Available online at www.sciencedirect.

com
ScienceDirect

Comput. Methods Appl. Mech. Engrg. 360 (2020) 112693

www.elsevier.com/locate/cma

Bayesian inference of non-linear multiscale model parameters

accelerated by a Deep Neural Network
Ling Wua ,∗, Kepa Zuluetab , Zoltan Majorc , Aitor Arriagab , Ludovic Noelsa
a University of Liege, Computational & Multiscale Mechanics of Materials, Allée de la découverte 9, B-4000 Liège, Belgium
b Leartiker Polymer R&D, Xemein Etorbidea 12A, 48270 Markina-Xemein, Bizkaia, Spain
c Johannes Kepler University Linz, Institute of Polymer Product Engineering, Altenbergerstrasse 69, 4040 Linz, Austria

Received 17 March 2019; received in revised form 24 June 2019; accepted 9 October 2019
Available online 31 October 2019

Abstract
We develop a Bayesian Inference (BI) of the parameters of a non-linear multiscale model and of its material constitutive laws
using experimental composite coupon tests as observation data. In particular we consider non-aligned Short Fibers Reinforced
Polymer (SFRP) as a composite material system and Mean-Field Homogenization (MFH) as a multiscale model. Although
MFH is computationally efficient, when considering non-aligned inclusions, the evaluation cost of a non-linear response for a
given set of model and material parameters remains too prohibitive to be coupled with the sampling process required by the
BI. Therefore, a Neural-Network (NNW) is first trained using the MFH model, and is then used as a surrogate model during
the BI process, making the identification process affordable.
⃝c 2019 Elsevier B.V. All rights reserved.

Keywords: Multiscale; Composites; Bayesian inference; Neural Network; Non-linear

1. Introduction
Short Fibers Reinforced Polymer (SFRP) composites are nowadays commonly used in several industrial
applications, increasing the need for computationally efficient modeling tools. Because of the heterogeneous nature
of the material, multiscale methods are favored [1], in particular in the non-linear range, in order to capture the
effects of the micro-structure geometrical parameters, such as the inclusions aspect ratio, orientation and spatial
distributions, and of the micro-constituents non-linear material responses.
Mean-Field Homogenization (MFH) is an efficient semi-analytical multiscale method which extends the Eshelby
single inclusion solution [2] to multiple-inclusion interactions, such as in the Mori–Tanaka (M-T) scheme [3]. In
the non-linear range, MFH revolves around the definition of a Linear Comparison Composite (LCC) [4–6] as a
virtual heterogeneous linear material system, allowing to extend linear theories while keeping a good to excellent
accuracy. Two-step MFH has also been developed in the context of non-aligned inclusions using an orientation
distribution function (ODF) to describe their misalignment. Homogenization is thus performed in two stages, first
∗ Corresponding author.
E-mail addresses: L.Wu@ulg.ac.be (L. Wu), kzulueta@leartiker.com (K. Zulueta), Zoltan.Major@jku.at (Z. Major),
aarriaga@leartiker.com (A. Arriaga), L.Noels@ulg.ac.be (L. Noels).

https://doi.org/10.1016/j.cma.2019.112693
0045-7825/⃝ c 2019 Elsevier B.V. All rights reserved.
2 L. Wu, K. Zulueta, Z. Major et al. / Computer Methods in Applied Mechanics and Engineering 360 (2020) 112693

on pseudo-grains of aligned inclusions, and then by weighting the pseudo-grain responses with the ODF [7,8]. Since
the manufacturing process induces a variation of the ODF along the sample thickness, this 2-step homogenization
has to be repeated on different layers to account for the so-called skin-core effect.
One of the difficulty with multiscale methods in general, and MFH in particular, is to identify the parameters
defining the microstructure such as fibers aspect ratio distribution, volume fraction, ODF, but also the material laws
parameters modeling the phases non-linear responses. On the one hand, fibers aspect ratio distribution, ODF, or again
volume fraction, can be experimentally measured [9] or predicted by a process numerical simulation [9]. However,
although the fiber aspect ratio distribution can be experimentally measured, when using the ODF in the context of
a two-step MFH, it is convenient to consider a unique “effective” aspect ratio, which has yet to be identified. On
the other hand, although fibers properties of SFRP are usually known, the matrix material properties are sensitive to
the process conditions and cannot always be directly identified because of the difficulty to produce samples whose
material response is exactly the same as the polymeric matrix of the composite material; the parameters of the
polymeric material models are thus generally obtained through an inverse identification process from composite
coupon tests.
Therefore, some micro-structure geometrical parameters, such as the effective aspect ratio, and some phases
material parameters, such as the matrix model parameters, should be inferred from composite experimental
responses. However, because of the number of parameters arising in the non-linear range, this identification requires
several loading conditions to be performed, and a unique set of parameters cannot reproduce all the experimental
tests because of the model limitations. Besides, the composite responses are inevitably entailed by experimental
errors. These difficulties can be circumvented by considering a Bayesian Inference (BI) [10] for which uncertainties
in the inferred model parameters arise from the identification process itself under the form of a so-called posterior
Probability Density Function (PDF). This posterior PDF of the parameters set is the correction of the initial belief
one has on the parameters set distribution, or prior distribution, by a likelihood function evaluated using different
observation data, i.e. here the experimental results.
BI was extensively applied to identify the parameters of either, possibly non-linear, homogeneous material
models [11–13, e.g.] or, generally linear, homogenized composite material models [14,15, e.g]. The inference of
multi-scale model parameters is more difficult because of its inherent computational cost making the evaluation of
the likelihood time consuming when considering a random sampling process. Although in [16] the authors have
inferred the two-step MFH parameters of SFRP, this was limited to the linear range because of the computational
cost bottleneck: when considering a two-step homogenization as a non-linear multiscale model, one simple analysis
needs around one minute of computation but hundreds of thousands of simulations are required for the sampling
process. In order to make BI computationally affordable in the non-linear range, in this work, the two-step MFH
model is substituted by a surrogate model during the random sampling process. The surrogate model can be any kind
of feasible high dimensional nonlinear mapping. Depending on the features of the input and output, typical surrogate
models can be constructed by linear combinations of well chosen high dimensional non-linear base functions.
However, a certain experience is required to choose the proper dimension of the function base and the order of the
non-linear functions. Neural Networks (NNW) can theoretically carry out any high dimensional non-linear mapping
if the NNW is designed with enough hidden layers and enough neurons on them. With the help of open NNW library,
the experience required to build a high dimensional non-linear surrogate model is minimized, which offers a more
user-friendly and general solution. Artificial Neural Networks (ANNW) have been used in the literature to reproduce
the homogenized behavior predicted by computational homogenization methods, either by approximating the strain
energy density surface [17,18, e.g.] or the stress–strain responses [19,20, e.g.]. The latter approach is chosen in the
paper, and a NNW is trained by the two-step MFH for different sets of micro-structure geometrical and material
parameters, and for different loading directions. The likelihood function is then constructed by considering Gaussian
noise [11–13] as an error function [15] evaluated from the experimental observation on 40% of weight GF reinforced
PA06 (PA06-GF40) coupon tests. The BI is then conducted using a Metropolis–Hastings (MCMC) random walk
during which the likelihood is evaluated using the NNW as surrogate.
The organization of the paper is as follows. The non-linear two-step MFH model is described in Section 2.
Section 3 details the construction of the NNW surrogate model and Section 4 summarizes the experimental tests
conducted on PA06-GF40 composite coupons. Finally, the BI is presented in Section 5 and the results analysis in
Section 6. Conclusions are drawn in Section 7.
 L. Wu, K. Zulueta, Z. Major et al. / Computer Methods in Applied Mechanics and Engineering 360 (2020) 112693 3

