Available online at www.sciencedirect.

com
ScienceDirect

Comput. Methods Appl. Mech. Engrg. 331 (2018) 159–183

www.elsevier.com/locate/cma

An efficient isogeometric solid-shell formulation for geometrically

nonlinear analysis of elastic shells
Leonardo Leonetti ∗, Francesco Liguori, Domenico Magisano, Giovanni Garcea
Dipartimento di Ingegneria Informatica, Modellistica, Elettronica e Sistemistica, Università della Calabria, 87036 Rende (Cosenza), Italy

Received 13 June 2017; received in revised form 15 November 2017; accepted 19 November 2017
Available online 2 December 2017

Abstract

In this work an isogeometric solid-shell model for geometrically nonlinear analyses is proposed. It is based on a linear
interpolation through the thickness and a NURBS interpolation on the middle surface of the shell for both the geometry and the
displacement field. The Green–Lagrange strains are linearized along the thickness direction and a modified generalized constitutive
matrix is adopted to easily eliminate thickness locking without introducing any additional unknowns and to model multi-layered
composite shells. Reduced integration schemes, which take into account the high continuity of the shape functions, are investigated
to avoid interpolation locking and to increase the computational efficiency. The relaxation of the constitutive equations at each
integration point is adopted in the iterative scheme in order to reconstruct the equilibrium path using large steps and a low number
of iterations, even for very slender structures. This strategy makes it possible to minimize the number of stiffness matrix evaluations
and decompositions and it turns out to be particularly convenient in isogeometric analyses.
⃝c 2017 Elsevier B.V. All rights reserved.

Keywords: Geometric nonlinearities; Isogeometric analysis; Composite shells; Reduced integration; MIP Newton

1. Introduction
In recent years an increasing amount of research has aimed at developing new efficient solid-shell finite elements
(FEs) [1–5] for nonlinear analysis of thin structures. This is due to the advantages of these kinds of elements in
comparison to classical shell ones. In particular, they allow the use of 3D continuum strain measures employing
translational degrees of freedom only, so avoiding complex and expensive rules for updating the rotations. Solid-shell
elements are often based on a linear displacement interpolation in order to achieve computational efficiency and
then exhibit shear locking, also present in traditional shell elements, and trapezoidal and thickness locking, typical
of solid elements [6]. These kinds of locking are usually sanitized by means of Assumed Natural Strain (ANS),
Enhanced Assumed Strain (EAS) [7,8] and mixed (hybrid) formulations [1,9,10]. Solid-shells have been used to
∗ Corresponding author.
E-mail addresses: [email protected] (L. Leonetti), [email protected] (F. Liguori), [email protected] (D.
Magisano), [email protected] (G. Garcea).

https://doi.org/10.1016/j.cma.2017.11.025
0045-7825/⃝ c 2017 Elsevier B.V. All rights reserved.
160 L. Leonetti et al. / Comput. Methods Appl. Mech. Engrg. 331 (2018) 159–183

model composites or laminated beams [3,9,11] and shell structures in both the linear [4,7,12] and nonlinear [1,2,8]
range. Among the most effective and interesting proposals are the mixed solid-shell elements of Sze and co-authors [1]
which extend the initial PT18β hybrid element of Pian and Tong [13] to thin shells and eliminate thickness locking by
means of a modified generalized constitutive matrix. This approach makes it possible, as opposed to EAS, to avoid the
introduction of additional degrees of freedom (DOFs) and to obtain good predictions for multi-layered composites.
Although there is the effective elimination of the interpolation locking, low order solid-shell elements exhibit a poor
behavior when analyzing curved geometries. High order Lagrangian interpolations, on the other hand, have been little
used due to the high number of DOFs and computational cost for the integration and assembly of the quantities [14].
The isogeometric analysis (IGA) [15,16] represents a good alternative to high order Lagrangian FEs. The main
reason for its success is, in our opinion, the way it makes it possible to elevate the order of the shape functions
while practically maintaining the same number of DOFs of linear Lagrangian interpolations. Another notable feature
is that the high order continuity of the shape functions allows the total number of integration points to be reduced
significantly as shown in [17,18] compensating for the computational cost of the assembly of the discrete operators.
Finally, the geometry is reproduced exactly, regardless of the mesh adopted and a simple link between CAD and
structural analysis is available.
These considerations make IGA very attractive, particularly in geometrically nonlinear analysis where a highly
continuous solution is often expected [19–21]. However, there are some difficulties associated with IGA with respect
to traditional finite elements. The use of very high order shape functions eliminates interpolation locking but, at the
same time, increases the computational effort for the integration and the assembly of the discrete quantities and for
the solution of the discrete problem because of the decrease in the stiffness matrix sparsity. For these reasons C 1
and C 2 NURBS interpolations are often preferred, even if they are not immune to locking phenomena. Due to the
inter-element continuity of the interpolation, element-wise reduced integrations and strategies, like ANS [22], only
alleviate, but do not eliminate locking, and so are not effective for very thin shells. For the same reason, mixed
formulations with stress shape functions defined at element level are not able to prevent locking. Conversely, mixed
formulations with continuous stress shape functions have been successfully proposed [23,24]. However, in this way
the total number of DOFs increases with respect to the initial displacement formulation and the static condensation
of the stress variables, usually employed in FE analysis and performed at the element level, can be carried out only
at patch level and as a result is not convenient because it produces a full condensed stiffness matrix. An interesting
alternative is the use of displacement formulations with patch-wise reduced integration rules [17]. These have been
shown to alleviate and, in some cases, eliminate interpolation locking in linear elastic problems [18] employing a
low number of integration points and so significantly improve the computational efficiency. This strategy seems more
attractive than the mixed formulation, since it preserves the stiffness matrix sparsity without introducing additional
unknowns and allows a more efficient integration.
However, when comparing mixed and displacement formulations in path-following geometrically nonlinear
analyses, many authors observed that the mixed ones can withstand much larger step sizes (increments) with a reduced
number of iterations to obtain an equilibrium point and then the equilibrium path. The reason for this is explained
in [25,26] where it is shown that the performances of the Newton method drastically deteriorate in displacement
formulations when the slenderness of the structure increases. Conversely, the Newton method in mixed formulations
is unaffected by this phenomenon, which depends only on the format of the iterative scheme adopted (mixed or purely
displacement based) and also holds when a mixed and a displacement formulation provide the same discrete accuracy.
To eliminate this inconvenience in displacement-based finite elements the Mixed Integration Point (MIP) strategy has
been recently proposed in [27]. It consists of the relaxation of the constitutive equations at each integration point
during the Newton iterative process.
In this work, we propose an isogeometric solid-shell formulation for geometrically nonlinear analyses of
homogeneous and composite multi-layered shells, which uses a linear through-the-thickness interpolation of geometry
and displacements. The nonlinear model is based on a Total-Lagrangian formulation adopting the Green–Lagrange
strain measure. A linearization of the strains and a pre-integration along the thickness direction allow the definition
of a modified generalized constitutive matrix, which effectively eliminates thickness locking without introducing any
additional through-the-thickness DOF [28] and leads to accurate predictions for composites. The displacement field
and the geometry are rewritten in terms of semi-sum and semi-difference of the top and bottom surface quantities.
The model so obtained allows a bidimensional description and interpolation of the geometry and displacements
using 2D NURBS of generic order and continuity. Each control point is equipped with six DOFs but, in contrast
 L. Leonetti et al. / Comput. Methods Appl. Mech. Engrg. 331 (2018) 159–183 161

to traditional shell models, no rotational DOF is employed. Shear and membrane locking, which already occur in
linear elastic problems for low order NURBS [29], are even heavier in the nonlinear range when large displacements
occur. Different patch-wise reduced integration rules [17,18], previously proposed for linear analyses, are investigated
in large deformation problems with the aim of eliminating interpolation locking and increasing the computational
efficiency in the proposed solid-shell model when C 1 and C 2 NURBS are adopted.
The displacement-based solid-shell formulation so obtained seems able to provide accurate solutions, practically
unaffected by locking, without the need of a mixed formulation and the corresponding, previously discussed,
drawbacks. However, it is still plagued, like any displacement formulation, by the slow convergence rate and the lack
of robustness of the Newton method when analyzing slender structures. The MIP strategy, which has been shown to
avoid this inconvenience in the FE context [27], is extended to the proposed IGA framework with the aim of reducing
the iterative effort and making it independent of the slenderness of the structure. Since the computational cost of
evaluating and decomposing the stiffness matrix represents, in IGA, a significant part of the total cost of the analysis,
the main goal is to exploit the high robustness of the MIP approach in order to minimize these operations by means
of a modified Newton scheme, which is usually prevented by the displacement formulations in large deformation
problems.
The paper is organized as follows: Section 2 presents the isogeometric solid-shell model for composite shells;
in Section 3 patch-wise reduced integrations are investigated for the elimination of the interpolation locking in the
nonlinear range and the MIP iterative strategy is illustrated; numerical tests are carried out in Section 4 to both validate
the accuracy of the proposed isogeometric model and highlight the benefits in using MIP in the IGA context; finally
the conclusions are discussed in Section 5.

2. The isogeometric solid-shell model

In this section the isogeometric solid-shell model for the geometrically nonlinear analysis of composite shells is
presented. Starting from the FE model proposed in [1], a solid-shell model with a NURBS interpolation of generic
order on the middle surface of the shell is derived. A Total Lagrangian formulation, based on a Green–Lagrange strain
measure, is adopted.

