Steinmann2012 Article HyperelasticModelsForRubber-li

Arch Appl Mech (2012) 82:1183–1217
DOI 10.1007/s00419-012-0610-z
O R I G I NA L
Paul Steinmann · Mokarram Hossain · Gunnar Possart
Hyperelastic models for rubber-like materials: consistent

tangent operators and suitability for Treloar’s data
Received: 3 August 2011 / Accepted: 8 January 2012 / Published online: 2 February 2012
© Springer-Verlag 2012
Abstract Rubber-like materials consist of chain-like macromolecules that are more or less closely connected
to each other via entanglements or cross-links. As an idealisation, this particular structure can be described as
a completely random three-dimensional network. To capture the elastic and nearly incompressible mechanical
behaviour of this material class, numerous phenomenological and micro-mechanically motivated models have
been proposed in the literature. This contribution reviews fourteen selected representatives of these models,
derives analytical stress–stretch relations for certain homogeneous deformation modes and summarises the
details required for stress tensors and consistent tangent operators. The latter, although prevalently missing
in the literature, are indispensable ingredients in utilising any kind of constitutive model for the numerical
solution of boundary value problems by iterative approaches like the Newton–Raphson scheme. Furthermore,
performance and validity of the models with regard to the classical experimental data on vulcanised rubber
published by Treloar (Trans Faraday Soc 40:59–70, 1944) are evaluated. These data are here considered as a
prototype or worst-case scenario of highly nonlinear elastic behaviour, although inelastic characteristics are
clearly observable but have been tacitly ignored by many other authors.
Keywords Rubber-like material · Hyperelasticity · Phenomenological model · Micro-mechanical model
1 Introduction and outline
Since the beginning of the twentieth century, the mechanics of rubber-like materials has been a very active
field of research due to their wide range of applications. In particular, the demand for predictive analysis tools
facilitating the economic design of complex devices for different loading conditions is continuously increasing.
Famous rubber applications are, for example, tires, seals, conveyor belts, base isolations of buildings and many
more. In the last few decades, developments in computational mechanics, especially in finite element methods,
have enabled three-dimensional, large strain analyses of complex elastomeric products to be an integral part of
design processes. This progress, vice versa, has led to a more critical assessment and further development of
constitutive models for rubber elasticity since an accurate reproduction of the three-dimensional stress–strain
behaviour is indispensable for any numerical simulation of complex deformations.
P. Steinmann (B) · M. Hossain · G. Possart

University of Erlangen-Nuremberg, Erlangen, Germany
E-mail: paul.steinmann@ltm.uni-erlangen.de
M. Hossain
E-mail: mokarram.hossain@ltm.uni-erlangen.de
G. Possart
E-mail: gunnar.possart@ltm.uni-erlangen.de
1184 P. Steinmann et al.
The starting point of hyperelastic material modelling is the formulation of a scalar strain (or stored or
potential) energy function. Such functions usually depend either on strain tensors like the right (or left) Cau-
chy-Green tensor C = F t F (or b = F F t ) with F = ∂ϕ/∂X being the material gradient of the nonlin-
ear deformation map ϕ, on invariants of these strain tensors, that is, I1 (C) = tr(C) = I1 (b) , I2 (C) =
2 [tr (C) − tr(C )] = I2 (b) , I3 (C) = det(C) = I3 (b), or directly on the principal stretches λ1 , λ2 , λ3 , which
1 2 2
are the square roots of the eigenvalues of C.
Every model for rubber-like materials can be classified as phenomenological or micro-mechani-
cal. The latter are derived from statistical mechanics arguments on networks of idealised chain mol-
ecules, whereas the former utilise more or less complex, frequently polynomial formulations in terms
of strain invariants or principal stretches. Although intrinsically tied to higher computational costs (and
homogenisation requirements), micro-mechanical approaches recently got more and more attention due
to the fact that their governing parameters—at least somehow—relate macroscopic mechanical behav-
iour to the causative physical/chemical structure. This property is a significant advantage compared to
phenomenological models, especially if, for example, gradients in the structural composition as observed
at phase boundaries (interphases) or temporal changes in the network structure (curing) do occur and
have to be described. Prominent examples for micro-mechanical models are the 3-chain, 4-chain, 8-chain
models as well as the unit sphere (21-chain) model which have all been proven to be appropriate for
moderate to large elastic deformations of rubber-like materials, cf. Arruda and Boyce [2,3] and Miehe
et al. [4–6].
Among the numerous phenomenological approaches, the class of the celebrated Ogden models [7] rep-
resents a very flexible ansatz in modelling rubber-like materials, in particular due to its modular polynomial
formulation in terms of principal stretches—which additionally makes it amenable to mathematical analy-
sis. Also well-known, widely used and of earlier origin are models of Mooney–Rivlin type, cf. e.g. [8] or
[9–11], which are formulated in strain invariants. Its simplest realisation, called neo-Hookean model, con-
stitutes a close link to micro-mechanical approaches since it coincides with a 3-chain model using Gauss-
ian chain statistics. The main challenge in designing invariant-based models is to choose an appropriate
(sub)set of invariants and to include sufficiently high orders of them into the strain energy function. As
pointed out, for example, by Yeoh [12], already cubic terms of I1 are able to reproduce the highly nonlinear,
S-shaped uniaxial behaviour of rubber, also at very large strains. Nonetheless, using only reduced sets of
invariants will always significantly impair the model’s behaviour at more complex, multiaxial deformation
modes.
Also hybrid models merging phenomenological and micro-mechanical approaches to benefit from the
advantages of both worlds have been proposed by different authors. Examples are Yeoh and Fleming [13] or
Wu and van der Giessen [14] who combine 3-chain and 8-chain models by a phenomenological adjustment
parameter. A different hybrid approach has been presented by Beda and Chevalier [15] who linked the Gent
and Thomas model [16] with that of Ogden [7] to have appropriate ansatzes for both the small and large strain
range. Additionally, a method to determine the material parameters has been proposed.
To assess the quality and accuracy that a particular model can achieve in reproducing real material behav-
iour, it is common practice to simulate homogeneous deformations and compare the results with experimental
data. Since experiments are time-consuming and cost intensive, it is desirable to minimise their number and
the complexity that is necessary to correctly determine the governing material parameters. Thus, the compet-
ing requirements read as follows: an ideal material model should (1) contain as few parameters as possible,
(2) be able to correctly reproduce arbitrarily complex deformation behaviour and (3) allow the determination
of its parameters for a specific material from fitting few simple experiments. Unfortunately, these properties are
often reconcilable in real-world applications and necessitate many compromises, as is documented by the vast
number of publications dealing with different aspects of this matter. Any bibliography on this field essentially
has to remain incomplete.
Concerning comparative studies, Marckmann and Verron [16], for example, analysed twenty different mod-
els and ranked them with respect to the classical Treloar data [1] indicating that the extended tube model [17],
the Shariff model [18], the unit sphere model [5] and the Ogden model [7] provide the best results. The same
approach was followed by Boyce and Arruda [3] who also compared several models using Treloar’s experi-
mental data for uniaxial, biaxial and pure shear deformations. Seibert and Schöche [20] compared six different
models with respect to uniaxial and biaxial experimental data of a carbon black-filled rubber and conclude
that higher order terms are essential to reproduce the typical increase in stiffness at large strains, even if such
terms are susceptible to instability problems. Furthermore, it is stated that the Arruda–Boyce model provides
reliable predictions of biaxial deformations although calibrated only with uniaxial data.
Hyperelastic models for rubber-like materials 1185
Parameter estimation for Rivlin and Saunders type hyperelasticity [21] has been examined in a series of
papers by Hartmann et al., who discussed aspects like optimisation constraints [22], polyconvexity [23] and
the identification from inhomogeneous experiments [24,25] on carbon black-filled rubber [26].
A higher order elasticity model, composed of a neo-Hookean-like compressible and a generalised
Mooney-type incompressible component, has been proposed by Attard and Hunt [27] and was proven
to perform excellent on experimental data of uniaxial and (equi)biaxial tension and pure shear from
seven different authors. In particular, it is pointed out that this formulation is capable to reproduce the
Valanis–Landel hypothesis [28] (experimentally approved by Vangerko and Treloar [29]), which states that
interactions of principal stresses are governed only by compressive deformation parts whereas the incom-
pressible strain energy density is represented by a superposition of terms dependent on the principal stretches.
Chagnon et al. [30] compared the models of Hart-Smith [31], Arruda–Boyce [2] and Gent [32] stating
similar responses and, furthermore, establishing equations to relate the material parameters. The similar-
ity in behaviour of Arruda–Boyce and Gent model was also demonstrated by Boyce [33]. Another con-
ceptual link between micro-mechanical and phenomenological models has been discussed by Horgan and
Saccomandi [34,35] who showed that the two-parameter Gent model is a very good approximation of
a non-Gaussian full-network model involving the inverse Langevin function. Furthermore, its phenome-
nological parameters can be formulated in terms of microscopic properties. A conceptually different and
interesting approach to compare constitutive models for elastomers was proposed by Currie [36]. It uses
the so-called attainable region in the invariant space of the free energy, which is a three-dimensional sur-
face computable from different homogeneous deformation experiments by methods introduced in Haines
and Wilson [37], to numerically compare different free energies. The advantage is that not only partic-
ular formulations or single deformation modes can be compared but the whole image spaces of specific
models.
In order to employ arbitrary constitutive models within a nonlinear finite element framework, it is essential
not only to derive the corresponding stress–strain relation but also to determine a consistent tangent operator
providing quadratic convergence while solving the governing nonlinear equations by iterative schemes like
Newton–Raphson. The theoretical background to derive stress tensors and tangent operators from arbitrary
strain energy functions for finite elasticity was, for example, given by Miehe [39], where main emphasis is
put on different mixed finite element formulations necessary to capture incompressibility. Frameworks for
the derivation of stress tensors and tangent operators for models based on invariants have been presented by
Liu et al. [40] and Kaliske and Rothert [41], where the former contains details for both reference and current
configurations, while the latter is restricted to the current configuration.
In most publications on rubber elasticity, authors propose new constitutive models and analyse their
performance by comparison with different, usually homogeneous experimental data like those from Treloar [1].
Unfortunately, only a minority of works provides both stress tensors and tangent operators for particular mod-
els, like, for example, Miehe et al. [5,6] did for the unit sphere model or Heinrich and Kaliske [17,42] who
introduced the extended tube model and provided details on the finite element implementation as well as a
model verification by (non)homogeneous numerical examples. To close this gap, the present contribution elab-
orately derives the tangent operators for a selection of both phenomenological and micro-mechanical models,
which is complemented by a comparative study of the model performances with respect to the classical Treloar
data.
The outline of the paper is as follows: in Sect. 2, we briefly review the general framework for stress
tensors and tangent operators for finite isotropic elasticity as developed in Holzapfel [38] or Miehe [39] and
summarise the details for the special case of invariant-based models as given by Liu et al. [40]. After these
preliminaries, Sects. 3 and 4 present a selection of fourteen constitutive models for elastic rubber-like materials
and derive the corresponding equations for stress tensor and tangent operator. The models considered cover
examples from both phenomenological (eleven) and micro-mechanical approaches (three) and are listed in
Fig. 1. Furthermore, Sects. 3 and 4 each contain details on the derivation of analytical stress–strain relations
from arbitrary free energy functions for the homogeneous cases of uniaxial tension, equibiaxial tension and
pure shear. These analytical formulations are then used to evaluate the performance of each model in repro-
ducing the experimental data provided by Treloar, cf. “Appendix A”. To this end, a standard fitting tool is
applied to calculate optimal model parameters wrt each set of Treloar’s data. Aiming at the above mentioned
model requirement (3), that is, the ability of a model to be correctly adjusted for all deformation modes by
as few as possible experiments, we further compare the errors that arise if each model together with the
optimal parameters for a particular deformation mode is used to simulate the remaining two deformation
modes.
Fig. 1 The here considered hyperelastic models for rubber-like materials and some corresponding references
2 Continuum mechanical preliminaries
The subsequent considerations are restricted to the case of isothermal isotropic hyperelastic material behaviour
in the finite strain regime. We review the steps to derive stress tensors and tangent operators from a given strain
energy density, for both formulations in terms of tensor-valued strain measures (cf. [39]) and for invariant-based
functions (cf. [40]).
2.1 Stress tensor and tangent operator derivation
The isothermal response of a hyperelastic material can be described by a strain energy function = (C),
which we assume to be dependent on the right Cauchy-Green tensor, that is, all subsequent considerations
take place in the material configuration. For a derivation of stresses and tangent operators in the spatial con-
figuration, we refer to, for example, [6,38] or [41] for the case of invariant-based free energy functions. The
second Piola–Kirchhoff stress tensor S is then given as twice the derivative of the free energy wrt C, and by
introducing the fourth-order elasticity tensor (or tangent operator) C, one can describe changes in stress S
that are caused by changes in deformation C via S = C : 21 C, that is,
∂ ∂S ∂ 2
S=2 , C=2 = 4 2. (1)
∂C ∂C ∂C
Since rubber-like materials usually exhibit a decoupled response to volumetric and deviatoric deformations, an
additive decomposition of the strain energy function into volumetric (shape preserving) and isochoric (volume
preserving but shape changing) parts can be motivated:

(C) = vol (J ) + iso C̄ , (2)
where J = det F and C̄ = J −2/3 C denote the isochoric right Cauchy-Green tensor. Decomposition (2) coin-
cides with the multiplicative split F = [J 1/3 I ]F̄ of the deformation gradient that goes back to Flory [50] and
satisfies the incompressibility condition det F̄ = 1. A combination of (1) and (2) yields the correspondingly
decoupled stress tensor and tangent operator:
∂vol ∂iso ∂S vol ∂S iso
S = S vol + S iso = 2 +2 , C = Cvol + Ciso = 2 +2 . (3)
∂C ∂C ∂C ∂C
From the chain rule and some algebraic manipulations, the volumetric stress tensor can be reformulated as
∂vol (J ) ∂ J
S vol = 2 = J p C −1 , (4)
∂J ∂C
where the hydrostatic pressure p = ∂vol (J )/∂ J has been introduced. In the case of incompressibility, J = 1
and p serves as a Lagrange multiplier to satisfy this kinematic constraint on the deformation. The hydrostatic
pressure can then only be calculated from the equilibrium equations together with boundary conditions. For
compressible materials, one usually prescribes some volumetric strain energy function in terms of a bulk mod-
ulus k0 and a parameter ξ like, for example, vol (J ) = k0 ξ −2 [J −ξ + ξ ln(J ) − 1] as given in [7]. Note that
this energy may not be suitable for all situations since convexity can be violated depending on the choice for
ξ , compare the discussion in [23].
To directly utilise formulations of iso in terms of the isochoric right Cauchy-Green tensor C̄, the fictitious
stress tensor S̄ = 2∂iso (C̄)/∂ C̄ is introduced and the isochoric deviatoric stress S iso is then obtained from a
fourth-order projection tensor P by
2 1
S iso = J − 3 P : S̄ with P = I − C −1 ⊗ C, Ii jkl = δik δ jl . (5)
3
From further evaluation of (3)2 , cf. e.g. [38] for details, the two parts of the tangent operator are obtained as

