Effective Simulation of Delamination in Aeronautical

Structures Using Shells and Cohesive Elements

Carlos G. Dávila1
NASA Langley Research Center, Hampton, VA, 23602

Pedro P. Camanho2
University of Porto, Porto, Portugal

and
Albert Turon3
University of Girona, Girona, Spain

A cohesive element for shell analysis is presented. The element can be used to simulate
the initiation and growth of delaminations between stacked, non-coincident layers of shell
elements. The procedure to construct the element accounts for the thickness offset by
applying the kinematic relations of shell deformation to transform the stiffness and internal
force of a zero-thickness cohesive element such that interfacial continuity between the layers
is enforced. The procedure is demonstrated by simulating the response and failure of the
Mixed Mode Bending test and a skin-stiffener debond specimen. In addition, it is shown that
stacks of shell elements can be used to create effective models to predict the inplane and
delamination failure modes of thick components. The results indicate that simple shell
models can retain many of the necessary predictive attributes of much more complex 3D
models while providing the computational efficiency that is necessary for design.

Nomenclature

a = crack length
A = area of integration of cohesive element
B = strain-displacement matrix
C = constitutive matrix
D = damage matrix
d = damage variable
e = length of lever arm in MMB test
F = applied force
FShell = internal force of shell cohesive element
F3D = internal force of zero-thickness cohesive element
GI, GII, GIII = mode I, II, and III energy release rates
Gc = critical energy release rate
GIc, GIIc = critical energy release rates in mode I, II
GT = total energy release rate
I = identity matrix
J = J-integral
K = penalty stiffness for cohesive law
KShell = stiffness matrix of shell cohesive element

1
Senior Aerospace Engineer, Durability, Damage Tolerance, & Reliability Branch, MS 188E.
2
Assistant Professor, DEMEGI- Faculdade de Engenharia.
3
Assistant Professor, Dept. d'Enginyeria Mecànica i de la Construcció Industrial.
K3D = stiffness matrix of zero-thickness cohesive element
lele = element length
Rij, R = shell transformation matrices
S = stress tensor
s, t = pseudo time factor
tF = pseudo time factor at the maximum load
tj = thickness of j-th ply
t u, t l = thickness of the plies above and below an interface
U 3j D , U j = vector of degrees of freedom at a node for 3D and shell cohesive elements
Shell
U3D , U Shell = vector of all degrees of freedom of 3D and shell cohesive elements
uj, vj, wj, θ xj , θ yj = translational and rotational nodal degrees of freedom
αij = coefficients of thermal expansion
Δ 0
= displacement jump for damage initiation
Δf = final displacement jump
Δ 3 , Δ1 , Δ 2 = mode I, mode II, and mode III displacement jumps
Δ shear = tangential displacement jump
ΔT = difference between the curing temperature and room temperature
δ ij = Kronecker delta
λ = equivalent displacement jump
ρ = viscoelastic stabilization factor
η = parameter of Benzeggagh-Kenane mixed-mode fracture criterion
τ, τi = interfacial tractions
τ 30 , τ shear
0
= nominal peel and shear strengths
τ 3* , τ shear
*
= reduced peel and shear strengths

I. Introduction

Delamination is one of the predominant modes of failure in laminated composites when there is no

reinforcement in the thickness direction. In order to design structures that are flaw- and damage-tolerant, analyses

must be conducted to study the propagation of delaminations. The calculation of delaminations can be performed

using cohesive elements1, 2, which combine aspects of strength-based analysis to predict the onset of damage at the

interface, and fracture mechanics to predict the propagation of a delamination.

Cohesive elements are normally designed to represent the separation at the zero-thickness interface between layers

of three-dimensional elements (3D). Unfortunately, the softening laws that govern the debonding of the plies often

result in extremely slow convergence rates for the solution procedure, and require very refined meshes in the region

ahead of the crack tip. In addition, it is often difficult to model zero-thickness interfaces using conventional finite

element pre-processors.

2
 The typical requirement of extremely fine meshes ahead of the crack tip1, 3, 4 has been recently relaxed by a

procedure that artificially reduces the interfacial strength while keeping the fracture energy constant in order to

increase the length of the cohesive zone. A longer cohesive zone allows the use of larger cohesive elements3.

However, the difficulties associated with the generation of complex meshes composed of stacks of 3D elements

joined together with zero-thickness cohesive elements often render the use of cohesive elements impractical outside

research environments.

Shell elements are more efficient for modeling thin-walled structures than 3D elements. Shell finite element

analysis has been used in conjunction with the Virtual Crack Closure Technique (VCCT) by Wang et al.5, 6
to

evaluate strain energy release rates for the damage tolerance analysis of skin-stiffener interfaces. Wang used a wall

offset to move the nodes from the reference surfaces to a coincident location on the interface between the skin and

the flange. Therefore, compatibility was ensured by equating the three translational degrees of freedom of both

shells with constraint equations and by allowing the rotations to be free. These models were shown to require a

fraction of the degrees of freedom than are needed for full 3D analyses while producing accurate values for mode I

and mode II strain energy release rates. However, Glaessgen et al.7 observed that shell VCCT models of debond

specimens in which the adherends are of different thickness predict the correct total energy release rate but that the

mode mixity does not converge with mesh refinement. This phenomenon is similar to the well known oscillatory

behavior of the in-plane linear elastic stress field near the tip of a bi-material interfacial crack8.

Very few references discuss the use of shell models in combination with cohesive elements and most of them are

designed for explicit time integration. A shell model for discrete delaminations under mode I loading using finite-

thickness volumetric elements connecting shell elements was proposed by Reedy9. Johnson et al.10 applied the

bilinear traction-displacement of the Crisfield11 cohesive model in conjunction with the Ladevèze mesomodel12 for

intralaminar damage to investigate wing leading edge impact and crash response of composite structures. A mixed-

mode adhesive contact shell element that accounts for the thickness offset and the rotational degrees of freedom in

the shells was developed by Borg13 for the explicit finite element code LS-DYNA®. Bruno et al.14 developed shell

models that use rigid links to offset the location of the cohesive element nodes from the shell reference surface to the

interface between the layers. Using this technique, multi-layered shell models were analyzed using the zero-

thickness cohesive elements available in the commercial finite element code LUSAS®.

