Available online at www.sciencedirect.

com
ScienceDirect

Comput. Methods Appl. Mech. Engrg. 405 (2023) 115848

www.elsevier.com/locate/cma

Geometrically exact isogeometric Bernoulli–Euler beam based on the

Frenet–Serret frame
A. Borkovića,b ,∗, M.H. Gfrerera , B. Marussiga
a Institute of Applied Mechanics, Graz University of Technology, Technikerstraße 4/II, 8010 Graz, Austria
b University of Banja Luka, Faculty of Architecture, Civil Engineering and Geodesy, Department of Mechanics and Theory of
Structures, 78000 Banja Luka, Bosnia and Herzegovina
Received 16 September 2022; received in revised form 18 November 2022; accepted 12 December 2022
Available online 22 December 2022

Abstract
A novel geometrically exact model of the spatially curved Bernoulli–Euler beam is developed. The formulation utilizes the
Frenet–Serret frame as the reference for updating the orientation of a cross section. The weak form is consistently derived and
linearized, including the contributions from kinematic constraints and configuration-dependent load. The nonlinear terms with
respect to the cross-sectional coordinates are strictly considered, and the obtained constitutive model is scrutinized. The main
features of the formulation are invariance with respect to the rigid-body motion, path-independence, and improved accuracy for
strongly curved beams. A new reduced beam model is conceived as a special case, by omitting the rotational DOF. Although
rotation-free, the reduced model includes the torsion of the beam axis, which allows simulations of spatial beams that are
predominantly bent with respect to the binormal. The applicability of the obtained isogeometric finite element is verified via
a set of standard academic benchmark examples. The formulation is able to accurately model strongly curved Bernoulli–Euler
beams that have well-defined Frenet–Serret frames.
© 2022 The Author(s). Published by Elsevier B.V. This is an open access article under the CC BY license
(http://creativecommons.org/licenses/by/4.0/).
Keywords: Spatial Bernoulli–Euler beam; Frenet–Serret frame; Rotation-free beam; Strongly curved beam; Geometrically exact analysis

1. Introduction
The aim of computational mechanics is to develop accurate and efficient models of various mechanical systems.
The most successful mechanical model for the simulation of slender bodies is beam. The first consistent beam
theories were developed in 18th century, and the search for a formulation with optimal balance between accuracy
and efficiency is still ongoing. To reduce the problem domain of slender bodies from 3D to 1D, the standard
assumption is that the cross sections are rigid, which results with the Simo–Reissner (SR) beam model. By additional
assumption that the cross sections remain perpendicular to the deformed axis, the Bernoulli–Euler (BE), also
known as Kirchhoff–Love, beam model follows. The subject of the presented research are large deformations of an
arbitrarily curved and twisted BE beam with an anisotropic solid cross section, without warping [1].
The nonlinear SR beam model has long been the main focus for researchers, partially because its spatial
discretization requires only C 0 -continuous basis functions, such as the Lagrange polynomials. As the name suggests,
∗ Corresponding author at: Institute of Applied Mechanics, Graz University of Technology, Technikerstraße 4/II, 8010 Graz, Austria.
E-mail addresses: [email protected], [email protected] (A. Borković).

https://doi.org/10.1016/j.cma.2022.115848
0045-7825/© 2022 The Author(s). Published by Elsevier B.V. This is an open access article under the CC BY license (http://creativecommons.
org/licenses/by/4.0/).
A. Borković, M.H. Gfrerer and B. Marussig Computer Methods in Applied Mechanics and Engineering 405 (2023) 115848

the SR theory was founded by Reissner [2], and later generalized by Simo [3], who conceived the term geometrically
exact beam theory. The main requirement of a geometrically exact formulation is that the relationship between the
configuration and the strain is consistent with the balance laws, regardless of the magnitude of displacements and
rotations. The adequate description of large rotations is one of the principal challenges since these are not additive
nor commutative and constitute nonlinear manifolds. This issue has been a driving force for the formulation of
various algorithms for the parameterization and interpolation of rotation [4–10]. A turning point in this development
was the finding by Crisfield and Jelenić that the interpolation of a rotation field between two configurations
cannot preserve objectivity and path-independence [11,12]. The reason is that incremental material rotation vectors,
at different instances, do not belong to the same tangent space of the rotation manifold [13]. An orthogonal
interpolation scheme that is independent of the vector parameterization of a rotation manifold is suggested in [11,12]
and several further strategies followed [14–16].
Although the geometrically exact formulations represent the state-of-the-art in beam modeling, their imple-
mentation is not straightforward and several alternatives exist, such as the corotational and the Absolute Nodal
Coordinate (ANC) approaches. The main idea of the corotational formulations is to decompose the deformation
into two parts. The first part is due to large rotations and the second is the local part, measured with respect to
the local co-rotated frame. It resembles the strategy employed in [11,12] and allows accurate simulation of large
deformations [17,18]. Recently, a novel approach has been developed using the incremental nodal rotation vectors
to define corotational nodal rotations that are then interpolated [19]. The obtained formulation is objective and path-
independent, regardless of the mesh and approximation functions. Regarding the ANC method, it can be considered
as a solid finite element for slender bodies. The method is well-suited for the implementation of 3D constitutive
models, but has issues with engineering structural analysis, where integration with respect to the cross-sectional
area is required [20,21].
The first BE beam models that are consistent with the geometrically exact theory are [22,23]. Meier et al. have
discussed the issues of objectivity and path-independence in [24], and proposed an orthogonal interpolation scheme
similar to that of Crisfield an Jelenić [11,12]. Membrane locking, contact, and reduced models are considered in
subsequent publications [25,26], followed by a comprehensive review [1]. An efficient BE beam model based on
the Cartan frame was developed in [27], where the position and the local frame are observed independently and
later related by the Lagrange multipliers.
The emergence of the spline-based isogeometric analysis (IGA) [28] has led to the development of a series of
SR beam models [29–36]. The formulation [37] stands out since it employs extensible directors and models various
couplings, thus bringing the beam theory and the complete 3D solution closer. An efficient and accurate isogeometric
version of [19] is given in [38]. The number of DOFs is almost halved, since only rotations are discretized. One
of the main features of IGA is the smoothness of utilized basis functions, a property that benefits the BE beam
due to its C 1 -continuity requirement. The first IGA BE beam models were introduced by Greco et al. in [39–42],
while the first nonlinear BE model was developed in [43]. Due to the reduction in number of DOFs, in comparison
with the SR model, multi-patch nonlinear analysis of BE beams has received special attention [44–47]. Invariance
of the geometric stiffness matrix in the frame of buckling analysis is considered in [48], while the effect of initial
curvature on the convergence properties of the solution procedure is considered in [49]. The first truly geometrically
exact IGA BE model that preserves objectivity and path-independence was developed in [50].
As emphasized in this brief literature review, the crucial issues of objectivity and path-independence in the
geometrically exact beam theory are related to the nonlinear nature of finite rotations. In order to obtain a generally
applicable formulation, the orthogonal interpolation schemes or similar procedures must be applied [24]. An
alternative approach is to utilize the Frenet–Serret (FS) triad as the reference frame for the update of rotation.
This frame does not depend on previous configurations, and the resulting formulation is expected to be objective
and path-independent. Although a natural choice, the FS frame is avoided for beam analysis since it is not defined
for straight segments of a curve. Furthermore, the FS frame exhibits significant rotation around the curve’s tangent
vector at inflection points. Due to these issues, the formulation based on the FS frame fails for arbitrary geometries.
Nevertheless, the derivation, implementation and verification of such a computational model is of fundamental
importance due to the intrinsic relation between the FS frame, the curve and the beam model. In this paper, we
develop a formulation of this kind. Configurations for which the FS frame is not well-defined are not generally
considered. An approximation of straight initial configuration will nevertheless be studied by imposing a small
curvature. Regarding the inflection points, it can be shown that for a regular analytic space curve, which is not a
2
A. Borković, M.H. Gfrerer and B. Marussig Computer Methods in Applied Mechanics and Engineering 405 (2023) 115848

straight line, a point with a zero-curvature is the point of analyticity of torsion [51]. This means that the torsional
angle is defined at inflection points. However, due to the large gradients of this angle, the FS frame exhibits
significant twisting and poor convergence is expected [52].
The calculation of torsion of the FS frame involves the third order derivative of the position vector, implying that
a C 2 -continuous discretization is required to obtain the torsion field. IGA allows high interelement continuities, up
to C p−1 , where p is the order of the basis functions. This feature makes IGA ideally suited for the implementation
of the beam model based on the FS frame. This fact was utilized in [52] for the linear analysis of such beam model.
There are several rotation-free formulations in the literature that model a spatial BE beam, e.g. [25,53]. However,
they are designed for specific cases of initially straight beams without torsional loading. Starting from the proposed
model based on the FS frame, on the other hand, we can derive a rotation-free model capable of approximating
some specific cases that are not torsion-free since it contains the torsion of the FS frame.
When a spatially curved beam exhibits large deformations, axial, torsional and bending actions become coupled
due to the nonlinear distribution of strain along the cross section. It is common to disregard these couplings when
modeling the BE beam [39,43]. Recently, axial-bending coupling was considered in the frame of linear [52,54] and
nonlinear analysis [50,55]. The curviness of a beam, K d, is introduced in [56,57] as a measure of this coupling.
Here, K is the curvature of beam axis and d is the maximum dimension of the cross section in the planes parallel
to the osculating plane. The curviness parameter allows classification of beams as small-, medium- or big-curvature
beams [58]. The axial-bending coupling is significant for K d > 0.1, and these beams belong to the category of
big-curvature, also known as strongly curved, beams [59]. In order to apply appropriate beam models, the current
curviness at each configuration must be observed [50,55]. The effect of large curviness is generally followed by
shear strains. However, the aim is to examine the impact of the higher-order metric on the simple BE model and
extend the approach to more comprehensive beam models.
The present research is based on the works [50,52,54,55,60] with the aim of extending the linear formulation
based on the FS frame to the geometrically exact setting and to improve the existing strongly curved BE model. To
summarize, this work makes three main contributions. The first is the derivation of the geometrically exact FS beam
formulation. It is geometrically exact in a sense that it can model arbitrary large deformations involving finite, but
small, strains. The restriction is that the beam must have a well-defined FS frame during the deformation, meaning
that straight segments and inflection points should not occur. It turns out that many academic examples satisfy this
requirement. The second contribution is the consideration of a special case for the proposed formulation that is
obtained by omitting the rotational DOF. Although rotation-free, this model includes one part of the torsion and
can give approximate results for specific deformation cases. The third contribution is a rational constitutive model
for strongly curved beams. The nonlinear terms of total strain with respect to the cross-sectional axes are taken
into account and simplified models are deduced. This approach improves upon the strongly curved beam models
considered in [50,55] where only the linear terms of incremental strain were considered.
The paper is organized as follows. The next section presents the basic relations of the beam metric and kinematic,
while the strain and stress measures are defined in Section 3. The finite element formulation is elaborated in Section 4
and numerical examples are presented in Section 5. The conclusions are delivered in the last section.

2. Configuration of Bernoulli–Euler beam

A spatial BE beam at an arbitrary reference configuration C is defined by the position of its axis and the
orientation of cross sections. The position of a spatial curve is a vector, while the orientation of the rigid cross
section is, in general, described with the rotation tensor. Since we are dealing with the BE beam, the cross sections
are perpendicular to the beam axis at each configuration, which leaves us to determine only the rotation in the
cross-sectional plane. Therefore, the metric of deformed configuration is defined analogously to the metric of an
arbitrary reference configuration.
Boldface letters are used for the notation of vectors and tensors. An asterisk sign is used to designate a current,
unknown, configuration. Greek index letters take the values of 2 and 3, while Latin ones take the values of 1, 2
and 3. Partial and covariant derivatives with respect to the mth coordinate of the convective frame (ξ, η, ζ ) are
˙ Other specific designations will be
designated with (•),m and (•)|m , respectively. Time derivative is marked as (•).
introduced as they appear in the text.
The details on the NURBS-based IGA modeling of curves are skipped for the sake of brevity since they are
readily found elsewhere [28,61].
3
A. Borković, M.H. Gfrerer and B. Marussig Computer Methods in Applied Mechanics and Engineering 405 (2023) 115848

Fig. 1. Reference and current configurations of a spatial BE beam. A configuration is defined with the position vector of beam axis and
the orientation of cross sections.

2.1. Metric of the beam axis

The beam axis is a spatial curve, defined with its position vector:
r = r(ξ ) = x m (ξ )im = x m im , (x 1 = x, x 2 = y, x 3 = z), (1)
where im are the base vectors of the Cartesian coordinate system, Fig. 1. A curve can be parameterized with either
the arc-length coordinate s ∈ [0, L] or some arbitrary parametric coordinate ξ ∈ [0, 1], where L is the length. For
every C 1 continuous curve, we can uniquely define a tangent vector g1 :
dr dr ds ds √ √ √
g1 = r,1 = m
= x,1 im = = t = gt, g = g1 · g1 , (2)
dξ ds dξ dξ
where t is the unit-length tangent of the beam axis.
There is an infinite set of local vector bases that can be defined to frame a curve, FS and Bishop frames being
the most prominent ones. The unique feature of the FS frame is that one of its base vectors is aligned with the
curvature of a line, while the Bishop frame is characterized with zero torsion [62]. One approach to the BE beam
modeling using the ANC formulation and the Bishop frame can be found in [63]. ( )
Let us now focus on the FS triad that consists of the tangent, normal and binormal t, n, b . The normal vector
is the unit vector of curvature, while the binormal is perpendicular to the osculating plane and completes the
orthonormal FS triad. Due to its intrinsic connection to the curvature, the FS frame cannot be defined for straight
lines and has sudden changes near the inflection points. These issues are readily discussed in the context of beam
formulations [24,52]. The derivatives of FS base vectors are defined with the well-known formulae:
⎡ ⎤ ⎡ ⎤⎡ ⎤
t,s 0 K 0 t
⎣n,s ⎦ = ⎣−K 0 τ ⎦ ⎣n ⎦ , (3)
b,s 0 −τ 0 b
where K is the curvature of a line while τ is the torsion of the FS frame.
The orientation of the cross section is here defined with two base vectors of unit length, g2 and g3 , that are
aligned with the principal axes of inertia of the cross section, Fig. 1. We will refer to the triad gi as the material
vector basis. If the cross sections and the beam axis are orthogonal, the orientation of the cross sections is uniquely
4
A. Borković, M.H. Gfrerer and B. Marussig Computer Methods in Applied Mechanics and Engineering 405 (2023) 115848

ref
defined with some reference basis vectors gα and an appropriate angle. To obtain the relation between components
ref
of curvature with respect to the FS and material vector bases, let us assume that the reference basis vectors gα
are the normal and binormal. Then, by introducing the angle θ r e f , we can define orientation of the cross section:
cos θ r e f sin θ r e f
[ ] [ ][ ]
g2 n
= . (4)
g3 − sin θ r e f cos θ r e f b
In this way, the material vector basis is completed, and its metric and reciprocal metric tensors are:
⎡ ⎤ ⎡ 11 ⎤
g11 0 0 g 0 0
1
gi j = ⎣ 0 1 0⎦ , g i j = ⎣ 0 1 0⎦ , det gi j = g11 = g = g1 · g1 , g 11 = .
( )
(5)
0 0 1 0 0 1 g11

Furthermore, the derivatives of material vector basis are:

⎡ ⎤ ⎡ 1 ⎤⎡ ⎤
g1,1 Γ11 K̃ 3 − K̃ 2 g1
gi,1 = Γi1k gk , ⎣g2,1 ⎦ = ⎣−K 3 0 K 1 ⎦ ⎣g2 ⎦ , (6)
g3,1 K 2 −K 1 0 g3
where Γi1k are the Christoffel symbols of the second kind, while K i are the components of the curvature vector
with respect to the reciprocal basis gi . In concrete terms, K 1 is the torsion of the material vector basis, while K α
are the appropriate components of curvature. Due to an arbitrary parameterization of the axis, K̃ α = gK α are the
components of curvature measured with respect to the parametric coordinate:
K 1 = g2,1 · g3 , K̃ 2 = −g1,1 · g3 , K̃ 3 = g1,1 · g2 . (7)
By using Eqs. (4), (6) and (7), the relation between components of curvature of the two local vector triads are:
1 1 ref ref
K 1 = √ τ + θ,1 H⇒ K 1 = g11 K 1 = τ̃ + θ,1 ,
g g (8)
K = K 2 = K sin θ , K = K 3 = K cos θ ,
2 ref 3 ref

√
where τ̃ = gτ is the FS torsion with respect to parametric coordinate. Finally, let us note the relation between
the normal and binormal, and the tangent vector g1 :
1 ( 1 (
1
g1 , b = √ g1 × n ,
) )
n= g1,1 − Γ11 (9)
K̃ g
where K̃ = K g is the modulus of curvature with respect to the parametric coordinate.