2.1. NURBS basics

A B-Spline curve is represented as

n
∑ p
u[ξ ] = Ni [ξ ]Pi = N[ξ ]P (1)
i=1
p
where Pi , i = 1 · · · n are control points and Ni (ξ ) are the set of B-Spline basis functions, which are piecewise poly-
nomial functions of order p. The latter are defined by a set of non-decreasing real numbers Ξ = [ξ1 , ξ2 , . . . , ξn+ p+1 ]
known as knot vector. More details on the B-Spline parametrization can be found in [30]. B-spline basis functions are
calculated recursively by using the formula
p ξ − ξi p−1 ξi+ p+1 − ξ p−1
Ni [ξ ] = N [ξ ] + N [ξ ]
ξi+ p − ξi i ξi+ p+1 − ξi+1 i+1
for p ≥ 1 and starting from piecewise constant functions ( p = 0 ) defined as
1, if ξi ≤ ξ ≤ ξi+1
{
Ni0 [ξ ] =
0, otherwise.
B-Spline basis functions have attractive properties: they satisfy the partition of unity that makes them suitable for
discretization methods, have a compact support and are non-zero and non-negative within the knot interval [ξi , ξi+ p+1 ].
The regularity r between two parametric or physical elements is described by the multiplicity of the associated knot
in Ξ . The regularity is given by r = p − s where p and s are the order used for the basis functions and the multiplicity
of the knot ξi respectively.
162 L. Leonetti et al. / Comput. Methods Appl. Mech. Engrg. 331 (2018) 159–183

Fig. 1. Undeformed and deformed shell geometry.

Since B-splines are polynomial functions they are not able to represent circular arcs and conic sections exactly. For
this reason NURBS extend the B-spline concept in order to represent these objects exactly. NURBS are obtained by a
projective transformation of B-splines extending Eq. (1) by using
p
p N [ξ ]wi
Ri [ξ ] = ∑n i p (2)
i Ni [ξ ]wi
as shape functions. It is worth noting that all properties of B-Splines are maintained and, in particular, B-Splines are
retrieved when all the weights are equal.
By applying the tensor product, the NURBS surface is constructed in a similar way to Eq. (1) as
n ∑
∑ m
p q
u[ξ, η] = Ri [ξ ]M j [η]Pi j = N[ξ, η]P (3)
i=1 j=1
p q
where Ξ = [ξ1 , ξ2 ...ξn+ p+1 ] and H = [η1 , η2 ...ηm+q+1 ] are two knot vectors, Ri and M j are the one-dimensional
basis functions over these knot vectors and Pi j defines a set of n × m control points. The tensor product of the knot
vectors Ξ and H defines a mesh of quadrilateral “isogeometric elements”.
Weights, as well as control points of the initial geometry, are provided by the CAD model while suitable algorithms
exist for the refinement required to approximate the unknown solution [15,30]. The geometry is always represented
exactly regardless of the mesh adopted.

2.2. Shell kinematics

We use a Total Lagrangian formulation to identify material points of the current configuration in terms of their
position vector X(ξ, η, ζ ) in the reference configuration and the displacement state d(ξ, η, ζ ), cf Fig. 1

x(ξ, η, ζ ) = X(ξ, η, ζ ) + d(ξ, η, ζ ) (4)

where ξ = [ξ, η, ζ ] denote convective curvilinear shell coordinates with (ξ, η) representing in-plane coordinates and
ζ being the shell thickness coordinate. The covariant basis vectors in the undeformed and deformed configuration are
obtained from the corresponding partial derivatives of the position vectors X and x, respectively

Gi = X,i , gi = x,i = Gi + d,i with i = 1, 2, 3 , (5)

where (),i denotes the partial derivative with respect to ith components of ξ . The contravariant basis vectors follow
j
from the dual basis condition: gi · g j = Gi · G j = δi and the metric coefficients are gi j = gi · g j and G i j = Gi · G j
with (i, j = 1, 2, 3). The motion of material points from the initial reference configuration to the current configuration
 L. Leonetti et al. / Comput. Methods Appl. Mech. Engrg. 331 (2018) 159–183 163

is described by the deformation map F : x → X

∂x
F= = gi ⊗ Gi (6)
∂X
and Einstein convention of summing on repeated indexes is adopted from now on.
Using the deformation gradient in Eq. (6) and the metric tensor coefficients gi j and G i j , the Green–Lagrange strain
tensor can be expressed as
1(
E = FT F − I = Ē i j Gi ⊗ G j
)
(7)
2
with
1( )
Ē i j = X,i · d, j + d,i · X, j + d,i · d, j with i, j = 1, 2, 3 (8)
2
where (·) means scalar product.
Assuming a linear through-the-thickness interpolation the position vector is expressed as

X = X0 [ξ, η] + ζ Xn [ξ, η] (9)

where
1 1
X0 = (X[ξ, η, 1] + X[ξ, η, −1]) , Xn = (X[ξ, η, 1] − X[ξ, η, −1])
2 2
with ζ = 1 and ζ = −1 identifying the top and the bottom surface of the shell respectively.
Similarly, the displacement field d = d0 [ξ, η] + ζ dn [ξ, η] is described as a combination of the displacements
1 1
d0 = (d[ξ, η, 1] + d[ξ, η, −1]) , dn = (d[ξ, η, 1] − d[ξ, η, −1]) .
2 2
The same convective coordinates ξ are used for expressing the interpolation of the discrete model.

2.3. The isogeometric solid-shell element

The kinematics of the solid-shell model derived in 2.2 allows a 2D description of the shell. Following the
isogeometric concept, geometry and displacement field are interpolated, over the element, as follows

X[ξ ] = Nd [ξ ]Xe , d[ξ ] = Nd [ξ ]de (10)

where de = [d0e , dne ] and Xe = [X0e , Xne ] collect the element control points for displacement and geometry. The
matrix Nd [ξ ] collects the interpolation functions

Nd [ξ ] ≡ N[ξ, η], ζ N[ξ, η]

[ ]
(11)
where ζ ∈ [−1, +1] and N[ξ, η] are bivariate NURBS (3), functions of the middle surface coordinates only.
Adopting a Voigt notation, the Green–Lagrange covariant strain components in Eq. (8) are collected in vector
Ē = [ Ē ξ ξ , Ē ηη , 2 Ē ξ η , Ē ζ ζ , 2 Ē ηζ , 2 Ē ξ ζ ]T , that, exploiting Eq. (10), becomes
( )
1
Ē = L[ξ ] + Q[ξ , de ] de , (12)
2
where L[ξ ] ≡ Q[ξ , Xe ] and Q is so defined
deT Nd ,ξT Nd ,ξ
⎡ ⎤

deT Nd ,ηT Nd ,η
⎢ ⎥
⎢ ⎥
⎢ ⎥
⎢dT (Nd ,T Nd , + Nd ,T Nd , )⎥
⎢ e ξ η η ξ ⎥
Q[ξ , de ] ≡ ⎢ ⎥. (13)
⎢
⎢ deT Nd ,ζT Nd ,ζ ⎥
⎥
⎣de (Nd ,ζ Nd ,η + Nd ,η Nd ,ζ )⎦
⎢ T T T ⎥

deT (Nd ,ξT Nd ,ζ + Nd ,ζT Nd ,ξ )

164 L. Leonetti et al. / Comput. Methods Appl. Mech. Engrg. 331 (2018) 159–183

The covariant strains are conveniently linearized with respect to ζ

ē[ξ, η] + ζ χ̄ [ξ, η]
⎡ ⎤

Ē ≈ ⎣ Ē ζ ζ [ξ, η] ⎦ (14)
γ̄ [ξ, η]
where
Ē ξ ξ [ξ, η, 0] Ē ξ ξ ,ζ [ξ, η, 0]
⎡ ⎤ ⎡ ⎤
[ ]
2 Ē ηζ [ξ, η, 0]
ē[ξ, η] ≡ ⎣ Ē ηη [ξ, η, 0] ⎦ χ̄ [ξ, η] ≡ ⎣ Ē ηη ,ζ [ξ, η, 0] ⎦ γ̄ [ξ, η] ≡
⎢ ⎥ ⎢ ⎥
2 Ē ξ ζ [ξ, η, 0]
2 Ē ξ η [ξ, η, 0] 2 Ē ξ η ,ζ [ξ, η, 0]
]T
are collected in the vector of generalized covariant strains ε̄[ξ, η] ≡ ē, Ē ζ ζ , χ̄ , γ̄ . In order to simplify the notation,
[

the dependence of the quantities on ξ, η will be omitted from now on, when clear.
The generalized stress components, once the kinematic model is assumed, are automatically given by assuring the
invariance of the internal work. By collecting the contravariant stress components S̄ ≡ [ S̄ξ ξ , S̄ηη , S̄ξ η , S̄ζ ζ , S̄ηζ , S̄ξ ζ ]T ,
the work conjugate variables with ε̄ are obtained by
∫ ∫ ( )
T T T
W= T
S̄ Ēd V = N̄ ē + M̄ χ̄ + s̄ζ ζ Ē ζ ζ + T̄ γ̄
V Ω
∫ (15)
= σ̄ ε̄dΩ
T
Ω
∫ξ ∫η
where, from now on, Ω (...) dΩ = 2 ξii+1 ηii+1 (...) det(J[ξ, η, 0])dξ dη and J denotes the Jacobian matrix
∫

