Journal Pre-proof

Deep learning-accelerated computational framework based on Physics

Informed Neural Network for the solution of linear elasticity

Arunabha M. Roy, Rikhi Bose

PII: S0893-6080(23)00131-4
DOI: https://doi.org/10.1016/j.neunet.2023.03.014
Reference: NN 5581

To appear in: Neural Networks

Received date : 5 July 2022

Revised date : 7 February 2023
Accepted date : 8 March 2023

Please cite this article as: A.M. Roy and R. Bose, Deep learning-accelerated computational
framework based on Physics Informed Neural Network for the solution of linear elasticity. Neural
Networks (2023), doi: https://doi.org/10.1016/j.neunet.2023.03.014.

This is a PDF file of an article that has undergone enhancements after acceptance, such as the
addition of a cover page and metadata, and formatting for readability, but it is not yet the definitive
version of record. This version will undergo additional copyediting, typesetting and review before it
is published in its final form, but we are providing this version to give early visibility of the article.
Please note that, during the production process, errors may be discovered which could affect the
content, and all legal disclaimers that apply to the journal pertain.

© 2023 Published by Elsevier Ltd.

 Journal Pre-proof

Deep learning-accelerated computational framework based on

Physics Informed Neural Network for the solution of linear

of
elasticity

Arunabha M. Roy∗1 and Rikhi Bose2

pro
1
Aerospace Engineering, University of Michigan, Ann Arbor, MI 48109,
U.S.A.
2
Mechanical Engineering, Johns Hopkins University, Baltimore, MD 21218,
U.S.A.

re- Abstract

The paper presents an efficient and robust data-driven deep learning (DL) computational
lP
framework developed for linear continuum elasticity problems. The methodology is based on
the fundamentals of the Physics Informed Neural Networks (PINNs). For an accurate repre-
sentation of the field variables, a multi-objective loss function is proposed. It consists of terms
corresponding to the residual of the governing partial differential equations (PDE), constitu-
rna

tive relations derived from the governing physics, various boundary conditions, and data-driven
physical knowledge fitting terms across randomly selected collocation points in the problem do-
main. To this end, multiple densely connected independent artificial neural networks (ANNs),
each approximating a field variable, are trained to obtain accurate solutions. Several bench-
mark problems including the Airy solution to elasticity and the Kirchhoff-Love plate problem
Jou

are solved. Performance in terms of accuracy and robustness illustrates the superiority of the
current framework showing excellent agreement with analytical solutions. The present work
combines the benefits of the classical methods depending on the physical information available
in analytical relations with the superior capabilities of the DL techniques in the data-driven
construction of lightweight, yet accurate and robust neural networks. The models developed
herein can significantly boost computational speed using minimal network parameters with
∗
Corresponding author

1
 Journal Pre-proof
2

easy adaptability in different computational platforms.

Keywords: Physics Informed Neural Networks (PINNs); Artificial neural networks (ANNs);
Linear elasticity; Bi-harmonic equations; Deep learning (DL)

of
1. Introduction :

pro
In recent years, driven by the advancement of bigdata-based architectures (Khan et al., 2022a),
deep learning (DL) techniques (LeCun et al., 2015) have shown great promises in computer
vision (Voulodimos et al., 2018; Roy and Bhaduri, 2021; Roy et al., 2022c; Roy and Bhaduri,
2022; Roy et al., 2022a), object detection (Zhao et al., 2019; Chandio et al., 2022; Roy et al.,
re-
2022b; Singh et al., 2023a), image classification (Rawat and Wang, 2017; Irfan et al., 2021;
Jamil et al., 2022; Khan et al., 2022b), damage detection (Guo et al., 2022; Glowacz, 2022,
2021) brain-computer interfaces (Roy, 2022b,a,c; Singh et al., 2023b) and across various scien-
tific applications (Butler et al., 2018; Ching et al., 2018; Bose and Roy, 2022).
lP
The success of these methods, such as various classes of Neural Networks (NNs), can be
largely attributed to their capacity in excavating large volumes of data in establishing com-
plex high-dimensional non-linear relations between input features and output (Kutz, 2017).
However, the availability of sufficient data is a major bottleneck for analyzing various com-
rna

plex physical systems (Butler et al., 2018; Ching et al., 2018). Consequently, the majority of
state-of-the-art machine learning algorithms lack robustness in predicting these systems. Upon
availability of sufficient data, these have also garnered considerable success in problems gov-
erned by physics, such as dynamical systems (Dana and Wheeler, 2020), geosciences (DeVries
et al., 2018; Bergen et al., 2019; Racca and Magri, 2021; Saha et al., 2021; Jahanbakht et al.,
Jou

2022), material science and informatics (Butler et al., 2018; Ramprasad et al., 2017; Batra
et al., 2021; Määttä et al., 2021), fluid mechanics (Kutz, 2017; Brunton et al., 2020), various
constitutive modeling (Tartakovsky et al., 2018; Xu et al., 2021), etc. Their applicability how-
ever may be further enhanced by utilizing physical information available by mathematical/
analytical means. The recent endeavor of scientific and engineering community has been in
attempting to incorporate such physical information within their predictive scheme in small
 Journal Pre-proof
3

data regimes.

The incorporation of physical information into the DL framework may have several advantages.
First, as previously mentioned, in absence of sufficient data, it may be possible to solely utilize

of
physical knowledge for solving such problems (Raissi et al., 2019), or to the least, enhance
solutions in a data-driven predictive scheme (Raissi et al., 2020; Karniadakis et al., 2021). For
example, in Sirignano and Spiliopoulos (2018), a high-dimensional Hamilton–Jacobi–Bellman

pro
PDE has been solved by approximating the solution with a DNN trained to satisfy the differ-
ential operator, initial condition, and boundary conditions. In incompressible fluid mechanics,
the use of the solenoidality condition of the velocity fields restricts the solution space of the
momentum equations. Therefore, this condition may be used as a constraint for solving the
re-
governing equations (conventional solvers are generally developed in a way to satisfy this con-
straint through the Poisson equation for pressure), or at least improve the predictions in a
data-driven approach. Second, physical systems are often governed by laws that must satisfy
certain properties, such as invariance under translation, rotation, reflection, etc. In a purely
lP
data-driven approach, it is almost impossible for a DL algorithm to inherit those properties
entirely from data without explicit external forcing. Embedding such properties in the DL
algorithm might automatically improve the accuracy of the predictions. For example, Ling
et al. (2016) used a Tensor-based Neural Network (TBNN) to embed Galilean invariance that
rna

improved NN models for Reynolds-averaged Navier Stokes (RANS) simulations for the pre-
diction of turbulent flows. And lastly, any scientific problem is governed by some underlying
mechanism dictated by physical laws. Neglect of such physical information in a purely data-
driven framework in the current state of affairs is, therefore, an unsophisticated approach, if
not an ignorant one.
Jou

Partial differential equations (PDEs) represent underlying physical processes governed by first
principles such as conservation of mass, momentum, and energy. In most cases, analytical so-
lutions to these PDEs are not obtainable. Various numerical methods, such as finite-difference
(Sengupta, 2013), finite element (FE) (Zienkiewicz and Taylor, 2005), Chebyshev and Fourier
spectral methods (Boyd, 2001), etc are used to obtain approximate solutions. However, such
 Journal Pre-proof
4

techniques are often computationally expensive and suffer from various sources of errors due to
the complex nature of the underlying PDEs, numerical discretization and integration schemes,
iterative convergence techniques, etc. Moreover, the solution of inverse problems is the cur-
rent endeavor of the engineering community which requires complex formulations and is often

of
prohibitively expensive computationally. The use of the NNs in solving/modeling the PDEs
governing physical processes in a forward/ inverse problem is an important challenge worth
pursuing, as these methods have the capacity to provide accurate solutions using limited com-

pro
putational resources in a significantly robust framework relative to the conventional methods.
In this paper, we explore the possibility of using NN to obtain solutions to such PDEs govern-
ing linear continuum elasticity problems applicable in solid mechanics.

re-
There has been a recent thrust in developing machine learning (ML) approaches to obtain
the solution of governing PDEs (Karniadakis et al., 2021; von Rueden et al., 2019). The idea is
to combine traditional scientific computational modeling with a data-driven ML framework to
embed scientific knowledge into neural networks (NNs) to improve the performance of learning
lP
algorithms (Lagaris et al., 1998; Raissi and Karniadakis, 2018; Karniadakis et al., 2021). The
Physics Informed Neural Networks (PINNs) (Lagaris et al., 1998; Raissi et al., 2019, 2020) were
developed for the solution and discovery of nonlinear PDEs leveraging the capabilities of deep
neural networks (DNNs) as universal function approximators achieving considerable success in
rna

solving forward and inverse problems in different physical problems such as fluid flows (Sun
et al., 2020; Jin et al., 2021), multi-scale flows (Lou et al., 2021), heat transfer (Cai et al., 2021;
Zhu et al., 2021), poroelasticity (Haghighat et al., 2022), material identification (Shukla et al.,
2021), geophysics (bin Waheed et al., 2021, 2022), supersonic flows (Jagtap et al., 2022), and
various other applications (Waheed et al., 2020; Bekar et al., 2022). Contrary to traditional
Jou

DL approaches, PINNs force the underlying PDEs and the boundary conditions in the solution
domain ensuring the correct representation of governing physics of the problem. Learning of
the governing physics is ensured by the formulation of the loss function that includes the un-
derlying PDEs; therefore labeled data to learn the mapping between inputs and outputs is no
more necessary. Such architectural construction can be utilized for complex forward and in-
verse (finding parameters) solutions for various systems of ODEs and PDEs (Karniadakis et al.,
 Journal Pre-proof
5

2021). Additionally, the feed-forward neural networks utilize graph-based automated differen-
tiation (AD) (Baydin et al., 2018) to approximate the derivative terms in the PDEs. Various
PINNs architectures notably self-adaptive PINNs (McClenny and Braga-Neto, 2020), extended
PINNs (XPINN) (Hu et al., 2021; De Ryck et al., 2022) have been proposed that demonstrated

of
superior performance. Moreover, multiple DNN-based solvers such as cPINN (Jagtap et al.,
2020), XPINNs (Jagtap and Karniadakis, 2021), and PINNs framework for solid mechanics
(Haghighat et al., 2021b) have been developed that provide important advancement in terms

pro
of both robustness and faster computation. In this regard, (Haghighat et al., 2020, 2021b)
have been the breakthrough works geared towards developing a DL-based solver for inversion
and surrogate modeling in solid mechanics for the first time utilizing PINNs theory. Addi-
tionally, PINNs have been successfully applied to the solution and discovery in linear elastic
re-
solid mechanics (Zhang et al., 2020; Samaniego et al., 2020; Haghighat et al., 2021a; Guo and
Haghighat, 2020; Vahab et al., 2021; Rezaei et al., 2022; Zhang et al., 2022), elastic-viscoplastic
solids (Frankel et al., 2020; Goswami et al., 2022; Arora et al., 2022), brittle fracture (Goswami
et al., 2020) and computational elastodynamics (Rao et al., 2021) etc. The solution of PDEs
lP
corresponding to elasticity problems can be obtained by minimizing the network’s loss function
that comprises the residual error of governing PDEs and the initial/boundary conditions. In
this regard, PINNs can be utilized as a computational framework for the data-driven solution
of PDE-based linear elasticity problems that can significantly boost computational speed with
rna

limited network parameters. The potential of the PINNs framework in achieving computa-
tional efficiency beyond the capacity of the conventional computational methods for solving
complex problems in linear continuum elasticity is the main motivation behind the present
work.
Jou

