Felippa Nonlinear Finite Element Method
Felippa Nonlinear Finite Element Method
Carlos A. Felippa
Department of Aerospace Engineering Sciences and
Center for Space Structures and Controls
University of Colorado
Boulder, Colorado 80309-0429, USA
August 2001
Material assembled from lecture notes for the course Nonlinear Finite Elements Methods, offered
since 1987 to date at the Aerospace Engineering Sciences Department of the University of Colorado
at Boulder.
Preface
This textbook presents an Introduction to the computer-based simulation of nonlinear structures by
the Finite Element Method (FEM). It assembles the still “unconverged” lecture notes of Nonlinear
Finite Element Methods or NFEM. This is an advanced graduate course offered in the Aerospace
Engineering Sciences of the University of Colorado at Boulder.
NFEM was first taught on the Spring Semester 1986 and has been repeated every two or three years.
Unlike the Introduction to Finite Element Methods (IFEM), NFEM is not a core course. It is
typicall taken by second year graduate students that are interested in the topic of nonlinear simulation
of mechanical systems.
Prerequisites for the course are an introductory course in finite elements such as IFEM, graduate-level
calculus, linear algebra, knowledge of structural mechanics at the Mechanics of Materials level, and
ability to program in a higher level language such as Matlab or Mathematica.
The course originally used Fortran 77 as computer implementation language. This has been gradually
changed to Mathematica since 1995. Unlike IFEM the changeover is not yet complete since the course
has been offered only twice since.
Book Objectives
(To be completed)
Book Organization
(To be completed)
Exercises
Each Chapter is followed by a list of homework exercises that pose problems of varying difficulty.
Each exercise is labeled by a tag of the form
[type:rating]
The type is indicated by letters A, C, D or N for exercises to be answered primarily by analytical
work, computer programming, descriptive narration, and numerical calculations, respectively. Some
exercises involve a combination of these traits, in which case a combination of letters separated
by + is used, e.g., A+N. For some problems heavy analytical work may be helped by the use of a
computer-algebra system, in which case the type is identified as A/C.
The rating is a number between 5 and 50 that estimates the degree of difficulty of an Exercise, in the
following “logarithmic” scale:
5 A simple question that can be answered in seconds, or is already answered in the text if the
student has read and understood the material.
10 A straightforward question that can be answered in minutes.
15 A relatively simple question that requires some thinking, and may take on the order of half to
one hour to answer.
i
20 Either a problem of moderate difficulty, or a straightforward one requiring lengthy computations
or some programming, normally taking one to six hours of work.
25 A scaled up version of the above, estimated to require six hours to one day of work.
30 A problem of moderate difficulty that normally requires on the order of one or two days of work.
Arriving at the answer may involve a combination of techniques, some background or reference
material, or lenghty but straightforward programming.
40 A difficult problem that may be solvable only by gifted and well prepared individual students,
or a team. Difficulties may be due to the need of correct formulation, advanced mathematics,
or high level programming. With the proper preparation, background and tools these problems
may be solved in hours or days, while remaining inaccessible to unprepared or average students.
50 A research problem, worthy of publication if solved.
Most Exercises have a rating of 15 or 20. Assigning three or four per week puts a load of roughly 5-10
hours of solution work, plus the time needed to prepare the answer material. Assignments of difficulty
25 or 30 are better handled by groups, or given in take-home exams. Assignments of difficulty over
30 are never assigned in the course, but provided as a challenge for an elite group.
Occasionally an Exercise has two or more distinct but related parts identified as items. In that case a
rating may be given for each item. For example: [A/C:15+20]. This does not mean that the exercise
as a whole has a difficulty of 35, because the scale is roughly logarithmic; the numbers simply rate
the expected effort per item.
(To be completed)
Acknowledgements
Thanks are due to students and colleagues who have provided valuable feedback on the original course
Notes, and helped metamorphosis into a textbook. Two invigorating sabbaticals in 1993 and 2001
provided blocks of time to develop, reformat and integrate material. The hospitality of Dr. P. G.
Bergan of Det Norske Veritas at Oslo, Norway and Professor E. Oñate of CIMNE/UPC at Barcelona,
Spain, during those sabbaticals is gratefully acknowleged.
ii
iii
Chapter Contents
Section
1 . . . . . . . . . . . . . . . . . . 1-1
2 . . . . . . . . . . . . . . . . . . 2-1
3 . . . . . . . . . . . . . . . . . . 3-1
4 . . . . . . . . . . . . . . . . . . 4-1
5 . . . . . . . . . . . . . . . . . . 5-1
6 . . . . . . . . . . . . . . . . . . 6-1
7 . . . . . . . . . . . . . . . . . . 7-1
8 . . . . . . . . . . . . . . . . . . 8-1
9 . . . . . . . . . . . . . . . . . . 9-1
10 . . . . . . . . . . . . . . . . . . 10-1
11 . . . . . . . . . . . . . . . . . . 11-1
12 . . . . . . . . . . . . . . . . . . 12-1
13 . . . . . . . . . . . . . . . . . . 13-1
14 . . . . . . . . . . . . . . . . . . 14-1
15 . . . . . . . . . . . . . . . . . . 15-1
16 . . . . . . . . . . . . . . . . . . 16-1
17 . . . . . . . . . . . . . . . . . . 17-1
18 . . . . . . . . . . . . . . . . . . 18-1
19 . . . . . . . . . . . . . . . . . . 19-1
20 . . . . . . . . . . . . . . . . . . 20-1
21 . . . . . . . . . . . . . . . . . . 21-1
22 . . . . . . . . . . . . . . . . . . 22-1
23 . . . . . . . . . . . . . . . . . . 23-1
24 . . . . . . . . . . . . . . . . . . 24-1
23 . . . . . . . . . . . . . . . . . . 23-1
24 . . . . . . . . . . . . . . . . . . 24-1
25 . . . . . . . . . . . . . . . . . . 25-1
26 . . . . . . . . . . . . . . . . . . 26-1
27 . . . . . . . . . . . . . . . . . . 27-1
28 . . . . . . . . . . . . . . . . . . 28-1
Appendices
A . . . . . . . . . . . . . . . . . . A-1
B . . . . . . . . . . . . . . . . . . B-1
C . . . . . . . . . . . . . . . . . . C-1
D . . . . . . . . . . . . . . . . . . D-1
H . . . . . . . . . . . . . . . . . . H-1
iv
v
14
.
The CR Formulation:
Space Bar
14–1
Chapter 14: THE CR FORMULATION: SPACE BAR 14–2
TABLE OF CONTENTS
Page
§14.1. Introduction 14–3
§14.2. Line Segment Moving in 3D 14–3
§14.2.1. Line Segment Derivatives . . . . . . . . . . . . . 14–4
§14.2.2. Derivatives of Length Functions . . . . . . . . . . . 14–5
§14.2.3. Mathematica Implementation and FD Verification . . . . . 14–6
§14.3. The CR Bar Element 14–9
§14.3.1. Internal Energy, Force and Stiffness . . . . . . . . . 14–9
§14.3.2. Matrices for Specific Strain Measures . . . . . . . . . 14–10
§14.3.3. Mathematica Implementation . . . . . . . . . . . 14–11
§14. Exercises . . . . . . . . . . . . . . . . . . . . . . 14–14
14–2
14–3 §14.2 LINE SEGMENT MOVING IN 3D
§14.1. Introduction
The expressions provided in the foregoing two Chapters and in Appendix R address the formation of
incremental equations for an arbitrary CR element. The main restriction is the assumption of small
local deformations, which permit the use of the linearized equations for the deformational energy.
Closed form of the incremental expressions become intractable for elements of arbitrary geometry,
such as curved shells or beams, because c and R are complicated functions of the element displacement
field, which is turn determined by u. Fortunately the CR approach is often used with elements of
simple geometry in which rotational freedoms, if any, may be ignored in defining the corotated
configuration. Under those conditions one can work out the base-to-deformed transformation arrays
directly from geometric arguments. The transformations may be then systematically applied to
existing linear elements through a modular interface, as illustrated in Figure 12.7.
In particular, for a simplex element (a constant strain element without rotational freedoms) it is
possible to work out all transformation from the intrinsic geometry of a line segment, triangle or
tetrahedron moving in 3D space. (By “intrinsic” is meant changes in edge lengths, face areas and
volume dimensions.) For the line segment modeling a space bar the formulas are worked out and
collected in the next sections. These apply to several types of finite elements, such as bars and cables.
The results may be specialized to two dimensions, if desired, by setting the third coordinate to zero.
It is important that the kinematic analysis be exact so that arbitrary rigid body motions can be
accomodated. Restrictions on local deformations in the motion from C R to C D can then be made
when considering specific elements, particularly those endowed with rotational DOFs.
Several of the following results are new. Their closed form derivation was made possible because of
the use of Mathematica to synthesize abtruse algebraic expressions containing symbolic terms.
uX2
Deformed uZ2
configuration CD 2 (x 2 ,y 2 ,z 2 )
1 (x1 ,y1 ,z 1 ) deformed length a
uX1 uZ1
Y, y uY2
uY1
10 (X1 ,Y1 ,Z 1 ) base length a 0
X, x
Z, z Global system
Base configuration C 0 2 0 (X2 ,Y2 ,Z2 )
Consider the line segment shown in Figure 14.1, defined by the end nodes 1-2. The segment moves
in three-dimensional space. The global axes will be denoted by {X, Y, Z } instead of {X 1 , X 2 , X 3 } so
14–3
Chapter 14: THE CR FORMULATION: SPACE BAR 14–4
a z = z21
a 2= a x2 + a y2 + a2z 2
a y = y21
1 a x = x21
Y, y
aX0 = X21
10
X, x aY0 = Y21
Z, z Global system a02 = 2
aX0 + 2
a Y0+ 2
a Z0
20
aZ0 = Z 21
that {X n , Yn , Z n } can be used for the coordinates of node n. Global axes {x, y, z}, which are used
for the deformed configuration, coalesce with {X, Y, Z },
The base line configuration C0 is specified by the coordinates (X 1 , Y1 , Z 1 ) and (X 2 , Y2 , Z 2 ) of the
line end nodes. The line moves to the deformed (current) configuration C D of length a defined by
coordinates {x1 = X 1 +u X 1 , y1 = Y1 +u Y 1 , z 1 = Z 1 +u Z 1 } and {x2 = X 2 +u X 2 , y2 = Y2 +u Y 2 , z 2 =
Z 2 + u Z 2 }, where u X 1 through u Z 2 are the node displacements. Node coordinate and displacement
differences are abbreviated by X 21 = X 2 − X 1 , x21 = x2 − x1 , u X 21 = u X 2 − u X 1 , etc. As illustrated
in Figure 14.2, the line lengths are given by
a02 = a 2X 0 + aY2 0 + a 2Z 0 = X 21
2
+ Y21
2
+ Z 21
2
, a 2 = ax2 + a y2 + az2 = x21
2
+ y21
2
+ z 21
2
. (14.1)
For further use define the following vectors
aX0 1 aX0 ax 1 ax u X 21
a0 = aY 0 , â0 = aY 0 , a = a y , â = a y , u21 = a − a0 = u Y 21 , (14.2)
aZ 0 a0 a a a a u Z 21
Z0 x z
Here â0 = a0 /a0 and â = a/a denote the direction cosine vectors of the base and current line
segment, respectively. This “hat convention” will be used to identify direction vectors normalized to
unit length. It is important not to confuse the 3-vector u21 with the 6-vector of node displacements
u = [ u X1 uY 1 u Z1 u X2 uY 2 u Z 2 ]T (14.3)
In fact, ∂u21 /∂u is the 3 × 6 matrix [ −I I ], where I is the 3 × 3 identity matrix.
§14.2.1. Line Segment Derivatives
Suppose that the node displacements are functions of two variables, generically denoted by and
: u X 1 = u X 1 (, ), etc. The partial derivative of a with respect to is
∂u
X 21
∂a ax ∂u X 21 a y ∂u Y 21 az ∂u Z 21 ∂
= + + = â u21, , in which u21, = ∂u Y 21 . (14.4)
T
∂ a ∂ a ∂ a ∂ ∂
∂u Z 21
∂
14–4
14–5 §14.2 LINE SEGMENT MOVING IN 3D
Note that the base configuration is only “remembered” through the displacements since the initial
length a0 does not appear explicitly.
To obtain the second derivative we take the partial of (14.4) with respect to the generic variable ,
which yields âT ∂ 2 u21 /(∂∂ ) + (∂ âT /∂ ) (∂u21 /∂). While the first term is easy, the second one
requires the derivatives of â. Since this involves the variation of a fixed-length (unit) vector, it can
be expected to involve the orthogonal projector associated with a. The final result can be presented
in the compact matrix form
∂ 2a
= u21,
T
H u21, + âT u21, , (14.5)
∂∂
in which
∂ 2 u X 21
∂∂ a y2 + az2 −ax a y −ax az
2 1 1 1
u21, = ∂ u Y 21 , H= −a y ax az2 + ax2 −a y az = (I − ââ ) = Pa . (14.6)
T
∂∂ a 3 a a
−az ax −az a y ax2 + a y2
∂ 2 u Z 21
∂∂
Here I denotes the identity matrix of order 3 whereas Pa = I − ââT is the orthogonal projector
associated with the direction a. (To show this, square Pa and verify that Pa2 = Pa .)
The results (14.4) and (14.5) can be specialized to various choices. For example, if the displacements
are viewed as functions of real time (or a time-like parameter) t, we have ≡ ≡ t, and
where a superposed dot denotes derivative with respect to t. Of more interest for element derivation
is to set and in turn to the six entries of the node displacement vector u arranged as (14.3). In
this case the term u21, in (14.5) vanishes, and we obtain
∂a 1 def ∂ 2a H −H def
= [ −ax −a y −az ax a y az ]T = [ −âT âT ]T = h, = = G.
∂u a ∂u ∂u −H H
(14.8)
Vector h and matrix G will appear in the derivation of the space bar element in §14..3. In addition
the following definition and relations are useful there:
def I −I 1
J = = aG + h hT , G = J − h hT . (14.9)
−I I a
Often the derivatives of a function F(a) are needed, for example for bar strain measures other than
engineering strains. The function F(a) is assumed twice differentiable respect to a. The a-derivative
abbreviations are
∂ F(a) ∂ 2 F(a)
F,a = , F,aa = , (14.10)
∂a ∂a 2
14–5
Chapter 14: THE CR FORMULATION: SPACE BAR 14–6
(Partials are used since F could be a function of other variables in addition to a.) Using the previous
results and the chain rule we obtain
∂ F(a) ∂ F(a) ∂a
= = F,a âT u21, ,
∂ ∂a ∂
∂ 2 F(a) ∂ F,a ∂a T ∂ 2a
= â u21, + F,a = u21,
T
F,aa ââT + F,a H u21, + F,a âT u21,
∂∂ ∂a ∂ ∂∂
(14.11)
For F(a) = a the previous results are recovered. The case of powers of a: F(a) = a n has immediate
applications. For n = 2, with and specialized first to t and then to u, we obtain the squared-length
derivatives:
˙ ¨
a 2 = 2a â u̇21 , a 2 = 2u̇T u̇21 + 2a ü21 (14.12)
∂a 2
∂ a
2 2
I −I
= 2a [ −âT âT ]T = 2a h, =2 = 2J. (14.13)
∂u ∂u ∂u −I I
The simplicity of the Hessian of a 2 will be a cogent argument for the use of the Green measure
of strain in bars. For n = −1, with and specialized as above, we obtain the inverse-length
derivatives:
˙ = −(1/a 2 ) â u̇,
1/a ¨ = (1/a 3 ) u̇T (3â âT − I) u̇ − (1/a 2 ) ü
1/a (14.14)
21 21 21
∂(1/a) ∂ 2 (1/a) 3â âT − I −3â âT + I
= −(1/a ) [ −â â ] ,
2
= (1/a )3
(14.15)
∂u ∂u ∂u −3â âT + I 3â âT − I
The foregoing equations may be specialized to two-dimensional motions of a segment moving in the
{X, Y } plane by setting the Z component to zero and then removing that component from vectors
and matrices.
All of the results given so far are geometrically exact and pose no limit on how much the segment
stretches or contracts. Restrictions in the form of small deformations will appear when the formulas
are applied to a bar element in §14.3.
14–6
14–7 §14.2 LINE SEGMENT MOVING IN 3D
LengthFunctionFirstDerivative[{fa_,a_},aXYZ_,u21_,Φ_]:=Module[
{nu=Length[u21],nΦ=Length[Φ],axyz,u21Φ}, axyz=aXYZ+u21;
u21Φ=Table[D[u21[[i]],Φ[[j]]],{i,1,nu},{j,1,nΦ}];
Return[(D[fa,a]*axyz/a).u21Φ] ];
LengthFunctionFirstDerivativeByFD[{fa_,a_},aXYZ_,u21_,δ _]:=
Module[{ax,ay,az,axyz,axyzp,axyzm,axp,ayp,azp,axm,aym,azm,
i,inc,dfada,da,d=Table[0,{6}]}, {ax,ay,az}=axyz=aXYZ+u21;
dfada=D[fa,a]/.a->Sqrt[ax^2+ay^2+az^2];
inc= δ*{{-1,0,0},{0,-1,0},{0,0,-1},{1,0,0},{0,1,0},{0,0,1}};
For [i=1,i<=6,i++, axyz=aXYZ+u21;
axyzp=axyz+inc[[i]]; axyzm=axyz-inc[[i]];
{axp,ayp,azp}=axyzp; {axm,aym,azm}=axyzm;
da=Sqrt[axp^2+ayp^2+azp^2]-Sqrt[axm^2+aym^2+azm^2];
d[[i]]=dfada*da/(2*δ );
]; Return[d]];
ClearAll[a,axyz,a,aX,aY,aZ,uX1,uY1,uZ1,uX2,uY2,uZ2,Φ,δ]; δ =1/100;
rep={aX->11,aY->10,aZ->2,uX1->3,uY1->-5,uZ1->-5,uX2->8,uY2->-4,
uZ2->1,a0->15,a->21};
aXYZ={aX,aY,aZ}; u21={uX2-uX1,uY2-uY1,uZ2-uZ1};
axyz=aXYZ+u21; {ax,ay,az}=axyz;
Print["aXYZ=",aXYZ/.rep," u21=",u21/.rep," axyz=",axyz/.rep,
" a0=",Sqrt[aX^2+aY^2+aZ^2]/.rep," a=",Sqrt[ax^2+ay^2+az^2]/.rep];
u={uX1,uY1,uZ1,uX2,uY2,uZ2};
For [if=1,if<=6,if++, fa={a,a^2,1/a,1/a^2,Sqrt[a],Log[a]}[[if]];
d=Simplify[LengthFunctionFirstDerivative[{fa,a},aXYZ,u21,u]];
Print["fa= ",fa,", d=",d,"\n = ",N[d/.rep]];
dFD=LengthFunctionFirstDerivativeByFD[{fa,a},aXYZ/.rep,u21/.rep, δ];
Print["d by FD= ", N[dFD/.rep]];
];
Figure 14.3. Mathematica implementation of first-derivative computations in (14.11) for F(a) and
numerical check by central finite differences.
fa= a , d=
aX − uX1 + uX2 aY − uY1 + uY2 aZ − uZ1 + uZ2 aX − uX1 + uX2 aY − uY1 + uY2 aZ − uZ1 + uZ2
− ,− ,− , , ,
2 a3/ 2 2 a3/2 2 a3/2 2 a3/2 2 a3/2 2 a3/2
= − 0.0831306, − 0.0571523, − 0.0415653, 0.0831306, 0.0571523, 0.0415653
√
Figure 14.4. Partial results from script of Figure 14.3 for F(a) = a and numerical values (14.16).
Displacement increment δ = 0.01 used for verification by central finite differences.
√
Figure 14.4 shows results for the case F(a) = a. Numerical results correspond to the data (14.16).
It can be seen that the finite difference values, obtained with a displacement increment of δ = 0.01,
agree to all places shown with the analytical results.
Scripts for the computation of second partial derivative of F(a) with respect to and are shown
in Figure 14.5. This is done by module LengthFunctionSecondDerivatives. The results when
both and are identified with the node displacement vector are verified through finite differences
with module LengthFunctionSecondDerivativesByFD. The driver code is√ shown at the bottom
of the figure. The computations are exercised for F(a) set to a, a 2 , 1/a, 1/a 2 , a and log(a) in the
loop shown at the bottom of Figure 14.5.
√
Figure 14.6 shows numerical results for F(a) = a and the data (14.16). Symbolic results are
omitted to reduce clutter. It can be seen that the finite difference values, obtained with a displacement
increment of δ = 0.01, agree to at least 5 places with the analytical results.
14–7
Chapter 14: THE CR FORMULATION: SPACE BAR 14–8
LengthFunctionSecondDerivative[{fa_,a_},aXYZ_,u21_,{Φ_,Ψ_}]:=Module[
{nc=Length[aXYZ],nu=Length[u21],nΦ=Length[Φ],nΨ=Length[Ψ],
u21Φ,u21Ψ,J,H,P,S,Q,dfda,d2fdada}, axyz=aXYZ+u21; S=Q=Table[0,{6},{6}];
u21Φ=Table[D[u21[[i]],Φ[[j]]],{i,1,nu},{j,1,nΦ}];
u21Ψ=Table[D[u21[[i]],Ψ[[j]]],{i,1,nu},{j,1,nΨ}];
J=Table[axyz[[i]]*axyz[[j]]/a^2,{i,1,nc},{j,1,nc}];
H=(IdentityMatrix[nc]-J)/a; Q=Table[0,{nΦ},{nΨ}];
For [i=1,i<=nc,i++, Q=Q+(axyz[[i]]/a)*
Table[D[u21[[i]],Φ[[j]],Ψ[[k]]],{j,1,nΦ},{k,1,nΨ}] ];
dfda=D[fa,a]; d2fdada=D[dfda,a];
S=Transpose[u21Φ].(d2fdada*J+dfda*H).u21Ψ;
Return[{S,Q}];
];
LengthFunctionSecondDerivativeByFD[{fa_,a_},aXYZ_,u21_,δ _]:=
Module[{aX,aY,aZ,uX21,uY21,uZ21,ax,ay,az,
i,inc,du,S=Table[0,{6},{6}]},
inc= δ*{{-1,0,0},{0,-1,0},{0,0,-1},{1,0,0},{0,1,0},{0,0,1}};
For [i=1,i<=6,i++, du=inc[[i]];
dp=LengthFunctionFirstDerivativeByFD[{fa,a},aXYZ,u21+du,δ ];
dm=LengthFunctionFirstDerivativeByFD[{fa,a},aXYZ,u21-du,δ ];
S[[i]]=(dp-dm)/(2* δ );
]; Return[S]];
√
Figure 14.6. Partial results from script of Figure 14.5 for F(a) = a and numerical values (14.16).
Symbolic results not shown to reduce clutter. Displacement increment δ = 0.01 used for verification
by central finite differences.
14–8
14–9 §14.3 THE CR BAR ELEMENT
− Length L
∼
y z
1 (x1 ,y1 ,z 1 )
∼
z Length L 0
Y, y
E, A constant
10 (X1 ,Y1 ,Z 1 )
20 (X 2 ,Y2 ,Z 2 )
Base configuration C 0
X, x x∼
Z, z Global system
The space bar element shown in Figure 14.7 will be presented here as an application of the foregoing
results on line segment moving through 3D space. The static EICR formulation is illustrated with a
linearly elastic, prismatic, 2-node bar element moving in 3D space, as depicted in that figure. The
element has six degrees of freedom collected in the 6-vector (14.3).
The area and length in the base configuration C0 are A0 and L 0 , respectively. These become A and
L in the deformed configuration C D . The element elongation is called d = L − L 0 . The elastic
modulus is E. In accordance with the usual assumptions of the CR description, bar deformations are
assumed to be small. This allows to carry all integrals over the initial volume A0 L 0 .
Application of the “best corotational fit criterion” shows that, as may be expected, the corotated
configuration C R is aligned with, and lies halfway from, the current end nodes. But this result will
not be required to develop the element equations.
The strain measure is taken to be a function e = e(d, L 0 ) = u(L , L 0 ) to be chosen later. Since
d = L − L 0 and L 0 is fixed, e,d = ∂e/∂d = ∂e/∂ L = e,L and likewise e,dd = ∂ 2 e/(∂d∂d) =
14–9
Chapter 14: THE CR FORMULATION: SPACE BAR 14–10
Table 14.1. Strain Measures for Bar Element and their Derivatives
∂ 2 e/(∂ L∂ L) = e,L L . The derivatives of e with respect to the nodal displacements DOF (14.3) are
obtained by the chain rule: ∂e/∂u = e,L (∂ L/∂u), etc. The necessary partials are already worked
out for the line segment in §14.2, with the substitutions a → L, a → L, â → L̂, etc. Recall that
∂L −L̂ ∂2 L H −H 1
=h= , =G= = (J − h hT ), (14.18)
∂u L̂ ∂u∂u −H H L
where L̂ is the â given in in (14.2), H is defined in (14.8) and J in (14.9), with the replacements
indicated above.
The internal force is the gradient of the internal energy with respect to the node displacements:
∂U ∂U ∂e ∂e ∂e ∂ L ∂e
p= = = L 0 (N0 + E A0 e) = L0 N = L0 N h. (14.19)
∂u ∂e ∂u ∂u ∂ L ∂u ∂L
The tangent stiffness is the Hessian of U :
∂ 2U ∂p ∂ N ∂e ∂ L ∂ 2e ∂ L ∂ L T ∂e ∂ 2 L
K= = = L0 + L0 N 2 + L0 N
∂u∂u ∂u ∂u ∂ L ∂u ∂ L ∂u ∂u ∂ L ∂u∂u
T
∂e ∂e ∂ L ∂ L ∂ 2e ∂ L ∂ L T
∂e ∂ 2 L
= L 0 E A0 + L0 N +
∂ L ∂ L ∂u ∂u ∂ L 2 ∂u ∂u ∂ L ∂u∂u
(14.20)
∂e 2 ∂ 2e ∂e
= E A0 L 0 h hT + N L 0 h h T
+ G
∂L ∂ L2 ∂L
∂e 2
2
1 ∂e ∂ e 1 ∂e
= E A0 L 0 h h + N L0
T
J+ − hh = K M + KG .
T
∂L L ∂L ∂ L2 L ∂L
Here K M and KG denote the material and geometric stiffness, a decomposition already encountered
in the TL description of Chapters 8–9.
§14.3.2. Matrices for Specific Strain Measures
Some specific strain measures and the values of their partials with respect to L are collected in Table
14.1. The appropriate choice should be replaced in (14.19) and (14.20) to get the final form of the
internal force vector and tangent stiffness matrix.
14–10
14–11 §14.3 THE CR BAR ELEMENT
E A0 T N
p = N h, KM = hh , KG = G, K = K M + KG . (14.21)
L0 L
L0 E A0 L 0 T N L0
p=N h, KM = hh , KG = (J − 2h hT ), K = K M + KG . (14.23)
L L2 L2
CRSpaceBar2InternalForce[ncoor0_,Em_,A0_,ue_,N0_,sm_,numer_]:=
Module[{X1,Y1,Z1,X2,Y2,Z2,LX,LY,LZ,LL0,L0,x1,y1,z1,x2,y2,z2,
Lx,Ly,Lz,LL,L,Lavg,d,EA,e,dedL,d2edL2,ND,h,p},
{{X1,Y1,Z1},{X2,Y2,Z2}}=ncoor0; {{uX1,uY1,uZ1},{uX2,uY2,uZ2}}=ue;
{LX,LY,LZ}={X2-X1,Y2-Y1,Z2-Z1}; LL0=LX^2+LY^2+LZ^2; L0=Sqrt[LL0];
{{x1,y1,z1},{x2,y2,z2}}=ncoor0+ue;
{Lx,Ly,Lz}={x2-x1,y2-y1,z2-z1}; LL=Lx^2+Ly^2+Lz^2; L=Sqrt[LL];
d=L-L0; Lavg=(L+L0)/2; e=Null;
If [sm=="eE", e=d/L0; dedL=1/L0; d2edL2=0];
If [sm=="eG", e=d*Lavg/L0^2; dedL=L/L0^2; d2edL2=1/L0^2];
If [sm=="eH", e=Log[L/L0]; dedL=1/L; d2edL2=-1/L^2];
If [sm=="eM", e=d/Lavg; dedL=L0/Lavg^2; d2edL2=-L0/Lavg^3];
If [e==Null,Print["CRSpaceBar2Force: Illegal sm arg"]; Return[Null]];
ND=N0+Em*A0*e; h={-Lx,-Ly,-Lz,Lx,Ly,Lz}/L;
p=ND*L0*dedL*h; If [numer,p=N[p]];
Return[p]];
ClearAll[ ]; ncoor0=N[{{0,0,0},{11,10,2}}];
Em=5000; A0=3; N0=10; numer=True;
For [iε=1, iε<=4, iε++, ={0,0.0001,0.01,0.1}[[iε]];
ue=N[{{2,3,-4},{-4,-5,8}}*(1+ε)^2];
For [ism=1,ism<=4,ism++, sm={"eE","eG","eH","eM"}[[ism]];
p=CRSpaceBar2InternalForce[ncoor0,Em,A0,ue,N0,sm,numer];
Print["sm: ",sm,", ε=", ε,", p=",p];
]];
Figure 14.8. Mathematica implementation of internal force vector calculation for 2-node space bar.
14–11
Chapter 14: THE CR FORMULATION: SPACE BAR 14–12
sm: eE, ε=0, p={− 3.33333, − 1.33333, − 9.33333, 3.33333, 1.33333, 9.33333}
sm: eG, ε= 0, p={− 3.33333, − 1.33333, − 9.33333, 3.33333, 1.33333, 9.33333}
sm: eH, ε= 0, p={− 3.33333, − 1.33333, − 9.33333, 3.33333, 1.33333, 9.33333}
sm: eM, ε= 0, p={− 3.33333, − 1.33333, − 9.33333, 3.33333, 1.33333, 9.33333}
sm: eE, ε=0.0001, p= {− 3.87431, − 1.54886, − 10.8525, 3.87431, 1.54886, 10.8525}
sm: eG, ε= 0.0001, p= {− 3.87476, − 1.54904, − 10.8538, 3.87476, 1.54904, 10.8538}
sm: eH, ε= 0.0001, p= {− 3.87386, − 1.54868, − 10.8513, 3.87386, 1.54868, 10.8513}
sm: eM, ε= 0.0001, p= {− 3.87386, − 1.54868, − 10.8513, 3.87386, 1.54868, 10.8513}
sm: eE, ε= 0.01, p={− 56.577, − 21.3257, − 165.128, 56.577, 21.3257, 165.128}
sm: eG, ε= 0.01, p={− 57.5008, − 21.6739, − 167.824, 57.5008, 21.6739, 167.824}
sm: eH, ε= 0.01, p={− 55.6686, − 20.9833, − 162.477, 55.6686, 20.9833, 162.477}
sm: eM, ε=0.01, p={− 55.6664, − 20.9824, − 162.47, 55.6664, 20.9824, 162.47}
sm: eE, ε=0.1, p={− 430.732, − 36.854, − 1902.59, 430.732, 36.854, 1902.59}
sm: eG, ε= 0.1, p={− 517.786, − 44.3025, − 2287.12, 517.786, 44.3025, 2287.12}
sm: eH, ε= 0.1, p={− 358.761, − 30.6961, − 1584.69, 358.761, 30.6961, 1584.69}
sm: eM, ε=0.1, p={− 356.998, − 30.5453, − 1576.9, 356.998, 30.5453, 1576.9}
14–12
14–13 §14.3 THE CR BAR ELEMENT
which is an adjustable parameter. If = 0 the bar moves in space but does not change length:
L 0 = L = 15. If > 0 the bar stretches by approximately L 0 .
The four strain measures e E , eG , e H and e M are exercised by cycling over sm = "eE", "eG", "eH",
and "eM", respectively, in the inner test loop. For values of : 0, 0.0001, 0.01 and 0.1 are tested in
the outer test loop.
As can be observed the internal force vectors are identical if = 0, since for zero strain it does not
matter which measure is used. The forces are very close for the different measures if = 0.0001,
which is very small strain (roughly 100 micros) but they differ substantially as gets large.
The implementation and verification of the tangent stiffness matrix is deferred to Exercises 14.3 and
14.4.
The formulation of the mass and damping matrices for dynamic analysis is not provided in this
Chapter.
14–13
Chapter 14: THE CR FORMULATION: SPACE BAR 14–14
EXERCISE 14.2 [A:15] Verify the expressions given in Examples 14.1–3, and append to these the p, K M and
KG matrices for the choice e = e M .
EXERCISE 14.3 [C:20] Implement the calculation of the tangent stiffness matrix for the space bar element
(any language is OK) to return K M and KG . If done in Mathematica, the code of Figure 14.8, which is posted
on the web site linked to the Chapter 14 Index, may be used as template. It is convenient to compute and return
K M and KG as two separate matrices as function value in Mathematica: Return[{ KM,KG }].
As numerical test, run the bar used in the test statements of that figure using = 0 and any strain measure. Print
out K M , KG , K = K M + KG and the eigenvalues of the three matrices. Validation check: the rank of K M , KG
and K should be 1, 3 and 3, respectively (except in the case of engineering strain, in which case the rank should
be 1, 2 and 3, respectively.)
EXERCISE 14.4 [C:20] Using the module developed in the previous Exercise, compute and show K M , KG and
K for the bar data used in the example of Figure 14.8. For the inner loop cycle over the 4 strain measure choices
used in the internal force test. For the outer test loop, cycle over equal to 0, 0.01 and −0.01. Compute and
show the eigenvalues of the three matrices. Validation check: the rank of K M , KG and K should be 1, 3 and 3,
respectively. For = −0.01 you may see some negative eigenvalues in KG and K; do not be alarmed.
14–14
13
.
CR Formulation
Overview II
13–1
Chapter 13: CR FORMULATION OVERVIEW II 13–2
TABLE OF CONTENTS
Page
§13.1. Internal Forces 13–3
§13.1.1. Force Transformations . . . . . . . . . . . . . . . 13–3
§13.1.2. Projector Properties . . . . . . . . . . . . . . . . 13–4
§13.2. Tangent Stiffness 13–5
§13.2.1. Definition . . . . . . . . . . . . . . . . . . . 13–5
§13.2.2. Material Stiffness . . . . . . . . . . . . . . . . 13–5
§13.2.3. Geometric Stiffness . . . . . . . . . . . . . . . . 13–6
§13.2.4. Consistency Verification . . . . . . . . . . . . . . 13–7
§13.3. Three Consistent CR Formulations 13–7
§13.3.1. Consistent CR formulation (C) . . . . . . . . . . . . 13–7
§13.3.2. Consistent Equilibrated CR Formulation (CE) . . . . . . . 13–8
§13.3.3. Consistent Symmetrizable Equilibrated CR Formulation (CSE) . . 13–8
§13.3.4. Formulation Requirements . . . . . . . . . . . . . 13–8
§13.3.5. Limitations of the EICR Formulation . . . . . . . . . . 13–9
§13.4. Conclusions 13–10
13–2
13–3 §13.1 INTERNAL FORCES
Consider an individual element e with N e nodes with six DOF (three translations and three rotations) at
each. Assume the element to be linearly elastic, undergoing only small deformations. Its internal energy
is assumed to be a function of the deformational displacements: U e = U e (v̄ed ), with array v̄ed organized
as shown in Table 13.2. U e is a frame independent scalar. The element internal force vector p̄e in the
CR frame is given by p̄e = ∂U e /∂ v̄ed . For each node a = 1, . . . N e :
∂U e
∂U e p̄eua ∂ ūda
p̄ae = e , or = (13.1)
∂ v̄da p̄eθa ∂U e
∂ θ̄da
where the second form separates the translational and rotational (moment) forces. To refer these to the
global frame we need to relate local-to-global kinematic variations:
∂ ūe ∂ ūedb
Ne db
δ ūeda δuae ∂uae ∂ωae
e = Jab , Jab = e e . (13.2)
δ θ̄da b=1
δωae ∂ θ̄db ∂ θ̄db
∂uae ∂ωae
e
From virtual work invariance, (p̄eu )T δ ūed + (p̄eθ )T δ θ̄d = (peu )T δue + (peθ )T δωe , whence
Ne
peua p̄eua
= T
Jab , a = 1, . . . N e . (13.3)
peθa b=1
p̄eθa
T T
It is convenient to split the Jacobian in (13.2) as Jab = H̄b P̄ab Ta and Jab
T
= TaT P̄ab H̄b . These matrices
are provided from three transformation stages, flowcharted in Figure 13.1:
δ ūedb I 0 δ ūeb
e
∂ θ̄db
e = , with H̄db = ,
δ θ̄db 0 H̄db δ ω̄eb ∂ ω̄edb
∂ ūe ∂ ūedb
db
∂ ūa ∂ ω̄ae
e
δ ūeb δ ūae
= P̄ab with P̄ab = , , (13.4)
δ ω̄eb δ ω̄ae
∂ ω̄edb ∂ ω̄edb
∂ ūe ∂ ω̄ae
e e e a
δ ūa δua TR 0 δua
= Ta = ,
δ ω̄ae δωae 0 TR δωae
The 3 × 3 matrix L is the Jacobian derivative already encountered in Chapter 12. An explicit expression
in terms of θ is given in Appendix R. To express compactly the transformations for the entire element it
is convenient to assemble the 6N e × 6N e matrices
13–3
Chapter 13: CR FORMULATION OVERVIEW II 13–4
P̄11 P̄12 . . . P̄1N e _ _
CR deformational HPT global total
P̄21 P̄22 . . . P̄2N e _
displac_& rotations
P̄ = , δud , δθd
displac & spins
δu, δω
... ... ... ... (13.5)
P̄ N e 1 P̄ N e 2 . . . P̄ N e N e _ rotation-to- global-to-CR T
T = diag [ T R T R . . . T R ] . H spin Jacobian frame rotator
The 6 × 6 matrix P̄ab in (13.4) extracts the deformational part of the displacement at node b in terms
of the total displacement at node a, both referred to the CR frame. At the element level, δ v̄ed = P̄ δ v̄e
extracts the deformational part by “projecting out” the rigid body modes. For this reason P̄ is called a
projector. As noted in Section 3.5, P̄ may be decomposed into a translational projector or T-projector P̄u
and a rotational projector or R-projector P̄ω , so that P̄ = P̄u + P̄ω . Each has a rank of 3. The T-projector
is a purely numeric matrix exemplified in Chapter 12. The R-projector can be expressed as P̄ω = S̄Ḡ,
where S̄ and Ḡ are defined in Chapter 12. Additional properties are studied below.
Remark 3. Rankin and coworkers [R.54,R.63–R.67] use an internal force transformation in which the incremental
−1
nodal rotations are used instead of the spins. This results in an extra matrix, H̄ appearing in the sequence (13.6).
The projector derived in those papers differs from the one constructed here in two ways: (1) only the R-projector is
considered, and (2) the origin of the CR frame is not placed at the element centroid but at an element node defined
by local node numbering. Omitting the T-projection is inconsequential if the element is “clean” with respect to
translational rigid body motions [R.30, Sec. 5].
P2ω = S (G S) G = S I G = S G = Pω . (13.7)
This assumes D = I; verification for non-identity D is immediate upon removal of zero rows and columns.
The orthogonality property Pu Pω = Pu S G = 0 follows by observing that Spin(xC ) = 0, where xC are
the coordinates of the element centroid in any frame with origin at C.
In the derivation of the consistent tangent stiffness, the variation of PT contracted with a force vector f,
where f is not varied, is required. The variation of the projector can be expressed as
For the tangent stiffness one needs δPT f. This vector can be decomposed into a balanced (self-
equilibrated) force fb = Pf and an unbalanced (out of equilibrium) force fu = (I − P)f. Then
δPT f = −(GT δST + δGT ST ) (fb + fu ) = −GT δST PT f − (GT δST + δGT ST ) fu
(13.9)
= −GT δST PT f + δPT fu .
13–4
13–5 §13.2 TANGENT STIFFNESS
where ST fb = 0 was used. This comes from the fact that the columns of S are the three rotational rigid
body motions, which do not produce work on an self-equilibrated force vector.
The term δPT fu will be small if element configurations C R and C D are close because in this case fu will
approach zero. If G has the factorizable form shown below, however, we can show that δPT fu = 0
identically, regardless of how close C R and C D are, as long as f is in translational equilibrium. Assume
that G can be factored as
G = Ξ Γ, with δG = δΞ Γ. (13.10)
where Ξ is a coordinate dependent invertible 3 × 3 matrix, and Γ is a constant 3 × 6N e matrix. Since
GS = I, Ξ−1 = Γ S, and δG S + G δS = δΞ Γ S + G δS = 0, whence δΞ = −G δS Ξ. Then
δPT fu = −(GT δST + ΓT δΞT ST ) fu = −(GT δST − ΓT ΞT δST , GT ST ) fu
= −(GT δST − GT δST PωT ) fu = −(GT δST (I − PωT ) fu = −GT δST (I − PωT ) fu (13.11)
= −GT δST (I − PωT )(I − PT ) f = 0,
if f is in translational equilibrium: f = PωT f. This is always satisfied for any element that represents rigid
body translations correctly [R.30, Sec. 5].
§13.2.1. Definition
The consistent tangent stiffness matrix Ke of element e is defined as the variation of the internal forces
with respect to element global freedoms:
def ∂pe
δpe = Ke δve , whence Ke = . (13.12)
∂ve
Taking the variation of pe in (13.6) gives rise to four terms:
T T T T T T T
δpe = δTT P̄ H̄ p̄e + TT δ P̄ H̄ p̄e + TT P̄ δHT p̄e + TT P̄ H̄ δ p̄e
(13.13)
= (KeG R + KeG P + KeG M + KeM ) δve .
The four terms identified in (13.13) receive the following names. K M is the material stiffness, KG M the
moment-correction geometric stiffness, KG P the equilibrium projection geometric stiffness, and KG R the
rotational geometric stiffness. If nodal eccentricities treated by rigid links are considered, one more term
appears, called the eccentricity geometric stiffness. This term is studied in great detail in [R.37].
The linear stiffness matrix in terms of the deformational freedoms in v̄ed is defined as the Hessian of the
internal energy:
e ∂ 2 Ū e ∂ p̄e
K̄ = e e = e (13.15)
∂ v̄d ∂ v̄d ∂ v̄d
13–5
Chapter 13: CR FORMULATION OVERVIEW II 13–6
m̄eN e 0 Spin(m̄eN e )
These are filled with the projection node forces p̄eP . Only the final form of the geometric stiffness
components is given below, omitting the detail derivations of [R.37]. The rotational geometric stiffness
T T
is generated by the variation of T: KeRG δve = δTT P̄ H̄ p̄e and can be expressed as
KeRG is the gradient of the internal force vector with respect to the rigid rotation of the element. This
interpretation is physically intuitive because a rigid rotation of a stressed element necessarily reorients
the stress vectors by that amount. Consequently the internal element forces must rigidly rotate to preserve
equilibrium.
The moment-correction geometric stiffness is generated by the variation of the jacobian H: KeRG δve =
T T
TTR P̄ δ H̄ p̄e . It is given by
T
KeG M = TT P̄ L̄ P̄T. (13.19)
where L is defined in Section 3.5.
The equilibrium projection geometric stiffness arises from the variation of the projector P̄ with respect
to the deformed element geometry: KeG P δve = TTR δPT HT p̄e . As in Section 4.2, decompose p̄e into a
T T
balanced (self-equilibrated) force p̄eb = P̄ p̄e and an unbalanced force p̄eu = p̄e − p̄eb . If δ P̄ p̄eu is either
identically zero or may be neglected as discussed in Section 4.2, KeG P is given by
T
KG P = −TT Ḡ F̄n P̄ T, (13.20)
13–6
13–7 §13.3 THREE CONSISTENT CR FORMULATIONS
KeRG expresses the variation of the projection of the internal force vector p̄e as the element geometry
changes. This can be interpreted mathematically as follows: In the vector space of element force vectors
the subspace of self-equilibrium force vectors changes as the element geometry changes. The projected
force vector thus has a gradient with respect to the changing self-equilibrium subspace, even though the
element force f̄e does not change.
The complete form of the element tangent stiffness, excluding correction terms (13.21) for highly warped
elements, is
T T e T T T e
Ke = TT P̄ H̄ K̄ H̄ P̄ + P̄ L̄ P̄ − F̄nm Ḡ − Ḡ F̄n P̄ T = TT K̄ P T, (13.22)
e
in which K̄ R , which is the local tangent stiffness matrix (the tangent stiffness matrix in the local CR
frame of the element) is given by the parenthesized expression.
From the foregoing unified forms of the internal force and tangent stiffness, three CR consistent for-
mulations can be obtained by making simplifying assumptions at the internal force level. These satisfy
self-equilibrium and symmetry to varying degree. The following subsections describe the three versions
in order of increasing complexity. For all formulations one can take in account DOFs at eccentric nodes
as described in [R.37].
13–7
Chapter 13: CR FORMULATION OVERVIEW II 13–8
e
approach the identity matrix. With small membrane strains K̄ is indifferent with respect to post-
e e
multiplication with P̄ because the C R and C configurations will be close and K̄ P → K̄ . The consistent
geometric stiffness is always unsymmetric, even at equilibrium. Because of this fact one cannot expect
quadratic convergence for this formulation if a symmetric solver is used.
This formulation may be unsatisfactory for warped quadrilateral shell elements since the C R and C D
reference configurations may be far apart. Only in the limit of a highly refined element mesh will the
C R and C D references in general be close, and satisfactory equilibrium ensured.
13–8
13–9 §13.3 THREE CONSISTENT CR FORMULATIONS
alter the equilibrium path since this is entirely prescribed by the foregoing equilibrium requirement.
However, lack of consistency may affect the location of bifurcation (buckling) points and the branch
switching mechanism for post-buckling analysis. In other words, an inconsistent tangent stiffness matrix
may detect (“see”) a bifurcation where equilibrium is not satisfied from the residual equation. Subsequent
traversal by branch-switching will then be difficult because the corrector iterations need to jump to the
secondary path as seen by the residual equation.
Invariance. This requirement refers to whether the solution is insensitive to internal choices that may
depend on node numbering. For example, does a local element-node reordering give an altered equi-
librium path or change the convergence characteristics for the analysis for an otherwise identical mesh?
The main contributor to lack of invariance is the way the deformational displacement vector is extracted
from the total displacements, if the extraction is affected by the choice of the local CR frame. If lack of
invariance is observed, it may be usually traced to the matrix G, which links the variation of the rigid
body rotation to that of the nodal DOF degrees of freedom.
Symmetrizability. This means that a symmetrized K can be used without loss of quadratic convergence
rate in a true Newton solver even when the consistent tangent stiffness away from equilibrium is not
symmetric. In the examples studied in Part II [R.38] this requirement was met when the material stiffness
of the formulation was rendered symmetric.
Element Independence. This is used in the sense of the EICR discussed in Section 2.5. It means that
the matrix and vector operations that account for geometrically nonlinear effects are the same for all
elements that possess the same node and DOF configuration.
Attributes of the C, CE and CSE formulations in light of the foregoing requirements are summarized in
Table 13.4.
13–9
Chapter 13: CR FORMULATION OVERVIEW II 13–10
respect to the deformed element geometry, effectively reduces the eigenvalues of the material stiffness
relative to the CR material stiffness K̄e before projection. This softening effect becomes significant if
the C R and C D geometries are far apart.
Such softening effects are noticeable in four-node initially-warped shell elements. Assume that the
element is initially warped with “positive” warping, and consider only the effect of P. The element
material stiffness of this initial positive warping is then K+ = P+ T
Ke P+ = Ke . Apply displacements
that switch this warping to the opposite of the initial one; that is, a “negative” warping. The new
element material stiffness then becomes K− = P− T
Ke P− = Ke One will intuitively want the two element
configurations to have the same rigidity in the sense of the dominant nonzero eigenvalues of the tangent
stiffness matrix. But it can be shown that the eigenvalues of the projected material stiffness matrix K−
can be significantly lower than those of the initial stiffness matrix K+ . If the element stiffness Ke is
referred to the flat element projection, one can restore symmetry of K+ and K− with respect to dominant
nonzero eigenvalues, but it is not possible to remove the softening effect.
This argument also carries over to higher order bar, arch and shell elements that are curved in the initial
reference configuration. It follows that the EICR is primarily useful for low-order elements of simple
geometry.
§13.4. Conclusions
This article presents a unified formulation for geometrically nonlinear analysis using the CR kinematic
description, assuming small deformations. Although linear elastic material behavior has been assumed
for brevity, extension to materially nonlinear behavior such as elastoplasticity and fracture within the
confines of small deformations, is feasible as further discussed below. All terms in the internal force and
tangent stiffness expressions are accounted for. It is shown how dropping selected terms in the former
produces simpler CR versions used by previous investigators.
These versions have been tested on thin shell and flexible-mechanism structures, as reported in Part II
[R.38]. Shells are modeled by triangle and quadrilateral elements. The linear stiffness of these elements
is obtained with the ANDES (Assumed Natural DEviatoric Strains) formulation of high performance
elements [R.27–R.32,R.53]. Test problems include benchmarks in buckling, nonlinear bifurcation and
collapse.
Does the unified formulation close the book on CR? Hardly. Several topics either deserve further
development or have been barely addressed:
(A) Relaxing the small-strain assumption to allow moderate deformations.
(B) Robust handling of extremely large rotations involving multiple revolutions.
(C) Integrating CR elements with rigid links and joint elements for flexible multibody dynamics.
(D) Using substructuring concepts for CR modeling of structural members with continuum elements.
(E) Achieving a unified form for CR dynamics, including nonconservative effects and multiphysics.
Topic (A) means the use of CR for problems where strains may locally reach moderate levels, say 1–10%,
as in elastoplasticity and fracture, using appropriate strain and stress measures in the local frame. The
challenge is that change of metric of the CR configuration should be accounted for, even if it means
dropping the EICR property. Can CR compete against the more established TL and UL descriptions?
It seems unreasonable to expect that CR can be of use in overall large strain problems such as metal
forming, in which UL reigns supreme. But it may be competitive in localized failure problems, where
most of the structure remain elastic although undergoing finite rotations.
13–10
13–11 §13.4 CONCLUSIONS
Some data points are available: previous large-deformation work presented in [R.46,R.47,R.50,R.51,R.57].
More recently Skallerud et al. reported [R.75] that a submerged-pipeline failure shell code using the AN-
DES CR quadrilateral of [R.37] plus elastoplasticity [R.74] and fracture mechanics [R.21] was able to
beat a well known commercial TL-based code by a factor of 600 in CPU time. This speedup is of obvious
interest in influencing both design cycle and deployment planning.
Topic (B) is important in applications where a floating (free-free) structure undergoes several revolutions,
as in combat airplane maneuvers, payload separation or orbital structure deployment. The technical
difficulty is that expressions presented in the Appendix cannot handle finite rotations beyond ±2π , and
thus require occasional resetting of the base configuration. While this can be handled via restarts for
structures such as full airplanes, it can be more difficult when the relative rotation between components
exceeds ±2π , as in separation, fragmentation or deployment problems.
Topics (C) and (D) have been addressed in the FEDEM program developed by SINTEF at Trondheim,
Norway. This program combines CR shell and beam elements of [R.37], grouped into substructures,
with kinematic objects typical of rigid-body dynamics: eccentric links and joints. Basic tools used in
FEDEM for combining joint models with flexible continuum elements are covered in a recent book
[R.73].
Finally, topic (E) is fertile ground for research. The handling of model components such as mass,
damping and nonconservative effects in fluid-structure interaction and aeroelasticity is an active ongoing
research topic. For example a recent paper [R.26] describes flight maneuver simulations of a complete
F-16 fighter using CR elements to model the aircraft. As in statics, a key motivation for CR in dynamics
is reuse of linear FEM force-stiffness libraries. Can that reuse extend to mass and damping libraries?
And how do standard time integration methods perform when confronted with unsymmetric matrices?
These topics have barely been addressed.
13–11
12
.
CR Formulation
Overview I
12–1
Chapter 12: CR FORMULATION OVERVIEW I 12–2
TABLE OF CONTENTS
Page
§12.1. Introduction 12–3
§12.2. The Emergence of CR 12–3
§12.2.1. Continuum Mechanics Sources . . . . . . . . . . . . 12–3
§12.2.2. FEM Sources . . . . . . . . . . . . . . . . . . 12–4
§12.2.3. Shadows of the Past . . . . . . . . . . . . . . . . 12–5
§12.2.4. Linking FEM and CR . . . . . . . . . . . . . . . 12–6
§12.2.5. Element Independent CR . . . . . . . . . . . . . . 12–7
§12.3. Corotational Kinematics 12–9
§12.3.1. Configurations . . . . . . . . . . . . . . . . . 12–9
§12.3.2. Coordinate Systems . . . . . . . . . . . . . . . . 12–9
§12.3.3. Coordinate Transformations . . . . . . . . . . . . . 12–11
§12.3.4. Rigid Displacements . . . . . . . . . . . . . . . . 12–11
§12.3.5. Rotator Formulas . . . . . . . . . . . . . . . . 12–12
§12.3.6. Degrees of Freedom . . . . . . . . . . . . . . . . 12–12
§12.3.7. EICR Matrices . . . . . . . . . . . . . . . . . 12–13
§12.3.8. Deformational Translations . . . . . . . . . . . . . . 12–16
§12.3.9. Deformational Rotations . . . . . . . . . . . . . . 12–17
12–2
12–3 §12.2 THE EMERGENCE OF CR
Note: This and the following Chapters on the CR formulation of geometrically nonlinear FEM are new.
References and Appendices for this material have been placed in Appendix R.
§12.1. Introduction
Three Lagrangian kinematic descriptions are in present use for finite element analysis of geometrically
nonlinear structures: (1) Total Lagrangian (TL), (2) Updated Lagrangian (UL), and (3) Corotational
(CR). The CR description is the most recent of the three and the least developed one. Unlike the others,
its domain of application is limited by a priori kinematic assumptions:
Displacements and rotations may be arbitrarily large, but deformations must be small. (12.1)
Because of this restriction, CR has not penetrated the major general-purpose FEM codes that cater to
nonlinear analysis. A historical sketch of its development is provided in Section 2.
As typical of Lagrangian kinematics, all descriptions: TL, UL and CR, follow the body (or element) as
it moves. The deformed configuration is any one taken during the analysis process and need not be in
equilibrium during a solution process. It is also known as the current, strained or spatial configuration
in the literature, and is denoted here by C D . The new ingredient in the CR description is the “splitting”
or decomposition of the motion tracking into two components, as illustrated in Figure 12.1.
1. The base configuration C0 serves as the origin
of displacements. If this happens to be one
actually taken by the body at the start of the Corotated C R
analysis, it is also called initial or undeformed. Deformed (current,
spatial) C
D
The name material configuration is used pri-
marily in the continuum mechanics literature.
Motion splits into Deformational
2. The corotated configuration C R varies from deformational and rigid motion
element to element (and also from node to
node in some CR variants). For each individ- Base (initial, undeformed,
ual element, its CR configuration is obtained material) configuration C 0
through a rigid body motion of the element
Rigid body motion
base configuration. The associated coordinate
system is Cartesian and follows the element Global
frame
like a “shadow” or “ghost,” prompting names
such as shadow and phantom in the Scandi-
Figure 12.1. The CR kinematic description.
navian literature. Element deformations are
Deformation from corotated to deformed (current)
measured with respect to the corotated con- configuration grossly exaggerated for visibility.
figuration.
In static problems the base configuration usually remains fixed throughout the analysis. In dynamic
analysis the base and corotated configurations are sometimes called the inertial and dynamic reference
configurations, respectively. In this case the base configuration may move at uniform velocity (a Galilean
inertial system) following the mean trajectory of an airplane or satellite.
From a mathematical standpoint the explicit presence of a corotated configuration as intermediary be-
tween base and current is unnecessary. The motion split may be exhibited in principle as a multiplicative
decomposition of the displacement field. The device is nonetheless useful to teach not only the physical
meaning but to visualize the strengths and limitations of the CR description.
12–3
Chapter 12: CR FORMULATION OVERVIEW I 12–4
12–4
12–5 §12.2 THE EMERGENCE OF CR
Corotated (a.k.a.
dynamic reference) frame
PR CR
P C
Corotated (shadow)
configuration C R
Deformed configuration C D
(Shown separate from C R for
visualization convenience)
P0
Base (a.k.a. inertial) frame
Figure 12.2. The concept of separation of base (a.k.a. inertial) and CR (a.k.a. dynamic)
configurations in aircraft dynamics. Deformed configuration (with deformations grossly exaggerated)
shown separate from CR configuration for visibility. In reality points C and C R coincide.
strain element formulations. Spins (instead of rotations) are used as incremental nodal freedoms. This
simplifies the EICR “front end” and facilitates attaining consistency.
Battini and Pacoste at KTH-Stockholm [R.2,R.3,R.58] have recently used the CR approach, focusing
on stability applications. The work by Teigen [R.79] should be cited for the careful use of offset nodes
linked to element nodes by eccentricity vectors in the CR modeling of prestressed reinforced-concrete
members.
12–5
Chapter 12: CR FORMULATION OVERVIEW I 12–6
;;;
(a)
Base
(b)
;;;; Base
velocities or accelerations of material points due to the deformations are minimum in some global sense.
If the body does not deform, any set of axes fixed into the body is of course a natural dynamic reference
frame.”
Clearly the focus of this article was on a whole structure, as illustrated in Figure 12.2 for an airplane.
This will be called the shadowing problem. A body moves to another position in space: find its mean
rigid body motion and use this information to locate and orient a corotated Cartesian frame.
Posing the shadowing problem in three dimensions requires fairly advanced mathematics. Using two
“best fit” criteria Fraeijs de Veubeke showed that the origin of the dynamic frame must remain at the
center of mass of the displaced structure: C R in Figure 12.2. However, the orientation of this frame
leads to an eigenvalue problem that may exhibit multiple solutions due to symmetries, leading to non
uniqueness. (This is obvious by thinking of the polar and singular-value decompositions, which were not
used in that article.) That this is not a rare occurrence is demonstrated by considering rockets, satellites
or antennas, which often have axisymmetric shape.
Remark 12.1. Only C D (shown in darker shade in Figure 12.2) is an actual configuration taken by the pictured
aircraft structure. Both reference configurations C and C are virtual in the sense that they are not generally
0 R
occupied by the body at any instance. This is in contrast to the FEM version of this idea.
The practical extension of Fraeijs de Veubeke’s idea to geometrically nonlinear structural analysis by
FEM relies on two modifications:
1. Multiple Frames. Instead of one CR frame for the whole structure, there is one per element. This
is renamed the CR element frame.
2. Geometric-Based RBM Separation. The rigid body motion is separated directly from the total
element motion using elementary geometric methods. For example in a 2-node bar or beam one
axis is defined by the displaced nodes, while for a 3-node triangle two axes are defined by the plane
passing through the points. See Figure 12.3.
The first modification is essential to success. It helps to fulfill assumption (12.1): the element defor-
mational displacements and rotations remain small with respect to the CR frame. If this assumption
is violated for a coarse discretization, break it into more elements. Small deformations are the key to
element reuse in the EICR discussed below. If intrinsically large strains occur, however, the breakdown
prescription fails. In that case CR offers no advantages over TL or UL.
12–6
12–7 §12.2 THE EMERGENCE OF CR
Deformational
Global element element
equations Incorporate equations Form element
Assembler rigid body mass, stiffness
motions & forces
System equations
of motion
Extract Evaluate
Solver deformational element
Total motions Deformational stresses
displacements displacements
Finite Element
CR "Filters" Library
Figure 12.4. The EICR as a modular interface to a linear FEM library. The
flowchart
is mainly conceptual. For computational efficiency the interface logic
may be embedded with each element through inlining techniques.
The second modification is inessential. Its purpose is to speed up the implementation of geometrically
simple elements. The CR frame determination may be refined later, using more advanced tools such as
polar decomposition and best-fit criteria, if warranted.
Remark 12.2. CR is ocassionally confused with the convected-coordinate description of motion, which is used in
branches of fluid mechanics and rheology. Both may be subsumed within the class of moving coordinate kinematic
descriptions. The CR description, however, maintains orthogonality of the moving frame(s) thus achieving an
exact decomposition of rigid-body and deformational motions. This property enhances computational efficiency as
transformation inverses become transposes. On the other hand, convected coordinates form a curvilinear system
that “fits” the change of metric as the body deforms. The difference tends to disappear as the discretization becomes
progressively finer, but the fact remains that the convected metric must encompass deformations. Such deformations
are more important in solid than in fluid mechanics (because classical fluid models “forget” displacements). The
idea finds more use in UL descriptions, in which the individual element metric is updated as the motion progresses.
As previously noted, one of the sources of the present work is the element-independent corotational
(EICR) description developed by Rankin and coworkers [R.54,R.63–R.67]. Here is a summary descrip-
tion taken from the Introduction to [R.54]:
“In the co-rotation approach, the deformational part of the displacement is extracted by purging the rigid
body components before any element computation is performed. This pre-processing of the displacements
may be performed outside the standard element routines and thus is independent of element type (except
for slight distinctions between beams, triangular and quadrilateral elements).”
Why is the EICR worth study? The question fits in a wider topic: why CR? That is, what can CR do
that TL or UL cannot? The topic is elaborated in the Conclusions section, but we advance a practical
reason: reuse of small–strain elements, including possibly materially nonlinear elements.
The qualifier element independent does not imply that the CR equations are independent of the FEM
discretization. Rather it emphasizes that the key operations of adding and removing rigid body motions
can be visualized as a front end filter that lies between the assembler/solver and the element library, as
sketched in Figure 12.4. The filter is purely geometric. For example, suppose that a program has four
different triangular shell elements with the same node and degree-of-freedom configuration. Then the
12–7
Chapter 12: CR FORMULATION OVERVIEW I 12–8
front end operations are identical for all four. Adding a fifth small-strain element of this type incurs
relatively little extra work to “make it geometrically nonlinear.”
This modular organization is of interest because it implies that the element library of an existing FEM
program being converted to the CR description need not be drastically modified, as long as the analysis
is confined to small deformations. Since that library is typically the most voluminous and expensive part
of a production FEM code, element reuse is a key advantage because it protects a significant investment.
For a large-scale commercial code, the investment may be thousands of man-years.
Of course modularity and computational efficiency can be conflicting attributes. Thus in practice the
front end logic may be embedded with each element through techniques such as code inlining. If so the
flowchart of Figure 12.4 should be interpreted as conceptual.
12–8
12–9 §12.3 COROTATIONAL KINEMATICS
This section outlines CR kinematics of finite elements, collecting the most important relations. Mathe-
matical derivations pertaining to finite rotations are consigned to Appendix A. The presentation assumes
static analysis, with deviations for dynamics briefly noted where appropriate.
§12.3.1. Configurations
To describe Lagrangian kinematics it is convenient to introduce a rich nomenclature for configurations.
For the reader’s convenience those used in geometrically nonlinear static analysis using the TL, UL
or CR descriptions are collected in Table 12.1. Three: base, corotated and deformed, have already
been introduced. Two more: iterated and target, are connected to the incremental-iterative solution
process covered in Part II [R.38]. The generic configuration is used as placeholder for any kinematically
admissible one. The perturbed configuration is used in variational derivations of FEM equations.
Two remain: reference and globally-aligned. The reference configuration is that to which element
computations are referred. This depends on the description chosen. For Total Lagrangian (TL) the
reference is base configuration. For Updated Lagrangian (UL) it is the converged or accepted solution
of the previous increment. For corotational (CR) the reference splits into CR and base configurations.
The globally-aligned configuration is a special corotated configuration: a rigid motion of the base that
makes the body or element align with the global axes introduced below. This is used as a “connector”
device to teach the CR description, and does not imply the body ever occupies that configuration.
The separation of rigid and deformational components of motion is done at the element level. As noted
previously, techniques for doing this have varied according to the taste and background of the investigators
that developed those formulations. The approach covered here uses shadowing and projectors.
12–9
Chapter 12: CR FORMULATION OVERVIEW I 12–10
uP
CR C
0
xPC P 0 uPR _
xP x2
C0 c
0 X 2,x2
xPR
xP b
a Corotated
x~1 O //x~1 ~
//x 2
X ,x _
element x~2 global frame 1 1 x 1 element
base frame (with material & rigid body
spatial coalesced) rotation CR frame
Note that the element index e has been suppressed to reduced clutter. That convention will be followed
throughout unless identification with elements is important. In that case e is placed as supercript.
The base frame {x̃i } is chosen according to usual FEM practices. For example, in a 2-node spatial beam
element, x̃1 is defined by the two end nodes whereas x̃2 and x̃3 lie along principal inertia directions. An
important convention, however, is that the origin is always placed at the element centroid C 0 . For each
deformed (current) element configuration, a fitting of the base element defines its CR configuration, also
known as the element “shadow.” Centroids C R and C ≡ C D coincide. The CR frame {x̄i } originates
at C R . Its orientation results from matching a rigid motion of the base frame, as discussed later. When
the current element configuration reduces to the base at the start of the analysis, the base and CR frames
coalesce: {x̃i ≡ x̄i }. At that moment there are only two different frames: global and local, which agrees
with linear FEM analysis.
Notational conventions: use of G, 0, R and D as superscripts or subscripts indicate pertinence to the
globally-aligned, base, corotated and deformed configurations, respectively. Symbols with a overtilde
or overbar are measured to the base frame {x̃i } or the CR frame {x̄i }, respectively. Vectors without a
superposed symbol are referred to global coordinates {xi ≡ xi }. Examples: x R denote global coordinates
of a point in C R whereas x̃G denote base coordinates of a point in CG . Symbols a, b and c = b − a are
abbreviations for the centroidal translations depicted in Figure 12.5, and more clearly in Figure 12.7(b).
A generic, coordinate-free vector is denoted by a superposed arrow, for example u , but such entities
rarely appear in this work.
The rotators T0 and T R are the well known local-to-global displacement transformations of FEM analysis.
12–10
12–11 §12.3 COROTATIONAL KINEMATICS
Figure 12.7. Further distillation to essentials of Figure 12.5. A bar moving in 2D is shown:
(a) Rigid motion from globally-aligned to base and corotated configurations; (b) key geometric
quantities that define rigid motions in 2D; (c) as in (a) but followed by a stretch from corotated
to deformed. The globally-aligned configuration is fictitious: only a convenient link up device.
12–11
Chapter 12: CR FORMULATION OVERVIEW I 12–12
Here I is the 3 × 3 identity matrix, whereas R̃0 = T0 R0 T0T and R̄0 = T R R0 TTR denote the C0 →C R
rotator referred to the base and corotated frames, respectively.
Traversing the links pictured in Figure 12.8 shows that any rotator
can be expressed in terms of the other two:
T
R̄0 = T0 TTR , R̄0 = T R T0T . (12.7) TT0 TR
T0 global TRT
Notice that T0 is fixed since C and C are fixed throughout the anal-
G 0 frame
ysis, whereas T R and R0 change. Their variations of these rotators X, x
are subjected to the following constraints:
Figure 12.8. Rotator frame links.
The last two come com the orthogonality conditions TTR T R = I and R0T R0 = I, respectively, and provide
δR0 = −R0 δR0T R0 , δR0T = −R0T δR0 R0T , etc.
We denote by ω and ω̄ the axial vectors of R0 and R̄0 , respectively, using the exponential map form of
the rotator described in Section A.10. The variations δω and δ ω̄ are used to form the skew-symmetric
T
spin matrices Spin(δω) = δR0 R0T = −Spin(δω)T and Spin(δ ω̄) = δ R̄0 R̄0 = −Spin(δ ω̄)T . These
matrices are connected by congruential transformations:
Using these relations the following catalog of rotator variation formulas can be assembled:
12–12
12–13 §12.3 COROTATIONAL KINEMATICS
δ v̄e = [ δ v̄e1 . . . δ v̄eN e ]T Local CR Element Localization of above to element e in CR frame. Trans-
e
δ ūa lations: δ ūae , spins: δ ω̄ae ; conjugate forces: n̄ae and m̄ae ,
with δ v̄ae =
δ ω̄ae respectively, for a = 1, . . . N e .
v̄ed = [ v̄ed1 . . . v̄ed N e ]T Local CR Element Deformational displacements and rotations at element
ūe e
with v̄eda = da
e
nodes. Translations: ūeda , rotations: θ̄da ; conjugate
θ̄da forces: n̄a and m̄a , respectively, for a = 1, . . . N e .
N = number of nodes in structure; N e = number of nodes in element e; a, b : node indices.
12–13
Chapter 12: CR FORMULATION OVERVIEW I 12–14
1. From the initial global nodal coordinates xae compute centroid position ae =
N e
e
xC0 = (1/N e ) a=1 xae . Form rotator Te0 as per element type convention.
Compute node coordinates in the element base frame: x̃ae = Te0 (xae − ae ).
3. Compute the deformational translations ūda = x̄ae − x̄eRa . R̃d = Tn Ra T0T Com-
e
pute the deformational rotator R̄da = Te Rae (Te0 )T . Extract the deformational
e
angles θ̄da
e
from the axial vector of R̄da .
For any N e ≥ 1 it is easy to verify that P2u = Pu , with 5N e unit eigenvalues and N e zero eigenvalues. Thus
Pu is an orthogonal projector. Physically, it extracts the deformational part from the total translational
displacements.
Matrix S is called the spin-lever or moment-arm or matrix. It is dimensioned 3N e × 3 and has the
configuration (written in transposed form to save space):
in which I is the 3 × 3 identity matrix and Sa are node spin-lever 3 × 3 submatrices. Let xa =
[ x1a x2a x3a ]T generically denote the 3-vector of coordinates of node a referred to the element centroid.
Then Sa = Spin(xa ). The coordinates, however, may be those of three different configurations: C0 , C R
and C D , referred to two frame types: global or local. Accordingly superscripts and overbars (or tildes)
are used to identify one of six combinations. For example
R D
0 −x3a 0
−a3 x2a
0
−a2 0 −x̄3a R
x̄2a 0 −x̄3aD
x̄2a
Sa0 = x3a −a1 , S̄a = x̄3a R
, S̄a = x̄3a D
R D
0
−a3 0 −x1a0 R
0 −x̄1a D
0 −x̄1a ,
−x2a −a2 x1a −a1
0 0
0 −x̄2a x̄1a
R R
0 −x̄2a x̄1a
D D
0
(12.13)
12–14
12–15 §12.3 COROTATIONAL KINEMATICS
def
def
N e
δω = G δve = G δvae ,
a a
δ ω̄ = Ḡ δ v̄e = Ḡ δ v̄ae ,
a a
with a
≡ a=1
. (12.15)
Here the spin axial vector variation δωe denotes the instantaneous rotation at the centroid, measured
in the global frame, when the deformed configuration is varied by the 6N e components of δve . When
referred to the local CR frame, these become δ ω̄e and δ v̄e , respectively. For construction, both G and
Ḡ may be split into node-by-node contributions using the 3 × 6 submatrices Ga and Ḡa shown above.
