Available online at www.sciencedirect.

com
ScienceDirect

Comput. Methods Appl. Mech. Engrg. 346 (2019) 152–179

www.elsevier.com/locate/cma

Gradient-enhanced damage modeling in Kirchhoff–Love shells:

Application to isogeometric analysis of composite laminates
M.S. Pigazzinia , D. Kamenskyc , D.A.P. van Ierselb , M.D. Alaydinc , J.J.C. Remmersb ,
Y. Bazilevsc ,∗
a Livermore Software Technology Corporation, 7374 Las Positas Road, Livermore, CA 94551, United States
b Department of Mechanical Engineering - Eindhoven University of Technology, 5600 MB, Eindhoven, Netherlands
c School of Engineering, Brown University, 184 Hope St., Providence, RI 02912, United States

Received 29 June 2018; received in revised form 30 October 2018; accepted 30 October 2018
Available online 29 November 2018

Highlights
• Developed a gradient-enhanced continuum damage model for thin shells.
• Tensor-valued strain smoothed by solving an elliptic PDE system.
• PDE system formulated to be independent of the coordinate choice.
• Numerical examples illustrate mesh-independence.
• Impact on geometrically complex composite laminates is simulated.

Abstract
We extend a recently-developed framework for isogeometric analysis of composite laminates to drive material damage evolution
with a smoothed strain field. This builds on ideas from gradient-enhanced continuum damage modeling, and is intended to limit
the dependence of damage predictions on the choice of discrete mesh. The resulting enhanced framework models each lamina of
a composite shell structure as a Kirchhoff–Love thin shell. To account for the anisotropic damage modes of laminae, we smooth
a tensor-valued strain by solving an elliptic partial differential equation (PDE) system on each lamina. This strain-smoothing
PDE system is formulated to be independent of the choice of coordinates and is applicable to general manifold shell geometries.
Numerical examples illustrate the enhanced damage model’s validity, mesh-independence, and applicability to complex industrial
geometries.
⃝c 2018 Elsevier B.V. All rights reserved.

Keywords: Nonlocal damage; Continuum damage; Gradient-enhanced model; Multilayer Kirchhoff–Love shell; Isogeometric analysis; NURBS

1. Introduction
Fiber reinforced composite materials show great promise in applications requiring both high strength and low
weight. However, the use of these materials in safety-critical applications is still limited, due to ongoing difficulties
∗ Corresponding author.
E-mail address: yuri_bazilevs@brown.edu (Y. Bazilevs).

https://doi.org/10.1016/j.cma.2018.10.042
0045-7825/⃝ c 2018 Elsevier B.V. All rights reserved.
 M.S. Pigazzini, D. Kamensky, D.A.P. van Iersel et al. / Computer Methods in Applied Mechanics and Engineering 346 (2019) 152–179 153

with prediction of their damage and fatigue behavior. These difficulties stem from the complex, multi-modal failure
mechanisms of fiber-reinforced composites [1]. Composite structures optimized for low-weight and high-strength
applications are often laminates, consisting of several layers of fiber-reinforced material, called laminae, or plies,
that are bonded together. Intralaminar damage may occur within a given lamina, and interlaminar damage, or
delamination may occur when bonds between laminae break down. We recently developed an approach for predicting
damage in composite laminates that models laminae as Kirchhoff–Love thin shells and interlaminar bonds as cohesive
interfaces [2–4]. Intralaminar damage in this framework evolves according to a continuum damage law accounting
for fiber and matrix failure modes. The cited references demonstrated the validity of this model by replicating several
laboratory experiments.
Despite these promising initial results, it is not assured that the approach of [2–4] will reliably work as-is for
predicting damage in objects with general geometries and/or highly non-uniform or unstructured discretizations.
The model of [2] relied on a local damage law, in which the damage computed at a material point depends on the
strain history only at that point. This, in combination with a strain-softening damage law, leads to the well-known
strain-localization phenomenon, in which strains become concentrated, non-uniquely, in sets of measure zero. The
ill-posedness of local models is attributable to loss of ellipticity of the governing partial differential equation (PDE)
system [5,6]. A consistent discretization scheme will inherit this ill-posedness, which manifests in finite element
models as mesh-dependent damage distributions that do not converge under refinement. The use of mesh-dependent
computational models that do not correspond to well-posed mathematical models is at odds with the principles for
verification and validation articulated in [7], in particular, with the assertions that mathematical models should be
well-posed, that exact solutions should be recoverable under refinement, and that “any validation process in which
computational parameters, such as mesh size, are varied to bring experimental observations into closer agreement
with numerical results are fundamentally flawed” [7, Footnote 4].
A way to improve the well-posedness of damage models is to introduce nonlocality, so that the damage state at a
material point depends on the strain history of not just that point, but neighboring points as well. A variety of nonlocal
models have been proposed over the past several decades, including integral convolutions [8,9], peridynamics [10],
phase-field fracture models [11], and gradient-enhanced continuum damage [12]. In this work, we focus on the last
of these, in which nonlocality is introduced by solving an elliptic PDE. Gradient-enhanced damage (or, simply,
“gradient damage”) models may involve a PDE that smooths strains, damage indices, or other quantities. Models
smoothing a damage index are closely related to phase-field fracture models [13], and, in some publications, the
terms “gradient-enhanced damage” and “phase-field fracture” are even used interchangeably. Phase-field fracture has
recently attracted much attention, and a number of authors have proposed such damage-smoothing formulations for
thin shell structures [14–17]. However, in the present work, we focus on the strain-smoothing family of gradient-
enhanced damage models. Our reasons for this are the following. We wish to adapt an existing, validated damage
model that we studied in [2,3]. When damage indices degrade material stiffness in the way stipulated by this model,
damage smoothing may result in a phenomenon called stress locking, in which the model is unable to represent full
material failure [18,19].1
A previous study adapted strain-smoothing gradient-enhanced damage modeling to continuum shell elements [20].
However, to our knowledge, very little (if any) prior work has applied this type of gradient damage modeling to
thin shell structures. The rise of Isogeometric Analysis (IGA) [21,22] brought renewed interest in computational
analysis of thin-shells due to the higher-order inter-element continuity which can be effortlessly achieved by using
spline shape functions. IGA is an emerging paradigm for computational mechanics, in which approximate solutions
to PDEs are represented in the same function space as the geometry of their domains. Geometries of engineered
objects are frequently designed by specifying surfaces using spaces of smooth spline functions such as NURBS [23],
or T-splines [24]. This makes IGA appealing for analysis of thin shells for two reasons. First, the PDE domain for thin
shell theory is itself a surface, and thus IGA may be applied directly to surface-based geometrical designs of shells.
Second, the spline spaces used to design surfaces typically have C 1 or higher continuity, which, unlike traditional C 0
finite element function spaces, allows for standard Bubnov–Galerkin-type discretizations of fourth-order thin shell
theories. IGA of thin shells began with the influential work by the authors of [25], on whose foundation the present
study builds.
1 Although the nonlocal model analyzed in [18] is not a gradient-enhanced model, our preliminary computations reported in Appendix A indicate
that some of the conclusions of [18] carry over.
154 M.S. Pigazzini, D. Kamensky, D.A.P. van Iersel et al. / Computer Methods in Applied Mechanics and Engineering 346 (2019) 152–179

We begin, in Section 2, by formulating a strain-smoothing model for the Green–Lagrange strain, for a general
3D continuum. In contrast to much prior work on gradient damage, we smooth the full tensor-valued strain (rather
than a scalar effective strain), to account for the anisotropic damage modes of fiber-reinforced composites. Further,
we formulate our model using the tensor generalization of the Laplace–Beltrami operator on manifolds and provide
formulas for arbitrary coordinate charts, to facilitate the reduction to a shell theory. Section 3 specializes this model
to the case of a thin shell by introducing kinematic assumptions, and Section 4 applies it to laminated composite
shells. The Isogeometric discretization for the gradient-enhanced constitutive equations is introduced in Section 5. In
Section 6, we provide several examples illustrating the validity and general applicability of the new gradient damage
model. Conclusions and future directions are summarized in Section 7.