3
 Cohesive elements have been found useful to study fracture along bi-metallic interfaces15-18. Cohesive laws

eliminate the stress singularity at the delamination front and, consequently, they prevent the oscillation in the stress

field and the corresponding non-convergence of fracture mode mixity18. Therefore, it is expected that the non-

convergence of the mode mixity observed by Glaessgen in shell VCCT models, which is also caused by a mismatch

of the elastic stiffnesses of the adherends, is also avoided in cohesive models. However, a study of the mesh

independence of mode mixity in cohesive shell models has not been attempted by the present authors.

The objective of this work is to present a method for the construction of shell cohesive elements for implicit

analysis. The formulation of an 8-node cohesive element for modeling delamination between shell elements on the

basis of a 3D cohesive element previously developed by the authors1, 2

is proposed. The formulation takes into

account the rotational degrees of freedom at the shell nodes and the relative location of the interface which allows

the construction of models with more than one layer of cohesive interfaces and does not require coincident nodes

along an interface nor does it require the use of rigid links. Therefore, the formulation allows the use of several

layers of shell/cohesive elements through the thickness of a thick laminate. Since the cohesive elements have the

topology of a standard 8-node element, the models are simpler to generate than those based on zero-thickness

interfaces.

In the following sections, the constitutive damage model previously developed by the authors will be outlined.

Then, a procedure to construct a shell cohesive element from that of any conventional zero-thickness cohesive

element will be described. Finally, three different examples of increasing structural complexity will be presented to

show that simple shell models based on the proposed cohesive element for shells can be used to represent the onset

and propagation of delamination in structural components.

II. Bilinear Cohesive Law

A cohesive constitutive law relates the traction, τ, to the displacement jumps, Δ, at the interface. The bilinear

softening model, which is chosen here for its simplicity, is shown in Fig. 1. One characteristic of all softening

models is that the cohesive zone can still transfer load after the onset of damage (Δ0 in Fig. 1). For pure mode I, II or

III loading, after the interfacial normal or shear tractions attain their respective interlaminar tensile or shear

strengths, the stiffnesses are gradually reduced to zero. The area under the traction-displacement jump curves is the

4
respective (mode I, II or III) fracture energy. Using the definition of the J integral, it can be shown that for small

cohesive zones19:

f
Δ
∫ 0 τ (Δ) dΔ = Gc (1)

where Gc is the total area under the traction-displacement law, is the critical energy release rate for a particular

mode, and Δ f is the displacement jump at failure, shown in Fig. 1. The penalty stiffness K is an arbitrarily large

number selected such that the presence of undamaged cohesive elements does not introduce appreciable compliance

to the structure3.

A bilinear cohesive law is assumed for each of the three fracture modes. It is assumed that the 3 direction is

normal to the interface and that the interlaminar shear strength τ shear
0
is independent of the shearing direction. Then,

the displacement jumps for damage initiation in each mode are simply

τ 30
Δ03 =
Mode I: K

τ shear
0
Δ02 =
Mode II: K (2)

τ shear
0
Δ01 =
Mode III: K

Similarly, the final displacement jumps are proportional to their corresponding toughness Gic as

GIc GIIc GIIIc

Δ 3f = 2 , Δ1f = 2 , and Δ 2f = 2 (3)
τ 30 τ shear
0
τ shear
0

Fig. 1. Bilinear cohesive law model.

5
A. Cohesive Law for Mixed-Mode Delamination

In structural applications of composites, delamination growth is likely to occur under mixed-mode loading.

Therefore, a general formulation for cohesive elements must deal with mixed-mode delamination growth problems.

Under pure mode I, II or III loading, the onset of damage at the interface can be determined simply by

comparing the tractions with their respective allowable. Under mixed-mode loading, however, damage onset may

occur before any of the traction components involved reach their respective allowable. Therefore, a mixed-mode

criterion must be established in terms of an interaction between components of the energy release rate.

The constitutive damage model used here, formulated in the context of Damage Mechanics (DM), was

previously proposed by the authors2. The damage model simulates delamination onset and delamination propagation

using a single scalar variable, d, to track the damage at the interface under general loading conditions. An initiation

criterion that evolves from the Benzeggagh–Kenane fracture criterion20 (B-K) ensures that the model accounts for

changes in the loading mode in a thermodynamically consistent way and avoids restoration of the cohesive state.

The constitutive model prevents interpenetration of the faces of the crack during closing, and a Fracture

Mechanics-based criterion is used to predict crack propagation. The parameter λ is the norm of the displacement

jump tensor (also called equivalent displacement jump norm), and it is used to compare different stages of the

displacement jump state so that it is possible to define such concepts as ‘loading’, ‘unloading’ and ‘reloading’. The

equivalent displacement jump is a non-negative and continuous function, defined as:

+ (Δ shear )2
2
λ= Δ3 (4)

where < ⋅ > is the MacAuley bracket defined as < x >= 12 (x + x ) , which sets any negative values to zero. The term Δ3

is the displacement jump in mode I, i.e., normal to the mid-plane, and Δshear is the tangential displacement calculated

as the Euclidean norm of the displacement jump in mode II and in mode III:

+ (Δ 2 )2
2
Δ shear = Δ1 (5)

A bilinear cohesive law for mixed-mode delamination can be constructed by determining the initial damage

threshold Δ0 from the criterion for damage initiation and the final displacement jump, Δf, from the formulation of the

6
propagation surface or propagation criterion2. For the B-K fracture criterion, the mixed-mode displacement jump for

damage initiation is2

η
Δ0 = (Δ )
0 2
3 (
+ ⎛⎜ Δ0shear
⎝
) − (Δ )
2 0 2 ⎞⎛ GII
3 ⎟⎜ ⎜
⎠⎝
+ GIII ⎞
⎟⎟ (6)
GT ⎠

where the B-K parameter η is obtained by curve-fitting the fracture toughness of mixed-mode tests and the fracture

mode ratio is2

⎛ GII + GIII ⎞ β2
⎜ ⎟= (7)
⎜ G ⎟ 1 + 2β 2 − 2β
⎝ T ⎠

and where the displacement jump ratio is defined as β = Δ shear / (Δ shear + Δ 3 ) .