J[ξ, η, ζ ] = [G1 , G2 , G3 ]T . ]T
The generalized contravariant stresses σ̄ ≡ N̄ , s̄ζ ζ , M̄, T̄ in Eq. (15) are then
[

1 1 1 1
∫ ∫
N̄ ≡ σ̄ p dζ M̄ ≡ ζ σ̄ p dζ
2 −1 2 −1
(16)
1 1 1 1
∫ ∫
s̄ζ ζ ≡ S̄ζ ζ dζ T̄ ≡ τ̄ dζ
2 −1 2 −1
with
⎡ ⎤
S̄ξ ξ [ ]
S̄ξ ζ
σ̄ p = ⎣ S̄ηη ⎦ τ̄ = .
S̄ηζ
S̄ξ η

2.4. The mapping between the parametric and the physical domains

The relation between the contravariant stresses and covariant strains in tensor notation and the corresponding
Cartesian ones is
E = J−1 ĒJ−T and S = JT S̄J, (17)
that in Voigt notation can be written as

E = T E Ē and S = T S S̄ (18)
with T S = T−T
E .
From (9) J = J0 [ξ, η] + ζ Jn [ξ, η] and its inverse can be linearized with respect to ζ as

J−1 [ξ, η] = J−1

0 [ξ, η] + ζ Jn [ξ, η]
−1
(19)

n by satisfying up to the first order in ζ the equation JJ

where it is possible to obtain J−1 −1
= I, that implies

n = J0 Jn J0 .
−1 −1
J−1
 L. Leonetti et al. / Comput. Methods Appl. Mech. Engrg. 331 (2018) 159–183 165

Substituting Eq. (19) in Eq. (18) and maintaining only the linear terms in ζ we obtain the linearized expression of
T E = T E0 + ζ T En . In particular letting
⎡ 0
Tee T0eζ T0eγ
⎡ n
Tee Tneζ Tneγ
⎤ ⎤

T E0 = ⎣ T0ζ e Tζ0ζ T0ζ γ ⎦ , T En = ⎣ Tnζ e Tζnζ Tnζ γ ⎦

⎢ ⎥ ⎢ ⎥

T0γ e T0γ ζ T0γ γ Tnγ e Tnγ ζ Tnγ γ

we have, maintaining the linear terms in ζ only,
e + ζχ
⎡ ⎤

E = T E Ē ≈ ⎣ E ζ ζ ⎦ . (20)
γ
Eq. (20) can be expressed in terms of the generalized strains as
ε = Tϵ ε̄ (21)
where ε = [e, E ζ ζ , χ, γ ] T
and
⎡ 0
Tee T0eζ T0eγ
⎤
03×3
⎢ 0
⎢ Tζ e Tζ0ζ 01×3 T0ζ γ ⎥
⎥
Tϵ = ⎢ ⎥.
⎢ Tn Tn
⎣ ee eζ T0ee Tneγ ⎥⎦
T0γ e T0γ ζ 02×3 0
Tγ γ
Combining Eqs. (12), (14) and (21) it is possible to obtain the Cartesian generalized strain–displacement
relationship
( )
1
ε = L + Q[de ] de . (22)
2

2.5. Modified generalized constitutive matrix

Multi-layered composites can be modeled using layer-wise interpolations [31,32] which provide accurate
interlaminar stress reconstructions or homogenization techniques usually more efficient and suitable for predicting
global behaviors accurately.
When a linear through-the-thickness interpolation is adopted, a generalized constitutive law of the multi-layered
composite can be obtained following [1,33]. It consists of a homogenization technique which imposes a constant with
ζ stress Sζ ζ in order to eliminate thickness locking and obtain an accurate prediction of stresses and displacements.
The material law of the generic lamina, assumed to be orthotropic elastic, can be conveniently expressed in a suitable
reference system {e1 , e2 , e3 } according to the fiber direction as
⎡ ⎤
Ĉ pp Ĉ pζ 0
⎢ T ⎥
Ŝ = ĈÊ with Ĉ = ⎢ ⎣Ĉ pζ Ĉζ ζ 0 ⎦
⎥ (23)
0 0 Ĉt
which furnishes, exploiting the decoupling of the transverse shear components, the inverse law as
] [ ][ ]
F̂ pp F̂ pζ
[
Ê p Ŝ p
= (24)
Ê ζ ζ F̂Tpζ F̂ζ ζ Ŝζ ζ
where symbol ( ˆ ) denotes Cartesian components expressed with respect to {e1 , e2 , e3 } with e3 aligned to ζ and
Ê p = ê + ζ χ̂ . Eq. (24) can be rewritten as
R = F̂ζ ζ + F̂Tpζ D
⎧
⎪
Ŝ p = SÊ p + D Ŝζ ζ
⎪
⎨
with D = −(F̂ pp )−1 F̂ pζ
Ê ζ ζ = −DT Ê p + R Ŝζ ζ ⎪
⎪
S = (F̂ pp )−1 .
⎩
166 L. Leonetti et al. / Comput. Methods Appl. Mech. Engrg. 331 (2018) 159–183

The constitutive law in terms of the quantities N̂ and M̂ is then obtained, integrating along ζ and imposing a
constant with ζ stress Ŝζ ζ = ŝζ ζ , as
S D ζS
⎡ ⎤ ⎡ ⎤ ⎡ ⎤
N̂ ∫ 1 ê
⎣ Ê ζ ζ ⎦ = 1 ⎣−D R −ζ D⎦ dζ ⎣ŝζ ζ ⎦
2 −1 χ̂
M̂ ζ S ζ D ζ 2S
⎡ ⎤⎡ ⎤
S0 D0 S1 ê
= ⎣−D0 R0 D1 ⎦ ⎣ŝζ ζ ⎦
S1 D1 S2 χ̂
which furnishes the thickness locking free generalized constitutive law

σ̂ [ξ, η] = Ĉϵ ε̂[ξ, η] (25)

where
S0 + D0 D0T /R0 D0 /R0 S1 + D0 D1T /R0
⎡ ⎤
0
⎢ D0T /R0 1/R0 D1T /R0 0 ⎥
Ĉϵ ≡ ⎢
⎣(S1 + D0 D1T /R0 )T D1 /R0 S2 + D1 D1T /R0
⎥
0 ⎦
0 0 0 Ĉt0
and
∫ 1
Ĉt0 = Ĉt dζ.
−1

The modified generalized constitutive matrix in the global system {X, Y, Z } is obtained as Cϵ = RϵT Ĉϵ Rϵ with Rϵ a
suitable rotation matrix.

3. Numerical integration and mixed integration points in IGA

In this section, after a brief description of the nonlinear analysis framework, a numerical investigation on the use
of patch-wise integration rules in large deformation problems is carried out. The Mixed Integration Point strategy is
then introduced to improve the performance of the Newton method with the aim of reducing the iterative effort.

3.1. Nonlinear analysis framework

The equilibrium of slender hyperelastic structures subject to conservative loads f [λ] proportionally increasing with
the amplifier factor λ is expressed by the virtual work equation

Φ[u]′ δu − λ f δu = 0 , u ∈ U , δu ∈ T (26)
where u ∈ U is the field of configuration variables, Φ[u] denotes the strain energy, T is the tangent space of U at u
and a prime is used to express the Fréchet derivative with respect to u. U is a linear manifold so that its tangent space
T is independent of u. The discrete counterpart of Eq. (26) is
s δu ≡ Φ ′ [u]δu
{ T
r[u, λ] ≡ s[u] − λ f = 0, with (27)
f T δu ≡ f δu
where r : R N +1 → R N is a nonlinear vectorial function of the vector z ≡ {u, λ} ∈ R N +1 , collecting the configuration
u ∈ R N and the load multiplier λ ∈ R, s[u] is the internal force vector and f the reference load vector. Eq. (27)
represents a system of N -equations and N + 1 unknowns and its solutions define the equilibrium paths as curves in
R N +1 from a known initial configuration u0 , corresponding to λ = 0. We also define the tangent stiffness matrix as

δuT K[u]ũ = Φ ′′ [u]ũδu , ∀ δu, ũ (28)

where δu and ũ are generic variations of the configuration field u and δu and ũ the corresponding discrete vectors.
 L. Leonetti et al. / Comput. Methods Appl. Mech. Engrg. 331 (2018) 159–183 167

The Riks approach [34] is the preferred strategy for solving Eq. (27) by adding a constraint of the shape
g[u, λ] − ξ = 0, which defines a surface in R N +1 . Assigning successive values to the control parameter ξ = ξ(k)
the solution of the nonlinear system
r[u, λ]
[ ]
R[ξ ] ≡ =0 (29)
g[u, λ] − ξ
defines a sequence of points (steps) z(k) ≡ {u(k) , λ(k) } belonging to the equilibrium path. Starting from a known
equilibrium point z0 ≡ z(k) , the new one z(k+1) is evaluated correcting a first extrapolation z1 = {u1 , λ1 } by a
sequence of estimates z j (loops) by a Newton iteration
J̄ż = −R j
{
(30a)
z j+1 = z j + ż
where R j ≡ R[z j ] and J̄ is the Jacobian of the nonlinear system (29) at z j or a suitable estimate. The simplest choice
for g[u, λ] is the linear constraint corresponding to the orthogonal hyperplane
{
nu ≡ M (u j − u(k) )
nu (u − u ) + n λ (λ − λ ) = ∆ξ where
T j j
(30b)
nλ ≡ µ (λ j − λ(k) )
M and µ being some suitable metric factors [26,35], ∆ξ an assigned increment of ξ and
∂R[z]
[ ] [ ]
K̄ −f
J̄ ≈ = T . (30c)
∂z z j nu n λ
The load-controlled scheme is obtained assuming g[u, λ] = λ (see [26] for further details) while keeping K̄ = K[u1 ]
we have the modified Newton scheme. The solution of Eq. (30) is conveniently performed as follows
T j
⎧
⎨ λ̇ = nu K̄r
⎪
n λ + nuT K̄f (31)
⎪
K̄u̇ = λ̇f − r . j
⎩

