Sunteți pe pagina 1din 13

Composites: Part B 32 (2001) 387±399

www.elsevier.com/locate/compositesb

Numerical modeling of composite laminated cylinders in compression


using a novel imperfections modeling method
N.G. Tsouvalis*, A.A. Zafeiratou, V.J. Papazoglou, N.C Gabrielides, P.D Kaklis
Department of Naval Architecture and Marine Engineering, National Technical University of Athens, 9 Heroon Polytechniou Ave.,
15773 Zografou, Athens, Greece
Received 5 June 2000; revised 14 March 2001; accepted 19 March 2001

Abstract
The present study presents the ®nite element modeling procedure of two composite laminated cylinders exhibiting initial geometric
imperfections. Using as input a set of experimental measurements of the cylinder geometry, the application of the skinning method leads to
the analytical representation of the cylinder imperfect internal, external and middle surfaces. A ®nite element mesh is then easily constructed
over these surfaces. The results of the analysis are in very good agreement with the experimental strains and buckling load measurements and
are used to estimate the knockdown effect of the imperfections on the cylinder buckling behaviour. They are also compared to results
obtained by other simpler ®nite element models, in an effort to evaluate the accuracy of various modeling simpli®cations. q 2001 Elsevier
Science Ltd. All rights reserved.
Keywords: B. Buckling; Composites

1. Introduction Composite laminated shell structures are normally quite


susceptible to buckling, due to their small thickness.
Composite materials and especially ®ber reinforced plas- Composite cylinders loaded by axial compressive loading
tics have long ago overcome the condition of being just are a classical example of this kind of structures, met as a
another very attractive construction material and are now structural member in almost all sectors of the civil engineer-
widely used in a plethora of applications, from the civil ing, offshore and aerospace industry. Determination of their
engineering sector to the marine, offshore and aerospace buckling performance, though, is not an easy task. The
industry. Their inherent advantages over the traditional design parameters of a composite laminated shell structure
construction materials (high strength to weight ratio, low are so many and complicated (to mention only the number
maintenance, very good fatigue characteristics, ability to of layers and stacking sequence), that the existing analytical
tailor their mechanical properties) have forced the scienti®c calculation methods cover a very small portion of their
community towards successfully solving the basic manufac- possible combinations, whereas, where they apply, they
turing, structural design and applications problems. Nowa- tend to signi®cantly overestimate buckling loads. In
days, attention is basically given towards expanding their addition, few of them are closed form solutions that can
use to additional, more complicated and challenging appli- be easily integrated into a common design procedure.
cations, as well as optimizing the structural design of The same happens also to the numerical analyses,
composite structures. Hence, among others, buckling failure although in a smaller degree, since parameters like bound-
has become a major concern for any new structural design, ary conditions can be modeled more accurately. Linear
particularly due to the low stiffness of composites with buckling loads calculated using numerical models are also
respect to the other structural materials. The design trends overestimated with respect to the experimental values,
to reduce dimensions through structural optimization and whereas the introduction of geometric nonlinearities in the
avoid stiffening members, since monocoque structures are analysis (large de¯ections) signi®cantly improves the
more cost effective, are additional factors in favour of results. The difference between recorded and numerically
buckling appearance. calculated buckling loads increases, as the deviations of
the actual structure from its nominal characteristics
* Corresponding author. Tel.: 1301-7721413; fax: 1301-7721412. (nominal thickness, nominal stacking sequence, nominal
E-mail address: tsouv@deslab.ntua.gr (N.G. Tsouvalis). material properties, etc.) increase. The cause of these
1359-8368/01/$ - see front matter q 2001 Elsevier Science Ltd. All rights reserved.
PII: S 1359-836 8(01)00023-3
388 N.G. Tsouvalis et al. / Composites: Part B 32 (2001) 387±399