As an example, G matrices for the space bar element shown in Figure 12.9(b) is 3 × 12. The spin-lever
matrix in C D referred to the local CR frame is
D 1 0 0 0 0 0 0 0 0 0 0 0 0
Ḡ = 0 0 1 0 0 0 0 0 −1 0 0 0 . (12.16)
L 0 −1 0 0 0 0 0 1 0 0 0 0
The first row is conventionally set to zero as the spin about the bar axis x̄1 is not defined by the nodal
freedoms. This “torsion spin” is defined, however, in 3D beam models by the end torsional rotations.
12–15
Chapter 12: CR FORMULATION OVERVIEW I 12–16
Unlike S, the entries of G depend not only on the element geometry, but on a developer’s decision: how
the CR configuration C R is fitted to C D . For the triangular shell element this matrix is given in Appendix
B. For quadrilateral shells and space beam elements it is given in Part II [R.38].
D D
Matrices S̄ and Ḡ satisfy the biorthogonality property
G S = D. (12.17)
where D is a 3 × 3 diagonal matrix of zeros and ones. A diagonal entry of D is zero if a spin component
D D
is undefined by the element freedoms. For instance in the case of the space bar, the product Ḡ S̄ of
(12.14) and (12.16) is diag(0, 1, 1). Aside from these special elements (e.g., bar, spars, shaft elements),
D = I. This property results from the fact that the three columns of S are simply the displacement
vectors associated with the rigid body rotations δ ω̄i = 1. When premultiplied by G one merely recovers
the amplitudes of those three modes.
The rotational projector or simply R-projector is generically defined as Pω = S G. Unlike the T-projector
Pu such as those in (12.11), the R-projector depends on configuration and frame of reference. Those are
R R R
identified in the usual manner; e.g., P̄ω = S̄ω Ḡω . This 6N e × 6N e matrix is an orthogonal projector of
rank equal to that of D = G S. If G S = I, Pr has rank 3. The complete projector matrix of the element
is defined as
P = Pu − Pω . (12.18)
This is shown to be a projector, that is P2 = P, in Section 4.2.
Two additional 6N e × 6N e matrices, denoted by H and L, appear in the EICR. H is a block diagonal
matrix built of 2N e 3 × 3 blocks:
Here Ha denotes the Jacobian derivative of the rotational axial vector with respect to the spin axial vector
evaluated at node a. An explicit expression of H(θ) is given in (R.48) of Appendix A. The local version
in the CR frame is
H̄ = diag [ I H̄d1 I H̄d2 . . . I H̄d N e ] , H̄da = H̄(θ̄da ), H̄(θ̄d ) = ∂ θ̄d /∂ ω̄d . (12.20)
where ma is the 3-vector of moments (conjugate to δωa ) at node a. The expression of L(θ, m) is
provided in (R.49) of Appendix A. The local form L̄ has the same block organization with La replaced
by L̄a = L(θ̄da , m̄a ).
Consider a generic point P 0 of the base element of Figure 12.5, with global position vector x0P . P 0 rigidly
moves to P R in C R with position vector x RP = x0P + u RP = x0P + c + x RPC . Next the element deforms to
occupy C D . P R displaces to P, with global position vector x P = x0P + u P = x0P + c + x RPC + ud P .
The global vector from C 0 to P 0 is x0P − a, which in the base frame becomes x̃0P = T0 (x0P − a). The
global vector from C R ≡ C to PR is x RP −b, which in the element CR frame becomes x̄ RP = T R (x RPC −b).
12–16
12–17 §12.3 COROTATIONAL KINEMATICS
But x̃0P = x̄ RP since the C0 →C R motion is rigid. The global vector from PR to P is ud P = x P − x RP ,
which represents a deformational displacement. In the CR frame this becomes ūd P = T R (x P − x RP ).
The total displacement vector is the sum of rigid and deformational parts: u P = ur P + ud P . The rigid
displacement is given by expressions collected in (12.5), of which ur P = (R0 − I) (x0P − a) + c is the
most useful. The deformational part is extracted as ud P = u P − ur P = u P − c + (I − R0 ) (x0P − a).
Dropping P to reduce clutter this becomes
The element centroid position is calculated by averaging its node coordinates. Consequently
N e
c = (1/N e ) u ,
b b
ua − c = U u ,
b ab b
with b
≡ b=1
(12.23)
in which Uab = (δab − 1/N e ) I is a building block of the T-projector introduced in the foregoing
subsection. Evaluate (12.22) at node a, insert (12.23), take variations using (12.10) to handle δR0 , use
(12.2) to map R0 (x0 − a) = x R − b, and employ the cross-product skew-symmetric property (R.3) to
extract δω:
δuda = δ(ua −c) − δR0 (xa0 −a) = U δub − Spin(δω) R0 (xa0 −a)
b ab
= Uab δub − Spin(δω) (x R
a − b) = U δub + Spin(xaR − b) δω
b ab
(12.24)
b
= U δub +
b ab
S R Gb δvb .
b a
Here matrices S and G have been introduced in (12.12)–(12.15). The deformational displacement in the
T
element CR frame is ūd = T R ud . From the last of (12.5) we get ūd = ū − c̄ − (I − R̄0 )x̄ R , where
R̄0 = T R R0 TTR . Proceeding as above one gets
D
δ ūda = U δ ūb +
b ab b
S̄a Ḡb δ v̄b . (12.25)
D
The node lever matrix SaR of (12.24) changes in (12.25) to S̄a , which uses the node coordinates of the
deformed element configuration.
The deformational rotation (12.26) is taken to be small but finite. Thus a procedure
to extract a rotation
axial vector θd from a given rotator is needed. Formally this is θ̄d = axial Loge (R̄d )], but this can be
prone to numerical instabilities. A robust procedure is presented in Section A.11. The axial vector is
evaluated at the nodes and identified with the rotational DOF.
12–17
Chapter 12: CR FORMULATION OVERVIEW I 12–18
Evaluating (12.26) at a node a, taking variations and going through an analysis similar to that carried
out in the foregoing section yields
∂θda
∂ωda
δθda = δω b = H a δab [ 0 I ] − Gb δvb .
∂ωda b ∂ω
b
b
(12.27)
∂ θ̄da
∂ ω̄da
δ θ̄da = δ ω̄b = H̄a δab [ 0 I ] − Ḡb δ v̄b .
∂ ω̄da b ∂ ω̄
b
b
12–18
11
.
11–1
Chapter 11: THE CORE CONGRUENTIAL FORMULATION: STIFFNESS EQUATIONS11–2
TABLE OF CONTENTS
Page
§11.1. DCCF Transformation to Physical Freedoms 11–3
§11.2. DCCF Transformation Examples 11–4
§11.2.1. The Bar Element . . . . . . . . . . . . . . . . 11–4
§11.2.2. Iso-P Plane Stress Element . . . . . . . . . . . . 11–4
§11.3. The Generalized CCF 11–5
§11.3.1. Generalized Coordinates as Generic Target . . . . . . . 11–6
§11.3.2. Algebraic Transformation . . . . . . . . . . . . . 11–6
§11.3.3. Differential Transformation . . . . . . . . . . . . . 11–7
§11.3.4. Multistage Transformation . . . . . . . . . . . . 11–7
§11.4. A 2-Node 2D Timoshenko Beam Element 11–8
§11.4.1. Generalized Coordinates and Stress Resultants . . . . . . 11–8
§11.4.2. Transformation Matrices . . . . . . . . . . . . . 11–9
§11.4.3. Internal Force Vector . . . . . . . . . . . . . . . 11–10
§11.4.4. Tangent Stiffness Matrix . . . . . . . . . . . . . 11–10
§11.4.5. Can a Secant Stiffness be Constructed? . . . . . . . . . 11–13
§11.5. A 2-Node 3D Timoshenko Beam Element 11–13
§11.5.1. Transformation to Generalized Gradients . . . . . . . . 11–13
§11.5.2. Transformation to the Rotational Vector . . . . . . . . 11–18
§11.5.3. Transformation to Finite Element Freedoms . . . . . . . 11–19
§11.6. Equivalence of DCCF and Standard TL Formulation 11–20
§11.7. References 11–22
§11. Exercises . . . . . . . . . . . . . . . . . . . . . . 11–24
11–2
11–3 §11.1 DCCF TRANSFORMATION TO PHYSICAL FREEDOMS
The present Chapter completes the development of the Core Congruential Formulation (CCF) of
Total Lagrangian (TL) elements. It present procedures to transform core equations to finite element
stiffness equations.
u = N v, (11.1)
where v now collects the element node-displacement degrees of freedom (DOFs) and N =
N(X 1 , X 2 , X 3 ) is a matrix of shape functions independent of v. Differentiating (11.1) with re-
spect to the X i and taking the first two v variations yields
(for the last one see Remark 10.2). Invariance of the strain energy variations δU and δ 2 U obtained
by integrating (10.14)-(10.15) over the element reference volume yields
K =
U T U
G S G d V, K =
r T r
G S G d V, K= GT SG d V, (11.3)
V0 V0 V0
f= GT Φ d V, p= GT Ψ d V, p0 = GT Ψ0 d V, (11.4)
V0 V0 V0
Although the dependency of Slevel and Ψ on g is not made implicit in these equations, it must be
remembered that the transformation g = Gv also appears there. Because of the ensuing algebraic
complexity, numerical integration is generally required unless the gradients are constant over the
element.
Often G is expressed as a chain of transformations, some of which are position dependent and
remain inside the element integral whereas others are not and may be taken outside. For example,
in the bar element treated below, G = T Ḡ, where Ḡ transforms g to local node displacements
while T transforms local to global node displacements.
11–3
Chapter 11: THE CORE CONGRUENTIAL FORMULATION: STIFFNESS EQUATIONS11–4
The core equations for a geometrically nonlinear TL bar were derived in Section 5.1. These equa-
tions are now applied to the formulation of a two-node, linear-displacement, prismatic TL bar
element. The element has constant reference area A0 and initial length L 0 . The two end nodes are
located at (X 1 , Y1 , Z 1 ) and (X 2 , Y2 , Z 2 ), respectively. The node displacements are (v X 1 , vY 1 , v Z 1 )
and (v X 2 , vY 2 , v Z 2 ). The element displacement field in local coordinates { X̄ , Ȳ , Z̄ } may be inter-
polated as
v̄
X1
ū X N1 0 0 N2 0 0 v̄Y 1
v̄
ū = ū Y = 0 N1 0 0 N2 0 Z 1 = N v̄, (11.5)
v̄ X 2
ū Z 0 0 N1 0 0 N2 v̄Y 2
v̄ Z 2
This transformation may be applied to the core matrices and vectors derived in Chapter 10. For
example, application to the core tangent stiffness (10.38) yields
1 T A0 T
K̄ = 2 Ḡ SḠ d V = Ḡ (EbbT + sH) Ḡ, (11.7)
L0 V0 L0
which is valid for both end nodes giving v̄i = Tv, i = 1, 2. Consequently the element tangent
stiffness matrix in local coordinates is given by
A0 TT 0 T T 0
K= Ḡ (Ebb + sH) Ḡ
T
. (11.9)
L0 0 TT 0 T
For this simple element all entries may be obtained in closed form and no numerical integration is
necessary.
11–4
11–5 §11.3 THE GENERALIZED CCF
11–5
Chapter 11: THE CORE CONGRUENTIAL FORMULATION: STIFFNESS EQUATIONS11–6
in which S and SG P are called the principal core tangent stiffness and principal geometric stiffness,
T
respectively. Second, transforming the term δ 2 g Φ to freedoms v produces a extra term in
accordance with the schematics
T
K = K M + KG P + KGC , S M → K M , SG P → KG P , δ 2 g Φ → δvT KGC δv, (11.14)
Here (Fδq) is the matrix Fi jk δqk = Fki j δqk ; F being a cubic array. The array G receives the name
tangent transformation matrix. The second term in the expansion of δ 2 gi vanishes because the qi
are assumed to be independent target variables.
Enforcing invariance of δ 2 U yields the tangent stiffness transformation
T
K= G (S M + SG P )G + Q d V = K M + KG P + KGC = K M + KG , (11.16)
V0
11–6
11–7 §11.3 THE GENERALIZED CCF
where the entries of Q are (cf. Remark 4.1) Q i j = Q ji = Fki j k with summation on k = 1, . . . n g .
Note that Q is symmetric because Fki j = Fk ji . Integration of Q over V0 yields the complementary
portion KGC of the geometric stiffness KG .
The internal, applied and prestress force vectors transform according to the formulas in (11.4) with
the G defined in (11.15):
f= T
G Φ d V, p= T
G Ψ d V, p =
0
GT Ψ0 d V. (11.17)
V0 V0 V0
What happens to KU and Kr ? They can be obtained, somewhat artificially, by constructing the
matrix equation
g = Wq, (11.18)
where W is called a secant transformation matrix. Generally this matrix is far from unique because
its n g × n q entries must satisfy only n g conditions. (Care has often to be given to the q j → 0 if
0/0 limits appear in W.) Using (11.18) we can proceed to form
K =
U T U
W S W d V, K =
r
GT Sr W d V. (11.19)
V0 V0
Remark 11.1. The AGGCF is applicable to finite elements with degrees of freedoms that include fixed-axis
rotations, because such rotations are integrable. Examples are provided by two-dimensional beams as well as
plane stress (membrane) elements with drilling freedoms if only in-plane motions are allowed.
Remark 11.2. Why is KGC called a geometric stiffness? Because it vanishes if the current configuration is
stress free, in which case the core internal force Φ vanishes and so does Q.
δgi = G i j δq j , or δg = G δq,
∂G i j (11.20)
δ 2 gi = δq j δqk = Fi jk δq j δqk , or δ 2 g = (Fδq) δq.
∂qk
The transformation equation (11.16) still applies for K whereas (11.17) holds for the force vectors.
But no integral g = g(q) as in the AGGCF exists. Consequently KU and Kr , which require a
secant matrix relation of the form (11.18), cannot be constructed. Furthermore Q is not necessarily
symmetric; a condition for that being Fki j = Fk ji or equivalently ∂G ki /∂q j = ∂G k j /∂qi .
Remark 11.3. For mechanical finite elements the DGCCF naturally arises when three-dimensional finite
rotations are present as nodal degrees of freedom, because such rotations are non-integrable.
11–7
Chapter 11: THE CORE CONGRUENTIAL FORMULATION: STIFFNESS EQUATIONS11–8
Remark 11.4. The relations (11.20) have points of resemblance with the case of non-holonomic constraints
in analytical dynamics.
11–8
11–9 §11.4 A 2-NODE 2D TIMOSHENKO BEAM ELEMENT
1 + g1 1 + − Y κ cos θ g3 − sin θ
g2 γ − Y κ sin θ 1 + g4 cos θ
b1 = = , b2 = = ,
0 0 1 + g1 1 + − Y κ cos θ
0 0 g2 γ − Y κ sin θ
(11.24)
The well known stress resultants of beam theory are the axial force N , transverse shear force V
and bending moment M. They are obtained by integrating the PK2 stresses over the beam cross
section:
N= s1 d A = E A + 12 ( 2 + γ 2 ) + 12 E I κ 2 + N 0 ,
A
0
V = s2 d A = G A s ω γ + V 0 , (11.25)
A0
M= s1 Y d A = −E I κω + M 0 ,
A0
11–9
11–10
Chapter 11: THE CORE CONGRUENTIAL FORMULATION: STIFFNESS EQUATIONS
−1 0 Y (cos θ + (L 0 − X )κ sin θ) 1 0 Y (− cos θ + X κ sin θ )
1 0 −1 Y (sin θ − (L 0 − X )κ cos θ) 0 1 Y (− sin θ − X κ cos θ )
G = G1 G2 =
L0 0 0 −(L 0 − X ) cos θ 0 0 −X cos θ
0 0 −(L 0 − X ) sin θ 0 0 −X sin θ
(11.30)
but it is more instructive (as well as conducive to higher efficiency in the computer implementation)
to perform the transformation phase in two stages.
Observe that the first transformation (from g to q) is nonlinear and algebraic whereas the second
one (from q to v) is linear. Consequently we have to use the AGCCF for the first transformation
but the second one can be done simply through the DCCF.
11–10
11–11 §11.4 A 2-NODE 2D TIMOSHENKO BEAM ELEMENT
QmatrixOf2DTimoBeamElement[eps_,gamma_,kappa_,theta_,Em_,Gm_,Y_]:=
Module[{g,h1,h2,H1,H2,e1,e2,s1,s2,b1,b2,phi,i,j,k},
q={eps,gamma,kappa,theta}; phi={1,1,1,1};
g={eps-Y*kappa*Cos[theta],gamma-Y*kappa*Sin[theta],
-Sin[theta],Cos[theta]-1};
gg={{g[[1]]},{g[[2]]},{g[[3]]},{g[[4]]}};
h1={{1},{0},{0},{0}}; h2={{0},{1},{1},{0}};
H1={{1,0,0,0},{0,1,0,0},{0,0,0,0},{0,0,0,0}};
H2={{0,0,1,0},{0,0,0,1},{1,0,0,0},{0,1,0,0}};
e1=(Transpose[h1].gg+(1/2)*Transpose[gg].H1.gg)[[1,1]];
e2=(Transpose[h2].gg+(1/2)*Transpose[gg].H2.gg)[[1,1]];
s1=Simplify[Em*e1]; s2=Simplify[Gm*e2];
11–11
11–12
Chapter 11: THE CORE CONGRUENTIAL FORMULATION: STIFFNESS EQUATIONS
b1={1+eps-Y*kappa*Cos[theta],gamma-Y*kappa*Sin[theta],0,0};
b2={-Sin[theta],Cos[theta],1+eps-Y*kappa*Cos[theta],
gamma-Y*kappa*Sin[theta]};
phi=Simplify[s1*b1+s2*b2];
Q=Table[0,{4},{4}];
For[i=1,i<=4,i++, For[j=1,j<=4,j++, For[k=1,k<=4,k++,
Q[[i,j]]=Q[[i,j]]+(D[D[g[[k]],q[[i]]],q[[j]]])*phi[[k]] ]]];
Return[Q]
];
The output of this module was integrated over the cross section and pattern matched with the
expression of the stress resultants (11.25)-(11.26) to produce
0 0 0 0
q 0 0 0
KGC = , (11.36)
0 −Mωγ
symm −V ωγ + Mκω − Cκ 2
which added to (11.35) yields the geometric stiffness
N 0 −M cos θ Mκ sin θ − V cos θ
q N −M sin θ −Mκ cos θ − V sin θ
KG = . (11.37)
C −Mωγ
symm −V ωγ + Mκω
q
Finally, the tangent stiffness in terms of q is Kq = K M + KG . Denoting the entries of Kq by K i j ,
i, j = 1, . . . 4 the tangent stiffness matrix K in terms of node displacements v is formed through
the DCCF transformation
q q q q
K 11 K 12 K 13 − 12 L 0 K 14
q q q
K 23 − 12 L 0 K 24
L0 K 22
q q q
K 33 − L 0 K 34 + 13 L 20 K 44
K= G2T Kq G2 d X =
0
symm
q q q q (11.38)
−K 11 −K 12 −K 13 − 12 L 0 K 14
−K 12
q
−K 22
q q
−K 23 − 12 L 0 K 24
q
q q
−K 13 + 12 L 0 K 14
q
−K 23 + 12 L 0 K 24
q q
−K 33 + 16 L 20 K 44
q
.
q q q q
K 13 + 12 L 0 K 14
K 11 K 12
q
K 22
q q
K 23 + 12 L 0 K 24
q q q
K 33 + L 0 K 34 + 13 L 20 K 44
The above rule can be applied to K M and KG should separate formation be desirable, as when
setting up a stability eigenproblem.
If the reference configuration is not aligned with X , the preceding expressions apply to the local
system { X̄ , Ȳ }. A final local-to-global transformation step, similar to that discussed for the 3D bar
11–12
11–13 §11.5 A 2-NODE 3D TIMOSHENKO BEAM ELEMENT
in §11.3, is then necessary. This step can be handled by a simple DCCF transformation, because
the finite rotation θ remains the same in global coordinates.
Remark 11.6. The foregoing exact expressions contain curvature-squared terms typically in the combination
I κ 2 . This can be shown to be of order (r/R)2 compared to other terms, where r is the radius of gyration
of the cross section and R = 1/κ the radius of curvature of the current configuration. For typical beams
(r/R)2 is 10−6 or less; consequently all such tiny terms may be dropped without visible loss of accuracy. For
highly-bent extremely-thin beams, however, that ratio may go up to 0.01 in which case the κ 2 terms might
have a noticeable though small effect if retained.
Table 11.1 Internal energy and its variations for 3D Timoshenko beam element
U = 12 gT SU g + gT Ψ0 — — —
δg = W δw δw = Z δz δz = Gz δv
where δΘ, defined in (10.68), measures the variation of angular orientation. Because this quantity
is not generally integrable for three-dimensional motions, it is not possible to express Θ as a unique
function of the displacements. The variation of g1 is
dδu0 T dδΘ T
ζ̃T κ,
δg1 = + RT ζ̃ + RT ζ̃ κ̃δΘ + RT δΘ (11.41)
dX dX
T
where we used the relation [1] δκ = dδΘ/d X + κ̃δΘ. On using the commutative law ãb = b̃ a
= ãb̃ − b̃ã we may rewrite (11.41) as
and Jacobi’s identity ãb
∂ δu0 T ∂ δΘ
δg1 = + RT ζ̃ + RT κ̃T ζ̃ δΘ (11.42)
∂X ∂X
11–14
11–15 §11.5 A 2-NODE 3D TIMOSHENKO BEAM ELEMENT
where I is the 3-by-3 identity matrix and Wi are 3-by-9 matrices. The second variation of g, which
is required for the complementary geometric stiffness, is
ζ̃ T d δΘ ζ̃T d δΘ
ζ̃T κ̃ δΘ + RT δΘ
δ 2 g1 = RT δΘ + RT δΘ
dX dX
ζ̃T κ̃δΘ + δ 2 RT ζ̃T κ + RT ζ̃T δ 2 κ,
+ RT δΘ (11.44)
δ 2 g2 = δ 2 RT i2 , δ 2 g3 = δ 2 RT i3
P = Aσb + P0 , sb = Eeb ,
Q = µs A + Q0 , τ = τ2 + τ3 , τ2 = Gγ2 h2 , τ3 = Gγ3 h3 ,
Mσ = E I S Ke + Mσ ,
0
IS = ζζT d A, Ke = φ̃ κ, (11.45)
A0
T
Mτ = µt G I P κ + Mτ ,
0
IP = ζ̃ ζ̃ d A.
A0
Here P, Q, Mσ and Mτ are axial forces, shear forces, bending moments and torsional moments,
respectively, at the current configuration C; P0 , Q0 , M0σ and M0τ are similar quantities at the
reference configuration C0 ; µs and µt are transverse-shear and torsion coefficients that account for
the actual shear stress distributions, respectively; and I S and I P are the cartesian and polar inertia
tensors, respectively, of the cross section. Should the axes Y and Z be aligned with the principal
inertia axes the latter simplified to
0 0 0 I22 + I33 0 0
IS = 0 I22 0 , IP = 0 I33 0 . (11.46)
0 0 I33 0 0 I22
Because the relation between g and w is of differential type the applicable transformation rules are
those the DGCCF, and no energy or secant stiffness survives. Thus only the internal force vector
R and tangent stiffness S associated with w are derived below.
11–15
11–16
Chapter 11: THE CORE CONGRUENTIAL FORMULATION: STIFFNESS EQUATIONS
where Rσ and Rτ are the contributions of the normal and shear stresses respectively. Detailed
calculations result [1] in the following exact expressions:
RT (P φ + κ̃M )
σ RT Q
Rσ = T
φ̃ Mσ , Rτ = Mτ . (11.48)
T
T
Ke Mσ φ̃ Q + κ̃ Mτ
T
These resemble the classic linearized theory equations. Furthermore observe that the term P RT φ
corresponds to the internal force of the TL 3D bar.
Tangent Stiffness. For the tangent stiffness we have the decomposition
S = S M + SG P + SGC . (11.50)
Material Stiffness. The generalized core material stiffness is given by the congruential transforma-
tion
SM = W SM W d A =
T
E i WT bi biT W d A = S Mσ + S Mτ . (11.53)
A0 i A0
11–16
11–17 §11.5 A 2-NODE 3D TIMOSHENKO BEAM ELEMENT
ART I⊥ R 0 ART I⊥ φ̃ 0 0 0
S Mτ = µG IP I P κ̃ , in which I⊥ = 0 1 0 . (11.55)
T
symm Aφ̃ I⊥ φ̃ + κ̃ I P κ̃
T 0 0 1
where F contains SGCq as q level kernel. Carrying out the algebraic manipulations one obtains
T
σ
PI RT M σ RT κ̃T M
SGwσ = SG Pσ + SGCwσ = 0 σ φ̃
M (11.58)
symm σ κ̃ + κ̃T M
φ̃M σ φ̃
Consequently
T
SGCqσ = V(φ̃ Mσ ) + U(κ̃Mσ ; φ). (11.60)
Because the next-level target variables q include the finite rotation parametrization, matrices V and
U depend on that choice. They are the source of unsymmetries in the stiffness matrices when certain
rotational parametrizations are adopted, such as the incremental rotation vector. If the rotational
vector is chosen these matrices are symmetric.
Geometric Stiffness due to Shear Stresses. The contribution of the shear stresses to the geometric
stiffness is
SGτ = SG Pτ + SGCwτ + SGCqτ = SGwτ + SGCqτ . (11.61)
The appropriate definitions are
SG Pτ = s12 (W1T HW2 + W2 HW1 ) + s13 (W1T HW3 + W3 HW1 ) d A
A
0 (11.62)
SGCτ = (s12 b2 + s13 b3 ) δ 2 g d A = δwT SGCwτ δw + Fτ (δ 2 R, δ 2 κ).
A0
11–17
11–18
Chapter 11: THE CORE CONGRUENTIAL FORMULATION: STIFFNESS EQUATIONS
Carrying out manipulations one obtains the surprisingly simple form for SGwτ
0 0 RT Q T
SGwτ = SG Pτ + SGCwτ = 0 0 . (11.63)
symm 0
The terms due to the second variation of g become
Fτ = QT δ 2 RΦ + MτT δ 2 κ. (11.64)
The kernel carried forward to the q level is
SGCqτ = V(Mτ ) + U(Q ; Φ). (11.65)
11–18
11–19 §11.5 A 2-NODE 3D TIMOSHENKO BEAM ELEMENT
where
τ
U = c1 τ ΦI + c2 τΦ + Φ τ + c3 τ Φαα + τ α̃ΦI + τ̃ Φα + αΦ τ̃
T T T TT T T T T
+ c5 αT Φ ταT + ατT + αT τ αΦT + ΦαT + τT ααT ΦI
+ c4 τT α̃ΦααT + c6 τT ααT ΦααT ,
T + c3 αMT + c5 ααT M
Vτ1 = c2 M τ + c7 Mτ αT + αT Mτ I + c8 αT Mτ ααT ,
τ τ
dα T
dα T
dα
T
dα
dα
Vτ2 = −c3 Mτ I − c4 Mτ ααT + c5 Mτ αT + α MτT + αT Mτ I
dX dX dX dX dX
T dα dα dα T
+ c6 α Mτ αα + c7
T
Mτ + Mτ
T
+
dX dX dX
dα T
dα T
+ c8 α αMτT + Mτ αT + αT Mτ I + c9 ααT Mτ ααT ,
dX dX
(11.73)
in which
sin α 1 − cos α sin α − α cos α
c1 = − , c2 = , c3 = ,
α α2 α3
c1 + 3c3 c1 + 2c2 c3 + 4c5
c4 = − , c5 = − , c6 = − , (11.74)
α2 α2 α2
1 + c1 3c3 − 2c2 c5 − 5c8
c7 = , c8 = , c9 = .
α2 α2 α2
A similar approach can be taken with (11.69), which defines Fσ . The tangent stiffness matrix can
be obtained by superposing all contributions.
11–19
11–20
Chapter 11: THE CORE CONGRUENTIAL FORMULATION: STIFFNESS EQUATIONS
The choice of shape functions for the rotational vector poses some subtle questions. In small-
deflection analysis it is common practice to select all Timoshenko beam shape functions to be linear
in X . This choice obviously enforces nodal compatibility while preserving constant curvature states.
But for finite deflections a linear interpolation for the rotational vector components cannot exactly
represent a constant curvature state unless the rotations are about a single axis (plane rotations).
The same is true if the rotation matrix R(X ) is interpolated linearly. On the other hand, linear
interpolation of Euler parameters does preserve the constant curvature state. This motivated the
development of an interpolation scheme that starts from the 4 Euler parameters i (X ), i = 0, 1, 2, 3,
i i = 1 that orient the normal of a cross section at X . These are collected in the 4-vector
2
= [ 0 1 2 3 ]T . Given the eight end values (0) and (L) the interpolation that can copy
a constant curvature vector κ is found to be [1]
tan(ζ ) sin(ζ )
(ζ ) = cos(ζ ) 1 − (0) + (L), (11.77)
tan(ζ L ) sin(ζ L )
√
where ζ = 12 κ X , ζ L = 12 κ L, κ = κT κ. The constant curvature vector can be extracted from the
end values through the formula
1
− 20 (0)I (0) ,
κ= (L) − 20 (L)I (0) − (0) (11.78)
β2 L
This interpolation is then transformed to the variations in terms of the rotational vector. Details are
provided in Reference [1].
§11.6. Equivalence of DCCF and Standard TL Formulation
The correspondence between the Direct Core Congruential Formulation (DCCF) and the Standard Formulation
(SF) of the Total Lagrangian (TL) kinematic description is established below for 3D continuum finite elements.
This connection was worked out in a course term project [4]. Such elements fit within the DCCF framework
because their physical DOFs (node displacements) are of translational type.
The Standard Formulation is based on the same scheme used for linear finite elements: first interpolate, then
vary. As in the linear case, the departure point is extremization of the Total Potential Energy functional (TPE)
over the element domain:
=U −W = e s0 d V +
T 1
2
e Ee d V −
T
u b dV −
T
uT t d S, (11.79)
V0 V0 V0 St0
where as usual conservative dead loading is assumed. In (11.79), b is the prescribed body force field, t are
surface tractions prescribed over portion St0 of the boundary in C0 , and other quantities are as defined in Section
4. The weak equilibrium equations are obtained on making (11.79) stationary:
δ = δU − δW = δe s0 d V +
T
δe Ee d V −
T
δu b d V −
T
δuT t d S = 0. (11.80)
V0 V0 V0 St0
The displacement and strain fields are interpolated in terms of the element degrees of freedom v:
where B = B(v) depends in v but N does not. Substituting these interpolations into (11.80) yields the residual
equilibrium equations
δ = δvT r = δvT (f − p) = 0. (11.82)
11–20
11–21 §11.6 EQUIVALENCE OF DCCF AND STANDARD TL FORMULATION
where
f= B (s0 + Ee) d V =
T T
B s d V, p= N b dV +
T
Ns d S, (11.83)
V0 V0 V0 St0
where f and p are the internal and external force vectors, respectively, and s = s0 + Ee are the PK2 stresses in
C. Because the variations δv are arbitrary, the residual-force nonlinear equilibrium equation is r = f − p = 0
or f = p. The tangent stiffness matrix is given by
∂r ∂f
K= = , (11.84)
∂v ∂v
because p (for conservative dead loading) does not depend on v. Splitting B = Bc + Bv (v), where Bc is
constant but Bv depends on v, gives the well known decomposition
K = K0 + K D + KG , (11.85)
where K0 , K D and KG denote the linear, initial-displacement and geometric stiffness matrices, respectively.
These are given by
K0 = BcT EBc d V,
V0
KG δv = δBT s d V.
V0
To correlate these standard forms with those produced by the DCCF, we note that the GL strains can be also
split as e = ec + ev , where ec and ev are linear and nonlinear in v, respectively. The latter may be expressed
in terms of the displacement gradients as
ev = 12 Ag, (11.87)
where A is the 6 × 9 matrix
gT 0 0
g g2 g3 0 0 0 0 0 0
1 1
0 g2T 0 0 0 0 g4 g5 g6 0 0 0
0 g3T 0 g9
A= = ,
0 0 0 0 0 0 g7 g8
0 0 (11.88)
g3T g2T 0 0 g7 g8 g9 g4 g5 g6
T
g3 0 g1T g7 g8 g9 0 0 0 g1 g2 g3
g2T g1T 0 g4 g5 g6 g1 g2 g3 0 0 0
in which the displacement gradients are vector-arranged as
∂u 1 ∂u 2 ∂u 2 ∂u 3 .
gT = [ g1 g2 ··· g8 g9 ] = ∂ X1 ∂ X1 ··· ∂ X3 ∂ X3 (11.89)
Comparing
δev = 1
2
δA g + 12 A δg = A δg, (11.90)
to the DCCF transformation relation δg = G δv, in which G is independent of v, we see that
Bv = AG. (11.91)
The other expression we require is δAT s, which appears in the geometric stiffness matrix contracted with δv:
KG δv = δB s d V =
T
GT δAT s d V. (11.92)
V0 V0
11–21
11–22
Chapter 11: THE CORE CONGRUENTIAL FORMULATION: STIFFNESS EQUATIONS
where I is the 3 × 3 identity matrix and si , i = 1, . . . 6 are components of the PK2 stress tensor ordered
s1 = s11 , s2 = s22 , . . . s6 = s23 . Using this relation, KG can be placed in the standard form
KG = GT MG d V, (11.94)
V0
The numerical D matrix can be easily related to the hi vectors introduced in (10.10). Because both D
and G are independent of v it follows that δec = DG δv and consequently Bc = DG. Partitioning A as
[ a1T a2T . . . a6T ] one easily finds that ai = Hi g. Now the following identities can be verified through
simple algebra:
DT ED = E i j hi hiT = S0 ,
DT EA = E i j hi aTj = E i j hi gT H j = S1 , AT ED = S1T ,
(11.96)
AT EA = E i j ai a j = E i j Hi gi gTj H j = S2 = S2T ,
M = si0 Hi + E i j hi gH j + 12 (gT Hi g)H j = si0 Hi + S∗1 + 12 S∗2 = si Hi .
KD = G (D EA + A ED + A EA)G d V =
T T T T
GT S D G d V, (11.97)
V0 V0
KG = G MG d V =
T
GT SG G d V,
V0 V0
which displays the equivalence of both formulations when no approximations are made. This proof may be
extended without difficulty to the AGCCF in which case G is a function of v, although as noted in the text that
situation is sometimes mishandled in the Standard Formulation through the introduction of a priori kinematic
approximations. The equivalence between DGCCF and SF is more difficult to prove because there is no TPE
functional from which the latter can be derived, and such connection should be regarded as an open problem.
11–22
11–23 §11.7 REFERENCES
§11.7. References
[1] L. A. Crivelli, A Total-Lagrangian beam element for analysis of nonlinear space structures, Ph. D.
Dissertation, Dept. of Aerospace Engineering Sciences, University of Colorado, Boulder, CO, 1990.
[2] C. A. Felippa and L. A. Crivelli, A congruential formulation of nonlinear finite elements, in Nonlinear
Computational Mechanics - The State of the Art, ed. by P. Wriggers and W. Wagner, Springer-Verlag,
Berlin, pp. 283–302, 1991.
[3] L. A. Crivelli and C. A. Felippa, A three-dimensional non-linear Timoshenko beam element based on
the core-congruential formulation, Int. J. Numer. Meth. Engrg., 36, pp. 3647–3673, 1993.
[4] F. Abedzadeh Anaraki, A. Barzegar Mehrabi and H. R. Lofti, Correspondence between CC-TL and
C-TL formulations, in Term Projects in Nonlinear Finite Element Methods, ed. by C. A. Felippa, Report
CU-CSSC-91-12, Center for Space Structures and Controls, University of Colorado, Boulder, CO, May
1991.
[5] O. C. Zienkiewicz, The Finite Element Method, 3r d ed., McGraw-Hill, London, 1976.
11–23
11–24
Chapter 11: THE CORE CONGRUENTIAL FORMULATION: STIFFNESS EQUATIONS
Not Assigned
EXERCISE 11.1
Consider a two-node geometrically TL nonlinear bar in three dimensions built with the CCF. Write subroutines
to compute the tangent stiffness matrix K and the PK2 stress s in the current configuration given the following
input data:
(1) The coordinates of both element nodes in the reference configuration global Cartesian system X, Y, Z .
(2) The X, Y, Z displacements v Xi , vY i , v Zi of the two element end nodes i = 1, 2.
(3) The elastic modulus E and the reference bar cross section A0 .
(4) The bar axial stress s 0 in the reference configuration.
The subroutines you have to write (in Fortran, C or C++) have the following names and calling sequence
interface:
BAR3F (X0, v, E, A0, s0, F, status)
BAR3K (X0, v, E, A0, s0, K, status)
BAR3S (X0, v, E, s0, s, status)
BAR3F and BAR3K compute the internal force vector f and tangent stiffness matrix K, respectively, in the current
configuration. Subroutine BAR3S computes the PK2 stress s in the current configuration.
Input variables (Fortran assumed in description):
11–24
11–25 Exercises
For intermediate manipulations, construct the local bar axis X joining nodes 1 to 2 in C0 . Then construct Y ,
and Z normal to X forming a right-handed system. As noted §11.2.1, there is some arbitrariness because Y , Z
may be “gyrated” around X without changing the final answer; select whatever orientation rule seem to be
more computationally “robust” in the sense that it should not fail for arbitrary bar orientations.
The answer for this exercise should be a listing of BAR3F, BAR3K and BAR3S.
EXERCISE 11.2
Explain the rule you chose to orient Y and Z . (Appropriate comments in the BARK source code should be
sufficient to answer this one.)
EXERCISE 11.3
Test the subroutines on the following input data:
To feed these values write a short test driver that calls the two subroutines in turn; a suggested driver is listed
under “Programming Recommendations.”
Compare the output of BAR3K with the following tangent stiffness matrix:
0.935471 0.097502 −0.071172 −0.935471 −0.097502 0.071172
0.097502 1.622137 −0.511148 −0.097502 −1.622137 0.511148
−0.071172 −0.511148 0.511148 −1.295011
K=
1.295011 0.071172
−0.935471 −0.097502 (E11.1)
0.071172 0.935471 0.097502 −0.071172
−0.097502 −1.622137 0.511148 0.097502 1.622137 −0.511148
0.071172 0.511148 −1.295011 −0.071172 −0.511148 1.295011
Compare the output of BAR3F to the following internal force vector:
−0.912675
−6.554668
4.784631
f=
0.912675
(E11.2)
6.554668
−4.784631
The computed PK2 stress s in the current configuration returned by BAR3S should be 5.603088.
11–25
10
.
The Core-Congruential
Formulation:
Core Equations
10–1
Chapter 10: THE CORE-CONGRUENTIAL FORMULATION: CORE EQUATIONS 10–2
TABLE OF CONTENTS
Page
§10.1. Introduction 10–3
§10.2. Overview 10–3
§10.2.1. Basic Concepts . . . . . . . . . . . . . . . . . 10–4
§10.2.2. Direct and Generalized CCF . . . . . . . . . . . . 10–4
§10.2.3. The CCF Philosophy: Divide and Conquer . . . . . . . 10–5
§10.3. Historical Background 10–6
§10.4. Core Stiffness Equations 10–8
§10.4.1. TL Description of Particle Motion . . . . . . . . . . 10–8
§10.4.2. Energy Variations . . . . . . . . . . . . . . . . 10–9
§10.4.3. Parametrized Forms . . . . . . . . . . . . . . . 10–9
§10.4.4. Spectral Forms . . . . . . . . . . . . . . . . . 10–11
§10.4.5. Generalization to H(g) . . . . . . . . . . . . . . 10–11
§10.5. Core Stiffness Derivation Examples 10–12
§10.5.1. Bar in 3D Space . . . . . . . . . . . . . . . . 10–12
§10.5.2. Plate in Plane Stress . . . . . . . . . . . . . . . 10–14
§10.5.3. Plate Bending . . . . . . . . . . . . . . . . . 10–16
§10.5.4. 2D Timoshenko Beam . . . . . . . . . . . . . . 10–18
§10.5.5. 3D Timoshenko Beam: Kinematics . . . . . . . . . . 10–20
§10.5.6. 3D Timoshenko Beam: Core equations . . . . . . . . 10–23
§10.6. References 10–24
§10. Exercises . . . . . . . . . . . . . . . . . . . . . . 10–25
10–2
10–3 §10.2 OVERVIEW
In Chapter 8 it was noted that two methods for developing geometrically nonlinear elements based
on the Total Lagrangian description exist: the Standard Formulation (SF) and the Core-Congruential
Formulation or CCF. Chapters 10 and 11 cover the second method. The present exposition is taken
largely from a recent survey article [1].
§10.1. Introduction
There is an elegant Total Lagrangian (TL) formulation of geometrically nonlinear mechanical finite
elements that has received little attention in the literature. This will be referred to as the Core-
Congruential Formulation, or CCF, in the sequel. The key concepts, presented by Rajasekaran and
Murray [2] in 1973, evolved from the analysis and reinterpretation of the pioneer work of Mallet
and Marcal [3] as well as Murray’s previous work in geometrically nonlinear finite element analysis
[4]. The discussion of Reference [2] by Felippa [5] provided parametric expressions for the stiffness
matrices that appear at various levels of the discrete governing equations. This work originated
what is called here the Direct Core Congruential Formulation, or DCCF.
In 1987 this course presented the derivation of several elements using the DCCF. Preparation
of homework assignments and feedback from students in this and follow-up offerings helped to
streamline the material. Subsequently Crivelli’s doctoral thesis [6] used the CCF in the systematic
development of a three-dimensional nonlinear Timoshenko beam element capable of undergoing
arbitrarily large rotations. Challenges posed by this application pushed this formulation beyond
frontiers hitherto deemed impassable by a TL element with rotational degrees of freedom. This
development was summarily reported in a survey article by Felippa and Crivelli [7] and explained
in more detail in a subsequent paper by Crivelli and Felippa [8].
A lesson gained from this research is that, when dealing with 3D finite rotations, the CCF should
be applied in a staged fashion that allows the systematic examination of additional terms arising in
the transformations to physical degrees of freedom. That transformation methodology gave rise to
what is here called the Generalized CCF, or GCCF.
Both DCCF and GCCF share the same “divide and conquer” philosophy. However, the core
equations as well as subsequent steps that transform those equations to physical freedoms vary
in complexity. To simplify the exposition while focusing on the essential aspects, Sections 10.3
through 10.7 focus on the DCCF. Examples of application to elements amenable to the direct
treatment are presented. The GCCF is discussed in Chapter 11, and illustrated with applications to
2D and 3D beam elements.
Remark 10.1. Several authors have expressed the belief that the approximation performance of TL-based
elements degrades beyond moderate rotations, and an updated Lagrangian or corotational description is nec-
essary for handling truly large motions. For example, in 1986 Mathiasson, Bengtsson and Samuelsson [9]
concluded that “The TL formulation can only be used in problems with small or moderate displacements.”
More recently Bergan and Mathisen [10] voice a similar opinion: “it is commonly known that in a step by step
TL formulation artificial strains easily arise in beam elements due to nonhomogeneities in the displacement
expansions in transverse and longitudinal directions.” Our experience shows that such limitations are not in-
herent in the TL description but instead emerge when a priori kinematic approximations are made to simplify
element derivations. The 3D beam element just cited exhibits computational and approximation performance
for very large rotations comparable to those based on the co-rotational and Updated Lagrangian descriptions
while retaining certain advantages listed in the Conclusions.
10–3
Chapter 10: THE CORE-CONGRUENTIAL FORMULATION: CORE EQUATIONS 10–4
§10.2. Overview
where S is the core stiffness matrix, K the physical stiffness in terms of the nodal degrees of
freedom v, G a core-to-physical-freedom transformation matrix assumed to be independent of
v, V0 the appropriate reference integration volume, and in which “level” identifies the governing
equation level at which the stiffness matrix is used.
The three variational levels of interest in practice are: energy (level 0), force equilibrium (level 1),
and first-order incremental equilibrium (level 2). Qualifiers “residual-force” and “secant-stiffness”
are also used for level 1, and “tangent-stiffness” used for level 2.
The core stiffness matrix is expressed in terms of the displacement gradients at each material point.
Displacement gradients g make a better choice of core variables than finite strains because for
elements with translational degrees of freedom (DOFs) they can be expressed linearly in terms of
node displacements v as g = Gv, a property that validates (10.1) for all levels. As discussed below,
such elements fall under the purview of the Direct CCF.
The qualifier “core” emphasizes the goal of independence of Slevel with respect to discretization
decisions such as element geometry, shape functions, and choice of nodal degrees of freedom.
Such a dependence is introduced by the congruential transformation indicated in (10.1) and the
integration over the element volume.
But this panoramic view needs to be rendered more precise. If the relation between core DOFs
(the displacement gradients g) and the physical DOFs (the node displacements v of a finite element
model) is linear, these transformations do not depend on level:
10–4
10–5 §10.2 OVERVIEW
In this diagram, numbers annotated within the “core box” denote the variational level of the gov-
erning equation in use. Internal force and secant stiffness are two alternative governing-equation
expressions at level 1. The energy level (level 0) may also be expressed in several ways, but this is
not shown in the diagrams to reduce clutter. Under the aforementioned assumption we obtain the
Direct Core Congruential Formulation, or DCCF.
If the relation between displacement gradients g and node displacements v is nonlinear, the trans-
formations sketched above are not only more complex but depend on variational level and possibly
the expression form used within a level. This complication arises when elements with rotational
degrees of freedom such as beams, plates and shells are considered. It gives rise to the Generalized
Core Congruential Formulation, or GCCF.
Two variants of the GCCF may be distinguished. If the relation between g and v is nonlinear
but algebraic, the transformation equations do vary with level but in principle are still possible as
illustrated in the following diagram.
S-Congruential
(0) Core Energy Stiffness ⇒ Transformation ⇒ Physical Energy Stiffness
(1) Core Secant Stiffness Equations Physical Secant Stiffness
T-Congruential
(1) Core Internal Force ⇒ Physical Internal Force
Transformation ⇒
(2) Core Tangent Stiffness Equations Physical Tangent Stiffness
Here “T-Congruential” and “S-Congruential” are abbreviations for “Tangent Congruential” and
“Secant-Congruential,” respectively. Such a distinction is elaborated upon in Chapter 11.
If the relation between g and v is nonlinear and can be expressed only in non-integrable differential
form, the “Secant Transformation Equations” of the preceding diagram do not generally exist, and
the diagram must be truncated:
T-Congruential
(1) Core Internal Force ⇒ Physical Internal Force
Transformation ⇒
(2) Core Tangent Stiffness Equations Physical Tangent Stiffness
These two variants of the GCCF are called Algebraic GCCF and Differential GCCF and denoted by
acronyms AGCCF and DGCCF, respectively, in the sequel. The main distinction between AGCCF
and DGCCF is that it makes no sense to talk about missing quantities, such as the physical secant
stiffness, with the latter.
The original development of the CCF outlined in the Introduction focused on elements with
translational-degree-of-freedom configurations. For such elements the Direct form of the CCF,
or DCCF, is sufficient. Sections §10.3 through §10.7 focus on that form, leaving the development
and application of the GCCF to the next Chapter.
10–5
Chapter 10: THE CORE-CONGRUENTIAL FORMULATION: CORE EQUATIONS 10–6
10–6
10–7 §10.3 HISTORICAL BACKGROUND
In 1968 Mallet and Marcal [3] attempted to establish a standard nomenclature for geometrically
nonlinear finite element structural analysis based on the Total Lagrangian (TL) kinematic descrip-
tion. Consider a discrete, finite element model of a static structural system under dead loading
with nodal displacement degrees of freedom collected in array v. Displacements are measured
from a fixed reference configuration C0 to a current configuration C. The virtual-work conjugate
forces, independent of v, are collected in array p. The system has a total potential energy function
= U − P that is the difference between the strain energy U and the loads potential P = pT v.
The residual node forces are r = ∂/∂v, and the symbol
denotes increment associated with
the variation of the current configuration. (In keeping up with the spirit of Reference [3] actual
variations are used below rather than virtual ones; the latter are identified by the usual δ prefix.)
Mallet and Marcal expressed the total potential energy, the residual (force-balance) equilibrium
equations, and the incremental equilibrium equations as follows:
= U − P = 12 vT K0 + 13 N1 + 16 N2 v − pT v, (10.2)
∂
r= = K0 + 12 N1 + 13 N2 v − p = 0, (10.3)
∂v
r = [K0 + N1 + N2 ]
v −
p = 0. (10.4)
Here K0 is the linear stiffness matrix evaluated at the reference configuration, whereas N1 and N2
are nonlinear stiffness matrices, also evaluated at the reference configuration, that depend linearly
and quadratically, respectively, on the node displacements v. The N matrices were said “to repeat”
in the foregoing expressions. (This old notation has not survived; presently symbol N is most
commonly used to identify matrices of element shape functions.)
Five years later Rajasekaran and Murray [2] examined more critically the structure of the matrices
that appear in the above equations. In that investigation they chose to start from the “core” stiffness
matrices corresponding to K, N1 and N2 expressed in terms of displacement gradients, and in
doing so laid down the main idea of the CCF. Working with specific elements they showed that
the nonlinear stiffness matrices N1 and N2 are not uniquely determined. Indeed (10.2)-(10.4) as
written are unique only for a single degree of freedom. They did not present, however, a general
expression valid for arbitrary elements. This was partly done by Felippa [5], who in the discussion
of Reference [2] considered again those equations, rewritten here in a more general and compact
form:
T
= 12 vT KU v + p0 − p v, (10.5)
r = Kr v + p0 − p = f − p = 0, (10.6)
r = K
v −
p = 0, (10.7)
in which the notation of this paper — rather than that of Reference [5] — is used. Here KU , Kr
and K denote the energy, secant and tangent stiffness matrices, respectively. (Energy and secant
stiffnesses are not denoted by Ke and Ks because such symbols are used for other purposes in
the finite element course noted in the Introduction.) In addition, p0 is the prestress force vector,
which vanishes if the reference configuration is stress free and was omitted in that discussion, [5]
and f = Kr v + p0 is the internal force vector. The tangent stiffness is of course fundamental in
incremental-iterative solution methods and stability analysis, while the secant stiffness (by itself or
10–7
Chapter 10: THE CORE-CONGRUENTIAL FORMULATION: CORE EQUATIONS 10–8
where si0 are stresses in the reference configuration (stresses that remain if ei = 0, also called
prestresses) and E i j are elastic moduli arranged as a n s × n s square array in the usual manner.
Let π, U, P, Ψ, Φ and Υ denote the analogues of , U , P, p, f and r, respectively, at the particle
level. (The first three acquire the meaning of energy densities, whereas Ψ is a dead-loading body
force density independent of u.) The strain energy density can be expressed as
10–8
10–9 §10.4 CORE STIFFNESS EQUATIONS
Next, introduce the n g displacement gradients gmn = ∂u m /∂ X n . These are subsequently identified
as gi (i = 1, 2, . . . n g ) so they can be conveniently arranged in a one-dimensional array g. Following
Rajasekaran and Murray [2] and Felippa [5] assume that the strains ei are linked to the displacement
gradients through matrix relations of the form
ei = hiT g + 12 gT Hi g , i = 1, 2, . . . n s (10.10)
where hi and Hi are arrays of dimension n g × 1 and n g × n g , respectively, with Hi symmetric.
In the original References [2,5] it was assumed that Hi is independent of g, which is the case for
the Green-Lagrange strain measure. This restriction, labeled (R3) in §10.3, will be enforced below
except in §10.4.5.
π = U − P = 12 gT SU g + (Ψ 0 − Ψ)T g, (10.11)
∂H
Υ= = Sr g + Ψ 0 − Ψ = Φ − Ψ = 0, (10.12)
∂g
Υ = S
g −
Ψ = 0. (10.13)
Here SU , Sr and S denote the energy, secant and tangent core stiffness matrices, and Ψ0 , which is
independent of g, is the core counterpart of p0 .
With this notation the first and second variations of the strain energy density can be expressed as
δU = δgT (SU g + Ψ0 ) + 12 gT δSU g = δgT Sr g + Ψ0 = δgT Φ, (10.14)
δ 2 U = δgT Sr δg + δgT δSr g + (δ 2 g)T Φ = δgT S δg + (δ 2 g)T Φ. (10.15)
These variational equations implicitly determine Sr , Φ and S from SU and Ψ0 . If the linearity
restriction (R4) holds, the term in δ 2 g drops out as explained in the Remark below, and
∂gi ∂ 2 gi ∂gi 2 0
δgi =
∂v j
δv j = G i j δv j , δ 2 gi =
∂v j ∂vk
δv j δvk +
∂v j
δ v j = Fi jk δv j δvk . (10.17)
Thus δg is still G δv but δ 2 g = (F δv) δv, where F is a cubic array. The presence of the term δ 2 g is taken into
account in the GCCF discussed in Chapter 11.
10–9
Chapter 10: THE CORE-CONGRUENTIAL FORMULATION: CORE EQUATIONS 10–10
S D = S1 + S1T + S2 , S M = S0 + S D ,
(10.21)
SG = S∗1 + 12 S∗2 + si0 Hi = si Hi .
These are the core versions of the initial-displacement, material and geometric stiffness, respec-
tively. The core tangent stiffness is S = S0 + S D + SG = S M + SG .
If the Generalized CCF is required for downstream element development as explained in Chapter
11, SG = si Hi is called the principal core geometric stiffness and is denoted by SG P . In this case
the combination
S = S M + SG P , (10.22)
receives the name principal core tangent stiffness.
10–10
10–11 §10.4 CORE STIFFNESS EQUATIONS
Remark 10.3. Finite element practicioners may be surprised at the nonuniqueness of SU and Sr . It appears to
contradict the fact that, given two square matrices A1 and A2 and an arbitrary nonzero test vector x, A1 x = A2 x
for all x implies A1 = A2 . But this is not necessarily true if A1 and A2 are functions of x. More precisely, the
energy core stiffness is not unique because
gT (S1 − S∗1 )g = 0, gT (S1T − S∗1 )g = 0, gT (S2 − S∗2 )g = 0, (10.23)
and the secant core stiffness is not unique because
(S1T − S∗1 )g = 0, (S2 − S∗2 )g = 0. (10.24)
Adding “gage terms” such as those of (10.24) multiplied by arbitrary coefficients does not change δU and
consequently the secant stiffness acquires two free parameters. Uniqueness holds for the tangent stiffness
because the test vectors are the virtual displacement gradient variations, and S is not a function of δg.
Remark 10.4. Because of (10.23), an additional free parameter appears in SU if unsymmetry is allowed. If
symmetry is enforced the first two gage expressions must be combined to read gT (S1 + S1T − 2S∗1 )g = 0.
10–11
Chapter 10: THE CORE-CONGRUENTIAL FORMULATION: CORE EQUATIONS 10–12
δ 2 U = δgT (S +
S) δg + (δ 2 g)T (Φ + Φ). (10.32)
S ,
where
r
are additional core terms that arise on account of the dependence of the Hi on g.
S and Φ
The parametrization and efficient characterization of such terms for several strain measures of
interest in practice, notably logarithmic and midpoint strains, are presently open problems. Such
topics would in fact be good candidates for term projects in advanced nonlinear finite element
courses.
Because the core equations reflect the motion of an individual particle, their form is primarily
determined by the choice of components of s, e and g that are retained in the strain energy density.
This choice is in turn a byproduct of the mathematical idealization of the actual structure or structural
component.
Several cases are worked out below to illustrate the basic steps. The core expressions developed in
these examples do not force commitment to specific elements, only to a mathematical model. For
example the bar core equations may be subsequently used to develop 2-node straight elements or
3-node curved ones. Some specific elements based on these equations are derived in Chapter 11.
The particle belongs to a bar moving in 3D space. The only energy contribution is due to the axial
(longitudinal) stress. We have n d = 3, n s = 1 and n g = 3. To simplify node subscripting, Cartesian
systems and displacement components will be denoted by {X, Y, Z }, {x, y, z} and {u X , u Y , u Z }
rather than {X 1 , X 2 , X 3 }, {x1 , x2 , x3 } and {u 1 , u 2 , u 3 }, respectively. In the reference configuration
C0 the bar is referred to a local Cartesian system { X̄ , Ȳ , Z̄ }, with X̄ located along the bar axis. See
Figure 10.1.
With reference to this local system, the motion of a particle initially at X̄ is defined by the displace-
ment components ū X = ū X ( X̄ ), ū Y = ū Y ( X̄ ) and ū Z = ū Z ( X̄ ). The three displacement gradients
that intervene in the definition of nonlinear strains are
g1 ∂ ū X /∂ X̄
g= g2 = ∂ ū Y /∂ X̄ . (10.33)
g3 ∂ ū Z /∂ X̄
10–12
10–13 §10.5 CORE STIFFNESS DERIVATION EXAMPLES
Z, z
2'
Y, y X, x C
Z̄ v̄Y 1 1'
v̄ Z 2
Ȳ
v̄ Z 1
1
v̄Y 2
C0
E, A0
L0
2 v̄ X 1
X̄
v̄ X 2
As uniaxial strain measure we adopt the Green-Lagrange (GL) axial strain, defined as
2
2
2
∂ ū X ∂ ū X ∂ ū Y ∂ ū Z
e ≡ e1 = + 12 + + = g1 + 12 (g12 + g22 + g32 )
∂ X̄ ∂ X̄ ∂ X̄ ∂ X̄
T T (10.34)
1 g1 g1 1 0 0 g1
= 0 g2 + 2 g2
1
0 1 0 g2 = hT g + 12 gT Hg.
0 g3 g3 0 0 1 g3
Thus for this choice of strain, h1T ≡ hT = [ 1 0 0 ] and H1 ≡ H is the 3 × 3 identity matrix.
The conjugate stress measure s1 ≡ s is the second Piola-Kirchhoff (PK2) axial stress. The stress-
strain relation is s = s 0 + Ee, where s 0 and s are PK2 axial stresses in the reference and current
configurations, respectively, and E is Young’s modulus.
Because H is independent of g, to form the core stiffnesses in local coordinates we can directly use
the spectral expressions (10.26)–(10.29). First construct the vectors
1 + 1g
2 1 1 + g1
c ≡ c1 = 12 g2 , b ≡ b1 = g2 , (10.35)
1
g3 g3
2
10–13
Chapter 10: THE CORE-CONGRUENTIAL FORMULATION: CORE EQUATIONS 10–14
3'
C
4'
C0 3
4 2'
1'
v̄Y 1
v̄ X 1
Y, y
1 2
X, x
(1 + 12 g1 )2 1
g (1 + 12 g1 )
2 2
1
g (1 + 12 g1 )
2 3 1 0 0
SU (1, 1) = E ccT + s 0 H = E 1 2
g
4 2
1
g g
4 2 3
+ s0 0 1 0 , (10.36)
symm 1 2
g 0 0 1
4 3
(1 + 12 g1 )2 1
g (1 + 12 g1 )
2 2
1
g (1 + 12 g1 )
2 3 1 0 0
Sr ( 12 , 1) = E ccT + s m H = E 1 2
g
4 2
1
g g
4 2 3
+ sm 0 1 0 , (10.37)
symm 1 2
g 0 0 1
4 3
(1 + g1 )2 g2 (1 + g1 ) g3 (1 + g1 ) 1 0 0
S = E bbT + sH = E g22 g2 g3 + s 0 1 0 . (10.38)
symm g32 0 0 1
10–14
10–15 §10.5 CORE STIFFNESS DERIVATION EXAMPLES
S = E i j bi biT + si Hi = S M + SG , (10.44)
where si = si0 + E i j e j , (i, j = 1, 2, 3), are the PK2 stresses in the current configuration.
In full and using the abbreviations a1 = 1 + g1 , a4 = 1 + g4 we get
E a 2 + 2E a g + E g 2 E 11 a1 g2 + E 13 (a1 a4 + g2 g3 ) + E 33 a4 g3
11 1 13 1 3 33 3
E 11 g22 + 2E 13 a4 g2 + E 33 a42
SM =
symm
(10.45)
E 12 a1 g3 + E 13 a12 + E 23 g32 + E 33 a1 g3 E 12 a1 a4 + E 13 a1 g2 + E 23 a4 g3 + E 33 g2 g3
E 12 g2 g3 + E 13 a1 g2 + E 23 a4 g3 + E 33 a1 a4 E 12 a4 g2 + E 13 g22 + E 23 a42 + E 33 a4 g2
E 22 g32 + 2E 23 a1 g3 + E 33 a12 E 22 a4 g3 + E 23 (a1 a4 + g2 g3 ) + E 33 a1 g2
E 22 a42 + 2E 23 a4 g2 + E 33 g22
10–15
Chapter 10: THE CORE-CONGRUENTIAL FORMULATION: CORE EQUATIONS 10–16
s1 0 s3 0
s1 0 s3
SG = . (10.46)
s2 0
symm s2
10–16
10–17 §10.5 CORE STIFFNESS DERIVATION EXAMPLES
h1: matrix([1],[0],[0],[0],[0],[0])$
h2: matrix([0],[0],[0],[0],[1],[0])$
h3: matrix([0],[1],[0],[1],[0],[0])$
g: matrix([g1],[g2],[g3],[g4],[g5],[g6])$
HH1:matrix([1,0,0,0,0,0],[0,1,0,0,0,0],[0,0,1,0,0,0],
[0,0,0,0,0,0],[0,0,0,0,0,0],[0,0,0,0,0,0])$
HH2:matrix([0,0,0,0,0,0],[0,0,0,0,0,0],[0,0,0,0,0,0],
[0,0,0,1,0,0],[0,0,0,0,1,0],[0,0,0,0,0,1])$
HH3:matrix([0,0,0,1,0,0],[0,0,0,0,1,0],[0,0,0,0,0,1],
[1,0,0,0,0,0],[0,1,0,0,0,0],[0,0,1,0,0,0])$
b1:h1+HH1.g$ b2:h2+HH2.g$ b3:h3+HH3.g$
SM:E11* b1.transpose(b1)+E22*b2.transpose(b2)+E33*b3.transpose(b3)
+ E12*(b1.transpose(b2)+b2.transpose(b1))
+ E13*(b1.transpose(b3)+b3.transpose(b1))
+ E23*(b2.transpose(b3)+b3.transpose(b2))$
ratvars(g6,g5,g4,g3,g2,g1,a5,a1,E11,E12,E13,E22,E23,E33)$
SM:ratsimp(SM)$
SG:ratsimp(s1*HH1+s2*HH2+s3*HH3)$
These matrices may be automatically converted to TEX by appropriate Macsyma statements (not
shown above). That output was reformatted by hand for inclusion here. For the core tangent
stiffness this semi-automated process yields
10–17
Chapter 10: THE CORE-CONGRUENTIAL FORMULATION: CORE EQUATIONS 10–18
Remark 10.6. If the plate element to which the particle belong has (as usual) rotational freedoms, an additional
geometric stiffness (the complementary geometric stiffness) appears in the transformation phase. Because of
this, the core geometric stiffness (10.52) has been relabeled as SG P , where subscript P means “principal.”
Remark 10.7. The core stiffness matrices may also be used for part of the formulation of thin-shell facet
elements, with the proviso that global reference axes {X, Y, Z } are to be replaced by a local coordinate system
{ X̄ , Ȳ , Z̄ } with Z̄ normal to the element midplane.
10–18
10–19 §10.5 CORE STIFFNESS DERIVATION EXAMPLES
θ2
θ
θ1 T
Rζ
x u u0
X ζ
1 2 X
X0
10–19
Chapter 10: THE CORE-CONGRUENTIAL FORMULATION: CORE EQUATIONS 10–20
10–20
10–21 §10.5 CORE STIFFNESS DERIVATION EXAMPLES
a2
a1
a3
n2 x0
u0
X, n1
X0
Z n3
where x0 describes the position of the centroid of the given cross-section, R is a 3-by-3 orthogonal
matrix function that orients the displaced cross section, and ζ is a cross-section position vector. The
displacement field is
u = x − X = u0 + (RT − I)ζ. (10.66)
where u0 (X ) = x0 (X ) − X0 (X ) is the centroidal displacement (see Figure 2).
10–21
Chapter 10: THE CORE-CONGRUENTIAL FORMULATION: CORE EQUATIONS 10–22
In the sequel 3 × 3 skew-symmetric matrices are consistently denoted by placing a tilde over their
axial 3-vector symbol; for example
0 a3 −a2 a1
ã = spin (a) = −a3 0 a1 , a = a2 = axial (ã). (10.67)
a2 −a1 0 a3
The skew-symmetric curvature matrix κ̃ is defined by κ̃ = R(dRT /d X ), which is the rate of change
of the orthogonal rotation matrix R with respect to the longitudinal coordinate. The curvature vector
is κ = axial (κ̃). We shall also require later the variation of angular orientation δΘ, defined as
the axial vector of the skew matrix R δRT :
= R δRT = −δR RT ,
δΘ
δΘ = axial (δΘ), (10.68)
All displacement gradients gi j appear in the GL strain measures. To maintain compactness the nine
gradients are partitioned into three 3-vectors:
∂u X /∂ X ∂u X /∂Y ∂u X /∂ Z
g1 = ∂u Y /∂ X , g2 = ∂u Y /∂Y , g3 = ∂u Y /∂ Z , (10.69)
∂u Z /∂ X ∂u Z /∂Y ∂u Z /∂ Z
The 9-component gradient vector is gT = [g1T g2T g3T ], but this symbol is not used directly here.
Also introduce the 3-vectors
1 0 0
h1 = 0 , h2 = 1 , h3 = 0 . (10.70)
0 0 1
With the help of these quantities, explicit expressions for the displacement gradient vectors g can
be given as
du0 du0 T
g1 = + RT κ̃ζ = + RT ζ̃ κ,
dX dX (10.71)
g2 = (R − I)h2 ,
T
g3 = (R − I)h3 .
T
where H is here the 3 × 3 identity matrix. Note that from the orthogonality of the rotation matrix
R we find
e22 = h2T g2 + 12 g2T g2
2
= R22 − 1 + 12 R21 + (R22 − 1)2 + R23
2
= R22 − 1 + 12 (2 − 2R22 ) = 0,
(10.73)
2e23 = h2T g3 + h3T g2 + g2T g3
= R32 + R23 + R21 R31 + R22 R32 − R32 + R23 R33 − R23 = 0,
10–22
10–23 §10.5 CORE STIFFNESS DERIVATION EXAMPLES
and similarly e33 = 0. This confirms that the only nonzero strains are (10.72).
The strains (10.72) may be rewritten in a more physically suggestive form:
Here eb , e f are stretching and flexural normal strains, γ2 and γ3 represent bending-induced shear
strains, and γ 2 , γ 3 are torsion-induced shear strains. The last term in e f represents a squared-
curvature contribution to flexure, which can usually be neglected (cf. Remark 9.2). The strain
energy stored in the current configuration is
U= U d A d X, with U = 12 Ee12 + 12 G e22 + e32 + s10 e1 + s20 e12 + s30 e3 . (10.75)
L 0 A0
where SU1 = c1 c1T , SU2 = c2 c2T , SU3 = c3 c3T , SU4 = c2 c1T and SU5 = c3 c1T . At the residual level
we obtain for Sr a form similar to (10.76) except that the prestresses si0 , i = 1, 2, 3 have to
be replaced by the midpoint stresses 12 (si0 + si ). The internal force vector conjugate to δg is
Φ = Sr g + Φ0 = Φσ + Φτ , in which
s 1 b1 s2 b2 + s3 b3
Φσ = 0 Φτ = s2 b1 , (10.77)
0 s3 b1
10–23
Chapter 10: THE CORE-CONGRUENTIAL FORMULATION: CORE EQUATIONS 10–24
where S1 = b1 b1T , S2 = b2 b2T , S3 = b3 b3T , S4 = b2 b1T and S5 = b3 b1T . The principal geometric
stiffness is
0
s1 H s2 H s3 H (s1 + Ee1 )H (s20 + Ge2 )H (s20 + Ge3 )H
S G P = s2 H 0 0 = (s20 + Ge2 )H 0 0 . (10.79)
s3 H 0 0 (s2 + Ge3 )H
0
0 0
T
The contribution of δ 2 g Φ to the complementary geometric stiffness depends on the target
variables in the ensuing transformation phase. Because this transformation requires the DGCCF, it
is taken up in Chapter 11.
§10.6. References
[1] C. A. Felippa, L. A. Crivelli and B. Haugen, A Survey of the Core-Congruential Formulation for Geo-
metrically Nonlinear TL Finite Elements, Archives of Computational Methods in Engineering, 1, 1994,
pp. 1–48
[2] S. Rajasekaran and D. W. Murray, Incremental finite element matrices, J. Str. Div. ASCE, 99, pp. 2423–
2438, 1973.
[3] R. H. Mallet and P. V. Marcal, Finite element analysis of nonlinear structures, J. Str. Div. ASCE, 94, pp.
2081–2105, 1968.
[4] D. W. Murray, Finite element nonlinear analysis of plates, Ph. D. Dissertation, Dept. of Civil Engineering,
University of California, Berkeley, California, 1967.
[5] C. A. Felippa, Discussion of Reference 1, J. Str. Div. ASCE, 100, pp. 2519–2521, 1974.
[6] L. A. Crivelli, A Total-Lagrangian beam element for analysis of nonlinear space structures, Ph. D.
Dissertation, Dept. of Aerospace Engineering Sciences, University of Colorado, Boulder, CO, 1990.
[7] C. A. Felippa and L. A. Crivelli, A congruential formulation of nonlinear finite elements, in Nonlinear
Computational Mechanics - The State of the Art, ed. by P. Wriggers and W. Wagner, Springer-Verlag,
Berlin, pp. 283–302, 1991.
[8] L. A. Crivelli and C. A. Felippa, A three-dimensional non-linear Timoshenko beam element based on
the core-congruential formulation, Int. J. Numer. Meth. Engrg., 36, pp. 3647–3673, 1993.
[9] K. Mathiasson, A. Bengtsson and A. Samuelsson, On the accuracy and efficiency of numerical algorithms
for geometrically nonlinear structural analysis, in Finite Element Methods for Nonlinear Problems, ed.
by P. G. Bergan, K. J. Bathe and W. Wunderlich, Springer-Verlag, Berlin, pp. 3–24, 1986.