In the present work, an efficient data-driven deep learning computational framework has been
presented based on the fundamentals of PINNs for the solution of the linear elasticity problem
in continuum solid mechanics. In order to efficiently incorporate physical information for the
elasticity problem, an improved multi-objective loss considering additional physics-constrained
terms has been carefully formulated that consists of the residual of governing PDE, various
boundary conditions, and data-driven physical knowledge fitting terms that demonstrate the
 Journal Pre-proof
6

efficacy of the model by accurately capturing the elasticity solution. Several benchmark prob-
lems including the Airy solution to an elastic plane-stress problem for an end-loaded cantilever
beam and simply supported rectangular Kirchhoff-Love thin plate under transverse sinusoidal
loading conditions have been solved which illustrates the superiority of the proposed model in

of
terms of accuracy and robustness by revealing excellent agreement with analytical solutions.
The employed models consist of independent multi-layer ANNs that are separately trained
on minimizing the prescribed loss function specific to the problem under consideration. The

pro
performance of PINNs has been evaluated for different activation functions and network archi-
tectures. Furthermore, we have illustrated the applicability of data-driven enhancement using
the smart initialization of a data-driven learning-based approach in reducing training time,
while simultaneously improving the accuracy of the model which is not possible in conventional
re-
numerical algorithms. Such an approach would be important in achieving computational effi-
ciency beyond the capacity of conventional computational methods for solving complex linear
elasticity problems. The present study also demonstrates the contribution of analytical solu-
tions for the data-driven construction of an accurate and robust PINNs framework that can
lP
significantly boost computational speed utilizing minimal trainable network parameters.
The paper is organized as follows: Section 2 introduces the background of PINNs the-
ory and the generalized idea of implementing multi-objective loss function into the PINNs
framework; In section 3, a brief overview of the theory of linear elasticity has been presented;
rna

Section 4 introduces the extension of the proposed PINNs framework for the Airy solution to
an elastic plane-stress problem for an end-loaded cantilever beam; in section 5, the proposed
PINNs framework has been extended to the solution of Kirchhoff–Love thin plate governed by
Biharmonic PDE; Section 7 deals with the relevant finding and prospects of the current work.
Finally, the conclusions have been discussed in section 8.
Jou

2. Physics-Informed Neural Networks :

The concept of training a NN in the PINNs framework is the construction of the loss function.
The loss function is intended to embed the underlying physics which is represented in mathe-
matical terms by the PDEs and the associated boundary conditions. In this section, we discuss
 Journal Pre-proof
7

the construction of the proposed multi-object loss functions for embedding a data-driven phys-
ical model that has been associated with the PINNs framework.
Let us consider a fully connected NN defined by

N k+1
Wk · N
(N k ) = κ k (W k
+ bk) (1)

of
where k ∈ {0, 1, · · · , N } represents the layer number of NN. N is a nonlinear map defined
by N m
(x̂ W m · x m + b m ) for mth -layer where W m and b m represents the weights
xm ) := κ m (W

pro
and biases of this transformation, respectively; κ(·) is the non-linear transformer or activation
function acting on a vector element-wise. Therefore, k = 0 represents the input layer of the
NN taking in the input x 0 .
Also consider a steady state general nonlinear partial differential operator G operated on
a scalar solution variable φ(~x) such that,
re-
G φ(~x) = 0 ~x ∈ Rndim (2)

Since G is a differential operator, in general, Eq. 2 is accompanied by appropriate boundary

conditions to ensure the existence and uniqueness of a solution. Let us assume, it is subjected to
lP
the boundary condition B φ(∂ ~Γ) = τ (∂~x) on the boundary ~Γ in domain Ω ∈ Rndim , ndim being
x), subjected to the
the spatial dimension. In a PINNs framework, the solution to Eq. 2, φ(x
aforementioned boundary condition may be approximated for an input x = ~x by constructing
rna

a feed-forward NN expressed mathematically as

φ̂ = N N
}N N −1
} · · · } N 0 (x
x) (3)

where φ̂ is the approximate solution to Eq. 2; } denotes the general compositional construction
0
of the NN; the input to the NN N := x 0 = ~x = (x1 , x2 , · · · xndim ) is the spatial coordinate
at which the solution is sought. Following Eq. 1 and Eq. 3, if W i and b i are all collected in
Jou

S
θ= N W i , b i ), the output layer N N contains the approximate solution φ̂(~x) to the PDE
i=0 (W

such that
k+1
N x, θ] = [φ̂1 , φ̂2 , ..., φ̂m ]
= φ̂ [x (4)

The spatial dependence of φ̂ is implicitly contained in the NN parameter θ. In the inter-

nal/ hidden layers of NN, several variations of nonlinear transformer or the activation func-
tion κ may be used, such as, the hyperbolic-tangent function tanh(ξ), the sigmoid function
 Journal Pre-proof
8

κ(ξ) = 1/(1 + e−ξ ), the rectified linear unit (ReLU) κ(ξ) = max(0, ξ), etc. The activation in
the final layer is generally taken to be linear for regression-type problems considered here.

2.1 Embedding constraints in NN :

of
This section briefly describes the general idea of embedding linear constraints into NN (Lagaris
et al., 1998; Du and Zaki, 2021). Let us consider U and A, two complete normed vector spaces,

pro
where NN function class M ⊂ U need to be constrained. A linear constraint on φ ∈ M can be
expressed as:
x) = 0,
Pφ(x φ∈M (5)

where, P : U → A expresses a linear operator on U. Generally, a such constraint can be

re-
realized for solving PDEs in most of the DL framework by minimizing the following functional

JA = kPφkA , φ∈M (6)

where k  kA denotes the norm corresponding to space A. It is noteworthy to mention that

lP
the aforementioned procedure approximately enforces linear constraint in Eq. 5. However, the
accuracy of the imposed constraint relies on the relative weighting between the constraint and
other objectives involved in the training include the satisfaction of the governing PDEs or the
rna

integration of data-driven schemes.

2.2 Multiple objective loss functions :

In order to incorporate physical information of the problem, one of the possibilities is to

Jou

impose Eq. 2 as a hard constraint in x ∈ Ω while training the NN on the physical data. Math-
ematically, such a condition is imposed by formulating a constrained optimization problem
which can be expressed as (Krishnapriyan et al., 2021),

x, θ) s.t. G φ(~x) = 0.
min ∆L (x (7)
θ

where ∆L represents data-driven physical knowledge fitting term which includes the imposed
initial and boundary conditions. G φ(~x) denotes the constraint corresponding to the residual
 Journal Pre-proof
9

PDE imposing the governing PDE itself. Thus, it is important to carefully impose appropriate
constraints for the NN to realize the underlying physics of the problem.
In the present work, we propose a multi-objective loss function that consists of residuals
of governing PDEs, various boundary conditions, and data-driven physical knowledge fitting

of
terms that can be expressed in the following general form:

x) − 0̂kΩ + βu kB Γu φ − g Γu kΓu + βt kB Γt φ − g Γt kΓt + αkφ − φ̂kΩ + · · · (8)

x, θ) = ϕkG φ(x
∆L (x

pro
where, ∆L (x x, θ) is the total loss function; the symbol k } k represents the mean squared error
J J
norm, i.e., k k = M SE( ) for regression type problem; kG φ(x x)− 0̂kΩ denotes the residual of
the governing differential relation in Eq. 2 for x ∈ Ω; Γu and Γt are the Dirichlet and Neumann
boundaries subjected to conditions B Γu φ = g Γu and B Γt φ = g Γt , respectively. The values
of g Γu and g Γt are specific to the problem under consideration, and therefore, pre-specified as
re-
inputs to the problem/ loss function. Note ϕ, βu , and, βt , are weights associated with each loss
term regularizing the emphasis on each term (the higher the relative value, the more emphasis
on satisfying the relation). The remaining task is to utilize standard optimization techniques
lP
x, θ)
to tune the parameters of the NN minimizing the proposed objective/ loss function ∆L (x
in Eq. 8.
However, even with a large volume of training data, such an approach may not guarantee
that the NN strictly obeys the conservation/governing equations in Eq. 2. Thus, additional
rna

loss terms to fit the observation data can be introduced. Hence, in the proposed objective loss
function, additional loss terms such as kφ − φ̄kΩ have been included that represent the data-
driven physical knowledge fitting term for the state variable φ(~x). Here, φ̄ is the true (target)
value of φ provided from either the analytical solution (if available), numerical simulation, or
experimental observations. α is the weight associated with the data-driven physical knowledge
Jou

fitting term for φ(~x). In the NN approximation, various degrees of differentials of the state
0 00
x) (i.e., φ (x
variable φ(x x), φ (x
x), · · · ) can also be included (if known) for stronger coupling in the
x) may be evaluated utilizing the graph-
data-driven approach. The partial differentials of φ(x
based automatic differentiation (Baydin et al., 2018) with multiple hidden layers representing
the nonlinear response in PINNs. Following the same steps, the initial conditions can also be
incorporated in Eq. 8. The loss from the initial conditions is not included herein due to the
quasi-static nature of the elasticity problem. In a more general case, the additional loss term
 Journal Pre-proof
10

kφ0 − φˆ0 kt=t

Ω
0
should be added for the loss contribution from the initial condition.
Finally, the optimal network parameters of NN θ̃ can be obtained by optimizing the loss
function in Eq. 8 as
θ̃ = arg mint ∆L (X̄
X̄, θ). (9)
θ⊂RN

of
SN i i
where, θ̃ := i=0 (W̃ , b̃ ) is the set of optimized network parameters; N t is the total number of
t
trainable parameters; and X̄ ∈ RNc ×N is the set of Nc collocation points used for optimization.

pro
3 Theory of linear elastic solid:

Consider an undeformed configuration B of an elastic body bounded in the domain Ω ⊂ Rndim

(1 ≤ ndim ≤ 3) with boundary Γ = Γu ∪ Γt where Γu 6= ∅ is the Dirichlet boundary, Γt is the
re-
Neumann boundary, and Γu ∩ Γt = ∅. With respect to the undeformed surface, the elastic
u on ΓD , and a prescribed surface traction
body can be subjected to a prescribed displacement ū
t̄t ∈ [L2 (Γt )]ndim . Additionally, a body force of density B ∈ [L2 (Ω)]ndim in Ω can be prescribed
with respect to the undeformed volume. Using a standard basis {eei } in Rndim , we can express
lP
u = 21 (ui,j + uj,i ) ei ⊗ ej ; where, ⊗ denotes the
the displacement, u = uiei , and its gradient, ∇u
tensor products. Second order symmetric tensors are linear transformations in S, defined as
  
S := ξ : Rndim → Rndim | ξ = ξ T with inner product ξ : ξ = tr ξ ξ T ≡ ξij ξij . Therefore,
the stress tensor can be expressed as σ := σij ei ⊗ ej . For infinitesimal strain, displacement
rna

 
gradient tensor ∇u u can be expressed as: ∇u u = ε + ω where ε := 12 ∇u u + ∇(uu)T is the
 
infinitesimal strain tensor with ∇ × ε = eijk εrj,i ek ⊗ er , and ω := 12 ∇u u)T is the in-
u − ∇(u
finitesimal rotation tensor.
Jou

3.1 Compatibility condition:

In the context of infinitesimal strain theory, we seek to find u : Ω → Rndim and corresponding
ε : Ω → Rndim ×ndim , and σ : Ω → Rndim ×ndim for a given infinite elastic solid satisfying the
following compatibility conditions (Marsden and Hughes, 1994):

R : = ∇ × (∇ × ε )T = 0 ; (10)
 Journal Pre-proof
11

where, R is Saint-Venant compatibility tensor. Alternatively, the elastic solid should satisfy
the Navier–Cauchy equations which can be expressed as (Lurie, 2010):

(λ + µ)∇(∇ · u) + µ∆u
u + B = 0, in Ω
(11)
u |ΓD = ū
ū;

of
where u = (u1 , u2 , ..., undim ) is the unknown displacement field; µ > 0 and λ > −µ are Lame
constants; ∇, ∆, and ∇ represent the gradient, the Laplacian, and the divergence operators,

pro
respectively. Equation 11 satisfies the continuity of the displacement field u and Dirichlet
boundary condition.