2. Gradient-enhanced damage on manifolds

Building on our prior work [2,26] on continuum damage in composite shell structures, we assume that the damage
law is driven by the Green–Lagrange strain, which encapsulates information about the changes in dot products
between tangent vectors to a deforming body. For an undeformed body modeled by manifold Ω0 , mapped to a
deformed body Ω by the motion φ, the Green–Lagrange strain E satisfies:
2E♭ = φ ∗ (g) − G , (1)
where g is the metric tensor on Ω , G is the metric tensor on Ω0 , φ ∗ is the pull-back by φ, and the superscript ♭ indicates
an associated tensor taking all vector arguments. The ♯ symbol will be used to denote an associated tensor taking all
co-vector arguments. For a review of mathematical prerequisites for large-strain elasticity, refer to [27, Chapter 1],
whose notational conventions we largely follow. In index notation, the pulled-back metric is given by:
φ (g) AB = gab F a A F b B ,
( ∗ )
(2)
where F = Dφ is the deformation gradient, raised indices indicate co-vector arguments, lowered indices indicate
vector arguments, and repeated indices denote contraction. In this notation, ♯ may be considered to “raise all indices”
and ♭(to) “lower all indices”, through contraction with the (inverse) metric tensor. For example, consider a tensor A of
type 32 ; some of its associated tensors are given in index notation as:
A ABC E
D = A ABC D F G F E , (3)
A AC D E = A AFC D E G F B , (4)
( ♭) B
A ABC D E = A ABC D E = A F G H D E G F A G G B G H C , (5)
( ♯ ) ABC D E
A = A ABC D E = A ABC F G G F D G G E , (6)
AB −1
and so on, where G is the index notation for G . We follow the convention that lower-case Roman-letter indices
correspond to tensor arguments in (co)tangent spaces of points in Ω , while upper-case Roman-letter indices refer
to Ω0 . The formula (1) becomes the familiar FT F − I in Cartesian coordinates, with gab = δab , G AB = δ AB , and
(·)T understood as matrix transposition (vs. [27, Proposition 3.4(ii), Chapter 1]), but we retain the generality of the
manifold viewpoint here, to facilitate reduction to a shell theory.
To avoid unstable strain localization, as discussed in Section 1, we propose to instead drive damage with a smoothed
Green–Lagrange strain. We build on the gradient-enhanced damage model of [12], in which a scalar strain measure
ε0 is smoothed to obtain ε̃0 satisfying the PDE:
ε̃0 − c∆ε̃0 = ε0 , (7)
where c is a scalar parameter with dimensions of area (i.e., length scale squared) and ∆ is the Laplace operator. By
analogy, we consider a smoothed Green–Lagrange strain Ẽ satisfying the PDE:
Ẽ − DIV c GRAD Ẽ = E ,
( )
(8)
where DIV and GRAD are the divergence and gradient operators on the reference configuration Ω0 , and the tensor
c permits anisotropy in the strain smoothing. The full tensor-valued strain E is smoothed because the damage model
of [2] requires anisotropic strain information. The smoothed strain Ẽ will then be used in place of the Green–Lagrange
strain to drive damage evolution. Multiplying (8) by a tensor-valued test function W, integrating by parts, and (again
 M.S. Pigazzini, D. Kamensky, D.A.P. van Iersel et al. / Computer Methods in Applied Mechanics and Engineering 346 (2019) 152–179 155
( )
following [12] by analogy) assuming natural boundary conditions (i.e., c GRAD Ẽ · N = 0, where N is the outward
normal to ∂Ω0 ), we obtain the weak form: Find Ẽ♭ ∈ V E such that for all W♭ ∈ V E :
∫ ∫ ∫
)♭
Ẽ♭ : W♯ dΩ0 + c GRAD Ẽ : (GRAD W)♯ dΩ0 = E♭ : W♯ dΩ0 ,
(
(9)
Ω0 Ω0 Ω0

where the tensor contraction operator “:” indicates contraction over all indices of its operands, i.e., (cf. [27, Definition
4.12, Chapter 1]):
S♭ : T♯ = S ABC... T ABC... . (10)
In index notation, Eq. (9) may be written unambiguously:
∫ ∫ ∫
Ẽ AB W AB dΩ0 + c F G Ẽ AB|C W D E|F G D A G E B G GC dΩ0 = E AB W AB dΩ0 , (11)
Ω0 Ω0 Ω0

where the vertical stroke “|” indicates covariant differentiation. (We assume a metric-compatible connection, for
which raising and lowering of indices commutes with covariant differentiation.) For a tensor A of type 02 , covariant
()

differentiation can be expressed in a coordinate chart {ξ A } as [27, Box 4.1, Chapter 1]:
(∇A) ABC = A AB|C = A AB,C − A D B Γ AC
D D
− A AD Γ BC , (12)
where the comma indicates partial differentiation and the Γ s are Christoffel symbols, which are defined such that:
C
∇G A G B = Γ AB GC , (13)
in which ∇w denotes covariant differentiation in the direction w and
∂
GA = (14)
∂ξ A
are the covariant basis vectors associated with the coordinates.

3. Specialization for thin shells

We now choose a particular coordinate chart tailored to the case of a thin shell structure whose thickness, h th , is
assumed to be much smaller than the overall size of the structure. With reference to Fig. 1, let {ξ α }2α=1 be coordinates
of the shell structure’s 2D un-deformed midsurface, Γ0 , and let the coordinate ξ 3 ∈ [−h th /2, h th /2] parameterize the
through-thickness direction orthogonal to Γ0 .2 Unless otherwise specified, we follow the notational convention that
Greek-letter indices have range 1, 2, while Roman-letter indices have range 1, 2, 3. Assuming standard Kirchhoff–
Love kinematics, we consider:
E α3 = E 3α = E 33 = 0 , (15)
and likewise for Ẽ and W. We make the further approximation that {Gα } do not depend on ξ 3 . This contradicts (14)
unless Γ0 is flat, but it is consistent with the assumptions frequently made to permit analytical through-thickness
integration of Kirchhoff–Love shell formulations, as in [25]; see [28, Section 4.4.4] for more detailed justification.
With this assumption, ∇G3 Gα = 0, or
B
Γα3 B
= Γ3α =0. (16)
Kirchhoff–Love thin shell kinematics allow for a separation of membrane and bending components of the in-plane
strain, so that:
Eαβ = εαβ + ξ 3 καβ , (17)
Ẽαβ = ε̃αβ + ξ κ̃αβ ,
3
(18)
Wαβ = δεαβ + ξ δκαβ , 3
(19)
where ε and κ are membrane and curvature-change strains (defined by [2, Section 2.1]), ε̃ and κ̃ are their smoothed
counterparts, and δε and δκ are the corresponding test functions. Due to our use of midsurface basis vectors {Gα }
2 In our previous work [2] the through-the-thickness basis vector was normalized such that ξ 3 ∈ [−1, 1].
156 M.S. Pigazzini, D. Kamensky, D.A.P. van Iersel et al. / Computer Methods in Applied Mechanics and Engineering 346 (2019) 152–179

Fig. 1. Parameterization of a thin shell as function of (ξ α ) in-plane and (ξ 3 ) through-the-thickness covariant coordinates. Curvilinear basis g A
are computed in the current configuration from (14). The shell reference plane Γ0 is represented as a NURBS surface highlighted in blue. (For
interpretation of the references to color in this figure legend, the reader is referred to the web version of this article.)

through the entire thickness, E♭ in (17) operates on tangent vectors of Γ0 for all ξ 3 . (We acknowledge the slight abuse
of notation, in using the same symbol for the 3D continuum strain, operating on tangent vectors of Ω0 .) Given the
strain decompositions (17)–(19), it is convenient to split the volumetric integrals in (11) into surface integrals, over
the shell midsurface Γ0 , and line integrals in the direction ξ 3 over the domain [−h th /2, h th /2]. This results into two
independent problems. First: Find ε̃ such that for all δε,
(∫ ∫ ) ∫
h th ε̃αβ δε αβ dΓ0 + cζ ν ε̃αβ|γ δεδϵ|ζ G δα G ϵβ G γ ν dΓ0 = h th εαβ δε αβ dΓ0 , (20)
Γ0 Γ0 Γ0

and, second: Find κ̃ such that for all δκ,

h 3th
(∫ ∫ )
αβ ζ δα ϵβ γ ν
κ̃αβ δκ dΓ0 + c ν κ̃αβ|γ δκδϵ|ζ G G G dΓ0
12 Γ0 Γ0
h3
∫ ∫
+ h th c3 3 κ̃αβ δκ αβ dΓ0 = th καβ δκ αβ dΓ0 . (21)
Γ0 12 Γ0

Terms coupling ε̃ and κ̃ have canceled, due to integration of odd powers of ξ 3 over the symmetric interval
[−h th /2, h th /2]. The “extra” term proportional to h th in (21) comes from smoothing of 3D strain in the ξ 3 direction:
∂ ε̃αβ + ξ 3 κ̃αβ
( )
Ẽαβ|3 = D
− Ẽ Dβ Γα3 D
− Ẽα D Γβ3 = κ̃αβ (22)
∂ξ 3
and, likewise,

Wαβ|3 = δκαβ , (23)

so that:

Ẽαβ|3 W αβ |3 G 33 = κ̃αβ δκ αβ , (24)

where we use assumptions (16) in the second equality of (22), and the fact that G 33 = 1 in (24). This additional
term will dominate the problem for κ̃ in the limit of h th → 0, driving κ̃ → 0. This is expected when smoothing the
odd-in-ξ 3 function ξ 3 κ̃ over a small interval about ξ 3 = 0.
 M.S. Pigazzini, D. Kamensky, D.A.P. van Iersel et al. / Computer Methods in Applied Mechanics and Engineering 346 (2019) 152–179 157

Remark 1. The use of covariant differentiation in the diffusive terms of (20) and (21) ensures that damage models
relying on the smoothed strain will be independent of the choice of coordinates on Γ0 . Most prior studies on gradient-
enhanced damage (and related phase field methods) smooth only a scalar-valued field, say, s, in which case s|α = s,α .
However, this is no longer true for higher-rank fields.

4. Application to continuum damage modeling for laminated composite shells

Following our earlier work [2], we model multi-layer laminated composite shell structures as assemblies of
Kirchhoff–Love thin shells coupled through cohesive interfaces, with one Kirchhoff–Love shell representing each
lamina. The problem (20) is used to determine the smoothed membrane strain in each of these layers. Based on the
assumption that the individual laminae are very thin, we assume that κ̃ = 0, as implied by the h th → 0 limit of (21).
This does not preclude damage due to bending of the full laminated shell, as laminae away from the neutral axis will
still have nonzero strains. For a single ply, then, we have Ẽ♭ = ε̃ for all ξ 3 ∈ [−h th /2, h th /2]. For the ply group
construction of [2, Section 2.2], in which a collection of lamina are represented by a single Kirchhoff–Love shell, but
may be damaged independently, we modify this to:
♭
Ẽk = ε̃ + ξk3 κ , (25)
♭
where Ẽkis the smoothed strain within the kth ply, ξk3
is the through-thickness coordinate of the midsurface of the
kth ply, ε̃ is the smoothed membrane strain of the entire group, and κ is the unsmoothed bending strain of the group.
This assumes that individual plies are thin enough to consider Ẽ constant through the thickness of each ply, but still
permits damage to be caused by pure bending movements of the ply group. The treatment of ply groups is an example
of an equivalent single layer theory, as developed in [29] for a broader class of elastic constitutive models.

