Radial Return

WORKSHOP:
Geometrical interpretation of
Radial and Oblique Return Methods
PLASTICITY 2000:
“The Eighth International Symposium on Plasticity
and its Current Applications”
Session FA I-1: Computational Plasticity and Viscoplasticity
Whistler, CANADA
July 17-21, 2000
R. M. Brannon
Computational Physics and Mechanics Department, 9232
Sandia National Laboratories*
P.O. Box 5800, MS 0820
Albuquerque, New Mexico 87185-0820
Sponsors: ARL, DOE
*Sandia is a multiprogram laboratory operated by Sandia Corporation, a Lockheed Martin

Company, for the United States Department of Energy under Contract DE-ACO4-94AL85000.
-64 of 93- http://me.unm.edu/~rmbrann/gobag.html

/home/rmbrann/Teach/MtlModels/RadialReturn/RadialReturn.vug
ABSTRACT: Return algorithms are probably the most popular means of numerically solving conven-
tional plasticity equations. The basic tenets of these techniques are here rigorously justified and
interpreted geometrically in 6D stress space. For any return algorithm, the first step is to tentatively
assume elastic behavior throughout a given time step. If the resulting “trial” stress is forbidden (i.e. if
it violates the yield condition), then the tentative assumption of elastic response is rejected. Even
when it is found to violate the yield condition, the trial stress is nevertheless useful because it can
then be projected back to the plastic yield surface to give the updated stress. The return algorithm is
called “normal” or “orthogonal” if the trial stress is projected directly to the nearest point on the yield
surface. The return method is called “radial” or “Prandtl” when the projection is accomplished by
reducing the magnitude of the trial stress deviator. Return algorithms are often wrongly regarded as
numerical “tricks” because they appear to be ad hoc means of keeping the stress on the yield surface.
It is natural to inquire whether other approaches might be more accurate for the same computational
cost, but it is shown here that return methods are rigorously justifiable and appear to correspond to
optimal numerical accuracy and efficiency. It is shown that issues such as plastic stability, dissipa-
tion, and convexity dictate appropriate choices for the quantities that are presumed known in the der-
ivation of return algorithms; it is not the return algorithm per se that addresses such physical
concerns. It is proved that the correct return direction is dictated by the governing equations and is
not aligned with the plastic strain rate except under certain conditions. Consequently, normality of
the plastic strain rate does not necessarily correspond to normality of the return direction, and vice
versa. These claims are proved first in the context of stationary yield surfaces and then generalized to
permit hardening or softening. The technical note is intended to provide nothing more than geometri-
cal insight into known results.

Tensors are vectors!
To a mathematician, a vector is a member of a set for which addition
and scalar multiplication satisfy certain rules.
Many familiar 3D vector concepts and theorems also apply to

tensors when regarded as 9D vectors.
3D vector operations 9D tensor operations

u = αc means u i = αc i U = αC means U ij = αC ij
˜ ˜ ˜ ˜
v = a + b means v i = a i + b i
˜ ˜ ˜ V = A + B means V ij = A ij + B ij
˜ ˜ ˜
3D inner product 9D inner product
3 3 3
r•s
˜ ˜
means ∑ ri si R:S
˜ ˜
means ∑ ∑ Rij Sij
i=1 i=1j=1

Projection operations
Orthogonal Oblique
projection projection
x x
˜ ˜
b
n ˜
˜ a
p ˜ p
˜ ˜
Plane perpendicular to n Plane perpendicular to b
˜
˜
a(b • x)
p = x – n(n • x) p = x – -----------------------
˜ ˜ ˜-
˜ ˜ ˜ ˜ ˜ ˜ ˜ a•b
˜ ˜
Note: b defines the target plane; a defines projection direction.
˜ ˜
A( B: X )
Analog for 9D tensor space: P ( X ) = X – ---------------------
˜ ˜ ˜
˜ ˜ A :B
˜ ˜
Projections are linear. . . P ( α 1 X 1 + α 2 X 2 ) = α 1 P ( X 1 ) + α 2 P ( X 2 )
˜ ˜ ˜ ˜

LEMMA
x
˜ If there is a β such that
x = y + βa , then P ( x )=P ( y ) .
y ˜ ˜ ˜ ˜ ˜
˜
Important: converse is true too!
a
˜
P ( x )=P ( y )
˜ ˜
Analog for tensors:
If X = Y + βA then P ( X )=P ( Y ) and vice versa.