deviations are geometric and material defects, which appear geometry were in the form of internal surface and thickness
very frequently in composite structures, depending on their data, but they can also be in the form of internal and external
manufacturing method. Shape and geometry imperfections, surface data. Moreover, experimental data can be captured
load eccentricities, as well as material faults like voids, ®ber in any mesh over the cylinder surfaces, either regular or
misalignments, resin rich or resin dry areas and especially even irregular. Based on these experimental measurements
delaminations, should be properly modeled, to reach a reli- of the geometry, various FE meshes of the cylinder middle
able numerical prediction of the buckling load. Once such surface are constructed, having different densities. The
modeling procedures are developed and their results vali- buckling load and the deformations time history calculated
dated through successful comparisons to experimental by applying this method are compared to the experimental
measurements, performance of extensive parametric studies recordings, as well as to results obtained by simpler ®nite
is an easy task and very important conclusions can be drawn element models, in order to evaluate the accuracy of various
without carrying out very costly experimental programs. modeling simpli®cations.
The effect of material defects on the buckling perfor-
mance of composite cylinders has been the subject of
several studies [1±4]. The consequent investigation of 2. Cylinder lay-up and geometry
geometrically imperfect cylindrical shells was initially
Among the various glass ®ber reinforced cylinders tested
based on that of isotropic shells [3,5]. Some initial efforts
within the framework of the experimental program
on comparing experimental, numerical and analytical buck-
described in Ref. [10], two are selected here for investiga-
ling loads [6,7] were followed by statistical studies regard-
tion, namely cylinder RVM05A and cylinder RVM07.
ing the effect of manufacturing on the imperfection patterns
These particular cylinders were selected as typical examples
[8] and the numerical calculation of the `knockdown'
of the two cylinder groups tested having different R/t ratios.
factors that have to be applied on the theoretical elastic
Both cylinders were manufactured on a cylindrical mould
critical buckling loads of GFRP curved panels to account
using the hand lay-up method, thus resulting in more intense
for the presence of imperfections [9].
imperfections with respect to other, common to cylindrical
The scope of the present study is twofold. First, to present
parts fabrication methods, such as ®lament winding. Fiber
a new method for modeling the initial geometric imperfec-
reinforcement was E-glass woven roving of type `Rovimat
tions of a composite laminated cylinder and second, to
1200', while the matrix was isophthalic polyester resin of
exploit this method for the construction, and ®nally the
type `Synolite 288'. `Rovimat 1200' had a chopped strand
evaluation, of a ®nite element mesh of the `actual' cylinder.
mat coating on one side, to facilitate proper bonding of the
Geometric imperfections can be of any type, such as
layers. Each layer has a nominal thickness of 1.38 mm and
ovalizations, deviations of the cylinder axis of symmetry
the material properties shown in Table 1 [12]. These proper-
from a straight line and thickness variations. The internal
ties have been experimentally measured by testing two ply
and external surfaces of the imperfect cylinder are modeled
¯at specimens with sequence [08]2. Young's moduli are
in the present paper by two smooth bicubic spline surfaces,
similar in tension and compression.
that interpolate the measurements data. These surfaces are
Cylinder RVM05A has a [08/908] stacking sequence, with
constructed with the aid of the so-called skinning method, a
the 08 layer on the inside and the 908 layer on the outside,
CAGD (Computer-Aided Geometric Design) technique
enabling the reconstruction of three-dimensional bodies Table 1
from cross-sectional data. Such a request arises in a wide Rovimat 1200/Synolite 288 layer properties
spectrum of applications, ranging from reverse engineering
Property Value
(stereolithography) to medical imaging. The obtained
numerical results for the geometry of the cylinder support Young's modulus in 14.7^0.1 GPa
the assertion that, in comparison to any best-®tting weft direction, E1
approach, the present one not only recovers satisfactorily Young's modulus in 13.2^0.4 GPa
warp direction, E2
the shape of the underlying surfaces, but pass through (inter-
Shear modulus, G12 3.5^0.1 GPa
polates) the measured data. Moreover, and in contrast to any Poisson's ratio, n 21 0.18^0.03
scheme relying on double Fourier analysis, the analytic Tensile strength in 199.6^11.5 MPa
structure of the present one (bicubic spline) is in full confor- weft direction, s ut,1
mity with that of the element which is used in the ensuing Tensile strength in 178.5^7.4 MPa
warp direction, s ut,2
®nite element analysis (16-node isoparametric shell
Compressive 249.6^38.9 MPa
element). strength in weft
The above skinning method is applied for the ®nite direction, s uc,1
element modeling of two GFRP hand lay-up composite Compressive 217.1^14.3 MPa
cylinders that have been experimentally tested and numeri- strength in warp
direction, s uc,2
cally modeled in concentric axial compression [10,11]. The
Shear strength, t u,12 103.6^7.1 MPa
available experimental measurements of the cylinder
N.G. Tsouvalis et al. / Composites: Part B 32 (2001) 387±399 389

Fig. 1. Cross section of cylinders RVM05A and RVM07 showing lamination overlaps.

while cylinder RVM07 has a [08/908/08] lay-up. The 08 for the buckling analyses [13]. Both cylinders are modeled
direction is the axial direction of the cylinder, coinciding using 16-node isoparametric shell elements, which have
to the weft direction of the layers. The two layers of cylinder proved their adequacy in modeling axially compressed
RVM05A consist of two separate sheets each, laminated as cylinders in earlier studies. These elements account for
shown in Fig. 1a, so as to form four overlap regions at 908 transverse shear deformation effects, incorporating Midlin's
intervals. Fig. 1b shows the corresponding lamination of formulation and are combined with the laminated material
cylinder RVM07, where again four overlaps at 908 intervals model, allowing modeling of the actual cylinder stacking
are formed, two at the middle layer and one at each of the sequence.
other two layers. Each one of these overlaps is 50 mm wide. Regarding boundary conditions, each cylinder is ®xed at
Both cylinders have a nominal internal radius of 300 mm one of its ends and loaded at its other end by an axial
and a total length equal to 800 mm, including two reinforced compressive force. Rigid links are used at the loaded end,
zones at the ends, 100 mm wide each (Fig. 2). These end rein- connecting all nodes of the end circumference to a central
forcements are added to the cylinder to avoid damage and fail- auxiliary node, on which the compressive force is applied.
ure at the ends of the cylinder during load application and to This auxiliary node has its axial translation and all rotations
ensure that this will happen at its monitored midpart. Accord- free (Fig. 3) to properly model the actual experimental
ing to the authors' experience, as well as other references [11], conditions [10]. The use of rigid links constrains the loaded
for geometries similar to those investigated here, the in¯uence end to remain ¯at and circular throughout the load applica-
of these reinforcements on buckling load is negligible. Thus, tion, maintaining its initial radius.
the clear (unreinforced) length of the cylinder which is Various ®nite element meshes are used in the present
600 mm, is used in all calculations performed. The nominal study. All these meshes are regular, denoted by two
thickness of cylinders (tnom, Fig. 2) is 2.8 mm for RVM05A and numbers C £ L, the ®rst indicating the number of elements
4.1 mm for RVM07, excluding the overlap zones, thus result- in the circumferential direction and the second the number
ing in R/t ratios equal to 108 and 72, respectively. of elements in the axial one. The initial application of an
analytical solution to the problem, often provides valuable
3. Modeling information, not only about the expected buckling load, but
also about the corresponding modeshape. Thus, having an
3.1. Mesh description idea of the expected modeshape, one can properly select the
density of the ®nite element mesh, in order to be able to
The general purpose ®nite element code SOLVIA is used obtain such a de¯ected shape. Although the existing

Fig. 2. Geometry of cylinders RVM05A and RVM07 (dimensions in mm).