2dp
Cvol = J p + J C −1 ⊗ C −1 − 2J p C −1 C −1 (6)
dJ
2 2 2 −1
Ciso = P : C̄ : PT + J − 3 S̄ : C P̄ − C ⊗ S iso + S iso ⊗ C −1 (7)
3 3
4 ∂ S̄ 4 ∂ iso (C̄)
2
C̄ = 2J − 3 = 4J − 3 (8)
∂ C̄ ∂ C̄
2
1
P̄ = C −1 C −1 − C −1 ⊗ C −1 (9)
3
1
[A B]i jkl = Aik B jl + Ail B jk (10)
2
after introduction of fictitious tangent operator C̄, modified projection tensor P̄ and tensor product . From
Eqs. (4)–(7), stress tensors and tangent operators can be calculated for arbitrary strain energy functions which
will later be applied to the different micro-mechanical chain models in Sect. 4.
2.2 Stress tensor and tangent operator derivation for invariant-based models
The general derivation of stress tensors and tangent operators is now particularised for the case of models
formulated in terms of strain invariants, that is, we consider = (I1 , I2 , I3 ) with Ii = Ii (C), i = 1, 2, 3
as given in Sect. 1. For the evaluation of the performance that the models attain in reproducing Treloar’s data,
it will later be of particular interest that the invariants of C are related to the principal stretches λi by
I1 = λ21 + λ22 + λ23 , I2 = λ21 λ22 + λ22 λ23 + λ21 λ23 , I3 = λ21 λ22 λ23 . (11)
Since C̄ = J −2/3 C, the invariants of C̄ and C are related by

I¯1 = J −2/3 I1 , I¯2 = J −4/3 I2 , I¯3 = det J −2/3 C = J −2 det C = J −2 I3 = 1, (12)
that is, they coincide for the case of incompressibility J = 1 and the third invariant then does not at all
contribute to the isochoric strain energy. To be able to just reuse the equations for stress tensor and tan-
gent operator summarised in the previous subsection, volumetric-isochoric strain energy decomposition (2) is
again assumed, and it is sufficient to determine fictitious stress tensor S̄ = 2∂iso /∂ C̄ and tangent operator
C̄ = 4J −4/3 ∂ 2 iso /∂ C̄∂ C̄, now from an isochoric strain energy function formulated in terms of the invariants
of the isochoric right Cauchy-Green tensor C̄, that is,

iso = iso I¯1 C̄ , I¯2 C̄ . (13)
Since ∂ tr(A)/∂A = I and ∂ tr(A2 )/∂A = 2At hold for any second order tensor, the chain rule yields for the
fictitious stress tensor:

∂iso ∂ I¯1 ∂iso ∂ I¯2 ∂iso ∂iso ¯
S̄ = 2 + =2 I +2 I1 I − C̄ , i.e.
¯
∂ I1 ∂ C̄ ¯
∂ I2 ∂ C̄ ¯
∂ I1 ∂I¯
2
∂iso ∂ ∂iso
+ I¯1
iso
S̄ = γ̄1 I + γ̄2 C̄ with γ̄1 = 2 and γ̄2 = −2 . (14)
¯
∂ I1 ¯
∂ I2 ∂ I¯2
The corresponding fictitious tangent operator can be derived from this according to Eq. (8), and after some
lengthy calculations omitted here, one obtains:
4 ∂ S̄ 4
C̄ = 2J − 3 = J − 3 δ̄1 I ⊗ I + δ̄2 I ⊗ C̄ + C̄ ⊗ I + δ̄3 C̄ ⊗ C̄ + δ̄4 I , where (15)
∂ C̄
2
∂ iso ¯ ∂ 2 iso ∂iso ¯2 ∂ iso
2
δ̄1 = 4 + 2 I1 + + I1 ,
∂ I¯1 ∂ I¯1 ∂ I¯1 ∂ I¯2 ∂ I¯2 ∂ I¯2 ∂ I¯2
2
∂ iso ∂ 2 iso ∂ 2 iso ∂iso
δ̄2 = −4 + I¯1 , δ̄3 = 4 , δ̄4 = −4 .
∂ I¯1 ∂ I¯2 ∂ I¯2 ∂ I¯2 ∂ I¯2 ∂ I¯2 ∂ I¯2
3 Phenomenological models: stress tensors, tangent operators and performance
Within continuum mechanics, the majority of phenomenological treatments of the elastic behaviour of rub-
ber-like materials depart from the assumption of isotropic hyperelasticity, that is, a strain energy function
usually depending on the deformation via principal stretches λi or strain invariants Ii = Ii (C) is considered.
Accordingly, phenomenological models are classified into principal stretch based and invariant-based models.
This section reviews eleven selected models of this type (cf. Fig. 1), provides the corresponding strain
energy functions as known from the literature and derives stress tensors and tangent operators according to
the equations summarised above. To evaluate the performance of the models, we resort to the classical exper-
imental Treloar data [1] for the three homogeneous deformation modes uniaxial and equibiaxial tension and
pure shear. First, some general details are stated on how to derive analytical stress–stretch relations for these
deformation modes for a particular constitutive model. With this at hand, optimal material parameters for each
model and deformation mode are determined (calibration) by a standard fitting procedure. The validity of
the models is then quantified by calculating the errors that arise if the two deformation modes not used for
parameter identification are simulated using the optimal parameters of the third mode.
3.1 Analytical stress formulations for uniaxial tension, equibiaxial tension and pure shear
The Treloar data used here to evaluate the performance of the polymer models are given in pairs of principal
stretches λi and principal nominal stresses Pi for the three cases of uniaxial and equibiaxial tension and pure
shear, cf. “Appendix A”. To determine optimal material parameters for each model and deformation mode by
a standard fitting procedure, it is necessary to derive analytical Pi (λ)-formulations from the particular free
energy functions. The material investigated by Treloar can be characterised as isotropic and incompressible, a
case for which principal stretches and nominal stresses are related by
∂ 1
Pi = − p, i = 1, 2, 3. (16)
∂λi λi
The details on how to derive this equation can be found, for example, in [38] or [51]. In the case of invariant-
based models, that is, = (I1 , I2 ), the chain rule provides
∂ ∂ I1 ∂ ∂ I2 1
Pi = + − p, i = 1, 2, 3, (17)
∂ I1 ∂λi ∂ I2 ∂λi λi
which subsequently serves to derive the desired analytical formulations for the three deformation modes.
3.1.1 Uniaxial tension (UT)
In uniaxial tension, the specimen is elongated in only one direction, for example, λ1 = λ. From incompress-
ibility, that is, I3 = λ21 λ22 λ23 = 1, and the assumption of isotropy, the complementary principal stretches follow
as λ2 = λ3 = λ−1/2 . Thus, the corresponding deformation gradient and the invariants, cf. Eq. (11), are
⎡ ⎤
λ 0 0
F UT = ⎣ 0 λ−1/2 0 ⎦ , I1UT = 2λ−1 + λ2 , I2UT = λ−2 + 2λ. (18)
0 0 λ−1/2
UT are zero and only P UT (λ) has to
Since the contraction is unhindered in the transversal directions, both P2,3 1
be determined. By setting Eq. (17) to zero, for example, for i = 2, and calculating the derivatives ∂ I1,2 /∂λ2
from Eq. (11), the resultant pressure is determined as

2 ∂ 1 ∂
p =
UT
+2 λ+ 2 . (19)
λ ∂ I1 λ ∂ I2
Inserting this into Eq. (17) for i = 1, we obtain the analytical formulation for the first principal stress:

∂ 1 ∂ 1
P1UT = 2 + λ− 2 . (20)
∂ I1 λ ∂ I2 λ
3.1.2 Equibiaxial tension (ET)
The specimen is equally stretched in two orthogonal directions, that is, λ1 = λ2 = λ. Again due to incom-
pressibility, the remaining principal stretch reads λ3 = λ−2 , and the corresponding deformation gradient and
invariants follow as
⎡ ⎤
λ0 0
F ET = ⎣ 0 λ 0 ⎦ , I1ET = λ−4 + 2λ2 , I2ET = 2λ−2 + λ4 . (21)
0 0 λ−2
The stresses in load directions are equal, while the third direction is stress-free due to unhindered contraction,
that is, P1ET = P2ET and P3ET = 0, respectively. The pressure is determined by setting Eq. (17) to zero for
i = 3:
2 ∂ 4 ∂
p ET = 4 + 2 , (22)
λ ∂ I1 λ ∂ I2
and by reinserting this into (17) for i = 1, we obtain the first and second principal stresses:

∂ 2 ∂ 1
P1 = P2 = 2
ET ET
+λ λ− 5 . (23)
∂ I1 ∂ I2 λ
3.1.3 Pure shear (PS)
The pure shear set-up of Treloar [52] utilises rectangular sheets having a much larger width than length to rea-
lise a zero deformation perpendicular to the loading direction λ1 = λ, that is, λ2 = 1 holds almost everywhere
except for the vicinity of the free edges. From incompressibility, the third principal stretch λ3 = λ−1 and the
corresponding deformation gradient and the invariants read
⎡ ⎤
λ0 0
F PS = ⎣ 0 1 0 ⎦ , I PS = I1PS = I2PS = λ2 + λ−2 + 1. (24)
0 0 λ−1
From unhindered contraction in the third direction, the pressure follows from setting (17) to zero for i = 3:

2 ∂ 1 ∂
p = 2
PS
+2 1+ 2 . (25)
λ ∂ I1 λ ∂ I2
Insertion of (25) into (17) yields the principal stresses in load direction and perpendicular to this:

∂ ∂ 1 ∂ ∂ 1
P1PS = 2 + λ − 3 , P2PS = 2 + λ2 1− 2 . (26)
∂ I1 ∂ I2 λ ∂ I1 ∂ I2 λ
3.2 Models formulated in principal stretches: Ogden models (1972)
A sophisticated class of models frequently used for the simulation of rubber-like materials goes back to
Ogden [7]. Although markedly modular and flexible, it is known that the determination of its material parame-
ters may become difficult if more complex versions are utilised [16]. The free energy density is formulated as
a finite sum of scaled powers of the principal stretches λi , which are the square roots of the eigenvalues of the
right Cauchy-Green tensor C, that is, λi2 Ni = C · Ni , i = 1, 2, 3. The eigenvectors Ni are also referred to as
principal referential directions. With regard to additive decomposition (2), we consider the following isochoric
free energy density

K
μk
iso = λ̄α1 k + λ̄α2 k + λ̄α3 k − 3 (27)
αk
k=1
in terms of the modified principal stretches λ̄i = J −1/3 λi , the square roots of the eigenvalues of C̄. Model (27)
is adjustable by 2K material parameters, the shear modulus like μk and the dimensionless exponents αk , for
which a consistency condition can be derived from comparison with the linear theory: k μk αk = 2μ, wherein
μ is the classical shear modulus, and μk αk > 0 have to be satisfied.
Models formulated in principal stretches intrinsically require to solve the eigenvalue problem of the con-
sidered strain tensor.
From the eigenvalues, the corresponding eigenbasis {N1 , N2 , N3 } for the spectral repre-
sentation C̄ = λ̄i2 Ni ⊗ Ni can be determined, for example, by Serrin’s formula. Since eigenbases of stress
and strain coincide for isotropic materials, the format of the stress tensor follows straightforward as

3
S iso = i
Siso Ni ⊗ Ni (28)
i=1
and only necessitates the computation of the eigenvalues Siso i . Due to the fact that
iso does not directly
depend on C̄, it provides no further advantage to use fictitious stress S̄ and the corresponding projection
operators, cf. Eqs. (5)–(9), to compute S iso and Ciso but to start immediately from Eq. (3). Recalling that
∂λi2 /∂C = Ni ⊗ Ni in the case of three distinct eigenvalues (cf. [53] or [54] on how to handle identical
eigenvalues), the chain rule applied to equation (3)1 yields

∂iso λ̄i 3
∂iso ∂λi2 3
1 ∂iso
S iso = 2 =2 = Ni ⊗ Ni , (29)
∂C ∂λi ∂C
2 λi ∂λi
i=1 i=1
i = λ−1 ∂ /∂λ from comparing coefficients with (28). This can be resolved further by exploiting
that is, Siso i iso i
∂ λ̄i /∂λ j = J −1/3 [δi j − 13 λ̄i λ̄−1
j ], and after some algebraic manipulations and insertion of ∂(27)/∂ λ̄i , the
eigenvalues completing spectral representation (28) of the isochoric stress tensor follow as
⎡ ⎤
αk
K 3 K
μ λ̄
= J −2/3 ⎣ μk λ̄iαk −2 −
k j ⎦
i
Siso , i = 1, 2, 3. (30)
3 λ̄i2
k=1 j=1 k=1
Concerning a detailed derivation of the corresponding spectral representation of the isochoric tangent operator
Ciso , the interested reader is referred to [38,53,55] or [56]:
3
1 ∂ Siso
i 3 j i
Siso − Siso
Ciso = Ni ⊗Ni ⊗N j ⊗N j + Ni ⊗N j ⊗Ni ⊗N j + Ni ⊗N j ⊗N j ⊗Ni . (31)
λ j ∂λ j λ − λi
2 2
i, j=1 i, j=1 j
i = j
Here, only the derivatives ∂ Sisoi /∂λ within the first sum term remain to be computed from Eq. (30).
j
To demonstrate consistency of tangent operator (31), we employ the local convergence test explained in
“Appendix B”, which iteratively determines the strain tensor corresponding to a prescribed stress tensor and
material parameters. In the case of the values published in [38] for Treloar’s uniaxial elastomer data, that is,
[μ1 , μ2 , μ3 ] = [0.63, 0.0012, −0.01] MPa, [α1 , α2 , α3 ] = [1.3, 5.0, −2.0], the test converges quadratically
to C ∞ = [1.51, 2.62, 1.28, 1.59, 0.0, 0.0], as the evolutions of the Euclidean norms of residual and strain
update indicate:
Iteration n 0 1 2 3 4
R(C n )2 2.01e−01 5.24e−01 4.15e−03 1.08e−05 3.46e−12
C n+1 − C n 2 2.03e−01 3.56e−02 1.809e−03 2.1e−06 1.06e−12
This table may be considered a desirable behaviour, the Matlab routine from “Appendix B” should converge
similarly for any realisation of a tangent operator given in this paper. Correctness of possible stress imple-
mentations can be verified by comparing the converged strain tensors C ∞ with those that are tabulated in
“Appendix B” and have been computed using the optimal UT-parameters as given the subsection of each
model.
Analytical Pi (λi )-relations for the deformation modes UT, ET and PS, which are required to check the
performance of the model on Treloar’s data, can be derived from (16) together with free energy density (27).
In analogy to Sects. 3.1.1–3.1.3, the pressure p has to be determined beforehand from the respective boundary
conditions, that is, from evaluating Eq. (16) for Pi = 0. The resulting stress–stretch relations read
UT
αk −1 − 21 αkUT −1
P1UT = μUT
k λ − λ , (32)
ET
λαk −1 − λ−2αk −1 ,
ET
ET
P1,2 = μET
k (33)