2.2. Metric of a generic point in beam continuum

In order to reduce the problem from 3D to 1D, the metric of a beam continuum will be represented via a finite
set of reference quantities. Let us define an equidistant line which is a set of points for which (η, ζ ) = const. Its
position and tangent base vectors are:
r̄ (ξ, η, ζ ) = r̄ = r + ηg2 + ζ g3 ,
(10)
ḡ1 (ξ, η, ζ ) = ḡ1 = r̄,1 = g1 + ηg2,1 + ζ g3,1 ,
¯ designates quantities at equidistant line. In order to enable a concise derivation, we will abuse the notation
where (•)
slightly by setting: ξ 2 = ξ2 = ζ and ξ 3 = ξ3 = −η. The base vectors of an equidistant line are now:
ḡ1 = g0 g1 − ζ K 1 g2 + ηK 1 g3 = g0 g1 − ξ α K 1 gα , ḡ2 = g2 , ḡ3 = g3 . (11)
The quantity g0 = 1 + ξ α K α is sometimes referred to as the shifter [64] or the curvature correction term [65]. The
metric tensor at an equidistant line is:
−ζ K 1 ηK 1
⎡ ⎤
ḡ11
0 ⎦ , det ḡi j = ḡ = g02 g11 , ḡ11 = g02 g11 + η2 + ζ 2 K 12 ,
( ) ( )
ḡi j = ⎣−ζ K 1 1 (12)
ηK 1 0 1
5
A. Borković, M.H. Gfrerer and B. Marussig Computer Methods in Applied Mechanics and Engineering 405 (2023) 115848

while its reciprocal counterpart is:

ζ K1
⎡ ⎤
1 −ηK 1
1
ḡ i j = ⎣ ζ K 1 ḡ11 − η2 K 12 −ηζ K 12 ⎦ . (13)
ḡ −ηK −ηζ K 2 ḡ11 − ζ 2 K 12
1 1
These rigorous relations show that the vector basis at an equidistant line, ḡi , is not orthogonal, which results with the
coupling of axial, bending and torsional actions. It is possible to analytically decouple these actions in the frame
of linear analysis [52], but, in the nonlinear analysis, the coupling is always present. A standard approach is to
simplify the beam model by setting g0 → 1 and neglect higher order terms [43]. Here, we will also decouple axial
and torsional actions by simplifying the metric tensors. However, we will not simplify the shifter g0 , which will
allow us to accurately model axial-bending coupling. In concrete terms, we are neglecting the higher order terms
with respect to the torsional curvature and the off-diagonal terms of reciprocal metric tensor:
⎡ ⎤
1/ḡ 0 0
ḡ11 ≈ g02 g11 , ḡ i j ≈ ⎣ 0 1 0⎦ . (14)
0 0 1

2.3. Kinematics of beam axis and cross section

As discussed at the beginning of this section, it is assumed that a cross section is rigid and remains perpendicular
to the beam axis in all configurations. The task is to find the material vector triad gi∗ at the current (unknown)
configuration by the update of some reference configuration, Fig. 1.
Finding the position and tangent of the beam axis at the current configuration are straightforward tasks:
r∗ = r∗ (ξ ) = r + u,
(15)
g∗1 = g∗1 (ξ ) = r∗,1 = g1 + u,1 ,
where u is the displacement vector of the beam axis. The other two basis vectors must be found via rotation:
g∗α = Rgα , (16)
where R is the rotation tensor. In general, this tensor is a member of the special orthogonal group SO(3) that
is a nonlinear manifold. This fact motivated the development of various strategies for the interpolation and
parameterization of the rotation tensor [9,66]. In essence, the rotation tensor must be linearized and this is achieved
via its linear tangent space so(3). In formal terms, we can choose to follow the velocity or the variation of the
current configuration, and the former option is adopted here.
Since the velocity of the beam axis is v = ṙ∗ = u̇, the velocity gradient with respect to the coordinate ξ follows
directly:
v,1 = ġ∗1 = u̇,1 . (17)
Let us adopt that g∗α
= gα + u,α , where vectors u,α represent increments of the basis vectors gα . We will refer to
their time derivatives formally as the velocity gradients along the material axes η and ζ , and express as a function
of the rotation tensor, Eq. (16):
v,α = u̇,α = ġ∗α = Ṙgα + R ġα = Ṙgα . (18)
The members of the so(3) group are skew-symmetric tensors (spinors) which allow an exponential mapping of
the elements of the SO(3) group [65]. In the case of the finite rotation tensor R, an appropriate spinor Φ is the
antisymmetric part of the displacement gradient, and its elements are infinitesimal rotations. Now, the exponential
mapping R = eΦ allows us to find time derivative of the rotation tensor and to calculate the velocity gradients in
Eq. (18):
Ṙ = Φ̇R H⇒ v,α = Φ̇g∗α . (19)
The components of the spinor Φ̇ are angular velocities:
1 1 1 1
ω̇i = ϵ i jk v̂k| j H⇒ ω̇1 = √ ∗ v̂3|2 , ω̇2 = − √ ∗ v̂3|1 , ω̇3 = √ ∗ v̂2|1 , (20)
2 g g g
6
A. Borković, M.H. Gfrerer and B. Marussig Computer Methods in Applied Mechanics and Engineering 405 (2023) 115848

where ϵ i jk is the Levi-Civita symbol and v̂k are the components of velocity with respect to the local material triad
v = v̂k g∗k . If we represent Φ̇ via its axial vector ω̇ = ω̇i gi∗ , Eq. (19) reduces to, [3]:
v,α = ω̇ × g∗α . (21)
Note that v̂3|2 = −v̂2|3 due to the assumption of the rigid cross sections, while the assumption of orthogonality of
the cross section and beam axis gives v̂2|1 = −v̂1|2 and v̂3|1 = −v̂1|3 . These relations allow the representation of
components ω̇2 and ω̇3 via the velocity of beam axis:
1 1
ω̇2 = − √ ∗ v̂3|1 = − √ ∗ g∗3 · v,1 ,
g g
(22)
1 1
ω̇3 = √ ∗ v̂2|1 = √ ∗ g∗2 · v,1 .
g g
Since these two components of angular velocity are not independent quantities, it follows that there is only one
independent component of the angular velocity of the BE beam:
1 1
ω̇1 = √ ∗ v̂3|2 = √ ∗ ω̇. (23)
g g
This quantity represents the angular velocity of a cross section with respect to the tangent of the beam axis, and it
is often referred to as the twist velocity. For simplicity, we will designate its physical counterpart with ω̇. In this
way, generalized coordinates of the BE beam are the components of the velocity of the beam axis and the twist
component of angular velocity of the cross section. In contrast to the SR beam model, the rotation of a cross section
of the BE beam belongs to the SO(2) group of in-plane rotations [24]. For another mathematically sound discussion
on the decomposition of the BE beam rotation, Ref. [47] is recommended.
Once the current triad gi∗ is found, we can define the complete metric of the current configuration by employing
the expressions (5), (11), (12) and (13). To obtain the relationship between the reference and current configurations,
we must represent quantities v,α as functions of the generalized coordinates [50]. By using Eqs. (21) and (22), we
find:
1 1 (
v,2 = Φ̇g∗2 = ω̇ × g∗2 = − ∗ v̂2|1 g∗1 + v̂3|2 g∗3 = − ∗ g∗2 · v,1 g∗1 + ω̇g∗3 ,
)
g g
(24)
1 1 ( ∗
v,3 = Φ̇g3 = ω̇ × g3 = − ∗ v̂3|1 g1 − v̂3|2 g2 = − ∗ g3 · v,1 g1 − ω̇g2 ,
∗ ∗ ∗ ∗
) ∗ ∗
g g
while their derivatives along the beam axis are:
{
1 ( ∗
g2 · v,1 g∗1,1 + K 1∗ g∗3 − K 3∗ g∗1 · v,1 g∗1
( ) ) [( ) ]
v,21 = v,2 ,1 = − ∗
g
}
+ g∗2 · v,11 g∗1 − 2 Γ11 ∗1
g2 · v,1 g∗1 +ω̇ K 2∗ g∗1 − K 1∗ g∗2 + ω̇,1 g∗3 ,
( ) ( ∗ ) ( )

{ (25)
1 ( ∗
g3 · v,1 g∗1,1 + K 2∗ g∗1 − K 1∗ g∗2 · v,1 g∗1
( ) ) [( ) ]
v,31 = v,3 ,1 = − ∗
g
}
+ g3 · v,11 g1 − 2 Γ11 g3 · v,1 g1 +ω̇ K 3∗ g∗1 − K 1∗ g∗3 − ω̇,1 g∗2 .
( ∗ ) ∗ ∗1
( ∗ ) ∗ ( )

2.4. Update of the local vector basis

In order to define the orientation of a cross section, the base vectors g∗α of the BE beam are found through the
ref
rotation of the reference basis vectors gα in the current cross-sectional plane, analogous to Eq. (4), see Fig. 2.
The definition of these reference basis vectors is not unique and three procedures are discussed here: the Smallest
Rotation (SR), the Nodal Smallest Rotation Interpolation (NSRISR), and the Frenet–Serret Rotation (FSR).
The SR mapping defines a reference vector basis by the rotation of the triad from the reference configuration such
that the tangents from both configurations align. The distinguishing property of the SR algorithm is that this rotation
angle is minimized, which gives the procedure’s name [10]. The SR procedure is readily used due to its simplicity
7
A. Borković, M.H. Gfrerer and B. Marussig Computer Methods in Applied Mechanics and Engineering 405 (2023) 115848

Fig. 2. In-plane rotation of a rectangular cross section (reference configuration — green, current configuration — red). The material basis
ref
vectors g∗α can be( updated) with respect to the reference material basis vectors gα using the total twist angle ω, or with respect to the
current FS basis n , b , using the independent twist angle θ . Additive decomposition of the total twist angle ω = ω F S + ∆θ is evident.
∗ ∗

(For interpretation of the references to color in this figure legend, the reader is referred to the web version of this article.)

and satisfactory accuracy [45]. Yet it is based on the rotation between the current and reference configurations, and
the interpolation of such rotations, in general, includes rigid-body motion [11]. This error is often ignored since it
mitigates with h-refinement [67].
In order to overcome the deficiency of the interpolation of the rotations between the current and reference
configurations, a linear interpolation of the relative rotation between the element nodes is suggested in [11].
Since this relative rotation is free from any rigid-body motion, the objectivity of the discretized strain measures
is preserved. One such algorithm is the NSRISR that is based on a specific double implementation of the SR
procedure [24]. The first step is similar to that of the standard SR algorithm, but new triads are defined only at the
SR SR ref
start and at the end of the finite element (gi,star t , gi,end ). The reference triad gi is then obtained by another SR
SR
mapping, but this time via the mapping of the triad at the start of element gi,star t along the length of finite element.
The reference triad that is free from the rigid-body motion is obtained by this means. In order to compensate for
the definition of the new reference frame, the twist angle DOF must be modified. The required correction angle
SR ref
equals the angle between the gi,star t and gi triads, and it is obtained through the linear interpolation between the
start and the end of the element. Since the NSRISR algorithm requires that the rotation field is C 0 continuous, this
approach cannot utilized increased smoothness provided by IGA [24,50].
Both SR and NSRISR procedures employ material basis vectors at some reference configuration and the total
twist angle ω. In order to apply the update, the basis vectors gα and (n, b) are first rotated to the current cross-
sectional plane without twist, using the SR algorithm. Let us observe one such current cross-sectional plane, where
the material and FS vectors in both configurations are designated, Fig. 2. Evidently, the total twist angle can be
the additively decomposed as:
ω = ω F S + ∆θ, (26)
where ω F S is the twist of FS frame, while ∆θ is the incremental change of the twist angle between the material and
FS vector bases, Fig. 2. Note that the twist of the FS frame can be found in several ways, e.g. from cos ω F S = n∗ ·n.
The decomposition (26) suggests that only the ∆θ part of the total twist represents the independent rotation field.
Following this observation, it is evident that the independent twist angle θ (or its increment ∆θ ) can be introduced
as the rotational DOF. The FSR formulation for the update of the local vector basis is introduced by this means. The
formulation utilizes the FS basis at the current configuration as the reference vector basis. This is a straightforward
approach that is frequently mentioned in the literature [24,68], but, to the best of our knowledge, was never
implemented in the context of the geometrically exact beam theory. The linear IGA was developed in [52], while
the nonlinear case is considered here.
8
A. Borković, M.H. Gfrerer and B. Marussig Computer Methods in Applied Mechanics and Engineering 405 (2023) 115848

3. Stress and strain

In this section, strain at an equidistant line is defined as a function of strains of the beam axis. Then, the strain
rates are introduced and the relations between the strain rates and generalized coordinates are set. Finally, the
constitutive relation is derived and section forces are defined. For convenience, we will assume that the reference
configuration is the initial stress-free configuration.

3.1. Strain measure

Components of the Green–Lagrange and Almansi strain tensors are the same. The axial strain component at an
equidistant line follows from Eqs. (12) and (14):
1( ∗ ) 1 ( ∗2 ∗ ) 1 [( )2 ]
1 + ξ α K α∗ g11 − 1 + ξ α K α g11
)2 ∗
ϵ̄11 = g0 g11 − g02 g11 =
(
ḡ11 − ḡ11 =
2 2 2 (27)
1 [( ∗
g11 − g11 + 2ξ K α g11 − K α g11 + ξ α ξ β K α∗ K β∗ g11
α
− K α K β g11 .
) ( ∗ ∗ ) ( ∗
)]
=
2
Recalling that for the definition of equidistant quantities, we abuse the notation by setting: ξ 2 = ξ2 = ζ and
ξ 3 = ξ3 = −η. Let us introduce the axial strain of the beam axis:
1( ∗
ϵ11 = g11 − g11 ,
)
(28)
2
and the changes of bending curvatures of beam axis with respect to the parametric convective coordinate:
κα = K α∗ g11
∗
− K α g11 = K̃ α∗ − K̃ α . (29)
Furthermore, we need to define changes of curvature with respect to convective arc-length coordinate [54]:
χα = K α∗ − K α , (30)
where these two measures of curvature change of the beam axis relate as:
κα = χα g11
∗
+ 2ϵ11 K α . (31)
Evidently, κα and χα differ due to the parameterization, the change of length of the beam axis and the initial
curvature. Let us now rewrite the last term in parentheses of Eq. (27):
K α∗ K β∗ g11
∗
− K α K β g11 = K α + χα K β∗ g11
∗
− K α K β g11 = K α κβ + χα K β∗ g11
∗
( )

= K α κβ + κα − 2K α ϵ11 K β + χβ
( )( )
(32)
= 2K α κβ − K β ϵ11 − χβ ϵ11 + κα χβ .
( )

By inserting Eqs. (32), (29), and (28) into Eq. (27), we obtain:
( )
α α β 1
ϵ̄11 = ϵ11 + ξ κα + ξ ξ K α κβ + χβ κα − ϵ11 K α K β − ϵ11 χβ K α
2
( ) (33)
1
= g0 (1 − ξ K α ) ϵ11 + ξ κα + ξ α ξ β χβ
α α
κα − K α ϵ11 .
[ ]
2
The first term of this expression corresponds to the linear analysis [52]. The second term is nonlinear with respect
to strains and it consists of two parts. The first is usually dominant as a product of curvature changes. The second
part consists of the product of curvature change and axial strain, and it only exist for initially curved beams. The
nonlinear relation between the equidistant axial strain and the changes of curvatures results with a strong coupling
between bending and axial actions, even for an initially straight beam:
1
ϵ̄11 = ϵ11 + ξ α κα + ξ α ξ β χβ κα . (34)
2
Relations for the equidistant shear strains due to the torsion are much simpler:
γ̄1α = 2ϵ̄1α = ḡ1α
∗
− ḡ1α = −ξ α κ1 (35)
where κ1 = K 1∗
− K 1 is the change of the beam’s torsional curvature. With Eqs. (33) and (35), the strain field of
the BE beam continuum is defined as a function of strains of the beam axis: ϵ11 , κ1 , κ2 , and κ3 . We will refer to
these quantities as the reference strains.
9
A. Borković, M.H. Gfrerer and B. Marussig Computer Methods in Applied Mechanics and Engineering 405 (2023) 115848

3.2. Strain rate

The strain rates are required, since we will be using the equation of virtual power. They follow as the time
derivatives of strain components ϵ̄˙i j = d̄i j . The rate of axial strain at an equidistant line is:
1 ˙∗
ḡ11 = g0∗ ġ0∗ g11
∗
+ g0∗ d11 = g0∗ ξ α K̇ α∗ g11
∗
+ 1 + ξ α K α∗ d11
( ) [ ( ) ]
d̄11 =
2 [ ( (36)
= g0∗ ξ α κ̇α − 2ξ α K α∗ + 1 + ξ α K α∗ d11 = g0∗ 1 − ξ α K α∗ d11 + ξ α κ̇α ,
) ( ) ] [( ) ]

where d11 and κ̇α are the respective rates of axial strain and curvature changes of beam axis. The obtained equidistant
strain rate, d̄11 , is analogous to the linear part of the strain in Eq. (33), but expressed with respect to the current
configuration. Additionally, the rates of equidistant shear strains are:
d̄1α = γ̄˙1α = 2ϵ̄˙1α = ḡ˙ 1α
∗
= −ξ α κ̇1 , (37)
where κ̇1 is the rate of change of torsional curvature.