390 N.G. Tsouvalis et al. / Composites: Part B 32 (2001) 387±399

other hand, the measuring grid for the thickness was not
regular, being more dense at the layer overlaps (Fig. 1).
Over the normal surface of the cylinder, the thickness
scan was performed on a 50 £ 50 mm grid, whereas at the
overlaps the circumferential spacing was reduced to 25 mm.
Thickness was measured over a cylinder length of 500 mm,
starting 50 mm away from one end and ®nishing 50 mm
before the other. Hence, with a cylinder length equal to
600 mm, the thickness scan covered the central 500 mm
of the specimen. Within the framework of the present
study and in order to analytically construct the full internal
and external surfaces of the cylinders, thickness measure-
Fig. 3. Loading and boundary conditions of the cylinders. ments at the ends of the cylinder were assumed equal to
those at the closest axial station. Note that the provided
analytical closed form solutions have a limited range of experimental data had already been manipulated to account
validity, in the sense that they do not cover all combinations for possible misalignments between the longitudinal axis of
of stacking sequences, boundary conditions, etc., their the specimen and the measuring frame [10].
results are still important as initial estimations. As an exam- In order to apply the ®nite element analysis for any user-
ple, in the present study, application of an analytical speci®ed mesh of the imperfect cylinder, it is necessary to
solution for cylinders with simply supported ends [14], possess a handy analytical representation for both its inter-
resulted in a critical buckling load of 304.4 kN for cylinder nal and external surfaces. These representations are built
RVM05A, with a modeshape consisting of six axial half here with the aid of the skinning method, which is widely
waves (m) and nine circumferential full waves (n). Very used in the area of CAGD for constructing the outer skin of
close was also the second buckling load, having a value of three-dimensional bodies from measurements taken on a
304.8 kN and a respective modeshape (m,n) ˆ (5,9). For family of, usually parallel to each other, axial stations
cylinder RVM07, the calculated critical buckling load was [15,16].
equal to 688.1 kN, with a (5,8) modeshape. Therefore, the Let us assume that we are given with a measurement set:
above results indicate that one needs a ®nite element mesh
[
N
with at least 25 nodes in the axial direction to be able to Dˆ Di ; Di ˆ {…zi ; uj ; rij †; j ˆ 1; ¼; M};
accurately obtain the six half waves and at least 72 nodes in iˆ1
the circumferential direction in order to obtain the nine full
providing the radius rij of the internal surface of the cylinder
waves (considering at least ®ve nodes per half wave). Note
at the axial station z ˆ zi, 0 # z # L, and the circumferential
that the above values should be increased, if one suspects
station u ˆ u j, 0 # u # 2p . Without loss of generality, we
that the buckles are not uniformly distributed over the length
shall henceforth assume that u 1 ˆ 0 and u M ˆ 2p . It is note-
of the cylinder, but are localized in a smaller area (i.e. near
worthy that the skinning method, which will be outlined in
its middle). The coarsest mesh used in this study consisted
the sequel, remains readily applicable also when the number
of 25 elements in the circumferential direction and eight
of circumferential stations varies along the axis of the cylin-
elements in the axial one (25 £ 8). This corresponds to 75
der. Then, a skinning surface that interpolates the data set D,
nodes in the circumference and 25 nodes in the axial direc-
is a vector-valued function Sint …u; z† [ IR3 that can be
tion for the 16-node elements used, values that, as lower
represented as:
limits, are in accordance with the analytical solution indica-
tions. The ®nest mesh used was a 40 £ 16 one. X
N
Sint …u; z† ˆ Li …z†Qi …u†; 0 # z # L; 0 # u # 2p: …1†
iˆ1
3.2. Modeling imperfections
Here Qi …u† are the so-called skeletal lines, i.e., spatial
The raw experimental measurements of the cylinders' curves lying on the axial planes z ˆ zi and interpolating
imperfect geometry [10] included data for their internal the planar point sets …uj ; rij † in the following sense:
surface and their thickness. An automated noncontact
Qi …uj † ˆ …xij ; yij ; zi †T ; j ˆ 1; ¼; M; …2†
laser device operating inside the specimen provided a
three dimensional scanning system for measuring the devia- where xij ; yij are the cartesian coordinates of the planar
tion of the internal shell wall from a perfectly cylindrical, point whose polar coordinates are equal to uj ; rij . Further-
arti®cial reference surface. The internal surface of the speci- more, Li(z) are scalar functions, referred to as the blending
mens was measured along the radial direction in 2232 functions, satisfying the constraints:
points. The measuring grid was perfectly regular, consisting
Li …zk † ˆ dik ; i; k ˆ 1; ¼; N; …3†
of 31 axial stations with an interval of 20 mm, and 72
circumferential stations with an interval of 58. On the d ik denoting the Kronecker's delta. By virtue of Eqs. (2) and (3),
N.G. Tsouvalis et al. / Composites: Part B 32 (2001) 387±399 391

