We recall that a homogeneous hyperelastic model is described by a strain-energy function
\(W(\mathbf{F})\) that depends on the deformation gradient tensor,
\(\mathbf{F}\), with respect to a fixed reference configuration, and is characterised by a set of
deterministic model parameters [
14,
39,
56]. In contrast, a stochastic homogeneous hyperelastic model is defined by a stochastic strain-energy function, for which the model parameters are
random variables that satisfy standard probability laws [
36,
51‐
53]. In this case, each model parameter is described in terms of its
mean value and its
variance, which contains information about the range of values about the mean value. While it is rarely possible if ever to obtain complete information about a random quantity in an elastic sample of material, the partial information provided by the mean value and the variance is the most commonly used in many practical applications [
8,
20,
29]. Here, we combine finite elasticity and information theory, and rely on the following general hypotheses [
35‐
37]:
(A1)
Material objectivity: The principle of material objectivity (frame indifference) states that constitutive equations must be invariant under changes of frame of reference. It requires that the scalar strain-energy function, \(W=W(\textbf{F})\), depending only on the deformation gradient \(\textbf{F}\), with respect to the reference configuration, is unaffected by a superimposed rigid-body transformation (which involves a change of position) after deformation, i.e., \(W(\textbf{R}^{T}\textbf{F})=W(\textbf{F})\), where \(\textbf{R}\in SO(3)\) is a proper orthogonal tensor (rotation). Material objectivity is guaranteed by considering strain-energy functions defined in terms of invariants.
(A2)
Material isotropy: The principle of isotropy requires that the strain-energy function is unaffected by a superimposed rigid-body transformation prior to deformation, i.e., \(W(\textbf{F}\textbf{Q})=W( \textbf{F})\), where \(\textbf{Q}\in SO(3)\). For isotropic materials, the strain-energy function is a symmetric function of the principal stretches \(\{\lambda _{i}\}_{i=1,2,3}\) of \(\textbf{F}\), i.e., \(W(\textbf{F})=\mathcal{W}(\lambda _{1},\lambda _{2},\lambda _{3})\).
(A3)
Baker-Ericksen inequalities: In addition to the fundamental principles of objectivity and material symmetry, in order for the behaviour of a hyperelastic material to be physically realistic, there are some universally accepted constraints on the constitutive equations. Specifically, for a hyperelastic body, the Baker-Ericksen (BE) inequalities, which state that
the greater principal (Cauchy) stress occurs in the direction of the greater principal stretch, are [
5]:
$$ (T_{i}-T_{j} ) (\lambda _{i}- \lambda _{j} )>0 \quad \mbox{if } \lambda _{i}\neq \lambda _{j},\quad i,j=1,2,3, $$
(1)
where
\(\{\lambda _{i}\}_{i=1,2,3}\) and
\(\{T_{i}\}_{i=1,2,3}\) denote the principal stretches and the principal Cauchy stresses, respectively, and “≥” replaces the strict inequality “>” if any two principal stretches are equal [
5,
28]. The BE inequalities (
1) take the equivalent form
$$ \biggl(\lambda _{1}\frac{\partial \mathcal{W}}{\partial \lambda _{1}}- \lambda _{2}\frac{\partial \mathcal{W}}{\partial \lambda _{2}} \biggr) (\lambda _{1}-\lambda _{2} )>0\quad \mbox{if } \lambda _{i}\neq \lambda _{j},\quad i,j=1,2,3, $$
(2)
where the strict inequality “>” is replaced by “≥” if any two principal stretches are equal.
(A4)
Finite mean and variance for the random shear modulus: We assume that, for any given deformation, the random shear modulus,
\(\mu \), and its inverse,
\(1/\mu \), are second-order random variables, i.e., they have finite mean value and finite variance [
51‐
53].
While (A4) contains physically realistic expectations on the random shear modulus, which will be drawn from a probability distribution, assumptions (A1)–(A3) are well-known principles in isotropic finite elasticity [
14,
39,
56].
Specifically, we focus our attention on homogeneous incompressible hyperelastic materials characterised by the following stochastic strain-energy function [
36,
51,
53],
$$ \mathcal{W}(\lambda _{1},\lambda _{2},\lambda _{3})=\frac{\mu _{1}}{2m ^{2}} \bigl(\lambda _{1}^{2m}+ \lambda _{2}^{2m}+\lambda _{3}^{2m}-3 \bigr) +\frac{\mu _{2}}{2n^{2}} \bigl(\lambda _{1}^{2n}+\lambda _{2}^{2n}+ \lambda _{3}^{2n}-3 \bigr), $$
(3)
where
\(m\) and
\(n\) are deterministic constants, and
\(\mu _{1}\) and
\(\mu _{2}\) are random variables following given probability distributions. In the deterministic elastic case,
\(\mu _{1}\),
\(\mu _{2}\),
\(m\) and
\(n\) are constants, and the model contains, as special cases, the neo-Hookean model, the Mooney-Rivlin model, and the one- and two-term Ogden models. In both the deterministic elastic and stochastic cases, the shear modulus for infinitesimal deformations of these models is defined as
\(\mu =\mu _{1}+\mu _{2}\) [
34,
36]. Note that we could easily extend our description to include
\(m\) and
\(n\) as stochastic variables as well. However, increasing the complexity in this way is not relevant for the present discussion. Including additional sources of randomness is an avenue of future research.
