Download as pdf or txt
Download as pdf or txt
You are on page 1of 15

Computational Mechanics (2019) 64:307–321

https://doi.org/10.1007/s00466-018-1643-0

ORIGINAL PAPER

A data-driven computational homogenization method based on


neural networks for the nonlinear anisotropic electrical response of
graphene/polymer nanocomposites
Xiaoxin Lu1 · Dimitris G. Giovanis2 · Julien Yvonnet1 · Vissarion Papadopoulos3 · Fabrice Detrez1 · Jinbo Bai4

Received: 28 July 2018 / Accepted: 22 September 2018 / Published online: 11 October 2018
© Springer-Verlag GmbH Germany, part of Springer Nature 2018

Abstract
In this paper, a data-driven-based computational homogenization method based on neural networks is proposed to describe
the nonlinear electric conduction in random graphene-polymer nanocomposites. In the proposed technique, the nonlinear
effective electric constitutive law is provided by a neural network surrogate model constructed through a learning phase on a
set of RVE nonlinear computations. In contrast to multilevel (FE2 ) methods where each integration point is associated with a
full nonlinear RVE calculation, the nonlinear macroscopic electric field-electric flux relationship is efficiently evaluated by the
surrogate neural network model, reducing drastically (by several order of magnitudes) the computational times in multilevel
calculations. Several examples are presented, where the RVE contains aligned graphene sheets embedded in a polymer matrix.
The nonlinear behavior is due to the modeling of the tunelling effect at the scale of graphene sheets.

Keywords Multiscale analysis · Data-driven analysis · Graphene nanocomposites · Homogenization · Electric behavior ·
Artificial neural network

1 Introduction [1,2] and independently developed by many other authors


under different names and later extended (see [3–12] and
Computational multiscale analysis is a powerful tool in engi- reviews in [13,14]). The basic idea is to conduct a nonlinear
neering to study the nonlinear behavior of structures at finite element calculation at the scale of the structure (macro
engineering scales based on the knowledge of the microstruc- scale) and to solve a nonlinear problem on an representative
ture in the material. In contrast to empirical models, such volume elements (RVE) attached to each integration point of
analysis allows to determine the sensitivity of macroscopic the macro mesh to evaluate the nonlinear constitutive rela-
properties with respect to microstructural constituents, and tionship. The main drawback of such technique is its inherent
offers new possibilities to design and propose new materials computational complexity, leading to untractable computa-
without costly experiments. Multilevel finite elements (FE2 ) tional times for large meshes at both scales, even though
analysis has been developed since the seminal papers of Feyel techniques like model reduction [9,15], hyper reduction [16],
or recent techniques like self-consistent clustering analysis
[17] can be employed to reduce the computational costs.
B Julien Yvonnet In [18,19], a new family of multiscale data-driven meth-
julien.yvonnet@univ-paris-est.fr ods has been initiated where a set of “Off-line calculations”
1
on RVE are first performed to construct a surrogate model
Laboratoire Modélisation et Simulation Multi Échelle,
which substitutes the RVE calculations in a multilevel anal-
MSME UMR 8208 CNRS, Université Paris-Est, 5 bd
Descartes, 77454 Marne-la-Vallée, France ysis. More specifically, in [18], the surrogate model was
2 constructed using high-order interpolation schemes, and in
Department of Civil Engineering, Johns Hopkins University,
Baltimore, USA [19] it was further extended to utilizing artificial neural net-
3 works, which is an important tool widely used for machine
Department of Civil Engineering, National Technical
University of Athens, Athens, Greece learning and artificial intelligence in data science. Artificial
4 neural networks (ANN), also called simply neural networks
Laboratoire de Mécanique des Sols, Structures et Matériaux,
UMR 8579 CNRS, Université Paris Saclay, Paris, France (NN) are computational models inspired by biological neu-

123
308 Computational Mechanics (2019) 64:307–321

ral networks. Provided examples (input-output relations), the 2 Multiscale modeling of the nonlinear
NN ’learns’ the data by creating a mapping between the input composite and effective quantities
and the output which can be used in order to predict the output
for any new input. In [20], such data-driven multiscale tech- 2.1 Miscroscale model
nique has been extended to stochastic models of materials
and to accelerated off-line simulations using self-consistent We consider an RVE with periodic microstructure defined in
clustering analysis. In [21], radial basis approximations have a domain Ω ⊂ R3 containing N planar multilayer graphene
been introduced to construct the surrogate model from the sheets randomly distributed (see Fig. 1).
RVE data. It is worth noting that these methods, relying Due to their extremely low thickness, the graphene sheets
on RVE calculations as data, differ from all data-driven are modeled as highly conducting imperfect surfaces [43].
techniques using neural networks or other machine learn- The related surfaces are collectively denoted by Γ . Following
ing methods where the data are obtained by experiments (see [42], the electric tunnelling effect, which has been shown to
e.g. [22–32]). In the present work, we extend the data-driven explain experimentally observed very low percolation thresh-
computational homogenization method based on neural net- olds and nonlinear behavior in nanocomposites (see [42,44]),
works developed in [19] to the nonlinear electric conduction is taken into account and described in the following. The elec-
in nanocomposites. tric power of the system is defined by
Graphene/polymer nanocomposites have recently
attracted a growing attention due to their interesting electrical  
properties [33–35]. Graphene sheets can provide percolated W = ω (x)dΩ +
b
ωs (x)dΓ , (1)
Ω Γ
pathways for electron transfer, making the composite electri-
cally conductive. These composites show a very low electric
conductivity at extremely small graphene volume fractions. where the density functions ωb and ωs are the bulk and sur-
However, increasing the graphene concentration, a sudden face density functions expressed by
rise of several magnitude in the electric conductivity at a
certain concentration is observed when the fillers form a 1 1
network, operating a transition from insulator to conduc- ωb (x) = j(x) · E(x), ωs (x) = js (x) · Es (x). (2)
2 2
tor [36,37]. The critical filler fraction is the percolation
threshold, which is presented to be as low as 0.07 vol%
due to the high aspect ratio of graphene sheets [38]. While In (2), E(x) and j(x) denote the electric field and cur-
below the percolation threshold the current–voltage relation rent density respectively, and E(x) is related to the electric
of the graphene nanocomposite is linear, the electric proper- potential φ by E(x) = −∇φ(x), where ∇(·) is the gradient
ties quickly become highly nonlinear above the percolation operator. Besides, Es (x) and js (x) are the surface electric
threshold [39]. The low percolation threshold and nonlinear field and surface current density with respect to the graphene
electric behavior can be explained by the tunneling effect sheets, where Es = P.E = −P∇φ with P = I − n ⊗ n a
phenomenon [40–42]. projector operator characterizing the projection of a vector
The layout of this paper is as follows. In Sect. 2, the non- along the tangent plane to Γ at a point x ∈ Γ and n is the
linear model describing the RVE at the scale of individual unit normal vector to Γ .
graphene sheets is described. In Sect. 3, the basics of artifi- The tunneling effect is introduced through the nonlin-
cal neural networks (ANN) is briefly reviewed. The strategy ear constitutive relationship relating j and E in the polymer
to construct the set of off-line data and the validation of the matrix as:
surrogate model is described in Sect. 4. Finally, numerical
 p
examples are presented, where the solution of the proposed k0 E if d(x) ≤ dcut ,
data-driven neural network computational homogenization j= (3)
G(E, d) |E|
E
if d(x) > dcut ,
(NN-CH) is compared to the solution of FE2 calculations
with regard to accuracy and efficiency.
where dcut is a cut-off distance above which the tunneling
p
effect can be neglected, and k0 is the electric conductiv-
ity tensor of the polymer matrix when neglecting tunneling
effect. In (3), d(x) is a distance function between the
graphene sheets, here defined as the sum of the two smallest
distances between a point x in the polymer matrix and the
two nearest neighbor graphene sheets (see more details in
[42]). The nonlinear term was provided by Simmons [45] as:

123
Computational Mechanics (2019) 64:307–321 309

 
2.2e3 E 2 8π 1 3
G (E, d) = exp − (2m) 2 Φ02 . . .
8π h p Φ0 2.96h p eE
 1
  
(2mΦ0 ) 2  2 4πd 1
+ 3· e/h p Ed exp − (2mΦ0 ) 2 .
2d hp
(4)

Here, the tunneling current G is a nonlinear function of


the electric field E, Φ0 and d denote to the barrier height and
barrier width respectively, e and m are the charge and the
effective mass of electron, respectively, and h p is the Planck
(a) (b)
constant.
The surface current density js of the graphene surface Γ
is related to the surface electric field, Es through (see e.g.,
[43,46]):

js (x) = ks Es , (5)

where the surface conductivity ks of the graphene can be


deduced from the bulk one (related to the multilayer graphene
with finite thickness h) as:

(c) (d) (k g n) ⊗ (k g n)
ks = hS∗ , S∗ = k g − . (6)
k g : (n ⊗ n)

Considering the constitutive equations above, and min-


imizing the electric power (1) with respect to the electric
potential field, the weak form which can be solved by finite
element method is obtained as
 
j(φ) · ∇(δφ)dΩ − P∇φ · ks P∇(δφ)dΓ = 0, (7)
Ω Γ

where φ ∈ H 1 (Ω), φ satisfying the Dirichlet boundary con-


(e) (f)
ditions over ∂Ω and δφ ∈ H 1 (Ω), δφ = 0 over ∂Ω. In
what follows, the (·) notations denote macro quantities, i.e.
related to the scale of the structure. The RVE is subjected to
a homogeneous electric field E, i.e.

1
E(x)dΩ = E, (8)
V Ω

where V is the volume of Ω, and which can be satisfied e.g.


by the follwing boundary condition:

(g) (h) φ(x) = −E · x + φ̃(x) on ∂Ω, (9)


Fig. 1 Realizations of microscopic RVE with various graphene volume
fraction. The graphene sheets are aligned parallel to X-Y plane, and where φ̃(x) is a periodic function over Ω. The effective elec-
the 2D meshes on graphene sheets are plotted. The microstructure is tric current J is defined according to (see [42]):
periodic, a graphene sheet crossing a boundary is extended on the RVE  

opposite side. a 0.53 vol%, b 0.66 vol%, c 0.79 vol%, d 0.92 vol%, e 1
1.05 vol%, f 1.19 vol%, g 1.32 vol% and h 1.58 vol% J= j(x)dΩ + js (x)dΓ . (10)
V Ω Γ

The effective (tangent) conductivity is defined as



 ∂J E
k E = . (11)
∂E

123
310 Computational Mechanics (2019) 64:307–321

2.2 Macroscopic problem estimation of J is directly provided by the surrogate model


based on Neural Networks. The tangent matrix k is evaluated
Considering a structure defined in a domain Ω ⊂ R3 at the from (11) by finite differences, i.e.:
macroscale with boundary ∂Ω, the equilibrium equation for
( j) 
the macroscopic problem reads as:   J i E + δE − Ji E
kT E  (21)
ij α
∇ · J = 0 in Ω, (12)
∗ ∗
φ = φ on ∂Ω φ , J · n = J n on ∂Ω J , (13) with
( j)
where ∇ · (·) is the divergence operator, and Ω φ and ∂Ω J δE = αe j , (22)
denote the Dirichlet and Neumann complementary bound-
∗ e j being a unitary vector basis and α a small perturbation.
aries, respectively. Assuming in the following that J n = 0,
the weak form corresponding to (12) is given by:

   3 Machine learning method based on neural
J φ · ∇ δφ dΩ = R φ = 0. (14)
Ω network
The problem being nonlinear, an iterative procedure is Artificial neural networks (ANN), commonly referred to as
used to solve it. In the following, a Newton method is neural networks, are massively parallel computing systems
employed. Using a Taylor expansion at first order of R(φ) inspired by the biological neural network of the human brain
k
around a known solution φ (e.g. provided by a previous i.e. an ANN model attempts to use some organizational prin-
iteration), we have: ciples used in a biological neural network like in the human
brain. Thus, ANN are information-processing mathematical
k k k
R φ + Δφ  R φ + DΔφ R φ , (15) models configured to perform tasks through a learning pro-
cess and have been extensively used by researchers across
where DΔφ R(φ) is the Gateaux derivative, given by: different disciplines in order to solve a variety of problems
in pattern recognition, prediction, optimization and cluster-
k d   ing.
DΔφ R φ = R φ + αΔφ . (16) In direct analogy to its biological counterpart, an ANN is
dα α=0
a cluster of simple processors called neurons. An artificial
Then we obtain the linearized problem: neuron, labeled k in Fig. 2, is a simple mathematical model
consisting of three essential parts: (i) the set of j = 1, . . . , n
k k synapses, each of which is characterized by its own weight
DΔφ R φ = −R φ , (17)
wk j . A signal x j at the input of synapse j is multiplied by
with the synaptic weight wk j ,1 (ii) the summation function which
sums the input signals, weighted by the respective synaptic

k  strengths of the neurons and (iii) the activation function ϕ(·)
DΔφ R φ =− k(φ k )∇ Δφ · ∇(δφ)dΩ. (18)
Ω which is used for limiting the amplitude of the output of a
neuron and its responsible for introducing nonlinearity to the
More details can be found in [42]. Discretizing (17) by network. Typical activation functions are the threshold func-
finite elements gives a linear system of equations in the form: tion and the sigmoid function. This neuron model is called
the perceptron.
k k
KT φ Δφ = −R φ . (19) In mathematical terms we can describe the neuron k of
Fig. 2 as the set of equations
Then, the electric potential is updated according to:

n
k+1 k vk = wk j x j (23)
φ = φ + Δφ (20)
j=1

and (19) is solved until a convergence criterion is reached. In and


FE2 , the estimation of J and k in (14) and (18) is performed
by solving the nonlinear RVE problem (19) at each Gauss yk = ϕ(vk + bk ) (24)
point, which induces a high level of complexity. In contrast,
in the data-driven approach proposed in the following, the 1 Synaptic weights may range in positive as well as negative values.

123
Computational Mechanics (2019) 64:307–321 311

Fig. 2 Nonlinear model of a neuron, labeled k

where x1 , x2 , . . . , x j are the input signals; wk1 , wk2 , . . . , wkn (a) (b)
are the respective synaptic weights of neuron k and bk is the
bias which has the effect of applying an affine transformation Fig. 3 Two simple feed-forward ANNs, a a single-layer network. The
to the output vk of Eq. (23). square nodes represent the input nodes while the circular nodes represent
the basic processing neurons, b a multi-layer network with with j = 6
The architecture of an ANN is defined by the structure of source nodes, h 1 = 4 neurons in the first hidden layer and q = 2
neurons in relation to each other. In general, an ANN consists neurons in the output layer (6–4–2). Square nodes represent the input
of the input layer, the hidden layer(s) and the output layer. nodes while the circular nodes represent the basic processing neurons
Depending on the presence of one or more hidden layers we
can have a single-layer network (no hidden layer) or a multi-
layer network (one or more hidden layers), also called in explained next. In the present work, the Matlab Neural
the literature “Deep learning” approaches. Figure 3 depicts Network Toolbox has been used.
the two types of networks; Fig. 3a presents a simple single-
layer network with four input source nodes and four output
neurons and Fig. 3b shows a multi-layer perceptron (MLP)
( j − h 1 − · · · − q) with with j = 6 source nodes, h 1 = 4 4 Construction of the surrogate model based
neurons in the first hidden layer and q = 2 neurons in the on ANN
output layer.
Moreover, the direction of the information through the 4.1 Off-line RVE calculations: strategy
network is used in order to distinguishes different types of
networks. ANNs in which information moves only from the The key idea of this work is to construct a surrogate
input to output and not vice-versa are called feed-forward model based on Neural Networks in a multilevel finite
networks while in any other case they are called recurrent. element framework. In the present method, called Neural
One of the most important features of an ANN is its abil- Network Computational Homogenization (NN-CH), we use
ity to learn the task for which is designed through examples; a multi-layer perceptron (MLP) to construct the nonlinear
provide the network with a set of input/output signals called relationship between the (input) effective electric field vector
training samples and use it to create a mapping between the E = [E x , E y , E z ] and the (output) effective electric current
input and the output. Basically, the training of the ANN con- density vector J = [J x , J y , J z ]. As this relationship is not
sists of ordinary steps targeting to tune the synaptic weights known in advance, each set of values for E requires to solve a
so that a map that fits closely the training set is achieved.2 nonlinear finite element problem to obtain the corresponding
There exist three major learning strategies: (i) supervised J as defined in Sect. 2. To avoid theses costly calculations
learning, in which each training sample is composed of the at the macro scale, we propose here to define this relation-
input signals and their corresponding outputs, (ii) unsuper- ship through the ANN surrogate model. For this purpose, a
vised learning in which prior knowledge of the respective set of data consisting of N couples of vectors E and their
outputs is not required and (iii) reinforced learning in which corresponding J are provided to the neural network.
the learning algorithms adjust the weights relying on infor- The procedure for constructing the ANN-based surrogate
mation acquired through the interaction with the model. multiscale model is illustrated in Figs. 4 and 5 and outlined
In this work we utilize a supervised learning strategy as in the following steps:

2 Training of the ANN can be considered as an optimization problem, Step 1 Generate N number of electric field vectors E =
where the design variables are the synaptic weights. [E x , E y , E z ] to be used for the MLP training.

