International Journal of Solids and Structures 277-278 (2023) 112319

Contents lists available at ScienceDirect

International Journal of Solids and Structures

journal homepage: www.elsevier.com/locate/ijsolstr

Application of Physics-Informed Neural Networks for forward and inverse

analysis of pile–soil interaction
M. Vahab a , B. Shahbodagh a ,∗, E. Haghighat b , N. Khalili a
a
School of Civil and Environmental Engineering, The University of New South Wales, Sydney 2052, Australia
b
Massachusetts Institute of Technology, Cambridge, MA, USA

ARTICLE INFO ABSTRACT

Keywords: The application of the Physics-Informed Neural Networks (PINNs) to forward and inverse analysis of pile–soil
Physics-Informed Neural Networks (PINNs) interaction problems is presented. The main challenge encountered in the Artificial Neural Network (ANN)
Deep learning modeling of pile–soil interaction is the presence of abrupt changes in material properties, which results in large
Pile–soil interaction
discontinuities in the gradient of the displacement solution. Therefore, a domain-decomposition multi-network
SciANN
model is proposed to deal with the discontinuities in the strain fields at common boundaries of pile–soil
regions and soil layers. The application of the model to the analysis and parametric study of single piles
embedded in both homogeneous and layered formations is demonstrated under axisymmetric and plane strain
conditions. The performance of the model in parameter identification (inverse analysis) of pile–soil interaction
is particularly investigated. It is shown that by using PINNs, the localized data acquired along the pile length
–possibly obtained via fiber optic strain sensing–can be successfully used for the inversion of soil parameters
in layered formations.

1. Introduction
The Finite Element Method (FEM) is ubiquitously used for the
computational analysis of various geotechnical engineering problems
(Zienkiewicz et al., 1999; Ghasemi-Fare and Basu, 2013; Rahmani and
Pak, 2012; Jafari et al., 2021; Versteijlen et al., 2018; Ebrahimian
et al., 2012). Consistently, in many classical studies, FEM has been the
natural choice for the design of inversion or back-analysis algorithms
for inferring important model parameters, like mechanical properties
of soils (Calvello and Finno, 2004). The state-of-the-art techniques in
this context mainly rely on minimization of the deviations between
the simulation responses (due to an adjustable set of model param-
eters) with respect to field and laboratory measurements. For this
purpose, a range of direct or gradient-based optimization techniques
are extended which exploit successive numerical solutions iteratively to
yield optimal precision for the sought-after parameters (Kabe, 1985).
Nonetheless, both approaches can be prohibitively computationally
expensive in particular for large-scale problems. Further challenges
encountered during numerical back analysis in engineering applications
have been due to non-uniqueness, material model limitations, and
disparate data sources (Vardakos et al., 2012; Walton and Sinha, 2022).
Deep Learning (DL) has proven to be a rigorous approach for
the forward and back analysis of geotechnical engineering problems
(Lefik and Schrefler, 2002; Gawin et al., 2001; Parish and Duraisamy,
2016; Kardani et al., 2020; Mei et al., 2022). Compared to traditional
computational mechanics solvers, (DL) has the advantage of offering
linear time complexity and surpassing them in performance beyond
training (Emu et al., 2020). In the analysis of piles, which is the focus of
this study, DL has been applied for the prediction of the shaft and tip re-
sistance of concrete piles by Momeni et al. (2015), for estimation of the
uplift resistance of screw piles by Moayedi and Mosallanezhad (2017),
and for evaluation of the lateral load bearing capacity of piles by Das
and Basudhar (2006) and Armaghani et al. (2017). It is noteworthy that
the enumerated studies emphasize on forward prediction of mechanical
response of piles itself. Based on the Universal approximation theo-
rem (a.k.a., Cybenko theorem, see Cybenko, 1989), neural networks
comprising of at least a single hidden layer can uniformly reconstruct
any function with arbitrary continuous non-linearity. Evidently, the
composite system of soil–piles involves a strong material discontinuity,
which in turn leads to a discontinuous strain field (i.e., the gradient
of displacement). This study aims to develop a DL framework capable
of handling the discontinuities induced by material interfaces within
soil–pile systems.
Early Artificial Intelligence (AI) approaches were developed in the
mid-20th century to tackle the intellectually difficult and complex-to-
articulate problems, which pertain to straight-forward solution algo-
rithms within relatively sterile environments (Bishop, 2006). DL aims

∗ Corresponding author.
E-mail address: b.shahbodagh@unsw.edu.au (B. Shahbodagh).

https://doi.org/10.1016/j.ijsolstr.2023.112319
Received 22 October 2022; Received in revised form 7 May 2023; Accepted 9 May 2023
Available online 19 May 2023
0020-7683/© 2023 Elsevier Ltd. All rights reserved.
M. Vahab et al. International Journal of Solids and Structures 277-278 (2023) 112319