3.3. Reference strain rates and generalized coordinates

The rates of reference strains allow us to find the strain rates at every point of the beam continuum. The relation
between the rate of reference axial strain and generalized coordinates is simple:
1 ∗
d11 = ġ11 = g∗1 · v,1 , (38)
2
while the rates of the curvature components are more involved. They follow from Eq. (7):
κ̇1 = K̇ 1∗ = v,21 · g∗3 + g∗2,1 · v,3 ,
κ̇2 = K̃˙ 2∗ = −v,11 · g∗3 − g∗1,1 · v,3 , (39)
κ̇ = K̃˙ ∗ = v · g∗ + g∗ · v ,
3 3 ,11 2 1,1 ,2
and by inserting Eqs. (24) and (25) into Eq. (39), the relations between the rates of curvatures and generalized
coordinates become:
κ̇1 = K 2∗ g∗2 · v,1 + K 3∗ g∗3 · v,1 + ω̇,1 ,
( ) ( )

κ̇2 = −g∗3 · v,11 − Γ11 ∗1

v,1 + K̃ 3∗ ω̇,
( )
(40)
κ̇3 = g2 · v,11 − Γ11 v,1 − K̃ 2 ω̇.
∗ ∗1 ∗
( )

These expressions employ velocity of the beam axis v and the total twist velocity of the cross section ω̇ as DOFs,
which is a standard approach for the BE beam [50].
In order to derive equations of the FSR formulation, we must additively decompose the total twist angular velocity
ω̇, analogously to Eq. (26):
ω̇ = ω̇ F S + θ̇ , (41)
where ω̇ F S is the twist angular velocity of the normal plane that is a function of velocity of the beam axis. It can
be found by the linearization of the normal and binormal [52]:
1 ∗ (
ω̇ F S = v̂3̄|2̄ = b∗ · v,2̄ = −v̂2̄|3̄ = −n∗ · v,3̄ = ∗1
v,1 .
)
b · v,11 − Γ11 (42)
K̃ ∗

Here, we have used indices with overbars (•)ᾱ,β̄ to distinguish the components and derivatives with respect to the
axes of the FS frame from those with respect to the axes of material basis gα . Furthermore, θ̇ in Eq. (41) represents
the part of total angular velocity that is independent of the velocity of the beam axis, and we will refer to it as the
independent twist angular velocity.
By introducing Eqs. (41) and (42) into Eq. (40), we obtain:
κ̇1 = T∗1 · v,1 − T∗2 · v,11 + T∗3 · v,111 + θ̇,1 ,
κ̇2 = n∗ · v,11 − Γ11 ∗1
v,1 sin θ ∗ + K̃ 3∗ θ̇ ,
( )
(43)
κ̇3 = n∗ · v∗,11 − Γ11∗1
v,1 cos θ ∗ − K̃ 2∗ θ̇ .
( )

10
A. Borković, M.H. Gfrerer and B. Marussig Computer Methods in Applied Mechanics and Engineering 405 (2023) 115848

where:
[ ( ) ]
1 K̃ ,1
T1 = 1
Γ11 τ̄ n + 1
Γ11 − 1
Γ11,1 + K K̃ b ,
K̃ K̃
[ ( ) ]
1 K̃ ,1
T2 = τ̄ n + Γ11
1
+ b , (44)
K̃ K̃
1
T3 = b.
K̃
With the decomposition of the total angular velocity, the reference strain rates are represented as a function of
the velocity of the axis v and the independent twist velocity θ̇. This approach comes at the cost of having to
introduce the third order derivatives of the velocity of the beam axis, which will require at least C 2 -continuous
spatial discretization. We should note that Eq. (43) can be derived directly from Eq. (8), as in [52]. A derivation of
this kind does not require the introduction and decomposition of the total twist velocity, and θ̇ naturally follows as
a generalized coordinate.
Note that we have omitted the configuration designation (asterisk sign) in definition (44). Such notation entails
that a vector Pt , in a configuration Ct , is expressed as a function of the geometric quantities (i.e. K t , τ t , gt1 , bt )
in the corresponding configuration Ct . The metric of an arbitrary configuration in Sections 2.1 and 2.2 is already
defined in that manner. This notation will allow us to simplify the writing, especially after we introduce a previously
calculated configuration in Section 4.
Next, let us define the variations of equidistant strain rates:

δ d̄11 = g0∗ 1 − ξ α K α∗ δd11 + ξ α δ κ̇α , δ d̄1α = −ξ α δ κ̇1 ,

[( ) ]
(45)

and the variations of reference strain rates:

δd11 = g∗1 · δv,1 ,
δ κ̇1 = K 2∗ g∗2 · δv,1 + K 3∗ g∗3 · δv,1 + δ ω̇,1 = T∗1 · δv,1 − T∗2 · δv,11 + T∗3 · δv,111 + δ θ̇,1 ,
( ) ( )
(46)
δ κ̇2 = −g∗3 · δv,11 − Γ11 ∗1
δv,1 + K̃ 3∗ δ ω̇ = n∗ · δv,11 − Γ11∗1
δv,1 sin θ ∗ + K̃ 3∗ δ θ̇ ,
( ) ( )

δ κ̇3 = g∗2 · δv,11 − Γ11∗1

δv,1 − K̃ 2∗ δ ω̇ = n∗ · δv∗,11 − Γ11
∗1
δv,1 cos θ ∗ − K̃ 2∗ δ θ̇ .
( ) ( )

3.4. Constitutive relation

We are considering hyperelastic St. Venant-Kirchhoff material:

S̄ i j = 2µϵ̄ i j + λδ i j ϵ mm , (47)

where µ and λ are Lamé material parameters while S̄ i j are contravariant components of the second Piola–Kirchhoff
stress tensor. From the conditions S̄ 22 = S̄ 33 = 0, we obtain:

S̄ i j = 2µ ϵ̄ i j + νδ i j ϵ̄ 11 = 2µ ḡ ik ḡ jl ϵ̄kl + νδ i j ḡ 1k ḡ 1l ϵ̄kl ,
( ) ( )
(48)

where ν is the Poisson’s ratio. By employing the simplification of the reciprocal metric tensor in Eq. (14), three
non-zero stress components of the BE beam are:

S̄ 11 = E ḡ 11 ḡ 11 ϵ̄11 and S̄ 1α = 2µḡ 11 ḡ αα ϵ̄1α = µḡ 11 ḡ αα γ̄1α (no summation over α) , (49)

where E is the Young’s modulus of elasticity.

11
A. Borković, M.H. Gfrerer and B. Marussig Computer Methods in Applied Mechanics and Engineering 405 (2023) 115848

Let us note that, for the BE beam, the relation between the components of the Cauchy stress σ̄ i j and the second
Piola–Kirchhoff stress S̄ i j are:
√
ḡ
σ̄ = √ S̄ i j .
ij
(50)
ḡ ∗
This relation follows from the fact that the area of the cross section does not change, and the change of the volume
element is only due to the length change along the direction of the beam axis.

3.5. Stress resultant and stress couples

The stress resultant is defined as the integral of the tractions at current configuration. For the BE beam, the stress
resultant has direction of the tangent of the beam axis:
∫ ∫ ∫
∗ √ √
∗
σ̄ t̄ dA =
∗
σ̄ ḡ1 ⊗ ḡ1 t dA = t
11
S̄ 11 g0∗ g0 g g ∗ dA ,
( ∗ ∗ ∗
) ∗
N = Nt =
A∗ ∫ A A
∫ (51)
11 ∗ √
S̄ g0 g0 g g dA ,
11
√ √ √
N= S̄ ∗
ḡ ḡ dA = ∗
A A
∗
where we note that dA = dA and t̄ = t∗ . Here, σ̄ is the Cauchy stress tensor while N is the physical normal
∗

force. Next, let us define the stress couples:

∫ ∫
ηg2 + ζ g∗3 × σ̄ k1 ḡ∗k ⊗ ḡ∗1 t∗ dA = M 1 t∗ + M α g∗α ,
∗
ρ × σ̄ t̄ dA∗ =
( ∗ ) ( )
M=
∗
A∫ A ∫ (52)
α1
M = − ξα S̄ ḡ dA , M α = ξ α S̄ 11 ḡ ∗ ḡ dA ,
1
√ √ √
A A
where M 1 is the physical torsional moment, while M α are the physical bending moments.
By the insertion of Eqs. (33), (35) and (49) into Eqs. (51) and (52), the stress resultant and stress couples can be
expressed as functions of the reference strains. However, due to the presence of shifters in both initial and current
configurations, g0 and g0∗ , the exact expressions are cumbersome. Our aim is to present a simplified relation between
the stress resultants and couples, and the reference strains. For this, the exact constitutive relation is approximated
using the Taylor series with respect to ξ α coordinates, and the higher order terms of strains and initial curvatures
are neglected. In this way, the following constitutive relation is obtained:
ST = N M 1 M 2 M 3 , ϵ = ϵ11 κ1 κ2 κ3 , S = D S F ϵ ,
[ ] [ ]
⎡ ⎤
A 0 Iζ ζ c13 Iηη c14
√ √
E g∗ ⎢ ⎢ 0 µg It /(E g ∗ ) 0 0 ⎥ ⎥,
DS F =
g g Iζ ζ c31
⎣ 0 Iζ ζ 0 ⎦
(53)
Iηη c41 0 0 Iηη
c13 = 1.5χ2 − K 2 , c14 = 1.5χ3 − K 3 , c31 = χ2 − 2K 2 , c41 = χ3 − 2K 3 ,
∫ ∫ ∫ ∫
dη dζ , Iζ ζ = ζ dη dζ , Iηη =
2
η2 dη dζ , η + ζ 2 dη dζ .
( 2 )
A= It =
A A A A
Axial and bending actions are evidently coupled, which mostly affects the stress resultant N . For circular cross
section, the term It reduces to the polar moment of area. However, for all the other cross-section shapes, the term
It equals so-called torsional constant which must be calculated approximately [45].

4. Finite element formulation

In this Section, the equation of motion of isogeometric spatial BE element based on the FS frame is derived,
spatially discretized, and linearized.

4.1. Principle of virtual power

The principle of virtual power represents the weak form of the equilibrium. It states that at any instance of time,
the total virtual power of the external, internal and inertial forces is zero for any admissible virtual state of motion.
12
A. Borković, M.H. Gfrerer and B. Marussig Computer Methods in Applied Mechanics and Engineering 405 (2023) 115848

If the inertial effects are neglected and the loads are applied with respect to the beam axis, the equation of motion
is:
∫ ∫
δ P = δ Pint + δ Pext = σ̄ : δ d̄ dV̄ − p · δv + m∗ · δ ω̇ g ∗ dξ = 0,
∗ ∗
( ∗ )√
(54)
V∗ ξ

where d is the strain rate tensor, while p and m are the vectors of external distributed line forces and moments,
respectively. All these quantities are defined at the current, unknown, configuration C∗ .
Eq. (54) is nonlinear, and we will solve it by Newton’s method which requires the linearization of the equation.
Assuming that the external load is configuration-independent, only the internal virtual power must be linearized:
L σ̄ ∗ = σ̄ ♯ + σ̄˙ ∆t ,
L δ d̄ = δ d̄♯ + ∆δ d̄, (55)
L σ̄ ∗ δ d̄ = σ̄ ♯ δ d̄♯ + σ̄˙ δ d̄♯ ∆t + σ̄ ♯ ∆δ d̄,
( )

where L marks the linearization, while (•)♯ designates sharp quantities, i.e., values from the previously calculated
configuration C♯ , which is generally not in equilibrium. σ̄˙ is the stress rate tensor which is calculated as the Lie
derivative of current stress. Since the components of the stress rate tensor are equal to the material time derivatives
of the components of the stress tensor, [64], the linearized form of the internal virtual power is:
∫ ∫ ∫
♯1k ♯ ♯
L δ Pint = σ̄ δ d̄1k dV̄ +
∗
σ̄ δ d̄1k dV̄ ∆t +
˙ 1k ∗
σ̄ ♯1k ∆δ d̄1k dV̄ ∗ . (56)
V∗ V∗ V∗

Using relations (50) and dV̄ ∗ = ḡ ∗ /ḡ dV̄ , we can switch to the components of the second Piola–Kirchhoff stress
√

and integrate with respect to the initial volume:

∫ ∫ ∫
♯1k ♯ ˙ ♯
L δ Pint = S̄ δ d̄1k dV̄ + S̄ δ d̄1k dV̄ ∆t +
1k
S̄ ♯1k ∆δ d̄1k dV̄ . (57)
V V V

By integrating Eq. (57) with respect to the area of the cross section, the integrals over the 3D volume reduce to
integrals along the beam axis:
∫ ( ∫ (
♯ ♯ √ ♯ ♯ √
) )
L δ Pint = ♯ ♯i
Ñ δd11 + M̃ δ κ̇i g dξ + Ñ˙ δd11 + M̃˙ i δ κ̇i g dξ ∆t
ξ ξ
∫ ( )√ (58)
+ Ñ ♯ ∆δd11 + M̃ ♯i ∆δ κ̇i g dξ ,
ξ
i
where Ñ and M̃ are stress resultant and stress couples that are energetically conjugated with the reference strain
rates of the beam axis, d11 and κ̇ j , while Ñ˙ and M̃˙ i are their respective rates. If we introduce the vectors:

fT = Ñ M̃ 1 M̃ 2 M̃ 3 , eT = d11 κ̇1 κ̇2 κ̇3 ,

[ ] [ ]
(59)
pT = p1 p2 p3 , mT = m̂ 1 m̂ 2 m̂ 3 ,
[ ] [ ]

linearized virtual power can be expressed in compact matrix form as:

∫ ∫ ∫ ∫ (
T √ T √ √ )√
f ♯ δe♯ g dξ + ḟ δe♯ gdξ ∆t + f ♯T ∆δe g dξ = p∗ T δv + m∗ T δ ω̇ g ∗ dξ . (60)
ξ ξ ξ ξ

