmetric coefficient matrix (cr. (8.41)), which is clearly an advantage in numerical
calculations. In general, the coefficient matrix (8.45) in the Galerkin method is not
symmetric, so that the advantage of using this method in the FE formulation may Basic steps in the FE formulation
not seem obvious at this point. However, the weighted residual methods discussed 1. Establish the strong formulation of the problem.
L in this chapter are based on the integral formulation (8.4) of the differential equation 2. Obtain the weak form of the problem.
I
(8.2). We have previously stressed that (8.4) is not a weak formulation. In a weak 3. Make an elementwise approximation over the entire body of the
I
formulation, an integration by parts is performed whereby the order of differentiation unknown function.
of the unknown function is decreased at the expense of the weight function being 4. Choose the weight function in accordance with the Galerkin method.
differentiated (see for instance (4.36)). As shown in the next chapter, this implies that
when the Galerkin method is used in combination with the weak formulation, a In order to illustrate these steps we shall consider the FE formulation of
symmetric coefficient matrix arises. This aspect as well as the applicability of this one-dimensional heat flow. Even though this constitutes a simple physical problem,
method to any differential equation suggest its fundamental importance in FE it will turn out that its FE formulation contains all the fundamental features of any
formulations. FE approach. Therefore, the reader is urged to read the present chapter especially
carefully. If the procedures described here are fully comprehended, it will be an easy
task to follow the remaining exposition which discusses the FE formulation of more
complicated physical problems , but where no essentially new FE concepts will be
introduced.

9.1 Global FE formulation


Consider one-dimensional heat flow in the fin shown in Figure 9.1. The strong form
of a quite similar problem was given by (4.8)-(4.10) and with the position of the fin

157
158 Introduction to th e fi nite element method FE formulation of I -D heat flo w 159

T The approximation for the temperature T is now introduced. Following the


summary of Chapter 7, the approximation over the entire region is in general written as

T=Na (9.4)

where N is the global shape function matrix and a contains the temperatures at the
Alx l
nodal points in the entire body (cr. (7.137) and (7.138)). That is, we have

T1
b
T,
N= [N , N, ... N.] ; a= (9.5)
LJ..-L--'--'---l...UU-l..-lJ Q I xl

Figure 9.1 Heat conduction in one-dimensional fin


where n denotes the number of nodal points for the entire body and N, = N,(x). As
as shown in the figure, the differential equation becomes a does not depend on x, (9.4) implies that

-d ( Ak -dT)
dx dx
+ Q = 0; a: x : b (9.1)
I dT = Ba I where ~ (9.6)
~
dx
where it is recalled that A(x) is the cross-sectional area of the fin, k(x) is the thermal
i.e,
conductivity and Q(x) is the heat supply per unit time and per unit length of the fin.
Q may be transferred to the fin across its outer surface or created internally, for
instance by electric heating. In order to solve (9.1), boundary conditions must be
B [dN dN, ... dN.]
= 1
(9.7)
dx dx dx
specified, for instance in the form given by (4.9) and (4.10), but to be able to deal
with completely arbitrary boundary conditions, we shall leave the boundary conditions in accordance with (7.140) and (7.141). Inserting (9.6) in (9.3) we obtain
unspecified until later.
To obtain the weak form of (9.1) we multiply by the arbitrary weight function
v(x) and integrate over the region of interest, i.e.
(f: ~: AkB dx)a = - [vAq]: + f vQ dx (9.8)

f {d~ ( Ak ::) +Q }x 0 =
It is important that the approximation for dT/dx as given by (9.6) is not inserted
into the boundary terms on the right-hand side of (9.8), which include the flux
q = - k d T / dx. The reason is that the boundary conditions either specify the flux or
The first term is integrated by parts in accordance with (4.33) to provide the temperature itself at the boundary and there is no reason to make an
approximation for matters that we know beforehand. We shall return to this topic
b dv Ak dT dx = [VAk dT]b + fb vQ dx (9.2) later on.
f adx dx dx.L a
The last step is the choice of weight function v. In accordance with the Galerkin
and using Fourier's law q = -k dT/dx, we get method, we choose the weight functions to be equal to the trial functions (cf. (8.42)
and (8.43)). In the present case the trial functions are the shape functions; that is,
Ia
-dv AkdT
dx
- dx = - [ vAq ]:
dx
+ I vQdx (9.3)
(8.13) yields

I I
II

v =Nc (9.9)
It appears that the integration by parts implies that the order of differentiation for
T has decreased at the expense of the weight function v being differentiated. Since v is an arbitrary function, the matrix c is arbitrary. Moreover, as v is one
160 Introd uction to the finite element method FE fo rmulation of I -D heat flow 161
number, we have v = VT , i.e. (9.9) can be written as and defining the force vector f by
v = cTN T (9.10)
(9.15)
which implies that
(9.14 ) can be written as
dN T
dv = cTBT where B T = __ (9.11)
dx dx I Ka = ( I (9.16)
Inserting (9.10) and (9.11) into (9.8) and using the fact that cT is independent of x
results in The temperatures at the nodal points given by a are obtained by solving this system
of linear equations. When a is known, the temperature T at an arbitrary point in the
eT[ U>T AkBdx)a + [NT Aq]~ - f NTQdx ] = 0
fin is given by (9.4) and the temperature gradient at an arbitrary point is given by
(9.6). From this temperature gradient, the flux q at any location in the fin is obtained
from Fourier's law (4.3). Therefore, when a has been determined from (9.16) all
As this expression should hold for arbitrary eT-matrices, it is concluded that
quantities of interest can be derived. Let us now evaluate the stiffness matrix K and
(f BTAkBdx)a = _[NT Aq]~ + f N TQ dx (9.12)
the force vector f in more detail.

From (9.13) and (2.30) it follows that


which is the required FE formulation. Here, we have derived (9.12) by means of(9.8)
and (9.9) and using the fact that v is arbitrary, and thus e is also arbitrary. Alternatively ,
in (9.8) we may directly choose n arbitrary weight functions v and first choose v = N "
KT = (f BTAkB dX)T = f BTAkB dx = K (9.17)

then v = N 2 and so on. Clearly, the result of this approach is again given by (9.12). i.e. the stiffness matrix K is symmetric. It is obvious that this symmet ry hinges on
This discussion is similar to that relating to (8.16). Here, we prefer the approach the use of the Galerkin method. If instead of this method (9.9) we had chosen v = Vc,
given by (9.9) as it facilitates a clear matrix formulation, especially in the more where Y N, then the corresponding stiffness matrix would have taken the form
complicated FE formulations that will be encountered later. However, one should
be aware of the equivalence of the two approaches. b dyT dy T
K= - AkB dx where _ BT
I I
In order to write (9.12) in a more compact fashion, we define the following J
a dx dx
matrices:
which implies that the symmetry of K would have been spoiled. This symmetry, the
applicability of the Galerkin method to any differential equation and the realistic
K= f BTAkBdx predictions provided by this method are the primary reasons for its adoption in the
FE formulation.
To evaluate the character of the stiffness mat rix K further we consider its
r, = - [N' Aq]~ (9.13)
components. From (9.13) and (9.7) it follows that
(, = f: NTQdX

Referring to (9.7) it is obvious that K is a square matrix with dimension n x n ; it is


f '-
dN-l A kdN'd
"dx
dN
-
dx
x

called the stiffness matrix. Likewise, both ( b and f. have the dimension n x I, and
they are called the boundary vector and load vector, respectively, since (b refers to
K= f b
- -2 A kdN'd
"dx
-
dx
x

conditions at the boundary, whereas f. considers the effect of the ' loading' Q. With
(9.13), (9.12) takes the form dN , kdN2d dN
- , A kdN'd
b b

J-- A - - x
"dx dx J "dx
-
dx
x

Ka = f b + (, I (9.14)
(9.18)
162 In tro d uc tion to th e tinite elem en t me tho d
FE form ulation of I -D heat flow 163
oj

Nj

.:>
'
Figure 9.3 Fin modelled by five linear elements
/
'

< /
Q
,

= x
stiffness matrix becomes banded with non-zero components clustered about the
diagonal of K. This has important numerical advantages as discussed in Chapter I L
We emphasize that the banded structure of the stiffness matrix K is a result of the
bJ
fundamental property that a global shape function differs from zero only in those
elements which contain the nodal point in question.
As an illustration of these features, consider the fin shown in Figure 9.3. The fin
is modelled by five elements and for convenience we make use of the simple linear
element of Chapter 7. With the adopted nodal numbering it follows that the stiffness
o o matrix K takes the following form :

2 3 4 5 6
I x x 0 0 0 0
Figure 9.2 Variation of global shape functions N; and N j for linear elements 2 x x x 0 0 0
3 0 x x x 0 0
K=4
The symmetry of K is also obvious from this expression. Moreover, consider a 0 0 x x x 0
component K ij of K, which can be written as 5 0 0 0 x x x
' dN s ,
K =
IJ
f -'Ak - Jdx
lldx dx
(9.19)
6 0 0 0 0 x x

In order to evaluate the force vector f given by (9.15), let us first consider the
The approximation of the temperature according to (9.4) is an elementwise
components of the boundary vector f b and the load vector f,. From (9.13) we'get
approximation, and in orde r to evaluate (9.19) we recall that one global shape funct ion
is related to each nodal point. The variation of the global shap e functions N, and
N is shown in Figure 9.2, where, for simplicity, it has been assumed that the simple
linear finite element has been used, i.e, N, and N j vary linearly (cf Figure 7.8).
[N ,AqJ: I N,Qdx

I i
However, the important point, which is common to all types of finite elements, is
tha t a global shape function N, differs from zero only in those elements which contain
the nodal point i. As shown in Figure 9.2, two principally different possibilities exist
fb = -
[N ,AqJ:
f, =
I N,Qdx
(9.20)
for the variation of the global shape functions N, and N j . In Figure 9.2(a), no elements
contain both nodal points i and j, whereas in Figure 92(b) one element contains
both nodal points i andj In Figure 9.2(a), N , is zero when N, is different from zero
and vice versa, i.e. when evaluating (9.19) the component K' j is zero. Only when the
[N,AqJ: f N,Q dx

two global shape functions N, and N, differ from zero at some common regions will
A typical term!" of the boundary vector f b is given by
K. . be different from zero. Such an example is shown in Figure 9.2(b) where N , and
N': only differ from zero simultaneously in the element given by nodal points i and fbi =- [ N ,AqJ: =- (N,Aq)x =b + (N,Aq)x=" (9.21)
j. In conclusion, the component K'j of the stiffness matrix K is zero unless both nodal
The positions x = a and x = b, i.e. the ends of the fin, clearly correspond to the
poin ts i and j are present in at least one element. This conclusio~ i.s si~ilar to the
position of two nodal points . For the sake of illustration assume that the nodal points
observation made in Chapter 3 (cf (3.70) and (3.71). Moreover, It implies tha t the
1 and n are locat ed at x = a and x = b, respectively. With the properties of a global
164 Introduction to the fini te element m eth od FE form ulation of l -D h eat fl o w 165
sha pe function in mind it follow s that 1 2 3 4
C

{ N 1 (x= a)= 1 if i = I (9.22) X : 2


N,(x = a) = 0
if i ~ I Figure 9.5 Three linear elements of equal length
and

N,(x = b) = 0
t.(X = b) = I if i = n
if i ~ n
(9.23)
where g, and h b ar e known qu antities. These boundar y cond itions corresp ond to a
presc ribed temperature at the left end of the fin and a prescribed flux q = - k dT/dx
at the right end . More over, let

With (9.20)-( 9.23), the boundary vecto r becomes g,= O'C; hb = 15 J/ m's (9.28)

( A q). ~,
Assume that we divide the fin into three elements of equal length and also ass ume
for con venience that we ado pt the simple linear element of Chapter 7. Th e nodal
o poin ts are indicated in Figure 9.5. The global shap e functions together with their
f, = (9.24~ derivatives are show n in Figure 9.6.
Let us first determ ine the stiffness matri x K. From (9.19) it follows tha t
o
- (Aq). ~ b
bd N , dN, f'd N , dN ,
K II = - Ak - dx = - Ak - dx
f , dx dx , dx dx
where q is the flux given by Fourier's law (4.3).

