NOVEMBER 1966 TECHNICAL NOTES 2035

Engineering Applications of Numerical - = node with eJ

— = node with 9 only
Integration in Stiffness Methods (7) = 7 degrees of freedom

BKUCE 1\L IRONS* Fig. 1 Assembly of

University of Wales, Swansea, Wales modular plane ele-
ments.
Nomenclature
a, 6, c — dimensionless coordinates in brick and quadrilateral
elements
D = modulus matrix
H = matrix converting deflections to strains
/ = Jacobian matrix; |/| = determinant of /
k = factor for numerical integration describe the curved boundaries shown in Fig. 1. A is a
LJ = area coordinates for a triangle; volume coordinates triangle with only three nodes, and the three nodal values
for a tetrahedron define 6 = K& + K^y + K 3 ; A is known as a grade 1 element
Pi — displacement function because 6 varies linearly along the sides. B is also grade 1
s = sampling point for numerical integration because the fourth degree of freedom is not allowed to inter-
u, v, w = deflections fere with the linearity along the sides. C is the first grade
x, y, z = Cartesian coordinates 2 element, whose six nodal values define 6 = KI -f- K%x +
K$y + K±x2 + K&y + K6y*. Because 6 is quadratic along
Downloaded by STANFORD UNIVERSITY on May 21, 2013 | http://arc.aiaa.org | DOI: 10.2514/3.3836

Introduction each side, the midside nodes are needed to insure conformity

I N a finite element solution, the first and crucial step is to

assume an interpolation formula expressing the deflections
at any point within the element in terms of the deflections
with neighboring elements. Veubeke published this element
in 1964, but the first published numerical results were by
Argyris in 1965.1 This solution is modified in a simple but
(and sometimes their derivatives) at the "nodes," which are far-reaching manner, by nominating as the midside nodal
usually at the corners of the element. It is then possible to variables the departure from linearity. If a midside variable
express the various energy densities needed in terms of the is ignored, 6 becomes linear along that side; thus element D
coefficients in the formula (known as "displacement func- can merge with a grade 1 element A along one side and with
tions" p) and their x, y, z derivatives. In the methods de- a grade 2 element E along another. It has been degraded by
scribed here, the displacement functions and their deriva- losing a degree of freedom along one side without affecting
tives are calculated explicitly in a subroutine, which can be the other two. E is the grade 2 quadrilateral that generalizes
checked numerically by comparing values at two points a B just as C generalizes A; again it is basically the grade 1
short distance apart, and by calculating values at the nodes, element B with grade 2 features introduced only by the mid-
etc. The integrations over the element are done numerically side nodes, and where these are removed (as in elements F)
in another subroutine, which can itself be checked using a it can blend with grade 1 elements.
simple element geometry for the purpose. Thus various The grade 3 elements derive from plate bending research.
element geometry subroutines can be plugged into programs Special "osculatory" nodes are introduced, having as nodal
solving different physical problems with some confidence, variables 6, d#/do;, c)0/dz/ (Ref. 9). Thus slope continuity is
giving new versatility to finite element techniques. The imposed at these nodes, wrhich are represented by circles in
pioneer work on stiffnesses was done by I. C. Taig but was Fig. 1. G is a grade 3 triangle with 6 defined by 9 nodal
not published, and the author recognizes Taig's quadrilateral variables; and 6 is a cubic along each side, defined by two
element of 1959 as the first of a powerful series of elements, slopes and two values of 6. The Zienkiewicz "nonconform-
whose higher members have curved edges and about which ing" cubic functions 2 are an obvious choice. Similarly H is
certain convergence theorems are here derived. a grade 3 quadrilateral, and several versions are known.3'4
The choice between the use of a few complicated elements As before, grade 2 and 3 elements are joined by degrading
made available by numerical integration, or many simple the higher, and J must have lost a degree of freedom because
elements, is not easily made. Sometimes physical features its boundary with C is grade 2. H would be degraded to J
dictate the use of elements smaller than those the engineer con- by writing the lost slopes in terms of the midside deflections,
siders desirable; sometimes the requirement of a fine mesh, and in addition to the remaining slopes and deflections, assuming
therefore very small elements, over a region of rapidly varying quadratic variation along the right-hand side and cubic along
stress leads to bad numerical conditioning so that larger and the other three. It is possible to degrade any side of an
more complicated elements are mandatory. element in this way without altering the grades of the other
Ideally the engineer should keep a well-graded stock of sides. K is a twice-degraded version of H, being grade 2
elements from the simplest to the most complicated; he along two adjacent sides.
should be able to combine them freely in a structure and he The functions for these elements are programed so as to
must be able to modify his choice promptly when a calcula- produce 6 and the first derivatives of 6 at a chosen point due
tion fails because of a misjudgement in idealization. This to each nodal deflection in turn. Then we can use numerical
paper suggests techniques based on numerical integration integration to solve the following twro-dimensional problems
which virtually achieve these goals. with little individual programing: 1) plane stress, plane
strain, including consistent mass matrices for dynamic
Plane Elements calculations, and the equivalent for stability; 2) torsion of a
Figure 1 shows an assembly of 16 plane elements of various shaft, a body of revolution of arbitrary shape; 3) body of
grades of complexity. In accordance with current practice, revolution with axisymmetric loading; 4) torsion, shearing
interpolation formulas express any deflection variable 6 in stresses and shear center of a prism, torsion bending con-
terms of values at the nodes of the element concerned. There- stant; 5) body of revolution with load varying sinusoidally
fore accurate interpolation requires an element with many around it. In this case there are three deflections at any
nodal variables; further, the nodal coordinates naturally point, giving nine variables at an osculatory node.
Nodal Efficiency
Received August 20, 1965; revision received April 11, 1966.
* Lecturer, University College of Swansea; formerly at We have described two essentially different approaches to
Rolls-Royce Ltd., Derby, England. higher elements, namely those with midside nodes and those
 2036 AIAA JOURNAL VOL. 4, NO. 11