to mitigate major sources of difficulty with conventional real-world

Machine Learning (ML) practice, in terms of data sensitivity, noises,
and other representation learning issues, through a nested hierarchy of
simpler representations/concepts (Goodfellow et al., 2016). Nowadays
DL algorithms are predominantly employed in an increasing number of
areas in engineering and science, including geotechnics (Zhang et al.,
2021), structural engineering (Azimi and Pekcan, 2020; Bao et al.,
2019), engineering mechanics (Li et al., 2019), reservoir engineer-
ing (Zhang et al., 2018), material science (Azimi et al., 2018), physical
modeling (De Bézenac et al., 2019), and earth sciences (Reichstein
et al., 2019), to name a few.
Most conventional DL algorithms rely on the availability of data
for training to enable the development of reliable predictive models.
However, in many engineering circumstances, the data acquisition costs
could be prohibitive. Partial information threatens the robustness of
machine learning to the extent that draws decision-making cumber-
some, if not impossible. In a diversity of engineering and scientific
applications, there exists a vast prior physical knowledge that can be
exercised as a regularization agent rendering the admissible solution
space into a manageable size (Karniadakis et al., 2021). These ap-
proaches could be considered as a subset of Reinforcement Learning Fig. 1. Schematic representation of a single pile in a layered soil medium.
(RL) (Sutton and Barto, 2018) in a sense that a range of incentives are
optimized rather than relying on the training dataset alone (Han et al.,
2018; Haghighat et al., 2021b). Concurrent efforts in recent years have
et al., 2020), which has been generalized for any PDE of choice in
been made to incorporate prior physical information in DL by Owhadi
extended PINNs (Jagtap and Karniadakis, 2021). The localized data
(2015), Han et al. (2018), Bar-Sinai et al. (2019), Rudy et al. (2017),
acquired along the pile length, in turn, is utilized for the inversion of
and Raissi et al. (2019).
key mechanical properties of soil in layered formations.
Physics-Informed Neural Networks (PINNs) are a class of deep
This paper is organized as follows: In Section 2, the equations
learning that incorporate a series of physical laws, frequently described
governing the pile–soil interactions are explained in detail. Section 3
in the form of partial differential equations (PDEs), to steer the learn-
is devoted to a brief introduction to the fundamentals of PINNs and its
ing towards the solution for sparse training data-sets, which could
application to the solution of pile–soil systems. The forward solution of
not be plausible with classic DL algorithms. The prosperity of PINNs
single piles –in both cylindrical and Cartesian coordinate systems –is
is attributed to substantial algorithmic advances (e.g., graph-based
explored in Section 4. In addition, the inverse analysis of soil mechan-
automated differentiation Baydin et al., 2018) and major software
ical properties is conducted for homogeneous and layered formations.
developments (e.g., TensorFlow Abadi et al., 2016, Keras Chollet et al.,
Concluding remarks are presented in Section 5.
2015). PINNs have been utilized with great success in a broad range
of engineering disciplines, including solid mechanics (Haghighat and
2. Problem statement and governing equations
Juanes, 2021; Haghighat et al., 2021b,a; Vahab et al., 2022; Khaleghi
et al., 2022), fluid mechanics (Mao et al., 2020; Sahli Costabal et al.,
Fig. 1 depicts the pile–soil system considered in this study. The
2020; Kharazmi et al., 2021), and thermo-mechanics (Niaki et al.,
system is modeled in both two-dimensional (2D) and three-dimensional
2021). Conventional data-driven inversion models frequently fail in
(3D) settings, representing a sheet-pile wall and a single cylindrical
generalization, owing to extrapolation or observational biases (Karni-
pile embedded in soil media, respectively. The Cartesian coordinate
adakis et al., 2021; Raissi et al., 2019). A beneficial remedy can be
system is used for the 2D plane strain analysis, whereas the cylindrical
offered by physics-informed deep learning, which provides the network
coordinate system is employed for the 3D analysis of pile–soil inter-
model along with the laws of physics governing the system. This can
action. In the figure, Ω𝑃 and Ω𝑆𝑘 (𝑘 = 1, … , 𝑁) denote the regions
rectify the issues associated with the missing ingredients induced by
of the space occupied by the pile and the 𝑘th soil layer, respectively,
the sparsity of data, uncertainties, data noises, or other less understood
where N is the number of the soil layers. The boundaries of the pile
factors (Khaleghi et al., 2022). Inverse physics-informed solutions have
and soil regions are designated by Γ𝑝 and Γ𝑆𝑘 , respectively. The pile is
been hitherto investigated in nano optics by Chen et al. (2020), in solid
assumed to be continuously bonded to and fully embedded in the soil
mechanics by Haghighat et al. (2021b), for conservation laws by Jagtap
medium. We leverage a generic representation of the pile deformation
et al. (2020), and for flow problems by Lou et al. (2021).
which enables the model to capture the non-uniform deformation along
In this study, we emphasize on a novel application of PINNs to
the cross-section of the pile, essential for the proper simulation of
the solution and inverse analysis of soil–pile interaction problems. For
short pile response and interactions at low pile–soil stiffness ratios. The
this sake, we use the longitudinal strain profile along the entire length
equilibrium equations governing the pile–soil system are expressed in
of piles, which could be obtained through the optical fiber strain-
the cylindrical coordinate system as
sensing technique in practice (e.g., see Mohamad et al., 2011, 2012).
𝛼 𝜎𝛼
Differential strains (i.e., at either side of piles) may also be extracted 1 𝜕 ( 𝛼 ) 1 𝜕𝜎𝑟𝜃 𝜕𝜎 𝛼
by the installation of a group of fibers so as to enable the monitor- 𝑟𝜎𝑟𝑟 + + 𝑟𝑧 − 𝜃𝜃 + 𝑓𝑟𝛼 = 0 ,
𝑟 𝜕𝑟 𝑟 𝜕𝜃 𝜕𝑧 𝑟
ing of both axial movements and bending (Mohamad et al., 2011).
Here, PINNs are first employed to construct efficient neural networks 𝛼 𝜕𝜎 𝛼 𝜎𝛼
1 𝜕 ( 𝛼 ) 1 𝜕𝜎𝜃𝜃 ( )
for the solution of soil–pile interactions in the absence of any data. 𝑟𝜎𝑟𝜃 + + 𝜃𝑧 + 𝑟𝜃 + 𝑓𝜃𝛼 = 0 , on Ω𝛼 𝛼 = 𝑃 , 𝑆𝑘
𝑟 𝜕𝑟 𝑟 𝜕𝜃 𝜕𝑧 𝑟
Salient features of these systems, namely strain discontinuity, contact
constraints, compatibility of stresses, and inhomogeneities, are incor- 𝛼
1 𝜕 ( 𝛼 ) 1 𝜕𝜎𝜃𝑧 𝜕𝜎𝑧𝑧
𝛼
porated through introducing a domain-decomposition multi-network 𝑟𝜎𝑟𝑧 + + + 𝑓𝑧𝛼 = 0 ,
model. The domain decomposition strategy is previously employed 𝑟 𝜕𝑟 𝑟 𝜕𝜃 𝜕𝑧
for the solution of conservation laws (see conservative PINNs Jagtap (1)

2
M. Vahab et al. International Journal of Solids and Structures 277-278 (2023) 112319

where 𝜎𝑖𝑗𝛼 (𝑖, 𝑗 = 𝑟, 𝜃, 𝑧) is the Cauchy stress tensor, 𝑓𝑖𝛼 is the body-force
density vector, and P and 𝑆𝑘 (𝑘 = 1, … , 𝑁) denote the corresponding
quantities of the pile region and the soil medium, respectively. The
strain–displacement relations are given by
( 𝛼 )
𝜕𝑢𝛼 1 𝜕𝑢𝜃 𝜕𝑢𝛼
𝜀𝛼𝑟𝑟 = 𝑟 , 𝜀𝛼𝜃𝜃 = + 𝑢𝛼𝑟 , 𝜀𝛼𝑧𝑧 = 𝑧 ,
𝜕𝑟 𝑟 𝜕𝜃 𝜕𝑧

( 𝛼 ( 𝛼 ))
1 1 𝜕𝑢𝑟 𝜕 𝑢𝜃
𝜀𝛼𝑟𝜃 = +𝑟 ,
2 𝑟 𝜕𝜃 𝜕𝑟 𝑟
( 𝛼 ) (2)
1 1 𝜕𝑢𝑧
𝛼 𝜕𝑢 ( )
𝜀𝛼𝑧𝜃 = + 𝜃 , on Ω𝛼 𝛼 = 𝑃 , 𝑆𝑘
2 𝑟 𝜕𝜃 𝜕𝑧

( )
1 𝜕𝑢𝛼𝑟 𝜕𝑢𝛼𝑧
𝜀𝛼𝑟𝑧 = + , Fig. 2. Standard PINN architecture, defining the mapping 𝒖 ∶ 𝒙 ↦ 𝑢 (𝒙; 𝐖, 𝐛).
2 𝜕𝑧 𝜕𝑟
where 𝜀𝛼𝑖𝑗 (𝑖, 𝑗 = 𝑟, 𝜃, 𝑧) is the strain tensor and 𝑢𝛼𝑖 is the displacement
vector. Using the indicial notation, Eqs. (1) and (2) can be written in
3.1. Physics-Informed Neural Networks
the Cartesian coordinate system (𝑖, 𝑗 = 𝑥, 𝑦, 𝑧) as
𝛼
𝜎𝑗𝑖,𝑗 + 𝑓𝑖𝛼 = 0 , ( ) Suppose 𝑢 (𝒙; 𝐖, 𝐛) is an 𝐿-layer neural network with 𝒙 and 𝒖 being
( ) on Ω𝛼 𝛼 = 𝑃 , 𝑆𝑘 (3)
1 𝛼 the input and output vectors, and 𝐖 and 𝐛 being weights and biases of
𝜀𝛼𝑖𝑗 = 𝑢𝑖,𝑗 + 𝑢𝛼𝑗,𝑖 ,
2 the transformation, respectively. As depicted in Fig. 2, a feed-forward
The elastic constitutive relation employed for the pile–soil system is network is employed to approximate the solution variable 𝒖 for given
given by inputs 𝒖 through the following transformations:
( )
𝜎𝑖𝑗𝛼 = 𝜆𝛼 𝜀𝛼𝑘𝑘 𝛿𝑖𝑗 + 2𝜇𝛼 𝜀𝛼𝑖𝑗 , on Ω𝛼 𝛼 = 𝑃 , 𝑆𝑘 (4) 𝒖 = 𝛴 𝐿 ◦𝛴 𝐿−1 ◦ … ◦𝛴 1 (𝒙) , (10)