Remark 2. The use of separate displacement fields for each ply (or ply group) effectively captures complex transverse
shear kinematics, even prior to delamination, through cohesive interface stiffness, as described in [2]. This is especially
relevant in composite layups with varying fiber directions and/or properties through the thickness. While capable of
reproducing the limit of a homogeneous thick shell ([2, Section 6.4.2]), the multi-layer framework is less efficient for
such situations than traditional thick-shell formulations, such as solid-like or continuum shells [30,31]. Continuum
shells can also be augmented to include delamination [32].
We re-use the damage model detailed in [2, Section 3], but alter the definitions of the “equivalent displacements”
that are used to drive the evolution of damage parameters associated with different damage modes. As in [2], we
consider each lamina to be composed of an orthotropic material. We consider four distinct modes of damage in such
materials:
• Mode 1T : Fiber breaking under tension.
• Mode 1C : Fiber buckling under compression.
• Mode 2T : Matrix cracking under transverse tension and shear.
• Mode 2C : Matrix cracking under transverse compression and shear.
These modes are associated with equivalent displacements δ̂1T , δ̂1C , δ̂2T , and δ̂2C . Equivalent displacements are
computed from dimensionless strains through multiplication by a length scale. In the case of local damage models, this
length scale is tied to the choice of discretization, and is chosen to be the smallest distance within which strains can
be localized in a discrete displacement field, in an attempt to prevent mesh-dependent energy dissipation. However,
the choice of length scale is by heuristic, at best, for problems discretized on general unstructured meshes in multiple
space dimensions. In the case of gradient-enhanced
√ damage modeling, the localization of strains is controlled by
the model parameter c, and we may consider c to be an appropriate mesh-independent length scale for computing
equivalent displacements from smoothed strains.
In particular, we compute the equivalent displacements by:
⟨ ⟩
δ̂1T = L 1 Ẽ11 (26)
⟨ ⟩
δ̂1C = L 1 −Ẽ11 (27)
√
⟨ ⟩ 2 ( )2
δ̂2T = L 2 Ẽ22 + Ẽ12 (28)
158 M.S. Pigazzini, D. Kamensky, D.A.P. van Iersel et al. / Computer Methods in Applied Mechanics and Engineering 346 (2019) 152–179
√
⟨ ⟩2 ( )2
δ̂2C = L 2 −Ẽ22 + Ẽ12 , (29)

where Ẽαβ are dimensionless components of the smoothed strain Ẽ♭ with respect to an orthonormal basis {eα } for the
co-tangent space of the lamina midsurface, in which e1 is parallel to fiber direction in the lamina (cf. [2, Section 2.1]).
Angle brackets ⟨·⟩ are Macaulay brackets, isolating the non-negative part of their argument. The coefficients {L α }
are the length scales used to convert from strains to equivalent displacements; in [2], we assumed L 1 = L 2 and
computed these length scales from the Jacobian of the mapping from a standard element to physical space, to limit
the effects of mesh-dependent damage localization. In the present work, we compute {L α } from c. First, we introduce
the assumption that c♭ is diagonal in the basis {eα }, then set:
√
L α = cαα , (30)
where the bar again indicates that components are in the fiber-aligned orthonormal basis, and the repeated index is
underlined to indicate that no summation takes place in this instance. As we do not presently have a method for
estimating appropriate anisotropy of c, we assume, for the remainder of this paper, that:
cα β = cδ α β , (31)
where the scalar parameter c has dimensions of area. The remainder of the shell structure mechanics, laminar damage
modeling, and interlaminar cohesive coupling from [2] is unchanged, and we refer the reader to the cited reference
for additional details.

Remark 3. The introduction of a new model parameter, c, may seem unappealing to practitioners working with local
models, who have no established protocol for calibrating a strain-smoothing length scale. However, mesh-dependence
of solutions to local models means that the number, positions, and connectivity of nodes in a finite element mesh must
all be considered free parameters. This makes calibration with specific geometries difficult, and renders intractable
the end goal of predicting damage in new scenarios, for which experimental data is not available.

5. Isogeometric discretization of strain smoothing

To apply the smoothed-strain damage model of Sections 3–4 in computer simulations, the numerical framework
of [2] is extended to solve for an approximate solution to (20) at each time (or load) step of a simulation. The
smoothed strain is nonlinearly coupled to the mechanics and damage evolution of the structure. However, the problem
of solving for ε̃ given a fixed structural deformation is linear, with only the source term depending on the deformation.
The relative straightforwardness of the fixed-displacement strain smoothing problem leads us to adopt a staggered
approach to solving the fully-coupled problem, in which we alternate between solving structural mechanics and strain
smoothing subproblems, holding the solution to the other fixed in each, until arriving at a fixed point. The precise
algorithm for applying this staggered solution procedure in conjunction with the generalized-α time integration used
in [2] is essentially identical to that given for phase field fracture analysis in [33, Section 3.3.2], with the strain
smoothing subproblem taking the place of the phase field subproblem.
The discretization of the structural mechanics subproblem is as specified in [2] and references therein. We discretize
the new strain smoothing subproblem (20) using a standard Bubnov–Galerkin approach, i.e., posing (20) over a
finite-dimensional subspace V Eh of V E . For convenience, each of the three unique components of the symmetric
tensor ε̃ (as expressed in the curvilinear basis {Gα }) is considered to be in the same finite-dimensional scalar-
valued space as the components of shell structure displacement. Test and trial spaces for the second-order strain
smoothing subproblem need only be H 1 (Γ0 )-conforming, but we use C 1 isogeometric spline functions, due to the more
stringent constraints on the displacement spaces, from the higher-order bending terms of the Kirchhoff–Love shell
formulation. In particular, we use B-spline and NURBS function spaces, and assume that the curvilinear coordinates
{ξ α } correspond to the parameter spaces of the spline patches.

Remark 4. The discretization of curvilinear (rather than Cartesian) tensor components is in contrast to some
prior finite element discretizations of surface differential operators acting on unknown fields of nonzero rank, e.g.,
[34–36]. The approaches of the cited references introduce redundant degrees of freedom by seeking higher-
dimensional tensor-valued solutions, and require extra mechanisms to enforce solution tangency, albeit with various
 M.S. Pigazzini, D. Kamensky, D.A.P. van Iersel et al. / Computer Methods in Applied Mechanics and Engineering 346 (2019) 152–179 159

advantages discussed in [36]. Our approach in the present work is analogous in some ways to the treatment of
isogeometric discrete differential forms in [37,38], where the coupled unknown scalar fields are associated with
parametric rather than spatial directions. In the same way that divergence-conforming B-spline discretizations of
incompressible flows fail to satisfy exact global momentum balance [39, Section 9.7], our smoothed strain field may
fail certain “patch tests” such as exactly reproducing spatially-uniform strains. However, the error converges under
refinement and is likely negligible for approximation spaces sufficient to represent accurate displacement solutions
to strain-softening damage models. Importantly, Eh = 0 still implies Ẽh = 0, so rigid-body motions cannot result in
spurious damage.
The only terms in the formulation (20) that are not readily available within prior frameworks for isogeometric
α
analysis of Kirchhoff–Love shell structures (as detailed in, e.g., [16,25,40]) are the Christoffel symbols, Γβγ . These
may be computed from the metric tensor G using the formula [27, Lemma 3.29, Chapter 1]:
α 1
= G αβ G γβ,δ + G δβ,γ − G δγ ,β .
( )
Γδγ (32)
2
The necessary partial derivatives of G can be obtained via:
G αβ,γ = X,α · X,β ,γ = X D ,α δ D E X E ,β ,γ = X D ,αγ δ D E X E ,β + X D ,α δ D E X E ,βγ ,
( ) ( )
(33)
where X : R2 → R3 maps from coordinates {ξ α } to embeddings of the corresponding points of Γ0 into a Cartesian
coordinate representation of 3D space.3 In standard isogeometric discretizations, position components X A are linear
combinations of scalar basis functions defined on the parameter space, and thus the derivatives X A ,α and X A ,αβ are
straightforwardly constructed as linear combinations of derivatives of shape functions.

6. Numerical results
Reinstatement of mesh objectivity is the primary objective of the gradient-enhanced method. Results presented
in the literature (i.e., see [12,13,31]) proved that the adoption of a gradient-enhanced strain regularization leads to
mesh-independent damage predictions.
The formulation presented in Section 3–4 is specifically developed for isogeometric applications, in which mesh
regularity is strictly connected to the NURBS representation of geometries. In this section, we verify the robustness
and convergence properties of the proposed isogeometric gradient-enhanced method with respect to mesh refinement
and distortion.
The gradient-enhanced damage model is then applied for impact simulations in combination with the full multi-
layer modeling approach [2]. First, the low velocity impact on a 24-plies flat laminate, presented in [4], is revisited
and results obtained using local and nonlocal damage formulations are compared. A second impact simulation shows
the application of nonlocal damage modeling to a low-velocity impact on a reinforced composite panel.
All graphical results presented in this paper are obtained by interpolating the smooth isogeometric solutions using
low-order finite-element meshes. This allows us to use convenient visualization tools developed for finite-element
applications.

