Engineering Applications of Arti cial Intelligence 148 (2025) 110347

Contents lists available at ScienceDirect

Engineering Applications of Artificial Intelligence

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

Using Physics-Informed neural networks for solving Navier-Stokes

equations in fluid dynamic complex scenarios
Tommaso Botarelli a , Marco Fanfani a , Paolo Nesi a,* , Lorenzo Pinelli b
a
DISIT Lab, Department of Information Engineering, University of Florence, Via S. Marta 3, 50139, Florence, Italy
b
T-Group, Department of Industrial Engineering, University of Florence, Via S. Marta 3, 50139, Florence, Italy

A R T I C L E I N F O A B S T R A C T

Keywords: Navier-Stokes equations used to model fluid dynamic processes are fundamental to address several real-world
Physics-Informed neural networks problems related to energy production, aerospace applications, automotive design, industrial process, etc.
Deep learning However, since in most cases they do not admit any analytical solution, numerical simulations are required in
Fluid dynamic
industrial contexts to assess fluid dynamic behaviors in specific setups. Computational Fluid Dynamics (CFD)
Partial differential equations
Navier-Stokes
methods, like those using finite volume or element approaches, are exploited to find Navier-Stokes solutions and
carry out simulations. However, such methods require expensive hardware resources, relevant computational
times, and manual efforts for the definition of dense meshes on which equations are evaluated iteratively for each
time step of the simulation. Physics-Informed Neural Networks (PINNs), which are deep neural networks where
physical laws are directly embedded into the training process, offer a promising approach for solving Navier-
Stokes equations, thus alleviating hardware and time requirements. PINNs bypass some CFD limitations by
using neural networks to produce solutions based on governing equations, thus reducing the need for large
datasets, dense meshing, and iterative estimation over time. This paper evaluates the application of PINNs in near
real-world scenarios, while considering various geometries. The study focuses on the achieved accuracy, by
comparing PINN estimates with CFD solutions obtained via OpenFOAM, and the required training times; this
includes evaluating different neural network architectures, activation functions, and numbers of sampling points.
Additionally, several training strategies such as fine-tuning, multi-resolution learning, and parametrized training
are proposed to enhance efficiency and obtain speed up. Results demonstrate that PINNs can achieve comparable
accuracy to CFD methods (with a velocity magnitude mean absolute error inferior to 10− 2 ) and significantly
reduce computational costs. Our findings demonstrated that with appropriate training techniques PINNs can be
effectively used in industrial applications requiring rapid and accurate fluid dynamic simulations, thus paving
the way for their broader adoption in practical engineering problems.

1. Introduction cars, wind turbines, etc.) to internal flow problems (aircraft engines,
piping, injectors, autoclave curing, etc.), as well as fluid mixture in­
Modelling Fluid Dynamic (FD) problems is fundamental in several vestigations (steam condensing) and reacting flows (combustion sys­
disciplines like astrophysics, chemistry, environmental sciences, civil, tems). Except for some simplified specific problems, the general form of
biomedical, and mechanical engineering. The goal is to simulate com­ N-S equations does not admit known analytical solutions and may be
plex processes to make predictions on specific scenarios, while avoiding solved by means of numerical approaches. Computational Fluid Dy­
the necessity of conducting expensive real-world experiments in the namic (CFD) methods based on finite elements/volumes/difference
design phase. FD problems are modeled by the Navier-Stokes (N-S) (Zienkiewicz et al., 2005; Szabó and Babuška, 2021) have been widely
equations, a system of partial differential equations (PDEs) describing used for solving N-S equations to estimate flow velocity and pressure
the momentum and energy balances, together with the conservation of fields given specific Boundary and Initial Conditions (BC and IC).
mass. The application of N-S equation to engineering studies can range However, such approaches have strong limitations, since they require
from external aerodynamic problems (aircraft fuselage and wing, racing considerable time for both their instantiation and execution (Pinelli

* Corresponding author.
E-mail address: [email protected] (P. Nesi).