[10] P. G. Bergan and K. M. Mathisen, Large displacement analysis of highly flexible offshore structures,
in Nonlinear Computational Mechanics - The State of the Art, ed. by P. Wriggers and W. Wagner,
Springer-Verlag, Berlin, pp. 303–331.
[11] S. Alexander, H. M. de la Fuente and B. Haugen, Correspondence between CC-TL and C-TL formulations
A 2D Timoshenko beam element using the Total-Lagrangian Core Congruential Formulation, in Term
Projects in Nonlinear Finite Element Methods, ed. by C. A. Felippa, Report CU-CSSC-91-12, Center
for Space Structures and Controls, University of Colorado, Boulder, CO, May 1991.
10–24
10–25 Exercises
EXERCISE 10.1
Show that the “core” forms gT SU g and Sr g, with SU and Sr given by (10.19), are independent of the coefficients
α, β, φ and ψ. It is sufficient to work this exercise for for i, j = 1, that is, only one stress, strain and modulus
is considered. Keep, however, g, h and H generic.
Note: This is a good exercise in matrix gymnastics. If you are “rusty” in matrix magic, reading the Addendum
below is strongly recommended.
In carrying out the manipulations required by Exercise 10.1, the following properties of matrix products ought
to be kept in mind.
(1) Let x and y be two conforming vectors. Since their inner product is a scalar, obviously
xT y = yT x.
(2) If x and y are two conforming vectors and A a conforming matrix, the quadratic form xT Ay is also a
scalar; consequently
xT Ay = yT AT x.
Furthermore, if A is square and symmetric, xT Ay = yT Ax.
(3) If x and y are conforming vectors, xyT and yxT are rank-one square matrices (the transpose of each
other). Furthermore, xxT is a symmetric matrix.
(4) Scalars can be moved to any position within a matrix product. If the scalar is in itself the result of a
vector or matrix product, the components may be transposed as per rules (1) and (2). For example, if A,
B and C are conforming matrices, and x and y are conforming vectors,
10–25
9
.
The TL
Timoshenko
Plane Beam
Element
9–1
Chapter 9: THE TL TIMOSHENKO PLANE BEAM ELEMENT 9–2
TABLE OF CONTENTS
Page
§9.1. Introduction 9–3
§9.2. Beam Models 9–3
§9.2.1. Basic Concepts and Terminology . . . . . . . . . . . 9–3
§9.2.2. Mathematical Models: Classical and Timoshenko . . . . . 9–4
§9.2.3. Finite Element Models . . . . . . . . . . . . . . 9–5
§9.2.4. Bernoulli-Euler versus Timoshenko Beam Elements . . . . 9–7
§9.3. X -Aligned Reference Configuration 9–9
§9.3.1. Element Description . . . . . . . . . . . . . . . 9–9
§9.3.2. Motion . . . . . . . . . . . . . . . . . . . 9–9
§9.3.3. Displacement Interpolation . . . . . . . . . . . . . 9–12
§9.3.4. Strain-Displacement Relations . . . . . . . . . . . 9–12
§9.3.5. *Consistent Linearization . . . . . . . . . . . . . 9–13
§9.4. Arbitrary Reference Configuration 9–14
§9.4.1. Strain-Displacement Matrix . . . . . . . . . . . . 9–14
§9.4.2. Constitutive Equations . . . . . . . . . . . . . . 9–15
§9.4.3. Strain Energy . . . . . . . . . . . . . . . . . 9–16
§9.5. The Internal Force 9–16
§9.6. The Stiffness Matrix 9–17
§9.6.1. The Material Stiffness Matrix . . . . . . . . . . . . 9–17
§9.6.2. Eliminating Shear Locking by RBF . . . . . . . . . 9–18
§9.6.3. The Geometric Stiffness Matrix . . . . . . . . . . . 9–19
§9.7. A Commentary on the Element Performance 9–21
§9.8. Summary 9–22
§9. Exercises . . . . . . . . . . . . . . . . . . . . . . 9–23
9–2
9–3 §9.2 BEAM MODELS
§9.1. Introduction
In the present Chapter the Standard Formulation of Total Lagrangian (TL) kinematics is used to
derive the finite element equations of a two-node Timoshenko plane beam element. This derivation
is more typical of the general case. It is still short, however, of the enormous complexity involved,
for instance, in the FEM analysis of nonlinear three-dimensional beams or shells. In fact the latter
are still doctoral thesis topics.
In the formulation of the bar element in Chapter 8, advantage was taken of the direct expression of
the axial strain in terms of reference and current element lengths. That shortcut bypasses the use
of displacement gradients and coordinate transformations. The simplification works equally well
for bars in three-dimensional space.
A more systematic but lengthier procedure is unavoidable with more complicated elements. The
procedure requires going through the displacement gradients to construct a strain measure. Some-
times this measure is too complex and must be simplified while retaining physical correctness.
Then the stresses are introduced and paired with strains to form the strain energy function of the
element. Repeated differentiations with respect to node displacements yield the expressions of the
internal force vector and tangent stiffness matrix. Finally, a transformation to the global coordinate
system may be required.
In addition to giving a better picture of the general procedure, the beam element offers an illustration
of the treatment of rotational degrees of freedom.
Beams represent the most common structural component found in civil and mechanical structures.
Because of their ubiquity they are extensively studied, from an analytical viewpoint, in Mechanics
of Materials courses. Such a basic knowledge is assumed here. The following material recapitulates
definitions and concepts that are needed in the finite element formulation.
A beam is a rod-like structural member that can resist transverse loading applied between its
supports. By “rod-like” it is meant that one of the dimensions is considerably larger than the other
two. This dimension is called the longitudinal dimension and defines the longitudinal direction
or axial direction. Directions normal to the longitudinal directions are called transverse. The
intersection of planes normal to the longitudinal direction with the beam are called cross sections,
just as for bar elements. The beam longitudinal axis is directed along the longitudinal direction and
passes through the centroid of the cross sections.1 .
Beams may be used as isolated structures. But they can also be combined to form framework
structures. This is actually the most common form of high-rise building construction. Individual
beam components of a framework are called members, which are connected at joints. Frameworks
1 If the beam is built of several materials, as in the case of reinforced concrete, the longitudinal axis passes through the
centroid of a modified cross section. The modified-area technique is explained in elementary courses of Mechanics of
Materials
9–3
Chapter 9: THE TL TIMOSHENKO PLANE BEAM ELEMENT 9–4
reference configuration
motion
current configuration
can be distinguished from trusses by the fact that their joints are sufficiently rigid to transmit bending
moments between members.
In practical structures beam members can take up a great variety of loads, including biaxial bending,
transverse shears, axial forces and even torsion. Such complicated actions are typical of spatial
beams, which are used in three-dimensional frameworks and are subject to forces applied along
arbitrary directions.
A plane beam resists primarily loading applied in one plane and has a cross section that is symmetric
with respect to that plane. Plane frameworks, such as the one illustrated in Figure 9.1, are assemblies
of plane beams that share that symmetry. Those structures can be analyzed with two-dimensional
idealizations.
A beam is straight if the longitudinal direction is a straight line. A beam is prismatic if the cross
section is uniform. Only straight, prismatic, plane beams will be considered in this Chapter.
9–4
9–5 §9.2 BEAM MODELS
θ Z (X ) ≡ θ(X )
Y, y
motion u Y (X )
X, x
u X (X )
Figure 9.2. Definition of beam kinematics in terms of the three displacement functions
u X (X ), u Y (X ) and θ (X ). The figure actually depicts the EB model kinematics.
In the Timoshenko model, θ (X ) is not constrained by normality (see next figure).
rotation occurs about a neutral axis that passes through the centroid of the cross section.
Timoshenko Model. This model corrects the classical beam theory with first-order shear deformation
effects. In this theory cross sections remain plane and rotate about the same neutral axis as the EB
model, but do not remain normal to the deformed longitudinal axis. The deviation from normality
is produced by a transverse shear that is assumed to be constant over the cross section.
Both the EB and Timoshenko models rest on the assumptions of small deformations and linear-
elastic isotropic material behavior. In addition both models neglect changes in dimensions of the
cross sections as the beam deforms. Either theory can account for geometrically nonlinear behavior
due to large displacements and rotations as long as the other assumptions hold.
To carry out the geometrically nonlinear finite element analysis of a framework structure, beam
members are idealized as the assembly of one or more finite elements, as illustrated in Figure 9.3.
The most common elements used in practice have two end nodes. The i th node has three degrees of
freedom: two node displacements u Xi and u Y i , and one nodal rotation θi , positive counterclockwise
in radians, about the Z axis. See Figure 9.4.
The cross section rotation from the reference to the current configuration is called θ in both models.
In the BE model this is the same as the rotation ψ of the longitudinal axis. In the Timoshenko
9–5
Chapter 9: THE TL TIMOSHENKO PLANE BEAM ELEMENT 9–6
motion
θ1 θ1
Y, y uY 2 uY 2
uY 1 uY 1
X, x
uX 1 uX 2 uX 1 uX 2
1 2 1 2
Figure 9.4. Two-node beam elements have six DOFs, regardless of the model used.
model, the difference γ̄ = ψ − θ is used as measure of mean shear distortion.2 These angles are
illustrated in Figure 9.5.
Either the EB or the Timoshenko model may be used as the basis for the element formulation.
Superficially it appears that one should select the latter only when shear effects are to be considered,
as in “deep beams” whereas the EB model is used for ordinary beams. But here a “twist” appear
because of finite element considerations. This twist is one that has caused significant confusion
among FEM users over the past 25 years.
9–6
9–7 §9.2 BEAM MODELS
normal to reference
θ beam axis X
ψ
−θ
normal to deformed
_ =ψ beam axis
γ
direction of deformed ψ
cross section
90◦ _
// X (X = X)
ds
_
Note: in practice γ << θ; typically 0.1% or less. The magnitude
_
of γ is grossly exaggerated in the figure for visualization convenience.
Figure 9.5. Definition of total section rotation θ and BE section rotation ψ in the
Timoshenko beam model. The mean shear deformation is γ̄ = θ − ψ, which
is constant over the cross section. For small deformations of typical engineering
materials γ̄ << 1; for example, typical values for |γ̄ | are O(10−4 ) radians
whereas rotations ψ and θ may be of the order of several radians.
Although the Timoshenko beam model appears to be more complex because of the inclusion of
shear deformation, finite elements based on this model are in fact simpler to construct! Here are
the two main reasons:
(i) Separate kinematic assumptions on the variation of cross-section rotations are possible, as
made evident by Figure 9.5. Mathematically: θ(X ) may be assumed independenltly of u X (X )
and u Y (X ). As a consequence, two-node Timoshenko elements may use linear variations in
both displacement and rotations. On the other hand a two-node EB model requires a cubic
polynomial for u Y (X ) because the rotation θ(X ) is not independent.
(ii) The linear transverse displacement variation matches that commonly assumed for the axial
deformation (bar-like behavior). The transverse and axial displacements are then said to be
consistent.
The simplicity is even more important in geometrically nonlinear analysis, as strikingly illustrated
in the two-node elements depicted in Figure 9.6. Although as shown in that figure both of these
elements have six degrees of freedom, the internal kinematics of the Timoshenko model is far
simpler.
9–7
Chapter 9: THE TL TIMOSHENKO PLANE BEAM ELEMENT 9–8
(a) (b)
C 1 element
with same DOFs
Figure 9.6. Sketch of the kinematics of two-node beam finite element models based on
(a) Euler-Bernoulli beam theory, and (b) Timoshenko beam theory. These
models are called C 1 and C 0 beams, respectively, in the FEM literature.
What would be the first reaction of an experienced but old-fashioned (i.e, “never heard about
FEM”) structural engineer on looking at Figure 9.6? The engineer would pronounce the C 0 element
unsuitable for practical use. And indeed the kinematics looks strange. The shear distortion implied
by the drawing appears to grossly violate the basic assumptions of beam behavior. Furthermore, a
huge amount of shear energy would be require to keep the element straight as depicted.
The engineer would be both right and wrong. If the two-node element of Figure 9.6(b) were
constructed with actual shear properties and exact integration, an overstiff model results. This
phenomenom is well known in the FEM literature and receives the name of shear locking. To avoid
locking while retaining the element simplicity it is necessary to use certain computational devices.
The most common are:
1. Selective integration for the shear energy.
2. Residual energy balancing.
These devices will be used without explanation in some of the derivations of this Chapter. For
detailed justification the curious reader may consult advanced FEM books such as Hughes’.3
In this Chapter the C 0 model will be used to illustrated the TL formulation of a two-node, geomet-
rically nonlinear beam element.
Remark 9.1. As a result of the application of the aforementioned devices the beam element behaves like a BE
beam although the underlying model is Timoshenko’s. This represents a curious paradox: shear deformation
9–8
9–9 §9.3 X -ALIGNED REFERENCE CONFIGURATION
is used to simplify the kinematics, but then most of the shear is removed to restore the correct stiffness.4 As a
result, the name “C 0 element” is more appropriate than “Timoshenko element” because capturing the actual
shear deformation is not the main objective.
Remark 9.2. The two-node C 1 beam element is used primarily in linear structural mechanics. (It is in fact
the beam model used in the “Introduction to FEM” course.) This is because some of the easier-construction
advantages cited for the C 0 element are less noticeable, while no artificial devices to eliminate locking are
needed. The C 1 element is also called the Hermitian beam element because the shape functions are cubic
polynomials specified by Hermite interpolation formulas.
We consider a two-node, straight, prismatic C 0 plane beam element moving in the (X, Y ) plane, as
depicted in Figure 9.7(a). For simplicity in the following derivation the X axis system is initially
aligned with the longitudinal direction in the reference configuration, with origin at node 1. This
assumption is relaxed in the following section, once invariant strain measured are obtained.
The reference element length is L 0 . The cross section area A0 and second moment of inertia I0
with respect to the neutral axis5 are defined by the area integrals
A0 = d A, Y d A = 0, I0 = Y 2 d A, (9.1)
A0 A0 A0
In the current configuration those quantities become A, I and L, respectively, but only L is frequently
used in the TL formulation. The material remains linearly elastic with elastic modulus E relating
the stress and strain measures defined below.
As in the previous Chapter the identification superscript (e) will be omitted to reduce clutter until
it is necessary to distinguish elements within structural assemblies.
The element has the six degrees of freedom depicted in Figure 9.4. These degrees of freedom and
the associated node forces are collected in the node displacement and node force vectors
u X1 f X1
uY 1 fY 1
θ1 fθ 1
u= , f= . (9.2)
u X2 f X2
uY 2 fY 2
θ2 fθ 2
4 The FEM analysis of plates and shells is also rife with such paradoxes.
5 For a plane prismatic beam, the neutral axis at a particular section is the intersection of the cross section plane X =
constant with the plane Y = 0.
9–9
Chapter 9: THE TL TIMOSHENKO PLANE BEAM ELEMENT 9–10
(a) (b)
L
_
Y, y θ=ψ−γ
θ(X)
2
2
x P(x,y) ψ
Figure 9.7. Lagrangian kinematics of the C 0 beam element with X -aligned reference
configuration: (a) plane beam moving as a two-dimensional body; (b) reduction
of motion description to one dimension measured by coordinate X .
§9.3.2. Motion
The kinematic assumptions of the Timoshenko model element have been outlined in §9.2.2. Ba-
sically they state that cross sections remain plane upon deformation but not necessarily normal to
the deformed longitudinal axis. In addition, changes in cross section geometry are neglected.
To analyze the Lagrangian kinematics of the element shown in Figure 9.6(a), we study the motion of
a particle originally located at P0 (X, Y ). The particle moves to P(x, y) in the current configuration.
The projections of P0 and P along the cross sections at C0 and C upon the neutral axis are called
C0 (X, 0) and C(xC , yC ), respectively. We shall assume that the beam cross section dimensions do
not change, and that the shear distortion γ << 1 so that cos γ can be replaced by 1. Then
.
x = xC − Y (sin ψ + sin γ cos ψ) = xC − Y [sin(ψ + γ ) + (1 − cos γ ) sin ψ] = xC − Y sin θ,
.
y = yC + Y (cos ψ − sin γ sin ψ) = yC + Y [cos(ψ + γ ) + (1 − cos γ ) cos ψ] = yC + Y cos θ.
(9.3)
But xC = X + u XC and yC = u XC . Consequently x = X + u XC − Y sin θ and y = u Y C + Y cos θ .
From now we shall call u XC and u Y C simply u X and u Y , respectively, so that the Lagrangian
representation of the motion is
x X + u X − Y sin θ
= (9.4)
y u Y + Y cos θ
9–10
9–11 §9.3 X -ALIGNED REFERENCE CONFIGURATION
in which u X , u Y and θ are functions of X only. This concludes the reduction to a one-dimensional
model, as sketched in Figure 9.7(b).
For future use below it is convenient to define an “extended” internal displacement vector w, and
its gradient or X derivative:
u X (X ) dw du X /d X uX
w = u Y (X ) , w = = du Y /d X = u Y , (9.5)
θ(X ) d X dθ/d X θ
in which primes denote derivatives with respect to X . The derivative θ is also as κ, which has the
meaning of beam curvature in the current configuration. Also useful are the following differential
relations
ds
1 + u X = s cos ψ, u Y = s sin ψ, s = = (1 + u X )2 + (u Y )2 , (9.6)
dX
in which ds is the differential arclength in the current configuration; see Figure 9.5. If u X and u Y
are constant over the element,
Remark 9.3. The replacement of 1 − cos γ by zero in (9.3) is equivalent to saying that Y (1 − cos γ ) can be
neglected in comparison to other cross section dimensions. This is consistent with the uncertainty in the cross
section changes, which would depend on the normal stress and Poisson’s ratio effects. The motion expression
(9.4) has the virtue of being purely kinematic and of leading to (exactly) eY Y = 0.
9–11
Chapter 9: THE TL TIMOSHENKO PLANE BEAM ELEMENT 9–12
∂ X ∂Y 1 + u X − Y κ cos θ − sin θ
F= = (9.10)
∂y ∂y u Y − Y κ sin θ cos θ
∂ X ∂Y
where primes denote derivatives with respect to X , and κ = θ . The displacement gradient matrix
is6
u X − Y κ cos θ − sin θ
GF = F − I = , (9.11)
u Y − Y κ sin θ cos θ − 1
from which the Green-Lagrange (GL) strain tensor follows:
eX X eX Y
e= = 12 (FT F − I) = 12 (G F + GTF ) + 12 GTF G F
eY X eY Y
6 This is denoted by G F to avoid clash with the shape function gradient matrix G = N introduced in (9.9).
9–12
9–13 §9.3 X -ALIGNED REFERENCE CONFIGURATION
below, it can be shown that under the small-strain assumptions made precise therein, the axial strain
e X X can be replaced by the simpler form
in which all quantities appear linearly except θ. The nonzero axial and shear strains will be arranged
in the strain vector
e1 eX X (1 + u X ) cos θ + u Y sin θ − Y θ − 1 e − Yκ
e= = = = . (9.14)
e2 2e X Y −(1 + u X ) sin θ + u Y cos θ γ
characterize axial strains, shear strains and curvatures, respectively. These are collected in the
following generalized strain vector:
hT = [ e γ κ] (9.16)
Because of the assumed linear variation in X of u X (X ), u Y (X ) and θ(X ), e and γ only depend on
θ whereas κ is constant over the element. Making use of the relations (9.7) one can express e and
γ in the geometrically invariant form
L cos γ̄ L sin γ̄
1 + e = s cos(θ − ψ) = , γ = −s sin(θ − ψ) = (9.17)
L0 L0
In theory one could further reduce e to L/L 0 and γ to L γ̄ /L 0 , but these “simplifications” actually
complicate the strain variations taken in the following Section.
§9.3.5. *Consistent Linearization
The derivation of the consistent linearization (9.14) is based on the following study, known in continuum
mechanics as a polar decomposition analysis of the deformation gradient. Introduce the matrix
cos α − sin α
Ω(α) = (9.18)
sin α cos α
which represents a two-dimensional rotation (about Z ) through an angle α. Since Ω is an orthogonal matrix,
ΩT = Ω−1 . The deformation gradient (9.9) can be written
s 0 −Y θ 0
F = Ω(ψ) + Ω(θ) . (9.19)
0 0 0 1
where s is defined in (9.6). Premultiplying both sides of (9.19) by Ω(−θ ) gives the modified deformation
gradient
s cos(θ − ψ) − Y θ 0
F̄ = Ω(−θ )F = (9.20)
−s sin(θ − ψ) 1
Now the GL strain tensor 2e = FT F − I does not change if F is premultiplied by an orthogonal matrix Ω
T
because FT ΩT ΩF = FT F. Consequently 2e = F̄ F̄ − I. But if the strains remain small, as it is assumed in
the Timoshenko model, the following are small quantities:
9–13
Chapter 9: THE TL TIMOSHENKO PLANE BEAM ELEMENT 9–14
9–14
9–15 §9.4 ARBITRARY REFERENCE CONFIGURATION
_
// γY
φ = ψ+ϕ _
θ2 ψ
// γX
ϕ
2(x2 ,y2) // γX
_ C
// γY
uY 2
θ1
_
1(x1 ,y1) Xγ
ϕ
2(X2, Y2 )
uY 1
Y, y _ uX 2
Yγ
C0
X, x
uX
1(X1, Y1 )
Figure 9.8. Plane beam element with arbitrarily oriented reference configuration.
in which ω = θ + ϕ, and where use is made of (9.24) in a key step. These derivatives were checked
with Mathematica. Collecting all of them into matrix B:
1 − cos ω − sin ω L 0 N1 γ cos ω sin ω L 0 N2 γ
B= sin ω − cos ω −L 0 N1 (1 + e) − sin ω cos ω −L 0 N2 (1 + e) . (9.26)
L0 0 0 −1 0 0 1
Here N1 = (1 − ξ )/2 and N2 = (1 + ξ )/2 are abbreviations for the element shape functions
(caligraphic symbols are used to lessen the chance of clash against axial force symbols).
Because the beam material is assumed to be homogeneous and isotropic, the only nonzero PK2
stresses are the axial stress s X X and the shear stress s X Y . These are collected in a stress vector s
related to the GL strains by the linear elastic relations
0
0
sX X s1 s1 + Ee1 s1 E 0 e1
s= = = 0 = 0 + = s0 + Ee, (9.27)
sX Y s2 s2 + Ge2 s2 0 G e2
where E is the modulus of elasticity and G is the shear modulus. We introduce the prestress
resultants
N =
0 0
s1 d A, V =
0 0
s2 d A, M =
0
−Y s10 d A, (9.28)
A0 A0 A0
9–15
Chapter 9: THE TL TIMOSHENKO PLANE BEAM ELEMENT 9–16
N
M
V
M0
N0
0
V
C0
which define the axial forces, transverse shear forces and bending moments, respectively, in the
reference configuration. We also define the stress resultants
N = N 0 + E A0 e, V = V 0 + G A0 γ , M = M 0 + E I0 κ. (9.29)
These represent axial forces, tranverse shear forces and bending moments in the current configura-
tion, respectively, defined in terms of PK2 stresses. See Figure 9.9 for signs. These are collected
in the stress-resultant vector
zT = [ N V M ] . (9.30)
Carrying out the area integrals while making use of (9.27) through (9.30), U can be written as the
sum of three length integrals:
U= (N e + 2 E A0 e ) d X̄ +
0 1 2
(V γ + 2 G A0 γ ) d X̄ +
0 1 2
(M 0 κ + 12 E I0 κ 2 ) d X̄ ,
L0 L0 L0
(9.32)
The three terms in (9.32) define the energy stored through bar-like axial deformations, shear dis-
tortion and pure bending, respectively.
9–16
9–17 §9.6 THE STIFFNESS MATRIX
Here h and B are defined in equations (9.16) and (9.26) whereas z collects the stress resultants in
C as defined in (9.28) through (9.30). Because δU = p T δu, we get
p= BT z d X̄ . (9.34)
L0
This expression may be evaluated by a one point Gauss integration rule with the sample point at
ξ = 0 (beam midpoint). Let θm = (θ1 + θ2 )/2, ωm = θm + ϕ, cm = cos ωm , sm = sin ωm ,
em = L cos(θm − ψ)/L 0 − 1, γm = L sin(ψ − θm )/L 0 , and
−cm −sm − 12 L 0 γm cm sm − 12 L 0 γm
1
Bm = B|ξ =0 = sm −cm 21 L 0 (1 + em ) sm −cm 12 L 0 (1 + em ) (9.35)
L0
0 0 −1 0 0 1
where subscript m stands for “beam midpoint.” Then
T
−cm −sm 1
L γ
2 0 m
cm sm 1
L γ
2 0 m N
p = L 0 BmT z = sm −cm − 2 L 0 (1 + em ) −sm
1
cm − 12 L 0 (1 + em ) V (9.36)
0 0 −1 0 0 1 M
This is again the sum of the material stifness K M and the geometric stiffness KG .
9–17
Chapter 9: THE TL TIMOSHENKO PLANE BEAM ELEMENT 9–18
This integral is evaluated by the one-point Gauss rule at ξ = 0. Denoting again by Bm the matrix
(9.35), we find
KM = BmT SBm d X̄ = KaM + KbM + KsM (9.40)
L0
where KaM , KbM and KsM are due to axial (bar), bending, and shear stiffness, respectively:
cm2 cm sm −cm γm L 0 /2 −cm2 −cm sm −cm γm L 0 /2
cm sm 2
sm −γm L 0 sm /2 −cm sm −sm2 −γm L 0 sm /2
E A0 −c
m γm L 0 /2 −γm L 0 sm /2 γm2 L 20 /4 cm γm L 0 /2 γm L 0 sm /2
γm2 L 20 /4
KaM =
L0 −cm2 −cm sm cm γm L 0 /2 cm2 cm sm cm γm L 0 /2
−cm sm −sm2
γm L 0 sm /2 cm sm sm2 γm L 0 sm /2
−cm γm L 0 /2 −γm L 0 sm /2 γm2 L 20 /4 cm γm L 0 /2 γm L 0 sm /2 γm2 L 20 /4
(9.41)
0 0 0 0 0 0
0 0 0 0 0 0
E I 0 0 0 1 0 0 −1
KbM = (9.42)
L0 0 0 0 0 0 0
0 0 0 0 0 0
0 0 −1 0 0 1
sm2 −cm sm −a1 L 0 sm /2 −sm2 c m sm −a1 L 0 sm /2
−cm sm cm2
cm a1 L 0 /2 cm sm −cm2 cm a1 L 0 /2
G A0
−a1 L 0 sm /2 cm a1 L 0 /2 a1 L 0 /4
2 2
a1 L 0 sm /2 −cm a1 L 0 /2 a12 L 20 /4
KsM =
L0 −sm2 cm sm a1 L 0 sm /2 sm2 −cm sm a1 L 0 sm /2
cm sm −cm2 −cm a1 L 0 /2 −cm sm cm2 −cm a1 L 0 /2
−a1 L 0 sm /2 cm a1 L 0 /2 a1 L 0 /4
2 2
a1 L 0 sm /2 −cm a1 L 0 /2 a12 L 20 /4
(9.43)
in which a1 = 1 + em .
9–18
9–19 §9.6 THE STIFFNESS MATRIX
much strain energy is taken by shear. The following substitution device, introduced by MacNeal,7
removes that deficiency in a simple way. The shear rigidity G A0 is formally replaced by 12E I0 /L 20 ,
and magically (9.44) becomes
EA
L
0 0 0 − ELA0 0 0
0 0
0 12E I0 6E I0 − 12E3 I0 6E I0
L 30 L 20
0
L 20
L0
6E I0 4E I0
0
L0 0 − 2I0
6E 2E I0
L0
L 20 L0
K̂ M =
E A0 E A0
.
(9.45)
− L 0 0 0 0
0 L0
0 − 12E3 I0 − 6E2I0 0 12E I0 6E
− 2 I 0
L0 L0 L 30 L0
0 6E I0 2E I0 0 − 6E2I0 4E I0
L 20 L0 L0 L0
This is the well known linear stiffness matrix of the C 1 (Hermitian) beam based on the Euler-
Bernoulli model. That substitution device is called the residual bending flexibility (RBF) correc-
tion.8 Its effect is to get rid of the spurious shear energy due to the linear kinematic assumptions. If
the RBF is formally applied to the nonlinear material stiffness one gets K̂ M = KaM + K̂ Mb , where
KaM is the same as in (9.43) (because the axial stiffness if not affected by the substitution), whereas
KbM and KsM merge into
12sm2 −12cm sm 6a1 L 0 sm −12sm2 12cm sm 6a1 L 0 sm
−12cm sm 12cm2 −6cm a1 L 0 12cm sm −12cm2 −6cm a1 L 0
b EI 6a L s −6cm a1 L 0 a2 L 20 −6a1 L 0 sm 6cm a1 L 0 a3 L 20
K̂ M = 3 1 0 2m
L 0 −12sm 12cm sm −6a1 L 0 sm 12sm2 −12cm sm −6a1 L 0 sm
12cm sm −12cm2 6cm a1 L 0 −12cm sm 12cm2 6cm a1 L 0
6a1 L 0 sm −6cm a1 L 0 a3 L 20 −6a1 L 0 sm 6cm a1 L 0 a2 L 20
(9.46)
in which a1 = 1 + em , a2 = 4 + 6em + 3em2 and a3 = 2 + 6em + 3em2 .
1 1 L 20
replace by + (9.47)
G A0 G As 12E I0
where G As is the actual shear rigidity; that is, As is the shear-reduced cross section studied in Mechanics
of Materials. The result of (9.47) is the C 1 Hermitian beam corrected by shear deformations computed
from equilibrium considerations.9 If the shear deformation is negligible, the right hand side of (9.47) is
approximately L 20 /(12E I0 ), which leads to the substitution used above.
7 R. H. MacNeal, A simple quadrilateral shell element, Computers and Structures, 8, 1978, pp. 175-183.
8 RBF can be rigurously justified through the use of a mixed variational principle, or through a flexibility calculation.
9 See, e.g., J. Przemieniecki, Theory of Matrix Structural Analysis, Dover, New York, 1968.
9–19
Chapter 9: THE TL TIMOSHENKO PLANE BEAM ELEMENT 9–20
The geometric stiffness KG comes from the variation of B while the stress resultants in z are kept
fixed. To get a closed form expression it is convenient to pass to indicial notation, reverting to
matrix notation later upon “index contraction.” Let the entries of KG , B, u and z be denoted as
K Gi j , Bki , u j and z k , where indices i, j and k range over 1–6, 1–6, and 1–3, respectively. Call
A j = ∂B/∂u j , j = 1, . . . 6. Then using the summation convention,
∂ Bki j
K Gi j δu j = δB z d X =
T
δu j z k d X = Aki z k d X δu j , (9.48)
L0 L0 ∂u j L0
whence
j
K Gi j = z k Aki d X̄ , (9.49)
L0
Note that in carrying out the derivatives in (9.49) by hand one must use the chain rule because B is
a function of e, γ and θ , which in turn are functions of the node displacements u j . To implement
this scheme we differentiate B with respect to each node displacement in turn, to obtain:
∂B 1 0 0 N1 sin ω 0 0 N2 sin ω
A1 = = 0 0 N1 cos ω 0 0 N2 cos ω ,
∂u X 1 L0 0 0 0 0 0 0
∂B 1 0 0 −N1 cos ω 0 0 −N2 cos ω
A2 = = 0 0 N1 sin ω 0 0 N2 sin ω ,
∂u Y 1 L0 0 0 0 0 0 0
∂B N1 sin ω − cos ω −N1 L 0 (1 + e) − sin ω cos ω −N2 L 0 (1 + e)
A3 = = cos ω sin ω −N1 L 0 γ − cos ω − sin ω −N2 L 0 γ ,
∂θ1 L0 0 0 0 0 0 0
(9.50)
∂B 1 0 0 −N1 sin ω 0 0 −N2 sin ω
A4 = = 0 0 −N1 cos ω 0 0 −N2 cos ω ,
∂u X 2 L0 0 0 0 0 0 0
∂B 1 0 0 N1 cos ω 0 0 N2 cos ω
A5 = = 0 0 −N1 sin ω 0 0 −N2 sin ω ,
∂u Y 2 L0 0 0 0 0 0 0
∂B N2 sin ω − cos ω −N1 L 0 (1 + e) − sin ω cos ω −N2 L 0 (1 + e)
A6 = = cos ω sin ω −N1 L 0 γ − cos ω − sin ω −N2 L 0 γ .
∂θ2 L0 0 0 0 0 0 0
as the entries of three 6 × 6 “weighting matrices” W N , WV and W M that isolate the effect of the
stress resultants z 1 = N , z 2 = V and z 3 = M. The first, second and third row of each A j becomes
9–20
9–21 §9.7 A COMMENTARY ON THE ELEMENT PERFORMANCE
Again the length integral should be done with the one-point Gauss rule at ξ = 0. Denoting again
quantities evaluated at ξ = 0 by an m subscript, one obtains the closed form
0 0 sm 0 0 sm
0 0 −cm 0 0 −cm
Nm sm −cm − 2 L 0 (1 + em ) −sm
1
cm − 2 L 0 (1 + em )
1
KG =
2 0 0 −sm 0 0 −sm
0 0 cm 0 0 cm
sm −cm − 2 L 0 (1 + em ) −sm
1
cm − 2 L 0 (1 + em )
1
(9.55)
0 0 cm 0 0 cm
0 0 sm 0 0 sm
Vm cm sm − 2 L 0 γm −cm −sm
1
− 2 L 0 γm
1
+ .
2 0 0 −cm 0 0 −cm
0 0 −sm 0 0 −sm
cm sm − 12 L 0 γm −cm −sm − 12 L 0 γm
in which Nm and Vm are N and V evaluated at the midpoint.
The material stiffness of the present element works fairly well once MacNeal’s RBF device is done.
On the other hand, simple buckling test problems, as in Exercise 9.3, show that the geometric
stiffness is not so good as that of the C 1 Hermitian beam element.10 Unfortunately a simple
10 In the sense that one must use more elements to get equivalent accuracy.
9–21
Chapter 9: THE TL TIMOSHENKO PLANE BEAM ELEMENT 9–22
Node Displacements u
Eq. (9.9)
Eq. (9.15)
Generalized strains h = [ e , γ, κ ]
T
Internal forces p
vary p: δp = ∫ L0
_
(B T δz + δBT z) dX δu = (KM + KG )
Eq. (9.40)
Figure 9.10. Main steps in the derivation of the C 0 plane beam element.
substitution device such as RBF cannot be used to improve KG , and the problem should be viewed
as open.
An intrinsic limitation of the present element is the restriction to small axial strains. This was
done to facilitate close form derivation. The restriction is adequate for many structural problems,
particularly in Aerospace (example: deployment). However, it means that the element cannot model
correctly problems like the snap-through and bifurcation of the arch example used in Chapter 8, in
which large axial strains prior to collapse necessarily occur.
§9.8. Summary
Figure 9.10 is a roadmap that summarizes the key steps in the derivation of the internal force and
tangent stiffness matrix for the C 0 plane beam element.
9–22
9–23 Exercises
EXERCISE 9.1 [A+N:20] Consider a plane prismatic beam of length L 0 , cross section area A0 and second
moment of inertia I0 . In the reference configuration the beam extends from node 1 at (X = Y = 0) to node
2 at (X = L 0 , Y = 0). The beam is under axial prestress force N 0
= 0 in the reference configuration, while
both V 0 and M0 vanish.
(a) Obtain the internal force p, the material stiffness matrix K M and the geometric stiffness matrix KG in
the reference configuration, for which u X 1 = u Y 1 = u X 2 = u Y 2 = θ1 = θ2 = 0.
(b) The beam rotates 90◦ rigidly to a current configuration for which u X 1 = u Y 1 = 0, u X 2 = −L 0 , u Y 2 = L 0 ,
θ1 = θ2 = 90◦ . Check that the stress resultants do not change (that is, N = N0 and V = M = 0 because
e = γ = κ = 0), and obtain K M and KG in that configuration.
EXERCISE 9.2
;; ;
[N:20] Analyze the pure bending of a cantilever discretized into an arbitrary number of elements.
;
reference configuration C0 Buckling sketch
P
L
Pcr = -λcr EI/L2
fixed end P
FEM discretization into
Ne equal length elements
EXERCISE 9.3
[N:25] This exercise and the next one pertain to the simple structure shown in Figure E9.1. It is a cantilever,
plane beam-column of length L, modulus E, area A and inertia I , loaded by an axial force P as shown. The
structure is discretized into Ne plane beam elements of equal length. The structures moves in the plane of the
figure. The objective is to compute the classical buckling load Pcr and compare with the known analytical
value.
In the classical buckling analysis, deformations prior to buckling are neglected. The structure stiffness K =
K M + KG is evaluated on the reference configuration, with KG evaluated from the internal force state. This
stability model is called linearized prebuckling or LPB, and is studied in detail in later Chapters. Under the
LPB assumptions K M is constant while KG linear in the applied forces. Critical point analysis leads to a linear
eigensystem called the buckling eigenproblem. The smallest eigenvalue characterizes the critical load, which
for the LPB model can be shown to produce bifurcation.
The classical buckling load (also called Euler load) for the configuration of Figure E9.1 is Pcr = λcr E I /L 2
with λcr = −π 2 /4 (negative because P has to compress the beam-column to achieve buckling). FEM results
for Ne = 1, 2, . . . can be obtained by running the code of Figure A9.7. This script assembles K taking in 0 with
E = I = A = L = 1 to simplify computations while leaving P = λ symbolic. The only nonzero internal
force is N0 = P. The determinant det(K) is formed explicitly as a polynomial in P, called the characteristic
polynomial. All of its roots are computed via NSolve. The root closest to zero defines the critical load.
9–23
Chapter 9: THE TL TIMOSHENKO PLANE BEAM ELEMENT 9–24
As answers to the exercise run the script until the largest Ne that can be reasonably handled by your Mathematica
version (this will depend on memory and CPU speed). Termination can be controlled by adjusting the loop
on k, which doubles Ne on each pass.
Report on the following items:
(a) Your computed λcr ’s for the Ne you were able to run. Comment on whether they converge toward −π 2 /4.
Is the convergence monotonic?
(b) Why does NSolve give 2Ne roots?11
(c) Why are the roots real?12
(d) Why do the computations get rapidly very slow as Ne increases?13
EXERCISE 9.4 [N:20] The brute-force technique used to find λcr in Exercise 9.3 is easy to implement but it
is extremely inefficient as the number of elements increase.14
A more effective tecnique is implemented in the script shown in Figure A9.8 of the Addendum. On Mathematica
this works fine on my Mac up to 128 elements, beyond which point memory and CPU time requirements grow
too large. As answer to this exercise:
(a) Describe what is going on in the script.
(b) Discuss why the solution is more efficient and robust than the previous script.15
9–24
9–25 Exercises
The Mathematica Notebook PlaneBeam.nb contains several modules and scripts that support Exercises 9.3
and 9.4. The Notebook file is posted on the course Web site. The important pieces of code are briefly described
in this Addendum.
Figure A9.1 shows module FormIntForceC0TwoNodePlaneBeam. This forms the internal force vector p of
a two-node, C 0 , plane beam formulated with the TL description. Although this module is not directly called
in Exercises 9.3 and 9.4, it is used later in the course. It is also used in the verifivation of the stiffness matrix
modules by finite differences.
FormIntForceC0TwoNodePlaneBeam[XY1_,XY2_,u1_,u2_,S0_,z0_]:=Module[
{X1,Y1,X21,Y21,L0,uX1,uY1,theta1,uX2,uY2,theta2,x21,y21,
thetam,ctheta,stheta,Lcpsi,Lspsi,N0,V0,M0,EA0,GA0,EI0,
cphi,sphi,cm,sm,Nm,Vm,Mm,kappa,p},
{X1,Y1}=XY1; {X2,Y2}=XY2; X21=X2-X1; Y21=Y2-Y1;
{uX1,uY1,theta1}=u1; {uX2,uY2,theta2}=u2;
x21=X21+uX2-uX1; y21=Y21+uY2-uY1;
L0=PowerExpand[Sqrt[X21^2+Y21^2]]; thetam=(theta1+theta2)/2;
ctheta=Cos[thetam]; stheta=Sin[thetam];
Lcpsi=Simplify[(X21*x21+Y21*y21)/L0];
Lspsi=Simplify[(X21*y21-Y21*x21)/L0];
em= (ctheta*Lcpsi+stheta*Lspsi)/L0-1;
gm=-(stheta*Lcpsi-ctheta*Lspsi)/L0;
kappa=(theta2-theta1)/L0;
{N0,V0,M0}=z0; {EA0,GA0,EI0}=S0;
Nm=Simplify[N0+EA0*em]; Vm=Simplify[V0+GA0*gm];
Mm=Simplify[M0+EI0*kappa];
cphi=X21/L0; sphi=Y21/L0;
cm=ctheta*cphi-stheta*sphi; sm=stheta*cphi+ctheta*sphi;
Bm= (1/L0)*{{-cm,-sm, L0*gm/2, cm,sm, L0*gm/2 },
{ sm,-cm,-L0*(1+em)/2,-sm,cm,-L0*(1+em)/2},
{ 0, 0, -1, 0, 0, 1 }};
p=L0*Transpose[Bm].{{Nm},{Vm},{Mm}};
Return[Simplify[p]] ];
p=FormIntForceC0TwoNodePlaneBeam[{0,0},{L/Sqrt[2],L/Sqrt[2]},
{0,0,Pi/2},{-2*L/Sqrt[2],0,Pi/2},
{EA,GA0,EI},{0,V0,0}];
Print["p=",p];
Figure A9.1. Evaluation of internal force vector p for plane beam element.
The 6 arguments of this module are XY1, XY2, u1, u2, S0 and z0. XY1 lists the coordinates {X 1 , Y2 } of node
1 whereas XY2 lists the coordinates {X 1 , Y2 } of node 2, in the reference configuration. u1 passes the three
displacements: {u X 1 , u Y 1 , θ1 } of node 1, and u2 does the same for node 2. S0 collects the section integrated
constitutive properties {E A0 , G A0 , E I0 }. To apply MacNeal’s RBF, G A0 should be replaced by 12E I0 /L 20
as discussed in Section 9.5. Finally z0 passes the internal forces {N0 , V0 , M0 } in the reference configuration;
9–25
Chapter 9: THE TL TIMOSHENKO PLANE BEAM ELEMENT 9–26
FormMatStiffC0TwoNodePlaneBeam[XY1_,XY2_,u1_,u2_,S0_,z0_]:=
Module[{X1,Y1,X21,Y21,L0,uX1,uY1,theta1,uX2,uY2,theta2,x21,y21,
thetam,ctheta,stheta,Lcpsi,Lspsi,N0,V0,M0,EA0,GA0,EI0,
cphi,sphi,cm,sm,a1,KM},
{X1,Y1}=XY1; {X2,Y2}=XY2; X21=X2-X1; Y21=Y2-Y1;
{uX1,uY1,theta1}=u1; {uX2,uY2,theta2}=u2;
x21=X21+uX2-uX1; y21=Y21+uY2-uY1;
L0=PowerExpand[Sqrt[X21^2+Y21^2]]; thetam=(theta1+theta2)/2;
ctheta=Cos[thetam]; stheta=Sin[thetam];
Lcpsi=Simplify[(X21*x21+Y21*y21)/L0];
Lspsi=Simplify[(X21*y21-Y21*x21)/L0];
em= (ctheta*Lcpsi+stheta*Lspsi)/L0-1;
gm=-(stheta*Lcpsi-ctheta*Lspsi)/L0;
cphi=X21/L0; sphi=Y21/L0;
cm=ctheta*cphi-stheta*sphi; sm=stheta*cphi+ctheta*sphi;
{N0,V0,M0}=z0; {EA0,GA0,EI0}=S0; a1=1+em;
KM = (EA0/L0)*{
{ cm^2,cm*sm,-cm*gm*L0/2,-cm^2,-cm*sm,-cm*gm*L0/2},
{ cm*sm,sm^2,-gm*L0*sm/2,-cm*sm,-sm^2,-gm*L0*sm/2},
{-cm*gm*L0/2,-gm*L0*sm/2,gm^2*L0^2/4,
cm*gm*L0/2,gm*L0*sm/2,gm^2*L0^2/4},
{-cm^2,-cm*sm,cm*gm*L0/2,cm^2,cm*sm,cm*gm*L0/2},
{-cm*sm,-sm^2,gm*L0*sm/2,cm*sm,sm^2,gm*L0*sm/2},
{-cm*gm*L0/2,-gm*L0*sm/2,gm^2*L0^2/4,
cm*gm*L0/2,gm*L0*sm/2,gm^2*L0^2/4}}+
(EI0/L0)*{{0,0,0,0,0,0}, {0,0,0,0,0,0},
{0,0,1,0,0,-1}, {0,0,0,0,0,0},
{0,0,0,0,0,0}, {0,0,-1,0,0,1}}+
(GA0/L0)*{
{sm^2,-cm*sm,-a1*L0*sm/2,-sm^2,cm*sm,-a1*L0*sm/2},
{-cm*sm,cm^2,cm*a1*L0/2,cm*sm,-cm^2,cm*a1*L0/2},
{-a1*L0*sm/2,cm*a1*L0/2, a1^2*L0^2/4,
a1*L0*sm/2,-cm*a1*L0/2, a1^2 L0^2/4},
{-sm^2,cm*sm, a1*L0*sm/2,sm^2,-cm*sm,a1*L0*sm/2},
{ cm*sm,-cm^2,-cm*a1*L0/2,-cm*sm,cm^2,-cm*a1*L0/2},
{-a1*L0*sm/2,cm*a1*L0/2,a1^2*L0^2/4,
a1*L0*sm/2,-cm*a1*L0/2, a1^2*L0^2/4}};
Return[KM] ];
Figure A9.2. Evaluation of material stiffness matrix K M for plane beam element.
these forces are assumed to be constant along the element. Thmodule forms the internal force vector p with a
one point integration rule as discussed earlier, and returns p as function value.
Figures A9.2 and A9.3 show modules FormMatStiffC0TwoNodePlaneBeam and
FormGeoStiffC0TwoNodePlaneBeam. As their name suggest, these form the material and geometric stiffness
components, respectively, of the plane beam element. They have been separated for convenience although for
most applications they are combined to form the tangent stiffness matrix. The 6 arguments are exactly the
9–26
9–27 Exercises
FormGeoStiffC0TwoNodePlaneBeam[XY1_,XY2_,u1_,u2_,S0_,z0_]:=
Module[{X1,Y1,X21,Y21,L0,uX1,uY1,theta1,uX2,uY2,theta2,
x21,y21,thetam,ctheta,stheta,Lcpsi,Lspsi,
N0,V0,M0,EA0,GA0,EI0,cphi,sphi,cm,sm,Nm,Vm,KG},
{X1,Y1}=XY1; {X2,Y2}=XY2; X21=X2-X1; Y21=Y2-Y1;
{uX1,uY1,theta1}=u1; {uX2,uY2,theta2}=u2;
x21=X21+uX2-uX1; y21=Y21+uY2-uY1;
L0=PowerExpand[Sqrt[X21^2+Y21^2]]; thetam=(theta1+theta2)/2;
ctheta=Cos[thetam]; stheta=Sin[thetam];
Lcpsi=Simplify[(X21*x21+Y21*y21)/L0];
Lspsi=Simplify[(X21*y21-Y21*x21)/L0];
em= (ctheta*Lcpsi+stheta*Lspsi)/L0-1;
gm=-(stheta*Lcpsi-ctheta*Lspsi)/L0;
kappa=(theta2-theta1)/L0;
{N0,V0,M0}=z0; {EA0,GA0,EI0}=S0;
Nm=Simplify[N0+EA0*em]; Vm=Simplify[V0+GA0*gm];
cphi=X21/L0; sphi=Y21/L0;
cm=ctheta*cphi-stheta*sphi; sm=stheta*cphi+ctheta*sphi;
KG = Nm/2*{{ 0, 0, sm, 0, 0, sm}, {0, 0, -cm, 0, 0,-cm},
{sm, -cm, -(L0/2)*(1+em), -sm, cm,-(L0/2)*(1+em)},
{0, 0, -sm, 0, 0, -sm}, {0, 0, cm, 0, 0, cm},
{sm, -cm, -(L0/2)*(1+em), -sm, cm,-(L0/2)*(1+em)}}+
Vm/2*{{0, 0, cm, 0, 0, cm}, {0, 0, sm, 0, 0, sm},
{cm, sm, -(L0/2)*gm, -cm, -sm,-(L0/2)*gm},
{0, 0, -cm, 0, 0, -cm}, {0, 0, -sm, 0, 0, -sm},
{cm, sm, -(L0/2)*gm, -cm, -sm,-(L0/2)*gm}};
Return[KG] ];
Figure A9.3. Evaluation of geometric stiffness matrix KG for plane beam element.
FormTanStiffC0TwoNodePlaneBeam[XY1_,XY2_,u1_,u2_,S0_,z0_]:=
Module[{},
KM=FormMatStiffC0TwoNodePlaneBeam[XY1,XY2,u1,u2,S0,z0];
KG=FormGeoStiffC0TwoNodePlaneBeam[XY1,XY2,u1,u2,S0,z0];
Return[KM+KG]];
ClearAll[EA,GA,EI];
KM=FormMatStiffC0TwoNodePlaneBeam[{0,0},{10,0},{0,0,0},{0,0,0},
{EA,GA,EI},{0,0,0}]; Print[KM];
KG=FormGeoStiffC0TwoNodePlaneBeam[{0,0},{4,3},{0,0,0},{0,0,0},
{1,1,1},{10,30,20}]; Print[KG];
Print[Chop[Eigenvalues[N[KG]]]];
Figure A9.4. Evaluation of tangent stiffness matrix K for plane beam element,
along with test statements.
9–27
Chapter 9: THE TL TIMOSHENKO PLANE BEAM ELEMENT 9–28
MergeElemIntoMasterIntForce[pe_,eftab_,pm_]:=
Module[{i,ii,nf=Length[eftab],p}, p=pm;
For[i=1, i<=nf, i++, ii=eftab[[i]];
If [ii>0,p[[ii,1]]+=pe[[i,1]]]
]; Return[p]
];
MergeElemIntoMasterStiff[Ke_,eftab_,Km_]:=
Module[{i,j,ii,jj,nf=Length[eftab],K}, K=Km;
For[i=1, i<=nf, i++, ii=eftab[[i]]; If[ii==0,Continue[]];
For[j=i, j<=nf, j++, jj=eftab[[j]];
If [ii>0 && jj>0,
K[[jj,ii]]=K[[ii,jj]]+=Ke[[i,j]]]
]
]; Return[K]
];
same as for the internal force module, Both matrices are formed with the one-point integration rule.and are
return as function values. The correctness of the implementation is checked with a finite difference technique
implemented in a script not listed here.
Figure A9.4 lists module FormTanStiffC0TwoNodePlaneBeam, which returns the tangent stiffness matrix
K = K M + KG . It simple calls the previous two modules and returns the matrix sum as function value.
Figure A9.5 lists two modules that merge the internal force vector and stiffness matrix, respectively, of one
individual element into the corresponding master quantities for the entire structure.
Figures A9.7 and A9.8 lists two scripts for use in Exercises 9.3 and 9.4, respectively.
9–28
9–29 Exercises
AssembleMasterIntForceOfCantBeam[{L_,P_,Ne_},{Em_,A0_,I0_},u_]:=
Module[{e,numnod,numdof,u1,u2,u3,Le,X1,X2,S0,z0,eftab,
pe,p},
numnod=Ne+1; numdof=3*numnod; Le=L/Ne;
p=Table[0,{numdof},{1}];
z0={P,0,0}; S0={Em*A0,12*Em*I0/Le^2,Em*I0}; (* RBF *)
For [e=1, e<=Ne, e++,
X1=(e-1)*Le; X2=e*Le;
If [e==1,u1={0,0,0}; u2=Take[u,{3*e-2,3*e}];
eftab={0,0,0,3*e-2,3*e-1,3*e}];
If [e>1, u1=Take[u,{3*e-5,3*e-3}]; u2=Take[u,{3*e-2,3*e}];
eftab={3*e-5,3*e-4,3*e-3,3*e-2,3*e-1,3*e}];
pe=FormIntForceC0TwoNodePlaneBeam[{X1,0},{X2,0},u1,u2,S0,z0];
(*Print["pe=",pe];*)
p=MergeElemIntoMasterIntForce[pe,eftab,p];
];
Return[Simplify[p]]
];
AssembleMasterStiffOfCantBeam[{L_,P_,Ne_},{Em_,A0_,I0_},u_]:=
Module[{e,numnod,numdof,u1,u2,u3,Le,X1,X2,S0,z0,eftab,
pe,p},
numnod=Ne+1; numdof=3*numnod-3; Le=L/Ne;
K=Table[0,{numdof},{numdof}];
z0={P,0,0}; S0={Em*A0,12*Em*I0/Le^2,Em*I0}; (* RBF *)
For [e=1, e<=Ne, e++,
X1=(e-1)*Le; X2=e*Le;
If [e==1,u1={0,0,0}; u2=Take[u,{3*e-2,3*e}];
eftab={0,0,0,3*e-2,3*e-1,3*e}];
If [e>1, u1=Take[u,{3*e-5,3*e-3}]; u2=Take[u,{3*e-2,3*e}];
eftab={3*e-5,3*e-4,3*e-3,3*e-2,3*e-1,3*e}];
Ke=FormTanStiffC0TwoNodePlaneBeam[{X1,0},{X2,0},u1,u2,S0,z0];
(*Print["Ke=",Ke];*)
K=MergeElemIntoMasterStiff[Ke,eftab,K];
];
Return[Simplify[K]]
];
Figure A9.6. Assembly modules for cantilevered beam-column under axial load.
9–29
Chapter 9: THE TL TIMOSHENKO PLANE BEAM ELEMENT 9–30
9–30
8
.
The TL
Bar Element
8–1
Chapter 8: THE TL BAR ELEMENT 8–2
TABLE OF CONTENTS
Page
§8.1. Introduction 8–3
§8.2. The Two-Dimensional Bar Element 8–3
§8.2.1. Element Kinematics . . . . . . . . . . . . . . . 8–3
§8.2.2. Strain Measure . . . . . . . . . . . . . . . . 8–4
§8.2.3. Stress Measure . . . . . . . . . . . . . . . . . 8–7
§8.2.4. Total Potential Energy and Residual Force Equations . . . . 8–8
§8.2.5. The Tangent Stiffness Matrix . . . . . . . . . . . . 8–8
§8.3. FEM Coding Using Mathematica 8–10
§8.3.1. The Example Structure . . . . . . . . . . . . . . 8–10
§8.3.2. Forming the Internal Force . . . . . . . . . . . . . 8–11
§8.3.3. The Equilibrium Equations . . . . . . . . . . . . 8–14
§8.3.4. Plotting the Equilibrium Paths . . . . . . . . . . . . 8–15
§8.3.5. Having Some Fun: Snapshots and Animation . . . . . . 8–15
§8.3.6. Forming the Tangent Stiffness Matrix . . . . . . . . . 8–18
§8.3.7. Critical Point Study . . . . . . . . . . . . . . . 8–18
§8. Exercises . . . . . . . . . . . . . . . . . . . . . . 8–25
8–2
8–3 §8.2 THE TWO-DIMENSIONAL BAR ELEMENT
§8.1. Introduction
In the present Chapter the key concepts of nonlinear continuum mechanics reviewed in Chapter 7
are applied to the development of governing equations of bar (truss) elements based on the Total
Lagrangian (TL) kinematic description. For brevity these will be referred to as “TL bar elements.”
There are two ways to construct TL elements:
1. The Standard Formulation (SF)
2. The Core Congruential Formulation (CCF).
The first method is easier to describe and will be presented in this Chapter through examples. The
second one is more flexible and powerful but it is more difficult to teach because it proceeds in
stages. As such it will be relegated to Chapters 10-11.
§8.2. The Two-Dimensional Bar Element
The element developed in this Chapter is a prismatic bar element that can be used to model pin-
jointed plane truss structures of the type sketched in Figure 8.1. These structures undergo large
displacements and rotations but their strains are assumed to remain small so that the material
behavior stays in the linear elastic range. These assumptions allows us to consider only geometric
nonlinear effects.
A two-node bar element appropriate to model members of such truss structures is shown in Figure
8.2. The element moves in the (X, Y ) plane. In the reference (base) configuration the element has
cross section area A0 (constant along the element) and length L 0 . In the current configuration the
cross section area and length become A and L, respectively. The material has an elastic modulus
E that links the axial-stress and axial-strain measures defined below.
Because this Chapter deals primarily with the formulation of an individual element, the identification
superscript (e) will be omitted to reduce clutter until assemblies are considered.
The element has four node displacements and associated node forces. These quantities are collected
in the vectors
u X1 fX1
u f
u = Y1 , f = Y1 , (8.1)
u X2 f X2
uY 2 fY 2
The loads acting on the nodes will be assumed to be conservative.
§8.2.1. Element Kinematics
In accordance with bar theory, to describe the element motion it is sufficient to consider a generic
point of coordinates X located on the longitudinal axis of the reference configuration C0 . That point
maps to point x in the current configuration C. The bar remains straight in any configuration. These
coordinates can be parametrically interpolated from the end nodes as
1
X (ξ ) 2
(1 − ξ )X 1 + 12 (1 + ξ )X 2
X(ξ ) = = 1
Y (ξ ) (1 − ξ )Y1 + 12 (1 + ξ )Y2
2
1 (8.2)
x(ξ ) (1 − ξ )x 1 + 1
(1 + ξ )x 2
x(ξ ) = = 21 2
y(ξ ) (1 − ξ )y1 + 2 (1 + ξ )y2
1
2
8–3
Chapter 8: THE TL BAR ELEMENT 8–4
Reference configuration
(same as base)
Motion
Current configuration
Here ξ is the usual isoparametric coordinate that varies from −1 at node 1 to +1 at node 2. The
displacement field is obtained by subtracting the foregoing position vectors:
1
u X (ξ ) 2
(1 − ξ )u X 1 + 12 (1 + ξ )u X 2
u(ξ ) = x(ξ ) − X(ξ ) = = 1 , (8.3)
u Y (ξ ) (1 − ξ )u Y 1 + 12 (1 + ξ )u Y 2
2
1 u X1
u X (ξ ) (1 − ξ ) 0 1
(1 + ξ) 0 uY 1
u(ξ ) = = 2 2
= N(ξ ) u.
u Y (ξ ) 0 1
2
(1 − ξ) 0 1
2
(1 + ξ ) u X 2
uY 2
(8.4)
The element kinematic defined by these equations is depicted in Figure 8.3.
As discussed in Chapter 7, in the Total Lagrangian (TL) description the Green-Lagrange (GL)
strains and the second Piola-Kirchhoff (PK2) stresses are frequently used as conjugate measures in
the formulation of the internal energy. The only GL strain that appears in the energy expression is
8–4
8–5 §8.2 THE TWO-DIMENSIONAL BAR ELEMENT
Area A0 , length L 0
Y, y 1(X 1 , Y1 )
uY 2
uY 1
fY 1
X, x Area A, length L
uX1
f X1 fY 2
1(x1 , y1 )
uX2
f X2
2(x2 , y2 )
Current configuration C
PK2 stress s, GL strain e
the axial strain e1 ≡ e which is most expediciously defined using the length change as
L 2 − L 20
e= , (8.5)
2L 20
rather than through displacement gradients. Because of the linear displacement assumptions (8.4)
the strain e is constant over the element.1
This expression can be maneuvered into a matrix function of the node displacements as follows. Let
X 21 = X 2 − X 1 , Y21 = Y2 − Y1 , a X = X 21 /L 0 , aY = Y21 /L 0 , u X 21 = u X 2 − u X 1 , u Y 21 = u Y 2 − u Y 1 ,
u mX = (u X 2 − u X 1 )/2 and u m
Y = (u Y 2 − u Y 1 )/2. Some of these quantities can be geometrically
interpreted as illustrated in Figure 8.4. Then
1 This is in fact the only use of the displacement interpolation (8.4) in the following derivations.
8–5
Chapter 8: THE TL BAR ELEMENT 8–6
_
X
1
X (ξ ) (1 − ξ )X 1 + 12 (1 + ξ )X 2
X(ξ ) = = 2
1(X 1 , Y1 ) Y (ξ ) 1
2
(1 − ξ )Y1 + 21 (1 + ξ )Y2
Y, y
1
u X (ξ ) (1 − ξ )v X 1 + 12 (1 + ξ )v X 2
u(ξ ) = x(ξ )−X(ξ ) = = 2
u Y (ξ ) 1
2
(1 − ξ )vY 1 + 12 (1 + ξ )vY 2
X, x
1
x(ξ ) (1 − ξ )x1 + 12 (1 + ξ )x2
x(ξ ) = = 2
y(ξ ) 1
2
(1 − ξ )y1 + 12 (1 + ξ )y2
1(x1 , y1 )
C
2(x2 , y2 )
L 2 = (X 21 + u X 21 )2 + (Y21 + u Y 21 )2 ,
L 2 − L 20 1 1 2
e= = (a X u X 21 + a Y u Y 21 ) + u X 21 + u 2
Y 21
2L 20 L0 2L 20
(8.6)
1 1
= [ −a X −aY a X aY ] u + 2 [ −u mX −u m Y u mX u mY ]u
L0 L0
= Bl + Bn (u) u.
Observe that the GL strain e has been separated into two parts: e = el + en , where
1. el = Bl u, where Bl is constant, depends linearly in the node displacements u. This is called
the linear part of the strain.
2. en = Bn u, in which Bn is a function of the node displacements, depends quadratically on
the displacements. This is called the nonlinear part of the strain, because Bn (u) vanishes if
u → 0. Note that both matrices are constant over the element.
8–6
8–7 §8.2 THE TWO-DIMENSIONAL BAR ELEMENT
_
X
2(X 2 , Y2 ) ψ0
X 21
C0
Y21 X 21
aX = = cos ψ0
L0
L0 Y21
aY = = sin ψ0
L0
Y, y 1(X 1 , Y1 )
x21 L
ax = = cos ψ
L0 L0
y21 L
X, x 1(x1 = X 1 + uX 1 , y1 = Y1 + uY 1 ) ay = = sin ψ
L0 L0
x21 = X 21 + u X 21
ψ
y21 = Y21 + uY 21
C
L 2(x2 = X 2 + uX 2 , y2 = Y2 + uY 2 )
Figure 8.4. Geometric interpretation of quantities used in the study of element kinematics.
∂(B + Bn u) 1
B= = [ −ax −a y ax ay ] (8.8)
∂u L0
in which
u X 21 x2 − x1 x21 u Y 21 y2 − y1 y21
ax = a X + = = , a y = aY + = = . (8.9)
L0 L0 L0 L0 L0 L0
8–7
Chapter 8: THE TL BAR ELEMENT 8–8
where X̄ is directed along the bar longitudinal axis in C0 , as shown in Figure 8.4. This energy
expression is separable because the internal energy depends only on u through e and not on λ.
The finite element residual equations are obtained by differentiating this equation or, equivalently,
making (8.13) stationary with respect to virtual displacement variations δ u:
δ = δU − δ P = ( p − f)T δu = 0. (8.14)
The first variation of U is given by
δU = p δu =
T
A0 (s0 δe + Eeδe) d X̄ = A0 s δe d X̄ = N0 B δu d X̄ = N L 0 B δu
L0 L0 L0
(8.15)
where N = A0 s is the axial force in the current configuration measured per area of the reference
configuration.2 It follows that the internal force vector is
−ax
−a
p = N0 L 0 BT = N y . (8.16)
ax
ay
This equation admits of a simple geometric interpretation; see Figure 8.5. The relation between
N0 = A0 s and the true axial force N = Aσ can be worked out from inspection of this diagram.
The load potential variation δ P simply generates f = λq as can be expected.
2 This is the PK2 axial force; cf (8.11).
8–8
8–9 §8.2 THE TWO-DIMENSIONAL BAR ELEMENT
f Y 1 = −N0 a y
f X 1 = −N0 ax 1
N0
2 f X 2 = N0 a x
f Y 2 = N0 a y
∂p ∂(N L 0 BT ) T ∂s ∂BT
K= = = A0 L 0 B + A0 L 0 s = K M + KG . (8.17)
∂u ∂u ∂u ∂u
The above expression shows that K splits naturally into two parts: K M and KG , which are called
the material stiffness matrix and geometric stiffness matrix, respectively, in the FEM literature.
To get K M note that
∂s ∂(s0 + Ee) ∂e
= =E = EB. (8.18)
∂u ∂u ∂u
Consequently
K M = E A0 L 0 BT B. (8.19)
Inserting the expression (8.8) for B yields
ax2 ax a y −ax2 −ax a y
E A0 ax a y a y2 −ax a y −a y2
KM = (8.20)
L0 −ax2 −ax a y ax2 ax a y
−ax a y −a y2 ax a y a y2
This component of K looks formally similar to the stiffness matrix of a linear bar element,3 except
that B now depends on u. The dependence of K M on the material properties (here the elastic
modulus E) explains the name “material stiffness” given in the FEM literature.
3 To which it reduces if u = 0. In that case ax and a y become the sine and cosine of the angle ψ0 shown in Figure 8.4.
8–9
Chapter 8: THE TL BAR ELEMENT 8–10
The other component can be obtained by differentiating B with respect to the node displacements,
the result being a constant 4 × 4 matrix:
1 0 −1 0
∂B T
1 0 1 0 −1
= 2 . (8.21)
∂u L 0 −1 0 1 0
0 −1 0 1
1 0 −1 0
N 0 1 0 −1
KG = . (8.22)
L 0 −1 0 1 0
0 −1 0 1
This component of K depends only on the stress state in the current configuration, because N = A0 s.
No material properties appear. Thus the name “geometric stiffness” applied to KG .4
Remark 8.1. Assuming that E, A0 and L 0 are nonzero, the rank of K M is obviosly one because B is a 1 × 4
matrix. On the other hand the rank of the numerical matrix in (8.21) is 2 (because its eigenvalues are 2, 2, 0, 0).
Consequently KG has rank 2 if s is nonzero and 0 otherwise. Combining these results it can be shown that the
rank of K = K M + KG is 1 if the configuration is stressed and 2 otherwise. In other words, the rank deficiency
is 3 and 2, respectively. The implications of this property in the analysis of stability are studied later.
Remark 8.2. The addition of KG increases the bar stiffness if the current configuration is in tension (s > 0),
but it reduces it if the current configuration is in compression (s < 0). This is in accord with physical intuition.
The main effect of this stiffness is on the rotational rigid-body motions of the bar about the Z axis.
This section illustrates the use of Mathematica as a quick way to do some simple problems with the TL bar
element. The use of a computer symbolic system as prototyping language has several advantages:
(i) The code is very compact, because of higher lever array notation. Because of its compactness, it can be
easily translated into other high-level languages such as, for example, Matlab or Maple.
(ii) Debugging is often quicker than with programming languages because of the interpretative nature of the
language front end. This is of course platform dependent; the implementation of Mathematica on a
Macintosh or PC is friendler than on a Unix workstation.
(iii) Graphic output is part of the language. On the Macintosh, graphics can be transported into documents
such as this one via PostScript and Adobe Illustrator.
(iv) The language can manipulate both algebraic (symbolic) and numerical expressions. The later may be
done exactly or in floating-point arithmetic. This flexibility is convenient for many situations, as the
examples below illustrate.
4 In the pre-1970 FEM literature, the name “initial stress stiffness” was used for KG by some authors.
8–10
8–11 §8.3 FEM CODING USING MATHEMATICA
fY = λ
Here λ > 0 means crown
load is applied upwards
uY
2 uX
E, A0 (1) (2) E, A0
Y, y
;; ;;
H
α
3
;; ;;
1 X, x
Figure 8.6. An arch as a two-bar FE model. This model has two degrees of
freedom: u X and u Y , which are the displacements of node 2. This
structure will be used as example for use of Mathematica in §8.3.
The two-spring arch studied in Exercise 6.3 will be used as a structure that illustrates the use of Mathematica
in conjunction with finite elements. The structure is shown in Figure 8.6. It is modeled with two TL bar
elements labeled (1) and (2), which join nodes 1-2 and 2-3, respectively. The modulus of elasticity and the
bar areas are unity. The arch geometry is defined by the span S and height H . The reference configuration is
assumed stress free.
Because nodes 1 and 3 are pinned to the ground, the only degrees of freedom are the horizontal and vertical
displacements of node 2. To avoid clutter these will be denoted in the sequel as u X and u Y rather than u X 2
and u Y 2 , respectively. The vertical motion u Y is positive upwards. The node-2 components of the applied
and internal force vector will be also denoted simply by ( f X , f Y ) and ( p X , pY ), respectively. Displacement and
force components for nodes 1 and 3 will be explicitly removed from the governing equations.
The arch is loaded by a vertical force f Y = λ applied at node 2, positive upwards. The applied horizontal
force f X will be always zero. Recall that in Exercise 6.3 only symmetric motions of the arch under a vertical
load were considered. That constraint effectively reduces the model to one degree of freedom: u Y , and allows
only limit point behavior. When two degrees of freedoms are considered, however, a richer set of possibilities
emerges; notably lateral bifurcation under a vertical load.
The Mathematica module FormIntForce2DTwoNodeBar listed in Figure 8.7 forms the internal force vector
of an individual 2D bar finite element handled by the TL description. The arguments of this module are
XY1 Coordinates {X 1 , Y1 } of end node 1 in C0 .
XY1 Coordinates {X 2 , Y2 } of end node 2 in C0 .
uX1 Displacements {u X 1 , u Y 1 } of end node 1.
8–11
Chapter 8: THE TL BAR ELEMENT 8–12
FormIntForce2DTwoNodeBar[XY1_,XY2_,uXY1_,uXY2_,Em_,A0_,s0_]:=
Module[{X1,Y1,X2,Y2,X21,Y21,uX1,uY1,uX2,uY2,uX21,uY21,L0,L,
e,s,ax,ay,ifv},
{X1,Y1}=XY1; {X2,Y2}=XY2; X21=X2-X1; Y21=Y2-Y1;
{uX1,uY1}=uXY1; {uX2,uY2}=uXY2; uX21=uX2-uX1; uY21=uY2-uY1;
L0=Sqrt[X21^2+Y21^2]; L=Sqrt[(X21+uX21)^2+(Y21+uY21)^2];
e=(L-L0)*(L+L0)/(2*L0^2); s=s0+Em*e;
ax=(X21+uX21)/L0; ay=(Y21+uY21)/L0;
ifv=A0*s*{{-ax}, {-ay}, {ax}, {ay}};
Return[Simplify[ifv]]
];
pe=FormIntForce2DTwoNodeBar[{2,3},{5,7},{1,0},{1,0}, 20,12, 0];
Print["Internal elem force pe=",pe];
pe=FormIntForce2DTwoNodeBar[{2,3},{5,7},{1,0},{1,0}, 20,12, s0];
Print["Internal elem force pe=",pe];
Figure 8.7. A Mathematica module that computes the internal force vector p of an individual
2-node 2D bar element. Output from two test statements is shown.