4.2. Calculation of the stress rate resultant and couples

Let us find the relation between the stress rate resultant and stress rate couples, and the reference strain rates.
This relation is required for the integration of internal virtual power over the cross-sectional area, see Eq. (58):
∫
♯ ♯ ♯
S̄˙ 1k δ d̄1k g0 dA = ḟ δe♯ = Ñ˙ δd11 + M̃˙ i δ κ̇i .
T
(61)
A
Variations of equidistant strain rates are given with (45), while the stress rates follow from (49). Similar to the stress
resultant and stress couples in Section 3.5, the exact expressions are cumbersome, and an analogous approximation
13
A. Borković, M.H. Gfrerer and B. Marussig Computer Methods in Applied Mechanics and Engineering 405 (2023) 115848

is made. The resulting simplified constitutive relation is symmetric, as required:

[ ]T
ḟ = Ñ˙ M̃˙ 1 M̃˙ 2 M̃˙ 3 , ḟ = D̃♯M e♯ ,
⎡ ⎤
A 0 Iζ ζ ã13 Iηη ã14
E ⎢ 0 µg It /E 0 0 ⎥ ⎥, (62)
D̃ M = 2 ⎢
g ⎣ Iζ ζ ã31 0 Iζ ζ 0 ⎦
Iηη ã41 0 0 Iηη
ã13 = ã31 = χ2 − 2K 2 , ã14 = ã41 = χ3 − 2K 3 ,

4.3. Calculation of internal forces

The correct calculation of the internal forces is crucial for accurate simulations. The internal forces follow from
Eqs. (49), (33), (35), and (57):
∫
♯ ♯ ♯
S̄ ♯1k δ d̄1k g0 dA = Ñ ♯ δd11 + M̃ ♯i δ κ̇i , (63)
A
and they allow the reduction given by Eq. (58). Again, after the approximation with Taylor series and by neglecting
higher order terms with respect to strains and initial curvatures, the relation between internal forces and the reference
strains is:
⎡ ⎤
A 0 Iζ ζ a13 Iηη a14
E ⎢ 0 µg It /E 0 0 ⎥
f ♯ = D̃♯ ϵ ♯ , D̃ = 2 ⎢ ⎥, (64)
g ⎣ Iζ ζ a31 0 Iζ ζ 0 ⎦
Iηη a41 0 0 Iηη
where the coupling coefficients a are:
a13 = 0.5χ2 − 2K 2 , a14 = 0.5χ3 − 2K 3 , a31 = χ2 − 2K 2 , a41 = χ3 − 2K 3 . (65)
In contrast to the matrix D̃ M , the constitutive matrix D̃ is not symmetric, since it relates total values of stress
resultant and couples, and reference strains. With the coupling coefficients of Eq. (65), the effect of strong curvature
is correctly captured.
Regarding the definition of the curviness parameter, it is evident from Eq. (65) that the initial curvature has
more influence on the axial-bending coupling than the change of curvature. This suggest that the current curviness
parameter K ∗ d = (χ + K ) d, as defined in the introduction, is not comprehensive and it can be refined. For
simplicity, we will keep the standard definition in this paper.
One of our aims is to examine the influence that strong curvature has on the response of the beam. The correct
constitutive relation for the calculation of internal forces is given by Eqs. (64) and (65), and we will designate
that model with D C . Furthermore, let us introduce two reduced constitutive models, as in [50]. The first one is the
standard decoupled model that ignores off-diagonal terms( in the constitutive
) relation (64) - D 0 model. The second
∗
reduced model is obtained by setting shifters to one g0 → 1, g0 → 1 and by neglecting the nonlinear terms in
Eq. (33). This approximation returns the constitutive model that restricts the change of the length of the axis due
to bending. We will refer to this constitutive model as the small-curvature model and designate it with D 1 . The
coupling coefficients in Eq. (64) for the D 1 model are:
a13 = −(χ2 + K 2 ) = −K 2∗ , a14 = −(χ3 + K 3 ) = −K 3∗ , a31 = −K 2 , a41 = −K 3 . (66)

4.4. Spatial discretization

Using IGA, both geometry and kinematics are here discretized with the same univariate NURBS functions R I :
N
∑ N
∑
r= R I (ξ )r I , θ= R I (ξ )θ I ,
I =1 I =1
N N
(67)
∑ ∑
v= R I (ξ )v I , θ̇ = R I (ξ )θ̇ I ,
I =1 I =1
14
A. Borković, M.H. Gfrerer and B. Marussig Computer Methods in Applied Mechanics and Engineering 405 (2023) 115848

where (•) I stands[ for the

] value of quantity at the I th control point. If we introduce a vector of generalized
coordinates, vTθ = vT θ̇ , and the matrix of basis functions N, the kinematic field of the beam can be represented
as:
vθ = Nq̇, (68)
where:
q̇ = q̇1 q̇2 ... q̇ I ... q̇ N , q̇ I = v 1I v 2I v 3I θ̇ I
[ ] [ ]
⎡ ⎤
[
N1 N2 ... N I ... N N
] RI 0 0 0 [ ]
(69)
T 03×1
N= , N = 0 RI 0 0⎦ , NθI = .
Nθ1 Nθ2 ... NθI ... NθN I
⎣
RI
0 0 RI 0
Here, 0m×n designates the m × n zero matrix.

4.5. Discrete equations of motion

To define spatially discretized equations of motion, we must relate the reference strain rates with the generalized
coordinates at the control points by using Eqs. (38), (44) and (69):
e = B∗L q̇ = H∗ Bq̇, (70)
where:
[ ]
T T
... ... BN , B I T = N I,1 T N I,11 T NθI NθI,1 ,
[ ]
B = B1 B2 BI N I,111 T
⎡ T ⎤
g1 01×3 01×3 0 0
T T
−T2 T T3 T 0 1⎥ (71)
⎥.
⎢ 1
H=⎢
⎣ −Γ 1 sin θ n T
11 sin θ n T 01×3 K̃ 3 0⎦
1
−Γ11 cos θ n T cos θ n T 01×3 − K̃ 2 0
The variation of the vector of reference strain rates is, cf. Eqs. (46) and (70):
δe♯ = H♯ Bδ q̇, (72)
while its linearized increment is:
∆δe = ∆ H∗ Bδ q̇ = ∆H∗ Bδ q̇,
( )
(73)
where the increment of the operator H∗ is quite involved, and it is given in detail in Appendix A. To continue the
derivation, let us refer to the matrix of basis functions BG and the matrix of generalized section forces G that are
defined in Appendix A by Eqs. (A.19) and (A.20). With these expressions, and Eqs. (62), (72), and (73), we can
rewrite the integrands on the left-hand side of Eq. (60) in a spatially discretized form:
T T T ♯
f ♯ δe♯ = f ♯ H♯ Bδ q̇ = f ♯ B L δ q̇,
T T ♯T ♯
ḟ δe♯ = e♯ D̃♯M δe♯ = q̇T B L D̃♯M B L δ q̇, (74)
T T
f ♯ ∆δe = f ♯ ∆H∗ Bδ q̇ = q̇T BG T G♯ BG δ q̇∆t.
Let us note that there is a virtual power that stems from the imposition of boundary conditions. These contributions
are discussed in Section 4.6 and Appendix B.
Regarding the external virtual power, it can be spatially discretized via the vector of the external load, Q:
∫ ( )√
p∗ T δv + m∗ T δ ω̇ g ∗ dξ = Q∗ T δ q̇, (75)
ξ

which is the same as in the linear analysis [52]. However, if the vector of external load depends on the configuration,
it must be linearized as well. For this case, a contribution to the geometric stiffness is derived in Appendix C.
15
A. Borković, M.H. Gfrerer and B. Marussig Computer Methods in Applied Mechanics and Engineering 405 (2023) 115848

With all these ingredients, the linearized equation of equilibrium is:

∫ ( )√ ∫
♯T ♯ T ♯√
q̇T B L D̃♯M B L + BG T G♯ BG g dξ δ q̇∆t = Q∗ T δ q̇ − f ♯ B L g dξ δ q̇, (76)
ξ ξ

which can be cast into the standard form:

♯
KT ∆q = Q∗ − F♯ , ∆q = q̇∆t ,
( )
(77)
where:
∫ ∫
♯ ♯ ♯ ♯T ♯√ √
KT = K M + KG = B L D̃♯M B L g dξ + BG T G♯ BG g dξ , (78)
ξ ξ

is the tangent stiffness matrix that consists of material and geometric parts, while:
∫
♯T √
F♯ = B L f ♯ g dξ , (79)
ξ

is the vector of internal forces. The vector ∆q in Eq. (77) contains increments of displacement and independent twist
angle at control points with respect to the previous configuration. This vector allows us to update the configuration
to check if the equilibrium is satisfied, that is, if the residual Ψ ♯ = Q∗ − F♯ is less than the prescribed error
tolerance. In addition to the standard Newton–Raphson method, the Arc-length method is also employed here in
order to simulate responses that include load limit points.

4.6. Imposition of boundary conditions

Kinematic boundary conditions with respect to the displacements can be implemented in a straightforward
manner. Regarding the rotations, the situation is more involved. The rotation components ω2 and ω3 are not
independent quantities and their values can be imposed at section ξ = ξc in a standard manner by using the
parent–child approach for constraining the DOFs. In essence, these conditions require that the tangent at ξc does
not rotate. For ξc = 0 and ξc = 1, the tangent is aligned with the control polygon, and only two control points
influence the rotation. The resulting constraint conditions are linear and straightforward to implement.
On the other hand, the twist angle ω1 = ω consists of two parts. One is the rotation of the normal plane ω F S
which depends on the displacement of the axis, while the other is the independent twist angle ∆θ. For simplicity, let
us consider the procedure required to impose homogeneous boundary condition at current configuration for ξ = 0,
ω∗ (0) = 0. The constraint equation is:
c∗ = ω∗ = ω∗F S + ∆θ = 0. (80)
Since ω∗F S = arccos n∗ · n , this constraint is nonlinear and the parent–child approach is not suitable. Therefore,
( )

we will implement the constraint (80) via Lagrange multiplier λ∗ , by requiring that:
λ∗ c∗ = 0. (81)

Remark. When the current and initial FS frames are known, and the cross section is fixed, the calculation of
the angle ω∗F S is trivial. However, the issue occurs if the cross section rotates arbitrarily. To find the twist of the
FS frame in this case, we must rotate the FS frame from the reference to the current configuration with the SR
algorithm, (see Section
) 2.4. (The obtained vectors are designated with (n S R , b S R ), and the angle ω∗F S follows as
ω F S = sgn n · b S R arccos n · n S R .
∗ ∗ ∗
)

The virtual rate of the condition (81) is added to the virtual power (54):
δ P + c∗ δ λ̇ + λ∗ δ ċ = 0. (82)
∗
Since the current configuration C is unknown, the next step is to linearize the constraint with respect to the previous
configuration C♯ . Let us represent ċ as a function of generalized coordinates, see Eq. (42):
1 ∗ ( ∗1
v,1 + θ̇ = H∗λ Bλ q̇λ = K∗λ q̇λ ,
)
ċ = b · v,11 − Γ11 (83)
K̃ ∗

16
A. Borković, M.H. Gfrerer and B. Marussig Computer Methods in Applied Mechanics and Engineering 405 (2023) 115848

where
Nλ1,1 03×1 Nλ2,1 03×1 ... NλI,1 03×1 ... NλN ,1 03×1
⎡ ⎤
0
Bλ = ⎣Nλ1,11 03×1 Nλ2,11 03×1 ... NλI,11 03×1 ... NλN ,11 03×1 0⎦ ,
0 1 0 0 ... 01×3 0 ... 01×3 0 0
⎡ 1×3 ⎤ 1×3 (84)
RI 0 0
bT [
NλI = ⎣ 0 R I 0 ⎦ , Hλ = 1
1 0 , q̇λ T = q̇ T λ̇ .
] [ ]
−Γ11
0 0 RI K̃
Linearization of the virtual power due to the constraint gives:
L c∗ δ λ̇ + λ∗ δ ċ = c♯ δ λ̇♯ + c♯ ∆δ λ̇ + δ λ̇♯ ∆c + λ♯ δ ċ♯ + λ♯ ∆δ ċ + δ ċ♯ ∆λ,
( )
(85)
where ∆λ = λ̇∆t and ∆c = ċ∆t. The terms of linearized virtual power in Eq. (85) can be represented in a spatially
discretized form as:
♯ ♯T ♯
c♯ δ λ̇♯ + λ♯ δ ċ♯ = δ λ̇♯ Kλ q̇λ + δ q̇λ T Kλ λ♯ = δ q̇λ T Fλ ,
♯
c♯ ∆δ λ̇ + λ♯ ∆δ ċ = 0 + δ q̇λ T KGλ q̇λ ∆t, (86)
♯ ♯T
δ λ̇♯ ∆c + δ ċ♯ ∆λ = δ λ̇♯ Kλ q̇λ ∆t + δ q̇λ T Kλ λ̇∆t,
♯
where Kλ is the vector that must be added to the material part of the tangent stiffness matrix as its row and column,
♯T
at the position corresponding to the λ̇ DOF. Vector Fλ = λ♯ K♯λ c♯ represents the contribution to the internal force
[ ]
vector that comes from the constraint, while KGλ is the contribution to the geometric stiffness. Detail derivation of
the matrix KGλ is given in Appendix B.

5. Numerical examples
The aim of the following numerical studies is to verify and benchmark the derived formulation. The Dirichlet
boundary conditions are imposed in a well-known manner where the rotations are treated with special care, cf.
Section 4.6. The global components of external moments are applied as force couples which must be updated at
each iteration [55]. Standard Gauss quadrature with p + 1 integration points per element are applied. All the results
are presented with respect to the load proportionality factor (LPF), rather than to the load intensity itself.
By removing the independent twist angular velocity θ̇ from the vector of unknowns, a novel rotation-free
formulation of the spatial BE beam is obtained. In contrast to existing reduced models [25,53], which completely
neglect the twist DOF, this formulation incorporates one part of the twist velocity - ω̇ F S . It should return better
results than existing reduced models because it does not completely neglect torsional stiffness. The formulation is
designated as the Frenet–Serret Rotation Twist-Free (FSR TF). Therefore, four element formulations are considered:
FSR, FSR TF, SR and NSRISR. For the interpolation of kinematic quantities, the highest available interelement
continuity is applied exclusively. The exception is the NSRISR formulation which employs C 0 continuity for the
approximation of twist.
Since the developed formulations are displacement-based, they are inevitably prone to locking. Hence, it is worth
noting that we do not utilize any specific procedure here to deal with the locking issues.
In some examples, the error of vector a is calculated using the following relative L 2 -error norm:
√
1 l
∫
1 2
∥e∥2 = ah − ar e f  ds, (87)
 
amax l 0
where l is the length of the beam and amax is the maximum component of the observed vector. ah represents the
approximate solution, while ar e f is the reference solution.
The convergence criteria for the nonlinear solvers is set with respect to the values of both displacement and force
error norms, as in [50]. The tolerance for these error norms is 10−4 in all examples.

5.1. Pre-twisted circular beam

5.1.1. Path-independence
Path-independence of a computational formulation can be analyzed in various ways. Some authors simply apply
different sizes of load increments [12], while the others change the order of the applied load [24,50]. Here, we
17
A. Borković, M.H. Gfrerer and B. Marussig Computer Methods in Applied Mechanics and Engineering 405 (2023) 115848

Fig. 3. Pre-twisted circular beam. (a) Geometry and load. (b) Path-dependence test for three formulations and cubic splines. Difference
between SUCXZ and SIM load cases for LPF = 1 vs. the number of elements.