https://doi.org/10.1016/j.engappai.2025.110347
Received 11 August 2024; Received in revised form 18 January 2025; Accepted 15 February 2025
Available online 7 March 2025
0952-1976/© 2025 The Authors. Published by Elsevier Ltd. This is an open access article under the CC BY-NC-ND license (http://creativecommons.org/licenses/by-
nc-nd/4.0/).
T. Botarelli et al. Engineering Applications of Arti cial Intelligence 148 (2025) 110347

et al., 2022). CFD methods need spatial and temporal domain dis­ low errors.
cretization (i.e., the definition of meshes or grids) on the basis of which
approximate solutions exploiting discrete models are computed. In the 2. Related works, paper aims and organization
case of very complex scenarios of the fluid domain, the mesh generation
may not be straightforward and is usually time-consuming, requiring Since the seminal work of Raissi et al. (2019), relevant attention has
manual efforts from expert engineers. Moreover, high fidelity CFD been given to PINNs, as a novel approach to solve PDE problems. In
simulations, based on Large Eddy Simulations (LES) or better Direct (Raissi et al., 2019), examples to show the applicability of PINNs
Numerical Simulations (DNS), are very computationally expensive, included the resolution of non-linear Schrödinger and Allen–Cahn
which means long execution times and expensive hardware like clusters equations in data-free setups. Some data from a spectral solver are
of several up-to-date CPUs. For these reasons, they are unfeasible for instead used together with the physical losses to address the Navier–­
many real-world applications where the availability of adequate hard­ Stokes (applied to the prototype problem of incompressible flow past a
ware is limited, and the use of highly specialized human resources is circular cylinder), Korteweg–de Vries, and Burger’s equations. In (Sun
extremely costly. Moreover, in Industry 4.0 contexts where fast simu­ et al., 2020), a physics-constrained deep learning approach in a data free
lations and quasi real time applications are required (Ferreira et al., setup was proposed to solve N-S equations, with 2D pipes, stenotic, and
2020), execution times of CFD simulation are impractical. aneurysmal flows as case studies. Even if they were inspired by
Machine Learning (ML) approaches have been proposed to address biomedical problems, the setups/scenarios were somewhat simple to
FD problems (Linse and Stengel, 1993; Faller and Schreck, 1997), and demonstrate the applicability of PINNs on real cases. Burger’s and 2D
with the increasing diffusion of Deep Learning (DL) (Tamil Thendral nonlinear diffusion–reaction equations were addressed in (Meng et al.,
et al., 2022; Cao et al., 2024a; Cao et al., 2024b) a renewed attention has 2020) and the paper provided an analysis on the usage of multiple time
been devoted to this topic in the last decade. By using physical mea­ windows to reduce total training time, however without giving clear
surements gathered during manufacturing process or numerical data indications about the relation among the number of epochs, the number
obtained through CFD simulations, ML models can be trained to predict of computation grid points, and the obtained error in field flow esti­
PDE governed processes (Parish and Duraisamy, 2016; Wang et al., mation. In (Xiang et al., 2021), solutions of N-S equations are searched
2017; Barzegar Gerdroodbary, 2020). Nevertheless, the difficulty of to solve 2D steady Kovasznay’s flow passing on a circular cylinder, and
acquiring sufficient data for training and the specificity of some pro­ as a Beltrami’s flow. In (Jin et al., 2021), the assessment of PINNs in
cesses represents a strong limitation to actual usage of such approaches. solving N-S equation incompressible flow on benchmark test cases
More interesting are the Physics-Informed Neural Networks (PINNs) (steady Kovasznay’s flow, flow passing in a circular cylinder, and un­
(Raissi et al., 2019), where the PDE solution is obtained by directly steady Beltrami’s flow) studying the impact of the number of grid
enforcing the equations into the training loss function of the neural sampling points, different architectures, optimizers, and loss weights
network, with Automatic Differentiation (Baydin et al., 2018) to be were proposed. A simple test on transfer learning was provided. How­
exploited to compute the derivatives. By minimizing the loss on a set of ever, no clear indications about training times have been reported and