λαk −1 − λ−αk −1 ,
PS PS
P1PS = μPS
k (34)
whereas λ denotes the isochoric stretch in the corresponding load direction as given in the previous subsec-
tions. Note that only the coefficients to αk differ between the deformation modes. By using Matlab’s Curve
Fitting Toolbox, Eqs. (32)–(34) can be fitted to the corresponding Treloar data (cf. “Appendix A”) to obtain
optimal parameters for three increasingly complex Ogden models with K = 1, 2, 3. To exploit the whole
experimental information at once, a fourth optimisation is carried out that computes material parameters under
simultaneous consideration of all UT, ET and PS deformation data. Thereby, the data have not been weighted,
that is, neither differing numbers of measurands nor different stretch ranges have been particularly accounted
for. The resulting parameters are:
K =1 K =2 K =3
μUT
1 = 0.0167 μUT = [0.3055,2.316e−6] μUT = [0.5649,3.856e−3,5.7e−13]
α1UT= 3.8540 α UT = [1.996,8.022] α UT = [1.297,4.342,15.13]
μET
1 = 0.2958 μET = [0.4856,1.965e−3] μET = [0.4848,1.918e−3,2.8e−14]
α1ET= 2.3660 α ET = [1.659,5.268] α ET = [1.662,5.281,2.1e−9]
μPS
1 = 0.3105 μPS = [0.4726,1.256e−3] μPS = [0.557,1.947e−2,1.3e−11]
α1PS= 2.0620 α PS = [1.57,4.869] α PS = [1.231,3.413,15]
μ1 = 0.0980 μ = [0.3528,1.032e−9] μ = [0.0662,5.875e−12,0.6249]
α1 = 2.9473 α = [2.05,11.771] α = [2.875,14.221,1]
Comparing UT-parameters for K = 3 with those from [38] cited above reveals significant qualitative and
quantitative differences—even though experimental data are reproduced almost perfectly by both parameter
sets, see Fig. 4 (left). These discrepancies are explicable by the high flexibility of large-K Ogden models
whose increasing number of parameters leads to a severe sensitivity of the optimisation method wrt the initial
values. In particular, one can observe that the fitting procedure optimises certain parameters close to zero, that
is, it reduces the complexity of the model by switching off certain terms. This behaviour reflects a general
problem reported also by other authors: reliable parameter identification becomes increasingly difficult, the
more flexible the considered models are. Consequently, an ideal model should not exceed the complexity of
the material behaviour it has to capture, which is important especially in designing phenomenological models
since these usually abdicate a direct relation to the causative material structure.
To validate the models, each set of optimal material parameters for UT, ET and PS is used to simulate
the other two deformation modes. The results are plotted in Figs. 2, 3, 4 which additionally contain the errors
between (a) each experiment and its optimal fit, for example, Error(UT-fit), and (b) the simulations of the other
deformation modes and their respective measurements, for example, Error(ET-sim). All errors are calculated
according to
1
M
2
Error2 (fit/sim) = Pfit/sim (λiTreloar ) − PTreloar (λiTreloar ) , (35)
M
i=1
UT fitted, ET/PS simulated ET fitted, UT/PS simulated PS fitted, UT/ET simulated
6
0.327 = Error(UT−fit) 0.076 = Error(ET−fit) 0.042 = Error(PS−fit)
0.708 = Error(ET−sim) 0.620 = Error(UT−sim) 1.335 = Error(UT−sim)
0.363 = Error(PS−sim) 0.449 = Error(PS−sim) 0.363 = Error(ET−sim)
5
4
P [MPa]
2
UT sim
ET sim
1 PS sim
UT data
ET data
0
1 2 3 4 5 6 7 8 1 2 3 4 5 6 7 8 1 2 3 4 5 6 7 PS data
8
λ[] λ[] λ[]
Fig. 2 Performance of the K = 1—Ogden model on Treloar’s data. Fitting quality and generalisation (capability to correctly simu-
late deformations not used for optimisation) are rather poor since the model cannot realise changes in curvature. All stress–stretch
relations yield almost identical curves for a certain parameter set. Error(UT/ET/PS-sim) = 0.434/0.362/0.218
6
5
4
P [MPa]
2
UT sim
ET sim
1 PS sim
UT data
ET data
0
1 2 3 4 5 6 7 8 1 2 3 4 5 6 7 8 1 2 3 4 5 6 7 PS data
8
λ[] λ[] λ[]
Fig. 3 Performance of the K = 2—Ogden model on Treloar’s data. The fitting quality is acceptable (UT) to perfect (ET, PS), but
the generalisation still is not sufficient since the small differences between the three stress–stretch relations do not impose curves
significantly differing for a particular parameter set. Error(UT/ET/PS-sim) = 0.132/0.269/0.123
6
0.289 = Error(UT−sim [38]−params.) 3.714 = Error(UT−sim) 8.841 = Error(UT−sim)
0.409 = Error(ET−sim) 0.579 = Error(PS−sim) 0.369 = Error(ET−sim)
5 0.047 = Error(PS−sim)
4
P [MPa]
2
UT sim
ET sim
1 PS sim
UT data
ET data
0
1 2 3 4 5 6 7 8 1 2 3 4 5 6 7 8 1 2 3 4 5 6 7 PS data
8
λ[] λ[] λ[]
Fig. 4 Performance of the K = 3—Ogden model on Treloar’s data. The fitting quality is close to perfection, the generalisation is
slightly better than for K = 2 but still far from satisfactorily. For comparison, the left figure also contains a UT-simulation (green
line) using the parameters published in [38]. Error(UT/ET/PS-sim) = 0.116/0.268/0.112. (Color figure online)
the sum of the squared differences between fitted or simulated and measured stresses, averaged by the number
of data points M (stretches) available for each deformation mode. The deviations between measurements and
simulations with parameters from the combined optimisation are given in the caption of each figure.
3.3 Models formulated in strain invariants
For the most general case, one can assume a strain energy function depending on the invariants of the right
Cauchy-Green tensor, that is, = (I1 (C), I2 (C), I3 (C)), which is continuously differentiable with respect
to all Ii . Then, this strain energy function can be expanded in an infinite Taylor series of the form
∞

= (I1 , I2 , I3 ) = cklm [I1 − 3]k [I2 − 3]l [I3 − 1]m , (36)
k,l,m=0
where cklm denote material parameters. The above form is termed as the general form of the Mooney–Rivlin
strain energy function [57,58]. In the case of incompressible material behaviour, det(F ) = λ1 λ2 λ3 = J = 1
and, thus, I3 = det(C) = det(J 2/3 C̄) = det(C̄) = I¯3 = 1, and the invariants Ii of C coincide with the
invariants I¯i of the isochoric right Cauchy-Green tensor C̄. Consequently, Eq. (36) can be reduced to the
isochoric strain energy function:
∞

k l
iso = iso (I1 (C), I2 (C)) = iso I¯1 C̄ , I¯2 C̄ = ckl I¯1 − 3 I¯2 − 3 . (37)
k,l=0
This provides the foundation for all invariant-based models a selection of which is subsequently reviewed.
3.3.1 Neo-Hooke model (1943)
The Neo-Hooke model constitutes the simplest specification of the Mooney–Rivlin model series since it only
considers c10 = 0 in Eq. (37), that is, the summation stops at k = 1, l = 0 and the additive constant c00 is set
to zero. The resulting isochoric free energy function reads
μ ¯
iso = I1 − 3 , (38)
2
where the only material parameter μ = 2c10 denotes the shear modulus. This model coincides with the 3-chain
model, cf. Sect. 4.1, which is a micro-mechanically motivated ansatz that uses Gaussian chains oriented along
the principal stretch directions. Evaluation of Eqs. (14) and (15) provides the corresponding fictitious stress
tensor and tangent operator:
S̄ = μI (39)
C̄ = 0 ⊗ 0. (40)
Note that only the fictitious tangent operator is zero here, the complete tangent according to Eqs. (6), (7) is
nonzero. The analytical Pi (λi )-relations for the deformation modes UT, ET and PS are obtained from Eqs. (20),
(23), (26) as

P1UT = μUT λ − λ−2 (41)

ET
P1,2 = μET λ − λ−5 (42)

P1PS = μPS λ − λ−3 . (43)
Fitting to Treloar’s data (cf. “Appendix A”) provides the following optimal material parameters:
μUT = 0.5673 MPa μET = 0.4104 MPa μPS = 0.3360 MPa μ = 0.5250 MPa
Each of the first three parameters is now used to simulate the experimental data of the other two deformation
modes. The results are plotted in Fig. 5, which additionally gives the errors that occur from the fitting proce-
dure and the simulation of the experiments that have not been used for parameter identification. Obviously,
the Neo-Hookean ansatz is not sufficient to correctly reproduce the experimental data. Especially, the charac-
teristic S-shape cannot be captured since the simple model structure does not allow for a change of curvature.
Merely fitting the PS-data yields an acceptable result. Nonetheless, if only small deformations with λ < 1.5
are considered, the Neo-Hooke model provides reasonable results, cf. also [59].
6
5
4
P [MPa]
2
UT sim
ET sim
1 PS sim
UT data
ET data
0
1 2 3 4 5 6 7 8 1 2 3 4 5 6 7 8 1 2 3 4 5 6 7 PS data
8
λ[] λ[] λ[]
Fig. 5 Performance of the Neo-Hooke model on the Treloar data. The minimum least squares fits to UT (left) and ET (middle)
yield unacceptable results due to the incapability of the model to reproduce the S-shape. The latter is less dominant for PS (right),
which leads to an acceptable fit quality in this case. Using the optimal parameter of a certain deformation mode to simulate
different modes is not advisable, as is indicated by the error values. Error(UT/ET/PS-sim) = 0.815/0.188/0.528
3.3.2 Mooney–Rivlin model (1940)
A more sophisticated, yet simple specification of Eq. (37) is called Mooney–Rivlin or Mooney model. It con-
siders c10 = 0 = c01 , that is, the summation stops at k, l = 1 and c00 , c11 are set to zero. This ansatz can also
be classified as a model of Ogden type, cf. e.g. [38] for the corresponding formulation in terms of principal
stretches. Its main advantage compared to the neo-Hooke model is the consideration of the second invariant:

iso = c10 I¯1 − 3 + c01 I¯2 − 3 . (44)
Here, 2c10 = : μ1 and 2c01 = : −μ2 denote material parameters of shear modulus type which are related to
the standard shear modulus via μ = μ1 − μ2 . Fictitious stress tensor, tangent operator and analytical Pi (λi )-
relations for UT, ET and PS follow again from evaluation of (14), (15) and (20), (23), (26), respectively:

S̄ = 2 c10 + c01 I¯1 I − 2c01 C̄ (45)
4
C̄ = 4J − 3 [c01 I ⊗ I − c01 I] (46)

P1UT = c10
UT
2λ − 2λ−2 + c01 UT
2 − 2λ−3 (47)
3
ET
P1,2 = c10
ET
2λ − 2λ−5 + c01 ET
2λ − 2λ−3 (48)

P1PS = c10
PS
2λ − 2λ−3 + c01PS
2λ − 2λ−3 . (49)
Fitting the latter three equations separately and at once, respectively, to Treloar’s data provides:
UT = 0.2588 MPa
c10 ET = 0.1713 MPa
c10 PS = 0.2348 MPa
c10 c10 = 0.2659 MPa
UT = −0.0449 MPa
c01 ET = 0.0047 MPa
c01 PS = −0.065 MPa
c01 c01 = −0.0017 MPa
It is worth noting that the identification of PS-parameters is rather unreliable since the two terms in (49) are
identical, that is, only one parameter is remaining. Furthermore, appropriate constraints to the optimiser have
to ensure that the above relation for the overall shear modulus is not violated: μ, as a modulus, should be
positive. To validate the model, each of the first three sets is used to simulate the other two deformation modes
and the results are plotted in Fig. 6. The model performs similar to the Neo-Hookean ansatz, both UT- and
PS-parameters do not provide acceptable generality. While UT is hardly reproduceable by fitting, the situation
is better for the much less curved PS. The remarks already given for the Neo-Hooke model remain valid, a
restriction to smaller stretches improves the accuracy significantly. A true advantage is obtained for ET, the
fitting of which clearly benefits from the extended complexity of the model allowing for changes in curva-
ture. Nonetheless, this flexibility is obviously not independent of the stretch range within which the curvature
is changing its sign. The model is still not complex enough to capture the pronounced S-shape of uniaxial
deformations at very large strains.
6 0.834 = Error(UT−fit) 0.052 = Error(ET−fit) 0.045 = Error(PS−fit)

5
4
P [MPa]
2
UT sim
ET sim
1 PS sim
UT data
ET data
0
1 2 3 4 5 6 7 8 1 2 3 4 5 6 7 8 1 2 3 4 5 6 7 PS data
8
λ[] λ[] λ[]
Fig. 6 Performance of the Mooney–Rivlin model on Treloar’s data. For UT (left) and PS (right), both fitting quality and simulation
errors are similarly insufficient as for the Neo-Hooke model. This is not the case for ET (middle) where the quality of the fit is
quite good and, at least, the simulation of PS seems to be reasonable. Error(UT/ET/PS-sim) = 0.806/0.181/0.537
3.3.3 Isihara model (1951)
The insufficient reproduction of S-shaped load-deformation curves by most of the earlier elastomer models can,
from a molecular point of view, be ascribed to the neglect of the non-Gaussian character in the configurational
entropy of a single chain. To overcome this, Isihara et al. [43] proposed a model incorporating a non-Gaussian
chain theory. The linearisation of the corresponding equations leads to a formulation of the Mooney–Rivlin
series type that contains three material parameters and, most important, a term nonlinearly coupling the two
invariants I¯1 and I¯2 . The free energy function reads, according to [16]:
2
iso = c10 I¯1 − 3 + c20 I¯1 − 3 c01 I¯2 − 3 , (50)
where c10 , c20 , c01 are the governing material parameters. From Eqs. (14) and (15), fictitious stress tensor and
the coefficients for the tangent are determined, the analytical Pi (λi )-relations follow from (20), (23), (26) and
with the here omitted insertion of the corresponding I¯1/2
UT/ET/PS
, cf. (18), (21), (24):
2 2
S̄ = 2c10 + 4c20 c01 I¯1 − 3 I¯2 − 3 + 2c20 c01 I¯1 I¯1 − 3 I − 2c20 c01 I¯1 − 3 C̄ (51)
2
δ̄1 = 8c20 c01 I¯2 − 3 + 16c20 c01 I¯1 I¯1 − 3 + 4c20 c01 I¯1 − 3 (52)
2
δ̄2 = −8c20 c01 I¯1 − 3 , δ̄3 = 0, δ̄4 = −4c20 c01 I¯1 − 3 (53)
2
UT UT ¯UT
P1UT = 2c10UT
+ 4c20 c01 I1 − 3 I¯2UT − 3 + c20 UT UT −1 ¯UT
c01 λ I1 − 3 λ − λ−2 (54)
2
ET ET ¯ET
ET
P1,2 = 2c10
ET
+ 4c20 c01 I1 − 3 I¯2ET − 3 + c20 ET ET 2 ¯ET
c01 λ I1 − 3 λ − λ−5 (55)
2
PS PS ¯PS
P1PS = 2c10PS
+ 6c20 c01 I − 3 λ − λ−3 since I¯1PS = I¯2PS (56)
Note that c20 and c01 could be combined into a single parameter to accelerate parameter identification. Opti-
misation with respect to Treloar’s data provides the following values:
UT = 0.1161 MPa
c10 ET = 0.1993 MPa
c10 PS = 0.1601 MPa
c10 c10 = 0.2617 MPa
UT = 0.0136 MPa
c20 ET = 0.0015 MPa
c20 PS = 0.0037 MPa
c20 c20 = 0.0969 MPa
UT = 0.0114 MPa
c01 ET = 0.0013 MPa
c01 PS = 0.0031 MPa
c01 c01 = 2.47e−6 MPa
Each of the first three sets is used to simulate the two complementary deformation modes, and results are
plotted in Fig. 7. The fit quality is very high for all deformation modes, especially for ET and PS, almost
perfect results are obtained. Isihara’s model can obviously capture the characteristic S-shape at large stretches,
6
5
4
P [MPa]
2
UT sim
ET sim
1 PS sim
UT data
ET data
0
1 2 3 4 5 6 7 8 1 2 3 4 5 6 7 8 1 2 3 4 5 6 7 PS data
8
λ[] λ[] λ[]
Fig. 7 Performance of the Isihara model on Treloar’s data. Fitting errors are low for all deformations, merely UT is not optimal.
The very small deviations for ET and PS can probably be ascribed to the nonsmoothness of the experimental data. Simulation
errors do not allow a final conclusion on the validity of the model. Error(UT/ET/PS-sim) = 0.815/0.191/0.524
whereas the nonlinearity of the model is not yet sufficient to simultaneously reproduce the high initial stiffness,
as reflected by the UT-curve. Furthermore, uniaxial tension is not recommendable to calibrate the model for
other deformation modes which is due to the dominance of I¯1 in the second term (compare the dubiously high
stress increases for ET and PS, Fig. 7, left). The situation is better for the other two cases but from comparison
of Fig. 7, middle vs. left part, the impression arises that this model is particularly sensitive to the curvature and
the stretch range of the data used for parameter identification.
3.3.4 Gent–Thomas model (1958)
Gent and Thomas [44] proposed another empirical two-parameter model similar to that of Mooney and Rivlin
but introducing a scaled logarithm of I¯2 . It can be interpreted as a mathematically convenient approximation of
a more complex model derived by Thomas [60] from a modified network theory. The corresponding isochoric
free energy density is given by
¯
I2
iso = c1 I¯1 − 3 + c2 ln , (57)
3
with material parameters c1 , c2 . Stress tensor, tangent operator and analytical Pi (λi )-relations follow as