9.1.1 Examp l e 1
= J: (-0.5)50( - 0.5) dx = 25

where N " i.e. dN , /d x, only differs from zero in the interval 2 ,:; x s; 4. Likewise
Having discussed the FE formulation of one-dimensional heat flow in general terms,
it is timely to investigate the solution of a specific example. Fo r simplicity, we assume - -, A k - - dx -_ f' -dN-, A k ~-
K 12 -- f bdN s, dN, dx
that the fin sho wn in Figure 9.4 has a constant cross -sectional ar ea A, a constant , dx dx , dx dx
thermal cond uctivity k and a constant hea t supply Q.
From F igur e 9.4 and (9.1) the differential equation is given by = J: ( -0.5)50(0.5) dx = - 25

dx
dT) +Q =O ;
-d ( Ak -
dx
(9.25) Here, we used the fact that N , and N , only differ from zero simultaneously in the
range 2 ,:; x ,:; 4. Since N 1 and N, have no common region where they differ from
zero, we ob tain
where
A =lOm' ; k = 5Jr Cm s; Q =IOOJ / sm ; a =2m ; b =8m (9.26) bdN , dN ,
K 13 = - Ak - dx = 0
f , dx dx
The boundary conditions are assumed to be of the form
and likewise
T (x = a) = g, ; q(x = b) = hb (9.27)
bdN , dN.
K 14 = - - Ak -- dx =0
f II dx dx
~ A' 1 0 . k ' 5
As K is symmetric (cf. (9.17)), we next con sider com ponent K" , i.e.
x
b dN , dN , f6 dN, dN , f'
K" = - Ak - dx = - Ak - dx = (0.5)50(0.5) dx

r
f , dx dx , dx dx z

+ (-0.5)50( - 0.5) dx = 50
Figure 9.4 Problem definition
166 Introduction to the finite elem en t method
FE formulati on of I-D heat fl ow
167
~6 ~8
~6 ~8

K" = f' dN , Ak dN , dx = f6(0.5) 10 x 5(0.5) dx


a dx dx 4

3
o
t;

x
0.5

- 0.5
+ J: ( -0.5) 10 x 5( - 0.5) dx = 50

K ,. =f' dN' Ak dN' dx= J ' ( - 0.5)10 X 5(0.5) dx =- 25


dNZ/d x dx
41 dx 6

K = f' dN. Ak dN. dx = J' (0.5)10 X 5(0.5) dx = 25


05
a dx dx 6
t:
o.-- x Collecting these results together, we ob tain the stiffness matrix K :
-05
25 - 25 0 0
- 25 50 - 25 0
N, d NJ'd x K= (9.29)
0 - 25 50 - 25

1 t 2~'O--- X
0 0
0.5 2 3
x
0 0 - 25 25
T he banded structure of the stiffness mat rix is obvious. Next, consider the boundary
- 0.5 vector fb Fro m (9.24) and (9.26)- (9.28) we obtai n

(Aq)x IOq( x = 2)
N, dN, ldx
0

,t
0
fb = = (9.30)
0 0
,
0
2
0
3L
e>--X
05 2 3 - (Aq) x=' - 150
-0 5 The load vecto r f, is given by (9.20), and from (9.26) and Figure 9.6 we find

Figure 9.6 Global shape functions and their derivatives r N ,Q dx J: N , dx 100

Mo reover, proceedin g in the same mann er we get I N 2Qd x f: N 2 dx + f : N 2 dx 200

K
2'
= f' dN 2 Ak dN , dx = f6dN 2
Ak dN , dx
f, =
I N , Q dx
=Q
f:N, dX + f:N , dX 200
(9.31)

ro dx dx dx dx

= (- 0.5)50(0.5) dx = - 25 I N . Q dx f : N . dx 100

and where aga in we utilized the fact that a global shape function N, differs from zero
s, only in those elements which contain the nodal point i.
K
2.
=
f ' dN 2
x
-
d
Ak -
dx
dx =
0
From (9.29)-(9.31), the final equation system (9.16) can be esta blished and since
the prescribed temperature T, is T, = T (x = a ) = g. = 0 (cf. (9.27) and (9.28 , th is
168 Introd uction to the fi ni te elem ent m eth od FE formulati on of I-D heat flow 169

implies that T

r:~'
25 - 25 0 0 0 100
20 e xa c t ~ .>
- 25 50 - 25 0 T2 200
+ (9.32)
0 -25 50 - 25 T, 200 Y
0 0 - 25 25 T. - 150 100
1
It can be noted that the largest computational effort is involved in determining 10 'I
the stiffness matrix K and that other flux boundary conditions and load s Q only '1 FE
solu tion
influence the right-hand side of (9.32). Considering the first row and the last three 'I
rows of (9.32) separately , we obtain 'I
V
- 25T, = 10q(x = 2) + 100 (9\ )
2
and
dT I dx
- 25
[ 50
O][T']
- 25 T, = [200]
- 2~ 50 200 (9.34)
10
-25 25 T. - 50 / exoct
By using the Gauss elimination technique presented in Chapter 2, it can easily be
shown that the solution of (9.34) becomes
5
solu ti on
(9.35)

With the known T,- value, the flux q at x = a = 2 is determined from (9.33) to provide
-1
qX.' =_(kdT
dx
)
.% = 2
= - 45 J/ m's (9.36)

i.e. heat must be extracted at the left end of the fin to keep the temperatur e at the
required value. Figure 9.7 Comparison of exact and FE solution
To evaluate this solution, we make comparisons with the exact solution. It can
easily be checked that by Hughes ( 1987, p. 27)). Besides, a fair agreement exists between the exact and the
FE solution and we may draw the general conclusion that , even though the
T = _x 2 + 13x - 22 (9.37)
temperatures ofthe F E solution are close to the exact values, the temperature gradi ent
is the exact solution which satisfies the differential equation (9.25) and the bound ary of the FE solution is in much less agreement with the exact temperature gradient.
conditions (9.27) with the specific data given by (9.26) and (9.28). In recognition of this, it may be of some surprise that the flux q at the left end of
The FE solution and the exact solution are compared in Figure 9.7, where due the fin as determined by (9.36) is in perfect agreement with the exact solution as is
to the simple element adopted, the temperature determined by the FE method varies easily checked from (9.37). The reason is because the temperature gradient d'T/d x
linearly between its values at the nodal points. It is of interest that the FE solution present in (9.36) is not determined from dT/d x = Ba, which would lead to an inferior
is exact at the nodal points. answer, but rather from the boundary vector fb where the exact q-value enters.
While this is certainly not a general conclusion, we note that this fortunate In a later section we shall prove that the balance principle for the entire body is
situati on occurs for some one-dimensional problems (a detailed discussion is presented fulfilled exactly by the FE solution, and as the loading due to Q is given, as is also
170 Introduc tion to the finite elemen t method
FE fo r mulation of I-D hea t flo w 171

the flux at the right end of the fin, the FE solution is forced to provide the exact flux To evaluate the properties ofthe stiffness matrix K further, we make a partitioning
of K a nd a (cf. Chapter 2). As an example we split K a nd a int o the followin g
at the left end of the fin.
submatrices:

(9.44)
9.1.2 Further properties of the stiffness matrix
In thi s partitioning ofK, use is made of the fact that K is symmetric and it also follows
Apart from the problem discussed in Example I, the boundary conditions relating that the submatrix !5 is symmetric, i.e.
to the differential equ at ion (9.1) ha ve not been discussed, i.e. the system of equation s
(9.16) holds for arbitrary bo undar y conditio ns. Clea rly, the boundary conditions
mu st be considered in order to solve a specific physical problem and in order to do
I !5 = !5
T
I (9.45)

so in a systematic manner we first need to establish some important properties of Using (9.44 ) the qu adrati c form (9.41) can be written as
the stiffness matrix K. \
We shall first show that
1= [ gT @T{~i ~,][:] :<: 0 (9.46)

I detK =O I (9.38)
where (9.43) has been used. We emphasize that the subma trices given by (9.44) mu st
have prope r dimen sions so that the ma trix multiplications indicated in (9.46 ) a re
To prove this, consider the following homogeneous equation system: permissible. Evalua ting (9.46) we obta in
Ka = 0 (9.39)
l =g T A lg +gT A ,@ +@ TA ~g +@T !5@ :<: O (9.47)
Referring to (2.55), we recall that a non-t rivial a-soluti on exists if and only if (9.38)
Tak ing account of (9.43), we observe that I = 0 if and only if all components of the

r
is true. As a is independent of x, use of (9.13) in (9.39) impl ies that a-matrix a re ident ical ; otherwise we have I > O. Therefore, choosing a T = [gT @T] =
[OT @T] it follows from (9.47) that
BTAkBa dx = 0 (9.40 )

All the finite elements considered are able to reflect a zero temperature gradient
@T!5@> 0 I (9.48)

exactly (cf. (7.2 . Therefore, if all the temperatures at the nodal points are equal then for a 01 O. This proves th at!5 is positive definite and from (2.67) it then follows tha t
dT jdx = Ba = Oeven though a 01 0, i.e. (9.40) and thus also (9.39) po ssess a non-trivial
a-solution . In turn, this proves that (9.38) is correct.
Next, consider the quanti ty I defined by the following quadratic form:
I det!5 i'O I (9.49 )

l=aTKa I (9.41)
The submatrix !5 considered above can be viewed as obtained fro m the stiffness
matrix K by deleting some rows and correspond ing colu mns in the manner indica ted
by the partitioning shown in (9.44). It then turns out that !5 possesses the three
With (9.13) and (9.6) we get fundament al properties given by (9.45), (9.48) and (9.49). However, by using exactly
the same type of arguments as presented above, it follows that any submatrix !5
1= ' aTBTAkBadx = f'dT- Ak -dT dx :<: O (9.42) obtained from the stiffness matrix K by deleting one or more rows and the
fa II dx dx corresponding columns possesses these fundamental properties. Th is con clusion ha s
where Ak > 0 and (d T jdx) ' :<: O. As 1 :<: 0 for arbitrary a-matrices different from important consequences when considering the boundar y conditions, as sho wn next.
zero, we conclude from (2.68) that the stiffness matrix K is a positive semi-definite
matrix. It also follows fro m (9.42) that 9.1.3 Systematic consideration of boundary conditions
1 > 0 ifdT j dx 01 0 (9.43) When the system of equ at ion s (9.16) was establi shed, no use was made of the boundary
1=0 ifdTjdx =O conditions. The differential equation (9.1) contains only the temperature gradient
172 Intr od uction to the finite elemen t method FE formulatio n of I-D heat flow 173
d T / dx. Let us assume th at the boundary condition s are prescribed in terms of given indicated in (9.54). Let us define the follo wing subm atrices of the stiffness matrix K:
fluxes at both ends of the fin, i.e.
A, = [K , , ] ; A2 = [K ' 2 K" K lo ]
q(x = a) = h. ; q(x = b) = h. (9.50) K 21 Kn K 2, K 2
where h. and h. are kn own qu antities and where it is recalled th at q = - k dT/dx. K" K' 2 K" K ,. (9.55 )
AI = 15=
In th is case , both the d ifferential equation and the boundar y conditions involve only
the temperature gradient dT/ dx . This means that if T(x ) is a solutio n to the problem,
K. , K., K. , K
then T (x) + C, where C is an arbitrary constant, is also a solution to the problem.
In this situ ation, we would expect that a unique solution to (9.16 ) does not exist. where the symmetry of K was used in the definit ion s of A 2 and AI- Likewise, we
Referring to (2.56), we would therefore expect th e determinant of the stiffness matrix define the following submatrices of a and f :
K to be zero and thi s is precisely the situation indicated by (9.38 ).
T2 12
Consequently, in o rder to obtain a un iqu e solutio n from (9.16) we must s~cify
at least one nodal temperature, and valid boundary conditions co uld therefore tlf of T, I,
the form g = [g. ] ; @. = r = [J,] ; f= (9.56)

q(x = a ) = h.; T(x =b)=g. (9.51) T" J.