points, the network is trained and converges to produce outputs that applications on different use cases closer to real scenarios were not
should be equal to the PDE model solution. Moreover, IC and BC can be provided. In (Eivazi et al., 2022), the authors evaluated PINNs for
included in the loss function as additional terms. Even if it is possible to solving Falkner–Skan boundary layer and 4 turbulent-flow cases. To
include numerical data in the PINN training process, such data needs on show how difficult PINN training can be in a data-free context, in
training are strongly reduced or better not required. Indeed, PINNs can (Chuang and Barba, 2022), some failure cases have been reported, by
potentially alleviate most limits of traditional CFD by providing more considering the Taylor-Green vortex problem and the flow passing into a
economic and fast solutions for FD problems. By exploiting consumer circular cylinder. Similar difficulties have been highlighted in (Wang
hardware as a single workstation equipped with a modern GPU, costs et al., 2023a), by providing tips and best practices about architectures
can be reduced; furthermore, since PINNs do not require complex finite and training strategies to improve PINN results in solving PDEs for a
difference computations and exploit highly optimized deep-learning number of equations and models such as: Allen-Cahn and Kur­
framework, simulations can be achieved with reduced times. amoto–Sivashinsky equations, lid-driven cavity flow, N-S flow in a torus
However, PINNs are relatively new techniques and solutions. Works and around a cylinder. The flow passing into a circular cylinder has been
on PINNs have emerged over the last few years (Lawal et al., 2022). Most addressed in (Hu and McDaniel, 2023), by considering an elliptical
of them have focused on proposing improvements on benchmark test particle in (Xu et al., 2023), in which they evaluated PINNs with limited
cases, such as: evaluating soft or hard BC and IC (Sun et al., 2020), data (Naderibeni et al., 2024). The Taylor–Green vortex test case has
training on multiple consecutive time windows (Meng et al., 2020), or been addressed in (De Ryck et al., 2024) siding PDE related losses of a
providing general best-practices (Wang et al., 2023a); there has been no PINN with data coming from CFD solvers.
attempt of actually evaluating the applicability of PINNs in real sce­ In Table 1, a summary of FD problems addressed in recent literature
narios. Most real-world applications have scenarios with specific setups, by PINN is provided. To summarize, most papers on PINNs were focused
for example, ventilation in an open pit (Flores et al., 2013), flow past on benchmark problems of PDE and FD. Moreover, several solutions
multiple rectangles (Ghamlouch et al., 2017; Fezai et al., 2020) or exploited numerical data, thus spoiling the assessment of PINNs as
diamond-shaped obstacles (Djeddi et al., 2013). Moreover, there is the substitutes of classical CFD solvers, and very few works offered detailed
need to assess the correlation between the result accuracy and the analysis on execution times with respect to estimation accuracies, by
training time, as well as the possibility to employ transfer learning or using transfer learning, fine-tuning, or other strategies.
fine-tuning techniques to reduce the time required for the training, when
passing from one scenario to another. 2.1. Paper aims and organization
In this paper, we present results about the usage of PINN in close to
real-world scenarios for the resolution of Navier-Stokes equations. In this paper, we have reported the results of the application of PINN
Complex domains with different geometries have been used to show on computing flows for cases close to real-world scenarios, while taking
performance of different neural architectures, activation functions, layer into account both estimation errors and execution times. The solutions
dimensions, network depths, while considering the impact of the num­ produced with PINN have been compared against CFD solutions carried
ber of sampling/collocation points. All results are discussed by taking out by using OpenFOAM, which is a well-known CFD solver suite
into account both achieved errors and required training times. More­ (OpenFoam soft.) exploiting the finite volume method (FVM) and vali­
over, we have proposed alternative learning strategies that can be dated by the authors on turbine and acoustic liner application
exploited to reduce training times and still obtain valid estimates with (Giaccherini et al., 2023).

