Compmrs & Stnfctims Vol. 8.

pp X5-369 00(5-%49/%10s0Lo365/w oolo

@ PetpmonPressLtd..1978. Pmted in GreatBritatn

VIRTUAL WORK

G. A. 0. DAVIES
Imperial College of Science and Technology. London, England

(Receiued October 19’77; received for publication S January 1978)

Abstract-The principle of virtual work in structural mechanics is reviewed, especially when applied to finite
element approximations with varying degrees of continuity. It is shown that displacement. equilibrium. mixed,
and hybrid formulations are simply and directly deduced without recourse to extremum principles and fictitious
energies, Lagrangemultipliers,or jump integrals.Continuity requirementsbecome self-evidentas does the nature
of the terms “pure”, “mixed” and “hybrid”. Two virtual work expressions are used which are dual but separate
with no fictional unification as is impfied in the general&d functional approach.

1. ~ODU~ON If we view eqn (1) as a “virtual conserva~on of work” it

In a classic series of articles in 1954[l] Argyris and may be deduced in a different fashion, for example
Kelsey presented a unified treatment of energy methods considering one stress component only, the creation of
including all the known special theorems, together with work in V should be balanced by the net e&x through
their generalisations, from the basic, and physically S,
natural, theorem of minimum potential energy to the
highly unnatural theorem of minimum complementary
potential energy of total deformation. However the use
of the title “Energy Theorems” was a misnomer, possibly
I $(ux.u)
v d.xdy dz =
I S
&uds,
& ax

to lure the conservative reader, because the complete where n is the outward normal. The summation of this
generalised discourse was based on the twin principles of equality for all nine stress~splacement products may be
virtual dispIacements and virtual forces first pro~unded written as
by Engesser 121in 1899. Since 1954Professor Argyris has
remained faithful to virtual work arguments, choosing
not to enter the race for functionals which took place in
a positive explosion of variational principles in the last
I
V
(u’Da + e’D*\l) do =
I s
u’(Dn)a ds (2)

decade [3,4]. It is a little curious that most finite element where the array of operators
models have been posed as the solution of a stationary
functional with constraints when work arguments yield
directly the Gale&m form which itself is readily extend-
able as a weighted residual technique[5] to other classes
of boundary value problems. Perhaps the universal prin- D=
ciple of m~mum potential energy has persuaded us
always to seek a grand unifyi~ theorem embracing all
field equations, boundary conditions, and continuity (3)
requirements, when in reality the demands of structural
mechanics (outside the stress-strain law) are the quite and (Dn) implies that all terms in the array operate on
simple and separate requirements of equilibrium and the same scalar R. Equation (2) is in fact a special form
compatibility. No generalised variation principle should of Gauss’ identity valid for finite D’u and Da, that is
cloud this fact. We now show briefly that familiar work piecewise-continuous stress and displacement fields. If
arguments are sufficient and more general than is con- we rearrange (2) as
ventionally supposed.

2. Virtual Worfi
The most impressive demons~tion is as follows. We
I uf
”
Dtu dv = -
I ”
ufDa du t
s
u’(Dn)u ds

suppose a body is subjected to surface and body force and compare with eqn (1X we may conclude that
vectors ps and py in S and in V, producing displacements
II and stresses and strains a and z, where it is convenient s=D’uinV
to adopt a Cartesian notation
Datp,=OinV
(Dn)a ‘ps on S.

We define, at this stage, virtual work as the product of Thus we appear to have extracted the equations of
force and displacement simply to avoid specifying the compatibility and the two sets of equilibrium equations
stress-strain law, and equate internal work to external, from a single work statement. (In non-Cartesian coor-
thus dinate systems the same ~gument leads to arrays of the
form D rt A where A contains only scalar multipliers [6]).
This demonstration is however only superficially im-
Iv~‘.dV=Ivp.‘udv+19p,‘uds. (I)
pressive, and eqn (1) is unusable until we define virtual
365
366 G. A. 0. DAVIES