T(x = a) = g.; q(x=b)= h. (9.52)
T his means th at the system of eq uation s (9.54) can be writte n as
T (x = a) = g.; T (x=b)=g. (9.53)
(9.57)
where h., h. and g. , g. a re given values of the fluxes an d temperatures, respectively.
We shall now prove that th ese boundary conditions lead to a modificati on of the
It appear s th at the only unkno wns in th is equation system are ~ and r. Perfo rming
system of equation s (9.16 ), which en sures a unique solution. T he proof follows the the matrix mul tipl icat ions ind icated in (9.57) we obtain
same lines irrespective of whether (9.51), (9.52) o r (9.53) is given and, as an example,
we assume boundary conditions in the form given by (9.52). Fo r convenience, we A,g + A ,~ = r
also as sume that nodal points I and n are located at x = a and x = b, respectively; AIg + !5@ = f
that is, at x = a we hav e T, = g . With (9.15) and (9.24), the system of equations
(9.16) can be written as which can be restated as

K I
l I I K 12
----
K
,-- -- -13- - - - - --
K
lo g. (A q). =.
-- - - - - - III -f-,-
!5~ = f - AIg I (9.58)

K 21 I Kn K" K ,. T, 0 1" 12
K 31 :K 32 K" K ,. T, 0 + 1" I, r = A,g + A2~ I (9.59)
I
I T he right-h and side of(9.58) is known and as 15 is obta ined fro m the stiffnes s matrix
K" : K "2 K., K T" -A(x = b)h. h. I. K by de leting a ro w and the corr esponding co lumn, 15 possesses the properties given
by (9.45), (9.48) and (9.49). In particular, det 15 # 0, i.e. a unique solution ~ is o btained
(9.54) from (9.58) and thi s solution d oes not depend on the unknown r-matrix. By insert ing
the ~-solutlO n int o (9.59), we can det erm ine the unknown r-matrix. Therefore a ll
All the components of the stiffness matrix are kn own and the same holds for the quantities of interest have been calcu lat ed in a uniq ue manner. We also take the
components of the load vector fl' However, the total amount of heat flow at x = a, opportunity to draw attention to th e similar observations made in Chapter 3 regarding
i.e. (Aq). =., is unknown, which means that the force component I, is unknown, the properties of K and 15 as well as the introduction of boundary conditions (cf. th e
whereas the components 12 ".f. are known. M ore over, component g. of a is known, discussion related to (3.25)-(3.32)).
whereas the component s T2 T" are unknown . A partitioning of the ma trices is The systema tic considerati on of the boundary conditions illustrated above proves
174 Introduction to the finite elemen t m e thod FE formulation of I -D heat flow 175

that, when at least one nodal temperature is prescribed, a unique solution of the
system of equations (9.16) follows. More over, this systematic app roach lends itself
directly to a computa tio nal implem entation of boundary conditions in an FE program. r = [I Oq(x = 2) + 100] ; f= [ ~:] (9.63)
With the discussion above we are even able to find all solutions to (9.16) when -50
the boundary conditions are given in terms of prescribed fluxes alone (cf. (9.50. For From (9.58) we then obtain
this purpose we prescribe an a rbitra ry temperature at one end of the fin, for instance

_~~ -;~ -2~][;'] = [ ~:] _ [-2~][0] = [ ~:]


at x = a; that is, we set T( x = a) = g. , where g. is arbitrary. Assuming that nodal
point 1 corresponds to x = a, we have T (x = a ) = g. = T" i.e. we again obtain the
system of equations (9.57). Now , the only unknown matrix is ~, but this matrix can [
be found by solving (9.58), i.e. all nodal temperatures are now available. If we denote o - 25 25 T. - 50 0 - 50
this solution by T, it is obvi ou s that T + C is also a solution, where C denotes an This system of equations is identical to (9.34), i.e. the solution becomes
arbitrary constant.

\ (9.64)
9.1.4 Example 2
From (9.59) and (9.61)-( 9.64) we get
In or der to illustrate the systema tic considera tion of boundary condition s as descr ibed

O{~~] =
above , we again con sider the prob lem of Example I. According to (9.14) we have
seen in general that r = [lOq(x = 2) + 100] = [25] [ 0] + [ -25 0 - 350 (9.65)
Ka = fb + f,
Fo r the problem of Example I this system of equation s take s the form (9.32), i.c, i.e.
I
25 1- 25 0 0 0 10qx. ' 100 q(x = 2) = - 45 (9.66)
- - - - r-- - - - - - - - - - - --- - - -
- 25 I 50 - 25 0 T, 0 200 The solutions (9.64) and (9.66) are evidently in accordance with (9.35) and (9.36) of
I (9.60)
+ Example I. Even though the analysis above may seem a little involved, the important
o I' - 25 50 - 25 T3 0 200
point is that we have illustrated how a systematic consideration of the boundary
0 : 0 - 25 25 T. - 150 100 conditions can be carried out. More over, this systematic approach appli es to general
where the specific boundary condition s were introduced. The unknowns of this situations, i.e. it lends itself directly to a n efficient computer implementation of the
equ ation a re T" T" T. and lOq(x = 2). To solve it in a systema tic mann er, we make FE method.
the partitionin g as shown abo ve, which is an alogous to that indicated by (9.54)-(9.57).
Tha t is,
9.1.5 Evaluation of the force vector - fulfilment of the global
A , = [2 5] ; A, = [ -25 0 0] balan ce principle
(9.61)

-2~]
-25
- 25] [50 It is an important characteristic of the FE meth od that even though it is an
AI = [ ~ ; ~ = - 2~ 50 approximate method, the balance principle for the entire body is fulfilled exactly . In
- 25 25 Figure 9.8 all heat quantities which enter and leave the one-dimensional fin a re sho wn.
and As the flux q given by Fourier ' s law (4.3) is positive in the x-direction, (Aq)x is
the heat per unit time which enters the fin at end point x = a and (Aq)X=b is the heat

g = [0] ; ~ [~]= (9.62)


per unit time which leaves the fin at end point x = b. We recall that Q is the heat
input per unit time and per unit length of the fin and is mea sured as positive if heat
is supplied to the fin.
176 Introduction to the fin i te el em ent m ethod
FE for m ula tio n of I -D heat fl ow 177
Thi s means that the exact fulfilment of the balance principle for the entire body is
IAQ )' oa-LI ] - I A qlnb
expressed by the fact that, in accordance with (9.68), the sum of all the components
of the force vector f is equ al to zero . As an illustration consider the solution obtained
in Exam ples I and 2. Fro m (9.60) an d (9.66) it follows directly th at the force vector
L LJ-'---'--J.-.U-lJ.... 1 I a I x l fulfils the balance principle of (9.68). Recalling that our FE equations were derived
from the balance principle itself, the property (9.68) is indeed not surprising.
I
a
,
b
9.1.6 E valua tion of the l oa d vect or - poin t sources
Figure 9.8 Heat supply and removal over the entire body
Let us con tinue the evalua tion of the force vecto r f. For this purpo se we will prov ide
As we are considering time-independent problems, the balance principle for the a physical inter pret ation of the boundar y vector fb an d the load vector f,. We ob serve
entire body sta tes that the heat entering the body equals the heat leaving the body, i.e. that they have the sa me dime nsion, and to establish this dimension it is convenient

(Aq)x =. + f Q dx = (Aq )x=,


. to consider f, as given by (9.13). The glob al shape function matrix N is dime nsionless,
and as Q is the hea t supply per unit time and per unit length of the fin, then f, and
thu s also fb have the dimen sion [l/ s], i.e, heat per unit time.
which can be written as In order to ob tain a physical int erp retation of the load vector f, we evaluate a

-[Aq]: + f Qdx=O (9.67)


f" = r
component f" of f,. F rom (9.20) such a com ponent is given by

N,Q dx (9.69)
Th is equation is the ba lance principle for the entire body. Let us consi der no",t he
force vector f which, according to (9.15), is given by It follows tha t if the load Q(x ) is prescribed over a regio n where N, is different from
J. = fb' + f,, ; i = I , ... , n zero then f" oF 0, otherwise f" = O.
Next let us consider the influence of a load Q in term s of a so-called point source.
This implies that T his means that the heat supply is concentrated to a point and at this point a heat
" " " supply Q. per unit time is prescribed. Q. is called the strength of the point source
"LJ.
i~1
= i "Lfb'
= 1
+ "Lf"
1 jz
and it ha s the dimension [l / s]. Using Dirac' s delt a function (cr. (8.23)- (8.25) , we
may expre ss the load Q as
Fro m (9.20) we get

J,"J. J,"[ N ,Aq]: + J,"f'N,Q dx


oo if x = c
Q = Q J( x - c) = { (9.70 )
s 0 oth erwise
= - a
where c is the po sition of the point source. By definition , we have that

= - [(t
1- 1
N ,) AqJ '
oJ
+ f' (i
a 1= 1
N ,)Q dx
ro Q dx = ro Q, J (x - c ) dx = Q, (9.71)
Use of (7.139) yields
Du e to (9.70), (9.71) may also be written as
,t/= - [ Aq ]:+ f Qdx
r.' Q,J(x - c) dx = Q. (9.72)
and from (9.67) we conclude tha t
where c + a nd c - denote x-values slightly larger than and smaller than c, respect ively.

I ,t,J. I = 0 (9.68)
t; = r
For such a poi nt sou rce, the component f" given by (9.69) becomes

N ,Q, J( x - c) dx = f-' N,Q . J(x - c) dx = N,(c ) f-' Q, J(x - c) dx


178 Intro du cti on to the finite ele ment method
FE formulation of I -D heat flow 179
i.e,
fin. From (9.15) we have

, fu = N ,(c) Q, I (9.73) L" J, = iL=1" fb' + i L=" 1 t; (9.76)


i= t
,
If the pos ition c of the point source coincides with the position of nodal point i, then
Using (9.75) a nd Figures 9.9 and 9.10,it is obvious that the balance principle states that
N ,(c) = N ,(x,) = 1, i.e. we have
n

ii, = Q. if the point source is located at nod al point i (9.74)


L J, = 0
i= 1

Referring .to (9.69) for an ar bit rary load Q, it follows that it is always possible to in accordance with (9.68).
devise point sources located at the nodal poin ts and with suc h stre ngths tha t ii,
obta ined from (9.74) equ als ii, obtained from (9.69), i.e.

ii, = f N, Q dx = Q.. (9.75)


9.2 Expanded FE formulation of one element -
where Q., is the strength of the nodal point source located at nodal point i. We may assembling process
therefore envision the components of the load vector f, as given by the strengths of
pomt so urces locat ed at the nodal points. This equivalence is show n schema tically Pre viously, we establi shed the FE formul at ion for the ent ire body , but for computer
in FIgure 9.9. As the load Q is positive when hea t is supplied to the fin, the same implementa tions it will turn ou t to be important also to kn ow the F E form ulatio n
hold s for the strengths Q., (cf. Figure 9.9). Occasionally, the components of the load of one element. The num ber of such an element is den oted by a an d the ends of the
vector f, fro m a distri buted load Q are ob tained by intui tive physical considerahons, element are located at x = x, and x = X j' as shown in Figu re 9.11. Fo r the linear
Ho wever, when determ ining f, using (9.69) one speaks of a consistent load vector element the only nodal poin ts a re th ose at x = x, and x = X j , but for higher-order
since thi s determ ination is consistent with the Galerkin meth od. ' elements we also have interior nodal points. T he space which element a occupies is
Let us now return to the boundary vecto r f b given by (9.24). In accordance with denoted by L, .
Fo urier's law (4.3), the flux q is mea sured as positive in the direct ion of the x-axis, The app roximation of the temperature given by (9.4) holds for the entire body,
That IS, thecomponents of f b given by (9.24) can be illustrated as in Figure 9.10. and therefore it also holds in the region L , . By treat ing the region L , as a sepa rate
With FIgures 9.9 and 9.10 we have now illustrat ed all the heat supplied to the body, we can make an FE formulation for thi s region and will clearl y end up with
a formulation simila r to (9.12). Not ing that the region L, of element a is x, ~ x ~ x j

(r r
(cf. Figure 9.11), it follows that
1

2
0 , 0
n
0 ~
1

2
0
, 0
n
0 BTAk B dx ) 3 = - [ NTAq]:: + N TQ dx (9.77)

OJIIID
a
OJ]] When performing integration s over a region L , which extends from x, to X j' it is
as! as, a,"
Figure 9.9 Equivalence between dist ributed load Q and point sources Qsi
reg io n L a

2
o , o
n
o o ~~j---_ x
el eme nt no.a.
>---+--
I - - - - - ---I-I - _ X
o b
r-
Figure 9.10 Illustrat ion of boundar y vector (b
Figure 9./l O ne element with ends at Xi and x j and with possible interior nodal points
\
180 .
In troductio n to the fin ite elem en t m eth od FE fo r m ula tion of I -D hea t fl ow 181

r
region La.. we therefore have

r
convenient to write this as

XI
H (x) dx = f L.
H( x) dx (9.78) H ( x ) dx = t (L H (x ) dX) (9.84 )

where H(x) denotes any function of x. Likewise, we adopt the notation where, again, H(x) is an arbitrary function. With (9.13), (9.80) and (9.84), we obtain

[NTAq ] ;: = [ N TAq ]L. (9.79)


If we define the following matrice s: (9.85)