The displacement jump for final fracture is also obtained from the critical displacement jumps as

η
⎛ G + GIII ⎞
Δ03Δ 3f + (Δ0shear Δ fshear − Δ03Δ 3f )⎜⎜ II ⎟⎟
Δf = ⎝ GT ⎠ (8)
Δ0

During overload, the state of damage d is a function of the current equivalent displacement jump λ :

d=
( )
Δ f λ − Δ0
λ (Δ − Δ )f 0
(9)

The corresponding tractions can be written as

⎡ ⎛ − Δ j ⎞⎟⎤
τ i = Dij Δ j = δ ij K ⎢1 − d ⎜⎜1 + δ 3 j ⎥Δ j (10)
⎢ Δ j ⎟⎥
⎣ ⎝ ⎠⎦

where the Kronecker δ ij is used to prevent the interpenetration of the surfaces of a damaged element when contact

occurs2.

In summary, by assuming that toughnesses and strengths in modes II and III are equal to each other, it is shown

that the mixed-mode constitutive equations of a cohesive element are defined by five material properties and three

displacement jumps: GIc, GIIc, τ 30 , τ shear

0
, K, Δ1, Δ 2 , and Δ3 . This formulation of the damage model allows an explicit

7
integration of the constitutive model and ensures consistency in the evolution of damage during loading, unloading,

and changes in mode mixity. The derivation and implementation of the cohesive damage model is described in Refs.

[2, 21] and the reader is referred to those sources for details that will not be repeated here.

B. Viscoelastic stabilization

Material models exhibiting softening behavior and sharp stiffness degradation typically exhibit severe

convergence difficulties in implicit analysis programs. To improve the convergence rate of the iterative procedure, a

viscous stabilization scheme was implemented as suggested by ABAQUS22 for its cohesive element and composite

damage models. The procedure is a generalization of the Duvaut-Lions regularization model which, for sufficiently

small time increments, maintains the tangent stiffness tensor of the softening material as positive definite. Defining

d as the “kinematic” damage that depends on the displacement variables and dv as the “stabilized” damage variable,

the rate of change in the damage variable dv is a function of the damping factor ρ:

d& v =
1
(
d − dv
ρ
) (11)

The stabilized damage variable at time t+Δt is updated as

Δt ρ
dv = ρ + Δt d t + Δt + ρ + Δt dv (12)
t + Δt t

It can be observed that the viscoelastic factor has the units of time and, therefore, its effect depends on the

pseudo-time load proportionality factor which is selected arbitrarily by the user. In the following sections, the effect

of the pseudo-time on the viscoelastic factor is eliminated by specifying it as a nondimensional value, ρ / t F where

t F is the estimated pseudo-time at the maximum load.

The response of the damaged material is evaluated using the stabilized damage variables. Using viscous

stabilization with a viscosity parameter that is much smaller than the time increment can improve the rate of

convergence in the presence of softening processes without compromising the accuracy of the results. However, the

use of larger values of the stabilization parameter can cause a toughening of the material response and produce a

stable propagation of delamination in problems with unstable delamination growth.

8
C. Finite Element Implementation of Zero-Thickness 3D Element

Zero-thickness cohesive elements as shown in Fig. 2a are used to simulate the response of the interface

connecting two laminae of a composite laminate.

Fig. 2. a) 3D cohesive elements b) shell cohesive elements.

The constitutive equation of zero-thickness cohesive elements is established in terms of displacement jumps and

tractions across the interface. The stiffness matrix for conventional 3D cohesive elements is constructed using

standard isoparametric linear Lagrangian interpolation functions. In a local coordinate frame with x-y-z aligned such

that the z-direction corresponds to the shell normal, the displacement jumps between the top and the bottom faces of

the element are

⎧Δ 3 ⎫ ⎧ w⎫ ⎧ w⎫
⎪ ⎪ ⎪ ⎪ ⎪ ⎪
⎨ Δ1 ⎬ = ⎨ u ⎬ − ⎨ u ⎬ = B U 3 D (13)
⎪Δ ⎪ ⎪ ⎪ ⎪ ⎪
⎩ 2 ⎭3D ⎩ v ⎭top ⎩ v ⎭bot

where B is the matrix relating the element’s degrees of freedom U3D to the displacements between the top and

bottom interfaces. The Principle of Virtual Work for 3D cohesive elements can be written as23

δ U T3D ∫A B T ((I − D)C)B dA U 3D = δ U T3D ∫A B T S dA (14)

where D is a diagonal matrix composed the damage state d at the interface. The term I represents the identity matrix,

and C is the undamaged constitutive matrix whose diagonal is composed of the penalty stiffnesses K, and S is the

9
stress tensor. The stiffness matrix K3D of the element is easily identified in the left hand side of Eq. 14 as the

integral over the area A of the element:

K 3D = ∫A BT ((I − D)C)B dA (15)

and the right hand side of Eq. 14 provides the element’s internal force:

F3D = ∫A B T S dA (16)

Because the stiffness K3D is constructed from the displacements U3D, only three degrees of freedom can be active

at a node, i.e., u, v, and w. The integration of the terms in Eqs. 15 and 16 is performed numerically using either

Newton-Cotes or Gaussian integration. The element proposed is implemented in ABAQUS22 as a user-defined

element.

III. Formulation Shell Cohesive Element

Shell cohesive elements have a finite non-zero thickness because they connect nodes on the reference surfaces of

two shells, as shown in Fig. 2b. In addition to the u, v, and w translational degrees of freedom, the nodes of a shell

element also possess three nodal rotations θx, θy, and θz. For kinematic continuity before delamination propagation,

all three translational degrees of freedom at the interface must be the same in both adjacent plies, as shown in Fig. 3.

Fig. 3. Kinematic continuity is achieved when the displacements at the interface are equal in both plies24.