2. Mean-field homogenization for non-aligned short fiber-reinforced composites

The finite element analysis of structures made of heterogeneous materials can be performed in a homogenization-
based multiscale approach, in which the relation between the macro-strains εM and stresses σM is transformed into a
relation between the averaged values of the local strain tensor εm and of the local stress tensor σm on a micro-scale
volume ω, with
εM = ⟨εm ⟩ω and σM = ⟨σm ⟩ω , (1)
where ⟨ f (xx )⟩ω = V1ω ω f (x)dV , with Vω the volume of ω.
∫

With a view to the homogenization of SFRP composites, the general equations for two-phase composites with
aligned uniform inclusions are first presented, the two-step homogenization method for non-aligned inclusions is
then summarized before being extended to account for skin-core effect. Finally, the material models used for the
different phases are summarized.

2.1. Mean-Field Homogenization (MFH) for two-phase composites

2.1.1. Mean-field equations for two-phase linear elastic materials

Considering a two-phase composite material with the respective volume fractions v0 +vI = 1, where the subscript
0 refers to the matrix and the subscript I to the aligned inclusions, the volume averages over the micro-scale volume
ω, Eqs. (1), can be explicitly expressed in terms of the volume averages over the two phases ω0 and ωI , as
εM = v0 ε0 + vI εI and σM = v0 σ0 + vI σI , (2)
where •i denotes the volume average over the phase ωi , i.e. ⟨•m ⟩ωi , for conciseness.
In the linear elastic range, the system of Eqs. (2) is completed by assuming a relationship between the average
strains of the different phases using a strain concentration tensor Bϵ , which is defined through the elastic tensors
Ciel in phase ωi , and reads
εI = Bϵ (I, Cel
0 , CI ) : ε 0 ,
el
(3)
where “I” represents the geometry of the inclusions. Using linear elastic constitutive laws σi = : εi , the set of
Ciel
Eqs. (2) and (3) can be rewritten in a general constitutive expression for linear elastic composites as
σM = Cel
M (I, C0 , CI , vI ) : εM ,
el el
(4)
ϵ ϵ −1
with CelM = vI CI : B (I, C0 , CI ) + v0 C0 : vI B (I, C0 , CI ) + v0 I , where (I)i jkl is the fourth-order identity
[ el el el el
] [ el el
]

tensor.

2.1.2. Mean-field equations for two-phase elasto-plastic materials

For the composites whose phases experience elasto-plastic deformations, MFH is carried out in an incremental
form through a so-called Linear Comparison Composite (LCC) [21,22]. The LCC is a virtual linear heterogeneous
material whose constituents behaviors are defined by virtual elastic operators matching the linearized behaviors of
the real composite material constituents at a given strain state. Therefore, by considering the virtual elastic operators
CLCC
0 of the matrix phase and CLCC
I of the inclusions phase, the MFH equations of a linear composite material can
be applied on the average incremental strains in the two phases; while the homogenized stress relation, Eq. (2)(a)
still holds, the homogenized strain relations, Eq. (2)(b) and Eq. (3), are rewritten
∆εM = v0 ∆ε0 + vI ∆εI , (5)
ϵ
∆εI = B (I, CLCC
0 , CLCC
I ) : ∆ε0 . (6)
Among the different linearization techniques developed in order to define the LCC, the incremental-secant
approach [6] is considered in this work. During a time increment [tn , tn+1 ], the composite material is first subjected
to a virtual elastic unloading from the configuration at time tn to reach a residual state so that σMres n = 0, where the
subscript “res” refers to the virtually unloaded state. Then, the composite material is loaded to the new configuration
at time tn+1 , see Fig. 1(a).
4 L. Wu, K. Zulueta, Z. Major et al. / Computer Methods in Applied Mechanics and Engineering 360 (2020) 112693

Fig. 1. Definition of the LCC in the incremental-secant method for elasto-plastic composites.

Since the virtual unloading is elastic, the residual state of the composite is fully determined by the linear elastic
MFH formula presented in Section 2.1.1. Then, from the residual state, εM res
n and σM n = 0, at tn , the secant
res
r
linearization of the non-linear composite material is carried out with the strain increment ∆εM defined such that
εMn+1 = εM
res
n
r
+ ∆εM , (7)
where εMn+1 is a known value of the macro-scale. Similarly, the phase strain increments ∆εir are defined such that
εin+1 = εires
n
+ ∆εir , (8)
as illustrated in Fig. 1(b). Finally, in each phase, an incremental-secant operator CiS is defined from the phase
residual stress–strain states (which do not necessarily vanish) such that
σin+1 = σires
n
+ ∆σir , and ∆σir = CiS : ∆εir . (9)
Therefore, the LCC is defined using the incremental-secant operators CiS , and the set of Eqs. (5)–(6) is thus
rewritten using CiS as LCC operator CiLCC . Finally, the incremental-secant form of Eq. (4) reads
σMn+1 = ∆σMr = CSM (I, CS0 , CSI , vI ) : ∆εMr
, with (10)
ϵ
] [ ϵ ]−1
CM = vI CI : B (I, C0 , CI ) + v0 C0 : vI B (I, CS0 , CSI ) + v0 I
S
.
[ S S S S
(11)
The resolution of the MFH equations (10)–(11) follows the iterative process detailed in [6]. We note for
completeness that the residual stress is canceled in the matrix phase, see details in [6].

2.2. Strain concentration tensor

The strain concentration tensors Bϵ (I, Cel el ϵ

0 , CI ) and B (I, C0 , CI ) are built upon assumptions. We consider the
S S

Mori–Tanaka [3] method (M-T) and Voigt model respectively for two-phase and multi-phase composites. In the
following expressions, Ci refers to either Ciel or CiS in the cases of linear elasticity and non-linear elasto-plasticity,
respectively.
• The M-T method assumes that the average strain in the matrix phase corresponds to the strain at infinity of
the single inclusion solution problem, i.e.
Bϵ (I, C0 , CI ) = {I + S : [(C0 )−1 : CI − I]}−1 , (12)
where the Eshelby tensor [2] S(I, C0 ) depends on the geometry of the inclusion “I” and on the material tensor
of the matrix phase C0 .
• The Voigt model assumes the same average strain in the different phases, i.e.
Bϵ = I . (13)
 L. Wu, K. Zulueta, Z. Major et al. / Computer Methods in Applied Mechanics and Engineering 360 (2020) 112693 5

Fig. 2. Two-step homogenization: the micro-scale volume element ω is decomposed into a set of pseudo-grains ω(k) homogenized using the
M-T assumption; the composite behavior is obtained using the Voigt assumption.