where 𝜆𝛼 and 𝜇𝛼 are the Lamé constants and 𝛿𝑖𝑗 is the Kronecker delta. where
The Lamé constants can be expressed in terms of the Young’s Modulus
𝛴 𝑙 (𝒛̂ 𝑙−1 ) ∶= 𝒛̂ 𝑙 = 𝜎 𝑙 (𝐖𝑙 ⋅ 𝒛̂ 𝑙−1 + 𝐛𝑙 ) , 𝑙 = 1, … , 𝐿 (11)
𝐸𝛼 and Poisson’s ratio 𝜈𝛼 as:
𝐸𝛼 𝜈 𝛼 𝐸𝛼 In the above relation, 𝑧0 ≡ 𝒙 and 𝑧𝐿 ≡ 𝒖 are the inputs and outputs
𝜆𝛼 = , 𝜇𝛼 = . (5)
(1 + 𝜈𝛼 )(1 − 2𝜈𝛼 ) 2(1 + 𝜈𝛼 ) of the model, with 𝜎 𝑙 s being the nonlinear activation functions and
◦ being the compositional construction of the network. Notably, the
These field equations are accompanied by the boundary conditions
hyperbolic-tangent is frequently employed as the activation function.
at the interfaces of pile and soil and soil layers, i.e.
( ) Consider a quasi-static partial differential operator  acting on the
𝛼 𝛽 solution variable 𝒖, as 𝒖(𝒙) = 𝑓 (𝒙). It is noteworthy that 𝒖 is the
𝜎𝑗𝑖 (𝒙) − 𝜎𝑗𝑖 (𝒙) 𝑛𝛼𝑗 = 0 ,
∀𝒙 ∈ (Γ𝛼 ∩ Γ𝛽 ), where 𝛼 ≠ 𝛽 (6) vector form of the displacement field described as 𝒖 = (𝑢𝑟 , 𝑢𝜃 , 𝑢𝑧 ) or 𝒖 =
𝑢𝛼𝑖 (𝒙) − 𝑢𝛽𝑖 (𝒙) = 0 , (𝑢𝑥 , 𝑢𝑦 , 𝑢𝑧 ) in cylindrical or Cartesian coordinate systems, respectively.
( )
in which 𝑛𝛼𝑗 is the unit outward normal vector of Γ𝛼 𝛼, 𝛽 = 𝑃 , 𝑆𝑘 . On The boundary conditions associated with the solution field can also be
the top surface, expressed versus partial differential operator  as 𝒖(𝜕𝒙) = 𝑔(𝜕𝒙), for
𝒙 ∈ R𝑑 , where 𝑑 is the spatial dimension of the problem. Suppose
𝑆
𝜎𝑗𝑖1 (𝑟 > 𝑎, 𝜃, 𝑧 = 0) 𝑛𝑗 1 = 0 ,
𝑆 ⋃
the network parameters are all collected in 𝜽 = 𝐿 𝑖 𝑖
𝑖=0 (𝐖 , 𝐛 ). In the
(7)
𝑃
𝜎𝑗𝑖 (0 ≤ 𝑟 ≤ 𝑎, 𝜃, 𝑧 = 0) 𝑛𝑃𝑗 = 𝑡𝑃𝑖 , context of PINNs, the network parameters are determined through
the minimization of a loss function constructed versus the equations
where 𝑡𝑃𝑖 is the surface traction applied to the pile head and 𝑎 is the ra- governing the problem on the domain and boundary conditions as
dius of the pile. Under vertical loading condition, 𝑡𝑃𝑖 = ∑
{ ( 2) }
𝑄∕ 𝜋𝑎 0 0 , where 𝑄 is the applied vertical load. For a single (𝒙; 𝜽) = 𝜆𝑖 𝑖 = 𝜆1 ‖𝐮 − 𝑓 ‖Ω + 𝜆2 ‖𝐮 − 𝑔‖𝜕Ω + ⋯ , (12)
pile embedded in a half-space, the regularity conditions at infinity are
with  being a loss function, and 𝜆𝑖 s being a selection of weights asso-
specified as
ciated with each loss term which are determined adoptively throughout
𝑆
√
𝜎𝑖𝑗𝑘 (𝑟, 𝜃, 𝑧) → 0 , as 𝑟2 + 𝑧2 → ∞ (8) the solution process.
The mean squared error norm is elaborated for the evaluation of the
For the case with underlying bedrock, the boundary condition at the loss function, i.e., ‖◦‖ = MSE(◦). In this fashion, the network parame-
bedrock level can be expressed as ters are determined by means of an optimization problem represented
𝑆
𝑢𝑖 𝑁 (𝑟, 𝜃, 𝑧 = 𝐻) = 0 , (9) by

where H is the depth to bedrock. 𝜽∗ = arg min (𝑿; 𝜽) , (13)

𝜽∈R𝐷

in which 𝐷 is the total number of trainable parameters, with 𝑿 ∈

3. PINN forward and inverse solvers
In this section, we briefly review the construction and training
process of Physics-Informed Neural Network (PINN) solvers. Next, we
implement the PINNs for the analysis of piles in homogeneous and
layered formations.
R𝑛×𝑑 being the set of 𝑛 collocation points used for the optimization
of the loss function. In this study, the construction and training of
PINNs is performed by taking advantage of the open-source python
API SciANN (Haghighat and Juanes, 2021), which is implemented on
reputed deep-learning packages TensorFlow (Abadi et al., 2016) and
Keras (Chollet et al., 2015).

3
M. Vahab et al. International Journal of Solids and Structures 277-278 (2023) 112319

3.2. Domain-decomposed PINNs for layered soils

The primary partial differential equation that governs the defor-

mation of pile–soil system is the equilibrium equation described in
Section 2. Provided that the material properties of the pile as well as
each layer of the surrounding soil could be significantly different, we
are dealing with inherent discontinuities within the strain field. Indeed,
the variations in material stiffness, that could differ by several orders
of magnitude, is responsible for the discontinuity in the derivatives of
the displacement field. As such, a domain-decomposed PINN solver is
a reasonable choice (Kharazmi et al., 2021; Niaki et al., 2021; Jagtap
et al., 2020; Jagtap and Karniadakis, 2021). We, therefore, introduce
a series of domain factors to distinguish one material domain from
another as
⎧1 , 𝒙 ∈ Ω𝛼
̄ ⎪
Π𝛼 (𝒙) = ⎨ 0 , 𝒙 ∈ Γ𝛼 (14)
⎪−1 , 𝒙 ∉ (Ω𝛼 ∪ Γ𝛼 )
⎩
in which 𝛼 = 𝑃 , 𝑆 for homogeneous domains, and 𝛼 = 𝑃 , 𝑆𝑘 (𝑘 = Fig. 3. The cylindrical pile embedded in homogeneous soil; problem definition and
boundary conditions.
1, … , 𝑁) in case of layered formations. In line with the definition of do-
main factors, it is required to define a distinct neural network for each
domain involved. This facilitates the presence of a weak discontinuity
in the displacement field across the material interfaces. Hence, the governing equations, the equilibrium equations are multiplied by 𝑟.
displacement field can be represented over the entire solution domain Thus, the simplified form of the equilibrium equations is expressed as
as 𝜕 ( 𝛼) 𝜕𝜎 𝛼
𝑟𝜎𝑟𝑟 + 𝑟 𝑟𝑧 − 𝜎𝜃𝜃
𝛼 𝛼 𝛼
≡ 𝑟𝑟 𝐮 =0 ,
∑
𝑁 𝜕𝑟 𝜕𝑧
𝑆
𝑢𝑖 ≃ 𝑢𝑃 (𝒙).Π𝑃 (𝒙) + 𝑢𝑖 𝑘 (𝒙).Π𝑆𝑘 (𝒙) , (15) (17)
𝑖
𝜕 ( 𝛼) 𝜕𝜎 𝛼
𝑘=1