I¯1 c2
S̄ = 2 c1 + c2 I − 2 C̄ (58)
I¯2 I¯2

− 43 I¯2 − I¯12 I¯1 1 1
C̄ = 4c2 J I ⊗ I + 2 I ⊗ C̄ + C̄ ⊗ I − 2 C̄ ⊗ C̄ − I (59)
I¯22 I¯2 I¯2 I¯2
2c1UT [λ3 − 1] 2c2UT [λ3 − 1]
P1UT = + (60)
λ2 2λ4 + λ
2c [λ − 1] 2c2 [λ6 − 1]
ET 6 ET
ET
P1,2 = 1 5 + (61)
λ λ7 + 2λ
2cPS [λ4 − 1] 2c2PS [λ4 − 1]
P1PS = 1 3 + 5 . (62)
λ λ + λ3 + λ
from (14), (15) as well as (20), (23), (26) together with inserting of I¯1,2
UT/ET/PS
as in (18), (21), (24). Fitting to
Treloar’s data provides the following optimal material parameters:
c1UT = 0.2837 MPa c1ET = 0.2052 MPa c1PS = 0.1629 MPa c1 = 0.2625 MPa
c2UT = 2.81e−11 MPa c2ET = 2.22e−14 MPa c2PS = 0.0376 MPa c2 = 2.22e−14 MPa
6
5
4
P [MPa]
2
UT sim
ET sim
1 PS sim
UT data
ET data
0
1 2 3 4 5 6 7 8 1 2 3 4 5 6 7 8 1 2 3 4 5 6 7 PS data
8
λ[] λ[] λ[]
Fig. 8 Performance of the Gent–Thomas model on Treloar’s data. Error(UT/ET/PS-sim) = 0.815/0.188/0.528
Figure 8 shows resulting fits, simulations and errors wrt Treloar’s data. Reproduction of experiments as well as
the validity of the parameters is almost identical to those of the Neo-Hookean model. Similarly to this, better
results can be obtained if the strain range used for optimisation does not require a change in curvature, which
the model is unable to capture since higher powers of I¯1 are missing.
3.3.5 Swanson model (1985)
An invariant-based model with a structure similar to that of Ogden’s series expansion has been proposed by
Swanson [45,61]. Its free energy contains two sums of weighted noninteger powers of the isochoric invari-
ants, thus allowing complexity, that is, number of parameters, and nonlinearity of the model to be arbitrarily
adjusted. The major advantage of Swanson’s approach is that neither eigenvalues nor -vectors of the strain
tensor need to be calculated as in the case of Ogden’s model. Nonetheless, due to the fast growing number of
parameters, similar difficulties in the determination of the material constants have to be expected [41,59]. The
isochoric strain energy function is given as
¯ 1+αi ¯ 1+β j
3 Ai 3 Bj
n n
I1 I2
iso = + , (63)
2 1 + αi 3 2 1 + βj 3
i=1 j=1
where the 4n material parameters separate into a group of non-negative shear moduli {Ai , B j } and a group
of exponents {αi , β j } having no real physical interpretation. Studies in [45] concerning a reasonable choice
of n reported that eight parameters pairs, that is, n = 4, are sufficient to obtain a perfect reproduction of
experiments, a result that has been approved on Treloar’s data by Böl [59]. Fictitious stress tensor, tangent
operator coefficients and Pi (λi )-relations follow as
⎡ ⎤
n ¯ αi
n ¯ β j
n ¯ β j
I I I2
S̄ = ⎣ + I¯1 ⎦I −
1 2
Ai Bj Bj C̄ (64)
3 3 3
i=1 j=1 j=1
¯ αi −1 ¯ β j ¯ β j −1
2 2 ¯2
n n n
I1 I2 I2
δ̄1 = Ai αi +2 Bj + I1 Bjβj (65)
3 3 3 3 3
i=1 j=1 j=1
¯ β j −1 ¯ β j −1 ¯ β j
2 2
n n n
I2 I2 I2
δ̄2 = − I¯1 Bjβj , δ̄3 = Bjβj , δ̄4 = −2 Bj (66)
3 3 3 3 3
j=1 j=1 j=1
⎡ αi ⎤
β j

I¯1UT B j I¯2UT
n n
⎣ ⎦ 1
P1UT = Ai + λ− 2 (67)
3 λ 3 λ
i=1 j=1
6
2370 = Error(ET−sim) 1.457 = Error(UT−sim) 1.191 = Error(UT−sim)
5
4
P [MPa]
2
UT sim
ET sim
1 PS sim
UT data
ET data
0
1 2 3 4 5 6 7 8 1 2 3 4 5 6 7 8 1 2 3 4 5 6 7 PS data
8
λ[] λ[] λ[]
Fig. 9 Performance of the n = 1−Swanson model. Fit quality is perfect for all deformations, and ET- and PS-errors are negligible.
None of the parameters can reproduce other modes satisfactorily. Error(UT/ET/PS-sim) = 0.245/0.103/0.111
⎡ α β ⎤
n ¯ET i
I
n ¯ET j
I 1

P1,2 =
ET ⎣ Ai 1
+ λ Bj
2 2 ⎦ λ− 5 (68)
3 3 λ
i=1 j=1
⎡ ⎤
n ¯PS αi n ¯PS β j
⎣ I I ⎦ 1
P1 =
PS
Ai + Bj λ− 3 . (69)
3 3 λ
i=1 j=1
The evaluation of Swanson’s model is restricted to n = 1, 2, mainly because a total number of twelve or
sixteen parameters cannot be considered manageable or efficient. The optimisation procedure yields:
1 = 4.287e−5
AUT 1 = 0.4209
AET 1 = 4.549e−3
APS A1 = 0.0297
B1UT = 0.4159 B1ET = 1.270e−3 B1PS = 0.3702 B1 = 0.4333
α1UT = 3.128 α1ET = −0.0936 α1PS = 1.529 α1 = 1.0771
β1UT = 1.085 β1ET = 0.4447 β1PS = −0.202 β1 = −0.9259
A UT
= [2.83e−3,2.82e−13] A ET
= [0.21,0.0074] A PS
= [0.0676,3.266e−11] A = [0.1922,4.6e−7]
B UT = [1.871e−13,0.4643] B ET = [0.1036,0.266] B PS = [0.2861,0.0267] B = [0.1834,0.0189]
α UT = [1.684,9.141] α ET = [−1833,1.429] α PS = [0.2687,9.131] α = [0.2547,4.6078]
β UT = [−0.4302,0.7882] β ET = [−6.634,−0.623] β PS = [−0.4683,0.7157] β = [−1.2548,
−0.2589]
Figures 9 and 10 depict the corresponding fits and simulations. Swanson’s model precisely captures the
experiments, already in its simplest n = 1—version with only four parameters. Increasing to n = 2 provides
no further improvement of fit quality, but requires much higher optimisation efforts. Still, none of the parameter
sets is really suited to calibrate the model. This picture changes dramatically if the optimiser is considering
all experimental data simultaneously, as can be concluded from the errors given in the captions of Figs. 9 and
10. Especially, the eight-parameter model then perfectly reproduces all deformations. In the case of n = 2,
it is, furthermore, important to note that the optimisation has reached quite a questionable sensitivity wrt the
initial values, small changes may already yield a completely different set of values. This is also explaining
the counterintuitive observation of a larger ET-fit error for n = 2, compared to n = 1. In this regard, the
parameters in the bottom part of the above table have to be considered with caution.
3.3.6 Yeoh model (1990)
Motivated by the experimental observation that the load-deformation curves of filled elastomers exhibit almost
zero values ∂/∂ I¯2 ≈ 0, Yeoh [12,46] proposed an accordingly adapted function belonging to the class of
6
5
4
P [MPa]
2
UT sim
ET sim
1 PS sim
UT data
ET data
0
1 2 3 4 5 6 7 8 1 2 3 4 5 6 7 8 1 2 3 4 5 6 7 PS data
8
λ[] λ[] λ[]
Fig. 10 Performance of the n = 2−Swanson model. To double the number of parameters hardly influences fit quality and validity,
compare Fig. 9. Optimal ET-parameters show some qualitative improvement, UT- and PS-simulations now reproduce at least the
experimental shapes. Error(UT/ET/PS-sim) = 0.060/0.020/0.019
Mooney–Rivlin models (37). It abstains from considering the second invariant in the free energy and includes
all uncoupled I¯1 terms up to the power of three. Consequently, the isochoric free energy density reads
2 3
iso = c1 I¯1 − 3 + c2 I¯1 − 3 + c3 I¯1 − 3 , (70)
where c1 , c2 , c3 denote material parameters of shear modulus type, and the second index for the neglected I¯2
terms is omitted. The typical nonlinear increase of the shear modulus at high strains is reproduced sufficiently
accurate by this model due to the third-order I¯1 terms. A very similar model additionally incorporating the first
I¯2 term has been proposed in 1958 by Biderman [62]. Fictitious stress tensor, tangent operator and analytical
stress–stretch relations read
2
S̄ = 2c1 + 4c2 I¯1 − 3 + 6c3 I¯1 − 3 I (71)
4
C̄ = J − 3 8c2 + 24c3 I¯1 − 3 I ⊗ I (72)

P1UT = 2c1 + 4c2 I¯1UT − 3 + 6c3 I¯1UT − 3
2
λ − λ−2 (73)
2
ET
P1,2 = 2c1 + 4c2 I¯1ET − 3 + 6c3 I¯1ET − 3 λ − λ−5 (74)

P1PS = 2c1 + 4c2 I¯PS − 3 + 6c3 I¯PS − 3
2
λ − λ−3 (75)
and the usual optimisation procedure yields the following optimal material parameters [MPa]:
c1UT = 0.1634 c1ET = 0.2059 c1PS = 0.1776 c1 = 0.1834

c2UT = −1.198e−3 c2ET = −7.124e−4 c2PS = −1.62e−3 c2 = −1.432e−3
c3UT = 3.781e−5 c3ET = 3.078e−5 c3PS = 5.033e−5 c3 = 3.951e−5
The corresponding curves are plotted in Fig. 11. The cubic character of Yeoh’s model is obviously suited to
reproduce the S-shape at large strains. Nonetheless, a comparison of the UT-fitting errors with those of Swan-
son’s models reveals that the restriction to integer exponents of I¯1 may not be the ideal choice to approximate
arbitrary changes in curvature. On the other hand, as indicated by the almost perfect approximation of the
ET- and PS-data, this may also depend on the range of stretches used for parameter optimisation. Furthermore,
it can be stated that each set of optimal parameters yields simulation results for the complementary deformation
modes that are qualitatively much more reasonable than for any of the previous models. This is particularly
remarkable in view of iso lacking I¯2 and the small number of only three material parameters. Interestingly, the
combined parameter set slightly deteriorates the accuracy of which the experiments are reproduced, compare
the single fitting errors with those from the caption of Fig. 11. Yeoh’s stress–stretch relations are obviously
not distinct enough to cover the spread in deformation behaviour with a single parameter set.
6
5
4
P [MPa]
2
UT sim
ET sim
1 PS sim
UT data
ET data
0
1 2 3 4 5 6 7 8 1 2 3 4 5 6 7 8 1 2 3 4 5 6 7 PS data
8
λ[] λ[] λ[]
Fig. 11 Performance of Yeoh’s model on Treloar’s data. The fit quality is good to excellent, especially in view of the low para-
meter number. UT- and PS-parameters are fairly coincident numerically and in underestimating ET. ET-parameters systematically
overestimate UT and PS. Error(UT/ET/PS-sim) = 0.14/0.19/0.027
3.3.7 Arruda–Boyce model (1993) (invariant form)
Arruda and Boyce [2] developed a highly regarded constitutive model for rubber-like materials by assuming
that the macromolecular polymer network can locally be represented by eight chains that are oriented from the
centre along the diagonals of a cube. The edges of this cube are thereby aligned with the eigendirections of C̄
and stretched by the corresponding eigenvalues. The resulting chain deformations lead to forces according to
an energy function that is governed by the inverse Langevin function, that is, by non-Gaussian statistics, cf. also
Sect. 4. Since there is no analytical inversion of Langevin’s function available, different types of approximation
have been applied providing different formats of the corresponding free energy density iso . In Sect. 4.2, the
micro-mechanical version usually referred to as eight chain model will be discussed, which makes use of
a Padé approximation [64] to substitute the inverse Langevin function. Another approach considers a series
expansion leading to the following convenient, invariant-based energy formulation

Ck ¯k
K
1 1 11 19 519
iso = μ I − 3 k
, [C 1 , C 2 , C 3 , C 4 , C 5 ] = , , , , , (76)
N k−1 1 2 20 1050 7000 673750
k=1
in which μ is denoting the shear modulus, N is the number of Kuhn [65] segments per polymer chain, and the
summation is stopped here after five terms. Only I¯1 is occurring in (76), which simplifies the determination of
stress tensor, tangent operator and analytical stress–stretch relations:

K
k Ck ¯k−1
S̄ = 2μ I I (77)
N k−1 1
k=1

K
[k 2 − k] C
I¯1k−2 I ⊗ I
4
C̄ = 4J − 3 μ
k
(78)
N k−1
k=1
K
2μ k CkUT
P1 =
UT ¯
[ I1 ]
UT k−1
λ − λ−2 (79)
[N ]
UT k−1
k=1
K
2μET k Ck
P1,2 =
ET
[ ¯1 ]
I ET k−1
λ − λ −5
(80)
[N ET ]k−1
k=1
K
2μPS k Ck
P1 =
PS
[ I¯ ]
PS k−1
λ − λ−3 . (81)
[N ]
PS k−1
k=1
6
5
4
P [MPa]
2
UT sim
ET sim
1 PS sim
UT data
ET data
0
1 2 3 4 5 6 7 8 1 2 3 4 5 6 7 8 1 2 3 4 5 6 7 PS data
8
λ[] λ[] λ[]
Fig. 12 Performance of the Arruda–Boyce model on Treloar’s data. The fitting quality is quite good, especially if compared to
models having similarly few parameters, for example, Neo-Hooke and Mooney–Rivlin. Identified parameters also have a more
general validity in the sense that they qualitatively reproduce the other deformation modes, although over and underestimations are
not satisfying but follow the same trends observed already for the previous models. Error(UT/ET/PS-sim) = 0.156/0.187/0.061
Fitting the latter equations to Treloar’s data yields the following material parameter sets:
N UT = 20.25 N ET = 27.73 N PS = 50.33 N = 21.49
Figure 12 depicts all fits, the simulations of complementary deformation modes and the deviations wrt to
Treloar’s data. Although experimental findings are reproduced well, some improvement regarding the general
validity of parameters would be desirable and can probably be realised by considering additional terms in the
free energy. In view of the structural interpretation of N as number of Kuhn segments, that is, chain length, it
can hardly be considered admissible that three different values for N are obtained since not the tested material
structure has been changed but only the deformation mode. Regarding the combined parameter optimisation,
the remarks stated for Yeoh’s ansatz are valid also for this model, cf. the last part of the previous subsection.
3.3.8 Gent model (1996)
Another sophisticated, phenomenologically motivated two-parameter model has been proposed by Gent [32].
Horgan and Saccomandi [34,35] have later derived a micro-mechanical re-interpretation and proved Gent’s
model to have some similarity to the inverse Langevin ansatz proposed already by James and Guth [48] and
Treloar [52]. Especially, finite chain extensibility is accounted for by the parameter Jm which ensures the
blow-up of the logarithmic energy function for limiting values of [ I¯1 − 3]:

μ I¯1 − 3
iso = −Jm ln 1 − . (82)
2 Jm
Again, μ is of shear modulus type and fictitious stress tensor, tangent operator and Pi (λi ) follow as

μJm
S̄ = I (83)
Jm − I¯1 + 3

− 43 2μJm
C̄ = J I ⊗I (84)
[Jm − I¯1 + 3]2

μUT J UT λ3 − 1
P1UT = UT m 3 (85)
λ λJm − λ + 3λ − 2

μET JmET λ6 − 1
P1,2 = 4 ET
ET (86)
λ λ Jm − 2λ6 + 3λ4 − 1
6
5
4
P [MPa]
2
UT sim
ET sim
1 PS sim
UT data
ET data
0
1 2 3 4 5 6 7 8 1 2 3 4 5 6 7 8 1 2 3 4 5 6 7 PS data
8
λ[] λ[] λ[]
Fig. 13 Performance of Gent’s model on Treloar’s data. Fit quality and validity of the parameters are very similar to those of the
Arruda–Boyce and Yeoh models. Error(UT/ET/PS-sim) = 0.125/0.192/0.065

μPS JmPS λ4 − 1
P1PS = . (87)
λ λ2 JmPS − λ4 + 2λ2 − 1
Optimisation with respect to Treloar’s data provides the following parameter values:
JmUT = 81.16 JmET = 111.9 JmPS = 237.7 Jm = 84.57
Figure 13 depicts the corresponding curves. The capability of the model to reproduce the experimental data
as well as the validity of the material constants is—even numerically—very close to those obtained for Arru-
da–Boyce’ and Yeoh’s models. Also the combined parameter optimisation is providing results similar to those
discussed already in the two previous subsections.
3.3.9 Yeoh–Fleming model (1997)
The model by Yeoh and Fleming [13] links concepts from statistical mechanics and phenomenological
approaches to rubber elasticity. It adapts features from the ansatzes proposed earlier by Gent [32] and Yeoh [12],
namely a logarithmic term ensuring energy blow-up for chains reaching their maximum stretch and another
exponential part capturing the nonlinearities at small strains. As its predecessors, the model does not consider
influences of the second invariant I¯2 . The isochoric free energy density according to [13]

A
¯
I¯1 − 3
iso = 1 − exp −B[ I1 − 3] − C[Im − 3] ln 1 − (88)
B Im − 3
is governed by the four material parameters A, B, C, Im . Stress tensor, tangent operator and Pi (λi ) are

Im − 3
¯
S̄ = 2 A exp −B[ I1 − 3] + 2C I (89)
Im − I¯1

Im − 3
C̄ = 4J − 43 ¯
−AB exp −B[ I1 − 3] + C I ⊗I (90)
[Im − I¯1 ]2

2 AUT [λ3 − 1] UT λ − 3λ + 2
3 2C UT [ImUT − 3][λ3 − 1]
P1UT = exp −B + (91)
λ 2 λ ImUT λ2 − λ4 − 2λ

2 AET [λ6 − 1] ET 2λ − 3λ + 1
6 4 2C ET [ImET − 3][λ6 − 1]
ET
P1,2 = exp −B + (92)
λ5 λ4 ImET λ5 − 2λ7 − λ

2 APS [λ4 − 1] λ4 − 2λ2 + 1 2C PS [I PS − 3][λ4 − 1]
P1PS = exp −B PS + PS 3 m 5 (93)
λ 3 λ 2 I m λ − λ − λ3 − λ
6
5
4
P [MPa]
2
UT sim
ET sim
1 PS sim
UT data
ET data
0
1 2 3 4 5 6 7 8 1 2 3 4 5 6 7 8 1 2 3 4 5 6 7 PS data
8
λ[] λ[] λ[]
Fig. 14 Performance of the Yeoh–Fleming model on Treloar’s data. Reproduction of experiments is consistently perfect, the
general validity of parameters is comparable to that of the previous models. Error(UT/ET/PS-sim) = 0.09/0.184/0.055
and optimal material parameter sets for Treloar’s data follow as:
AUT = 0.0517 AET = 0.0467 APS = 0.0512 A = 0.0601

B UT = 0.2362 B ET = 0.1303 B PS = 0.1976 B = 0.0124
C UT = 0.1235 C ET = 0.1635 C PS = 0.1350 C = 0.1
ImUT = 83.23 ImET = 93.35 ImPS = 94.13 Im = 78.26
Figure 14 depicts the respective results. As a consequence of the model’s extreme nonlinearity, together with its
significant four-parameter flexibility, the optimisation procedure is even more sensitive wrt the initial values
than that of the second Swanson model. In particular, the combined parameter identification did not at all
provide any stable result, the given values are neither reliable nor do they allow for further conclusions.
3.3.10 Carroll model (2011)
Carroll [47] recently proposed a phenomenological approach which, interestingly, is based on a successive
extension of the free energy by terms that are chosen such that they reduce the errors that remain in the stress
response of the previous terms, compared to measurements. Firstly, Treloar’s UT-data are fitted with a Neo-
Hookean function. Its stress response is then subtracted from the measured curve and the remaining difference
is fitted once more, now with a term proportional to I¯14 . Both terms together are used to simulate ET and the
deviation from Treloar’s ET-data is finally approximated by a last summand proportional to the square root of
I¯2 . The isochoric free energy density thus reads

iso = a I1 + b I1 + c I¯2
¯ ¯4
(94)
and is adjusted by three stiffness-like parameters a, b, c. Carroll identified [a, b, c] = [0.15, 3.1e−7, 0.095]
MPa, which is, of course, a result based on the simultaneous optimisation wrt two deformation experiments.
The corresponding curves are depicted in Fig. 15 (left) and show the expected accuracy. Also PS is captured
very precisely by these parameters, whose general validity is, altogether, unreached by any other of the here
regarded models. Fictitious stress tensor, tangent operator coefficients and stress–stretch relations follow as
−1/2
S̄ = 2a + 8b I¯13 I + c[ I¯1 I − C̄] I¯2 (95)
−1/2 −3/2 −3/2 −3/2 −1/2
δ̄1 = 48b I¯12 + 2c I¯2 − c I¯12 I¯2 , δ̄2 = c I¯1 I¯2 , δ3 = −c I¯2 , δ4 = −2c I¯2 (96)
3
−1/2
P1UT = 2a + 8b 2λ−1 + λ2 + c 1 + 2λ3 λ − λ−2 (97)
3 −1/2
ET
P1,2 = 2a + 8b λ−4 + 2λ2 + cλ2 2λ−2 + λ4 λ − λ−5 (98)
3 −1/2
P1PS = 2a + 8b λ2 + λ−2 + 1 + c λ2 + λ−2 + 1 λ − λ−3 . (99)
Carroll’s parameters ET fitted unconstrained (c = − 0.1099), UT/PS simulated
6
0.092 = Error(UT−fit) 0.017 = Error(ET−fit)
0.025 = Error(ET−sim) 1.136 = Error(UT−sim)
0.023 = Error(PS−sim) 0.452 = Error(PS−sim)
5
4
P [MPa]
0
1 2 3 4 5 6 7 8 1 2 3 4 5 6 7 8
λ[] λ[]
Fig. 15 Performance of Carroll’s model on Treloar’s data using parameters from [47] (left) and from unconstrained fitting of
ET-data (right), which provides a negative value for c
6
5
4
P [MPa]
2
UT sim
ET sim
1 PS sim
UT data
ET data
0
1 2 3 4 5 6 7 8 1 2 3 4 5 6 7 8 1 2 3 4 5 6 7 PS data
8
λ[] λ[] λ[]
Fig. 16 Performance of the Carroll model on Treloar’s data. Error(UT/ET/PS-sim) = 0.083/0.024/0.02
Fitting to Treloar’s data yields the following material parameter sets:
a UT = 0.1481 MPa a ET = 0.1988 MPa a PS = 0.1297 MPa a = 0.1433 MPa

bUT = 3.024e−7 MPa bET = 3.141e−7 MPa bPS = 4.91e−7 MPa b = 3.17e−7 MPa
cUT = 0.06623 MPa cET = 2.2e−14 MPa cPS = 0.1876 MPa c = 0.1118 MPa
The corresponding curves are depicted in Fig. 16 and reveal a remarkable model performance. Both fitting
quality and validity of the parameter sets are very good to perfect. The errors remaining from UT-based sim-
ulations are as convincingly low as those in Fig. 15, left, although the latter are based on parameters derived
with the additional help of ET-data. Optimisation wrt PS-data provides parameters slightly overestimating
complementary deformation modes. Caution has to be exercised when fitting to ET-data: without enforcing
physically reasonable ranges for the parameters, [a, b, c] result as [0.2514, 3.405e−7, −0.1099], that is, c
becomes negative and both UT and PS are overestimated significantly, compare Fig. 15, right. This situation
improves when c > 0 is enforced, and UT- and PS-simulations then are more reasonable (Fig. 16, center) since
the I¯2 -term is just switched of numerically. Simultaneous fitting to all deformation data provides, as expected,
almost perfect results, cf. the error values in the caption of Fig. 16. Only Swanson’s eight-parameter model
shows a slightly better performance.
4 Micro-mechanical models: stress tensors, tangent operators and performance
Micro-mechanical approaches to rubber elasticity usually depart from a description of the deformation behav-
iour of a single polymer chain. The macromolecule constituting the chain is itself built from a certain number
of chemically identical repeat units, the so-called monomers. This structure is frequently modelled as a chain
of N rigid beams, each of length l, which are allowed to be arbitrarily oriented with respect to each other
(freely-jointed chain). The beams are commonly denoted as Kuhn segments, and the assumption of free rota-
tion usually requires that each segment comprises multiple monomers since chemical bond angles can take
only certain admissible values. The maximum distance between the two chain ends equals its contour length
rmax = Nl, that is, in case the chain is fully elongated and all segments are aligned identically. From statisti-
cal considerations
√ (random walk chain), the end-to-end distance of a stress-free undeformed chain results as
r0 = Nl, cf. e.g. [65], which motivates the introduction of the following chain stretch:
r r √
= =√ ∈ [0, N ]. (100)
r0 Nl
Similar to macroscopic continua, the elastic behaviour of a single chain is described in terms of a scalar free
energy function, which will here be denoted by ψ = ψ(). Two prominent examples are the so-called Gauss
and Langevin chains, respectively, whose energy functions are given by
3
ψ Gauss () =
k B 2 + ψ0 (101)
2 √
⎡ ⎛ ⎞⎤
L −1 N −1
√
ψ Langevin () = k B N ⎣ √ L−1 N −1 + ln ⎝ √ ⎠⎦ + ψ0 (102)
N sinh L −1 N −1
and wherein k B , and L−1 denote Boltzmann’s constant, absolute temperature and inverse of the Langevin
function L(•) = coth (•) − (•)−1 , respectively. Different other approaches are discussed in the literature, for
example, the worm-like chain [63], which abandons the rigidity of the beams and assumes that successive
segments are oriented in √
roughly the same direction. Note that the Gaussian chain is valid only for low to
moderate stretches
N since the corresponding chain force
∂ψ Gauss
f Gauss () = = 3k B (103)
∂
is a linear function and does not adequately reflect finite chain extensibility, that is, the dramatic force increase
that is observed if the chain approaches its maximal end-to-end distance rmax . In numerical applications, the
inverse of Langevin’s function is usually substituted by the following Padé approximation [64]:
√ 3N − 2
L−1 N −1 ≈ N −1 . (104)
N − 2
With the above type of chain description, macroscopic material models are derived by averaging the energies
of a certain ensemble of chains, which ideally is chosen such that it reflects the mechanical behaviour of the
true polymer network as realistic as possible:
n
K

iso = iso C̄ := n ψ ≈ ψ(k ). (105)
K
k=1
We remain here with the assumption of incompressibility and the corresponding energy decomposition of the
previous sections and consider purely isochoric behaviour. The number of chains K as well as their spatial
orientations are decisive with regard to qualitative behaviour and (an)isotropy of the model, whereas the chain
density n quantitatively relates micro- and macro-stiffnesses. Furthermore, some relation between macroscopic
deformation C̄ and chain stretches k has to be defined, for which nonaffine as well as affine approaches like
the following have been proposed in the literature: If we assume the initially unstretched chain to be aligned
√
with a unit normal vector t 0k , its end-to-end vector reads r 0k = Nlt 0k and a deformation gradient F̄ implies
r k = F̄ r 0k for the end-to-end vector of the correspondingly stretched chain. Consequently,
!
√ T
t 0k , F̄ F̄ t 0k
F̄ r 0 , F̄ r 0 Nl
r k k k
k = 0 = √ = √ = t 0k , C̄t 0k = C̄ : [t 0k ⊗ t 0k ] (106)
r Nl Nl
k
as required to evaluate Eq. (105). Three different micro-mechanical models, which result from certain choices
for chain energy ψ, chain number K , initial chain orientations t 0k and micro- and macro-stretch relation

k = k C̄ , will be discussed in the following subsections.
For macroscopic energy densities defined as in Eq. (105), the corresponding isochoric stress tensors and
tangent operators required for finite element implementation can again be calculated from Eq. (5) together
with S̄ = 2∂iso (C̄)/∂ C̄ and Eq. (7), respectively.
Regarding the analytical stress–stretch relations for UT, ET and PS, which are essential to evaluate the
performance of also the micro-mechanical models on Treloar’s data, we consider a re-parametrisation of iso
in terms of the isochoric principal macro-stretches λi . From Eq. (16), the pressure p follows as
∂iso
p = λj (107)
∂λ j
from a nominal principal stress P j being zero due to the type of deformation, for example, P2 = 0 in the case
of uniaxial tension. Reinserted into (16), utilisation of chain energy average (105) yields
∂iso λ j ∂iso
Pi = − , i ∈ {1, 2, 3} , j such that P j = 0
∂λi λi ∂λ j
n ∂ψ ∂k λ j n ∂ψ ∂k
K K
= − . (108)
K ∂k ∂λi λi K ∂k ∂λ j
k=1 k=1
Therein, the derivatives of chain stretches k wrt principal macro-stretches λi follow from corresponding
relations like (106), while the chain forces f k = ∂ψ(k )/∂k are given by (103) for the Gaussian case and
∂ψ Langevin √ √
= k B N L−1 k N −1 , (109)
∂k
for the Langevin chain, respectively, compare, for example, [6].
4.1 Three chain model (1943)
The three chain model [48,49] locally substitutes the true polymer network by K = 3 chains and assumes each
of them to be aligned with one of the eigenvectors N i of the isochoric right Cauchy-Green tensor, cf. Fig. 17.
As before, only the isochoric part of the deformation is considered and modelled, that is, all overbars are again
omitted for simplicity.
ˆˆ
Fig. 17 Three chain model: initial and deformed chain orientations and stretches
It is obvious from the particular choice of their orientations that the stretch of each chain is equal to the
corresponding isochoric principal stretch, that is,
λk r0 N k
k = = λk , k = 1, 2, 3. (110)
r0 ek
Regarding the free chain energy ψ, both the Gaussian and Langevin ansatzes have been discussed in the
literature. While insertion of (101) with (110) into (105) yields a macroscopic model behaving identically to
the Neo-Hookean, the incorporation of the more sophisticated Langevin chain provides a rubber model which
√ to the frequently used definitions for shear modulus μ = nk B
is valid also at very large strains. Resorting
and inverse Langevin function γk = L−1 ( N −1 k ), the macroscopic free energy density is given as
3
n Langevin μN −1
3
γk
iso = ψ (k ) = N λk γk + ln (111)
3 3 sinh (γk )
k=1 k=1
according to Eq. (105) together with (110) in (102). Fictitious stress tensor S̄ = 2∂iso /∂ C̄ then follows from
the chain rule, Eq. (109) with Padé approximation (104) and in view of
∂k (110) ∂λk 1 −1
= = λ [N k ⊗ N k ] , (112)
∂ C̄ ∂ C̄ 2 k
and depends on the isochoric principal stretches and the eigenvectors of C̄:
μ 3N − λ2k
3
S̄ = [N k ⊗ N k ] , (113)
3
k=1
N − λ2k