2.3. MFH for multi-phase composite materials

For short-fiber reinforced composites, the composite material cannot be treated as being two-phase in the MFH
process because of the misalignment and of the variation in aspect ratio of the fibers. When considering such a
material with inclusions having different orientations or shapes, a two-step homogenization strategy [7,8] can be
adopted.

2.3.1. Two-step homogenization

For a short-fiber reinforced composite material, whose volumes of matrix and short fibers are denoted by V0
and VI , respectively, the volume fraction of matrix reads v0 = V V+V 0
I
. First, it is assumed that all the short fibers
0
are straight and that the aspect ratio of a fiber is defined by its length l over its diameter d, ar = dl . Besides, it is
further assumed that all the fibers have the same aspect ratio Ar , which is an effective value and serves as a material
parameter. Finally, the fiber orientation is characterized by a unit vector p oriented along its axis. Therefore, the
two-step homogenization process of short fiber reinforced composites can be achieved as illustrated in Fig. 2 and
is summarized as follows
• The composite material is treated as an aggregate of pseudo-grains, ω(k) (k = 1, 2, . . .), in which the inclusions
“I(k) ” have the same aspect ratio Ar , and the same orientation defined by the direction p(k) . All these
pseudo-grains see the same volume fraction of matrix v0 ;
• The homogenization is first performed on each pseudo-grain ω(k) , with the set of Eqs. (5)–(6) rewritten as
⟨∆ε⟩ω(k) = v0 ⟨∆ε0 ⟩ω(k) + vI ⟨∆εI ⟩ω(k) , (14)
⟨σ ⟩ω(k) = v0 ⟨σ0 ⟩ω(k) + vI ⟨σI ⟩ω(k) , (15)
⟨∆ε I ⟩ω(k) = Bϵ I(k) , CS0 , CSI : ⟨∆ε0 ⟩ω(k) .
( )
(16)
• The homogenization on the aggregate of pseudo-grains, ω(k) (k = 1, 2, . . .) is then achieved using Voigt strain
concentration tensor (13), in which case the set of Eqs. (5)–(6) is rewritten as
⟨∆ε⟩ω(k) = ⟨∆ε⟩ω = ∆εM ∀ω(k) , (17)
with the homogenized stress evaluated by
∑
σM = ⟨σ ⟩ω = vω(k) ⟨σ ⟩ω(k) , (18)
k

where vω(k) is the volume fraction defined by the volume of fibers oriented along a direction p(k) , VIω(k) ,
VI (k)
over the total volume of fibers VI , e.g. vω(k) = ω
VI
. The values of vω(k) can be approximated using a fibers
Orientation Distribution Function.

2.3.2. Orientation distribution function (ODF)

For a collection of fibers, the complete description of their orientations is represented by a probability density
function π P ( p), also called Orientation Distribution
∮ Function (ODF), such that π P ( p) d p is the probability of a
fiber to be oriented between p and p + d p with π P ( p) d p=1. It is convenient to write the ODF in the spherical
coordinates as
∫ π ∮ 2π
π P ( p(θ, φ)) sin(θ ) dφ dθ = 1 , (19)
θ =0 φ=0
6 L. Wu, K. Zulueta, Z. Major et al. / Computer Methods in Applied Mechanics and Engineering 360 (2020) 112693

Fig. 3. Two-step homogenization with skin-core effect: the micro-scale volume element ω is decomposed into layers (l); each layer is
decomposed into pseudo-grains ω(k; l) and its homogenized behavior is obtained using the 2-step homogenization; the composite behavior is
eventually obtained using Voigt assumption on the different layers.

where θ is the polar angle and φ is the azimuthal angle. In practice, the ODF π P ( p) is not always directly available,
and it is commonly constructed through a second-order orientation tensor, which reads [8,23],
∮
a= p ⊗ pπ P ( p) d p . (20)

More details on the construction of π P ( p(θ, φ)) from a can be found in [8,16]. Since a constant fiber aspect
ratio Ar is assumed, the value of vω(k) , in Eq. (18), can be approximated by the volume fraction of fibers whose
orientations are within [ p(k) − 21 ∆ p , p(k) + 12 ∆ p]. Using the expression of ODF in the spherical coordinates,
Eq. (19), we have
∫ θ (k) + ∆θ ∮ φ (k) + ∆φ
2 2
vω(k) ≈ π P ( p(θ, φ)) sin(θ) dφ dθ , (21)
θ (k) − ∆θ
2 φ (k) − ∆φ
2

where θ (k) and φ (k) are respectively the polar and azimuthal angles of orientation p(k) . The angle increments
∆θ = π/N π2 , and ∆φ = π/Nθ with Nθ = N π2 sin(θ ), are chosen such that the surface of the unit sphere is
subdivided into facets of almost equal areas [8].

2.3.3. Skin-core effect

During the injection molding process, there exists a skin-core effect, in which case the fiber orientation
distribution is not uniform across the plate thickness. In this work, the plate thickness is divided into Nl layers,
and the 2-step homogenization is applied in each layer (l) characterized by a distribution π P(l) ( p) and an inclusions
volume fraction vI(l) . This 2-step homogenization is followed by a Voigt homogenization of the different layers
responses, see Fig. 3. This second Voigt assumption is justified because of the similar out-of-plane behavior of the
different layers due to the fact that the fibers are mostly oriented in the plane.
With ω(k; l) the pseudo-grain (k) of the layer (l), the set of Eqs. (17)–(18) finally becomes,
⟨∆ε⟩ω(k; l) = ⟨∆ε⟩ω = ∆εM ∀ω(k; l) , (22)

Nl
∑ ∑
σM = ⟨σ ⟩ω = vω(k; l) ⟨σ ⟩ω(k; l) , (23)
l k
 L. Wu, K. Zulueta, Z. Major et al. / Computer Methods in Applied Mechanics and Engineering 360 (2020) 112693 7

Fig. 4. (a) An artificial neuron; (b) An artificial Neural Network.

where vω(k; l) is obtained from Eq. (21) as

∫ θ (k) + ∆θ ∮ φ (k) + ∆φ
1 2 2
vω(k; l) ≈ π P(l) ( p(θ, φ)) sin(θ ) dφ dθ . (24)
Nl θ (k) − ∆θ
2 φ − 2
(k) ∆φ

Remarks:. In the considered short-fiber reinforced material, the fiber aspect ratio is not a uniform value. It has
however been shown in [16] that the distribution of aspect ratio along the injection direction (θ = π2 , φ = 0)
is representative of the in-plane directions, while the content of fibers along out-of-plane direction is neglectable.
Therefore, the approximation in Eq. (21) still holds providing that an effective aspect ratio is used in the model.