˜ ˜ ˜ ˜ ˜
Corollary: P ( P ( X ) ) = P ( X ) (projecting twice makes no change).

˜ ˜

Nonhardening plasticity
Known:
B , gradient of yield function ( B ij = ∂f ⁄ ∂σ ij ).
˜
ε˙ , total strain rate.
˜
E , fourth-order elastic tangent stiffness tensor.
˜˜
M , direction of the plastic strain rate.
˜
Unknown:
σ˙ , rate of stress
˜
ε˙ e , elastic part of the strain rate
˜
ε˙ p , plastic part of the strain rate.
˜
λ̇ , magnitude of the plastic part of the strain rate.

Governing equations
ε˙ = ε˙ e + ε˙ p strain rate decomposition

˜ ˜ ˜
ε˙ p = λ̇M plastic strain direction is known
˜ ˜˜
σ˙ = E :ε˙ e stress rate linear in elastic strain rate
˜ ˜˜ ˜
B : σ˙ = 0 stress stays on (nonhardening) yield surface
˜ ˜
Solution:
Note ε˙ e = ε˙ – ε˙ p = ε˙ – λ̇M , so σ˙ = E : ( ε˙ – λ̇M ) . For convenience,
˜ ˜ ˜ ˜ ˜ ˜ ˜˜ ˜ ˜
define σ˙ trial = E :ε˙ and A = E :M . Then σ˙ = σ˙ trial – λ̇ A
˜ ˜˜ ˜
˜˜ ˜ ˜ ˜ ˜ ˜
Enforce last equation to get B : [ σ˙ trial – λ̇ A ] = 0 . Solve for λ̇ and back
˜ ˜ ˜
 B : σ˙ trial
substitute to get solution for stress rate: σ˙ = σ˙ trial –  -----------------
˜ ˜ - A .
˜ ˜  B :A ˜
˜ ˜ 

Geometrical interpretation
Slightly rearrange solution to final form:
A( B: X )
σ˙ = P ( σ˙ trial ) where P ( X ) = X – ---------------------
˜ ˜ ˜
˜ ˜ ˜ ˜ A :B
˜ ˜
A = E :M
˜ ˜˜ ˜˜
B
˜ ε˙ p =λ̇M
˜ ˜
yield σ˙ trial ∆t
surface ˜
σ˙ ∆t e
σ σ˙ ∆t
p
˜
˜
˜
Numerical solution: σ = σ trial + β A . Find β by f ( σ trial + β A ) = 0 .

˜ ˜ ˜ ˜ ˜

Discussion
The return direction is...

• coaxial with A (= E :M ).
˜ ˜˜ ˜
• not generally normal to the yield surface.
• not generally aligned with the plastic strain rate.
• not dictated by physical considerations such as positive dissipation,
yield surface convexity, or plastic stability. (Such concerns dictate
appropriate values for “known” quantities.)
• “radial” if and only if the material is plastically incompressible.
• An algorithm that returns normal to the yield surface (i.e., A = αB )
–1 ˜ ˜
is implicitly using a plastic strain rate direction M = αE :B .
˜ ˜˜ ˜
The above analysis can be generalized (see web document) to
include hardening/softening. Projection of the trial stress back to the
current yield surface remains valid even though the stress rate is no
longer a projection of the trial stress rate.

Equivalent plastic strain
Many constitutive models use yield surface evolution laws that
depend on the so-called “equivalent plastic strain,” which is defined
2˙ p ˙ p 2
γp ≡ ∫ --- ε ′: ε ′ dt =
3˜ ˜
---
3 ∫ ε˙ p ′ dt
˜
The best method uses the definition directly:
S˙
2 2
∆γ p ≡ --- ε˙ ′ – ε˙ e ′ ∆t , or, for isotropic, ∆γ p ≡ --- ε˙ ′ – -------
˜ - ∆t .
3 ˜ ˜ 3 ˜ 2G
For a finite time step ∆t ,

2G ( ε˙ ∆t p )
˜
S new – S old
2G ( ε˙ ∆t e ) γ new ≡ γ old + 2
--- ε˙ ′ ∆t – ---------------------------
˜ ˜ -
˜ S* p p 3 2G
˜ ˜
S old S new
˜ ˜
(...better suited for partially plastic intervals.)