8–12
8–13 §8.3 FEM CODING USING MATHEMATICA
MergeElemIntoMasterIntForce[pe_,eftab_,p_]:=
Module[{i,ii,neldof,fmaster}, fmaster=p;
neldof=Dimensions[eftab][[1]];
For[i=1, i<=neldof, i++, ii=eftab[[i]];
If [ii>0,fmaster[[ii,1]]+=pe[[i,1]]]
]; Return[fmaster]
];
AssembleMasterIntForceOfTwoBarArch[span_,height_,Em_,A0_,uX_,uY_]:=
Module[{f1,f2,fmaster},
fmaster=Table[0,{2},{1}];
f1=FormIntForce2DTwoNodeBar[{-span/2,0},{0,height},
{0,0},{uX,uY},Em,A0,0];
fmaster=MergeElemIntoMasterIntForce[f1,{0,0,1,2},fmaster];
f2=FormIntForce2DTwoNodeBar[{0,height},{span/2,0},
{uX,uY},{0,0},Em,A0,0];
fmaster=MergeElemIntoMasterIntForce[f2,{1,2,0,0},fmaster];
Return[Simplify[fmaster]]
];
ClearAll[Em,A0,S,H,uX,uY,f];
p=AssembleMasterIntForceOfTwoBarArch[2,2.5,10,.75,-0.4,0.25];
Print["Master Int Force p= ",p];
p=AssembleMasterIntForceOfTwoBarArch[S,H,Em,A0,uX,uY];
Print["Master Int Force p= ",p];
Figure 8.8. Two Mathematica modules that assemble the internal force vector p
for the two-bar arch structure of Figure 8.6. Two test statements, one
numeric and the other symbolic, are shown along with output.
Two test statements are also shown in the figure. The first one is purely numeric and sets the following
properties for the structure:
S = 2, H = 2.5, E = 10, A0 = 0.75, u X = 0.4, u Y = −0.25 (8.23)
and the module AssembleMasterIntForceOfTwoBarArch returns
pX −0.53365
p= = (8.24)
pY 1.55576
The second one specifies symbolic values for all of the arguments: S for span S, H for the height H , etc. The
result (displayed in InputForm so it can be cut and pasted easily) gives the internal force for an arbitrary
8–13
Chapter 8: THE TL BAR ELEMENT 8–14
structure:
pX 4E A0 u X (S 2 + 2u 2X + 4H u Y + 2u 2Y )
p= = . (8.25)
pY (4H 2 + S 2 )3/2 2(H + u Y )(u 2X + 2H u Y + u 2Y )
It follows that p X = 0 and pY = λ. Expression (8.26) shows that p X = 0 is a cubic equation that has the trivial
root u X = 0 in addition to the two roots of the quadratic S 2 + 2u 2X + 4H u Y + 2u 2Y = 0. Rather than solving
these by hand, let us exemplify the use of Mathematica to solve a polynomial equation using the built-in
Solve function. These roots are then substituted, using the /. operator, into the other equation λ = pY to get
the equations of the 3 equilibrium paths. The appropriate statements and output are shown in Figure 8.9.
ClearAll[Em,A0,S,H,uX,uY,pX,pY,lambda];
pX = (4*A0*Em*uX*(S^2+2*uX^2+4*H*uY+2*uY^2))/(4*H^2+ S^2)^(3/2);
pY = (8*A0*Em*(H+uY)*(uX^2+2*H*uY+uY^2))/(4*H^2+S^2)^(3/2);
roots=Solve[pX==0, uX];
Print["roots of pX=0 are ",roots];
Print["lambda=",Simplify[pY/.roots]];
Figure 8.9. Illustration of use of Mathematica built-in Solve function to obtain the
roots of p X = 0 in terms of u X , where p X is given by (8.25). These roots
are then substituted into pY = λ with pY from the same equation,
to get the expression of the 3 equilibrium paths in the form λ = λ(u Y ).
The definition of pX in Figure 8.9 was directly set via cut-and-paste from the output shown in Figure 8.8. From
the printed solution it is obvious that the three roots in terms of u X are
p
u X = 0, u s1,s2
X = ± −S 2 /2 − 2H u Y − u 2Y , (8.27)
p
The solution u X , where superscript p stands for “primary,” corresponds to a symmetric deformation of the
arch in its primary equilibrium path. This represents the kind of response studied in Exercise 6.3.
8–14
8–15 §8.3 FEM CODING USING MATHEMATICA
8–15
Chapter 8: THE TL BAR ELEMENT 8–16
Figure 8.10. Statements to generate λ = λ(u Y ) response plots for four geometries.
√
For all cases E = A√ ◦
0 = 1 and
√ S = 2. The height H is set to 3/3 (α = 30 ,
as in Exercise 6.3), 2/2, 3 and 3.
lambda lambda
Case 1: H=Sqrt[3]/3 Case 2: H=Sqrt[2]
0.2
0.4
0.1
0.2
vY vY
-1 -0.8 -0.6 -0.4 -0.2 -2.5 -2 -1.5 -1 -0.5
-0.2
-0.1
-0.4
-0.2
lambda lambda
Case 3: H=Sqrt[3] Case 4: H=3
0.4 0.3
0.2
0.2
0.1
vY vY
-3.5 -3 -2.5 -2 -1.5 -1 -0.5 -6 -5 -4 -3 -2 -1
-0.1
-0.2
-0.2
-0.4 -0.3
Figure 8.11. Plots of the arch equilibrium paths projected on the λ versus u Y plane
for the four cases described in Figure 8.10.
8–16
8–17 §8.3 FEM CODING USING MATHEMATICA
ClearAll[S,H,Em,A0,uX,uY]; H=Sqrt[3.];
lambdamax[S_,H_,Em_,A0_]:=3.0792*A0*Em*H^3/(4*H^2+S^2)^(3/2);
primarypath[S_,H_,Em_,A0_,uY_]:=
{0,uY,8*A0*Em*uY*(H+uY)*(2*H+uY)/(4*H^2+S^2)^(3/2)};
secondarypath1[S_,H_,Em_,A0_,uY_]:=Module[{c,uYB1,uYB2},
uYB1=Re[N[-H+Sqrt[H^2-S^2/2]]]; uYB2=Re[N[-H-Sqrt[H^2-S^2/2]]];
c=4*A0*Em*S^2/(4*H^2+S^2)^(3/2);
If [N[uY]>=uYB1, Return[{0,uYB1,-c*(H+uYB1)}]];
If [N[uY]<=uYB2, Return[{0,uYB2,-c*(H+uYB2)}]];
Return[{Sqrt[-S^2/2-2*H*uY-uY^2],uY,-c*(H+uY)}]
];
secondarypath2[S_,H_,Em_,A0_,uY_]:=Module[{c,uYB1,uYB2},
uYB1=Re[N[-H+Sqrt[H^2-S^2/2]]]; uYB2=Re[N[-H-Sqrt[H^2-S^2/2]]];
c=4*A0*Em*S^2/(4*H^2+S^2)^(3/2);
If [N[uY]>=uYB1, Return[{0,uYB1,-c*(H+uYB1)}]];
If [N[uY]<=uYB2, Return[{0,uYB2,-c*(H+uYB2)}]];
Return[{-Sqrt[-S^2/2-2*H*uY-uY^2],uY,-c*(H+uY)}]
];
lambdarange=1.1*{-lambdamax[2,H,1,1],lambdamax[2,H,1,1]};
pp=ParametricPlot3D[primarypath[2,H,1,1,uY], {uY,0,-2*H},
PlotPoints->201, PlotRange->{{-H,H},{0,-2*H},lambdarange},
BoxRatios->{1,2,1},AxesLabel->{"uX","uY","lambda"},
DisplayFunction->Identity];
ps1=ParametricPlot3D[secondarypath1[2,H,1,1,uY], {uY,0,-2*H},
PlotPoints->201, PlotRange->{{-H,H},{0,-2*H},lambdarange},
BoxRatios->{1,2,1},AxesLabel->{"uX","uY","lambda"},
DisplayFunction->Identity];
ps2=ParametricPlot3D[secondarypath2[2,H,1,1,uY], {uY,0,-2*H},
PlotPoints->201, PlotRange->{{-H,H},{0,-2*H},lambdarange},
BoxRatios->{1,2,1},AxesLabel->{"uX","uY","lambda"},
DisplayFunction->Identity];
Show[pp,ps1,ps2,ViewPoint->{3,1,2},DisplayFunction->$DisplayFunction];
Figure 8.12. Statements used to generate response plots in the three-dimensional state-control
space (u X , u Y , λ), such as the one shown in the next figure.
To illustrate what happens to the arch as it traverses its equilibrium path, configuration “snapshots” and
particularly animations are invaluable and fun to do. The program shown in 8.14 does the first kind of display,
showing configurations as the arch traverses the equilibrium path by incrementing u Y . The program also
depicts (very roughly) the magnitude of the applied vertical force. The result of running this program on an
arch with S = 2, H = 3, E = A0 = 1 is shown in Figure 8.15.
A plot such as that of Figure 8.15 is a bit confusing because all configurations are in the same cell, which
appears on the screen as a flash. More instructive is the use of animations. Mathematica animations are
much like movies or videos: they are a sequence of pictures that, when displayed in rapid succession, appear
to move. In Mathematica versions endowed with a Notebook interface, any sequence of graphic cells can
8–17
Chapter 8: THE TL BAR ELEMENT 8–18
-1
uX 0
0.2
λ 0
-0.2
-3
-2
-1
uY
0
be animated by simply doubly clicking the mouse on one of the pictures (any one will do). The cell group
containing the pictures will be selected and the animation started automatically.
The program shown in Figure 8.16 is a variant of the program of Figure 8.14. The logic is rearranged to
produce a sequence of graphic cells, one per configuration, and so make the animation possible. Unfortunately
the result cannot be shown on the old-fashioned medium of these Notes. But it should be possible in the next
millennium.
The tangent stiffness matrix derived in §8.2 is computed by the Mathematica module shown in Figure 8.17,
along with some test statements. The arguments of this module, called FormTangentStiff2DTwoNodeBar
are the same as those of the internal force subroutine shown in Figure 8.7. The module returns the 4 × 4
K = K M + KG matrix for the element. Figure 8.18 lists two modules: MergeElemIntoMasterStiff and
AssembleMasterStiffOfShallowArch, which together assemble the 2 × 2 master stiffness matrix for the
example arch structure.
The main use of the stiffness matrix in this and subsequent Chapters is the investigation of critical points,
because solution argorithms based on K have not yet been described.
A detailed symbolic analysis, not reported here, for arbitrary S, H , E, A0 , u X and u Y shows that under a
vertical load the determinant of K can only vanish if u X = 0. Thus it is enough to restrict consideration to
symmetric motions. The stiffness matrix for u X = 0 and its determinant are produced by the test statements
shown in Figure 8.19 along with their output. It is seen that
8–18
8–19 §8.3 FEM CODING USING MATHEMATICA
8–19
Chapter 8: THE TL BAR ELEMENT 8–20
Figure 8.16. A variant of the program of Figure 8.14. This program generates a
sequence of configuration plots (frames) that can be used for animation.
8E A0 S 2 /2 + 2H u Y + u 2Y 0
K(S, H, E, A0 , u X = 0, u Y ) = , (8.31)
(4H 2 + S 2 )3 0 2H 2 + 6H u Y + 3u 2Y
Therefore
det(K) = 32E 2 A20 (S 2 + 4H u Y + 2u 2Y )(2H 2 + 6H u Y + 3u 2Y )/ (4H 2 + S 2 )3 . (8.32)
Because K is diagonal if u X = 0, det(K) factors out into
√ two quadratic polynomials in u Y . Consequently there
are four critical points (C.P.s). The roots −1 ± H/ 3 of 2H 2 + 6H u Y + 3u 2Y = 0 are always real and can
be easily checked to correspond to limit points. The other two, which are roots of S 2 /2 + 2H u Y + u 2Y = 0,
correspond to bifurcation points and are real if 2H 2 ≥ S 2 , as already announced in §8.3.4.
The expression of the four roots of det(K ) = 0 in terms of u Y is symbolically obtained using the statements
listed in Figure 8.19. These roots are substituted, again using the /. operator, into the equations λ(u Y ) of the
primary equilibrium path to get the four values of λ shown in the output cell of Figure 8.19:
√
16E A0 H 3 2 2 E A0
λ L1 = −λ L2 = − √ , λ B1 = −λ B2 = − S 2 2H 2 − S 2 . (8.33)
3 3 (4H 2 + S 2 )3 (4H 2 + S 2 )3
The subscripts L and B stand here for limit and bifurcation point, respectively. That classification is readily
done given the diagonal nature of K shown in (8.31).
Figure 8.20 shows statements used to produce plots of λ L1 and λ B1 versus H upon setting E = A0 = 1
and S = 2. Thus H = tan α. Negative λ (downward load) is plotted upwards for convenience. The plot
8–20
8–21 §8.3 FEM CODING USING MATHEMATICA
FormTangentStiff2DTwoNodeBar[XY1_,XY2_,uXY1_,uXY2_,Em_,A0_,s0_]:=
Module[{X1,Y1,X2,Y2,X21,Y21,uX21,uY21,L0,L,e,s,ax,ay,Ke},
{X1,Y1}=XY1; {X2,Y2}=XY2; X21=X2-X1; Y21=Y2-Y1;
{uX1,uY1}=uXY1; {uX2,uY2}=uXY2; uX21=uX2-uX1; uY21=uY2-uY1;
L0=Sqrt[X21^2+Y21^2]; L=Sqrt[(X21+uX21)^2+(Y21+uY21)^2];
e=(L-L0)*(L+L0)/(2*L0^2); s=s0+Em*e;
ax=(X21+uX21)/L0; ay=(Y21+uY21)/L0;
Ke=(Em*A0/L0)*{{ ax*ax, ax*ay,-ax*ax,-ax*ay},
{ ax*ay, ay*ay,-ay*ax,-ay*ay},
{-ax*ax,-ay*ax, ax*ax, ay*ax},
{-ax*ay,-ay*ay, ay*ax, ay*ay}} +
(A0*s/L0)*{{1,0,-1,0},{0,1,0,-1},{-1,0,1,0},{0,-1,0,1}};
Return[Simplify[Ke]]
];
Ke=FormTangentStiff2DTwoNodeBar[{0,-S+H},{0,H},{0,0},{0,uY},Em,A0,0];
Ke=Simplify[Ke/.{Sqrt[S^2]->S,(1/Sqrt[S^2])->1/S}];
Print["Ke=",Ke//InputForm];
Ke=FormTangentStiff2DTwoNodeBar[{-4,0},{0,3},{0,0},{0,0}, Em,A0,s0];
Print["Ke=",Ke];
Print[Eigenvalues[Ke]];
is shown in Figure 8.21 upon some labeling √ and beautification via Adobe Illustrator. The coalescence
of the bifurcation and limit points for H = 3 or α = 60◦ is clear in the plot. This figure also shows that
for fixed S, E and A0 the maximum vertical load carrying capacity for this structure is obtained for that rise
angle. This optimality criterion (coalescence of critical points gives the strongest structure) is typical of those
encountered in optimal design with stability constraints.
8–21
Chapter 8: THE TL BAR ELEMENT 8–22
MergeElemIntoMasterStiff[Ke_,eftab_,Km_]:=
Module[{i,j,ii,jj,neldof,K}, K=Km;
neldof=Length[eftab];
For[i=1, i<=neldof, i++, ii=eftab[[i]];
For[j=i, j<=neldof, j++, jj=eftab[[j]];
If [ii>0 && jj>0,
K[[jj,ii]]=K[[ii,jj]]+=Ke[[i,j]]]
]
]; Return[K]
];
AssembleMasterStiffOfShallowArch[S_,H_,Em_,A0_,uX_,uY_]:=
Module[{K1,K2,K},
K=Table[0,{2},{2}];
K1=FormTangentStiff2DTwoNodeBar[{-S/2,0},{0,H},
{0,0},{uX,uY},Em,A0,0]; (*Print[K1//TableForm];*)
K=MergeElemIntoMasterStiff[K1,{0,0,1,2},K];
K2=FormTangentStiff2DTwoNodeBar[{0,H},{S/2,0},
{uX,uY},{0,0},Em,A0,0]; (*Print[K2//TableForm];*)
K=MergeElemIntoMasterStiff[K2,{1,2,0,0},K];
Return[Simplify[K]]
];
ClearAll[S,H,Em,A0,uX,uY];
K=AssembleMasterStiffOfShallowArch[S,H,Em,A0,uX,uY];
Print["Master stiffness matrix = ",K//InputForm];
Figure 8.18. Two modules that assemble the tangent stiffness matrix K for the arch
structure of Figure 8.6.
8–22
8–23 §8.3 FEM CODING USING MATHEMATICA
ClearAll[H,S,Em,A0,uY];
detK=(32*A0^2*Em^2*(S^2 + 4*H*uY + 2*uY^2)*
(2*H^2 + 6*H*uY + 3*uY^2))/(4*H^2 + S^2)^3;
sol=Simplify[Solve[detK==0,uY]];
Print[sol//InputForm];
lambdaCP=(8*A0*Em*uY*(H + uY)*(2*H + uY))/(4*H^2 + S^2)^(3/2)/.sol;
lambdaCP=Simplify[Expand[lambdaCP]];
Print["lambdaCP=",lambdaCP//InputForm];
Figure 8.19. Obtaining the critical points as roots of the tangent stiffness determinant.
8–23
Chapter 8: THE TL BAR ELEMENT 8–24
ClearAll[H,S,Em,A0,uY];
lambdaL=(16*A0*Em*H^3)/(3*3^(1/2)*(4*H^2 + S^2)^(3/2));
lambdaB=(2*2^(1/2)*A0*Em*S^2*(2*H^2 - S^2)^(1/2))/(4*H^2 + S^2)^(3/2);
set={Em->1,A0->1,S->2};
pL=Plot[lambdaL/.set,{H,0,5},DisplayFunction->Identity,
AxesLabel->{"H","lambda at CPs"},
PlotLabel->"lambdaCP vs H (S=2,E=A0=1)"];
pB=Plot[lambdaB/.set,{H,Sqrt[2.00001],5},DisplayFunction->Identity];
Figure 8.20. Program to produce the critical loads plots of next Figure.
lambda at CPs
lambdaCP vs H (S=2,E=A0=1)
0.35 √ λ L1
H= 3
0.3
0.15
λ B1
λ L1
0.1 λ B1
0.05
H
1
√ 2 3 4 5
2
Figure 8.21. Plot of maximum load capacity as defined by λ (negative upwards) at the
critical points for arches with different heights H (S = 2, E = A0 = 1).
8–24
8–25 Exercises
EXERCISE 8.1 [D:10] The α = 30◦ (case 1) plot in Figure 8.11 shows a limit-point λ of about 0.048. Explain
why that value could be different from that of the “exact solution” plot of Figure E6.3 for the same structure.
EXERCISE 8.2 [A/C:20] A problem in optimal design. For the 2-bar arch example structure, the plot in
Figure 8.21 makes evident that, given the span S, modulus E and cross section area A0 , the√ largest vertical-
downward-load capacity is obtained if the rise angle is α = 60◦ , or H = (S/2) tan α = S 3/2. Suppose
that the design objective is to get the largest possible λmax = |λ| per unit volume of material, with respect to
the total volume. Mathematically: maximize λmax /V0 with respect to V0 , where λmax is the largest downward
load sustained before failure by snap-through or bifurcation. What would be the best α?
fY = λ
uY
2 uX
E, A0 (1) (2) E, A0
H Y, y
;; ;;
α
1 X, x 3
S
S
(3)
E, A0
;; 4
EXERCISE 8.3 [A/C:25] You go to work as a nonlinear-FEM engineer for a car company. Your supervisor
assigns you the job of designing a component of a wheel suspension system that can be modeled by the 3-bar
structure depicted in Figure E8.1. The model has the dimensions and properties shown and is only subjected
to vertical loads at node 2. The length S, bar section areas A0 and elastic modulus E are known, but the rise
angle α > 0 is a design variable. Find the largest α for which bifurcation, which is bad for the wheel,
√ cannot
occur. (For the 2-bar arch example structure that maximum α was shown to be defined by tan α ≤ 2/2.)
8–25
Chapter 8: THE TL BAR ELEMENT 8–26
;;
Y, y
(1)
Gravity field g
uY
1 uX X, x
Point mass m
EXERCISE 8.4 [A:C:20] Although this course focuses on statics, this exercise deals with the effect of the
geometric stiffness on vibrations. Consider the pendulum configuration idealized in Figure E8.2. A lumped
mass m is suspended by a weightless elastic string. The string is modeled as a 2-node bar element. This
element is under a tensile prestress s0 = mg/A0 , where g is the accelaration of gravity. The tangent stiffness
matrix for the cable element in the reference configuration is K = K M + KG , which is 2 × 2 upon removing
the degrees of freedom at the fixed node 2. Because of the prestress the geometric stiffness does not vanish.
The order-2 vibration eigenproblem is
where i is the mode index, ωi is the i th circular frequency in radians per second, zi the associated eigenvector
that include the horizontal and vertical displacements of node 1, and the mass matrix is
m 0
M= (E8.2)
0 m
Compute the two frequencies ω1 and ω2 . One of them, say ω1 , describes pendulum motions while the other
one pertains to a “bar mode” associated with axial motions. Discuss what happens to ω1 and ω2 if E → ∞,
which characterizes
√ the “inextensional string” limit, and whether the classical pendulum small-oscillations
frequency ω P = g/L is correct.
EXERCISE 8.5 (Requires knowledge of continuum mechanics.) [A:15] Suppose that the bar-element material
is linear isotropic, with elastic modulus E and ν is Poisson’s ratio ν. Find the relation between the true (Cauchy)
axial stress σ = σx x in the bar and the PK2 axial stress s = s X X . Hint: study the change in cross section areas.
8–26
7
.
Review of
Continuum
Mechanics
7–1
Chapter 7: REVIEW OF CONTINUUM MECHANICS 7–2
TABLE OF CONTENTS
Page
§7.1. The FEM Discretization 7–3
§7.2. Notation: Continuum vs. Discrete Mechanics 7–3
§7.3. Particles, Motions, Displacements, Configurations 7–4
§7.3.1. Distinguished Configurations . . . . . . . . . . . . . 7–4
§7.3.2. Kinematic Descriptions . . . . . . . . . . . . . . 7–6
§7.3.3. Coordinate Systems . . . . . . . . . . . . . . . . 7–7
§7.3.4. Configurations and Staged Analysis . . . . . . . . . . 7–8
§7.4. Kinematics 7–8
§7.4.1. Deformation and Displacement Gradients . . . . . . . . . 7–8
§7.4.2. Stretch and Rotation Tensors . . . . . . . . . . . . . 7–9
§7.4.3. Green-Lagrange Strain Measure . . . . . . . . . . . . 7–10
§7.4.4. Strain-Gradient Matrix Expressions . . . . . . . . . . 7–11
§7.4.5. Pull Forward and Pull Back . . . . . . . . . . . . . . 7–12
§7.5. Stress Measure 7–12
§7.6. Constitutive Equations 7–13
§7.7. Strain Energy Density 7–14
§7. Exercises . . . . . . . . . . . . . . . . . . . . . . . 7–15
7–2
7–3 §7.2 NOTATION: CONTINUUM VS. DISCRETE MECHANICS
In Chapters 3 through 6 we have studied some general properties of the governing force-equilibrium
equations of geometrically nonlinear structural systems with finite number of degrees of freedom (DOF).
The DOFs are collected in the state variable vector u. Those residual equations, being algebraic, are
well suited for numerical computation.
Mathematical models of real structures, however, possess an infinite number of DOFs. As such they
cannot be handled by numerical computations. The reduction to a finite number is accomplished by
discretization methods. As noted in Chapter 1, for nonlinear problems in solid and structural mechanics
the finite element method (FEM) is the most widely used discretization method.
This section provides background material for the derivation of geometrically nonlinear finite elements.
The material gives a review of kinematic, kinetic and constitutive concepts from the three-dimensional
continuum mechanics of an elastic deformable body, as needed in following Chapters. Readers familiar
with continuum mechanics should peruse it to grab notation.
Continuum mechanics deals with vector and tensor fields such as displacements, strains and stresses.
Four types of notation are in common use:
1. Indicial Notation. Also called component notation. This notation uses indexed components along
with abbreviation rules such as commas for partial derivatives and Einstein’s summation convention.
It is a powerful notation, and as such is preferred in journals and monographs. It has the advantage
of readily handling arbitrary tensors of any order, arbitrary coordinate systems and nonlinear relations.
It sharply distinguishes between covariant and contravariant quantities, which is necessary in non-
Cartesian coordinates. Tends to conceal or mask intrinsic properties, however, and as such is not
suitable for basic instruction.
2. Direct Notation. Sometimes called algebraic notation. Vectors and tensors are represented by single
symbols, usually bold letters. Has the advantage of compactness and quick visualization of intrinsic
properties. Some operations correspond to matrix notations while others do not. This fuzzy overlap
can lead to confusion in FEM work.
3. Matrix Notation. This is similar to the previous one, but entities are rearranged as appropriate
so that only matrix operations are used. It can be translated directly to discrete equations as well as
matrix-oriented programming languages such as Matlab. It has the disadvantage of losing contact with
the original physical entities along the way. For example, stress is a symmetric second-order tensor that
is rearranged as a 6-component vector for FEM developments. This change loses essential properties.
For instance it makes sense to talk of principal stresses as eigenvalues of the stress tensor. But those
get lost (as or least moved to the background) when rearranged as a stress vector.
4. Full Notation. In the full-form notation every term is spelled out. No ambiguities of interpretation
can arise; consequently this works well as a notation of last resort, and also as a “comparison template”
against one can check out the meaning of more compact expressions. It is also useful for programming
in low-order languages.
As an example, consider the well known dot product between two physical vectors in 3D space, a =
(a1 , a2 , a3 ) and b = (b1 , b2 , b3 ) done in the four different notations:
ai bi =
a.b =
aT b = a1 b1 + a2 b2 + a3 b3 . (7.1)
indicial direct matrix full
7–3
Chapter 7: REVIEW OF CONTINUUM MECHANICS 7–4
In the following review the direct, matrix and full notation are preferred, whereas the indicial notation
is avoided. Usually the expression is first given in direct form and confirmed by full form if feasible.
Then it is transformed to matrix notation for later used in FEM developments. The decision leads to
possible ambiguities against reuse of vector symbols in two contexts: continuum mechanics and FEM
discretizations. Such ambiguities are resolved in favor of keeping FEM notation simple.
In the present section a structure is mathematically treated as a continuum body B. The body is
considered as being formed by a set of points P called particles, which are endowed with certain
mechanical properties. For FEM analysis the body is divided into elements.
Particles displace or move in response to external actions characterized by control parameters i or,
following the stage reduction discussed in Chapter 3, the single stage parameter λ. A one-parameter
series of positions occupied by the particles as they move in space is called a motion. The motion may
be described by the displacement u(P) ≡ u(x) of the particles with respect to a base or reference state
in which particle P is labelled P0 . [The underlining is used to distinguish the physical displacement
vector from the finite element node displacement array, which is a computational vector.]
The displacements of all particles u(x, y, z) such that x ≡ x, y, z ∈ B, constitutes the displacement
field.
The motion is said to be kinematically admissible if:
1. Continuity of particles positions is preserved so that no gaps or voids appear. (The mathematical
statement of this condition is given later.)
2. Kinematic constraints on the motion (for example, support conditions) are preserved.
A kinematically admissible motion along a stage will be called a staged motion. For one such motion
the displacements u(x) characterize the state and the stage control parameter λ characterizes the control
or action. Both will be generally parametrized by the pseudo-time t introduced in Chapter 3, so that a
staged motion can be generally represented by
A particular characteristic of geometrically nonlinear analysis is the need to carefully distinguish among
different configurations of the structure.
As noted above, each set of kinematically admissible displacements u(x) plus a staged control parameter
λ defines a configuration. This is not necessarily an equilibrium configuration; in fact it will not usually
7–4
7–5 §7.3 PARTICLES, MOTIONS, DISPLACEMENTS, CONFIGURATIONS
be one. It is also important to realize that an equilibrium configuration is not necessarily a physical
configuration assumed by the actual structure.1
Some configurations that are important in geometrically nonlinear analysis receive special qualifiers:
admissible, perturbed, deformed, base, reference, iterated, target, corotated, aligned
7–5
Chapter 7: REVIEW OF CONTINUUM MECHANICS 7–6
x = x(t)
Z, z
u=x−X
V
X, x Y, y X ≡ x0 P0 (X)
Base configuration C0
(often same as reference)
7–6
7–7 §7.3 PARTICLES, MOTIONS, DISPLACEMENTS, CONFIGURATIONS
Remark 7.2. The TL formulation remains the most widely used in continuum-based finite element codes. The CR
formulation is gaining in popularity for structural elements such as beams, plates and shells. The UL formulation
is primarily used in treatments of vary large strains and flow-like behavior.
2 In dynamic analysis the global frame may be moving in time as a Galilean or inertial frame. This is convenient to track the
motion of objects such as aircraft or satellites.
3 The choice between {X 1 , X 2 , X 3 } versus {X, Y, Z } and likewise {x1 , x2 , x3 } versus {x, y, z} is a matter of convenience. For
example, when developing specific finite elements it is preferable to use {X, Y, Z } or {x, y, z} so as to reserve coordinate
subscripts for node numbers.
7–7
Chapter 7: REVIEW OF CONTINUUM MECHANICS 7–8
Remark 7.3. Variations of this notation scheme are employed as appropriate to the subject under consideration.
For example, the coordinates of P in a target configuration Cn may be called (xn , yn , z n ).
Remark 7.4. In continuum mechanics, coordinates (X, Y, Z ) and (x, y, z) are called material and spatial coor-
dinates, respectively. In general treatments both systems are curvilinear and need not coalesce. The foregoing
relation (7.4) is restrictive in two ways: the base coordinate systems for the reference and current configurations
coincide, and that system is Cartesian. This assumption is sufficient, however, for the problems treated in this
course.
Remark 7.5. The dual notation (X, Y, Z ) ≡ (x 0 , y0 , z 0 ) is introduced on two accounts: (1) the use of (x 0 , y0 , z 0 )
sometimes introduces a profusion of additional subscripts, and (2) the notation agrees with that traditional in
continuum mechanics for the material coordinates as noted in the previous remark. The identification X ≡ x0 ,
Y ≡ y0 , Z ≡ z 0 will be employed when it is convenient to consider the reference configuration as the initial target
configuration (cf. Remark 7.1).
§7.3.4. Configurations and Staged Analysis
The meaning of some special configurations can be made more precise if the nonlinear analysis process is viewed
as a sequence of analysis stages, as discussed in Chapter 3. We restrict attention to the Total Lagrangian (TL) and
Corotational (CR) kinematic descriptions, which are the only ones covered in this course. In a staged TL nonlinear
analysis, two common choices for the reference configuration are:
(1) Reference ≡ base. The base configuration is maintained as reference configuration for all stages.
(2) Reference ≡ stage start. The configuration at the start of an analysis stage, i.e. at λ = 0, is chosen as reference
configuration.
A combination of these two strategies can be of course adopted. In a staged CR analysis the reference is split
between base and corotated. The same update choices are available for the base. This may be necessary when
rotations exceed 2π; for example in aircraft maneuvers.
The admissible configuration is a “catch all” concept that embodies all others as particular cases. The perturbed
configuration is an admissible variation from a admissible configuration. An ensemble of perturbed configurations
is used to establish incremental or rate equations.
The iterated and target configurations are introduced in the context of incremental-iterative solution procedures
for numerically tracing equilibrium paths. The target configuration is the “next solution”. More precisely, an
equilibrium solution (assumed to exist) which satisfies the residual equations for a certain value of the stage control
parameter λ. While working to reach the target, a typical solution process goes through a sequence of iterated
configurations that are not in equilibrium.
The corotated configuration is a rigid-body rotation of the reference configuration that “follows” the current
configuration like a “shadow”. It is used in the corotational (CR) kinematic description of nonlinear finite elements.
Strains measured with respect to the corotated configuration may be considered “small” in many applications, a
circumstance that allows linearization of several relations and efficient treatment of stability conditions.
§7.4. Kinematics
This section cover the essential kinematics necessary for finite displacement analysis.
§7.4.1. Deformation and Displacement Gradients
The derivatives of (x, y, z) with respect to (X, Y, Z ), arranged in Jacobian format, constitute the so-
called deformation gradient matrix:
∂x ∂x ∂x
∂ X ∂Y ∂Z
∂(x, y, z) ∂y ∂y ∂y
F= = . (7.5)
∂(X, Y, Z ) ∂ X ∂Y ∂Z
∂z ∂z ∂z
∂ X ∂Y ∂Z
7–8
7–9 §7.4 KINEMATICS
The inverse relation gives the derivatives of (X, Y, Z ) with respect to (x, y, z) as
∂X ∂X ∂X
∂x ∂y ∂z
∂(X, Y, Z )
F−1 = = ∂Y ∂Y ∂Y
. (7.6)
∂(x, y, z) ∂x ∂y ∂z
∂Z ∂Z ∂Z
∂x ∂y ∂z
These matrices can be used to relate the coordinate differentials
dx dX
dx = dy = F dY = F dX, dX = F−1 dx. (7.7)
dz dZ
Similarly, the displacement gradients with respect to the reference configuration can be presented as
the 3 × 3 matrix
∂x ∂x ∂ x ∂u X ∂u X ∂u X
∂X − 1 ∂Y ∂Z ∂X ∂Y ∂Z
∂y ∂y ∂y
G=F−I= = ∂u Y ∂u Y ∂u Y = ∇u. (7.8)
∂X ∂Y − 1 ∂Z ∂X ∂Y ∂Z
∂z ∂z ∂z ∂u Z ∂u Z ∂u Z
∂X ∂Y ∂Z − 1 ∂X ∂Y ∂Z
Displacement gradients with respect to the current configuration are given by
∂u X ∂u X ∂u X
1 − ∂∂ X
x
∂X
∂y
∂Z
∂x ∂x ∂y ∂z
∂Y 1 − ∂Y ∂Y ∂u Y ∂u Y ∂u Y
J = I − F−1 = ∂x ∂y ∂z = ∂x ∂y ∂z . (7.9)
∂Z ∂Z 1 − ∂∂zZ ∂u Z ∂u Z ∂u Z
∂x ∂y ∂x ∂y ∂z
For the treatment of the Total Lagrangian description it will found to be convenient to arrange the
displacement gradients of (7.8) as a 9-component vector (printed as row vector to save space):
gT = [ g1 g2 g3 g4 g5 g6 g7 g8 g9 ]
∂u Z . (7.10)
= ∂u X ∂u Y ∂u Z ∂u X ∂u Y ∂u Z ∂u X ∂u Y
∂X ∂X ∂X ∂Y ∂Y ∂Y ∂Z ∂Z ∂Z
Remark 7.6. For arbitrary rigid-body motions (motions without deformations) FT F = FFT = I, that is, F is an
orthogonal matrix.
Remark 7.8. In nonlinear continuum mechanics, displacement gradients play an important role that is absent in
the infinitesimal theory. This is even more so in the Total-Lagrangian core-congruential formulation covered in
Chapters 8–11.
7–9
Chapter 7: REVIEW OF CONTINUUM MECHANICS 7–10
Remark 7.9. The ratio between infinitesimal volume elements d V = d x d y dz and d V0 = d X dY d Z in the current
and reference configuration appears in several continuum mechanics relations. Because of (7.7) this ratio may be
expressed as
dV ρ0
= = det F, (7.12)
d V0 ρ
where ρ and ρ0 denote the mass densities in the current and reference configuration, respectively. This equation
expresses the law of conservation of mass.
Tensors F and G are the building blocks of various deformation measures used in nonlinear continuum
mechanics. The whole subject is dominated by the polar decomposition theorem: any particle defor-
mation can be expressed as a pure deformation followed by a rotation, or by a rotation followed by a
pure deformation. Mathematically this is written as multiplicative decompositions:
F = RU = VR. (7.13)
Here R is an orthogonal rotation tensor, whereas U and V are symmetric positive definite matrices
√ rotation, U = V = I.
called the right and left stretch tensors, respectively. If the deformation is a pure
Premultiplying (7.13) by FT = URT gives U2 = FT F and consequently√ U = FT F. Postmultiplying
(7.13) by FT = RT V gives V2 = FFT and consequently V = FFT . Upon taking the square roots, the
rotation is then computed as either R = FU−1 or R = V−1 F. Obviously U = RVRT and V = RT UR.
The combinations C R = FT F and C L = FFT are symmetric positive definite matrices that are called the
right and left Cauchy-Green strecht tensors, respectively. To get U and V as square roots it is necessary
to solve the eigensystem of C R and C L , respectively.
To convert a stretch tensor to a strain tensor one substracts I from it or takes its log, so as to have a
measure that vanishes for rigid motions. Either U − I or V − I represent proper strain measures. These
are difficult, however, to express analytically in terms of the displacement gradients because of the
intermediate eigenproblem. A more convenient strain measure is described next.
A convenient finite strain measure is the Green-Lagrange4 strain tensor. Its three-dimensional expression
in Cartesian coordinates is
eX X eX Y eX Z
e= 1
2
FT F − I = 12 (G + GT ) + 12 GT G = eY X eY Y eY Z , (7.14)
eZ X eZ Y eZ Z
4 A more proper name would be Green-St.Venant strain tensor. In fact Lagrange never used it but his name appears because of
its strong connection to the Lagrangian kinematic description. Many authors call this measure simply the Green strain tensor.
7–10
7–11 §7.4 KINEMATICS
Remark 7.10. Several other finite strain measures are used in nonlinear continuum mechanics. The common
characteristic of all measures is that they must predict zero strains for arbitrary rigid-body motions, and must
reduce to the infinitesimal strains if the nonlinear terms are neglected. This topic is further explored in Exercise
7.5.
7–11
Chapter 7: REVIEW OF CONTINUUM MECHANICS 7–12
ei = hiT g + 12 gT Hi g, (7.18)
where hi are sparse 9 × 1 vectors:
1 0 0 0 0 0
0
0
0 0 0 1
0 0 0 0 1 0
0 0 0 0 0 1
h1 = 0 , h2 = 1 , h3 = 0 , h4 = 0 , h5 = 0 , h6 = 0 , (7.19)
0 0 0 1 0 0
0 0 0 0 1 0
0 0 0 1 0 0
0 0 1 0 0 0
and Hi are very sparse 9 × 9 symmetric matrices:
1 0 0 0 0 0 0 0 0
0 1 0 0 0 0 0 0 0
0 0 1 0 0 0 0 0 0
0 0 0 0 0 0 0 0 0
H1 = 0 0 0 0 0 0 0 0 0, etc. (7.20)
0 0 0 0 0 0 0 0 0
0 0 0 0 0 0 0 0 0
0 0 0 0 0 0 0 0 0
0 0 0 0 0 0 0 0 0
Remark 7.11. For strain measures other than Green-Lagrange’s, expressions similar to (7.14) may be constructed.
But although the hi remain the same, the Hi become complicated functions of the displacement gradients.
7–12
7–13 §7.6 CONSTITUTIVE EQUATIONS
~
Y ~ Deformed
X
Corotated _
θB y
Base
θ
Y
_
Globally X x
aligned
Figure 7.2. Rigid motion of bar in 2D illustrating concept of globally aligned configuration.
Remark 7.12. The physical meaning of the PK2 stresses is as follows: si j are stresses “pulled back” to the reference
configuration C 0 and referred to area elements there.
Remark 7.13. The PK2 stresses are related to the Cauchy (true) stresses σi j through the transformation
∂X ∂X ∂X ∂X ∂X ∂X ∂X ∂X ∂X ∂X ∂X ∂X
∂x ∂x ∂y ∂y ∂z ∂z ∂ y ∂z ∂z ∂ x ∂x ∂y
∂Y ∂Y ∂Y ∂Y ∂Y ∂Y ∂Y ∂Y ∂Y ∂Y ∂Y ∂Y
s
∂x ∂x ∂y ∂y ∂z ∂z ∂ y ∂z ∂z ∂ x ∂ x ∂ y σX X
XX
sY Y ∂ Z ∂ Z ∂ Z ∂ Z ∂ Z ∂ Z ∂ Z ∂ Z ∂ Z ∂ Z ∂ Z ∂ Z σY Y
s Z Z ρ0 ∂ x ∂ x ∂ y ∂ y ∂z ∂z ∂ y ∂z ∂z ∂ x ∂x ∂y
σZ Z
sY Z = ρ ∂Y ∂ Z ∂Y ∂ Z ∂Y ∂ Z ∂Y ∂ Z ∂Y ∂ Z ∂Y ∂ Z σY Z , (7.23)
∂x ∂x ∂y ∂y ∂z ∂z ∂ y ∂z ∂z ∂ x
∂ x ∂ y σZ X
sZ X
∂Z ∂X ∂Z ∂X ∂Z ∂X ∂Z ∂X ∂Z ∂X ∂Z ∂X
sX Y σX Y
∂ x ∂ x ∂ y ∂ y ∂z ∂z ∂ y ∂z ∂z ∂ x ∂x ∂y
∂ X ∂Y ∂ X ∂Y ∂ X ∂Y ∂ X ∂Y ∂ X ∂Y ∂ X ∂Y
∂x ∂x ∂y ∂y ∂z ∂z ∂ y ∂z ∂z ∂ x ∂x ∂y
∂x ∂x ∂x ∂x ∂x ∂x ∂x ∂x ∂x ∂x ∂x ∂x
∂ X ∂ X ∂Y ∂Y ∂ Z ∂ Z ∂Y ∂ Z ∂ Z ∂ X ∂ X ∂Y
σ ∂y ∂y ∂y ∂y ∂y ∂y ∂y ∂y ∂y ∂y ∂y ∂y
XX ∂ X ∂ X ∂Y ∂Y ∂ Z ∂ Z ∂Y ∂ Z ∂ Z ∂ X ∂ X ∂Y s X X
σY Y ∂z ∂z ∂z ∂z ∂z ∂z ∂z ∂z ∂z ∂z ∂z ∂z sY Y
σZ Z ρ ∂ X ∂ X ∂Y ∂Y ∂ Z ∂ Z ∂Y ∂ Z ∂ Z ∂ X ∂ X ∂Y
= sZ Z .
σY Z ρ0 ∂ y ∂z ∂ y ∂z ∂ y ∂z ∂ y ∂z ∂ y ∂z ∂ y ∂z
YZ
s
(7.24)
∂ X ∂ X ∂Y ∂Y ∂ Z ∂ Z ∂Y ∂ Z ∂ Z ∂ X ∂ X ∂Y s Z X
σZ X
σX Y ∂z ∂ x ∂z ∂ x ∂z ∂ x ∂z ∂ x ∂z ∂ x ∂z ∂ x s X Y
∂ X ∂ X ∂Y ∂Y ∂ Z ∂ Z ∂Y ∂ Z ∂ Z ∂ X ∂ X ∂Y
∂x ∂y ∂x ∂y ∂x ∂y ∂x ∂y ∂x ∂y ∂x ∂y
∂ X ∂ X ∂Y ∂Y ∂ Z ∂ Z ∂Y ∂ Z ∂ Z ∂ X ∂ X ∂Y
The density ratios that appears in these equations may be obtained from (7.13). If all displacement gradients are
small, both transformations reduce to the identity, and the PK2 and Cauchy stresses coalesce.
7–13
Chapter 7: REVIEW OF CONTINUUM MECHANICS 7–14
where ei and si denote components of the strain and stress vectors defined by (7.16) and (7.22), respec-
tively, si0 are stresses in the reference configuration (also called prestresses) and E i j are constant elastic
moduli with E i j = E ji . In full matrix notation,
0
s1 s1 E 11 E 12 E 13 E 14 E 15 E 16 e1
s2 s2 E 12
0
E 22 E 23 E 24 E 25 E 26 e2
0
s3 s3 E 13 E 23 E 33 E 34 E 35 E 36 e3
= 0+ , (7.26)
s4 s4 E 14 E 24 E 34 E 44 E 45 E 46 e4
0
s5 s5 E 15 E 25 E 35 E 45 E 55 E 56 e5
s6 s60 E 16 E 26 E 36 E 46 E 56 E 66 e6
or in compact form,
s = s0 + Ee. (7.27)
Remark 7.14. For an invariant reference configuration, PK2 and Cauchy (true) prestresses obviously coincide (see
Remark 7.14). Thus σ0 ≡ s0 in such a case. However if the reference configuration is allowed to vary often, as in
the UL description, things get more complicated.
We conclude this review by giving the expression of the strain energy density U in the current
configuration reckoned per unit volume of the reference configuration:
U = si0 (hiT g + gT Hi g) + 1
2
(gT hi + 12 gT Hi g)E i j (hTj g + 12 gT H j g) . (7.30)
Because hi and Hi are constant, this relation shows that the strain energy density is quartic in the
displacement gradients collected in g.
The strain energy in the current configuration is obtained by integrating this energy density over the
reference configuration:
U= U d X dY d Z . (7.31)
V0
This expression forms the basis for deriving finite elements based on the Total Lagrangian description.
7–14
7–15 Exercises
EXERCISE 7.2 [A:15] Derive (7.28) by integrating si dei from C0 (ei = 0) to C (ei = ei ) and using (7.25).
EXERCISE 7.3 [A:20] A bar of length L 0 originally along the X ≡ x axis (the reference configuration C0 ) is
rigidly rotated 90◦ to lie along the Y ≡ y axis while retaining the same length (the current configuration C). Node
1 at the origin X = Y = 0 stays at the same location.
(a) Verify that the motion from C0 to C is given by
x = −Y, y = X, z = Z. (E7.1)
(b) Obtain the displacement field u, the deformation gradient matrix F, the displacement gradient matrix G and
the Green-Lagrange axial strain e = e X X . Show that the Green-Lagrange measure correctly predicts zero
axial strain whereas the infinitesimal strain measure = X X = ∂u X /∂ X predicts the absurd value of −100%
strain.
EXERCISE 7.4 [A:20] Let L 0 and L denote the length of a bar element in the reference and current configurations,
respectively. The Green-Lagrange finite strain e = e X X , if constant over the bar, can be defined as
L 2 − L 20
e= . (E7.2)
2L 20
Show that the definitions (E7.2) and of e = e X X in (7.15) are equivalent. (Hint: express L 0 and L in terms of the
coordinates and displacements in the bar system.)
EXERCISE 7.5 [A:25] The Green-Lagrange strain measure is not the only finite strain measure used in structural
and solid mechanics. For the uniaxial case of a stretched bar that moves from a length L 0 in C0 to a length L in C,
some of the other measures are defined as follows:
(a) Uniaxial Almansi strain:
L 2 − L 20
eA = . (E7.3)
2L 2
(b) Uniaxial Hencky strain, also called logarithmic or “true” strain:
e H = log(L/L 0 ), (E7.4)
where log denotes the natural logarithm.
(c) Uniaxial midpoint strain5
L 2 − L 20
eM = . (E7.5)
2[(L + L 0 )/2]2
(d) Uniaxial engineering strain:
L − L0
eE = = . (E7.6)
L0
If L = (1 + )L 0 , show by expanding e A , e H and e M in Taylor series in (about = 0) that these measures, as
well as the Green-Lagrange axial strain (E7.2), agree with each other to first order [i.e., they differ by O( 2 )] as
→ 0.
5 The midpoint strain tensor, which is a good approximation of the Hencky strain tensor but more easily computable, is frequently
used in finite element plasticity or viscoplasticity calculations that involve large deformations, for example in metal forming
processes.
7–15
Chapter 7: REVIEW OF CONTINUUM MECHANICS 7–16
EXERCISE 7.6 [A:30] (Advanced). Extend the definition of the Almansi, Hencky, midpoint and midpoint strains
to a three dimensional strain state. Hint: use the spectral decomposition of FT F and the concept of function of a
symmetric matrix.
EXERCISE 7.7 [A:35] (Advanced). Extend the definition of engineering strain to a three-dimensional strain
state. The resulting measures (there are actually two) are called the stretch tensors. Hint: use either the spectral
decomposition of FT F, or the polar decomposition theorem of tensor calculus.
EXERCISE 7.8 [A:40] (Advanced). Define the stress measures conjugate to the Almansi, Hencky, midpoint and
engineering strains.
7–16
6
.
Conservative
Systems
6–1
Chapter 6: CONSERVATIVE SYSTEMS 6–2
TABLE OF CONTENTS
Page
§6.1. Introduction 6–3
§6.2. The Load Potential 6–3
§6.2.1. Concentrated Dead Loads . . . . . . . . . . . . . 6–4
§6.2.2. Distributed Dead Loads . . . . . . . . . . . . . 6–5
§6.3. The Internal Energy: A Linear Spring 6–6
§6.4. The Internal Energy: How Geometric Nonlinearities Arise 6–6
§6.5. Internal Energy: Additivity Property 6–8
§6. Exercises . . . . . . . . . . . . . . . . . . . . . . 6–9
6–2
6–3 §6.2 THE LOAD POTENTIAL
§6.1. Introduction
It was noted in previous Chapters that if a structure and the forces acting upon it collectively form a
conservative mechanical system, the residual force vector r may be expressed as the gradient of the
total potential energy with respect to the state vector:
∂
r= . (6.1)
∂u
∂U ∂P
p= , f= . (6.2)
∂u ∂u
where p and f are the internal and external forces, respectively, U is the internal energy — which
reduces to the strain energy in the problems considered in this course — and P is the potential of the
applied loads, the negative of which is called the external work function W .
The force equilibrium equations r = 0 or f = p express the fact that the total potential energy is
stationary with respect to variations of the state vector when the structure is in static equilibrium.
Mathematically:
∂ T
δ = r δu =
T
δu = 0. (6.3)
∂u
where δu denotes a virtual displacement, δ being the variation symbol. Since δu is arbitrary, (6.3)
implies that r = 0.
If the structural system is conservative there are substantial advantages in taking advantage of that
property:
(1) If discrete force equilibrium equations are worked out by hand (either for complete structures
or finite elements) derivation from a potential is usually simpler than direct use of equilibrium,
because differentiation is a straightforward and less error prone operation, especially as regards
signs. Exercise 6.3 gives an example of this.
(2) The transformation of residual equations to different coordinate systems is simplified because of
the invariance properties of energy functions.
(3) The conventional finite element discretization method relies on the availability of an internal
energy functional.
(4) The tangent stiffness matrix is symmetric. Consequently equation solvers (and eigensolvers) can
take advantage of this property.
(5) Loss of stability can be assessed by the singular stiffness criterion, which is static in nature. If the
system is nonconservative, loss of stability may have to be tested by a dynamic criterion, which
is always more difficult and computationally expensive.
This Chapter introduces the concepts of internal and external potential for systems with finite degrees of
freedom. The presentation is not general in nature but relies on a few simple examples complemented
with exercises. The material is intended to serve as a “bridge” to the formulation of geometrically
nonlinear finite elements, which starts in the next Chapter.
6–3
Chapter 6: CONSERVATIVE SYSTEMS 6–4
Deformed
;
u
;;; ;
F
Undeformed
;;; No change in load
magnitude or direction
The concept of load potential is the easiest to understand. This function, called P, is the potential of
the work done by the applied or prescribed forces working on the displacements of the points on which
those forces act. The negative of this potential W = −P is called the work function, but this function
will not be used in the present course.
In general these forces will be defined by their three components along the axes x, y, z and are more
properly represented by vectors fk . For example, if at location k = 3 we have a force F3 acting in the
y-direction,
0
f3 = F3 . (6.6)
0
Likewise, the displacement of points of application of fk is denoted by vector uk . The vector general-
ization of (6.5) is then the sum of n inner products:
n
P= fkT uk . (6.7)
k=1
6–4
6–5 §6.2 THE LOAD POTENTIAL
f y (x)
Deformed A B
u y (x)
y
;
;;; ;
x
Undeformed
;;;
Figure 6.2. Structure under under line load f y (x)
(directed upward) over segment AB.
Finally, if all applied force components are collected in the external force vector f (augmented with
zero entries as necessary to be in one-to-one correspondence with the state vector u) then we have the
compact inner-product expression
P = fT u. (6.8)
where u y (x) is the y-displacement component of points on segment (A,B). A similar technique can
be used for volume (body) forces as illustrated in Exercise 6.1.
6–5
Chapter 6: CONSERVATIVE SYSTEMS 6–6
; k δ
;
x
1 2
Remark 6.1. Substantial mathematical complications arise if some forces are functions of the displacements.
For example, in slender structures under aerodynamic pressure loads the change of direction of the forces as the
structure deflects may have to be considered in the stability analysis. These so-called “follower” forces, which
introduce force B.C. nonlinearities, are considered later in the course. Suffices to say here that no loads potential
P generally exist in such cases and the system is nonconservative.
The internal energy, called U , is the recoverable mechanical work “stored” in the material of the
structure by virtue of its elastic deformation. When this work is expressed in terms of strains and
stresses, as in following Chapters, it is called the strain energy. Note that only flexible bodies can store
strain energy; a rigid body cannot.
We shall illustrate the internal energy concept here by considering the simplest of all structural elements
already encountered in linear finite element analysis: a linear spring of stiffness k, illustrated in
Figure 6.3. Generalization to more complicated structures and structural components will be made in
subsequent Chapters.
If the spring is undeformed, its internal energy U can be conventionally taken as zero (because an
energy function can be adjusted by an arbitrary constant without changing its gradients). Now let the
spring deform slowly (to avoid inertial effects) such that its two ends separate by a distance δ called
the elongation. The internal spring force f¯ for an intermediate elongation 0 ≤ δ̄ ≤ δ is f¯ = k δ̄. An
elementary result of mechanics is that the strain energy taken up by the spring in its deformed state is
δ δ
U= spring-force × d(elongation) = (k δ̄) d δ̄ = 12 kδ 2 . (6.10)
0 0
Suppose that the spring is fixed at end 1 and that end 2 can move only along the x axis, as in Figure
6.3. Call u the x displacement of end 2. Then δ = u − 0 = u and the strain energy is U = 12 ku 2 .
According to (6.2) the internal force, which in this case is just the spring axial force p, is the derivative
of U with respect to u:
∂U
p= = ku. (6.11)
∂u
This is linear in the displacement u so nothing has changed so far with respect to linear finite element
analysis.
6–6
6–7 §6.4 THE INTERNAL ENERGY: HOW GEOMETRIC NONLINEARITIES ARISE
Deformed
uy2
y k
x
ux2
Undeformed
2(x2 ,y2)
Now suppose that the spring can move arbitrarily on the plane x, y, as depicted in Figure 6.4. The
position of the deformed spring is completely defined by the four displacement components u x1 , u y1 ,
u x2 and u y2 , which we collect in the state vector
u x1
u
u = y1 . (6.12)
u x2
u y2
Let and d denote the spring lengths in the undeformed and deformed configurations, respectively.
The elongation δ is given by
δ = d − = ( x +
x ) + ( y +
y ) − 2x + 2y ,
2 2 (6.13)
where
x = u x2 − u x1 ,
y = u y2 − u y1 , x = x2 − x1 , y = y2 − y1 , in which x1 , y1 , x2 and y2 denote
the x, y coordinates of the end nodes of the undeformed spring. Consequently
U = 12 kδ 2 = 12 k( 2 + 2d − 2 d )
(6.14)
= 12 k (2 2 + 2 x
x +
2x + 2 y
y +
2y − 2 ( x +
x )2 + ( y +
y )2 .
6–7
Chapter 6: CONSERVATIVE SYSTEMS 6–8
Even for this simple case the exact nonlinear equations are quite nasty, involving irrational functions
of the displacements. The second property, however, shows that approximations to the exact nonlinear
equations may be made when the change in direction is “small” in some sense. This feature is illustrated
in Exercise 6.3.
If the structure consists of m linear springs, each of which absorbs an internal energy Uk , the total
internal energy is the sum of the individual spring energies:
U = U1 + U2 + . . . + Um . (6.16)
This additivity property is of course general because energies are scalar quantities. It applies to arbitrary
structures decomposed into structural components such as finite elements. Furthermore, (6.16) is not
affected by whether the structure is linear or nonlinear.
The last property explains why finite element equations should be derived from energy functions if
such functions exist. That is not, however, always possible.
6–8
6–9 Exercises
EXERCISE 6.1 [A:15] A body of volume V and density ρ is in an uniform gravity field g acting along the
−z axis. The body displaces to another position defined by the small-displacement field u(x, y, z). Find the
expression of the load potential P as an integral over the body if the change in shape of the body is negligible.
EXERCISE 6.2 [A:20] Work out the expression of the internal forces for (6.15). Then extend this relation to the
three-dimensional case in which the ends of the spring move by u x1 , u y1 , u z1 , u x2 , u y2 , u z2 in the x, y, z space.
EXERCISE 6.3 [A+N/C:30] Consider the shallow arch model shown in Figure E6.1. This consists of two
identical linear springs of axial stiffness k pinned to each other and to unmoving pinned supports as shown. The
springs are assumed able to resist both tensile and compressive forces. The distance between the supports is 2L.
The undeformed springs form an angle α with the horizontal axis.
The central pin in loaded by a dead vertical force of magnitude f , positive downwards, which is parametrized
as f = λk L. Only symmetrical deformations of the arch are to be considered for this Exercise. Consequently
the system has just one degree of freedom which we take to be the displacement u under the load, also positive
downwards. The response of this system exhibits the snap-through behavior sketched in Figure E6.2.
(a) Show that the internal energy U and load potential P of the two-spring system are given by
2
1 1
U = kL 2
− , P = f u, (E6.1)
cos α cos θ
where θ is the angle shown in Figure E6.1, which is linked to u by the relation tan θ + u/L = tan α.
(c) Derive the exact equation for the limit load parameters
∂λ(µ)
= 0. (E6.5)
∂µ µ=µ
L ,λ=λ L
(Hint: the exact equation in terms of the angular coordinate θ is cos3 θ L = cos α). Solve this trigonometric
equation1 for the limit-load parameters λ L1 and λ L2 and the dimensionless displacements µ L1 and µ L2 at
those points assuming that α = 30◦ .
1 Equation (E6.5) is equivalent to det K = 0 because for a one-DOF system det K = K = ∂λ/∂µ.
6–9
Chapter 6: CONSERVATIVE SYSTEMS 6–10
u k
k
;; ;;
θ
α
L L
L1
u or µ
L2
(d) If the arch initially is and remains sufficiently “shallow” throughout its snap-through behavior, we may make
the small-angle approximations,
Recast the energy, equilibrium equations, and limit load equations in terms of these approximations, obtaining
U as a quartic polynomial in θ , r as a cubic polynomial in θ, etc, then replace in terms of µ. As a check,
the residual equation in terms of λ and µ should be given by (4.16). Calculate the limit load parameters λ L1
and λ L2 , and the dimensionless displacements
√ µ L1 and µ L2 at those loads. Verify that these displacements
correspond to the angles θ L = ±α/ 3.
(e) Draw the control-state response curves r (µ, λ) = 0, derived using the exact nonlinear equations and those
from the small-angle approximations on the λ, µ plane (as in the sketch of Figure E6.2, going up to µ ≈ 2.5)
for α = 30◦ .
EXERCISE 6.4 [A+N:15] Derive the current stiffness parameter κ defined in Equation (5.8) for the approximate
(small-angle) model of the two-spring arch of Exercise 6.3. Plot the variation of κ(µ) as µ varies from 0 to µ L1
at the first limit point, with µ along the horizontal axis. Does κ vanish at the limit point?
6–10
5
.
Critical Points
and Related
Properties
5–1
Chapter 5: CRITICAL POINTS AND RELATED PROPERTIES 5–2
TABLE OF CONTENTS
Page
§5.1. Classification of Critical Points 5–3
§5.2. Limit Point Sensors 5–6
§5.3. *Turning Points 5–6
§5.4. *Derivatives of Energy Functions 5–6
§5.5. *Energy Increments 5–8
§5. Exercises . . . . . . . . . . . . . . . . . . . . . . 5–10
5–2
5–3 §5.1 CLASSIFICATION OF CRITICAL POINTS
This Chapter provides additional material on properties of the one-parameter force residual equa-
tions. It begins with a study of critical points, which are classified into limit and bifurcation points.
Limit point “sensors” and turning points are briefly described. The section concludes with some
mathematical derivations for conservative systems, which will be of use in Chapters dealing with
incremental solution methods.
Throughout this Chapter it is assumed that the structural system is conservative1 and consequently
K is symmetric.
Response points at which K becomes singular are of great interest in the applications because of
their intimate connection to structural stability. These are called critical points, and also nonregular
or singular points. At these points the velocity vector v is not uniquely determined by q from (4.8).
Physically this means that the structural behavior cannot be controlled by the parameter λ.
It is convenient to distinguish the following types of critical points:
1. Isolated limit points, at which the tangent (4.17) to the equilibrium path is unique but normal
to the λ axis so v becomes infinitely large.
2. Multiple limit points, at which there the tangent lies in the null space of K and is not unique
but still normal to the λ axis.
3. Isolated bifurcation points, also called branch points or branching points, from which two
equilibrium path branches emanate and so there is no unique tangent. The rank deficiency of
K is one.
4. Multiple bifurcation points, from which more than two equilibrium path branches emanate.
The rank deficiency of K is two or greater.
A critical point that is both a limit and a bifurcation point is classified as a multiple bifurcation
point.
Figures 5.1, 5.2 and 5.3 illustrate isolated limit points (identified as L 1 , L 2 , . . .) and bifurcation
points (identified as B1 , B2 , . . .).
To classify critical points we proceed as follows. Let z be a null right eigenvector of K at a critical
point, that is,
Kz = 0. (5.1)
Since K is assumed symmetric, zT K = 0; that is, z is also a left null eigenvector. The parametric
differential equation of the equilibrium path is Ku̇ = q λ̇, which multiplied through by dt becomes
K du = q dλ. (5.2)
zT q dλ = 0. (5.3)
1 A property elaborated upon in Chapter 6. This property is important in that a symmetric K is guaranteed to have a full
set of eigenvectors. Furthermore left and right eigenvectors coalesce.
5–3
Chapter 5: CRITICAL POINTS AND RELATED PROPERTIES 5–4
λ L1
B
L2
u1
Figure 5.1. Limit points L 1 and L 2 and bifurcation point (B) for a two degree of freedom system
(u 1 , u 2 ) shown on the u 1 versus λ plane. Limit point (“snap through” behavior) occurs
before bifurcation. Full lines represent physically “preferred” paths.
then dλ must vanish, and we have a limit point. The point is isolated if z is the only null eigenvector
and multiple otherwise.
On the other hand, if
zT q = 0, (5.5)
then we have a bifurcation or branching point. The point is isolated if z is the only null eigenvector
and multiple otherwise. The key physical characteristic of a bifurcation point is an abrupt transition
from one deformation mode to another mode; the latter having been previously “concealed” by virtue
of being orthogonal to the incremental load vector.
Remark 5.1. If K is not symmetric, several changes must be made in the previous assumptions and derivations.
These are explained in the Chapters that deal with nonconservative systems (29-30) and the possible loss of
stability by growing dynamic oscillations (flutter).
Remark 5.2. If λ is an applied load multiplier, a limit point such as L 1 in Figure 5.1 is called a snap-through
point because the structure “snaps” dynamically to another equilibrium position. The term collapse applies to
critical points beyond which the structure becomes useless.
Remark 5.3. As an isolated limit point is approached, v tends to become parallel to z whereas its magnitude
goes to ∞; that is
v
→ z. (5.6)
|v|
Consequently the normalized v may be a good eigenvector estimate if K has been factored near the limit point.
(This is nothing more that a statement of the well known inverse iteration process for finding eigenvectors.)
5–4
5–5 §5.1 CLASSIFICATION OF CRITICAL POINTS
λ L1
B1
B2
L2
u1
Figure 5.2. Similar to Figure 5.1, but here the bifurcation point B1 implying “buckling”
behavior occurs before the limit point L 1 , which is physically unreachable.
A more realistic three-dimensional view of this case is shown in Figure 5.3.
λ L1
B1
B2
L2
u2 u1
Figure 5.3. The equilibrium path of Figure 5.2 shown in the 3D space (u 1 , u 2 , λ).
This is the type of response exhibited by a uniformly pressurized deep
arch, for which u 1 and u 2 are amplitudes of the symmetric and antisymmetric
deformation mode, respectively, and λ is a pressure multiplier.
Remark 5.4. The set of control parameters for which det K = 0 while r(u, λ) = 0 is sometimes called the
bifurcation set in the mathematical literature. The name is misleading, however, in that the set may include
limit points; the name critical set would be more appropriate.
Remark 5.5. Showing bifurcation points on the λ versus u plane as in Figures 5.1 and 5.2 may be misleading, as
it conceals the phenomenon of transition from one mode of deformation to another. A more realistic picture is
provided in Figure 5.3, which shows the classical bifurcation behavior for a symmetrically loaded shallow arch;
here u 1 and u 2 measure amplitude of symmetric and antisymmetric displacement shapes, respectively. At B1
the arch, which had been previously deforming symmetrically, takes off along an antisymmetric deformation
mode; at B2 the latter disappears and the arch rejoin the symmetric path.
5–5
Chapter 5: CRITICAL POINTS AND RELATED PROPERTIES 5–6
Remark 5.6. Physically the distinction between the two types of critical points is not so marked, inasmuch as
imperfect structures display limit-point behavior. A bifurcation point may be viewed as the limit of a sequence
of progressively sharper limit points realized as the structure strives towards mathematical perfection.
Remark 5.7. For a readable introduction to elastic structural stability along the lines of classical perturbation
theory, the monographs by Thompson and Hunt [181,183] are still unsurpassed. The treatise by Knops and
Wilkes [100] goes deeper into various mathematical questions, but is computationally useless. Brush and
Almroth [37] give more information on computational methods. The survey and book by Bushnell [40] has
more physics (e.g., temperature, plasticity and creep effects) and a wider selection of practical problems. The
connection between potential-based structural stability and modern catastrophe theory is presented in a highly
readable manner by Poston and Steward [141] and Thompson [182].
Scalar estimates of the overall stiffness of the structure as the control parameter varies are useful
as limit points sensors. The following estimator is based on the Rayleigh quotient approximation
to the fundamental eigenvalue of K:
xT Kx
kx = T , (5.7)
x x
where x is an arbitrary nonnull vector, and K is evaluated at an equilibrium position u(λ). An
“equilibrium-path stiffness” estimator is obtained by taking x to be v = K−1 q, in which case
qT v
k = kv = . (5.8)
vT v
This value of course depends on λ. It is convenient in practice to work with the dimensionless ratio
κ = k(λ)/k(0), (5.9)
which takes the value 1 at the start of an analysis stage, and goes to zero as a limit point is approached.
A stiffness estimator with this behavior (although computed in a different way) was introduced by
Bergan and coworkers under the name current stiffness parameter. It should be noted, however,
that no estimator of this type can reliably predict the occurrence of a bifurcation point. Sensors for
such points are described later in the context of augmented equations.
§5.3. *Turning Points
Turning points are regular points at which the tangent is parallel to the λ axis so that v = 0. The unit tangent
takes the form
0
tu = (5.10)
±1
Although these points generally do not have physical meaning, they can cause special problems in path-
following solution procedures because of “turnback” effects.
To detect the vicinity of a turning point one can check the two mathematical conditions: v becomes orthogonal
to q and u tends to zero faster than q. For example:
where κ is the current stiffness parameter. Typical values may be δ = 0.01, κmin = 100.
5–6
5–7 §5.4 *DERIVATIVES OF ENERGY FUNCTIONS
If the residual r(u, λ) is derivable from a total potential energy function (u, λ) as in (3.2), then the stiff-
ness matrix and incremental load vector appear naturally as components of the following matrix of second
derivatives:
∂ 2 ∂ 2
∂u∂u ∂u ∂λ K −q
2 T = (5.12)
∂ ∂ 2 −qT a
∂λ∂u ∂λ∂λ
where a = ∂ 2 /∂λ2 has not been introduced previously. Obviously the tangent stiffness matrix K (the Hessian
of ) is now symmetric. Note also that
∂q ∂ 3 ∂ ∂ 2 ∂K
= = = = Kλ , (5.13)
∂u ∂u ∂λ∂u ∂λ ∂u∂u ∂λ
is a symmetric matrix.
The complementary energy function ∗ may be defined from the dual Legendre transformation (see e.g.,
Chapter 2.5 of Sewell’s book [163]) as
∂
+ ∗ = u i = uT r = rT u. (5.14)
∂u i
This gives ∗ (r, λ) = rT u − with u eliminated from r(u, λ) = 0, so now the residual forces are the active
variables. Obviously
∂∗ ∂∗
u= , or u i = . (5.15)
∂r ∂ri
The matrix of second derivatives of ∗ is
∂ 2 ∗ ∂ 2 ∗
∂r∂r ∂r∂λ
F v
T 2 ∗
= T (5.16)
∂ 2 ∗ ∂ v b
∂λ∂r ∂λ∂λ
These are linked to the quantities that appear in (5.12) by the matrix relations
The tangent flexibility matrix F = K−1 (the Hessian of K ) is now symmetric. Note also that
∂u ∂ 3 ∗ ∂ ∂ 2 ∗ ∂F
= = = = Fλ , (5.19)
∂r ∂r ∂λ∂r ∂λ ∂r∂r ∂λ
is a symmetric matrix.
5–7
Chapter 5: CRITICAL POINTS AND RELATED PROPERTIES 5–8
Remark 5.8. The following matrix appears (as amplification matrix) in the study of the stability of incremental methods:
∂v ∂(Fq) ∂F ∂q ∂F ∂K
A= = = q+F = Kq + F . (5.20)
∂u ∂u ∂u ∂u ∂r ∂λ
Although A is unsymmetric, under some general conditions it has real eigenvalues. To show that we express A as the product
of two symmetric matrices:
∂v ∂v ∂r ∂F
A= = = K = Fλ K, (5.21)
∂u ∂r ∂u ∂λ
where the relation (5.19) has been used. If Fλ is nonsingular, the eigensystem Axi = µi xi can be transformed to the
generalized symmetric eigenproblem
Kxi = µi F−1λ xi . (5.22)
If K is positive definite this system has nonzero real roots µi . If Fλ is singular but K positive definite, consideration of the
alternative eigensystem
Fλ yi = µi K−1 yi = µi Fyi , (5.23)
shows that such a singularity contributes only zero roots.