Fig. 4. Path-independence of a pre-twisted beam. Deformed configurations of the beam for two different loading orders and four values of
LPFs: (a) SUCZX, (b) SUCXZ.

employ the latter approach and analyze a quarter-circle cantilever beam loaded with two forces at the free end, as
shown in Fig. 3a. A special feature of this example is that the beam is pre-twisted with an angle of θ P T = π ξ/2.
Three cases of the application of load are considered. First, both F X and F Z are applied simultaneously — SIM
case. For the other two cases, the forces are applied successively, one for 0 < LPF < 0.5 and the other for 0.5 <
LPF < 1. The case when the F X is applied first is designated with SUCXZ while the other case is marked with
SUCZX. For all cases, the load is applied in 20 increments.
For a path-independent solution, the final configurations must be the same, that is invariant to the load order.
The difference of position between the SIM and SUCZX loadings are calculated for LPF = 1 using Eq. (87). The
local vector basis is updated incrementally, with respect to the previously converged configuration. The results for
different cubic NURBS meshes are shown in Fig. 3b for all three formulations. The results indicate that the presented
FSR formulation is indeed path-independent since the observed difference is practically zero. This is expected since
there is no interpolation hidden in the history of evolution of the independent twist angle [11]. Additionally, it is
confirmed that the NSRISR formulation is path-independent, while the SR is not [50].
For visualization purposes, the deformed beam configurations for SUCXZ and SUCZX, and the four characteristic
LPFs are shown in Fig. 4. Apparently, both load cases yield similar final configurations in visual terms, but each
with a different deformation history.

5.1.2. Objectivity
The invariance of a computational formulation with respect to the rigid-body motions is designated as objectivity.
This means the structure subjected to a rigid-body motion should not be strained. Invariance with respect to the
translations of beams is readily satisfied while the invariance with respect to the rotation requires special attention
[69].
18
A. Borković, M.H. Gfrerer and B. Marussig Computer Methods in Applied Mechanics and Engineering 405 (2023) 115848

Fig. 5. Objectivity of a pre-twisted beam. (a) Visualization of deformed configurations. (b) Evolution of internal strain energy with respect
to the number of rotations. (c) Internal strain energy at final configuration for different formulations using quintic elements.

The present example is based on [24,50] where a quarter-circular cantilever beam is rotated ten times around
its clamped end with respect to the x-direction, see Fig. 3a. For a deformation case of this kind, an objective
formulation should not produce any internal strain energy. In contrast to the previous studies [24,50], the beam is
pre-twisted here.
The beam is discretized with two elements, and two different NURBS orders are considered, p = 4 and p = 5.
The non-homogeneous boundary condition, ϕx = 20π , is applied in 100 increments. Characteristic configurations
during the first cycle of rotation are visualized in Fig. 5a. The internal strain energy in the final configuration is
plotted in Fig. 5b with respect to the number of rotations. These results suggest that the FSR formulation is indeed
objective since the internal strain energy is practically equal to zero. Additionally, it is confirmed that the NSRISR
formulation is objective while the SR is not. These observations are invariant with respect to the NURBS order.
The same energy is observed as a function of mesh density in Fig. 5c. The results indicate that the problem with
the representation of rigid-body motion mitigates for the SR formulation when the number of elements is increased.
This is a well-known fact which sometimes justifies the application of non-objective formulations in quasi-static
analyses [67]. Furthermore, our implementation of the NSRISR method shows an increase of the strain energy,
similar to that in [50]. Nevertheless, if the scaling factor is included here, as in [24,50], the normalized internal
energy would equal zero up to the machine precision. Regarding the FSR formulation, the results are completely
invariant with respect to the number of rotations and are equal zero.
Let us note in passing that the rotation-free FSR TF formulation can describe rigid-body motion of this beam
by default.

5.1.3. Convergence
Next, we examine the convergence behavior of the FSR formulation, again applied to the pre-twisted beam
example. The reference solution for the SIM load case is obtained using a quintic NURBS mesh with 128 elements.
The position of the axis and three reference strains are observed at the final configuration, Fig. 6. The theoretical
convergence rates are min [ p + 1, 2 ( p − m + 1)], where m is the highest derivative appearing in the weak form
[24]. Since m = 3 for the FSR formulation, the expected convergence rates for the position using the cubic,
quartic, and quintic NURBS are 2, 4, and 6, respectively. The obtained rates in Fig. 6 are generally in-line with
these predictions, while the quintic mesh slightly exceeds theoretical expectations. Note that the rates for torsional
curvature are aligned with the estimated values ( p + 1 − 3). A potential explanation for these observations is that
the third-order derivative of basis functions influences the torsion directly, contrary to other quantities that are
influenced indirectly.
Next, we investigate the performance of our nonlinear solvers. The number of required increments and iterations
for the convergence of five quintic meshes are shown in Table 1. The first increment is applied as LPF = 0.01. An
19
A. Borković, M.H. Gfrerer and B. Marussig Computer Methods in Applied Mechanics and Engineering 405 (2023) 115848

Fig. 6. Convergence test of a pre-twisted beam. Number of elements vs. relative L 2 -error of: (a) position, (b) axial strain, (c) torsional
strain, (d) bending strain.

Table 1
Number of required increments/iterations for the convergence using quintic elements.
n el = 5 n el = 10 n el = 20 n el = 40 n el = 80
NSRISR 29/158 29/158 28/152 28/152 28/152
Newton–Raphson FSR 27/146 27/146 27/146 36/202 59/345
FSR TF 24/130 24/127 24/130 30/165 44/254
NSRISR 21/102 20/98 19/94 18/90 18/92
Arc-length FSR 23/106 23/106 23/106 29/146 38/204
FSR TF 21/94 21/94 21/94 24/117 31/162

automatic incrementation algorithm is used and the desired number of iterations per increment is set as n d = 6.
The new increment for the Newton–Raphson solver is calculated by scaling the previous one with n d /n c , where
n c is the number of iterations required for the convergence of previous increment. For the Arc-length procedure,
√
the arc-length of the predictor is calculated by scaling the previous one with n d /n c . The results suggest that our
implementation of the Arc-length has superior convergence over the Newton–Raphson implementation. This is due
both to the flexibility of the Arc-length method, where the increment size is varied during one load step, and also
to the specific automatic incrementation setup.
The NSRISR formulation returns the most consistent results and the convergence properties do not change
significantly with the mesh density. The opposite holds for the FSR and FSR TF formulations. For n el > 20, the
required increments and iterations increase with a decrease in element size. To gain more insight, Fig. 7 illustrates
the condition number of the linear stiffness matrix for different meshes and NURBS orders. It is evident that the
condition number and its increase rate are significantly larger for FSR in comparison with NSRISR. These facts
provide a rationale for the increase of the required increments and iterations for FSR. We attribute this behavior to
20
A. Borković, M.H. Gfrerer and B. Marussig Computer Methods in Applied Mechanics and Engineering 405 (2023) 115848

Fig. 7. Pre-twisted beam. The condition number of the linear stiffness matrix vs. the number of elements.

Fig. 8. A pre-twisted beam. (a) Comparison of displacement components vs. LPF using the FSR, FSR TF, and Abaqus. (b) Comparison of
deformed configurations for LPF = 1, using the FSR and FSR TF.

the presence of the third order derivatives of basis functions in the FSR formulation. The results for FSR TF are
indistinguishable from those of FSR, and they are thus omitted.

5.1.4. Twist-free model

In order to assess the influence of the independent twist angle, the pre-twisted beam is analyzed with the
FSR TF model and the results are compared in Fig. 8a, while the deformed configurations are visualized in
Fig. 8b. Additionally, the FSR is verified by a comparison with the results obtained using Abaqus and a mesh
of 314 B31 elements. Although the B31 element models shear deformable beams [70], the results are practically
indistinguishable from those of the FSR model. Interestingly, the cubic element B33 that models BE beams could
not converge for such large deformations.
It is evident that the FSR TF formulation returns erroneous results in this case. The value of the independent twist
angle depends on the boundary conditions and load, and the usage of the FSR TF cannot be generally recommended.
For large load values in this example, the error decreases as the axial strain of the beam axis increases.

5.2. Pure bending of a cantilever beam

This is a standard benchmark example for in-plane nonlinear beam formulations. A cantilever is loaded with a
tip moment, causing the state of pure bending, Fig. 9. Since the beam deforms in a plane, there is no twisting,
and we can apply the rotation-free plane IGA model, see [55]. The purpose of the example is to investigate the
influence of large curviness on the beam response. If the condition of inextensibility of the beam axis is enforced,
the beam deforms into a circle with curvature χ = 2nπ/L (n = 1, 2, . . .). For this case, the analytical solution is
straightforward [37]. However, if the nonlinear distribution of axial strain along the cross section is considered,
21
A. Borković, M.H. Gfrerer and B. Marussig Computer Methods in Applied Mechanics and Engineering 405 (2023) 115848

Fig. 9. Pure bending of a cantilever beam. Geometry and applied load.

Fig. 10. Pure bending of a cantilever beam. Displacement of the tip for n = 2: (a) h = 0.1, (b) h = 0.2.

coupling between the bending and axial actions occurs. The results for the displacement components of the tip for
n = 2, and cross-sectional heights h = 0.1 and h = 0.2 are compared in Fig. 10. The values of curviness at the
final configuration are 0.126 and 0.251, for cases h = 0.1 and h = 0.2, respectively. Evidently, all constitutive
models are in agreement with analytical predictions for small load values. As the load and curviness increase, the
differences in displacement components become apparent, as emphasized in the zoomed parts of the graphs. The
constitutive model D 1 is fully aligned with the analytical solution, which suggests that Eq. (66) indeed results with
near-zero axial strain of the beam axis.
Furthermore, due to the pure bending conditions, this example is ideally suited for validating the strongly curved
beam model. The expressions for physical normal force and bending moment are given by Eq. (53). For an in-plane
beam they reduce to:
√ √
E g∗ ( E I g∗ (
Aϵ11 + 1.5I χ κ χ ϵ11 + κ .
) )
N= and M = (88)
g g g g
It is reasonable to assume that both, the axial strain and the change of curvature, are constant ( along) the beam
axis. By imposing the conditions N = 0 and M = Mext , and having the relation κ = χ 2ϵ11 + g in mind,
Eqs. (88) reduce to a system of two nonlinear algebraic equations with two unknowns. This allows us to calculate
the reference strains of the beam axis and to test our computational isogeometric model. For this, a dense mesh
of 64 quartic elements is used and the equilibrium paths of the reference axial strain are shown in Fig. 11a for
the three constitutive models and h = 0.1. This evolution is nonlinear for the D C and D 0 models, while the D 1
model returns effectively zero axial strain. Using the D C model, we have obtained the reference axial strain of
ϵ(11)
∗
= −0.0004948 for n = 1, which is in agreement with the values presented in [37]. It is interesting that the
axis compresses, but the ends of the beam overlap. This is due to the fact that the curvature at the final configuration
is not exactly 2nπ/L, but is slightly larger. In concrete terms we have obtained K ∗ = χ ∗ = −0.62956, for n =
1 and h = 0.1, while 2nπ/L = 0.2π = 0.62832. At first, this behavior is counter-intuitive, and it can result in a
misinterpretation of the sign for axial strain as in [55].
22
A. Borković, M.H. Gfrerer and B. Marussig Computer Methods in Applied Mechanics and Engineering 405 (2023) 115848

Fig. 11. Pure bending of a cantilever beam. (a) Axial strain of the beam axis for three constitutive models (h = 0.1, n = 2). (b) L 2 norm
of a difference between the axial strain of beam axis calculated by solving Eqs. (88) and by the D C model.

Fig. 12. Circular ring subjected to twisting. (a) Load and geometry. (b) Comparison of LPF vs. rotation at the point of the application of
load.

Additionally, we have obtained the internal normal force using the equations of the D C model (64) and (65). Its
tensorial value is Ñ g = −395.7, which is again in agreement with the value reported in [37] where this quantity
is named effective axial stress resultant.
Finally, we have compared the values of reference axial strain for n = 2, and h = 0.1 and h = 0.2. Relative
differences of results calculated by the IGA D C model and Eqs. (88) are displayed in Fig. 11b. The differences are
small, but increase with both the load and curviness. Further test, which we have left out of the manuscript, show
that this difference reduces with h-refinement.

5.3. Circular ring subjected to twisting

A circular ring that is subjected to symmetrical twisting is a well-known test for the verification of formulations
involving large rotations and small strains of spatial beams [19,24,50]. The geometry and load are displayed in
Fig. 12a. Here, the external twist is applied through a pair of concentrated moments M = E I /R. Due to the
symmetry of the load and the geometry, only a quarter of the ring is modeled [50,71]. The equilibrium path of the
external angle of twist is commonly observed for the verification of computational models. The results obtained
with a quintic mesh with 16 elements are compared with the reference solution from [72] in Fig. 12b. The FSR and
reference results are in full agreement. As expected, the response obtained with the FSR TF formulation differs.
However, for approximately ϕ ∈ (50, 100◦ ), the equilibrium path is well-aligned with the exact one. The zoomed
part in Fig. 12b shows that the FSR TF passes LPF = 0 twice, near 162◦ and 180◦ . While the FSR TF formulation
disagrees with standard formulations, it is astonishing that this simplified rotation-free model can approximate such
complex behavior.
Our tests show that there is no noticeable difference in displacements between the different constitutive models
since the beam has small curviness [50]. The reference axial strain and normal force, however, are affected by the
constitutive relation as will be discussed later.
23
A. Borković, M.H. Gfrerer and B. Marussig Computer Methods in Applied Mechanics and Engineering 405 (2023) 115848

Fig. 13. Circular ring subjected to twisting. Deformed configurations calculated with the FSR and FSR TF formulations.

Fig. 14. Circular ring subjected to twisting. Comparison of twist angles. (a) Total twist ω of section B for the FSR and FSR TF formulations.
(b) Independent twist angle θ at section A and B using the FSR formulation.

A characteristic feature of this example is that after the external twisting of ϕ = 180◦ , the ring deforms into a
smaller ring, with a diameter reduced by a factor of three. Additional application of the external twisting returns
the ring into its original configuration for ϕ = 360◦ . A graphical representation is given in Fig. 13 for both FSR
and FSR TF formulations. This visualization confirms that the FSR TF formulation fairly approximate the exact
behavior of this beam for some values of the external load.
In order to closely examine this phenomena, the twist angles of cross sections A and B (marked in Fig. 12a)
are observed for ϕ = 360◦ . The total twist angle ω of the cross section B is displayed in Fig. 14a where similar
equilibrium paths for both formulations are observed. A cause for the partial agreement of the two formulations is
revealed in Fig. 14b, where the equilibrium paths of the independent twist angle θ at sections A and B is given.
This angle is relatively small in this case, which allows approximate modeling with the FSR TF formulation.
Further investigation is made by comparing the values of section forces using the D C model and a dense mesh
of 32 quintic elements, see Fig. 15. These results suggest that the FSR TF vaguely follows the true equilibrium
paths but gives erroneous values. The error is particularly pronounced for the moment M2 , due to the fact that the
material basis vectors of the FSR TF model are aligned with the normal and binormal. The effect of this error on
the structural response is partly compensated with the opposite sign of M3 for the FSR and FSR TF before the
load limit point, and by the linear response of M2 . The value of M2 is large because Iζ ζ = 9Iηη , but the changes
of curvature χ2 and χ3 are both around 0.01 near the load limit point, after which χ2 returns to zero, while χ3
shows snap-through behavior for both formulations. All in all, this example shows that the FSR TF formulation can
approximate some parts of the equilibrium path for specific deformation cases.
Next, the influence of constitutive relations is assessed. The results for the reference axial strain ϵ(11) at section
A are given in Fig. 16a for different constitutive models using 32 quintic elements. Interesting equilibrium paths
are obtained and the values differ significantly. The D 1 model returns extension, while the D 0 and D C return
compression. As a check, for the angle of ϕ = 180◦ , there should be no normal force in the ring. In this
configuration, the reference strains of the D C model are ϵ(11) = −0.0000926 and χ3 = 0.100014 which give nearly
zero value of the normal force, cf. Eq. (53). This validates the presented D C model and its usage is recommended
instead of the D 0 , D 1 , and the models suggested in [50]. Furthermore, the normal force distribution along the
24
A. Borković, M.H. Gfrerer and B. Marussig Computer Methods in Applied Mechanics and Engineering 405 (2023) 115848

Fig. 15. Circular ring subjected to twisting. Comparison of the FSR and FSR TF formulations. Stress resultant and stress couples at section
A: (a) normal force, (b) torsional moment, (c) bending moment M2 , (d) bending moment M3 .