123
312 Computational Mechanics (2019) 64:307–321

Fig. 4 Off-line computational


in the data-driven computational
homogenization framework

Learning
step

...
Surrogate model
RVE (neural network)
nonlinear local N preliminary
problem calculations
on the RVE

Fig. 5 Online computational in the data-driven computational homog-


enization framework

Step 2 Run N FEM nonlinear RVE calculations, for applied Fig. 6 Illustration of the LHS sampling technique for two parameters
 (i)
effective electric field vectors E , i = 1, 2, . . . , N
and compute the corresponding effective current den-
 (i)
sities J , i = 1, 2, . . . , N as described in Sect. 2.
that we used for the effective electric field are provided in
Step 3 Train the MLP using {Ei }i=1 N as inputs and {J } N as
i i=1
the following.
the corresponding outputs.
The quality in the prediction obtained by the the trained
Step 4 Check for accuracy of the trained MLP.
ANN with respect to the number of data is strongly affected
Step 5 If the error (using appropriate metrics) is larger than
by the sampling strategy. In this work, we have used the Latin
a tolerance, add new input (through steps 1–2) to the
Hypercube Sampling (LHS) [47] strategy to define the initial
set of data and perform step 3–5 again.  (i)
set of components for the vectors E , i = 1, 2, . . . , N .
LHS is a statistical method for generating a near-random
The definition of the domain in which the values of sample of parameter values from a multidimensional distri-
the electric field vectors E are determined is important. bution. LHS has the advantage to converge faster than the
The graphene nanocomposites exhibit a linear behavior for crude Monte Carlo simulations and reduces the variance of
low amplitudes of applied effective electric field, nonlinear the statistical estimates.
behavior for larger values of the effective electric field, and An illustration of the LHS is provided in Fig. 6 for two
breakdown at extremely high values of effective electric field. parameters E x and E y . First, the parameter space is dis-
Thus the domain of the electric field vectors should contain cretized into P × P discrete values. Then a set of indices are
the nonlinear region but avoid the electric breakdown, which generated randomly such that for a new set (i, j) of indices,
is not included in our analysis. This step has to be checked i∈/ Si , j ∈/ S j , where S i and S j are sets of indices already
by preliminary simulations on the RVE. The range of values generated.

123
Computational Mechanics (2019) 64:307–321 313

Table 1 Properties of the trained ANN for each study case


Case (vol%) Architecture Number of hidden layers Number of neurons/layer Number of epochs MSE (validation error)

0.53 Feedforward (3–9–3) 1 9 283 4.66 × 10−5


0.66 Feedforward (3–11–3) 1 11 256 2.64 × 10−5
0.79 Feedforward (3–13–3) 1 13 447 9.64 × 10−4
0.92 Feedforward (3–15–3) 1 15 495 8.19 × 10−4
1.05 Feedforward (3–13–3) 1 13 546 9.94 × 10−3
1.19 Feedforward (3–17–3) 1 17 885 2.24 × 10−3
1.32 Feedforward (3–18–3) 1 18 546 5.32 × 10−3
1.58 Feedforward (3–22–3) 1 22 885 8.04 × 10−2

It is worth noting that in the present study, the electric regression technique. As a result, an output J can be gener-
response ANN surrogate model was constructed separately ated for any input E belonging to the above ranges of values
for the different volume fractions, even though the volume by the ANN.
fraction could have been considered as an additional input Table 1 depicts the properties of the trained ANN for each
in the ANN model. This point will be considered in a future case study. One common line along these ANN is the selec-
study. tion of the architecture for which we selected a feedforward
architecture with one hidden layer. The justification is moti-
4.2 Numerical parameters vated by the universal approximation theorem [49] which
states that a feed-forward network, with a single hidden layer,
A series of realizations are taken for graphene volume frac- containing a finite number of neurons, can approximate con-
tion ranging between 0.53 vol% and 1.58 vol% as shown tinuous functions accurately. Moreover, since the first version
in Fig. 1. The graphene sheets are perpendicular to the z- of this theorem was applied for sigmoid activation functions
direction within the RVE, and their positions are randomly we have also selected this type of activation function
distributed with uniform probability law, resulting in an
2
anisotropic effective electric behavior. The size of RVE is 80 f (x) = −1 (26)
nm, which has been shown in a previous study [42] as being 1 + e−2x
a converged size with respect to the effective properties. The which is nonlinear in nature as well as its combinations. We
graphene sheets all have the same dimensions, and are par- used the Levenberg–Marquardt algorithm for the training,
allelepipedic domains with dimensions 15 × 15 × 0.3 nm3 , which is a blend of the steepest descent method and the
leading to an aspect ratio of 50. The barrier height between Gauss-Newton algorithm, and which converges much faster
graphene and polymer Φ0 is 0.17 eV, e = 1.60 × 10−19 C, than these two methods. Table 1 shows the optimal number
m = 9.11×10−31 kg, h p = 6.62×10−34 J s in (4). The elec- of neurons of the hidden layer for each case study, the num-
tric conductivity of the polymer is taken as k0 = 1 × 10−10
p
ber of epochs (one forward pass and one backward pass of
S/m and the conductivity of graphene is given by [48] as: all the training examples), as well as the mean squared error
⎡ ⎤ (MSE) of the trained ANN over the validation data. It is clear
8.32 × 104 0 0 from these results that the ANN were not over-fitted since the
k =
g ⎣ 0 8.32 × 104 0 ⎦ S/m. (25) MSE error is significantly small.
0 0 83.2 In what follows, we validate the accuracy of the ANN
surrogate model by comparing the values of J obtained by
4.3 Validation of the surrogate model based on ANN the ANN and the corresponding values computed by full RVE
calculations.
For each realization, 500 applied electric field vectors E
are randomly generated by LHS algorithm, where E x ∈ 4.3.1 Unidirectional electric field
[−0.012, 0.012] V/nm, E y ∈ [−0.012, 0.012] V/nm, and
E z ∈ [0, 0.012] V/nm. The corresponding effective current First, a uniaxial effective electric field is considered, i.e. the
density J are calculated by finite element method following components of effective electric field E y and E z are set to be
the procedure described in Sect. 2 for each set of effective 0, and E x ranges from 0 to 0.012 V/nm with a step of 0.0003
electric field components. The 500 training data (E and J ) V/nm. For each set of values, we compute both the refer-
are used to train the ANN through the multilayer perceptron ence solution by FEM and the solution obtained by the ANN

123
314 Computational Mechanics (2019) 64:307–321

surrogate model. Results are provided in Figs. 7 and 8. For −7 0.53 vol%
x 10
2.5
small graphene volume fraction (0.53 vol% and 0.66 vol%), FEM calculations
ANN model
the electric behavior is linear, while after reaching the per- 2
colation threshold, the relation between current density and