2
 T. Botarelli et al.
Table 1
summary of FD N-S problems addressed with PINNs in the literature. Symbols: Y = yes, N = no, L = limited; C=CFD simulation, A = analytic solution. Most state-of-the-art related works focused on benchmark problems
with simple geometries and some of them used data in training, partially spoiling the potentiality of PINN. Moreover, few of them presented analysis for the transient phase of the F-D processes, evaluated the trade-off
between estimation accuracy and training times, carried out an assessment of neural architectures, or proposed training strategies. Differently the proposed study addresses all these analyses using as test cases complex
geometry setups closer to real-world applications.
Paper Channel Geometry Piece Geometry Equation/ Validation Time Transient Data Time vs Test Test Test Comparison of different
Problem dependency analysis free errors architectures activations collocation training strategies
points

Raissi et al. 2D Pipe Circular cylinder Navier–Stokes C Y N N N N N N None

(2019)
Sun et al. (2020) 2D Pipe No piece Poiseuille flow C N N Y Y Y Y Y Hard BC
2D stenosis
Meng et al. 1D Closed Pipe No piece Burgers C Y Y Y Y N N N Hierarchical time
(2020) equations windows
Xiang et al. 2D Rectangular No piece Kovasznay flow A Y N Y N N N N Loss balancing
(2021) periodic BC
2D Pipe Circular cylinder Navier–Stokes C Y N N
3D Box periodic BC No piece Beltrami flow A Y N Y
Jin et al. (2021) 2D Rectangular No piece Kovasznay flow A Y N Y L Y Y Y Fine-tuning
periodic BC
2D Pipe Circular cylinder Navier–Stokes C Y N Y
3D Box periodic BC No piece Beltrami flow A Y Y Y
3D Channel No piece Navier–Stokes C Y Y Y
Eivazi et al. 2D Pipe Falkner–Skan Navier–Stokes C N N N N N N N None
(2022) boundary layer
2D Pipe ZPG turbulent Navier–Stokes C N N N
boundary layer
3

2D Pipe APG turbulent Navier–Stokes C N N N

boundary layer
2D Pipe NACA4412 airfoil Navier–Stokes C N N N
2D Pipe Periodic hill Navier–Stokes C N N N
Chuang and 2D Box periodic BC No piece Taylor–Green A Y Y Y Y N N N None
Barba (2022) vortex
2D Pipe Circular cylinder Navier–Stokes C Y Y Y
Wang et al. 2D Pipe Circular cylinder Stokes flow A N N Y Y Y N N Curriculum training
(2023) 2D square cavity None Lid-driven cavity C N N Y

Engineering Applications of Arti cial Intelligence 148 (2025) 110347

flow
2D Torus None Navier–Stokes C Y N Y
2D Pipe Circular cylinder Navier–Stokes C Y N Y
Hu and 2D Pipe Circular cylinder Navier–Stokes C N N Y N N N N None
McDaniel 2D Pipe Elliptical particle Navier–Stokes C N N Y
(2023)
Xu et al. (2023) 2D Pipe Circular cylinder Navier–Stokes C Y Y N N N N N Effects of external (CFD)
data in training
Naderibeni et al. 2D Pipe Circular cylinder Navier–Stokes C Y Y N N N N N None
(2024)
De Ryck et al. 2D Box periodic BC No piece Taylor–Green A Y N N N Y N Y None
(2024) vortex
Proposed 2D Pipe, with Single rectangular Navier–Stokes C Y Y Y Y Y Y Y Fine-tuning
small inlet/outlet box Multiresolution learning
2D Pipe, with Circular cylinder Navier–Stokes C Y Y Y Parametrized learning
small inlet/outlet
2D Pipe, with 3 rectangular boxes Navier–Stokes C Y Y Y
small inlet/outlet
T. Botarelli et al. Engineering Applications of Arti cial Intelligence 148 (2025) 110347