3.2 Equilibrium condition:

re-
In addition, the equilibrium condition and the Neumann boundary condition should be satisfied
which can be expressed as (Marsden and Hughes, 1994):

∇· σ + B = 0 , in Ω
(12)
t := Tu
u = t̄t, on Γt σ |Γt n̂ = t̄
lP
where, t̄t is a prescribed function on Γt ; n̂ is the field normal to Γt . Equation 12 satisfies the
momentum equation and the Neumann boundary condition where T follows the conformal
derivative operator such that (Atkin and Fox, 2005)
rna

u
∂u
Tu
u = λ(∆ u ) · n̂ + 2µ + µ n̂ × (∇ × u ) (13)
n̂
∂n̂

3.3 Constitutive relation:

Jou

Subsequently, the elastic constitutive relation can be expressed from generalized Hooke’s law
(Timoshenko, 1970) as:
σ =C :ε (14)

where, the fourth-order stiffness tensor C = Cijkl ei ⊗ ej ⊗ ek ⊗ el denotes the constitutive

u to the Cauchy stress tensor σ . For an
relation that maps the displacement gradient ∇u
isotropic linearly elastic material, Cijkl = λδij δkl + µ(δik δjl + δil δjk ) where δij is the Kronecker
 Journal Pre-proof
12

delta. The components of the stress tensor σ , and the strain tensor ε , are expressed as :
n
Xdim  
1 ∂ui ∂uj
u) = λδij
σij (u u) + 2µεij (u
εkk (u u), u) =
εij (u + , i, j = 1, 2, ..., ndim .
k=1
2 ∂xj ∂xi
(15)

of
Note that σ is the Cauchy stress tensor in linear elasticity applicable under small deformation.
The constitutive relation in terms of strain can be alternatively expressed as,

εij,kl + εkl,ij − εik,jl − εjl,ik = 0 i, j, k, l ∈ 1, 2, ..., ndim . (16)

pro
Equations governing a linear elastic boundary value problem (BVP) are defined by Eqs. 11–16
where the field variables u , σ, ε can be obtained for given material constants (Atkin and Fox,
2005; Lurie, 2010).

re-
4. PINNs formulation for continuum linear elasticity:

The proposed PINNs framework is applied to linearly elastic solids. A two-dimensional (ndim =
2) problem is considered. The input features (variables) to the models are the spatial coordi-
lP
nates x = (x, y). A separate NN is used to approximate each output field variable. As shown in
x), stress σ (x
Fig. 1, displacement u (x x), and strain ε (x
x) fields are obtained by densely connected
x),
independent ANNs. For ndim = 2, considering symmetry of the stress and strain tensors, u (x
x), and ε (x
σ (x x) fields can be approximated as:
rna

 
NN ũNN x)
x (x
x) ' Ξu (x
u (x x) = NN (17)
x)
ũy (x
 NN NN x
  NN 
NN x) σ̃xy
σ̃xx (x (x ) NN x) ε̃NN
ε̃xx (x x)
xy (x
x) ' Ξσ (x
σ (x x) = NN NN x ; x) ' Ξε (x
ε (x x) = NN (18)
x) σ̃xy
σ̃yx (x (x ) x) ε̃NN
ε̃yx (x x)
xy (x

NN x NN x
Here Ξu (x ), Ξσ (x ), and ΞεNN (x
x) denote the NN approximations for u (x
x), σ (x
x), and ε (x
x),
Jou

respectively.

4.1 Loss function:

To define the loss function for the linear elasticity problem, governing equations including
compatibility conditions, equilibrium conditions, constitutive relations, and boundary condi-
tions that fully describe the problem have been considered. Additionally, as in a data-driven
 Journal Pre-proof
13

of
pro
re-
Figure 1: PINNs network architecture for solving linear elasticity problem consisting of multi-
ANN (NNi ∀ i = 1, k) for each output variables ũNN x), ũNN
x (x
NN x
x), σ̃xx
y (x
NN x
(x ), σ̃yy NN x
(x ), σ̃xy (x ), ε̃NN x),
xx (x
NN x NN x
ε̃yy (x ), and ε̃xy (x ), with independent variable x = (x, y) as input features.

approach, the field variables in Eq. 8 have been included. The generalized mutli-objective loss
lP
functional ∆L can be expressed as:

x, θ) = ϕ ∆Ω
∆L (x e c Γu Γt u σ ε
L + ϕe ∆L + ϕc ∆L + βu ∆L + βt ∆L + αu ∆L + ασ ∆L + αε ∆L (19)
rna

where, ∆eL , ∆cL , and ∆Ω

L are the loss components from the equilibrium condition (Eq. 12),

constitutive relation (Eq. 14), and the compatibility condition (Eq. 15), respectively; ∆ΓLu and
∆ΓLt represent the loss components computed at the Dirichlet boundary Γu , and the Neumann
boundary Γ (Eq. 11), respectively; ∆u , ∆σ , and ∆ε are the loss components for the fields
t L L L

x), σ (x
u (x x), and ε (x
x), respectively, when a data driven approach is pursued. The coefficients
Jou

ϕ, ϕe , ϕc , βu , βt , αu , ασ , and αε are the weights associated with each loss term that dictates the
emphasis on each penalty term. Evidently, the terms in the cost function are the measures of
the errors in the displacement and stress fields, the momentum balance, and the constitutive
 Journal Pre-proof
14

x, θ) is,
law. The explicit expression for each term in ∆L (x

Nc Ω
1 X
∆Ω
L = Ω
k∇ · ΞNN xl|Ω ) + B (x
σ (x xl|Ω )k (20)
Nc l=1
Nc Ω
1 X  

of
NN NN
∆cL = Ω
kΞσ xl|Ω ) − C ∇ · Ξu
(x xl|Ω ) k
(x (21)
Nc l=1
Γu
Nc
1 X
∆ΓLu = kΞNN (x
x ) − ū xk|Γu )|k
ū(x (22)
NcΓu k=1 u k|Γu

pro
Γt
Nc
1 X
∆ΓLt = kΞNN (x
xj|Γt )n̂
n̂ − t̄t̄(x
xj|Γt )k (23)
NcΓt j=1 σ
Nc Ω
1 X
∆u
L = kΞNN (x
x ) − û
u(xxl|Ω )k (24)
NcΩ l=1 u l|Ω

σ
∆L =
1 X
Nc re-
NcΩ l=1 σ l|Ω
Ω

kΞNN (x
x ) − σ̂
σ (x
xl|Ω )k (25)

Nc Ω
1 X
∆εL = kΞNN (x
xl|Ω ) − ε̂ε(x
xl|Ω )k (26)
NcΩ l=1 ε
lP
 n o
where, x 1|Ω , ..., x NcΩ |Ω are randomly chosen collocation points over the domain Ω; x 1|Γu , ..., x NcΓu |Γu
n o
and x 1|Γt , ..., x NcΓt |Γt are those chosen randomly along the boundaries Γu and Γt , respectively.
u(x
The terms û xl|Ω ), σ̂
σ (x
xl|Ω ), and ε̂ε(x
xl|Ω ) represent the true (target) value obtained by means of
rna

analytical solution or high-fidelity simulation. The weights ϕ, ϕe , ϕc ∈ R+ are the weights

corresponding to the compatibility, equilibrium, and constitutive relations, respectively. In
general, these coefficients can be prescribed as 1 for solving a relatively less complex problem,
whereas, βu and βt are the binary (i.e., either 0 or 1) integers. The weights αi = 1; ∀ i = u , σ , ε
x), σ (x
for a complete data driven approach for u (x x), and ε (x
x), respectively at the collocation
Jou

points NcΩ . However, we prescribe αi = 0 ∀ (i = u , σ , ε ) as labeled training data is unavailable,

which may not guarantee the accuracy of PINNs solutions.
The forward problem is studied herein, where the displacement, stress, and strain fields
are obtained as the PINNs solutions assuming material properties λ and µ remain constant.
However, the loss functional in Eq. 19 can also be utilized in an inverse problem for parameter
identification, where λ and µ can be treated as network outputs which may vary during train-
ing (Fig. 1). For the network construction in the PINNs framework, SciANN (Haghighat and
 Journal Pre-proof
15

of
pro
Figure 2: (a) Elastic plane-stress problem for an end-loaded cantilever beam of length L,
height 2a and out-of-plane thickness b which has been clamped at x = L; (b) distributions of
total collocations points Nc = 5, 000 on the problem domain and various boundaries during
PINNs training.

re-
Juanes, 2021), a convenient high-level Keras (Chollet et al., 2015) wrapper for PINNs is used.

4.2 Solution for linear elasticity problem :

lP
For this study, an end-loaded isotropic linearly elastic cantilever beam of height 2a, length
L, thickness b (assuming b  a) has been considered to ensure a state of plane-stress condition
as shown in Fig. 2. The left edge of the beam is subjected to a resultant force P . Whereas,
the right-hand end is clamped. The top and bottom surfaces of the beam, y = ±a are traction
rna

free. An approximate solution to the problem can be obtained from the Airy function discussed
next.

4.2.1 The Airy solution to the end-loaded cantilever beam:

Jou

The Airy solution in Cartesian coordinates Ω ⊂ R2 can be found from the Airy potential
φ(x, y) that satisfies (Bower, 2009),

∂ 4φ ∂ 4φ ∂ 4φ ∂bx ∂by
∇φ = 4
+ 2 2 2
+ 4
= C(ν)( + ) (27)
∂x ∂x ∂y ∂y ∂x ∂y

where,  1−ν
1−2ν
(plane strain)
C(ν) = 1 (28)
1−ν
(plane stress)
 Journal Pre-proof
16

∂Ω ∂Ω
Here, the body forces bx , by have the form ρ0 bx = ∂x
, ρ 0 by = ∂y
; Ω(x, y) is the positional scalar
function. The solution of the Airy function can be expressed in the polynomial form φ(x, y) =
P∞ P∞ m n
m=0 n=0 Amn x y . For m+n ≤ 3, the terms automatically satisfy the biharmonic equation

for any Amn . Additionally, φ must satisfy the following traction boundary conditions on Ω.

of
∂ 2φ ∂ 2φ ∂ 2φ ∂ 2φ
n x − ny = tx ; ny − ny = ty (29)
∂y 2 ∂x∂y ∂x2 ∂x∂y

Here, (nx , ny ) are the components of a unit vector normal to the boundary. For the end-loaded

pro
cantilever beam, the Airy function can be formulated as,
3P P
φ=− xy + 3 xy 3 (30)
4ab 4a b
∂2φ ∂2φ ∂ φ 2
where, σxx = ∂y 2
− Ω; σyy = ∂x2
− Ω; σxy = σyx = − ∂x∂y with Ω = 0. At the clamped end,
x1 = L, displacement boundary conditions are ux = uy = ∂uy /∂x = 0. The top and bottom
re-
surfaces of the beam (i.e., y = ±a) are traction free, σij ni = 0, that requires σyy = σxy = 0.
Whereas, the resultant of the traction acting on the surface at x = 0 is −P ey with traction
2
vector ti = σij nj = −σxy δiy = − 4ab3P
(1 − ay2 )δiy . The resultant force can be obtained as :
Ra 2
3P
Fi = b −a − 4ab (1 − ay2 )δiy dx2 = −P δiy . On satisfaction of the aforementioned conditions,
lP
approximate analytical solutions for the displacements ux , uy , the strain fields εxx , εyy , εxy and
the stress fields σxx , σyy , σxy can be expressed as:
3P 2 P 3 P a2 3P L2
ux = x y − (2 + µ) y + 3(1 + µ) y − y (31)
4Ea3 b 4Ea3 b 2Ea3 b 4Ea3 b
rna

3µP 2 P 3 3P L2 P L3
uy = − xy − x + x − (32)
4Ea3 b 4Ea3 b 4Ea3 b 2Ea3 b  
3P 3P µ 3P (1 + µ) y2
εxx = xy; εyy = − xy; εxy = 1− 2 (33)
2Ea3 b 2Ea3 b 4Eab a
 2

3P 3P y
σxx = 3
xy; σyy = 0; σxy = 1− 2 (34)
2a b 4ab a
Jou

x), σ (x
These analytical solutions for u (x x), and ε (x
x) have been used as û
u(xxl|Ω ), σ̂
σ (x
xl|Ω ), and
ε̂ε(x
xl|Ω ) at the collocation points for data-driven enhancement in Eqs. 24-26, respectively, for
solving the field variables in the proposed PINNs framework.