Jx (A/cm2)
1.5
electric field turns out to be nonlinear. We can note that in all
cases, the agreement between the finite element result and 1
the ANN solution is good.
0.5

4.3.2 Bidirectional electric field 0


0 0.002 0.004 0.006 0.008 0.01 0.012
E (V/nm)
In a second example, we prescribe a bidirectional effective x

electric field to the RVE. More specifically, E z is set to be (a)


zero, and E x = E y ranges from 0 to 0.012 V/nm simulta- x 10
−7 0.66 vol%
neously with a step of 0.0003 V/nm. As shown in Figs. 9 6
FEM calculations
and 10, the predictions of current density by the ANN sur- 5 ANN model
rogate model are still in very good agreement in this second
4

Jx (A/cm2)
case.
3

2
5 Numerical examples
1

In the following presented examples, a composite material 0


0 0.002 0.004 0.006 0.008 0.01 0.012
whose microstructure is characterized by an RVE containing Ex (V/nm)
randomly distributed, aligned graphene sheets, is consid-
(b)
ered (see Fig. 1). By structural calculation we do not mean
0.79 vol%
mechanical calculations, but only study the electric distri- 0.8
bution within the structure at the macro scale. The structure FEM calculations
ANN model
at the macroscale is discretized with linear finite elements. 0.6
In the FE2 method, each integration point of the structure
Jx (A/cm2)

corresponds to an RVE. Four volume fractions of graphene 0.4


sheets are considered: 0.53 vol%, 0.79 vol%, 0.92 vol% and
1.58 vol%. At the microscale, the nonlinear model described 0.2
in Sect. 2.1 is considered. The volume fraction 0.53 vol% is
below the percolation threshold, and exhibits only a linear 0
0 0.002 0.004 0.006 0.008 0.01 0.012
behavior. The other volume fractions are around or above
E (V/nm)
x
the percolation threshold, and involve a nonlinear behavior
(c)
(see Figs. 7, 8, 9, 10). Each RVE contains around 40,000
tetrahedral elements. Off-line computations have been car- 0.92 vol%
0.8
ried out according to Sect. 4 to create the initial set of data. FEM calculations
ANN model
Then, the learning phase has been carried out according to 0.6
the procedure described in Sect. 3. The off-line calculation
Jx (A/cm2)

times for each volume fraction are indicated in Table 2. It 0.4


is worth noting that the off-line computation time increases
with the graphene volume fraction. This phenomenon results 0.2
from two reasons: firstly, as the volume fraction increases, a
finer mesh is required to discretize the RVE; secondly, as the 0
volume fraction increases, the localization of electric field 0 0.002 0.004 0.006 0.008 0.01 0.012
Ex (V/nm)
around graphene sheets increases and the nonlinear behav-
ior is more pronounced, inducing slower convergence of the (d)
Newton algorithm. However, (i) the data obtained from RVE Fig. 7 Comparisons between direct simulations obtained by nonlinear
calculations are independent and can be reduced drastically FEM calculations on the RVE and the surrogate ANN model: values of
by distributing them on as many processors as available; (b) J x as a function of a unidirectional effective electric field E x , E y =
once obtained, the surrogate model can be used for many E z = 0; a 0.53 vol%, b 0.66 vol%, c 0.79 vol% and d 0.92 vol%

123
Computational Mechanics (2019) 64:307–321 315

1.05 vol% −7
x 10 0.53 vol%
40 2.5
FEM calculations FEM calculations
ANN model ANN model
2
30
Jx (A/cm2)

Jx (A/cm2)
1.5
20
1

10
0.5

0 0
0 0.002 0.004 0.006 0.008 0.01 0.012 0 0.002 0.004 0.006 0.008 0.01 0.012
Ex (V/nm)
Ex (V/nm)
(a)
(a)
−7 0.66 vol%
1.19 vol% x 10
15 5
FEM calculations FEM calculations
ANN model ANN model
4
10

Jx (A/cm2)
3
Jx (A/cm2)

5
2

0 1

0
−5 0 0.002 0.004 0.006 0.008 0.01 0.012
0 0.002 0.004 0.006 0.008 0.01 0.012
E (V/nm)
x
Ex (V/nm)
(b)
(b)
0.79 vol%
1.32 vol% 2
80 FEM calculations
FEM calculations ANN model
ANN model 1.5
60
Jx (A/cm2)
Jx (A/cm2)

40 1

20
0.5
0

0
−20 0 0.002 0.004 0.006 0.008 0.01 0.012
0 0.002 0.004 0.006 0.008 0.01 0.012
Ex (V/nm)
Ex (V/nm)
(c)
(c)
0.92 vol%
1.58 vol% 40
300 FEM calculations
FEM calculations ANN model
250 30
ANN model
200
Jx (A/cm2)

20
Jx (A/cm2)

150
10
100

50 0

0
−10
0 0.002 0.004 0.006 0.008 0.01 0.012
−50
0 0.002 0.004 0.006 0.008 0.01 0.012 E (V/nm)
x
E (V/nm)
x
(d)
(d)
Fig. 9 Comparisons between direct simulations obtained by nonlinear
Fig. 8 Comparisons between direct simulations obtained by nonlinear FEM calculations on the RVE and the surrogate ANN model: values of
FEM calculations on the RVE and the surrogate ANN model: values of J x as a function of a bi-directional effective electric field E x = E y and
J x as a function of a unidirectional effective electric field E x , E y = E z = 0; a 0.53 vol%, b 0.66 vol%, c 0.79 vol% and d 0.92 vol%
E z = 0; a 1.05 vol%, b 1.19 vol%, c 1.32 vol% and d 1.58 vol%

123
316 Computational Mechanics (2019) 64:307–321

1.05 vol%
Table 2 Off-line computational times for the different volume fractions
40
FEM calculations
ANN model
30 Volume fraction (vol%) Off-line computation time (h)
Jx (A/cm2)

20 0.53 12.5
0.79 47.2
10
0.92 48.9
0 1.58 120

−10
0 0.002 0.004 0.006 0.008 0.01 0.012
Ex (V/nm)
(a)
1.19 vol%
120
FEM calculations
100
ANN model