3.2. Displacement-based isogeometric formulations

∑
The strain energy can be expressed as a sum of element contributions Φ[u] ≡ e Φe [u]
∫ ( )
1 T
Φe [u] ≡ ε Cε dΩe (32)
Ωe 2
where Ωe is the element domain. The first variation of the generalized strains in Eq. (22) can be written as
δε = B[de ]δde with B[de ] = L + Q[de ]
and, then, the first variation of the strain energy is
∫
Φe [u] δu ≡
′
δε Cϵ ε dΩe
( T )
Ωe
∫ (33)
δdeT B[de ]T Cϵ ε[de ] dΩe = δdeT se [de ]
( )
=
Ωe
where se [de ] is the element internal force vector. The second variation of the strain measure is
δ ε̃ = Q[d̃e ]δde = Q[δde ]d̃e
and its kth component can be evaluated as
δ ε̃k = d̃eT Ψk δde .
Letting σ [de ] = Cϵ ε[de ], the following expression holds
∑
σ T δ ε̃ ≡ σk δ ε̃k = d̃eT G[σ [de ]]δde
k
168 L. Leonetti et al. / Comput. Methods Appl. Mech. Engrg. 331 (2018) 159–183

with
∑
G[σ [de ]] = σk [de ]Ψk . (34)
k

The second variation of the strain energy is

∫ ( )
Φe′′ [u]δu ũ ≡ δε T Cϵ ε̃ + δ ε̃ T σ [de ] dΩe = δdeT Ke [de ]d̃e (35)
Ωe
with the element tangent stiffness matrix defined as
∫
B[de ]T Cϵ B[de ] + G[σ [de ]] dΩe .
( )
Ke [de ] ≡ (36)
Ωe

3.3. Locking and patch-wise reduced integration in geometrically nonlinear analysis

The high continuity of the interpolation functions used for the approximation of the displacement field does not
make the formulation immune to interpolation locking phenomena (shear and membrane locking) when low order
NURBS, the most used in practical applications, are employed.
Many strategies for resolving locking phenomena in Lagrangian FEM have been proposed over the years. Among
them, element-wise reduced integrations [36,37], ANS [38,39] and mixed formulations [1] are widely employed.
Unfortunately, all these element-wise approaches are not able to eliminate lockings in the context of IGA, because of
the inter-element high continuity of the NURBS basis.
On the other hand, mixed formulations [23,40] with continuous shape functions for the stresses have been
successfully proposed, providing locking-free models. However, in this way, the total number of unknowns
significantly increases due to the stress variables, which cannot be condensed at element level as is usual in the
FEM context. A patch-wise condensation is still possible, but this does not seem a convenient choice, because it leads
to a full condensed stiffness matrix with negative effects in terms of memory and computational efficiency.
Recently, patch-wise integration rules, which take into account the inter-element high continuity of the displace-
ment interpolation have been proposed [17,18] and applied to linear elastic problems. In our opinion, these works
represent an important development in IGA. The d-dimensional target space of order p and regularity r , labeled as
p
Sr , is exactly integrated by a number of ≈ (( p − r )/2)d integration points per element, distributed over the patch,
significantly lower than in standard Gauss quadrature rules. Their positions and weights are not equal for each element,
but are evaluated, once and for all, in a pre-processing phase and depend on r , p and patch mesh. The algorithms which
provide these kinds of integration rules can be found in [17,18] and are very efficient. Their computational burden is
just a small fraction of the total cost of a linear analysis and negligible compared to a nonlinear analysis.
p
The patch-wise exact integration of a given space Sr also opens up new possibilities for patch-wise reduced
integration schemes. In fact p and r can be selected by the user and are not required to be those for the exact integration
of the problem space. If the integration space presents spurious modes a certain number of quadrature points are added
near the boundary elements in order to remove them and the approximation space is said to be over-integrated and
p
labeled as S̄r [17,18]. With respect to the element-wise reduced integrations, an appropriate selection of the patch-
wise reduced integration rules makes it possible to avoid spurious modes, alleviate or eliminate interpolation locking
in the linear elastic range and further reduce the number of integration points.
This strategy, in our opinion, seems preferable to mixed formulations with continuous stress interpolation since it
does not increase the number of unknowns, preserves the sparsity of the stiffness matrix and makes the integration
efficient. The last one represents a significant part of the total cost of the analysis, in IGA much more than in FEM
formulations, and the reduction of integration points drastically increases the computational efficiency.
In the following we carry out a numerical investigation on different patch-wise integration rules for the proposed
solid-shell formulation in large deformation problems to look for an optimal solution in terms of accuracy, efficiency
and robustness. Remembering that in patch-wise rules the number of integration points n can be different element-by-
element, the strain energy can then be evaluated as
n
1∑
Φe [de ] ≡ ε g [de ]T Cg ε g [de ] wg (37)
2 g=1
 L. Leonetti et al. / Comput. Methods Appl. Mech. Engrg. 331 (2018) 159–183 169

Fig. 2. Curved bar: geometry and loads.

the internal force vector as

n
∑
Bg [de ]T Cg ε g [de ] wg
( )
se [de ] = (38)
g

while the tangent stiffness matrix is

n
∑
Bg [de ]T Cg Bg [de ] + G g [σ g [de ]] wg
( )
Ke [σ g [de ], de ] = (39)
g

where subscript g denotes quantities evaluated at the integration point ξ g , wg is the product of the corresponding
weight and the determinant of the Jacobian matrix J evaluated at the integration point, Cg is Cϵ at the integration
point.
Note that Ke [de ] is written as Ke [σ g [de ], de ] as a reminder of the way it is computed.

3.3.1. Interpolation locking tests

In the linear elastic range, low order NURBS interpolations usually exhibit shear and membrane locking. In the
nonlinear range, when a Total Lagrangian formulation is adopted, other similar locking phenomena occur due to the
different approximation of the linear and quadratic part of the Green–Lagrange strain components. This means that
locking occurs in nonlinear analyses, even if the initial geometry is flat. In both linear and nonlinear cases, locking is
related to the slenderness of the shell.
A series of patch-wise exact and reduced integration schemes for C 1 and C 2 NURBS basis are employed and
compared. For the C 1 interpolation the comparison also includes element-wise reduced integration and the Assumed
Natural Strain technique (ANS) [39], proposed for IGA in [22,29].
Linear analysis. The first test is a classical benchmark to address membrane locking in the linear elastic range. It
regards the clamped curved beam in Fig. 2, which is considered a severe test to assess discrete formulations [22].
The normalized displacement at point A is reported in Fig. 3 for an increasing slenderness and for the different
interpolations and integration schemes. The reference value u A,r e f , from Bernoulli beam theory, is 0.942. A mesh of
10 × 1 elements is employed.
170 L. Leonetti et al. / Comput. Methods Appl. Mech. Engrg. 331 (2018) 159–183

Fig. 3. Curved bar: linear displacement for different slenderness ratio.

Fig. 4. Cantilever beam under two load cases: geometry and loads.

Concerning the C 1 interpolation, the full S04 integration presents a very strong locking and provides bad results also
for R/t = 100 and completely wrong results for R/t = 1000. The ANS technique gives good results for R/t = 100
slightly alleviating locking but it is not satisfactory for R/t = 1000. Furthermore the 2 × 2 Gauss element-wise
reduced integration shows the identical results as ANS, which employs a grid of 3 × 3 Gauss points per element. The
S̄02 reduced integration has the best performance and is almost insensitive to locking, a part from the extremely slender
case R/t = 10 000. From the computational point of view, it is worth noting that S̄02 uses about one integration point
per element and is then more efficient than ANS and 2 × 2 reduced integration.
For the C 2 interpolation, the full integration S16 is clearly affected by locking. On the contrary, both the integration
schemes S̄13 and S14 provide excellent results and are practically insensitive to locking. It is worth noting that S̄13
requires about one integration point per element, while S14 about 2.25 integration points per element, so that both
strategies are very efficient compared with Gauss rules.
Nonlinear analysis. In order to show the performances of the different strategies in dealing with locking, the simple
cantilever beam depicted in Fig. 4 is analyses with the proposed solid-shell model, for different values of the
slenderness parameters k = L/t and under two different load conditions.
For the shear load case, Fig. 5 shows the equilibrium paths, up to the maximum value of the load λmax = 4·107 /k 3 ,
obtained with the C 1 interpolation for two different values of k = 100 and k = 1000 and different meshes. The full
S04 integration scheme provides bad results also for the smallest value of k, unless a large number of elements is used,
and completely wrong results for k = 1000. The ANS technique gives good results for k = 100 slightly alleviating
 L. Leonetti et al. / Comput. Methods Appl. Mech. Engrg. 331 (2018) 159–183 171

Fig. 5. Cantilever beam under shear force: equilibrium path for C 1 and L/ h = 100 (left) and L/ h = 1000 (right).

