Professional Documents
Culture Documents
1 s2.0 S0020768323002160 Main
1 s2.0 S0020768323002160 Main
1 s2.0 S0020768323002160 Main
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 2016; Kardani et al., 2020; Mei et al., 2022). Compared to traditional
computational mechanics solvers, (DL) has the advantage of offering
The Finite Element Method (FEM) is ubiquitously used for the linear time complexity and surpassing them in performance beyond
computational analysis of various geotechnical engineering problems training (Emu et al., 2020). In the analysis of piles, which is the focus of
(Zienkiewicz et al., 1999; Ghasemi-Fare and Basu, 2013; Rahmani and this study, DL has been applied for the prediction of the shaft and tip re-
Pak, 2012; Jafari et al., 2021; Versteijlen et al., 2018; Ebrahimian sistance of concrete piles by Momeni et al. (2015), for estimation of the
et al., 2012). Consistently, in many classical studies, FEM has been the uplift resistance of screw piles by Moayedi and Mosallanezhad (2017),
natural choice for the design of inversion or back-analysis algorithms and for evaluation of the lateral load bearing capacity of piles by Das
for inferring important model parameters, like mechanical properties and Basudhar (2006) and Armaghani et al. (2017). It is noteworthy that
of soils (Calvello and Finno, 2004). The state-of-the-art techniques in the enumerated studies emphasize on forward prediction of mechanical
this context mainly rely on minimization of the deviations between response of piles itself. Based on the Universal approximation theo-
the simulation responses (due to an adjustable set of model param- rem (a.k.a., Cybenko theorem, see Cybenko, 1989), neural networks
eters) with respect to field and laboratory measurements. For this
comprising of at least a single hidden layer can uniformly reconstruct
purpose, a range of direct or gradient-based optimization techniques
any function with arbitrary continuous non-linearity. Evidently, the
are extended which exploit successive numerical solutions iteratively to
composite system of soil–piles involves a strong material discontinuity,
yield optimal precision for the sought-after parameters (Kabe, 1985).
which in turn leads to a discontinuous strain field (i.e., the gradient
Nonetheless, both approaches can be prohibitively computationally
of displacement). This study aims to develop a DL framework capable
expensive in particular for large-scale problems. Further challenges
of handling the discontinuities induced by material interfaces within
encountered during numerical back analysis in engineering applications
soil–pile systems.
have been due to non-uniqueness, material model limitations, and
disparate data sources (Vardakos et al., 2012; Walton and Sinha, 2022). Early Artificial Intelligence (AI) approaches were developed in the
Deep Learning (DL) has proven to be a rigorous approach for mid-20th century to tackle the intellectually difficult and complex-to-
the forward and back analysis of geotechnical engineering problems articulate problems, which pertain to straight-forward solution algo-
(Lefik and Schrefler, 2002; Gawin et al., 2001; Parish and Duraisamy, 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
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
3
M. Vahab et al. International Journal of Solids and Structures 277-278 (2023) 112319
where 𝑖 = (𝑟, 𝜃, 𝑧) or (𝑥, 𝑦, 𝑧) for the Cylindrical and Cartesian coordi- where 𝛼 = 𝑃 , 𝑆𝑘 . The above equations are accompanied by the
nates, respectively. constitutive relation given by Eq. (4).
The compatibility constraint required for the interface formed at the
intersection of pile and soil is expressed by
4. PINNs solution and parametric study
⎧ 𝑢𝑃𝑟 (𝒙) − 𝑢𝑆𝑟 (𝒙) = 0 ,
⎪
In this section, application of the proposed framework is demon- ⎪ 𝑃 𝑆
⎨ 𝑢𝑧 (𝒙) − 𝑢𝑧 (𝒙) = 0 , (19)
strated for the solution and parametric study of piles in both homoge- ⎪
neous and layered formations. For this sake, the governing equations ⎪
⎩ ∀𝒙 ∈ (Γ𝑃 ∩ Γ𝑆 )
presented in Section 2 are employed for the solution of pile–soil systems
under axisymmetric and plane strain conditions. The performance of As depicted in Fig. 3, suppose a cylindrical pile with the slender-
the framework for parameter identification in layered soils is explored ness ratio of 𝓁0 ∕𝑑0 = 5, which is subject to the vertical loading of
in the final example. 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
4.1. Forward solution of cylindrical piles in homogeneous soils 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 =
In this example, the PINNs solution of an axisymmetric cylindrical 2.5, 5, 10 GPa; Poisson’s ratio, 𝜈𝑃 = 0.25. In this fashion, the problem is
pile under vertical loading is investigated. The general form of the studied for the stiffness ratios 𝜂 = 𝐸𝑃 ∕𝐸𝑆 = 25, 50 100.