4.2.2 PINNs solutions for linear elasticity problem:

For the benchmark, end-loaded cantilever beam problem, L = 3 m, a = 0.5 m, and b = 0.001
 Journal Pre-proof
17

m have been considered. The material properties are, Young’s modulus E = 1 GPa, and the
Poisson ratio ν = 0.25 as shown in Fig. 2 -(a). Unless otherwise stated, a total of Nc = 5, 000
randomly distributed collocation points over the domain and boundaries have been used for
training the PINNs model as shown in Fig. 2 -(a). During training, the optimization loop was

of
run for 500 epochs using the Adam optimization scheme with a learning rate of 0.001, and a
batch size of 32 for optimal accuracy and faster convergence.
The Airy solutions for various fields including displacements ux , uy , stresses σxx , σyy , σxy ,

pro
and strains εxx , εyy , εxy as in Eqs. 31-34 are shown in Fig. 3-(a). The corresponding PINNs
approximations using the tanh activation function are shown in Fig. 3 -(b). Additionally, in
Fig. 3 -(c), the absolute error between the Airy solutions and PINNs predictions for each field
variable is shown. The overall results from PINNs are in excellent agreement with the Airy
re-
solutions. The PINNs approximations attained satisfactory accuracy with low absolute errors
for all field variables. For the displacement fields, the absolute error is relatively high near to
clamped edge for ux . For uy , the absolute error is maximum at the midsection and near the
horizontal edges as shown in Fig. 3 -(c). This is due to the approximate nature of the Airy solu-
lP
tions at clamped end x1 = L for the displacement boundary conditions ux = uy = ∂uy /∂x = 0.
Such differences also propagate through the solutions of stress and strain fields, where PINNs
predictions slightly deviate from the Airy solutions, in particular, near the free vertical and
horizontal edges as shown in Fig. 3 -(c). However, according to Saint-Venant’s principle, these
rna

deviations do not sufficiently influence the solution far from the end, which is reflected in the
result. Overall, the proposed PINNs model can capture the distributions of various fields ac-
curately from the solution of the Airy stress function.

4.2.3 Suitable activation function :

Jou

The impact of the use of various activation functions on training the PINNs models in pre-
dicting field variables and the epoch evolution of various components of the loss function is
explored. The ReLU, sigmoid, and tanh activation functions are compared; the network archi-
tecture remains the same: the number of neurons in each layer N = 20 with the total number
of hidden layers Ln = 5 in the PINNs model. The evolution of the total loss ∆L , and the
 Journal Pre-proof
18

of
pro
re-
lP
rna
Jou

Figure 3: (a) The Airy solutions for displacements ux , uy , stresses σxx , σyy , σxy , strains εxx ,
εyy , εxy ; (b) corresponding PINNs solutions for ũNN NN NN NN NN NN NN NN
x , ũy , σ̃xx , σ̃yy , σ̃xy , ε̃xx , ε̃yy , and ε̃xy ;
(c) absolute error between the Airy solutions and PINNs predictions associated with each field
variables for an end-loaded cantilever beam.
 Journal Pre-proof
19

of
pro
Figure 4: Comparison of (a) total loss ∆Ω Ω
L ; (b) constitutive loss ∆L for tanh, sigmoid and
ReLU activation functions for network parameters N = 20, Ln = 5.

constitutive loss ∆Ω
L are depicted in Fig. 4. Additionally, values of the various loss components
re-
and training times ttr at the end of training are compared in Table. 1. Evidently, the tanh
activation provides the best performance in terms of the value of the total loss at the end of
training. The final constitutive loss with tanh activation is significantly lower compared to
the other two activations illustrating the suitability of the use of the tanh activation for the
lP
PINNs model for solving the elasticity problem herein. In addition, all other loss components
obtained are lowest upon using the tanh activation as shown in Table 1.
Comparing the evolution of ∆L , the convergence characteristics for the ReLU activation are
rna

better compared to the tanh with fewer fluctuations and rapid decrease in loss values as shown
in Fig. 4-(a). However, the tanh illustrates better adaptability in the constitutive loss with
an excellent convergence rate in Fig. 4-(b). Out of the three activations, ReLU performs the
worst possibly due to its derivative being discontinuous. However, the total loss for all three
activations is negligible (loss value in the range below 10−4 to 10−5 ) within 200 epochs indi-
Jou

cating the adaptability of the proposed PINNs framework to any of these activations provided
the models are trained sufficiently long. In comparing the training time, the tanh activation
takes longer for the same number of epochs compared to the other two. This coincides with
the fact that the evolution of the total loss has a higher degree of discontinuity. However, the
model with the ReLU activation trains the fastest possibly due to its linear nature. From the
comparison, it can be concluded that although tanh is the best in terms of accuracy, however,
ReLU can be an optimal choice of activation considering both accuracy and training time for
 Journal Pre-proof
20

Table 1: Influence of different activation functions on the final values of various loss components
(in 10−09 ) and training times ttr in the proposed PINNs model for solving linear elastic beam
problem.

Activation Function ∆Ω
L ∆cL ∆ΓLu ∆ΓLt ∆u
L ∆σ
L ∆εL ∆L ttr

of
(min)
ReLU 107.16 43.43 14.51 36.75 24.97 1.07 5.48 233.37 9.4
Sigmoid 30.96 54.33 517.38 126.14 37.85 124.51 592.82 1483.99 13.8

pro
tanh 4.56 0.73 31.47 25.64 3.11 9.60 10.45 85.56 15.7

Table 2: Influence of network parameters N and Ln on training times ttr and final values
various loss components (in 10−09 ) for tanh activation.

Network np re-ttr ∆Ω
L ∆cL ∆ΓLu ∆ΓLt ∆u
L ∆σ
L ∆εL ∆L
identifier (min)
N-1 (N = 20, Ln = 5) 22,706 15.7 4.56 0.73 31.47 25.64 3.11 9.60 10.45 85.56
N-2 (N = 40, Ln = 5) 113,530 23.8 2.21 90.39 77.73 59.58 4.29 24.16 78.39 336.75
N-3 (N = 20, Ln = 10) 54,494 18.3 6.89 0.89 12.73 65.42 13.01 17.19 4.67 120.8
lP
N-4 (N = 40, Ln = 10) 272,472 32.3 2.78 3.67 18.78 12.63 24.19 43.10 2.49 107.64

solving elasticity equation in the proposed PINNs framework.

rna

4.2.4 Influence of network complexity:

It is worth mentioning that the PINNs approximations are sensitive to network architecture
including the depth of the hidden layer and the number of network parameters. In this section,
Jou

the influence of network architecture parameters, i.e., the number of neurons in each hidden
layer N , and the number of hidden layers Ln on the accuracy and the efficiency of the PINNs
solution are explored. Since the tanh activation performs the best in terms of accuracy (see
previous section), it is chosen as the activation for different networks used in the following
experiments.
In the current study, four different networks considering the combinations N = 20, 40, and
Ln = 5, 10 are tested, and values of different loss components at the end of the training, train-
 Journal Pre-proof
21

of
pro
Figure 5: Comparison of (a) total loss ∆Ω Ω
L ; (b) constitutive loss ∆L for various combinations
of network parameters N and Ln considering tanh activation function.

ing duration (ttr ), along with model complexities in terms of network parameters (np ) for these
re-
architectures are presented in Table. 2. For fair comparison, Nc = 5, 000 for all experiments.
The evolution of the total loss ∆L and the constitutive loss ∆Ω
L for these networks are shown in

Fig. 5. From the comparisons, for the chosen number of collocation points relatively shallow
network N = 20, Ln = 5 provides the best performance in terms of ∆L and ∆Ω
L at the end
lP
of training. Additionally, the time required for training is faster due to a significantly lower
number of network parameters. However, for a relatively deeper network, N = 20, Ln = 10
with increased network complexity, the performance of the model degrades with respect to loss
rna

values as shown in Table. 2 possibly due to an increase in variability and reduction in bias.
Interestingly, an increase in the number of neurons N = 40 while maintaining the depth of
the network (Ln = 5) leads to the worst performance which can be attributed to over-fitting
(Bilbao and Bilbao, 2017; Jabbar and Khan, 2015). The epoch evolution of the loss for various
network architectures demonstrates the efficacy of a relatively shallow network with signifi-
Jou

cantly faster training for solving elasticity problems in the proposed PINNs framework.

5. PINNs formulation for linear elastic plate theory :

In this section, the PINNs framework is expanded for the solution of the classical Kirchhoff-
Love thin plate (Timoshenko and Woinowsky-Krieger, 1959) subjected to a transverse loading
in linearly elastic plate theory. In the subsequent section, the Kirchhoff-Love theory has been
 Journal Pre-proof
22

briefly described; PINNs formulation for solving the governing fourth-order biharmonic partial
differential equation (PDE) for the solution of the thin plate is elaborated. For a benchmark
problem, the proposed PINNs approach is applied for the solution of a simply supported rect-
angular plate under a transverse sinusoidal loading condition.

of
5.1 Kirchhoff-Love thin plate theory :

pro
Thin plates are structurally planar elements that have small thickness relative to their in-
plane dimensions which can be simplified as a two-dimensional plate problem. According to
the Kirchhoff-Love theory, the kinetics of a thin plate under the effect of a distributed trans-
verse loading q = q(x, y) can be described by a fourth-order differential equation (Timoshenko
and Woinowsky-Krieger, 1959; Reddy, 2006).re- ∆(D∆w) = q (35)

When the elastic plate is bounded in the domain Ω ⊂ R2 , Eq. 35 is known as the Kirchhoff-
lP
Love equation. In Cartesian coordinates, w = w(x, y) represents the transverse displacement
field, D = D(x, y) is the bending stiffness of the plate, and ∆ = ∂ 2 /∂x2 + ∂ 2 /∂y 2 is the
Laplace operator. Considering a homogeneous and isotropic plate (i.e., D ≡ constant ), Eq.
35 becomes the biharmonic equation (Timoshenko and Woinowsky-Krieger, 1959; Szilard and
rna

Nash, 1974)  
2 ∂ 4w ∂ 4w ∂ 4w
D∆ w = D + 2 + =q (36)
∂x4 ∂x2 ∂y 2 ∂y 4

Under appropriate boundary conditions, and with D(x, y) > 0 and q(x, y) ≥ 0, both being
known, the problem possesses a unique solution for the displacement w(x, y). The set of
Jou

solution variables includes the primitive variable deflection w, and the derived quantities,
moments Mxx , Myy , Mxy = −Myx , and shearing forces Qxx , Qyy . The expressions for the
derived fields are,
 2     
∂ w ∂ 2w ∂ 2w ∂ 2w ∂ 2w
Mxx = −D +ν 2 ; Myy = −D + ν ; Mxy = −D(1 − ν) (37)
∂x2 ∂y ∂y 2 ∂x2 ∂x∂y

   
∂Myx ∂Mxx ∂ ∂ 2w ∂ 2w ∂Myy ∂Mxy ∂ ∂ 2w ∂ 2w
Qxx = + = −D + ; Qyy = − = −D + (38)
∂y ∂x ∂x ∂x2 ∂y 2 ∂y ∂x ∂y ∂x2 ∂y 2
 Journal Pre-proof
23

of
pro
re-
Figure 6: PINNs network architecture for solving Kirchhoff-Love thin plate problem governed
by biharmonic equation consisting of multi-ANN (NNi ∀ i = 1, k) for each field variables w̃NN (x x),
NN x NN x NN x NN x NN x
M̃xx (x ), M̃xy (x ), M̃yy (x ), Q̃xx (x ), and Q̃yy (x ) with independent variable x = (x, y) as input
features.
lP

5.2 PINNs formulation for the Biharmonic equation:

rna

For solving the Biharmonic equation using the PINNs framework, the input features are the
x), M (x
spatial coordinates x := (x, y); the field variables, w(x x), and Q (x
x) are obtained using
multiple densely connected independent ANNs, with each network approximating one of the
outputs (Fig. 7). Different field variables approximated by the NNs are as follows:
Jou

w(xx) ' ΞNNw = w̃

NN
x)
(x (39)
 NN   NN 
x) M̃NN
M̃xx (x x)
xy (x x)
Q̃xx (x
x) ' ΞNN
M (x = ; Q (x NN
x) ' Ξ = (40)
M NN x NN x
M̃yx (x ) M̃yy (x ) Q NN x)
Q̃yx (x

where, ΞNN NN NN
w , ΞM , and ΞQ are the neural network appoximations. From the NN approximations

x, θ) can be defined as:

of the fields, the muti-objective loss function ∆L (x
Γu Γt Q
x, θ) = ϕ ∆Ω
∆L (x w M
L + βu ∆L + βt ∆L + αw ∆L + αM ∆L + αQ ∆L (41)
 Journal Pre-proof
24