one can readily show that Eq. (1) interpolates the given The construction of the skinning surface Sext …u; z†; that
measurement set D. represents the external surface of the cylinder, requires the
Thus, the analytical structure of the interpolant Sint …u; z†, materiallization of a straightforward preprocessing step.
that represents the internal surface of the cylinder, depends This is due to the fact that we are not given measurements
on the manner the skeletal lines Qi …u† and the blending on the external surface of the cylinder but, instead, measure-
functions Li(z), are to be constructed. In our case, both ments of the
0
wall thickness tij over a set of angles u 0j ,
have been chosen to be twice continuously differentiable j ˆ 1,¼,M , at each
0
axial station z ˆ zi. Note that, neither
(C 2) cubic splines. This choice secures that the resulting the cardinality M nor the angles u 0ij need to coincide with
skinning surface Sint …u; z† possesses suf®cient geometric those used for measuring the internal surface. Then, the
smoothness, namely its tangent plane along with the preprocessing step simply consists in producing the required
Gaussian and mean curvatures vary continuously, while external-surface measurements via the formula:
its degree within each polynomial patch Sint …u; z†,
uj # u # uj11 , zi # z # zi11 , is ®xed and low (bicubic). Sext …u 0j ; zi † ˆ Sint …u 0j ; zi † 1 …tij cosu 0j ; tij sinu 0j ; zi †T ;
The latter also guarantees full compatibility between the …6†
adopted analytical representation of the cylinder boundary i ˆ 1; ¼; N; j ˆ 1; ¼; M 0 :
surfaces with the ensuing ®nite element modeling, that, as
Next, the skinning surface Sext …u; z†, u 01 # u # u 0M 0 ,
stated in the beginning of the previous subsection, uses 16-
0 # z # L, that interpolates the data set in Eq. (6), can be
node isoparametric shell elements, which are indeed
constructed by employing the technique, described just
bicubic.
above, for deriving the internal surface Sint …u; z†. Obviously
In view of the above, computing Sint …u; z† is equivalent
u 0M 0 2 u 01 ˆ 2p; whereas u 01 does not, in general, coincide
to computing the C 2 cubic splines Qi …u† and Li …z†,
with u 1 ˆ 0.
i ˆ 1,¼,N. This is done by endowing the interpolation
Once we possess the analytic representation of both the
constraints (2) and (3) for Qi …u† and Li(z), respectively,
internal, Sint …u; z†, and the external, Sext …u; z†, cylinder
with appropriate boundary conditions, leading to well
surface, we need to evaluate its midsurface Smid …u; z† for,
posed interpolation problems. As for Qi …u†, which are
in the subsequent ®nite element analysis, a shell element is
closed (Qi …0† ˆ Qi …2p†), these conditions are the well
determined by its nodes on Smid …u; z† and the wall thickness
known periodicity conditions:
at each one of these nodes. Taking into account the u -
dr Qi …0† dr Qi …2p† domain de®nitions of Sint …u; z† and Sext …u; z†, we have the
r ˆ ; r ˆ 1; 2; …4† following formula for the intermediate surface:
du d ur
guaranteeing that Qi …u† will be twice continuously differen- 1
tiable over the entire interval [0,2p ]. Then, after selecting a Smid …u; z† ˆ S …u; z†
2 int
suitable basis for the cubic polynomial segments of Qi …u†, ( )
Eqs. (2) and (4) lead to an (M 2 1) £ (M 2 1) linear system 1 Sext …2p 1 u; z† if 0 # u # u 01
1 :
for the weights of the basis functions. Since the matrix of 2 Sext …u; z† if u 01 # u # 2p
this system is positive de®nite, the system can be robustly
…7†
solved with a cost of O(M 3) operations. Detailed expres-
sions for both the left- and the right-hand side matrices of The numerical performance of the adopted skinning method
this system can be found in the Appendix; see Eq. (A.11) is tested against several experimental data sets. A typical
and the ensuing comments therein. example of such a comparisson is shown in Fig. 4, where the
As for the cubic spline blending functions Li(z), deviation of cylinder RVM05A internal surface from its
i ˆ 1,¼,N, we impose the so-called natural boundary nominal radius is plotted versus the circumference angle
conditions: at two different axial stations of the cylinder, at distance
200 and 300 mm from one of its ends. Positive deviations
d2 Li …z1 † d 2 Li …zN †
2
ˆ ˆ 0: …5† are outwards from the nominal radius, while negative devia-
dz dz2 tions are inwards. Fig. 4 con®rms our numerical experience
Then, combining the interpolation constraints of Eq. (3) that, between the experimental points which are by
with the above boundary conditions, and adopting the construction interpolated, the constructed surface exhibits
cubic basis already chosen for the skeletal lines Qi …u†, we a very good behaviour by conforming to the geometry
are led to a linear system for the weights of the basis func- (monotonicity and convexity) information of the data.
tions in all cubic segments of Li(z), 0 # z # L. Structurally, This can be attributed, on one hand, to the low degree
the left-hand side matrix of this system is almost identical to (cubic) of the component splines in representation formula
that of the skeletal lines, the only difference being that now (Eq. 1) and, on the other hand, to the direct geometry signif-
the matrix is tridiagonal as well, which guarantees its robust icance of the parameters z and u appearing therein. The
solution after O(N) computations; see Eq. (A.14) and the same good behaviour is obtained for cylinder RVM07 too.
ensuing comments in the Appendix. Fig. 5 presents a developed view of the cylinder RVM05A
392 N.G. Tsouvalis et al. / Composites: Part B 32 (2001) 387±399

Fig. 4. Comparison of internal surface for cylinder RVM05A (shown developed).

internal surface as constructed after applying the skinning pattern, which is maximum at one end of the cylinder and
method to the experimental data, whereas Fig. 6 shows a gradually becomes less intense towards the other end. The
side view of both internal and external surfaces of this cylin- same happens for cylinder RVM07. The maximum devia-
der. The presence of the four overlaps is obvious in this last tion of this ovalization scheme from the nominal radius is
®gure. All calculated results in Figs. 4±6 are based on a equal to about 50% of cylinder RVM05A nominal thickness
38 £ 12 16-node shell element mesh, consisting of 114 (40% of its average thickness) and about 20% of cylinder
equally spaced nodes in the circumferential and 37 equally RVM07 nominal thickness (16% of its average thickness),
spaced nodes in the axial direction. that is, global shape imperfections are much more intense in
Regarding a qualitative and quantitative evaluation of the the thinner cylinder RVM05A. Regarding thickness imper-
imperfections, Figs. 4 and 5 indicate that the internal surface fections, Table 2 presents the average of the experimental
of cylinder RVM05A presents a characteristic ovalization thickness measurements at the four overlap and the four

Fig. 5. Developed internal surface of cylinder RVM05A.


N.G. Tsouvalis et al. / Composites: Part B 32 (2001) 387±399 393

Fig. 6. Developed internal and external surface of cylinder RVM05A.