Table 1
Cantilever-beam: normalized end displacement at λ/λmax = 1 for different interpolations and slenderness.
4 elm. 8 elm. 16 elm.
L/ h S̄02 S03 S04 ANS S̄02 S03 S04 ANS S̄02 S03 S04 ANS
C1 102 0,767 0,784 0,610 0,873 0,981 0,953 0,896 0,982 0,992 0,989 0,983 0,993
103 0,343 0,315 0,164 0,661 0,952 0,511 0,385 0,835 0,972 0,828 0,722 0,961
L/ h S̄13 S14 S15 S16 S̄13 S14 S15 S16 S̄13 S14 S15 S16
C2 102 0,977 0,992 0,988 0,961 0,999 1,000 1,000 0,998 1,000 1,000 1,000 1,000
103 0,941 0,988 0,916 0,797 0,998 0,990 0,991 0,946 1,000 1,000 1,000 0,997

locking but it is not satisfactory for k = 1000. Furthermore, the 2 × 2 Gauss element-wise reduced integration shows
the identical results as ANS (3 × 3 Gauss points per element) also in nonlinear context. The S̄02 reduced integration
seems the best choice being almost insensitive to locking effects, except for the coarsest mesh, which is penalized by
the over-integration required to avoid singularities. The general recommendation is to use it with at least 5 elements.
S̄02 is also far more efficient than ANS and 2 × 2 reduced integration.
For the C 2 interpolation, the equilibrium paths of the cantilever beam under shear load discretized with 4 and 8
elements are reported in Fig. 6. Also in this case, the full integration S16 exhibits locking. On the contrary, both the
integration schemes S̄13 and S14 provide very good results. S14 is practically insensitive to locking effects for every mesh,
while S̄13 is slightly penalized for the coarsest mesh due to the over integration required to avoid singularities.
In Table 1 the results previously described are summarized reporting the value of the end beam displacement w A
ref
corresponding to a unitary load normalized with respect to the reference values w A obtained with C 2 interpolation,
4
32 elements and a S1 integration. The table makes the comparison of the different strategies easy and highlights the
great accuracy and insensitivity to locking of the C 2 interpolation when integrated with S̄13 and S14 schemes and the
enormous qualitative leap when passing from C 1 to C 2 . Since the number of DOFs of the C 1 and the C 2 interpolation
as well as the number of integration points, using the same mesh, is almost the same, the cost of the C 2 interpolation
is just slightly higher than the C 1 one.
The second test regards the same cantilever beam under compression, i.e. a standard Euler cantilever beam. A
very small shear imperfection load is added to avoid the jump of the bifurcation. The equilibrium path for different
discretizations, integration schemes and slenderness ratios are reported in Fig. 7 for the C 1 interpolation and in
Fig. 8 for the C 2 one. The load factor is normalized with respect to analytical buckling load λb . Similar comments
172 L. Leonetti et al. / Comput. Methods Appl. Mech. Engrg. 331 (2018) 159–183

Fig. 6. Cantilever beam under shear force: equilibrium path for C 2 and 4 e 8 elements.

Fig. 7. Euler beam under compression force: equilibrium path for C 1 and L/ h = 100 (left) and L/ h = 1000 (right).

to the previous test hold. In particular, the reduced integration schemes S̄02 for C 1 and S̄13 and S14 for C 2 provide
good predictions. However, as in the previous load case only the C 2 interpolation with S14 integration is practically
insensitive to k even for a very coarse mesh, where, conversely, the over-integrated schemes are penalized. Finally,
the C 2 interpolation outperforms the C 1 one again in terms of accuracy, using the same mesh, and then employing a
similar number of DOFs and integration points.
Generally the C 2 interpolation seems preferable to the C 1 due to the possibility of also using coarse meshes,
especially when integrated with the S14 scheme, which is insensitive to locking, also for very slender structures and,
in our opinion, is a more robust choice with respect to the S̄13 scheme. For these reasons, we recommend it among the
strategies investigated. Other numerical tests will be presented in the next section to further validate this proposal.
 L. Leonetti et al. / Comput. Methods Appl. Mech. Engrg. 331 (2018) 159–183 173

Fig. 8. Euler beam under shear force: equilibrium path for C 2 and 4 e 8 elements.

3.4. The iterative scheme with mixed integration points

The isogeometric solid-shell model proposed in Section 2 with the patch-wise reduced integration described in
Section 3.3 is very accurate and efficient and represents a reliable choice from the point of view of the discrete
approximation and the efficiency of the integration. However, the efficiency and the robustness of a nonlinear analysis
do not only depend on the number of unknowns and integration points, but also on the iterative effort, that is on
the capability of the Newton method to converge using a low number of iterations and to withstand large step sizes
(increments). In [25,26], it is shown that the Newton method exhibits a slow convergence and requires a small step
size for slender elastic structures undergoing large displacements when any purely displacement-based formulation
is adopted. This could be considered as a sort of “locking” of the Newton method, since its performance gets worse
when the slenderness of the structures increases. This fact is unrelated to the accuracy of the interpolation and always
occurs in displacement formulations where the stresses σ g [de ], used to evaluate the tangent matrix Ke [σ g [de ], de ],
are forced to satisfy the constitutive equations at each iteration.
Conversely, mixed (stress–displacement) formulations are not affected by this phenomenon, because the stresses
are directly extrapolated and corrected in the iterative process, allowing a faster convergence of the Newton method
and very large steps, independently of the slenderness of the structure. We refer readers to [25,26] for further details
on this phenomenon.
In [27] a strategy called Mixed Integration Point (MIP) has been proposed in order to overcome these limitations in
standard displacement-based FE problems. The approach, however, seems general and it is now extended and tested
in the proposed displacement-based isogeometric formulation.
The fundamental idea of the MIP Newton scheme is to relax the constitutive equations at the level of each
integration point during the iterations. This is made by rewriting the total energy in a pseudo Hellinger–Reissner
form on the element
σ1
⎡ ⎤
⎢ .. ⎥
Πe [ue ] ≡ Φe [ue ] − deT fe with ue = ⎢ . ⎥ (40)
⎢ ⎥
⎣σ n ⎦
de
174 L. Leonetti et al. / Comput. Methods Appl. Mech. Engrg. 331 (2018) 159–183

where fe is the element counterpart of the load vector f and the pseudo “mixed” strain energy Φe [ue ] is obtained by
rewriting Eq. (37) as
n ( )
∑ 1 T −1
Φe [ue ] ≡ σ g ε g [de ] − σ g Cg σ g wg
T
(41)
g=1
2
in which the stresses at each integration point σ g are now independent variables.
The first variation of (41) is
n [ ]T [ ]
∑ δσ g sgσ
Φe δu =
′
wg (42)
δde sgd
g=1

with
sgσ ≡ ε g [de ] − C−1
g σg
{
(43)
sgd ≡ Bg [de ]T σ g .
The second variation of (40) and (41) is
]T [
σ˜g
n
[ ][ ]
∑ δσ g −C−1
g Bg
Φe δu ũ =
′′
wg (44)
δde
g=1
T
Bg G g d˜e
where G g ≡ G e [σ g ] is the matrix G e evaluated at the integration point g, that is now a function of the independent
stresses σ g only.
The linear system in Eq. (31), to be solved at each Newton iteration, can then be rewritten at the element level as
⎤j ⎤j
1 w1 B1 w1 s1σ w1
⎡
−C−1
⎡
.. .. ⎥ σ̇ 1 ..
⎡ ⎤ ⎡ ⎤
0
.
⎢
. .
⎢ ⎥
⎥ ⎢ .. ⎥ ⎢ .. ⎥ ⎢
⎢ ⎥ ⎥
. .
⎢
−Cn wn Bn wn ⎥ ⎢ ⎥ = (λ + λ̇) ⎢ ⎥ − ⎢
−1
w
j
⎢ ⎥
⎢ ⎥ ⎢ ⎥ ⎢ ⎥ ⎢ s nσ n ⎥ (45)
⎣σ̇ n ⎦
⎢
⎢ n ⎥ ⎣0⎦ ⎢ n ⎥
⎥
⎢ ⎥ ∑
BgT σ g wg
∑
⎣ BT w1 . . . BT wn G g wg ⎦ ḋe
( )
fe
⎣ ⎦
1 n
g g

where the superscript on matrices denotes that they are evaluated during the iterative process at the current
j
estimate ue .
By performing a static condensation of the stress correction σ̇ g , locally defined at the level of the integration point,
we obtain
σ̇ g = Cg Bgj ḋe + Cg sgσ
j
= Cg Bgj ḋe + Cg ε gj − σ gj (46)
j j
and, letting rce [de ] = sce [de ] − λ j pe , the linear system in the condensed form becomes
Ke [uej ]ḋe = −rce [dej ] + λ̇fe (47)
with
n ( )
∑ T
Ke [σ gj , dej ] = Bg [dej ] Cg Bg [dej ] + G g [σ gj ] wg (48)
g=1

the condensed tangent stiffness matrix, that has the same expression as the classical displacement based one (39).
However, this time it also depends on the independent stresses at the integration points, which are now directly
extrapolated and corrected during the iterations.
j
Conversely, note that the condensed internal forces sce [de ]
n ( )
∑ T
sce [dej ] = Bgj Cg ε gj wg
g

coincide exactly with the internal forces of the displacement-based formulation in Eq. (38) and then the equilibrium
path reconstructed is exactly the same as the initial displacement formulation.
 L. Leonetti et al. / Comput. Methods Appl. Mech. Engrg. 331 (2018) 159–183 175

Table 2
Schematic description of the principal points of the algorithms: the differences between the standard
Newton and the MIP Newton are marked in red.
Newton MIP Newton
Predictor d1 = d(k) + ∆d d1 = d(k) + ∆d
λ1 = λ(k) + ∆λ λ1 = λ(k) + ∆λ
σ g [d1 ] = Cg ε g [d1 ]
Iteration matrix K[σ g [d j ], d j ]
Residual vector s[d j ] − λ j f s[d j ] − λ j f
New estimate d j+1 = dj + ḋ d j+1 = d j + ḋ
λ j+1 = λ j + λ̇ λ j+1 = λ j + λ̇
j+1
σ g = Cg ε g [d j+1 ]

This iterative scheme is then very close to the standard Newton one for purely displacement models as it is
highlighted in Table 2. The main difference consists of the different value of the stresses at the integration points
used for the evaluation of the tangent stiffness matrix. This means that the computational cost of a MIP iteration is
practically the same as a standard one and only a few changes are required to transform a standard displacement-based
Newton iteration into a MIP one. It is worth noting that the MIP strategy, compared to the mixed formulation, does not
require the definition of shape functions for the stresses and does not modify the expression of the condensed internal
force vector and tangent matrix, preserving the sparsity of the initial displacement formulation and making the static
condensation inexpensive.