PDEs governing the deformation of the pile–soil system is given by the In order to perform the PINNs solution, as explained in Section 3,
Equilibrium Eqs. (1). Considering the axisymmetricity of the problem, we need to use multiple neural networks proportional to the number of
all derivatives with respect to 𝜃 and the terms related to shear stresses materials existing throughout the entire domain. As such, two distinct
exerted in direction 𝜃 are vanished. To avoid the singularity of the 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 .
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.
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. International Journal of Solids and Structures 277-278 (2023) 112319
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 this sake, the interaction of a single pile with the surrounding homo-
networks consists of 4 hidden layers with 20 neurons each, where geneous soil medium under the plane strain condition is considered
hyperbolic-tangent is used as the activation function. The loss terms first. The material properties considered in here are: Young’s Modulus
of the total cost function in here are defined identically to the previous of Elasticity, 𝐸𝑃 = 5 GPa, 𝐸𝑆 = 0.5, 0.1, 0.02 GPa; Poisson’s ratio,
example (see Eq. (21)), except that the indices 𝑟 are now replaced by 𝜈𝑆 = 𝜈𝑃 = 0.25, stiffness ratios 𝜂 = 𝐸𝑃 ∕𝐸𝑆 = 10, 50, 250. The
𝑥. The training is performed by means of 6000 sampling points, with input data-set is due to the longitudinal strain (or equivalently stress)
3000 points assumed for each of Ω𝑃 and Ω𝑆 domains. NTK adaptive
profile along the pile, which is obtained by means of optical fiber strain-
weighting is applied for the training in conjunction with a learning rate
sensing technique in practice (e.g., see Mohamad et al., 2011, 2012).
of 0.003.
The training history of the normalized loss versus epochs and time In lieu of field data, a synthetic data-set is generated through the high-
is reported in Fig. 8. As can be seen, the loss function has immediately fidelity FEM solution of the same problem using COMSOL software. To
reached below the relative error norm of 10−5 within 500 epochs. The emulate the limitations encountered in practice, merely a 1D profile
improved performance in terms of convergence rate in comparison to along the center line of the pile is extracted from the FEM solution.
the previous example is attributed to the increased simplicity of the This input data, in turn, is employed for the inversion of the Young’s
governing equations in the Cartesian system of coordinates. Finally, in Modulus of the surrounding soil (i.e., 𝐸𝑆 ). Sampling grid involves 3000
Fig. 9, contours of the normalized vertical displacement field 𝑢𝑧 ∕𝓁1 is sampling points over each solution domain Ω𝑃 and Ω𝑆 , in conjunction
presented for all the cases considered based on the PINNs solution and with an input data consisting of 2000 points (8000 points in total).
a reference FEM using COMSOL Muiltiphysics. The excellent agreement It is noteworthy that a similar study is not feasible by the use of
between the PINNs results and the reference solution indicates the conventional deep learning considering the sparsity of the input data in
robustness of the extended framework in the study of pile–soil systems
this problem. Here, we demonstrate the versatility of PINNs in handling
under the plane strain condition.
such study with extremely limited data-set.
4.3. Identification of material parameters by inverse analysis The same governing equations presented in the previous example
are applied in here for the inverse analysis of piles in homogeneous
In this example, we evaluate the performance of PINNs for the formations (i.e., Eqs. (22)–(24)). However, the loss function is now up-
identification of model parameters involved in pile–soil systems. For dated by the inclusion of an extra term corresponding to the available
7
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).