Remark 5.9. Another quantity that appears in the analysis of incremental methods is the vector
∂v ∂v ∂u
v = = = Av = Fλ Kv = Fλ q. (5.24)
∂λ ∂u ∂λ
Remark 5.10. Two other Legendre transforms may be constructed: X (δ, u) and Y (δ, r), in which δ = ∂/∂λ (a generalized
displacement if λ is a load multiplier) is the active variable and either u or r take the role of passive variables. X and Y
together with and K form a closed chain of Legendre transformations. The functions X and Y are, however, of limited
interest in the present context.
+ F(λ), (5.26)
does not change the equilibrium equations or rate forms. To second order in the increments we get
with
∂ ∂ 2
A=, a= , (5.28)
∂λ ∂λ2
evaluated at (u, λ). But we can always adjust F(λ) in (5.25) so that A = a = 0. Furthermore at an equilibrium
position r = 0, and along the equilibrium path u = K−1 q λ = uT λ. Substituting we find for the energy
increment
= U − W = 12 qT u(λ)2 − qT u(λ)2 = − 12 qT u(λ)2 . (5.29)
This formula displays the important function of the product qT u in the energy increment. By extension we
may call
W = qT u(λ)2 (5.30)
5–8
5–9 §5.5 *ENERGY INCREMENTS
the external work increment even if r does not derive from a potential.
To fix the ideas assume that r derives from a quadratic potential
= 12 uT Ku − qT uλ + Cλ + D, (5.31)
where C and D are arbitrary constants. Then the increment from an equilibrium position (u, λ) that
satisfies the linear relation Ku = qT λ, to an arbitrary configuration (u + u, λ + λ) is
5–9
Chapter 5: CRITICAL POINTS AND RELATED PROPERTIES 5–10
u 1 = 2, u 2 = 1, λ = 12 , (E5.2)
EXERCISE 5.2 Show that all critical points of (E5.1) satisfy either of the equations
63 − u 1 − 36u 2 = 0, 5 − 2u 1 − 3u 2 + u 1 u 2 = 0 (E5.3)
called critical point surfaces, and that the only intersection of these surfaces and the equilibrium path is at
(E5.2).
EXERCISE 5.3 Show that the critical point surface defined by det(K) = 0 is independent of λ if the residual
force system is separable.
EXERCISE 5.4 Show that qT z is independent of λ if the residual force system is separable and the load is
proportional.
EXERCISE 5.5 (Advanced, requires knowledge of matrix eigensystem theory). If K is not symmetric, the
critical point classification argument based on qT z fails. Explain why.
5–10
4
.
One-Parameter
Residual Equations
4–1
Chapter 4: ONE-PARAMETER RESIDUAL EQUATIONS 4–2
TABLE OF CONTENTS
Page
§4.1. Introduction 4–3
§4.2. Rate Forms and Incremental Velocity 4–3
§4.3. Separable Residuals and Proportional Loading 4–3
§4.4. Response Visualization by Incremental Flow 4–4
§4.4.1. Diagrams for One Degree of Freedom . . . . . . . . . 4–4
§4.4.2. Diagrams for Multiple Degrees of Freedom . . . . . . . 4–7
§4.5. Intrinsic Geometry of Incremental Flow 4–8
§4.5.1. Tangent Vector . . . . . . . . . . . . . . . . . 4–8
§4.5.2. Normal Hyperplane and Flow-Orthogonal Envelope . . . . 4–8
§4.5.3. ArcLength Distance . . . . . . . . . . . . . . . 4–8
§4.6. *State Vector Scaling 4–11
§4. Exercises . . . . . . . . . . . . . . . . . . . . . . 4–13
4–2
4–3 §4.3 SEPARABLE RESIDUALS AND PROPORTIONAL LOADING
§4.1. Introduction
This Chapter continues on the topic of residual equations introduced in Chapter 3. The general
residual force equation presented there is specialized, through the concept of staging introduced in
§3.4, to the one-parameter form in which r is a function of u (the state) and λ (the control). Together
these form the control-state space. The separable case in which u and λ can be segregated to both
sides of the residual equations, is described.
Further insight into the structural response may be achieved with the help of constant-residual
incremental flows. Paths and orthogonal hypersurfaces are introduced and interpreted geometrically.
Finally, the concepts of arclength and scaling are discussed.
In this section we study further the one-parameter residual equation (3.17), reproduced below for
convenience:
r(u, λ) = 0. (4.1)
The corresponding residual-derivative equations are
∂r ∂r
K= , q=− (4.4)
∂u ∂λ
where K is the tangent stiffness matrix introduced in §3.3, and q is the incremental load vector. The
latter is the specialization of the control matrix Q defined in §3.3, to the one-parameter case. These
equations will be used in the sequel instead of the more general (3.14)–(3.15) unless otherwise
noted.
Rate forms of r(u, λ) = 0 are obtained by equating the above derivatives to zero:
4–3
Chapter 4: ONE-PARAMETER RESIDUAL EQUATIONS 4–4
Remark 4.2. If a separable system derives from a total potential energy = U − P, then the external work
potential P must be linear in the state parameters u i . Furthermore for the loading to be proportional, P must
also be linear in λ.
r(u, λ) = rc , (4.13)
4–4
4–5 §4.4 RESPONSE VISUALIZATION BY INCREMENTAL FLOW
Response curve r = 0
λ +
− Pseudo-time t increasing
Pseudo-time t decreasing
where rc is a constant vector. This is the general solution of ṙ = 0. Additional information can
be conveyed by drawing the solutions of (4.13) for various values of the right-hand side near zero.
This produces constant-residual paths as illustrated in Figure 4.3. Collectively these paths form the
incremental flow whose differential equation is either ṙ = 0, or, if we take λ ≡ t :
∂r
r = = 0, (4.14)
∂λ
where primes denote derivatives with respect to λ. This can also be presented as
∂u ∂r
r = K + = Ku − q = 0. (4.15)
∂λ ∂λ
4–5
Chapter 4: ONE-PARAMETER RESIDUAL EQUATIONS 4–6
P(u, λ ) t+
positive tangent at P
λ
r =0
normal "hyperplane" at P
P(u, λ ) t+
positive tangent at P
λ
r =0
Figure 4.4. Incremental flow (full curves) and the flow-orthogonal envelope
(dashed curves). This envelope reduces here to a family of curves
because there is only one degree of freedom u.
4–6
4–7 §4.4 RESPONSE VISUALIZATION BY INCREMENTAL FLOW
Example 4.1. For simple one-degree of freedom systems it is easy to plot the incremental flow using standard
graphic packages. As an example consider the following residual equation, which is obtained as solution of
one of the Exercises of Chapter 6:
Here µ is a dimensionless state parameter and α an angle in radians characterizing the reference position of
the structure. The following Mathematica program produces the incremental flow plot for α = 30◦ using the
ContourPlot function:
0.1
0.05
-0.05
-0.1
0 0.5 1 1.5 2
Figure 4.5. Incremental flow plot for the residual (4.16) produced
by Mathematica via its ContourPlot function.
Examination of Figure 4.5 shows that the r = const curves are simply translations of each other along the λ
axis because λ appears simply as −λ. This is typical of proportional loading situations.
4–7
Chapter 4: ONE-PARAMETER RESIDUAL EQUATIONS 4–8
where v = K−1 q is the incremental velocity vector (4.8). Tangent vectors are illustrated in Figures
4.3 and 4.8 for one and two degrees of freedom, respectively.
The tangent vector normalized to unit length is
v/ f
tu = , (4.18)
1/ f
where f is the scaling factor
f = |t| = + ||t||2 = + 1 + vT v. (4.19)
The positive tangent direction and the positive unit tangent are defined as
+ v + t+ v/ f
t =± , tu = =± . (4.20)
1 f 1/ f
The positive tangent direction points in the positive sense of path traversal, as defined in §4.2 and
Figure 4.2.
§4.5.2. Normal Hyperplane and Flow-Orthogonal Envelope
The hyperplane N P normal to t at P(u, λ) has the equation
vT
u +
λ = 0, (4.21)
where
u = u − u P and
λ = λ − λ P are increments from P. Dividing these increments by
t
and passing to the limit one obtains
vT u̇ + λ̇ = 0. (4.22)
For a one degree of freedom u the hyperplane reduce to a line in (u, λ) space, as illustrated in
Figure 4.4. For two degrees of freedom the normal hyperplane is an ordinary plane in the 3D space
(u 1 , u 2 , λ), as illustrated in Figure 4.8.
For one degree of freedom (4.22) is the differential equation of a flow orthogonal to the incremental
flow, as illustrated in Figure 4.3; this flow is the envelope of the normals. For two degrees of
freedom (4.22) represents a family of surfaces, see Figure 4.6. For more degrees of freedom (4.22)
is a family of hypersurfaces. The orthogonality property plays an important role in corrective
solution methods.
4–8
4–9 §4.5 INTRINSIC GEOMETRY OF INCREMENTAL FLOW
r =0
u2
u1
Figure 4.6. An incremental flow response diagram for two degrees of freedom.
The plane paths of Figure 4.3 now become space curves.
Only a few paths are shown to reduce clutter.
r =0
u2
u1
Figure 4.7. A response diagram for two degrees of freedom, showing some
members of the flow-orthogonal envelope. Only the primary
equilibrium path r = 0 is shown to reduce clutter.
4–9
Chapter 4: ONE-PARAMETER RESIDUAL EQUATIONS 4–10
Normal hyperplane
v T u̇ + λ̇ = 0
P t+
λ
+ sense of increasing t
P̄
+
r =0
t+
u2
u1
4–10
4–11 §4.6 *STATE VECTOR SCALING
Normal hyperplane at P
s
Q
P
λ
Positive tangent direction
Path
u2
u1
du v dλ 1
= , = . (4.26)
ds f ds f
Remark 4.5. In the mathematical literature the incremental flow projected on the u state space is sometimes
called a Davidenko flow in honor of the father of continuation methods, should λ be interpreted as a continuation
parameter.
Remark 4.6. An alternative to plotting (4.13) for response visualization, is to consider the use of the constant-
residual-norm equation
||r(u, λ)|| = C, (4.27)
where ||r|| denotes a vector norm such as, for instance, the Euclidean norm ||r||2 = rT r, and C is a nonnegative
numeric constant. This relation does not generally represent a family of curves but a family of tube-like
hypersurfaces that for sufficiently small C “wrap around” equilibrium paths, as illustrated in Figure 4.10.
Because of the visual clutter evident in that figure, equation (4.27) is less suitable than (4.13) to study what
happens in the neighboorhood of equilibrium paths.
4–11
Chapter 4: ONE-PARAMETER RESIDUAL EQUATIONS 4–12
||r|| = C2 > C1
r =0
u2
u1
Figure 410. For Remark 4.6: illustrating that the constant residual-norm
equation ||r|| = const generally represents a family of
tube-like surfaces “wrapping around” the equilibrium paths.
4–12
4–13 Exercises
EXERCISE 4.3 [A:20] Suppose the first residual force above is replaced by r1 = u 1 + 3u 22 − 221 .
(a) Is the system still separable?
(b) For the same two stages of the previous exercise, is the loading proportional?
EXERCISE 4.4 [A:25] For stage 1 of Exercise 4.2, write down the analytical expressions of the incremental
velocity, the tangent vectors t and tu , the normal hyperplane equation, and the differential equations of the
flow-orthogonal envelope. Note: explicit inversion of K−1 may be done using the formulas to invert a 2 × 2
matrix.
EXERCISE 4.5 [A:25] Verify the assertion of Remark 4.6 by using the Euclidean norm ||r|| = rT r of the
residual vector.
EXERCISE 4.6 [A:25] Explain whether the unnormalized tangent vector t introduced in §4.5.1 may be defined
as
u̇
t= , (E4.2)
λ̇
and whether this definition is more general than (4.17).
4–13
Chapter 2: A TOUR OF NONLINEAR ANALYSIS 2–10
Application Explanation
Strength analysis How much load can the structure support before
global failure occurs?
Deflection analysis When deflection control is of primary importance
6 the exclusion of constitutive or material nonlinearities does not imply that there are less important than the others. Quite
the contrary. But the topic is covered in separate courses offered in Civil Engineering departments.
7 These are the Tonti diagrams introduced in the IFEM course.
2–10
2
.
A Tour of
Nonlinear
Analysis
2–1
Chapter 2: A TOUR OF NONLINEAR ANALYSIS 2–2
TABLE OF CONTENTS
Page
§2.1. Introduction 2–3
§2.2. Equilibrium Path and Response Diagrams 2–3
§2.2.1. Load-deflection response . . . . . . . . . . . . . 2–3
§2.2.2. Terminology . . . . . . . . . . . . . . . . . 2–3
§2.3. Special Equilibrium Points 2–4
§2.3.1. Critical points . . . . . . . . . . . . . . . . . 2–4
§2.3.2. Turning points . . . . . . . . . . . . . . . . 2–5
§2.3.3. Failure points . . . . . . . . . . . . . . . . . 2–5
§2.4. Linear Response 2–5
§2.5. Tangent Stiffness and Stability 2–6
§2.6. Generalized Response 2–7
§2.7. Response Flavors 2–8
§2.8. Engineering Applications 2–9
§2.9. Sources of Nonlinearities 2–9
§2.10. Geometric Nonlinearity 2–11
§2.11. Material Nonlinearity 2–12
§2.12. Force BC Nonlinearity 2–13
§2.13. Displacement BC Nonlinearity 2–14
§2. Exercises . . . . . . . . . . . . . . . . . . . . . . 2–15
§2. Solutions to .Exercises
. . . . . . . . . . . . . . . . . . . . . 2–19
2–2
2–3 §2.2 EQUILIBRIUM PATH AND RESPONSE DIAGRAMS
§2.1. Introduction
This chapter reviews nonlinear structural problems by looking at the manifestation and physical
sources of nonlinear behavior.
We begin by introducing response as a pictorial characterization of nonlinearity of a structural
system. Response is a graphical representation of the fundamental concept of equilibrium path.
This concept permeates the entire course because of both its intrinsic physical value and the fact
that incremental solution methods (mentioned in Chapter 1) are based on it.
Finally, nonlinearities are classified according to their source in the mathematical model of con-
tinuum mechanics and correlated with the physical system. Examples of these nonlinearities in
practical engineering applications are given.
The concept of equilibrium path plays a central role in explaining the mysteries of nonlinear
structural analysis. This concept lends itself to graphical representation in the form of response
diagrams. The most widely used form of these pictures is the load-deflection response diagram.
Through this representation many key concepts can be illustrated and interpreted in physical,
mathematical or computational terms.
The gross or overall static behavior of many structures can be characterized by a load-deflection
or force-displacement response. The response is usually drawn in two dimensions as a x-y plot
as illustrated in Figure 2.1. In this figure a “representative” force quantity is plotted against a
“representative” displacement quantity. If the response plot is nonlinear, the structure behavior is
nonlinear.
Remark 2.1. We will see below that a response diagram generally depicts the relationship between inputs and
outputs. Or, in more physical terms, between what is applied and what is measured. For structures the most
common inputs are forces and the most common outputs are displacements or deflections1
Remark 2.2. The qualifier “representative” implies a choice among many possible candidates. For relatively
simple structures the choice of load and deflection variables is often clear-cut from considerations such as
the availability of experimental data. For more complex structures the choice may not be obvious, and many
possibilities may exist. The load is not necessarily an applied force but may be an integrated quantity: for
example the weight of traffic on a bridge, or the lift on an airplane wing.
Remark 2.3. This type of response should not be confused with what in structural dynamics is called the
response time history. A response history involves time, which is the independent variable, plotted usually
along the horizontal axis, with either inputs or outputs plotted vertically.
1 A deflection is the magnitude or amplitude of a displacement. Displacements are vector quantities whereas deflections
are scalars.
2–3
Chapter 2: A TOUR OF NONLINEAR ANALYSIS 2–4
Representative
load
Equilibrium path
Representative
deflection
Reference state
§2.2.2. Terminology
A smooth curve shown in a load-deflection diagram is called a path.2 Each point in the path
represents a possible configuration or state of the structure. If the path represents configurations
of static equilibrium it is called an equilibrium path. Each point in an equilibrium path is called an
equilibrium point. An equilibrium point is the graphical representation of an equilibrium state or
equilibrium configuration.
The origin of the response plot (zero load, zero deflection) is called the reference state because it is
the configuration from which loads and deflections are measured. However, the reference state may
be in fact chosen rather arbitrarily, and this freedom is exploited in some nonlinear formulations
and solution methods, as we shall see later.
For problems involving perfect structures3 the reference state is unstressed and undeformed, and is
also an equilibrium state. This means that an equilibrium path passes through the reference state,
as in Figure 2.1.
The path that crosses the reference state is called the fundamental equilibrium path or fundamental
path for short. (Many authors also call this a primary path.) The fundamental path extends from
the reference state up to special states called critical points informally described in §2.3. Any path
that is not a fundamental path but connects with it at a critical point is called a secondary path. See
Figure 2.2.
§2.3. Special Equilibrium Points
Certain points of an equilibrium path have special significance in the applications and thus receive
special names. Of interest to our subject are critical, turning and failure points.
2 The terms branch and trajectory are also used. “Branch” is commonly used in the treatment of bifurcation phenomena,
whereas “trajectory” has temporal or historical connotation.
3 A concept to be explained later in connection with stability analysis. A perfect structure involves some form of idealization
such as perfectly centered loads or perfect fabrication. An imperfect structure is one that deviates from that idealization
in measurable ways.
2–4
2–5 §2.4 LINEAR RESPONSE
Representative
load Critical point
Secondary path
Fundamental or
primary path
Initial linear
response Representative
deflection
Reference state
Critical points are characterized mathematically in later chapters. It is sufficient to note here that
there are two types:
1. Limit points, at which the tangent to the equilibrium path is horizontal, i.e. parallel to the
deflection axis, and
Points at which the tangent to the equilibrium path is vertical, i.e. parallel to the load axis, are called
turning points. These are not critical points and have less physical significance, but are of interest
for some structures. They have some computational significance, however, because they can affect
the performance of certain solution methods.
Points at which a path suddenly stops or “breaks” because of physical failure are called failure
points. The phenomenon of failure may be local or global in nature. In the first case (e.g, failure of
a noncritical structure component) the structure may regain functional equilibrium after dynamically
“jumping” to another equilibrium path. In the latter case the failure is catastrophic or destructive
and the structure does not regain functional equilibrium.
In the present exposition, bifurcation, limit, turning and failure points are often identified by the
letters B, L, T and F, respectively.
Equilibrium points that are not critical are called regular.
2–5
Chapter 2: A TOUR OF NONLINEAR ANALYSIS 2–6
Representative
load
goes on forever
Representative
deflection
Figure 2.3. The response diagram for a purely linear structural model.
1. A linear structure can sustain any load whatsoever and undergo any displacement magnitude.
2. There are no critical, turning or failure points.
3. Response to different load systems can be obtained by superposition.
4. Removing all loads returns the structure to the reference position.
These assumptions are not only physically unrealistic but mutually contradictory. For example, if
the deformations are to remain infinitesimal for any load, the body must be rigid rather than elastic,
which contradicts the first assumption. Thus, there are necessarily limits placed on the validity of
the linear model.
Despite these obvious limitations, the linear model can be a good approximation of portions of the
nonlinear response. In particular, the fundamental path response in the vicinity the reference state.
See for instance Figure 2.2. Because for many structures this segment represents the operational
or service range, the linear model is widely used in design calculations. The key advantage of
this idealization is that the superposition-of-effects principle applies. Practical implications of the
failure of the superposition principle are further discussed in Chapter 3.
2–6
2–7 §2.6 GENERALIZED RESPONSE
Control parameter λ
Equilibrium path
State parameter µ or u
2–7
Chapter 2: A TOUR OF NONLINEAR ANALYSIS 2–8
R R R
Figure 2.5. Basic flavors of nonlinear response: (a) Linear until brittle failure,
(b) Stiffening or hardening, (c) Softening.
Remark 2.4. The interpretation of the tangent-to-the-path as stiffness discussed in §2.5 does not necessarily
carry over to more general control-state diagrams. Similarly, the interpretations of the sign of the tangent and
of the enclosed-area in terms of stability indicator and stored work, respectively, do not necessarily hold. This
is because control and state are not necessarily conjugate in the virtual work sense.
The response diagrams in Figure 2.5 illustrate three “monotonic” types of response: linear, hard-
ening, and softening. Symbols F and L identify failure and limit points, respectively.
A response such as in (a) is characteristic of pure crystals, glassy, and certain high strength composite
materials.
A response such as in (b) is typical of cable, netted and pneumatic (inflatable) structures, which may
be collectively called tensile structures. The stiffening effect comes from geometry “adaptation”
to the applied loads. Some flat-plate assemblies also display this behavior initially.
A response such as in (c) is more common for structure materials than the previous two. A linear
response is followed by a softening regime that may occur slowly or suddenly. More “softening
flavors” are given in Figure 2.6.
The diagrams of Figure 2.6 illustrate a “combination of basic flavors” that can complicate the
response as well as the task of the analyst. Here B and T denote bifurcation and turning points,
respectively.
The snap-through response (d) combines softening with hardening following the second limit point.
The response branch between the two limit points has a negative stiffness and is therefore unstable.
(If the structure is subject to a prescribed constant load, the structure “takes off” dynamically when
the first limit point is reached.) A response of this type is typical of slightly curved structures such
as shallow arches.
The snap-back response (e) is an exaggerated snap-through, in which the response curve “turns
back” in itself with the consequent appearance of turning points. The equilibrium between the
two turning points may be stable and consequently physically realizable. This type of response is
exhibited by trussed-dome, folded and thin-shell structures in which “moving arch” effects occur
2–8
2–9 §2.9 SOURCES OF NONLINEARITIES
L L B B
T
F B
T T
L F
L F L F
R R R R
Figure 2.6. More complex response patterns: (d) snap-through, (e) snap-back,
(f) bifurcation, (g) bifurcation combined with limit points and snap-back.
following the first limit point; for example cylindrical shells with free edges and supported by end
diaphragms.
In all previous diagrams the response was a unique curve. The presence of bifurcation (popularly
known as “buckling” by structural engineers) points as in (f) and (g) introduces more features. At
such points more than one response path is possible. The structure takes the path that is dynamically
preferred (in the sense of having a lower energy) over the others. Bifurcation points may occur in
any sufficiently thin structure that experiences compressive stresses.
Bifurcation, limit and turning points may occur in many combinations as illustrated in (g). A
striking example of such a complicated response is provided by thin cylindrical shells under axial
compression.
Nonlinear Structural Analysis is the prediction of the response of nonlinear structures by model-
based simulation. Simulation involves a combination of mathematical modeling, discretization
methods and numerical techniques. As noted in Chapter 1, finite element methods dominate the
discretization step.
Table 2.1 summarizes the most important applications of nonlinear structural analysis.
A response diagram characterizes only the gross behavior of a structure, as it might be observed
simply by conducting an experiment on a mechanical testing machine. Further insight into the source
of nonlinearity is required to capture such physical behavior with mathematical and computational
models for computer simulation.
For structural analysis there are four sources of nonlinear behavior. The corresponding nonlinear
effects are identified by the terms material, geometric, force B.C. and displacement B.C., in which
B.C. means “boundary conditions.” In this course we shall be primarily concerned with the last
2–9
Chapter 2: A TOUR OF NONLINEAR ANALYSIS 2–10
Application Explanation
Strength analysis How much load can the structure support before
global failure occurs?
Deflection analysis When deflection control is of primary importance
6 the exclusion of constitutive or material nonlinearities does not imply that there are less important than the others. Quite
the contrary. But the topic is covered in separate courses offered in Civil Engineering departments.
7 These are the Tonti diagrams introduced in the IFEM course.
2–10
2–11 §2.10 GEOMETRIC NONLINEARITY
Displacement
Prescribed BCs
displacements Displacements Body forces
Kinematic Equilibrium
equations equations
Force
Constitutive (Traction)
equations Prescribed
Strains Stresses
BCs tractions
or forces
linked in a “duality” sense, and so the term geometric nonlinearities applies collectively to both sets
of relations. The force BC nonlinearities couple displacements and applied forces (surface tractions
and/or body forces) and thus bring the additional links drawn in Figure 2.8.
In the following sections these sources of nonlinearities are correlated to the physics in more detail.
u=^
u
u^ on Su
u b
T
e=Du D σ+ b = 0
in V in V
σ =Ee σn = ^t
e σ ^t
in V on S t
Physical source
Change in geometry as the structure deforms is taken into account in setting up the strain-
displacement and equilibrium equations.
Applications
2–11
Chapter 2: A TOUR OF NONLINEAR ANALYSIS 2–12
Force B.C.
Displacement B.C. nonlinearities
nonlinearities
^
u u b
Geometric
nonlinearities
Material
nonlinearities
e σ ^t
Slender structures in aerospace, civil and mechanical engineering applications. Tensile structures
such as cables and inflatable membranes. Metal and plastic forming. Stability analysis of all types.
Mathematical source
Strain-displacement equations:
e = Du (2.1)
The operator D is nonlinear when finite strains (as opposed to infinitesimal strains) are expressed
in terms of displacements. Internal equilibrium equations:
b = −D∗ σ (2.2)
In the classical linear theory of elasticity, D∗ = DT is the formal adjoint of D, but that is not
necessarily true if geometric nonlinearities are considered.
Remark 2.5. The term geometric nonlinerities models a myriad of physical problems:
Large strain. The strains themselves may be large, say over 5%. Examples: rubber structures (tires, mem-
branes), metal forming. These are frequently associated with material nonlinearities.
Small strains but finite displacements and/or rotations. Slender structures undergoing finite displacements
and rotations although the deformational strains may be treated as infinitesimal. Example: cables, springs,
arches, bars, thin plates.
Linearized prebucking. When both strains and displacements may be treated as infinitesimal before loss of
stability by buckling. These may be viewed as initially stressed members. Example: many civil engineering
structures such as buildings and stiff (non-suspended) bridges.
2–12
2–13 §2.12 FORCE BC NONLINEARITY
Physical source
Material behavior depends on current deformation state and possibly past history of the deformation.
Other constitutive variables (prestress, temperature, time, moisture, electromagnetic fields, etc.)
may be involved.
Applications
Structures undergoing nonlinear elasticity, plasticity, viscoelasticity, creep, or inelastic rate effects.
Mathematical source
The constitutive equations that relate stresses and strains. For a linear elastic material
σ = Ee (2.3)
where the matrix E contains elastic moduli. If the material does not fit the elastic model, general-
izations of this equation are necessary, and a whole branch of continuum mechanics is devoted to
the formulation, study and validation of constitutive equations.
Remark 2.6. The engineering significance of material nonlinearities varies greatly across disciplines. They
seem to occur most often in civil engineering, that deals with inherently nonlinear materials such as concrete,
soils and low-strength steel. In mechanical engineering creep and plasticity are most important, frequently
occurring in combination with strain-rate and thermal effects. In aerospace engineering material nonlinearities
are less important and tend to be local in nature (for example, cracking and “localization” failures of composite
materials).
Remark 2.7. Material nonlinearities may give rise to very complex phenomena such as path dependence,
hysteresis, localization, shakedown, fatigue, progressive failure. The detailed numerical simulation of these
phenomena in three dimensions is still beyond the capabilities of the most powerful computers.
Physical Source
Applied forces depend on deformation.
Applications
The most important engineering application concerns pressure loads of fluids. These include
hydrostatic loads on submerged or container structures; aerodynamic and hydrodynamic loads
caused by the motion of aeriform and hydroform fluids (wind loads, wave loads, drag forces). Of
more mathematical interest are gyroscopic and non-conservative follower forces, but these are of
interest only in a limited class of problems, particularly in aerospace engineering.
Mathematical source
The applied forces (prescribed surface tractions
t and/or body forces b) depend on the displacements:
2–13
Chapter 2: A TOUR OF NONLINEAR ANALYSIS 2–14
t =
t(u), b = b(u), (2.4)
the former being more important in practice.
Physical source
Displacement boundary conditions depend on the deformation of the structure.
Applications
The most important application is the contact problem,8 in which no-interpenetration conditions
are enforced on flexible bodies while the extent of the contact area is unknown. Non-structural
applications of this problem pertain to the more general class of free boundary problems, for
example: ice melting, phase changes, flow in porous media. The determination of the essential
boundary conditions is a key part of the solution process.
Mathematical source
For the contact problem: prescribed displacements
d depend on internal displacements u:
d =
d(u) (2.5)
More complicated dependencies can occur in the free-boundary problems mentioned above.
8 Contact-impact in dynamics.
2–14
2–15 Exercises
EXERCISE 2.1 [D:10 Explain the difference, if any, between a load-deflection response and a control-state
response.
EXERCISE 2.2 [D:20] Can the following occur simultaneously: (a) a limit and a bifurcation point, (b) a
bifurcation and a turning point, (c) a limit and a turning point, (d) two bifurcation points coalescing into one.
If you answer “yes” to an item, sketch a response diagram to justify that reply.
EXERCISE 2.3 [D:25] In §2.10–13, nonlinearities are classified according to physical source into geomet-
ric, material, force boundary conditions, and displacement boundary conditions. For each of the following
mechanical systems indicate the source(s) of nonlinearity that you think are significant; note that there may
be more than one. (If you are not familiar with the underlying concepts, read those sections.)
(a) a long, slender elastic pipe bent under end couples while the pipe material stays elastic. See Figure E2.1.
(b) an inflating balloon. See Figure E2.2.
(c) a cable deflecting under action of wind forces while its material stays elastic. See Figure E2.3.
(d) a forming process in which hot metal is extruded through a rigid die. See Figure E2.4.
(e) a metal anchor is drilled into the soil to serve as a cable support; the hole is then filled with concrete.
See Figures E2.5 and E2.6. The question refers to the soil-drilling process, ignoring dynamics.
(f) a hefty bird — say a condor — sucked into an aircraft jet engine. Ignore dynamics; engine is the structure,
bird the load.
EXERCISE 2.4 [D:15] Can you think of a mechanical component that has the load-deflection response diagram
pictured in Figure E2.7? (Explain why). Hint: Think of a helicoidal spring.
2–15
Chapter 2: A TOUR OF NONLINEAR ANALYSIS 2–16
Figure E2.1. Slender elastic pipe bent under end couples for Exercise 2.3(a).
2–16
2–17 Exercises
Cable
Figure E2.3. Cable deflecting under wind forces for Exercise 2.3(c).
Die
Hot metal
Figure E2.4. Hot metal extruded trough a rigid die for Exercise 2.3(d).
2–17
Chapter 2: A TOUR OF NONLINEAR ANALYSIS 2–18
;;;;;;
;
Figure E2.5. Drill element of a cable anchor, for Exercise 2.3(e).
;;;;;;
;
;;;
;;; ;
(a) (b)
hole
;;;;;;
;
concrete grouting
;;;;;;
;
soil
Figure E2.6. Configuration of cable anchor after drilling in the soil, for Exercise 2.3(e).
Axial force B
II
III
R
Axial deflection (shortening)
2–18
1
.
Overview
1–1
Chapter 1: OVERVIEW 1–2
Theoretical
Mechanics Applied (1.1)
Computational
Theoretical mechanics deals with fundamental laws and principles of mechanics studied for their
intrinsic scientific value. Applied mechanics transfers this theoretical knowledge to scientific and
engineering applications, especially as regards the construction of mathematical models of physical
phenomena. Computational mechanics solves specific problems by simulation through numerical
methods implemented on digital computers.
REMARK 1.1
Paraphrasing an old joke about mathematicians, one may define a computational mechanician as a person who
searches for solutions to given problems, an applied mechanician as a person who searches for problems that
fit given solutions, and a theoretical mechanician as a person who can prove the existence of problems and
solutions.
Nanomechanics deals with phenomena at the molecular and atomic levels of matter. As such it
is closely interrelated with particle physics and chemistry. Micromechanics looks primarily at the
1–2
1–3 §1.2 WHERE THIS MATERIAL FITS
crystallographic and granular levels of matter. Its main technological application is the design and
fabrication of materials and microdevices.
Continuum mechanics studies bodies at the macroscopic level, using continuum models in which
the microstructure is homogenized by phenomenological averages. The two traditional areas of
application are solid and fluid mechanics. The former includes structures which, for obvious
reasons, are fabricated with solids. Computational solid mechanics takes a applied-sciences ap-
proach, whereas computational structural mechanics emphasizes technological applications to the
analysis and design of structures.
Computational fluid mechanics deals with problems that involve the equilibrium and motion of
liquid and gases. Well developed related areas are hydrodynamics, aerodynamics, atmospheric
physics, and combustion.
Multiphysics is a more recent newcomer. This area is meant to include mechanical systems that
transcend the classical boundaries of solid and fluid mechanics, as in interacting fluids and structures.
Phase change problems such as ice melting and metal solidification fit into this category, as do the
interaction of control, mechanical and electromagnetic systems.
Finally, system identifies mechanical objects, whether natural or artificial, that perform a distin-
guishable function. Examples of man-made systems are airplanes, building, bridges, engines,
cars, microchips, radio telescopes, robots, roller skates and garden sprinklers. Biological systems,
such as a whale, amoeba or pine tree are included if studied from the viewpoint of biomechanics.
Ecological, astronomical and cosmological entities also form systems.1
In this progression of (1.2) the system is the most general concept. Systems are studied by de-
composition: its behavior is that of its components plus the interaction between the components.
Components are broken down into subcomponents and so on. As this hierarchical process continues
the individual components become simple enough to be treated by individual disciplines, but their
interactions may get more complex. Thus there is an art in deciding where to stop.2
In dynamics the time dependence is explicitly considered because the calculation of inertial (and/or
damping) forces requires derivatives respect to actual time to be taken.
Problems in statics may also be time dependent but the inertial forces are ignored or neglected.
Static problems may be classified into strictly static and quasi-static. For the former time need not
1 Except that their function may not be clear to us. “The usual approach of science of constructing a mathematical model
cannot answer the questions of why there should be a universe for the model to describe. Why does the universe go to
all the bother of existing?” (Stephen Hawking).
2 Thus in breaking down a car engine, say, the decomposition does not usually proceed beyond the components you can
buy at a machine shop.
1–3
Chapter 1: OVERVIEW 1–4
be considered explicitly; any historical time-like response-ordering parameter (if one is needed)
will do. In quasi-static problems such as foundation settlement, creep deformation, rate-dependent
plasticity or fatigue cycling, a more realistic estimation of time is required but inertial forces are
still neglected.
Linear static analysis deals with static problems in which the response is linear in the cause-and-
effect sense. For example: if the applied forces are doubled, the displacements and internal stresses
also double. Problems outside this domain are classified as nonlinear.
A final classification of CSM static analysis is based on the discretization method by which the
continuum mathematical model is discretized in space, i.e., converted to a discrete model of finite
number of degrees of freedom:
Finite Element Method (FEM)
Boundary Element Method (BEM)
Finite Difference Method (FDM)
Spatial discretization method (1.4)
Finite Volume Method (FVM)
Spectral Method
Mesh-Free Method
For linear problems finite element methods currently dominate the scene, with boundary element
methods posting a strong second choice in specific application areas. For nonlinear problems the
dominance of finite element methods is overwhelming.
Classical finite difference methods in solid and structural mechanics have virtually disappeared
from practical use. This statement is not true, however, for fluid mechanics, where finite difference
discretization methods are still dominant. Finite-volume methods, which address finite volume
method conservation laws, are important in highly nonlinear problems of fluid mechanics. Spectral
methods are based on transforms that map space and/or time dimensions to spaces where the problem
is easier to solve.
A recent newcomer to the scene are the mesh-free methods. These are finite different methods on
arbitrary grids constructed by a subset of finite element techniques
The term Finite Element Method actually identifies a broad spectrum of techniques that share
common features outlined in introductory FEM textbooks. Two subclassifications that fit well
1–4
1–5 §1.3 THE FEM ANALYSIS PROCESS
VERIFICATION
Mathematical Discretization + solution error
model
IDEALIZATION
REALIZATION
FEM
SOLUTION
generally irrelevant
Figure 1.1. The Mathematical FEM. The mathematical model (at top) is the source of the process.
Discrete model and solution follow from it. The ideal physical system is inessential.
Using the foregoing classification, we can state the topic of this book more precisely: the compu-
tational simulation of nonlinear static structural problems by the Finite Element Method. Of the
variants listed in (1.5), emphasis is placed on the displacement formulation and stiffness solution.
This combination is called the Direct Stiffness Method or DSM.
3 The distinction between strong, weak and variational forms is discussed in advanced FEM courses. In the present book
such forms will be stated as recipes.
1–5
Chapter 1: OVERVIEW 1–6
generally
Ideal irrelevant
Mathematical
model
CONTINUIFICATION
SOLUTION
VALIDATION
Figure 1.2. The Physical FEM. The physical system is the source of
the process. The ideal mathematical model is inessential.
On the left Figure 1.1 shows an ideal physical system. This may be presented as a realization of
the mathematical model; conversely, the mathematical model is said to be an idealization of this
system. For example, if the mathematical model is the Poisson’s equation, realizations may be a
heat conduction or a electrostatic charge distribution problem. This step is inessential and may be
left out. Indeed FEM discretizations may be constructed without any reference to physics.
The concept of error arises when the discrete solution is substituted in the “model” boxes. This
replacement is generically called verification. The solution error is the amount by which the
discrete solution fails to satisfy the discrete equations. This error is relatively unimportant when
using computers, and in particular direct linear equation solvers, for the solution step. More
relevant is the discretization error, which is the amount by which the discrete solution fails to
satisfy the mathematical model.4 Replacing into the ideal physical system would in principle
quantify modeling errors. In the mathematical FEM this is largely irrelevant, however, since the
ideal physical system is merely that: a figment of the imagination.
4 This error can be computed in several ways, the details of which are of little practical importance.
1–6
1–7 §1.3 THE FEM ANALYSIS PROCESS
others, notably complex engineering systems, it makes no sense. Indeed FEM discretizations may
be constructed and adjusted without reference to mathematical models, simply from experimental
measurements.
The concept of error arises in the physical FEM in two ways, known as verification and validation,
respectively. Verification is the same as in the Mathematical FEM: the discrete solution is replaced
into the discrete model to get the solution error. As noted above this error is not generally important.
Substitution in the ideal mathematical model in principle provides the discretization error. This is
rarely useful in complex engineering systems, however, since there is no reason to expect that the
mathematical model exists, and if it does, that it is more relevant than the discrete model. Validation
tries to compare the discrete solution against observation by computing the simulation error, which
combines modeling and solution errors. Since the latter is typically insignificant, the simulation
error in practice can be identified with the modeling error.
5 This interplay is not exactly a new idea: “The men of experiment are like the ant, they only collect and use; the reasoners
resemble spiders, who make cobwebs out of their own substance. But the bee takes the middle course: it gathers its
material from the flowers of the garden and field, but transforms and digests it by a power of its own.” (Francis Bacon).
6 Of course at the atomic and subatomic level quantum mechanics works for everything, from landing gears to passengers.
But it would be slightly impractical to model the aircraft by 1036 interacting particles.
7 A substructure is a part of a structure devoted to a specific function.
1–7
Chapter 1: OVERVIEW 1–8
ENT
M PON
at ical CO EVEL
hem L
Matmodel
ent
ponns
ary Com atio
Libr en t equ
FEM Com
pon e
TEM
ret
discdel
mo
SYS EL
LEV e
plet
Comution
s o l
em
Systrete
discdel
m o
l
sica
Phy tem
sys
Figure 1.3. Combining physical and mathematical modeling through multilevel FEM.
Only two levels (system and component) are shown for simplicity.
REMARK 1.2
More intermediate decomposition levels are used in some systems, such as offshore and ship structures, which
are characterized by a modular fabrication process. In that case the decomposition mimics the way the system is
actually constructed. The general technique, called superelements, is outlined in the IFEM book. In nonlinear
analysis the analysis procedure is more complex since the static condensation technique of linear analysis no
longer works.
REMARK 1.3
There is no point in practice in going beyond a certain component level while considering the complete model,
since the level of detail can become overwhelming without adding significant information. Further refinement
or particular components is done by the so-called global-local analysis techniques. This technique is an
instance of multiscale analysis.
For sufficiently simple structures, passing to a discrete model is carried out in a single idealization
and discretization step, as illustrated for the truss roof structure shown in Figure 1.4. Multiple
levels are unnecessary here. Of course the truss may be viewed as a substructure of the roof, and
the roof as a a substructure of a building.
1–8
1–9 §1.4 INTERPRETATIONS OF THE FINITE ELEMENT METHOD
member
support
joint
Physical System
IDEALIZATION
;;
;;
;;
;;
;;
;;
Figure 1.6. The idealization process for a simple structure. The physical system,
here a roof truss, is directly idealized by the mathematical model: a
pin-jointed bar assembly. This coalesces with the discrete model.
The physical interpretation is aligned with the view of Figure 1.2. This interpretation has been
shaped by the discovery and extensive use of the method in the field of structural mechanics. This
relationship is reflected in the use of structural terms such as “stiffness matrix”, “force vector” and
“degrees of freedom.” This terminology carries over to non-structural applications.
The basic concept in the physical interpretation is the breakdown (≡ disassembly, tearing, partition,
separation, decomposition) of a complex mechanical system into simpler, disjoint components
called finite elements, or simply elements. The mechanical response of an element is characterized
in terms of a finite number of degrees of freedom. These degrees of freedoms are represented as
the values of the unknown functions as a set of node points. The element response is defined by
algebraic equations constructed from mathematical or experimental arguments. The response of
the original system is considered to be approximated by that of the discrete model constructed by
connecting or assembling the collection of all elements.
The breakdown-assembly concept occurs naturally when an engineer considers many artificial and
natural systems. For example, it is easy and natural to visualize an engine, bridge, aircraft or
skeleton as being fabricated from simpler parts.
The underlying theme is divide and conquer. If the behavior of a system is too complex, the
recipe is to divide it into more manageable subsystems. If these subsystems are still too complex
the subdivision process is continued until the behavior of each subsystem is simple enough to fit a
mathematical model that represents well the knowledge level the analyst is interested in. In the finite
element method such “primitive pieces” are called elements. elements The behavior of the total
system is that of the individual elements plus their interaction. A key factor in the initial acceptance
of the FEM was that the element interaction can be physically interpreted and understood in terms
that were eminently familiar to structural engineers.
1–9
Chapter 1: OVERVIEW 1–10
FACT
The key difficulty is tied to the essentially obscure nature of general nonlinear systems, about which
very little can be said in advance. And you can be sure that Murphy’s law8 works silently in the
background.
One particularly vexing aspect of dealing with nonlinear systems is the solution morass. A deter-
minate system of 1, 1000, or 1000000 linear equations has, under mild conditions, one and only
one solution. The computer effort to obtain this solution can be estimated fairly accurately if the
sparseness (or denseness) of the coefficient matrix is known. Thus setting up linear equation solvers
as “black-box” stand-alone functions or modules is a perfectly sensible thing to do.
By way of contrast, a system of 1000 cubic equations has 31000 ≈ 10300 solutions in the complex
plane. This is much, much larger than the number of atoms in the Universe, which is merely 1050
1–10
1–11 §1.6 HISTORICAL BACKGROUND
give or take a few. Suppose just several billions or millions of these are real solutions. Which
solution(s) have physical meaning? And how do you compute those solutions without wasting time
on the others?
This combinatorial difficulty is overcome by the concept of continuation, which engineers also call
incremental analysis. Briefly speaking, we start the analysis from an easily computable solution
— for example, the linear solution — and then try to follow the behavior of the system as actions
applied to it are changed by small steps called increments. The previous solution is used as a starting
point for the iterative solution-search procedure. The underlying prescription: follow the physics.
This technique is interwined with the concept of response explained in Chapter 2.
REMARK 1.4
Not surprisingly, incremental analysis was used by the aerospace engineers that first used the finite element
method for geometrically nonlinear analysis in the late 1950s. Techniques have been considerably refined
since then, but the underlying idea remains the same.
We conclude this overview with a historical perspective on nonlinear finite element methods in
solid and structural mechanics, along with a succint bibliography.
1. Finite deflections (geometric nonlinearities). Nonlinear effects arise from strain-displacement equations,
which are well behaved for all strain measures in practical use.
1–11
Chapter 1: OVERVIEW 1–12
3. Follower forces (e.g., pressure loading). External forces are smooth nonlinear functions of displacements.
A unifying characteristic of this problem is that nonlinearities are of equality type, i.e., reversible, and these
relations are continuous at each point within the structure. Mathematicians call these smooth mappings.
It is important to point out, however, that the overall structural behavior is not necessarily smooth; as witnessed
by the phenomena of buckling, snapping and flutter. But at the local level everything is smooth: nonlinear
strain-displacement equations, nonlinear elasticity law, follower pressures.
Methods for solving this class of problems are highly developed, and have received a great deal of attention
from the mathematical and numerical analysis community. This research has directly benefitted many areas
of structural analysis.
Let us consider finite deflection problems as prototype. Within the finite element community, these were
originally treated by purely incremental (step-by-step) techniques; but anomalies detected in the mid-1960’s
prompted research into consistent linearizations. A good exposition of this early work is given in the book by
Oden [123]. Once formulation questions were settled, investigators had correct forms of the “residual” out-
of-balance forces and tangent stiffness matrix, and incremental steps began to be augmented with corrective
iterations in the late 1960s. Conventional and modified Newton methods were used in the corrective phase.
These were further extended through restricted step (safeguarded Newton) and, more recently, variants of the
powerful conjugate-gradient and quasi-Newton methods.
But difficulties in detecting and traversing limit and bifurcation points still remained. Pressing engineering
requirements for post-buckling and post-collapse analyses led to the development of displacement control,
alternating load/displacement control, and finally arclength control. The resultant increment-control methods
have no difficulty in passing limit points. The problem of reliably traversing simple bifurcation points without
guessing imperfections remains a research subject, while passing multiple or clustered bifurcation points
remains a frontier subject. A concerted effort is underway, however, to subsume these final challenges.
These reliable solution methods have been implemented into many special-purpose finite element programs,
and incorporation into general-purpose programs is proceeding steadily.
REMARK 1.5
As noted above, incremental methods were the first to be used in nonlinear structural analysis. Among the pre-1970
contributions along this line we may cite Argyris and coworkers [9, 10], Felippa [62], Goldberg and Richard [82], Marcal,
Hibbitt and coworkers [89,108,109], Oden [122], Turner, Martin and coworkers [110,187,188],
REMARK 1.6
The earliest applications of Newton methods to finite element nonlinear analysis are by Oden [122], Mallet and Marcal [107],
and Murray and Wilson [117,118]. During the early 1970s Stricklin, Haisler and coworkers at Texas A&M implemented
and evaluated self-corrective, pseudo-force, energy-search and Newton-type methods and presented extensive comparisons;
see Stricklin et. al. [175–178], Tillerson et. al. [185], and Haisler et. al. [85]. Almroth, Brogan, Bushnell and coworkers
at Lockheed began using true and modified Newton methods in the late 1960s for energy-based finite-difference collapse
analysis of shells; see Brogan and Almroth [34], Almroth and Felippa [5], Brush and Almroth [37], and Bushnell [38–39].
By the late 1970s Newton-like methods enjoyed widespread acceptance for geometrically nonlinear analysis.
REMARK 1.7
Displacement control strategies for finite element post-buckling and collapse analysis were presented by Argyris [11] and
Felippa [62] in 1966, and generalized in different directions by Sharifi and Popov [165,166] (fictitious springs), Bergan et.
al. [25,26], (current stiffness parameter), Powell and Simons [142] and Bergan and Simons [27] (multiple displacement
controls). A modification of Newton’s method to traverse bifurcation points was described by Thurston [184]. Arclength
control schemes for structural problems may be found in the following source papers: Wempner [194], Riks [154], Schmidt
1–12
1–13 §1.6 HISTORICAL BACKGROUND
[160], Crisfield [48,49], Ramm [146], Felippa [65–66], Fried [73], Park [133], Padovan [130,131], Simo et.al. [167], Yang
and McGuire [199], Bathe and Dvorkin [20]. Other articles of particular interest are Bathe and Cimento [18], Batoz and
Dhatt [22], Bushnell [39], Bergan [27], Geradin et al. [79,80]. Meek and Tan [112], Ramm [146,147], Riks [155–157], Sobel
and Thomas [170], Zienkiewicz [200,201,203]. Several conferences have been devoted exclusively to nonlinear problems
in structural mechanics, for example [12,28,17,125, 179,180,198]. Finite element textbooks and monographs dealing rather
extensively with nonlinear problems are by Oden [123], Bathe [19], Bushnell [40], White [196] and Zienkiewicz [202].
REMARK 1.8
In the mathematical literature the concept of continuation (also called imbedding) can be traced back to the 1930s. A survey
of the work up to 1950 is given by Ficken [70]. The use of continuation by parameter differentiation as a numerical method
is attributed to Davidenko [54]. Key papers of this early period are by Freudenstein and Roth [72], Deist and Sefor [58] and
Meyer [114], as well as the survey by Wasserstrom [191]. This early history is covered by Wacker [190].
REMARK 1.9
Arclength continuation methods in the mathematical literature are generally attributed to Haselgrove [87] and Klopfestein
[99] although these papers remained largely unnoticed until the late 1970s. Important contributions to the mathematical
treatment are by Abbott [1], Anselone and Moore [8], Avila [14], Brent [31], Boggs [29], Branin [30], Broyden [35,36],
Cassel [43], Chow et. al. [45], Crandall and Rabinowitz [47], Georg [77,78], Keller and coworkers [44,56,57,93–96],
Matthies and Strang [111], Moore [115,116], Pönish [139,140], Rheinboldt and coworkers [59,113, 149–150], Watson
[192] and Werner and Spence [195]. Of these, key contributions in terms of subsequent influence are [45,94,149]. For
surveys and edited proceedings see Allgower [2,3], Byrne and Hall [42], Küpper [104,105], Rall [145], Wacker [190], and
references therein. Textbooks and monographs dealing with nonlinear equation solving include Chow and Hale [46], Dennis
and Schnabel [61], Kubı́ček and Hlaváček [102], Kubı́ček and Marek [103], Ortega and Rheinboldt [128], Rabinowitz [143],
Rall [144], Rheinboldt [153], and Seydel [164]. Of these, the book by Ortega and Rheinboldt remains a classic and an
invaluable source to essentially all mathematically oriented work done prior to 1970. The book by Seydel contains material
on treatment of conventional and Hopf bifurcations not readily available elsewhere. Nonlinear equation solving is interwined
with the larger subject of optimization and mathematical programming; for the latter the textbooks by Gill, Murray and
Wright [81] and Fletcher [71] are highly recommended.
Rough nonlinearities are characterized by discontinuous field relations, usually involving inequality constraints.
Examples: flow-rule plasticity, contact, friction. The local response is nonsmooth.
Solution techniques for these problems are in a less satisfactory state, and case-by-case consideration is called
for. The local and overall responses are generally path-dependent, an attribute that forces the past response
history to be taken into account.
The key difficulty is that conventional solution procedures based on Taylor expansions or similar differential
forms may fail, because such Taylor expansions need not exist! An encompassing mathematical treatment
is lacking, and consequently problem-dependent handling is presently the rule. For this class of problems
incremental methods, as opposed to incremental-iterative methods, still dominate.
REMARK 1.10
Earliest publications on computational plasticity using finite element methods are by Gallagher et. al. [74], Argyris [9,10],
Marcal [108], Pope [138] and Felippa [62]. By now there is an enormous literature on the numerical treatment of inelastic
processes, especially plasticity and creep. Fortunately the survey by Bushnell [40], although focusing on plastic buckling,
contains over 300 references that collectively embody most of the English-speaking work prior to 1980. Other important
surveys are by Armen [13] and Willam [197]. For contact problems, see Oden [126], Bathe and Chaudhary [19], Kikuchi
and Oden [97,98], Simo et. al. [167] Stein et. al. [173], Nour-Omid and Wriggers [121], and references therein.
1–13
Chapter 1: OVERVIEW 1–14
Smooth + Rough
This separation is usually done at the force level. Then the smooth-nonlinear term is treated by conventional
techniques whereas the rough-nonlinear term is treated by special techniques. This scheme can be particularly
effective when the rough nonlinearity is localized, for example in contact and impact problems.
Nesting may be used when a simple additive separation is not available. This is best illustrated by an actual
example. In the early 1970s, some authors argued that Newton’s method would be useless for finite-deflection
elastoplasticity, as no unique Jacobian exists in plastic regions on account of loading/unloading switches. The
argument was compelling but turned out to be a false alarm. The problem was eventually solved by “nesting”
geometric nonlinearities within the material nonlinearity, as illustrated in Fig. 1.1.
In the inner equilibrium loop the material law is “frozen”, which makes the highly effective Newton-type meth-
ods applicable. The non-conservative material behavior is treated in an outer loop where material properties
and constitutive variables are updated in an incremental or sub-incremental manner.
Another application of nesting comes in the global function approach (also called Rayleigh-Ritz or reduced-
basis approach), which is presently pursued by several investigators. The key idea is to try to describe the
overall response behavior by a few parameters, which are amplitudes of globally defined functions. The small
nonlinear system for the global parameters is solved in an inner loop, while an external loop involving residual
calculations over the detailed finite element model is executed occasionally.
Despite its inherent implementation complexity, the global function approach appears cost-effective for smooth,
path-independent nonlinear systems. This is especially so when expensive parametric studies are involved, as
in structural optimization under nonlinear stability constraints.
REMARK 1.11
For geometric-material nesting and subincremental techniques see Bushnell [38–41], and references therein. The global-
function approach in its modern form was presented by Almroth, Stern and Brogan [7] and pursued by Noor and coworkers
under the name of reduced-basis technique; see Noor and Peters [119] and Noor [120] as well as the chapter by Noor in this
volume. For perturbation techniques see the survey by Gallagher [75].
1–14
30
.
Dynamic
Stability
Analysis
30–1
Chapter 30: DYNAMIC STABILITY ANALYSIS 30–2
TABLE OF CONTENTS
Page
§30.1. Introduction 30–3
§30.2. The Linearized Equations of Motion 30–3
§30.3. The Characteristic Problem 30–4
§30.3.1. Connection with the Free-Vibration Eigenproblem . . . . . 30–5
§30.4. Characteristic Exponents and Stability 30–5
§30.4.1. Negative Real Case: Harmonic Oscillations . . . . . . . 30–5
§30.4.2. Positive Real Case: Divergence . . . . . . . . . . . 30–6
§30.4.3. Complex Case: Flutter . . . . . . . . . . . . . . 30–7
§30.4.4. Stable and Unstable Regions in the Complex Plane . . . . . 30–8
§30.5. Graphical Representations 30–8
§30.5.1. Root locus plots . . . . . . . . . . . . . . . . 30–8
§30.5.2. Amplitude Plots . . . . . . . . . . . . . . . . 30–10
§30.6. Regression to Zero Frequency and Static Tests 30–11
§30. Exercises . . . . . . . . . . . . . . . . . . . . . . 30–12
§30. Solutions to .Exercises
. . . . . . . . . . . . . . . . . . . . . 30–16
30–2
30–3 §30.2 THE LINEARIZED EQUATIONS OF MOTION
§30.1. Introduction
If the loading is nonconservative the loss of stability may not show up by the system going into
another equilibrium state but by going into unbounded motion. To encompass this possibility we
must consider the dynamic behavior of the system because stability is essentially a dynamic concept
(recall the definition in §25.1).
The essential steps are as follows. We investigate the motion that occurs after some initial per-
turbation is applied to the equilibrium state being tested, and from the properties of the motion
we can infer or deny stability. It if turns out that the perturbed motion consists of oscillations of
increasing amplitude, or is a rapidly increasing departure from the equilibrium state, the equilibrium
is unstable; otherwise it is stable.
The practicality of this approach depends crucially on the linearization of the equations of motion of
the perturbation. Thus we avoid having to trace the ensemble of time histories for every conceivable
dynamic departure from equilibrium — which for a system with many degrees of freedom would
clearly be a computationally forbidding task.
By linearizing we can express the perturbation motion as the superposition of complex exponential
elementary solutions. The characteristic exponents of these solutions can be determined through
a characteristic value problem or eigenproblem. This problem includes the free-vibration natural
frequency eigenproblem as particular case when the system is conservative and the tangent stiffness
matrix is symmetric. Through the stability criterion discussed in §30.3, the set of characteristic ex-
ponents gives complete information on the linearized stability of the system at the given equilibrium
configuration.
In practical studies the characteristic exponents are functions of the control parameter λ. Assuming
that the system is stable for sufficiently small λ values, say λ = 0, we are primarily concerned with
finding the first occurrence of λ at which the system loses stability. The transition to instability may
occur in two different ways, which receive the names divergence and flutter, respectively.1
The distinction between divergence and flutter instability is important in that the singular-stiffness
test discussed in Chapter 26 remains valid if the stability loss occurs by divergence, although of
course the tangent stiffness is not necessarily symmetric. Therefore it follows that in that case
we may fall back upon the static criterion, which is simpler to apply because it does not involves
information about mass and damping. Such a regression is not possible, however, if the loss of
stability occurs by flutter.
The structure is in static equilibrium under a given value of the control parameter λ. The equilibrium
state is defined by the state vector u. At time2 τ = 0 apply a dynamic input (e.g., an impulse) to this
configuration and examine the subsequent motion of the system. Roughly speaking if the motion
is unbounded (remains bounded) as τ tends to infinity the system is dynamically unstable (stable).
1 These names originated in aeronautical engineering applications, more specifically the investigation of sudden airplane
“blow ups” during the period 1910-1930. In the mathematical literature flutter goes by the name ‘Hopf bifurcation.’
2 The symbol τ denotes real time because t is used throughout the course to denote a pseudo-time parameter. Only real
time is considered in this Chapter.
30–3
Chapter 30: DYNAMIC STABILITY ANALYSIS 30–4
As noted in the Introduction, to simplify the mathematical treatment we consider only the local
stability condition, in which the imparted excitation is so tiny that the subsequent motion can be
viewed as a linearizable perturbation. We are effectively dealing with small perturbations about
the equilibrium position.
Let M be the symmetric mass matrix, which is assumed positive definite, and K the tangent
stiffness matrix, which is real but generally unsymmetric because of load nonconservativeness. The
perturbation motion is denoted as
d(τ ) = u(τ ) − u(0), τ ≥ 0+ (30.1)
The discrete, unforced, undamped governing equations of motion are
Md̈ + Kd = 0, (30.2)
in which a superposed dot — unlike previous Chapters — denotes differentiation with respect to
real time. The ordinary differential equations (30.2) express the linearized dynamic equilibrium
between stiffness and inertial forces. The stiffness forces generally include nonconservative loading
effects.
Remark 30.1. In structural with rotational DOFs, M might be only nonnegative definite because of the presence
of zero rotational masses. If so it is assumed that those DOFs have been eliminated by a static condensation
process.
The assumption of positive definiteness also excludes the presence of Lagrange multipliers in the state vector
u, because the associated masses of such degrees of freedom are zero. Again the stability criteria can be
extended by eliminating the multipliers in the linearized equation of motion.
(1) The effect of diagonalizable, light viscous structural damping does not generally affect stability results
(it certainly does not when stability loss is by divergence). See also Remark 30.4.
(2) The effect of more complicated nonlinear damping mechanisms such as dry friction may not be amenable
to linearization.
Thus cases when damping effects are significant lead to mathematics beyond the scope of this course. Readers
interested in pursuing this topic are referred to the vast literature on the subject of dynamic stability.
where i ranges over the number of degrees of freedom (number of state parameters). The pi are
generally complex numbers called the characteristic exponents whereas the corresponding column
vectors zi are the characteristic modes or characteristic vectors.3
3 In his classical treatise Nonconservative Problems of the Theory of Elastic Stability, (Pergamon, 1963), Bolotin employs
s for what we call here p, and so do many other authors. This notation connects well to the common use of the Laplace
transform to do more complicated systems. However, we have already reserved s for Piola-Kirchhoff stresses as well as
arclength.
30–4
30–5 §30.4 CHARACTERISTIC EXPONENTS AND STABILITY
(K + pi2 M) zi = 0, (30.4)
which is the characteristic problem or eigenproblem that governs dynamic stability. This equation
befits the generalized unsymmetric eigenproblem of linear algebra
pi = ± jωi , (30.6)
√
where j = −1, and the nonnegative real numbers ωi are the natural frequencies of free vibration.
Because p 2 = −ωi2 , (30.4) reduces to the usual vibration eigenproblem
(K − ωi2 M) zi = 0. (30.7)
Thus for the conservative case we regress to a well studied problem. In such a case the system will
simply vibrate, that is, perform harmonic oscillations about the equilibrium position because each
root is associated with the solution
The presence of positive damping will of course damp out these oscillations and the system even-
tually returns to the static equilibrium position.
pi = αi + jωi , (30.9)
√
where αi and ωi are real numbers, and j = −1. The component representation of the square of
pi is
pi2 = (αi2 − ωi2 ) + 2 jαi ωi , (30.10)
The exponential of a complex number has the component representation
On the basis of this representation we can classify the growth behavior of the subsequent motion
and consequently the stability of the system as examined in the next 3 subsections.
30–5
Chapter 30: DYNAMIC STABILITY ANALYSIS 30–6
di
2π/ωi
Figure 30.1. Harmonic oscillatory motion for the case where root pi2 of (30.4) is negative
real. Equivalently, pi = ± jωi where ωi is the circular frequency.
If
pi = ± jωi , d(τ ) = di (τ ), di (τ ) = Ai cos ωi τ + Bi sin ωi τ. (30.12)
where Ai and Bi are determined by initial conditions. The motion di associated with ± jωi is
harmonic and bounded, as illustrated in Figure 30.1. The system is dynamically stable for this
individual eigenvalue.
If all eigenvalues are negative real and distinct, the system is dynamically stable because any
superposition of harmonic motions of different periods is also a harmonic motion. If two or more
eigenvalues coalesce the analysis becomes more complicated because of the appearance of secular
terms that grow linearly in time. These effects can be studied in more detail in treatises in mechanical
vibrations.
di
eαi τ
e−αi τ
Figure 30.2. Aperiodic, exponentially growing motion for the real root case pi2 = αi2 ,
pi = ±αi . Transition to this kind of instability is called divergence.
30–6
30–7 §30.4 CHARACTERISTIC EXPONENTS AND STABILITY
di
eαi τ
2π/ωi
−eαi τ
Figure 30.3. Periodic, exponentially growing motion for case pi = +αi ± jωi with
nonzero αi . Transition to this kind of instability is called flutter.
30–7
Chapter 30: DYNAMIC STABILITY ANALYSIS 30–8
Remark 30.3. Frequency coalescence is necessary but not sufficient for flutter. It is possible for two frequencies
to pass by other “like ships crossing in the night” without merging. This happens if there is no mechanism by
which the two associated eigenmodes can exchange energy.
Remark 30.4. The fact that all characteristic motions are either harmonic or exponentially growing is a
consequence of the neglect of damping in setting up the stability problem. As noted in Remark 30.2, the
presence of damping or, in general, dissipative forces, introduces additional mathematical complications that
will not be elaborated upon here. Suffices to say that the addition of damping to a conservative system has
always a stabilizing effect (Rayleigh’s theorem). For non-conservative systems, the preceding statement is no
longer true, and indeed several counterexamples involving destabilizing damping have been constructed over
the past 40 years. In spite of this the effect is not often observed in practice.
Remark 30.5. The occurrence of flutter requires the coalescence of two natural frequencies. Consequently,
flutter cannot occur in systems with one degree of freedom (“it takes two to flutter”). The physical interpretation
of the flutter phenomenon is that one vibration mode absorbs energy and feeds it into another; this transference
or “energy resonance” becomes possible when the two modes have the same frequency.
Stable Unstable
ω2 − α 2 α
Stable Unstable
Figure 30.4. Stable and unstable regions in (a) the complex p 2 plane, (b) the
complex p plane. For the latter the stable region is the left-half
plane α =
( p) ≤ 0. For (a) it is the negative real axis.
α =
( p) ≤ 0. (30.16)
which includes the imaginary axis α = 0 as stability boundary. The right-hand p-plane α > 0 is
unstable. See Figure 30.4(b).
30–8
30–9 §30.5 GRAPHICAL REPRESENTATIONS
p 2 plane p plane
j pi (λ)
root trajectory
pi2 (λ)
Divergence
root trajectory Divergence − j pi (λ)
Figure 30.5. Root locus plots on the complex p 2 and p planes for divergence instability.
p 2 plane p plane
j p1 (λ) trajectory of
interacting roots
− j p1 (λ)
Figure 30.6. Root locus plots on the complex p 2 and p planes for flutter instability.
4 The word root in root-locus is used as abbreviation for characteristic root or eigenvalue
30–9
Chapter 30: DYNAMIC STABILITY ANALYSIS 30–10
| p|
| p1 | = |ω1 |
Divergence
λcr
Figure 30.7. Root amplitude plot illustrating loss of stability by divergence at λcr .
| p|
| p1,2 |
| p2 | = |ω2 |
| p1 | = |ω1 |
Flutter
λ
λcr
Figure 30.8. Root amplitude plot illustrating loss of stability by flutter at λcr .
Figure 30.6 illustrates loss of stability by flutter. As λ is varied, two interacting eigenvalues, labeled
as p12 and p22 , coalesce on the negative real axis of the p 2 plane and branch out into the unstable
region. The right-hand diagram depicts the same phenomenon on the p plane for the interacting
roots, which appears in complex-conjugate pairs.
30–10
30–11 §30.6 REGRESSION TO ZERO FREQUENCY AND STATIC TESTS
more precisely than with a locus or trajectory plot, (b) all related square roots such as ±αi ± ωi
“collapse” into a single value, and (c) the variation of several important roots (for several values of
i) may be shown without cluttering the picture.
Figures 30.7 and 30.8 illustrate typical root-amplitude plots in loss of stability by divergence and
flutter, respectively.
(−ωi2 M + K) zi = 0. (30.17)
But if ωi = 0 and M is positive definite, which we assume, then K must be singular. Therefore we
can regress to the static criterion or singular tangent stiffness test
which allows us to discard the mass matrix. This regression may be useful if one is solving a series
of closely related problems, for example during the design of a structure which is known a priori
to become unstable by divergence.
It should be cautioned, however, that the tangent stiffness matrix K for nonconservative systems
is generally unsymmetric (Chapter 29), and that the test for singularity must take account of that
property.
30–11
Chapter 30: DYNAMIC STABILITY ANALYSIS 30–12
EXERCISE 30.1 (A+C:25) This Exercise studies the stability of the “follower load” nonconservative system
shown in Figure E30.1.
ux
λP
90◦
2
L
;;
;;
Bar (1) uy
1 2 1
k1 90◦
C0
y
Bar (2) k2
L C
;; ;;
x
3 3
Two elastic bars, (1) and (2), are supported at 1 and 3 and hinged at 2. The bars have length L, axial stiffnesses
k1 and k2 , respectively, and can only move in the x, y plane. Bar (1) is loaded at node 2 by a force λP1 ,
directed upwards, that stays normal to bar (1) as it displaces. Bar (2) is loaded at node 2 by a force λP2 ,
directed leftwards, that stays normal to bar (2) as it displaces.
For the present exercise set P1 = P2 = P. Furthermore the following simplifying assumptions are to be made:
(A1) The displacements from the reference configuration are so small that C ≡ C0 insofar as setting up the
stability eigensystem5
(a) Show that under the simplifying assumptions (A1)–(A2), the tangent stiffness at C ≡ C0 in terms of the
two degrees of freedom u x = u x2 and u y = u y2 , is
λP
k1 0 0 1
K= + . (E30.1)
0 k2 L −1 0
5 This is similar to LPB (Chapters 24-25), but here a dynamic analysis is involved.
30–12
30–13 Exercises
The first component of K is the material stiffness whereas the second component is the load stiffness.
Hint for the latter: use the results of Remark 30.4
(b) The linearized dynamic eigenproblem (30.4) is
The exponents pi (the square roots of pi2 ) are generally complex numbers:
pi = αi + jωi , (E30.3)
where α and ω are the real and imaginary part of pi , respectively, zi are associated eigenmodes, and M
is the diagonal mass matrix
M 0
M= , (E30.4)
0 M
where M is the lumped mass at node 2 (half of the sum of the bar masses). By appropriate normalization
show that the eigenproblem can be reduced to the dimensionless form
1 0 κ 0 0 1
p̄ 2 + + λ̄ z̄i = 0, (E30.5)
0 1 0 1 −1 0
(c) Show that the critical positive λ̄cr at which the eigenvalues p̄i2 coalesce is given by the relation
|1 − κ|
λ̄cr = . (E30.6)
2
Further show that if λ̄ > λ̄cr the roots p̄i become complex and hence explain whether loss of stability
occurs. Is it divergence or flutter?
(d) For κ = 0.01, 1.0, 4.0 and 100 plot the dependence of | p̄i | (i = 1, 2) (where |.| denotes the modulus of
a complex number) on λ using
√ √
| p̄|/ κ, λ/ κ, (E30.7)
as vertical and horizontal axes, respectively. Go from λ = 0 up to 2λcr or 1.0, whichever is greater, and
use sufficient steps to get reasonable graphical accuracy.
EXERCISE 30.2 (A+C:25) Do the previous exercise removing assumption (A2), that is, considering now the
effect of the geometric stiffness KG but still assuming C ≡ C0 . Is there any difference with the critical load
result (E30.6)?
EXERCISE 30.3 (A+C:30) Beck’s column6 is the simplest follower-load problem involving a cantilevered
beam-column.7 This problem is shown in Figure E30.2.
The beam-column has length L, elastic modulus E and smallest moment of inertia I . It is loaded by a
compressive force λP which after deformation rotates with the end section of the column and remains tangential
to its deformed axis (see Figure above). The mass M (half of the column mass) is lumped at its free end.
6 M. Beck, Die Knicklast des eiseiting eigenspannen, tangential gedrückten Stabes, Z. angew. Math. Phys., 3, No. 3, 1952.
7 It is sometimes used as a very simple model to illustrate stability analysis of rockets against the “pogo” effect.
30–13
Chapter 30: DYNAMIC STABILITY ANALYSIS 30–14
x θ
M λP
1
C0 C
L
;;;
E, A, I constant
2 y
If this problem is treated by the static criterion (Euler’s method) one erroneously concludes that the beam
column cannot lose stability for any value of the load λP 8 A dynamic stability analysis, first carried out by
Beck (loc.cit.), shows that stability is lost by flutter at the critical load
EI
λPcr = 20.05093 . (E30.8)
L2
(a) Find the critical dynamic load given by the finite element method if one Euler-Bernoulli beam-column
element is used along the length of Beck’s column. Lateral displacements may be considered infinites-
imal; hence sin θ ≈ θ, cos θ ≈ 1, and the axial force is simply λP. The degrees of freedom are u x1 ,
u y1 and θz1 . Use the material and geometric stiffness matrices given in equations (E24.2) and (E24.3),
respectively, to which an unsymmetric load stiffness matrix K L , which couples the θz1 and u y1 degrees
of freedom, should be added.