Supplemental topic:
invariant yield functions
Tresca: Stress state is below yield if and only if
Good 1
f ( σ ) = --- max ( σ 1 – σ 2 , σ 2 – σ 3 , σ 3 – σ 1 ) – k < 0 (1)
˜ 2
Some authors (e.g. Fung, 1965, Lubliner 1990) wrongly claim that
an acceptable alternative Tresca yield function is
BAD
F ( σ ) = [ ( σ 1 – σ 2 ) 2 – 4k 2 ] [ ( σ 2 – σ 3 ) 2 – 4k 2 ] [ ( σ 3 – σ 1 ) 2 – 4k 2 ] . (2)
˜
This is intoxicating because it can be written with invariants as
WORSE
F ( σ ) = 4J 23 – 27J 32 – 36k 2 J 22 + 96k 4 J 2 – 64k 6 (3)
˜
FATAL FLAW: If stress is below yield, then F ( σ ) ≤ 0 , but converse

˜
is false! A return algorithm using F might wrongly think a plastic trial
stress is below yield. For example, σ 1 = σ 2 = 3k and σ 3 = 0 is
correctly identified to be above yield by f ( σ ) , but not by F ( σ ).
˜ ˜

Plot of (bad) invariant Tresca function
Under the assumption of plane stress where σ 3 = 0 , regions where
F ( σ ) > 0 are shown in black. A valid yield function should be black
˜
everywhere outside the yellow Tresca hexagon. The invariant F ( σ )
˜
is invalid!
σ 2 ⁄ 2k
σ 1 ⁄ 2k

Supplemental topic: 9D vector basis
Recall that tensors are 9D vectors, so we may define a 9 × 1
component array for them: T 1o, T 2o, T 3o, T 4o, T 5o, T 6o, T 7o, T 8o, T 9o =
{ T 11, T 21, T 31, T 12, T 22, T 32, T 13, T 23, T 33 } ,
3D vector basis expansion 9D tensor expansion

v = v1 e1 + v2 e2 + v3 e3 T = T 11 e 1 e 1 + T 12 e 1 e 2 + T 13 e 1 e 3
˜ ˜ ˜ ˜ ˜ ˜ ˜ ˜ ˜ ˜ ˜
+ T 21 e2 e1 + T 22 e2 e2 + T 23 e2 e3
˜ ˜ ˜ ˜ ˜ ˜
+ T 31 e3 e1 + T 32 e3 e2 + T 33 e3 e3
˜ ˜ ˜ ˜ ˜ ˜
Summation form Summation form
3 3 3 9
v =
˜ ∑ vk e˜ k T =
˜ ∑ ∑ T ij e˜ i e˜ j = ∑ o ξo
TK
˜K
k=1 i=1j=1 K =1
where T 1o = T 11 , ξ o = e 1 e 1 , T 2o = T 21 , ξ o = e 2 e 1 , etc.
˜1
˜ ˜ ˜ ˜2 ˜ ˜

Subspace of symmetric tensors
Suppose that a physical problem involves a plane even if there are

some non-planar aspects of the motion (e.g., oblique impact of a
projectile onto a slab of armor). For solving the problem, any
sensible engineer would line up a basis with the plane: all base
vectors are either in the plane or normal to the plane.
The set of all symmetric tensors forms a subspace, which is

analogous to a plane. The “normal” to the plane is the set of all
skew-symmetric tensors. If you add two vectors in a plane, the result
is also in the plane. Analogously, if you form any linear combination
of symmetric tensors, the result is also symmetric.
Yield functions are defined for stress, which is symmetric. Our

constitutive modelling problems intimately involve symmetric
tensors, so it makes sense to use a basis for tensor space such that
all base tensors are either purely symmetric or purely skew-
symmetric.

Voigt vs. Mandel — Introduction
Voigt components: { T } v = { T 11, T 22, T 33, T 23, T 31, T 12 }

Then R : S equals R 1v S 1v + R 2v S 2v + R 3v S 3v + 2 ( R 4v S 4v + R 5v S 4v + R 6v S 6v )
˜ ˜
Note the ungainly factor of 2 needed because the off diagonal

components contribute twice in the expression R : S = R ij S ij .
˜ ˜
Mandel components: { T } m = { T 11, T 22, T 33, 2T 23, 2T 31, 2T 12 }

Then R : S equals R 1m S 1m + R 2m S 2m + R 3m S 3m + R 4m S 4m + R 5m S 4m + R 6m S 6m
˜ ˜
Ah! much more intuitive! The Mandel approach incorporates the
factor of 2 inside the definition of the components.
Q: Is the Mandel convention just a “trick” likely to bite us some day?

A: NO! Voigt components are the dangerous choice — they are
referenced to an irregular basis for symmetric tensors. Mandel
components are referenced to the same — but normalized — basis!