resented by using the low-order integrating rule. In the

limit, the stress, elastic properties, etc. are sensibly constant
over any element, so the representation is adequate and
the solution still converges, although not monotonically.
(See also Ref. 2, addendum.) It is found that grade 1 ele-
ments require at least a 1 X 1 Gauss rule, grade 2 a 2 X 2
rule, and grade 3 a 3 X 3 rule.

Solid Elements
The tetrahedron is the simplest solid element, and the
Fig. 2 Grade 2 solid elements.
grade 1 version has displacement functions pi = Liy the four
volume coordinates in turn. (This is an unsuccessful ele-
ment.) The grade 2 tetrahedron with midside nodes has a
with osculatory nodes. But it is possible to dispense with typical p for midside nonlinearity = 4L2L3. Essentially the
osculatory nodes by introducing two ordinary nodes at -£ and same element has been used with great success by Argyris,1
f of the distance along each side, and at first sight this seems but it remains a relatively inefficient and cumbersome element.
preferable. It is not. We define "nodal valency" as the One osculatory node could be introduced if the three op-
number of elements attaching to a node. posite midside nodes were retained; but we prefer elements
all elements based on the rectangular brick, at whose eight vertices the
dimensionless coordinates a, &, and c take values dbl. The
Downloaded by STANFORD UNIVERSITY on May 21, 2013 | http://arc.aiaa.org | DOI: 10.2514/3.3836