2.4. Phases materials

The material system is a short E-glass fibers reinforced PA06 polyamide.

• The short glass fibers are modeled using a linear elastic constitutive law with Young’s modulus E I = 72.0 GPa
and Poisson coefficient νI = 0.22.
• The PA06 matrix material is modeled as an elasto-plastic material with Poisson coefficient ν0 = 0.4. Its
Young’s modulus E 0 and initial yielding stress σY0 are unknown at this stage and should be inferred. The
matrix plastic flow is assumed to follow an isotropic hardening law in terms of the accumulated plastic strain
p, which reads
R( p) = hp m 1 (1 − exp(−m 2 p)) , (25)
where h, m 1 and m 2 are unknown hardening parameters to be inferred.
The unknown material parameters E 0 , σY0 , h, m 1 , m 2 and the effect of the effective fiber aspect ratio Ar will
be identified by Bayesian Inference (BI) using experimental tests conducted on composite coupons. Although the
presented two-step homogenization is rather efficient compared to computational homogenization, running this
homogenization process during a BI process with a Metropolis–Hastings (MCMC) random walk is only affordable
in the linear elastic case. Therefore, in order to carry out the BI in the non-linear range, a Deep Neural Network is
adopted as a surrogate model of the two-step homogenization.

3. Deep neural network

Artificial Neural Network (ANNW), or Neural Network (NNW) in short, is a network constructed ∑n 0 by artificial
neurons, see Fig. 4(a). Artificial neurons first perform a weighted sum operation ∑ on input (w0 + k=1 wk x k ), and
then produce output through an activation function of the weighted sum f ( ). NNW models can be viewed
as mathematical models, with a set of algorithms, defining a function F : X → Y , see Fig. 4(b), in which
X = [x1 , x2 , . . . , xn 0 ] and Y = [y1 , y2 , . . . , yn N ]. ’Deep neural network’ is the name used for ’stacked neural
8 L. Wu, K. Zulueta, Z. Major et al. / Computer Methods in Applied Mechanics and Engineering 360 (2020) 112693

Table 1
The ranges of strain and unknown material properties.
ϑ εM [-] E 0 [GPa] σY0 [MPa] h [MPa] m 1 [-] m 2 [-] Ar [-]
Ranges [0 ∼ 0.04] [2 ∼ 7] [0.01 ∼ 70] [0.01 ∼ 120] [0.0001 ∼ 1] [1 ∼ 800] [5 ∼ 25]

network’, that is, networks composed of more than one hidden layer, such as in Fig. 4(b) which presents a NNW
with N − 1 hidden layers.
The training of the NNW refers to regression as supervised learning. Based on a given set of input–output pairs,
supervised learning refers to training the weight parameters wki j , with i = 1, . . . , N ; k = 1, . . . , n i−1 and j = 1,
. . . , n i , and bias w0i j , with i = 1, . . . , N and j = 1, . . . , n i , see the notations in Fig. 4(b), and making the NNW
being able to map input to corresponding output. A loss function is defined to measure the difference between the
predicted output Ŷ Y and given output Y , such as f loss = 12 ∥Ŷ Y − Y ∥2 , where ∥ • ∥ refers to the Frobenius norm.
Deep learning can be carried out by back-propagation, which updates the values of wik j iteratively to minimize the
loss function f loss . To this end, the Python library ’Scikit-learn’ for NNW regression [24] is used in this work.
The architecture of the NNW is obtained by following a trying process: i) starting from a simple architecture; ii)
increasing the depth of the network (the number of hidden layers) progressively; and iii) monitoring the improvement
of the NNW. The hyperbolic tangent function ‘tanh’ is chosen as the activation function.
In this work we consider uniaxial tension performed on composite coupons and then use the homogenized strain
history (along the loading direction) and material parameters as input, X = [εM , ϑ ], and as signal output Y = [σM ],
the unidirectional homogenized tensile stress at a given strain εM . The material parameters ϑ correspond to the
unknown properties of the elasto-plastic matrix, E 0 , σY0 , h, m 1 , m 2 , see Eq. (25), completed by the effective short
fiber aspect ratio Ar . Therefore, we have n 0 = 7 input for n N = 1 output, see Fig. 4(b). Finally, a NNW of five
hidden layers, with 17 neurons in each hidden layer, is adopted.
In order to substitute the multiscale model by a NNW surrogate, the training and testing data of the NNW are
achieved using unidirectional tensile simulations obtained with the two-step MFH method presented in Section 2.3
for different realizations of the material parameters ϑ . For each simulation, the strain εM increases from zero to a
given value and the material parameters are taken randomly from the given ranges listed in Table 1. The bounds
of E 0 and Ar are chosen with respect to the identification process conducted in [16] for the same material system,
but limited to the linear range, in which the inferred values are within these ranges. The bounds of the other
parameters correspond to extreme values by lack of prior knowledge. During each simulation, 70 strain-stress points
are recorded: together with the material parameters, they correspond to 70 input and output data pairs of the NNW.
Only a few hundreds of simulations (typically < 500) were needed for the designed NNW to be trained (i.e. the
mean squared error is lower than 0.5 × 10−4 ). Compared to the hundreds of thousands of simulations required for
a MCMC random walk, the total computation time is reduced drastically.

4. Experimental tests
The 40% of weight fraction short E-glass fibers (GF) reinforced PA06 composite (PA06-GF40) plates of
304.8×101.6×3.65± 0.02 mm3 were manufactured by injection molding with an Engel Insert v200 Single machine.
The nozzle temperature has been set at 270 ◦ C and the mold’s temperature at 90 ◦ C. The plates have been injected
in 2 s under a pressure of 110 bar and the 50-bar packing pressure has been applied during 50 s. Coupons were
then cut from the plates along 3 different Directions, 0◦ , 45◦ and 90◦ , and centered on the plate as illustrated, with
their main dimensions adapted from the ASTM D638, in Fig. 5(a).
Quasi-static uniaxial tensile tests have been performed in a MTS Insight electromechanical actuator at 23 ◦ C and
at a strain rate of 1 mm·min−1 . The strain measurements were carried out by means of a MTS Clip On extensometer
with the error of resolution being below 0.009%. The resulting uniaxial tensile strain-stress curves are reported in
Fig. 5(b).
The fibers orientations have been characterized by Computed Tomography (CT) technique on volume samples of
2 × 2 × 3.2 mm3 at the plate center while the volume fraction has been estimated from the mass fraction measured
using the pyrolysis technique, see details in [16]. The second-order orientation tensors athickness and the volume
fraction vIthickness when considering the full thickness are reported in Table 2. The CT-scan measurements are then
used to evaluate the orientation tensor a(l) and volume fraction vI(l) on 11 layers across the thickness to account for
the skin-core effect.
 L. Wu, K. Zulueta, Z. Major et al. / Computer Methods in Applied Mechanics and Engineering 360 (2020) 112693 9

Fig. 5. (a) Locations and geometries (in mm) of the tested coupons along 0◦ , 45◦ and 90◦ Directions; (b) Unidirectional tensile results of
0◦ , 45◦ and 90◦ .