Change of basis for tensors
The basis expansion of any tensor may be written

T = T 11 e1 e1 +T 12 e1 e2 +T 13 e1 e3 +T 21 e2 e1 +T 22 e2 e2 +T 23 e2 e3 +T 31 e3 e1 +T 32 e3 e2 +T 33 e3 e3
˜ ˜ ˜ ˜ ˜ ˜ ˜ ˜ ˜ ˜ ˜ ˜ ˜ ˜ ˜ ˜ ˜ ˜ ˜
= T ( 11 ) e 1 e 1 + T ( 22 ) e 2 e 2 + T ( 33 ) e 3 e 3
˜ ˜ ˜ ˜ ˜ ˜
+ T ( 23 ) ( e 2 e 3 + e 3 e 2 ) + T ( 31 ) ( e 3 e 1 + e 1 e 3 ) + T ( 12 ) ( e 1 e 2 + e 2 e 1 )
˜ ˜ ˜ ˜ ˜ ˜ ˜ ˜ ˜ ˜ ˜ ˜
+ T [ 32 ] ( e 3 e 2 – e 2 e 3 ) + T [ 13 ] ( e 1 e 3 – e 3 e 1 ) + T [ 21 ] ( e 2 e 1 – e 1 e 2 )
˜ ˜ ˜ ˜ ˜ ˜ ˜ ˜ ˜ ˜ ˜ ˜
where
1
T ( ij ) ≡ --- ( T ij + T ji ) and T [ ij ] ≡ 1--- ( T ij – T ji )
2 2
If the tensor is symmetric, the last three terms are all zero. If the
tensor is skew-symmetric, then the first six terms are all zero and the
last three terms are the components of the axial vector.

Voigt sym-dev basis
T = T ( 11 ) e 1 e 1 + T ( 22 ) e 2 e 2 + T ( 33 ) e 3 e 3
˜ ˜ ˜ ˜ ˜ ˜ ˜
+ T ( 23 ) ( e 2 e 3 + e 3 e 2 ) + T ( 31 ) ( e 3 e 1 + e 1 e 3 ) + T ( 12 ) ( e 1 e 2 + e 2 e 1 )
˜ ˜ ˜ ˜ ˜ ˜ ˜ ˜ ˜ ˜ ˜ ˜
+ T [ 32 ] ( e 3 e 2 – e 2 e 3 ) + T [ 13 ] ( e 1 e 3 – e 3 e 1 ) + T [ 21 ] ( e 2 e 1 – e 1 e 2 )
˜ ˜ ˜ ˜ ˜ ˜ ˜ ˜ ˜ ˜ ˜ ˜
Traditional Voigt:
T 1v =T ( 11 ) , T 2v =T ( 22 ) , T 3v =T ( 33 ) , T 4v =T ( 23 ) , T 5v =T ( 31 ) , ...
ξ v =e 1 e 1 , ξ v =e 2 e 2 , ξ v =e 3 e 3 , ξ v = ( e 2 e 3 + e 3 e 2 ) , ξ v = ( e 3 e 1 + e 1 e 3 ) , ...
˜1 ˜ ˜ ˜2 ˜ ˜ ˜3 ˜ ˜ ˜4 ˜ ˜ ˜ ˜ ˜5 ˜ ˜ ˜ ˜
9
Then T =
˜ ∑ v ξv .
TK
˜K
K =1
For symmetric, T ( ij ) = T ij and T [ ij ] = 0 , so the sum may be

truncated at six terms.
MAJOR PROBLEM: Voigt basis is not normalized!

Voigt basis is not normalized
Consider the inner product:
9 9 9 9
   
R:S = 
˜ ˜ K
∑ v ξ v  :
RK ∑
˜ K J = 1
S Jv ξ v  =
˜ J
∑ ∑ R Kv SJv ( ξ˜ Kv :ξ˜ Jv )
=1 K = 1J = 1
The Voigt basis is orthogonal:

ξ v :ξ v = 0 if K ≠ J .
˜K ˜J
The first three Voigt base tensors are normalized:

ξ v :ξ v = 1 , ξ v :ξ v = 1 , and ξ v :ξ v = 1 , but the remaining base tensors
˜1 ˜1 ˜2 ˜2 ˜3 ˜3
are not normalized. They all have a magnitude of 2 . Thus
9
R:S =
˜ ˜ ∑ v Sv ξ
RK K K
˜
2 = R vS v + R vS v + R vS v + 2( R vS v) + 2( R vS v) + …
1 1 2 2 3 3 4 4 5 5
K =1