2 −4 3
2N 3N − λ2k ∂ (N k ⊗ N k )
C̄ = μJ 3 2 [N k ⊗ N k ⊗ N k ⊗ N k ] + . (114)
3
k=1 N − λ2k N − λ2k ∂ C̄
Fictitious tangent C̄ required another derivation wrt C̄, details on the computation of ∂ (N k ⊗ N k ) /∂ C̄ can
be found in [54]. Analytical stress–stretch relations can be derived from Eq. (108), which simplifies consider-
ably since ∂k /∂λi = δki due to (110). After insertion of the deformation-specific principal stretches λ1,2,3 ,
cf. Sects. 3.1.1–3.1.3, and together with Padé approximation (104) of (109), one finally obtains
UT − λ−1
1 UT 3λN UT − λ3 −2 3N
P1 = μ
UT
−λ (115)
3 N UT − λ2 N UT − λ−1
ET − λ−4
1 3λN ET − λ3 −5 3N
ET
P1,2 = μET − λ (116)
3 N ET − λ2 N ET − λ−4
PS − λ−2
1 PS 3λN PS − λ3 −3 3N
P1 = μ
PS
−λ , (117)
3 N PS − λ2 N PS − λ−2
√
which become singular for λ → N . The usual optimisation provides the following material parameter sets:
N UT = 77.29 N ET = 45.41 N PS = 165.3 N = 82.1
Figure 18 depicts the corresponding curves; all experiments are perfectly reproduced, whereas the general
validity of parameters cannot convince. The values for μ are reasonable, that is, ET is stiffer than PS, itself
being stiffer than UT, which is not the case for the numbers of Kuhn segments N . These are neither identical,
as one would expect due to the physical interpretation, nor does their order reflect the assumption of softer
material behaviour due to longer chains. Additionally, all stress–stretch relations yield almost identical curves,
their structure is not suited to distinguish different deformation modes. This has already been observed, for
example, for the Neo-Hooke or Gent–Thomas model, the consequences regarding the quality of parameters
obtained from combined optimisation are similar. Even more important is the fact that optimal ET-parameters
cannot be used to simulate UT since the necessary stretch range exceeds the corresponding locking stretch
6
5
4
P [MPa]
2
UT sim
ET sim
1 PS sim
UT data
ET data
0
1 2 3 4 5 6 7 8 1 2 3 4 5 6 7 8 1 2 3 4 5 6 7 PS data
8
λ[] λ[] λ[]
Fig. 18 Performance of the three chain model on Treloar’s data. Fitting quality is convincing, especially if compared to models
having similarly few parameters like Neo-Hooke or Mooney–Rivlin. Optimised parameters lack a general validity and care must
be taken in view of segment numbers and locking stretches. Error(UT/ET/PS-sim) = 0.148/0.333/0.075
√
N ET = λET locking = 6.74 < 7.69 = λmax , which produces an unphysical pole. To constrain the optimisation
UT
by fixing N at a value suited to capture the whole stretch range is not advisable. The resulting increase of fitting
errors is not accompanied by a higher validity of parameters due to the likeness between stress relations, also
compare the low impact of the second terms in Eqs. (115–117).
4.2 Eight chain model (1993)
Similar to the three chain model, the eight chain approach proposed by Arruda and Boyce [2] is resorting to
a cuboid spanned by the principal directions of the isochoric right Cauchy-Green tensor. The local polymer
network is—as the name indicates—then approximated by an ensemble of K = 8 chains, each of which being
oriented along one of the half diagonals
√ of the cuboid, cf. Fig. 19. The edge length of the undeformed
√ cube
immediately follows as a0 = 2 3−1r0 from basic geometric considerations, wherein r0 = Nl is again
denoting the end-to-end distance of an unstretched chain. All chains are equally elongated
if the surrounding
√
box is deformed with the isochoric principal stretches λi . This leads to rk = r0 3 −1 λ21 + λ22 + λ23 for the
new end-to-end distances and one finally obtains
"

rk 1 I¯1
= k = = √ λ21 + λ22 + λ23 = , k = 1, . . . , 8 , (118)
r0 3 3
that is, a unique stretch which is valid for all chains regardless of their orientation within the cuboid.
Fig. 19 Eight chain model: initial and deformed chain orientations and stretches
If Langevin chain behaviour is assumed, evaluation of (105) provides

n Langevin
8
(118) γ
iso = ψ (k ) = nψ Langevin () = μN N −1 γ + ln (119)
8 sinh (γ )
k=1
√
for the macroscopic free energy density, wherein μ = nk B is the shear modulus, and γ = L−1 ( N −1 )
again abbreviates the inverse Langevin function. Fictitious stress tensor S̄ = 2∂iso /∂ C̄ is then computed to

μ 3N − 2 μ 9N − λ21 − λ22 − λ23
S̄ = I= I (120)
3 N − 2 3 3N − λ21 − λ22 − λ23
via the chain rule, applying Padé approximation (104) to Eq. (109) and with ∂/∂ C̄ = (6)−1 I since
∂ I¯1 /∂ C̄ = I . According to (8), the fictitious tangent then follows from another derivative wrt C̄:
4
4 4 1 4μN J − 3
C̄ = μN J − 3 2 I ⊗ I = 2 I ⊗ I (121)
9 N − 2 3N − λ21 − λ22 − λ23
Analytical stress–stretch relations are again derived from (108), which can be simplified beforehand due to
identical chain stretches k = , since ∂/∂λi = λi /3 holds according to (118), and with the usual Padé
approximation of Langevin chain force (109):

μ 3N − 2 λ2j
Pi = λi − , i ∈ {1, 2, 3} , j such that P j = 0. (122)
3 N − 2 λi
With deformation specific, squared chain stretches 2UT,ET,PS = I¯1UT,ET,PS /3 as following from (18), (21) and
(24) one obtains

μUT 9N UT − λ2 − 2λ−1
P1UT = −1
λ − λ−2 (123)
3 3N − λ − 2λ
UT 2
ET
μ ET 9N − 2λ2 − λ−4 −5

ET
P1,2 = λ − λ (124)
3 3N ET − 2λ2 − λ−4
PS
μ PS 9N − λ2 − λ−2 − 1
P1PS = −2
λ − λ−3 (125)
3 3N − λ − λ − 1
PS 2
and fitting to Treloar’s data provides the following parameter sets:
N UT = 25.84 N ET = 30.32 N PS = 55.55 N = 26.54
Figure 20 depicts the corresponding curves. Reproduction of experimental data via fitting is as perfect as for
the three chain model while the general validity of parameters is better, although not yet satisfying, and can be
compared to that of Yeoh’s model. Optimised segment numbers are again neither identical nor do they satisfy
the intuitive expectations concerning the stiffnesses of the different deformation modes. Compared to the three
chain model, simultaneous parameter identification yields better results, but individually optimised values still
provide lower deviations wrt the measurements. Equations (123–125) cannot sufficiently distinguish defor-
mations with only a single parameter set. Note that the eight chain model is much less sensitive wrt locking
stretches, that is, all segment numbers are valid for the whole stretch range, cf. also [5].
6
5
4
P [MPa]
2
UT sim
ET sim
1 PS sim
UT data
ET data
0
1 2 3 4 5 6 7 8 1 2 3 4 5 6 7 8 1 2 3 4 5 6 7 PS data
8
λ[] λ[] λ[]
Fig. 20 Performance of the eight chain model on Treloar’s data. The fitting quality is convincing, but all parameters still lack
sufficient generality. Comparable to Yeoh’s ansatz, the model better distinguishes deformation modes than the three chain version
and is not sensitive with respect to locking stretches. Error(UT/ET/PS-sim) = 0.099/0.188/0.055
4.3 Unit sphere model (2002)
Different to the above discussed three and eight chain approaches, the unit sphere model proposed by Miehe
et al. [4–6] does not consider the C̄-eigenvectors cuboid as the confining volume of the local polymer network.
It instead resorts to a unit sphere in which the chains are assumed to be oriented along radius vectors from the
centre to the surface. In the most general case, this would require to extend average (105) to an integration
over the energies of chains aligned with all possible unit vectors. To avoid a complicated analytical evaluation,
a numerical scheme inspired by a contribution of Bažant and Oh [66] on the discrete integration over spheres
is applied. Similar to a Gauss integration, K = 21 unit vectors t 0k and weight factors wk are chosen such
that an approximately uniform distribution of the chains across the sphere is realised, which ensures isotropic
behaviour of the local network. If Langevin chain behaviour is assumed, the macroscopic isochoric free energy
density follows as

21
21
γk
iso = n wk ψ Langevin
(k ) = μN wk N −1 k γk + ln , (126)
sinh (γk )
k=1 k=1
√
wherein μ = nk B and γk = L−1 ( N −1 k ) denote inverse Langevin function of the chain stretches as
before. Due to the particular choice of initial chain orientations t 0k , the stretches k must be computed here
according to Eq. (106)

k C̄ = t 0k · C̄ · t 0k , (127)
that is, they do not follow directly in closed form via geometric considerations as has been the case for the
three (110) or eight (118) chain model. Coordinates of the t 0k and the weight factors wk can be found, for
example, in [5]. Similar to the derivation for the three chain model, fictitious stress tensor S̄ follows from
applying the chain rule to Eq. (126)1 , insertion of Padé approximation (104) of Eq. (109) and together with
relation ∂k /∂ C̄ = (2k )−1 [t 0k ⊗ t 0k ]:

21
∂ψ Langevin ∂k
21
3N − 2k 0
S̄ = 2n wk =μ wk t k ⊗ t 0k (128)
∂k ∂ C̄ N − k
2
k=1 k=1
4
21
4N 0
C̄ = μJ − 3 wk 2
t k ⊗ t 0k ⊗ t 0k ⊗ t 0k . (129)
k=1 N − k 2
The corresponding fictitious tangent operator simply required another derivative wrt C̄. To derive analytical
stress–stretch relations for UT, ET and PS, we depart from
21
∂ψ Langevin ∂k λ j ∂k
Pi = n wk − , j such that P j = 0, (130)
∂k ∂λi λi ∂λ j
k=1
as obtained by insertion of energy density (126)1 into (108)1 . The chain forces ∂ψ Langevin /∂k contained are
again substituted by Padé approximation (104) of (109) into which the deformation-specific chain stretches

t t t
2 t + k2 + k3 , ET = λ2 t + λ2 t + k3 , PS = λ2 t + t + k3
t
UT
k = λ k1 k k1 k2 k k1 k2 (131)
λ λ λ4 λ2
in terms of applied macro-stretch λ have to be inserted. These are easily obtained by evaluating (127) with the
deformation mode dependent strain tensors C̄, which in turn follow from the particular deformation gradients
given in Eqs. (18)1 ,(21)1 and (24)1 . For the sake of simplicity, we have introduced
tki := [ei · t 0k ]2 (132)
to abbreviate the squared Cartesian coordinates of the initial chain orientation vectors. The derivatives of chain
stretches wrt principal stretches appearing in (130) require the consideration of spectral decomposition

3
C̄ = λa2 N a ⊗ N a . (133)
a=1
Inserted into (127), we find

∂k ∂ 1 ∂ 0 3
= t 0k · C̄ · t 0k = t k · a=1 λa2 N a ⊗ N a · t 0k
∂λi ∂λi 2k ∂λi
1 ∂ 3 3
∂ 2
= λ 2 t 0 · [N ⊗ N ] · t 0 = 1 λ N · t 0 2 , (134)
a=1 a k a a a
2k ∂λi k
2k ∂λi a k
a=1
that is, the general computation of these derivatives necessitates the eigenvectors N a of C̄. Since the here
considered UT, ET and PS involve diagonal deformation gradients only, the eigenvectors are not rotated from
their reference orientation, and we may identify N a = ea , which allows to further simplify Eq. (134) towards
1 ∂ 2 2 (132) 1
3 3
∂k λi tki
= λa ea · t 0k = 2λa δai tka = . (135)
∂λi 2k ∂λi 2k k
a=1 a=1
If we now reformulate (130) with (104) of (109) and the squares of chain stretches (131), with (135) and
λUT UT = λ−3/2 , λET /λET = λ−3 , λPS /λPS = λ−2 (cf. Sects. 3.1.1–3.1.3), the desired stress–stretch
2 /λ1 3 1 3 1
relations finally read:

21
3N UT − λ2 tk1 − λ−1 [tk2 + tk3 ] tk2
P1 = μ
UT UT
wk UT λtk1 − 2 (136)
N − λ2 tk1 − λ−1 [tk2 + tk3 ] λ
k=1
21
3N ET − λ2 [tk1 + tk2 ] − λ−4 tk3 tk3
ET
P1,2 =μ ET
wk ET λtk1 − 5 (137)
N − λ2 [tk1 + tk2 ] − λ−4 tk3 λ
k=1
21
3N PS − λ2 tk1 − tk2 − λ−2 tk3 tk3
P1PS = μPS wk λt k1 − . (138)
N PS − λ2 tk1 − tk2 − λ−2 tk3 λ2
k=1