When using shell elements based on First Order Shear Deformation, the three displacements at the interface are

calculated from the nodal degrees of freedom as23:

10
 ⎧u ⎫
⎡ tj ⎤⎪ j ⎪
⎢1 0 0 0 ± 0⎥ v j
⎧ uk ⎫ 2 ⎪ ⎪
⎢ ⎥⎪ ⎪
⎪ ⎪ tj ⎪w j ⎪
⎨ vk ⎬ = ⎢0 1 0 m 0 0⎥ ⎨ ⎬ or U 3k D = R kj U Shell
j
(17)
⎢ 2 ⎥ θ xj
⎪w ⎪ ⎢0 0 1 ⎪ ⎪
⎩ k ⎭3D 0 0 0⎥ ⎪θ ⎪
⎢ ⎥ yj
⎢⎣ ⎥⎦ ⎪⎪θ ⎪⎪
⎩ zj ⎭ Shell

where tj is the thickness of the shell at node j, and its sign depends on whether the shell node described the

displacements of the upper or the lower ply. For arbitrary orientations of a shell in space, the kinematic

transformation matrix Rkj is rotated such that the z-direction is aligned with the normal direction defined by two

connected shell nodes. A transformation matrix R for the entire set of degrees of freedom in an element can be

assembled using the transformations Rkj such that

U 3 D = R U Shell and δ U 3 D = R δ U Shell (18)

The stiffness matrix Kshell of a cohesive element for shell analysis can be constructed by substituting Eqs. 18 into

the expression of the Principle of Virtual Work for a 3D cohesive element (Eq. 12):

δ U TShell R T ∫A B T ((I − D)C)B dA R U Shell = δ U TShell R T ∫A B T S dA (19)

By comparing Eqs. 14 and 19, it is clear that the stiffness and internal force of a shell cohesive element can be

constructed from any standard cohesive element formulation using the following expressions:

K Shell = R T K 3 D R and FShell = R T F3 D (20)

Eqs. 20 show that a cohesive element for shells is easily constructed from any 3D cohesive element by simple

matrix multiplication. An 8-node cohesive element for shells was implemented as an ABAQUS UEL (user element)

subroutine22 by using a zero-thickness 3D element previously developed by the authors2 as a kernel.

A. Spatial Orientation of the Shell Normals

The transformations shown in Eqs. 20 are only valid when the normals to the upper and the lower shells are

aligned with the z-direction of the global reference frame. In the case of arbitrary orientation of the shells in space, it

is necessary to define local reference frames at each corner of the shell cohesive elements. In the present model, the

local normal direction (z') is defined by the vector between the lower and upper nodes at a corner. Using a rotation

11
tensor, the normal and tangential components of the displacement jump can be expressed in terms of the

displacement field in global coordinates.

B. Special Configurations

It was shown in Eqs. 17 and 20 that the element formulation is based on the thickness t u and t l of the upper and

lower shells being connected together by a cohesive element. Therefore, the values of the thicknesses must be

provided for each corner of every cohesive element in the model by using element property data cards. In models

where the thicknesses of the adherends vary across the model, different property cards would have to be generated

for every cohesive element. To improve the ease of generating a model where the shell thicknesses are non-uniform,

it is convenient to define common geometric configurations that allow the use of single property cards for large

areas of cohesive elements. Two such configurations are shown in Fig. 4. The configuration shown in Fig. 4a, which

is typical of skin/flange constructions, can be defined by specifying the thickness t l of the skin, which is constant for

all cohesive elements in the feature. For configurations that are symmetric with respect to the interface, as in the

example shown in Fig. 4b, the ratio t u/t l is constant. In either case, the two thickness t u and t l can be determined

within the element with the additional knowledge that the distance t u + t l is equal to twice the distance between the

upper and the lower nodes and a single property card can be used for all elements in the feature.

Fig. 4. Special configurations with constant thickness properties.

12
 IV. Example and Verification Problems

A. Mixed Mode Bending (MMB) Specimen

The MMB test offers the possibility of measuring different mixed-mode ratios, ranging from pure Modes I to II,

by changing the length e of the loading lever shown in Fig. 5. The loading arm is modeled with a multi-point

constraint equation between the displacements at the load application point and the two loading points on the upper

arm of the specimen.

The MMB test of a 24-ply unidirectional AS4/PEEK (APC2) carbon fiber reinforced composite was simulated

with a simple model composed of two shell layers, and the results were compared with the results of a 3D model

previously published by the authors2. The specimen is 102 mm-long, 25.4mm- wide, with two 1.56 mm-thick arms.

The initial delamination length is a0=34mm, and the length e of the lever is 43 mm, which was calculated to provide

a mode mixity GI/GT=50%. The material properties of the specimen are shown in Table 1, and the properties of the

interface are shown in Table 2. It should be noted that the parameter η was calculated from a fit of experimental data

corresponding to a single test specimen at each mode ratio25.

F e Loading arm

Specimen

Base

Fig. 5. MMB test configuration and finite element models.

13
 Table 1. Material Properties of AS4/Peek25.
E11 (GPa) E22=E33 (GPa) ν12=ν13 ν23 G12=G13 (GPa) G23 (GPa)
122.7 10.1 0.25 0.45 5.5 3.7

Table 2. Properties of the interface1, 25.

GIc (N/mm) GIIc (N/mm) τ 30 (MPa) τ shear
0
(MPa) η
0.969 1.719 80 100 2.28

The experimental results shown in Fig. 6 were performed by Reeder25. They relate the load to the displacement

of the point of application of the load F on the lever. The analytical solutions shown in the figure were calculated

using the equations from Reeder26 with Gc =1.123 N/mm, which corresponds to the point where the delamination

growth was first observed. This point is the second experimental data point in Fig. 6. Therefore, this point lies on the

analytical constant-G curve. It can be observed that the test data indicates an R-curve effect consisting of an increase

in toughness that ranges from an initiation value of 1.123 N/mm to a propagation value of approximately 1.27

N/mm.

The shell finite element model shown in Fig. 5 was developed using two layers of shell elements. Each layer is

composed of a mesh of 2 X 150 elements. Therefore, the size of the elements in the direction of crack propagation is

0.68mm. The shell model only represents 1/10 of the specimen width, so the predicted applied load was multiplied

by a factor of 10. The loading lever was modeled as a rigid link using a multi-point equation. To improve the

convergence rate, a viscoelastic factor of ρ / t F = 2.0 ⋅10 −4 was used, where t F is the estimated pseudo-time at the

maximum load. A penalty stiffness of K= 105 N/mm3 was used. Recommendations on the determination of K are

provided in Ref. 4. The corresponding load-displacement results are shown in Fig. 6. For comparison, the results of

a simulation using a three-dimensional model by the authors2 is also shown in Fig. 6. It can be observed that both

models predict the analytical peak load and load-displacement curve accurately. The correlation of the analytical and

finite element models with the test results could be improved using the toughness for propagation rather than the

value for initiation. For an applied displacement larger than 7 mm, the delamination front proceeds under the mid-

span load point and the present analytical solution is no longer valid. At that point, the finite element models diverge

from the analytical solution.