MANDEL basis
Obvious thing to do ... normalize the basis.
ξv
Mandel basis: ξ m ≡ ------------ ˜K -.
˜K ξv
˜K
T 1m =T ( 11 ), T 2m =T ( 22 ), T 3m =T ( 33 ), T 4m = 2T ( 23 ) , T 5m = 2T ( 31 ) , ...
( e2 e3 + e3 e2 ) ( e3 e1 + e1 e3 )
ξ m =e 1 e 1 , ξ m =e 2 e 2 , ξ m =e 3 e 3 , ξ = -----------------------------------
m ˜ ˜ ˜ ˜ , ξ = -----------------------------------
m ˜ ˜ ˜ ˜ , ...
˜1 ˜ ˜
˜ ˜2 ˜ ˜
˜ ˜3 ˜ ˜
˜ ˜
˜ 4 2 ˜
˜ 5 2
9 9
Then T =
˜ ∑ mξm ,
TK
˜K
ξ m :ξ m = δ KJ ,
˜K ˜J
and R:S =
˜ ˜ ∑ m Sm
RK K
K =1 K =1
With this orthonormal Mandel basis, the tensor inner product takes a
form that is a direct analog of the ordinary 3D vector inner product
formula that applies when the basis is orthonormal.

Mandel basis for symmetric tensors
Dropping “m” identifier, the Mandel basis for 9D full tensor space is
ξ1 = e1 e1 , ξ2 = e2 e2 , ξ3 = e3 e3
˜ ˜ ˜ ˜ ˜ ˜ ˜ ˜ ˜
1 1 1
ξ 4 = ------- ( e 2 e 3 + e 3 e 2 ) , ξ 5 = ------- ( e 3 e 1 + e 1 e 3 ) , ξ 6 = ------- ( e 1 e 2 + e 2 e 1 )
˜ 2 ˜ ˜ ˜ ˜ ˜ 2 ˜ ˜ ˜ ˜ ˜ 2 ˜ ˜ ˜ ˜
1 1 1
ξ 7 = ------- ( e 3 e 2 – e 2 e 3 ) , ξ 8 = ------- ( e 1 e 3 – e 3 e 1 ) , ξ 9 = ------- ( e 2 e 1 – e 1 e 2 )
˜ 2 ˜ ˜ ˜ ˜ ˜ 2 ˜ ˜ ˜ ˜ ˜ 2 ˜ ˜ ˜ ˜
The basis is orthogonal because ξ : ξ = 0 if K ≠ J . The basis is

K J
˜ ˜
normalized (i.e., ξ : ξ = δ KJ ) because of the factors of 2 .
K J
˜ ˜
Just as an ordinary vector has components v k = v • e k , the Mandel

˜ ˜
components of a tensor T are T K = T : ξ .
˜ ˜ ˜K

Related topic: isomorphic stress space
Stress: σ
˜
1 1
Mean stress: p = --- trσ = --- I :σ (positive in tension)
3 ˜ 3˜ ˜
Stress deviator: S = σ – pI
˜ ˜ ˜
Magnitude of the stress deviator: τ ≡ S : S

˜ ˜
S S S
ˆ
Unit tensor in the direction of S : S ≡ --------- = -------------- = ---˜-
˜ ˜
˜ ˜ S S:S τ
˜ ˜ ˜
Then σ = τSˆ + pI .
˜ ˜ ˜
We now show that non-intuitive factors appear because the identity

I is not a unit tensor. Specifically, I = I :I = 3 .
˜ ˜ ˜ ˜

Motivational example
A popular simplified yield criterion assumes that the yield function
depends only on τ and p . F ( σ ) = f ( τ, p ) . The yield surface defined
˜
by F ( σ ) = 0 is a hyper-cylinder in stress space — it is a surface of
˜
revolution about the isotropic axis.
∂f dτ ∂f dp ∂f ˆ ∂f 1
-------- = ------  ------- + ------  ------- = ------ ( S
Gradient of yield: B = dF ) + ------  --- I
˜ dσ ∂τ  dσ ∂p  dσ ∂τ ˜ ∂p  3 ˜
˜ ˜ ˜ ˜ ˜ ˜
Let σ t = τ t Sˆ + p t I denote a trial (t) elastic stress.
˜ ˜ ˜
Let σ n = τ n Sˆ + p n I denote the new (n) updated stress on the yield