These equations are not very handy but can be shortened before implementation by exploiting tki = 1 valid
for the initial chain orientations and further by collecting the summands having identical tki and wk values
into nine groups. Optimisation wrt Treloar’s data then yields
N UT = 63.74 N ET = 38.02 N PS = 101.3 N = 64.52
6
5
4
P [MPa]
2
UT sim
ET sim
1 PS sim
UT data
ET data
0
1 2 3 4 5 6 7 8 1 2 3 4 5 6 7 8 1 2 3 4 5 6 7 PS data
8
λ[] λ[] λ[]
Fig. 21 Performance of the unit sphere model on Treloar’s data. Error(UT/ET/PS-sim) = 0.125/0.263/0.101
for the material parameter sets. Figure 21 shows the corresponding curves. Reproduction of experimental data
by fitting is quite acceptable, although the errors are slightly higher than for some of the phenomenologi-
cal models. Unconstrained optimisation induces the already discussed
√ inconsistencies concerning segment
number N , fitting ET-data in particular yields a locking stretch N = 6.17 far to small for the UT strain
ET
range. Furthermore, the structural differences between analytical relations (136–138) are again not sufficient
to reproduce all deformation modes with only one parameter set. This situation improves significantly if the
affinity assumption contained in (106) is substituted by more sophisticated approaches to link chain- and
macro-stretches, like in the eight chain model or the general ansatz proposed in [5]. Further improvement is
obtained by so-called tube constraints which additionally account for an increasingly hindered chain stretch
due to the lateral contraction of the whole network, also cf. [5] or [6].
5 Conclusion
This contribution presents a survey on fourteen representatives out of the numerous hyperelastic models for
rubber-like materials. A first focus lies on the derivation of accurate tangent operators that are indispensable
to ensure quadratic convergence when boundary value problems are solved by Newton-like iterative schemes.
Such derivations traditionally are quite time-consuming for a beginner in this field, and we hope that these
pages may occasionally be of some assistance. Especially, the test procedure in “Appendix B” together with the
material parameters given for every model is suited to check the correctness of own codes for stress tensors and
tangent operators. To evaluate the performance of a particular model constitutes a second focus and is similarly
challenging, especially if the experimental data available are limited. In this context, it is usually helpful to
quickly get an impression on the general behaviour of a model by optimising it with respect to one-dimensional
measurements, some of which can be obtained easily in many cases. The ingredients necessary for such fitting
procedures are derived for all models, and the application to Treloar’s elastomer data may provide some hints
concerning choice or design of a model for a particular application.
Appendix A: Treloar Data
To evaluate the performance of particular constitutive models for elastomers, it is common practice to test their
ability to reproduce the classic experimental data of Treloar [1]. Most publications only mention that these
data have been used, some others like Arruda and Boyce [2] at least plot the data used, but numbers cannot be
found in the literature so far. To remedy this deficiency, we give the quantitative details as derived from the
original paper of Treloar [1], who extensively investigated the behaviour of vulcanised rubber containing 8% of
sulphur. A standard graphics software has been used to extract pairs of stretch values and corresponding Piola
stresses from the curves for uniaxial tension (UT, p. 63), pure shear (PS, p. 66) and equibiaxial tension (ET,
p. 62) given in [1]. For the latter case, the originally plotted stress values additionally had to be multiplied by
the stretch to obtain the Piola stress. Different to various other authors, we use the complete set of experimental
data for fitting, that is, deviations in the material parameters obtained here and elsewhere are explicable by
differences in the chosen (sub)sets of experimental data as well as in numeric deviations caused by the data
extraction.
Experimental data for UT, PS and ET of vulcanised rubber containing 8% sulphur, as extracted from Treloar [1]:
uniaxial tension pure shear equibiaxial tension
[] P [MPa] [] P [MPa] [] P [MPa]
6
1.00 0.00 1.00 0.00 1.00 0.00
1.01 0.03 1.06 0.07 1.04 0.09
1.12 0.14 1.14 0.16 1.08 0.16 uniaxial tension (UT)
1.24 0.23 1.21 0.24 1.12 0.24 pure shear (PS)
5
1.39 0.32 1.32 0.33 1.14 0.26 equibiaxial tension (ET)
1.61 0.41 1.46 0.42 1.20 0.33
1.89 0.50 1.87 0.59 1.31 0.44
1. PK stress [MPa]
2.17 0.58 2.40 0.76 1.42 0.51 4
2.42 0.67 2.98 0.93 1.69 0.65
3.01 0.85 3.48 1.11 1.94 0.77
3.58 1.04 3.96 1.28 2.49 0.96
4.03 1.21 4.36 1.46 3.03 1.24 3
4.76 1.58 4.69 1.62 3.43 1.45
5.36 1.94 4.96 1.79 3.75 1.72
5.76 2.29 4.03 1.96
6.16 2.67 4.26 2.22 2
6.40 3.02 4.44 2.43
6.62 3.39
6.87 3.75
7.05 4.12 1
7.16 4.47
7.27 4.85
7.43 5.21
7.50 5.57 0
7.61 6.30 A total of 56 measurements. 1 2 3 4 5 6 7 8
λ[ ]
It is important to note that these deformations are by far not reversible, but, in fact, are revealing significant
inelastic characteristics if stretches λ ≥ 3 . . . 4 are applied, cf. the original plots in [1] or [52]. This behaviour,
which is frequently explained by strain-induced crystallisation, had been paid particular attention already by
Treloar himself. Recently, Gent et al. [67] have pointed out again that it is, therefore, not appropriate to use
strain energy functions to model rubber-like materials, at least not for the very large strain data on lightly
cross-linked natural rubber as published by Treloar. It was further concluded from a universal relation between
torsional stiffness and tensile stress derived by Rivlin [11], that it is at least doubtful to apply hyperelasticity
to ‘a typical elastic solid’ strained more than 300%. As a method to determine the range of deformation that is
valid for a material to be described by a strain energy function the so-called Mooney–Rivlin plot comparing
torsional and tensile behaviour is recommended.
Despite the limited reversibility of Treloar’s stress–strain data, the whole strain range is employed here to
identify the material parameters and to evaluate the models for the following reason. Various other elastomers
are known to deform reversibly while simultaneously exhibiting certain characteristics observable in Treloar’s
data, for example, large strains and the typical S-shape. We, therefore, consider these data as a prototype or
worst-case scenario of highly nonlinear behaviour, which one might want to reproduce by using hyperelastic
models, and tacitly ignore that they originate from an inelastic material. Thus, all parameters given in the
previous sections have to be understood as valid for a fictitious, highly nonlinear elastic material rather than
for slightly cross-linked natural rubber as used by Treloar. To assess the performance of the models with respect
to this particular material, all fitting procedures have to be repeated on the subset of reversible deformations,
which is postponed to a later contribution since it would inadmissibly extend the scope of this work.
Appendix B: A local check for quadratic convergence
A desirable feature of numerical solutions of continuum mechanical boundary value problems by iterative
techniques is to obtain quadratic convergence. This property is given if the tangent operator used to update the
stress state in consequence of a new strain increment is consistently derived. To prove quadratic convergence for
the tangent operators given in this contribution, that is, to assure properness of derivation and implementation,
we use a simple but effective local method described in the following.
Both the 2nd Piola–Kirchhoff stress S as well as the material tangent operator C are given by the free
energy according to Eq. (1), that is, they depend on the right Cauchy-Green tensor C. By prescribing a
particular stress state S ∞ = const., we can formulate the computation of the corresponding strain state C ∞
of a particular constitutive model as to calculate the root of the nonlinear tensor-valued residual
∂(C) !
R(C) := S ∞ − 2 = S ∞ − S(C) = 0. (139)
∂C
This can be solved using Newton’s method, that is, by incrementally updating the strain tensor according to
R (C n ) : C = −R(C n ). The required 4th-order Jacobian R coincides with the tangent operator:
# #
∂R(C) ## 1 ∂ 2 (C) ## 1
R (C n ) = =− 4 = − C(C n ), (140)
∂C #C n 2 ∂C 2 #
Cn 2
that is, one has to solve

1
C(C n ) : C n+1 − C n = S ∞ − S(C n ). (141)
2
Using symmetry and Voigt’s notation, this reduces to a matrix-vector equation, that is, a six-dimensional system
of equations has to be solved. If the initial strain C 0 is chosen sufficiently close to the solution, the iteration
will converge quadratically. This procedure provides a simple but general method to check accuracy and con-
sistency of arbitrary tangent operators—without a complete finite element implementation. It is realised by
the subsequent Matlab code which calls the functions Stress and Tangent for the calculation of S(C n )
and C(C n ) according to the analytical formulations derived in the particular subsection of each model.
function [ni] = check(C_0,S_inf,it)

C_n = C_0; C ∞ = [·x x , · yy , ·zz , ·x y , ·x z , · yz ]
Jac = zeros(6,6); Neo-Hooke [1.58, 2.46, 1.22, 1.53, 0.0, 0.0]
Tol = 1.0e-8; Mooney–Rivlin [1.64, 2.57, 1.18, 1.62, 0.0, 0.0]
for i = 1:it Isihara [1.53, 2.38, 1.23, 1.46, 0.0, 0.0]
C_old = C_n;
Gent–Thomas [1.69, 2.69, 1.19, 1.72, 0.0, 0.0]
R = Stress(C_n)-S_inf; cf. (139)
Swanson* [1.57, 2.44, 1.23, 1.51, 0.0, 0.0]
Jac = Tangent(C_n); cf. (140)
Yeoh [1.57, 2.44, 1.22, 1.51, 0.0, 0.0]
Jac(1:6,4:6) = 2*Jac(1:6,4:6);
Delta = Jac\(-R’); Arruda–Boyce [1.54, 2.39, 1.23, 1.47, 0.0, 0.0]
C_n = C_n + Delta; Gent [1.54, 2.39, 1.23, 1.47, 0.0, 0.0]
Diff = norm(C_n-C_old,2); Yeoh–Fleming [1.55, 2.41, 1.23, 1.48, 0.0, 0.0]
if Diff <= Tol Carroll [1.57, 2.44, 1.23, 1.52, 0.0, 0.0]
break;end Three chains [1.55, 2.41, 1.23, 1.48, 0.0, 0.0]
if i == it Eight chains [1.55, 2.40, 1.23, 1.48, 0.0, 0.0]
fprintf(’No convergence...’); Unit sphere [1.57, 2.44, 1.22, 1.51, 0.0, 0.0]
break;end
*Optimal Treloar parameters for n = 4 as in [59] are used.
ni = i;
end
For the prescribed stress state we have chosen S ∞ = [6.55, 4.3, 3.5, −3.9, 0.0, 0.0] while for the initial
strain state C 0 = [1.55, 2.5, 1.2, 1.5, 0.1, 0.1] is used. The table above lists the final strain states, where-
upon the respective optimal UT-parameters as stated in the subsection of each model were used. Thus, the
interested reader may check the correctness of own stress tensor implementations via comparison. Quadratic
convergence, that is, a faultless tangent operator implementation, is indicated if the number of correct digits is
doubled in every iteration—similarly as in the example table of Sect. 3.2, which shows the evolutions of the
Euclidean norms of strain updates C n+1 − C n 2 and residuals R(C n )2 .
Appendix C: Correlation matrices of parameter sensitivities
Further insight to the roles particular parameters play in reproducing the material behaviour can be obtained
from analysing the correlation of their responses sensitivities. A method to compute corresponding correlation
matrices has been proposed in [68] and was applied successfully by, for example, [69] to discern the individual
relevance of parameters in complex models. As in [69], we compute the vectors of relative response sensitivity
dR
r j = κj
, no sum on j (142)
dκ j
$ % $ %T
R = P1UT λiUT ; κ , {P1ET λiET ; κ }, P1PS λiPS ; κ ∈ R56 (143)
requiring response vector R, which compiles the simulation output of the model under consideration for UT,
ET and PS at Treloar’s measuring points. Thereby, the parameters κ from combined optimisation wrt to all
experiments as given in each subsection have been used. Note that the derivatives of all stress–stretch relations
wrt every material parameter are required. The entries of the correlation matrices given below for each model
then follow according to
rm · rn
cmn = ∈ [−1, 1]. (144)
r m 2 r n 2
If cmn → ±1, the parameters κm , κn correlate, that is, one of them is dispensible, or the variety of defor-
mation modes considered is not large enough to separate the influence of the response from each parameter.
The latter option is not important here since rather different deformations are simulated. An observation made
for several models is clearly corroborated by the appendant correlation matrices: stress–stretch relations which
are not different or complex enough to discriminate the deformation modes with a single parameter set are
frequently revealing closely correlated parameters.
Ogden 1 Ogden 2 Ogden 3 Mooney–Rivlin Isihara
μ 1 α1 μ 1 α1 μ2 α2 μ 1 α1 μ2 α2 μ3 α3 c10 c01 c10 c20 c01
μ1 1 0.993 μ1 1 0.982 0.702 0.695 μ1 1 0.993 0.783 0.777 0.815 0.923 c10 1 −0.378 c10 1 0.384 0.384
α1 1 α1 1 0.789 0.782 α1 1 0.834 0.828 0.744 0.870 c01 1 c20 1 1
μ2 1 1 μ2 1 1 0.445 0.573 c01 1
α2 1 α2 1 0.441 0.568
μ3 1 0.972
α3 1
Gent–Thomas Swanson 1 Swanson 2

c1 c2 A 1 α1 B1 β1 A 1 α1 B1 β1 A2 α2 B2 β2
c1 1 0.554 A1 1 0.996 0.313 −0.349 A1 1 0.984 0.375 −0.548 0.784 0.772 0.411 −0.347
c2 1 α1 1 0.262 −0.297 α1 1 0.252 −0.439 0.857 0.845 0.354 −0.311
B1 1 −0.743 B1 1 0.670 0.106 0.103 0.342 −0.197
β1 1 β1 1 −0.165 −0.158 −0.759 0.631
A2 1 1 0.094 −0.073
α2 1 0.085 −0.065
B2 1 −0.982
β2 1
Yeoh Arruda–Boyce Gent Yeoh–Fleming Carroll

c1 c2 c3 μ N μ Jm A B C Im a b c
c1 1 −0.916 0.834 μ 1 −0.946 μ 1 −0.923 A 1 −0.869 0.850 −0.606 a 1 0.786 0.528
c2 1 −0.980 N 1 Jm 1 B 1 −0.983 0.858 b 1 0.175
c3 1 C 1 −0.924 c 1
Im 1
3-chains 8-chains Unit sphere