2_] (degrees of freedom/element) =

grade 1 brick has pt = (1 ± a)(l ± 6)(1 ± c)/8. (See
all nodes Appendix 1.) The grade 2 brick has 20 nodes including
zL (variables/node) X (nodal valency) = those at mid-edge, whose p are typified by (1 — a 2 )(l ± b)
(number of variables) X (mean nodal valency) (1 ± c)/4. Its more efficient form has alternate osculatory
and ordinary nodes. Let a = b = c = — 1 be an osculatory
Thus the cost of a banded solution varies approximately as node. The three basic slope functions giving unit a, 6, and
(mean nodal valency) ~3. A corner node usually has a valency c derivatives are pa = (1 — a2) (I — b)(l — c)/8, pb =
of 4 or 6, whereas a midside node usually has a valency of (1 - a)(l - 6 2 )(1 - c)/8, and pc = (1 - a)(l - 6)(1 -
2, so that tradition is correct in favoring corner nodes. In c2)/8. The function giving unit x derivative is px = padx/da
particular, osculatory corner nodes will cost less than midside + pbdx/db + pcd£/dc, where the x derivatives are at the
nodes. osculatory node in question and are tabulated in the general-
The grade 2 quadrilateral exists in a more efficient form, ized nodal coordinates xif A typical function for unit dis-
using osculatory nodes at two opposite corners and ordinary placement is p = (1 — a)(l — b)(l — c)/8 — (pa + pb +
nodes at the others; along any boundary the two types of P<)/2.
node alternate. The mean nodal valency for an extensive The grade 3 brick is regarded as the most promising solid
mesh is increased by 50% from f to 4. Meshes exist into element, with a 96 X 96 stiffness matrix. Following a pre-
which these elements do not fit, but this is not an over-riding cedent from plate bending theory, it is observed that poly-
objection. The greatest difficulty with osculatory nodes nomial terms like 1, #, xy, xyz, x2, x2y, x2yz, £3, x3y, and x*yz
arises when the moduli change discontinuously across a obey the biharmonic equation, and by good fortune there are
bound ary. 32 of them. (This is not a coincidence. The 4-point plane
and the 8-point solid grade 1 elements are "harmonic" in
Curved Boundaries and Convergence Theorems the same sense, and the degrees of freedom are correct for
Let the displacement functions be pi so that the deflections "triharmonic" elements with nodes osculating to second
Uj v may be expressed in terms of the nodal deflections u = derivatives.) Replacing x, y, z in the biharmonic solution
2 ptUi, v = 2 piVi. It is assumed that the pt are functions by the dimensionless coordinates a, b, c, the eight nodes may
of subsidiary coordinates a, b. Even in the most complicated be written (a,-, fe», d) = (±1, ±1, ±1), i = I to 8. If a0
elements discussed here it is possible to write the coordinates = aa,i, etc., the response at a point (a, 6, c) to unit nodal de-
x, y themselves in terms of the same pi and with somewhat flection Ui is
artificial values of the "nodal coordinates" Xi yi} so that
x = 2 piXi and y = ^ piyf. This definition permits curved u = (OQ + l)(6o + l)(co + 1)(2 + a0 + fro + a0
boundaries, and in every case the boundaries conform before
and after deformation. The solutions are independent of c02)/16
choice of origin and rotation of axes. From this, with the
definition of x and yy follows the proposition that rigid body
Table 1 Gauss and Radau integrating constants
motions and states of constant strain are faithfully repro-
duced if the nodal deflections require it. The subdivision Si ki Si* ki*
criterion of Melosh5 is stronger and it also holds, except n = 1 0. 5 1. 0 0.3333333333 0..75
that subdivisions must be applied in pairs in the grade 2 (1. 0) (0.25)
cases with alternate osculating nodes. Convergence is amply n = 2 0. 2113248654 0. 5 0. 1550510257 0. 3764030627
0. 7886751346 0. 5 0. 6449489743 0. 5124858262
proved, except perhaps in pathological cases; for example, the (1. 0) (0. 1111111111)
Jacobian determinant \J\ must not change sign. n = 3 0. 1127016654 0. 2777777778 0. 0885879595 0. 2204622112
0. 5 0. 4444444444 0. 4094668644 0. 3881934688
If \J\ is a polynomial in a and 6, a certain order of integra- 0. 8872983346 0. 2777777778 0..7876594618 0. 3288443200
tion rule will always give the area of the element accurately. (1. 0)
0. 0571041961
(0. 0625)
0. 1437135608
In general, this rule does not give accurate elastic stiffnesses. n = 4 0. 0694318442 0. 1739274226
0. 3300094782 0. 3260725774 0..2768430136 0. 2813560151
But if, for example, the integral giving a typical nodal load 0. 6699905218 0. 3260725774 0. 5835904324 0. 3118265230
for an element with the x tensile stress constant and unity 0..9305681558 0.,1739274226 0..8602401357 0. 2231039011

0..1184634425
(1..0)
0..0398098571
(0..04)
0. 1007941926
n = 5 0..0469100770
Pi = // da db 0..2307653449 0..2393143353 0..1980134179 0.. 2084506672
0.5 0. 2844444444 0.4379748102 0..2604633916
is expanded, it is found to contain the same powers of a and 0.7692346551 0..2393143353 0. 6954642734 0..2426935942
b as // | J\ da db. It follows that a small element under 0.9530899230 0.1184634425 0.9014649142 0. 1598203766
constant stress gives correct nodal loads even if it is misrep-
(1 .0) (0..0277777778)
 NOVEMBER 1966 TECHNICAL NOTES 2037

and the response to unit dw/da at node i is ume

N N N
u = at-(oo + 1)(60 + l)(cb stiffness matrix = ^ £ X)
Finite element theory is too often suffocated by algebra, but 1 1 1
this last example, perhaps the ultimate in complication, where the k and the values of a, 6, and c are taken from a
shows how numerical integration genuinely simplifies things. table of Gauss rules (e.g., Ref. 7, Appendix 4.1). Virtual
The basic displacement functions for the 96 X 96 brick work calculations of equivalent nodal forces are also con-
element are calculated, with their derivatives, in a subroutine veniently done by numerical integration. For thermal
of only 37 FORTRAN statements. loads it is easiest to append a column (— aT, —aT, —aT,
0, 0, 0)r to Hj where a is the coefficient of expansion and T
Appendix 1. Details of Grade 1 Brick the temperature; the extra column of the stiffness matrix
This element is the three-dimensional equivalent of gives the load contribution.
Taig's quadrilateral element, which he integrated analytically An efficient method of calculating HTDH has been de-
in 1958 and reported in 1961.6 It exemplifies most of the scribed.8 It is equally efficient to reduce the factored D by
processes described in general terms in the body of this the Choleski square root process, D (factored) = LU, to
paper. The nodes are ordered as in Fig. 2b so that (at-, 6», over-write H with UH, and to calculate the required scalar
cO = (-1, -1, - !),'(!, -1, -1), (1, 1, -1), (-1, 1, -1), products of columns. Both methods could fail with the
(-1, -1, 1), (1, -1, 1), (1, 1, 1), and (-1, 1, 1), for i = 1 indefinite matrices arising in stability calculations.
to 8.
Appendix 2 Integration Rules for Triangles
)(l + cc<)/8
Downloaded by STANFORD UNIVERSITY on May 21, 2013 | http://arc.aiaa.org | DOI: 10.2514/3.3836