3.4.1. The MIP modified Newton

The modified Newton method evaluates and decomposes the iteration matrix at the first extrapolation (predictor)
of each step and represents an attractive choice if the method converges in a reasonable number of iterations.
Unfortunately, in geometrically nonlinear analyses, displacement formulations usually prevent the use of the modified
scheme. In fact, failure easily occurs even for the full Newton and is very likely for the modified method unless a very
small step size and a very large number of iterations are employed.
This is not the case with the previously proposed MIP strategy which, even in the modified Newton version is
more robust and requires a lower number of iterations compared to the traditional full Newton method. We refer
to [27] for further details. In IGA, the modified version of the method seems even more attractive then in low order FE
formulations for two reasons. The first one is that the number of stiffness matrix evaluations, which involve integration
and assembly, becomes much lower and, in IGA, this represents a significant part of the cost of the analysis. The
second reason is that the ratio between the cost of an iteration performed with a new matrix and its cost using an
already decomposed one is directly proportional to the band. This means that the modified Newton becomes more and
more convenient as the order of the NURBS, and then the band of the matrix, increases.

3.4.2. Remarks on the computational cost

The computational cost of the iterative method is directly proportional to the number of iterations for the full
versions and almost directly proportional to the number of steps for the modified version. When the number of DOFs
increases, the cost of an entire step of the modified version tends to the cost of a single iteration of the full method.

3.4.3. MIP newton convergence tests

To highlight the robustness and the efficiency of the proposed MIP strategy, the equilibrium path of the example
in Fig. 6 has been reconstructed using a load-controlled scheme. The maximum load value has been subdivided in
N steps equal load increments. The total number of iterations required to evaluate the equilibrium path with different
strategies, but under the same convergence criteria, are reported in Table 3.
The performances of the standard Newton method clearly depend on the slenderness of the beam. In particular,
its robustness, in terms of increment size worsens when the ratio k = L/t increases and the method is not able to
converge for k = 104 and requires 10 increments to converge for k = 103 . Conversely, the MIP Newton is able to
evaluate the equilibrium point corresponding to λ = λmax with just a single load increment and only 5 iterations.
When the number of load steps N steps increases, the number of iterations per step gets smaller, but the total number
176 L. Leonetti et al. / Comput. Methods Appl. Mech. Engrg. 331 (2018) 159–183

Table 3
Cantilever beam under shear force (C 2 − S14 , L/t = 102 , 103 , 104 ): total number of iterations for the evaluation of the equilibrium path vs the
number of load subdivisions.
Nsteps Newton MIP Newton MIP M. Newton
k k k
102 103 104 102 103 104 102 103 104
1 14 Fails Fails 5 5 5 15 15 15
5 43 Fails Fails 16 16 16 21 21 21
10 62 77 Fails 30 30 30 32 32 32
20 103 121 Fails 51 51 51 51 51 51

Fig. 9. Circular ring: geometry.

of iterations increases. However, even for the smallest step size the MIP Newton is more than twice as efficient as
the standard Newton. Finally, even the modified MIP Newton withstands the largest step size and requires a number
of iterations tending towards that of the full MIP Newton when the step size decreases. In this way the modified
method represents a very good choice, considering that its computational cost is dominated by the number of matrix
decompositions and so of increments, not of iterations. Finally, it is interesting to note that the performances of both
the full and the modified MIP Newton are unrelated to the slenderness k and, in our opinion, this represents the main
advantage of the MIP strategy.

4. Numerical results
In this section, the accuracy of the proposed isogeometric solid-shell model with C 2 interpolation and S14
patch-wise integration, labeled as C 2 -S14 , is tested as well as the performances of the MIP strategy. Geometrically
nonlinear problems are considered for shell structures in both isotropic and composite multi-layered materials. Some
comparisons with FE results are reported. In particular we adopt the well-established hybrid stress linear FE of Sze [1]
in the implementation proposed in [33], based on the Green–Lagrange strains, in order to avoid differences due to the
strain measure. It is labeled as C 0 -H S.

4.1. Slit annular plate subjected to line force

The first test is a circular ring undergoing large displacements, a very popular benchmark in geometrically nonlinear
analysis [1,22]. Geometry, load and boundary conditions are reported in Fig. 9. Fig. 11 shows the equilibrium path
of the ring obtained using C 0 -H S and C 2 -S14 . Three meshes are considered for C 0 -H S: 10 × 6 (420 DOFs), 20 × 6
(840 DOFs) and 30 × 6 (1260 DOFs) elements. The FE needs the finest mesh to obtain a converged curve, while
C 2 -S14 provides the same curve with a mesh of 8 × 3 (576 DOFs). This is mainly due to an exact description of the
circular geometry provided by the isogeometric formulation regardless of the mesh adopted. Conversely, the C 0 -H S,
in the case of curved shell, suffers when coarse meshes are employed, because of the linearized geometry. Observing
the equilibrium path in Fig. 11, obtained using a load-controlled scheme, as well as the evolution of the deformed
 L. Leonetti et al. / Comput. Methods Appl. Mech. Engrg. 331 (2018) 159–183 177

(a) λ = 0.00. (b) λ = 0.04. (c) λ = 0.36. (d) λ = 0.64.

Fig. 10. Circular ring: evolution of the deformed shape.

Fig. 11. Circular ring: equilibrium path.

Table 4
Slit annular plate: total number of iterations for the evaluation of the equi-
librium path vs the number of load subdivisions.
N steps Newton MIP Newton MIP M. Newton
Iters Iters Iters
1 Fails 8 Fails
5 Fails 19 47
10 Fails 33 41
20 Fails 55 60
30 202 73 74

configuration depicted in Fig. 10, this nonlinear problem seems easy to solve. However, if we look at Table 4, reporting
the total number of iterations vs the number of equal load increments N steps used to reach the maximum load value,
it is clear that the standard full Newton method is unable to converge unless a large number of load subdivisions is
employed. On the other hand, the MIP Newton easily converges even if the maximum load is reached using just one
step. The MIP modified Newton fails for the largest step size, but is much more robust than the standard full Newton.
Furthermore, when N steps increases, the total number of iterations of the MIP modified Newton is practically the
same as the standard Newton and, so, the modified version actually becomes the most convenient. Finally, even for
the largest value of N steps the MIP Newton is about three times more efficient than the standard Newton.
178 L. Leonetti et al. / Comput. Methods Appl. Mech. Engrg. 331 (2018) 159–183

Fig. 12. The pinched cylinder: geometry.

Table 5
The pinched cylinder: total number of steps, iterations and normalized elapsed time for the evaluation of the equilibrium path with C 2 -S14 .
Mesh Newton MIP Newton MIP M. Newton
Steps Iters Steps Iters Elapsed timea Steps Iters Elapsed timea
50 × 50 127 507 61 235 0.47 89 352 0.24
a Normalized with respect to Newton elapsed time.

4.2. The pinched cylinder

Another interesting test regarding large deformations is the pinched cylinder depicted in Fig. 12, that has been
studied by several authors [1,28]. Exploiting the problem symmetries only an eighth of the cylinder is analyzed using
C 2 -S14 and C 0 -H S. The equilibrium path of the cylinder is reported in Fig. 13. Three uniform meshes are considered
for C 2 -S14 . The coarsest one 30 × 30 (6208 DOFs) already furnishes a good curve, which, however, is not smooth but
exhibits fluctuations. This phenomenon is already known in literature in both the FE [41] and IGA [28] context when
coarse meshes are employed. It is due to wrinkles developing and moving during the loading process, as can be noted
looking at the evolution of the deformed configuration in Fig. 14. The second mesh adopted for C 2 -S14 is 40 × 40
(10 668 DOFs), which provides a smoother curve that is practically coincident with that provided by the 50 × 50
mesh (16 328 DOFs). Finally the C 0 -H S results obtained with two meshes is also reported. The 40 × 40 mesh (9680
DOFs) gives a good prediction but presents a clear discretization error, which slowly decreases by refining the mesh.
In fact, the curve given by the 75 × 75 mesh (33 900 DOFs) tends towards the isogeometric curve. Again, as in the
previous test, C 2 -S14 converges quickly to the most likely solution because of the exact geometry, while C 0 -H S is
probably penalized by the linearized representation of the curved geometry.
The equilibrium path is obtained using an arc-length path-following analysis with the Riks constraint and an
adaptive step size. The total number of steps and iterations required by the different iterative strategies are illustrated in
Table 5. It is also reported the elapsed time, normalized with respect to that required by the standard Newton strategy.
Even in this test, the MIP Newton outperforms the standard Newton, particularly in the modified version that is clearly
the most efficient choice.
 L. Leonetti et al. / Comput. Methods Appl. Mech. Engrg. 331 (2018) 159–183 179

Fig. 13. Pinched cylinder: equilibrium path.