normal sectors of the cylinders, along with their coef®cient overlaps and the normal sectors, as in Table 2. Model 2 is of
of variation (numbers in parentheses). It is reminded that course unrealistic, since it does not model the outward
each overlap is approximately 50 mm wide (arc length). In eccentricity of the overlaps with respect to the midsurface.
the same table, the nominal thickness values and the overall However, as it will be shown later, it is very useful as a
average thicknesses are also provided. It is evident from this reference model, since the only imperfections it represents
table that both cylinders are thicker than they were intended are those due to the thickness changes at the overlaps.
to be, especially at their normal sectors. Model 3: Imperfect midsurface, considering the nominal
In order to investigate the in¯uence of the out-of-round- value of thickness (t ˆ 2.8 mm for RVM05A and
ness and thickness variations, as well as how accurate some t ˆ 4.1 mm for RVM07), ignoring overlaps. Model 3
simple models of these variations are with respect to the represents the pure out-of-roundness imperfections of the
actual situation, the following ®nite element models were midsurface.
developed. In all cases the cylinder length was 600 mm. Model 4: Imperfect midsurface, considering a uniform
thickness for the whole cylinder (thus ignoring overlaps),
equal to the overall average (3.5 mm for cylinder RVM05A
Model 1: Perfect midsurface, considering the nominal
and 5.0 mm for cylinder RVM07). Same to model 3, with
value of the cylinder internal radius (Rin ˆ 300 mm) and
different thickness.
the nominal value of thickness (t ˆ 2.8 mm for RVM05A
Model 5: Imperfect midsurface, considering the different
and t ˆ 4.1 mm for RVM07), ignoring overlaps. Model 1
average thicknesses of Table 2 at the overlaps and the
corresponds to the nominal perfect geometry of the two
normal sectors. Model 5 is a very close approximation of
cylinders and is thus used as the basic reference model.
the actual cylinder, since its only assumption is the homo-
For this case, the ®nite element analysis was performed
geneous but different thickness in each sector.
for various meshes.
Model 6: Imperfect midsurface and actual thickness
Model 2: Perfect midsurface, considering the nominal
distribution.
value of the cylinder internal radius (Rin ˆ 300 mm) and
Table 3 summarizes the basic characteristics of the above
taking into account the different average thicknesses at the
six FE models. A comparison of model 1 to model 3 reveals
Table 2 the sole in¯uence of the out-of-roundness imperfections
Thicknesses of the normal and the overlap sectors, in mm on the buckling loads. The same happens by comparing
model 2 to model 5, for the most complicated situation of
Cyl. Sector RVM05A RVM07
eight different thicknesses. The sole in¯uence of thickness
No. 1 (overlap) 4.2 (6.7) a 5.8 (7.3)
No. 2 (normal) 3.1 (7.3) 4.8 (9.8)
Table 3
No. 3 (overlap) 3.9 (14.9) 6.3 (10.9)
No. 4 (normal) 3.2 (9.0) 4.9 (11.6) Basic characteristics of the FE models developed
No. 5 (overlap) 4.0 (7.0) 5.9 (8.8)
Model Midsurface Overlaps Thickness
No. 6 (normal) 3.3 (8.1) 4.7 (5.0)
No. 7 (overlap) 4.6 (6.2) 5.2 (7.2) 1 Perfect Ignored Nominal
No. 8 (normal) 3.1 (7.8) 4.6 (4.3) 2 Perfect Considered Sector average
Overall average thick. 3.5(15.5) 5.0 (13.3) 3 Imperfect Ignored Nominal
Normal nominal thick. 2.8 4.1 4 Imperfect Ignored Overall average
Overlap nominal thick. 4.1 5.5 5 Imperfect Considered Sector average
a 6 Imperfect Considered Actual distribution
Numbers in parentheses indicate the coef®cient of variation.
394 N.G. Tsouvalis et al. / Composites: Part B 32 (2001) 387±399

imperfections is revealed by comparing model 2 to model 1 Table 4


and model 5 to model 3. Finally the results of models 4 and Eigenvalue buckling loads for models 1 and 2, in kN
5 show the accuracy of their assumptions, when compared Geometry Mesh (C£L) RVM05A RVM07
to the results of the closest to reality model 6.
Model 1 Analytical [14] 304.8 688.1
25£8 ˆ 200 307.2 679.4
24£10 ˆ 240 308.0 681.2
4. Results 36£12 ˆ 432 300.3 674.5
38£12 ˆ 456 300.0 674.1
A linearized buckling analysis was performed for the 40£16 ˆ 640 299.6 673.6
Model 2 38£12 ˆ 456 408.3 889.2
models with perfect midsurface, whereas a geometrically
nonlinear buckling analysis was performed for all models.
The ®rst type of analysis leads to the critical buckling load
modeshape, compared to the (5 £ 8) one predicted by the
and its associated modeshape of the cylinder, whereas the
closed form solution.
second one provides the load-de¯ection curve, strain and
Table 4 shows that the analytical solution predicted the
stress histories of the cylinder and the consequent nonlinear
buckling load very accurately, since its difference from the
buckling load. For cylinder RVM05A, an experimental
best ®nite element result (that corresponding to the ®nest
buckling load equal to 290 kN was measured, accompanied
mesh) is only about 2%. The same Table indicates also that
by local buckling at midlength of the cylinder and an
there is very good convergence of the results for the model 1
indication of twisting [10]. For the thicker cylinder
meshes analyzed. Therefore, taking into account that in
RVM07, a failure load equal to 704 kN was recorded,
order to properly model geometry models 2 and 5 with differ-
with signi®cant material failure accompanying buckling
ent thickness at the overlaps, a circumferential subdivision
[11]. It was not clear from the test whether local buckling
which is at least equal or submultiple of the overlap width is
resulted in material failure or vice versa, but the ®nal
needed, the 38 £ 12 mesh was selected for further consid-
load-axial shortening recording is in favour of the ®rst
eration in the remaining eigenvalue and nonlinear analyses.
option, since almost all of the cylinder load carrying capa-
Table 4 presents also the linearized eigenvalue buckling
city was lost suddenly, without any progressive evolution of
loads for model 2. The increased average thicknesses with
the failure [10].
respect to the nominal values, as well as the stiffening effect
of the overlaps which are taken into account in model 2,
4.1. Eigenvalue analysis result in model 2 eigenvalue buckling loads which are
increased about 30±35% with respect to the model 1
The effect of mesh magnitude was investigated ®rst, by loads. Signi®cant differences appear also in the mode-
performing eigenvalue analyses of model 1 for various shapes. Whereas model 1 buckles with a normal modeshape
meshes. The results are shown in Table 4 for both cylinders, as shown in Fig. 7 for cylinder RVM05A, the modeshape of
along with the analytical prediction. All model 1 meshes of model 2 presents local buckling deformations, more concen-
cylinder RVM05A result in a (5 £ 9) modeshape, whereas trated at the midlength of the cylinder and at that portion of
the analytical solution predicted a (6 £ 9) one. For cylinder its circumference having minimum thickness. This can be
RVM07, all model 1 meshes of Table 4 result in a (4 £ 8) seen in Fig. 8, presenting the critical buckling modeshape

Fig. 7. Buckling modeshape of cylinder RVM05AÐmodel 1.