Table 2
Orientation tensor a and fiber volume fraction vI at center location for the PA06-GF40 plate: global values
(l)
athickness and vIthickness when considering the full thickness and the layer values a(l) and vI when considering
11 layers to account for the skin-core effect.
a= [a11 a22 a33 a12 a23 a13 ] vI
athickness = [0.747 0.204 0.049 0.031 −0.0032 −0.005] 0.2214
a(1) = [0.851 0.115 0.033 0.0045 0.001 −0.0005] 0.2343
a(2) = [0.867 0.096 0.037 0.0181 0.0006 −0.005] 0.2323
a(3) = [0.844 0.109 0.047 0.0261 0.0005 −0.008] 0.2271
a(4) = [0.817 0.134 0.049 0.0211 0.0009 −0.012] 0.2227
a(5) = [0.680 0.260 0.060 0.0360 0.0017 −0.036] 0.2035
a(6) = [0.227 0.674 0.099 0.125 −0.0134 −0.0092] 0.1950
a(7) = [0.648 0.294 0.059 0.0358 −0.0098 0.0237] 0.2032
a(8) = [0.818 0.139 0.044 0.0063 −0.0040 0.0026] 0.2240
a(9) = [0.857 0.1054 0.038 0.0195 −0.0027 −0.0003] 0.2308
a(10) = [0.878 0.092 0.031 0.0126 −0.0016 −0.0022] 0.2364
a(11) = [0.835 0.134 0.030 0.0145 −0.0068 −0.0036] 0.2261

5. Bayesian inference of the multiscale parameters and matrix properties

Bayesian Inference is a statistical analysis approach based on Bayes’ theorem which states that the posterior
probability of a random parameters vector α ∈ Rn for given observations of another random vector β ∈ Rm ,
π (α|β), is proportional to the prior probability πprior (α) multiplied by the likelihood of β for given observations of
α, where π (•) (π (•|•)) denotes a (conditional) Probability Density Function (PDF).
The prior distribution πprior (α) reflects the initial belief or knowledge one has on α. The likelihood function is
defined from observations of β by a conditional PDF π (β|α), which is constructed from the different observation
data. The conditional PDF πpost (α|β) is the posterior distribution of the random vector α that accounts for the
observation data. Since π (β) is a constant for given observations, we simply write
πpost (α|β) ∝ π (β|α)πprior (α) . (26)

5.1. Stress noise-based inference

For uni-axial tensile loading, the predictions of two-step MFH model with skin-core effect ΦMFH (εM , ϑ ) reads
σM (εM , ϑ ) = ΦMFH (εM , ϑ ) , (27)
10 L. Wu, K. Zulueta, Z. Major et al. / Computer Methods in Applied Mechanics and Engineering 360 (2020) 112693

Table 3
Experimental strain–stress observations.
Direction Strain ε M [-] Stress Σ [MPa]
ε10 0.0025 Σ10 28.3 29.6 30.2
ε20 0.005 Σ20 53.0 55.6 56.0
0◦ ε30 0.01 Σ30 93.8 97.0 96.3
ε40 0.015 Σ40 120.6 123.8 122.1
ε50 0.02 Σ50 137.3 140.5 138.1
ε145 0.005 Σ145 30.1 30.9 36.3
ε245 0.01 Σ245 53.0 53.5 60.7
45◦ ε345 0.02 Σ345 79.6 79.8 85.3
ε445 0.03 Σ445 91.0 92.7 95.4
ε545 0.04 Σ545 96.2 98.8 –
ε190 0.0045 Σ190 25.2 24.8 26.2
ε290 0.009 Σ290 44.3 44.0 45.6
90◦ ε390 0.018 Σ390 66.1 66.3 68.0
ε490 0.027 Σ490 75.2 75.8 76.6
ε590 0.036 Σ590 79.5 80.6 –

where ϑ represents the material parameters, which include the matrix material properties and the effective short
fiber aspect ratio, to be inferred. The relation between a stress measurement Σ and the model response is written
as
Σ = σM + ωΣ = ΦMFH (εM , ϑ ) + ωΣ , (28)
where ωΣ is the stress noise assumed to follow a Gaussian distribution N(x|0, sΣ
2
) with
[ ( )2 ]
1 1 x−y
2
N(x|y, sΣ )= √ exp − . (29)
sΣ 2π 2 sΣ
Therefore, the conditional distribution of stress reads
π (Σ |εM , ϑ ) = N(Σ − ΦMFH (εM , ϑ )|0, sΣ
2
). (30)

5.2. Bayesian inference using the surrogate model

Since the computational time of the two-step MFH process makes a MCMC sampling unaffordable, in particular
when accounting for the skin-core effect, NNW is used as a surrogate ΦNNW of MFH model ΦMFH in order to
perform the BI, see Fig. 6.
In order to minimize the number of training data for NNW, three NNWs are trained for the three loading
Directions, 0◦ , 45◦ , 90◦ , respectively, as summarized in Box 1 of Fig. 6. The observations are the unidirectional
tensile experimental data points; practically, only the experimental data points marked in Fig. 5(b) are used and are
reported in Table 3. Therefore, depending on the loading Direction, Eq. (30) becomes
( )
π (Σid |εid , ϑ ) = N Σid − ΦNNW
d
(εid , ϑ )|0, sΣ
2
d , (31)
i

where the superscript d = 0, 45, 90 refers to the tensile loading and coupon Direction, and where the subscript i
is used to indicate the stress variance sΣ d obtained from the experimental curves at a given strain εid , the subscript
i
“M” being omitted for conciseness.