displacements or forces; for consider the principle of where

virtual displacements, writing eqn (1) as
K = X a,’ ksas and R = 2 a:R, ;
8

and so because all components of f are virtual,

We denote virtual displacements by a bar since it is quite
unnecessary to consider smaller displacements 61 unless R=Kr. (8)
the real deformations are gross. If Gauss’ identity (2) is
inserted into (4) it becomes Apart from summing (6) to obtain the total solution (8)
we have done more as may be deduced if we write the
equivalent statement (5) as a summation,
(I’@ - D’ii) du - ii’(Datp,)du
Y I ”

t
Is
u’[(Dn)u-ps]du=O. v,
C’(Du + py) do -
s,
u’(Dn)u ds

+I,, ii’p, ds] = 0. (9)

We see now that the principle of virtual displacements
works provided we enforce compatibility on the virtual
displacements, 4 = D’g, whence we are left with Our approximate solution is clearly a weighted residual
technique[5] in V, where Da t pv hopefully converges
to 0 the more degrees of freedom we include in the
ii’(Dotp,)du- ii’[(Dn)u - ps] ds = 0 (5)
I V I s weight C. If we select ii = wp and Do +p” =
DKD~CCJ~ t p” then the method is a true Galerkin ap-
and provided that ii is arbitrary we may conclude that proximation. Further the contour integral on S,, when
summed, will pair with contiguous elements so that at an
Dutp,=OinVand(Dn)u=p,onS. interface between elements “i” and “j” the combined
integral will be
Notice that although it is usual to enforce E = DLuupon
the real displacements to obtain a displacement solution,
the above principle does not require this (e.g. the struc-
ture may be statically determinate). Secondly it is also
I SR
ii’[(Dn)ui - (Dn)uil ds

usual to restrict the surface integral to that part of the provided that fi is continuous-a kinematically conformal
surface S, where the forces ps are prescribed by choos- element. Thus we see why tractions (Dn)u need not be
ing ii = 0 on S. where the displacements are prescribed. continuous and why then interface equilibrium is
This is convenient but, unlike the variational approach, is satisfied only “in the mean”-indeed it is consistent to be
clearly not necessary; it is permissible to use (4) to no more demanding across S, than in V,. This “relaxed
deliver unknown reactions ps on S, where u = 0. and in continuity condition” was first pointed out by PragerI71
an approximate displacement solution this may be much who also proposed an additional term in the appropriate
more accurate than evaluating (Dn)u from the strains functional to allow discontinuous tractions.
E = D’u.
4.THE FINITRELEMENT FORCE METROD
3.TRE FlNlTJl
ELEMENT DISPLACEMENT METHOD We now turn to the dual of the principle of virtual
Suppose we briefly recall the finite element ap- displacements-the principle of virtual forces. Thus, in-
proximation, thus in the gth element we expand u = wp, stead of (4) we simply write
in V, in terms of nodal displacements ps so that the
virtual work (4) for a single element is
j-vBtedu-~vp,,?tdu-j-sp.(Ods=O. (10)
(6)
In an analogous way, it is permissible to allow virtual
where forces 8%on both S, and S,, indeed the most convenient
way of obtaining u on S, from a known strain field is to
u=ice,ks= “, (DL~)‘~(D’w)du, allow pI to induce an equilibrating field d and then use
(lo)-the generalised unit load method[l]. However it is
again convenient to eliminate unknown displacements on
and S, by stipulating that pd = 0 leaving only the prescribed
displacements i on S,,. Using the Gauss identity (2) on
(10) the principle of virtual forces becomes
R, = ~‘pu du t o’ps du,
I “a I %iJ

and we have now restricted the element surface integral

to that part S,, where forces are prescribed. By relat-
I V
b’(e-D*u)du-
V
u’(D~tp,)dV

ing nodal displacements to a global list r as ps = a,r we

may sum (6) for the whole structure as
t
I s u’(Dn)ti ds -
I
s, ps’ii ds = 0.