6.1. Convergence of strain smoothing under h-refinement

The method of manufactured solutions is used to verify the convergence of the Galerkin discretization of (20)
under h-refinement. We use a geometry that is flat, so the source term may be derived in a convenient global Cartesian
basis. However, we use a distorted parameterization of this flat geometry, so the metric tensor components and
Christoffel symbols are nontrivial spatially-varying functions. The manufactured strain field Ẽ is defined by specifying
its components in the global Cartesian basis {eα } on the domain Γ0 = (−1, 1)2 . Its components are chosen such that
the flux-free natural boundary condition is identically satisfied on ∂Γ0 :
αβ
Ẽ♯ = ε̃ eα ⊗ eβ = cos π X 1 m αβ eα ⊗ eβ ,
( )
(34)
where X 1 is the coordinate associated with the global basis vector e1 , and m is a constant second order tensor with
m 11 = 1, m 22 = 3, and m 12 = m 12 = 2. The components of the source term ε in (20) are straightforwardly computed
3 The Kronecker delta δ AB gives the components of the 3D metric tensor in Cartesian coordinates.
160 M.S. Pigazzini, D. Kamensky, D.A.P. van Iersel et al. / Computer Methods in Applied Mechanics and Engineering 346 (2019) 152–179

Fig. 2. Convergence (in L 2 ) of the strain-smoothing method, using quadratic NURBS basis functions.

in Cartesian coordinates from (8), then transformed into covariant components εαβ for consistency with Eq. (20). The
components cα β of the smoothing tensor are set to δ α β .
Convergence testing is performed using both regular and skewed parameterizations of Γ0 . For the current test, we
employ quadratic NURBS shape functions. The L 2 norm of the error ∥Ẽ − Ẽh ∥ is reported in Fig. √ 2, showing that the
error converges at the optimal cubic rate with respect to the characteristic element size lele = 2 ∥G1 ⊗ G2 ∥ averaged
over the mesh.
A visual comparison of the numerical results obtained with different distorted meshes is shown in Fig. 3. The mesh
skew factor is computed for quadrilateral elements as SF = max ∥ cos α∥, where α is an approximation of the angle
between G1 and G2 at the element center.

6.2. Uniaxial tensile test

The second benchmark simulates a uniaxial tensile test on a slender bar [12]. The purposes of this test are:
• To investigate damage localization using NURBS shape functions of increasing polynomial order and inter-
element continuity, in combination with local damage modeling;
• To verify that, if nonlocal regularization is employed, the numerical solution converges under mesh refinement.
The coupon measures 100 mm and 10 mm in the longitudinal and lateral directions, as shown in Fig. 4. The bar has a
constant thickness of 1 mm except for a central section, which extends for a length of 10 mm, where the thickness is
reduced to 0.9 mm. A clamped boundary condition is imposed at X 1 = 0 mm, while uniform displacement is imposed
at X 1 = 100 mm. The test is performed under displacement control to ensure stability of the simulation during the
softening damage process. The material is isotropic with a Young’s modulus E = 20 GPa. Poisson’s ratio is set to
ν = 0 to simulate a uniaxial stress state.
For this problem, we adopt an isotropic damage model based on a residual-stiffness constitutive law:
S = (1 − diso ) CE , (35)
where S denotes the second Piola–Kirchhoff stress tensor, diso is a scalar isotropic damage variable, and C is the
fourth-order tensor of elastic moduli. A bilinear damage law is introduced to compute the isotropic damage variable
diso as a function of a strain history variable κ:
if κ ≤ κ 0 ,
⎧
⎪ 0
κ κ −κ
⎪
⎨ F 0
diso (κ) = if κ 0 ≤ κ ≤ κ F , (36)
⎩ κ κ −κ
⎪ F 0
⎪
1 if κ > κ F ,
where κ 0 = 1.0−4 corresponds to the onset of damage and κ F = 1.25 × 1.0−2 is the strain corresponding to failure.
The history variable κ will be defined as the axial component of either the strain or the smoothed strain in different
tests below.
 M.S. Pigazzini, D. Kamensky, D.A.P. van Iersel et al. / Computer Methods in Applied Mechanics and Engineering 346 (2019) 152–179 161

Fig. 3. Convergence test performed by using the manufactured solution method. Plot contours represent (a) the distribution of the skew factor of
h
the mesh and (b) the Ẽ11 component referred to the global Cartesian basis {eα }.

Fig. 4. Schematic depiction of the uniaxial tensile test simulation. Section view in the longitudinal direction.

6.2.1. Damage localization under k-refinement

Isogeometric analysis introduces the concept of k-refinement (cf. [22]), which simultaneously elevates polynomial
order and inter-element continuity of the approximation space. We illustrate the effect of k-refinement on damage
localization using three different B-spline discretizations: (a) 160 linear elements with C 0 inter-element continuity;
(b) 160 quadratic C 1 elements; (c) 160 cubic C 2 elements. Element size in the longitudinal direction is constant and
equal to 0.625 mm, while one element is used in the lateral direction. (Note that linear C 0 B-splines correspond to
linear finite elements.) For these simulations, κ in (36) is the local (i.e., non-smoothed) axial component of strain, E11 .
Force–displacement curves from numerical simulations are shown in Fig. 5. Results obtained using linear C 0 shape
functions show that, once permanent damage begins, the specimen instantaneously loses its stiffness without energy
dissipation. This paradoxical result stems from the discontinuous nature of the strain field computed from linear
C 0 shape functions. Strain localizes in the element where the damage initiation criterion is initially met, while the
162 M.S. Pigazzini, D. Kamensky, D.A.P. van Iersel et al. / Computer Methods in Applied Mechanics and Engineering 346 (2019) 152–179

Fig. 5. Uniaxial tensile test: Local damage model. Force–displacement curves obtained from k-refinement on a 160-element mesh.

Fig. 6. Uniaxial tensile test: Local damage model. Simulation results obtained using a discretization of 160 (a) linear C 0 , (b) quadratic C 1 and (c)
cubic C 2 elements. Contour plot of the local axial strain component E11 . Results obtained for an applied displacement ∆ = 0.045 mm. View from
top.

remainder of the bar experiences no damage. Similar results are reported in [41] for a local damage theory in which
a length scale is omitted in the definition of the damage law. In this case, damage localization instability leads, for
sufficiently refined meshes and linear C 0 shape functions, to failure without energy dissipation, which is unphysical.
The problem of damage localization is partially alleviated, as shown in Fig. 6, by employing higher-order shape
functions with higher inter-element continuity. Axial strain does not localize in the single element where damage
initiates. Instead, it localizes over a domain corresponding to the support region of those shape functions which also
have support in the element where permanent damage is initially detected.
It is worth noting that, if local effects related to the discontinuity of the thickness are neglected, the nominal
axial strain is constant in the weak section of the bar. Therefore damage may, theoretically, initiate anywhere in the
region x ∈ [45, 55] mm. However, if at least C 1 shape functions are employed, damage initiation is triggered in
the proximity of the thickness discontinuity, due to overshoot from the approximation of discontinuous strain by
continuous derivatives of the smooth displacement approximation.
Fig. 7 shows the evolution of the local strain axial component E11 on a logarithmic scale as a function of the material
coordinate X 1 and of the applied displacement ∆. The use of linear C 0 shape functions led to strain localization in a
single element while the remainder of the bar unloads instantaneously. Solutions obtained using quadratic and cubic
NURBS shape functions, with higher inter-element continuity, show that, while imposed displacement increases, axial
strain localizes in a narrow sub-domain.
 M.S. Pigazzini, D. Kamensky, D.A.P. van Iersel et al. / Computer Methods in Applied Mechanics and Engineering 346 (2019) 152–179 163

Fig. 7. Uniaxial tensile test: Local damage model. Simulation results obtained using a discretization of 160 (a) linear C 0 , (b) quadratic C 1 and (c)
cubic C 2 elements. Evolution of the local axial strain component E11 along the X 1 coordinate as a function of the applied displacement ∆.

6.2.2. Convergence using the gradient-enhanced nonlocal model

We now investigate the effect of the gradient-enhanced method under h-refinement, taking the parameter κ in (36)
to be the smoothed axial component of strain, Ẽ11 . Convergence is investigated by using four progressively-refined
meshes, containing 40, 80, 160 and 320 quadratic C 1 B-spline elements, uniformly-distributed in the longitudinal
direction. The smoothing parameter c is set equal to 1 mm2 . Force–displacement curves plotted in Fig. 8 show that
the dissipated energy quickly converges to a mesh-independent value under h-refinement.
Fig. 9 shows the distribution of the isotropic damage variable diso as a function of the longitudinal coordinate.
The model based on a 40-element discretization fails to predict a symmetric distribution of the damage variable.
However, results obtained by using progressively-refined meshes show that the damage distribution converges to a
symmetrical solution. The distributions of local axial strain and isotropic damage variable are shown in Fig. 10 and
Fig. 11, respectively, for the 40 and 320 element mesh.

6.3. Three-point bending test

We now consider a three-point bending test as in [13]. We restrict the loading and deformation to the planar
case, which results in a 2D plane-stress formulation. The dimensions of the beam are 2000 mm and 300 mm in the
164 M.S. Pigazzini, D. Kamensky, D.A.P. van Iersel et al. / Computer Methods in Applied Mechanics and Engineering 346 (2019) 152–179

Fig. 8. Uniaxial tensile test: Gradient-enhanced nonlocal regularization. Force–displacement curves obtained by using quadratic C 1 NURBS shape
functions, progressively refined discretization and a smoothing parameter c = 1 mm2 .

Fig. 9. Uniaxial tensile test: Gradient-enhanced nonlocal regularization. Distribution of the isotropic damage variable obtained by using quadratic
C 1 NURBS shape functions, progressively refined discretization and a smoothing parameter c = 1 mm2 . Results obtained for an applied
displacement ∆ = 0.04 mm.

Fig. 10. Uniaxial tensile test: Gradient-enhanced nonlocal regularization. Contour plot of the local axial strain component E11 obtained by using
quadratic C 1 NURBS shape functions, smoothing parameter c = 1 mm2 , (a) 40 and (b) 320 elements. Results obtained for an applied displacement
∆ = 0.04 mm. View from top.

longitudinal and lateral directions, respectively, as shown in Fig. 12. The thickness is constant and equal to 50 mm.
Simply-supported boundary conditions are imposed on the lower corners at X 1 ± 1000 mm and X 2 = 0 mm, while a
prescribed-displacement boundary condition is imposed on the upper edge, over a domain that extends for 100 mm.
The material is assumed to be linear isotropic with a Young’s modulus E = 20 GPa and a Poisson’s ratio ν = 0.2. The
Poisson’s ratio is assumed to remain constant during the damage process. The isotropic softening constitutive model
is defined by (35). For this simulation, the isotropic damage variable diso is computed using an exponential damage
 M.S. Pigazzini, D. Kamensky, D.A.P. van Iersel et al. / Computer Methods in Applied Mechanics and Engineering 346 (2019) 152–179 165

Fig. 11. Uniaxial tensile test: Gradient-enhanced nonlocal regularization. Contour plot of the isotropic damage variable diso obtained by using
quadratic C 1 NURBS shape functions, smoothing parameter c = 1 mm2 , (a) 40 and (b) 320 elements. Results obtained for an applied displacement
∆ = 0.04 mm. View from top.