Γu Γt
where, ∆Ω
L , ∆L , ∆L are the losses in the domain Ω, and along the boundaries Γu and Γt ,

respectively. Their expressions are,

Nc Ω
1 X q̂
∆Ω
L = Ω
k∇2 ∇2 w − k (42)
Nc l=1 D

of
Γu
Nc
1 X
∆ΓLu = kΞNN (x
x xk|Γu )|k
) − w̄(x (43)
NcΓu k=1 w k|Γu
Γt
Nc
1 X

pro
∆ΓLt = kΞNN (x
x ) − M̄ (x
xj|Γt )|k (44)
NcΓt j=1 M j|Γt

 n o n o
where, x 1|Ω , ..., x NcΩ |Ω , x 1|Γu , ..., x NcΓu |Γu , x 1|Γt , ..., x NcΓt |Γt are the collocation points over
the domain Ω, and along the boundaries Γu and Γt , respectively; ϕ ∈ R+ is the penalty
coefficient for imposing the biharmonic relation in Eq. 36. Additionally, data driven estimates
x), M (x
of w(x x), and Q (x
re-
x) at the collocation points across Ω are used to define ∆L (x

1 X
Nc Ω
x, θ).

∆w
L = kΞNN (x
x ) − ŵ(x
xl|Ω )k (45)
NcΩ l=1 w l|Ω
Nc Ω
1 X
lP
∆M
L = kΞNN (x
x ) − M̂ (x
xl|Ω )k (46)
NcΩ l=1 M l|Ω
Nc Ω
1 X
∆Q = kΞNN (x
x ) − Q̂ xl|Ω )k
Q̂(x (47)
L
NcΩ l=1 Q l|Ω
rna

xl|Ω ), M̂ (x
Here, ŵ(x xl|Ω ), and Q̂ xl|Ω ) are obtained by means of analytical or high-fidelity nu-
Q̂(x
merical solutions. Note, αi = 1; ∀ i = w, M, Q for data-driven enhancement coupled with
physics-informed regression by forcing the PDE constraints in Eqs. 36-38. Whereas, αi = 0
switches off the data-driven enhancement of accuracy of the NN approximations. The loss
x), M (x
function in Eq. 41 can either be used for obtaining PINNs approximations of w(x x),
Jou

x) (i.e., forward problem ), or identification of model parameters λ and µ (i.e., inverse

and Q (x
problem ).

5.3 Simply supported Kirchhoff-Love plate:

A simply supported rectangular plate of size (a×b) under a sinusoidal load q(x, y) = q0 sin πx
a
sin πy
b

is considered in Cartesian coordinates as shown in Fig. 7. Here, q0 is the intensity of the load at
 Journal Pre-proof
25

of
pro
Figure 7: Benchmark problem setup for Kirchhoff-Love plate: (a, b) simply supported rect-
angular plate of a = 200 cm and b = 300 cm with thickness t = 1 cm subjected to transverse
sinusoidal loading of intensity q0 = 9.806 × 10−4 MPa; (b) distributions of total collocations
points Nc = 10, 000 on the problem domain and various boundaries during PINNs training.
re-
the center of the plate. The following boundary conditions are applied at the simply supported
(SS) edges:
∂ 2w
w = 0; =0 for x = 0 and x = a (48)
∂x2
lP
∂ 2w
w = 0; =0 for y = 0 and y = b (49)
∂y 2

5.3.1 Analytical solution: Along with the governing equation in Eq. 36 and the bound-
ary conditions in Eqs. 48- 49, the analytical solutions of w are obtained as:
rna

q0 πx πy
w= sin sin (50)
π 4 ( a12 + 1 2
b2
) a b

Utilizing Eqs. 37-38, analytical solutions for the moments Mxx , Myy , Mxy and the shearing
forces, Qxx , Qyy are obtained as:
 
q0 1 ν πx πy
Jou

Mxx =
+ sin sin (51)

1 1 2 a2 b 2 a b
π2 a2
+ b2
 
q0 ν 1 πx πy
Myy =
2 2
+ 2 sin sin (52)
π 2 a12 + b12 a b a b
 
q0 (1 − ν) ν 1 πx πy
Mxy =
2 2
+ 2 cos cos (53)
π 2 a12 + b12 ab a b a b
q0 πx πy
Qxx = 1 1

cos sin (54)
πa a2 + b2 a b
q0 πx πy
Qyy =
sin sin (55)
πa a12 + b12 a b
 Journal Pre-proof
26

x), M (x
These analytical solutions, w(x x), and Q (x
x) have been utilized as ŵ(x
xl|Ω ), M̂ (x
xl|Ω ), and
Q̂ xl|Ω ) for data driven enhancement in Eqs. 45-47, respectively for the PINNs approximations
Q̂(x
of the field variables.

of
5.4 PINNs solutions for the Biharmonic equation:

For the benchmark problem, a rectangular plate (a = 200 cm, b = 300 cm) with thickness

pro
t = 1 cm is considered with the following material properties: Young’s modulus of elasticity
E= 202017.03 MPa, Poisson’s ratio ν = 0.25, and flexural rigidity D= 17957 N-m. The sinu-
soidal load intensity q0 = 9.806 × 10−4 MPa is presribed as shown in Fig. 7. A similar problem
has been also solved in the recent work(Vahab et al., 2021). Unless otherwise stated, the total
re-
number of randomly distributed collocation points, Nc = 10, 000 is used during the training of
the PINNs model. Additionally, a learning rate of 0.001, and a batch size of 50 were prescribed
for optimal accuracy and faster convergence of the optimization scheme. For better accuracy
during training, the Adam optimization scheme is employed with 1000 epochs. In the present
lP
study, three different activation functions were tested (see section 5.4.1).
In Fig. 8(a–f), the analytical solution for various fields including plate deflection w, mo-
ments Mxx , Myy , Mxy , and shearing forces Qxx , and Qyy in Eqs. 50-55 are shown. Corre-
sponding approximations from PINNs for various activation functions are shown in Fig. 8
rna

(a–f) which illustrate the efficacy of the proposed model in terms of accuracy and robustness
as excellent agreement with the analytical solutions is evident.

5.4.1 Influence of the activation function:

Jou

The accuracy of the field variables and epoch evolution of the loss functions are explored
for various activation functions for solving the fourth-order biharmonic PDE. To this end,
three different activations, i.e., ReLU, sigmoid, and tanh are selected; the network used is
defined by N = 20, Ln = 5. The corresponding results are depicted in Fig. 8 (g–l). Based on
the results, all the activations perform well as the NN approximations are in good agreement
with the analytical solutions both qualitatively and quantitatively. For further insight into the
 Journal Pre-proof
27

of
pro
re-
lP
rna
Jou

Figure 8: Solution of field variables obtained from (a-f) analytical solutions (left to right): w,
Mxx , Myy , Mxy , Qxx , and Qyy ; (g-l) proposed PINNs results (left to right): w̃NN , M̃NN NN
xx , M̃xy ,
M̃NN NN NN
yy , Q̃xx , and Q̃yy for activation functions (i) ReLU, (ii) sigmoid, and (iii) tanh.
 Journal Pre-proof
28

of
pro
re-
lP
Figure 9: Absolute error of field variables between analytical solution and PINNs results (a)
|w − w̃NN |; (b) |Mxx − M̃NN NN NN NN
xx |; (c) |Myy − M̃yy |; (d) |Mxy − M̃xy |; (e) |Qxx − Q̃xx |; and (f)
|Qyy − Q̃NN
yy | for activation functions (i) ReLU, (ii) sigmoid, and (iii) tanh.
rna
Jou

Figure 10: Comparison of (a) total loss ∆L ; (b) constitutive loss ∆Ω

L during training for tanh,
sigmoid and ReLU activation functions for network parameters N = 20, Ln = 5.
 Journal Pre-proof
29

Table 3: Influence of different activation functions on the final values of various loss components
(in 10−05 ) and training times ttr in the proposed PINNs model for solving biharmonic PDE.

Activation Function ∆Ω
L ∆ΓLt ∆ΓLu ∆w
L ∆M
L ∆Q
L ∆L ttr
(min)

of
ReLU 5.34 132.31 1672.91 278.43 498.76 101.36 2689.11 23.1
Sigmoid 63.07 980.67 4601.60 1707.50 987.89 117.56 8458.29 25.8
tanh 0.12 7138.43 9807.31 6809.34 397.89 500.37 24653.46 34.6

pro
influence of an activation function on the accuracy of the solutions, the absolute error between
the analytical solutions and the PINNs approximations for each field variable is compared for
the solutions obtained with different activations in Fig. 9 (a–f). From the comparison, ReLU
re-
provides the least absolute error distributions in solving the Biharmonic equation for the simply
supported plate. Although, the sigmoid activation provides the best result for |Mxy − M̃NN
xy |, the

absolute error for the rest of the fields is higher compared to the solutions obtained with ReLU.
Because of the sinusoidal nature of the solution, it was expected that tanh activation might
lP
be specifically suitable for this problem. Surprisingly, tanh provides worse results compared to
ReLU and sigmoid activations. This can be due to the complex nature of the solution space,
where ReLU can provide better adaptability during training. Furthermore, in Fig. 10, the
epoch evolution of the total loss ∆Ω Ω
L , and constitutive loss ∆L is compared for different activa-
rna

tion functions. For a particular epoch, ReLU performs better than the other two activations
for ∆L . For ∆Ω
L , tanh activation shows better convergence and the lowest loss value at the end

of training due to the sinusoidal nature of the solution of the Biharmonic PDE. However, the
fluctuations in the loss curve for tanh have a relatively higher variance compared to ReLU and
sigmoid. As reported in Table 3, overall, performance in terms of various loss components at
Jou

the end of training is superior for the ReLU activation for solving the Biharmonic PDE using
the proposed PINNs framework. Additionally, the model with the ReLU activation requires
the least training time ttr , indicating better convergence and faster computation of the forward
and backpropagation steps.

5.4.2 Influence of network parameters:

 Journal Pre-proof
30

As was found for the linear elasticity problem, PINNs solutions are sensitive to the NN ar-
chitecture. Various parameters that influence the NN architectures, the number of neurons in
each hidden layer N , and the total number of hidden layers Ln , on the accuracy of the model

of
and the efficiency of training the model have been explored herein. Because of its superior
performance for the problem, ReLU is chosen as the activation function. Four different net-
works with combinations N = 20, 40, and Ln = 5, 10 were trained. Corresponding network

pro
parameters (np ), model training time (ttr ), and values of different loss components at the end
of training have been presented in Table. 4. The comparisons of the absolute error between
the analytical solutions and the PINNs approximations for each field are shown in Fig. 11.
Comparisons of the total loss ∆L , the constitutive loss ∆Ω
L for various combinations of network

re-
parameters, N and Ln are shown in Fig. 12.
Based on the comparisons shown in Fig. 11, increased network depth improves the accuracy
of the PINNs approximations for all variables. Predictions by both networks with Ln = 10
are superior compared to the analytical solutions for the chosen number of collocation points.
lP
On the other hand, an increase in the number of neurons in each layer increases model predic-
tion variance which is reflected in the higher absolute error comparisons for N = 20, 40 and
Ln = 10. Similar conclusions may be drawn based on Fig. 12 and Table. 4. The total and
constitutive losses are minimum for N = 40 and Ln = 10 at the end of training. However, the
rna

approximations by this model have higher variance. Expectedly, more complex models (higher
Ln ), or with larger np , require longer training time ttr . For the chosen number of collocation
points, Ln = 10 is optimal.