80
Jx (A/cm2)

60 (a) (b)
40 Fig. 11 Validation example: structure geometry and boundary condi-
tions
20

0 Φ (V)
0 0.002 0.004 0.006 0.008 0.01 0.012 Z
X 22.5
E (V/nm)
x
16.875
(b)
1.32 vol% 11.25
200
FEM calculations Y 5.625
ANN model
150 0
Jx (A/cm2)

Fig. 12 Electric potential field in the homogeneous structure with


100
potential difference on the corresponding end surfaces [φ(x = 0 µm) =
0 V, φ(x = 2.5 µm) = 22.5 V] by NN-CH approach. J = 0 on the
50 other boundary surfaces. Graphene volume fraction is 0.79 vol% in the
microscopic RVE
0
0 0.002 0.004 0.006 0.008 0.01 0.012
E (V/nm) analyses on different macroscopic structures, as shown in
x
(c) the next examples.
1.58 vol%
500
FEM calculations
ANN model 5.1 Validation and comparison between FE2 and
400
NN-CH
Jx (A/cm2)

300
In this first example a parallelepipedic structure, is consid-
200 ered. The geometry and boundary conditions are described
100
in Fig. 11a. The dimensions are L x = 2.5 µm, L y =
L z = 0.5 µm. A potential difference is applied on the cor-
0
0 0.002 0.004 0.006 0.008 0.01 0.012
responding end surfaces such as φ(x = 0 µm) = 0 V,
E (V/nm)
x
φ(x = 2.5 µm) = 22.5 V. The current density J = 0 on
(d) the other boundary surfaces.
The structure is discretized with 64 tetrahedral elements.
Fig. 10 Comparisons between direct simulations obtained by nonlinear
We apply both FE2 and NN-CH to solve the nonlinear mul-
FEM calculations on the RVE and the surrogate ANN model: values of
J x as a function of a bi-directional effective electric field E x = E y and tiscale problem. The obtained electric potential obtained by
E z = 0; a 1.05 vol%, b 1.19 vol%, c 1.32 vol% and d 1.58 vol% the NN-CH is shown in Fig. 12, for an RVE with 0.79 vol%.
The local electric field in an RVE is shown in Fig. 13, where
we can observe the concentration of electric field between

123
Computational Mechanics (2019) 64:307–321 317

Fig. 13 Local electric field in the micro-scaled RVE of an element with


0.79 vol% graphene. The values below 0.02 V/nm is set to be zero (as
shown in scale bar) (a)

Table 3 Comparison of computation times and macroscopic current


density component J˜x between FE2 and NN-CH, for the solid sample
in Fig. 12

Volume fraction Method Computational Effective J˜x


(vol%) time (A/cm2 )

0.53 FE2 6.9 h 1.792e–7


NN-CH 10.73 s 1.789e–7
0.79 FE2 32.2 h 0.139 (b)
NN-CH 10.83 s 0.153
0.92 FE2 35.5 h 0.098
NN-CH 11.80 s 0.107
1.58 FE2 42.1 h 39.897
NN-CH 10.08 s 44.586

(c)

Fig. 14 Structural problem: geometry, boundary conditions and mesh

the neighboring graphene sheets. A comparison


 of the effec-
tive electric current density J˜x = J x (x) for both methods
is provided in Table 3 for the 4 considered volume fractions.
We can note that that accuracy of the NN-CH is very good (d)
with respect to the reference solution at low graphene volume
fraction (0.53 vol%), which is provided by the FE2 method. Fig. 15 NN-CH simulation results compared with FE2 results for the
structural composite structure, with 0.53 vol% of graphene within the
As for the other samples with higher volume fractions, the
RVE. a Electric potential field, b Electric field component E x on (z =
NN-CH results present the same magnitude with FE2 method 1 µm) slice obtained by NN-CH simulation. c Electric potential field
and the error is about 10%. Next, the computational times and d electric field component E x on (z = 1 µm) slice obtained by FE2
are compared in Table 3. if we do not count the off-line cal- computation

123
318 Computational Mechanics (2019) 64:307–321

Table 4 Comparison of computation time between conventional FE2


and NN-CH for the complex structure
Volume fraction (vol%) Method Computation time

0.53 FE2 126 h


NN-CH 267.45 s
0.79 FE2 –
NN-CH 345.56 s
(a)
0.92 FE2 –
NN-CH 366.08 s
1.58 FE2 –
NN-CH 391.49 s

(b)

(a)

(c)

(b)
Fig. 17 Current density component Jx field for the composite with
0.79 vol% of graphene. a Overall view. b View on (x = 8) µm, (y =
4) µm and (z = 1) µm slices

(d) be solved without new calculations on the RVE and taking


Fig. 16 NN-CH simulation results compared with FE2 results for the advantage of the off-line calculation costs. In addition, due
structural composite structure, with 0.53 vol% of graphene within the to its efficiency which is close to classical FEM, meshes with
RVE. a Overall current density component Jx field, b Jx on slices arbitrary sizes can be considered in 3D, which is nowadays
(x = 8 µm, y = 4 µm, z = 1 µm) computed by NN-CH simulation.
not tractable for FE2 methods. In this next example, we con-
c Overall current density component Jx field and d Jx on slices (x =
8 µm, y = 4 µm, z = 1 µm) computed by FE2 simulation sider the structure described in Fig. 14. The side lengths of
the domain are L x = 16 µm, L y = 8 µm, L z = 2 µm. The
culation times, we can note that drastic computational time diameter of the hole in the middle of the cuboid is D = 4 µm.
reduction can be obtained: for example for the case 1.58 %, A mesh of 1820 tetrahedral elements is constructed. The
the FE2 calculation takes 42.1 h against 10s for the NN-CH, boundary condition is φ(x = 0 µm) = 0 V, φ(x = 16 µm) =
representing a speedup of the order of 104 . 160 V on the corresponding end surfaces. The current density
J = 0 on the other boundary surfaces.
5.2 Structural application Results for an RVE of 0.53 vol% are shown in Figs. 15
and 16. More specifically, the potential field and electric field
One advantage of the NN-CH approach is that once the surro- component E x by NN-CH method are shown in Fig. 15a, b.
gate model is constructed, then many structural problems can The current density on the surface is shown in Fig. 16a. In