5.2.1. Likelihood function

In order to construct the likelihood, it is assumed that each tensile test is independent from the other ones. This is
justified by the fact that the samples are extracted from different composite plates. Using the observations reported
L. Wu, K. Zulueta, Z. Major et al. / Computer Methods in Applied Mechanics and Engineering 360 (2020) 112693 11

Fig. 6. The Bayesian inference of the multiscale model ΦMFH parameters accelerated by the surrogate model ΦNNW .
12 L. Wu, K. Zulueta, Z. Major et al. / Computer Methods in Applied Mechanics and Engineering 360 (2020) 112693

Fig. 7. Distribution of E 0 obtained by MCMC sampling and its approximated beta distribution.

in Table 3 and the theory of conditional independence [25], the final likelihood function is obtained from Eq. (31)
and reads1
∏ ( )
Σ |εε M , ϑ ) =
π (Σ N Σi,d j − ΦNNW
d
(εid , ϑ )|0, sΣ
2
d , (32)
i
d, i, j
po po
where ε M and Σ represent the strain, εid , and stress, Σi,d j , observations listed in Table 3, with i = 1, . . . , n d , n d
being the considered points number on a stress–strain curve, and j = 1, . . . , n d,i , n d,i being the number of stress
ob ob

observations corresponding to the strain εid . The standard deviations are obtained from the 43 pairs of observations
(εid , Σi,d j ) listed in Table 3 as

d, i n ob d, i n ob
1 ∑ ( ))2 1 ∑ d
2
Σ .
( d
Σi, j − E Σid with E Σid = ob
( )
sΣ d = ob (33)
i n d, i − 1 j=1 n d, i j=1 i, j

5.2.2. Prior distribution

The prior distribution of ϑ = [E 0 , σY0 , h, m 1 , m 2 , Ar ] is constructed from 6 independent distributions, as
ϑ ) = Be E0 × UσY0 × Uh × Um 1 × Um 2 × U Ar .
πprior (ϑ (34)
For all the parameters, at the exception of E 0 , by lack of information we use uncorrelated uniform distributions U
whose bounds are the ones used in the NNW training and were justified in Section 3, see Table 1. For the matrix
Young’s modulus E 0 , we consider a beta distribution Be since in [16] such a distribution was used to describe the
material discrepancy in the linear range, resulting in a conditional beta distribution Be(E 0 ; α, β, a, b) in which
α, β are its shape parameters and a, b are its lower and upper bounds, respectively. Based on this conditional
distribution of E 0 , we have
∫
π (E 0 ) = Be(E 0 ; α, β, a, c)π (α, β, a, b)dα dβ da db , (35)
α, β, a, c
where the distributions of α, β, a, b were inferred in [16] from experimental tests. Their integration in Eq. (35) is
carried out using a MCMC sampling, and the resulting marginal distribution is eventually approximated by a beta
distribution Be(E 0 |6.4, 11.3, 2.4 GPa, 6.5 GPa), see Fig. 7, which serves as the prior distribution of E 0 .
1 po
For the n d data points extracted from one tensile sample of a defined unique loading path, the measured stresses Σid at different strain
levels εid are correlated: the correlation among the experimental measurements results from the unique material properties of a given tested
sample. In all generality, the likelihood function is built, using the data points from one sample, as
π(Σ1d , . . . , Σndpo |ε1d , . . . , εndpo , ϑ ).
d d

Since the correlations among the measured stresses are assumed to result solely from the shared material properties, when these material
properties are given, the probability distributions of the stresses become independent. As strain εid has no effect on stress Σ dj , i ̸= j, using
∏n po
the theory of conditional independence, the likelihood function can be factorized into i=1 d
π(Σid |εid , ϑ ).
 L. Wu, K. Zulueta, Z. Major et al. / Computer Methods in Applied Mechanics and Engineering 360 (2020) 112693 13

Fig. 8. Trace of the inferred parameters with respect to the MCMC iteration.

5.2.3. Posterior distribution

Finally, the posterior distribution
ϑ |εε M , Σ ) ∝ π (Σ
πpost (ϑ Σ |εε M , ϑ )πprior (ϑ
ϑ) , (36)
is evaluated using a MCMC technique, which is a random walk in the parameter space ϑ ∈ R6 , as summarized in
Box 2 of Fig. 6. The adaptive variant [26] of the Metropolis algorithm [27] is used, see also [16] for details.

6. Results and discussion

In this section, we first ascertain the convergence of the BI before providing the posterior distribution obtained
by the NNW-accelerated BI. We also show that considering the NNW as surrogate during the BI does not impact
on the accuracy. Finally we show that, while the NNW-accelerated BI is manageable, using directly the two-step
MFH during the BI is computationally unaffordable.

6.1. Convergence

The BI process described in Section 5.1 has been conducted to generate 20000 data samples from which the first
2000 samples were discarded. The convergence of the MCMC algorithm is studied in Fig. 8 in terms of the trace,
i.e. the realizations in terms of the iteration number. After a few thousands of iterations, the trace has the typical
fuzzy shape of a converged analysis.

6.2. Accuracy verification

The posterior distribution of the random material properties ϑ and their correlations can be seen in Fig. 9. The
three data points marked in Fig. 9 were picked among the obtained random 18000 samples: the point “H”, “M” or
“L” represents a sample with the Highest, Moderate or Lowest (among the 18000 samples) probability, respectively.
For each of them, the strain-stress curves predicted by the NNW ΦNNM and the two-step MFH with skin-core effect
ΦMFH are presented in Figs. 10(a) and 10(b), respectively. The colors of the curves correspond to the colors of the
markers of Fig. 9. The experimental measurements are also reported for comparison purpose.
A good accuracy of surrogate ΦNNM is observed in Fig. 10. Among the obtained random data samples, the
three samples “H”, “M” and “L” exhibit a good agreement with the experimental measurements. Because of the
good repeatability of the experimental measurements, the variances of the stress noise functions ωΣ , Eq. (28), are
rather small. As a consequence, any sample obtained by the MCMC can serve as a good solution for the required
material parameters: when using different material parameters, the discrepancy observed between the strain-stress
curves is comparable to the accuracy of the trained surrogate ΦNNM . On the one hand, rigorously speaking, the
inferred distributions of material properties do not seem useful at first since the different observations provide
accurate predictions. Nevertheless, on the other hand, compared to a parameters identification based on optimization
14 L. Wu, K. Zulueta, Z. Major et al. / Computer Methods in Applied Mechanics and Engineering 360 (2020) 112693

Fig. 9. Distributions of material properties from the MCMC sampling; three marked points [4.16 GPa, 7.14 MPa, 97.3 MPa, 0.216,
606.7, 13.4], [4.36 GPa, 5.51 MPa, 99.7 MPa, 0.212, 504.7, 13.2], and [4.42 GPa, 1.15 MPa, 112.0 MPa, 0.221, 626.1, 14.3] are extracted
and correspond to realizations of ϑ of the Highest, Moderate and Lowest (among the 18000 samples) probabilities.