(b) Repeat the analysis for two and four elements of equal length along the column. For two elements the
three nodes are 1 (top), 2 (middle of column) and 3 (root). Use lumped masses with M X 2 = MY 2 equal
to one half of the total column mass and M X 1 = MY 1 = M X 2 /2 = MY 2 /2. For four elements there are
five nodes, etc. Use of Mathematica or a similar program is recommended.
30–14
30–15 Exercises
+P 0 0 1 uy = 0
0 0 0 θz 0
where for simplicity u x = u x1 , u y = u y1 , θz = θz1 . The first dynamic equation in u x uncouples and has no effect in the
analysis. The last equation is static in nature because the rotational mass is zero. Thus, we can solve for θz in terms of u y :
P
6E I
− +
L 210 N
θz = uy = uy (E30.10)
4E I 4P L D
−
L 30
where N and D denote the numerator and denominator, respectively, of the relation that links θz to u y . The eigenvalue
equation becomes 12E I 6E I N 36P N P N
p2 M + − − + 3L + P u y = 0. (E30.11)
L3 L2 D 30L D 30L D
One of the bending eigenvalues p 2 of (E30.9) is always ∞ because the rotational mass is zero. Flutter occurs when the two
beding eigenvalues coalesce at infinity. The finite p 2 becomes infinite if D = 0 while N = 0. Thus the critical load for
“flutter at infinity” is
30E I
Pcr = (E30.12)
L2
which is about 50% in error with respect to the analytical value 20.05093E I /L 2 quoted in the exercise statement.
EXERCISE 30.4 (A:25) Do the previous exercise for a one-element discretization if the line of action of
the applied end load is forced to pass through the cantilever root (point 2). Does the structure loses stability
dynamically or statically?
30–15
29
.
Nonconservative
Loading
29–1
Chapter 29: NONCONSERVATIVE LOADING 29–2
TABLE OF CONTENTS
Page
§29.1. Introduction 29–3
§29.2. Potential Force Example: Gravity 29–3
§29.3. Follower Load and Associated Load Stiffness 29–5
§29.4. General Characterization of the Load Stiffness 29–6
§29.5. Forces Produced by Fluid Motion 29–7
§29.6. Load Stiffness For 2D Fluid Motion 29–9
§29. Exercises . . . . . . . . . . . . . . . . . . . . . . 29–11
29–2
29–3 §29.2 POTENTIAL FORCE EXAMPLE: GRAVITY
§29.1. Introduction
In Chapter 5 a mechanical system was defined to be conservative when both external and internal
forces are derivable from a potential. In this course we consider only elastic systems; conse-
quently the internal forces are derivable from an strain (internal) energy potential U . Thus the
conservative/nonconservative character depends on whether the external loads are conservative or
nonconservative.
Conservative applied forces f may be derived from the external loads potential V by differentiating
with respect to the state variables:
∂V
f= . (29.1)
∂u
Nonconservative forces, on the other hand, are not expressable as (29.1). They have to be worked
out directly at the force level.
In the present Chapter we will give examples of both force types in conjunction with the TL-
formulated two-node bar element The main result is that consideration of nonconservative loads
contributes an unsymmetric component, called load stiffness, to the tangent stiffness matrix. Treat-
ing this effect in stability analysis requires a dynamic criterion, which is covered in Chapter 30.
Remark 29.1. The chief sources of nonconservative forces in various branches of engineering are:
1. Aerodynamic forces (aerospace, civil); hydrodynamic forces (mechanical, marine, chemical); aircraft
and rocket propulsion forces (aerospace); frictional forces (mechanical, civil).
2. Gyroscopic forces (aerospace, electrical).
3. Active control systems (aerospace, electrical, mechanical).
In this Chapter we consider only hydrodynamic (fluid motion) forces as prototype example.
29–3
Chapter 29: NONCONSERVATIVE LOADING 29–4
x̄
g = directed along − z 2
C
Z, z
1
x̄0
C0
Y, y
E, A0 , L 0
X, x
Figure 29.1. TL bar element displacing in a gravity field g.
x̄
2
C
g Z, z
1
vz x̄0
C0
Y, y
ρ A0 d x̄0
X, x
Figure 29.2. Calculation of external potential.
(As usual in the TL kinematic description, all quantities are referred to C0 .) It follows that the
external force vector for the element is
∂ V /∂u x1 0
∂ V /∂u y1 0
∂V
∂ V /∂u z1 1
fg = = = − 2 ρ A0 L 0 .
1
(29.4)
∂u ∂ V /∂u x2 0
∂ V /∂u y2 0
∂ V /∂u z2 1
This can also be derived through basic principles of statics. Note that this vector is independent of
29–4
29–5 §29.3 FOLLOWER LOAD AND ASSOCIATED LOAD STIFFNESS
L x̄
ȳ
pd
θ
C
uX1 uY2
Y, y uY1
pd
1 2 x̄0 ≡ X
uX2
C0
L0
X, x
Figure 29.3. 2D bar under constant “follower” pressure pd .
the current configuration. This is a distinguishing feature of external work potentials that depend
linearly on the displacements, such as (29.3).
§29.3. Follower Load and Associated Load Stiffness
To illustrate the concept of load stiffness with a minimum of mathematics, let us consider a two-
dimensional specialization. The bar element originally lies along the x axis in the reference con-
figuration C0 and moves in the (x, y) plane to C, which forms an angle θ with x. The bar is under a
a constant pressure pd that is always normal to the element as it displaces, as shown in Figure 29.3.
This kind of applied force is called a follower load in the literature.1
From statics the external force vector is obviously
− sin θ
cos θ
0
f = 2 pd L
1
(29.5)
− sin θ
cos θ
0
From geometry
L 0 + u X 21 u Y 21
cos θ = , sin θ = , with u X 21 = u X 2 − u X 1 , u Y 21 = u Y 2 − u Y 1 , (29.6)
L L
1 Such loads are often applied by fluids at rest or in motion. The latter case is studied in Sections 29.4-5.
29–5
Chapter 29: NONCONSERVATIVE LOADING 29–6
Consequently
−u Y 21
L 0 + u X 21
0
f = 12 pd . (29.7)
−u Y 21
L 0 + u X 21
0
Take now the partial of the negative of this external load vector with respect to u. The result is a
matrix with dimensions of stiffness, denoted by K L :
0 −1 0 0 1 0
1 0 0 −1 0 0
∂f
0 0 0 0 0 0
KL = − = 2 pd
1
. (29.8)
∂u 0 −1 0 0 1 0
1 0 0 −1 0 0
0 0 0 0 0 0
K L is called a load stiffness matrix. It arises from displacement-dependent loads.2 We can see
from this example that K L is unsymmetric. A consequence of this fact is that (29.2) does not have
a potential V that is a function of the node displacements.3
K = K M + KG + K L . (29.13)
29–6
29–7 §29.5 FORCES PRODUCED BY FLUID MOTION
∂r
K= = K M + KG + K Lc + K Ln . (29.15)
∂u
The nonconservative load stiffness matrix, K Ln , is unsymmetric.
Remark 29.2. In practice one derives the total force f from statics, as in the example of §29.3, and obtains
K L by taking the partials with respect to the displacements in u. If the resulting stiffness is unsymmetric the
load is nonconservative. The splitting of K L into a symmetric matrix K Lc and unsymmetric part K Ln can be
done in a variety of ways. (If the unsymmetric part is required to be antisymmetric, however, the splitting is
unique.)
For simplicity in the formulation below, we further assume that the velocity field is uniform, i.e.,
does not depend upon (X, Y, Z ), and that it is directed along the x axis:
ufX
uf = 0 , (29.17)
0
4 The symbol u and its vector counterparts u and u are commonly used in fluid mechanics to denote velocities rather
than displacements as in structural and solid mechanics. In fact displacements are rarely used in fluids. Subscript f is
introduced here to lessen the risk of confusion with structural displacements.
5 For arbitrary cross sections, the fluid motion exerts drag and lift forces, the latter being normal to the bar axis and to the
normal velocity vector. Lift forces are what makes airplanes fly. This more general situation is dealt with in treatises on
aerodynamics, wind forces and hydraulics.
29–7
Chapter 29: NONCONSERVATIVE LOADING 29–8
z̄
ȳ
Y, y fluid velocity vector
v f X in (x̄, X ) plane
X, x
Z, z
x̄
C0
x̄0
For slow (laminar) flow the drag force is proportional to the magnitude of the normal velocity
component whereas if the motion is fully turbulent it is proportional to the square of that velocity.
We assume here the latter case. Other drag-velocity dependencies can be similarly treated.
Consider the bar in the (x̄, X ) plane as illustrated in Figure 29.5, and let ȳ be defined as the normal
to the element axis x̄ that is located in this plane and forms an acute angle θ with x. The drag force
on the element per unit length is directed along ȳ and has the value
pd = 12 Cd ρ f d u 2f n (29.18)
where Cd is the drag coefficient,6 ρ f the fluid mass density, d the “exposed width” (for a bar
of circular cross-section, its external diameter), and u f n the fluid-normal velocity u f X cos θ (see
Figure). The total force on the element is pd L, where L is the current length, and this force “lumps”
into 12 pd L at each node.
In order to refer these forces to the global X, Y, Z axes, we need to know the direction cosines t21 ,
t22 and t23 of ȳ with respect to x, y, z. Then the hydrodynamic node force vector in the (X, Y, Z )
29–8
29–9 §29.6 LOAD STIFFNESS FOR 2D FLUID MOTION
ȳ, n
paper is plane (x̄, X ) ≡ plane (x̄, ȳ)
1 vfn
θ
X, x vfX
C
2
x̄
Figure 29.5. Normal fluid velocity component in the current bar configuration
system is
t21
t22
t
f = 2 pd L 23
1
(29.19)
t21
t22
t23
To compute these direction cosines, one proceeds as follows:
(1) Compute the direction z̄ by taking the cross product of x̄ and X .
(2) Compute the direction ȳ by taking the cross product of z̄ and x̄.
If x̄ and X are parallel, step (1) does not define z but then the fluid flow occurs along the element
axis and the pressure pd vanishes.
Remark 29.3. If the fluid flow is uniform with speed u f j along a general direction j ≡ j, the preceding
derivation must be modified by taking z̄ = x̄ × j, ȳ = z̄ × x̄,
θ = angle( ȳ, j). Observe that it would be
incorrect to decompose u f j onto its components in the X , Y and Z directions and superpose associated forces,
because the drag force is nonlinear in the velocity.
Remark 29.4. If the flow is steady but nonuniform, numerical integration over elements is generally required.
For this simple elkement integration with the flow velocity evaluated at the element center is often sufficient.
To show what kind of load stiffness is produced by fluid drag forces, consider again the case of
Figure 29.4 but now make pd depend on the “tilt” θ as explained in §29.3; see Figure 29.6. Since
a turbulent-motion-induced drag force is proportional to the square of u f n = u f X cos θ, it may be
expressed as
pd = pd0 cos2 θ (29.20)
where pd0 is pd for θ = 0 (bar normal to fluid motion).
29–9
Chapter 29: NONCONSERVATIVE LOADING 29–10
L x̄
2 θ
ȳ c os
pd 0
=
pd
θ
C
uX1 uY2
uY1
Y, y
pd0
1 2 x̄0 ≡ X
uX2
C0
L0
X, x
Figure 29.6. Follower pressure pd on a 2D bar that depends on the “tilt angle” θ.
29–10
29–11 Exercises
EXERCISE 29.1 (A:20) Work out fd for the case of a uniform flow of speed u f j in a general direction j as
described in Remark 29.2.
EXERCISE 29.2 (A:15) Specialize the result of Exercise 29.1 to the two dimensional case (bar and flow in
the x, y plane). Differentiate to obtain K L , comparing with (29.20).
EXERCISE 29.3 (A:20) In the previous exercise take into account the effect of friction forces exerted on the
bar by the flow. Use the linear model: the tangential friction force pt per unit length of the bar is directed
along x̄ and has the value C f au f t , where C f is a friction coefficient, a is the “exposed perimeter” of the bar
(for a circular cross section, a = 2π d), and u f t = u f j sin θ is the tangential velocity (fluid velocity projected
on the current bar direction, with proper sign).
EXERCISE 29.6 (A:30) A simple example of a gyroscopic force is a torsional moment M̄x directed along the
longitudinal axis x̄ of a beam-column element, which keeps pointing in that direction as the element moves
and rotates. Obtain the gyroscopic force vector fn and associated load stiffness K Ln for a three-dimensional
beam column of length L currently directed along the global x axis. The element degrees of freedom are
For this “moment tilting” analysis it is sufficient to assume that: (a) node 1 stays fixed, (b) the element remain
straight, and (c) any deviations from the current x direction are infinitesimal.
29–11
28
.
Imperfections
28–1
Chapter 28: IMPERFECTIONS 28–2
TABLE OF CONTENTS
Page
§28.1. No Body is Perfect 28–3
§28.2. The Imperfect Hinged Cantilever 28–3
§28.2.1. Equilibrium Analysis . . . . . . . . . . . . . . . 28–3
§28.2.2. Critical Point Analysis . . . . . . . . . . . . . . 28–3
§28.2.3. Discussion . . . . . . . . . . . . . . . . . . 28–4
§28.3. The Imperfect Propped Cantilever 28–5
§28.4. Parametrizing Imperfections 28–7
§28.5. Imperfection Sensitivity at Critical Points 28–8
§28.5.1. Limit Point . . . . . . . . . . . . . . . . . 28–9
§28.5.2. Asymmetric Bifurcation . . . . . . . . . . . . . . 28–9
§28.5.3. Stable Symmetric Bifurcation . . . . . . . . . . . 28–10
§28.5.4. Unstable Symmetric Bifurcation . . . . . . . . . . . 28–10
§28.6. Extensions: Multiple Bifurcation, Continuous Systems 28–10
§28. Exercises . . . . . . . . . . . . . . . . . . . . . . 28–12
28–2
28–3 §28.2 THE IMPERFECT HINGED CANTILEVER
28–3
Chapter 28: IMPERFECTIONS 28–4
ε p
L rigid
;;;;
k
K θ̇ − q λ̇ = 0, (28.5)
where
∂r ∂r
K = = k(1 − λ cos θ ), q= = k sin θ. (28.6)
∂θ ∂θ
We have stability if K > 0, that is
1 − λ cos θ > 0, (28.7)
1
λcr = . (28.9)
cos θ
On equating this value of λ with that given by the equilibrium solution (28.4) we obtain
θ − = tan θ. (28.10)
This relation characterizes the locus of critical points as is varied. It is not difficult to show that these critical
points are limit points if = 0 (imperfect systems) and a bifurcation point if and only if = 0 (perfect system).
28–4
28–5 §28.3 THE IMPERFECT PROPPED CANTILEVER
1.25
0.75 Stable
<0 >0
0.5
0.25
θ (rad)
-2 -1 0 1 2
§28.2.3. Discussion
28–5
Chapter 28: IMPERFECTIONS 28–6
u = L sin θ
;; k p
L(1 − cos θ)
; θ
L
;;
rigid column
;;
Figure 28.3. The perfect propped cantilever.
L u = L sin θ
; k p
;
L(1 − cos θ)
θ
L
;;
Figure 28.4. The imperfect propped cantilever. The imperfection
parameter is , where L is the displacement from the
vertical at which the rectilinear spring is unstrained.
where u = L sin θ is the total horizontal displacement from the vertical, and a constant term has been dropped
from V . It is convenient to take the ratio λ = f L/k as dimensionless control parameter and µ = u/L as the
28–6
28–7 §28.4 PARAMETRIZING IMPERFECTIONS
1.4
>0 <0
1.2
Unstable
1
0.8
Stable
0.6
0.4
<0 >0
0.2
µ
-1 -0.5 0 0.5 1
dimensionless state variable. Then the potential energy, upon dividing by k L 2 , becomes
(µ, λ) = 12 (µ − )2 − λ 1 − µ2 (28.12)
Carrying out the analysis as in the previous section, one finds the response paths depicted in Figure 28.5.
The important difference is that the bifurcation point is now of unstable-symmetric type. The equilibrium
paths that emerge from the unloaded state of the imperfect systems are no longer rising but exhibit limit-point
maxima that may be viewed as failure loads. These limit points occur at lower loads than the bifurcation load
of the perfect structure.
Therefore, the load-carrying capacity of the propped cantilever is adverseley affected by the imperfection, and
the structure is said to be imperfection-sensitive.
The treatment of imperfections in the foregoing two examples illustrates many features that recur
in more complex cases. Imperfect systems are derived as perturbations of the perfect system.
Imperfections in real structures are seldom known precisely. They are usually random quantities
that can be rigorously treated only by stochastic techniques. Such a treatment, however, would be
hopelessly expensive in nonlinear systems. A more practical deterministic approach consists of
looking at a parametrized family of imperfect systems characterized by a dimensionless imperfection
parameter .
28–7
Chapter 28: IMPERFECTIONS 28–8
λ λcr
<0
L
>0 =0
µ
The effect of imperfections on the load-carrying capacity of a structure that fails at a critical point
may drastically vary according to the type of critical point. Structures whose failure loads are
substantially reduced by imperfections are called imperfection sensitive. In this Chapter we review
the question of sensitivity for the four types of critical points introduced in Chapter 23, using typical
response plots.
In the following two-dimensional response plots, the dimensionless control parameter λ is plotted
against a representative state parameter. This λ is assumed to be a load multiplier or load factor that
characterizes the strength of the structure. Following the same conventions as in the example of
§25.2 heavy lines represent the equilibrium path of theperfect system while light lines represent the
equilibrium paths of imperfect systems. Furthermore continuous lines represent stable equilibrium
path segments whereas broken lines represent unstable equilibrium segments.
28–8
28–9 §28.5 IMPERFECTION SENSITIVITY AT CRITICAL POINTS
λ λcr
>0
<0
B
<0
>0
µ
28–9
Chapter 28: IMPERFECTIONS 28–10
λ
λcr
>0 <0
= 0
µ
critical load level of the perfect system is reached. It follows that there is no local branch of the
λcr () curve for < 0.
Typical pictures for an stable symmetric point of bifurcation are shown in Figure 28.8. The general
behavior is similar to that encountered in the hinged-cantilever example of §25.2. We see that
imperfections play here a relatively minor role in changing the response of the system. Small
positive and small negative imperfections have similar effects, each yielding a continuously stable
and rising equilibrium paths as shown. Therefore, imperfect systems of this type display no sharp
failure load, “buckling” being simply characterized by a more rapid growth of the deflections as the
critical load of the perfect system is approached.
Finally, typical pictures for an unstable symmetric point of bifurcation are shown in Figure 28.9. We
see that imperfections play here a significant role in modifying the behavior of the system although
the effect is not so drastic as in the asymmetric bifurcation case. Small positive and small negative
imperfections have symmetrical effects, each now inducing failure at a limit point that corresponds
to a considerably reduced value of λcr (). The variation of λcr with is shown in the right-hand
diagram. For both positive and negative imperfections the function follows locally (that is, near
= 0) the so-called “two-thirds law:”
λ ∝ 2/3 , (28.14)
discovered originally by Koiter in the 1940s. This law yields a sharp “cusp“ at = 0 as shown. We
can summarize this case as being one of high imperfection sensitivity to both positive and negative
imperfections.
28–10
28–11 §28.6 EXTENSIONS: MULTIPLE BIFURCATION, CONTINUOUS SYSTEMS
λ λcr
>0 <0
B
<0 >0
µ
The preceding discussion, as well as the treatment of Chapter 24, pertains to limit points and isolated
bifurcation points of discrete structural systems. This rises questions as to what happens at multiple
or compound bifurcation points, or critical points of continuous systems.
At a multiple bifurcation point of order k (order being the rank deficiency of the tangent stiffness
matrix there) a second-order analysis similar to that carried out in §27.4 show that one can expect
in general 2k crossing branches. For an isolated bifurcation point k = 1 and 2k = 2 as previously
found. But if k ≥ 2 much greater complexity of behavior can be expected. At a multiple bifurcation
point the effect of initial imperfections is typically far more severe than at single bifurcation points
if one or more of the emanating branches are unstable, as is usually the case. The systematic
investigation of these effects remains a frontier research subject, which is nonetheless gaining in
importance because of its application to stability-optimized aerospace structures that are designed
to simultaneously fail in several buckling modes.
As for continuous systems, the aforementioned two-thirds power law for the imperfection sensitivity
at an unstable-symmetric bifurcation point carries over to continuous systems if the geometric
imperfection is assumed to have the shape of the buckling mode. If this common assumption is not
made new power laws may emerge from the continuous analysis. As for asymmetric buckling, it
remains poorly understood in the continuous case.
28–11
Chapter 28: IMPERFECTIONS 28–12
EXERCISE 28.1 Work out the details of the analysis of the imperfect propped cantilever described in §28.3.
In particular, verify that the diagram shown in Figure 28.5 is correct. Obtain the equation of the imperfection
sensitivity diagram λcr () where λcr are load limits (limit points) obtained when = 0. Plot this diagram for
values of = 0 to 1; note vertical-tangent “cusp” at = 0! [Qualitatively, the diagram should look like the
one in Figure 28.9.]
EXERCISE 28.2 A highly simplified, one-DOF “beer-can-like” structure has the total potential function
(u, f, ) = U − V, U = k L 2 ( 1 + (u/L) − 1 − γ )2 , V = f L 1 + (u/L) − 1 − (u/L)2 (E28.1)
where u is the state variable, f is the applied load, k, L are structure property constants with dimensions of
spring constant and length, respectively, γ is a dimensionless constant, and is a dimensionless geometric
imperfection parameter. Reduce (E28.1) to a dimensionless form
(µ, λ, ) (E28.2)
by defining the state parameter µ = u/L and the control parameter λ = f /(k L), and dividing the whole
thing through by k L 2 . From then on take γ = 0. Form the equilibrium equations and generate a response
diagram for the imperfect structure similar to those shown in Figures 28.2 and 28.5, varying in the range
(-1,1), µ in the range (-0.99,0.99), and λ in (0,1.5). From visual inspection conclude whether the structure
experiences asymmetric bifurcation (that is, it has a preferred buckling direction) or a symmetric one. Draw
the imperfection sensitivity diagram of λcr versus . What is the load capacity drop for imperfections of
magnitude 0.01, 0.1 and 1?
Use of built-in graphic tools such as those provided in Matlab, Mathematica or Maple can speed up signifi-
cantly the generation of response diagrams for Exercises 28.1 and 28.2. For example, an initial version of the
diagram shown in Figure 28.2 was produced by the following Mathematica script:
lam[theta_,eps_]:=(theta-eps)/Sin[theta];
p1=Plot[{lam[theta,0.01],lam[theta,-0.01],lam[theta,0.1],
lam[theta,-0.1],lam[theta,0.2],lam[theta,-0.2],lam[theta,0.5],lam[theta,-0.5]},
{theta,-Pi/1.2,Pi/1.2},PlotRange->{0,2},DisplayFunction->Identity];
p2=Plot[lam[theta,0],{theta,-Pi/1.2,Pi/1.2},PlotRange->{0,2},DisplayFunction->Identity];
p3=Plot[1/Cos[theta],{theta,-1.5,1.5},PlotRange->{0,2},DisplayFunction->Identity];
Show[Graphics[Thickness[0.002]],p1, Graphics[Thickness[0.004]],p2,
Graphics[Thickness[0.004]],Graphics[AbsoluteDashing[{5,5}]],p3,
PlotRange->{0,2},Axes->True,AxesLabel->{"theta","lambda"},
DisplayFunction->$DisplayFunction];
The plot cell was then converted and saved as an Adobe Illustrator 88 file, picked up by Adobe Illustrator 6.0
and “massaged” for bells and wistles such as Greek labels, dashed lines, shading of unstable region, etc.
28–12
27
.
Nonlinear
Bifurcation
Analysis
27–1
Chapter 27: NONLINEAR BIFURCATION ANALYSIS 27–2
TABLE OF CONTENTS
Page
§27.1. Introduction 27–3
§27.2. Levels of Bifurcation Analysis 27–3
§27.3. Recapitulation of Governing Equations 27–3
§27.3.1. Residual and Rate Equations . . . . . . . . . . . . 27–3
§27.3.2. Stiffness and Load Rates . . . . . . . . . . . . . 27–5
§27.3.3. Limitations of λ-Parametrized Forms . . . . . . . . . 27–5
§27.4. A Deeper Look at Bifurcation 27–5
§27.4.1. State Decomposition . . . . . . . . . . . . . . 27–6
§27.4.2. Failure of first-order rate equations at bifurcation . . . . . 27–6
§27.5. Branch Analysis of Simple Bifurcation 27–6
§27.5.1. State Decomposition . . . . . . . . . . . . . . 27–6
§27.5.2. Finding σ . . . . . . . . . . . . . . . . . . 27–8
§27.6. The Hinged Cantilever 27–10
§27.6.1. Finding the Critical Point . . . . . . . . . . . . . 27–11
§27.6.2. Branching Analysis . . . . . . . . . . . . . . . 27–11
§27. Exercises . . . . . . . . . . . . . . . . . . . . . . 27–14
27–2
27–3 §27.3 RECAPITULATION OF GOVERNING EQUATIONS
§27.1. Introduction
We initiated our study of stability of conservative systems in Chapters 24-25 by using the simplified
model of linearized prebuckling (LPB). This was followed in Chapter 26 by a qualitative study of
the more general stability model that led to classifying isolated critical points into four types: limit
point, asymmetric bifurcation, stable-symmetric bifurcation, and unstable-symmetric bifurcation.
This classification is especially helpful in understanding the effect of imperfections on stability.
This Chapter presents a more detailed mathematical analysis of the phenomenon of bifurcation by
studying equilibrium branches in the vicinity of an isolated bifurcation point. The topic is covered
under the name nonlinear bifurcation to emphasize that we are dealing with the general case as
opposed to the LPB model. A simple example involving a one-degree of freedom system is then
worked out in some detail. The next Chapter takes up the subject of how physical or numerical
imperfections affect structural behavior as regards both limit and bifurcation points.
Nonlinear bifurcation analysis can be carried out at different levels of detail, as demanded by
application needs.1 Four levels of increasing detail are schematized in Figure 27.1, which assume
the occurrence of an isolated bifurcation point B.
1. Locate: find where B occurs while tracing a response. Can be done by monitoring changes
of sign of the determinant of K or equivalently tracing the sign of factorization pivots (See
Chapter 21).
2. Determine subspace: having located B, determine vectors y (particular solution) and z (null
eigenvector for isolated bifurcation) that together with λ form an intrinsic subspace “where
the action is.” Requires a partial eigensolution; more precisely getting the null eigenvector(s).
3. Branching analysis: having located B, and computed y and z, find the directions u̇1 and u̇2 of
tangents to the equilibrium paths (branches) that pass through B. Requires an analysis of the
second order rate equations r̈ = 0.
4. Branch curvature analysis: having located B and determined y, z, u̇1 and u̇2 , find the curvatures
of the equilibrium paths (branches) passing through B. Requires an analysis of the third order
...
equation r = 0.
The information necessary for level 3 is quite difficult to obtainfrom a general purpose finite element
programs, while that needed for level 4 is truly inaccesible. For this reason most FE programs can
provide only levels 1 and 2 on a routine basis. In the present Chapter we study up to level 3
(branching analysis), but the practical difficulties of implementing that level should be kept in
mind.
Below we recapitulate discrete governing equations derived in Chapters 3 and 4, and introduce
additional nomenclature required for the branching analysis carried out in §27.4.
1 For example in preliminary design only the location of the first bifurcation point would be of interest.
27–3
Chapter 27: NONLINEAR BIFURCATION ANALYSIS 27–4
y
B Determine
90 ◦
1 (Eigensolution)
active subspace
z
u̇2
Branching
B 2 if isolated
u˙ 1 analysis
Branch
B curvature 3 if isolated
analysis
Figure 27.1. The four levels of information for nonlinear bifurcation analysis.
r(u, λ) = 0, (27.1)
where λ is the stage control parameter and u is the state vector. Solutions of this equation may be
conveniently represented in parametric form
where t is a dimensionless path parameter. Two important special choices for pseudotime are
t = λ, t = s, (27.3)
27–4
27–5 §27.4 A DEEPER LOOK AT BIFURCATION
Rate equations are systems of ordinary differential equations obtained by successive differentiation
of (27.1) with respect to t. Recall that K and q denote the tangent stiffness matrix and incremental
load vector, respectively, whose entries are given by
∂ri ∂ri
Ki j = , qi = . (27.4)
∂u j ∂λ
Using superposed dots to denote t-differentiation we obtain
∂ 2 ri ∂ 2 ri ∂ Ki j ∂ 2 ri ∂qi
L i jk = , Ni j = = , ai = = . (27.9)
∂u j ∂u k ∂u j ∂λ ∂λ ∂λ∂λ ∂λ
Remark 27.1. Note that L is a three-dimensional array which may be called a cubic matrix to distinguish it
from an ordinary square matrix. (Also referred to as a third order tensor.) Postmultiplying a cubic matrix by
a vector yields an ordinary matrix. For example Lu̇ is a matrix.
27–5
Chapter 27: NONLINEAR BIFURCATION ANALYSIS 27–6
At critical points K becomes singular and therefore possesses at least one null eigenvector, which
as usual is called z. This eigenvector will be normalized to unit length. As discussed in Chapter 5,
if the eigenvector is orthogonal (non-orthogonal) to the incremental load vector, the critical point
is a bifurcation (limit) point.
Kz = 0, z = zT z = 1, qT z = 0. (27.13)
u = (σ z + y) λ, (27.14)
where y is the particular solution introduced in §25.2, and σ is the buckling mode amplitude.
Dividing by t and passing to the limit t → 0 we obtain the rate form of the above equation:
u̇ = (σ z + y)λ̇. (27.15)
This decomposition of u̇ in the y, z plane is depicted in Figure 27.2, a duplicate of Figure 25.1.
At bifurcation points the first-order rate equations (27.5) yield no information on the buckling
mode amplitude. This is worked out in Exercise 25.3, which shows that σ = 0/0 and is therefore
indeterminate. To get deterministic information in the vicinity of a bifurcation point it is necessary
to use information from higher-order rate equations. This is covered in the following subsection
for a isolated (simple, distinct) bifurcation point. For this case the second-order rate equations
(27.6) are usually sufficient.
27–6
27–7 §27.5 BRANCH ANALYSIS OF SIMPLE BIFURCATION
zT y = 0
||z||2 = 1
σ zλ̇ u̇
B y
yλ̇
Figure 27.2. State decomposition at isolated bifurcation point B, depicted in the (y, z) plane.
27–7
Chapter 27: NONLINEAR BIFURCATION ANALYSIS 27–8
Branch 2
2
λ
1
nt
ch
ge
an
n
Ta
Br
y
u˙ 2 λ̇ 1
Tangent
B λ̇
u˙ 1
yλ̇
z
Remark 27.3. The decomposition (27.16) is analogous in many respects to the decomposition of element
motions into purely-deformational and rigid-body, studied in Chapter 10. Here z take the role of rigid body
mode. The decomposition (27.16) is, however, expressed in terms of rates because it is local: it is restricted
to the vicinity of the bifurcation point.
§27.5.2. Finding σ
The first-level information on the equilibrium branches at B is given by their tangents at B. Because
we can obtain y and z from the first-order rate equations, these tangents are fully determined if in
addition we know σ , u̇ and λ̇ at B.
But as noted previously the first-order rate equations (27.5) do not provide information on the
buckling mode amplitude σ . To get that information it is necessary to go to the second-order
system (27.6), which is repeated here for convenience:
Kü + K̇u̇ − qλ̈ − q̇λ̇ = 0 (27.18)
Premultiplying both sides by zT and taking account of the bifurcation conditions (27.13) we get at
B the scalar equation
zT K̇u̇ − zT q̇λ̇ = 0 (27.19)
Replacement of K̇ and q̇ by the expressions (27.8) gives
zT (Lu̇ + Nλ̇)u̇ + zT (Nu̇ + aλ̇)λ̇ = 0. (27.20)
Finally, substitution of u̇ by its homogeneous-plus-particular decomposition (27.16) yields
zT L(y + σ z)λ̇ + Nλ̇ (y + σ z)λ̇ + zT N(y + σ z)λ̇ + aλ̇ λ̇ = 0. (27.21)
27–8
27–9 §27.5 BRANCH ANALYSIS OF SIMPLE BIFURCATION
Removing the common differential factor (λ̇)2 and collecting terms in σ we arrive at the quadratic
equation
aσ 2 + 2bσ + c = 0, (27.22)
in which
a = zT Lzz, b = zT [Lzy + Lyz + 2Nz] , c = zT [Lyy + 2Ny + a] . (27.23)
This quadratic equation generally provides two roots: σ1 and σ2 . In what follows we shall assume
that these two roots are real (see Remark 27.5 below).
Substitution of σ1 and σ2 into (27.16) furnishes the branching directions at the bifurcation point:
u̇1 = (y + σ1 z)λ̇, u̇2 = (y + σ2 z)λ̇. (27.24)
These are sketched in Figure 27.3 in the three-dimensional space (y, z, λ) with origin at B. Figure
27.4 projects this picture onto the (y, z) plane for additional clarity.
y
Ta
ng
ent
1
nt
2
ge
Bra
n
nch
Ta
2
1
σ2 zλ̇ σ1 zλ̇ Branch
u̇2 u˙ 1
yλ̇
B z
Figure 27.4. Same as Figure 27.3 but looking down the λ axis onto the (y, z) plane.
The key result of this subsection is that there are at most two branches emanating from a simple
bifurcation point. The classification of such points into asymmetric and symmetric bifurcation
points according to the values of σ1 and σ2 appears in the Exercises. In the following section an
illustrative example is worked out by hand.
27–9
Chapter 27: NONLINEAR BIFURCATION ANALYSIS 27–10
P=λq
A
θ
L
rigid
;;;;
k
Remark 27.4. If a = 0 one root, say σ2 , becomes infinite while the other is σ1 = −c/(2b), assuming b = 0.
Then u̇2 becomes aligned with z. Only in this case it is justified to call z the “buckling mode.”
Remark 27.5. Intuitively it appears that the two roots of (27.22) must be real. The argument goes as follows:
one of the two branches is supposed to exist since B has been located by hypothesis on the equilibrium path.
Its tangent at B must therefore correspond to one of the roots of (27.22). Since one of the roots is by hypothesis
real, the other must also be real because a, b and c are real coefficients.
This indirect proof is not intellectually satisfying, especially to a mathematician. It would be preferable to
prove the root reality by direct reasoning. However the writer has not been able to find such a proof in the
literature, and personal efforts (one hour trying) have been so far unrewarding.
Remark 27.6. If a = b = c = 0 the second-order rate form (27.6) does not provide any local information
as regards branches at B. Then one must continue to the third order rate form (27.7). This will give a cubic
equation in σ with four real coefficients. Since such an equation can have one or three real roots, things get far
more complicated. If all four coefficients vanish, one must go to the fourth-order rate form, and so on. (For a
mathematician specialized in this kind of analysis, hell is a place where the first one million rate forms yield
no information.)
27–10
27–11 §27.6 THE HINGED CANTILEVER
∂
r= = k(θ − λ sin θ ) = 0, (27.26)
∂θ
This has the two solutions
θ
θ = 0, λ= , (27.27)
sin θ
which pertain to the primary (vertical or untilted) and secondary (tilted) equilibrium paths, respectively. The
two paths intersect at λ = 1, which is therefore a bifurcation point.
λ = 1, or P = k/L . (27.30)
On this path the stiffness is positive (negative) if λ < 1 (λ > 1), respectively. On the secondary path,
λ = θ/ sin θ, the stiffness is given by
θ cos θ
K =k 1− , (27.31)
sin θ
which vanishes at θ = 0 because θ/ sin θ → 1 as θ → 0. If θ = 0, K > 0. The various cases as regards the
sign of K are summarized in Figure 27.6. Because K is a scalar, positive and negative values corresponds to
stable and unstable equilibrium, respectively, with neutral stability at B. Stable (unstable) paths are showns
with full (dashed) lines.
It is seen that θ = 0◦ and λ = 1 is the only point at which K vanishes, and consequently is the only critical
point. Let us verify now that the critical point is a bifurcation point. Since the system has only one degree of
freedom, the normalized null eigenvector is simply the scalar z = 1, and the inner product zT q reduces to
zq = q = k sin θ (27.32)
In this problem the particular solution y vanishes because there is only one degree of freedom. We may
therefore take
θ̇ = σ zλ̇ = σ λ̇ (27.33)
The second-order rate equation is
λ sin θ θ̇ θ̇ − 2 cos θ θ̇ λ̇ = 0,
which upon substituting θ̇ = σ λ̇ yields the quadratic equation (27.22) with a = λ sin θ, b = −2, c = 0. At
the bifurcation point (λ = 1, θ = 0) we get
27–11
Chapter 27: NONLINEAR BIFURCATION ANALYSIS 27–12
K <0
Secondary path
B K >0
K>0
K =0
K >0
Primary (fundamental) path
R
θ
Figure 27.6. The sign of the stiffness coefficient K for the hinged cantilever response.
σ1 = 0, σ2 = ∞, (27.35)
θ̇ = 0, λ̇ = 0. (27.36)
These branches are the tangents to the primary (vertical bar) and secondary (tilted bar), respectively, at the
bifurcation point. See Figure 27.7.
This Figure also sketches the post-buckling response, which for this problem is easily obtained from the
exact equilibrium solutions (27.27). According to the qualitative classification of Chapter 26, the bifurcation
point is of stable-symmetric type. This subclassification of a symmetric bifurcation point into stable and
unstable cannot be discerned, however, from the branch-tangent analysis, because it requires information on
the curvature of the z-directed branch.
27–12
27–13 §27.6 THE HINGED CANTILEVER
Tangent 1: θ̇ = 0
Secondary path
Tangent 2: λ̇ = 0
R
θ
Figure 27.7. The two branch directions at bifurcation point of the hinged cantilever.
27–13
Chapter 27: NONLINEAR BIFURCATION ANALYSIS 27–14
∂K ∂K ∂q
L= , N= , a=− . (E27.1)
∂u ∂λ ∂λ
Are these relations true?
EXERCISE 27.2 [A:20] If a → 0 in the quadratic equation (27.22) while b = 0, one of the roots, say σ1 , goes
to ∞ whereas the other one becomes σ2 = −c/2b. This is called a symmetric bifurcation. Show that in such
a case the branch direction corresponding to σ1 coincides with the buckling mode z, and draw a bifurcation
diagram similar to Figure 27.1.
EXERCISE 27.3 [A:40] Algebraically prove that the roots of the quadratic equation (27.5) are real2
EXERCISE 27.4 [A:25] The LPB first order rate equations are ṙ = Ku̇ − qλ̇ = 0, in which K = K0 + λK1
and where K0 , K1 and q are constant. Using Exercise 25.3 (posted solution), show that LPB can only predict
symmetric bifurcation. What wonderful thing happens if K1 y = 0?
EXERCISE 27.5 [A:20] The propped cantilever shown in Figure 28.3 consists of a rigid bar of length L
pinned at A and supported by a linear extensional spring of stiffness k. The spring is assumed to be capable
of resisting both tension and compression and retains its horizontal orientation as the system deflects. The bar
may rotate all the way around the pin. The rigid bar is subjected to a vertical dead load P that remains vertical.
Define dimensionless control and state parameters as
P
λ= , µ = sin θ. (E27.2)
kL
Analyze the stability of the propped cantilever in a manner similar to §27.5. Show that the secondary equi-
librium path is the circle λ2 + µ2 = 1 and sketch the response paths showing the complete circle. From this
diagram, can you tell whether the bifurcation point at λ = 1 is stable-symmetric or unstable-symmetric? How
about the one at λ = −1?
2 A very difficult assignment worth of a paper. I am not aware of anybody that has done for the general case.
27–14
26
.
Qualitative
Analysis of
Critical Points
26–1
Chapter 26: QUALITATIVE ANALYSIS OF CRITICAL POINTS 26–2
TABLE OF CONTENTS
Page
§26.1. General Notion of Stability 26–3
§26.2. Stability of a Discrete Conservative System 26–3
§26.3. Stability Transformation at a Limit Point 26–4
§26.4. Stability Exchange at Bifurcation Points 26–7
§26.4.1. Asymmetric bifurcation . . . . . . . . . . . . . . 26–7
§26.4.2. Stable-symmetric bifurcation . . . . . . . . . . . . 26–8
§26.4.3. Unstable-symmetric bifurcation . . . . . . . . . . . 26–8
26–2
26–3 §26.2 STABILITY OF A DISCRETE CONSERVATIVE SYSTEM
In Chapter 24 stability was informally defined as the ability of a physical system to return to
equilibrium when disturbed. If the equilibrium is static in nature, we speak of static stability. For
a more precise definition concerning a mechanical system, let’s hear Dirichlet:1
“The equilibrium [of a mechanical system] is stable if, in displacing the points of the system from
their equilibrium positions by an infinitesimal amount and giving each one a small initial velocity,
the displacements of different points of the system remain, throughout the course of the motion,
contained within small prescribed limits”
As discussed in Chapter 24, the static stability of a conservative mechanical system can be tested
completely using a static criterion. Such criterion, often referred to as the Euler stability test, the
energy test, or the method of adjacent states, relates to the positive definiteness character of the
second variation of the potential energy.
We know that a stationary value of the total potential energy with respect to the state variables is
necessary and sufficient for the equilibrium of the system. Proceeding one step further, a complete
relative minimum of the total potential energy is necessary and sufficient for the stability of an
equilibrium state.
For a discrete system with a finite number of degrees of freedom the criterion can be enunciated in
terms of the definiteness of the tangent stiffness matrix K if all state variables are of displacement
type, which we assume in the sequel (see Remark below). For a conservative system we know that
K is a symmetric matrix.
1 As it appears in his Appendix to the German translation of Lagrange’s Mécanique Analytique (1853).
26–3
Chapter 26: QUALITATIVE ANALYSIS OF CRITICAL POINTS 26–4
Remark 26.1. The restriction to systems with displacement state variables aims to exclude those in which
Lagrange multipliers are carried along as degrees of freedom. For such systems the criterion applies upon
eliminating the multipliers, but such elimination is often messy and would complicate the exposition.
The stability criterion for a conservative system is summarized in the following table.
If the eigenvalues of K are easily available a test for stability is immediate.2 If all eigenvalues
are greater than zero, the matrix is positive definite and the equilibrium is stable. If one or more
eigenvalues are zero and the rest positive, the equilibrium is neutrally stable. If one or more
eigenvalues are negative, the equilibrium is unstable.
In practice an eigenvalue test can be recommended only for small matrices, say of order less than 20
or so. For larger matrices the same information can be obtained more economically by decomposing
K using triangular factorization or Gauss elimination, as discussed in Remark 24.2. If all pivots
are positive, the equilibrium is stable. If at least one pivot is negative, the equilibrium is unstable.
The border case of neutral stability is more difficult to detect in the presence of rounding errors.
26–4
26–5 §26.3 STABILITY TRANSFORMATION AT A LIMIT POINT
L1
( u1 , u 2 , λ1 )
u2 L2
u1
26–5
Chapter 26: QUALITATIVE ANALYSIS OF CRITICAL POINTS 26–6
Secondary path
(u, λcr )
λcr λ (u, λ1 ), λ1 < λcr
prescribed value of λ is increased the maximum and minimum approach each other and coalesce
when λ = λcr . At this critical point the total potential energy (u, λ) has an horizontal point of
inflexion. At a higher value of λ, say λ2 > λcr , there are no local equilibrium states and the total
potential energy (u, λ2 ) has no stationary point. The critical equilibrium state is seen to be itself
unstable, and the absence of local equilibrium states at values of λ greater than λcr implies that a
physical system under slowly increasing λ will eventually snap-through dynamically. Limit points
are generally insensitive to imperfections.
A more general schematic diagram is shown in three dimensions in Figure 26.2 on a plot of λ
against two state parameters, v1 and v2 . This plot includes a remote rising region of the equilibrium
path since this is often encountered with this type of behavior. A total potential energy surface
(v1 , v2 , λ) is drawn for a fixed value of λ < λcr . As λ is slowly increased through its critical
value the system will “snap through” dynamically, eventually stop, and initiate a large amplitude,
nonlinear vibration about the remote stable equilibrium path. In the presence of some damping the
system will eventually rest on that path.
These figures illustrate the physics well but if we are dealing with a system with many degrees of
freedom care must be taken in drawing conclusions from these schematic figures. On an actual plot
of λ against one of the vi ’s, the limit point is normally seen as a smooth maximum, but it must be
realized that for a certain choice of the state parameter vi the point might appear as a sharp cusp.
The smooth maximum of a path in three-dimensional space can for example be seen as a cusp if
the eye is directed along the horizontal tangent to the path.
26–6
26–7 §26.4 STABILITY EXCHANGE AT BIFURCATION POINTS
Secondary path
(u, λ2 ), λ2 > λcr
(u, λcr )
B
(u, λ1 ), λ1 < λcr
λcr
λ
Primary (fundamental) path
After the limit point we consider bifurcation or branching points. We cover the three most common
types of bifurcation: asymmetric, stable-symmetric, and unstable symmetric.
26–7
Chapter 26: QUALITATIVE ANALYSIS OF CRITICAL POINTS 26–8
(u , λcr )
Symmetric bifurcation points are characterized by the fact that the intersecting path has zero slope
with respect to the control parameter at B. These points may be categorized into stable and unstable,
depending on whether the intersecting post-buckling) path is “rising” or “falling”.
Figure 26.4 depicts the case of an stable-symmetric point of bifurcation. Here a fundamental
equilibrium path rising monotonically from the reference state is seen to intersect a stable rising
secondary (post-buckling) path that passes smoothly through the critical equilibrium state with zero
slope. The continuation of the fundamental path beyond B is unstable. The total potential energy
(u, λ1 ), where λ1 < λcr , has a single stationary value with respect to u, namely the minimum
on the stable region of the fundamental path, and as the value of λ is increased this minimum is
transformed into two minima and one maximum. The critical equilibrium state is neutrally stable
and the secondary path is stable, so a physical system under slowly increasing λ would exhibit no
dynamic snap but would follow the stable rising post-buckling path, the direction taken depending
on the small disturbances or imperfections which are inevitably present.
Critical points of this type are insensitive to the presence of imperfections.
The last configuration examined here is the unstable-symmetric point of bifurcation, shown in
Figure 26.5. Here the fundamental path intersects an unstable falling path which as in the previous
case has a zero slope at the critical equilibrium state. At a prescribed value of λ = λ1 < λcr
the total potential energy (u, λ1 ) has now three stationary values with respect to u, namely two
26–8
26–9 §26.4 STABILITY EXCHANGE AT BIFURCATION POINTS
maxima on the unstable post-buckling or secondary path, and a minimum on the stable region of
the fundamental path. As the figure shows, these three stationary points transform into a single
maximum with increasing λ. The critical equilibrium state is seen to be unstable, so a physical
system would snap dynamically from the critical equilibrium state, the direction taken depending
on the postulated small disturbances or imperfections.
Critical points of this type are highly sensitive to the presence of structural or loading imperfections.
Sometimes the sensitivity is extreme, as in the classical case of the axially compressed cylindrical
shell discussed in §25.4.
26–9
25
.
Bifurcation:
Linearized
Prebuckling II
25–1
Chapter 25: BIFURCATION: LINEARIZED PREBUCKLING II 25–2
TABLE OF CONTENTS
Page
§25.1. Introduction 25–3
§25.2. State Decomposition at Bifurcation Point 25–3
§25.3. LPB Assumptions 25–3
§25.4. Limitations of LPB 25–5
§25.4.1. When LPB Works . . . . . . . . . . . . . . . . 25–6
§25.4.2. And When It Doesn’t . . . . . . . . . . . . . . 25–7
§25.4.3. How to Extend the Applicability of LPB . . . . . . . . 25–8
§25. Exercises . . . . . . . . . . . . . . . . . . . . . . 25–9
25–2
25–3 §25.3 LPB ASSUMPTIONS
§25.1. Introduction
This Chapter continues with the subject of linearized prebuckling (LPB) bifurcation analysis. It
goes deeper than Chapter 24 in that it probes the assumptions (so far stated without proof) behind
LPB, and the practical modeling implications that emanate from these assumptions.
To present some of the derivations in mathematical terms it is necessary to introduce the concept
of state decomposition at the bifurcation point and to define homogeneous and particular solutions.
This is done briefly in §24.2 primarily as a means of introducing notation for S24.3 and following.
The detailed mathematical analysis of this decomposition is relegated to Chapter 26.
Recall from previous Chapters that an isolated bifurcation point at λcr is characterized by a singular
tangent stiffness at the equilibrium configuration,
and by the normalized null eigenvector (buckling mode) z = 0, z = 1, being orthogonal to the
incremental load vector:
qT z = zT q = 0. (25.2)
Assume that we have located a bifurcation point B and computed the buckling mode z. Our next
task is to examine the structural behavior in the neighborhood of B. We shall be content with looking
at the so-called branching direction information. This information characterizes the tangents to the
equilibrium branches that cross at B.
To carry out this task we borrow from algebraic ODE theory. Consider the variation in the state
vector u measured from its value u B at buckling:
u = u − u B (25.3)
Divide this increment by t, t being the timelike parameter introduced in Chapter 3, and pass to
the limit:
u
u̇ = lim . (25.4)
t→0 t
This variation rate u̇ from the bifurcation point can be decomposed into a homogeneous solution
component σ z in the buckling mode direction, and a particular solution component y, which is
orthogonal to z:
u̇ = (y + σ z)λ̇, yT z = zT y = 0, (25.5)
The particular solution solves the system
Ky = q, yT z = 0, (25.6)
which is simply the first-order incremental equation Ku̇ = qλ̇ augmented by a normality con-
straint. Imposing this constraint removes the singularity (rank deficiency) of K. The geometric
interpretation of this decomposition on the y, z plane is shown in Figure 24.1.
25–3
Chapter 25: BIFURCATION: LINEARIZED PREBUCKLING II 25–4
zT y = 0
||z||2 = 1
σ zλ̇
B y
yλ̇
p = q0 + λq. (25.7)
and the structure is linearly elastic. Inother words, the residual equations are derivable from
a potential energy function.
(II) The displacements and displacement gradients prior to the critical state are negligible in the
sense that (a) the material stiffness matrix can be evaluated at the reference configuration,
and (b) the geometric stiffness is proportional to the control parameter λ:
K M ≡ K0 , KG ≡ λK1 , (25.8)
in which K0 is the material matrix evaluated at the reference configuration, also called the
linear stiffness, and K1 is the reference geometric stiffness. As discussed in the previous
Chapter the singular stiffness criterion det K = 0 leads to the eigenproblem
25–4
25–5 §25.4 LIMITATIONS OF LPB
25–5
Chapter 25: BIFURCATION: LINEARIZED PREBUCKLING II 25–6
;;;; ;; Figure 25.2. Structures that are adequately modeled by LPB assumptions.
negligible deformation
prior to buckling
R v
loses meaning. The topic of inelastic buckling (in particular creep and plastic buckling) is an
enormous subject that falls outside the scope of this course.
5. Applied loads should not depend nonlinearly on the displacements. Such a dependence usually
introduces nonconservative effects, thus voiding the conservative-loading assumptions. Even
is the loads remain conservative, the reference geometric stiffness would depend on the load
level, thus leading to a nonlinear eigenproblem.
6. The effect of imperfections is negligible. Some structures are highly imperfection sensitive in
that the first critical load is strongly affected by the presence of imperfections. In such cases
obviously LPB is of limited value or outright irrelevant.
25–6
25–7 §25.4 LIMITATIONS OF LPB
(a) (b)
λ L λ B
LPB
LPB Actual
B L Actual
R v R v
Figure 25.4. Two structures that fit the LPB assumptions poorly.
The systems that best fit the LPB model are symmetrically loaded structures such as straight columns
and in-plane-loaded plates (laminas) which are not excessively thin. See Figure 25.2. The lateral
buckling of such structures occurs following very small deformations, as typified by the response
sketch in Figure 25.3.
Two examples of structures that are not properly treated by the LPB model are shown in Figure
25.4. The LPB predictions are way off in both cases, but for different reasons.
Case (a) is an axially compressed cylindrical shell made up of almost flat panels joined by curved
panels, forming like a “curved triangle” cross section seen in some combat helicopters and the
Space Shuttle fuselage. There is a substantial redistribution of stresses due to changes on geometry.
The structure eventually collapses at a limit point substantially over the predicted LPB load. The
latter is therefore overly safe.
25–7
Chapter 25: BIFURCATION: LINEARIZED PREBUCKLING II 25–8
On the other hand, the axially compressed circular cylinder of case (b) is highly imperfection-
sensitive structure that fails at a substantially lower load than that predicted by LPB. Consequently
the LPB prediction is highly unsafe.
3 As naturally done in the CR description, in which the deformational displacements are measured from a continuously
varying configuration, and also in the Updated Lagrangian description.
25–8
25–9 Exercises
EXERCISE 25.1 [A:15] Find the particular solution y at the lowest bifurcation load of the two-bar example
of Chapter 24.
EXERCISE 25.2 [A:15] Find the particular solution y at the symmetric and antisymmetric bifurcation loads
of the one-element Euler column example of Exercise 24.4.
EXERCISE 25.3 [A:25] The first order residual rate equations is ṙ = 0, where ṙ is given by
(E25.1) holds at a bifurcation point where K and q are the tangent stiffness matrix and incremental load vector,
respectively, at bifurcation. Decompose u̇ = (y + σ z)λ̇, where y is the particular solution and z = 0 the
buckling mode normalized to length one. Show that the first-order differential equation system (E25.1) cannot
give information on the “buckling mode amplitude” σ because one gets σ = 0/0. (Hint: premultiply that
equation by an appropriate vector.)
25–9
24
.
Bifurcation:
Linearized
Prebuckling I
24–1
Chapter 24: BIFURCATION: LINEARIZED PREBUCKLING I 24–2
TABLE OF CONTENTS
Page
§24.1. Introduction 24–3
§24.2. Loss of Stability Criteria 24–4
§24.2.1. Static criterion . . . . . . . . . . . . . . . . . 24–4
§24.2.2. Dynamic criterion . . . . . . . . . . . . . . . 24–4
§24.3. The Tangent Stiffness Test 24–5
§24.4. Linearized Prebuckling 24–6
§24.5. The LPB Eigensystem 24–7
§24.6. Solving the Stability Eigenproblem 24–7
§24.7. LPB Analysis Example 24–8
§24.8. Summary of LPB Steps 24–11
§24. Exercises . . . . . . . . . . . . . . . . . . . . . . 24–12
24–2
24–3 §24.1 INTRODUCTION
§24.1. Introduction
This Chapter starts a systematic study of the stability of elastic structures. We shall postpone the
more rigorously mathematical definition of stability (or lack thereof) until later because the concept
is essentially dynamic in nature. For the moment the following physically intuitive concept should
suffice:
Note that this informal definition is dynamic in nature, because the words “returns” and “upon”
convey a sense of history. But it does not imply that the inertial and damping effects of true dynamics
are involved. So real time is not involved in the static case.
A structure that is initially stable may lose stability as it moves to another equilibrium position
when the control parameter(s) change. Under certain conditions, that transition is associated with
the occurrence of a critical point. These have been classified into limit points and bifurcation points
in Chapter 5.
For the slender structures that occur in aerospace, civil and mechanical engineering, bifurcation
points are more practically important than limit points. Consequently, attention will be initially
directed to the phenomena of bifurcation or branching of equilibrium states, a set of phenomena
also informally known as buckling. The analysis of what happens to the structure after it crosses a
bifurcation point is called post-buckling analysis.
The study of bifurcation and post-buckling while carrying out a full nonlinear analysis is a math-
ematically demanding subject. But in important cases the loss of stability of a geometrically
nonlinear structure by bifurcation can be assessed by solving linear algebraic eigenvalue problems
or “eigenproblems” for short. This eigenanalysis provides the magnitude of the loads (or, more
generally, of the control parameters) at which buckling is expected to occur. The analysis yields no
information on post-buckling behavior. Information on the buckling load levels is often sufficient,
however, for design purposes.
The present Chapter covers the source of such eigenproblems for conservatively loaded elastic
structures. Chapters 26 through 28 discuss stability in the context of full nonlinear analysis. The
two final Chapters (29–30) extend these concepts to structures under nonconservative loading.
Following a brief review of the stability assessment criteria the singular-stiffness test is described.
Attention is then focused on the particular form of this test that is most used in engineering practice:
the linearized prebuckling (LPB) analysis. The associated buckling eigenproblem is formulated.
The application of LPB on a simple problem is worked out using the bar element developed in
the previous three sections. The assumptions underlying LPB and its range of applicability are
discussed in the next Chapter.
24–3
Chapter 24: BIFURCATION: LINEARIZED PREBUCKLING I 24–4
For elastic, geometrically-nonlinear structures under static loading we can distinguish the following
techniques for stability assessment.
Static criterion (Euler method): singular stiffness
Conservative
Dynamic criterion: zero frequency
Loading
zero frequency (divergence)
Nonconservative Dynamic criterion
frequency coalescence (flutter)
The static criterion is also known as Euler’s method, since Euler introduced it in his famous
investigations of the elastica published in 1744. Other names for it are energy method and method
of adjacent states. To apply this criterion we look at admissible static perturbations of an equilibrium
position1 . These perturbations generate adjacent states or configurations, which are not generally
in equilibrium.
Stability is assessed by comparing the potential energy of these adjacent configurations with that
of the equilibrium position. If all adjacent states have a higher potential energy, the equilibrium is
stable. If at least one state has a lower (equal) potential energy the equilibrium is unstable (neutrally
stable). This comparison can be expressed in terms of the second variation of the potential energy
and hence can be reduced to the assessment of the positive definite character of the tangent stiffness
matrix.
Although stability is a dynamic phenomenon, no true-dynamics concepts such as mass or damping
are involved in the application of the static criterion, which is a key reason for its popularity. But
the reasoning behind it makes it strictly applicable only to conservatively loaded systems, because
a load potential function is assumed to exist.
The dynamic criterion looks at dynamic perturbations of the static equilibrium position. In informal
terms, “give the structure a (little) kick and see how it moves.” More precisely, we consider small
oscillations about the equilibrium position, and pose an eigenproblem that determines characteristic
exponents and associated eigenmodes. The characteristic exponents are generally complex num-
bers. If all characteristic exponents have no positive real components the equilibrium is dynamically
stable, and unstable otherwise.
These exponents change as the control parameter λ is varied. For sufficiently small values the
structure is stable. Loss of stability occurs when a characteristic exponent enters the right-hand
complex plane. If that happens, the associated mode viewed as a displacement pattern will amplify
exponentially in the course of time. A deeper study of the stable-to-unstable transition mechanism
1 “Admissible” in the sense of the Principle of Virtual Work: variations of the state parameters that are consistent with the
essential boundary conditions (kinematic constraints)
24–4
24–5 §24.3 THE TANGENT STIFFNESS TEST
reveals two types of instability phenomena, which are associated with the physically-oriented terms
terms divergence and flutter.
Divergence occurs when the characteristic exponent enters the right-hand plane through the origin,
and it can therefore be correlated with the zero frequency test and the singular stiffness test.
The dynamic criterion is applicable to both conservative and nonconservative systems. This wider
range of application is counterbalanced by the need of incorporating additional information (mass
and possibly damping) into the problem. Furthermore, unsymmetric eigenproblems arise in the
nonconservative case, and these are the source of many computational difficulties.
§24.3. The Tangent Stiffness Test
The stability of conservative systems can be assessed by looking at the spectrum2 of the tangent
stiffness matrix K. Let µi denote the i th eigenvalue of K. The set of µi ’s are the solution of the
algebraic eigenproblem
Kzi = µi zi . (24.1)
Since K is real symmetric3 all of its eigenvalues are real. Thus we can administer the following
test:
In engineering applications one is especially interested in the behavior of the structure as the stage
control parameter λ is varied, and so
K = K(λ). (24.2)
Given this dependence, a key information is the transition from stability to instability at the value
of λ closest to stage start, which is usually λ = 0. This is called the critical value of λ, which we
shall denote as λcr .
If the entries of K depend continuously on λ the eigenvalues of K also depend continuously4 on
λ, although the dependence is not necessarily continuously differentiable. It follows that transition
from strong stability — case (I) — to instability — case (III) — has to go through case (II), i.e. a
zero eigenvalue. Thus a necessary condition is that K be singular, that is
det K(λcr ) = 0, (24.3)
or, equivalently,
K(ucr , λcr )z = 0, (24.4)
where z = 0 is the buckling mode introduced in Chapters 4–5, where it was called a null eigenvector.
Equation (24.3) or (24.4) is the expression of the static test for finding a stability boundary.
24–5
Chapter 24: BIFURCATION: LINEARIZED PREBUCKLING I 24–6
Remark 24.1. Equation (24.3) is a nonlinear eigenvalue problem because: (a) K has to be evaluated at an
equilibrium position, and (b) K is a nonlinear function of u, which in turn is a nonlinear function of λ as defined
by the equilibrium path. It follows that in general a complete response analysis has to be conducted to solve
(24.3). Such techniques were called “indirect methods” in the context of critical point location methods in
Chapter 23. This involves evaluating K at each computed equilibrium position, and then finding the spectrum
of K. An analysis of this nature is obviously computationally expensive. One way of reducing part of the cost
is noted in the following remark.
Remark 24.2. If K is known at a given λ, an explicit solution of the eigenproblem (24.1) is not necessary for
assessing stability. It is sufficient to factor K as
K = LDLT (24.5)
where L is unit lower triangular and D is diagonal. The number of negative eigenvalues of K is equal to
the number of negative diagonal elements (“pivots”) of D. Matrix factorization is considerably cheaper than
carrying out a complete eigenanalysis because sparseness can be exploited more effectively.
Remark 24.3. The condition (24.3) is not sufficient for concluding that a system that is stable for λ < λcr will go
unstable as λ exceeds λcr . A counterexample is provided by the stable-symmetric bifurcation point discussed
in later Chapters. The Euler column furnishes a classical example. At such points (24.3) holds implying
neutral stability but the system does not lose stability as the bifurcation state is traversed. Nonetheless the
displacements may become so large that the structure is practically rendered useless.
We investigate now the first critical state of an elastic system if the change in geometry prior to it
can be neglected. We shall see that in this case the nonlinear equilibrium equations can be partly
linearized, a process that leads to the classical stability eigenproblem or buckling eigenproblem.
The eigenstability analysis procedure that neglects prebuckling displacements is known as lin-
earized prebuckling (LPB). The modeling assumptions that are tacitly or explicitly made in LPB
are discussed in some detail in the next Chapter, as well as the practical limitations that emanate
from these assumptions. In the present Chapter we discuss the formulation of the LPB eigenproblem
and illustrate these techniques on a simple problem using the bar elements developed in previous
Chapters.
24–6
24–7 §24.6 SOLVING THE STABILITY EIGENPROBLEM
The two key results from the LPB assumptions (which are studied in the next Chapter) can be sum-
marized as follows. Recall from Chapters 8–10 that the tangent stiffness matrix can be decomposed
as the sum of material and geometric stiffness matrices:
K = K M + KG . (24.6)
(1) The material stiffness is the stiffness evaluated at the reference configuration:
K M = K0 . (24.7)
KG = λK1 . (24.8)
Now the stability test (24.3) requires that K be singular, which leads to the stability eigenproblem
In the following Chapter we shall prove that under certain restrictions the critical states deter-
mined from this eigenproblem are bifurcation points and not limit points. That is, they satisfy the
orthogonality test
zT q = 0. (24.10)
Ax = λBx, (24.11)
where both matrices A ≡ K0 and B ≡ −K1 are real symmetric, and x ≡ z are the buckling mode
eigenvectors. If (as usual) the material stiffness K0 is positive definite, eigensystem theory says
that all eigenvalues of (24.11) are real. We cannot in general make statements, however, about the
sign of these eigenvalues. That will depend on the physics of the problem as well as on the sign
conventions chosen for the control parameter(s).
24–7
Chapter 24: BIFURCATION: LINEARIZED PREBUCKLING I 24–8
p = λq
Y, y k (1)
2 (2)
L (2)
0
;; k (2) 3
L (1
0
)
(1)
;;
X, x k (2) << k (1)
;;
1
Figure 24.1. LPB example involving two bar elements displacing on the X ≡ x, Y ≡ y plane.
Azi = µi zi (24.12)
This is a standard algebraic eiegnproblem, which can be solved by standard library routines for the
eigenvalues µi and eigenvectors zi . For example, EigenSystem in Mathematica or Eig in Matlab.
The µi farthest away from zero gives the λi = −1/µi closest to zero.
One disadvantage of this reduction is that A is unsymmetric even if K0 and K1 are. There are
more complicated reduction methods that preserve symmetry. These may be studied in standard
numerical analysis textbooks covering linear algebra; for example Golub and Van Loan.
§24.7. LPB Analysis Example
To illustrate the application of LPB to a very simple example, the 2-bar assembly shown in Figure 24.1 is chosen.
The bars can only displace on the x, y plane, thus the problem is two dimensional. The equivalent-spring
stiffness of the bars is denoted by
24–8
24–9 §24.7 LPB ANALYSIS EXAMPLE
The figure shows the reference configuration C0 for the two bars. That configuration is taken when the applied
load is zero, that is, λ = 0. We shall assume that the stiffness of bar 1 is much greater than that of bar 2, i.e.,
k (1) >> k (2) and is such that the vertical displacement u Y 2 of node 2 under the load is very small compared to
the dimensions of the structure.
Now let the load p = λq be gradually applied by increasing λ. The structure assumes a deformed current
configuration in equilibrium, that is, a target configuration C. According to the LPB basic assumption, the
displacements prior to the buckling load level characterized by λcr are negligible. Therefore C ≡ C0 as long
as |λ| < |λcr |.
The linear finite element equations for the example problem are as follows. For element (1):
0 0 0 0 u X1 0
0 1 0 −1 u Y 1 0
k (1) = . (24.14)
0 0 0 0 u X2 0
0 −1 0 1 uY 2 −λq
The axial linear strain and Cauchy (true) stress developed in element (1) are
uY 2 q q
(1) = =λ , σ (1) = E
(1) = −λ . (24.18)
L (1)
0 E A(1)
0 A(1)
0
According to the assumptions stated above the change in geometry prior to buckling is neglected. Consequently
24–9
Chapter 24: BIFURCATION: LINEARIZED PREBUCKLING I 24–10
For element (2) we have the same linear matrix equations as before because its geometric stiffness vanishes.
Assembling and applying displacement boundary conditions we get the equations
λq
k (2) − 0
L (2) u X2 = 0 .
0
λq (24.20)
0 k (1) − (1) uY 2 −λq
L0
One now regards K in (24.20) as unaffected by the displacements u X 2 and u Y 2 , which is consistent with the
assumption that the change of geometry prior to buckling is neglected. This having being done, setting the
determinant of K to zero yields the buckling eigenproblem:
λq
k (2) − 0
L (1)
det 0
λq
= 0. (24.21)
0 k (1) −
L (1)
0
This matrix is singular if either diagonal element vanishes, which yields the two eigenvalues
as critical values of the load parameter. Since k (2) << k (1) the lowest critical load will be
24–10
24–11 §24.8 SUMMARY OF LPB STEPS
The foregoing example illustrates the key steps of LPB analysis. These are summarized below for
completeness.
1. Assemble the linear stiffness K0 and solve the linear static problem
K0 u = q0 λ, (24.24)
for λ = 1 and obtain the internal force (stress) distribution. Note: In statically
determinate structures, such as Exercise 24.4, the internal forces and stresses may
be obtained directly from equilibrium. However K0 is still necessary for step 3.
2. Form the reference geometric stiffness K1 for that internal force distribution. The
geometric stiffness is KG = λK1 .
3. Solve the stability eigenproblem
The eigenvalue λi closest to zero is the critical load multiplier, and the associated
eigenvector zi gives the corresponding buckling mode.
24–11
Chapter 24: BIFURCATION: LINEARIZED PREBUCKLING I 24–12
EXERCISE 24.1 [A:15] Find the buckling mode (null eigenvector) z (normalized to unit length) for the
example problem of §24.6, and verify the orthogonality condition zT q = 0.
EXERCISE 24.2 [A:15] Find the buckling load for the two-bar problem if bar (2) forms an angle 0 ≤ ϕ < 90◦
with the x axis. Assume still that the system displaces only on the x − y plane and that k (2) << k (1) so that
the stress in bar (2) can be neglected in forming the geometric stiffness. Determine z and verify orthogonality.
EXERCISE 24.3 [A:20] Suppose the load of the two-bar example problem of §24.6 depends on the vertical
displacement of point 2 as p = −λcu 2Y 2 , where c is a constant with dimensions of stress. Show that even if
prebuckling deformations are neglected, the singular stiffness test leads to a nonlinear eigenvalue problem.
EXERCISE 24.4 [A:25] The “Euler column” shown in Figure E24.1 is modelled by one 2-node Euler-Bernoulli
beam column element along its length:
Y, y
;
E, I constant
; ;;
1 2
z X, x λP
The state parameters are the nodal displacements degrees of freedom arranged as
u
X1
uY 1
θz1
u=
u X2 (E24.1)
uY 2
θz2
where θz1 and θz2 are (to first order) the end rotations, positive counterclockwise about z.
The linear material matrix in the reference (undeformed) configuration is
24–12
24–13 Exercises
EA
L 0 0 − ELA 0 0
12E I 6E I − 12E3 I 6E I
0
L3 L2 L L2
4E I − 6E2I 2E I
L 0 L
K0 = L (E24.2)
EA 0 0
L
12E I
− 2I
6E
L3 L
symm 4E I
L
in which E is the elastic modulus, L the element length, A the cross section area, and I the moment of inertia
of the cross section about the z neutral axis.
The exactly-integrated geometric stiffness at the reference configuration is1
0 0 0 0 0 0
36 3L 0 −36 3L
P 4L 2 0 −3L −L 2
KG = λK1 , K1 = (E24.3)
30L 0 0 0
36 −3L
symm 4L 2
where N is the axial force in the element (here obviously equal to the applied force λP because the structure
is statically determinate.)