(a) λ = 1334. (b) λ = 2243. (c) λ = 12 434.

Fig. 14. Pinched cylinder: evolution of the deformed configuration.

4.3. Clamped semi-cylinder

While the results presented so far regard isotropic materials, this benchmark tests the proposed solid-shell model
and the MIP Newton in the case of a composite multi-layered shell. The structure is a semi-cylinder loaded by a
concentrated force at the middle of one of the curved edges, while the other one is clamped. The vertical displacement
of the straight edges is constrained. In Fig. 16, the geometry and the boundary conditions are depicted. Due to its
symmetry, only a half of the structure is analyzed. Two cases are considered: isotropic material, characterized by
E = 2068.50 and ν = 0.3, and a composite multi-layered material. The local reference system, used for defining
the material proprieties, has the direction 1 aligned with the y of the global system and the direction 3 is the
normal to the surface from inside out. The stacking sequences of the laminated material are [90/0/90] and [0/90/0],
measured with respect to the direction 1 of the local reference system and the material properties are E 1 = 2068.50,
E 2 = E 3 = 517.125, ν12 = ν23 = ν13 = 0.3 and G 12 = G 23 = G 13 = 759.58.
Fig. 15 shows the equilibrium paths obtained using the element C 2 -S14 and for the different material cases analyzed.
Two uniform meshes of 20 × 20 elements (2948 DOFs) and 30 × 30 elements (6208 DOFs) are used. The results
180 L. Leonetti et al. / Comput. Methods Appl. Mech. Engrg. 331 (2018) 159–183

Fig. 15. Clamped semi-cylinder: equilibrium paths with element C 2 -S14 for different meshes and layups.

Fig. 16. Clamped semi-cylinder: geometry and deformed configuration at the last evaluated equilibrium point for [90/0/90].

of the coarse mesh are practically identical to those obtained with the finer one, except for the case [90/0/90] which
exhibits small fluctuations, similar to the previously analyzed pinched cylinder, which disappear when the finer mesh
is employed. This behavior is again related to the development of wrinkles as can be observed in the deformed shape
at the last evaluated equilibrium point, pictured in Fig. 16. The results, in both the isotropic and composite cases,
can be compared with the solutions obtained by Abaqus, reported in [42], which are the same as the present ones.
Also in this benchmark the robustness of the MIP strategy is evident. Table 6 shows how the MIP strategy drastically
reduces the number of iterations required to trace the equilibrium path and how the MIP modified Newton is the most
convenient choice in terms of computational time.
 L. Leonetti et al. / Comput. Methods Appl. Mech. Engrg. 331 (2018) 159–183 181

Table 6
Clamped semi-cylinder: total number of steps and iterations for the evaluation of the equilibrium path using 30 × 30 C 2 -S14 elements.
Layup Newton MIP Newton MIP M. Newton
Steps Iters Steps Iters Elapsed timea Steps Iters Elapsed timea
Isotropic 95 382 37 138 0.36 55 216 0.20
[0/90/0] 64 253 32 113 0.44 51 195 0.27
[90/0/90] 92 380 36 142 0.37 62 255 0.23
a Normalized with respect to Newton elapsed time.

(a) N1 , C 2 -S14 30 × 30 elements. (b) M1 , C 2 -S14 30 × 30 elements.

(c) N1 , C 0 -H S 60 × 60 elements. (d) M1 , C 0 -H S 60 × 60 elements.

Fig. 17. Clamped semi-cylinder: generalized stresses at λ = 2000, layup [0/90/0].

Lastly, Fig. 17 shows two significant generalized stress components evaluated with a mesh of 30 ×30 C 2 -S14
elements (6208 DOFs) compared with a reference solution obtained with C 0 -H S and a mesh of 60 × 60 elements
(21 720 DOFs). The concentrated force causes a singularity in the 3D continuum model. The maximum value of the
color map of N1 is then limited to make the comparison clearer over the structure by leaving the singular values just
under the force out, which are also mesh-dependent.
182 L. Leonetti et al. / Comput. Methods Appl. Mech. Engrg. 331 (2018) 159–183

5. Conclusion
In this paper an isogeometric solid-shell formulation for geometrically nonlinear analyses has been proposed. A
linear through-the-thickness interpolation has been adopted for the geometry and the displacement field. This allows
the solid model to be rewritten in a bidimensional way in terms of middle surface quantities, semi-sum and semi-
difference of the top and bottom surface ones. These are interpolated using 2D NURBS shape functions with each
control point equipped with six DOFs, as in standard shell models but without employing rotational DOFs. Multi-patch
structures can be easily modeled because only C 0 continuity is required by the continuum model. A Total Lagrangian
description is used exploiting the Green–Lagrange strain measure, which is linearized through the thickness of the
shell in order to define the generalized quantities. This allows the use of a modified generalized constitutive matrix
which prevents thickness locking and produces accurate results for multi-layered composites without introducing any
additional DOF. A series of patch-wise integrations for C 1 and C 2 NURBS have been investigated in large deformation
problems, in order to obtain an optimal solution in terms of accuracy, efficiency and robustness. The C 2 -S14 formulation
has proved to be the best choice among those analyzed, being practically immune to locking and also accurate for
very coarse meshes and thin shells. With respect to a mixed formulation with continuous stress shape functions, the
proposal seems more attractive because it preserves the stiffness matrix sparsity, does not require any additional DOF,
and just 2.25 integration points per element are needed. However the proposed displacement-based model, like any
displacement formulation, is plagued, in geometrically nonlinear analyses, by a slow convergence of the Newton
method used in reconstructing the equilibrium path. To avoid this drawback a Mixed Integration Point strategy has
been adopted, which gives superior performances with respect to the standard Newton. In particular, MIP allows very
large steps without any loss in convergence and makes it possible to reduce the iterative effort. It has been shown
that the main feature of the MIP strategy is its insensitivity to the slenderness of the structure which, on the contrary,
heavily affects the performance of the displacement formulation. This strategy is so robust that a modified version of
the iterative method, which evaluates and decomposes the stiffness matrix only at the first iteration of each step, can
be conveniently adopted. The proposed formulation is characterized by a high efficiency from the point of view of
the discrete approximation, the numerical integration and the iterative effort, which are all crucial in geometrically
nonlinear analyses. As future work, it would be interesting to extend the proposal to nonlinear dynamics, where the
MIP strategy is expected to have the same impact on the efficiency as that demonstrated in the quasi-static case. In
this context, an appropriate selection of the integration scheme for the mass matrix should be investigated.

References
[1] K. Sze, W. Chan, T. Pian, An eight-node hybrid-stress solid-shell element for geometric non-linear analysis of elastic shells, Internat. J.
Numer. Methods Engrg. 55 (7) (2002) 853–878. http://dx.doi.org/10.1002/nme.535.
[2] S. Klinkel, F. Gruttmann, W. Wagner, A robust non-linear solid shell element based on a mixed variational formulation, Comput. Methods
Appl. Mech. Engrg. 195 (1–3) (2006) 179–201. http://dx.doi.org/10.1016/j.cma.2005.01.013.
[3] M. Schwarze, S. Reese, A reduced integration solid-shell finite element based on the EAS and the ANS conceptGeometrically linear problems,
Comput. Methods Appl. Mech. Engrg. (80) (2009) 1322–1355. http://dx.doi.org/10.1002/nme.
[4] L. Vu-Quoc, X.G. Tan, Optimal solid shells for non-linear analyses of multilayer composites. I Statics, Comput. Methods Appl. Mech. Engrg.
192 (9–10) (2003) 975–1016. http://dx.doi.org/10.1016/S0045-7825(02)00435-8.
[5] Q. Li, Y. Liu, Z. Zhang, W. Zhong, A new reduced integration solid-shell element based on EAS and ANS with hourglass stabilization,
Internat. J. Numer. Methods Engrg. (2015) 1885–1891. http://dx.doi.org/10.1002/nme. arXiv:1010.1724.
[6] K. Sze, Three-dimensional continuum finite element models for plate/shell analysis, Prog. Struct. Engrg. Mater. 4 (2002) 400–407.
[7] S. Klinkel, F. Gruttmann, W. Wagner, Continuum based three-dimensional shell element for laminated structures, Comput. Struct. 71 (1)
(1999) 43–62. http://dx.doi.org/10.1016/S0045-7949(98)00222-3.
[8] M. Schwarze, S. Reese, A reduced integration solid-shell finite element based on EAS and the ANS concept: Large deformation problems,
Internat. J. Numer. Methods Engrg. (85) (2011) 289–329. http://dx.doi.org/10.1002/nme.
[9] K. Sze, A. Ghali, Hybrid hexahedral element for solids, plates, shells and beams by selective scaling, Internat. J. Numer. Methods Engrg.
36 (9) (1993) 1519–1540.
[10] L. Vu-Quoc, X. Tan, Efficient Hybrid-EAS solid element for accurate stress prediction in thick laminated beams, plates, and shells, Comput.
Methods Appl. Mech. Engrg. 253 (2013) 337–355. URL http://dx.doi.org/10.1016/j.cma.2012.07.025.
[11] K. Sze, X. Liu, S. Lo, Hybrid-stress six-node prismatic elements, Internat. J. Numer. Methods Engrg. 61 (9) (2004) 1451–1470. http:
//dx.doi.org/10.1002/nme.1118.
[12] K. Sze, L. Yao, A hybrid stress ANS solid-shell element and its generalization for smart structure modelling. Part I - Solid-shell element
formulation, Internat. J. Numer. Methods Engrg. 48 (4) (2000) 545–564.
[13] T.H.H. Pian, C.-C. Wu, Hybrid and Incompatible Finite Element Methods, Chapman & All, CRC, New-York, 1969.
[14] K. Sze, S.-J. Zheng, S. Lo, A stabilized eighteen-node solid element for hyperelastic analysis of shells, Finite Elem. Anal. Des. 40 (3) (2004)
319–340. http://dx.doi.org/10.1016/S0168-874X(03)00050-7.
 L. Leonetti et al. / Comput. Methods Appl. Mech. Engrg. 331 (2018) 159–183 183