Fig. 10. The comparison of the experimental measurements to the numerical prediction of (a) the NNW and (b) the two-step homogenization
for the three parameters observations extracted from Fig. 9.

techniques, instead of obtaining a unique solution, the BI reveals the non-linear relationship among the material
properties, or the manifold of the material properties. Besides, it naturally allowed considering several loading cases
altogether, which could lead to three different sets of parameters for a deterministic inverse identification.
Nevertheless, in this paper, the micro-mechanical model is seen as deterministic: although some uncertainties
exist in the identified parameters the multiscale model is not able to represent the observed dispersion in the
materials. However, in the case of polymeric-based composites, such dispersion can be important: in [28], the PA06
tensile modulus measured at constant temperature and strain rate ranges from 1200 to 3400 MPa, and tensile tests
conducted on PA06-GF30 lead to a Young’s modulus ranging from 6200 to 9500 MPa. To capture this dispersion,
properties spatial distributions can be inferred [29], or BI can be adapted in order to infer the parameters of an
assumed distribution of the material properties instead of the material properties themselves [30]. This so-called
distribution-based BI approach requires a double MCMC sampling process in the non-linear range, which is the
reason why in [16] the authors have inferred the distribution of the linear material constants only. However, because
of the high efficiency, see next subsection, of the NNW, such an approach would become affordable to infer the
parameters distribution of elasto-(visco)-plastic composites multiscale models.
 L. Wu, K. Zulueta, Z. Major et al. / Computer Methods in Applied Mechanics and Engineering 360 (2020) 112693 15

Table 4
Comparison of computational time between the NNW-accelerated BI and the direct BI. Simulations were performed on a laptop.
Direct BI NNW-accelerated BI
Offline Number of MFH 500 × 3 dir.
simulations
Time of MFH 500 × 3×
simulations 100 s = 42 h
Training time 8 min × 3 dir.
∼ 0.5 h
Online Number of samples 20000 Number of samples 20000
Number of rejected [10–20000] Number of rejected [10–20000]
iterations per sample iterations per sample
Time for one iteration 100 s × 3 dir. Time for one iteration 0.005 s × 3 dir.
Estimated time for BI ∼ 300 × 20000 × 3 Time for BI ∼ 300 × 20000 × 3
×100 s ≃ 57 years ×0.005 s ≃ 1 day

6.3. Computational efficiency

Table 4 compares the computational time needed to perform a direct BI, which uses the 2-step MFH model ΦMFH
to evaluate the likelihood, to the time to perform the NNW-accelerated BI, which uses the surrogate model ΦNNW
when evaluating the likelihood. For the NNW-accelerated BI, an offline computation is required. Data needs first
to be generated, which consists in running 500 times the 2-step MFH model ΦMFH along the 3 loading directions.
Although this requires roughly 2 days of computations, this can be performed in parallel since the simulations
are independent. The NNW then needs to be trained for each loading direction, requiring half an hour. The total
offline time is thus about 2 days. To perform the BI, for each accepted sample, the MCMC algorithm rejects
between 10 and 20000 samples. It is estimated that 6 millions evaluations of the likelihood are required to reach
the converge posterior (we note that a rejected sample requires the evaluation of the likelihood). Since evaluating the
surrogate model ΦNNW is almost instantaneous, the NNW-accelerated BI remains affordable and requires roughly
1 day of computation. However, the direct BI cannot be considered, since the evaluation of one likelihood requires
the evaluations of the 2-step MFH model ΦMFH along the 3 directions, corresponding to more than 300 sec per
iteration, which would lead to several years of computation. Besides, the iterations in the MCMC algorithm are
dependent and the parallelization scalability is limited since the first thousands of iterations need to be burned.

7. Conclusions
In order to model Short Fiber Reinforced Polymer responses, a prevalent method is to adopt a multiscale modeling
strategy, for which the material properties of each phase should be inferred. However, although the behavior of
the composite material seen as a homogenized medium can be experimentally measured, the identification of the
matrix material properties from the composites response requires a special identification process. Indeed, applying
classical identification processes, such as optimization and Bayesian Inference, is limited by the complexity of
the micromechanics models, especially for non-linear cases in which multi-parameters are involved and for which
the computation cost is not negligible. In particular, when considering a two-step homogenization as a non-linear
multiscale model, the MCMC sampling of a BI process becomes unaffordable.
In this paper, a NNW was adopted as a surrogate model to replace the expensive micromechanics model during
the material parameters identification process. The accuracy of NNW was verified from the comparison between
the strain stress curves obtained with NNW and those obtained by the two-step homogenization. The matrix
material properties and the effective fiber aspect ratio were then identified by the BI process from the experimental
measurements on composite coupons. In particular, it was shown that the inferred properties could predict the
composite material response in agreement with the experimental curves.
Although the methodology was developed for a particular homogenization method, i.e. MFH, it can be extended
to other ones, such as computational homogenization, or to more complex identification tests. For example, the
method can be applied for multi-axial stress–strain cases, assuming a monotonic and proportional loading, in which
case a strain vector – of size up to 6– serves as inputs, together with the material parameters, of the NNW instead of a
16 L. Wu, K. Zulueta, Z. Major et al. / Computer Methods in Applied Mechanics and Engineering 360 (2020) 112693

single strain variable; the corresponding – up to 6 – outputs will be used to represent the stress vector. Consequently,
since the number of inputs is increased, the adopted structure of the NNW needs to be enhanced by more hidden
layers and neurons to be able to represent a more complicated high dimensional non-linear mapping. As a result of
the enhancement of the NNW, an increased number of training data would probably be required. On the aspect of
BI process, the original uni-variate Gaussian distribution used to construct the likelihood function has to be replaced
by a multivariate Gaussian distribution of higher dimension — up to 6, and the different importance of the stress
entries can be addressed by using a modified precision matrix in the multivariate Gaussian distribution.
Finally, in order to introduce some stochasticity in the multiscale model, it is foreseen to infer the parameters of an
assumed distribution of the material properties instead of the material properties themselves. Because this so-called
distribution-based BI approach requires a double MCMC sampling process in the non-linear range, the method is
not practical with direct BI for a complex micro-mechanical model. However, because of the high efficiency of the
NNW, this becomes possible in the cases of elasto-(visco)-plastic composites.

Acknowledgments
The research has been funded by the Walloon Region, Belgium under the agreement no 1410246-STOMMMAC
(CT-INT 2013-03-28), by the Gaitek 2015 program of the Basque Government, Spain, and by the Austrian Research
Promotion Agency (ffg), Austria under the agreement no 850392 (STOMMMAC) in the context of the M-ERA.NET
Joint Call 2014. Special appreciation goes to the CT Group of the University of Applied Sciences (Wels, A) for
conducting the CT experiments.