To verify the effects of mesh size and viscoelastic stabilization, a new shell model was created with 300

elements in the propagation direction, or twice the number of elements of the base model. The model was analyzed

14
with ρ=0. The predicted load with the refined mesh model is less than 1% lower than the load obtained from the

previously discussed coarser model with ρ / t F = 2. ⋅ 10−4 , as can be observed in Fig. 7. However, the refined mesh

model failed to converge after a short amount of crack propagation.

An alternative to using fine meshes of cohesive elements was proposed by Turon et al.3 The approach consists of

reducing the strengths τ 30 and τ shear

0
such that the length of the cohesive zone increases to cover at least three

elements. According to Turon, the strengths that are required to achieve convergence using an element size of 0.68

mm are:

τ 3* 1 9π E33 GIc
= 0 = 0.63
τ3 τ3
0
32 × 5 lele
(21)
τ shear
*
1 9π E33 GIIc
= 0 = 0.67
τ shear
0
τ shear 32 × 5 lele

Equations 21 indicate that the strength reduction needed for mesh convergence is greater for peel than for shear.

Since both strengths should be reduced by the same factor to maintain a constant mode mixity, the reduced strengths

τ 3* and τ shear
*
that ensure that the cohesive zone is covered by at least five elements are:

τ 3* = 0.63 × τ 30 = 50 .4 MPa ; τ shear

*
= 0.63 × τ shear
0
= 63 .0 MPa .

350

Linear
300 Gc = 1.123 N/mm

Test [Reeder, 2002]

250 3D Analysis
Applied load, N.

Turon 2006

200
Shell analysis

150

100

50

0
0 1 2 3 4 5 6 7 8

Applied displacement, mm.

Fig. 6. Predicted and experimental response of 50% mode mix MMB test.

15
 320

280 Linear

240

Applied load, N.
200

160 Gc =cst.

120

Original mesh, ρ /t =0.0002

F

80
Refined mesh, ρ =0
Original mesh, ρ =0, reduced strengths τ*
40

0
0 1 2 3 4 5 6 7 8

Applied displacement, mm.

Fig. 7. Effect of mesh density and viscoelastic stabilization on predicted load displacement response of
50% mode mix MMB test.

An analysis was performed with the original (coarse) mesh model with reduced strengths and no viscoelastic

stabilization. The results shown in Fig. 7 indicate that the results from the original (coarser) mesh using reduced

strengths are virtually identical to those using a refined mesh. In the present problem, the use of lowered strengths

reduces the occurrence of divergence in the iterative solution and the use of viscoelastic stabilization is not required.

B. Debonding of Skin-Stiffener Specimen

Most composite components in aerospace structures are made of panels with co-cured or adhesively bonded

frames and stiffeners. Testing of stiffened panels has shown that bond failure at the tip of the stiffener flange is an

important and very likely failure mode. Comparatively simple specimens consisting of a stringer flange bonded onto

a skin have been developed by Minguet et al.27 to study skin/stiffener debonding. The configuration of the

specimens shown in Fig. 8 was analyzed by Cvitkovich et al.28 The specimens are 203 mm-long, 25.4 mm-wide.

Both skin and flange were made from IM6/3501-6 graphite/epoxy prepreg tape with a nominal ply thickness of

0.188 mm. The skin lay-up consisting of 14 plies was [0/45/90/-45/45/-45/0]s, and the flange lay-up consisting of 10

plies was [45/90/-45/0/90]s.

16
 The properties of the unidirectional IM6/3501-6 unidirectional graphite/epoxy tape reported by Cvitkovich28 are

shown in Table 3. However, the interface properties for IM6/3501-6 were not available. Instead, the interface

properties of AS4/3501-6 are used, as suggested by Cvitkovich28. The values of GIc, GIIc and the parameter η=1.75

for the B-K criterion shown in Table 4 are taken from test data for AS4/3501-629.

Fig. 8. Skin-stiffener specimen configuration28.

Table 3. Material Properties of IM6/3501-6 unidirectional graphite/epoxy tape [Ref. 28]

E11 (GPa) E22=E33 (GPa) ν12=ν13 ν23 G12=G13 (GPa) G23 (GPa)
144.7 9.65 0.3 0.45 5.2 3.4

Table 4. Properties of the interface [Refs. 28,29,30].

GIc (N/mm) GIIc (N/mm) τ 30 (MPa) τ shear

0
(MPa) η
0.0816 0.554 61 68 1.75

A complete analysis of the delamination growth in the tension specimen requires a high degree of complexity.

For instance, Krueger31 developed highly detailed two-dimensional models using up to four elements per ply

thickness. In previously published work30, the authors determined that it is possible to predict the debond load of the

specimen using cohesive elements using a relatively coarse three-dimensional model. To keep the modeling

difficulties low and the approach applicable to larger problems, the 3D model was developed using only two brick

elements through the thickness of the skin, and another two through the flange. The complete model consists of

17
1,002 three-dimensional 8-node C3DI elements and 15,212 degrees of freedom. This model did not contain any pre-

existing delaminations.

To account for the effect of residual thermal strains, a thermal analysis step was conducted before the

mechanical loads were applied. The same coefficients of thermal expansion (α11=-2.4x10-8/ºC and α22=3.7x10-5/ºC)

are applied to the skin and the flange, and the temperature difference between the curing and room temperature is

ΔT= -157 ºC [Krueger31]. The flange has more 90º plies than 0º plies and the skin is quasi-orthotropic, so that

residual thermal stresses develop at room temperature at the skin/flange interface.

In contrast to previous models of the skin-stiffener specimen, an extremely simple shell model was developed

that uses one layer of shell elements to represent the skin and another layer to represent the flange of the stiffener.