5.4.3 Smart initialization of data-driven enhancement:

Jou

In this section, we explore the applicability of data-driven enhancement in the proposed PINNs
framework to improve the accuracy of the solution. Initially, the network is trained with rel-
atively low Nc = 10, 000. The pre-trained model is then trained for the higher number of
collocation datasets Nc = 15, 000 and Nc = 20, 000 to further improve the model accuracy.
 Journal Pre-proof
31

of
pro
re-
lP

Figure 11: Absolute error of field variables between analytical solution and PINNs results
(a) |w − w̃NN |; (b) |Mxx − M̃NN NN NN NN
xx |; (c) |Myy − M̃yy |; (d) |Mxy − M̃xy |; (e) |Qxx − Q̃xx |; and (f)
rna

|Qyy − Q̃NN
yy | for various network parameters (i) N = 20, Ln = 5, (ii) N = 40, Ln = 5, (iii)
N = 20, Ln = 10, and (iv) N = 40, Ln = 10.
Jou

Figure 12: Comparison of (a) total loss ∆Ω Ω

L ; (b) constitutive loss ∆L for various combinations
of network parameters N and Ln considering ReLU activation.
 Journal Pre-proof
32

Table 4: Influence of network parameters N and Ln on training times ttr and final values of
various loss components (in 10−05 ) for tanh activation.

of
Network np ttr ∆Ω
L ∆ΓLu ∆ΓLt ∆w
L ∆M
L ∆Q
L ∆L
identifier (min)
N-1 (N = 20, Ln = 5) 12,940 23.1 5.34 132.31 1672.91 278.43 498.76 101.36 2689.11

pro
N-2 (N = 40, Ln = 5) 52,760 29.8 0.47 35.13 467.34 128.38 198.11 40.29 869.72
N-3 (N = 20, Ln = 10) 32,056 31.7 0.07 82.15 86.84 77.82 298.01 10.17 555.06
N-4 (N = 40, Ln = 10) 126,224 42.8 0.009 0.67 5.12 4.21 0.53 0.17 10.709

re-
lP
rna
Jou

Figure 13: Influence of smart initialization of data-driven enhancement on (a) total loss
∆Ω Ω
L ; (b) constitutive loss ∆L for increasing Nc considering ReLU activation; (c) Absolute
error of field variables between analytical solution and PINNs results for (i) Nc = 10, 000, (ii)
Nc = 15, 000 TL, and Nc = 20, 000 TL.
 Journal Pre-proof
33

Table 5: Network parameters, training time, and the component of loss for different smart
initialization of data-driven enhancement models.

Network Nc Epochs ∆Ω
L ∆ΓLt ∆ΓLu ∆w
L ∆M
L ∆Q
L ∆L ttr
identifier (min)

of
N-1 10000 1000 5.34 132.31 1672.91 278.43 498.76 101.36 2689.11 23.1
N-TL1 15000 250 0.025 1.31 17.34 1.43 13.11 9.89 43.11 5.1
N-TL2 20000 250 0.005 0.71 2.96 2.01 2.56 0.87 9.11 7.2

pro
The idea is to speed up the training by utilizing pre-trained weights; the initial states of
the PINNs models in the later phases of training are not random anymore. The speed-up is
reflected in Figs. 13-(a, b) when the convergence of the loss curves (∆L and ∆Ω
L ) for the pre-

re-
trained models corresponding to Nc = 15, 000 and Nc = 20, 000 are much improved compared
to the first training phase with Nc = 10, 000. In Fig. 13-(c), the absolute errors between the
approximations and analytical solutions are shown which demonstrate significant improvement
of the PINNs approximations with the increase in Nc . Additionally, parameters related to the
lP
efficiency of the network training processes with initialization of data-driven enhancement are
reported in Tab. 5. The loss terms quickly reduce by orders of magnitude in the second train-
ing phase which indicates that for the considered network architecture, Nc = 15000 is possibly
optimal.
rna

6. Discussions :

In the current study, a generalized PINNs framework for solving problems in linear contin-
uum elasticity in the field of solid mechanics is presented. The fundamentals of the PINNs
Jou

framework involve a construction of the loss function for physics-informed learning of the NNs
through the embedding of the linear constraint during training. Following the PINNs philos-
ophy to solve the linear elastic problem accurately, a multi-objective loss function has been
formulated and implemented. The proposed multi-objective loss function consists of the resid-
ual of the governing PDE, various boundary conditions, and data-driven physical knowledge
fitting terms. Additionally, weights corresponding to the terms in the loss function dictate
 Journal Pre-proof
34

the emphasis on satisfying the specific loss terms. To demonstrate the efficacy of the frame-
work, the Airy solution to an end-loaded cantilever beam and the Kirchhoff-Love plate theory
governed by fourth-order Biharmonic PDE has been solved. The proposed PINNs framework
is shown to accurately solve different fields in both problems. Parametric investigations on

of
activation functions and network architectures highlight the scope of improvement in terms of
solution accuracy and performance. Data-driven enhancement of the PINNs approximations
using analytical solutions significantly boosts accuracy and speed only using minimal network

pro
parameters. Therefore, such an approach can be employed to enhance solution accuracy for
complex PDEs. Additionally, the applicability of a smart initialization of data-driven en-
hancement learning-based approach quickening the training process and also improving model
accuracy have been illustrated. Such an approach would be key in achieving computational
re-
efficiency beyond conventional computational methods for solving linear continuum elasticity.
The proposed PINNs elasticity solvers utilize Tensorflow as the backend which can be easily de-
ployed in CPU/ GPU clusters, whereas, conventional algorithms lack such adaptability. Thus,
it opens new possibilities for solving complex elasticity problems that have remained unsolved
lP
by conventional numerical algorithms in the regime of continuum mechanics. It is however
worth noting that exploitation of the computational advantages of the PINNs framework de-
pends on various factors including the choice of the network architectures, hyperparameter
tuning, sampling techniques (distribution) of collocation points, etc. It has been shown that
rna

appropriate combinations of such factors significantly improve the training process and the
trained models.
In the present study, random sampling of the collocation points has been considered which
is simple, yet powerful, that can lead to a significantly better reconstruction of the elastic
fields. Importantly, this approach does not increase computational complexity, and it is easy to
Jou

implement. However, in elastic/elastoplastic PDE problem which exhibits local behavior (e.g.,
in presence of sharp, or very localized, features) or problems with singularities the performance
of PINNs may vary drastically with various sampling procedures (Daw et al., 2022; Leiteritz
and Pflüger, 2021). To overcome such an issue, a failure-informed adaptive enrichment strategy
such as failure-informed PINNs (FI-PINNs) can be employed that adopts the failure probability
as the posterior error indicator to generate new training points in the failure region (Gao et al.,
 Journal Pre-proof
35

2022). Furthermore, the basic resampling scheme can be further improved with a gradient-
based adaptive scheme to relocate the collocation points through a cosine-annealing to areas
with higher loss gradient, without increasing the total number of points that demonstrated
significant improvement under relatively fewer number of collocation points and sharper forcing

of
function (Subramanian et al., 2022). In addition, the evolutionary sampling (Evo) method
(Daw et al., 2022) that can incrementally accumulate collocation points in regions of high
PDE residuals can be an efficient choice for solving various time-dependent PDEs with little

pro
to no computational overhead. Instead of using a random approach such as Latin Hypercube
sampling, in the future, different deterministic and pseudo-random sampling strategies such as
Sparse Grid sampling or Sobol Sequences can be employed to further improve the performance
of the model.
re-
Furthermore, it is critical to obtain the statics of saturation along different parts of the
solution domain during the training of DNNs (Glorot and Bengio, 2010; Rakitianskaia and
Engelbrecht, 2015b). The saturation occurs when the hidden units of a DNN predominantly
output values close to the asymptotic ends of the activation function range which reduces the
lP
particular PINNs model to a binary state, thus limiting the overall information capacity of
the NN (Rakitianskaia and Engelbrecht, 2015a; Bai et al., 2019). The saturated units can
make gradient descent learning slow and inefficient due to small derivative values near the
asymptotes which can hinder the training PINNs efficiently (Bai et al., 2019). Thus, in the
rna

future, NN saturation can be studied quantitatively in relation to the ability of NNs to learn,
generalize, and the degree of regression accuracy. In addition, various weighting coefficients of
the loss terms in Eq. 8 and implementation of second-order optimization techniques (Tan and
Lim, 2019) can accelerate the training significantly. Based on the performance of the PINNs
framework herein, further studies quantifying the computational gains of the PINNs approach
Jou

compared to conventional numerical methods are in order. The proposed approach can be
extended to the solution in various computational mechanics problems such as soil plastic-
ity (Chen and Baladi, 1985; Bousshine et al., 2001), strain-gradient plasticity (Guha et al.,
2013, 2014), composite modeling (Roy, 2021c) etc. Furthermore, the present model can be
employed to predict microstructure evolution in Phase-field (PF) approach including various
solid-solid phase transitions (PTs) (Levitas et al., 2013; Levitas and Roy, 2015; Roy, 2020c,a,b),
 Journal Pre-proof
36

solid-solid PT via intermediate melting (Levitas and Roy, 2016; Roy, 2021a,f,e,d,b, 2022d), etc.

7. Conclusions :

of
Summarizing, the current work presents a deep learning framework based on the fundamen-
tals of PINNs theory for the solution of linear elasticity problems in continuum mechanics. A
multi-objective loss function is proposed for the linear elastic solid problems that include gov-

pro
erning PDE, Dirichlet, and Neumann boundary conditions across randomly chosen collocation
points in the problem domain. Multiple deep network models trained to predict different fields
result in a more accurate representation. Traditional ML/ DL approaches that only rely on fit-
ting a model that establishes complex, high-dimensional, non-linear relationships between the
re-
input features and outputs, are unable to incorporate rich information available through gov-
erning equations/ physics-based mathematical modeling of physical phenomena. Conventional
computational techniques on the other hand rely completely on such physical information for
prediction. The PINNs approach combines the benefits of the DL techniques in the extraction
lP
of complex relations from data with the advantages of the conventional numerical techniques
for physical modeling. The proposed method may be extended to nonlinear elasticity, vis-
coplasticity, elastoplasticity, and various other mechanics and material science problems. The
present work builds a solid foundation for new promising avenues for future work in machine
rna

learning applications in solid mechanics.

Acknowledgements: The support of the Aeronautical Research and Development Board

(Grant No. DARO/08/1051450/M/I) is gratefully acknowledged.
Jou

Competing interests: The author declares no competing interests.

References
Arora, R., Kakkar, P., Dey, B., and Chakraborty, A. (2022). Physics-informed neural networks
for modeling rate-and temperature-dependent plasticity. arXiv preprint arXiv:2201.08363.

Atkin, R. J. and Fox, N. (2005). An introduction to the theory of elasticity. Courier Corporation.
 Journal Pre-proof
37

Bai, W., Zhou, Q., Li, T., and Li, H. (2019). Adaptive reinforcement learning neural net-
work control for uncertain nonlinear system with input saturation. IEEE transactions on
cybernetics, 50(8):3433–3443.

of
Batra, R., Song, L., and Ramprasad, R. (2021). Emerging materials intelligence ecosystems
propelled by machine learning. Nature Reviews Materials, 6(8):655–678.

Baydin, A. G., Pearlmutter, B. A., Radul, A. A., and Siskind, J. M. (2018). Automatic

pro
differentiation in machine learning: a survey. Journal of machine learning research, 18.

Bekar, A. C., Madenci, E., Haghighat, E., Waheed, U. b., and Alkhalifah, T. (2022). Solving the
eikonal equation for compressional and shear waves in anisotropic media using peridynamic
differential operator. Geophysical Journal International, 229(3):1942–1963.
re-
Bergen, K. J., Johnson, P. A., Maarten, V., and Beroza, G. C. (2019). Machine learning for
data-driven discovery in solid earth geoscience. Science, 363(6433).

Bilbao, I. and Bilbao, J. (2017). Overfitting problem and the over-training in the era of
lP
data: Particularly for artificial neural networks. In 2017 eighth international conference on
intelligent computing and information systems (ICICIS), pages 173–177. IEEE.

bin Waheed, U., Alkhalifah, T., Haghighat, E., and Song, C. (2022). A holistic approach to
rna

computing first-arrival traveltimes using neural networks. In Advances in Subsurface Data

Analytics, pages 251–278. Elsevier.

bin Waheed, U., Haghighat, E., Alkhalifah, T., Song, C., and Hao, Q. (2021). Pinneik: Eikonal
solution using physics-informed neural networks. Computers & Geosciences, 155:104833.
Jou

Bose, R. and Roy, A. (2022). Accurate deep learning sub-grid scale models for large eddy
simulations. Bulletin of the American Physical Society.