K" = L. T
B AkB dx

~"= - [NTAq k (9.80) where K:e denotes the expanded element stiffness matrix for element C( and fe' is the
expanded load vector for element ex.
ff" =f NTQdx Therefore, the (global) stiffness matri x K is the sum of all the expanded element
L, stiffnesses K:". Likewise, the (global) load vector f1 is the sum of all the expanded
element load vectors Ii". We emphasize that K and K: " have the same dimension
then (9.77) can be written as n x n, where n is the total number of nod al points in the body. Likewise f1 and rr
have the same dimension n x 1.
K"a = f~" + ff" (9.81) With reference to Figure 9.12, let us consider two neighbouring elements with
~ numbers ex and p. It appears that
Fo r reasons that will be explained later , superscript ee denotes ' expanded element'.
That is, K" is the ex panded element stiffness matrix for element ex, f~" is the expanded _ [ N TAq] ;: - [NTAq] ;; = - [ NT A q ] ;: (9.86)
element boundar y vector for element ex and ff" is the ex panded element load vect or for where q = - k dTldx . We have seen that, in genera l, the temperature gradient dTldx
element ex. Moreover, if we define the ex panded element force vector f" for element varies in a discontinuous manner over element boundaries; see for instance
ex according to Figure 9.7. Considering point Xj of Figure 9.12,it may therefore be somewhat surprising
that the boundary term from region L, cancels the corresponding boundary term
(9.82) from region L p However, the impo rtant fact is that the boundary fluxes q entering
the boundary terms in (9.86) are the exact ones, since no approximation of the
(9.81) ma y be written as temperature grad ient dTldx was used for these boundary fluxes (ef. the discussion
of (9.8)). Let us now determine the sum of the boundary term s - [ N TAq ] ;: for all
Keca = fCC (9.83) the elements of the body.
From (9.86) and with the notation of (9.79), we obtain
This system of equations co nstitutes the FE formulation of one element. We remark
that as a in (9.83) is the same as that present in the system of equations (9.16), which - [NTAq] : = - I'" [ N TAq] L. (9.87)
applies to the entire body, the dimension of a is n x 1, where n is the total number Cl "" l

of nodal points in the body. This mean s that the matrices K" and f" have the same
dimens ions as those of K and fin (9.16), respectively. As (9.83) has been formulated ste rnen t a. eleme nt 13
in terms of all the degrees of freedom for the entire body a nd not only tho se related I I
to the specific element, (9.83) is referred to as the ex panded FE fo rmulation of one
element. Examples of such a formulation are given by (3.15) and (3.17). ~--'i=
a
=
X' ,
=
Xj

= = =
b
}- x

Now let n., deno te the total number ofelements for the entire body which occupies
the region a '" x '" b; cf. Figure 9.1. An integra tion over th is region can always be
expressed as a sum of integr at ions over each element. With element ex occupying the Figure 9./ 2 Two neighbouring elements with regions La. and Lp
FE formula ti on of I-D heat flo w 183
182 In trod uction to the fi ni te elemen t method

Using (9.80) an d (9.13), (9.87) can be written as 9.2.1 Example 3


In order to illustrate how th e F E form ulation of the en tire bod y ca n be established
(9.88) as a sum of the contributions of all elemen ts in accorda nce with (9.85), we co nside r
again the problem descr ibed in Example 1.
With thi s procedure we ha ve to identify the number of the eleme nts ; thus
where f1: denotes the element boundary vecto r for element IX. Fig ure 9.5 is red ra wn in Figure 9.13. As we have three elemen ts, i.e. n" = 3, th e
Fro~ (9.82) it follows that element numbers become I, 2 an d 3, bu t apart from that the specific numberi ng of
".1 ".. no. the elements if of course imm aterial.
L:
a =1
f;' = L: r;,: + L:
~= l <1 =1
rr We recall th at the bou ndary vector f. is established in the same m anner a s before ;
th at is, from (9.30) we get
i.e. (9.85), (9.88) and (9.15) imply th at IOq(x = 2)
o (9.91)
(9.89) o
- 150
We a re no w in a posi tron to esta blish a fundamen tal property fo r the FE Next let us establish the expanded element stiffness mat rix and expa nded element
formulat ion . Based o n th e expanded FE formula tio n of o ne element, expression (9.83), loa d vector for each elemen t. F ro m (9.80), we have
we sum over all eleme nts of the body to o bta in
.. K" = f B T AkB d x = f T
dN Ak d N dx
".. ".1 L. L. dx dx
L K:Ca = L r: c

<1 "' 1 /1= 1 i.e. the com ponent (K" )ij of K " is
dN s ,
and with (9.85) and (9.89 ) this expressio n become s

Ka = f (9.90)
(K " ) . =
IJ f I
_ ' A k _J d x
dx dx
Likewise, we have from (9.80) that a co mponent (fn, of 1' is
(9.92 )

which is precisely the FE formu lat ion of the enti re body (cf. (9.16)). T his sho ws th at
the F E equation s for the ent ire bod y can be established as th e sum of the contributio ns
(1'), = f L.
N ,Q dx (9.93)
of all elements and th e process of making thi s summation is called assembling. This
essenti al result is an alogou s to the assembling procedure discussed in Cha pter 3 Co nsider first elemen t I, where 2 ,,; x,,; 4 corresponds to th e region L t , i.e.
(see for ins ta nce the discussion of (3.23) and (3.26)). 4 dN dN
We emphasize that we need nol establish the FE equat ions for the entire body (K ") .. =
f - ' Ak _ J dx
dx

r
I I) 2 dx
as a sum of the co ntributio ns of all elements. We can just as well use the formulation
(9.16) directly, but it will turn out later tha t a slight reformulation of the process
given by (9.85) and (9.89 ) provides essential adv ant age s when constructing a computer (ffa = N ,Q dx
program for F E an alys is. We also note that obt aini ng the boundary vecto r f. as the
sum of the expan ded element bo undary vectors f1: in accordance with (9.88) does
not provide an y adv antages. In fact, from (9.24) it follo ws th at the boundary vect or
f. only posse sses two non-zero components and it is therefore much more
ad vantageo us to establish the boundar y vector f. dir ectly from (9.24). Moreover, we
em phasize tha t the bou nd ar y co nditions of an elem en t are in fact unknown, except
whe n th e elemen t bo un da ry coi ncides with the bo unda ry of the bod y. Figure 9.13 Numbering of elements
184 Introduction to the fi nite element method
FE formulation of I -D heat flow 185
According to (9.26) we have. A k = 50 and Q = 100. More over, referring to
FIgure 9.6, the only shape functions which differ from zero in the region 2 < x < 4 The expanded element stiffness matrix K'" and the expanded element load vector
are N I and N, . It follows tri vially that - - f~ 'are of dimensions n x n an d n x I , respectively, where n is the total number of
noda l point s. The nodal temperat ures, which influence the temperature appro ximation
25 -25 0 0 100 within an element, are those related to the nodal points belongin g to the element.
- 25 25 0 0 100 Therefore, if nodal point j does not belong to the element, the nodal temperat ure Tj
KCC_
1 - (If = (9.94) has no influence on the results and , consequently, the component ( K'")ij of the
0 0 0 0 0 expanded element stiffness matrix is zero. As K'" is symmetric, we conclude th at a
0 0 0 0 0 comp onent ( K'" )'j is zero unless both nodal point i and nodal point j belong to the
element a considered. Therefore, if the element contains n, nodal points then only
In the same manner for element 2, where 4 S; x S; 6 corresponds to the region L "c x ne co mponents of the expanded element stiffness matrix KC Cwill be non- zero.
we obtain 2,
As K '" has the dimension n x n, this implies that man y components of K'" are zero
0 0 0 0 0 (see for instance (9.94)- (9.96)). This will be especially pron oun ced if the bod y is
0 25 - 25 0 divided into many elements.
,
Kce -_
0 - 25
fH =
100
(9.95)
As an example, assume that simple linear elements are used, i.e. n, = 2; then the
25 0 100 number of non-zero components in each expanded element stiffness matri x is
0 0 0 0 2 x 2 = 4. Assume that the total number of nodal points in the bod y is n = 30. As
0
the dimension of K'" is n x n = 30 x 30 = 900, the number of zero components in
For element 3 with the region L , corre spond ing to 6 S; x S; 8, we find that K '" is 900 - 4 = 896. This implies that a computer program based on the use of
o 0 0 0 0 expanded element stiffness matrice s would be highly inefficient, as it would perform
a sequence of operations with numbers equa l to zero. The same arguments hold for
'" _ 0 0 0 0 0 the expanded element load vecto r.
KJ - ff3 =
o 0 25 - 25 100 (9.96) It appears that it would be much more efficient to establish the stiffness matrix
and the load vecto r for an element using only those degrees of freedom that are
o 0 -25 25 100
related to the specific element in question.
From (9.85), we obtain with (9.94)-(9.96) When (9.83) was establi shed for element a, use was made of the approximation
for the temperature T given by (9.4) and valid for the entire body . Ob viously, as we
25 - 25 0 0 100 only consider one element we may equally well consider the approximation for T
K= -25 50 -25 0 200 applicable for this specific element only.This approximat ion is given by (7.127), i.e.
f, = (9.97)
o -25 50 - 25
o 0 - 25 25
200
T = N' a ' I (9.98)
100
As expected, K and f, esta blished in this manner correspond exactly to the expressions where N ' is the element shape function matri x and a ' contains the tempe ratures at
given In Example I (cf. (9.29) and (9.31)). Again we note the analogy of the assembling the nodal point s of the element. Thus
process used above to that discussed In Chapter 3 (see for instance (3.23l and (3.26)). T1
T,
N' = [ Nl Nj ... N;.J; a' = (9.99)
9.3 FE formulation of one e lement 7;"
We have seen t~at the FE equation for the entire body can be established as the sum where n, denotes the num ber of nodal points for the element. From (9.98) we ob tai n
of the FE equanons for each element. We also noted that this summation was made
only for the stiffness mat rix K and the load vector f" since it is more conven ient to
establi sh the boundary vector fb directly from (9.24). T SJN'
[EJ
_ = Beat
dx
where BC = --
dx
(9.100)
186 Introducti on to the finite element meth od FE form ula ti on of 1-D heat flow 187

i .e, T " Na K
/ I g lobal

B' = [dN~
dx
ddN
x
, ... dN:.J
dx (9.101)
0 0 0 0 0
for mulation
L
Use of(9.100) in the weak form (9.3) and adoption of the Galerkin method evidently
result in
K"
(1.. B' TA k B' d x )a' = - [ N ' TAq]" + 1.. N 'TQ dx (9 .102)
0
T = No

0
I 0
/ 0 : ==''2':: = :3
expanded el em ent

formu lat ion


where. as before, L, is the region of the element a considered and where the notations
La.
(9.78) and (9.79) have been utilized. If we define the following matrices:

K' = f B,TAkB' d x T = Nl!! a e-


I
K'
,/
L. elem en t
0
r:. = - [N'T Aq] L. (9.103)
0
formu lation

II = 1.. N'T Q dx
La.

.. Figure 9.14 The three different FE formulations

then (9.102) can be written as toc al numbers : 2

(9.104) glob al numbers :


o
2

3

5

6

7

B

9

K' is the element stiffne ss matrix for element a, f~ is the element boundar y vector for ------------
\
elem ent consider ed
element ~ and ff is the element load vector for element a. Moreover, if we define the
element f orce vector f" for element rx as Figure 9.15 Local and global nodal point numbering

I f' = r:. + 11 I (9.105)


formulation, one element is considered, but the global appro ximation T = Na is still
then (9.104) may be written as used. Finally, in the element formulation, one element is considered and the
approximation T = N ' a', applicable to this particular element, is used.
(9.106) Therefore, to obtain an efficient FE computer progr am the element stiffness
matrix K' is derived for each element. Then, instead of using the inefficient formulation
This system of equations constitutes the FE formulation of one element. Contrary (9.85), each term of K' is added directly to the proper component of the (global)
to its expanded form, (9.106) contains only those degrees of freedom which belong stiffness matri x K. In a similar manner, each term of the element load vector ff is
to the element considered and it follows that the non-zero components of K" and added directly to the proper component of the (global) load vector fl ' To describe
II' are those given by K' and ff, respectively. We recall that the dimensions of K' this procedure, we first note that in order to obtain a standa rdized determinat ion of
and If are " c x nc and nc x 1, respectively. the FE equations (9.104) for an element, it is advantageous to use a local nodal point
The three different FE formulations - global formul ation, expanded element numbering which should be distinguished from the global nodal point numbering.
formulation and element formulation - are illustrated in Figure 9.14. We recall tha t, To firm up these ideas, consider the body shown in Figure 9.15, where,
in the global formul ation, the entire body is considered and the appro ximation for convenience, simple linear elements are used. The element given by the global
T = Na, applicable to the entire body, is ad opted . In the expanded element numbers 5 and 6 is considered. Using the local numb ering, the FE equat ions (9.104)
FE formulation of I -D hea t flo w 189
188 In troducti on to the fi nite el em ent method
loca l no dal point s
take the form
I
1
\2
[~i: ~i:J[~:J =~:J +~:J (9.107) x= 2

=4
x

el.1
As local point number 1 corresponds to global point number 5 and local point
number 2 corresponds to globa l point number 6, we obtain the following procedure: Figure 9./7 Local nodal point numbering of element 1