123
Computational Mechanics (2019) 64:307–321 319

is provided in Figs. 15c, d and 16c, d, showing a very good


agreement.
The results for the other volume fractions 0.79 vol%,
0.92 vol% and 1.58 vol% are provided in Figs. 17, 18 and 19.
For these calculations, we did not perform FE2 calculations
due to predicted unreasonable computational times. Compu-
tational times for NN-CH are provided in Table 4.
(a)

6 Conclusion

In this work, we have proposed a data-driven multiscale


method based on Neural Networks to drastically reduce the
computational times in nonlinear multiscale analyses. The
main idea is to compute a set of preliminary, off-line calcula-
tions which are used to construct a surrogate model provided
by a Neural Network. More specifically, a series of nonlinear
(b) electric conduction calculation have been performed on RVE
Fig. 18 Current density component Jx field for the composite with
containing graphene sheets embedded in a polymer matrix.
0.92 vol% graphene. a Overall view. b View on (x = 8) µm, (y = The nonlinear effects are due to a model of electric tunnelling
4) µm and (z = 1) µm slices effect at the scale of graphene sheets. Then, for a series of
applied effective electric fields, the corresponding effective
electric current densities are computed numerically by finite
elements on the RVE. The whole set of data are then used to
train a Neural Network which is able to provide the effective
electric current density for a given applied effective electric
field. At the macroscopic scale, the nonlinear problem can
then be solved without any new RVE calculation to provide
the constitutive relationship at each Gauss point, by using
the surrogate model instead. We have shown than drastic
(a) computational savings (of the order of 104 ) can be obtained
during the online calculations, reducing the FE2 calculations
times to the ones of a simple FEM analysis with equivalent
accuracy.

Acknowledgements The financial support from Institut Universitaire


de France (IUF) is gratefully acknowledged for JY.

References
(b)
1. Feyel F (1999) Multiscale FE2 elastoviscoplastic analysis of com-
Fig. 19 Current density component Jx field for the composite with posite structure. Comput Mater Sci 16(1–4):433–454
1.58 vol% graphene. a Overall view. b View on (x = 8) µm, (y = 2. Feyel F, Chaboche JL (2000) FE2 multiscale approach for mod-
4) µm and (z = 1) µm slices elling the elastoviscoplastic behaviour of long fibre SiC/Ti compos-
ite materials. Comput Methods Appl Mech Eng 183(3–4):309–330
3. Smit R, Brekelmans W, Meijer H (1998) Prediction of the mechan-
ical behavior of nonlinear heterogeneous systems by multi-level
Fig. 16b, a local view on slices in three different directions is finite element modeling. Comput Methods Appl Mech Eng 155(1–
provided. For this case, we have compared the computational 2):181–192
4. Miehe C, Schröder J, Schotte J (1999) Computational homogeniza-
times for both FE2 and NN-CH. For NN-CH, the calculation
tion analysis in finite plasticity simulation of texture development in
took 267.45 s with 12 parallel workers, against 126 h for FE2 polycrystalline materials. Comput Methods Appl Mech Eng 171(3–
(see Table 4). A comparison of the different fields by FE2 4):387–418