˜ ˜ ˜
surface obtained by returning to the nearest point on the yield
surface in stress space. (We now know this doesn’t necessarily mean that the plastic
strain rate is normal to the yield surface — we use a normal return direction to illustrate a different point
–1
here. A normal return direction implies a direction of plastic strain rate parallel to E :B ).
˜˜ ˜

Normal projection (cont’d)
Suppose we wish to return nearest point on yield surface. Then

σ t – σ n is normal to the yield surface. There’s a scalar β such that
˜ ˜
σ t – σ n = βB , or ˆ + ( p t – p n )I = β  ∂f
( τ t – τ n )S
∂f 1
------ ( Sˆ ) + ------  --- I 
˜ ˜ ˜ ˜ ˜  ∂τ ˜ ∂p  3 ˜ 
τt – τn ∂f ⁄ ∂τ τ
Therefore ------------------
pt – pn
- = 3  ∂f
---------------- .
⁄ ∂p
trial stress
( p t, τ t )
Thus, to project normal to the yield surface in

wrong answer
stress space, you must project using a slope correct answer
( p n, τ n )
3 times steeper than the normal in τ vs. p GOAL: RETURN TO NEAREST POINT
ON THE YIELD SURFACE in stress space! p
space. This counterintuitive behavior arises That means project obliquely in this τ v.s. p space.
because τ and p are not isomorphic to stress

space. The base tensors Sˆ and I , while orthogonal, are not normalized. We
˜ ˜
should instead use Iˆ ≡ I ⁄ I = I ⁄ 3 with an appropriately modified
˜ ˜ ˜ ˜
measure of mean stress. Namely, p̂ = 3 p .

Rendulic plane
The Rendulic plane plots a “shear stress” versus a “mean stress.”
Engineer’s choice View in 3-D

physical space Cylindrical
z vs. r plane.
zê z
˜ x
“shear stress:” τ = S :S , and x =rê r + zê z
˜ ˜ ˜ ˜
˜ ˜
1
“mean stress:” p = --- trσ . Then σ = S + pI . ê z
˜
3 ˜ ˜ ˜ ˜ ê r rê r
Problem: This τ vs. p space isn’t ˜ ˜
isomorphic to stress space. For example,

σ :σ ≠ τ 2 + p 2 . Importantly, the normal to the τ
˜ ˜
yield surface in τ vs. p space is not normal
isomorphic
to the yield surface in stress space. engineering
p or p̂
Mathematician’s (isomorphic) choice: “shear stress”

ˆ and “mean 1
τ = S :S = σ :S stress” p̂ = ------- trσ = σ :Iˆ = 3 p . Then
˜ ˜ ˜ ˜ 3 ˜ ˜ ˜
ˆ + p̂Iˆ . The normalized
σ = τS Iˆ is like the ê z cylindrical base vector.
˜ ˜ ˜ ˜ ˜

Supplemental Topic:
Anisotropic yield surfaces
For elastically anisotropic material, a very common “first-cut” best
guess at the plastic yield surface is a Tsai-Wu ellipsoid of the form
f ( σ ) = ( σ – σ∗ ): L : ( σ – σ∗ ) – 1 , (contrary to Walker’s recent claims, this form is perfectly capable

˜ ˜ ˜ ˜˜ ˜ ˜ of modelling even highly anisotropic media.)
where L shares the same anisotropy with the stiffness E .

˜˜ ˜˜
Elastic constants may be nondestructively measured, but the yield

L ijkl parameters are more difficult since a fresh sample must be
used to measure each component. Thus, data are often lacking.
Proposal: Face with a dearth of data, assume that E and L have

˜˜
the same eigenprojectors, a term which we now define... ˜˜

What are eigenprojectors?
To illustrate, consider simpler 3D space. Here’s a sample tensor
17 – 2 – 2 λ1 = 9 v 1 = 1--3- { 1, 2, 2 }
˜ 1
[ A ] = – 2 14 – 4 , which has eigenpairs λ 2 = 18 v 2 = ------5- { – 2, 0, 1 }
˜ ˜
– 2 – 4 14 λ 3 = 18 1
v 3 = ---------- { – 2, 5, – 4 }
˜ 3 5
In spectral form, A = λ 1 v 1 v 1 + λ 2 v 2 v 2 + λ 3 v 3 v 3
˜ ˜ ˜ ˜ ˜ ˜ ˜
= 9v 1 v 1 + 18 ( v 2 v 2 + v 3 v 3 )
˜ ˜ ˜ ˜ ˜ ˜
P P unique!
˜1 ˜2
With respect to the principal basis,
x
9 0 0 100 000 P •x ˜
A = 0 18 0 , P = 0 0 0 , and P 2 = 0 1 0 ˜1 ˜
˜ ˜1 ˜ P •x
0 0 18 000 001 ˜2 ˜