Fig. 16. Circular ring subjected to twisting. (a) Comparison of the reference axial strain at section A using the different constitutive models
and FSR formulation. (b) Distribution of normal force for different meshes using the D C model.

modeled part of a ring is given in Fig. 16b. Clearly, the normal force shows oscillatory behavior that mitigates with
h-refinement, which suggests presence of membrane locking.
The final benchmark is related to the path-independence. Due to the cyclic response of this ring, path-dependence
can be easily detected [19,50]. Let us observe the torsional strain at point A while the ring is twisted six times
(ϕ = 12π). It is known that the path-dependence mitigates with the increase in the mesh density [50]. Therefore, for
this test a sparse mesh with 8 quintic elements is used. The results are calculated with the SR and FSR formulations
and compared in Fig. 17. This test confirms the previous observation. The FSR formulation is path-independent
while the SR formulation with incremental update of the local vector basis is not.

5.4. Straight beam bent to helix

In this example, we have considered the response of an initially straight cantilever beam loaded with two end
moments, as indicated in Fig. 18a. Since the FSR formulation requires a well-defined FS frame, the beam axis is
defined with a quadratic spline, and a small initial curvature is imposed by moving the third control point by 2
25
A. Borković, M.H. Gfrerer and B. Marussig Computer Methods in Applied Mechanics and Engineering 405 (2023) 115848

Fig. 17. Circular ring subjected to twisting. Equilibrium path of the torsional curvature at point A during six cycles of twisting: (a) FSR;
(b) SR.

Fig. 18. Straight beam bent to helix. (a) Geometry and load. (b) Deformed configurations.

Fig. 19. Straight beam bent to helix. Comparison of the z-component of the tip displacement for different constitutive models and two
formulations. Zoomed part of the equilibrium path is shown on the right.

along the y-direction. In this way, the initial curvature of the beam analyzed with the FSR approach is practically
constant with the value of K 3 ≈ 4 · 10−6 . The NSRISR model is used for comparison, but without the initial
curvature. The beam is discretized with 30 quintic elements and subjected to the tip moments Mx = Mz = 20. The
three characteristic configurations plotted in Fig. 18b reveal the complex response of this beam that deforms into
a near-perfect helix. Fig. 19 illustrates the z-component of the tip displacement for different constitutive models
26
A. Borković, M.H. Gfrerer and B. Marussig Computer Methods in Applied Mechanics and Engineering 405 (2023) 115848

Fig. 20. Straight beam bent to helix. (a) Distributions of the reference axial strain of the beam axis for LPF = 1 using the NSRISR and
FSR formulations and different constitutive relations. (b) Distribution of the normal force N for LPF = 1 and D C model using the NSRISR
and FSR formulations for three mesh densities of quintic elements.

and formulations. Additionally, the analytical solution for small-curvature beams is employed for the comparison
[24,50].
Regarding the different constitutive models, all return same equilibrium paths for small LPF. As the LPF
increases, the curviness also increases and the differences of displacement become evident. As anticipated, the
D 1 model is aligned with the analytical solution that assumes inextensibility of the beam axis. The results of the
fully uncoupled D 0 model are between the strong- and small-curvature models.
The comparison of the different formulations shows that the FSR approach returns almost identical results as
the NSRIS. The differences are most prominent at displacement limit points and they are attributed to the initial
curviness applied for the FSR model, which does not exist for the NSRISR model. This example shows that the
problem of a non-existent FS frame for straight configurations can be alleviated by imposing the small curvature
without significantly affecting the response.
An interesting aspect of this example is that the beam is in a state of pure bending. The fact that the axial strain
exists while there is no normal force is discussed in [50] and ameliorated results are presented here. Similar to the
example of pure in-plane bending of a beam, cf. Section 5.2, axial strain can be obtained from the equilibrium
conditions with respect to the stress resultant and stress couples, Eq. (53):
√
E g∗ (
Aϵ11 + 1.5Iζ ζ χ2 κ2 + 1.5Iηη χ3 κ3 ,
)
N=
g g
√
E Iζ ζ g ∗ (
χ2 ϵ11 + κ2 ,
)
M2 = (89)
g g
√
E Iηη g ∗ (
χ3 ϵ11 + κ3 .
)
M3 =
g g

By assuming that the axial strain is constant along the beam, we can solve these equations at some fixed section.
For example at the clamped end, we have N (0) = M2 (0) = 0 and M3 (0) = 20, and the resulting axial strain is
ϵ(11) = −0.00754. This result is compared in Fig. 20a with the numerical results obtained by the NSRISR and FSR
formulations using 40 quintic elements. Both formulations return the same results since the deformed configurations
for LPF = 1 match, as shown in Fig. 19. It is important that the values calculated with the D C model correspond
to the solution of equilibrium equations (89). Again, the D 1 model returns zero axial strain.
The issue of oscillatory normal force along the length of the beam is emphasized in Fig. 20b where the results of
both formulations are compared for three mesh densities. Evidently, the NSRISR and FSR return relatively similar
results. The oscillations of the stress resultant reduce with h-refinement, as in the previous example. Evidently, the
D C model correctly predicts zero normal force in this example but requires a dense mesh.
27
A. Borković, M.H. Gfrerer and B. Marussig Computer Methods in Applied Mechanics and Engineering 405 (2023) 115848

6. Conclusions

A geometrically exact isogeometric formulation of the spatial Bernoulli–Euler (BE) beam based on the Frenet–
Serret (FS) frame is presented. This new formulation, designated as the Frenet–Serret Rotation (FSR), employs
an additive decomposition of the total twist angle into the twist of the FS frame and an independent twist angle.
The weak form is rigorously derived and linearized, including the contributions from the constraints and external
loads. Nonlinear terms of strain with respect to the material axes are considered and the effect of strong curvature
is captured through axial-bending coupling. Two simplified constitutive models for the calculation of internal forces
are derived and compared with the exact one.
The FSR formulation is well-suited for the analysis of beams that undergo large rigid-body motions, because
it is objective by definition. Since the FSR requires C 2 continuity, the NURBS-based IGA is an ideal framework
for its implementation. By the consistent treatment of the virtual power and the finite element implementation, the
formulation returns results that are indistinguishable from standard methods. The main shortcoming of the FSR
approach is that it fails for configurations that do not have an uniquely defined FS frame. If such a configuration is
encountered, the implementing of a switch to some other formulation would be a straightforward task. A switch of
this kind was not required, however, in any of the presented standard numerical examples. Moreover, it is shown
that a straight configuration can be approximated with a slightly curved geometry without significantly affecting
the beam’s response. All in all, the FSR formulation presents a significant contribution to the theory of beams,
while having a potential to efficiently and accurately simulate specific mechanical systems composed of deformable
slender bodies that undergo large rigid-body motions.
A nonlinear rotation-free model of the spatial BE beam follows as a special case of the FSR formulation by
omitting the independent twist angle from the DOFs. Importantly, this reduced model employs the FS part of
torsion, which makes it unique compared to the existing rotation-free formulations of spatial beams. In particular,
it can provide approximate solutions for spatial beams that have well-defined FS frames and are predominantly
bent with respect to the binormal of the beam axis. The structural response of such beams can be approximated
by neglecting the independent twist angle, as demonstrated by the behavior of the ring subjected to symmetrical
twisting. However, the formulation cannot be directly applied to general problems since it is difficult to estimate
the value of the independent twist angle in advance.
The definition of the BE beam’s metric requires several assumptions in order to decouple axial and torsional
effects. The aim to capture the effect of the axial-bending coupling is fulfilled by the consistent derivation of
the axial strain at an arbitrary point. Through the strict considerations of the current and initial beam metric, a
computational model suitable for the nonlinear analysis of strongly curved BE beams is obtained. The influence
that large curviness has on axial-bending coupling is evident for standard academic examples.
Since the BE assumption is unsuitable for thick beams, the next step is to derive a strongly curved model of the
shear deformable beams. Also, locking issues should be eliminated, for example using the approach suggested in
[38]. Another interesting topic is to further assess the relationship between the total torsional angle and the FS one
in the context of the torsion-free model. Further work aims to extend the developed formulation to the dynamic
analysis of curved slender bodies with large rigid-body motion.

Declaration of competing interest

The authors declare that they have no known competing financial interests or personal relationships that could
have appeared to influence the work reported in this paper.

Data availability

Data will be made available on request.

Acknowledgment

We acknowledge the support of the Austrian Science Fund (FWF), Austria: M 2806-N.
28
A. Borković, M.H. Gfrerer and B. Marussig Computer Methods in Applied Mechanics and Engineering 405 (2023) 115848

Appendix A. Geometric stiffness matrix

Since the linear increments of the material base vectors are:

∆g∗m = ∆u,m = v,m ∆t, (A.1)
the linearized increments of the virtual reference strain rates are:
∆δd11 = ∆g∗1 · δv,1 = v,1 · δv,1 ∆t,
∆δ κ̇1 = ∆ T∗1 · δv,1 − T∗2 · δv,11 + T∗3 · δv,111 + δ θ̇,1
( )

= ∆T∗1 · δv,1 − ∆T∗2 · δv,11 + ∆T∗3 · δv,111 ,

[ ]
∆δ κ̇2 = ∆ n∗ · δv,11 − Γ11 ∗1
δv,1 sin θ ∗ + K̃ 3∗ δ θ̇ = ∆n∗ · δv,11 − Γ11 ∗1
δv,1 sin θ ∗
( ) ( )
(A.2)
+ n∗ · −∆Γ11 ∗1
δv,1 sin θ ∗ + n∗ · δv,11 − Γ11 ∗1
δv,1 ∆ sin θ ∗ + ∆ K̃ 3∗ δ θ̇ ,
( ) ( )
[ ]
∆δ κ̇3 = ∆ n∗ · δv,11 − Γ11 ∗1
δv,1 cos θ ∗ − K̃ 2∗ δ θ̇ = ∆n∗ · δv,11 − Γ11 ∗1
δv,1 cos θ ∗
( ) ( )

+ n∗ · −∆Γ11 ∗1
δv,1 sin θ ∗ + n∗ · δv,11 − Γ11 ∗1
δv,1 ∆ cos θ ∗ − ∆ K̃ 2∗ δ θ̇ .
( ) ( )

The task is to express these linearized strains as a function of DOFs. In the following, we will adopt ∆t = 1, for
brevity.
The increment of the Christoffel symbol is:
( ∗
g1,1 · g∗1
)
1 [( ♯ ♯ ♯1 ♯
)
♯
]
∗1
∆Γ11 =∆ = K̃ n − Γ g · v,1 + g · v,11 . (A.3)
g∗ g♯ 11 1 1

The increments of the gradients of velocities equal the increments of basis vectors and they follow from Eq. (24):
1 ( ♯ )
♯ ♯
∆v,2 = − ♯ g2 · v,1 g1 + g3 ω̇,
g
(A.4)
1 ( ♯ )
♯ ♯
∆v,3 = − ♯ g3 · v,1 g1 − g2 ω̇.
g
Using the additive decomposition of total angular velocity, Eq. (41), these expressions can be written as a function
of the independent twist velocity θ̇ :
♯1
( )
1 ♯ ♯ Γ11 ♯ ♯ 1 ♯ ♯
∆v,2 = − ♯ g1 ⊗ g2 − g3 ⊗ b v,1 + g3 ⊗ b♯ v,11 + g3 θ̇ ,
g K̃ ♯ K̃ ♯
( ♯1
) (A.5)
1 ♯ ♯ Γ11 ♯ ♯ 1 ♯ ♯ ♯
∆v,3 = − ♯ g1 ⊗ g3 + g2 ⊗ b v,1 − g2 ⊗ b v,11 − g2 θ̇ .
g K̃ ♯ K̃ ♯
Now, the increments of curvature components follow from Eq. (7):
♯ ♯ ♯1 ♯
( )
∆ K̃ 2∗ = ∆ −g∗1,1 · g∗3 = −v,11 · g3 − g1,1 · ∆v,3̄ = n♯ · v,11 − Γ11 v,1 sin θ ♯ + K̃ 3 θ̇ ,
( )
(A.6)
♯ ♯ ♯1 ♯
( )
∆ K̃ 3∗ = ∆ g∗1,1 · g∗2 = v,11 · g2 + g1,1 · ∆v,2̄ = n♯ · v,11 − Γ11 v,1 cos θ ♯ − K̃ 2 θ̇ ,
( )

and they equal the increments of curvature changes. Furthermore, let us define the following increments:
1 2 ♯ 1 1 ♯ ( ♯1
)
∆ ∗ = − ♯2 g1 · v,1 , ∆ =− n · v,11 − Γ11 v,1 . (A.7)
g g K̃ ∗ K̃ ♯2
Now, the linearized increments of the normal and binormal follow from Eqs. (9), (A.3) and (A.7), [52]:
( ♯1 )
Γ11 ♯ ♯ 1 ♯ ♯ 1 ♯
∗
∆n = − b ⊗ b + ♯ g1 ⊗ n v,1 + b ⊗ b♯ v,11 ,
K̃ ♯ g K̃ ♯
( ♯1 ) (A.8)
Γ 11 ♯ 1 ♯ 1 ♯ ♯
∆b∗ = n ⊗ b♯ − ♯ g1 ⊗ b♯ v,1 − n ⊗ b♯ v,11 ,
K̃ ♯ g K̃ ♯

29
A. Borković, M.H. Gfrerer and B. Marussig Computer Methods in Applied Mechanics and Engineering 405 (2023) 115848

while the increments of the sine and cosine of the angle θ are:

∆ sin θ ∗ = θ̇ cos θ ♯ and ∆ cos θ ∗ = −θ̇ sin θ ♯ . (A.9)

These expressions allow us to write the increments of virtual curvature rates as:

♯ ♯ ♯ ♯ ♯ ♯
( ) ( ) ( )
∆δ κ̇2 = v,1 · K̃ 2 W11 δv,1 − v,1 · K̃ 2 W12 δv,11 − v,1 · K̃ 3 W14 δ θ̇
♯ ♯ ♯ ♯
( ) ( )
− v,11 · K̃ 2 W21 δv,1 + v,11 · K̃ 2 W22 δv,11
♯ ♯ ♯ ♯ ♯ ♯ ♯
( ) ( ) ( )
+ v,11 · K̃ 3 W24 δ θ̇ − θ̇ K̃ 3 W41 · δv,1 + θ̇ K̃ 3 W42 · δv,11 − θ̇ K̃ 2 δ θ̇ ,
(A.10)
♯ ♯ ♯ ♯ ♯ ♯
( ) ( ) ( )
∆δ κ̇3 = v,1 · K̃ 3 W11 δv,1 − v,1 · K̃ 3 W12 δv,11 + v,1 · K̃ 2 W14 δ θ̇
♯ ♯ ♯ ♯
( ) ( )
− v,11 · K̃ 3 W21 δv,1 + v,11 · K̃ 3 W22 δv,11
♯ ♯ ♯ ♯ ♯ ♯ ♯
( ) ( ) ( )
− v,11 · K̃ 2 W24 δ θ̇ + θ̇ K̃ 2 W41 · δv,1 − θ̇ K̃ 2 W42 · δv,11 − θ̇ K̃ 3 δ θ̇ ,

where the following designations are introduced:

( )
1 1
Γ11 Γ11 g K̃
W11 = b ⊗ b − 1 n ⊗ n + g1 ⊗ n + n ⊗ g1 ,
g K̃ K̃ Γ11
1
Γ11 1
W12 = W21 T = b⊗b+ n ⊗ g1 , (A.11)
K̃ 2 g K̃
1
Γ11 1 1
W14 = W41 = n, W22 = b ⊗ b, W24 = W42 = n.
K̃ K̃ 2 K̃

The linearization of the virtual rate of torsion is straightforward, but significantly more involved. We will
rearrange the expression in Eq. (43) as:

G∗
( )
1 ( 1 ∗
κ̇1 = ∗1
v,1 − v,11 · G ∗2 b∗ + G ∗3 n∗ + b · v,111 − ∗1 v,1 + θ̇,1 ,
) ( )
Γ11 (A.12)
K̃ ∗2 K̃ ∗ g

where the following components of g1,11 vector are introduced:

[ ( 1 )2 ]
1
G 1 = g1,11 · g1 = g Γ11,1 + Γ11 − K K̃ ,
1
G 2 = g1,11 · n = Γ11 K̃ + K̃ ,1 , (A.13)
G 3 = g1,11 · b = K̃ τ̃ .