̄ 𝛼 + 1)∕2. 𝑟𝜎𝑟𝑧 + 𝑟 𝑧𝑧 ≡ 𝑧𝑧

𝛼 𝛼
𝐮 =0 .
where Π𝛼 = (Π 𝜕𝑟 𝜕𝑧
As mentioned in Section 2, in order to avoid non-physical over- The strain field is described by Eq. (2), except that all terms involv-
lap/separation of the displacement fields associated with different so- ing 𝑢𝜃 or derivatives with respect to 𝜃 are vanished. In addition, the
lution domains, displacement constraints need to be imposed along singularities due to the presence of the term 1∕𝑟 at the origin (i.e., 𝑟 = 0)
the material interfaces. For the sake of simplicity, it is also stipulated are relieved through the addition of an small amount 𝜖, ̄ as 1∕𝑟 ≃ 1∕(𝑟+𝜖) ̄
that no-slip may occur along any material interface. As such, the (here, 𝜖̄ = 0.001). Therefore, it follows that
compatibility constraints corresponding to the displacement field across 𝜕𝑢𝛼𝑟 𝑢𝛼𝑟 𝜕𝑢𝛼𝑧
the whole domain can be expressed by 𝜀𝛼𝑟𝑟 = , 𝜀𝛼𝜃𝜃 = , 𝜀𝛼𝑧𝑧 = ,
𝜕𝑟 (𝑟 + 𝜖)
̄ 𝜕𝑧
⎧ 𝑁𝑢𝛼 (𝒙) − 𝑁𝑢𝛽𝑖 (𝒙)
=0 , (18)
⎪ 𝑖 ( )
⎨ (16) 1 𝜕𝑢𝛼𝑟 𝜕𝑢𝛼𝑧
⎪ ∀𝒙 ∈ (Γ ∩ Γ ) , where 𝛼 ≠ 𝛽 𝜀𝛼𝑟𝑧 = + ,
⎩ 𝛼 𝛽 2 𝜕𝑧 𝜕𝑟

where 𝑖 = (𝑟, 𝜃, 𝑧) or (𝑥, 𝑦, 𝑧) for the Cylindrical and Cartesian coordi-
nates, respectively.
4. PINNs solution and parametric study
In this section, application of the proposed framework is demon-
strated for the solution and parametric study of piles in both homoge-
neous and layered formations. For this sake, the governing equations
presented in Section 2 are employed for the solution of pile–soil systems
under axisymmetric and plane strain conditions. The performance of
the framework for parameter identification in layered soils is explored
in the final example.
4.1. Forward solution of cylindrical piles in homogeneous soils
In this example, the PINNs solution of an axisymmetric cylindrical
pile under vertical loading is investigated. The general form of the
PDEs governing the deformation of the pile–soil system is given by the
Equilibrium Eqs. (1). Considering the axisymmetricity of the problem,
all derivatives with respect to 𝜃 and the terms related to shear stresses
exerted in direction 𝜃 are vanished. To avoid the singularity of the
where 𝛼 = 𝑃 , 𝑆𝑘 . The above equations are accompanied by the
constitutive relation given by Eq. (4).
The compatibility constraint required for the interface formed at the
intersection of pile and soil is expressed by
⎧ 𝑢𝑃𝑟 (𝒙) − 𝑢𝑆𝑟 (𝒙) = 0 ,
⎪
⎪ 𝑃 𝑆
⎨ 𝑢𝑧 (𝒙) − 𝑢𝑧 (𝒙) = 0 , (19)
⎪
⎪
⎩ ∀𝒙 ∈ (Γ𝑃 ∩ Γ𝑆 )
As depicted in Fig. 3, suppose a cylindrical pile with the slender-
ness ratio of 𝓁0 ∕𝑑0 = 5, which is subject to the vertical loading of
𝑄 = 100 kN. The domain consists of a homogeneous soil layer that
is extended for the normalized radius of 𝑟𝑇 ∕𝑑0 = 10 and length of
𝓁𝑇 ∕𝓁0 = 2. The material properties for the soil are assumed as: Young’s
Modulus of Elasticity, 𝐸𝑆 = 100 MPa; Poisson’s ratio, 𝜈𝑆 = 0.25. The
material properties of the pile are: Young’s Modulus of Elasticity, 𝐸𝑃 =
2.5, 5, 10 GPa; Poisson’s ratio, 𝜈𝑃 = 0.25. In this fashion, the problem is
studied for the stiffness ratios 𝜂 = 𝐸𝑃 ∕𝐸𝑆 = 25, 50 100.
In order to perform the PINNs solution, as explained in Section 3,
we need to use multiple neural networks proportional to the number of
materials existing throughout the entire domain. As such, two distinct
neural networks are introduced as per component of the displacement

4
M. Vahab et al. International Journal of Solids and Structures 277-278 (2023) 112319

Fig. 4. The network training history for the cylindrical pile in homogeneous domains for different values of the stiffness ratio 𝜂 = 𝐸𝑃 ∕𝐸𝑆 .

Fig. 5. Contours of the normalized vertical displacement 𝑢𝑧 ∕𝓁𝑇 (10−2 ) for the cylindrical pile in homogeneous domains for different values of the stiffness ratio 𝜂 = 𝐸𝑃 ∕𝐸𝑆 .

field as where  𝛼 , 𝛼 and 𝑔 𝛼 are differential operators associated with the

equilibrium equations, boundary conditions, and preassigned boundary
𝑢𝑃𝑟 ≃ 𝑢𝑃 (𝑟, 𝑧) , 𝑢𝑃𝑧 ≃ 𝑢𝑃 (𝑟, 𝑧) ,
𝑟 𝑧
(20) values of the problem, respectively, in the directions 𝑟 and 𝑧.
𝑢𝑆𝑟 ≃ 𝑢𝑆 (𝑟, 𝑧) , 𝑢𝑆𝑧 ≃ 𝑢𝑆 (𝑟, 𝑧) . The training is performed by using 3000 sampling points over each
𝑟 𝑧
of domains Ω𝑃 and Ω𝑆 (i.e., 6000 in total), whereas at least 50% of
4 hidden layers with 20 neurons in each layer are considered in all the sampling points are clustered over the boundaries Γ𝑃 and Γ𝑆 . The
the cases. Hyperbolic-tangent is also used as the activation function. use of non-uniform sampling grid ensures that the complex boundary
Our selection of neural network architecture is underpinned by our conditions introduced along the pile–soil interface as well as the ex-
extensive prior experience in working with the equilibrium equation ternal boundaries are properly satisfied. Furthermore, Neural Tangent
PDE (see Haghighat et al., 2021b). The physics-informed loss terms of Kernel (NTK) adaptive weighting is employed to determine 𝜆s, which
the total cost function are expressed as guarantees all loss terms are calibrated proportionally throughout the
training process (Wang et al., 2022). Here, the Adam optimization
𝑇 = Ω + ΓB.C. + ΓCont , scheme is adopted for the training with the learning rate of 0.003.
‖ 𝑃 𝑃‖ ‖ 𝑆 𝑆‖ In Fig. 4, the evolution of normalized loss versus epochs and time
Ω = 𝜆1 ‖𝑟𝑟 𝐮 ‖ + 𝜆2 ‖𝑟𝑟 𝐮 ‖
‖ ‖on Ω𝑃 ‖ ‖on Ω𝑆 is illustrated. Evidently, in all cases, the designated architecture con-
‖ 𝑃 𝑃‖ ‖ 𝑆 𝑆‖ verges rapidly to the relative error of 10−6 within less than 1000 epochs.
+ 𝜆3 ‖𝑧𝑧 𝐮 ‖ + 𝜆4 ‖𝑧𝑧 𝐮 ‖ ,
‖ ‖on Ω𝑃 ‖ ‖on Ω𝑆
Meanwhile, the training process is accomplished relatively fast (total
‖ 𝑃‖ ‖ 𝑆‖
ΓB.C. = 𝜆5 ‖𝑃𝑟𝑟 𝐮𝑃 − 𝑔𝑟𝑟 ‖ + 𝜆6 ‖𝑆𝑟𝑟 𝐮𝑆 − 𝑔𝑟𝑟 ‖ (21) duration ≈ 3000 s). Contours of the normalized vertical displacement
‖ ‖on Γ𝑃 ⧵Γ𝑆 ‖ ‖on Γ𝑆 ⧵Γ𝑃
‖ 𝑃 𝑃 𝑃 ‖ ‖ 𝑆 𝑆 𝑆‖ field 𝑢𝑧 ∕𝓁𝑇 is shown for all the cases in Fig. 5. A reference solution
+ 𝜆7 ‖𝑧𝑧 𝐮 − 𝑔𝑧𝑧 ‖ + 𝜆8 ‖𝑧𝑧 𝐮 − 𝑔𝑧𝑧 ‖ ,
‖ ‖on Γ𝑃 ⧵Γ𝑆 ‖ ‖on Γ𝑆 ⧵Γ𝑃 obtained using the FEM software COMSOL Multiphysics (Multiphysics,
‖ ‖ ‖ ‖ 1998) is also presented for the sake of comparison. Additionally, in
ΓCont = 𝜆9 ‖𝐮𝑃 − 𝐮𝑆 ‖ + 𝜆10 ‖𝐭 𝑃 − 𝐭 𝑆 ‖ ,
‖ ‖on Γ𝑃 ∩Γ𝑆 ‖ ‖on Γ𝑃 ∩Γ𝑆 Fig. 6 the profile of normalized vertical displacement 𝑢𝑧 ∕𝓁𝑇 along

5
M. Vahab et al. International Journal of Solids and Structures 277-278 (2023) 112319

Table 1
The computational complexity  of the proposed PINNs solution and the traditional FEM analysis of the cylindrical pile in homogeneous
domains.
DOFs/samplings PINNs (4 × 20) Iterative solvers (IS)a IS ∕PINN Direct solvers (DS)b DS ∕PINN
1000 3.2E+5 1.0E+7 31.2 5.0E+8 1562.5
6000c 1.9E+6 3.6E+8 187.5 1.1E+11 56250
10,000 3.2E+6 1.0E+9 312.5 5.0E+11 156250
a
IS ≈ 𝛼𝑛2 , with 𝛼 = 10.
b
DS ≈ 𝛽𝑛3 , with 𝛽 = 1.
c
Present solution.