When subjected to tensile loads, the primary mechanism for debonding is shear lag. This fracture mechanism is

captured in the model by using a layer of shell cohesive elements to bond the two shell layers together. The model is

composed of six elements across the width, and the mesh in one of the two flange tip regions is refined. The element

size in the propagation direction is 0.5 mm, and the rest of the flange is meshed with 1.0 mm elements. A penalty

stiffness of K= 105 N/mm3 was used for this analysis. The model is composed of 624 shell and cohesive elements

and only 3192 degrees of freedom. Since this particular problem is dominated by the initiation of the debond rather

than its propagation, the nominal strengths were not reduced, and an accurate calculation of the energy release rates

during propagation was not attempted.

A deformed plot of the finite element model immediately after flange separation is shown in Fig. 9.

Shell decohesion elements

at skin-flange interface

Fig. 9. Two-shell model of skin-stiffener specimen.

18
 In previous work by the authors, it was stated that eigenmode analysis of the element stiffness matrices has

shown that Gaussian integration can cause undesired spurious oscillations of the traction field when large traction

gradients are present over an element and, consequently, that Newton–Cotes integration is superior to Gaussian

integration in cohesive elements1,32. However, in the analyses of the skin-stringer specimen, it was found that the

Newton-Cotes integration causes odd wrinkling of the thin flange tip during the onset of delamination, as illustrated

in Fig. 10a. The wrinkling deformations were found to delay the full separation of the flange tip. On the other hand,

the wrinkling deformations were virtually eliminated by using Gaussian integration, as shown in Fig. 10b.

Consequently, previously established conclusions on the preferred integration scheme may need to be re-examined.

a) Newton-Cotes integration b) Gaussian integration

Fig. 10. Onset of delamination with two different integration schemes (10X deformation).

The predicted load-extensometer response for two values of viscoelastic stabilization are shown in Fig. 11. The

initiation of delamination predicted using a viscoelastic stabilization factor of ρ / t F = 2. ⋅ 10−4 is 17.2 kN, which is

24% below the mean debond load measured by Cvitkovich28 (22.7 kN). The unstable nature of delamination

propagation can be deduced from the reduction of the applied load during the propagation of the delamination. The

predicted delamination initiation load using ρ / t F = 2. ⋅10 −3 is 17.5 kN. However, delamination propagation is

stabilized by the use of a larger factor, as can be observed in Fig. 11. The advantage of using a higher amount of

stabilization is that the solution is obtained much faster than with a lesser amount of stabilization: the model with a

stabilization factor of 2. ⋅ 10 −4 took 6 minutes of CPU time on a 3.20GHz Xeon-based PC, while the model with a

factor of 2. ⋅ 10 −3 took less than 3 minutes.

19
 The previously published results of a 3D model are also shown in Fig. 11. The slope of the force-extensometer

curves for the 3D and shell models are different because nodal displacements were used to calculate the

extensometer measurement. For the 3D model, the nodes are located on the surface of the laminates, while for the

shell models, the nodes are located on the reference surfaces. It can be observed that the 3D model predicts a more

accurate debond load of 21.9 kN, which is probably a result of the higher fidelity of the flange tip region in the 3D

model.

25 +2σ

Mean experimental
debond load -2σ
20
ρ/tF=0.002
3D Model Instability
Camanho 2004
Reaction Force, kN

ρ/tF=0.0002
15

10

0
0 0.05 0.1 0.15 0.2 0.25
Predicted extensometer measurement, mm.

Fig. 11. Predicted and experimental load-extensometer response for skin-stiffener specimen in tension.

C. Strength Analysis of Thick Composite Lug

Shell finite elements are typically reserved for the analysis of thin-section structures such as fuselage skins.

Since shell elements either neglect or simplify the transverse deformations, they are not usually applied to model

components in which the thickness dimension is of the order of magnitude of other characteristic dimensions of the

component. However, computationally efficient models of thick components have occasionally been developed by

stacking layers of shell elements. For instance, the complex through-the-thickness deformations of thick sections of

20
the Composite Armored Vehicle, which includes layers of glass/epoxy, ceramic tiles, and rubber, were successfully

modeled by tying four or more layers of shell elements using multi-point constraint equations24.

In addition to the computational efficiency that results from using shell elements, the use of shell cohesive

elements to model layered structures provides the added benefit of simplifying the construction of complex models

in several ways. Firstly, shell cohesive elements eliminate the laborious task of constructing multi-point constraints

to tie the layers together. Secondly, unlike conventional cohesive elements, shell cohesive elements have a non-zero

thickness and they have the same topology and connectivity as a standard volume element. Therefore, they can be

constructed using any finite element pre-processor. Finally, shell cohesion elements do not require the use of wall

offsets. Therefore, the construction of shell models is simplified and, more importantly, the shell reference surfaces

do not have to be shifted to a common interface, which allows stacks of multiple shell layers.

During the course of a failure analysis of a composite aircraft fin undertaken at NASA Langley33, a layered shell

element of the critical attachment lug area was developed. The model detail shown in Fig. 12 illustrates how

cohesive elements for shells can be used to construct models of large and geometrically complex thick composite

structures by stacking layers of shell elements. The thick lug fittings were modeled with 14 layers of shell elements,

which were connected with shell cohesive elements. All other regions of the model were modeled with a single layer

of shell elements. The pin assembly was modeled as a rigid surface with a diameter equal to that of the lug hole.

Frictionless contact equations were prescribed between the edge of the shells around the bolt hole and the rigid

surface. This model has 20886 nodes with 34524 elements. The response of the layered-shell model was compared

to a 3D model (no delamination) and it was found that the strain distributions predicted by the two models were in

close agreement33.

21
 Fig. 12. Layered shell model of composite lug with ply damage and cohesive elements.

In addition to delamination at any interface between the 14 layers of shells, the model uses a progressive damage

model in which the matrix and fiber damage is simulated by degrading the material properties. An example of

damage analysis is shown in Fig. 12, where the regions in red represent areas where damage is predicted. The

analyses for intra-ply damage were developed within ABAQUS using a USDFLD user-written subroutine. The finite

element program calls this subroutine at all integration points of elements that have material properties defined in

terms of the field variables. The subroutine call provides access points to a number of variables such as stresses,

strains, material orientation, current load step, and material name, all of which can be used to compute the field

variables. Stresses and strains are calculated at each incremental load step, and evaluated by the failure criteria to

determine the occurrence of failure and the mode of failure.

The results of a failure analysis indicate that a cleavage-type failure of the lug, in which an initial fracture

develops in a direction parallel to the orientation with the greatest number of fibers. This direction is nearly aligned

with the load direction, as shown in Figs. 13.