[15] J.A. Cottrell, T.J.R. Hughes, Y. Bazilevs, Isogeometric Analysis: Toward Integration of CAD and FEA, Wiley, ISBN: 978-0-470-74873-2,
2009.
[16] V.P. Nguyen, C. Anitescu, S.P. Bordas, T. Rabczuk, Isogeometric analysis: An overview and computer implementation aspects, Math. Comput.
Simulation 117 (2015) 89–116. http://dx.doi.org/10.1016/j.matcom.2015.05.008.
[17] K.A. Johannessen, Optimal quadrature for univariate and tensor product splines, Comput. Methods Appl. Mech. Engrg. 316 (2017) 84–99.
http://dx.doi.org/10.1016/j.cma.2016.04.030. Special Issue on Isogeometric Analysis: Progress and Challenges.
[18] C. Adam, T. Hughes, S. Bouabdallah, M. Zarroug, H. Maitournam, Selective and reduced numerical integrations for NURBS-based
isogeometric analysis, Comput. Methods Appl. Mech. Engrg. 284 (2015) 732–761. http://dx.doi.org/10.1016/j.cma.2014.11.001.
[19] G. Garcea, R. Gonçalves, A. Bilotta, D. Manta, R. Bebiano, L. Leonetti, D. Magisano, D. Camotim, Deformation modes of thin-walled
members: A comparison between the method of Generalized Eigenvectors and Generalized Beam Theory, Thin-Walled Struct. 100 (2016)
192–212. http://dx.doi.org/10.1016/j.tws.2015.11.013.
[20] G. Garcea, L. Leonetti, D. Magisano, R. Gonçalves, D. Camotim, Deformation modes for the post-critical analysis of thin-walled compressed
members by a Koiter semi-analytic approach, Int. J. Solids Struct. 110–111 (2017) 367–384. http://dx.doi.org/10.1016/j.ijsolstr.2016.09.010.
[21] G. Garcea, F.S. Liguori, L. Leonetti, D. Magisano, A. Madeo, Accurate and efficient a posteriori account of geometrical imperfections in
Koiter finite element analysis, Internat. J. Numer. Methods Engrg. 112 (9) (2017) 1154–1174. http://dx.doi.org/10.1002/nme.5550. nme.5550.
[22] J. Caseiro, R. Valente, A. Reali, J. Kiendl, F. Auricchio, R. Alves de Sousa, Assumed natural strain NURBS-based solid-shell element
for the analysis of large deformation elasto-plastic thin-shell structures, Comput. Methods Appl. Mech. Engrg. 284 (2015) 861–880.
http://dx.doi.org/10.1016/j.cma.2014.10.037.
[23] R. Echter, B. Oesterle, M. Bischoff, A hierarchic family of isogeometric shell finite elements, Comput. Methods Appl. Mech. Engrg. 254
(2013) 170–180. http://dx.doi.org/10.1016/j.cma.2012.10.018. cited By 61.
[24] R. Bouclier, T. Elguedj, A. Combescure, Efficient isogeometric NURBS-based solid-shell elements: Mixed formulation and B-method,
Comput. Methods Appl. Mech. Engrg. 267 (2013) 86–110. http://dx.doi.org/10.1016/j.cma.2013.08.002. cited By 20.
[25] D. Magisano, L. Leonetti, G. Garcea, Advantages of the mixed format in geometrically nonlinear analysis of beams and shells using solid
finite elements, Internat. J. Numer. Methods Engrg. 109 (9) (2017) 1237–1262. http://dx.doi.org/10.1002/nme.5322.
[26] G. Garcea, G. Trunfio, R. Casciaro, Mixed formulation and locking in path-following nonlinear analysis, Comput. Methods Appl. Mech.
Engrg. 165 (1–4) (1998) 247–272.
[27] D. Magisano, L. Leonetti, G. Garcea, How to improve efficiency and robustness of the Newton method in geometrically non-linear
structural problem discretized via displacement-based finite elements, Comput. Methods Appl. Mech. Engrg. 313 (2017) 986–1005.
http://dx.doi.org/10.1016/j.cma.2016.10.023.
[28] S. Hosseini, J.J.C. Remmers, C.V. Verhoosel, R. de Borst, An isogeometric solid-like shell element for nonlinear analysis, Internat. J. Numer.
Methods Engrg. 95 (3) (2013) 238–256 URL http://dx.doi.org/10.1002/nme.4505.
[29] J.F. Caseiro, R.A.F. Valente, A. Reali, J. Kiendl, F. Auricchio, R.J. Alves de Sousa, On the Assumed Natural Strain method to alleviate locking
in solid-shell NURBS-based finite elements, Comput. Mech. 53 (6) (2014) 1341–1353. http://dx.doi.org/10.1007/s00466-014-0978-4.
[30] W.T. Les Piegl, The NURBS book, Springer, 1997. http://dx.doi.org/10.1007/978-3-642-59223-2.
[31] C.H. Thai, H. Nguyen-Xuan, S.P.A. Bordas, N. Nguyen-Thanh, T. Rabczuk, Isogeometric analysis of laminated composite plates using the
higher-order shear deformation theory, Mech. Adv. Mater. Struct. 22 (6) (2015) 451–469. http://dx.doi.org/10.1080/15376494.2013.779050.
[32] Y. Guo, M. Ruess, A layerwise isogeometric approach for NURBS-derived laminate composite shells, Compos. Struct. 124 (2015) 300–309.
http://dx.doi.org/10.1016/j.compstruct.2015.01.012.
[33] D. Magisano, L. Leonetti, G. Garcea, Koiter asymptotic analysis of multilayered composite structures using mixed solid-shell finite elements,
Compos. Struct. 154 (2016) 296–308. http://dx.doi.org/10.1016/j.compstruct.2016.07.046.
[34] E. Riks, An incremental approach to the solution of snapping and buckling problems, Int. J. Solids Struct. 15 (7) (1979) 529–551.
http://dx.doi.org/10.1016/0020-7683(79)90081-7.
[35] G. Garcea, G.A. Trunfio, R. Casciaro, Path-following analysis of thin-walled structures and comparison with asymptotic post-critical solutions,
Internat. J. Numer. Methods Engrg. 55 (1) (2002) 73–100.
[36] O.C. Zienkiewicz, R.L. Taylor, J.M. Too, Reduced integration technique in general analysis of plates and shells, Internat. J. Numer. Methods
Engrg. 3 (2) (1971) 275–290. http://dx.doi.org/10.1002/nme.1620030211.
[37] T. Belytschko, B.L. Wong, H.-Y. Chiang, Advances in one-point quadrature shell elements, Comput. Methods Appl. Mech. Engrg. 96 (1)
(1992) 93–107. http://dx.doi.org/10.1016/0045-7825(92)90100-X.
[38] T. Hughes, T. Tezduyar, Finite elements based upon Mindlin plate theory with particular reference to the four-node bilinear isoparametric
element, J. Appl. Mech. Trans. ASME 48 (3) (1981) 587–596.
[39] M.L. Bucalem, K.-J. Bathe, Higher-order MITC general shell elements, Internat. J. Numer. Methods Engrg. 36 (21) (1993) 3729–3754.
http://dx.doi.org/10.1002/nme.1620362109.
[40] R. Bouclier, T. Elguedj, A. Combescure, Efficient isogeometric NURBS-based solid-shell elements: Mixed formulation and -method, Comput.
Methods Appl. Mech. Engrg. 267 (2013) 86–110. http://dx.doi.org/10.1016/j.cma.2013.08.002.
[41] R. Hauptmann, K. Schweizerhof, A systematic development of solid-shell element formulations for linear and non-linear analyses employing
only displacement degrees of freedom, Internat. J. Numer. Methods Engrg. 42 (1) (1998) 49–69. http://dx.doi.org/10.1002/(SICI)1097-
0207(19980515)42:1<49::AID-NME349>3.0.CO;2-2.
[42] K. Sze, X. Liu, S. Lo, Popular benchmark problems for geometric nonlinear analysis of shells, Finite Elem. Anal. Des. 40 (11) (2004)
1551–1569. http://dx.doi.org/10.1016/j.finel.2003.11.001.