Fig. 6. The profile of normalized vertical displacement 𝑢𝑧 ∕𝓁𝑇 along the center line of cylindrical piles in homogeneous domains for different values of the stiffness ratio 𝜂 = 𝐸𝑃 ∕𝐸𝑆 .
The solid lines represent the PINN solutions, while the marks signify the COMSOL solutions.

the center line of the domain is illustrated. An excellent agreement

is observed between the PINNs solution and the results of the FEM
analysis. This demonstrates the validity of the proposed framework for
the analysis of pile–soil systems under the axisymmetric condition. At
the end of this example, we have conducted a time complexity analysis
to compare the performance of the proposed PINNs solution with
respect to the traditional finite element method. As shown in Table 1,
once the training is completed, the PINNs solution is significantly
faster than what is achievable with the computational mechanics. This
demonstrates that PINNs have the potential to significantly reduce
the computational complexity compared to traditional FEM analysis,
especially as the scale of the problem increases. This further highlights
the potential of using deep learning-based methods in the real-time
analysis of complex solid mechanics problems.

4.2. Forward solution of piles in homogeneous soils under plane strain Fig. 7. The sheet-pile wall in layered formation; problem definition and boundary
conditions.
condition

This example investigates the PINN solution of a sheet-pile wall where 𝛼 = 𝑃 , 𝑆𝑘 . Here, the compatibility constraint for the material
subject to vertical loading in homogeneous soils. The response of the interface of pile–soil system is expressed as
wall is governed by the equilibrium Eq. (3) in Cartesian coordinate sys-
tems. However, as sheet-piles are relatively long structural members in ⎧ 𝑢𝑃𝑥 (𝒙) − 𝑢𝑆𝑥 (𝒙) = 0 ,
⎪
nature, simplified plane strain description of the equilibrium equation ⎪ 𝑃 𝑆
⎨ 𝑢𝑧 (𝒙) − 𝑢𝑧 (𝒙) = 0 , (24)
is typically applied for the analysis of their response in soil medium. ⎪
In this respect, all derivatives with respect to 𝑦 are vanished. In the ⎪
⎩ ∀𝒙 ∈ (Γ𝑃 ∩ Γ𝑆 )
absence of body forces, the equilibrium equation under the plane-strain
condition can be described as Consider a sheet-pile wall with the slenderness ratio of 𝓁0 ∕𝑑0 = 5
𝛼
𝜕𝜎𝑥𝑥 𝛼
𝜕𝜎𝑧𝑥 subject to a vertical line load of 𝑄 = 10, 000 kN/m, as shown in Fig. 7. In
𝛼 𝛼
+ ≡ 𝑥𝑥 𝐮 =0 , this example, it is assumed that the surrounding soil is homogeneous
𝜕𝑥 𝜕𝑧 (22)
𝛼
𝜕𝜎𝑥𝑧 𝛼
𝜕𝜎𝑧𝑧 with the depth ratio of 𝓁1 ∕𝓁0 = 2. The stiffness ratios 𝜂 = 𝐸𝑃 ∕𝐸𝑆 =
𝛼 𝛼
+ ≡ 𝑧𝑧 𝐮 =0 . 10, 25, 50 with 𝐸𝑃 = 5 GPa and Poisson’s ratio 𝜈𝑆 = 𝜈𝑃 = 0.25
𝜕𝑥 𝜕𝑧
are considered in this problem. The PINNs solution of this problem is
In the plane-strain regime, the strain terms manifest in Eq. (3) are
performed by using the below set of neural networks
further simplified as
( 𝛼 ) 𝑢𝑃𝑥 ≃ 𝑢𝑃 (𝑥, 𝑧) , 𝑢𝑃𝑧 ≃ 𝑢𝑃 (𝑥, 𝑧) ,
𝜕𝑢𝛼 𝜕𝑢𝛼 1 𝜕𝑢𝑥 𝜕𝑢𝛼 𝑥 𝑧
(25)
𝜀𝛼𝑥𝑥 = 𝑥 , 𝜀𝛼𝑧𝑧 = 𝑧 , 𝜀𝛼𝑥𝑧 = 𝜀𝛼𝑧𝑥 = + 𝑧 , (23) 𝑢𝑆𝑥 ≃ 𝑢𝑆 (𝑥, 𝑧) , 𝑢𝑆𝑧 ≃ 𝑢𝑆 (𝑥, 𝑧) .
𝜕𝑥 𝜕𝑧 2 𝜕𝑧 𝜕𝑥 𝑥 𝑧

6
M. Vahab et al.
Fig. 8. The network training history for the sheet-pile wall in homogeneous soils for different values of the stiffness ratio 𝜂 = 𝐸𝑃 ∕𝐸𝑆 .
Fig. 9. Contours of the normalized vertical displacement 𝑢𝑧 ∕𝓁1 (10−2 ) for the sheet-pile wall in homogeneous domains for different values of the stiffness ratio 𝜂 = 𝐸𝑃 ∕𝐸𝑆 (forward
solution).
In resemblance to the previous example, the architecture of all neural
networks consists of 4 hidden layers with 20 neurons each, where
hyperbolic-tangent is used as the activation function. The loss terms
of the total cost function in here are defined identically to the previous
example (see Eq. (21)), except that the indices 𝑟 are now replaced by
𝑥. The training is performed by means of 6000 sampling points, with
3000 points assumed for each of Ω𝑃 and Ω𝑆 domains. NTK adaptive
weighting is applied for the training in conjunction with a learning rate
of 0.003.
The training history of the normalized loss versus epochs and time
is reported in Fig. 8. As can be seen, the loss function has immediately
reached below the relative error norm of 10−5 within 500 epochs. The
improved performance in terms of convergence rate in comparison to
the previous example is attributed to the increased simplicity of the
governing equations in the Cartesian system of coordinates. Finally, in
Fig. 9, contours of the normalized vertical displacement field 𝑢𝑧 ∕𝓁1 is
presented for all the cases considered based on the PINNs solution and
a reference FEM using COMSOL Muiltiphysics. The excellent agreement
between the PINNs results and the reference solution indicates the
robustness of the extended framework in the study of pile–soil systems
under the plane strain condition.
4.3. Identification of material parameters by inverse analysis
In this example, we evaluate the performance of PINNs for the
identification of model parameters involved in pile–soil systems. For
7
International Journal of Solids and Structures 277-278 (2023) 112319
this sake, the interaction of a single pile with the surrounding homo-
geneous soil medium under the plane strain condition is considered
first. The material properties considered in here are: Young’s Modulus
of Elasticity, 𝐸𝑃 = 5 GPa, 𝐸𝑆 = 0.5, 0.1, 0.02 GPa; Poisson’s ratio,
𝜈𝑆 = 𝜈𝑃 = 0.25, stiffness ratios 𝜂 = 𝐸𝑃 ∕𝐸𝑆 = 10, 50, 250. The
input data-set is due to the longitudinal strain (or equivalently stress)
profile along the pile, which is obtained by means of optical fiber strain-
sensing technique in practice (e.g., see Mohamad et al., 2011, 2012).
In lieu of field data, a synthetic data-set is generated through the high-
fidelity FEM solution of the same problem using COMSOL software. To
emulate the limitations encountered in practice, merely a 1D profile
along the center line of the pile is extracted from the FEM solution.
This input data, in turn, is employed for the inversion of the Young’s
Modulus of the surrounding soil (i.e., 𝐸𝑆 ). Sampling grid involves 3000
sampling points over each solution domain Ω𝑃 and Ω𝑆 , in conjunction
with an input data consisting of 2000 points (8000 points in total).
It is noteworthy that a similar study is not feasible by the use of
conventional deep learning considering the sparsity of the input data in
this problem. Here, we demonstrate the versatility of PINNs in handling
such study with extremely limited data-set.
The same governing equations presented in the previous example
are applied in here for the inverse analysis of piles in homogeneous
formations (i.e., Eqs. (22)–(24)). However, the loss function is now up-
dated by the inclusion of an extra term corresponding to the available
M. Vahab et al. International Journal of Solids and Structures 277-278 (2023) 112319