Fig. 12. Schematic depiction of the three-point bending test simulation.

law that is a function of a history variable δ̂:

)]} if δ̂ ≤ δ̂ ,
{ 0
0
diso (δ̂) = { [ ( (37)
1 − δ̂δ̂0 (1 − α) + α exp β δ̂ 0 − δ̂ if δ̂ > δ̂ 0 ,

where α and β are material parameters, and δ̂ 0 is a critical value of the history variable corresponding to damage
initiation. For this simulation, the onset of permanent damage is determined by δ̂ 0 = 10−4 , while material parameters
are set to α = 0.99 and β = 500. Eqs. (35) and (37) are complemented by a damage loading function
( )
f η, δ̂ = η − δ̂ , (38)

where η denotes an equivalent strain measure. The damage loading function f and the rate of the history variable δ̂˙
satisfy the Kuhn–Tucker conditions which ensure irreversibility of damage:
f δ̂˙ = 0 , f ≤ 0 , δ̂˙ ≥ 0 . (39)
( )
The equivalent strain measure η used for the damage loading function f η, δ̂ is:
√
( ) ⟨ ⟩2 ⟨ ⟩2
η Ẽ = Ẽ1 + Ẽ2 , (40)

where Ẽ1 and Ẽ2 are the in-plane principal strains expressed in the global orthogonal coordinate system. Macaulay
brackets isolate positive components of strain allowing for damage growth only under tensile condition.
The three-point bending test is simulated using three different meshes, shown in Fig. 13. The baseline model uses
a regular mesh with rectangular elements whose edges are oriented parallel to global axes. Progressively-distorted
discretizations are considered to verify the robustness of the gradient-enhanced method. Simulations are performed
by setting the parameter c = 50 mm2 . The distribution of the damage variable is reported in Fig. 14 for different
meshes.
It can be observed that differences in the damage distributions computed using regular and skew symmetrical
meshes are negligible, while minor discrepancies can be observed at the boundaries of the damage zone in the skew
166 M.S. Pigazzini, D. Kamensky, D.A.P. van Iersel et al. / Computer Methods in Applied Mechanics and Engineering 346 (2019) 152–179

Fig. 13. Three-point bending test: Meshes used for the simulation. Contour plot of the mesh skew factor for the (a) Regular, (b) Skewed symmetrical
and (c) Skewed asymmetrical mesh.

Fig. 14. Three-point bending test: Gradient-enhanced nonlocal regularization. Contour plots of isotropic damage variable diso for three different
discretizations: (a) Regular mesh, (b) Skewed symmetrical mesh and (c) Skewed asymmetrical mesh. Results are obtained for a prescribed
displacement ∆ = 1.0 mm and a parameter c = 50 mm2 .

asymmetrical case. However, it is worth noting that the nonlocal modeling approach leads to a symmetrical distribution
of damage even if a non symmetrical discretization is adopted.
Fig. 15 shows force–displacement curves obtained from numerical simulations. No significant discrepancies can be
observed between the baseline and the skewed symmetrical model. Very small differences in the maximum predicted
force can be observed in simulations performed using a skewed asymmetrical mesh.

6.4. Composite plate with a hole under tension

We now consider a benchmark problem using the full constitutive model for a fiber-reinforced composite with
intralaminar damage. The problem consists of a single orthotropic composite ply subjected to tensile load. The
 M.S. Pigazzini, D. Kamensky, D.A.P. van Iersel et al. / Computer Methods in Applied Mechanics and Engineering 346 (2019) 152–179 167

Fig. 15. Three-point bending test: Force–displacement curves computed for three different discretization using the gradient-enhanced nonlocal
regularization. Results obtained for a smoothing parameter c = 50 mm2 .

Fig. 16. Schematic depiction of the tensile test on a composite plate with a circular hole. Uniform displacement boundary condition is applied on
the right edge of the specimen.

specimen and the applied essential boundary conditions are depicted in Fig. 16. Material properties used for the
simulations are reported in Table 1, where X T,C , YT,C and Z L represent material strength in the fiber, matrix and in-
plane shear direction, respectively. Subscripts T and C refer to tensile and compressive loading condition, respectively.
Fracture energies in the fiber and matrix directions are denoted as G X,Y respectively. The reinforcement fibers are
oriented at 45◦ with respect to the global Cartesian axis x1 .
The problem is discretized in space using four quadratic NURBS patches. The baseline mesh, denoted as M1 , is
constructed in analogy with the example reported in [21, Figure 17]. Four progressively refined meshes are considered
in order to perform a convergence study under h-refinement, as shown in Fig. 17. We perform N − 1 uniform h-
refinements of mesh M1 to obtain meshes M N , for N ∈ {2, 3, 4}. The number of displacement degrees of freedom is
2028 for mesh M1 , 7116 for mesh M2 , 26,508 for mesh M3 and 102,156 for mesh M4 . All the meshes are conforming
on the connected edges. Furthermore, the parameterization of the multi-patch geometry is G 1 at patch interfaces. This
is necessary in order to ensure continuity of the smooth strain field Ẽ as a tensor when the function space for its
components is assumed to be continuous.
For the gradient-enhanced nonlocal damage formulation, we adopt isotropic smoothing with a scalar parameter
c = 0.1 mm2 , independent of the mesh. We perform a quasi-static implicit analysis in order to neglect inertial effects.
Viscous regularization of the intralaminar damage variables is enabled, as described in [2, Section 3.3]). The viscous
parameter is ρ mat = 10−4 s. For this example we adopt the complete composite damage model described in Section 4.
The imposed displacement ∆ is increased from zero to 0.1 mm in 100 constant steps. This range of displacements
is parameterized over the interval [0, 1.0]s in terms of a pseudo-time variable. Both the local and nonlocal models
168 M.S. Pigazzini, D. Kamensky, D.A.P. van Iersel et al. / Computer Methods in Applied Mechanics and Engineering 346 (2019) 152–179

Table 1
Material properties for the T800/3900-2 car-
bon/epoxy composite material.
E 1 [GPa] 159.9
E 2 [GPa] 8.96
G 12 [GPa] 6.205
ν12 0.27
X T [MPa] 2840
X C [MPa] 1551
YT [MPa] 55.2
YC [MPa] 165.5
Z L [MPa] 88.2
G X T [N/mm] 179.7
G X C [N/mm] 179.7
G YT [N/mm] 0.419
G YC [N/mm] 1.187

predict damage initiation in the matrix phase on the boundary of the hole. Matrix damage propagates diagonally
away from the hole as the prescribed displacement increases. The final extent of matrix damage is shown, for all
the meshes, using the local damage model in Fig. 18. Results obtained using the nonlocal damage model are shown
in Fig. 19. It appears that damage localization occurs if the local model is adopted. This leads to mesh-dependent
damage distributions. The damage field obtained using the nonlocal model converges quickly, under h-refinement, to
a mesh-independent distribution.
Quantitative comparison of reaction force curves, expressed as a function of the applied displacement, is reported
in Fig. 20. The force–displacement curves from the local model exhibit a clear mesh-dependence, with the precipitous
softening of the specimen as a whole occurring later in finer discretizations. Inspection of the damage distribution
indicates that damage initiates around the same time across the three meshes (starting only slightly earlier on the
more refined meshes). However, damage evolves and spreads more slowly on the more refined meshes. It is worth
noting that the finest local model failed to converge for an applied displacement ∆ ≥ 0.06 mm. Results obtained using
the gradient enhanced nonlocal regularization converge under h-refinement both in terms of damage initiation and
dissipated energy.
Results obtained using the gradient-enhanced nonlocal regularization are affected by the choice of the smoothing
parameter, hence c can be effectively interpreted as a constant of the constitutive material model. Therefore, the
selection of a proper smoothing parameter is beyond the scope of this work as it should be considered part of
the material characterization procedure. In particular, we do not claim that the present choice of c = 0.1 mm2 is
necessarily appropriate for modeling damage in composite materials; the purpose of this section is to illustrate that
nonlocality enables mesh-independence of results.

Remark 5. It is worth noting that function spaces for displacement components are only C 0 continuous at the
patch interfaces, while the constitutive equations of the Kirchhoff–Love shell require, at least, C 1 continuity of the
displacement field. However, in this problem, the curvature-change components of the strain tensor are identically
null. It follows that C 0 continuity is sufficient in order to ensure the compatibility of the discretized displacement field
everywhere in the domain.

6.5. Low-velocity impact on a flat composite laminate

This section revisits the simulation from [4] of a low-velocity impact on a flat composite laminate. We now simulate
this scenario using nonlocal damage modeling and compare with the results of [4]. The rectangular plate measures
152.4 mm × 101.6 mm. Clamped boundary conditions are enforced on all sides over a domain that extends 12.7 mm
from the edges. The effective test window at the center of the coupon measures 127 mm × 76.2 mm.
The plate is made of 24 unidirectional T800/3900-2 carbon/epoxy plies and the total thickness is 4.809 mm.
Laminae are stacked according to the lamination sequence [0/45/90/-45]3s . Material properties used for the simulations
are the same as those used for the benchmark of Section 6.4, given in Table 1. The mass density of the composite
 M.S. Pigazzini, D. Kamensky, D.A.P. van Iersel et al. / Computer Methods in Applied Mechanics and Engineering 346 (2019) 152–179 169

Fig. 17. Tensile test on a composite plate with a circular hole: Progressive h-refinement used for convergence study.