What are eigentensors?
We seek tensors Y and scalars λ such that E :Y = λY . The major

˜ ˜˜ ˜ ˜
and minor symmetries of E allow this to be written as an ordinary
˜˜
6 × 6 matrix eigenproblem:
E 1111 E 1122 E 1133 2E 1123 2E 1131 2E 1112 Y 11 Y 11

E 2211 E 2222 E 2233 2E 2223 2E 2231 2E 2212 Y 22 Y 22
E 3311 E 3322 E 3333 2E 3323 2E 3331 2E 3312 Y 33 Y 33

= λ
2E 2311 2E 2322 2E 2333 2E 2323 2E 2331 2E 2312 2Y 23 2Y 23
2E 3111 2E 3122 2E 3133 2E 3123 2E 3131 2E 3112 2Y 31 2Y 31
2E 1211 2E 1222 2E 1233 2E 1223 2E 1231 2E 1212 2Y 23 2Y 23
An eigensolver will output a set of six orthonormal 6-dimensional

eigenvectors. Each of these correspond to symmetric eigentensors.

If λ has multiplicity of 1, then P ijkl = Y ij Y kl is the
corresponding eigenprojector. When it operates on any
tensor, the result is the part of that tensor in the direction of Y ij .
EXAMPLE: For isotropy, 3K is an eigenvalue of multiplicity 1. The

normalized eigentensor is I ⁄ 3 . The projector is 1--3- δ ij δ kl , which
˜
merely returns the isotropic part of any tensor it operates on.
If λ has multiplicity of 2, then the eigentensors Y ( 1 ) and Y ( 2 ) are not

˜ ˜
unique. Instead, the eigenprojector, P ijkl = Y ij( 1 ) Y kl
( 1 ) + Y ( 2 ) Y ( 2 ) is
ij kl
unique. When it operates on an arbitrary tensor, the result is the part
of the tensor in the subspace. Higher multiplicities are similar.
EXAMPLE: For isotropy, 2G is an eigenvalue of multiplicity 5. The

eigenprojector (constructed by summing dyads of the five
orthonormalized eigenprojectors) returns the deviator of any tensor it
operates on. Thus, ANY DEVIATORIC TENSOR is an eigentensor
for isotropy.

Back to anisotropic yield...
Recall f ( σ ) = ( σ – σ∗ ): L : ( σ – σ∗ ) – 1 . If the material is transverse, the

˜ ˜ ˜ ˜˜ ˜ ˜
Mandel eigenproblem is of the form
E0 E2 E3 0 0 0 Y 11 Y 11
E2 E0 E3 0 0 0 Y 22 Y 22
E3 E3 E1 0 0 0 Y 33 Y 33
= λ ,
0 0 0 E4 0 0 2Y 23 2Y 23 where
0 0 0 0 E5 0 2Y 31 2Y 31 E 1 = E 3333 , E 2 = E 1122 , E 3 = E 1133 ,
0 0 0 0 0 E5 2Y 23 2Y 23 E 4 = 2E 2323 , E 5 = 2E 1212 , E o = E 2 + E 5
There are five independent stiffnesses, but only four independent

eigenvalues (and therefore only four independent eigenprojectors).
Forcing L to have the same eigenprojectors gives a formula for the
˜
elusive L˜1133 value that couples lateral and axial response.

Conclusions
This presentation covered many applications that illustrate the
usefulness of regarding tensors as higher-dimensional vectors.
Key points were

• For radial and oblique return models, the stress may be returned to
the yield surface via a projection operation that is analogous to
projecting a simple vector onto a plane.
• Symmetric tensors are analogous to planes. The Mandel convention
for symmetric tensor components correspond to an orthonormal
basis for symmetric tensors.
• The invariant form of the Tresca yield criterion is invalid because
negative values of that “yield function” do not necessarily
correspond to stresses that are below yield.
• The isomorphic stress measures are a more accurate representation
of stress space that is analogous to viewing the stress “vector” in the
“plane” formed by the isotropic tensor and the stress itself.
• Anisotropic yield may be coupled to elastic isotropy via the elastic
eigenprojectors.