P’(Kr-R)=O (7) We now enforce equilibrium on the virtual stresses, thus

 Virtual work 367

Db + pu = 0 in V and (Dn)6 = pS on S, and zero on S,, though the real displacements can never be extracted
and are left with from the integral

I ”
b’(r - D’u) du t
I
su [(Dn)cF]‘(u -ii) ds = 0
(11)
I s, [(Dn)eF]‘uds = Pip,.

With
Equation (11) is clearly the dual to (5) as we would
expect, and the dual to (9) becomes p” = 0, 6 = /3PN,

and
C?(E- D’u) du t s, [(Dn)&]’ u ds
“8
PS = BNPN;
- 1,.[(Dn)61’ii ds] = 0. (12)
so that

This time we deduce that provided the tractions (not the

stresses) are continuous across interfaces, then dis- I, Bsluds =I, [(Dn)@]‘u ds = PN*BN’pg;
placements may be discontinuous, but will be continuous
“in the mean” as will be the interior equations of and the element virtual work (12) becomes
compatibility. It is possible to go a stage further and
allow interface displacements to differ from the boun- PN’[fN PN + fo - B~‘pgl
dary values of the fields in both contiguous elements [S].
We simply apply our summation (12) over existing ele- where
ments and an additional interface element of infinites-
simal volume, as shown in the figure. One half of the
contour integral over the interface element is then in-
corporated into the element term (12). Such relaxed
continuity has been exploited successfully in a few and for interest we have added initial strains R into
cases [ 141. E = #a t R, As the PN are element variables only, the
The construction of equilibrium elements is well above work must be zero and hence
known[9] and can be formed directly from (10) putting
py = 0 to eliminate displacement variables. We satisfy PN = fNv-‘Bjv’& - frv--‘fo
equilibrium directly by obtaining suitable interpolation giving
functions /I from D/l = 0 and add to this the particular
solutions fl,, of D/S = -pv. Thus, P,=k&-BNfr,-‘fotBo,

where
fJ=BPN+&l. (13)
k, = BNfrv-‘Brv’.
(The suffix “N” is chosen for the unknowns PN because
these fields are clearly “natural” stress tields[lO] static- It is interesting to note that this expression for the
ally compatible with a set of self-equilibrating surface element stiffness is identical to that advocated by
forces). It is generally agreed the equilibrium models Argyris[lO] in 1964 for the displacement method since
should be acceptable to displacement finite element the “natural” stiffness kN = fNm’ is both invariant and
systems, and with this in mind it is necessary to sum- invaluable when forming a true Eulerian description of
marise element surface stresses as selected nodal stiffness after large rotations.
“forces” PB which must be statically equivalent to (usu- Element surface tractions pS are now integrated and
ally) piecewise integrals of I( ds, so on using (13) collected as a global list of statically equivalent forces R,
we may always obtain an equation of the form at the same nodes as P, so that equilibrium may be
written as
P8 = BNPN +B,,. (14)

Corresponding nodal displacements pS now follow, al- a,‘P, = R,,

Fig. 1.
368 G. A. 0. DAVIES

whence where R, is as before,

Kr = R, + 2 ai(BNfN-‘fO - Bo).