For this problem:
(a) Check that K0 and K1 satisfy translational infinitesimal rigid body motion conditions u X ≡ 1 and u y ≡ 1
if the six degrees of freedom are left unconstrained. (Convert those modes to node displacements, then
premultiply by the stiffness matrices.)
buckling mode
θz1
1 2
θz2 = −θz1
24–13
Chapter 24: BIFURCATION: LINEARIZED PREBUCKLING I 24–14
(c) Justify that freedom u X 2 can be isolated from the eigenproblem, and proceed to drop it to reduce the
eigenproblem to 2 × 2.
(d) Reduce the 2 × 2 eigenproblem to a scalar one for the symmetric buckling mode sketched in Figure
E24.2. by setting θz1 = −θz2 (see Figure), and get the first critical load parameter λ1 .
(e) Repeat (e) for the antisymmetric buckling mode sketched in Figure E24.3 by setting θz1 = θz2 (see
Figure) and obtain the second critical load parameter λ2 .
θz1
θz2 = θz1
1 2
buckling mode
(f) Compare the results of (d)–(e) to the exact critical load values
EI EI
P1E = −π 2 , P2E = −4π 2 , (E24.5)
L2 L2
The first one was determined by Euler in 1744 and therefore is called the Euler critical load. For P1E the
FEM result should be within 25%, which is good for one element.
(g) Repeat the calculations of λ1 and λ2 with the following reduced-integration geometric stiffness matrices6
0 0 0 0 0 0
9 3L 0 −9 3L
P L2 0 −3L −L 2
K∗1 = (E24.6)
8L 0 0 0
9 −3L
symm L2
0 0 0 0 0 0
24 0 0 −24 0
P 2L 2 −2L 2
K∗∗ = 0 0 (E24.7)
1
24L 0 0 0
24 0
symm 2L 2
and comment on the relative accuracy obtained against the exact values.
(h) Repeat the calculations of steps (b) through (e) for a two equal-element discretization. Verify that
the symmetric-mode buckling load is now −10E I /L 2 , which is (surprisingly) close to Euler’s value
P1E = −π 2 E I /L 2 .
6 These matrices are obtained by one-point and two-point Gauss integration, respectively, whereas the K1 of (E24.3) is
obtained by either 3-point Gauss or analytical integration.
24–14
24–15 Exercises
EXERCISE 24.5 [A:25] This is identical in all respects to Exercise 24.4, except that the 2-node Timoshenko
beam model is used, with Mac Neals’s RBF correction for the material stiffness matrix. The net result is
that K0 is identical to (E24.2) but K1 is different. In fact λK1 is given by Equation (9.45) where V = 0
and N = λP. Note: dont be surprised if the one-element results are poor when compared to the analytical
buckling load.
EXERCISE 24.6 [A/C: 20] The column shown in Figure E24.4 consists of 3 rigid bars of equal length L
connected by hinges and stabilized by two lateral springs of linear stiffness k. The applied axial load is
λP, where λ > 0 means compression. Compute the two buckling loads λ1 P and λ2 P in terms of k and L,
assuming the LPB model of infinitesimal displacements from the initial state. Show that one load corresponds
;;;;
to a symmetric buckling mode and the other to an antisymmetric buckling mode, and find which one is critical.
λP
;; ;;
A
;; k
2
L
;;
B
Hint. The two degrees of freedom are the small lateral displacements u 1 and u 2 of hinges 1 and 2, where
u 1 << L, u 2 << L. Write the total potential energy as
λP 2
= U − W, U = 12 ku 21 + 12 ku 22 , W = u 1 + (u 1 − u 2 )2 + u 22 . (E24.8)
2L
Explain where the expression of W comes from. Once is in hand, it is smooth sailing.
24–15
23
.
Detecting
and Traversing
Critical Points
23–1
Chapter 23: DETECTING AND TRAVERSING CRITICAL POINTS 23–2
TABLE OF CONTENTS
Page
§23.1. Introduction 23–3
§23.2. Direct Methods and Test Functions 23–3
§23.2.1. Determinant as CPTF . . . . . . . . . . . . . . . 23–4
§23.2.2. Lowest Eigenvalue as CPTF . . . . . . . . . . . . 23–4
§23.2.3. Solving the Minimum-Eigenvalue CPTF System . . . . . . 23–5
§23.3. Indirect Methods 23–5
§23.3.1. Refined Determination of a Critical Point . . . . . . . . 23–6
§23.4. Traversing Critical Points: Difficulties 23–7
§23.5. Newton-Raphson near Critical Points 23–9
§23.5.1. Injecting Perturbations . . . . . . . . . . . . . . 23–9
§23.5.2. Penalty Spring Stabilization . . . . . . . . . . . . 23–9
§23.5.3. Thurston’s Equivalence Transformation . . . . . . . . . 23–10
§23.6. Crivelli’s Procedure 23–11
§23.7. Predictor Stabilization 23–13
§23.8. Critical-Point Test Functions 23–13
§23.9. Branch switching 23–15
§23.9.1. Tangent Predictors . . . . . . . . . . . . . . . 23–15
§23.9.2. Buckling Mode Injection . . . . . . . . . . . . . 23–17
§23.10.Correctors 23–18
§23.11.Treating Bifurcation by Perturbation 23–20
§23. Exercises . . . . . . . . . . . . . . . . . . . . . . 23–21
23–2
23–3 §23.2 DIRECT METHODS AND TEST FUNCTIONS
§23.1. Introduction
Critical points were defined and mathematically characterized in Chapter 5. Chapter 18 explains
how solution procedures of purely incremental type can be adjusted to traverse limit points, and
difficulties in traversing bifurcation points are noted. Chapters 24 and follwing discuss critical
points in the context of stability analysis.
The present Chapter deals with the determination or detection of critical points in statically nonlinear
analysis and with procedures for traversal of such points with incremental/corrective techniques.
Emphasis is given to handling isolated bifurcation points, which offer moderate degree of difficulty:
tougher than limit points, easier than multiple bifurcation. Much of the material that follows is
taken from Chapter 5 of Crivelli’s thesis.1
There are two basic approaches to the problem of detecting critical points. We may attempt to find
critical points by solving an algebraic equation which has those point as a root, or we can try to
detect critical points as we are marching along an equilibrium path.
The first approach embodies what are collectively called direct methods. As the name implies,
these methods look for critical points without being concerned, at least theoretically, with tracing
equilibrium paths up to those points.
The second approach embodies the so-called indirect methods. In these methods the detection of a
critical point is intimately related to the continuation procedure.
Remark 23.1. The simplest example of a direct method is the Linearized PreBuckling (LPB) analysis discussed
in Chapters 24 and 25. This procedure, when applicable, sets up directly an eigenproblem in the reference
configuration and thus avoids tracing the nonlinear response.
A direct method for calculating bifurcation points consists of formulating a suitable set of equations
which has the critical points as solutions. In the context of general nonlinear analysis (that is, when
no a priori simplifying assumptions are made as in the LPB analysis) this set of equations should
include the equilibrium equation because only critical points on the equilibrium path are of interest.
It follows that direct formulations are achieved by augmenting the residual force equilibrium equa-
tion with a set of constraints that characterize the critical point. The main characteristic of this
procedure is that it does not require, at least theoretically, equilibrium path tracing. Thus, ideally it
should be possible to obtain all the critical points first and then join them with equilibrium paths.
Except for certain specialized situations this is not possible, however, because the solution of the
nonlinear equations resulting from this approach requires a good initial guess. Since these equations
are nonlinear, they cannot be solved directly, and linearization coupled with an iterative procedure
is required. This linearization accounts for most of the cost of direct methods.
As mentioned above, the procedure consists of augmenting the original residual force equation by
a set of constraints that define a critical point. This characterization of critical points is intimately
1 L. A. Crivelli, A Total-Lagrangian Beam Element for the Analysis of Nonlinear Space Structures, Ph. D. Dissertation,
Department of Aerospace Engineering Sciences, University of Colorado, Boulder, CO, April 1991.
23–3
Chapter 23: DETECTING AND TRAVERSING CRITICAL POINTS 23–4
related to the idea of constructing a critical point test function or CPTF. By definition, a CPTF is a
function whose zeros are the critical points of the response.2
2 The name “branching test function” is sometimes restricted to test functions that catch only bifurcation points. Here we
shall use the generic term “critical test function” to embody all critical points.
3 That of the tangent stiffness K, plus N tangent stiffnesses corresponding to the perturbation of each state vector component.
4 R. Seydel, From Equilibrium to Chaos, Elsevier, New York, 1988.
23–4
23–5 §23.3 INDIRECT METHODS
where r is the residual from the equilibrium equations, rλ is the residual from the eigenvalue problem
and φ measures the violation of the normalization condition.
This approach has been explored by Wriggers and Simo5 in the context of nonlinear structural
analysis. As can be seen from the previous equation the Jacobian of this augmented system is
neither sparse nor symmetric, and has large storage requirements. In addition, the coefficient
matrix becomes singular at bifurcation points, and hence ill-conditioned in their neighborhood.
It is desirable to solve (21.3) while taking advantage of the symmetry of K. This can be done by
using the concept of auxiliary systems. Assuming K is nonsingular, consider the five auxiliary
systems:
φ − aT dhr
η=− , d = dr + ηdq , h = dh − dhr + ηdhq . (23.6)
aT dhq
The solution of these five systems at each corrective iteration is expensive. However, the main
disadvantage of this approach is having to compute the u-derivative of the stiffness matrix. Since an
exact derivation of Kv is only practically feasible for simple nonlinear problems, some investigators
(e.g. Seydel and Wriggers-Simo, loc. cit.) propose using numerical differentiation to compute Kv .
In this context it should be observed that only the product Kv z (a square matrix) is required.
Numerical experiments reported in Seydel’s book show that the convergence rate of this approach
depends on whether the critical point is a limit or a bifurcation point. At limit points, the Jacobian of
the augmented equation is nonsingular, which guarantees fast local convergence. At a bifurcation
point, the vector [ zT 0T 0 ]T is a left null eigenvector, which shows that the Jacobian becomes
singular. Thus, slower convergence rates may be expected at bifurcation points. However, experi-
ence shows that for some problems the convergence rate is good for the λ component when a not
very stringent accuracy is required, as can be seen from the examples presented in Wriggers and
Simo.
Remark 23.2. If the critical point is known to be a limit point, the penalty spring augmentation method
described below is much simpler and inexpensive. And even the simplistic random perturbation scheme
described there often works. Thus it appears that the direct method is advantageous in special instances, for
example when direct determination of the CP locations as stability envelope, is desirable, and a good estimation
of that location is available. This occurs in some applications such as optimization with stability constraints.
5 P. Wriggers and J.C. Simo, A General Procedure for the Direct Computation of Turning and Bifurcation Points, Int. J. Nu-
mer. Meth. Engrg., 30, 155–176, 1990.
23–5
Chapter 23: DETECTING AND TRAVERSING CRITICAL POINTS 23–6
The underlying idea behind indirect methods is to recognize the occurrence of a critical point while
following an equilibrium path by continuation. The key to success is having a good CPTF τ . We
defer the discussion on how to construct such a CPTF to §21.4, and for the moment assume that
such a function is available.
These methods benefit from data collected from the continuation procedure. A critical point is
characterized by being a zero of the CPTF. Because hitting an exact zero is highly unlikely, we
check for a critical point by invoking the condition
This is called a bracketing or straddling condition. It means that a critical point occurs in the
interval or bracket [λk ≤ λ ≤ λk+1 ]. The next task is to “zero in” this point by reducing the bracket
size.
Assume a critical point has been located by verifying bracketing condition (21.7) on the CPTF. Further, assume
that the test function is a continuous function of λ and has only one isolated zero in the interval. This condition
requires using a sensible incremental step — which is not known a priori. However, the problem of determining
an appropriate step may be solved by a combination of intuition and an adaptive incremental control strategy,
which monitors the convergence behavior of the solution.
Once the algorithm finds that a critical point is bracketed or straddled by two equilibrium solutions, we may
obtain a first approximation to such point by using linear interpolation:
λk+1 − λk k
λc λk + τ (23.8)
τ k − τ k+1
We present below an adaptive method developed in Crivelli’s thesis (loc. cit. to improve the initial approxi-
mation (21.8). It makes use of iterative inverse interpolation and divided differences.6 We have computed a
set of m values of the CPTF at m unequally spaced values of the control parameter λ. We seek a polynomial
interpolation for τ of the form
τ (λ) − τ (λ0 )
τ [λ0 ; λ] =
λ − λ0
.. ..
. . (23.10)
τ [λ0 , . . . , λk−2 ; λ] − τ [λ0 , . . . , λk−2 ; λk−1 ]
τ [λ0 , . . . , λk−1 ; λ] =
λ − λk−1
6 For background in such techniques, see for example G. Dahlquist and A. Björck, Numerical Methods, Prentice-Hall,
Englewood Cliffs, 1974.
23–6
23–7 §23.4 TRAVERSING CRITICAL POINTS: DIFFICULTIES
c0 = τ (λ0 )
c1 = τ [λ0 ; λ1 ]
.. .. (23.11)
. .
ck = τ [λ0 , . . . , λk−1 ; λk ]
Observe that this formula is recursive and a new point can be added without great difficulty. Now the problem
of locating a critical point is reduced to finding a zero of (21.9).
Remark 23.3. Observe that we may as well use equation to interpolate λ as a function of τ and obtain λc just
by setting τ = 0. However, both procedures are fundamentally different and may give different results. For
instance, for very sharp limit points and for bifurcation points the test function could stay almost constant for
most of the continuation and change to a steep slope while very close to the limit point. This type of behavior
can be regarded as an exponential relation between τ and λ. Such an exponential can be well approximated
by a polynomial, whereas its inverse —a logarithmic function— may pose severe difficulties. Thus we may
expect faster convergence from the former procedure.
Based on the foregoing remark we choose inverse interpolation for finding a zero of (21.9). For
this we recast that equation into the form
λc = ψ(λc ) (23.12)
1
λc = λ0 − [c0 + · · · + cm (λc − λ0 ) . . . (λc − λm−1 )] (23.13)
c1
23–7
Chapter 23: DETECTING AND TRAVERSING CRITICAL POINTS 23–8
In the remainder of this section we consider the problem of traversing critical points using incremen-
tal/corrective methods. Corrective solution methods, of which Newton-Raphson is the prototype,
may run into difficulties at or near critical points. Essentially those are caused by the fact that the
tangent stiffness matrix becomes singular at such points and consequently ill-conditioned in their
neighborhood. This ill-conditioning can introduce noise which may render the solution procedure
unstable.
Sometimes the problems associated with noise are easily circumvented, whereas in some cases
the problems may become computationally overwhelming and may require either specialized tech-
niques or intensive human intervention to proceed. Generally speaking, the degree of difficulty is
influenced by the following factors:
1. Bifurcation points are more difficult to handle than limit points.
2. Isolated critical points (at which the rank deficiency is one) are easier to handle than multiple
or compound critical points.
3. Dynamic critical points of flutter type are more difficult to handle than static ones. Ultimate
in nastiness (sort of like the triple jump in ice skating) are problems of fracture or localiza-
tion where an equililibrium path suddenly terminates, and dynamic methods are required for
proceeding further.
In what follows we focus on the characteristics of Newton’s method near an isolated critical point of
static type. Even under those relatively benign conditions the topic is still an area of active research.
23–8
23–9 §23.5 NEWTON-RAPHSON NEAR CRITICAL POINTS
23–9
Chapter 23: DETECTING AND TRAVERSING CRITICAL POINTS 23–10
Here ei is the elementary vector of order N all of whose entries are zero except the i th entry which
is one, and Ei = ei eiT . Some possible choices for the index i are discussed in that article (handed
out in class). One advantage to this approach is that the symmetry and sparseness of K is not
affected by this diagonal correction. The solution of the original Newton-Raphson system (17.13),
reproduced here for convenience:
K −q d r
=− , (23.15)
a T
g η c
can be expressed as a linear combination of the solutions of the three auxiliary problems defined in
(21.14):
d = drs + σq dqs + σe dse ,
. (23.16)
η = − c + aT d /g,
Here the coefficients σq and σe are obtained by requiring that d and η solve the original problem.
The following 2 × 2 unsymmetric system results:
g + aT dqs aT dse σq −c − aT drs
= (23.17)
−eiT dqs 1 − eiT dse σe eiT drs
The computation of η from this equation breaks down if g = 0. In such a case, it is recommended
that η be recovered from the equivalent equation:
σe di − di2
η = σq + s (23.18)
dT q
The value of s in equation is irrelevant as long as it is sufficiently large to stabilize K.
An advantage of this procedure is that it avoids the need for partitioning the stiffness matrix proposed
by Rheinboldt (loc. cit.), which requires special treatment of the elements in the i th row and column
of K. Furthermore, none of the right-hand sides of the auxiliary systems in requires access to
elements of K.
Disadvantages of this procedure are as follows:
1. It breaks down at bifurcation points, where the coefficient matrix of the 2 × 2 system becomes
singular.
2. Some overhead is necessary to trace the rate of change of the components of the state vectors,
which is used as a criterion in the selection of the index i.
3. The equation solver should be able to undo part of the reduction, from the current value of the
row index to the selected value of i.
9 G. A. Thurston, F. A. Brogan and P. Stehlin, Postbuckling Analysis Using a General Purpose Code, AIAA Journal, 24,
1013–1020, 1986.
23–10
23–11 §23.6 CRIVELLI’S PROCEDURE
23–11
Chapter 23: DETECTING AND TRAVERSING CRITICAL POINTS 23–12
linearly dependent. Denote by S = {i 1 , . . . , i k } a set of indices such that S ⊂ {1, . . . , n}; this set S
is a subset of the column indices of K. This set is chosen so that there is a linear combination of
columns of K satisfying
βik kik = 0 (23.24)
i k ⊂S
where ki is the i th column of K. Such a set exists because K is singular. Choose an index i from
such that βi = 0. Column ki can be then expressed as a linear combination of the other columns
of K. Apply now the procedure with the index i chosen as described. It is immediately seen that
kiT z = 0; therefore z is the null eigenvector of K.
If K is almost singular, then z is an approximation to the null eigenvector. It can be shown that if K
is continuously differentiable and K̂ is nonsingular, then z approaches z as (u, λ) approach (uc , λc ).
To solve the original problem, define the following auxiliary systems:
d = dr + σq dq + σz z (23.26)
and
Kdq = q + (kiT dq − qi )ei (23.28)
Also observe that
T
kiT dr = drT (Pi K̂ẑ + K ii ei ) = d̂r K̂ẑ
(23.29)
= −rT Pi ẑ = −rT (z + ei ) = −rT z − ri
Thus
kiT dr + ri = −rT z (23.30)
Similarly
kiT dq − qi = −qT z (23.31)
From previous equations it is immediately seen that
c + aT d
Kd − qη = −r + σq + q + kiT zσz + qT zσq − rT z ei (23.32)
g
To recover the original system, we require the terms inside squared brackets to vanish, which leads
to the auxiliary system
g + aT dq aT z σq −c − aT dr
= (23.33)
qT z kiT z σz rT z
23–12
23–13 §23.8 CRITICAL-POINT TEST FUNCTIONS
η = σq (23.34)
This approach has some advantages over the one pertaining to the penalty spring method. First, if
σz = 0 it reduces to the one used by classical continuation. Second, it enforces prior equations.
Closer inspection of tells us that σe = di if exact arithmetic is used. Thus, the second term vanishes.
When finite precision arithmetic is used, rounding errors may prevent this term from vanishing.
If in addition, a sufficiently large value for the penalty spring s is chosen, could produce some
artificially large value of η with the consequence that the corrector step may move the solution
away from convergence.
Several remarks may be made in regard to the preceding derivation. At a limit point, neither qT z
nor aT z vanish and the equations are well behaved even though kiT z is zero. Thus this procedure can
satisfactory traverse limit points. At a bifurcation point the last row of equation is identically null,
thus the value of σz is not defined. This is consistent with the definition of bifurcation point, since at
these points more than one solution is possible and we cannot continue tracing any branch without
introducing additional information. Thus when a bifurcation point is detected, this procedure breaks
down and we have to resort to a branch switching algorithm as described in the following sections.
Thus
Ku(0) − qλ(0) = λ(0) (kiT q − qi ) + σz kiT z ei (23.36)
Requiring the term in brackets to vanish and using equations (17) the load increment is obtained as:
(0) kiT z
λ = T σz (23.37)
q z
Assuming we use the global hyperelliptic constraint we can obtain σq as
σz = (23.38)
2 T 2 2
a ki z k T
z
± 2 dqT Sdq + zT Sz + i
b2
v qT z qT z
23–13
Chapter 23: DETECTING AND TRAVERSING CRITICAL POINTS 23–14
In this section we investigate several candidates for CPTF. A CPTF is a function that allows
monitoring the continuation procedure for occurrence of critical points. By definition, critical
points should be the only zeros of a CPTF. As mentioned before, critical points are characterized
by the tangent stiffness K being singular. Thus, test functions essentially monitor K to obtain an
estimate of its proximity to a critical point.
A highly accurate estimator of how singular is K is its smallest eigenvalue. Since K is real and
symmetric, its eigenvalues are all real; in addition, if K varies continuously, its eigenvalues also
vary continuously and singularity is simply detected by checking for a null eigenvalue. Thus, the
smallest eigenvalue of K makes a very attractive choice for a CPTF. However, an eigenvalue analysis
at each increment is prohibitively expensive.
Another possible test function is the determinant of K. As noticed by Abbott10 the order of mag-
nitude of the determinant may be computationally inconvenient, especially for large stiff systems,
when the solution is computed far away from bifurcation. Scaling K by 10 changes the determinant
by 10 N ! This difficulty can be overcome by computing the determinant as a pair of numbers, a
characteristic normalized between −1 and 1 and a mantissa. This decomposition is available in
several direct solvers. A drawback of this approach occurs, however, if the null eigenvalue has an
even multiplicity, since in this case the determinant will not change sign and the tester will fail to
detect a critical point.
Two additional choices are considered in the following. Using previous equations the following
relation results:
zT Kz kiT z
= (23.39)
zT z zT z
If we regard the left-hand side of this equation as a Rayleigh quotient, and taking into account that
z is a close approximation to the smallest eigenvector of K when the solution is sufficiently close to
the limit point, we may use the right-hand side of equation to monitor the smallest eigenvalue of K.
Thus, ki z is a candidate for a CPTF. A similar test function is proposed by Seydel (loc.cit.). This
approach is expensive if it is to be carried out at every increment, because the decomposition is only
computed near critical points. However, if by some other means we have detected the proximity
of a critical point, this computation may be obtained as a by-product of the stabilization procedure
described in the previous Chapter. In such case, it becomes an attractive CPTF. We therefore look
for a cheaper alternative that can provide a good estimate of the singularity of K.
Since we use full Newton we generally have a decomposition of K of the form:
K = LT DL (23.40)
10 J. P. Abbott, An Efficient Algorithm for the Determination of Certain Bifurcation Points, J. Comput. Appl. Math., 4,
19–27, 1978.
23–14
23–15 §23.9 BRANCH SWITCHING
also monitors the smallest eigenvalue of K. Moreover, τ = 0 when K is singular and τ < 0 if
K is indefinite. Thus, τ defined is readily available without any overhead and is the test function
adopted here. Once it has been determined that there is a critical point lying between two solutions
we may switch to the test function.
23–15
Chapter 23: DETECTING AND TRAVERSING CRITICAL POINTS 23–16
Unacceptable
iteration
Predicted P
λ solution
Acceptable
Unknown branch
C B iteration
known solution
Known branch
have assumed that the bifurcation point is simple, there can be at most two intersecting branches.
Since q is in the range of Kc , the system Kw = q is consistent at a bifurcation point. Hence, there
is a unique solution w lying in the range of K(uc , λc ) that is, satisfying the relation wT z = 0. Now
the state variation rate u̇ at a bifurcation point can be decomposed into a homogeneous solution
component σ z in the buckling mode direction, and a particular solution component, orthogonal to
z, which is furnished by w:
u̇ = (w + σ z) λ̇ (23.42)
The next step is to find σ . It was pointed out that the first-order rate equations do not provided enough
information for computing σ . To obtain the required information we must resort to the second-
order system. Premultiplying equation by zT and taking into account the bifurcation condition 3 in
Definition 2 we get the scalar equation
The rate matrices K̇ and q̇ may be written as linear combinations of u̇ and λ̇ as:
K̇ = Lu̇ + Nλ̇, q̇ = − Nu̇ + jλ̇ (23.44)
Then
zT Lu̇ + Nλ̇ u̇ + zT Nu̇ + jλ̇ λ̇ = 0 (23.45)
The replacement of u̇ by its decomposition gives
zT L (w + σ z) λ̇ + Nλ̇ (w + σ z) λ̇ + zT N (w + σ z) λ̇ + jλ̇ λ̇ = 0 (23.46)
23–16
23–17 §23.9 BRANCH SWITCHING
d = w − u, λ=λ (23.49)
Observe that the two tangents at the bifurcation point obtained by replacing the two roots span a
plane. In a sufficiently small neighborhood of the bifurcation point (uc , λc ) both branches lie on
that plane. If λ is sufficiently close to λc it would be possible to approximate the vector d by a vector
almost parallel to this tangent plane. If d could be obtained exactly it would be possible to switch
branches in a straightforward manner. However, in most practical problems only an approximation
to d is possible. Such approximation is generally split into two steps
1. Find a direction pointing to the other branch
2. Find a step along this direction so that the distance between the two solutions is minimized in a
given norm.
The problem is then reduced to finding d and δ such that
The first condition can be viewed in the more general context of a constraint relation between the
solutions on both branches, thus generalizing constraints.
The first subproblem can be solved by taking d proportional to z since z lies on the plane defined
by the two tangents. This can easily be seen just by taking the difference between the two tangents
and observing that this difference is a multiple of z. This procedure can be viewed as an injection
of the buckling mode into the known solution.
23–17
Chapter 23: DETECTING AND TRAVERSING CRITICAL POINTS 23–18
The solution to the second subproblem is more difficult. It does not seem to be possible to find a δ
that satisfies the previous equation for a given value of λ without resorting to information provided
by second or higher-order rate equations. We cope with this problem by shifting it to the corrector,
which will adjust λ instead of δ in such a way that a constraint condition between the solutions on
both branches is satisfied. The problem is so restated because the corrector is better equipped to
deal with a problem posed in terms of changes in the state vector and control parameters, because
an incremental equation for δ is not available. Thus, at the predictor level, we will be content with
an estimation of δ. Some values dictated by experience are given in Seydel (loc.cit.).
Note that two predictors are in fact available:
w1,2 = u ± δz (23.51)
For perfect structures undergoing symmetric —pitchfork— bifurcation, the sign choice is irrelevant,
but for asymmetric —transcritical— bifurcation, one branch will be energy preferred. This difficulty
can be overcome by calculating solutions on both half-branchs emanating from the bifurcation point
and choosing the one with least energy.
The success of the whole procedure strongly depends on an appropriate corrector. Correctors should
be carefully designed in order to display selective properties. This is meant as the capability of the
corrector to avoid converging towards the known branch. The design of such correctors is dealt
with in the next section.
§23.10. Correctors
One alternative is to enrich the displacement increment with the buckling mode. Felippa (loc.cit.)
modifies equations so that
d = drs + σq dqs + σe dse ± σz z (23.52)
To provide sufficient equations a length constraint in injected, zT d = δ. Again, the unknown
coefficients are obtained by requiring that d verifies the incremental equation which leads to the
following 3 × 3 unsymmetric auxiliary system:
g + aq ae az σq −c − ar
−dqi 1 − dei l σe = dri (23.53)
sq se 1 ±σz δ − sr
in which
az = zT a, dz = zT d, sr = zT drs , sq = zT dqs , se = zdse (23.54)
The selective properties of this corrector are not clearly established.
To obtain a corrector with enhanced selective properties, Skeie and Felippa11 propose using a
constraint that has greater affinity for the emanating branch than for the known branch. For this
purpose, they look at the limit behavior of the local hyperelliptic constraint. They show that this
constraint reduces to a narrow cylinder around the known branch when a/ b/. To obtain a
robust corrector it is necessary to inject information from both branches into the constraint equation.
11 G. Skeie and C.A. Felippa, Detecting and Traversing Bifurcation Points in Nonlinear Structural Analysis, Report CU-
CSSC-89-23, Center for Space Structures and Controls, University of Colorado, Boulder, 1989.
23–18
23–19 §23.10 CORRECTORS
This can easily be achieved if the constraint is expressed in terms of the difference between the
solutions on two different branches, as shown below.
In the correction step, can be generalized to
c(w − u, λ − λ) = 0 (23.55)
Increments to both w and u and to λ are computed so that this constraint is enforced. What we
are trying to avoid is an iteration converging towards the known branch. Experience shows that
this frequently happens when λ = λc is kept fixed while trying to obtain a state vector u satisfying
the residual equation. For example, consider the case of symmetric —pitchfork— bifurcation. In
this case, the crossing branch exists for values of λ either greater or smaller than λc . Since we
do not know beforehand on which side of the bifurcation point the crossing branch lies, there is
a 50% chance that no emanating branch exist for the chosen value of λ. Besides, it is possible
that the emanating branch could not be parametrized by λ close to the bifurcation point and a local
parametrization may be required.
Once again we can expand the solution space. A point in this expanded space is defined by (u, w, λ)
and a parametrization for the solution in this new space is sought. In the case of prior equations the
parameter was the arc-length ; here δ plays the same role. The only difference is that we replace
u, which represents an increment of the solution on the known branch, by w − u which measures
the difference between two solutions lying on different branches. For example, displacement control
results in
wk − u k = δ (23.56)
In summary, we look for two state corrections dw and dv and a load correction η that simultaneously
satisfy
r(uk + dkv , λk + ηk ) = r(wk + dkw , λk + ηk ) = 0 (23.57)
and the constraint condition
c(uk + dkv , wk + dkw ; λk + ηk ) = 0 (23.58)
These corrections are obtained from
Kv 0 −q dv rv
0 Kw −q dw = − rw (23.59)
a T
−aT g η c
where Kv = K(uk , λk ) and Kw = K(wk , λk ).
Again it is possible to solve the previous equation while preserving the symmetry and sparseness
of K. If K is nonsingular we may write
dr u = −K−1
v rv , dr w = −K−1
w rw , dqu = K−1
v q, dqw = K−1
w q (23.60)
Thus, the two incremental displacements and the load increment may obtained as:
dv = dr u + ηdqu
dw = dr w + ηdqw
(23.61)
c + aT (dr u − dr w )
η=−
g + aT (dqu − dqw )
23–19
Chapter 23: DETECTING AND TRAVERSING CRITICAL POINTS 23–20
By using the procedure described in previous equations we require much less storage than when
using the original system, because Kv and Kw can overwrite each other. The extra storage re-
quirement is for four N -dimensional auxiliary vectors. However, it is still demanding in terms of
computer time, because two linear systems of equations with different coefficient matrices must be
solved at each iteration.
As an alternative, we may keep the solution on the known branch fixed and search for another
equilibrium solution that also verifies the constraint equation. That is, solve
r(w, λ)
=0 (23.62)
c(u; dw , λ)
This equation resembles the original continuation equation; the main difference being a slight
modification of the constraint condition. However, there is no guarantee that a solution of this
equation will lie on the emanating branch; observe that u and w correspond to two different values
of the load parameter λ and it is possible that two solutions on the same branch satisfy the constraint
equation.
In the examples presented in Crivelli’s thesis the first approach is used, since we are willing to
pay the additional price to guarantee the success of the branch switching. Furthermore this cost is
modest compared to that of the entire continuation process.
A remark is in order here. The stiffness matrices Kv and Kw may become ill-conditioned and a
procedure similar to that leading to equations may be developed. However, note that we are not
looking for a highly accurate solution on the known branch, since we are using it just to repel an
iteration from the crossing branch back to the known branch. Such a process is not required to
provide an highly accurate solution on the crossing branch either, because that solution is just used
to obtain a starting point on the new branch. We can thus apply the transformation to K to avoid
numerical instability, but regarding dr u and dr w as sufficiently good approximations to dr u and dr w ,
respectively. Once a point on the new branch is obtained, the solution can be further improved by
switching back to the continuation procedure used for regular points.
The last approach to be mentioned here is treating bifurcation by perturbation. The idea is to perturb
the residual equation r(u, λ) in such a way that the underlying regularity intrinsic to a bifurcation
point is destroyed. This is accomplished by introducing physical or numerical imperfections. In the
mathematical literature, this approach is referred to as unfolding. The perturbed system displays
limit point behavior rather than bifurcation point behavior. The key difficulty with this approach is
to find a perturbation parameter and an imperfection distribution that is critical to the underlying
regularity. For simple structures, this parameter may be guessed from physical considerations.
Once an appropriate value is found, the branch can be traced by the standard continuation procedure
and the problem is reduced to limit point traversal.
23–20
23–21 Exercises
EXERCISE 23.1
[C:25] For the steep 2-bar arch with S = 2 and arbitrary H and subject to a vertical load −λq, find the locus
of all critical points directly by setting up and solving the r = 0 system augmented by det(K) = 0 as CPTF.
Hint: you may try the effect of
ClearAll[Em,A0,S,vX,vY,H,lambda];
Em=1; A0=1; S=2;
Do [
rX = (4*A0*Em*vX*(S^2 + 2*vX^2 + 4*H*vY + 2*vY^2))/(4*H^2 + S^2)^(3/2);
rY = (8*A0*Em*(H + vY)*(vX^2 + 2*H*vY + vY^2))/(4*H^2 + S^2)^(3/2)-lambda;
K={{(4*A0*Em*(S^2 + 6*vX^2 + 4*H*vY + 2*vY^2))/(4*H^2 + S^2)^(3/2),
(16*A0*Em*vX*(H + vY))/(4*H^2 + S^2)^(3/2)},
{(16*A0*Em*vX*(H + vY))/(4*H^2 + S^2)^(3/2),
(8*A0*Em*(2*H^2 + vX^2 + 6*H*vY + 3*vY^2))/(4*H^2 + S^2)^(3/2)}};
sol=NSolve[{rX==0,rY==0,Det[K]==0},{lambda,vX,vY}];
Print["solution for H=",H," is ",({lambda,vX,vY}/.sol)//InputForm],
{H,1/2,3,1/2} ];
where v ≡ θ is the only degree of freedom, provides what is perhaps the simplest example of bifurcation.
This has two solution paths, θ = 0 (the fundamental path) and λ = θ/ sin θ (the post-buckling path), which
intersect at B (λ = 1, θ = 0.
Solve this equation using the purely incremental arclength control method programmed in Chapter 18 and
devise a procedure by which the program, starting from λ = θ = 0 and moving up the fundamental path, detects
B and (without cheating) follows one of the post-buckling paths. Probably the simplest one to implement is
perturbation.
EXERCISE 23.3 [C:25] Solve the steep arch H = 2S under a vertical load using the program of Exercise
20.3 and observe whether the program senses the bifurcation point or just marches along the fundamental path.
Experiment with perturbing the problem by inserting a small horizontal load to see whether you can coax it
into the secondary path.
23–21
22
.
Accelerators
and Line Search
22–1
Section 22: ACCELERATORS AND LINE SEARCH 22–2
TABLE OF CONTENTS
Page
§22.1. Introduction 22–3
§22.2. General Comments 22–3
§22.3. Accelerator Models 22–3
§22.3.1. Tangent Accelerators for Fixed λ . . . . . . . . . . . 22–3
§22.4. References 22–10
22–2
22–3 §22.3 ACCELERATOR MODELS
§22.1. Introduction
Newton-like methods were initially used in geometrically nonlinear analysis with a unit h. This is
optimal for conventional Newton (CNM) near a solution, but not necessarily so otherwise. Since
a key reason for letting h vary is to accelerate convergence to equilibrium, schemes that utilize
nonunitary stepsizes are called accelerators. Another application is diagnosing and overcoming
divergence in “difficult” regions of the response. The latter application is tied up with the line search
technique that comes from the field of optimization. In the present Chapter we study acceleration
formulas and line search schemes based on the relaxation equations.
22–3
Section 22: ACCELERATORS AND LINE SEARCH 22–4
22–4
22–5 §22.3 ACCELERATOR MODELS
Remark 22.2. For the quadratic scalar equation r = u 2 = 0, u (0) = 0, we have ṙ = 2u u̇, K̇ = 2u̇, ν = −0.5
and h CNM = 2, which yields the exact solution u = 0 in one step. Similarly, for the equation r = u m = 0 we
get h CNM = m and again we arrive at the solution in one step.
Remark 22.3. At noncritical points the performance of h CNM is not so dramatic as above; in fact it tends to
“overshoot” the solution by the same amount as h = 1 tends to undershoot it. This is the rationale behind the
Halley accelerator.
T
u̇(k) ṙ(k)
h MNM = T (k) T
(22.9)
u̇(k) K̇ u̇(k) + ṙ(k) F ṙ(k)
T
∗ u̇(k) ṙ(k)
h MNM = T
(22.10)
ṙ(k) F ṙ(k)
which is more easily calculable given the information available in a MNM iteration. (A similar
simplification in (80) would be too drastic, for it would reduce it to h CNM = 1; but K̇u̇ is more
easily calculable in CNM.) Accelerators for the damped Newton method (73) can be constructed
in a similar manner.
Formulas such as (79) and (82) are called tangent accelerators as they depend only on the state at
pseudotime tk and t-derivatives there. But the presence of rates such as K̇ and ṙ causes computational
difficulties because the rates can be readily calculated only in simple problems. The problem can
be overcome by using secant information as discussed later.
Remark 22.4. Stepsize Matrix. A second approach to solving (78) is to transmute h into a “stepsize matrix”
H given by
(k)
u(k+1) = u(k) − (K̇ + W(k) K(k) )−1 W(k) r(k) , (22.12)
into which we may replace K̇ by the rank-two estimate derived from Quasi-Newton formulas in Felippa [19].
This approach is presently unexplored.
22–5
Section 22: ACCELERATORS AND LINE SEARCH 22–6
Remark 22.5. Diagonalised Stepsize Matrix. Yet a third approach that merits attention is to transmute h into
a diagonal matrix, H D . This effectively assigns stepsizes component by component, which is in line with the
empirical “alpha method” of Nayak and Basu [20]. To get the stepsize h i associated with the i th component
of u, premultiply both sides by eiT , which is a row vector with unit i th component and zero otherwise. The
result is
T
ei(k) ṙ(k)
hi = T (k) T
(22.13)
ei(k) K̇ u̇(k) + ei(k) W(k) ṙ(k)
Appropiate W may now be inserted and rates replaced by secant relations.
Secant Accelerators
The presence of rates such as K̇ and ṙ in the tangent accelerators gives rise to computational
difficulties in complex nonlinear problems. More practical is to use secant-type (finite difference)
information to obtain secant accelerators. The idea is to formally replace
in the tangent accelerators. This formal procedure, however, has the disadvantage that K is not only
unwieldy to calculate in CNM but unavailable in MNM. Fortunately in the tangent accelerators K̇
appears always in the combination K̇u̇; consequently K is always postmultiplied by d in its secant
counterparts. A convenient expression for K d can be obtained from the Quasi-Newton formulas,
which are multidimensional generalizations of the scalar secant method. Pertinent expressions for
the Broyden family [21] are worked out in Fletcher [19], from where one gets
22–6
22–7 §22.3 ACCELERATOR MODELS
g(k) d(k)
T
h̄ = T
(22.19)
g(k) T W(k) d(k) + d(k) (g(k) − K(k) d(k) )
On replacing the appropriate W, secant counterparts of h CNM , h MNM and h ∗MNM are obtained as
T
1 d(k) K(k) d(k)
h̄ CNM = , ν= , (22.20)
1+ν g(k) T d(k)
g(k) d(k)
T
h̄ MNM = T
. (22.21)
g(k) T Fg(k) + d(k) (g(k) − K(k) d(k) )
g(k) d(k)
T
∗
h̄ MNM = , (22.22)
g(k) T Fg(k)
Remark 22.6. The important equation (22.18) is not sensitive to the choice of Quasi-Newton formula. In fact
one always has
K(k+1) d(k) = (K(k) + K(k) ) d(k) = g(k) , (22.23)
which is called the Quasi-Newton condition and serves as a point of departure for defining such formulas.
3. Calculate h̄. If this value differs substantially from the assumed one (say by more than ±50%)
repeat steps 2 and 3 but no more than a fixed number of times (typically 1).
4. If a h̄ has been accepted, use to calculate u(k+1) . Otherwise branch to a line search procedure.
Many stylistic variations are possible. For example, one might decide never to repeat steps 2–3.
This is the same as extrapolating h̄ from the previous-step information, which is economical but
potentially dangerous.
22–7
Section 22: ACCELERATORS AND LINE SEARCH 22–8
where s(k) is defined by (65). Note that for variable λ vector c(k+1) = v̄ λ(k+1) depends indirectly
on h. The stepsize h is to be chosen so that a norm of r is minimized. Most commonly used is the
scaled Euclidean norm:
r = 12 rT Dr, (22.25)
where D is a positive-definite scaling matrix (usually diagonal) A non-identity diagonal D is useful
when the residual components have different physical units. The choice of D ≡ F = K−1 (if
positive definite) is also of some interest because the residual norm becomes an energy norm.
The general approach to minimizing r (h) is through a line-search method as described in the next
subsection. But if the dependence of r on h is mildly nonlinear, an approximate minimizer can be
obtained by assuming a linear dependence of r on h:
where r(0) and r(1) denote the residual vectors evaluated at h = 0 and h = 1, respectively; if k > 0
then r(0) should be available from the previous iteration. The minimizing condition ∂r/∂h = 0
yields
r(0)T Dr(0) − r(1)T Dr(0)
ĥ = (22.27)
r(0)T Dr(0) − 2r(1)T Dr(0) + r(1)T Dr(1)
Inasmuch as the assumption (22.26) is only strictly valid for linear response behavior, it is recom-
mended to check whether
r (ĥ) < r (1) and r (ĥ) < r (0) (22.28)
holds before accepting ĥ. If this condition is not verified, a line search may be called for.
Remark 22.7. Unlike the accelerators based on the zero-velocity model, the formula (22.27) depends on the
corrective method only indirectly through r(1). This gives it an implementation edge over secant accelerators
in the sense that the stepsize calculations are less method dependent. On the other hand, its performance is
not usually as good as that of secant accelerators.
Remark
22.8. If D is the identity matrix, ξ is the residual magnitude ratio ξ = r(1)/r(0) =
r(1) r(1)/r(0)T r(0) and ϕ is the angle subtended by r(1) and r(0), the above formula can be presented as
T
1 − ξ cos ϕ
ĥ = (22.29)
1 − 2ξ cos ϕ + ξ 2
which admits of a geometric interpretation studied in Felippa [22].
Remark 22.9. Is it worthwhile accepting a non-unitary ĥ? If so, an extra residual evaluation at ĥ would
be required. A detailed analysis by Felippa [22] indicates that ĥ should be accepted only if it is near 0.5
(underrelaxation) or near 2.0 (overrelaxation).
22–8
22–9 §22.3 ACCELERATOR MODELS
Line Search
A line search is a systematic procedure to find a stepsize h that approximately minimizes a residual-
magnitude measure, which for definiteness shall be assumed to be (98). Vector s(k) is called the
search direction.
Line search procedures are highly developed in optimization work, where they are essential com-
ponents of general-purpose function minimization programs: a line search is carried out at each
iteration. On the other hand, their role in nonlinear equation solving is secondary. Because line
search is fairly expensive in terms of residual evaluations, it is invoked only when a solution
procedure runs into severe difficulties.
The acceptance tests stated below involve the directional derivative g = ∂r (h)/∂h along s. If D is
constant,
∂r ∂r ∂λ
g(h) = = rT D = rT DK(s(k) + v̄ ) (22.30)
∂h ∂h ∂h
The last expression is computationally cumbersome because it involves the tangent stiffness matrix.
More practical is to use finite differences to estimate the factor ∂r/∂h.
Line search algorithms “backtrack” h in a systematic fashion until an acceptance condition is
verified. A widely used acceptance condition in optimization programs of the the early 1970s is
the Goldstein-Armijo (GA) rule which may be stated as
where η1 and η2 are scalars that satisfy 0 < η1 ≤ η2 < 1; typical values being η1 ≈ 0.1 and
η2 ≈ 0.9. The upper and lower bounds in (104) insure that h is neither “too large” nor “too small”.
In more modern optimization work [23], the GA rule is replaced by slope conditions of the form
g(h) − g(h )
1
|g(h)| ≈ ≤ −ηg(0), (22.32)
h − h1
r (0) − r (h) ≥ −ξ hg(0). (22.33)
where 0 ≤ η < 1 and 0 < ξ ≤ 0.5, ξ ≤ η, and 0 ≤ h 1 < h. A value of η ≈ 0.8, which gives a not
very restrictive line search, can be recommended.
22–9
Section 22: ACCELERATORS AND LINE SEARCH 22–10
There are many variations on this theme and readers interested in additional details are referred to
the abundant literature on practical optimization methods. The following algorithm is meant only
to illustrate the basic procedural steps.
1. Estimate g(0) by finite differences, for example
(r(0.1) − r(0))
g(0) ≈ r(0)T D (22.34)
0.1
If g(0) ≥ 0 exit with h = 0, which should trigger a refactoring. Else set h 1 = 0, h 2 = 10
(say), h = 1.
2. Evaluate r (h).
3. Test for (22.32). If not satisfied, compute ĥ by restricted interpolation from data at h 1 and h;
set h 2 = h, h = ĥ, and return to 2.
4. Test for (22.31). If satisfied, exit. If not, compute ĥ by restricted interpolation from data at h 1
and h; reset h 1 = h, h = ĥ and go to 2.
If the exit h < 0.1 (say), the iteration step should be abandoned and the stiffness matrix refactored
at h = 0.
As stressed, a line search procedure should be initiated only when erratic or divergent behavior is
suspected. This being the case, the initial steps of the search procedure should take advantage of
previously computed information such as r(0) and r(1). Similarly (22.27), suitably safeguarded,
can be utilized for the interpolation process in steps 3 and 4; its key advantage being that it does
not require g derivatives.
§22.4. References
[22.1] OTTER, J. R. H., “Computations for Prestressed Concrete Reactor Pressure Vessels Using Dynamic
Relaxation,” Nucl. Struct. Engng., 1, 61–75, 1965.
[22.2] UNDERWOOD, P. G., “Dynamic Relaxation — A Review,” Ch. 5 in Computational Methods
for Transient Dynamic Analysis, (eds. T. Belytschko and T. J. R. Hughes), North-Holland,
Amsterdam, 1983.
[22.3] PARK, K. C., “A Family of Solution Algorithms for Nonlinear Structural Analysis Based on the
Relaxation Equations,” Int. J. Num. Meth. Engng., 18, 1337–1347, 1982.
[22.4] RIKS, E., “The Application of Newton’s Method to the Problem of Elastic Stability,”, J. Appl.
Mech., 39, 1060–1065, 1972.
[22.5] CRISFIELD, M. A., “Incremental/Iterative Solution Procedures for Nonlinear Structural Analy-
sis,” in Numerical Methods for Nonlinear Problems, Vol. 1, (eds. C. Taylor, E. Hinton and
D. R. J. Owen), Pineridge Press, Swansea, U. K., 261–290, 1980.
[22.6] CRISFIELD, M. A., “An Incremental-Iterative Algorithm that Handles Snap-Through,” Computer
& Structures, 13, 55–62, 1981.
[22.7] KUBÍČEK, M. and HLAVÁČEK, V., Numerical Solution of Nonlinear Boundary Value Prob-
lems with Applications, Prentice-Hall, Englewood Cliffs, New Jersey, 1983.
22–10
22–11 §22.4 REFERENCES
22–11
21
.
Newton-Like
Methods
21–1
Chapter 21: NEWTON-LIKE METHODS 21–2
TABLE OF CONTENTS
Page
§21.1. Introduction 21–3
§21.2. Newton Iteration as a Dynamical System 21–3
§21.2.1. Corrective Process for Fixed λ . . . . . . . . . . . . 21–4
§21.2.2. Corrective Process for Varying λ . . . . . . . . . . 21–6
§21.3. Relaxed Newton Methods 21–6
§21.4. Damped Newton Methods 21–6
§21.5. Chord and Modified Newton Methods 21–7
§21.6. Quasi-Newton Methods 21–8
§21.7. *Convergence of Modified Newton 21–8
§21. Exercises . . . . . . . . . . . . . . . . . . . . . . 21–10
21–2
21–3 §21.2 NEWTON ITERATION AS A DYNAMICAL SYSTEM
§21.1. Introduction
The conventional Newton method (CNM) described in Chapter 20 is hindered by two major short-
comings:
High cost. The tangent stiffness matrix Kk = K(uk , λk ) has to be formed and factored at each
iteration step.
Low Reliability. Convergence to the desired solution is not guaranteed unless the initial estimate
is sufficiently close. The method may diverge, or converge to an unwanted solution. This is quite
likely in the vicinity of bifurcation points.
Because of these shortcomings many variations of CNM, collectively called Newton-like methods,
have been proposed and implemented over the past four decades, with varying degree of success.
Among the most important are
1. Relaxed Newton Methods (RNM): for reliability
2. Damped Newton Methods (DNM): for reliability
3. Modified Newton Methods (MNM): for efficiency
4. Quasi-Newton Methods (QNM): for efficiency
As can be seen by the large number of variations, no modification can be said to be uniformly superior
to others. The above list covers the most important so-called Newton-like methods. Question arise,
however, as to how far the offsprings can deviate from the parent and still be called Newton-like.
Authors have different opinions in this matter. To further complicate things, problem-adaptive
combinations of these techniques are often used in advanced nonlinear solvers.
Some of these variants, notably the Relaxed Newton methods (RNM), are more easily derived
by interpreting the Newton method in the context of a dynamical system. This interpretation is
discussed next.
Figure 21.1 sketches what the chief goal of the Newton method as corrector is: to allow large
incremental steps by eliminating the drift error.
As discussed in previous Chapters, the incremental phase is driven by the first order rate form ṙ = 0,
in which the pseudo-time t is measured by an “increment clock.” Penalize the drift error by adding
a term proportional to the residual r:
ṙ + W r = 0, (21.1)
in which W is a positive-definite residual weighting matrix, which for the moment is left arbitrary.
This is called a first order corrective form, and also a first order relaxation form. It obviously
reduces to ṙ = 0 on an equilibrium path r = 0. The job of the penalty term W r is to force the
solution trajectories of (21.1) to approach r = 0 as the pseudotime t runs along a “corrective clock”
See Figure 21.2.
Figures 21.1 and 21.2 are a bit deceptive in that they depict corrective processes for a one-DOF
problem. A more realistic state of affairs can be observed in Figure 21.3, which depicts trajectories
of a corrective process on a constraint surface for the case of two DOFs.
21–3
Chapter 21: NEWTON-LIKE METHODS 21–4
λ 2 3 Equilibrium Path
Co ∆λ1
rre
ct
1 ∆u1
t
dic
Pre
Increment Control
∆λ0 Constraint
0 ∆u0
u
R
To bring explicitly the stiffness matrix and incremental load vector into play, insert ṙ = Ku̇ − qλ̇
into the above and transfer qλ̇ to the right hand side:
Remark 21.1. A second order corrective form generalizes (21.1) by taking the second order differential
equation
r̈ + W1 ṙ + W2 r = 0, (21.3)
in which W1 and W2 are weighting matrices. These have different function: W1 provides damping while W2
is a “conditioner.” This more general form is not analyzed here.
Suppose that we are at {uk , λk } at which the tangent stiffness Kk is nonsingular. We want to move
to a new state {uk+1 , λk } closer to r = 0 while keeping λ = λk fixed. This can be done by treating
(21.2) with the Forward Euler integrator
uk+1 = uk − h Fk Wk rk , (21.5)
Kk d = −h Wk rk (21.6)
If now we take
W = I, h=1 for any k, (21.7)
21–4
21–5 §21.2 NEWTON ITERATION AS A DYNAMICAL SYSTEM
λ
corrective clock
t=t 0 r=0
t=t K
k
loc
al c
ent
rem
inc
0
u
R
Figure 21.2. Figure 21.2. Pseudo time t running along an “incremental clock”
interspersed by a“corrective clock.” K is the total number of corrective iterations.
λ Incremental contraint
surface c = 0
r=0
u2
u1
Figure 21.3. The corrective process for two degrees of freedom. The challenge is to
end at a solution no matter where one starts on the constraint surface.
we obtain the conventional Newton method (CNM) for fixed λ, as can be easily verified. A variant
called Relaxed Newton, discussed below, results by letting h be adjustable.
21–5
Chapter 21: NEWTON-LIKE METHODS 21–6
Suppose next that λ is to be let vary while satisfying the scalar constraint c( u, λ) = 0 that
controls the increment size. The corrective equation can be generalized as
ṙ + Wr 0
= (21.8)
ċ 0
Inserting ṙ = Ku̇ − qλ̇ and ċ = aT u̇ + g λ̇, in which a = ∂c/∂u and g = ∂c/∂λ, one gets
Ku̇ + Wr qλ̇
= (21.9)
a u̇ + g λ̇
T
0
Integrating (21.9) with (21.10), followed by setting W = I and h = 1, yields the conventional
Newton method for general increment control treated in the previous Chapter. The verification is
the matter of an exercise.
One commonly used variant of CNM aims to increase the reliability but not necessarily lower the
cost per iteration. This is done by deriving CNM from the dynamical process described in the
previous section, and letting the steplength h be a variable. The
K −q d r
=− , (21.11)
a T
g η c
where supercript k is suppressed from K, q, etc., to reduce clutter. Solving for d and η as explained
in the previous Chapter, one then corrects
1 Some authors called this the damped Newton-Raphson method but that name is reserved here for the variant discussed
in the next section.
21–6
21–7 §21.5 CHORD AND MODIFIED NEWTON METHODS
The Relaxed Newton Methods provide gains in reliability as long as the stiffness matrix is not
singular or ill-conditioned. But it does not help in the vicinity of critical points. For example if
K is exactly singular, system (21.11) is not solvable by the double RHS method discussed in the
previous Chapter, and the variable steplength device does not help.
Critical points may come in many flavors. In order of increasing traversal difficulty: isolated limit
points, isolated bifurcation points, initially singular structures, and clustered limit and/or bifurcation
points.
For the less difficult cases, moving away slightly from the singularity often works. Much tougher
is the case when stiffness matrix at the start of the analysis, or of an analysis stage, may be highly
singular. This happens, for instance, in some cable, pneumatic and biological structures that are
mechanisms in the reference configuration and acquire stiffness as they deform. For such cases a
variants collectively known as the Damped Newton Method or DNM, can be effective at the coist
of programming complexity. [The name of Regularized Newton is also used.]
DNM overcomes the singularity problem by adding a diagonal correction to the stiffness matrix.
Instead of (21.10) one solves
K + γ D −q d r
= , (21.13)
a g η c
If the damping coefficient γ is zero and h is unity, the CNM results. As γ is increased the method
approaches steepest descent if D = I and scaled steepest descent for general D. This has been useful
in conjunction with cable net structures that must traverse highly singular regions. Two choices for
D tried in that case are
rT Kr
D = βI, β= T
r r
T (21.15)
r K Kr K
D = βD K , β=
rTK r
where D K = diag K and r K = D−1
K r. Practical values for the damping coefficient γ may be
characterized as follows:
21–7
Chapter 21: NEWTON-LIKE METHODS 21–8
21–8
21–9 §21.7 *CONVERGENCE OF MODIFIED NEWTON
where x = u − u(∞) is the distance to the equilibrium solution at t = ∞. This can be modally decomposed
as
ẏ = −µy (21.19)
in which µ are the roots of the symmetric eigenproblem K0 z = Kz and y are modal amplitudes. The
appropriate eigenvalue for the Newton direction u̇ can be estimated by the Rayleigh quotient
u̇T Ku̇
µ= (21.20)
u̇T Ku̇
Remark 21.2. To reduce the variation of µ, one may reduce the incremental step, or proceed to reform the
stiffness matrix. In some programs the strategy control attempts to cut the incremental steplengths; after two
or three unsuccesful attempts the stiffness is reformed.
21–9
Chapter 21: NEWTON-LIKE METHODS 21–10
EXERCISE 21.1 [A:10] (Very easy, just to get acquanted with the relaxation equation). Verify that (21.4) and
(21.5) followed by W = I and h = 1 lead to the Conventional Newton Method (20.25) for load control.
EXERCISE 21.2 [A:20] Starting from (21.9) and (21.10) derive the general form of the Relaxed Newton
method. Verify that if W = I and h = 1 this reduces to the Conventional Newton system for general
increment control defined by (20.12)-(20.14).
EXERCISE 21.3 [A:20] Find out which W leads to the Damped Newton system (21.13).
21–10
20
.
Conventional
Newton
Methods
20–1
Chapter 20: CONVENTIONAL NEWTON METHODS 20–2
TABLE OF CONTENTS
Page
§20.1. Introduction 20–3
§20.2. Stage Analysis Review 20–3
§20.3. Problem Statement 20–3
§20.4. The Corrective Phase 20–4
§20.5. Solving the Newton Systems 20–5
§20.6. Termination Tests 20–5
§20.7. The Ordinary Newton Method 20–6
§20. Exercises . . . . . . . . . . . . . . . . . . . . . . 20–10
20–2
20–3 §20.3 PROBLEM STATEMENT
§20.1. Introduction
In the overview of solution methods given in Chapter 16 it was noted that solution methods based on
continuation generally included two phases: incremental and corrective. In the purely incremental
methods covered in Chapter 17–19 the corrective phase is absent. If the corrective phase is present,
the incremental formula simply functions as a predictor that provides a starting point for the
corrective iteration. The purpose of this iteration is to eliminate (or at least reduce) the drifting
error by moving towards the equilibrium path along the constraint hypersurface.
Solution methods that include a corrective phase will be collectively called corrective methods,
although perhaps a more appropriate name would be predictor-corrector methods. There are purely
corrective methods that lack a predictor phase entirely (for example, the orthogonal trajectory
accession method) but they have not proven important in practical applications.
The most important class of corrective methods pertains to the Newton-Raphson method and its
numerous variants: modified, modified-delayed, damped, quasi, and so forth. These are collectively
called Newton-like methods, and only require access to the past solution. In the present section we
study the conventional Newton method under general increment control.
r(u, λ) = 0 (20.1)
over a loading stage as the control parameter is incremented from 0. As previously discussed the
additional equation that makes (20.1) determinate is the increment constraint equation
c(u, λ) = 0. (20.2)
that characterizes numerically the response u = u(λ) while satisfying the residual equations (20.1)
within prescribed accuracy.
The purely incremental methods covered in the three previous Chapters compute a sequence of
values such as (20.3) by direct integration of the first-order rate equations Ku = q or Kr e f u =
q − f . The methods considered here implement a corrective phase in which one iterates for
equilibrium while satisfying the increment constraint. The starting point for the corrective phase
is the solution predicted by the incremental method. Consequently, these methods are often called
incremental-iterative methods.
20–3
Chapter 20: CONVENTIONAL NEWTON METHODS 20–4
r(un+1 , λn+1 ) = 0,
(20.6)
c(un , λn ) = 0,
where
un = un+1 − un , λn = λn+1 − λn (20.7)
Although the above increment constraint is a special case of (20.2), it befits those most commonly
used in practical calculations.
The predicted solution
u0n , λ0n , (20.8)
is typically obtained by performing an incremental step as described in Chapters 17 and 18.
uk , λk , (20.10)
20–4
20–5 §20.6 TERMINATION TESTS
where
∂r ∂r ∂c ∂c
K= , q=− , aT = , g= , (20.14)
∂u ∂λ ∂u ∂λ
and all known quantities are evaluated at uk , λk . Note that for notational simplicity this superscript
has been kept out of d, η, r, etc, unless it is desirable to make the dependency on the iteration index
k explicit. If the tangent stiffness matrix K is of order N , the coefficient matrix of the linear system
(20.13) has order N + 1. This matrix is called the augmented stiffness matrix.
Note that although generally K is symmetric and sparse, the augmented stiffness is generally un-
symmetric (but see Exercise 20.2), and its sparseness is detrimentally affected by the augmentation.
It is therefore of interest to treat the linear system (20.13) with techniques that preserve those at-
tributes. The solution procedures described below make use of auxiliary systems of equations to
achieve that goal. The number of auxiliary systems depends on whether the tangent stiffness K
is nonsingular (regular points) or singular (critical points). For the latter we have to distinguish
between limit points and bifurcation points. In the present section we shall concentrate on the
treatment of regular points.
Recall from Chapter 4 that regular points of the system (20.1) are equilibrium solutions (u, λ) at
which the tangent stiffness matrix K is nonsingular. If this property holds, we can perform forward
Gauss elimination on (20.13) to get rid of d and produce the following scalar equation for η:
Then
c + aT dr
η=− , d = dr + ηdq . (20.17)
g + aT dq
It is seen that two right hand sides, r and q, have to be generally solved for at each Newton step. The
number reduces to one for k > 1, however, if modified Newton is used so that K is held fixed for
several steps and q does not vary. The last assumption holds in structural mechanics applications
if the loading is conservative and proportional. (The modified Newton method is described in the
next Chapter).
20–5
Chapter 20: CONVENTIONAL NEWTON METHODS 20–6
2. Residual convergence test. Since the residual r measures the departure from equilibrium,
another appropriate convergence test is
r ≤ r . (20.19)
|rT d| ≤ d r (20.20)
2. Since d and r have usually physical dimensions, so do necessarily d and r . For a general
purpose implementation of Newton iteration this dependency on physical units is undesirable
and it is more convenient to work out with ratios that render the r and d dimensionless. For
example:
r
≤ r (20.21)
r0
where r0 is the residual after the predictor step; now r can be dimensionless. A similar ratio
can be used for the displacement convergence test, but here the reference value should be a
total or accumulated displacement; for example:
d
≤ d (20.22)
u0
3. Divergence Safeguards. The Newton iteration is not guaranteed to converge. There should
therefore be divergence detection tests that will cause the iteration to be interrupted. For ex-
ample, turning the above ratios around, divergence may be diagnosed if either of the following
inequalities occur:
r d
≥ gr , ≥ gd (20.23)
r
0 u0
where gr and gd are “dangerous growth” factors, for example gr = gd = 1000.
Occasionally the Newton iteration will neither diverge not converge but just “bounce around”
(oscillatory behavior). To avoid excessive wheel spinning in such cases it is always a good
practice to put a maximum number of iterations per step in the program. Typical limits might
be 20 to 50.
20–6
20–7 §20.7 THE ORDINARY NEWTON METHOD
f[x_,a_]:=x^2-a; Df[x_]:=2*x;
a=3.; xk=1.; Print["x0=",xk];
For[k=0, k<6, k++, xkp1=xk-f[xk,a]/Df[xk];
Print["x",k+1,"=",xkp1//InputForm]; xk=xkp1];
x0=1.
x1=2.
x2=1.75
x3=1.732142857142857143
x4=1.732050810014727541
x5=1.732050807568877295
x6=1.732050807568877294
√
After six cycles the iteration yields 16 places of accuracy for 3. Note that the number of exact digits roughly
doubles from k = 2 onwards. This is typical of the Newton iteration once it “locks in” a root because the
process has asymptotically quadratic convergence.
The numerical process shown above is actually that used by mathematical software libraries of languages like
Fortran or C for the computation of the square root function. However the initial value is determined by a
scaled rational interpolant that gives 2-4 digits of accuracy for x0 ; as a result only 2 or 3 cycles are needed to
achieve double-precision accuracy for most inputs.
20–7
Chapter 20: CONVENTIONAL NEWTON METHODS 20–8
Example 20.2. This example is more typical of a structural application. Consider the residual equation for a
2-DOF system
r1 u 1 + 2u 31 − u 22 − 2λ
r(u, λ) = = (20.27)
r2 3u 2 − 2u 1 u 2 − λ
The tangent stiffness matrix is
∂r
1 + 6u 21 −2u 2
K= = (20.28)
∂u −2u 2 3 − 2u 1
For λ = 1 the residual equations have three real roots which to six digits of accuracy are
1 1.18400 1.66726
u1 = , u2 = , u3 = . (20.29)
1 1.58227 −2.98938
The following Mathematica programs starts from the initial values u 01 = u 02 = 0.8 and quickly finds the
nearest root u 1 = u 2 = 1, delivering 16 digits of accuracy after 5 cycles:
ClearAll[r,u1,u2,lambda];
lambda=1;
r[u1_,u2_]:={{u1+2*u1^3-u2^2-2*lambda},{3*u2-2*u1*u2-lambda}};
Kt[u1_,u2_]:={{1+6*u1^2,-2*u2}, {-2*u2,3-2*u1}}
uk={{0.8},{0.8}};
Print["Starting v0=",uk//InputForm];
For [k=0, k<5, k++,
{u1,u2}={uk[[1,1]],uk[[2,1]]};
ukp1 = uk - Inverse[Kt[u1,u2]].r[u1,u2];
Print["Cycle k= ",k," u",k+1,"=",ukp1//InputForm];
uk=ukp1 ];
{u1,u2}={uk[[1,1]],uk[[2,1]]};
Print["Final residual=",r[u1,u2]];
But changing the initial values to u 0 = 0.8 and u 1 = 1.1, which is even closer to the u 1 = u 2 = 1 root, the
process converges to the second root in (20.29), reaching 16 digits of accuracy after 8 cycles:
ClearAll[r,u1,u2,lambda];
lambda=1;
r[u1_,u2_]:={{u1+2*u1^3-u2^2-2*lambda},{3*u2-2*u1*u2-lambda}};
Kt[u1_,u2_]:={{1+6*u1^2,-2*u2}, {-2*u2,3-2*u1}}
uk={{0.8},{1.1}};
Print["Starting u0=",uk//InputForm];
For [k=0, k<5, k++,
20–8
20–9 §20.7 THE ORDINARY NEWTON METHOD
{u1,u2}={uk[[1,1]],uk[[2,1]]};
ukp1 = uk - Inverse[Kt[u1,u2]].r[u1,u2];
Print["Cycle k= ",k," u",k+1,"=",ukp1//InputForm];
uk=ukp1 ];
{u1,u2}={uk[[1,1]],uk[[2,1]]};
Print["Final residual=",r[u1,u2]];
This illustrates the “finicky” nature of Newton iteration. It can do (and often does) the unexpected, such as
diverging or converging to the “wrong” root. In fact the a whole subset of fractal or chaotic mathematics
is devoted to the understanding of “domains of attraction” of roots. Because of this capricious behavior, in
practical use of the Newton corrector numerous safeguards are implemented to avoid surprises. But the whole
subject is too lenghty for coverage in an introductory treatment.
20–9
Chapter 20: CONVENTIONAL NEWTON METHODS 20–10
EXERCISE 20.1 (A:15) If the case of load control, the incremental stepsize constraint reduces to
c = λn − n = 0 (E20.1)
where
n is prescribed. After the predictor step, c = 0. Show that the Newton method reduces to a standard
form that requires only the solution of one of the auxiliary systems (20.16).
EXERCISE 20.2 (A:15) What algebraic constraint c = 0 would make the augmented tangent stiffness sym-
metric?
EXERCISE 20.3 (C:25) Enlarge the program that solves Exercise 18.2 by using conventional Newton as
corrector, the incremental stepsize constraint
c = |sn | − n = 0, (E20.2)
(where
=
n is kept fixed throughout) and the termination tolerance |rk /r0 | ≤ 10−3 or 20 iterations, whichever
occurs first. Suggested values for
n for all n is 0.01, but the results should be insensitive to that choice.
EXERCISE 20.4 (A:30) (Advanced, paper level) Obtain the expression of the augmented stiffness if the
orthogonal trajectory accession method is used and write out the Newton algorithm for this case.
20–10
19
.
Pseudo-Force
Incremental
Methods
19–1
Chapter 19: PSEUDO-FORCE INCREMENTAL METHODS 19–2
TABLE OF CONTENTS
Page
§19.1. Pseudo Force Formulation 19–3
§19.2. Computing the Reference Stiffness and Internal Force 19–4
§19.3. Integration of Pseudo-force Rate Equation 19–5
§19.3.1. Forward Euler Integration under Load Control . . . . . . 19–5
§19.3.2. Pseudo-Force Extrapolation . . . . . . . . . . . . 19–6
§19.3.3. Iterative Improvement . . . . . . . . . . . . . . 19–6
§19.4. Numerical Stability 19–7
§19.5. Accuracy Control 19–8
§19.6. Secant Estimation of n 19–8
§19.7. General Increment Control 19–8
19–2
19–3 §19.1 PSEUDO FORCE FORMULATION
The disadvantages of purely incremental methods in terms of solution “drift” and high computational
expense were recognized as applications of nonlinear finite element analysis expanded to cover
wider classes of problems. Many techniques aimed at avoiding these difficulties were proposed
and tested during the late 1960s and early 1970s. From the mass of experience accumulated during
this period, two principal strategies emerged.
For path-independent, smooth nonlinearities typified by finite deflection and nonlinear-elastic be-
havior, the power of Newton-like corrective methods was eventually recognized. In such problems,
incremental methods were relegated to the secondary role of predictors for starting a corrective
process.
For path-dependent material nonlinearities, however, purely incremental methods have remained
important because of the reasons noted in Chapter 18. These methods underwent modifications
aimed primarily at reducing the computational expense while retaining the advantages of numerical
stability, implementation simplicity and physical transparency. Unfortunately these goals, being
largely contradictory, can only be met half-way. The most successful attempt in this direction has
been the development of pseudo-force incremental methods, which are covered below.
In the pseudo-force reformulation of incremental methods the pervasive role that the tangent stiffness
matrix plays in the methods discussed in Chapters 17-18 is relaxed. Instead the deviation from a
“reference linear response” is collected in a pseudo-force vector. This approach allows a reference
stiffness matrix to be reused over many incremental steps. Since this avoids having to repeatedly
assemble and factor the tangent stiffness matrix, the gain in speed per step over the conventional
incremental methods may be very substantial in two- and three-dimensional problems. There is no
free lunch, however, for the speed gain is counterbalanced by two disadvantages:
1. Pronounced accuracy loss as nonlinearities become severe, which may force extremely small
increments to be taken.
f = Kr e f (u − ur e f ) + n, (19.2)
where Kr e f is a nonsingular reference stiffness matrix that is kept fixed as long as possible and ur e f
is the state at λ = λr e f (usually 0). The deviation
n = f − Kr e f (u − ur e f ), (19.3)
19–3
Chapter 19: PSEUDO-FORCE INCREMENTAL METHODS 19–4
is called the pseudo-force vector. Inserting (19.2) in the residual expression (19.1) we get
r = Kr e f (u − ur e f ) + n − p = 0. (19.4)
r = Kr e f u + n − q = 0, (19.5)
where q = ∂p/∂λ is the incremental load vector. Since Kr e f is assumed nonsingular, solving for
u yields
u = Kr−1
e f (q − n ) = v ,
f
(19.6)
Vector v f is called the pseudo incremental velocity. It plays a similar role to that of v = K−1 q in
the conventional incremental methods discussed in Chapters 14 and 15.