5 6 K ~ , isaddedto K"
From (9.103) the element stiffness matrix K' of element I becomes
K ~2 is added to K , .
K ' l is added to K.,
(9.108)
K' = J: BoT AkB' dx (9.109)
K ' 2 is added to K
Accordin g to (9.26) we have Ak = 50 and using the B'-matri x as given by (7.35), we
If, is added to 1" obtain
5[ ff,J
6 If,
=>
1f2 is added to 1,.
K' = f4 [ -I /LJAk[ - ~ J-
-I dx - Ak
L'
f4 [ I - I J dx
I
It appears that the identification of proper components of K and f, as given by (9.108) 2 IlL L L 2 - I
is purely a geometrical or to pological scheme. In a general situation, tbe information (9.110)
on how local numbers communicate with global numbers is given by the so-called = 50[ 2 - 2J
to pology data . 22 - 2
Let us summarize the discussion abo ve. The FE equations can alwaoys be i.e.
established directly from (9.16), (9.15) and (9.13). The stiffness matrix K and the load
vector f, may also be derived according to (9.85) using the expanded form of the FE K' = [ 25 - 25J (9.111)
equations for the individua l elements. However, in order to construct an efficient FE -25 25 ...
computer program, the stiffness matri x K and the load vector f, are established using The element load vector If is given by (9.103) and as the load Q = 100 (cf. (9.26
the element stiffness K' and element load vector ff given by (9.103), where the we obtain for element I
component s of K' and If are added to K and I, using the topology data . In any case, 4
the boundar y vecto r I, is most conveniently obtai ned by using (9.13). I' =
'
f4N" Q dx =f1.f
2 L 2
[-x + 4J dX = l OO f 4[ - X +4JdX
x -2 2 2 x -2
(9.112)

9.3.1 Example 4 where (7.29) has been used. The result is that

We shall now show the manner in which the FE equations for the entire body can f' = [IOOJ (9.113)
, 100
be established using the element stiffness K' and element load vector If given by
(9.103). Again the problem discussed in Examples I, 2 and 3 is considered. For According to Figures 9.16 and 9.17 the local nodal point I corresponds to the global
convenience, the element numbering and global nodal point numbering are shown nodal point I and the local nodal point 2 corresponds to the global nodal pomt 2.
again in Figure 9.16. The cont ributi ons of (9.111) and (9.113) are added to the (global) stiffness ~atrix K
We consider element I first with the local nodal points as shown in Figure 9.17. and the (global) load vecto r I, and as these are originally zero and of dimension
4 x 4 and 4 x I, respectively, we obta in the intermediate result
25 - 25 0 0 100
e L1 el.2 e1.3
1 2 3 4 - 25 25 0 0 100 (9.\14)
---4 0 0 0 x
K= I, =
x-2 =4 =6 =6 , 0 0 0 0 0
Figure 9.16 Global nodal points and element numbering 0 0 0 0 0
FE formulation of I-D heat flow 191
190 Introduction to the finite element method

loca l nodal points the following final result for the glob al stiffness matrix K and the global load vector f,:

1
/ \2 25 - 25 0 0 100
c 0
200
- 25 50 ,- 25 0
x =4 =6
K= f, = (9.116)
el.2 0 - 25 50 - 25 200
Figure 9./8 Local nodal point numbering of element 2 0 0 -25 25 100

l ocal nodal points As expected, these K- and f,-mat rices corr espond exactly to th ose obtain ed in
/ \ Examples I and 2 (cf. (9.29), (9.31) and (9.97)), respectively. Fin ally, we recall that
1 2 th e boundary vector f b is obtained as in Example I , i.e. it is given by (9.30).
o
013
Figure 9./9 Local nodal point numbering of element 3 9.3.2 Use of local coordinate systems

We next cons ider element 2 with the local nodal points as shown in Figure 9.18. In order to standa rdize fur ther the determinati on of th e element stiffness matrix K '
and the element load vector f, we may use a convenient local coordinate system.
From (9.103) the element stiffness matrix K' of element 2 is

K' = f: B'T AkB' dx


Since K ' and f1 are obtained by integration (cf. (9.103 , it is obvious that we may
choose any coordinate transformation when these integrations are carried out,
To show this, we ad opt the linear tr ansformation
With (7.35) and as Ak = 50 we o bta in . x = c~ +d (9.117)

where c and d are param eters a nd ~ is the new variable. As sho wn in Figure 9.20,
K'=f6[ -I /LJAk[ - ~ ~J dX
4 IIL LL let the coordinates of the left and right ends of the element be given by Xi and Xj'
respectively, i.e. the length L of the element becomes L = xj - X/ .
and a compa rison with (9.110) shows that the element stiffnesses K' for elements I Now we choose the coordinate transformation (9.117), for instance, such tha~
and 2 are identical. This is certainly not surprising, since Ak and L are identical for ~ = _ I and ~ = I at the left end and right end of the element, respectively. That is,
the two elements. Moreo ver, it can ea sily be shown that the element load vector f' in the local ~-coo rdi nate system, th e element of F igure 9.20 tak es the form shown in
for element 2 is equ al to that of element I, i.e. it is given by (9.113). From Fig ure 9.21.
Figure s 9.16 and 9.18 the top ology da ta of element 2 is that local nodal point I
corresponds to global nodal poin t 2 and local nod al point 2 correspo nds to globa l
nod al point 3. Th erefore , adding (9.111) a nd (9.113 ) to the proper components of
(9.114) result s in possi bl e interio r noda l point s
\ /
25 - 25 0 0 100 o o ~:g.--- "

-25 50 - 25 0 200
(9.115) L
K= f, =
0 - 25 25 0 100
Figure 9.20 Geometry of one element in global co ordinate system
0 0 0 0 0
We then con sider element 3 with the local nodal points as shown in Figure 9.19.
It can easily be shown that the element stiffness matrix K' and the element load possible int er ior noda l points
vector f1 ar e again given by (9.111) and (9.113), respectively. According to
FIgure s 9.16 and 9.19 the topology data for element 3 is that local point I corresponds
~--iC
-1
Wo :::]~--g
1
to global nod al point 3 and local nod al poin t 2 corresponds to glob al nodal p6int 4.
Th erefore, add ing (9.111) and (9. / 13) to the proper compo nents of (9.115) gives Figure 9.21 Geo metry of one element in local { -coordinate system
192 In trod uction to the fin i te element method FE formulati on of I -D heat fl o w 193
This transformation is easily show n to be given by can be written as

x =
L
-~
2
I
+ -2(x'. + x ).) (9.118) Ii =
I
x )

x,
NoTQ dx = -II' [1 + ~ J
2 - I 1 ~
L ~
Q -d
2
(9. 125)

which gives and if Q = con stant we get


L
dx = -dI' (9.119)
2 . (9.126)

As an exam ple, con sider the simple linear elemen t of Chapter 7. From (7.28) and
(7.29) we ha ve It appea rs from (9. 124) an d (9.126) that for constant A k- and Q-value s the element
stiffness matrix K' and the element load vector Ii can be establi shed dire ctly using
w = [N~ Nj ] (9. 120) closed-form expressions. In an FE program such expre ssions clearl y facilitate
where computati onal efficiency a nd similar closed-form solution s may be establ ished for a
seq uence of typical forms of Ak and Q. Here we have only considered the simple
N ,,( x ) = - -
1 ( x - Xj) = - -I -~
L L2
[L + -1 (x. + x .) -
2 ' ) )
J1 1- ' )
x , = -(
2 '
linear element, but similar closed-form expression s may be deriv ed for higher-order
element s.

,
Nj(x)= I
-(x
L L2 2 )
[L
-x,)= -1 -~+ -1 ( XI + X . ) - X . = -(
,
I 1+')
2 '
J (9.121)

9.4 Axially loaded elastic bar


Moreover, from (7.33) we have that
Consider the axially loaded elastic bar shown in Figure 9.22, where A(x) is the
(9.122) cross-sectional area, E( x) is Young's modulus and hex) is the body force per unit
axial length.
With u being the displacement measured positive in the x-direction we have
in accordance with (7.35). F ro m (9.119) and (9.122) it follows that the element stiffness
matrix K' given by (9.103) is
I I (9.127)

K' =
Jx,
r B' TAkB' dx
X

) = I I [ - I I LJAk[ -
- I IlL L
2. 2.J!:: d~
L 2
u=Na

where a now cont ains the displacements at the nodal points. The strain e is given by

= -
2L
I II [I -IJ
_I
Ak
- 1 1
d~
I e = du = Ba
dx
I (9.128)

i.e,
In Chapter 4 it was shown that the equations for such a problem are similar to those
K' = ~[
2L
1
- I
-IJI'
1 - I
Ak d~ (9.123)
of heat flow ; cf. (4.8)- (4.10) with (4.18) and (4.20), (4.21). Referring to (9.12) and

Elx l Alx}
If Ak = constant we get

~ / j
i.......
- - - -1----
,
(9.124)
f - - -- +
blx}
I - -- - -- - - - -+-
I - - x
a b
From (9.119) and (9.121) it follows that the element load vector Ii .given by (9.103) Figure 9.22 Axially loaded elastic bar
194 In troduction to the fi ni te element m e th od
FE form ula tion of I ~ D h eat fl o w 195
Table 4.1 we can therefore write the F E formulation directly as

(J: r
B AEB dx)a = [NT Au]: + J: N Tb dx (9.129)
~=bL[IJ
1 2 I
(9.136)

where a is the normal stress. Care should be taken not to confuse the axial load b where k is the stiffness of the bar (cr. (3.38).
with the position x = b at the right end of the bar. The force N in the bar is given by For the heat flow problem, the global stiffness matrix K is singular (cf. (9.38)),
whereas any submatrix K of K is non-singular (cr. (9.49)). The introduction of the
du
N=Au' u=Ee=E - (9.130) boundary conditions by which at least one nodal temperature is specified implies
, dx
that K could be established from K. The present case of an axially loaded bar is quite
N is seen to be positive if a tensile stress exists in the bar. Let us define the following similar and the stiffness matrix K of (9.131) is also singular. The introduction of
matrices: boundary conditions in terms of the specification of at least one nodal displacement
is now the tool which enables us to obtain a non-singular submatrix K from K. This

K= J: BTAEBdx
can be proved formally just as for heat flow, but the need to specify at least one
nodal displacement is physically obvious since it removes possible rigid-body motions.
We recall that a rigid-body motion does not create any strains.
r, = [NT Au]: (9.131 )

f1 = l' NTb dx
.. 9.4.1 Example 5

With the discussion of the element stiffness matrix K ' and the element load vector
as well as f,and the introduction of a convenient local coordinate system for the establishment
of these matrices, we are in a position to illustrate the details of the efficient
I f = fb + f1 (9.132) implementation of the FE equations into a computer program. For this purpose
consider the axially loaded elastic bar shown in Figure 9.23. The bar consists of two
Then (9.129) can be written as parts, I and 2, each having a constant Young's modulus E, constant cross-sectional
"'fIa A and constant external load b.
I Ka = f I (9.133) This problem is analyzed using the two simple linear elements shown in
Figure 9.24.
where it is easily seen that the force vector f has the dimension of force, i.e. [N].
Next let us consider the FE formulation of one element. The element stiffness E,A, E2 A2
matrix K ' and the element load vector ff now take the form 1- - -:- -, - - -I x
\b 'b ,
K' = f L,
B'T AEB' dx
x =0
1
x=b

F, = f L,
N'Tb dx
(9.134) l,
Figure 9.23 Configuration of axially loaded elastic bar