Fig. 18. Tensile test on a composite plate with a circular hole: Contour plot of the matrix damage variable d2 obtained using the local damage
model. Results showed are obtained for an applied displacement ∆ = 0.06 mm.
170 M.S. Pigazzini, D. Kamensky, D.A.P. van Iersel et al. / Computer Methods in Applied Mechanics and Engineering 346 (2019) 152–179

Fig. 19. Tensile test on a composite plate with a circular hole: Contour plot of the matrix damage variable d2 obtained using the gradient enhanced
nonlocal regularization. Results showed are obtained for an applied displacement ∆ = 0.1 mm.

Fig. 20. Tensile test on a composite plate with a circular hole: Force–displacement curves obtained using progressively refined meshes in
combination with (top) local and (bottom) gradient-enhanced nonlocal damage models.

Table 2
Cohesive-interface properties.
Kcoh [N/mm3 ] t0n [N/mm2 ] t0τ [N/mm2 ] G Cn [N/mm] G Cτ [N/mm] η
1 × 104 20.0 36.0 0.6 2.1 1.45

material is assumed to be ρ = 1550 kg m−3 . Cohesive properties are reported in Table 2, where Kcoh is the stiffness of
the cohesive interface, while t0 and G C represent the maximum cohesive traction and the critical energy release rate,
respectively. Subscripts n and τ refer to normal and in-plane opening modes, respectively.
The impacting device has a hemispherical head with a radius of 25.4 mm. The impact energy is 25.1 J. No
ply-grouping is employed; the plate is modeled ply-by-ply using 24 rectangular NURBS surfaces, connected by 23
 M.S. Pigazzini, D. Kamensky, D.A.P. van Iersel et al. / Computer Methods in Applied Mechanics and Engineering 346 (2019) 152–179 171

Fig. 21. Impact on a flat laminate: Simulation result. Correlation of impact-force time history obtained for a 25.1 J impact simulation. Results
obtained using the local damage model [4] are reported for comparison.

cohesive interfaces. The gradient-enhanced method for strain smoothing is enabled on all the NURBS surfaces. A
penalty contact interface [3] is defined between the impactor and the first lamina on top of the composite plate.
For this simulation, we use the same discretization adopted in [4]. The characteristic element size in the center
of the plate is 1 mm and it gradually increases to 1.5 mm closer to the clamped edges. The cohesive model imposes
restrictions on the mesh size. As discussed by Turon et al. [42], a fine mesh resolution is required in order to correctly
capture the propagation of the delamination front. The local damage model [2] is replaced with the nonlocal version
presented in Section 4. Isotropic smoothing is employed due to lack of theoretical evidence supporting an anisotropic
smoothing model. As discussed in Section 4 we relate the smoothing parameter c and length-scale L α of the damage
model. However, length scales implied by dimensional analysis of fracture energy and critical strength in this material
are too small to resolve on practical meshes, and we employ the ad hoc choice of setting c equal to the element
length squared. (Mesh-dependent length scales have previously been used with some success to reduce the resolution
requirements of some phase-field models [43].) The model has a total number of 872,430 Degrees Of Freedom
(DOFs), equally-distributed between structural displacements DOFs and smooth-strain DOFs. The generalized-α
method [44] is used for time integration, with a constant time step of 2 µs.
The impact-force time history, obtained using the gradient-enhanced nonlocal regularization method, is shown in
Fig. 21. Results obtained from experimental testing [4], as well as from a baseline isogeometric model based on
local damage modeling, are also reported for comparison. Numerical results obtained using the gradient-enhanced
regularization exhibit an excellent correlation with experimental data; the peak force is over-predicted by 0.3%, while
no differences can be observed in the impact duration. As discussed in our recent work [4], the global stability of
the analysis performed using a local damage model was not affected by local instabilities associated with material
strain-softening. Differences between local and nonlocal damage modeling stem from regularization of the strain field
used to compute damage variables.
An interesting side-effect of gradient-enhanced damage is that it improves the convergence rate of the nonlinear
solver. Normalized nonlinear residuals, computed at every time step after a fixed number of nonlinear iterations, are
reported in Fig. 22. Gradient-enhanced damage modeling improves convergence by as much as a factor of 10 at some
points relative to local damage modeling.
Ply-by-ply distributions of matrix damage are shown in Fig. 23. Results obtained using a local damage model are
reported for comparison purposes, exhibiting spurious, asymmetrical damage features. With nonlocal regularization,
damage patterns reflect the geometric and material symmetry of the problem data, as expected from a stable,
deterministic model.
Nonlocal damage modeling for in-ply damage also leads to a symmetric distribution of predicted delamination,
as shown in Fig. 24. While in-ply and interlaminar damage modes are not explicitly coupled through constitutive
equations, the interaction of these damage modes is implicitly embedded in the damage mechanics of layered
composites. This indirect influence is correctly captured by the multi-layer analysis framework.
172 M.S. Pigazzini, D. Kamensky, D.A.P. van Iersel et al. / Computer Methods in Applied Mechanics and Engineering 346 (2019) 152–179

Fig. 22. Impact on a flat laminate: Comparison of normalized residual obtained for a constant number of nonlinear iterations at every integration
time-step. Simulation performed using local damage model is used as reference.

Fig. 23. Impact on a flat laminate: Simulation results. Ply-by-ply contour plots of matrix damage: colored areas represent, for each ply, the sub-
domain where d2 ≥ 0.5. Comparison of results obtained using the (a) nonlocal and the (b) local damage model. A clamped boundary condition is
enforced on elements highlighted in red. (For interpretation of the references to color in this figure legend, the reader is referred to the web version
of this article.)

6.6. Low-velocity impact on a stiffener-reinforced composite panel

The last example considered in this paper is the simulation of a low-velocity impact on a stringer-reinforced
composite panel. Numerical results obtained from a simplified model of the reinforced panel were originally reported
in [4]. However, in [4], we used the ply-grouping technique, which improves computational efficiency but, at the
same time, precludes delamination within groups of laminae. This simplification often results in an over-prediction
 M.S. Pigazzini, D. Kamensky, D.A.P. van Iersel et al. / Computer Methods in Applied Mechanics and Engineering 346 (2019) 152–179 173

Fig. 24. Impact on a flat laminate: Simulation results. Interface-by-interface contour plots of inter-laminar damage: colored areas represent, for
each interface, the sub-domain where dcoh ≥ 0.5. Comparison of results obtained using the (a) nonlocal and the (b) local damage model. A
clamped boundary condition is enforced on elements highlighted in red. (For interpretation of the references to color in this figure legend, the
reader is referred to the web version of this article.)

of impact forces. In this section, a full multi-layer model of the stringer-reinforced panel is analyzed, limiting ply
grouping to plies with the same fiber orientation, which are unlikely to delaminate.
The model comprises a flat skin panel and a hat-shaped stringer. An additional single lamina, oriented at 90◦ , is
placed between the other parts. The panel measures 127 mm in the longitudinal direction of the stringers and 203.2 mm
in the lateral direction. The distance between the stringer centerlines is 260.35 mm. The impactor has a hemispherical
head with a radius of 12.7 mm. The total impacting mass is 5.067 kg and the initial kinetic energy is 30 J. For this full
multi-layer simulation, we consider the scenario of an impact occurring on the top of the stringer cap.
The baseline NURBS representation, as shown in Fig. 25, contains:
• 9 NURBS patches to represent the stringer;
• 5 NURBS patches to represent the skin;
• 3 NURBS patches to represent the single 90◦ ply;
• 1 NURBS patch to represent the impactor.
Bending strips are employed to connect NURBS patches that belongs to the same lamina. The bending strip technique
(i.e., see [45]) was developed for IGA purposes, in combination with the thin-shell element model, in order to transfer
bending moment across the boundaries of NURBS patches. The bending strips constitute penalty enforcement of the
kinematic constraint that NURBS patches representing a contiguous structure meet at specified angles. This constraint
applies in both damaged and un-damaged states, so no damage is applied to the bending strips.
Starting from the baseline geometry, the multi-layer model is constructed using the commercial CAD soft-
ware Rhinoceros3D [46], by off-setting NURBS surfaces. The lamination sequence for the skin and stringer is
[45/-45/0/45/90/-45/0/90]s .
The complete multi-layer model is comprised of:
• 213 NURBS surfaces to represent skin, stringer and extra 90◦ ply;
174 M.S. Pigazzini, D. Kamensky, D.A.P. van Iersel et al. / Computer Methods in Applied Mechanics and Engineering 346 (2019) 152–179

Fig. 25. Impact on a reinforced panel: Baseline representation of the stringer-reinforced panel. Colored patches identify (green) skin, (blue) extra
90◦ ply, (purple) stringer and (red) impactor. The narrow yellow bands represent bending strips . (For interpretation of the references to color in
this figure legend, the reader is referred to the web version of this article.)

Fig. 26. Impact on a reinforced panel: Multi-layer representation of the experimental setup for the impact simulation.

• 182 Bending strips.

A total of 201 zero-thickness cohesive interfaces are employed to connect laminae of individual parts, as well as to
connect the skin to the 90◦ ply and the 90◦ ply to the stringer. Material and cohesive properties used for the simulation
are reported in Tables 1 and 2. A clamped boundary condition is imposed on the long sides of the panel, over a domain
that extends for 2.5 mm from the edges. The characteristic element size in the impact area is 1.2 mm. The multi-layer
model, shown in Fig. 26 has a total number of 2,475,726 DOFs, equally distributed between structural displacements
and smoothed strains.
Damage predictions for the low-velocity impact case are reported in Figs. 27 and 28. These figures show contour
plots of maximum matrix and interlaminar damage obtained from the impact simulation. The damage distribution
obtained using the gradient-enhanced damage model is continuous across NURBS patches, where displacement field
is only C 0 continuous. This stems from the definition of the regularized strain as an independent unknown field. C 0
continuity of the smoothed strain is strictly enforced between NURBS patches by merging the corresponding degrees
of freedom on connected patches.