Thus we simulate a displacement model except that the

loading due to surface and body forces is accumulated in
a different way.
We are obliged to have continuous displacements and so
ps = a8r, as in the displacement model, however it is
5. rlNITEELEMENT MIXED METHODS permissible to enforce continuity of P8 whilst allowing
Mixed methods, in which both equilibrium and discontinuous (Dn)a-the stress field may be statically
compatibility are indirectly satisfied in the mean, and non-conforming. Continuity of P8 does not involve the
hybrid methods recently enjoyed a vogue at least kinematically equivalent loads R, so that if we list P=
academic if not commercial. Certainly the number of {P,P, . . . . P8 . . .} the connectivity matrix ba in P8 = b,P
published functionals available for variation became is a sparse Boolean array. We can now sum (15) over all
embarrassingly large although the useful methods are all elements to form
derivable from two work equations as we will show. The
incentive for mixed models arises largely out of the
ditIiculty sometimes encountered in obtaining an element (16)
displacement or stress field which is isotropic and con-
formal (kinematically or statically) and which leads to a where
convenient pattern of nodal unknowns for the chosen
geometry and idealisation. There is also a sporting
chance that a mixed model may be more efficient, having A= z biQsas, F = x b:f,b,, H = 2 b,‘H,.
a mean stiffness somewhere between the upper and
lower bounds of the displacement and equilibrium This mixed formulation has been popular for various
models. plate and shell elements where high order (normal) dis-
Mixed models may be constructed simply using our placement fields are necessary in the displacement
existing two work eqns (9) and (12), with the term i model, but if nodes are confined to the computationally
conveniently dropped since non-zero displacements are attractive corners, this means using displacement
rarely prescribed between elements. The two equations derivatives in pn which are of a higher order than necessary
perform separate duties-of equilibrium and compati- and which produce undesirable force continuity at
bifity-and, apart from the stress-strain law, are linked geometrical discontinuities [ 121. Although (16) allows us
only when the virtual weights ii and ti are selected as the to increase the degrees of freedom above those required
same set of interpolation functions used in the real for continuity in u = up-indeed the forces P improve
variables u and (r. There Seems little point in uniting (9) stress continuity-this approach has not been com-
and (12); however their sum does happen to be entirely mercially popular because it is not acceptable in a dis-
equivalent to the modified Helhnger-Reissner variational placement system. We therefore settle for equilibrium
theorem [ 111.Because ii and 6 are separate weights we continuity in the mean and decouple PB between ele-
now see why our original work eqn (1) cannot deliver ments, so from (15) P8 = f,-‘Q& - fg-‘Hg and, using the
both equilibrium and compatibility equations, as we first term in (15) to sum, we find,
earlier surmised, since to insert both ii and Q into (1)
would leave no real variables to form residuals. Consider
now the simultaneous use of (9) and (12) remembering that (2 aiQ,‘f[‘Q,a.)r = 2 a,‘(R, + fg_‘Hg).
if tractions (or displacements) are discontinuous across
element interfaces then we must drop the weights (Dn)ti Thus the mixed model simulates a displacement element
(of ii) on S,. of stiffness k, = Qif[‘Qg, even though the chosen
Firstly the tractions may be discontinuous, but stresses “forces” and displacements do not correspond, i.e.
will be continuous within the element so we may apply
Pg# h&g.
(2) to transform the interface integral in (9) and our two
Secondly we could form the dual mixed method and
equations become allow discontinuous displacements, drop the second term
in (9) and again transform the surface integral in (12)
v, (pu’ii - a’D’ii) du + p,‘ii ds leaving us with
I I %P
ii’@a+p,)dv+ ii’p, ds
and I “, I %P
and
@‘(E- D’u) du. v (e*e + u’Db) du.
I V, I

We now assume fields for both II= ops and u = @PB Upon expanding u and CTas before we obtain identical
(not a natural stress field) where P, are nodal stresses or expressions to (15) except that now Q8 = -IV, w’DB du.
equivalent measures. The two equations then become We are committed to continuous tractions but suppose
respectively the element is kinematically non-conformal, that is ps =
agr but u = UP, is not continuous. Putting PB = b,P as
B’(Rg - QgfPAand P,‘(f,P, - Q& + HA (15) before and summing we naturally obtain (16) again as a
 Virtualwork 369