For constant AE- and b-values, these expression can, analogous to (9.124) and (9.126),
~~~u .
, 3

K' = k[ - I1 -IJ.
I '
k = AE
L
(9.135)
\ l,

Figure 9.24 Two simple linear elements used to model the problem of Figure 9.23
196 In trod ucti on to the finite elemen t method FE form ulation of I -D heat flow 197

Let us cons ider first the boundary vector f b given by (9.131). As in (9.24) we obtain It appears that NX ~" is that ten sile force - namely, the reaction - which must be
applied at x = a in order to ma intain equilibrium of the bar.
;/
I (9.137)

Next we evaluate the element with length L , . From (9.135) and (9.136) we have 9.5 Basic features of an FE computer prog ram yl
K ' -_ k, [ I
- I
-IJ. I"
f' = ~[IJ
2 I
We are now in a position to sketch the comput er fl ow diagram for an F E program
as follow s:

where k, = A,E ,jL,. The topology data follows from F igure 9.24 ; that is, after the
contributions for the element with length L " K and f, are given by Input data :
N umber of equat ions n

K = -~
k, - k,0]
~'~; f, =
[b 'L ']
~ b,~,
Number nel and type of elements
Po sition of nodal points
[ Topology data
Loading data
Material data
F or the element with length L 2 we obtain
I
-IJ. f,=b L
"
2 2[IJ
K' = k 2 [
- I
I
I " 2 1 I K ~ O; f, ~ 0 I
where k 2 = A 2E2 /L 2 . With Figur e 9.24, after the contributions of this element we I
obtain For i = 1 until i = ne\ do
Calculate Ke and f 1forelement j
k, - k, add K ' and fi to proper positions of K and f
K = - k, k, + k2 ]
- k2 ; f, = ~ [b'
b,L , +L,b2 L 2 ] (9.138)
I
[
- k2 k2 b, L 2

r
Inp ut data :
From (9.137) and (9.138) th e final FE equations are Boundary co nditions

[ -{Ao-)X~"] I
k, - k,

[
- k, k, + k, - k, u, = + ~ [b,L,s.z.,
+ b, L
]
2 (9.139) I Calcu late Co I

- k2 k, u, (AO')X ~b b,L 2
I
The boundary condition s are now introd uced. We assume that the bar is fixed I Establish Ka ~ r, + f,
at x = a and that the right end of the bar is free, i.e.
(9.140) I
The solution becomes Modify system of equations according to the boundary conditions
and solve the modified system of equations
I I
u, = - (b, L , + 2b,L,); u, = - - [ k, b, L , + (k , + 2k2)b2L,J (9.141)
I
2k, 2k,k,
I Calculate and write results (temperatures, fluxes, etc.) I
(9.142)
198 Introduction to the finite element method FE formulation of I-D heat flo w 199

First a sequence of input data is read. Then the total stiffness matrix K and the total / Al x)
load vector ) are initialized, i.e. the matrix dimensions are identified and all their
components are set to zero . Fo r each elemen t, the element stiffness matrix Ke and p
the element load vector ff are calculated and added to the proper components of K
and f, using the topology data. Th is is the assembling process. The boundary conditions
are read and the boundary vector fb is calculated. Those components of f b that are Q lx )
unknown (i.e. the 'reactions') are included in the ma trix r of( 9.57). The ent ire system
a b
of equations is established, and after the introduction of the boundary conditions, I I
the modified system of equations is solved according to (9.58). The component r of
f which was originally unknown is then calculated using (9.59), i.e. the 'reactions ' Figure 9.25 Convection at the right end of the fin
can now also be calculated. Finally, the results in terms of temperature, fluxes and
'reactions ' are calculated and written out.
In practice, a so-called pre-processor is often used in connection with the input
to the program. This pre-processor may perform an automatic division into elements fin as shown in Figure 9.25;cording to (9.14) the F E formulation is
following some rules specified by the user, and it may also provide graphical
information on the element mesh, nodal points, etc. This facilitates the possibilities Ka = f b + f, (9.144)
for ensuring that correct input data is used. Likewise, a so-called post-processor is The boundary vector fb given by (9.13) can be written as
often used to provide graphical information on the results , for instance in terms of
contour curves for the temperature etc. fb = - [NT Aq]: = _ (N TAq)x: ' + (NTAq )x:o (9.145)
As we have one-dimensional heat flow, the flux q. of Figure 9.25 is q. = q(x = b),
i.e. (9.143) takes the following form of a boundary condition:

9.6 Heat flow with convection (q) x: ' = ~(T, :, - Too) (9.146)
We note that if convection occurs at the left end of the fin, then q. = - q( x = a). In
Generally, fluid motion around a body is a very effective means of transporting heat general, we have the approximation T = Na, i.e. we may write the unknown
between the body and the fluid. This type of heat transport is called c0"l!ection. temperature TX=b as
Assume that the fluid far away from the body has the uniform temperature T~he
temperature at the surface of the body is given by T and it is natural to expect that T.: =b = N X =b 3
the temperature difference T - Too controls the amount of heat flow between the This implies that (9.146) becomes
body and the fluid. This lead s to Newton's convect ion boundary condition expressed
through (q)x :' = ~Nx :,a - ~ Too
,.--- -- --
The boundary vector f, given by (9.145) can then be written as
q. = ~(T - Too) I (9.143)
f b = - ~ ( ANTN )x : , a + ~Too(ANT)x :' + (NT Aq)x :,
where q. is the flux through the boundary with the outer unit normal vector n (cf. Use of this expression in (9.144) gives
Figure 6.1 and (6.3)) and a is a parameter, the so-called convection coeffi cient with
the dimension [J im's C ]. We note the similarity of (9.143) and (6.31). The (K + KJa = ~Too(ANT)x :' + (NT Aq)x: o + f,
convection coefficient C( depends on a variety of circumstances such as fluid velocity, (9.147)
where K, = ~( ANTN)x : '
type of fluid, surface conditions, etc. but for our present purpose we assume that the
convection coefficient 0: is known. Moreover, 0: is a positive quantity, implying that
when the temperature difference T - Too is positive, the flux q. as given by (9.143) It may be of interest to evaluate the expression for K, . The number of nodal
becomes positive. This is in accordance with the definition of the flux q. (cf. points for the ent ire body is given by n. Assuming that nodal point n is located at
Figure 6.1), since a positive q. means that heat leaves the body. x = b we obtain
As an example of (9.143) assume that convection occurs at the right end of the
200 Introductio n to th e fin ite element method FE f or m ulati on of l-D heat [Jaw 201
i.e. not truly one-dimensional. Actually, this situation is similar to the discussi on related
to the heat supply Q, a nd referring to the discussion on page 48 we accept the
o 0 o engineering approximation of a o ne-dimensional heat flow in the x-direction provided
o 0 o that the temperature vari ati on s along the fin are much larger th an temperature
(9.148) variations normal to th e fin. Th is assumption is usually reali stic if the fin is th in.
With this discus sion it is o bvio us th at we may consider th e heat flux q, due to
o 0 convection as a mod ificat ion of t he~t supply given by Q. N ow Q is th e heat supply
Expression (9.147) shows that the convection condition implies a modification per unit time and per unit length of the fin. Moreover, Q is positive when hea t is
? f the stiffness matrix. The new stiffness matrix is symmetric j ust like K, bill it is supplied to the fin. In contrast, q, is the heat per unit time and per unit a rea and q,
Important to observe that in contrast to K the mod ified stiffness matrix is non-singular, is po sitive when heat lea ves the fin. Let P be the perimeter of th e fin ; th at is, q,P is
i.e. the heat per un it time and per unit length of the fin. It appears that we may treat
the convection as a modi ficati on of Q according to the scheme
I det(K + Kol i' 0 I (9.149) (9.152)

To sho w this, consider the quadratic form I given by which mean s that Q sho uld be replaced by the term Q - q, P. The F E formulation
is again given by (9.14), i.e,
I = a T(K + Kola = aTka + a TK,a (9.150)
Referring to (9.41) and (9.43) we have aTKa = 0 if and only if dTjdx = 0, i.e. if the
Ka = fb + f, (9.153 )
components of a are equal ; otherwise aTKa > O. From (9.150) and (9.148) it therefore where the load vector f, is given by (9.13), i.e.

f
follows that

[ >0 for all a se O (9.151) f, = NTQdx (9.154)

That is, the modified stiffness matrix is positive defin ite and, according to (2.67). it
is then non-singular. This proves that (9.149) is correct and implies that the system Combining (9.152) and (9.143) gives
of equations (9.147) can be solved directly irrespective of the boundary condition' at Q1r"Q - aPT + aPToo
the left end of the fin. \......
As another important engineering problem of convection, assume that lateral The unknown temperature T can be replaced by T = Na to obtain
convection occurs along the outer surface of the fin as shown in Figure 9.26. This Q -"..Q - ",PNa + aPToo (9.155)
type of convection may cause conceptual difficulties because it can only occur if heat
flows in directions perpendicular to the x-axis and this means that the heat flow is With (9.153)-(9.155) we obtain

(K + K,)a = r, + f, + f ",PToo N T dx

where K, = f: ",P N TN dx
(9.156 )

and f b and f l are given in the usual manner by (9.13). Agai n it appea rs th at the
convection result s in a mod ificati on of the stiffness matrix. T he modified stiffness
Q Ix J matrix is symmetric, and in a manner simila r to (9.150) and (9.151) it can easily be
pro ved that this matri x is non -sing ular, i.e.
o b
,
I - - -+-- - -- - - _ _ +-_ _ x
(9.157)
Figure 9.26 Lateral convection along the outer surface of the fin
202 Introduction to the fin i te element m ethod FE formulati on of I -D heat flo w 203
Hence the system of equ at ions (9. 156) can be solved directly irrespective of the ,..--- ele ment boundary
boundary conditions imp osed on the ends of the fin. I
oJ I
We fina lly obs erve that with the results above it is straightforwa rd to combine "smoot hing" J""e
the convect ion cases of Fig ures 9.25 and 9.26. N,
<.
L - - -4
I
+-- - - - -- x
9.7 CO-continuity bJ ,, ,
I

We have already established criteria in Chapter 7 which ens ure that the FE formulation i'"
conv erges towards the exact so lution for decreas ing element size. According to page I------'-~e___-------x

93, these .convergence crit: ria consist of the fulfilment of comp leteness and
compa tibility, T he compatibility or conforming requi rement sta ted that the
approximation of the temperature also varies in a continuous manner over element cJ
boundaries. .
Let us now derive this con tinuity requirement in a different manner. It is not
surprising that the fulfilment of the co nvergence criteria hinges on the exp ressions
chose n for the app roximation, i.e. on the global shape functio n matrix N. Referring
to the global stiffness matrix K, the bo undary vector fb and the load vector f, given
"
by (9.13), It appears that K places the most severe conditions on N since the deriva tive "
B = dN/ dx enters the expression for K. A compo nent K/j of K is given by (9.19), i.e. "
: ~ I
I I
' dN I s,
, I
K Ij =
f -
dx
Ak - dx
dx
(9.158)
"-
Figure 9.27 Differentiation of shape function with CO-continuity
!fthe term s Ak and dN,/dx are continuous functions, the integration given by (9.1fs)
can be earned out direc tly. Ho wever, if discont inui ties occur we may run int o trouble.
In order to eva luate thi s situa tion, ass ume tha t the globa l sha pe functio n N . varies Here we conclude that CO.conti nuity is sufficient for thc globa l sha pe functions of
m the manner shown in Figure 9.27. It a ppea rs that N . is continuous whereas dN ./ dx heat flow problems, but it is obv ious that the continuity requ ired depends on the
is discon tinuo us over the element bo unda ry. Mo reo;er, even tho ugh dNJdx v~ries differentia l eq uations involved. We shall return to this to pic in late r chapters.
discontinu ou sly, an integ ration which involves dN J dx is still well defined.
To investigate whether it is also possible to ma ke a well-defined integration of
d N J dx , ~e replace N , by the smoo th dashed curve sho wn in Figure 9.27( a ). Th e
2 2

correspo ndmg varranons of d N,/dx and d 2 N J dx 2 are show n in Figure 9.27(b) and 9.8 Concluding remarks
(c), respect ively, and It follows that an integr ation in volving d 2N ,/ dx 2 is not,.in
general, well defined. We may draw the conclusion that in order for the stiffness In principle, even though thi s chapter has been devoted to simple phy sical problems,
matri x to be established in a well-defined manner, we are led to the compatibility it contains all the ingred ients of an FE formulation . In later chapters more complicated
requi rement stat ing that the approxima tio n of the tempe rature must vary in a physical problem s will be treated, but the fundament al FE result s a re analogo us to