7. Conclusions
The novel multi-layer isogeometric analysis framework for progressive damage simulation of composite laminates
from [4] has been extended to incorporate a nonlocal strain regularization technique. The gradient-enhanced method
aims to re-establish mesh objectivity by alleviating material instabilities, which are related to the strain-softening
constitutive law adopted for the analysis. This is achieved by using a spatially-smoothed strain measure to drive the
continuum damage model.
 M.S. Pigazzini, D. Kamensky, D.A.P. van Iersel et al. / Computer Methods in Applied Mechanics and Engineering 346 (2019) 152–179 175

Fig. 27. Impact on a reinforced panel: Simulation results. Ply-by-ply contour plot of matrix damage variable d2 obtained using the nonlocal damage
model. A clamped boundary condition is enforced on elements highlighted in red. (For interpretation of the references to color in this figure legend,
the reader is referred to the web version of this article.)

Fig. 28. Impact on a reinforced panel: Simulation results. Interface-by-interface contour plot of cohesive damage variable dcoh obtained using the
nonlocal damage model. A clamped boundary condition is enforced on elements highlighted in red. (For interpretation of the references to color in
this figure legend, the reader is referred to the web version of this article.)

The gradient-enhanced formulation presented in this paper is specialized for thin shell elements. Several benchmark
simulations are presented to assess the convergence and accuracy of the isogeometric gradient-enhanced damage
formulation under h- and k-refinement. Results show that nonlocal damage modeling allows mesh-objective damage
prediction. Nonlocal regularization also improves the stability of the analysis, leading to faster convergence of
nonlinear solvers. This compensates for the additional computational cost of computing the smoothed strain field.
Due to the implicit coupling of matrix damage and delamination, regularization of in-ply damage also leads to a more
regular distribution of interlaminar damage. Finally, the introduction of the smoothed strain as an additional unknown
field allows us to obtain a continuous damage distribution across connected NURBS patches.
An important topic for future work is the calibration of the length scale used to smooth strains in the damage
model. As shown in Section 6.2.2, this parameter affects the mesh-independent fracture energy to which the model
converges under refinement, and should be viewed as a constitutive parameter. We may also explore the possibility of
introducing anisotropy into the strain smoothing through more general choices of the tensor c in the formulation of
Section 2.

Acknowledgments
This work was supported by NASA Advanced Composites Project No. 15-ACP1-0021. D. Kamensky was partially
supported through AFOSR Award No. FA9550-16-1-0131. We thank Drs. F. Leone and C. Rose from the NASA
176 M.S. Pigazzini, D. Kamensky, D.A.P. van Iersel et al. / Computer Methods in Applied Mechanics and Engineering 346 (2019) 152–179

Fig. A.29. Uniaxial tensile test: Gradient-enhanced nonlocal regularization. Distribution of the local (diso ) and nonlocal (d̃iso ) isotropic damage
variable obtained on a mesh of 160 elements. The applied displacement is ∆ = 0.045 mm. Results from the strain smoothing method are reported
for comparison.

Fig. A.30. Uniaxial tensile test: Gradient-enhanced nonlocal regularization. Force–displacement curves obtained using the damage-smoothing
method. Results from the strain-smoothing method are reported for comparison.

Langley Research Center for their valuable comments and suggestions. We also thank Prof. H. Kim and A. Ellison
from the Department of Structural Engineering at the University of California, San Diego for providing their expertise
and experimental data for comparison of low-velocity impact simulations.

Appendix A. Comparison of strain-smoothing and damage-smoothing formulations

The purpose of this appendix is to elaborate on our rationale for smoothing strains, instead of damage indices.
Given the current popularity of phase-field fracture modeling, in which a damage-like field satisfies an elliptic PDE,
one might reasonably consider a variant of gradient-enhanced damage which smooths damage indices rather than
strain. This alternative is discussed in [13], and compared with phase-field fracture modeling. However, as suggested
by earlier work on strongly-nonlocal models [18], damage smoothing may give rise to a “stress locking” phenomenon,
in which the material is incapable of complete failure, thus limiting the model’s range of applicability. For additional
discussion on strong and weak models and effects of different smoothing schemes, see [19]. The results presented in
this appendix indicate that the conclusions of [18] apply also to weakly-nonlocal damage-smoothing models.
In particular, we consider a damage-smoothing formulation that softens material using smoothed damage indices
{d̃ℵ }, ℵ ∈ {1, 2, 6}, satisfying the PDEs: Find d̃ℵ ∈ Vd , ℵ ∈ {1, 2, 6}, such that, for all sℵ ∈ Vd ,
∫
cd̃ℵ,α G αβ sℵ,β + d̃ℵ − dℵ sℵ dΓ0 = 0 ,
( ( ) )
(A.1)
Γ0
 M.S. Pigazzini, D. Kamensky, D.A.P. van Iersel et al. / Computer Methods in Applied Mechanics and Engineering 346 (2019) 152–179 177

Fig. B.31. Three-point bending test: Force–displacement curves computed for two different discretizations. Results obtained using the complete
gradient-enhanced (G-E) formulation and the simplified counterpart in which all the terms associated with Christoffel symbols are neglected.

where {dℵ } are the locally-computed damage indices from the formulation of [2] and the space Vd of trial and test
functions is scalar-valued, i.e., ℵ does not refer to a 2D (Greek letter) or 3D (Roman letter) tensor index. As in the
strain-smoothing case, we discretize the PDEs (A.1) by Galerkin’s method and couple them with the displacement
subproblem in a staggered fashion.
When damage smoothing is employed, Cartesian components of the local strain tensor Eαβ replace their smoothed
counterpart in the definition of equivalent displacements δ̂i in (26). Instead, the smoothed damage is used to modify
the definition of tangent moduli in the residual-stiffness approach [2, Section 3], and the material stiffness matrix C̃,
expressed in local material axes, is given by
⎡ ⎤
E 1 ν21 1 − d̃1 1 − d̃2
( ) ( )( )
E 1 1 − d̃1 0
1 ⎢
⎣ E 1 ν21 1 − d̃1 1 − d̃2 ⎦ ,
( )( ) ( )
C̃ = (A.2)
⎥
E 2 1 − d̃2 0
D ( )
0 0 G 12 D 1 − d̃6
where D = 1 − ν21 ν12 1 − d̃1 1 − d̃2 . When using an isotropic constitutive law, all components of the tensor of
( )( )

tangent moduli are scaled by a single isotropic damage variable:

S = 1 − d̃iso CE .
( )
(A.3)
To study stress locking effects in this model, we apply it to the uniaxial tensile test described in Section 6.2. Two
discretizations of 80 and 160 elements are considered, and quadratic C 1 NURBS are used as shape functions. The
smoothing parameter c is set to 1 mm2 . Results are shown in Figs. A.30 and A.29. The distribution of d̃iso as a function
of the axial coordinate shows that, unlike its local counterpart, the smoothed damage variable does not reach the limit
of one anywhere in the bar. Although the damage area is larger with damage smoothing, the force–displacement curve
does not exhibit the expected strain-softening response. These results corroborate the stress-locking effect predicted
for damage-smoothing models in [18].

Appendix B. Sensitivity of the solution with respect to mesh orientation

The presence of Christoffel symbols in the gradient-enhanced formulation assumes relevance when the orientation
of covariant basis vectors is not constant over the domain. We revisit the three-point bending test simulation using a
simplified version of the gradient-enhanced method, in which all terms associated with the Christoffel symbols are
neglected. This tests the sensitivity of the solution to the partial derivatives of local curvilinear basis vectors with
respect to parametric coordinates. Two meshes are considered: A regular mesh and an asymmetrical skewed mesh, as
shown in Fig. 13(a) and (c), respectively. Simulations are performed by setting the parameter c = 50 mm2 .
Fig. B.31 shows the comparison of force–displacement curves obtained using both the complete and the simplified
gradient-enhanced formulation on two different meshes. In case of regular discretization, the local curvilinear
coordinate system is aligned with the global Cartesian axes. Therefore, the Christoffel symbols are identically null and
178 M.S. Pigazzini, D. Kamensky, D.A.P. van Iersel et al. / Computer Methods in Applied Mechanics and Engineering 346 (2019) 152–179

Fig. B.32. Three-point bending test: Gradient-enhanced nonlocal regularization. Contour plots of isotropic damage variable diso , at ∆ = 0.65 mm,
obtained using a skewed asymmetrical mesh, the (a) complete and (b) the simplified gradient-enhanced formulation.

results coincide. However, significant discrepancies can be observed by comparing results obtained from simulations
performed using the asymmetrical skewed mesh. Differences are related to the incompleteness of the simplified
formulation, which introduces inaccuracies in the covariant differentiation of the strain field (12).
Fig. B.32 shows distributions of damage variable diso at ∆ = 0.65 mm obtained using the complete and the
simplified gradient-enhanced formulations on the skewed asymmetrical mesh. It can be observed that the simplified
formulation introduces significant errors, related to mesh dependency, which manifest when damage spreads in the
area corresponding to the severely distorted mesh.