Fig. 10. The network training history for the inverse analysis of soil–pile interaction in homogeneous domains for different values of the stiffness ratio 𝜂 = 𝐸𝑃 ∕𝐸𝑆 .

Fig. 11. Contours of the normalized vertical displacement 𝑢𝑧 ∕𝓁𝑇 (10−2 ) for the sheet-pile wall in homogeneous domains for different values of the stiffness ratio 𝜂 = 𝐸𝑃 ∕𝐸𝑆 (inverse
analysis).

input data as Table 2

Inversion of soil Young’s Modulus for the pile embedded in homogeneous
𝑇 = Ω + ΓB.C. + ΓCont + Ω 𝑃
, domains (unit: GPa).
𝜎𝑧𝑧
‖ 𝑃 (26) pre
𝐸𝑆exact
𝑃 ‖
Analysis 𝐸𝑃 𝐸𝑆
Ω = 𝜆11 ‖ 𝜎𝑧𝑧 − 𝜎̄ 𝑧𝑧 ‖ ,
𝑃
𝜎𝑧𝑧 ‖ ‖on Ω𝑃 Analysis 1. 5.0 0.0213 0.02
Analysis 2. 5.0 0.105 0.10
in which 𝜎̄ 𝑧𝑃 is the input data-set. Analysis 3. 5.0 0.525 0.5
In Fig. 10, the training history of the normalized loss versus epochs
and time is presented for the identification study. As can be seen, the
relative error norm has reached below 10−4 within 1000 epochs. This
extended by inclusion of both soils layers (i.e., 𝛼 = 𝑆1 , 𝑆2 ). In particular,
indicates a substantial reduction in convergence rate in comparison to
the contact constraints (i.e., Eq. (24)) need to be imposed across the
the forward study, which is expected in inverse solutions. In Fig. 11,
interface of the pile with each soil layer as well as between the soil
contours of the normalized vertical displacement field 𝑢𝑧 ∕𝓁𝑇 are pre-
layers itself (see Eq. (16)). Such derivations are straightforward task,
sented for the backward solution and compared to the reference FEM
solution using COMSOL. Evidently, the PINNs results are in excellent which is not presented here for the sake of brevity.
agreement with the reference solution. In Table 2 the exact values of the PINNs are employed to extract the Young’s moduli of both soils
Young’s modulus of the soil inverted in each case are reported. Notably, (i.e., 𝐸𝑆1 and 𝐸𝑆2 ). This inversion is again carried out by using a 1D
the deduced values in all cases lie within 5% variation from the exact data-set involving the strain (stress) profile across the center line of the
values. pile. Six neural networks are elaborated for the PINNs solution in this
In the last example, the application of PINNs to identify the ma- example as
terial parameters of layered soils based on the inverse analysis of the
𝑢𝑃𝑥 ≃ 𝑢𝑃 (𝑥, 𝑧) , 𝑢𝑃𝑧 ≃ 𝑢𝑃 (𝑥, 𝑧) ,
axial strain profile measured/obtained along the pile is demonstrated. 𝑥 𝑧
𝑆 𝑆 𝑆 𝑆
Consider a two-layered soil with the thickness ratios of 𝓁1 ∕𝓁𝑡 = 0.25 𝑢𝑥 1 ≃ 𝑢𝑥1 (𝑥, 𝑧) , 𝑢𝑧 1 ≃ 𝑢𝑧1 (𝑥, 𝑧) , (27)
and 𝓁2 ∕𝓁𝑡 = 0.75. The material properties of the domain are: Young’s 𝑆 𝑆 𝑆 𝑆
𝑢𝑥 2 ≃ 𝑢𝑥2 (𝑥, 𝑧) , 𝑢𝑧 2 ≃ 𝑢𝑧2 (𝑥, 𝑧) ,
Moduli of Elasticity, 𝐸𝑃 = 5 GPa, 𝐸𝑆1 = 0.1 GPa, 𝐸𝑆2 = 0.02 GPa;
Poisson’s ratio, 𝜈𝑃 = 𝜈𝑆1 = 𝜈𝑆2 = 0.25. Here, the equations governing which consist of the same architecture as the previous case. The loss
the response of homogeneous soils (i.e., Eqs. (22)–(24)) need to be function is extended to incorporate both layers of the soil in conjunction

8
M. Vahab et al. International Journal of Solids and Structures 277-278 (2023) 112319

Table 3
Inversion of soil Young’s Modulus for the pile embedded in layered formation (unit:
GPa).
pre pre
Analysis 𝐸𝑃 𝐸𝑆 𝐸𝑆excat 𝐸𝑆 𝐸𝑆excat
1 1 2 2

5.0 0.095 0.1 0.023 0.02

the training history of the normalized loss is presented. As can be

seen, the convergence rate versus epochs has been relatively fast. Still,
given the number of losses has expanded dramatically, the training is
conducted significantly slower with respect to the case of inversion in
homogeneous soils. Contours of the normalized vertical displacement
field 𝑢𝑧 ∕𝓁𝑇 are depicted in Fig. 13 for both PINNs and reference
solution by COMSOL. Evidently, excellent agreement is observed be-
tween the PINNs results and FE simulation. Finally, in Table 3 the
inverted Young’s modulus for each soil layer is reported. The inverted
values are in very good agreement with the precise amounts. This
further demonstrates the excellent performance of PINNs in complex
parametric studies pertaining to extremely limited input data.
Fig. 12. Training data history for the pile in layered soil.
5. Conclusions