Now, the linearized increment of the virtual rate of torsional curvature can be expressed as:
( )
1 ♯1
( ) (
♯ ♯
) 1 ( ) ( ♯ ♯
)
∆δ κ̇1 = ∆ Γ11 δv,1 − δv,11 · G 2 b♯ + G 3 n♯ + ∗1
∆Γ11 δv,1 · G 2 b♯ + G 3 n♯
K̃ ∗2 K̃ ♯2
1 (
♯1
) (
♯ ♯
)
+ Γ11 δv,1 − δv,11 · ∆G ∗2 b♯ + G 2 ∆b∗ + ∆G ∗3 n♯ + G 3 ∆n∗ (A.14)
K̃ ♯2
♯
( ) ( ) ( )
1 ♯ 1 G 1 ♯ ∗ 1 ♯ 1
+ ∆ b + ∆b · δv,111 − ♯ δv,1 −
∗ 1
b · ∆G 1 ♯ + G 1 ∆ ∗ δv,1 .
K̃ ∗ K̃ ♯ g K̃ ♯ g g
30
A. Borković, M.H. Gfrerer and B. Marussig Computer Methods in Applied Mechanics and Engineering 405 (2023) 115848

Moreover, we need the following linearized increments:

( ♯ )
G1 ♯ ♯ ♯ ♯ ♯ ♯
∗
∆G 1 = g + G 2 n + G 3 b · v,1 + g1 · v,111 ,
g♯ 1
♯
G (
♯1
)
♯
∆G ∗2 = 3 b♯ · v,11 − Γ11 v,1 − G 1 n♯ · v,1 + n♯ · v,111 , (A.15)
K̃ ♯
♯
G (
♯1
)
♯
∆G ∗3 = − 2 b♯ · v,11 − Γ11 v,1 − G 1 b♯ · v,1 + b♯ · v,111 .
K̃ ♯
By inserting Eqs. (A.3), (A.8), and (A.15) into Eq. (A.14), we obtain:
♯ ♯ ♯ ♯
∆δ κ̇1 = v,1 · X11 δv,1 + v,1 · X12 δv,11 + v,1 · X13 δv,111 + v,11 · X21 δv,1
♯ ♯ ♯ ♯ (A.16)
+ v,11 · X22 δv,11 + v,11 · X23 δv,111 + v,111 · X31 δv,1 + v,111 · X32 δv,11 ,

where:
⎡ ( )2 ⎤
1
Γ11 1 τ̃ Γ11
1 (
= τ̄ ⎣2 + ⎦ (n ⊗ n − b ⊗ b) −
)
X11 g1 ⊗ n + n ⊗ g1
K̃ g g K̃
( )[ ]
1 1
Γ11 G2 G1 Γ11 1 (
(n ⊗ b + b ⊗ n) − g1 ⊗ b + b ⊗ g1 ,
)
+ − 2
K̃ g K̃ 2 g K̃
( )
τ̃ Γ111
1 G1 1
Γ11 G2
X12 = X21 = 2T
(b ⊗ b − n ⊗ n) + −2 (b ⊗ n + n ⊗ b)
K̃ 2 K̃ 2 g K̃
G2 τ (A.17)
+ b ⊗ g1 + n ⊗ g1 ,
g K̃ 2 g K̃
Γ1 1
X13 = X31 T = 11 (b ⊗ n + n ⊗ b) − b ⊗ g1 ,
K̃ 2 g K̃
τ̄ G2
X22 =2 (n ⊗ n − b ⊗ b) + 2 (n ⊗ b + b ⊗ n) ,
K̃ 2 K̃ 3
1
X23 = X32 T = − (n ⊗ b + b ⊗ n) .
K̃ 2
Let us introduce the following submatrices:
( )
G11 = N I3×3 + M1 X11 + M2 K̃ 2 + M3 K̃ 3 W11 ,
( )
G12 = M1 X12 − M2 K̃ 2 + M3 K̃ 3 W12 ,
( )
G13 = M1 X13 , G14 = −M2 K̃ 3 + M3 K̃ 2 W14 ,
( ) (A.18)
G22 = M1 X22 + M2 K̃ 2 + M3 K̃ 3 W22 ,
( )
G23 = M1 X23 , G24 = M2 K̃ 3 − M3 K̃ 2 W24 ,
G33 = 03×3 , G34 = 03×1 , G44 = −M2 K̃ 2 − M3 K̃ 3 ,

which constitute the matrix of generalized section forces:

⎡ ⎤
G11 G12 G13 G14
⎢ G22 G23 G24 ⎥
⎥.
G=⎢ (A.19)
⎣ G33 G34 ⎦
sym. G44
31
A. Borković, M.H. Gfrerer and B. Marussig Computer Methods in Applied Mechanics and Engineering 405 (2023) 115848

Additionally, we need the matrix of basis functions BG :

BG = BG1 BG2 ... BG I ... BG N ,
[ ]
[ ] (A.20)
BG I T = N I,1 T N I,11 T N I,111 T NθI T .
Now, the part of the virtual power due to the known stress and the increment of the virtual strain rate in Eq. (60)
can be expressed as:
∫ ∫ ∫
√ √ √ ♯ T
f ♯T ∆δe g dξ = f ♯T ∆H∗ B g dξ δ q̇ = q̇T BG T G♯ BG g dξ δ q̇ = q̇T KG δ q̇. (A.21)
ξ ξ ξ

Note that the derived geometric stiffness matrix KG is symmetric. This confirms that the energetically conjugated
pairs are correctly adopted.

Appendix B. Linearization of constraint equation

In order to linearize the constraint condition (82), the increment of the constraint rate is required. This increment
equals the increment of the total twist velocity ω̇:
( )
1 (
♯1
) 1 (
♯1
) 1 ♯ (
b♯ · δv,11 − Γ11 δv,1 + ∆b∗ · δv,11 − Γ11 δv,1 + ∗1
δv,1
)
∆δ ċ = ∆ b · −∆Γ11
K̃ ∗ K̃ ♯ K̃ ♯ (B.1)
♯ ♯ ♯ ♯
= v,1 · Ḡλ11 δv,1 + v,1 · Ḡλ12 δv,11 + v,11 · Ḡλ21 δv,1 − v,11 · Ḡλ22 δv,11 .
where:
( )2
1 ( 1
Γ11 Γ11 1
(n ⊗ b + b ⊗ n) −
)
Ḡλ11 = g1 ⊗ b + b ⊗ g1 − b ⊗ n,
g K̃ K̃ g
1
Γ11 1 (B.2)
Ḡλ12 = Ḡλ21 T = (n ⊗ b + b ⊗ n) − b ⊗ g1 ,
K̃ 2 g K̃
1
Ḡλ22 = − (n ⊗ b + b ⊗ n) .
K̃ 2
Now, if we introduce the following matrices:
[ ]
Ḡλ11 Ḡλ12
Ḡλ = ,
Ḡ Ḡλ22
[ λ21 (B.3)
Bλ = Bλ1 Bλ2 ... BλI ... BλN ,
]

BλI T = N I,1 T N I,11 T .

[ ]

♯
the increment of the constraint rate is ∆δ ċ = δ q̇λ T Bλ T Gλ Bλ q̇λ , and we can derive Eq. (86):
♯ ♯
λ♯ ∆δ ċ = δ q̇λ T Bλ T λ♯ Gλ Bλ q̇λ = δ q̇λ T KGλ q̇λ . (B.4)
The constraint condition and its contribution to the tangent matrix must be evaluated at the fixed coordinate
ξ = ξc , where the boundary condition is defined. Obviously, the resulting geometric stiffness matrix KGλ is not
symmetric due to the one term in Ḡλ11 . However, our simulations show that this term has negligible influence on
the performance of nonlinear solver.

Appendix C. Linearization of external virtual power

Let us consider the contribution to the tangent stiffness that comes from the configuration-dependent concentrated
moment, since it was the load used in the numerical analysis. If the external moment is constant with respect to the
global coordinate system, the load vector must be updated at every time step due to the change of the local vector
basis. The increment of the external virtual power is:
♯
∆δ Pext = −∆ m∗ · δ ω̇ = −∆ m∗ · gi∗ δωi = −m∗ · ∆gi∗ δωi − m∗ · gi ∆δωi ,
( ) ( )
(C.1)
32
A. Borković, M.H. Gfrerer and B. Marussig Computer Methods in Applied Mechanics and Engineering 405 (2023) 115848

while the increments of virtual angular velocities follow from Eqs. (22) and (B.1). The increment of the virtual
twist velocity is:
[ ]
1 ( 1 ( 1
∆δ ω̇ = ∆ √ ∗ δ ω̇ F S + δ θ̇ = ∆ √ ∗ δ ω̇ F S + δ θ̇ + √ ∆δ ω̇ F S
1
) )
g g g♯
⎧ ⎡(
♯1 ♯1 2
) ( )
1 ⎨ Γ11 ( ♯ ♯ ♯ ♯
) Γ11 ( ♯
b ⊗ n♯ + n♯ ⊗ b♯
)
=√ v,1 · ⎣ 2g1 ⊗ b + b ⊗ g1 −
g♯ ⎩ g ♯ K̃ ♯ K̃ ♯
] [ ]
1 ♯ 1
− ♯ n ⊗ b♯ δv,1 − v,11 · n♯ ⊗ b♯ + b♯ ⊗ n♯ δv,11
( )
(C.2)
g K̃ ♯2
[ ♯1 ]
Γ11 ( ♯ ♯ ♯ ♯ 1 ( ♯ ♯ ♯ ♯
)
δv,11
)
+ v,1 · n ⊗b +b ⊗n − b ⊗ g1 + g1 ⊗ b
K̃ ♯2 g ♯ K̃ ♯
[ ♯1 ] ( ♯) }
Γ11 ( ♯ 1 ♯ g
n ⊗ b♯ + b♯ ⊗ n♯ − g1 ⊗ b♯ δv,1 − θ̇ 1
δv,1 ,
)
+v,11 ·
K̃ ♯2 g ♯ K̃ ♯ g♯

while the increments of the other two components are:

♯1
( ) [ ]
1 ♯ 1 (
♯ ♯ ♯ ♯
) Γ ♯
∆δω2 = ∆ − √ g3 · δv,1 = v,1 · g ⊗ g1 + g1 ⊗ g3 − √ 11 b♯ ⊗ g2 δv,1
g♯ g ♯3/2 3 g ♯ K̃ ♯
1 (
♯
) 1 ♯
+√ v,11 · b♯ ⊗ g2 δv,1 + √ θ̇ g2 · δv,1 ,
g ♯ K̃ ♯ g♯
( ) [ ♯1
] (C.3)
1 ♯ 1 ( ♯ ♯ ♯ ♯
) Γ11 ♯ ♯
∆δω = ∆ √ g2 · δv,1 = −v,1 ·
3
g ⊗ g1 + g1 ⊗ g2 + √ b ⊗ g3 δv,1
g♯ g ♯3/2 2 g ♯ K̃ ♯
1 (
♯
) 1 ♯
+√ v,11 · b♯ ⊗ g3 δv,1 + √ θ̇ g3 · δv,1 .
♯
g K̃ ♯ g♯

The required increments of the basis vectors are given with Eqs. (A.1) and (A.5). By the insertion of Eqs. (C.2),
(C.3) and (A.5) into Eq. (C.1), we obtain:

♯ ♯ ♯ ♯
−∆δ Pext = v,1 · Ḡ11 T δv,1 + v,11 · Ḡ12 T δv,1 + v,1 · Ḡ21 T δv,11 − v,11 · Ḡ22 T δv,11
♯
(C.4)
+ v,1 · Ḡ41 T δ θ̇ ,

where:
⎡ ( )2
1
1( Γ11
∗
(b ⊗ n + n ⊗ b)
( ) )
Ḡ11 = m · t ⎣ g2 ⊗ g3 − g3 ⊗ g2 +
g K̃
]
1 (
Γ11 ) 1 1 (
g1 ⊗ b + 2b ⊗ g1 + b ⊗ n − 3/2 m∗ · g2 g3 ⊗ g1 + g1 ⊗ g3
)( )
−
g K̃ g g
1 ( Γ1
g2 ⊗ g1 + g1 ⊗ g2 + √ 11 b ⊗ m∗ ,
m∗ · g3
)( )
+
g 3/2 g K̃
[ ]
1
( ∗ ) 1 Γ11
Ḡ12 = m ·t b ⊗ g1 − (b ⊗ n + n ⊗ b) ,
g K̃ K̃ 2
33
A. Borković, M.H. Gfrerer and B. Marussig Computer Methods in Applied Mechanics and Engineering 405 (2023) 115848

[ ]
1
1 ( Γ11 1
Ḡ21 = m∗ · t (b ⊗ n + n ⊗ b) − √ b ⊗ m∗ ,
( ) )
g1 ⊗ b + b ⊗ g1 −
g K̃ K̃ 2 g K̃
1 ( ∗ )
Ḡ22 = m · t (b ⊗ n + n ⊗ b) , (C.5)
K̃ 2
1 [( ∗ ) T ]
Ḡ41 =√ m · t t − m∗ ,
g
[ ]
define non-zero submatrices of the matrix Ḡ = Ḡmn with m, n = 1, 2, 3, 4. Now, the geometric stiffness matrix
♯Q
due to the configuration-dependent load can be written as KG = BG T Ḡ♯ BG . This matrix must be evaluated at the
point of the application of load and added to the tangent stiffness matrix in Eq. (78).