References
[1] R. Talreja, Damage development in composites: Mechanisms and modelling, J. Strain Anal. Eng. Des. 24 (4) (1989) 215–222.
[2] Y. Bazilevs, M.S. Pigazzini, A. Ellison, H. Kim, A new multi-layer approach for progressive damage simulation in composite laminates
based on isogeometric analysis and kirchhoff–love shells. Part I: Basic theory and modeling of delamination and transverse shear, Comput.
Mech. 62 (2018) 563–585.
[3] M.S. Pigazzini, Y. Bazilevs, A. Ellison, H. Kim, A new multi-layer approach for progressive damage simulation in composite laminates
based on isogeometric analysis and kirchhoff–love shells. Part II: Impact modeling, Comput. Mech. 62 (2018) 587–601.
[4] M.S. Pigazzini, Y. Bazilevs, A. Ellison, H. Kim, Isogeometric analysis for simulation of progressive damage in composite laminates, J.
Compos. Mater. 52 (2018) 3471–3489.
[5] R.H.J. Peerlings, W.A.M. Brekelmans, R. de Borst, M.G.D. Geers, Gradient-enhanced damage modelling of high-cycle fatigue, Internat. J.
Numer. Methods Engrg. 49 (12) (2000) 1547–1569.
[6] R.H.J. Peerlings, R. deBorst, W.A.M. Brekelmans, M.G.D. Geers, Localisation issues in local and nonlocal continuum approaches to fracture,
Eur. J. Mech. A Solids 21 (2) (2002) 175–189.
[7] I. Babuška, J.T. Oden, Verification and validation in computational engineering and science: basic concepts, Comput. Methods Appl. Mech.
Engrg. 193 (36–38) (2004) 4057–4066.
[8] G. Pijaudier-Cabot, Z.P. Bazant, Nonlocal damage theory, J. Eng. Mech. 113 (1987) 1512–1533.
[9] Z.P. Bazant, G. Pijaudier-Cabot, Nonlocal continuum damage, localization instability and convergence, J. Appl. Mech. 55 (1988) 287–293.
[10] S.A. Silling, Reformulation of elasticity theory for discontinuities and long-range forces, J. Mech. Phys. Solids 48 (1) (2000) 175–209.
[11] M. Ambati, T. Gerasimov, L.D. Lorenzis, A review on phase-field models of brittle fracture and a new fast hybrid formulation, Comput.
Mech. 55 (2) (2015) 383–405.
[12] R.H.J. Peerling, R. De Borst, W.A.M. Brekelmans, J.H.P. De Vree, Gradient enhanced damage for quasi-brittle materials, Internat. J. Numer.
Methods Engrg. 39 (19) (1996) 3391–3403.
[13] R. de Borst, C.V. Verhoosel, Gradient damage vs phase-field approaches for fracture: Similarities and differences, Comput. Methods Appl.
Mech. Engrg. 312 (2016) 78–94.
[14] A.D. Nguyen, M. Stoffel, D. Weichert, A gradient-enhanced damage approach for viscoplastic thin-shell structures subjected to shock waves,
Comput. Methods Appl. Mech. Engrg. 217–220 (Supplement C) (2012) 236–246.
[15] F. Amiri, D. Millán, Y. Shen, T. Rabczuk, M. Arroyo, Phase-field modeling of fracture in linear thin shells, Theor. Appl. Fract. Mech. 69
(Supplement C) (2014) 102–109, Introducing the new features of Theoretical and Applied Fracture Mechanics through the scientific expertise
of the Editorial Board.
[16] J. Kiendl, M. Ambati, L. De Lorenzis, H. Gomez, A. Reali, Phase-field description of brittle fracture in plates and shells, Comput. Methods
Appl. Mech. Engrg. 312 (Supplement C) (2016) 374–394, Phase Field Approaches to Fracture.
[17] P. Areias, T. Rabczuk, M. Msekh, Phase-field analysis of finite-strain plates and shells including element subdivision, Comput. Methods
Appl. Mech. Engrg. 312 (Supplement C) (2016) 322–350, Phase Field Approaches to Fracture.
[18] M. Jirásek, Nonlocal models for damage and fracture: Comparison of approaches, Int. J. Solids Struct. 35 (31) (1998) 4133–4145.
 M.S. Pigazzini, D. Kamensky, D.A.P. van Iersel et al. / Computer Methods in Applied Mechanics and Engineering 346 (2019) 152–179 179

[19] Z.P. Bažant, M. Jirásek, Nonlocal integral formulations of plasticity and damage: Survey of progress, J. Eng. Mech. 128 (11) (2002)
1119–1149.
[20] S. Hosseini, J.J.C. Remmers, R. de Borst, The incorporation of gradient damage models in shell elements, Comput. Methods Appl. Mech.
Engrg. 98 (2014) 391–398.
[21] T.J.R. Hughes, J.A. Cottrell, Y. Bazilevs, Isogeometric analysis: CAD, finite elements, NURBS, exact geometry, and mesh refinement,
Comput. Methods Appl. Mech. Engrg. 194 (2005) 4135–4195.
[22] J.A. Cottrell, T.J.R. Hughes, Y. Bazilevs, Isogeometric Analysis. Toward Integration of CAD and FEA, Wiley, 2009.
[23] L. Piegl, W. Tiller, The NURBS Book (Monographs in Visual Communication), second ed., Springer-Verlag, New York, 1997.
[24] T. Sederberg, J. Zheng, A. Bakenov, A. Nasri, T-splines and T-NURCCS, ACM Trans. Graph. 22(3) (2003) 477–484.
[25] J. Kiendl, K.-U. Bletzinger, J. Linhard, R. Wüchner, Isogeometric shell analysis with Kirchhoff–Love elements, Comput. Methods Appl.
Mech. Engrg. 198 (49) (2009) 3902–3914.
[26] X. Deng, A. Korobenko, J. Yan, Y. Bazilevs, Isogeometric analysis of continuum damage in rotation-free composite shells, Comput. Methods
Appl. Mech. Engrg. 284 (2015) 349–372.
[27] J.E. Marsden, T.J.R. Hughes, Mathematical Foundations of Elasticity, in: Dover Civil and Mechanical Engineering Series, Dover, ISBN:
9780486678658, 1994.
[28] M. Bischoff, K. Bletzinger, W. Wall, E. Ramm, Models and finite elements for thin-walled structures, Encyclopedia Comput. Mech. 2 (2004)
59–137.
[29] F. Roohbakhshan, R.A. Sauer, Isogeometric nonlinear shell elements for thin laminated composites based on analytical thickness integration,
J. Micromech. Mol. Phys. 01 (03n04) (2016) 1640010.
[30] S. Hosseini, J.J.C. Remmers, C.V. Verhoosel, R. de Borst, An isogeometric solid-like shell element for non-linear analysis, Internat. J. Numer.
Methods Engrg. 95 (2013) 238–256.
[31] S. Hosseini, J.J.C. Remmers, C.V. Verhoosel, R. de Borst, An isogeometric continuum shell element for non-linear analysis, Comput. Methods
Appl. Mech. Engrg. 271 (2014) 1–22.
[32] S. Hosseini, J. Remmers, C. Verhoosel, R. de Borst, Propagation of delamination in composite materials with isogeometric continuum shell
elements, Internat. J. Numer. Methods Engrg. 102 (2015) 159–179.
[33] M.J. Borden, C.V. Verhoosel, M.A. Scott, T.J.R. Hughes, C.M. Landis, A phase-field description of dynamic brittle fracture, Comput.
Methods Appl. Mech. Engrg. 217–220 (2012) 77–95.
[34] M.E. Rognes, D.A. Ham, C.J. Cotter, A.T.T. McRae, Automating the solution of PDEs on the sphere and other manifolds in FEniCS 1.2,
Geosci. Model Dev. 6 (6) (2013) 2099–2119.
[35] P. Hansbo, M.G. Larson, K. Larsson, Analysis of Finite Element Methods for Vector Laplacians on Surfaces, ArXiv e-prints, 2016.
[36] T. Jankuhn, M.A. Olshanskii, A. Reusken, Incompressible fluid problems on embedded surfaces: Modeling and variational formulations,
ArXiv e-prints, 2017.
[37] A. Buffa, G. Sangalli, R. Vázquez, Isogeometric analysis in electromagnetics: B-splines approximation, Comput. Methods Appl. Mech.
Engrg. 199 (17) (2010) 1143–1152.
[38] A. Buffa, J. Rivas, G. Sangalli, R. Vázquez, Isogeometric discrete differential forms in three dimensions, SIAM J. Numer. Anal. 49 (2) (2011)
818–844.
[39] J.A. Evans, Divergence-free B-spline Discretizations for Viscous Incompressible Flows (Ph.D. thesis), University of Texas at Austin, Austin,
Texas, United States, 2011.
[40] J. Kiendl, M.-C. Hsu, M. Wu, A. Reali, Isogeometric Kirchhoff–Love shell formulations for general hyperelastic materials, Comput. Methods
Appl. Mech. Engrg. 291 (2015) 280–303.
[41] I. Lapczyk, J.A. Hurtado, Progressive damage modeling in fiber-reinforced materials, Composites A 38 (2007) 2333–2341.
[42] A. Turon, C. Dàvila, P. Camanho, J. Costa, An engineering solution for mesh size effects in the simulation of selamination using cohesive
zone models, Eng. Fract. Mech. 74 (2007) 1665–1682.
[43] H. Gomez, T.J.R. Hughes, X. Nogueira, V.M. Calo, Isogeometric analysis of the isothermal Navier–Stokes–Korteweg equations, Comput.
Methods Appl. Mech. Engrg. 199 (25) (2010) 1828–1840.
[44] J. Chung, G.M. Hulbert, A time integration algorithm for structural dynamics with improved numerical dissipation: The generalized-α
method, J. Appl. Mech. 60 (1993) 371–375.
[45] J. Kiendl, Y. Bazilevs, M.-C. Hsu, R. Wüchner, K.-U. Bletzinger, The bending strip method for isogeometric analysis of Kirchhoff–Love
shell structures comprised of multiple patches, Comput. Methods Appl. Mech. Engrg. 199 (2010) 2403–2416.
[46] R. McNeel, Associates, Rhinoceros 5 Users Guide. Seattle, WA, USA, 2014.