A physics-informed deep learning framework is presented for the

with the additional contact constraints as
analysis of pile–soil interaction under vertical loading. In the frame-
𝑇 = Ω + ΓB.C. + ΓCont + Ω 𝑃
, work, a domain-decomposition multi-network model is introduced to
𝜎𝑧𝑧
‖ 𝑃 𝑃‖ ‖ 𝑆 ‖ ‖ 𝑆 ‖ deal with the sharp discontinuities in the strain field at the interfaces of
Ω = 𝜆1 ‖𝑥𝑥 𝐮 ‖ + 𝜆2 ‖𝑥𝑥1 𝐮𝑆1 ‖ + 𝜆3 ‖𝑥𝑥2 𝐮𝑆2 ‖
‖ ‖on Ω𝑃 ‖ ‖on Ω𝑆1 ‖ ‖on Ω𝑆2 pile–soil regions and soil layers. The framework is trained by minimiz-
‖ 𝑃 𝑃‖ ‖ 𝑆1 𝑆1 ‖ ‖ 𝑆2 𝑆2 ‖ ing the loss function defined in terms of the equilibrium equations and
+ 𝜆4 ‖𝑧𝑧 𝐮 ‖ + 𝜆5 ‖𝑧𝑧 𝐮 ‖ + 𝜆6 ‖𝑧𝑧 𝐮 ‖ ,
‖ ‖on Ω𝑃 ‖ ‖on Ω𝑆1 ‖ ‖on Ω𝑆2 the boundary conditions governing the pile–soil interaction problem.
‖ 𝑃 ‖ Several examples are provided to demonstrate the performance of the
ΓB.C. = 𝜆7 ‖𝑃𝑥𝑥 𝐮𝑃 − 𝑔𝑥𝑥 ‖
‖ ‖on Γ𝑃 ⧵(Γ𝑆1 ∪Γ𝑆2 ) framework in the analysis of single piles embedded in homogeneous
‖ 𝑆1 𝑆1 𝑆 ‖ and layered soils under axisymmetric and plane strain conditions.
+ 𝜆8 ‖𝑥𝑥 𝐮 − 𝑔𝑥𝑥1 ‖
‖ ‖onΓ𝑆1 ⧵(Γ𝑃 ∪Γ𝑆2 ) Essential features of the framework are validated by comparing the
‖ 𝑆 𝑆 ‖ PINN results with the results obtained from the finite element analysis.
+ 𝜆9 ‖𝑥𝑥2 𝐮𝑆2 − 𝑔𝑥𝑥2 ‖
‖ ‖onΓ𝑆2 ⧵(Γ𝑃 ∪Γ𝑆1 ) Good agreement is observed between the PINN and FEM results in
‖ 𝑃 ‖
+ 𝜆10 ‖𝑃𝑧𝑧 𝐮𝑃 − 𝑔𝑧𝑧 ‖ all the cases considered. The application of the model for the inverse
‖ ‖on Γ𝑃 ⧵(Γ𝑆1 ∪Γ𝑆2 )
analysis and parameter identification of pile–soil interaction is also
‖ 𝑆1 𝑆1 𝑆1 ‖ (28)
+ 𝜆11 ‖𝑧𝑧 𝐮 − 𝑔𝑧𝑧 ‖ presented. In the examples provided, the localized data acquired along
‖ ‖on Γ𝑆1 ⧵(Γ𝑃 ∪Γ𝑆2 )
the pile length – possibly obtained via fiber optic strain sensing – is used
‖ 𝑆 𝑆 ‖
+ 𝜆12 ‖𝑧𝑧2 𝐮𝑆2 − 𝑔𝑧𝑧2 ‖ , for the inversion of soil parameters in both homogeneous and layered
‖ ‖on Γ𝑆2 ⧵(Γ𝑃 ∪Γ𝑆1 )
formations. As expected, a substantial reduction in the convergence
‖ ‖ ‖ ‖
ΓCont = 𝜆13 ‖𝐮𝑃 − 𝐮𝑆1 ‖ + 𝜆14 ‖𝐮𝑃 − 𝐮𝑆2 ‖ rate is observed in the inverse analysis in comparison to the forward
‖ ‖on Γ𝑃 ∩Γ𝑆 ‖ ‖on Γ𝑃 ∩Γ𝑆
1 2 study. However, it is seen that the proposed PINN framework is able to
‖ ‖
+ 𝜆15 ‖𝐮𝑆1 − 𝐮𝑆2 ‖ identify the material parameters quite efficiently.
‖ ‖on Γ𝑆 ∩Γ𝑆
1 2
‖ ‖ ‖ ‖
+ 𝜆16 ‖𝐭 𝑃 − 𝐭 𝑆1 ‖ + 𝜆17 ‖𝐭 𝑃 − 𝐭 𝑆2 ‖ Declaration of competing interest
‖ ‖on Γ𝑃 ∩Γ𝑆 ‖ ‖on Γ𝑃 ∩Γ𝑆
1 2
‖ ‖
+ 𝜆18 ‖𝐭 𝑆1 − 𝐭 𝑆2 ‖ , The authors declare that they have no known competing finan-
‖ ‖on Γ𝑆 ∩Γ𝑆
1 2 cial interests or personal relationships that could have appeared to
‖ 𝑃 𝑃 ‖ influence the work reported in this paper.
Ω = 𝜆19 ‖𝜎𝑧𝑧 − 𝜎̄ 𝑧𝑧 ‖ ,
𝑃
𝜎𝑧𝑧 ‖ ‖on Ω𝑃

Sampling grid involves 3000 sampling points over Ω𝑃 , 2000 points Data availability
in each of the soil layers Ω𝑆1 and Ω𝑆2 , and 2000 points due to the
input data-set, which is summed at 9000 points in total. In Fig. 12, Data will be made available on request.

Fig. 13. Contours of the normalized vertical displacement 𝑢𝑧 ∕𝓁𝑇 (10−2 ) for the pile in layered soil.