μ N μ N μ N
μ 1 −0.899 μ 1 −0.903 μ 1 −0.813
N 1 N 1 N 1
References
1. Treloar, L.R.G.: Stress-strain data for vulcanised rubber under various types of deformation. Trans. Faraday Soc. 40,
59–70 (1944)
2. Arruda, E.M., Boyce, M.C.: A three-dimensional constitutive model for the large stretch behavior of rubber elastic materi-
als. J. Mech. Phys. Solids 41, 389–412 (1993)
3. Boyce, M.C., Arruda, E.M.: Constitutive models of rubber elasticity: a review. Rubber Chem. Technol. 73, 504–523 (2000)
4. Lulei, F.: Mikromechanisch motivierte Modelle zur Beschreibung finiter Deformationen gummiartiger Polymere: Physika-
lische Modellbildung und numerische Simulation. Ph.D. Thesis, University of Stuttgart, Germany (2002)
5. Miehe, C., Göktepe, S., Lulei, F.: A micro-approach to rubber-like materials - Part I: The non-affine micro-sphere model of
rubber elasticity. J. Mech. Phys. Solids 52, 2617–2660 (2004)
6. Göktepe, S.: Micro-macro approaches to rubbery and glassy polymers: predictive micromechanically-based models and
simulations. Ph.D. Thesis, University of Stuttgart, Germany (2007)
7. Ogden, R.W.: Large deformation isotropic elasticity—on the correlation of theory and experiment for incompressible rub-
berlike solids. Proc. R. Soc. Lond. A Math. Phys. Sci. 326, 565–584 (1972)
8. Mooney, M.: A theory of large elastic deformation. J. Appl. Phys. 11, 582–596 (1940)
9. Rivlin, R.S.: Large elastic deformations of isotropic materials. IV. Further developments of the general theory. Philos. Trans.
R. Soc. A 241, 379–397 (1948)
10. Rivlin, R.S.: Large elastic deformations of isotropic materials. V. The problem of flexure. Proc. R. Soc. Lond. A Math. Phys.
Sci. 195, 463–473 (1949)
11. Rivlin, R.S.: Large elastic deformations of isotropic materials. VI. Further results in the theory of torsion, shear and
flexure. Philos. Trans. R. Soc. A 242, 173–195 (1949)
12. Yeoh, O.H.: Some forms of the strain energy function for rubber. Rubber Chem. Technol. 66, 754–771 (1993)
13. Yeoh, O.H., Fleming, P.D.: A new attempt to reconcile the statistical and phenomenological theories of rubber elasticity.
J. Polym. Sci. Polym. Phys. 35, 1919–1931 (1997)
14. Wu, P.D., van der Giessen, E.: On improved 3-D non-Gaussian network models for rubber elasticity. Mech. Res. Com-
mun. 19, 427–433 (1992)
15. Beda, T., Chevalier, Y.: Hybrid continuum model for large elastic deformation of rubber. J. Appl. Phys. 94, 2701–2706
(2003)
16. Marckmann, G., Verron, E.: Comparison of hyperelastic models for rubber-like materials. Rubber Chem. Technol. 79,
835–858 (2006)
17. Kaliske, M., Heinrich, G.: An extended tube-model for rubber elasticity: Statistical mechanical theory and finite element
implementation. Rubber Chem. Technol. 72, 602–632 (1999)
18. Shariff, M.H.B.M.: Strain energy function for filled and unfilled rubber-like material. Rubber Chem. Technol. 73, 1–18 (2000)
19. Bergström, J.S., Boyce, M.C.: Constitutive modeling of the large strain time-dependent behavior of elastomers. J. Mech.
Phys. Solids 46, 931–954 (1998)
20. Seibert, D.J., Schöche, N.: Direct comparison of some recent rubber elasticity models. Rubber Chem. Technol. 73,
366–384 (2000)
21. Rivlin, R.S., Saunders, D.W.: Large elastic deformations of isotropic materials. VII. Experiments on the deformation of
rubber. Philos. Trans. R. Soc. A 243, 251–288 (1951)
22. Hartmann, S.: Parameter estimation of hyperelasticity relations of generalized polynomial-type with constraint condi-
tions. Int. J. Solids Struct. 38, 7999–8018 (2001)
23. Hartmann, S., Neff, P.: Polyconvexity of generalized polynomial-type hyperelastic strain energy functions for near-incom-
pressibility. Int. J. Solids Struct. 40, 2767–2791 (2003)
24. Hartmann, S.: Numerical studies on the identification of the material parameters of Rivlin’s hyperelasticity using tension-
torsion tests. Acta Mech. 148, 129–155 (2001)
25. Hartmann, S., Tschöpe, T., Schreiber, L., Haupt, P.: Finite deformations of a carbon black-filled rubber. Experiment, optical
measurement and material parameter identification using finite elements. Eur. J. Mech. A-Solid 22, 309–324 (2003)
26. Haupt, P., Sedlan, K.: Viscoplasticity of elastomeric materials: experimental facts and constitutive modelling. Arch. Appl.
Mech. 71, 89–109 (2001)
27. Attard, M.M., Hunt, G.W.: Hyperelastic constitutive modeling under finite strain. Int. J. Solids Struct. 41, 5327–5350
(2004)
28. Valanis, K.S., Landel, R.F.: The strain-energy function of a hyperelastic material in terms of the extension ratios. J. Appl.
Phys. 7, 2997–3002 (1967)
29. Vangerko, H., Treloar, L.R.G.: The inflation and extension of rubber tube for biaxial strain studies. J. Phys. D Appl.
Phys. 11, 1969–1978 (1978)
30. Chagnon, G., Marckmann, G., Verron, E.: A comparison of the Hart-Smith model with Arruda–Boyce and Gent formulations
for rubber elasticity. Rubber Chem. Technol. 77, 724–735 (2004)
31. Hart-Smith, L.J.: Elasticity parameters for finite deformations of rubber like materials. Z. Angew. Math. Phys. 17,
608–626 (1966)
32. Gent, A.N.: A new constitutive relation for rubber. Rubber Chem. Technol. 69, 59–61 (1996)
33. Boyce, M.C.: Direct comparison of the Gent and Arruda–Boyce constitutive models of rubber elasticity. Rubber Chem.
Technol. 69, 781–785 (1996)
34. Horgan, C.O., Saccomandi, G.: Simple torsion of isotropic, hyperelastic, incompressible materials with limiting chain
extensibility. J. Elast. 56, 159–170 (1999)
35. Horgan, C.O., Saccomandi, G.: A molecular-statistical basis for the Gent constitutive model of rubber elasticity. J.
Elast. 68, 167–176 (2002)
36. Currie, P.K.: Comparison of incompressible elastic strain energy functions over the attainable region of invariant space. Math.
Mech. Solids 10, 559–574 (2005)
37. Haines, D.W., Wilson, W.D.: Strain-energy density function for rubberlike materials. J. Mech. Phys. Solids 27, 345–
360 (1979)
38. Holzapfel, G.A.: Nonlinear Solid Mechanics. Wiley, Chichester (2001)
39. Miehe, C.: Aspects of the formulation and finite element implementation of large strain isotropic elasticity. Int. J. Numer.
Methods Eng. 37, 1981–2004 (1994)
40. Liu, C.H., Hofstetter, G., Mang, H.A.: 3D finite element analysis of rubber-like materials at finite strains. Eng.
Comput. 11, 111–128 (1994)
41. Kaliske, M., Rothert, H.: On the finite element implementation of rubber-like materials at finite strains. Eng. Comput. 14,
216–232 (1997)
42. Heinrich, G., Kaliske, M.: Theoretical and numerical formulation of molecular based constitutive tube-model of rubber
elasticity. Comput. Theor. Polym. Sci. 7, 227–241 (1997)
43. Isihara, A., Hashitsume, N., Tatibana, M.: Statistical theory of rubber-like elasticity. IV. (Two-dimensional stretching). J.
Chem. Phys. 19, 1508–1512 (1951)
44. Gent, A.N., Thomas, A.G.: Forms for the stored (strain) energy function for vulcanized rubber. J. Polym. Sci. 28,
625–628 (1958)
45. Swanson, S.R.: A constitutive model for high elongation elastic materials. J. Eng. Mater. Trans. ASME 107, 110–115 (1985)
46. Yeoh, O.H.: Characterization of elastic properties of carbon-black-filled rubber vulcanizates. Rubber Chem.
Technol. 63, 792–805 (1990)
47. Carroll, M.M.: A strain energy function for vulcanized rubbers. J. Elast. 103, 173–187 (2011)
48. James, H.M., Guth, E.: Theory of the elastic properties of rubber. J. Chem. Phys. 11, 455–481 (1943)
49. Wang, M.C., Guth, E.: Statistical theory of networks of non-Gaussian flexible chains. J. Chem. Phys. 20, 1144–1157 (1952)
50. Flory, P.J.: Thermodynamic relations for high elastic materials. Trans. Faraday Soc. 57, 829–838 (1961)
51. Wriggers, P.: Nonlinear Finite Element Methods. Springer, Berlin (2008)
52. Treloar, L.R.G.: The Physics of Rubber Elasticity. Clarendon Press, Oxford (1975)
53. Ogden, R.W.: Non-linear Elastic Deformations. Dover, New York (1997)
54. Miehe, C.: Computation of isotropic tensor functions. Commun. Numer. Methods Eng. 9, 889–896 (1993)
55. Reese, S., Govindjee, S.: A theory of finite viscoelasticity and numerical aspects. Int. J. Solids Struct. 35, 3455–3482 (1998)
56. Simo, J.C., Taylor, R.L.: Quasi-incompressible finite elasticity in principal stretches. Continuum basis and numerical algo-
rithms. Comput. Methods Appl. Mech. Eng. 85, 273–310 (1991)
57. Başar, Y., Weichert, D.: Nonlinear Continuum Mechanics of Solids. Fundamental Mathematical and Physical Con-
cepts. Springer, Berlin (2000)
58. Bonet, J., Wood, R.D.: Nonlinear Continuum Mechanics for Finite Element Analysis. Cambridge University Press, Cam-
bridge (1997)
59. Böl, M.: Numerische Simulation von Polymernetzwerken mit Hilfe der Finite-Elemente-Methode. Ph.D. thesis, Ruhr-Uni-
versity Bochum, Germany (2005)
60. Thomas, A.G.: The departures from the statistical theory of rubber elasticity. Trans. Faraday Soc. 51, 569–582 (1955)
61. Swanson, S.R., Christensen, L.W., Ensign, M.: Large deformation finite element calculations for slightly compressible
hyperelastic materials. Comput. Struct. 21, 81–88 (1985)
62. Biderman, V.L.: Calculation of Rubber Parts (in Russian). Rascheti na Prochnost, Moscow (1958)
63. Kratky, O., Porod, G.: Röntgenuntersuchung gelöster Fadenmoleküle. Recl. Trav. Chim. Pay-B 68, 1106–1122 (1949)
64. Cohen, A.: A Padé approximant to the inverse Langevin function. Rheol. Acta 30, 270–273 (1991)
65. Kuhn, W.: Über die Gestalt fadenförmiger Moleküle in Lösungen. Kolloid Z. 68, 2–15 (1942)
66. Bažant, Z.P., Oh, B.H.: Efficient numerical integration on the surface of a sphere. Z. Angew. Math. Mech. 66, 37–49 (1986)
67. Nah, C., Lee, G.B., Lim, J.Y., SenGupta, R., Gent, A.N.: Problems in determining the elastic strain energy function for
rubber. Int. J. Non-Linear Mech. 45, 232–235 (2010)
68. Fossum, A.F.: Parameter estimation for an internal variable model using nonlinear optimization and analytical/numerical
response sensitivities. J. Eng. Mater. Trans. ASME 119, 337–345 (1997)
69. Ekh, M.: Thermo-elastic-viscoplastic modeling of IN792. J. Mech. Behav. Mater. 12, 359–387 (2001)

Steinmann2012 Article HyperelasticModelsForRubber-li

Uploaded by

Document Information

Copyright

Available Formats

Share this document

Share or Embed Document

Sharing Options

Did you find this document useful?

Is this content inappropriate?

Copyright:

Available Formats

Steinmann2012 Article HyperelasticModelsForRubber-li

Uploaded by

Copyright:

Available Formats

Arch Appl Mech (2012) 82:1183–1217

Paul Steinmann · Mokarram Hossain · Gunnar Possart

Hyperelastic models for rubber-like materials: consistent

1 Introduction and outline

P. Steinmann (B) · M. Hossain · G. Possart

2 Continuum mechanical preliminaries

2.1 Stress tensor and tangent operator derivation

Since C̄ = J −2/3 C, the invariants of C̄ and C are related by

3 Phenomenological models: stress tensors, tangent operators and performance

3.1.1 Uniaxial tension (UT)

3.1.2 Equibiaxial tension (ET)

3.1.3 Pure shear (PS)

3.2 Models formulated in principal stretches: Ogden models (1972)

UT fitted, ET/PS simulated ET fitted, UT/PS simulated PS fitted, UT/ET simulated

UT fitted, ET/PS simulated ET fitted, UT/PS simulated PS fitted, UT/ET simulated

UT fitted, ET/PS simulated ET fitted, UT/PS simulated PS fitted, UT/ET simulated

3.3 Models formulated in strain invariants

3.3.1 Neo-Hooke model (1943)

UT fitted, ET/PS simulated ET fitted, UT/PS simulated PS fitted, UT/ET simulated

3.3.2 Mooney–Rivlin model (1940)

UT fitted, ET/PS simulated ET fitted, UT/PS simulated PS fitted, UT/ET simulated

6 0.834 = Error(UT−fit) 0.052 = Error(ET−fit) 0.045 = Error(PS−fit)

3.3.3 Isihara model (1951)

UT fitted, ET/PS simulated ET fitted, UT/PS simulated PS fitted, UT/ET simulated

3.3.4 Gent–Thomas model (1958)

UT fitted, ET/PS simulated ET fitted, UT/PS simulated PS fitted, UT/ET simulated

Fig. 8 Performance of the Gent–Thomas model on Treloar’s data. Error(UT/ET/PS-sim) = 0.815/0.188/0.528

3.3.5 Swanson model (1985)

UT fitted, ET/PS simulated ET fitted, UT/PS simulated PS fitted, UT/ET simulated

3.3.6 Yeoh model (1990)

UT fitted, ET/PS simulated ET fitted, UT/PS simulated PS fitted, UT/ET simulated

c1UT = 0.1634 c1ET = 0.2059 c1PS = 0.1776 c1 = 0.1834

UT fitted, ET/PS simulated ET fitted, UT/PS simulated PS fitted, UT/ET simulated

3.3.7 Arruda–Boyce model (1993) (invariant form)

UT fitted, ET/PS simulated ET fitted, UT/PS simulated PS fitted, UT/ET simulated

3.3.8 Gent model (1996)

UT fitted, ET/PS simulated ET fitted, UT/PS simulated PS fitted, UT/ET simulated

3.3.9 Yeoh–Fleming model (1997)

UT fitted, ET/PS simulated ET fitted, UT/PS simulated PS fitted, UT/ET simulated

AUT = 0.0517 AET = 0.0467 APS = 0.0512 A = 0.0601

3.3.10 Carroll model (2011)

Carroll’s parameters ET fitted unconstrained (c = − 0.1099), UT/PS simulated

UT fitted, ET/PS simulated ET fitted, UT/PS simulated PS fitted, UT/ET simulated

Fig. 16 Performance of the Carroll model on Treloar’s data. Error(UT/ET/PS-sim) = 0.083/0.024/0.02

Fitting to Treloar’s data yields the following material parameter sets:

a UT = 0.1481 MPa a ET = 0.1988 MPa a PS = 0.1297 MPa a = 0.1433 MPa

4 Micro-mechanical models: stress tensors, tangent operators and performance

4.1 Three chain model (1943)

UT fitted, ET/PS simulated ET fitted, UT/PS simulated PS fitted, UT/ET simulated

4.2 Eight chain model (1993)

If Langevin chain behaviour is assumed, evaluation of (105) provides

and fitting to Treloar’s data provides the following parameter sets:

UT fitted, ET/PS simulated ET fitted, UT/PS simulated PS fitted, UT/ET simulated

4.3 Unit sphere model (2002)

UT fitted, ET/PS simulated ET fitted, UT/PS simulated PS fitted, UT/ET simulated

Appendix A: Treloar Data

Appendix B: A local check for quadratic convergence

that is, one has to solve

function [ni] = check(C_0,S_inf,it)

Appendix C: Correlation matrices of parameter sensitivities

Gent–Thomas Swanson 1 Swanson 2

Yeoh Arruda–Boyce Gent Yeoh–Fleming Carroll

3-chains 8-chains Unit sphere

You might also like