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
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. International Journal of Solids and Structures 277-278 (2023) 112319
References Khaleghi, M., Haghighat, E., Vahab, M., Shahbodagh, B., Khalili, N., 2022. Fracture
characterization from noisy displacement data using artificial neural networks. Eng.
Abadi, M., Barham, P., Chen, J., Chen, Z., Davis, A., Dean, J., Devin, M., Ghe- Fract. Mech. 271, 108649.
mawat, S., Irving, G., Isard, M., et al., 2016. Tensorflow: A system for large-scale Kharazmi, E., Zhang, Z., Karniadakis, G.E., 2021. hp-vpinns: Variational physics-
machine learning. In: 12th {USENIX} Symposium on Operating Systems Design and informed neural networks with domain decomposition. Comput. Methods Appl.
Implementation. ({OSDI} 16), pp. 265–283. Mech. Engrg. 374, 113547.
Armaghani, D.J., Raja, R.S.N.S.B., Faizi, K., Rashid, A.S.A., et al., 2017. Develop- Lefik, M., Schrefler, B.A., 2002. Artificial neural network for parameter identifications
ing a hybrid PSO–ANN model for estimating the ultimate bearing capacity of for an elasto-plastic model of superconducting cable under cyclic loading. Comput.
rock-socketed piles. Neural Comput. Appl. 28 (2), 391–405. Struct. 80 (22), 1699–1713.
Azimi, S.M., Britz, D., Engstler, M., Fritz, M., Mücklich, F., 2018. Advanced steel Li, X., Liu, Z., Cui, S., Luo, C., Li, C., Zhuang, Z., 2019. Predicting the effective
microstructural classification by deep learning methods. Sci. Rep. 8 (1), 1–14. mechanical property of heterogeneous materials by image based modeling and deep
Azimi, M., Pekcan, G., 2020. Structural health monitoring using extremely compressed learning. Comput. Methods Appl. Mech. Engrg. 347, 735–753.
data through deep learning. Comput.-Aided Civ. Infrastruct. Eng. 35 (6), 597–614. Lou, Q., Meng, X., Karniadakis, G.E., 2021. Physics-informed neural networks for
Bao, Y., Tang, Z., Li, H., Zhang, Y., 2019. Computer vision and deep learning–based solving forward and inverse flow problems via the Boltzmann-BGK formulation.
data anomaly detection method for structural health monitoring. Struct. Health J. Comput. Phys. 110676.
Monit. 18 (2), 401–421. Mao, Z., Jagtap, A.D., Karniadakis, G.E., 2020. Physics-informed neural networks for
Bar-Sinai, Y., Hoyer, S., Hickey, J., Brenner, M.P., 2019. Learning data-driven dis- high-speed flows. Comput. Methods Appl. Mech. Engrg. 360, 112789.
cretizations for partial differential equations. Proc. Natl. Acad. Sci. 116 (31), Mei, J., Ma, G., Wang, Q., Wu, T., Zhou, W., 2022. Micro-and macroscopic aspects of
15344–15349. the intermittent behaviors of granular materials related by graph neural network.
Baydin, A.G., Pearlmutter, B.A., Radul, A.A., Siskind, J.M., 2018. Automatic Int. J. Solids Struct. 111763.
differentiation in machine learning: a survey. J. Mach. Learn. Res. 18. Moayedi, H., Mosallanezhad, M., 2017. Uplift resistance of belled and multi-belled piles
Bishop, C.M., 2006. Pattern Recognition and Machine Learning. springer. in loose sand. Measurement 109, 346–353.
Calvello, M., Finno, R.J., 2004. Selecting parameters to optimize in model calibration Mohamad, H., Soga, K., Bennett, P.J., Mair, R.J., Lim, C.S., 2012. Monitoring twin
by inverse analysis. Comput. Geotech. 31 (5), 410–424. tunnel interaction using distributed optical fiber strain measurements. J. Geotech.