The main contributions of the paper are. analysis. In Section 5, the proposed PINN based approach for PDE so­
lution is described. Section 6 presents the experimental results that have
• Complex setups/scenarios. As highlighted above and summarized been carried out to identify the best configuration on the basis of
in Table 1, most of the work on PINN has considered benchmark different architectures, multiple activation functions, various numbers
setups very far from real world application. On the contrary, we have of neurons per layer, and different number of layers. A rectangular box-
considered a 2D closed rectangular domain, which can be regarded like obstacle has been used as a first case. Then novel trainings have
as pipe segment or tunnel, with small apertures for both inlet and been conducted on different obstacle geometries to confirm the validity
outlet flows, with the aim to get closer to real-world applications. of the defined setup. An evaluation of the impact of using a different
Such a setup/scenario is inspired by industrial autoclaves in which number of points to assess the trade-off between errors and training
controlled flows are used to transport the heating energy required to times is reported at the end of the section. Then, Section 7 describes
cure composite materials (Fernlund et al., 2018; Clyne and Hull, possible solutions devised to reduce training times by exploiting alter­
2019): in fact, an accurate understanding of the time varying flow native training strategies, such as: (A) fine-tuning, (B) multi-resolution
field with respect to the immersed piece is fundamental to obtain learning, and (C) parametrized learning. Finally, conclusions are
correct curing processes. Therefore, the first contribution of this drawn in Section 8.
paper is the assessment of PINN performance on more complex
setups, not yet addressed in the literature. 3. Modeling PDEs and problem definition
• Evaluation of neural architectures (accuracy and training
times). During PINN training, a self-supervised approach has been This section provides a brief introduction on N-S equations and the
exploited relying solely on physical losses without any data. A mul­ description of the test cases closer to real scenarios considering different
tiple time-window technique has been used to optimize the results. pieces inside a channel with small inlet/outlet apertures (not a simple
Tests on different neural architectures, number of neurons per layer pipe). Then, details on both the CFD solver used to produce validation
(i.e., layer width), number of layers (i.e., network depth), and acti­ data, and on the PINN architectures and implemented characteristics are
vation functions have been carried out to assess PINN performance in reported in the following sections.
producing flow field results. Moreover, particular attention has been According to N-S equation, the dynamic of a fluid is governed by
devoted to training times, which is something not sufficiently mass, momentum and energy conservation. For non-reacting incom­
considered in literature. In this context, specific tests to observe the pressible flows, mass and momentum equations are decupled from en­
impact of different numbers of sampling points have been conducted ergy equation. In the tensorial notation with Einstein rule, equations
to show the achievable trade-off between times and accuracy. have the following form:
• Multiple piece/obstacle geometries. Case studies have been
∂ui
addressed always considering multiple geometries for the pieces =0 (1)
∂xi
placed inside the channel, an aspect that is different from the ge­
ometries used in state-of-the-art works: the flow field around a single ∂uj ∂ui uj 1 ∂p 1 ∂τij
rectangular piece has been initially studied, then spherical and + =− + (2)
∂t ∂xi ρ ∂xj ρ ∂xi
multiple rectangular pieces have been analyzed.
• Training strategies. Results for a set of alternative training strate­ where: ui is the ith component of the velocity, t is the time, ρ the fluid
gies aimed at the reduction of training times have been issued: (A) density, p the pressure, and τij the fluid viscous stress tensor that for
techniques of fine tuning applied to the multiple time-windows and
Newtonian flow is given by
to transfer the trained model in different setups, a critical aspect for [ ]
massive usage of the solution in differentiated productions, (B) a ∂u ∂uj
τij = μ i + (3)
newly devised multi resolution training approach, and (C) a ∂xj ∂xi
parametrized training to compute the solving model once and reuse
it in execution with a large range of difference pieces. where μ is the dynamic viscosity of the fluid.
In this paper, we have addressed the problem of solving velocity and
The paper is organized as follows and it is described in Fig. 1. In pressure fields evolution in a closed box-channel with small apertures C
Section 3, Navier-Stokes PDE equations are briefly described, with sce­ (shown in Fig. 2) with FVM and PINN approaches. The channel length
nario configuration being defined. Section 4 presents the CFD method and height go from − 0.5 to 0.5 and from − 0.25 to 0.25, respectively,
used to produce the PDE results employed as reference in the following according to a normalized shape dimension. Inlet (Γinlet ) and outlet