References
[1] C. Meier, A. Popp, W.A. Wall, Geometrically exact finite element formulations for slender beams: Kirchhoff–Love theory versus
Simo–Reissner theory, Arch. Comput. Methods Eng. 26 (1) (2019) 163–243, http://dx.doi.org/10.1007/s11831-017-9232-5.
[2] E. Reissner, On finite deformations of space-curved beams, Z. Angew. Math. Phys. 32 (6) (1981) 734–744, http://dx.doi.org/10.1007/
BF00946983.
[3] J.C. Simo, A finite strain beam formulation. The three-dimensional dynamic problem. Part I, Comput. Methods Appl. Mech. Engrg.
49 (1) (1985) 55–70, http://dx.doi.org/10.1016/0045-7825(85)90050-7.
[4] J.C. Simo, L. Vu-Quoc, A three-dimensional finite-strain rod model. part II: Computational aspects, Comput. Methods Appl. Mech.
Engrg. 58 (1) (1986) 79–116, http://dx.doi.org/10.1016/0045-7825(86)90079-4.
[5] A. Cardona, M. Geradin, A beam finite element non-linear theory with finite rotations, Internat. J. Numer. Methods Engrg. 26 (11)
(1988) 2403–2438, http://dx.doi.org/10.1002/nme.1620261105.
[6] J.C. Simo, L. Vu-Quoc, On the dynamics in space of rods undergoing large motions – A geometrically exact approach, Comput.
Methods Appl. Mech. Engrg. 66 (2) (1988) 125–161, http://dx.doi.org/10.1016/0045-7825(88)90073-4.
[7] M. Iura, S.N. Atluri, On a consistent theory, and variational formulation of finitely stretched and rotated 3-D space-curved beams,
Comput. Mech. 4 (2) (1988) 73–88, http://dx.doi.org/10.1007/BF00282411.
[8] A. Ibrahimbegović, On finite element implementation of geometrically nonlinear Reissner’s beam theory: three-dimensional curved
beam elements, Comput. Methods Appl. Mech. Engrg. 122 (1) (1995) 11–26, http://dx.doi.org/10.1016/0045-7825(95)00724-F.
[9] A. Ibrahimbegovic, On the choice of finite rotation parameters, Comput. Methods Appl. Mech. Engrg. 149 (1) (1997) 49–71,
http://dx.doi.org/10.1016/S0045-7825(97)00059-5, Containing papers presented at the Symposium on Advances in Computational
Mechanics.
[10] M.A. Crisfield, Non-Linear Finite Element Analysis of Solids and Structures, Volume 2, Advanced Topics – Wiley. Vol. 2, Wiley,
1997.
[11] M. Crisfield, G. Jelenic, Objectivity of strain measures in the geometrically exact three-dimensional beam theory and its finite-element
implementation, Proc. R. Soc. Math. Phys. Eng. Sci. 455 (1983) (1999) 1125–1147, http://dx.doi.org/10.1098/rspa.1999.0352.
[12] G. Jelenić, M.A. Crisfield, Geometrically exact 3D beam theory: implementation of a strain-invariant finite element for statics and
dynamics, Comput. Methods Appl. Mech. Engrg. 171 (1) (1999) 141–171, http://dx.doi.org/10.1016/S0045-7825(98)00249-7.
[13] J. Mäkinen, Rotation manifold SO(3) and its tangential vectors, Comput. Mech. 42 (6) (2008) 907, http://dx.doi.org/10.1007/s00466-
008-0293-z.
[14] S. Ghosh, D. Roy, A frame-invariant scheme for the geometrically exact beam using rotation vector parametrization, Comput. Mech.
44 (1) (2009) 103, http://dx.doi.org/10.1007/s00466-008-0358-z.
[15] I. Romero, The interpolation of rotations and its application to finite element models of geometrically exact rods, Comput. Mech. 34
(2) (2004) 121–133, http://dx.doi.org/10.1007/s00466-004-0559-z.
[16] E. Zupan, M. Saje, D. Zupan, On a virtual work consistent three-dimensional Reissner–Simo beam formulation using the quaternion
algebra, Acta Mech. 224 (8) (2013) 1709–1729, http://dx.doi.org/10.1007/s00707-013-0824-3.
[17] K.M. Hsiao, J.Y. Lin, W.Y. Lin, A consistent co-rotational finite element formulation for geometrically nonlinear dynamic analysis of
3-D beams, Comput. Methods Appl. Mech. Engrg. 169 (1) (1999) 1–18, http://dx.doi.org/10.1016/S0045-7825(98)00152-2.
[18] T.-N. Le, J.-M. Battini, M. Hjiaj, A consistent 3D corotational beam element for nonlinear dynamic analysis of flexible structures,
Comput. Methods Appl. Mech. Engrg. 269 (2014) 538–565, http://dx.doi.org/10.1016/j.cma.2013.11.007.
[19] D. Magisano, et al., A large rotation finite element analysis of 3D beams by incremental rotation vector and exact strain measure with
all the desirable features, Comput. Methods Appl. Mech. Engrg. 361 (2020) 112811, http://dx.doi.org/10.1016/j.cma.2019.112811.
[20] I. Romero, A comparison of finite elements for nonlinear beams: the absolute nodal coordinate and geometrically exact formulations,
Multibody Syst. Dyn. 20 (1) (2008) 51–68, http://dx.doi.org/10.1007/s11044-008-9105-7.
[21] J. Gerstmayr, H. Sugiyama, A. Mikkola, Review on the absolute nodal coordinate formulation for large deformation analysis of
multibody systems, J. Comput. Nonlinear Dyn. 8 (3) (2013) http://dx.doi.org/10.1115/1.4023487.
[22] H. Weiss, Dynamics of geometrically nonlinear rods: I. Mechanical models and equations of motion, Nonlinear Dynam. 30 (4) (2002)
357–381, http://dx.doi.org/10.1023/A:1021268325425.
34
A. Borković, M.H. Gfrerer and B. Marussig Computer Methods in Applied Mechanics and Engineering 405 (2023) 115848

[23] F. Boyer, et al., Geometrically exact Kirchhoff beam theory: Application to cable dynamics, J. Comput. Nonlinear Dyn. 6 (041004)
(2011) http://dx.doi.org/10.1115/1.4003625.
[24] C. Meier, A. Popp, W.A. Wall, An objective 3D large deformation finite element formulation for geometrically exact curved Kirchhoff
rods, Comput. Methods Appl. Mech. Engrg. 278 (2014) 445–478, http://dx.doi.org/10.1016/j.cma.2014.05.017.
[25] C. Meier, A. Popp, W.A. Wall, A locking-free finite element formulation and reduced models for geometrically exact Kirchhoff rods,
Comput. Methods Appl. Mech. Engrg. 290 (2015) 314–341, http://dx.doi.org/10.1016/j.cma.2015.02.029.
[26] C. Meier, W.A. Wall, A. Popp, A unified approach for beam-to-beam contact, Comput. Methods Appl. Mech. Engrg. 315 (2017)
972–1010, http://dx.doi.org/10.1016/j.cma.2016.11.028.
[27] N. Jamun Kumar, et al., A novel four-field mixed FE approximation for Kirchhoff rods using Cartan’s moving frames, Comput. Methods
Appl. Mech. Engrg. (2022) 115094, http://dx.doi.org/10.1016/j.cma.2022.115094.
[28] 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 (39) (2005) 4135–4195, http://dx.doi.org/10.1016/j.cma.2004.10.008.
[29] E. Marino, Isogeometric collocation for three-dimensional geometrically exact shear-deformable beams, Comput. Methods Appl. Mech.
Engrg. 307 (2016) 383–410, http://dx.doi.org/10.1016/j.cma.2016.04.016.
[30] E. Marino, Locking-free isogeometric collocation formulation for three-dimensional geometrically exact shear-deformable beams with
arbitrary initial curvature, Comput. Methods Appl. Mech. Engrg. 324 (2017) 546–572, http://dx.doi.org/10.1016/j.cma.2017.06.031.
[31] O. Weeger, S.-K. Yeung, M.L. Dunn, Isogeometric collocation methods for Cosserat rods and rod structures, Comput. Methods Appl.
Mech. Engrg. 316 (2017) 100–122, http://dx.doi.org/10.1016/j.cma.2016.05.009, Special Issue on Isogeometric Analysis: Progress and
Challenges.
[32] O. Weeger, B. Narayanan, M.L. Dunn, Isogeometric collocation for nonlinear dynamic analysis of Cosserat rods with frictional contact,
Nonlinear Dynam. 91 (2) (2018) 1213–1227, http://dx.doi.org/10.1007/s11071-017-3940-0.
[33] E. Marino, J. Kiendl, L. De Lorenzis, Isogeometric collocation for implicit dynamics of three-dimensional beams undergoing finite
motions, Comput. Methods Appl. Mech. Engrg. 356 (2019) 548–570, http://dx.doi.org/10.1016/j.cma.2019.07.013.
[34] D. Vo, P. Nanakorn, T.Q. Bui, A total Lagrangian timoshenko beam formulation for geometrically nonlinear isogeometric analysis of
spatial beam structures, Acta Mech. 231 (9) (2020) 3673–3701, http://dx.doi.org/10.1007/s00707-020-02723-6.
[35] A. Tasora, et al., A geometrically exact isogeometric beam for large displacements and contacts, Comput. Methods Appl. Mech. Engrg.
358 (2020) 112635, http://dx.doi.org/10.1016/j.cma.2019.112635.
[36] M.-J. Choi, S. Cho, Isogeometric configuration design sensitivity analysis of geometrically exact shear-deformable beam structures,
Comput. Methods Appl. Mech. Engrg. 351 (2019) 153–183, http://dx.doi.org/10.1016/j.cma.2019.03.032.
[37] M.-J. Choi, R. Sauer, S. Klinkel, An isogeometric finite element formulation for geometrically exact Timoshenko beams with extensible
directors, Comput. Methods Appl. Mech. Engrg. 385 (2021) 113993, http://dx.doi.org/10.1016/j.cma.2021.113993.
[38] D. Magisano, L. Leonetti, G. Garcea, Isogeometric analysis of 3D beams for arbitrarily large rotations: Locking-free and path-
independent solution without displacement DOFs inside the patch, Comput. Methods Appl. Mech. Engrg. 373 (2021) 113437,
http://dx.doi.org/10.1016/j.cma.2020.113437.
[39] L. Greco, M. Cuomo, B-spline interpolation of Kirchhoff-Love space rods, Comput. Methods Appl. Mech. Engrg. 256 (2013) 251–269,
http://dx.doi.org/10.1016/j.cma.2012.11.017.
[40] L. Greco, M. Cuomo, An implicit G1 multi patch B-spline interpolation for Kirchhoff–Love space rod, Comput. Methods Appl. Mech.
Engrg. 269 (2014) 173–197, http://dx.doi.org/10.1016/j.cma.2013.09.018.
[41] L. Greco, M. Cuomo, An isogeometric implicit G1 mixed finite element for Kirchhoff space rods, Comput. Methods Appl. Mech.
Engrg. 298 (2016) 325–349, http://dx.doi.org/10.1016/j.cma.2015.06.014.
[42] L. Greco, M. Cuomo, Consistent tangent operator for an exact Kirchhoff rod model, Contin. Mech. Thermodyn. 27 (4) (2015) 861–877,
http://dx.doi.org/10.1007/s00161-014-0361-x.
[43] A.M. Bauer, et al., Nonlinear isogeometric spatial Bernoulli beam, Comput. Methods Appl. Mech. Engrg. 303 (2016) 101–127,
http://dx.doi.org/10.1016/j.cma.2015.12.027.
[44] A.M. Bauer, R. Wüchner, K.-U. Bletzinger, Weak coupling of nonlinear isogeometric spatial Bernoulli beams, Comput. Methods Appl.
Mech. Engrg. 361 (2020) 112747, http://dx.doi.org/10.1016/j.cma.2019.112747.
[45] D. Vo, P. Nanakorn, T. Bui, Geometrically nonlinear multi-patch isogeometric analysis of spatial Euler–Bernoulli beam structures,
Comput. Methods Appl. Mech. Engrg. 380 (2021) 113808, http://dx.doi.org/10.1016/j.cma.2021.113808.
[46] L. Greco, A. Scrofani, M. Cuomo, A non-linear symmetric G1-conforming Bézier finite element formulation for the analysis of
Kirchhoff beam assemblies, Comput. Methods Appl. Mech. Engrg. 387 (2021) 114176, http://dx.doi.org/10.1016/j.cma.2021.114176.
[47] L. Greco, et al., An updated Lagrangian Bézier finite element formulation for the analysis of slender beams, Math. Mech. Solids
(2022) http://dx.doi.org/10.1177/10812865221101549.
[48] Y.B. Yang, Y.Z. Liu, Invariant isogeometric formulation for the geometric stiffness matrix of spatial curved Kirchhoff rods, Comput.
Methods Appl. Mech. Engrg. 377 (2021) 113692, http://dx.doi.org/10.1016/j.cma.2021.113692.
[49] S. Herath, G. Yin, On the geometrically exact formulations of finite deformable isogeometric beams, Comput. Mech. (2021)
http://dx.doi.org/10.1007/s00466-021-02015-3.
[50] A. Borković, B. Marussig, G. Radenković, Geometrically exact static isogeometric analysis of an arbitrarily curved spatial
Bernoulli–Euler beam, Comput. Methods Appl. Mech. Engrg. 390 (2022) 114447, http://dx.doi.org/10.1016/j.cma.2021.114447.
[51] R.A. Hord, Torsion at an inflection point of a space curve, Am. Math. Mon. 79 (4) (1972) 371–374, http://dx.doi.org/10.2307/2978088.
[52] G. Radenković, A. Borković, Linear static isogeometric analysis of an arbitrarily curved spatial Bernoulli–Euler beam, Comput. Methods
Appl. Mech. Engrg. 341 (2018) 360–396, http://dx.doi.org/10.1016/j.cma.2018.07.010.
[53] S.B. Raknes, et al., Isogeometric rotation-free bending-stabilized cables: Statics, dynamics, bending strips and coupling with shells,
Comput. Methods Appl. Mech. Engrg. 263 (2013) 127–143, http://dx.doi.org/10.1016/j.cma.2013.05.005.
35
A. Borković, M.H. Gfrerer and B. Marussig Computer Methods in Applied Mechanics and Engineering 405 (2023) 115848

[54] G. Radenković, A. Borković, On the analytical approach to the linear analysis of an arbitrarily curved spatial Bernoulli–Euler beam,
Appl. Math. Model. 77 (2020) 1603–1624, http://dx.doi.org/10.1016/j.apm.2019.09.012.
[55] A. Borković, B. Marussig, G. Radenković, Geometrically exact static isogeometric analysis of arbitrarily curved plane Bernoulli–Euler
beam, Thin-Walled Struct. 170 (2022) 108539, http://dx.doi.org/10.1016/j.tws.2021.108539.
[56] A. Borković, et al., Rotation-free isogeometric analysis of an arbitrarily curved plane Bernoulli–Euler beam, Comput. Methods Appl.
Mech. Engrg. 334 (2018) 238–267, http://dx.doi.org/10.1016/j.cma.2018.02.002.
[57] A. Borković, et al., Rotation-free isogeometric dynamic analysis of an arbitrarily curved plane Bernoulli-Euler beam, Eng. Struct. 181
(2019) 192–215, http://dx.doi.org/10.1016/j.engstruct.2018.12.003.
[58] V. Slivker, Mechanics of Structural Elements: Theory and Applications, Springer, 2010.
[59] A. Cazzani, et al., Constitutive models for strongly curved beams in the frame of isogeometric analysis, Math. Mech. Solids (2015)
http://dx.doi.org/10.1177/1081286515577043.
[60] G. Radenković, Finite Rotation and Finite Strain Isogeometric Structural Analysis, Faculty of Architecture Belgrad, 2017, (in Serbian).
[61] L. Piegl, W. Tiller, The NURBS Book, in: Monographs in Visual Communication, Springer-Verlag, Berlin Heidelberg, 1995,
http://dx.doi.org/10.1007/978-3-642-97385-7.
[62] R.L. Bishop, There is more than one way to frame a curve, Am. Math. Mon. 82 (3) (1975) 246–251, http://dx.doi.org/10.2307/2319846.
[63] M. Ebrahimi, A. Butscher, H. Cheong, A low order, torsion deformable spatial beam element based on the absolute nodal coordinate
formulation and Bishop frame, Multibody Syst. Dyn. 51 (3) (2021) 247–278, http://dx.doi.org/10.1007/s11044-020-09765-7.
[64] G. Radenković, A. Borković, B. Marussig, Nonlinear static isogeometric analysis of arbitrarily curved Kirchhoff-Love shells, Int. J.
Mech. Sci. 192 (2021) 106143, http://dx.doi.org/10.1016/j.ijmecsci.2020.106143.
[65] R.K. Kapania, J. Li, On a geometrically exact curved/twisted beam theory under rigid cross-section assumption, Comput. Mech. 30
(5) (2003) 428–443, http://dx.doi.org/10.1007/s00466-003-0421-8.
[66] M. Géradin, D. Rixen, Parametrization of finite rotations in computational dynamics: a review, Rev. Eur. Élém. Finis 4 (5–6) (1995)
497–553, http://dx.doi.org/10.1080/12506559.1995.10511200.
[67] S.G. Erdelj, G. Jelenić, A. Ibrahimbegović, Geometrically non-linear 3D finite-element analysis of micropolar continuum, Int. J. Solids
Struct. 202 (2020) 745–764, http://dx.doi.org/10.1016/j.ijsolstr.2020.05.025.
[68] Y.B. Yang, Y. Liu, Y.T. Wu, Invariant isogeometric formulations for three-dimensional Kirchhoff rods, Comput. Methods Appl. Mech.
Engrg. 365 (2020) 112996, http://dx.doi.org/10.1016/j.cma.2020.112996.
[69] F. Armero, J. Valverde, Invariant hermitian finite elements for thin Kirchhoff rods. II: The linear three-dimensional case, Comput.
Methods Appl. Mech. Engrg. 216 (2012) 458–485, http://dx.doi.org/10.1016/j.cma.2011.05.014.
[70] M. Smith, ABAQUS/Standard User’s Manual, Version 6.9, 2009.
[71] G. Yoshiaki, et al., Elastic buckling phenomenon applicable to deployable rings, Int. J. Solids Struct. 29 (7) (1992) 893–909,
http://dx.doi.org/10.1016/0020-7683(92)90024-N.
[72] P.F. Pai, A.N. Palazotto, Large-deformation analysis of flexible beams, Int. J. Solids Struct. 33 (9) (1996) 1335–1353, http:
//dx.doi.org/10.1016/0020-7683(95)00090-9.

36