N.G. Tsouvalis et al. / Composites: Part B 32 (2001) 387±399 395

Fig. 8. Buckling modeshape of cylinder RVM05AÐmodel 2.

for model 2 of cylinder RVM05A. Similar results were model 1 to model 2 and model 3 to model 5 for both cylin-
obtained for cylinder RVM07 too. ders. The resulting increase of buckling loads for models 2
The model 2 results of Table 4 are in good agreement and 5 varies between 35 and 38% and is due to the increased
with other relative results in Ref. [11], the ones presented thickness and the stiffening effect of the overlaps, which are
here being closer to the experimental buckling loads. The taken into account in these models. This stiffening effect
small differences noticed, are justi®ed by the slightly differ- remains the same regardless of whether the midsurface is
ent FE models considered in the two studies. geometrically perfect or not. A comparison between model
3 and model 4 veri®es the expected result drawn out also in
4.2. Nonlinear analysis Ref. [11], that the critical buckling load of composite cylinders
under compressive loading is proportional to the square of
In an attempt to evaluate the validity of each geometry
their thickness, just as in the case of isotropic cylinders.
model, the nonlinear analysis was performed for all models.
The predictions of the closest to reality model 6 are very
The nonlinear buckling loads were obtained using the
good, differing from the experimental buckling loads by
following process: a large deformation nonlinear analysis
only 4% for both cylinders. Relatively good are also the
was ®rst performed, with a linearly increasing load and an
predictions of model 5, which results in differences of
automatically adjusted time step. For various time steps of
14% for cylinder RVM05A and only 4% for cylinder
this analysis, the model data accumulated until the current
RVM07. This model is simpler than model 6 and therefore
time step was saved. Then, using as input (i.e. as initial
more ease to develop, since it does not account for the actual
conditions) these saved data, an eigenvalue analysis was
thickness distribution, but only for uniform sector average
performed to obtain the buckling load. This procedure was
thickness distributions. All other models do not give results
repeated by performing the eigenvalue analysis at various
of acceptable accuracy with respect to the experimental
steps of the nonlinear analysis, until convergence of the
measurements, since it appears that, for the cylinders exam-
buckling load was obtained. For the performance of the
ined, the effect of both the imperfections of the midsurface
nonlinear analysis of model 1 an arti®cial initial imper-
and the overlaps are too strong to be neglected.
fection is needed to initiate buckling, since model 1 is
geometrically perfect. Thus, a deformation scheme similar
Table 5
to the ®rst eigenvalue buckling modeshape was considered, Experimental and nonlinear buckling loads, in kN
having a pre-de®ned maximum value equal to 0.001 £ t.
Table 5 presents the results of all nonlinear calculations, Geometry RVM05A RVM07
together with the experimentally measured buckling loads.
Model 1 292.7 654.7
Comparing model 1 to model 3 and model 2 to model 5, we Model 2 404.5 878.8
see a reduction of the buckling load in both cylinders, owing Model 3 245.4 534.9
solely to the imperfections of the midsurface. It is important Model 4 382.3 826.0
to note that this reduction is about the same for both cylin- Model 5 331.8 734.7
Model 6 301.9 728.0
ders and both comparisons, varying from 216 to 218%.
Experiment Refs. [10,11] 290.0 704.0
The same consistent behaviour is found, when one compares
396 N.G. Tsouvalis et al. / Composites: Part B 32 (2001) 387±399

Fig. 9. Axial shortening versus load for cylinder RVM05AÐmodel 6.

The cylinder axial shortening predicted by model 6 is local buckling of cylinder RVM07 might be the initial
plotted against the experimental measurements in Fig. 9 cause of the observed material failure.
for cylinder RVM05A and in Fig. 10 for cylinder RVM07. A material failure analysis was performed for both cylin-
The comparison is quite good, indicating that the axial stiff- ders, calculating the Tsai-Hill failure criterion value in
ness of the cylinders was accurately modeled. Good corre- various steps of the nonlinear response for model 6. At the
lation also exists between the calculated and measured nonlinear buckling load level, the maximum criterion values
strains. Figs. 11 and 12 present the axial and circumferential were found to be quite low, equal to 0.22 for cylinder
strains at the cylinder midlength, plotted versus the applied RVM05A and 0.43 for cylinder RVM07, thus indicating
load for cylinders RVM05A and RVM07, respectively. no material failure before buckling failure. These results
Numerical strains are average values over the circumference show that stress levels in cylinder RVM07 are higher than
of the cylinders, whereas the experimental data were taken those in cylinder RVM05A, thus making the buckling load of
in three positions, spaced 1208 between each other. The the former cylinder closer to its material failure load than the
clear minimum exhibited by one of the axial strains experi- respective relation in the latter cylinder. It must be noticed,
mental curves near the failure load for cylinder RVM07 however, that the calculated failure criterion values increase
(Fig. 12) strengthens the previously expressed view that very fast near the buckling region of the nonlinear analysis.

Fig. 10. Axial shortening versus load for cylinder RVM07Ðmodel 6.


N.G. Tsouvalis et al. / Composites: Part B 32 (2001) 387±399 397

Fig. 11. Axial and circumferential strains versus load for cylinder RVM05AÐmodel 6.