global solution. However if we drop continuity of nodal assembly of the stiffness matrix is identical to the mixed
displacements so that fig is only an element variable, we displacement model as we see when we compare (17)
find we cannot eliminate the nodal forces from (15) with (15).
because Q, in general cannot be inverted. It is hardly The hybrid stress model is a useful compromise in
surprising that an element is statically indeterminate but stiffness and continuity between the equilibrium and the
it is tiresome that the duality we have come to expect of mixed formulation. The order of @ should exceed o by
virtual work now fails and we are unable to simulate this at least the number of rigid body modes to avoid rank
mixed model as a displacement element. deficiency, at least at element level. Also the order may
be increased to complete the polynomial and obtain
isotropic results, however it is found that the behaviour
6.FINITEELEMENT HYBRID METHODS of higher order hybrids tends to that of the straightfor-
The hybrid approach is to use (9) and (12) again but ward displacement model of the same order.
to satisfy some equilibrium or compatibility equations In conclusion it should be said that element flexibility
directly, thus some weighted residuals will disappear. inversion is not commercially attractive, and neither is
The various permutations have now become large, global solution of non-positive definite matrices.
considering that there are three equations or conditions However if a mixed formulation of some sort is desired,
of both equilibrium and compatibility inside and on the there seems no reason why virtual work whould not be
surface of an element. The choice of a permutation used to formulate it.
depends very much on the specific structure and the field REFERENCES
idealisation-such as thinwalled structures-so there
1. J. H. Argyris and S. Kelsey, Energy theorems and structural
may be considerable variation in the types of equation analysis. Aircrafr Engng M (1954) and 27 (1955).
and their acceptable errors. Without making a sweeping 2. F. Z. Engesser, Zeitschrift des Architekten und Ing.-
generalisation, it does seem that Pian’s “hybrid stress” is Vereins zu Hannover 35 (1899).
probably most useful[l3]. In this we satisfy internal 3. T. H. H. Pian, Formulations of finite element methods for
equilibrium directly, that is u = /~PN + PO as before, but solid continua. In Recenr Advances in Matrix Methods in
we do not here satisfy interface equilibrium, instead we Structural Analysis. University of Alabama Press (1971).
choose to make interface displacements continuous as in 4. T. H. H. Pian, Finite element methods by variational prin-
ciples with relaxed continuity requirements. Proc. Conf. on
the displacement method. We therefore lose the first
Variational Meth. in Engng, University of Southampton
term in (9) and drop (Dn)6 in (12). It is consistent to (1972).
expand displacements only at the interface and this is 5. B. A. Finlayson and L. E. Striven, The method of weighted
possible since, noting that DC?= 0 in V, we can use (2) to residuals-a review. Applied Mech. Reviews 19 (Sept. 1966).
transform (12) to 6. A. S. L. Ghan, Imperial College, London (private com-
munication, 1973).
7. W. Prager. Variational principles of linear elastostatics for
I,, &E dv -1, u’(Dn)C ds. discontinuous displacements, strains and stresses. Recent
Progress in Applied Mechanics, The Folke Odauist Volume.
Wiley. StocklGlm, pp. 463-474 (1%7). ’
(In practice most hybrid stress models still employ
8. K. Wash& Variational Methods in Elasticity and Plasticity.
boundary values of single element field expansions). The Pergamon Press, London (1968).
work equations now become 9. B. Fraeijs de Veubeke, Displacement and equilibrium models
in the finite element method. Stress Analysis (Edited by 0.
&‘[R, - B,‘PN] and &‘[fNPN + fo- B&J (17) C. Zienkiewicz and G. S. Holister). Wiley (1%5).
IO. J. H. Argyris, Recent Advances in Matrix Methods of Struc-
where this time tural Analysis. Pergamon Press, London (1!%4).
II. T. H. H. Pian and P. Tong, Basis of finite element methods
for solid continua. Int. 1. for Numer. Meth. in Engng 1, 3-28
R, = wtps ds - o’VW3o ds, (1%9).
I %P I SK 12. A. S. L. Chan and V. M. Trbojevic, Thin shell finite element
by the mixed method formulation. Comp. Meth. in Appl.
B,’ = w’(Dn)@ ds, Mech. and Engng, 9.337-367 (1976).
I% 13. T. H. H. Pian, Derivation of element stiffness matrices by
assumed stress distributions, AIAA J. 2, 1333-1336 (1964).
and fw, f. are as in the equilibrium model; however the 14: D. J. Allman, A simple cubic displacement element for plate
bending. RAE. TR 74176 (1974).