Bousshine, L., Chaaba, A., and De Saxce, G. (2001). Softening in stress–strain curve for
drucker–prager non-associated plasticity. International Journal of Plasticity, 17(1):21–46.

Bower, A. F. (2009). Applied mechanics of solids. CRC press.

Boyd, J. P. (2001). Chebyshev and Fourier spectral methods. Courier Corporation.

 Journal Pre-proof
38

Brunton, S. L., Noack, B. R., and Koumoutsakos, P. (2020). Machine learning for fluid me-
chanics. Annual Review of Fluid Mechanics, 52:477–508.

Butler, K. T., Davies, D. W., Cartwright, H., Isayev, O., and Walsh, A. (2018). Machine
learning for molecular and materials science. Nature, 559(7715):547–555.

of
Cai, S., Wang, Z., Wang, S., Perdikaris, P., and Karniadakis, G. E. (2021). Physics-informed
neural networks for heat transfer problems. Journal of Heat Transfer, 143(6).

pro
Chandio, A., Gui, G., Kumar, T., Ullah, I., Ranjbarzadeh, R., Roy, A. M., Hussain, A., and
Shen, Y. (2022). Precise single-stage detector. arXiv preprint arXiv:2210.04252.

Chen, W.-F. and Baladi, G. Y. (1985). Soil plasticity: theory and implementation. Elsevier.

re-
Ching, T., Himmelstein, D. S., Beaulieu-Jones, B. K., Kalinin, A. A., Do, B. T., Way, G. P.,
Ferrero, E., Agapow, P.-M., Zietz, M., Hoffman, M. M., et al. (2018). Opportunities and
obstacles for deep learning in biology and medicine. Journal of The Royal Society Interface,
15(141):20170387.
lP
Chollet, F. et al. (2015). keras.

Dana, S. and Wheeler, M. F. (2020). A machine learning accelerated fe homogenization algo-

rithm for elastic solids. arXiv preprint arXiv:2003.11372.
rna

Daw, A., Bu, J., Wang, S., Perdikaris, P., and Karpatne, A. (2022). Rethinking the importance
of sampling in physics-informed neural networks. arXiv preprint arXiv:2207.02338.

De Ryck, T., Jagtap, A. D., and Mishra, S. (2022). Error estimates for physics informed neural
networks approximating the navier-stokes equations. arXiv preprint arXiv:2203.09346.
Jou

DeVries, P. M., Viégas, F., Wattenberg, M., and Meade, B. J. (2018). Deep learning of
aftershock patterns following large earthquakes. Nature, 560(7720):632–634.

Du, Y. and Zaki, T. A. (2021). Evolutional deep neural network. Phys. Rev. E, 104:045303.

Frankel, A., Tachida, K., and Jones, R. (2020). Prediction of the evolution of the stress field
of polycrystals undergoing elastic-plastic deformation with a hybrid neural network model.
Machine Learning: Science and Technology, 1(3):035005.
 Journal Pre-proof
39

Gao, Z., Yan, L., and Zhou, T. (2022). Failure-informed adaptive sampling for pinns. arXiv
preprint arXiv:2210.00279.

Glorot, X. and Bengio, Y. (2010). Understanding the difficulty of training deep feedforward
neural networks. In Proceedings of the thirteenth international conference on artificial intel-

of
ligence and statistics, pages 249–256. JMLR Workshop and Conference Proceedings.

Glowacz, A. (2021). Fault diagnosis of electric impact drills using thermal imaging. Measure-

pro
ment, 171:108815.

Glowacz, A. (2022). Thermographic fault diagnosis of shaft of bldc motor. Sensors, 22(21):8537.

Goswami, S., Anitescu, C., Chakraborty, S., and Rabczuk, T. (2020). Transfer learning en-
hanced physics informed neural network for phase-field modeling of fracture. Theoretical
re-
and Applied Fracture Mechanics, 106:102447.

Goswami, S., Yin, M., Yu, Y., and Karniadakis, G. E. (2022). A physics-informed variational
deeponet for predicting crack path in quasi-brittle materials. Computer Methods in Applied
lP
Mechanics and Engineering, 391:114587.

Guha, S., Sangal, S., and Basu, S. (2013). Finite element studies on indentation size effect
using a higher order strain gradient theory. International Journal of Solids and Structures,
50(6):863–875.
rna

Guha, S., Sangal, S., and Basu, S. (2014). On the fracture of small samples under higher order
strain gradient plasticity. International Journal of Fracture, 187(2):213–226.

Guo, M. and Haghighat, E. (2020). An energy-based error bound of physics-informed neural

network solutions in elasticity. arXiv preprint arXiv:2010.09088.
Jou

Guo, X., Liu, X., Królczyk, G., Sulowicz, M., Glowacz, A., Gardoni, P., and Li, Z. (2022).
Damage detection for conveyor belt surface based on conditional cycle generative adversarial
network. Sensors, 22(9):3485.

Haghighat, E., Amini, D., and Juanes, R. (2022). Physics-informed neural network simulation
of multiphase poroelasticity using stress-split sequential training. Computer Methods in
Applied Mechanics and Engineering, 397:115141.
 Journal Pre-proof
40

Haghighat, E., Bekar, A. C., Madenci, E., and Juanes, R. (2021a). A nonlocal physics-informed
deep learning framework using the peridynamic differential operator. Computer Methods in
Applied Mechanics and Engineering, 385:114012.

of
Haghighat, E. and Juanes, R. (2021). Sciann: A keras/tensorflow wrapper for scientific com-
putations and physics-informed deep learning using artificial neural networks. Computer
Methods in Applied Mechanics and Engineering, 373:113552.

pro
Haghighat, E., Raissi, M., Moure, A., Gomez, H., and Juanes, R. (2020). A deep learning
framework for solution and discovery in solid mechanics. arXiv preprint arXiv:2003.02751.

Haghighat, E., Raissi, M., Moure, A., Gomez, H., and Juanes, R. (2021b). A physics-informed
deep learning framework for inversion and surrogate modeling in solid mechanics. Computer
re-
Methods in Applied Mechanics and Engineering, 379:113741.

Hu, Z., Jagtap, A. D., Karniadakis, G. E., and Kawaguchi, K. (2021). When do ex-
tended physics-informed neural networks (xpinns) improve generalization? arXiv preprint
arXiv:2109.09444.
lP
Irfan, M., Iftikhar, M. A., Yasin, S., Draz, U., Ali, T., Hussain, S., Bukhari, S., Alwadie,
A. S., Rahman, S., Glowacz, A., et al. (2021). Role of hybrid deep neural networks (hdnns),
computed tomography, and chest x-rays for the detection of covid-19. International Journal
rna

of Environmental Research and Public Health, 18(6):3056.

Jabbar, H. and Khan, R. Z. (2015). Methods to avoid over-fitting and under-fitting in su-
pervised machine learning (comparative study). Computer Science, Communication and
Instrumentation Devices, 70.
Jou

Jagtap, A. D. and Karniadakis, G. E. (2021). Extended physics-informed neural networks

(xpinns): A generalized space-time domain decomposition based deep learning framework
for nonlinear partial differential equations. In AAAI Spring Symposium: MLPS.

Jagtap, A. D., Kharazmi, E., and Karniadakis, G. E. (2020). Conservative physics-informed

neural networks on discrete domains for conservation laws: Applications to forward and
inverse problems. Computer Methods in Applied Mechanics and Engineering, 365:113028.
 Journal Pre-proof
41

Jagtap, A. D., Mao, Z., Adams, N., and Karniadakis, G. E. (2022). Physics-informed neural
networks for inverse problems in supersonic flows. arXiv preprint arXiv:2202.11821.

Jahanbakht, M., Xiang, W., and Azghadi, M. R. (2022). Sediment prediction in the great

of
barrier reef using vision transformer with finite element analysis. Neural Networks, 152:311–
321.

Jamil, S., Abbas, M. S., and Roy, A. M. (2022). Distinguishing malicious drones using vision

pro
transformer. AI, 3(2):260–273.

Jin, X., Cai, S., Li, H., and Karniadakis, G. E. (2021). Nsfnets (navier-stokes flow nets):
Physics-informed neural networks for the incompressible navier-stokes equations. Journal of
Computational Physics, 426:109951.
re-
Karniadakis, G. E., Kevrekidis, I. G., Lu, L., Perdikaris, P., Wang, S., and Yang, L. (2021).
Physics-informed machine learning. Nature Reviews Physics, 3(6):422–440.

Khan, W., Kumar, T., Cheng, Z., Raj, K., Roy, A. M., and Luo, B. (2022a). Sql and nosql
lP
databases software architectures performance analysis and assessments–a systematic litera-
ture review. arXiv preprint arXiv:2209.06977.

Khan, W., Raj, K., Kumar, T., Roy, A. M., and Luo, B. (2022b). Introducing urdu digits
rna

dataset with demonstration of an efficient and robust noisy decoder-based pseudo example
generator. Symmetry, 14(10):1976.

Krishnapriyan, A., Gholami, A., Zhe, S., Kirby, R., and Mahoney, M. W. (2021). Charac-
terizing possible failure modes in physics-informed neural networks. Advances in Neural
Information Processing Systems, 34.
Jou

Kutz, J. N. (2017). Deep learning in fluid dynamics. Journal of Fluid Mechanics, 814:1–4.

Lagaris, I. E., Likas, A., and Fotiadis, D. I. (1998). Artificial neural networks for solving
ordinary and partial differential equations. IEEE transactions on neural networks, 9(5):987–
1000.

LeCun, Y., Bengio, Y., and Hinton, G. (2015). Deep learning. nature, 521(7553):436–444.
 Journal Pre-proof
42

Leiteritz, R. and Pflüger, D. (2021). How to avoid trivial solutions in physics-informed neural
networks. arXiv preprint arXiv:2112.05620.

Levitas, V. I. and Roy, A. M. (2015). Multiphase phase field theory for temperature-and

of
stress-induced phase transformations. Physical Review B, 91(17):174109.

Levitas, V. I. and Roy, A. M. (2016). Multiphase phase field theory for temperature-induced
phase transformations: Formulation and application to interfacial phases. Acta Materialia,

pro
105:244–257.

Levitas, V. I., Roy, A. M., and Preston, D. L. (2013). Multiple twinning and variant-variant
transformations in martensite: phase-field approach. Physical Review B, 88(5):054113.

Ling, J., Kurzawski, A., and Templeton, J. (2016). Reynolds averaged turbulence modelling
re-
using deep neural networks with embedded invariance. Journal of Fluid Mechanics, 807:155–
166.

Lou, Q., Meng, X., and Karniadakis, G. E. (2021). Physics-informed neural networks for
lP
solving forward and inverse flow problems via the boltzmann-bgk formulation. Journal of
Computational Physics, 447:110676.

Lurie, A. I. (2010). Theory of elasticity. Springer Science & Business Media.

rna

Määttä, J., Bazaliy, V., Kimari, J., Djurabekova, F., Nordlund, K., and Roos, T. (2021).
Gradient-based training and pruning of radial basis function networks with an application
in materials physics. Neural Networks, 133:123–131.

Marsden, J. E. and Hughes, T. J. (1994). Mathematical foundations of elasticity. Courier

Jou

Corporation.

McClenny, L. and Braga-Neto, U. (2020). Self-adaptive physics-informed neural networks using

a soft attention mechanism. arXiv preprint arXiv:2009.04544.

Racca, A. and Magri, L. (2021). Robust optimization and validation of echo state networks
for learning chaotic dynamics. Neural Networks, 142:252–268.
 Journal Pre-proof
43

Raissi, M. and Karniadakis, G. E. (2018). Hidden physics models: Machine learning of non-
linear partial differential equations. Journal of Computational Physics, 357:125–141.

Raissi, M., Perdikaris, P., and Karniadakis, G. E. (2019). Physics-informed neural networks: A

of
deep learning framework for solving forward and inverse problems involving nonlinear partial
differential equations. Journal of Computational Physics, 378:686–707.