123
320 Computational Mechanics (2019) 64:307–321

5. Terada K, Kikuchi N (2001) A class of general algorithms for multi- 26. Peherstorfer B, Willcox K (2015) Dynamic data-driven reduced-
scale analysis of heterogeneous media. Comput Methods Appl order models. Comput Methods Appl Mech Eng 291:21–41
Mech Eng 190:5427–5464 27. Kirchdoerfer T, Ortiz M (2016) Data-driven computational
6. Kouznetsova V, Brekelmans W, Baaijens F (2001) An approach to mechanics. Comput Methods Appl Mech Eng 304:81–101
micro–macro modeling of heterogeneous materials. Comput Mech 28. Versino D, Tonda A, Bronkhorst CA (2017) Data driven modeling
27(1):37–48 of plastic deformation. Comput Methods Appl Mech Eng 318:981–
7. Ghosh S, Lee K, Raghavan P (2001) A multi-level computational 1004
model for multi-scale damage analysis in composite and porous 29. Kirchdoerfer T, Ortiz M (2017) Data driven computing with noisy
materials. Int J Solids Struct 38(14):2335–2385 material data sets. Comput Methods Appl Mech Eng 326:622–641
8. Ibrahimbegović A, Markovič D (2003) Strong coupling meth- 30. Ibañez R, Abisset-Chavanne E, Aguado JV, Gonzalez D, Cueto E,
ods in multi-phase and multi-scale modeling of inelastic behavior Chinesta F (2018) A manifold learning approach to data-driven
of heterogeneous structures. Comput Methods Appl Mech Eng computational elasticity and inelasticity. Arch Comput Methods
192(28–30):3089–3107 Eng 25(1):47–57
9. Yvonnet J, He QC (2007) The reduced model multiscale method 31. Nguyen LTK, Keip MA (2018) A data-driven approach to nonlinear
(R3M) for the non-linear homogenization of hyperelastic media at elasticity. Comput Struct 194:97–115
finite strains. J Comput Phys 223(1):341–368 32. Waszczyszyn Z, Ziemiański L (2001) Neural networks in mechan-
10. Ibrahimbegovic A, Papadrakakis M (2010) Multi-scale models and ics of structures and materials—new results and prospects of
mathematical aspects in solid and fluid mechanics. Comput Meth- applications. Comput Struct 79(22–25):2261–2276
ods Appl Mech Eng 199:1241–1241 33. Zhao H, Bai J (2015) Highly sensitive piezo-resistive graphite
11. Fullwood DT, Niezgoda SR, Adams BL, Kalidindi SR (2010) nanoplatelet-carbon nanotube hybrids/polydimethylsilicone com-
Microstructure sensitive design for performance optimization. posites with improved conductive network construction. ACS Appl
Progr Mater Sci 55(6):477–562 Mater Interfaces 7(18):9652–9659
12. Papadopoulos V, Impraimakis M (2017) Multiscale modeling of 34. Park OK, Jeevananda T, Kim NH, Kim S, Lee JH (2009) Effects of
carbon nanotube reinforced concrete. Compos Struct 182:251–260 surface modification on the dispersion and electrical conductivity
13. Geers MG, Kouznetsova VG, Brekelmans W (2010) Multi-scale of carbon nanotube/polyaniline composites. Scr Mater 60(7):551–
computational homogenization: trends and challenges. J Comput 554
Appl Math 234(7):2175–2182 35. Jia J, Sun X, Lin X, Shen X, Mai YW, Kim JK (2014) Exceptional
14. Geers M, Yvonnet J (2016) Multiscale modeling of microstructure- electrical conductivity and fracture resistance of 3d interconnected
property relations. MRS Bull 41(08):610–616 graphene foam/epoxy composites. ACS Nano 8(6):5774–5783
15. Hernández J, Oliver J, Huespe AE, Caicedo M, Cante J (2014) 36. Wang T, Liang G, Yuan L, Gu A (2014) Unique hybridized
High-performance model reduction techniques in computational graphene and its high dielectric constant composites with enhanced
multiscale homogenization. Comput Methods Appl Mech Eng frequency stability, low dielectric loss and percolation threshold.
276:149–189 Carbon 77:920–932
16. Zahr MJ, Avery P, Farhat C (2017) A multilevel projection-based 37. Wei T, Luo G, Fan Z, Zheng Yan J, Yao C, Li W, Zhang C (2009)
model order reduction framework for nonlinear dynamic multiscale Preparation of graphene nanosheet/polymer composites using in
problems in structural and solid mechanics. Int J Numer Methods situ reduction-extractive dispersion. Carbon 47(9):2296–2299
Eng 112(8):855–881 38. Pang H, Chen T, Zhang G, Zeng B, Li ZM (2010) An electrically
17. Liu Z, Bessa M, Liu WK (2016) Self-consistent clustering anal- conducting polymer/graphene composite with a very low percola-
ysis: an efficient multi-scale scheme for inelastic heterogeneous tion threshold. Mater Lett 64(20):2226–2229
materials. Comput Methods Appl Mech Eng 306:319–341 39. Wang J, Yu S, Luo S, Chu B, Sun R, Wong CP (2016) Investiga-
18. Yvonnet J, Gonzalez D, He QC (2009) Numerically explicit poten- tion of nonlinear I–V behavior of CNTs filled polymer composites.
tials for the homogenization of nonlinear elastic heterogeneous Mater Sci Eng B 206:55–60
materials. Comput Methods Appl Mech Eng 198(33–36):2723– 40. Martin-Gallego M, Bernal MM, Hernandez M, Verdejo R, López-
2737 Manchado MA (2013) Comparison of filler percolation and
19. Le BA, Yvonnet J, He QC (2015) Computational homogenization mechanical properties in graphene and carbon nanotubes filled
of nonlinear elastic materials using neural networks. Int J Numer epoxy nanocomposites. Eur Polym J 49(6):1347–1353
Methods Eng 104(12):1061–1084 41. Zeng X, Xu X, Kovalev PMSE, Baudot C, Mathews N, Zhao Y
20. Bessa M, Bostanabad R, Liu Z, Hu A, Apley DW, Brinson C, (2011) Characteristics of the electrical percolation in carbon nan-
Chen W, Liu WK (2017) A framework for data-driven analysis of otubes/polymer nanocomposites. J Phys Chem C 115(44):21685–
materials under uncertainty: countering the curse of dimensionality. 21690
Comput Methods Appl Mech Eng 320:633–667 42. Lu X, Yvonnet J, Detrez F, Bai J (2017) Multiscale modeling of
21. Fritzen F, Kunc O (2017) Two-stage data-driven homogenization nonlinear electric conductivity in graphene-reinforced nanocom-
for nonlinear solids using a reduced order model. Eur J Mech A posites taking into account tunnelling effect. J Comput Phys
Solids 69:201–220 337:116–131
22. Ghaboussi J, Garrett J Jr, Wu X (1991) Knowledge-based modeling 43. Yvonnet J, He QC, Toulemonde C (2008) Numerical modelling of
of material behavior with neural networks. J Eng Mech 117(1):132– the effective conductivities of composites with arbitrarily shaped
153 inclusions and highly conducting interface. Compos Sci Technol
23. Furukawa T, Yagawa G (1998) Implicit constitutive modelling for 68(13):2818–2825
viscoplasticity using neural networks. Int J Numer Methods Eng 44. Lu X, Yvonnet J, Detrez F, Bai J (2018) Low electrical percolation
43(2):195–219 thresholds and nonlinear effects in graphene-reinforced nanocom-
24. Lefik M, Schrefler B (2003) Artificial neural network as an incre- posites: a numerical analysis. J Compos Mater 52(20):2767–2775
mental non-linear constitutive model for a finite element code. 45. Simmons JG (1963) Electric tunnel effect between dissimilar elec-
Comput Methods Appl Mech Eng 192(28–30):3265–3283 trodes separated by a thin insulating film. J Appl Phys 34(9):2581–
25. Yun GJ, Ghaboussi J, Elnashai AS (2008) A new neural network- 2590
based model for hysteretic behavior of materials. Int J Numer 46. Gu S, He QC (2011) Interfacial discontinuity relations for cou-
Methods Eng 73(4):447–469 pled multifield phenomena and their application to the modeling

123
Computational Mechanics (2019) 64:307–321 321

of thin interphases as imperfect interfaces. J Mech Phys Solids 49. Cybenko G (1989) Approximation by superpositions of a sigmoidal
59(7):1413–1426 function. Math Control Signals Syst 2(4):303–314
47. Eglajs V, Audze P (1977) New approach to the design of multifactor
experiments. Probl Dyn Strengths 35(1):104–107
48. Stankovich S, Dikin DA, Dommett GHB, Kohlhaas KM, Zimney
Publisher’s Note Springer Nature remains neutral with regard to juris-
EJ, Stach EA, Piner RD, Nguyen ST, Ruoff RS (2006) Graphene-
dictional claims in published maps and institutional affiliations.
based composite materials. Nature 442(7100):282–286

123

You might also like