For the stochastic materials described by (
3), condition (A4) is guaranteed by setting the following mathematical expectations [
36,
37,
51‐
53]:
$$\begin{aligned} \textstyle\begin{cases} E [\mu ]=\underline{\mu }>0,&\\ E [\log \ \mu ]=\nu ,&\mbox{such that $|\nu |< +\infty $}. \end{cases}\displaystyle \end{aligned}$$
(4)
Then, under the constraints (
4), the random shear modulus,
\(\mu \), with mean value
\(\underline{\mu }\) and standard deviation
\(\|\mu \|=\sqrt{\operatorname{Var}[\mu ]}\), defined as the square root of the variance,
\(\operatorname{Var}[\mu ]\), follows a Gamma probability distribution [
48,
49], with hyperparameters
\(\rho _{1}>0\) and
\(\rho _{2}>0\) satisfying
$$ \underline{\mu }=\rho _{1}\rho _{2}, \qquad \|\mu \|=\sqrt{\rho _{1}}\rho_{2}. $$
(5)
The corresponding probability density function takes the form [
1,
25]
$$ g(\mu ;\rho _{1},\rho _{2})= \frac{\mu ^{\rho _{1}-1}e^{-\mu /\rho _{2}}}{ \rho _{2}^{\rho _{1}}\varGamma (\rho _{1})}, \quad \mbox{for}\ \mu >0\ \mbox{and}\ \rho _{1}, \rho _{2}>0, $$
(6)
where
\(\varGamma :\mathbb{R}^{*}_{+}\to \mathbb{R}\) is the complete Gamma function
$$ \varGamma (z)= \int _{0}^{+\infty }t^{z-1}e^{-t} \mathrm{d}t. $$
(7)
For technical convenience, we set a finite constant value
\(b>-\infty \), such that
\(\mu _{i}>b\),
\(i=1,2\) (e.g.,
\(b=0\) if
\(\mu _{1}>0\) and
\(\mu _{2}>0\), but
\(b\) is not unique in general), and introduce the auxiliary random variable [
36]
$$ R_{1}=\frac{\mu _{1}-b}{\mu -2b}, $$
(8)
such that
\(0< R_{1}<1\). Consequently, we can equivalently express the random model parameters
\(\mu _{1}\) and
\(\mu _{2}\) as follows,
$$ \mu _{1}=R_{1}(\mu -2b)+b, \qquad \mu _{2}=\mu -\mu _{1}=(1-R_{1}) (\mu -2b)+b. $$
(9)
It is reasonable to assume [
36,
51‐
53]
$$\begin{aligned} \textstyle\begin{cases} E [\log \ R_{1} ]=\nu _{1},& \mbox{such that $|\nu _{1}|< +\infty $}, \\ E [\log (1-R_{1}) ]=\nu _{2},& \mbox{such that $|\nu _{2}|< +\infty $}, \end{cases}\displaystyle \end{aligned}$$
(10)
in which case, the random variable
\(R_{1}\), with mean value
\(\underline{R}_{1}\) and variance
\(\operatorname{Var}[R_{1}]\), follows a standard Beta distribution [
1,
25], with hyperparameters
\(\xi _{1}>0\) and
\(\xi _{2}>0\) satisfying
$$ \underline{R}_{1}=\frac{\xi _{1}}{\xi _{1}+\xi _{2}}, \qquad \operatorname{Var}[R_{1}]=\frac{\xi _{1}\xi _{2}}{ (\xi _{1}+\xi _{2} ) ^{2} (\xi _{1}+\xi _{2}+1 )}. $$
(11)
The associated probability density function is
$$ \beta (r;\xi _{1},\xi _{2})= \frac{r^{\xi _{1}-1}(1-r)^{\xi _{2}-1}}{B(\xi _{1},\xi _{2})}, \quad \mbox{for}\ r\in (0,1)\ \mbox{and}\ \xi _{1}, \xi _{2}>0, $$
(12)
where
\(B:\mathbb{R}^{*}_{+}\times \mathbb{R}^{*}_{+}\to \mathbb{R}\) is the Beta function
$$ B(x,y)= \int _{0}^{1}t^{x-1}(1-t)^{y-1}\mathrm{d}t. $$
(13)
Thus, for the random coefficients given by (
9), the corresponding mean values take the form,
$$ \underline{\mu }_{1}=\underline{R}_{1}( \underline{\mu }-2b)+b, \qquad \underline{\mu }_{2}=\underline{\mu }- \underline{\mu }_{1}=(1-\underline{R}_{1}) (\underline{\mu }-2b)+b, $$
(14)
and the variances and covariance are, respectively,
$$\begin{aligned} &\operatorname{Var} [\mu _{1} ] =(\underline{\mu }-2b)^{2} \operatorname{Var}[R_{1}]+(\underline{R}_{1})^{2} \operatorname{Var}[\mu ]+\operatorname{Var}[ \mu ]\operatorname{Var}[R_{1}], \end{aligned}$$
(15)
$$\begin{aligned} &\operatorname{Var} [\mu _{2} ] =(\underline{\mu }-2b)^{2} \operatorname{Var}[R_{1}]+(1-\underline{R}_{1})^{2} \operatorname{Var}[\mu ]+ \operatorname{Var}[\mu ]\operatorname{Var}[R_{1}], \end{aligned}$$
(16)
$$\begin{aligned} &\operatorname{Cov}[\mu _{1},\mu _{2}]=\frac{1}{2} \bigl( \operatorname{Var}[\mu ]- \operatorname{Var}[\mu _{1}]-\operatorname{Var}[\mu _{2}] \bigr). \end{aligned}$$
(17)
It should be noted that the random variables
\(\mu \) and
\(R_{1}\) are independent, depending on parameters
\((\rho _{1},\rho _{2})\) and
\((\zeta _{1},\zeta _{2})\), respectively, which are derived by fitting distributions to given data. However,
\(\mu _{1}\) and
\(\mu _{2}\) are dependent variables as they both require
\((\mu ,R_{1})\) to be defined. Explicit derivations of the probability distributions for the random parameters when stochastic isotropic hyperelastic models are calibrated to experimental data are presented in [
36,
53].