5. Conclusions The results obtained from the various ®nite element


models developed are very useful in many ways. First
The present study and its results reveal ®rst that the use of they provide credibility to the modeling procedure followed,
an analytical representation of an imperfect surface for since they agree very well with the experimental measure-
developing an as accurate as possible ®nite element mesh ments. Second, they make evident that a nonlinear analysis
is a very useful analysis tool, leading to a very good predic- is needed to come up with an accurate prediction of the
tion of the buckling behaviour, compared to the experimen- buckling load. Additionally, the present results give
tal data. The particular computer aided geometric design valuable information about the degree of accuracy of each
method used here, the skinning method, proved its validity modeling, starting from the more simple (model 1) and
and can be safely used further as a parametric study tool, proceeding to the most complex (model 6).
producing surfaces with predetermined types and sizes of Apart from the above, a comparison between eigen-
geometric imperfections. value results for models with perfect geometry and the

Fig. 12. Axial and circumferential strains versus load for cylinder RVM07Ðmodel 6.
398 N.G. Tsouvalis et al. / Composites: Part B 32 (2001) 387±399

experimentally measured buckling loads leads to the estima- and:


tion of the knockdown factors. These factors express the
Li …zk † ˆ dik ; i; k ˆ 1; ¼; N; …A:2†
reduction of the buckling load due to the geometric (and
possibly the material) imperfections. In the present case, respectively; see Eqs. (2) and (3). In addition, appro-
the complete perfect geometry is expressed only by model priate boundary conditions have to be ful®lled, namely
1, which has no midsurface imperfections and a uniform periodic boundary conditions for the skeletal lines:
constant thickness. However, the resulting knockdown
dr Qi …0† dr Qi …2p†
factors of 1.0 for cylinder RVM05A and 1.04 for cylinder ˆ ; r ˆ 1; 2; i ˆ 1; ¼; N; …A:3†
RVM07 are completely misleading and should not be taken d ur d ur
into account, since model 1 is not at all representative of the and natural boundary conditions for the blending func-
actual cylinders geometry. The two basic simpli®cations of tions:
model 1, namely the uniform nominal thickness which is
20±25% less than the actual thickness and the absence of d2 Li …z1 † d2 Li …zN †
ˆ ˆ 0; i ˆ 1; ¼; N; …A:4†
overlapping, signi®cantly in¯uence the cylinder buckling dz2 dz2
behaviour, leading to lower eigenvalue buckling loads and see Eqs. (6) and (7).
thus compensating the buckling load reduction due to the We shall now show that, combining Eqs. (A.1) and (A.3)
actual midsurface and thickness imperfections. Therefore, with our continuity (C 2) and degree (cubic) assumptions,
the agreement of experimental data and model 1 results in Qi …u† can be fully determined at the computational cost of
the present case is accidental and misleading and the result- solving a linear system. Naturally, an analogous result holds
ing knockdown factors are meaningless. On the contrary, true for Eqs. (A.2) and (A.4) that are associated with the
model 2 eigenvalue results can be safely used for the esti- blending functions Li(z).
mation of the knockdown factors, since in this model, both To start with, since Qi …u†, 0 # u # 2p , is assumed to be a
perfect geometry is preserved (no midsurface imperfections C 2 cubic spline, its restriction on the interval [uj ; uj11 ] is a
and no thickness imperfections within each sector) and the cubic polynomial that can be represented as:
stiffening effect of the overlaps is taken into account. The
resulting knockdown factors of 0.71 for cylinder RVM05A Qi …u† ˆ …1 2 t†aij 1 tbij 1 cij F…1 2 t† 1 dij F…t†;
and 0.79 for cylinder RVM07 are very typical for problems …A:5†
of this type and additionally very close to other relevant uj # u # uj11 ;
results [10]. where t ˆ …u 2 uj †=hj , hj ˆ uj11 2 uj , is the so-called local
variable, ranging in [0,1], while F…t† ˆ …t3 2 t†=6.
Obviously, since the function set {…1 2 t†; t; F…1 2
Acknowledgements
t†; F…t†} comprises four linearly independent polynomials
of degree at most three, it constitutes a basis in the space of
The results presented in this paper are based on work
cubic polynomials. Next, we set:
carried out within the framework of EU Contract No.
BEU2-CT94-0926 `Design and Validation of Imperfec-  ij ; dij ˆ h2j Q
aij ˆ Iij ; bi;j11 ˆ Ii;j11 ; cij ˆ h2j Q  i;j11 ; …A:6†
tionÐTolerant Shell Structures', whose ®nancial support
is thankfully acknowledged. The authors would like to where the dot denotes differentiation with respect to u and:
thank all partners in this project and especially the collea- 2
gues from Imperial College of Science, Technology and  ij ˆ d Qi …uj †
Q …A:7†
Medicine, Dept. of Civil and Environmental Engineering, d u2
for providing all experimental data. Substituting Eq. (A.6) into Eq. (A.5), we arrive at the
following representation for:

Appendix A. Determining the skeletal lines and the  ij F…1 2 t† 1 h2j Q


Qi …u† ˆ …1 2 t†Iij 1 tIi;j11 1 h2j Q  i;j11 F…t†;
blending functions of a C 2 bicubic interpolating skinning
surface uj # u # uj11 ; (A.8)

As discussed in Subsection 3.2, the measurement set D is which ful®lls the interpolation conditions (Eq. A.1) (C 0
interpolated by the skinning surface Sint …u; z† (Eq. (1)), continuity) and, moreover, ensures continuity between the
provided that the tensor-product components of the latter, left- and the right-hand second-order nodal derivatives, i.e.:
namely its skeletal lines Qi …u†, 0 # u # 2p, and blend-  i …uj 1† ˆ Q
 i …uj 2† ˆ Q  ij :
Q …A:9†
ing functions Li(z), 0 # z # L, ful®l the interpolation
constraints: 2
Then, in order to achieve C continuity, it suf®ces to
ensure ®rst-order derivative continuity at the nodes, that is:
Qi …uj † ˆ Iij ˆ …xij ; yij ; zi †T ; j ˆ 1; ¼; M; i ˆ 1; ¼; N;
…A:1† Q _ i …uj 1†; j ˆ 1; ¼; M:
_ i …uj 2† ˆ Q …A:10†
N.G. Tsouvalis et al. / Composites: Part B 32 (2001) 387±399 399