22
 ε22 0°
Load

2 1

Tension Load
Load
Compression
Failure mode: cleavage

Fig. 13. Progressive failure analysis predicts a two-step cleavage failure of the composite lug33.

The greatest advantage of the layered shell model is its ability to predict the initiation of delamination, as well as

the component’s tolerance to delamination flaws. A study was undertaken to evaluate the damage tolerance of the

lug in the presence of a pre-existing delamination. The flaw was introduced in the regions of the model shown in red

in Figs. 14 by setting the region of shell cohesive elements shown in red to “failed” by defining the damage variable

of these elements as unity. Failed cohesive elements do not have stiffness in tension or shear, but they prevent

interpenetration of the crack surfaces in compression.

Fig. 14. Configuration of delaminated lug and model for damage tolerance analysis.

The strength analyses with and without a pre-existing delamination were performed under tensile loading. For

this loading condition, the analyses did not predict delamination initiation nor propagation of the delaminated

region. The load-displacement results of the two analyses are shown in Fig. 15. It can be observed that the initial

flaw does not appreciably affect the overall stiffness or strength of the composite lug. The predicted loads for both

23
simulations fall between the load measured for the subcomponent test performed in 2003 and the W375 tension load

that is believed to have caused the in-flight failure of the lug33.

1200

Subcomponent test, 2003

1000
Reaction Force, kN W375 load at failure

800 w/ initial delamination

no delamination

600

Design Limit Load

400

200

0
0 0.5 1 1.5 2 2.5 3 3.5 4

Pin displacement, mm.

Fig. 15. Predicted lug strength with and without initial flaw.

V. Conclusions

The formulation of a cohesive element for shell analysis that can be used to simulate the initiation and growth of

delaminations between stacked, non-coincident layers of shell elements was presented. The procedure to construct

the element accounts for the thickness offset by applying the kinematic relations of shell deformation to transform

the stiffness and internal force of a zero-thickness cohesive element such that interfacial continuity between shell

layers is enforced. The internal load vector and the stiffness matrix of a cohesive element connecting shells can be

calculated in a very simple way by means of a transformation matrix, independently of the cohesive constitutive

models.

Stacks of shell elements can be used to create effective models that predict the inplane and delamination failure

modes of thick components. The formulation allows the use of multiple layers of shells through the thickness and it

does not require offsetting nodes to a common reference surface. In addition, shell cohesive elements have the same

24
topology as three-dimensional elements so they can be generated using any finite element pre-processor while zero-

thickness cohesive elements for three-dimensional elements cannot be generated with most pre-processors.

Using representative examples of increasing structural complexity, it was shown that the proposed cohesive

element for shells can be used to represent the propagation of a pre-existing delamination, as well as the onset and

propagation of delamination in components that do not contain pre-existing cracks. The analysis of a 50% mode-mix

MMB test indicates that the initiation and propagation of delamination under mixed mode can be predicted

accurately. It was found that the convergence difficulties that are typical of cohesive laws can be mitigated with the

use of viscoelastic stabilization or by reducing the interfacial strength to enlarge the process zone. The analysis

results of a two-layer model of a skin-stiffener debond indicate that such a simplified model can produce results that

may be sufficiently accurate for design and down-select analyses. For better predictive accuracy, models in which

the flange taper is represented more accurately would be required. Finally, the analysis results of a large fin lug

component were presented to illustrate how layered shell models with cohesive elements can be used to predict the

strength of thick composite structures. This combination of cohesive elements and shells represents an effective way

of simulating complex aeronautical structures that incorporate other non-linearities such as ply damage propagation,

contact, and finite rotations. However, for the loading conditions that were analyzed, the cohesive elements in the

model were not subjected to loads sufficiently large to cause any delamination growth.

References

[1] Camanho, P.P., Dávila, C.G., and de Moura, M.F.S.F., "Numerical Simulation of Mixed-Mode Progressive
Delamination in Composite Materials," Journal of Composite Materials, Vol. 37, No. 16, 2003, pp. 1415-
1438.
[2] Turon, A., Camanho, P.P., Costa, J., and Davila, C.G., "A Damage Model for the Simulation of
Delamination in Advanced Composites under Variable-Mode Loading," Mechanics of Materials, Vol. 38,
No. 11, 2006, pp. 1072-1089.
[3] Turon, A., Dávila, C.G., Camanho, P.P., and Costa, J., "An Engineering Solution for Mesh Size Effects in
the Simulation of Delamination Using Cohesive Zone Models," Engineering Fracture Mechanics, Vol. 74,
No. 10, 2007, pp. 1665-1682.
[4] Yang, Q., and Cox, B.N., "Cohesive Models for Damage Evolution in Laminated Composites," International
Journal of Fracture, Vol. 133, No. 2, 2005, pp. 107-137.
[5] Wang, J.T., and Raju, I.S., "Strain Energy Release Rate Formulae for Skin-Stiffener Debond Modeled with
Plate Elements," Engineering Fracture Mechanics, Vol. 54, No. 2, 1996, pp. 211-228.