References
[1] K. Matous, M.G. Geers, V.G. Kouznetsova, A. Gillman, A review of predictive nonlinear theories for multiscale modeling of
heterogeneous materials, J. Comput. Phys. 330 (Supplement C) (2017) 192–220, http://dx.doi.org/10.1016/j.jcp.2016.10.070.
[2] J.D. Eshelby, The determination of the elastic field of an ellipsoidal inclusion, and related problems, Proc. R. Soc. A 241 (1226)
(1957) 376–396.
[3] T. Mori, K. Tanaka, Average stress in matrix and average elastic energy of materials with misfitting inclusions, Acta Metall. 21 (5)
(1973) 571–574.
[4] R. Hill, Continuum micro-mechanics of elastoplastic polycrystals, J. Mech. Phys. Solids 13 (2) (1965) 89–101, http://dx.doi.org/10.
1016/0022-5096(65)90023-2.
[5] D.R.S. Talbot, J.R. Willis, Variational principles for inhomogeneous non-linear media, IMA J. Appl. Math. 35 (1) (1985) 39–54.
[6] L. Wu, L. Noels, L. Adam, I. Doghri, A combined incremental–secant mean–field homogenization scheme with per–phase residual
strains for elasto–plastic composites, Int. J. Plast. 51 (2013) 80–102, http://dx.doi.org/10.1016/j.ijplas.2013.06.006.
[7] C.W. Camacho, C.L. Tucker, S. Yalva c, R.L. McGee, Stiffness and thermal expansion predictions for hybrid short fiber composites,
Polym. Compos. 11 (4) (1990) 229–239, http://dx.doi.org/10.1002/pc.750110406.
[8] I. Doghri, L. Tinel, Micromechanics of inelastic composites with misaligned inclusions: numerical treatment of orientation, Comput.
Methods Appl. Mech. Engrg. 195 (13) (2006) 1387–1406, http://dx.doi.org/10.1016/j.cma.2005.05.041.
[9] M. Vincent, T. Giroud, A. Clarke, C. Eberhardt, Description and modeling of fiber orientation in injection molding of fiber reinforced
thermoplastics, Polymer 46 (17) (2005) 6719–6725.
[10] J.L. Beck, L.S. Katafygiotis, Updating models and their uncertainties. I: Bayesian statistical framework, J. Eng. Mech. 124 (1998)
455–461.
[11] T. Most, Identification of the parameters of complex constitutive models: least squares minimization vs. Bayesian updating, in: D.
Straub (Ed.), Reliability and Optimization of Structural Systems, 2010.
[12] S. Madireddy, B. Sista, K. Vemaganti, A Bayesian approach to selecting hyperelastic constitutive models of soft tissue, Comput.
Methods Appl. Mech. Engrg. 291 (2015) 102–122.
[13] H. Rappel, L. Beex, J. Hale, L. Noels, S. Bordas, A tutorial on Bayesian inference to identify material parameters in solid mechanics,
Arch. Comput. Methods Eng..
[14] T.C. Lai, K. Ip, Parameter estimation of orthotropic plates by Bayesian sensitivity analysis, Compos. Struct. 34 (1) (1996) 29–42,
http://dx.doi.org/10.1016/0263-8223(95)00128-X.
[15] F. Daghia, S. de Miranda, F. Ubertini, E. Viola, Estimation of elastic constants of thick laminated plates within a Bayesian framework,
Compos. Struct. 80 (3) (2007) 461–473.
[16] M. Mohamedou, K.Z. Uriondo, C.N. Chung, H. Rappel, L. Beex, L. Adam, Z. Major, L. Wu, L. Noels, Bayesian identification of
mean-field homogenization model parameters and uncertain matrix behavior in non-aligned short fiber composites, Compos. Struct.
(2019) (submitted for publication).
[17] B.A. Le, J. Yvonnet, Q.-C. He, Computational homogenization of nonlinear elastic materials using neural networks, Internat. J. Numer.
Methods Engrg. 104 (12) (2015) 1061–1084, http://dx.doi.org/10.1002/nme.4953.
[18] M. Bessa, R. Bostanabad, Z. Liu, A. Hu, D.W. Apley, C. Brinson, W. Chen, W.K. Liu, A framework for data-driven analysis of
materials under uncertainty: countering the curse of dimensionality, Comput. Methods Appl. Mech. Engrg. 320 (2017) 633–667,
http://dx.doi.org/10.1016/j.cma.2017.03.037.
 L. Wu, K. Zulueta, Z. Major et al. / Computer Methods in Applied Mechanics and Engineering 360 (2020) 112693 17

[19] J.F. Unger, C. Könke, Coupling of scales in a multiscale simulation using neural networks, Comput. Struct. (ISSN: 0045-7949) 86
(21) (2008) 1994–2003, http://dx.doi.org/10.1016/j.compstruc.2008.05.004.
[20] F. Fritzen, M. Fernández, F. Larsson, On-the-fly adaptivity for nonlinear twoscale simulations using artificial neural networks and
reduced order modeling, Front. Mater. (ISSN: 2296-8016) 6 (2019) 75, http://dx.doi.org/10.3389/fmats.2019.00075.
[21] D.R.S. Talbot, J.R. Willis, Bounds and self-consistent estimates for the overall properties of nonlinear composites, IMA J. Appl. Math.
39 (3) (1987) 215–240, http://dx.doi.org/10.1093/imamat/39.3.215.
[22] P. Ponte Castañeda, A new variational principle and its application to nonlinear heterogeneous systems, SIAM J. Appl. Math. 52 (5)
(1992) 1321–1341.
[23] B. Weber, B. Kenmeugne, J. Clement, J. Robert, Improvements of multiaxial fatigue criteria computation for a strong reduction of
calculation duration, Comput. Mater. Sci. 15 (4) (1999) 381–399, http://dx.doi.org/10.1016/S0927-0256(98)00129-3.
[24] F. Pedregosa, G. Varoquaux, A. Gramfort, V. Michel, B. Thirion, O. Grisel, M. Blondel, P. Prettenhofer, R. Weiss, V. Dubourg, J.
Vanderplas, A. Passos, D. Cournapeau, M. Brucher, M. Perrot, E. Duchesnay, Scikit-learn: machine learning in python, J. Mach. Learn.
Res. 12 (2011) 2825–2830.
[25] C.M. Bishop, Pattern Recognition and Machine Learning, in: Information Science and Statistics, Springer-Verlag, New York, USA,
ISBN: 978-0-387-31073-2, 2006.
[26] H. Haario, E. Saksman, J. Tamminen, Adaptive proposal distribution for random walk metropolis algorithm, Comput. Statist. 14 (3)
(1999) 375–395, http://dx.doi.org/10.1007/s001800050022.
[27] W. Gilks, S. Richardson, D. Spiegelhalter, Markov chain Monte Carlo in practice, in: Chapman & Hall/CRC Interdisciplinary Statistic,
Chapman and Hall, Weinheim, Germany, 1995.
[28] H. Nouri, Modélisation et identification de lois de comportement avec endommagement en fatigue polycyclique de matériaux composite
à matrice thermoplastique (Ph.D. thesis), Arts et Métiers ParisTech, Metz (France), 2009.
[29] A. Vigliotti, G. Csányi, V. Deshpande, Bayesian inference of the spatial distributions of material properties, J. Mech. Phys. Solids 118
(2018) 74–97, http://dx.doi.org/10.1016/j.jmps.2018.05.007.
[30] H. Rappel, L. Beex, Estimating fibres’ material parameter distributions from limited data with the help of Bayesian inference, Eur. J.
Mech. A Solids 75 (2019) 169–196.