9
M. Vahab et al.
References
Abadi, M., Barham, P., Chen, J., Chen, Z., Davis, A., Dean, J., Devin, M., Ghe-
mawat, S., Irving, G., Isard, M., et al., 2016. Tensorflow: A system for large-scale
machine learning. In: 12th {USENIX} Symposium on Operating Systems Design and
Implementation. ({OSDI} 16), pp. 265–283.
Armaghani, D.J., Raja, R.S.N.S.B., Faizi, K., Rashid, A.S.A., et al., 2017. Develop-
ing a hybrid PSO–ANN model for estimating the ultimate bearing capacity of
rock-socketed piles. Neural Comput. Appl. 28 (2), 391–405.
Azimi, S.M., Britz, D., Engstler, M., Fritz, M., Mücklich, F., 2018. Advanced steel
microstructural classification by deep learning methods. Sci. Rep. 8 (1), 1–14.
Azimi, M., Pekcan, G., 2020. Structural health monitoring using extremely compressed
data through deep learning. Comput.-Aided Civ. Infrastruct. Eng. 35 (6), 597–614.
Bao, Y., Tang, Z., Li, H., Zhang, Y., 2019. Computer vision and deep learning–based
data anomaly detection method for structural health monitoring. Struct. Health
Monit. 18 (2), 401–421.
Bar-Sinai, Y., Hoyer, S., Hickey, J., Brenner, M.P., 2019. Learning data-driven dis-
cretizations for partial differential equations. Proc. Natl. Acad. Sci. 116 (31),
15344–15349.
Baydin, A.G., Pearlmutter, B.A., Radul, A.A., Siskind, J.M., 2018. Automatic
differentiation in machine learning: a survey. J. Mach. Learn. Res. 18.
Bishop, C.M., 2006. Pattern Recognition and Machine Learning. springer.
Calvello, M., Finno, R.J., 2004. Selecting parameters to optimize in model calibration
by inverse analysis. Comput. Geotech. 31 (5), 410–424.
Chen, Y., Lu, L., Karniadakis, G.E., Dal Negro, L., 2020. Physics-informed neural
networks for inverse problems in nano-optics and metamaterials. Opt. Express 28
(8), 11618–11633.
Chollet, F., et al., 2015. Keras.
Cybenko, G., 1989. Approximation by superpositions of a sigmoidal function. Math.
Control Signals Systems 2 (4), 303–314.
Das, S.K., Basudhar, P.K., 2006. Undrained lateral load capacity of piles in clay using
artificial neural network. Comput. Geotech. 33 (8), 454–459.
De Bézenac, E., Pajot, A., Gallinari, P., 2019. Deep learning for physical processes:
Incorporating prior scientific knowledge. J. Stat. Mech. Theory Exp. 2019 (12),
124009.
Ebrahimian, B., Noorzad, A., Alsaleh, M.I., 2012. Modeling shear localization along
granular soil–structure interfaces using elasto-plastic cosserat continuum. Int. J.
Solids Struct. 49 (2), 257–278.
Emu, M., Yan, P., Choudhury, S., 2020. Latency aware VNF deployment at edge
devices for IoT services: An artificial neural network based approach. In: 2020 IEEE
International Conference on Communications Workshops. ICC Workshops, IEEE, pp.
1–6.
Gawin, D., Lefik, M., Schrefler, B., 2001. ANN approach to sorption hysteresis within a
coupled hygro-thermo-mechanical FE analysis. Internat. J. Numer. Methods Engrg.
50 (2), 299–323.
Ghasemi-Fare, O., Basu, P., 2013. A practical heat transfer model for geothermal piles.
Energy Build. 66, 470–479.
Goodfellow, I., Bengio, Y., Courville, A., Bengio, Y., 2016. Deep Learning, Vol. 1. MIT
press Cambridge.
Haghighat, E., Bekar, A.C., Madenci, E., Juanes, R., 2021a. A nonlocal physics-informed
deep learning framework using the peridynamic differential operator. Comput.
Methods Appl. Mech. Engrg. 385, 114012.
Haghighat, E., Juanes, R., 2021. Sciann: A keras/tensorflow wrapper for scientific
computations and physics-informed deep learning using artificial neural networks.
Comput. Methods Appl. Mech. Engrg. 373, 113552.
Haghighat, E., Raissi, M., Moure, A., Gomez, H., Juanes, R., 2021b. A physics-informed
deep learning framework for inversion and surrogate modeling in solid mechanics.
Comput. Methods Appl. Mech. Engrg. 379, 113741.
Han, J., Jentzen, A., Weinan, E., 2018. Solving high-dimensional partial differential
equations using deep learning. Proc. Natl. Acad. Sci. 115 (34), 8505–8510.
Jafari, A., Vahab, M., Khalili, N., 2021. Fully coupled XFEM formulation for hydraulic
fracturing simulation based on a generalized fluid leak-off model. Comput. Methods
Appl. Mech. Engrg. 373, 113447.
Jagtap, A.D., 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. pp. 2002—2041.
Jagtap, A.D., Kharazmi, E., Karniadakis, G.E., 2020. Conservative physics-informed
neural networks on discrete domains for conservation laws: Applications to forward
and inverse problems. Comput. Methods Appl. Mech. Engrg. 365, 113028.
Kabe, A.M., 1985. Stiffness matrix adjustment using mode data. AIAA J. 23 (9),
1431–1436.
Kardani, N., Zhou, A., Nazem, M., Shen, S.-L., 2020. Estimation of bearing capacity of
piles in cohesionless soil using optimised machine learning approaches. Geotech.
Geol. Eng. 38 (2), 2271–2291.
Karniadakis, G.E., Kevrekidis, I.G., Lu, L., Perdikaris, P., Wang, S., Yang, L., 2021.
Physics-informed machine learning. Nat. Rev. Phys. 3 (6), 422–440.
10
International Journal of Solids and Structures 277-278 (2023) 112319
Khaleghi, M., Haghighat, E., Vahab, M., Shahbodagh, B., Khalili, N., 2022. Fracture
characterization from noisy displacement data using artificial neural networks. Eng.
Fract. Mech. 271, 108649.
Kharazmi, E., Zhang, Z., Karniadakis, G.E., 2021. hp-vpinns: Variational physics-
informed neural networks with domain decomposition. Comput. Methods Appl.
Mech. Engrg. 374, 113547.
Lefik, M., Schrefler, B.A., 2002. Artificial neural network for parameter identifications
for an elasto-plastic model of superconducting cable under cyclic loading. Comput.
Struct. 80 (22), 1699–1713.
Li, X., Liu, Z., Cui, S., Luo, C., Li, C., Zhuang, Z., 2019. Predicting the effective
mechanical property of heterogeneous materials by image based modeling and deep
learning. Comput. Methods Appl. Mech. Engrg. 347, 735–753.
Lou, Q., Meng, X., Karniadakis, G.E., 2021. Physics-informed neural networks for
solving forward and inverse flow problems via the Boltzmann-BGK formulation.
J. Comput. Phys. 110676.
Mao, Z., Jagtap, A.D., Karniadakis, G.E., 2020. Physics-informed neural networks for
high-speed flows. Comput. Methods Appl. Mech. Engrg. 360, 112789.
Mei, J., Ma, G., Wang, Q., Wu, T., Zhou, W., 2022. Micro-and macroscopic aspects of
the intermittent behaviors of granular materials related by graph neural network.
Int. J. Solids Struct. 111763.
Moayedi, H., Mosallanezhad, M., 2017. Uplift resistance of belled and multi-belled piles
in loose sand. Measurement 109, 346–353.
Mohamad, H., Soga, K., Bennett, P.J., Mair, R.J., Lim, C.S., 2012. Monitoring twin
tunnel interaction using distributed optical fiber strain measurements. J. Geotech.
Geoenviron. Eng. 138 (8), 957–967.
Mohamad, H., Soga, K., Pellew, A., Bennett, P.J., 2011. Performance monitoring
of a secant-piled wall using distributed fiber optic strain sensing. J. Geotech.
Geoenviron. Eng. 137 (12), 1236–1243.
Momeni, E., Nazir, R., Armaghani, D.J., Maizir, H., 2015. Application of artificial neural
network for predicting shaft and tip resistances of concrete piles. Earth Sci. Res. J.
19 (1), 85–93.
Multiphysics, C., 1998. Introduction to Comsol Multiphysics® , Vol. 9. COMSOL
Multiphysics, Burlington, MA, p. 2018, Accessed Feb.
Niaki, S.A., Haghighat, E., Campbell, T., Poursartip, A., Vaziri, R., 2021. Physics-
informed neural network for modelling the thermochemical curing process of
composite-tool systems during manufacture. Comput. Methods Appl. Mech. Engrg.
384, 113959.
Owhadi, H., 2015. Bayesian numerical homogenization. Multiscale Model. Simul. 13
(3), 812–828.
Parish, E.J., Duraisamy, K., 2016. A paradigm for data-driven predictive modeling using
field inversion and machine learning. J. Comput. Phys. 305, 758–774.
Rahmani, A., Pak, A., 2012. Dynamic behavior of pile foundations under cyclic loading
in liquefiable soils. Comput. Geotech. 40, 114–126.
Raissi, M., Perdikaris, P., Karniadakis, G.E., 2019. Physics-informed neural networks:
A deep learning framework for solving forward and inverse problems involving
nonlinear partial differential equations. J. Comput. Phys. 378, 686–707.
Reichstein, M., Camps-Valls, G., Stevens, B., Jung, M., Denzler, J., Carvalhais, N., et
al., 2019. Deep learning and process understanding for data-driven earth system
science. Nature 566 (7743), 195–204.
Rudy, S.H., Brunton, S.L., Proctor, J.L., Kutz, J.N., 2017. Data-driven discovery of
partial differential equations. Sci. Adv. 3 (4), e1602614.
Sahli Costabal, F., Yang, Y., Perdikaris, P., Hurtado, D.E., Kuhl, E., 2020. Physics-
informed neural networks for cardiac activation mapping. Front. Phys. 8,
42.
Sutton, R.S., Barto, A.G., 2018. Reinforcement Learning: An Introduction. MIT Press.
Vahab, M., Haghighat, E., Khaleghi, M., Khalili, N., 2022. A physics-informed neural
network approach to solution and identification of biharmonic equations of
elasticity. J. Eng. Mech. 148 (2), 04021154.
Vardakos, S., Gutierrez, M., Xia, C., 2012. Parameter identification in numerical
modeling of tunneling using the differential evolution genetic algorithm (DEGA).
Tunn. Undergr. Space Technol. 28, 109–123.
Versteijlen, W., de Oliveira Barbosa, J., van Dalen, K., Metrikine, A., 2018. Dynamic
soil stiffness for foundation piles: Capturing 3D continuum effects in an effective,
non-local 1D model. Int. J. Solids Struct. 134, 272–282.
Walton, G., Sinha, S., 2022. Challenges associated with numerical back analysis in rock
mechanics. J. Rock Mech. Geotech. Eng.
Wang, S., Yu, X., Perdikaris, P., 2022. When and why PINNs fail to train: A neural
tangent kernel perspective. J. Comput. Phys. 449, 110768.
Zhang, W., Li, H., Li, Y., Liu, H., Chen, Y., Ding, X., 2021. Application of deep learning
algorithms in geotechnical engineering: a short critical review. Artif. Intell. Rev.
1–41.
Zhang, D., Lin, J., Peng, Q., Wang, D., Yang, T., Sorooshian, S., Liu, X., Zhuang, J.,
2018. Modeling and simulating of reservoir operation using the artificial neural
network, support vector regression, deep learning algorithm. J. Hydrol. 565,
720–736.
Zienkiewicz, O.C., Chan, A., Pastor, M., Schrefler, B., Shiomi, T., 1999. Computational
Geomechanics, Vol. 613. Citeseer.