Substituting Eq. (A.8) into Eq. (A.10) and taking into The matrix of the above system is tridiagonal, symmetric
account the boundary conditions (Eq. A.3) for r ˆ 1, we and strictly diagonally dominant. Since it is tridiagonal, it
arrive at the following linear system for the second-order admits a very simple factorization enabling its solution at a
nodal derivatives Q ij ; j ˆ 1; ¼; M; of the i-th skeletal line cost of 3(N 2 3) additions, 3(N 2 3) multiplications and
Qi …u†: 2(N 2 2) divisions, which constitutes a considerable reduc-
tion in comparison with the computational cost of the
hM21 1 h1  h  h  i;M21 ˆ bi1 2 biM ; Cholesky factorization. Obviously, the method will be
Qi1 1 1 Q 1 M21 Q
3 6 i2 6 also robust, for the matrix is positive de®nite.
hj21  hj21 1 hj  hj 
Qi;j21 1 Qij 1 Q ˆ bi;j11 2 bij ;
6 3 6 i;j11
References
j ˆ 2; ¼; M 2 2
[1] Poggi C, Taliercio A, Capsoni A. Fibre orientation effects on the
hM21  h  i;M22 1 hM22 1 hM21 Q
 i;M21 buckling behaviour of imperfect composite cylinders. In: Jullien JF,
Qi;1 1 M22 Q editor. Buckling of shell structures, on land, in the sea and in the air.
6 6 3
London/New York: Elsevier Applied Science, 1991. p. 114±123.
ˆ biM 2 bi;M21 (A.11) [2] Davies P, Rannou F. The effect of defects in tubes: Part 1. Mode I
delamination resistance. Applied Composite Materials 1995;1:333±49.
[3] Simitses GJ. Buckling and postbuckling of imperfect cylindrical
where: shells: A review. Applied Mechanics Reviews 1986;10:1517±24.
[4] Simitses GJ, Chen ZQ, Sallam S. Delamination buckling of cylind-
Iij 2 Ii;j21 rical laminates. Thin Walled Structures 1991;11:25±41.
bij ˆ …A:12† [5] Chryssanthopoulos MK, Baker MJ, Dowling PJ. Imperfection model-
hj21
ing for buckling analysis of stiffened cylinders. Journal of Structural
Engineering 1998;117(7):2017.
The matrix of system (Eq. A.11) is cyclic tridiagonal,
[6] Giavotto V, Poggi C, Chryssanthopoulos M. Buckling of imperfect
symmetric and strictly diagonally dominant. The latter composite shells under compression and torsion. Proceedings of the
two properties induce that the matrix is positive de®nite, American Helicopter Society Intl Meeting on Rotorcraft Basic
which in turn implies that (Eq. A.11) can be solved robustly Research, Atlanta, Georgia 1991:43-1±43-13.
via the Cholesky factorization. This method requires [7] Li YW, Elishakoff I, Starnes JH, et al. Effect of the thickness variation
and initial imperfection on buckling of composite cylindrical shells:
(M 2 1) 3/6 additions, (M 2 1) 3/6 multiplications,
2 Asymptotic analysis and numerical results by BOSOR4 and
(M 2 1) /2 divisions and (M 2 1) root extractions, that PANDA2. International Journal of Solids and Structures
compares very favourably with the (M 2 1) 3/3 additions, 1997;34(28):3755±67.
(M 2 1) 3/3 multiplications and (M 2 1) 2/2 divisions of the [8] Chryssanthopoulos MK, Giavotto V, Poggi C. Statistical imperfection
Gaussian elimination. models for buckling analysis of composite shells. In: Jullien JF,
editor. Buckling of shell structures, on land, in the sea and in the
As for the cubic-spline blending functions Li(z), we adopt,
air. London/New York: Elsevier Applied Science, 1991. p. 43±52.
in direct analogy with the skeletal lines Qi …u†, the following [9] Chryssanthopoulos MK, Kim KD, Dowling PJ. Buckling behaviour of
representation: composite GFRP cylindrical panels under axial compression. In: Patel
MH, Gibbins R, editors. Proceedings of 6th International Conference
Li …z† ˆ dik …1 2 s† 1 di;k11 s 1 z2k L 00ik F…1 2 s† 1 z2k L 00i;k11 F…s†; on the Behaviour of Offshore Structures. London: Bentham Press,
1992. p. 733±43.
zk # z # zk11 ; (A.13) [10] Elghazouli AY, Chryssanthopoulos MK, Esong IE. Buckling of
woven GFRP cylinders under concentric and eccentric compression.
Composite Structures 1999;45:13±27.
where s ˆ …z 2 zk21 †=zk , zk ˆ zk11 2 zk and L 00ik denotes
[11] Chryssanthopoulos MK, Elghazouli AY, Esong IE. Validation of FE
the value of the second-order derivative L 00i …z† of Li(z) models for buckling analysis of woven GFRP shells. Composite
at the node z ˆ zk. Then, appealing to the interpolation Structures 2000;49:355±67.
constraints (Eq. A.2), the boundary conditions (Eq. A.4) [12] Poggi C. Characterization of materials. BRITE±EURAM Project
(L 00i1 ˆ L 00iN ˆ 0) and working as in the case of the skele- DEVILS, Report No. WP04.DR/PM(1), Politecnico di Milano,
Italy, October 1996.
tal lines, we arrive at the following (N 2 2) £ (N 2 2) linear
[13] SOLVIA-PRE 95.0, Users manual for stress analysis. Report SE 95-1,
system for the interior second-order nodal derivatives L 00ik , Solvia Engineering AB, Sweden, April 1996.
k ˆ 2,¼,N 2 1, of each one of the blending function Li(z): [14] Vinson JR. The behavior of shells composed of isotropic and compo-
site materials. The Netherlands: Kluwer Academic Publishers, 1993.
zk21 00 z 1 zk 00 z
L 1 k21 L ik 1 k L 00i;k11 [15] Woodward CD. Skinning techniques for interactive B-spline surface
6 i;k21 3 6 interpolation. CAD 1988;20(8):441±51.
[16] Wu SC, Abel J, Greenberg D. An interactive computer graphics
di;k11 2 dik d 2 di;k21
ˆ 2 ik ; k ˆ 2; ¼; N 2 1: …A:14† approach to surface representation. Communications ACM
zk zk21 1977;20(10):703±12.

S-ar putea să vă placă și