Chen, Y., Lu, L., Karniadakis, G.E., Dal Negro, L., 2020. Physics-informed neural Geoenviron. Eng. 138 (8), 957–967.
networks for inverse problems in nano-optics and metamaterials. Opt. Express 28 Mohamad, H., Soga, K., Pellew, A., Bennett, P.J., 2011. Performance monitoring
(8), 11618–11633. of a secant-piled wall using distributed fiber optic strain sensing. J. Geotech.
Chollet, F., et al., 2015. Keras. Geoenviron. Eng. 137 (12), 1236–1243.
Cybenko, G., 1989. Approximation by superpositions of a sigmoidal function. Math. Momeni, E., Nazir, R., Armaghani, D.J., Maizir, H., 2015. Application of artificial neural
Control Signals Systems 2 (4), 303–314. network for predicting shaft and tip resistances of concrete piles. Earth Sci. Res. J.
Das, S.K., Basudhar, P.K., 2006. Undrained lateral load capacity of piles in clay using 19 (1), 85–93.
artificial neural network. Comput. Geotech. 33 (8), 454–459. Multiphysics, C., 1998. Introduction to Comsol Multiphysics® , Vol. 9. COMSOL
De Bézenac, E., Pajot, A., Gallinari, P., 2019. Deep learning for physical processes: Multiphysics, Burlington, MA, p. 2018, Accessed Feb.
Incorporating prior scientific knowledge. J. Stat. Mech. Theory Exp. 2019 (12), Niaki, S.A., Haghighat, E., Campbell, T., Poursartip, A., Vaziri, R., 2021. Physics-
124009. informed neural network for modelling the thermochemical curing process of
Ebrahimian, B., Noorzad, A., Alsaleh, M.I., 2012. Modeling shear localization along composite-tool systems during manufacture. Comput. Methods Appl. Mech. Engrg.
granular soil–structure interfaces using elasto-plastic cosserat continuum. Int. J. 384, 113959.
Solids Struct. 49 (2), 257–278. Owhadi, H., 2015. Bayesian numerical homogenization. Multiscale Model. Simul. 13
Emu, M., Yan, P., Choudhury, S., 2020. Latency aware VNF deployment at edge (3), 812–828.
devices for IoT services: An artificial neural network based approach. In: 2020 IEEE Parish, E.J., Duraisamy, K., 2016. A paradigm for data-driven predictive modeling using
International Conference on Communications Workshops. ICC Workshops, IEEE, pp. field inversion and machine learning. J. Comput. Phys. 305, 758–774.
1–6. Rahmani, A., Pak, A., 2012. Dynamic behavior of pile foundations under cyclic loading
Gawin, D., Lefik, M., Schrefler, B., 2001. ANN approach to sorption hysteresis within a in liquefiable soils. Comput. Geotech. 40, 114–126.
coupled hygro-thermo-mechanical FE analysis. Internat. J. Numer. Methods Engrg. Raissi, M., Perdikaris, P., Karniadakis, G.E., 2019. Physics-informed neural networks:
50 (2), 299–323. A deep learning framework for solving forward and inverse problems involving
Ghasemi-Fare, O., Basu, P., 2013. A practical heat transfer model for geothermal piles. nonlinear partial differential equations. J. Comput. Phys. 378, 686–707.
Energy Build. 66, 470–479. Reichstein, M., Camps-Valls, G., Stevens, B., Jung, M., Denzler, J., Carvalhais, N., et
Goodfellow, I., Bengio, Y., Courville, A., Bengio, Y., 2016. Deep Learning, Vol. 1. MIT al., 2019. Deep learning and process understanding for data-driven earth system
press Cambridge. science. Nature 566 (7743), 195–204.
Haghighat, E., Bekar, A.C., Madenci, E., Juanes, R., 2021a. A nonlocal physics-informed Rudy, S.H., Brunton, S.L., Proctor, J.L., Kutz, J.N., 2017. Data-driven discovery of
deep learning framework using the peridynamic differential operator. Comput. partial differential equations. Sci. Adv. 3 (4), e1602614.