It is natural to treat the triangle as a degenerate case of

x = (1 - a)(l - - c)xi/S 4- the quadrilateral, either with two coincident vertices, or with
the fourth vertex at a midside of the triangle. In either
case, one degree of freedom is lost as compared with the
x = + • • • + PsX% quadrilateral; but the accuracy can be restored in the case
with coincident vertices by using a Radau rule instead of
y Gauss, introducing an integrating point at the coincident
a,-(l + &&<)(! + cc<)/8 vertices where the combined factor is zero. The required
rules have been derived, and are given in Table 1 for 1, 4,
(1 + aai)bi(l + cd)/8 9, 16, and 25 integrating points within the triangle. They
are accurate to 1st, 3rd, 5th, 7th, and 9th powers of x and y,
or of the area coordinates LI, L2, L3. (The ninth powers, for
fbx/da example, include xHf or L-fL^L^.} In Ref. 1 the 16-point
[J] = 5x/56 double Gauss rule was adequate for a triangular element with
|_d x/'dc dt//dc singularities at the vertices, so it is expected that these rules
will suffice for practical purposes.

ff /(Li, L2) L3) dxdy = Z E fc.

= JJ
i = l j = l

. . . dps/()c_ where
Li = s^ L2 = Sj(l — LI) L3 = 1 — LI — L2
k = 2fo*fc y (area)(l - Li)
References
1
The deflections - at the nodes give deflections else- Argyris, J. H.; "Continua and discontinua," Opening paper,
where. Conference on Matrix Methods in Structural Analysis, Wright-
Patterson Air Force Base (October 1965).
2
Bazeley, G. P., Cheung, Y. K., Irons, B. M., and Zienkiewicz,
0. C., "Triangular elements in bending— conforming and non-
The strains in #, #, z axes can be written conforming solutions," Conference on Matrix Methods in
Structural Analysis, Wright-Patterson Air Force Base, Session

" dp/cta
0
0
bp/5^/
0
0
-i »!
Ui 3, Paper 7 (October 1965).
3
Zienkiewicz, O. C. and Cheung, Y. K., "The finite element
method of analysis for arch dam shells and comparison with
finite difference procedures," International Symposium on
Ui Theory of Arch Dams, Southampton, Session 2, Paper 2, (April
0 0 dp/52
dp/dy dp/do; 0 ' • 1964); also Theory of Arch Dams (Pergamon Press Ltd., London,
1965).
bp/cte 0 dp/do;
t-
4
Sander, G., "Upper and lower bounds in matrix analysis of
^8 plates in bending," University of Liege, Belgium, Rept. SA-2
IMS (November 1964).
5
Melosh, R. J., "Basis for derivation of matrices for the direct
[H = } (i = 1 to 8) stiffness method," AIAA J. 1, 1790-1795 (1963).
6
Taig, I. C., "Structural analysis by the displacement
Thus at a point in the element, [H] is assembled term by method," English Electric Aviation Ltd., Rept. SO 17 (1961);
term from the subroutine output data. Matrix notation is unpublished.
7
inappropriate, especially, for example, in the case of sinu- Kopal, Z., Numerical Analysis (Chapman and Hall, London,
1955).
soidal loading of a body of revolution. [H] may include the 8
Irons, B. M., "Comment on 'Distributed mass matrix for plate
Pi themselves, or second derivatives in bending problems. element bending/ " AIAA J. 4, 189 (1966).
The strain.energy density can be expressed as %(ui . . . w?8) 9
Ergatoudis, J. G., "Quadrilateral elements in plane analysis
HTDH(ui . . . tpg)r, where D here is a familiar 6 X 6 modulus and introduction to solid analysis," M. Sc. Thesis, University of
matrix. Differentiating twice and integrating over the vol- Wales, Swansea, Wales (June 1966).