con tmuo us manner over the element boundari es. A functio n which is contin uo us but those discussed here. It is for this reason that we ha ve presented a rather detailed
which has a discontinuous first derivative is termed a C O-continuous f unction. In discussion of variou s ty pical issues related to the FE formulation. As these matters
general, we have the following sta tement : are easily generalized, we will ado pt in later cha pters a much more compac t meth od
of presentation for establ ishing the different FE form ulations.
en-con tinuity mean s that the n-derivati ve is continuous. Let us therefore review some of the funda mental facts. The FE formulation can
always be given in the form of(9.16), (9.15) and (9.13) a nd this is typically the form at
204 Introduction to the finite element method FE formulation of I -D heat flow 205
method and this fact is illustrated in Figure 9.28. We also note that~ addition to
Gcler ki n
being more general, the Galerkin method often provides a more straightforward
derivation of the FE equations than the Rayleigh-Ritz method .
We finally draw the reader's attention to the comprehen sive textbook s by Becker
et al. (1981), Hughes (1987) and Stasa (1985) that also treat one-dimensional heat flow.

Figure 9.28 Relation between the Rayleigh-Ritz method and the Galerkin method

chosen when discussing theoretical aspects. Alternatively, the global stiffness matrix
K and the global load vector f, may be derived as the sum of the expanded element
stiffness matrix K" and the expanded element load vector C;' , respectively, for all the
individual elements (cf. (9.85 . Conceptually (9.85) is of fundamental importance,
but when establishing a computer program for the FE method, these expanded forms
are extremely inconvenient to work with, so, instead, the element stiffness matrix KC
and the element load vector C; are derived using only those degrees of freedom related
to each individual element (cf. (9.103 . The contributions ofK' and C; are then added
to the global stiffness matrix K and the global load vector f" respectively, u~ng the
topology dat a. We also note that the boundary vector fb is always derived uyng the
form given by (9.13).
It appears that the FE method is extremely wellsuited for efficientimplementation
on a computer, which enables us to deal with a variety of problems.
As a global shape function N, only differs from zero in those elements which
contain the nodal point i, the global stiffness matrix K was found to be banded .
Moreover, the Galerkin method implies that K is symmetric. We found that K is
singular and positive semi-definite; see (9.38), (9.41) and (9.43). We proved that any
submatrix ~ obtained from K by deleting one or more rows and the corresponding
columns is symmetric, non-singular and positive definite; see (9.45), (9.48) and (9.49).
These aspects have import ant consequences when the boundary conditions are
prescribed. We also showed that, apart from problem s involving convection , the
specification of at least one nodal temperature is a necessity in order to obtain a
unique solution of the FE equations. More over, the balance equation for the entire
body is fulfilled through relation (9.68).
In this textbook , the Galerkin method is used universally, but we remark that,
as an alternative, one may also adopt the so-called Rayleigh-Ritz method. This
method relies on the construction of certain functionals for which a minimum is
sought using the calculus of variations. If such a functional can be established, it turns
out that the Rayleigh- Ritz and Galerkin methods result in identical F E formulations.
For a discussion of these topics the reader is referred to Hughes (1987), Strang and
Fix ( 1973) and Zienkiewicz and Taylor (1989). However, the important point is that
whereas the Rayleigh-Ritz method can be used only for certain types of differential
equations, the Galerkin method is applicable to arbitrary differential equations. This
means that the Rayleigh-Ritz method can be viewed as a subset of the Galerkin
...

FE fo rm ulation of 2-D and 3-D h eat fl ow 207


10 ---, y
~

FE formulation of two- and


three-dimensional heat flow
L x

Figure 10.1 Two-dimensional region A with boundary !l' = ::Rft + !l',I

Moreover, the flux q, at the boundary of the region is given by Figure 6.1 and (6.3), i.e.
Following the detailed discussion of the FE formulation for one-dimensional heat (10.5)
flow, it is an easy task to treat two- and three-dimensional heat flow. We emphasize
tha t even though we focus on hea t flow here , a sequence of other important physical and q, is positive when heat leaves th e body. As shown in Figure 10.1, the boundary
problems is governed by similar equations (see Table 6.1). Therefore, the FE !!' of the region A consists of .5t; and !.f,. On .5t; we have q, = h where h is a known
formulation presented in this chapter has significance in a variety of other fij dS. quantity and on!.f, we have T = g, where g is a known quantity. With these preliminary
remarks we are in a position to obtain the FE formul ation,
The temperature T is approximated by mean s of (7.137), i.e.
T =Na ( 10.6)
where N is the global shape function matrix and a contains the temperatures at the
10.1 Two-dimensional heat flow nodal points in the entire body. This means that

Considering two-dimensional heat flow first, the strong form was given by
(6.32)-( 6.34) and the corresponding weak form by

-f, f f
N.[N, N, ... N.J .. m (10.7)

f.4
(V v)TtDVTdA =

T = g
~
vht d!!' -
~
vq, t d!!' +
It
vQt dA (10.1)
where n is the number of nodal points for the entire bod y and a component N ,
on!.f, ( 10.2) x
depends on and i.e. y, = N, N,(x, y).
F rom ( 10.6) we obtain
(cf. (6.46) a nd (6.47 . In (10.1 ) and (10.2) we recall that v is the arbitrary weight ( 10.8)
VT = Ba where B = VN
function, D the constitutive matrix, T the temperature, t the thickness and Q the heat
supply per unit time and per unit volume of the body. The flux vector q is determined which implies that
from Fourier's law (6.4), i.e.
aN1- aN
-2 '' ' - aN,]
q = -DVT (10.3 )
B=
ax ax ax (10.9)
where the temperature gradient VT is
[-en, e,
-
oN,
-
oy oy . oy

vr [1~] (10.4)
in accordance with (7.140) and (7.142). Inserting (10.~) into (10.1) gives

(f A
(Vv)TD Bt dA)a = -f ~
vht d!!' - f
~
vq, t d!!'
..
+ f
A
vQt dA (10.10)

206
208 In troduction to the finite element method FE fo r mulation of 2-D and 3-D heat flow 209
The final step is to ch oose the arbitrary weight function v. In accordance with i.e. ( 10.16) becomes
the G alerkin method we set
/" v = Nc (10.11) I Ka = f I (10.18 )
/
Since v is arbitrary, the matrix c is arbitrary. F ro m (10.11) we obtain J ust like o ne-dimensional heat flo w, the force vecto r f has the d imen sion of [J/ s]
and a comparison with th e corresponding formulat ion for one-d imensiona l heat flow
V v = Be ( 10.12)
(cf. (9.13)- (9.16) shows tha t a co mplete similarity exists.
As v = vr, (10.11) can al so be written as Ow ing to the symmetry of the co nstitutive matrix 0 (cf. (6.10)) it foll ows.from
v = cTNT ( 10.13) (10.15) tha t K is symmet ric '"

Insert ing (10.12) and ( 10.13) into (10.10), and noting th at c is indepen de nt of posit ion ,
gives
I K.= K
T
I ( 10.19)

Similar to the eval uat ion of (9.38) - (9.40) it follows th at K is sing ula r, i.e.

I det K = 0 I ( 10.20 )
As thi s expression should hold for arbitrary cT-matrices, we conclude that
M oreover, since 0 is positi ve definite (cf. (6.8)), we derive in a simila r manner to

(f A
BTDBt dA )a = -f 2'1.
N Thl d S!' - f
:e,
N Tq. 1 d S!' + f
A
N TQt d A
\.
(10.14)
(9.41)- (9.43) th at K is positive semi-definite, i.e.

I aTKa ~O ( 10.2 1)
which is the FE formulation so ugh t.
To write (10 .14) in a mo re compact fashion, we define the following matrices : for all a # 0 and th at o nly if the temperature gradient is zero , i.e. VT = Ba = 0, is
the quadratic form a bo ve equal to zero. Finally, using the a ppro ac h discu ssed in

K= L BTDBtdA
relation to (9.44) -(9.49 ) it can easily be shown that any submatrix !5 o btained from
K by deleting one or more row s and the corresponding co lumns is symmetric,
non-sing ular and po sitive definite . With refere nce to Chapter 9, we conclude that at

r, = -f ~
NTht d S!' - f
~
N Tq.t d S!' (10.15)
leas t on e nodal temperature has to be p rescribed in order to obtain a unique solu tio n
of the FE equations. A systematic con sideration of the boundary condition s is again
given by the approach o utlined in (9.57)- (9.59) .
Let us now prove that th e co mpo nen ts of the fo rce vector again fulfil relation
fl = LNTQt dA
(9.68). As we consider a region A with thickness I, the balance principle sta tes that

As 0 has the dimension 2 x 2 and B the dimension 2 x n, it follows that K is a
sq ua re matrix with d imension n x n and it is the stiffness matrix. Likewise, both fb
L Qt dA = t.l"" " ( 10.22)

(cf. (6.19)). Let us write (10.17) as


a nd fl ha ve the dimen sion n x I a nd they are termed the boundary veclor and load
vector, respectively. With ( 10.15), ( 10.14 ) can be written as J. = fb' + I.,; i = 1, . .. , n
which leads to
I Ka = f b + fl ( 10.16)
L J. = iL= l f b' + ; L= 1fH (10 .23)
i=l
We define the f orce vector f by Acco rding to ( 10.15) we ha ve tha t

(10.17) t; = -f .2;;
N,hl d S!' - f :e,
N ,q. 1d S!' ( 10.24 )
210 Introducti on to the tini te element method FE fo r mulation of 2-D and 3-D h eat flo w 211
We recall that the boundary conditions specify the flux q, = h along 2;, whereas the Fo r such a line source, the compon ent j., of the load vector given by (l 0.26) becom es
flux q, along .lf, is unspecified beforehand. Th erefore, (l 0.24) may be rewritten as

fb' = -t.. N, q, t dZ (l0.25)


.!., = L N,(x, y) Qt dA = r:r: N ,(x, Y)Q,~(x - a)~(y - b )t dx dy

Fro m ( 10.15) a component of th e load vecto r is given by = N,(a , b )t(a, b)Q, ( 10.32)

f li = LN, Qr dA ( 10.26)
If the position (a, b) of the line so urce coincides with the position of nod al point i,
th en N,( a, b) = N,(x" y,) = I, i.e. we obtai n I
Using (10.25) and ( 10.26) in ( 10.23) lead s to [ = Q,t (x" y,) if the line sou rce is located at nodal point i ( 10.33)

fi = _,I: ( N,) q, t dZ +f ( N,) QtdA As in the one-dimensiona l case of Cha pter 9, it follows th at an y di stributed load Q
1"'1 j,2' 1=1 A 1''' 1
may be replaced by line sources locat ed at the nodal points.
which with (7.139) results in The glob al stiffness matrix K and the global load vector f, given by (10.15) are
obtained by integration o ver the entire region A. These integrat ion s may be o btai ned
fi = _,I: q,rdZ +f Qtd A ( 10.27) as a summatio n of integrations over each element. In thi s way we are led to the
1"" 1 Jz A
expanded element stiffness matrix K " for element IX and the ex panded element load
A co mpariso n with (l0.22) shows th at vector t:' for element IX given by

I fi=O
i= 1 I (l0.28)
K" = f A,
BTDBt dA

in accordance with (9.68) . This means that the balance principle for the body is
expressed by the fact that the sum of the components of the force vector f is equal
t:' = f A,
N TQt dA
(10 .34)

to zero . We emphasize that (10.28) holds exactly even though the FE method is an
approximate approach. wher e A, is the region of element IX. Moreover, it follows directly that
The components of the load vect or f, are given by (l0.26). Let us assume now
th at the load Q is given in terms of a line source, where the heat supply is concentrated

to a point in the xy-plane and at this point a heat supply Q, per unit time and per K = '" K:e
L
<ll = 1
unit thickness of the region is prescribed. Q, is the strength of the line source and ( 10.35)
has the dimension [J/s m]. In such a situation, we may express Q using Dirac's delta
",
fl = L~:
function in a similar way to (9.70)- (9.72), i.e. :1 = 1

Q =Q.~ (x _a)~ (Y _b ) = { OO if (x ,y)=(a, b) ( 10.29) where nel denotes the total number of elements and where K:e and ff: refer to the
o otherwise pertinent quantities for element IX. Conceptuall y, relations ( 10.34) and ( 10.35) are of
where (a, b) is th e position of th e line source. By definition, we have that fundamental importance, but just as for the one-dimensional FE formulation these
relation s lead to a very inefficient co mputer implementation . In o rder to identify the
rJ~~ Q dx dy = rJ~~ Q,~(x - a)~(y - b) dx dy = Q, ( 10.30 ) non-zero components of K" and ff' directly , we consider th e F E formulation for an
element using on ly th ose degrees of freedo m which belong to th at specific element.
As a result of ( 10.29 ), (l 0.30) may also be writt en as The approx imation of the temperature over each element is given by

Jb- L
fb+ a - Q, ~ ( x - a ) ~ (y -b )dxdY =Q, ( 10.31)
where N' is the elem ent sha pe function matrix and a' co ntains the temperatures at

( 10.36)