Fig. 1. Block diagram of experimental activity and paper organization. After defining the governing PDEs, the problem geometry, and the initial and boundary
conditions (IC and BC) in Section 3, the CDF FMV solution used as benchmark comparison and the PINN solution with the implemented loss function were described
in Sections 4 and 5, respectively. Then, Section 6 was devoted to assessing the best neural configuration by means of a sequence of tests considering different ar­
chitectures (Section 6.2), different activation functions (Section 6.3), different network widths and depths (Section 6.4), different piece geometries (Section 6.5), and
different numbers of points (Section 6.6). With the outcomes of Section 6, in Section 7 different training strategies aimed at reducing training times were evaluated:
fine-tuning (Section 7.1), multi-resolution training (Section 7.2), and parametrized learning (Section 7.3). For all the tests, accuracies in terms of mean average errors
between the velocity and pressure estimates outputted by the CFD/FVM method and by PINN, and PINN training times were reported.

4
T. Botarelli et al. Engineering Applications of Arti cial Intelligence 148 (2025) 110347

Fig. 2. Graphical representations of the geometrical setups used. The channel is defined as a closed box with small apertures for the inlet and the outlet (indicated in
red and blue, respectively). Different pieces have been placed inside the channel and are represented with green borders: (a) a single rectangular box; (b) a circular
cylinder; and (c) three small boxes. Dimensions are in meters, normalized. (Best viewed in color). (For interpretation of the references to color in this figure legend,
the reader is referred to the Web version of this article.)

(Γoutlet ), represented with red and blue lines in Fig. 2, are centered 1.51 × 10− 5 m2 /s as the fluid is air, whereas the reference velocity (U)
around the longitudinal axis with a width of 0.1 (from − 0.05 to 0.05). In can be directly derived from the Re number selected for simulations.
this container for fluid flow, three different kinds of piece have been Note that, in the FVM approach, the equation nondimensionalization is
considered, represented in green in Fig. 2: a single solid box (see Fig. 2 internally performed by the solver returning simulation results in a
(a)) with width of 0.5 (from − 0.3 to 0.2) and height of 0.2 (from − 0.1 to dimensional form. On such grounds, for a direct comparison against
0.1); a circular cylinder (see Fig. 2(b)) centered in (0,0) and with radius PINN nondimensional results, the scaling factors reported in equation
of 0.1; and three small solid boxes (see Fig. 2(c)) with each box having (7) are applied.
width 0.5 and height 0.04 spaced on the vertical axis by 0.04.
In order to solve N-S equations (1) and (2), IC and BC have been 4. Solving N-S PDE via finite volume method (FVM)
specified as follows for the whole temporal domain from 0 to 3,
addressing the transient fluid flow from the instants when the fluid 4.1. Background theory of FVM
enters from the inlet. On the inlet, where the flow enters the channel, the
flow velocity is instantaneously set to FVM basically consists of four main key-steps, discussed in the
( )
ux (x, y, t), uy (x, y, t) = (1, 0) ∀ t ∈ [0, 3] ∧ ∀ (x, y) ∈ Γinlet (4) following from (i) to (iv). (i) Spatial discretization: the fluid domain
must be discretized into a finite number of volumes (or cells) usually
where: ux (x, y, t) is the longitudinal component, and uy (x, y, t) is the organized according to structured grid-like mesh. (ii) Integral form of
vertical component of the velocity. For the outlet, where the flow exits PDEs: the N-S equations are transformed from their differential form to
the channel, the longitudinal velocity u has been set as an integral representation, more suitable to be solved by the FVM. For
example, the continuity equation for mass conservation of incompress­
ux (x, y, t) = 1 ∀ t ∈ [0, 3] ∧ ∀ (x, y) ∈ Γoutlet (5) [ ]⊤
ible flows – equation (1) – ∇u = 0, where u = ux , uy , is rewritten as
Note that no constraints on pressure have been used, leaving the ∫
solution to manifest different pressions conditions along with the fluid ∇u dV = 0 (10)
flow solution. For all the wall surfaces, both channel (Γchan ) and piece(s) Vi