25
[6] Wang, J.T., Raju, I.S., Dávila, C.G., and Sleight, D.W., "Computation of Strain Energy Release Rates for
Skin-Stiffener Debonds Modeled with Plate Elements," Proceedings of the 34th AIAA/ASME/ASCE/AHS
Structures, Structural Dynamics and Materials Conference, LaJolla, CA, April 19-21, 1993.
[7] Glaessgen, E.H., Riddell, W.T., and Raju, I.S., "Nodal Constraint, Shear Deformation and Continuity Effects
Related to the Modeling of Debonding of Laminates Using Plate Elements," CMES, NASA Center for
AeroSpace Information (CASI), Vol. 3, Hampton, VA, 2002.
[8] Raju, I.S., Crews, J., J. H., and Aminpour, M.A., "Convergence of Strain Energy Release Rate Components
for Edge-Delaminated Composite Laminates," Engineering Fracture Mechanics, Vol. 30, No. 3, 1988, pp.
383-396.
[9] Reedy, E.D., Mello, F.J., and Guess, T.R., "Modeling the Initiation and Growth of Delaminations in
Composite Structures," Journal of Composite Materials, Vol. 31, 1997, pp. 812-831.
[10] Johnson, A.F., and Pickett, A.K., "Impact and Crash Modelling of Composite Structures: A Challenge for
Damage Mechanics," Proceedings of the European Conference on Computational Mechanics, Munich, Sept
1999.
[11] Crisfield, M.A., Mi, Y., Davies, G.A.O., and Hellweg, H.-B., "Finite Element Methods and the Progressive
Failure-Modelling of Composites Structures," Computational Plasticity: Fundamentals and Applications '97,
Edited by D.R.J. Owen, Vol. 1, CIMNE, Barcelona, 1997, pp. 239-254.
[12] Ladevèze, P., Allix, O., Deü, J.-F., and Lévêque, D., "A Mesomodel for Localisation and Damage
Computation in Laminates," Computer Methods in Applied Mechanics and Engineering, Vol. 183, No. 1-2,
2000, pp. 105-122.
[13] Borg, R., Nilsson, L., and Simonsson, K., "Simulating DCB, ENF and MMB Experiments Using Shell
Elements and a Cohesive Zone Model," Composites Science and Technology, Vol. 64, No. 2, 2004, pp. 269-
278.
[14] Bruno, D., Greco, F., and Lonetti, P., "Computation of Energy Release Rate and Mode Separation in
Delaminated Composite Plates by Using Plate and Interface Variables," Mechanics of Advanced Materials
and Structures, Vol. 12, No. 4, 2005, pp. 285-304.
[15] Jin, Z.H., and Sun, C.T., "Cohesive Zone Modeling of Interface Fracture in Elastic Bi-Materials,"
Engineering Fracture Mechanics, Vol. 72, No. 12, 2005, pp. 1805-1817.
[16] Tvergaard, V., "Influence of Plasticity on Interface Toughness in a Layered Solid with Residual Stresses,"
International Journal of Solids and Structures, Vol. 40, No. 21, 2003, pp. 5769-5779.
[17] Tvergaard, V., "Predictions of Mixed Mode Interface Crack Growth Using a Cohesive Zone Model for
Ductile Fracture," Journal of the Mechanics and Physics of Solids, Vol. 52, No. 4, 2004, pp. 925-940.
[18] Zhang, W., and Deng, X., "Asymptotic Fields around an Interfacial Crack with a Cohesive Zone Ahead of
the Crack Tip," International Journal of Solids and Structures, Vol. 43, No. 10, 2006, pp. 2989-3005.
[19] Camanho, P.P., Dávila, C.G., and Ambur, D.R., "Numerical Simulation of Delamination Growth in
Composite Materials," NASA/TP-2001-211041, Hampton, VA, August 2001.

26
[20] Benzeggagh, M.L., and Kenane, M., "Measurement of Mixed-Mode Delamination Fracture Toughness of
Unidirectional Glass/Epoxy Composites with Mixed-Mode Bending Apparatus," Composites Science and
Technology, Vol. 56, No. 4, 1996, pp. 439-449.
[21] Turon, A., "Simulation of Delamination in Composites under Quasi-Static and Fatigue Loading Using
Cohesive Zone Models," PhD Dissertation, Dept. d'Enginyeria Mecànica i de la Construcció Industrial,
Universitat de Girona, Girona, Spain, 2006.
[22] ABAQUS 6.6 User's Manual, ABAQUS Inc., Providence, RI, USA, 2006.
[23] Dávila, C.G., "Solid-to-Shell Transition Elements for the Computation of Interlaminar Stresses," Computing
Systems in Engineering, Vol. 5, 1994, pp. 193-202.
[24] Dávila, C.G., and Chen, T.-K., "Advanced Modeling Strategies for the Analysis of Tile-Reinforced
Composite Armor," Applied Composite Materials, Vol. 7, No. 1, 2000, pp. 51-68.
[25] Reeder, J.R., "Load-Displacement Relations for DCB, ENF, and MMB AS4/Peek Test Specimens," Private
Communication, 2002.
[26] Reeder, J.R., Demarco, K., and Whitley, K.S., "The Use of Doubler Reinforcement in Delamination
Toughness Testing," Composites Part A: Applied Science and Manufacturing, Vol. 35, No. 11, 2004, pp.
1337-1344.
[27] Minguet, P.J., Fedro, M.J., O'Brien, T.K., Martin, R.H., and Ilcewicz, L.B., "Development of a Structural
Test Simulating Pressure Pillowing Effects in Bonded Skin/Stringer/Frame Configuration," Proceedings of
the Fourth NASA/DoD Advanced Composite Technology Conference, Salt Lake City, Utah, June 7-11,
1993.
[28] Cvitkovich, M.K., Krueger, R., O'Brien, T.K., and Minguet, P.J., "Debonding in Composite Skin/Stringer
Configurations under Multi-Axial Loading," Proceedings of the American Society for Composites - 13th
Technical Conference on Composite Materials, Baltimore, MD, September 21-23, 1998, pp. 1014-1048.
[29] Reeder, J.R., "3D Mixed-Mode Delamination Fracture Criteria–an Experimentalist’s Perspective,"
Proceedings of the 21st Annual Technical Conference of the American Society for Composites, Dearborn,
MI, September 17-20, 2006.
[30] Camanho, P.P., Dávila, C.G., and Pinho, S.T., "Fracture Analysis of Composite Co-Cured Structural Joints
Using Decohesion Elements," Fatigue & Fracture of Engineering Materials & Structures, Vol. 27, No. 9,
2004, pp. 745-757.
[31] Krueger, R., París, I.L., O'Brien, T.K., and Minguet, P.J., "Fatigue Life Methodology for Bonded Composite
Skin/Stringer Configurations," NASA/TM-2001-210842, Hampton, April 2001.
[32] Schellekens, J.C.J., and de Borst, R., "On the Numerical Integration of Interface Elements," International
Journal for Numerical Methods in Engineering, Vol. 36, 1993, pp. 43-66.
[33] Raju, I.S., Glaessgen, E.H., Mason, B.H., Krishnamurthy, T., and Dávila, C.G., "NASA Structural Analysis
Report on the American Airlines Flight 587 Accident - Local Analysis of the Right Rear Lug," Proceedings
of the 46th AIAA/ASME/ASCE/AHS/ASC Structures, Structural Dynamics and Materials Conference,
Austin, TX, April 18-20, 2005.

27