Remark 19.1. The pseudo-force vector n may be viewed as a “force deviation from the linear reference
response” Kr e f (u − ur e f ). Now if the structural response is linear so that
choosing Kr e f equal to the linear stiffness K0 and ur e f = 0 gives n = 0. For this reason many authors call n
the nonlinear force vector.
Remark 19.2. Differentiating (19.3) with respect to u gives, for a separable residual,
∂r ∂n
= K = Kr e f + Knon , Knon = . (19.8)
∂u ∂u
If Kr e f is kept equal to the linear stiffness throughout, Knon is called the nonlinear stiffness matrix.
r = K0 (u − u0 ) + n − p = 0. (19.9)
Response calculation procedures based on the choice (19.9) are sometimes called the initial stiffness
method.
Another common strategy is to start with the initial stiffness and continue with as many incremental
steps as possible, resetting Kr e f = Kn if numerical stability or accuracy problems are detected at
the (n + 1)th step. Still a third approach is to keep Kr e f for a preassigned number of incremental
steps, say m, and update Kr e f at steps 0, m, 2m, . . ., unless numerical stability or accuracy problems
are encountered.
The effectiveness of pseudo-force methods in finite element programming depends largely on the
ability to compute the internal force vector f directly on an element-by-element basis. Recall, for
example, that in the core-congruential formulation of the Total Lagrangian (TL) description,
f= GT si bi d V0 , (19.10)
V0
19–4
19–5 §19.3 INTEGRATION OF PSEUDO-FORCE RATE EQUATION
where V0 is the reference volume, G the transformation between displacement gradients and physical
degrees of freedom, si are PK stresses in the current configuration, and bi are the vectors defined
in §8.5. Note that expression does not explicitly involve material properties, and is consequently
applicable to problems with material nonlinearities. More precisely: the only requirement for using
(19.9) is the availabilibity of the stresses si in the current configuration whereas the procedure by
which such stresses are obtained is irrelevant.
Remark 19.3. In finite element work the synonyms initial force method, initial stress method and initial strain
method have been associated with restricted versions of what we call here pseudo-force methods. These names
focus attention on various physical interpretations of the calculation of the f term. A heated controversy as to
which version was the best took place in the late sixties; from current perspective such arguments have only
historical interest.
nn = fn − Kr e f (un − ur e f ),
nn − nn−1
nn = ,
λn−1 (19.11)
un = Kr−1
e f (qn − nn ) λn = vn λn ,
f
un+1 = un + un , λn+1 = λn + λn .
19–5
Chapter 19: PSEUDO-FORCE INCREMENTAL METHODS 19–6
r = 5u − u 3 − λ = 0, for λ = 4, (19.12)
using the pseudo-force method (19.11) with two λ increments of 2.0, K r e f = K 0 = 5 and u r e f = u 0 = 0. The
exact solution on the fundamental path is v(λ) = v(4) = 1. Since q = −∂r/∂λ = 1, and n = 5u − u 3 − 5u =
−u 3 , the rate form is r = 5u − (u 3 ) − 1 = 0.
The first increment, with λ0 = 2 specified, is
n 0 = 0, u 0 = K r−1
e f (q0 − n 0 ) λ0 = 0.4, u 1 = u 0 + u 0 = 0.4, λ1 = λ0 + λ0 = 2.
u 2 = u 1 + u 1 = 0.8128, λ2 = λ1 + λ1 = 4.
Repeating these computations with 4, 8 and 17 equal increments of λ gives 0.8461, 0.8841 and 0.9190,
respectively. As can be observed the accuracy attained is low. Table 19.1 compares these values with those
obtained with other methods.
Another way to improve accuracy while avoiding the reset of Kr e f is to iterate on un+1 while keeping λn
fixed. To derive an iteration scheme, write the residual form (19.4) at n + 1 and n, subtract, and solve for un :
un = Kr−1
e f (nn+1 − nn ) − qn λn (19.15)
Let k be an iteration step index and u1n+1 be the value obtained from the increment equation. The resulting
iterative scheme is
ukn = Kr−1e f (nn+1 − nn ) − qn λn ,
k
k = 1, . . . (19.16)
n+1 = un + un .
uk+1 k
19–6
19–7 §19.4 NUMERICAL STABILITY
Using the fact that q is independent of u, this can be rewritten in the “subincremental” form
−1 k
n+1 = un+1 − Kr e f rn+1 .
uk+1 (19.17)
k
If Kr e f ≡ Kn , this is precisely the modified Newton-Raphson (MNR) method with unit steplength. This
shows that the the iterated pseudo-force incremental method is a MNR method with an arbitrary selection of
reference stiffness. The properties of these methods are investigated in Chapters that deal with Newton-Raphson
corrective methods.
This condition is independent of the stepsize λ. It is seen that the key for numerical stability is that the
reference stiffness “dominates” the nonlinear stiffness in the sense (19.20). That is, from that standpoint it is
better to overestimate K r e f .
The generalization of (19.19) to N degrees of freedom is
∂n
u = Kr−1 −1
e f u = Kr e f u = Kr−1
e f Knon = Au . (19.20)
∂u
19–7
Chapter 19: PSEUDO-FORCE INCREMENTAL METHODS 19–8
This condition holds regardless of the stepsize λ. Likewise, the iteration (19.17) converges only if the
condition (19.22) holds.
A practical estimator for this eigenvalue is
|Kr−1
e f (nn − nn−1 )|
µ≈ . (19.23)
|un − un−1 |
Remark 19.4. This result is another aspect of the close relationship between pseudo-force incremental methods
and modified-Newton corrective methods. If nonlinearities are substantial, the method diverges regardless of
the increment length used.
Remark 19.5. It is beneficial from a stability standpoint to have Kr e f “dominate” Knon . This happens in
softening structures when the elastic stiffness is selected as Kr e f , and explains the success of the method in
plasticity analysis. On the other hand, if the structure hardens as λ increases (examples: cable and pneumatic
structures), the stability condition is easily violated.
Remark 19.6. The simplest cure to numerical instability is to recompute the reference stiffness. Another (as
yet unexplored) possibility is to correct Kr e f with a rank-one matrix.
19–8
19–9 §19.7 GENERAL INCREMENT CONTROL
Note that the residual rs is computed by holding λ constant and equal to λn and is not rn − rn−1 .
Two choices for the inverse stiffness secant approximant are the Davidon-Fletcher-Powell (DFP)
rank-two update formula
us usT Fr e f rs rsT Fr e f
Fs = Fr e f + − , (19.28)
usT rs rsT Fr e f rs
( us − Fr e f rs )( us − Fr e f rs )T
Fs = Fr e f + . (19.29)
( us − Fr e f rs )T rs
These formulas are collectively called Quasi-Newton updates in the numerical analysis literature,
although to be more precise what we have shown above is just the first member of such updates.
The preceding developments assume that λn is prescribed. But we can readily extend the pseudo-
force technique to general increment control by following the procedures discussed in Chapter 18.
For that we must replace vn by
vnf = Kr−1
e f (qn − ṅn ). (19.30)
where the superposed dot denotes derivative respect to the pseudotime parameter t pertinent to the
increment strategy chosen. Of particular importance is arclength control, in which t becomes the
arclength s. Criteria for stability and accuracy are readily converted to this case.
19–9
18
.
Purely Incremental
Methods:
General Control
18–1
Chapter 18: PURELY INCREMENTAL METHODS: GENERAL CONTROL 18–2
In this Chapter we continue the development of the purely incremental methods under general
incremental control conditions. A general nonlinear transformation of state and control parameters
to pseudo-time is introduced. Computable forms of this constraint are specialized to arclength
and hyperelliptic control. The computational implications of these decisions are discussed, and a
practical implementation of hyperplane-distance arclength control outlined.
In previous Chapters it was noted that continuation solution methods under λ (load) control have
difficulty traversing limit points. This shortcoming can be circumvented through the use of more
general increment control schemes. This generalization can be practically effected by adjoining
algebraic constraint equations such as those listed in Chapter 16.
These more general forms of increment control can be described on a uniform basis as follows.
Because the main effect of enforcing the constraint c(u, λ) = 0, is to link the increments of u
and λ, we express the response in the pseudo time parametric form
u = u(λ), (18.2)
of the equilibrium path, and nothing would be gained. But the pseudotime t in (18.1) is now at our
disposal and we can try to do better. Differentiating (18.1) with respect to t we get
§18.1.1. Requirements
The incremental path equation becomes
∂r(u, λ) ∂r ∂r
ṙ = = u̇ − λ̇ = Ku̇ − q/ f = 0. (18.5)
∂t ∂u ∂λ
At limit points dλ = 0. By “smooth traversal” of a limit point is meant that t varies regularly as a
“vehicle odometer” as that limit point is crossed. Consequently f = dt/dλ must go to infinity at
a limit point. This reasoning shows why t ≡ λ does not work, because if so f = 1 everywhere. It
also follows that the relation between λ and a useful t must be necessarily nonlinear, for if we take
t = c1 λ + c2 with c1 = 0, f is always 1/c1 .
It is computationally desirable, however, that in “almost linear” portions of the response f ap-
proaches a finite value, because in the limit we would like to recover the conventional methods of
linear structural analysis.
18–2
18–3 §18.1 PARAMETRIC FORM
Now which quantity becomes infinite at limit points? An obvious choice is the incremental velocity
v = K−1 q or, more precisely, some norm of it. A particularly attractive choice in light of its
geometric significance is
f = 1 + vT v, (18.6)
This corresponds to taking
dt ≡ ds = f dλ = v du + dλ, (18.7)
As shown in Chapter 4, ds is the differential arclength of the response curve in state-control space.
Consequently s is the arclength traversed along this curve measured from some arbitrary point,
such as the last solution. Note the following attractive features of this choice:
(d) The exact arclength sn from the last computed solution at Pn (un , λn ) to another point P(u, λ)
on the equilibrium path is given by the path integral
sn = ds = 1 + vT v ds. (18.8)
Pn →P Pn →P
But as written this expression is not directly computable because it requires knowledge of
v = K−1 q along the equilibrium path that emanates from the last solution, which is precisely
what we want to compute. Consequently an increment control constraint such as
sn =
n , (18.9)
(e) The choice (18.6) intermixes state parameters, which have generally physical dimension (of
displacement), with the dimensionless scalar 1. This mixture can introduce difficulties in
that the computed solutions are not invariant with respect to the choice of state parameter
dimensions. It can be corrected by employing appropriate state-scaling techniques as discussed
in Chapter 4.
(f) Traversal of bifurcation points remain difficult or impossible without additional “tricks”. The
key roadblock is that the continuous parametrization (18.1) breaks down at bifurcation points,
18–3
Chapter 18: PURELY INCREMENTAL METHODS: GENERAL CONTROL 18–4
because it cannot simultaneously represent the two or more branches that intersect there.1
Circumventing this difficulty within the context of purely incremental methods is not easy,
and consideration is deferred to later Chapters.
We now study two computable approximations to the arclength parametrization.
in which
n , which controls the magnitude of the n th incremental step, is either specified or auto-
matically adjusted through some accuracy control rule, as discussed below.
Substituting un = vn λn into the above we get the Forward Euler incremental scheme
n
vn = K−1
n qn , f n = 1 + vnT vn , λn = ± ,
fn (18.11)
un+1 = un + vn λn , λn+1 = λn + λn .
2
n−1
n = (18.12)
an−1
n ≥
min , (18.13)
which avoids “getting stuck” at limit points under certain conditions, as well as a maximum arclength
travel distance condition
n ≤
max , (18.14)
to avoid surprises in hitting rapidly changing portions of the response.
1 A second order rate form is in fact required in the vicinity of a simple bifurcation point.
18–4
18–5 §18.3 GLOBAL HYPERELLIPTIC CONSTRAINT
Remark 18.2. In a poorly scaled problem the previous scheme should be modified by replacing f by its scaled
equivalent:
fn = 1 + vnT S2 vn , (18.16)
where S2 is a diagonal scaling matrix. Some appropriate choices for S2 are discussed in the next Chapter.
18–5
Chapter 18: PURELY INCREMENTAL METHODS: GENERAL CONTROL 18–6
Remark 18.3. Equation (18.19) represents a “closed” constraint surface (a hyperellipse in state-control space)
and thus it may be thought of as “safer” than the hyperplane distance constraint, which is “open”. Although
this consideration has some merit if a corrective phase that moves on that surface follow, it has no weight
in a purely incremental method. One drawback of the hyperelliptic form is the need to choose an and bn in
addition to
n at each step. Although this adds flexibility, it can complicate the implementation and require
more run-time decisions.
Remark 18.4. The scalar uT K u is not necessarily positive if K is indefinite; hence the need for taking
the absolute value of qT v in (18.20).
Remark 18.5. Diagonal scaling on displacement increments provides an intermediate choice between unscaled
and K-scaled forms. That is, choose S2 = diag(K).
ds ∂u ∂λ
√
in which v f = v/ f = v/ 1 + vT v. The amplification matrix A is that shown in brackets above. So far this
problem appears to be analytically intractable and no general conclusions may be drawn so far. However,
numerical evidence show that arclength-control incremental methods are as stable as those using load control.
Thus the empirical rule of Chapter 14 may be used; namely if
is adjusted for accuracy, setting < 1 takes
care of stability.
18–6
18–7 §18.8 ASSESSMENT OF PURELY INCREMENTAL METHODS
Suppose that the computed solution process approaches an isolated limit point at which K has the
null eigenvector z normalized to unit length, that is
Kz = 0, zT z = 1. (18.26)
As the critical point is approached, it can be shown (using the theory of inverse iteration) that
That is, v tends to become parallel to z although its norm goes to (plus or minus) infinity. But the
f -normalized v does approach the normalized eigenvector:
v
v f = 1/ f v = √ → ±z. (18.28)
1 + vT v
and the incremental equations approach
u ±z
=
(18.29)
λ 0
Therefore the f -normalization of v automtically take care of aligning the incremental direction
normally to the λ axis.
Remark 18.6. Some minor safeguards remain: if the tangent stiffness is evaluated at or very near the limit
point, K(u) may be numerically singular, in which case a simple remedy is to change u by a tiny amount and
try again (a more theoretically sound approach based on penalty spring stabilization is discussed later). This
simple technique is used in the computer program for Exercises 18.1–18.5. Also if
is automatically adjusted
by accuracy requirements it must not be allowed to fall under a minimum value.
We are now in a position to give the outline of a arclength-controlled incremental algorithm essen-
tially based on the scheme described previously. This is done in Table 18.1. The steps listed therein
are for a single stage. (In programs that accept multistage analysis, “stop” is replaced by a save,
exit, check error conditions and recover if possible, reset and restart process with the next stage.
Although programming that sequence can be elaborate, it does not involve conceptual difficulties.)
The simplest load-controlled incremental scheme described in Chapter 17 has been extensively
used as a stand-alone method (that is, not combined with equilibrium corrections) in early imple-
mentations (1955-1965) of finite-element-based nonlinear structural analysis. It survives in this
primitive form in a surprisingly large proportion of finite element computer programs; particularly
those aimed at treating highly nonlinear material behavior. From current perspective it suffers from
two serious disadvantages.
Drift Error. The residual force vector r never enters the calculations. Consequently, the deviation
from the equilibrium path due to the propagation and accumulation of local integration errors cannot
18–7
Chapter 18: PURELY INCREMENTAL METHODS: GENERAL CONTROL 18–8
Step 1. Form and factor stiffness matrix Kn . If the factorization fails on account of
singularity, perturb u by a tiny amount and repeat. If this failure repeats after
a certain number of tries, stop with appropriate error message.
side qn and solve Kn vn = qn for the incremental velocity vn .
Step 2. Form right-hand
Form f n = 1 + vnT vn .
Step 3. If an adaptive stepsize scheme is used, adjust
n as per §18.2, else keep
n =
0 .
Set λn =
n / f n , and give it the sign of qnT vn .
Step 4. Compute un = vn λn . Advance un+1 = un + un and λn+1 = λn + λn .
If adaptive
control is used, save v f to be used in the estimation of an−1 .
Step 5. If |un+1 | exceeds u max , or |λn+1 | exceeds λmax , or n exceeds n max , stop. Else
set n ← n + 1 and return to Step 1.
If scaling of v is introduced because of the physical dimensionality discrepancy between
λ and u, the simplest implementation is to define a reference length L r e f as part of the
inputs. Then f n = 1 + ṽnT ṽn , in which ṽn = vn /L r e f .
be eliminated by corrective iteration. In practical terms, this means that realistic estimates of the
response tracing accuracy can be obtained only by rerunning the problem with several increment
sizes.
Computational Expense. To keep the drifting error down, many small steps may be required,
particularly in “difficult” regions of the response. But at each step the stiffness matrix must be
formed and factored. This can be an expensive proposition in two- and three-dimensional problems.
To reduce the stiffness recalculation cost, the “pseudo-force” methods discussed in Chapter 19 have
been used extensively in plasticity and viscoelasticity calculations. These methods are affected,
however, by serious numerical stability difficulties.
What are the advantages? First, given the early applications of finite element methods (see Remark
below), incremental methods are quite easy to program as extensions of linear analysis codes. The
absence of the residual vector is in fact helpful, since linear analysis does not require it whereas
the stiffness matrix is always available. And the underlying concept (follow the physics) is readily
understood by practicing engineers. The virtues of simplicity and physical transparency should
18–8
18–9 §18.8 ASSESSMENT OF PURELY INCREMENTAL METHODS
never be underestimated!
Second, in problems that exhibit strong path-dependency, as typified by flow plasticity models,
the many-small-steps requirement is not necessarily a hindrance, since the increments have to be
kept small anyway to avoid unacceptable element-level errors (for example, yield surface drift) in
material-law calculations.
Finally, the high frequency of stiffness matrix assembly and refactorization has a silver lining: it is
difficult to miss critical points. The fact that purely incremental methods have trouble tracing post-
buckling branches beyond bifurcation points is irrelevant to situations in which the determination
of such points, rather than traversing them, is the main objective of the analysis.
Remark 18.7. The preference of incremental over corrective methods in the early implementation of nonlinear
finite element analysis has historical roots. Finite element methods were invented in the aircraft/aerospace
industry where linear analysis dominates. (This also helps to explain the initial popularity of the force method,
which is ill-suited for nonlinear analysis.) Furthermore, early excursions in the nonlinear world involved fairly
mild nonlinearities. Thus, developers of displacement-based finite element codes passed naturally from the
linear stiffness equations
Ku = q (18.30)
to the incremental form
K u = q,
(18.31)
unew = uold + u
18–9
Chapter 18: PURELY INCREMENTAL METHODS: GENERAL CONTROL 18–10
Exercises 18.1 and 18.2 pertain to the response analysis by purely incremental methods, of the nonlinear
response of the two-bar arch example structure of Chapter 8, which was treated by the TL description.
The structure properties are: span S = 2, height H = 1 (hence the rise angle is α = 45◦ ), elastic modulus
E = 1 and cross section area A0 = 1. The applied loads are f X = 0 and f Y = λF where F = −1 is constant.
This gives a downward vertical applied load if λ > 0. For this rise angle, the first limit point of the Total
Lagrangian model analytically occurs at load level λ L ≈ 0.136 and a vertical deflection of u Y L ≈ −0.425.
Traversal of limit points is done through a positive-work advancing criterion: qT v > 0.
Exercises 18.1 and 18.2 should be done with the Mathematica Notebook called IncSolTwoBarArch.nb,
which has been posted on the course Web site. This is a very rough conversion from the ancient Fortran code
used in previous offerings. [The conversion took much longer than anticipated despite the simplicity of the
code.]
Despite the roughness of this implementation, it can be used to illustrate the behavior of two integrators:
Forward Euler (FE) and Midpoint Rule (MR), combined with three increment control strategies: Load Control
(LC), Displacement Control (DC) and Arclength Control (AC). [The Classical Runge Kutta (RK4) integrator
will be added later as an Exercise.] One significant advantage of Mathematica over Fortran is the availability
of built-in graphics. Thus response plots, for example, can be immediately generated as part of the output,
helping quick visualization of method performance.
EXERCISE 18.1 [D:15] Prepare a hierarchical diagram of Cells 1-12 of the IncSolTwoBarArch.nb Note-
book, beginning with the main program given in Cell 12. Note which module calls which and write down the
purpose of each module in one or two lines along the module name. Return this diagram as answer to the
homework.*
EXERCISE 18.2 [C:20] Run the four scripts (main programs) in Cells 12 through 15 of IncSolTwoBarArch.nb.
Briefly explain what they do, and how the four method combinations driven by them stack up in terms of (i)
robustness in traversing limit points, (ii) accuracy in locating the first limit point (analytical values are given
above), and (iii) accuracy in crossing λ = 0 at u Y = −1. Please attach the four response plots of u Y versus λ
to your returned homework (do not bother with printouts). (I forgot to include method labels in the ListPlot
commands to identify which method is which; these may be written by hand).
Note: before you can run those scripts, Cells 1-11 should be initialized. A quick way to accomplish that in
version 3.0 is to click Kernel → Evaluate → Evaluate Initialization. Do this twice to get rid of error message
boxes.
* If you are not sure of what a hierachical diagram is, go to the IFEM Web page:
http://caswww.colorado.edu/courses.d/IFEM.d/Home.html
and look up Exercise 15.7 in Chapter 15.
18–10
18–11 Exercises
EXERCISE 18.3 [C:25] Consider the one-DOF nonlinear problem (not a structure) governed by the force
residual equation
(λ − 1)2 + (u + 1)2 − 2 = 0 (E18.1)
√
This represents a circle of radius 2 and center (1, −1) in the (λ, u) plane. Suppose that one starts at the
reference state u = λ = 0 and tries to trace the response, with a constant stepsize, by going around the circle
counterclockwise and eventually returning to the origin. In doing so you need to cross two turning points and
two limit points. While traversing limit points is easy using Arclength Control and a postive work criterion,
turning points are trouble.
Program from scratch an incremental solution to this problem using the Midpoint Rule, Arclength Control,
constant
, and the angle criterion2 to keep a positive traversal direction, and try it.3 If your implementation
works all the way you are ahead of all commercial nonlinear FEM codes in this regard.
EXERCISE 18.4 [C:25] The fixed-step classical 4th order Runge Kutta integrator (RK4) is implemented in
Cell 10 of the IncSolTwoBarArch.nb Notebook.4 It is identified as integ="RK4". Check visually if the
implementation is correct by comparing to the description, for example, in Numerical Recipes.
Run the same arch problem as in Exercise 8.2 using RK4 combined with Load Control and Arclength Control;
experiment with ell and nmax and report on whether you can obtain high accuracy (see 8.2 for how to assess
it) with a fairly large stepsize, say ell=0.1.
EXERCISE 18.5 [C:25] It was √ shown in Chapter 8 that if the rise angle α of the two-bar arch example structure
exceeds 60◦ , i.e. tan α > 3, it will fail by bifurcation first. For example S = 2 and H = 2 would do it
because tan α = 2. However if you set those inputs and run the programs in IncSolTwoBarArch.nb the
bifurcation point will be completely masked; you will see only the symmetric solution passing two limit points.
One simple technique to make “dumb incremental solvers,” like those provided in the Notebook, pay attention
to bifurcation points is to inject artificial imperfections. This can be done, for example, by putting a fictitious
but tiny load system that disturbs the symmetric response. For example, define the reference crown load in
force as {-0.001,-1} instead of {0,-1}. Set S = H = 2 and play with the tiny X -force, the increment
length ell and number of steps nmax until you see a decent tracing of bifurcation post-buckling: at a certain
load level u X will increase rapidly, signaling that the arch is buckling horizontally. To see that better, do the
ListPlot of u X versus λ collected in list uXvslambda found near the bottom of the driver cell. Comment on
what combination of method and solution parameters let you succeed.
Note (1). Bifurcation experiments can be found in the Notebook IncBifSolTwoBarArch.nb also posted on
the web site.
Note (2). If you take the response long enough you may be able to have the structure return to the primary
symmetric path upon passing through the second bifurcation point, but that may take lots and lots of steps
since the implementation uses a constant stepsize
.
2 The computation of the angle between vn and vn−1 is illustrated in the posted Notebook. It is saved as part of the solution
table, although it is not used in the solution procedure therein. The variable is called a or an and is actually the cosine
of that angle, which is simply the dot product of those velocity vectors normalized to unit length.
3 The positive-work criterion qT v > 0 fails because one needs to release work along some parts of the response trajectory.
In fact the net work on doing a complete circle is zero.
4 For a description of the RK4 algorithm see any book on numerical methods for ODEs. For example the widely used
Numerical Recipes in Fortran; it is presented in Section 16.1 of the second edition.
18–11
17
.
Purely Incremental
Methods:
Load Control
17–1
Chapter 17: PURELY INCREMENTAL METHODS: LOAD CONTROL 17–2
TABLE OF CONTENTS
Page
§17.1. Governing Differential Equation 17–3
§17.2. Forward Euler Integration 17–3
§17.3. More Accurate Integration 17–4
§17.4. Numerical Stability of Forward Euler 17–5
§17.5. Accuracy Monitoring 17–7
17–2
17–3 §17.2 FORWARD EULER INTEGRATION
Incremental methods calculate the nonlinear response through the numerical integration of a rate
form of the equilibrium equations as the stage control parameter λ is varied. In the nomenclature
introduced in Chapter 16, we can characterize these as predictor-only methods: no corrective
iterations to recover equilibrium are performed. They are also known as step-by-step, initial-
value or marching methods in the engineering literature. The qualifier “purely” distinguishes these
incremental methods from those that make use of the pseudo-force concept, and which are covered
in Chapter 19.
The present Chapter emphasizes purely incremental methods in which the first-order rate equations
are integrated by a forward Euler scheme. Furthermore, for simplicity we focus on the simplest
increment control strategy: load control, in which λ is treated as an independent variable. This
restriction allows subjects such as stability and accuracy to be discussed in a straightforward manner.
An arclength-parametrized version, which allows the introduction of more robust increment control
techniques and the automatic traversal of limit points, is presented in the following Chapter.
r = Ku − q = 0, (17.1)
where primes denote differentiation with respect to λ. If the stiffness matrix is nonsingular, this
equation uniquely relates the differential of u to that of λ:
du
u = = K−1 q = v, (17.2)
dλ
where as usual v denotes the incremental velocity vector. Purely incremental methods with λ as
independent variable are based on the numerical integration of (17.2) to generate an approximate
response u = u(λ) given the initial condition
u = u0 at λ = 0. (17.3)
Remark 17.1. The exact integral of (17.2) with the initial conditions (17.3) is
r(u, λ) = r0 (17.4)
where r0 = r(u0 , 0). Thus an initial equilibrium error does not decay even if the integration were carried out
exactly. This is the source of the drifting error that afflicts purely incremental methods. The error committed
at each step moves the equilibrium point to a neighboring curve in the incremental flow (see Figure 16.2).
Consequently the solution may “drift away” quickly when the incremental flow paths “flare out” from the
equilibrium path.
17–3
Chapter 17: PURELY INCREMENTAL METHODS: LOAD CONTROL 17–4
In the remaining subsections of this Chapter we consider that the incrementation process is controlled
directly by varying the stage parameter λ, which thus assumes the role of independent variable.
This is tantamount to using the λ-control increment discussed in Chapter 16. This restriction is
removed in the next Chapter.
The simplest incrementation scheme is obtained by using the forward Euler integrator
is the stage parameter stepsize. Treating (17.2) with this integrator yields the scheme
To increase accuracy, more refined integration formulas have been proposed. An attractive second-
order choice is the explicit midpoint rule (also called Heun’s rule by some authors):
un+1/2 = un + 12 K−1
n qn λn ,
def def
Kn+1/2 = K(un+1/2 ), qn+1/2 = q(un+1/2 ), (17.8)
un+1 = un + K−1
n+1/2 qn+1/2 λn .
This scheme was used in the author’s thesis1 to treat problems with combined geometric and
material nonlinearities. The midpoint rule has attractive features for flow-plasticity studies, since
local elastic unloading can be detected during the first “trial” step and Kn+1/2 adjusted accordingly.
The same feature can be used to advantage in bifurcation analysis if a stiffness-determinant change
is detected between Kn and Kn+1/2 . But note that the stiffness matrix has to be formed and factored
twice per incremental step.
1 C. A. Felippa, Refined Finite Element Analysis of Linear and Nonlinear Two-dimensional Structures, Ph.D. thesis, Dept.
of Civil Engineering, University of California, Berkeley (1966)
17–4
17–5 §17.4 NUMERICAL STABILITY OF FORWARD EULER
Natural extensions of (17.8) are third and fourth-order Runge-Kutta (RK) formulas, which require
three and four stiffness evaluations and factorizations per step, respectively. These more refined
methods, however, are rarely used in structural mechanics for the amount of work per step is
considerable. Remark 17.8, however, indicates a possible niche for the classical fourth-order RK
in nonconservative problems.
Can the integration process (17.7) become numerically unstable? The subject is rarely mentioned
in the finite element literature. For simplicity we begin with the one-degree-of-freedom counterpart
u = v of u = v. The right-hand side v = K −1 q is Taylor-series expanded in u = u − u n about
u n as
∂v
u = u n + µ u + O(u 2 ) with µ = . (17.9)
∂u
For the linearized stability analysis only the homogeneous part of (17.9) is retained, which yields
the model equation
u = µu. (17.10)
Consider the case in which µ is negative real and h = λ > 0. Then the solution u = u(λ) of
the model equation is exponentially decreasing as λ increases. The forward Euler integration is
absolutely stable2 if
|1 + hµ| ≤ 1, or h ≤ −2/µ. (17.11)
If h exceeds this value, the computed solution exhibits oscillatory instability. If µ is positive real the
solution of the model equation grows exponentially as λ increases and the forward Euler integration
is “relatively stable” for all h > 0.
Now if λ decreases so that h = λ < 0 the roles are reversed (cf. Remark 17.3). The stability
condition is h ≥ −2/µ if µ > 0. If λ is a load parameter, loading and unloading sequences may
be viewed as equally likely; consequently a safe stability constraint is
2
|λ| ≤ . (17.12)
|µ|
For the general system (17.2), let µi (i = 1, 2 . . . N , N being the number of degrees of freedom)
be the eigenvalues of the so-called amplification matrix
∂v ∂(K−1 q)
A= = . (17.13)
∂u ∂u
It is shown in Remark 5.8 that this matrix, although generally unsymmetric, has real eigenvalues if
the problem is conservative, i.e. K is the Hessian of a potential (u, λ) for fixed λ, and K is positive
2 C. W. Gear, Numerical Initial Value Problems in Ordinary Differential Equations, Prentice-Hall, Englewood Cliffs, N.J.
(1971)
L. Lapidus and J. H. Seinfield, Numerical Solution of Ordinary Differential Equations, Academic Press, New York
(1971).
17–5
Chapter 17: PURELY INCREMENTAL METHODS: LOAD CONTROL 17–6
definite. The eigenvalues are given by the eigensystems (5.22) or (5.23). Under such conditions a
safe increment is given by (17.12), where now
If K does not depend on λ (as in the linear case) all µi vanish and the increment is unrestricted. If
K depends midly upon λ, eigensystems (5.21)–(5.22) show that the largest |µi | are associated with
the smallest eigenvalues of K, i.e. the fundamental stiffness modes (see also Remark 17.4).
Of course the actual calculation of all µi at each step would be a formidable computational task.
But the following finite-difference “path” estimate is easily obtained:
vn+1 − vn vn+1 − vn 1 an
µ≈ = = , (17.15)
un+1 − un vn λn λn
where
√ an = vn+1 − vn /vn and denotes the 2-norm or Euclidean norm of a vector: x =
xT x. (an may be viewed as a kind of “incremental acceleration”.) Unfortunately this quantity
is not available until the n th step is completed, and to get a practical estimate we replace it by the
previous step estimate:
an−1 vn − vn−1
µ≈ , an−1 = . (17.16)
λn−1 vn−1
Insertion into (17.12) yields the stability condition
2 |λn−1 |
|λn | ≤ . (17.17)
an−1
If A can have complex eigenvalues, however, this simple rule does not apply (see Remark 17.8).
Remark 17.3. The fact that h = λ can be either positive or negative is a distinguishing feature of incremental
static analysis. In the numerical integration of actual dynamical systems, the time increment h = τ is never
negative; therefore stability results from conventional ODE theory should be used with caution.
Remark 17.4. Having the largest µi associated with the smallest eigenvalues of K represents another note-
worthy difference with dynamic problems. In the latter, the stability limit of explicit integrators such as (17.7)
is determined by the highest frequencies of the system. As discussed in Chapters dealing with dynamic relax-
ation methods, the difference is due to the fact that the matrix multiplying the highest derivative is the mass
(stiffness) in a dynamic (static) problem.
Remark 17.5. If the midpoint rule (17.8) is used, the stability limit remains the same for real µi .
Remark 17.6. Replacing vn−1 by vn in (17.16) is inconsequential, as it is only a gross estimate. In fact,
choosing the smallest of the two norms would be the more conservative policy. A more serious objection is
the choice of the 2-norm unless the problem is well scaled and all degrees of freedoms have common physical
dimension. Otherwise a diagonal scaling matrix may have to be introduced as discussed in Chapter 4; for
example
|(vn − vn+1 )T S2 (vn − vn+1 )|
2
an−1 = . (17.18)
min(|vnT S2 vn |, |vn−1
T
S2 vn−1 |)
17–6
17–7 §17.5 ACCURACY MONITORING
Remark 17.7. As K approaches singularity, v → ∞ and the stable λ approaches zero. This is an
indication of the problems encountered by this type of incremental method at critical points. “Flat” limit
points can be traversed using the step-controlled parametric formulation discussed in the next Chapter. At
those points λ changes sign. But as the limit point becomes progressively sharper, numerical difficulties
increase. Bifurcation points, which in some sense may be viewed as infinitely sharp limit points, are difficult
to traverse without resort to either perturbation or special techniques that necessarily involve buckling mode
estimation, as discussed later.
2 |λn−1 |
|λn | ≤ . (17.23)
an−1
3 P. Henrici, Error Propagation for Difference Methods, Wiley, New York (1963).
17–7
Chapter 17: PURELY INCREMENTAL METHODS: LOAD CONTROL 17–8
This increment size must also be subjected to other bounds provided by rules discussed later. Similar
accuracy monitoring techniques can be devised for more complicated integration schemes such as
the midpoint rule.
Remark 17.9. Comparing (17.23) with (17.17) furnishes a simple rule: choosing < 1 takes care of stability
if the roots of A are real. In any case the similarity between the stability and accuracy control rules is striking.
Remark 17.10. For poorly scaled problems the use of a weighted norm, as in (17.18), is advisable.
Remark 17.11. Given bounds on e at each step, bounds on the accumulated drift error can be obtained but
they are usually so pessimistic as to be useless, unless some detailed problem information is available; see
e.g. Gear (loc.cit. in footnote 2). The only reliable way to assess global accuracy is to rerun the problem with
several values of , for example = 0.2, 0.1 and 0.05.
Remark 17.12. This truncation error monitoring technique works in static nonlinear analysis because e
is controlled by physically relevant low-frequency modes. It fails on direct time integration of dynamical
problems — see e.g. Park4 — because e is then controlled by physically irrelevant high-frequency modes. In
numerical analysis parlance, problems in structural dynamics are said to be stiff.
4 K. C. Park and C. A. Felippa, Direct Time Integration Methods in Nonlinear Structural Dynamics, Comp. Meth. Appl.
Mech. Engrg., 17/18, pp. 277–313 (1979)
K. C. Park, Time Integration of Structural Dynamics: A Survey, Ch. 4.2 in Pressure Vessels and Piping Design Technology
— A Decade of Progress ASME, New York (1982).
17–8
16
.
Overview of
Solution Methods
16–1
Chapter 16: OVERVIEW OF SOLUTION METHODS 16–2
TABLE OF CONTENTS
Page
§16.1. Introduction 16–3
§16.1.1. Stages, Increments and Iterations . . . . . . . . . . . 16–3
§16.1.2. Why Incrementation? . . . . . . . . . . . . . . 16–4
§16.2. Advancing the Solution: Increment Control 16–5
§16.3. Advancing the Solution: Prediction 16–5
§16.4. Advancing the Solution: Correction 16–7
§16.5. Traversing Equilibrium Path in Positive Sense 16–7
§16.5.1. Positive External Work . . . . . . . . . . . . . . 16–8
§16.5.2. Angle Criterion . . . . . . . . . . . . . . . . 16–8
§16.6. Constraint Strategy 16–9
§16.6.1. λ Control . . . . . . . . . . . . . . . . . . . 16–9
§16.6.2. State Control . . . . . . . . . . . . . . . . . 16–9
§16.6.3. Arclength Control . . . . . . . . . . . . . . . . 16–11
§16.6.4. (Global) Hyperelliptic Control . . . . . . . . . . . 16–11
§16.6.5. Local Hyperelliptic Control . . . . . . . . . . . . . 16–12
§16.7. Practical Solution Requirements 16–12
§16.7.1. Tracing the Response . . . . . . . . . . . . . . 16–12
§16.7.2. Finding a Nonlinear Solution . . . . . . . . . . . . 16–13
§16.7.3. Stability Assessment . . . . . . . . . . . . . . 16–13
§16.7.4. Post-buckling and Snap-through . . . . . . . . . . . 16–13
§16.7.5. Multiple Load Parameters . . . . . . . . . . . . . 16–14
§16. Exercises . . . . . . . . . . . . . . . . . . . . . . 16–15
16–2
16–3 §16.1 INTRODUCTION
In previous Chapters we have covered the governing equations of geometrically nonlinear structural
analysis and the discretization of those equations by finite element methods. The result is a set of
parametrized nonlinear algebraic equations called residual force equations.
The solution of these equations as the control parameters are varied varied provides the equilibrium
response of the structure. In this Chapter we begin the coverage of solution methods suitable for
digital computation.
§16.1. Introduction
It was noted in Chapter 1 that all solution procedures of practical importance are strongly rooted in
the idea of “advancing the solution” by continuation. The basic idea is to follow the equilibrium
response of the structure as the control and state parameters vary by small amounts. The motivation
in terms of circumventing the “solution morass” is described in that Chapter.
This overarching framework gives rise to many variants called solution schemes. A common feature
is that continuation is a multilevel process, as illustrated in Figure 16.1. The process involves
a hierarchical breakdown into stages, incremental steps, and iterative steps. The middle level:
incrementation, is always present. Staging may be missing if there is only one control parameter.
Iteration may be missing if there if no correction process.
In the present Chapter multilevel continuation is described in general terms, with the goal of
maintaining independence from specific solution schemes. The final subsections describe how the
general procedure is adapted to the analysis of problems encountered in engineering practice.
16–3
Chapter 16: OVERVIEW OF SOLUTION METHODS 16–4
Stages
Increments
Iterations
the corrective process. This superscript is enclosed in parentheses if there is potential confusion
with exponents. Iterative changes in u and λ are often shortened to d and η, respectively.
Remark 16.1. Solutions accepted after each increment following completion of the corrective process, are
often of interest to users because they represent approximations to equilibrium states. They are therefore saved
as they are computed. On the other hand, intermediate results of iterative processes are rarely of interest unless
one is studying the “insides” of solution processes. Hence most production programs discard them.
Remark 16.2. The terminology of nonlinear static analysis is far from standardized. Despite their practical
importance, few authors recognize the existence of stages. Many use the term step to mean incremental step
whereas the terms substep, subincrement and cycle are used for the iteration level. There is more uniformity
in dynamic analysis, possibly because there is only one advancing level: step is universally used to denote the
change over a time increment.
The use of increments may seem at first sight unnecessary if one is interested primarily in the final
solution. But breaking up a stage into increments may serve other purposes:
Helping convergence. Success in a corrective process done by a Newton-like method may hinge on
having a good initial guess supplied by the predictor, since such methods are notoriously finicky.
The quality of this guess can be improved by reducing the increment.
Sidestepping extraneous roots. Incrementation helps the solution procedure from falling into the
“root morass” discussed in Chapter 1.
Gaining insight into structural behavior. As noted in Remark 16.1, programs often save converged
solutions after each increment and for a good reason: a response plot can teach the engineer more
about the structural behavior than simply knowing the final solution.
Avoiding surprises. Critical points may occur before the stage ends. There are problems in which
such points, notably bifurcation, may be masked if coarse increments are taken.
16–4
16–5 §16.3 ADVANCING THE SOLUTION: PREDICTION
Alleviating path dependence. Although the focus of this course is on path-independent problems,
it should be noted that the presence of path-dependent effects severely restricts increment sizes
because of history-tracing constraints. For example, in plasticity analysis stress states must not be
allowed to stray too far outside the yield surface.
A nonlinear analysis program is “marching” along a stage. Assume that n incremental steps have
been completed. The last accepted solution is un , which corresponds to λn . Performing the (n+1)th
step entails the calculation of the increments
that satisfy the residual equilibrium equations r(u, λ) = 0 to requested accuracy. As stated the task
is not fully defined because there are less equations than unknowns, which makes the increment
sizes indeterminate. The problem is closed by adopting an increment control strategy. The strategy
may be expressed in general form as a constraint condition:
aT u̇ + g λ̇ = 0, (16.3)
where
∂c ∂c
aT = , g= . (16.4)
∂u ∂λ
Remark 16.3. The addition of the constraint equation serves two purposes: it makes the algebraic problem
determinate, and it can be used to control the increment size directly or indirectly to enhance robustness and
convergence.
Remark 16.4. Note that the constraint c = 0 is expressed in terms of the increments {un , λn } from the last
solution and not in terms of the total values. This localization condition is essential to maintaining invariance
with respect to the origin chosen for u and λ. Once a step is finished, the constraint is reset for the next one
by moving its origin.
Remark 16.5. Specific choices for (16.4) are discussed in §164 below but for some developments it is possible
to keep c arbitrary. Furthermore, it is also possible to specify the constraint directly in the rate form (16.3)
without an explicit integral. An example is Fried’s orthogonal trajectory accession method.1
1 I. Fried, Orthogonal Trajectory Accession to the Nonlinear Equilibrium Curve, Comp. Meth. Appl. Mech. Engrg., 47,
283–297, (1984).
16–5
Chapter 16: OVERVIEW OF SOLUTION METHODS 16–6
Having decided upon an increment control strategy, to start up the (n+1)th incremental step, an
initial approximation
u0n , λ0n , (16.5)
to the increments (16.1) is calculated by a prediction step. These values are called the predicted
increments and the formula used is called a predictor or extrapolator.
Most predictors are based on the first-order path equation derived in Chapter 4 and repeated here
for convenience:
ṙ = 0, or Ku̇ = q λ̇, (16.6)
Assuming K to be nonsingular, the forward Euler method furnishes the simplest predictor:
u0n = K−1
n qn λn = vn λn ,
0 0
(16.7)
in which v is the incremental velocity vector defined in Chapter 4. The process is completed by
selecting an increment control strategy through the constraint (16.2). Two examples follow.
u0n = vn
n , λn =
n . (16.9)
This formula obviously fails when Kn is singular, i.e. at critical points, because there vn becomes either infinite
(at limit points) or nonunique (at bifurcation points). This suggests that the solution process will break down
at those points.
Example 16.2. For the arclength strategy in which the absolute value of the distance (4.23) is specified to be
n f n
n
n vn
n
λ0n = = =± , u0n = ± . (16.11)
±(vn vn +
T 1) ± vnT vn + 1 fn fn
In this case two signs for the increment are obtained. The proper one is obtained by applying one of the “path
advancing” criteria discussed below.
Note also that (16.10) does not fail at isolated limit points if one properly passes to the limit v/|v| → z, as per
Remark 4.2. This limit process yields
λ0n = 0, u0n = ±
n z (16.12)
The normalized v near the limit point serves as a good approximation for z. It should be noted, however, that
the formula fails at multiple limit points and at bifurcation points; thus the arclength strategy is no panacea.
16–6
16–7 §16.5 TRAVERSING EQUILIBRIUM PATH IN POSITIVE SENSE
Computed solutions
λ
Drift error
v
Figure 16.2. Drift error in purely incremental solution procedure.
2 M. A. Crisfield, An Incremental-Iterative Algorithm that Handles Snap-Through, Computer & Structures, 16, 55–62
(1981)
M. A. Crisfield, An Arc-Length Method Including Line Searches and Accelerations, Int. J. Num. Meth. Engrg., 19,
1269–1289 (1983).
16–7
Chapter 16: OVERVIEW OF SOLUTION METHODS 16–8
qT v = 0, (16.17)
because then the condition (16.16) is vacuous. This happens in the following cases.
Bifurcation points. As a bifurcation point B is approached, v/|v| → z, achieving equality at B.
Since qT z = 0, it follows that (16.16) fails at B.
Incremental velocity reversal. If the structure becomes “infinitely stiff” at a point in the equilibrium
path v vanishes. This case is rarer than the previous one, but may arise in the vicinity of turning
points.
Bifurcation points demand special treatment and cannot be easily passed through simple predictor
methods. One way out is to insert artificial purtuebations that transformperturbations are inserted.
However, the case v → 0 can be overcome by a modification of the previous rule.
16–8
16–9 §16.6 CONSTRAINT STRATEGY
To get over this difficulty a condition on the angle of the prediction vector is more effective. Let
tn−1 be the tangent at the previous solution. Then chose the positive sense so that
Once the “ping-pong” region is crossed, the work criterion should be reversed so the external work
is negative.
Remark 16.7. Other geometric criteria are given by Crisfield (loc. cit. in footnote 2) and Skeie and Felippa4
So far the form of the constraint equation, (16.2) or (16.3), has been left arbitrary. In the sequel we
list, roughly in order of ascending complexity, instances that are either important in the applications
or have historical interest. In what follows
is always a dimensionless scalar that characterizes the
size of the increment. Six constraints are pictured in Figure 16.3. In this figure, c is the constraint
curve, S is the last solution point, P the predicted point and C the converged solution.
§16.6.1. λ Control
At each step λn =
n , where
is a dimensionless scalar. The constraint equation is (16.8) listed
in Example 16.1. This is generally called λ-control. Often the parameter λ is associated with a
loading amplitude, in which case this is called load control. The physical analogy would be a test
machine in which the operator increases the load to specific values.
The differential form (16.3) has
a = 0, g = 1. (16.19)
As noted in Example 16.1, this constraint form fails as critical points are approached.
where u is a reference value with dimensions of displacement, which is introduce for scaling
purposes. An alternative way of doing that consists of using the scaled increment of §4.6:
unT
un −
2n = 0. (16.21)
aT = 2un , g = 0. (16.22)
4 G. Skeie and C. A. Felippa, A Local Hyperelliptic Constraint for Nonlinear Analysis, Proceedings of NUMETA’90
Conference, Swansea, Wales, Elsevier Sci. Pubs, 1990.
16–9
Chapter 16: OVERVIEW OF SOLUTION METHODS 16–10
c=0
c=0 P P P
C C C
S S S
c=0
P P P
C C C
S S S
c=0
c=0
c=0
Figure 16.3. Geometric representation of constraint equations for a one-dof problem, with
state u and control parameter λ plotted horizontally and vertically, respectively.
(a) load control, (b) state control, (c) arclength control, (d) hyperspherical control,
(e) global hyperelliptical control, and (f) local hyperelliptic control.
Remark 16.8. In the finite element literature the term displacement control has been traditionally associated
with the case in which the magnitude of only one of the components of u, say u i , is specified, which is
tantamount to choosing a special infinity norm of u. This old technique was used in the mid-1960s by Argyris
and Felippa.5 There is a generalization of single displacement control in which several reference displacements
are used. This multiple dimensional hyperplane control has been investigated by Powell, Bergan and others.6
5 J. H. Argyris, Continua and Discontinua, in Proceedings Conference on Matrix Methods in Structural Engineering,
AFFDL-TR-66-80, Wright-Patterson AFB, Dayton, Ohio, 11–189 (1966).
C. A. Felippa, Refined Finite Element Analysis of Linear and Nonlinear Two-dimensional Structures, Ph.D. thesis, Dept.
of Civil Engrg, University of California, Berkeley (1966).
6 G. H. Powell and J. Simons, Improved Iteration Strategy for Nonlinear Structures, Int. J. Num. Meth. Engrg., 17, 1655–
1667 (1981)
P. G. Bergan, G. Horrigmoe, B. Krakeland and T. H. Søreide, Solution Techniques for Nonlinear Finite Element Problems,
Int. J. Num. Meth. Engrg., 12, 1677–1696 (1978)
P. G. Bergan, Solution Algorithms for Nonlinear Structural Problems, Computers & Structures, 12 497–509 (1980)
P. G. Bergan and J. Simons, Hyperplane Displacement Control Methods in Nonlinear Analysis, in Innovative Methods
for Nonlinear Problems, ed. by W. K. Liu, T. Belytschko and K. C. Park, Pineridge Press, Swansea, U.K., 345–364
(1984)
16–10
16–11 §16.6 CONSTRAINT STRATEGY
a T = vn / f n , g = 1/ f n . (16.24)
Remark 16.9. The “orthogonal trajectory” constraint discussed by Fried (see footnote 1) may be regarded as
a generalization of the arclength constraint in which a traversal orthogonality condition is applied throughout
the corrective phase. This differential constraint is interesting in that it does not fit the form (16.2) and may
in fact be followed independently of the the predictor and past solution. But following the trajectory depends
on v = K−1 q being frequently updated and is practical only with a true Newton corrector.
There is a wide family of constraints that combine the magnitude of λn and a norm of un . A
frequently used combination is the hyperelliptic constraint
an2
unT
un + bn2 (λn )2 =
2n , (16.27)
7 E. Riks, The Application of Newton’s Method to the Problem of Elastic Stability, Trans. ASME, J. Appl. Mech., 39,
1060–1065 (1972)
G. A. Wempner, Discrete Approximations Related to Nonlinear Theories of Solids, Int. J. Solids Structures, 7, 1581–1599
(1971).
16–11
Chapter 16: OVERVIEW OF SOLUTION METHODS 16–12
by Padovan and Park.8 The expression was rendered dimensionless by Felippa9 who introduced
scaling parameters and and discussed appropriate choices. If a = b = 1 in the unscaled form
(16.27) we recover the hyperspherical constraint proposed (but not used) by Crisfield (loc. cit. in
footnote 4).
The constraint gradients are
a = 2a 2 u, g = 2b2 λ. (16.28)
This is a variation of the previous one in which we take a combination of λ and a norm of u,
where λ and u are to be determined according to a local coordinate system at S(un , λn ):
c(u, λ) = a 2 (u − un )T S(u − un ) + b2 (λ − λn )2 −
2n = 0, (16.29)
Scaling of this constraint to achieve consistency is discussed by Skeie and Felippa (work cited in
footnote 4), where additional computational details may be found. It turns out that this constraint
can include all ones previously discussed as special regular or limit cases.
10
Remark 16.10. Another interesting strategy: the work constraint of Bathe and Dvorkin limits the total
external work spent during the corrective phase.
Remark 16.11. In path-independent problems that involve only geometric or conservative boundary-condition
nonlinearities, it is generally best to maximize step lengths subject to stability and equilibrium accuracy
constraints. Stability depends on the curvature of the response path, presence of critical points, and solution
method used. Equilibrium accuracy depends chiefly on whether a corrective process is applied.
8 J. Padovan and S. Tovichakchaikul, Self-Adaptive Predictor-Corrector Algorithm for Static Nonlinear Structural Analysis,
Computers & Structures, 15, 365–377 (1982).
K. C. Park, A Family of Solution Algorithms for Nonlinear Structural Analysis Based on the Relaxation Equations, Int.
J. Num. Meth. Engrg., 18, 1637–1647 (1982).
9 C. A. Felippa, Dynamic Relaxation under General Increment Control, in Innovative Methods for Nonlinear Problems,
ed. by W. K. Liu, T. Belytschko and K. C. Park, Pineridge Press, Swansea, U.K., 103–163 (1984).
10 K. J. Bathe and E. Dvorkin, On the Automatic Solution of Nonlinear Finite Element Equations, Computers & Structures,
17, 871–879 (1983).
16–12
16–13 §16.7 PRACTICAL SOLUTION REQUIREMENTS
The remaining subsections describe various types of nonlinear structural analyses encountered in
engineering practice, and the requirements they pose on solution procedures.
“Tracing the response” is of interest for many nonlinear problems. For a typical stage, perform a
sequence of incremental steps to find equilibrium states
un , λn , n = 1, 2, . . .
in sufficient number to ascertain the response u = u(λ) of the structure within engineering require-
ments.
If the control parameter is associated with a fundamental load system, the response path is known
as the fundamental equilibrium path, as it pertains to the service range in which the structure is
supposed to operate.
One class of problems that fit this requirement is that in which structural deflections, rather than
strength, are of primary importance in the design. For example, some large flexible space structures
must meet rigorous “dimensional stability” tolerances while in service.
A variant of the foregoing occurs if the primary objective of the analysis is to find a solution u
corresponding to a given λ (for example, λ = 1), whereas tracing of the response path is in itself
of little interest.
Very flexible structures that must operate in the nonlinear regime during service fit this problem
class. The example of the suspension bridge under its own weight, discussed in §3.4, provides
a good illustration. The undeflected “base” configuration u = 0 is of little interest as it has no
physical reality and the bridge never assumes it. It is merely a reference point for measuring
deflections.
Under such circumstances, the chief consideration is that the accuracy with which the response path
is traced is of little concern. Getting the final answer is the important thing. Once this reference
configuration is obtained, “excursions” due to live loads, temperature variations, wind effects and
the like may be the subject of further analysis staging.
This is perhaps the most important application of nonlinear static analysis. The analyst is concerned
with the value (or values) of λ closest to 0 at which the structure behavior is not uniquely determined
by λ. These are the critical points discussed in Chapter 5. In physical terms, the system becomes
uncontrollable and may “take off” dynamically.
Problem of this nature arise in stability design. The determination of limit points is called collapse
or snapping analysis. The determination of bifurcation points is called buckling analysis.
16–13
Chapter 16: OVERVIEW OF SOLUTION METHODS 16–14
16–14
16–15 Exercises
EXERCISE 16.1
[C:20] Consider the following residual equilibrium equation:
r (ψ, λ) = sec(α−ψ) sec α (2+λ sin ψ) − 2 sec(α−ψ) tan(α−ψ) − λ cos ψ sec α = 0, (E16.1)
in which α is a problem parameter, λ the control parameter, and ψ the only degree of freedom. This r comes
from the 2-bar arch problem already studied in Exercise 6.2. Here α is the initial arch rise angle whereas
ψ = α − θ is the angle change from the reference state, at which ψ0 = 0 and λ0 = 0. A plot of the exact
λ(ψ) for ψ = [0, 60◦ ] = [0, π/3] is shown in Figure E16.1(a); the fundamental path ends at limit point L.
λ
(a) λ (b) 0.08 B
0.06 Exact response for Exact response for
L 0.06
Exercise 16.1 Exercise 16.2
0.04 0.04
0.02 ψ (rad)
ψ (rad) 0.02
−0.04 −0.04
−0.06
−0.06
(a) Derive the first-order rate form K ψ̇ = q λ̇ by taking t ≡ λ, and convert to u̇ = du/dλ = v. (Recall that
K = ∂r/∂ψ, q = −∂r/∂λ, and v = K −1 q.)
(b) Integrate numerically the rate equation u̇ = v found in (a) by the purely incremental, forward-Euler
method with load control over λ = [0, 0.1]. Start from λ0 = 0 and ψ0 = 0. (All angles should be in
radians.) Use α = 30◦ = π/6 as arch rise angle and take 10 load increments of
n = 0.01 (same for all
steps). Are you able to detect and traverse the limit point L?
(c) Repeat the run twice, each time cutting
n by 1/4 and quadrupling the number of steps. Is limit point
detection and traversal improved?
Hints. If using Mathematica the 10-step forward Euler script followed by the response plot could be imple-
mented as
in which r̂ is the residual (E16.1). The fundamental path now ends at a bifurcation point B, as pictured in
Figure E16.1(b). Use α = 30◦ and same solution method. Are you able to detect the bifurcation point?
16–15
15
.
The CR Description:
C1 Plane Beam
15–1
Chapter 15: THE CR DESCRIPTION: C1 PLANE BEAM 15–2
TABLE OF CONTENTS
Page
§15.1. Introduction 15–3
§15.2. CR Beam Kinematics 15–3
§15.2.1. Coordinate Systems . . . . . . . . . . . . . . . 15–3
§15.2.2. Degrees of Freedom . . . . . . . . . . . . . . . 15–3
§15.2.3. Partial Derivatives . . . . . . . . . . . . . . . . 15–5
§15.2.4. Arbitrary Initial Configuration . . . . . . . . . . . 15–7
§15.2.5. Stress Resultants . . . . . . . . . . . . . . . . 15–8
§15.3. The Deformational Strain Energy 15–8
§15.4. Internal Force Vector and Tangent Stiffness Matrix 15–9
§15.4.1. Internal Force Vector . . . . . . . . . . . . . . 15–9
§15.4.2. Material Stiffness Matrix . . . . . . . . . . . . . 15–9
§15.4.3. Geometric Stiffness Matrix . . . . . . . . . . . . 15–10
§15. Exercises . . . . . . . . . . . . . . . . . . . . . . 15–11
15–2
15–3 §15.2 CR BEAM KINEMATICS
§15.1. Introduction
In this Chapter we use the CR description to construct a geometrically nonlinear, 2-node Bernoulli-
Euler plane beam. Unlike Chapter 9 we will do a C 1 (Hermitian) beam from the start, since with CR
it is as easy to do C 1 or C 0 , and the former has a much better geometric stiffness matrix.
The CR formulation of the beam motion is quite similar to that of the bar element in many respects,
and much of the development can be reused. Only the major differences will be noted here.
As in Chapter 9, we consider a plane, straight, prismatic beam element with two nodes. The element
is initially aligned with the global X axis in the initial configuration C0 , with the origin O0 located
at the element midpoint. This configuration is assumed to be straight and undeformed although it
may be under initial uniform axial stress with resultant N 0 . The bar properties include the elastic
modules, E, the cross section area A0 and the moment of inertia I0 about the neutral axis. The length
in C0 is L 0 .
The motion on the {X, Y } plane carries it to the current configuration C. The corotated configuration
C R is selected as depicted in Figure 15.1:
1. The longitudinal axis passes through the current position of the end nodes. This defines the
local axis x e . The origin of {x e , y e } is placed halfway between the nodes. This forms an angle
ψ with X .
2. The C R nodes are placed at an equal distance from the C nodes. Hence the corotated axes
{x Re , y Re }, including origin, coincide with {x e , y e }.
The new ingredient is the rotation angle θ about Z or z. With C R chosen as indicated, the deformation
part of these rotations is easily extracted: θ̄ = θ − ψ.
Other possibilities for selecting C R are possible. The foregoing choice has the advantage of being
compatible with that of the bar element discussed in the previous Chapter.
The beam element has six degrees of freedom, which are placed in the vectors
1
u X1 ū eX 1 −2d
uY 1 u eY 1 0
θ θ̄ e θ̄
1 1 1
u= , ūe = e = 1 . (15.1)
u X2 ū X 2 2 d
e
uY 2 ū Y 2 0
θ2 θ̄2e θ̄2
See Figure 15.2 for a picture of the global displacements and Figure 15.3 for the deformation dis-
placements.
15–3
Chapter 15: THE CR DESCRIPTION: C1 PLANE BEAM 15–4
_
// Y
φ = ψ+ϕ_
θ2 ψ
// X
ϕ
2(x2 ,y2) // X
Current C
_
// Y
uY2
θ1
Corotated CR_
1(x1 ,y1) X
Base or initial configuration ϕ
C0 uY1
2(X2, Y2 )
Y, y _ uX2
Yγ
X, x
uX1
1(X1, Y1)
θ2
2 //X, x
C
θ1
//X, x u Y2
u X1
1
uY1
C0 uX2
10 20
ye
_
−d/2 θ1 d/2 _
CR θ2
1 xe
2
L0 C
L
15–4
15–5 §15.2 CR BEAM KINEMATICS
Proceeding as in the general formulation specialized to the 2D case, we can obtain the following
relation:
ū e 1 L (1 − c )
x1 cψ sψ 0 0 0 0 u X1 − u X0 2 0 ψ
ū y1 −sψ cψ 0
e
0 0
uY 1 − uY 0
1
0 L s
2 0 ψ
θ̄ e 0
1 0 1 0 0 0 θ −ψ
ū = e =
e
1
+1 (15.2)
ū x2 0 0 0 cψ sψ 0 u X 2 − u X 0 2 L 0 (cψ − 1)
e
ū y2 0 0 0 −sψ cψ 0 uY 2 − uY 0 − 2 L 0 sψ
1
θ̄ e 0 0 0 0 0 1 θ2 −ψ
2
Here cψ and sψ and the angle ψ are implicitly defined by the displacements through the trigonometric
relations
Ly Lx Ly
sψ = sin ψ = , cψ = cos ψ = , ψ = arctan (15.3)
L L Lx
where L x = L 0 + u X 2 − u X 1 , L y = u Y 2 − u Y 1 , and
L = L 2x + L 2y (15.4)
is the bar length in the current configuration, ignoring the bending deformation.
We note the following relations
∂L ∂L ∂L ∂L ∂L ∂L
=− = cψ , =− = sψ , = = 0,
∂u X 2 ∂u X 1 ∂u Y 2 ∂u Y 1 ∂θ1 ∂θ2
∂cψ ∂cψ sψ2 ∂cψ ∂cψ sψ cψ ∂cψ ∂cψ
=− = , =− =− , = = 0,
∂u X 2 ∂u X 1 L ∂u Y 2 ∂u Y 1 L ∂θ1 ∂θ2
(15.5)
∂sψ ∂sψ sψ cψ ∂sψ ∂sψ cψ2 ∂sψ ∂sψ
=− =− , =− = , = = 0,
∂u X 2 ∂u X 1 L ∂u Y 2 ∂u Y 1 L ∂θ1 ∂θ2
∂ψ ∂ψ sψ ∂ψ ∂ψ cψ ∂ψ ∂ψ
=− = , =− =− , = = 0.
∂u X 2 ∂u X 1 L ∂u Y 2 ∂u Y 1 L ∂θ1 ∂θ2
which are useful in the calculations that follow.
The first and second partial derivatives of the deformations d, θ̄1 and θ̄2 with respect to the node
displacements are necessary for the computations of internal forces and stiffness matrices.
Using (15.5) and Mathematica, one obtains for the first derivatives:
δ ū e
x1
1
c
2 ψ
1
s
2 ψ
0 − 12 cψ − 12 sψ 0 δu X 1
δ ū y1
e −sψ cψ L 0 /L cψ2 L 0 /L 0 sψ cψ L 0 /L −cψ2 L 0 /L 0
δu Y 1
δ θ̄ e −s /L cψ /L sψ /L −cψ /L 0 δθ1
1 ψ 1
e = (15.6)
δ ū x2 − 12 cψ − 12 sψ 0 1
c
2 ψ
1
s
2 ψ
0 δu
e X2
δ ū y2 sψ cψ L 0 /L −cψ2 L 0 /L 0 −sψ cψ L 0 /L cψ L 0 /L
2
0 δu Y 2
δ θ̄2e −sψ /L cψ /L 0 sψ /L −cψ /L 1 δθ2
15–5
Chapter 15: THE CR DESCRIPTION: C1 PLANE BEAM 15–6
_
// Y
φ = ψ+ϕ _
ψ
θ2 // X
ϕ
2(x2 ,y2 ) // X
_
// Y
C
uY 2
θ1
1(x1 ,y1) _
X
ϕ
2(X2, Y2 )
uY 1
Y, y _ uX 2
Y C0
X, x
uX1
1(X1, Y1 )
δu X 1
δu Y 1
δd −cψ −sψ 0 cψ sψ 0 δθ
δ θ̄1 = −sψ /L 1
cψ /L 1 sψ /L −cψ /L 0 (15.7)
δu X 2
δ θ̄2 −sψ /L cψ /L 0 sψ /L −cψ /L 1
δu Y 2
δθ2
The second derivatives of deformation variables are
sψ2 −sψ cψ 0 −sψ2 sψ cψ 0
−sψ cψ cψ2 0 sψ cψ −cψ2 0
∂ 2d 1 0 0 0 0 0 0
=
(15.8)
∂u ∂u L −s 2
ψ sψ cψ 0 sψ2 −sψ cψ 0
s c −cψ2 0 −sψ cψ cψ2 0
ψ ψ
0 0 0 0 0 0
−2sψ cψ cψ2 − sψ2 0 2sψ cψ sψ2 − cψ2 0
c2 − s 2 2sψ cψ 0 sψ2 − cψ2 −2sψ cψ 0
ψ ψ
∂ 2 θ̄1 1
0 0 0 0 0 0
= 2 (15.9)
∂u ∂u L 2sψ cψ sψ − cψ2
2
0 −2sψ cψ cψ − sψ2
2
0
s 2 − c2 −2sψ cψ 0 cψ2 − sψ2 2sψ cψ 0
ψ ψ
0 0 0 0 0 0
15–6
15–7 §15.2 CR BEAM KINEMATICS
−2sψ cψ cψ2 − sψ2 0 2sψ cψ sψ2 − cψ2 0
c2 − s 2 2sψ cψ 0 sψ2 − cψ2 −2sψ cψ 0
ψ ψ
∂ 2 θ̄2 1 0 0 0 0 0 0
= 2 (15.10)
∂u ∂u L 2sψ cψ sψ − cψ2
2
0 −2sψ cψ cψ − sψ2
2
0
s 2 − c2 −2sψ cψ 0 cψ2 − sψ2 2sψ cψ 0
ψ ψ
0 0 0 0 0 0
15–7
Chapter 15: THE CR DESCRIPTION: C1 PLANE BEAM 15–8
N
M2
V
M1
V
M20
N N0
V0
C0
V0
N0
M10
Figure 15.5. Beam stress resultants depicting positive sign conventions. Axial forces N
and transverse shear forces V are constant along the length, but the bending
moments M vary linearly. Hence two nodal values of M are required.
−2sφ cφ cφ2 − sφ2 0 2sφ cφ sφ2 − cφ2 0
c2 − s 2 2sφ cφ 0 sφ2 − cφ2 −2sφ cφ 0
φ φ
∂ 2 θ̄2 1 0 0 0 0 0 0
= 2 (15.15)
∂u ∂u L 2sφ cφ sφ − cφ2
2
0 −2sφ cφ cφ − sφ2
2
0
s 2 − c2 −2sφ cφ 0 cφ2 − sφ2 2sφ cφ 0
φ φ
0 0 0 0 0 0
The next step in the CR formulation is to work out the deformational strain energy of the beam. The
basic choices are:
15–8
15–9 §15.4 INTERNAL FORCE VECTOR AND TANGENT STIFFNESS MATRIX
1. A linear beam
2. A nonlinear TL beam
The strain energy of the beam for small strains can be written
U = Ua + Ub + Ug (15.17)
where U a , U b and U g are the energy taken by axial (bar) deformation, bending deformation, and
initial-stress geometric effects, respectively. We adopt the following energy expressions:
U a = N 0 d + 12 (N − N 0 )d 2 = N 0 L 0 e + 12 E A0 L 0 e2 ,
T
1 θ̄1 E I0 4 2 θ̄1
U = M2 θ̄2 − M1 θ̄1 + 2
b 0 0
,
θ̄2 L0 2 4 θ̄ 2 (15.18)
T 0
1 θ̄1 N L 0 4 −1 θ̄1
U =2
g
.
θ̄2 30 −1 4 θ̄2
The 2 × 2 matrices appearing in U b and U g may be derived from those given in Chapters 5 and 15,
respectively, of Przemieniecki’s book.1 This book, howevr, omits the initial stress terms.
§15.4. Internal Force Vector and Tangent Stiffness Matrix
The internal force vector and tangent stiffness matrix of the corrotational element are then obtained
by the usual formulas:
∂U ∂p
p= , K= = K M + KG (15.19)
∂u ∂u
To develop these quantities it is necessary to find the first and second partial derivatives of d, θ̄1 and
θ̄2 in terms of the node displacements.
Using the partial derivatives compiled above and Mathematica, one obtains the following expression
for the internal forces.
p = pa + pb + pg (15.20)
where
∂U a
pa = = N [ −cφ −sφ 0 cφ sφ 0 ]T
∂u
∂U b
pb = = [ V sφ −V cφ −M1 −V sφ V cφ M2 ]T
∂u (15.21)
∂U g N 0 L0
pg = = [ −3sφ (θ̄1 + θ̄2 )/L 3cφ (θ̄1 + θ̄2 )/L 4θ̄1 − θ̄2
∂u 30
3sφ (θ̄1 + θ̄2 )/L −3cφ (θ̄1 + θ̄2 )/L 4θ̄2 − θ̄1 ]T
15–9
Chapter 15: THE CR DESCRIPTION: C1 PLANE BEAM 15–10
15–10
15–11 Exercises
EXERCISE 15.1 Complete the derivation of p for the 2-node C 1 beam element and implement in Mathematica,
using the same inputs as in Chapter 9 Addendum. (Implemented and posted on Web)
EXERCISE 15.2 Complete the derivation of K for the 2-node C 1 beam element and implement in Mathematica,
using the same inputs as in Chapter 9 Addendum. (Implemented and posted on Web)
EXERCISE 15.3 A plane 2-node C 1 beam element has properties L 0 = 6, E = 3000, A0 = 2, I0 = 12, N 0 = 5
◦
in the initial state C0 along X , with node 1 at (0,0) and node 2 at (L 0 , 0). The beam rotates by 45
√ about the origin
√
so that at the current configuration C node 1 stays at {0, 0} while node 2 moves to {(L 0 + d)/ 2, (L 0 + d)/ 2},
where d = L 0 /1000. The rotational freedoms at C are θ1 = θ2 = 45◦ = π/4 radians. Compute p, K M and
KG at the current configuration, and compare those quantities with those of the C 0 beam element presented in
Chapter 9, using RBF for the latter.
Note: A Mathematica implementation of this C 1 element has been posted on the Web as a Mathematica 4.1
Notebook PlaneBeamC1.nb. The element checks out when moving about the reference configuration C0 . It
gives excellent buckling values for the problem of Exercise 9.3. More tests are needed, however, for an arbitrary
configuration to make sure the internal force vector and the tangent stiffness are consistent.
EXERCISE 15.4 Confirm the previous statement by repeating the buckling calculations of Exercise 9.3 using the
CR beam element provided in the Mathematica Notebook mentioned above (extract the material and stiffness
matrices, ignore the rest). Compare the speed of convergence of the CR and TL element for the cantilever
buckling problem.
15–11