(Γpiece ), no slip condition has been defined – i.e., solid walls that do not
Where: Vi is a control volume. Using the Divergence Theorem, it is
admit flow penetration:
possible to convert the volume integral to surface integral, that is
( )
ux (x, y, t), uy (x, y, t) = (0, 0) ∀ t ∈ [0, 3] ∧ ∀ (x, y) ∈ Γchan , Γpiece (6) ∫ ∮
∇u dV = un dA (11)
The PINN formulation may be used to solve the nondimensional/ Vi ∂Vi

normalized form of the incompressible N-S equations, that can be ob­

(
tained by defining a reference length (L) and velocity vector U = ux ,
[ where: ∂Vi is the surface of the control volume, n is the outward pointing
]) normal vector on this surface, and A is the area of the surface of the
uy . By using these two reference values, quantities of equations (1) and
control volume. In practice, equation (11) states that the net mass flow
(2) can be made nondimensional by setting:
in and out the control volume must be zero for incompressible flow. A
x u p U similar formulation can be obtained for the momentum and energy
x́ = ú = ṕ = t́ = t (7)
L U ρU 2 L equation (2). Applying the integral form of these equations ensures the
mass, momentum, and energy conservation in each control volume.
and thus, the N-S equations without dimensions can be rewritten as: Integral form requires also a discretization of the equations, in order to
be numerically solved in the following steps. For example, equation (11)
∂úi
=0 (8) can be discretized with the following approximation
∂x́i
∮ ∑
un dA ≈ uf nf ΔAf = 0 (12)
∂új ∂úi új ∂ṕ 1 ∂2 új
+ =− + (9) ∂Vi f
∂t́ ∂x́i ∂x́j Re ∂x́i x́i
where: the summation works on all the volume faces f, while uf is the
where: ν is the kinematic viscosity, and Re is the Reynolds number
velocity at the face, nf is the normal, and ΔAf the face area. In a 2D
defined as:
regular grid, with cell size Δx × Δy, equation (12) becomes
ρUL UL ( )
Re = = (10) ( e )
μ ν ux − uwx Δy + uny − usy Δx = 0 (13)

As to the problem under investigation, the reference length (L) is

where: e, w, n, s indicate the east, west, north, and south cell faces. (iii)
equal to the box x dimension (1 m). The kinematic viscosity is set to
Flux approximation: using numerical approximations (Hirsch, 2007),