Radial Return

Enviado por

Dados do documento

Direitos autorais

Formatos disponíveis

Compartilhar este documento

Compartilhar ou incorporar documento

Opções de compartilhamento

Você considera este documento útil?

Este conteúdo é inapropriado?

Direitos autorais:

Formatos disponíveis

Radial Return

Enviado por

Direitos autorais:

Formatos disponíveis

WORKSHOP:

*Sandia is a multiprogram laboratory operated by Sandia Corporation, a Lockheed Martin

-64 of 93- http://me.unm.edu/~rmbrann/gobag.html

-65 of 93- http://me.unm.edu/~rmbrann/gobag.html

Many familiar 3D vector concepts and theorems also apply to

3D vector operations 9D tensor operations

3D inner product 9D inner product

-66 of 93- http://me.unm.edu/~rmbrann/gobag.html

-67 of 93- http://me.unm.edu/~rmbrann/gobag.html

Analog for tensors:

If X = Y + βA then P ( X )=P ( Y ) and vice versa.

Corollary: P ( P ( X ) ) = P ( X ) (projecting twice makes no change).

-68 of 93- http://me.unm.edu/~rmbrann/gobag.html

-69 of 93- http://me.unm.edu/~rmbrann/gobag.html

ε˙ = ε˙ e + ε˙ p strain rate decomposition

-70 of 93- http://me.unm.edu/~rmbrann/gobag.html

Numerical solution: σ = σ trial + β A . Find β by f ( σ trial + β A ) = 0 .

-71 of 93- http://me.unm.edu/~rmbrann/gobag.html

The return direction is...

-72 of 93- http://me.unm.edu/~rmbrann/gobag.html

The best method uses the definition directly:

For a finite time step ∆t ,

-73 of 93- http://me.unm.edu/~rmbrann/gobag.html

FATAL FLAW: If stress is below yield, then F ( σ ) ≤ 0 , but converse

-74 of 93- http://me.unm.edu/~rmbrann/gobag.html

-75 of 93- http://me.unm.edu/~rmbrann/gobag.html

3D vector basis expansion 9D tensor expansion

-76 of 93- http://me.unm.edu/~rmbrann/gobag.html

Suppose that a physical problem involves a plane even if there are

The set of all symmetric tensors forms a subspace, which is

Yield functions are defined for stress, which is symmetric. Our

-77 of 93- http://me.unm.edu/~rmbrann/gobag.html

Voigt components: { T } v = { T 11, T 22, T 33, T 23, T 31, T 12 }

Note the ungainly factor of 2 needed because the off diagonal

Mandel components: { T } m = { T 11, T 22, T 33, 2T 23, 2T 31, 2T 12 }

Q: Is the Mandel convention just a “trick” likely to bite us some day?

-78 of 93- http://me.unm.edu/~rmbrann/gobag.html

The basis expansion of any tensor may be written

-79 of 93- http://me.unm.edu/~rmbrann/gobag.html

For symmetric, T ( ij ) = T ij and T [ ij ] = 0 , so the sum may be

MAJOR PROBLEM: Voigt basis is not normalized!

-80 of 93- http://me.unm.edu/~rmbrann/gobag.html

The Voigt basis is orthogonal:

The first three Voigt base tensors are normalized:

-81 of 93- http://me.unm.edu/~rmbrann/gobag.html

-82 of 93- http://me.unm.edu/~rmbrann/gobag.html

The basis is orthogonal because ξ : ξ = 0 if K ≠ J . The basis is

Just as an ordinary vector has components v k = v • e k , the Mandel

-83 of 93- http://me.unm.edu/~rmbrann/gobag.html

Magnitude of the stress deviator: τ ≡ S : S

We now show that non-intuitive factors appear because the identity

-84 of 93- http://me.unm.edu/~rmbrann/gobag.html

Let σ n = τ n Sˆ + p n I denote the new (n) updated stress on the yield

-85 of 93- http://me.unm.edu/~rmbrann/gobag.html

Suppose we wish to return nearest point on yield surface. Then

Thus, to project normal to the yield surface in

because τ and p are not isomorphic to stress

-86 of 93- http://me.unm.edu/~rmbrann/gobag.html

Engineer’s choice View in 3-D

isomorphic to stress space. For example,

to the yield surface in stress space. engineering

Mathematician’s (isomorphic) choice: “shear stress”

-87 of 93- http://me.unm.edu/~rmbrann/gobag.html

f ( σ ) = ( σ – σ∗ ): L : ( σ – σ∗ ) – 1 , (contrary to Walker’s recent claims, this form is perfectly capable