Methods Appl. Mech. Engrg. 385, 114012. Sahli Costabal, F., Yang, Y., Perdikaris, P., Hurtado, D.E., Kuhl, E., 2020. Physics-
Haghighat, E., Juanes, R., 2021. Sciann: A keras/tensorflow wrapper for scientific informed neural networks for cardiac activation mapping. Front. Phys. 8,
computations and physics-informed deep learning using artificial neural networks. 42.
Comput. Methods Appl. Mech. Engrg. 373, 113552. Sutton, R.S., Barto, A.G., 2018. Reinforcement Learning: An Introduction. MIT Press.
Haghighat, E., Raissi, M., Moure, A., Gomez, H., Juanes, R., 2021b. A physics-informed Vahab, M., Haghighat, E., Khaleghi, M., Khalili, N., 2022. A physics-informed neural
deep learning framework for inversion and surrogate modeling in solid mechanics. network approach to solution and identification of biharmonic equations of
Comput. Methods Appl. Mech. Engrg. 379, 113741. elasticity. J. Eng. Mech. 148 (2), 04021154.
Han, J., Jentzen, A., Weinan, E., 2018. Solving high-dimensional partial differential Vardakos, S., Gutierrez, M., Xia, C., 2012. Parameter identification in numerical
equations using deep learning. Proc. Natl. Acad. Sci. 115 (34), 8505–8510. modeling of tunneling using the differential evolution genetic algorithm (DEGA).
Jafari, A., Vahab, M., Khalili, N., 2021. Fully coupled XFEM formulation for hydraulic Tunn. Undergr. Space Technol. 28, 109–123.
fracturing simulation based on a generalized fluid leak-off model. Comput. Methods Versteijlen, W., de Oliveira Barbosa, J., van Dalen, K., Metrikine, A., 2018. Dynamic
Appl. Mech. Engrg. 373, 113447. soil stiffness for foundation piles: Capturing 3D continuum effects in an effective,
Jagtap, A.D., Karniadakis, G.E., 2021. Extended physics-informed neural networks non-local 1D model. Int. J. Solids Struct. 134, 272–282.
(XPINNs): A generalized space-time domain decomposition based deep learning Walton, G., Sinha, S., 2022. Challenges associated with numerical back analysis in rock
framework for nonlinear partial differential equations. In: AAAI Spring Symposium: mechanics. J. Rock Mech. Geotech. Eng.
MLPS. pp. 2002—2041. Wang, S., Yu, X., Perdikaris, P., 2022. When and why PINNs fail to train: A neural
Jagtap, A.D., Kharazmi, E., Karniadakis, G.E., 2020. Conservative physics-informed tangent kernel perspective. J. Comput. Phys. 449, 110768.
neural networks on discrete domains for conservation laws: Applications to forward Zhang, W., Li, H., Li, Y., Liu, H., Chen, Y., Ding, X., 2021. Application of deep learning
and inverse problems. Comput. Methods Appl. Mech. Engrg. 365, 113028. algorithms in geotechnical engineering: a short critical review. Artif. Intell. Rev.
Kabe, A.M., 1985. Stiffness matrix adjustment using mode data. AIAA J. 23 (9), 1–41.
1431–1436. Zhang, D., Lin, J., Peng, Q., Wang, D., Yang, T., Sorooshian, S., Liu, X., Zhuang, J.,
Kardani, N., Zhou, A., Nazem, M., Shen, S.-L., 2020. Estimation of bearing capacity of 2018. Modeling and simulating of reservoir operation using the artificial neural
piles in cohesionless soil using optimised machine learning approaches. Geotech. network, support vector regression, deep learning algorithm. J. Hydrol. 565,
Geol. Eng. 38 (2), 2271–2291. 720–736.
Karniadakis, G.E., Kevrekidis, I.G., Lu, L., Perdikaris, P., Wang, S., Yang, L., 2021. Zienkiewicz, O.C., Chan, A., Pastor, M., Schrefler, B., Shiomi, T., 1999. Computational
Physics-informed machine learning. Nat. Rev. Phys. 3 (6), 422–440. Geomechanics, Vol. 613. Citeseer.
10