Raissi, M., Yazdani, A., and Karniadakis, G. E. (2020). Hidden fluid mechanics: Learning

pro
velocity and pressure fields from flow visualizations. Science, 367(6481):1026–1030.

Rakitianskaia, A. and Engelbrecht, A. (2015a). Measuring saturation in neural networks. In

2015 IEEE symposium series on computational intelligence, pages 1423–1430. IEEE.

Rakitianskaia, A. and Engelbrecht, A. (2015b). Saturation in pso neural network training:

re-
Good or evil? In 2015 IEEE Congress on Evolutionary Computation (CEC), pages 125–
132. IEEE.

Ramprasad, R., Batra, R., Pilania, G., Mannodi-Kanakkithodi, A., and Kim, C. (2017). Ma-
lP
chine learning in materials informatics: recent applications and prospects. npj Computational
Materials, 3(1):1–13.

Rao, C., Sun, H., and Liu, Y. (2021). Physics-informed deep learning for computational
rna

elastodynamics without labeled data. Journal of Engineering Mechanics, 147(8):04021043.

Rawat, W. and Wang, Z. (2017). Deep convolutional neural networks for image classification:
A comprehensive review. Neural computation, 29(9):2352–2449.

Reddy, J. N. (2006). Theory and analysis of elastic plates and shells. CRC press.
Jou

Rezaei, S., Harandi, A., Moeineddin, A., Xu, B.-X., and Reese, S. (2022). A mixed formu-
lation for physics-informed neural networks as a potential solver for engineering problems
in heterogeneous domains: comparison with finite element method. Computer Methods in
Applied Mechanics and Engineering, 401:115616.

Roy, A. M. (2020a). Effects of interfacial stress in phase field approach for martensitic phase
transformation in nial shape memory alloys. Applied Physics A, 126(7):1–12.
 Journal Pre-proof
44

Roy, A. M. (2020b). Evolution of martensitic nanostructure in nial alloys: tip splitting and
bending. Material Science Research India (Online), 17(special 1):03–06.

Roy, A. M. (2020c). Influence of interfacial stress on microstructural evolution in nial alloys.

of
JETP Letters, 112(3):173–179.

Roy, A. M. (2021a). Barrierless melt nucleation at solid-solid interface in energetic nitramine

octahydro-1, 3, 5, 7-tetranitro-1, 3, 5, 7-tetrazocine. Materialia, 15:101000.

pro
Roy, A. M. (2021b). Energetics and kinematics of undercooled nonequilibrium interfacial
molten layer in cyclotetramethylene-tetranitramine crystal. Physica B: Condensed Matter,
615:412986.

Roy, A. M. (2021c). Finite element framework for efficient design of three dimensional multi-
re-
component composite helicopter rotor blade system. Eng, 2(1):69–79.

Roy, A. M. (2021d). Formation and stability of nanosized, undercooled propagating inter-

mediate melt during β→ δ phase transformation in hmx nanocrystal. Europhysics Letters,
lP
133(5):56001.

Roy, A. M. (2021e). Influence of nanoscale parameters on solid–solid phase transformation in

octogen crystal: Multiple solution and temperature effect. JETP Letters, 113(4):265–272.
rna

Roy, A. M. (2021f). Multiphase phase-field approach for solid–solid phase transformations via
propagating interfacial phase in hmx. Journal of Applied Physics, 129(2):025103.

Roy, A. M. (2022a). Adaptive transfer learning-based multiscale feature fused deep convolu-
tional neural network for eeg mi multiclassification in brain–computer interface. Engineering
Jou

Applications of Artificial Intelligence, 116:105347.

Roy, A. M. (2022b). An efficient multi-scale CNN model with intrinsic feature integration for
motor imagery EEG subject classification in brain-machine interfaces. Biomedical Signal
Processing and Control, 74:103496.

Roy, A. M. (2022c). A multi-scale fusion cnn model based on adaptive transfer learning for
multi-class mi-classification in bci system. BioRxiv.
 Journal Pre-proof
45

Roy, A. M. (2022d). Multiphase phase-field approach for virtual melting: a brief review. Roy
AM Multiphase Phase-Field Approach for Virtual Melting: A Brief Review. Mat. Sci. Res.
India, 18(2).

Roy, A. M. and Bhaduri, J. (2021). A deep learning enabled multi-class plant disease detection

of
model based on computer vision. AI, 2(3):413–428.

Roy, A. M. and Bhaduri, J. (2022). Real-time growth stage detection model for high degree

pro
of occultation using densenet-fused YOLOv4. Computers and Electronics in Agriculture,
193:106694.

Roy, A. M., Bhaduri, J., Kumar, T., and Raj, K. (2022a). A computer vision-based object
localization model for endangered wildlife detection. Ecological Economics, Forthcoming.
re-
Roy, A. M., Bhaduri, J., Kumar, T., and Raj, K. (2022b). Wildect-yolo: An efficient and
robust computer vision-based accurate object localization model for automated endangered
wildlife detection. Ecological Informatics, page 101919.
lP
Roy, A. M., Bose, R., and Bhaduri, J. (2022c). A fast accurate fine-grain object detection
model based on YOLOv4 deep neural network. Neural Computing and Applications, pages
1–27.
rna

Saha, P., Dash, S., and Mukhopadhyay, S. (2021). Physics-incorporated convolutional recur-
rent neural networks for source identification and forecasting of dynamical systems. Neural
Networks, 144:359–371.

Samaniego, E., Anitescu, C., Goswami, S., Nguyen-Thanh, V. M., Guo, H., Hamdia, K.,
Zhuang, X., and Rabczuk, T. (2020). An energy approach to the solution of partial differen-
Jou

tial equations in computational mechanics via machine learning: Concepts, implementation

and applications. Computer Methods in Applied Mechanics and Engineering, 362:112790.

Sengupta, T. (2013). High accuracy computing methods: fluid flows and wave phenomena.
Cambridge University Press.

Shukla, K., Jagtap, A. D., Blackshire, J. L., Sparkman, D., and Karniadakis, G. E. (2021).
A physics-informed neural network for quantifying the microstructural properties of poly-
 Journal Pre-proof
46

crystalline nickel using ultrasound data: A promising approach for solving inverse problems.
IEEE Signal Processing Magazine, 39(1):68–77.

Singh, A., Raj, K., Kumar, T., Verma, S., and Roy, A. M. (2023a). Deep learning-based

of
cost-effective and responsive robot for autism treatment. Drones, 7(2):81.

Singh, A., Ranjbarzadeh, R., Raj, K., Kumar, T., and Roy, A. M. (2023b). Understanding eeg
signals for subject-wise definition of armoni activities. arXiv preprint arXiv:2301.00948.

pro
Sirignano, J. and Spiliopoulos, K. (2018). Dgm: A deep learning algorithm for solving partial
differential equations. Journal of computational physics, 375:1339–1364.

Subramanian, S., Kirby, R. M., Mahoney, M. W., and Gholami, A. (2022). Adap-
tive self-supervision algorithms for physics-informed neural networks.
re- arXiv preprint
arXiv:2207.04084.

Sun, L., Gao, H., Pan, S., and Wang, J.-X. (2020). Surrogate modeling for fluid flows based on
physics-constrained deep learning without simulation data. Computer Methods in Applied
lP
Mechanics and Engineering, 361:112732.

Szilard, R. and Nash, W. (1974). Theory and analysis of plates, classical and numberical
methods.
rna

Tan, H. H. and Lim, K. H. (2019). Review of second-order optimization techniques in ar-

tificial neural networks backpropagation. In IOP conference series: materials science and
engineering, volume 495, page 012003. IOP Publishing.

Tartakovsky, A. M., Marrero, C. O., Perdikaris, P., Tartakovsky, G. D., and Barajas-Solano,
Jou

D. (2018). Learning parameters and constitutive relationships with physics informed deep
neural networks. arXiv preprint arXiv:1808.03398.

Timoshenko, S. (1970). Theory of elastic stability 2e. Tata McGraw-Hill Education.

Timoshenko, S. and Woinowsky-Krieger, S. (1959). Theory of plates and shells.

 Journal Pre-proof
47

Vahab, M., Haghighat, E., Khaleghi, M., and Khalili, N. (2021). A physics-informed neural
network approach to solution and identification of biharmonic equations of elasticity. Journal
of Engineering Mechanics, 148(2):04021154.

of
von Rueden, L., Mayer, S., Beckh, K., Georgiev, B., Giesselbach, S., Heese, R., Kirsch,
B., Pfrommer, J., Pick, A., Ramamurthy, R., et al. (2019). Informed machine learning–
a taxonomy and survey of integrating knowledge into learning systems. arXiv preprint

pro
arXiv:1903.12394.

Voulodimos, A., Doulamis, N., Doulamis, A., and Protopapadakis, E. (2018). Deep learning
for computer vision: A brief review. Computational intelligence and neuroscience, 2018.

Waheed, U., Haghighat, E., Alkhalifah, T., Song, C., and Hao, Q. (2020). Eikonal solution
re-
using physics-informed neural networks. In EAGE 2020 Annual Conference & Exhibition
Online, volume 2020, pages 1–5. European Association of Geoscientists & Engineers.

Xu, K., Huang, D. Z., and Darve, E. (2021). Learning constitutive relations using symmetric
positive definite neural networks. Journal of Computational Physics, 428:110072.
lP
Zhang, E., Dao, M., Karniadakis, G. E., and Suresh, S. (2022). Analyses of internal struc-
tures and defects in materials using physics-informed neural networks. Science advances,
8(7):eabk0644.
rna

Zhang, E., Yin, M., and Karniadakis, G. E. (2020). Physics-informed neural networks for non-
homogeneous material identification in elasticity imaging. arXiv preprint arXiv:2009.04525.

Zhao, Z.-Q., Zheng, P., Xu, S.-t., and Wu, X. (2019). Object detection with deep learning: A
review. IEEE transactions on neural networks and learning systems, 30(11):3212–3232.
Jou

Zhu, Q., Liu, Z., and Yan, J. (2021). Machine learning for metal additive manufacturing:
predicting temperature and melt pool fluid dynamics using physics-informed neural networks.
Computational Mechanics, 67(2):619–635.

Zienkiewicz, O. C. and Taylor, R. L. (2005). The finite element method for solid and structural
mechanics. Elsevier.
 Journal Pre-proof

Title page information

of
Title : Deep learning-accelerated computational framework based on
Physics Informed Neural Network for solution of Linear Elasticity

pro
First author: Arunabha Mohan Ro y , Ph.D.
Affiliation: University of Michigan, Aerospace Engineering, Ann Arbor, MI 48109,
U.S.A.
re-
Present address: 2032 François-Xavier Bagnoud Aerospace Building
1320 Beal Avenue Ann Arbor, MI 48109-2140, USA
lP
Second author: Rikhi Bose, Ph.D.
Affiliation: Johns Hopkins University, Mechanical Engineering, Baltimore, MD 21218,
U.S.A.
rna

Corresponding Author: Arunabha Mohan Roy

Jou

Email ID: [email protected]

 Journal Pre-proof

The major highlights of this work are the followings:

 An efficient and robust data-driven deep learning accelerated computational framework based on the
fundamentals of the Physics Informed Neural Networks (PINNs) has been developed for linear continuum

of
elasticity problems.

 A novel multi-objective loss functional is proposed, when minimized, results in the model accurately capturing
physics in the elasticity solution.

pro
 Several benchmark problems including the Airy solution to elasticity and Kirchhoff-Love plate problems are
solved illustrating the superiority of the current framework showing excellent agreement with analytical solutions

 Current study illustrates the applicability of data-driven enhancement using the transfer learning-based
approach in reducing training time, simultaneously improving the accuracy of the model


re-
The present study combines the benefits of the classical methods depending on the physical information
available in analytical relations with the superior capabilities of the deep learning techniques in the data-driven
construction of lightweight, yet accurate and robust neural networks.
lP
rna
Jou
 Journal Pre-proof

Declaratio if ioterettt

☒ The authors declare that they have no known competng fnancial interests or personal relatonships
that could have appeared to infuence the work reported in this paper.

of
☐The authors declare the following fnancial interestsppersonal relatonships which may be considered
as potental competng interests:

pro
re-
lP
rna
Jou