212 In trodu cti on to th e fin i te element method FE fo rm ulation of 2-D an d 3-D heat flo w 213

the nodal points of the eleme nt, i.e.

i,
N' = [N1 N'2 .. . N']'
n., a' = [i,] (10.37)
I '
,I
- - - t- - - j-- -
r
,I
I

- l-- - - global

1;" I, A II

- - - -I- - - t-- - '!"' - - form ula tion


where n, denotes the nu mber of nodal points for the element. F ro m (10.36) we obtain r I I
I I I
"I T = B'a' where B' = "IN' ( 10.38) I r '

i.e,

ON1 aN, ... ON~] r--- --- -- - -,


I I
Be =
ax ax ax ( 10.39 ) I I
I I ex pa nded elem ent
aN'_I aN' en:
[_
~
_ _ 2 ----.!:!!
oy oy oy I
I
Ao. I'
I fo r mulat ion
A com pa riso n with (10 .14) -(10.18) shows directly that the FE formu latio n for one
element is given by
I
I
~
T~ N Q K" i
'
J

I K'a' = r ( 10.40)
ele men t
where the element fo rce vector r is given by

fo rmu la tion
( 10.41)

Moreover, the element stiffness matrix K' , the element boundary vector ~ and the
Figure 10.2 The three different FE formulations
element load vector t: for element a ar e given by

K' = f A,
B' TDB't d A approximation T = N a, a pplicable to the entire bod y, is ado pted. In the expa nded
elemen t formulat ion , one element is considered, but the global ap proxi matio n T = Na

~= -f N 'Thr d.5l' -
go""
f 9',..
N'T q t d.5l' ( 10.42)
is used. Finally, in the element formula tio n, one element is consi dered and the
ap proximatio n T = N ' a', applicable to this particula r eleme nt, is used.
In a computer program , the element sti/fness K' and the element load vecto r t:
t: = f A.
N'TQt d A
are determined by (10.42) and the contri butio ns to the glob al stiffness matrix K and
the global load vect or f, a re obtained using the topology dat a. The boundar y vecto r
, fb is alwa ys derived using the form given by (10.15). It appears that the approach
where A, is the region of element a, .!ih, is that part of the element boundary on described is completely similar to that discussed for the F E formulation of
which the flux q. = h is kn own, whereas SI'g, is the remaining part of the element one-dimensio nal heat flow. We also obser ve that just as for one-dimensional heat
boundary. We recall that the boundary conditions along an element boundary are flow, the stiffness matrix K given by ( 10.15) contains first derivatives of the global
unknown, except when an element boundary coincides with the boundary of the body. shape functions , i.e. we again require that these shape functions fulfil CO-continuity.
Similar to Figure 9.14, the thre e different FE formulations - global formulation, When the nodal temperature a ha s been determined by ( 10.18), the temperature
expanded elemen t formulation a nd element formul at ion - are illustrated in Figure T at an arbitr ary point in an element is given by ( 10.36) and the temperature gradient
' 10.2. We recall that in the global formulation the entire body is considered and the "I T is given by (10.38). F rom the temperature gradient, the flux vector q at any place
214 Introduct ion to th e fini te elemen t method FE fo r mulatio n of 2-D and 3-D h eat fl ow 215

in the body is obtained from Fourier's law (10.3). Therefore, when a has been y y
01 / q , = h =30 bl
determined from (10.18) all quantities of interest can be derived.
y ' o ttIlIe
k=4
10.1.1 Example Q =45
'-.,.- t = 1 -rx--: x
CD
I""-------b-- _ x
Consider the square panel shown in Figure 10.3, where the boundary along the y-axis 8 2
is insulated (i.e. q, = h = 0) and a constant flux q, = h = 30 Ji m' s is prescribed
along the boundaries y = I m and y = - I m. A constant temperature T = 10 ' C is Figure 10.4 (a ) Problem reformulation ; (b) finite element mesh
prescribed along x = 2 m and a constant heat supply Q = 45 Jim's is applied all
over the panel. The thickness t is constant and is equal to I m. The material is assumed
to be isotropic and homogeneous, i.e. according to (6.13) and (6.14) the constil~e
matrix D can be written as
h

(10.43)

where the thermal conductivity is given by k = 4 J r C m s. "


It is obvious that the heat flow in the panel is symmetric about the x-axis, i.e.
no heat flows across the x-axis. We may therefore reformulate the problem as shown
in Figure 1O.4(a).
In order to solve this problem by the FE method using hand calculations only, Figure 10.5 Body with triangular base and height h
we adopt the extremely simple FE model consisting of two three-node triangles as
shown in Figure 10.4(b). The element numbers and the global nodal numbers are
also indicated on this figure.
triangle (cf. (7.99)), we obtain
The element stiffness matrix K' and the element load vector Ii are given by
(10.42). With (10.43) and noting that B' is a constant matrix for the three-node Ke = kBeTBetA.x (10.44)
where

B' - I [Y j - Y, Yk - Y, Y, - Yj J (10.45)
y
q n = h = 30 2A(J Xk - xj Xi - Xk Xj - Xi

-: As both the load Q and the thickness t are constants, (10.42) yields
7
/

~/
I I I I I

kd
....-T= 10 Ii ~ Qt f A.
N' TdA (10.46)

in sul at e-d Q=45 x 2


~ The element shape function matrix N' is given by (7.92) and (10.46) can be derived
t =1 using standard rules for area integration. In the present case, however, the shape
~ functions vary linearly, which means that (10.46) may be evaluated in a very simple
/
I I I I j manner. For this purpose consider the body shown in Figure 10.5. The volume Vof
~ q =h = 30 this body is known to be given by v
<>,
2
V = thA (10.47)
where h is the height and A the area of the triangle which forms the base of the body.
Figure 10.3 Square panel with heat flow Reference to Figure 7.23 shows that the element shape functions vary in the
216 Introduction to the finite element method FE lormulation of 2-D and 3-D heat flow 217

y Next we consider element 2 of Figure 10.6(b ) where i = I,j = 2 and k = 3, i.e.


01 bJ Y

3 3 Xj= Yi =O; xj= 2, Yj= 1; xlt = O, Yt = l to _ I


12 .1)
The area A , is A , = I and B' as g iven ~\ ( 10.45) becomes l

(j)
/,<::::..----~-_ x bC'-- - - - - - x B' = -
l[ 0 I - I J ~C.\ ' 0" -.:
t 'f
. Q
0
I
-4
- (10.51 )
2 2 -2 0 2 _ ~ .
_. \ A
~ J . -<
Figure 10.6 Local nodal points for element I and 2 From ( 10.44) and (10.48) the results for the second element are then given by

man ner given by Figure 10.5, i.e. from ( 10.46) we obtain with ( 10.47)
x- { _: =:J H{]
~ ~/A.[i]
= (10.48)
Comparing Figures 1O.4(b) and 1O.6(b), the to pology data for this element is th ~
the local nod al points 1, 2, 3 correspond to the global nodal point s 1,3,4 respectively.
This result is certainly not sur prising as it shows that a constant load Q contributes Fo r clarity, and using this to pology data, the expanded element stiffness matrix K"
equal amounts to each nodal point. and the expanded element load vecto r ~' are
The local nodal points of the two elements are shown in Figure 10.6 and in
4 0 0 - 4
accordance with Chapter 7 these nodal points are listed in the counter-clockwise
direction 1, 2, 3. We first consider element 1 where i = I , j = 2 and k = 3, i.e. 0 0 0 0 0
K;c = m= 15
1).....2 3' 'l x, = Y, = 0; Xj = 2, Yj = 0; x. = 2, Y. = I 0 0 I - I I

With ,these values and as the area A , = I, B' as given by (10.45) becomes - 4 0 - 1 5

- I. 0l B, ='~ [ - I, . I OJ ~ ''1 0, -ite (.) Adding these results to (10.50) we then obt ain
' -2 . 2 O. - 2 2 ....,. 1.,0) I"C; _ 'I (10.49)
5 - I 0 - 4 2
Frdfu ( 10.44) and ( 10.48) and as the thickness t = I, the results for the first element
are given by 'f . ' l . - I 5 ~4 0 I
v
K= f, = 15 ( 10.52)
0 - 4 5 -I 2

~
-I
5 - 4 0 - I 5
K' = [ - - \ o
- 4 5 4 Let us now evaluate the boundary vector fb given by ( 10.15). From Figure 10.4
. F' ", - 4 Ll- we conclude that
Companng igures 1O.4(b) and 1O.6(a), local nodal points 1,2,3 correspond to the
global nodal points I, 2, 3 respectively. That is, after the contributions from the first
element, the global stiffness matrix K and the global load vector f, are given by

I - I 0 0 i.e.

K = Kit =
- I
0 - 4
5 - 4
4
0
0
f, = ~; = 15
I
I
( 10.50 ) r, = -I J .Poe
NTh dY - t J .?oc
NTq.dY ( 10.53)

0 0 0 0 0 The only shape functions which differ from zero along the bound ary DC are N 3 and
218 Int ro d uction to the tini te element method FE formu lation of 2-D and 3-D heat flo w 219
N 4 . Moreover, as they vary linearly and as h = 30 along DC we have . Use of this solution in the second and third row of ( 10.57) implies that
o o ( 10.59)
f
)
N2Q. d2' = 25; f N 3 Q. d2' = 5
o o f -SfIlC
fPoe
N Th d2' = 30
Zoe N3
d2' = 30
I
I
( 10.54)
!t'BC

As N 2 and N 3 are positive functions, (10.59) implies that Q. is positive along the
boundar y BC, i.e. we conclude that in order to maintain the temperature at T = 10
along the bou nda ry BC, heat must be extracted from the body along this bou nda ry.
Likewise, it follows tha t
With the result of (10.59) it follows from (10.57) that the balanceprinciple for the
o body is satisfied in accorda nce with (10.28).
Let us finally determine the flux vecto r q in each element. Fro m ( 10.3) and (1O.3ll'T
N 2q.d2' we get
f .roc
( 10.55) q = - DVT = - kV T = - kB' a' 1
f .roc
N 3q. d2'

o
where k = 4. For element I we obta in with (10.49) that

- I 1
Using( 10.53)- ( 10.55) and as the thickness t = I, the boundary vector fb is given by q= - 4 x t[ 0
- 2
o
whereas for element 2, where ( 10.51) holds, we get

fb = -
f
.r,c
N2q . d2'

f
( 10.56) I
q= - 4 x t[ 0
30 + N 3 q. d2' - 2 o
.r,c
30
The FE equations for the body are o btained from (10.52) and (10.56), i.e. 10.1.2 Two-dimension al heat fl ow with convection
0 Previously, we assumed that the bound ary 2' consists of the two part s, 2'" and Sf"

- 1
5 - I
5 - 4
0 - 4
0
T,
10
f .roc
N 2 q. d2'
2
I
where the flux Q. and the temperature T are prescribed, respectively. Let us now also
assume that convection occurs on part of the boundary and let .II; denote that part .

f
= - + 15 ( 10.57) Referring to (10.15) this affects only the boundary vector f b , which now becomes
0 - 4 5 - I 10 2
30 + N 3 Q. d2'
- 4 0 - I 5 T4 .roc fb = - f N Th td 2' - f N T Q. t d2'- f N T Q. t d2' ( 10.60)
2'/, !e, .!f'e
30
where 2'", Sf, and .II; now comprise the entire boundary, as shown in Figure 10.7.
where T2 = T3 = 10. From the first and four th row we obtain
The convection along .II; is again given by Newton's convection bound ary condition
- 4] [ T1 ] = [ IO- 0 + 30]= [ 40] (9.143), i.e.
5 T4 10 -30 +1 5 -5 Q.= ~ (T -Too ) ( 10.61)
with the solution To consider the influence of convection, the procedure discussed in Chapter 9 is
followed, i.e. the unknown temperature Tin (10.61)is replaced by T = Na. This gives
(10.58)
220 FE for m ulatio n of 2-D an d 3-D heat flow 221
Introduction to the fin i te element method

where V is the volume of the body and S is the bo undary surface which con sists of
y S. and Sg, where the flux q. = h and the temperature T = g are prescri bed, respectively.
The temperature is agai n approximated by mea ns of
(10.65)
T = Na
q .h

m
n where

'-- - - - -- - - - - ,
N - [ N, N, ... N." a ~ ( 10.66)

Figure 10.7 Two-dimensional region with boundary !t' = !t'll + !l', + !,
and n is the number of nodal points for the enti re bod y and a com ponent N , now
and use of this expression in ( 10.60) result s in depen ds on x, y and z, i.e. N, = N,(x, y,z). F rom ( 10.65) we o btain

r, = -f NTht d2' -
..r"
f.:t;
NTq,t d2' - (f ~NTNt
s;
d2')a + Too f NT~t
s;
d2' i.e.
VT = Ba where B = VN (10.67 )

With this expression , the FE formulation of ( 10.16) takes th e form


aN, aN, ou,
ax ax ax
(K + K, )a = -f "'I
NTht d2' - f-'4
NTq. t d2' + Too f NT,t d2'
g;
+ f,
(10.62)
B=
aN, -aN,-
ay ay
aN.
ay
( 10.68)

whe re K, = fg; a NTN t d2'


aN, aN,
az az
aN.
az
The der ivati on of the F E formulation from ( 10.64 ) using the approximation
It appears that a modification of the stiffness matrix occurs. Thi s new stiffness ma trix
above in combination with the Galerkin method follows th e same procedure as for
is symmetric, just like K, and similar to the discussion of (9.149) it follows that the
mod ified stiffness matrix is non- singul ar , i.e. two-dimensional heat flow. We may th erefore write the result directly as

I det (K + K, ) "" 0 I ( 10.63) I Ka = f I ( 10.69)

We fina lly mention tha t we are a ble to consider con vection occurring along the where
lateral surfaces of the region A, i.e. convection normal to the xy-plane. T his situ ation (10.70)
ca n be dea lt with by a mod ification of the load Q which lead s to a modification of f = fb + f,
the load vecto r f, as in th e procedure for one-dimensional hea t flow (cf. (9.152),
(9.154) and (9.155). Thi s modifica tion is straightforward and is left as an exercise and
for the interested reader.
K = LBTDBdV
10,2 Three-dimensional heat flow
The wea k form of three-dimensional heat flow is given by (6.49) and (6.50), i.e. f b = - f N ThdS - f N Tq. dS ( 10.71)

iv (V v)TDV T d V= - r vh dS - Js,
.,
r vq. dS+ Jv
r vQ d V
( 10.64 )
s,

f, = LNTQ d V
s,

T = g on S,
2N