5
T. Botarelli et al. Engineering Applications of Arti cial Intelligence 148 (2025) 110347

fluxes at the boundaries (faces) of each control volume are estimated where: the superscript 0 stands for the value at the previous time, and 00
through interpolation of neighboring cells. For example, considering the is the value two time-steps ago. Convective terms are discretized with a
( ) (
Central Differencing Scheme, we get uex = uPx + uEx /2, uwx = uW x + second order upwind scheme (Hirsch, 2007), while viscous (diffusive)
) ( ) ( )
P
P n P N s S
ux /2, uy = uy + uy /2, and uy = uy + uy /2, where P is the cell terms are linear corrected. The time step for the transient evolution has
centroid, and E, W, N, and S are the east, west, north, and south been set to 5e-4s in order to maintain a CFL (Courant-Friedrichs-Lewy)
neighboring cells. By substituting these values in (13), we finally obtain number around 1. Inlet and outlet conditions impose constant velocities
( ) on the respective patches, while no slip conditions are selected for
( E )
uNy − uSy external box and obstacle surface.
ux − uW
x
Δy + Δx = 0 (14) Each simulation has been run in a parallel CPU-based environment
2 2
and took around 6 h (8 cores of a CPU Intel Core i9-11900K at 3.5 GHz)
that is a finite difference in the spatial domain. A similar derivation to perform 30000 cells on 200s duration of simulation with time-step
can be applied to the momentum and energy equations including both (5e-4s) used to keep the algorithm stable.
spatial and temporal domains and considering also second order de­
rivatives (see for example the second order backward scheme applied in 5. Using PINN for CFD/FVM simulation/solution
OpenFoam, described in the following). (iv) Equation resolution: a
system of algebraic equations is defined gathering any contributions 5.1. Background theory
from all the cells to obtain velocity and pressure fields.
Neural networks (NNs) can be regarded as a sort of function
4.2. FVM implementation details approximators and our objective is to train a NN N taking as input
spatial and temporal coordinates and as output the respective velocity
( )
In this work, we considered three different piece geometries, namely and pressure, i.e., N (x,y,t; θ) = ux ,uy ,p , where θ indicates the network
single-box, circular cylinder, and three-boxes (see Fig. 2), and thus weights and biases. While this kind of problem could be addressed by
different spatial domain discretization are required according to such using classical NN and a huge amount of data in training, the idea
different piece geometry. The defined grid meshes for FVM solution are behind PINNs is to directly exploit the equations governing the physical
shown in Fig. 3. As to the single-box and the three-boxes, meshes are phenomena to train the network, without relying on training data. This
regular rectangular grids. On the other hand, a more complex grid has can be achieved by introducing in the NN training the so-called residual
been defined for the circular cylinder test case, in order to accurately network that starting from the NN outputs (velocity and pressure) ap­
cover the areas around the rounded shape. More in detail, the grid plies Automatic Differentiation (AD) to compute first and second order
meshes for finite element computation have been generated with the derivatives and obtain the PDE residual. Differently from the numerical
blockMesh utility of the OpenFOAM suite and consist of body-fitted un­ differentiation used in FVM (described in Section 4), AD computes de­
structured multi-block grids with wall clustering for boundary layer rivatives by breaking down functions in simple operations and exploit­
resolution (y+ ≈ 1). The mesh generation process needed a parametric ing the chain-rule. This is possible since in the computational graph
definition of the domain performed by the user, and a successive mesh defined by the NN, any intermediate values obtained during the forward
tuning to achieve better mesh quality standards in term of skewness and ( )
passing are recorded and used. Once the ux , uy , p outputs are available,
non-orthogonality. Meshes are three-dimensional with one cell along the AD is used to compute the derivatives of the PDEs with respect to the
third dimension to solve a 2D problem. About 30000 hexahedral ele­ output layer. Then, such derivatives are propagated back through other
ments have been defined for each test case. layers, until the first layer is reached and the PDE derivatives with
In order to compute the solution, addressing the fourth step of the respect to the inputs (x, y, t) are obtained. Derivatives are used to
FVM pipeline, the icoFoam solver has been selected. icoFoam solves the compute any physical residuals of the network loss function, also
incompressible laminar N-S equations by using the PISO algorithm including IC and BC losses. In addition, other terms related to external
(Pressure-Implicit with Splitting Operators), developed for non-iterative data coming from simulations or real computations could be used.
computation of unsteady flows. The code is inherently transient and can Please note that in this work external data are not used in training, thus
handle mesh non-orthogonality by means of non-orthogonality itera­ putting the solver in the worst possible case.
tions. PISO involves predictor and corrector steps to satisfy mass con­ Compared to numerical derivative computations, AD is more accu­
servation: the number of PISO corrections can be also set by the user. rate, since it should not suffer from truncation or round-off errors.
Concerning problems under investigation, 1 non-orthogonal iteration Moreover, the PINN approach compared to FVM avoids the need to
and 2 PISO corrections are selected. For time discretization a second define huge dense mesh grids and related computation at each iteration
order backward scheme has been selected, therefore the temporal de­ in space and time, thus reducing time and complexity for the experi­
rivative of equation (2) is computed as follows mental setup preparation. When all the loss components are ready, the
∂uj 3uj − 4u0j + u00 derivatives of the loss function are computed exploiting the AD, per­
j
≈ (15) forming the well-known backpropagation based on which the network
∂t 2Δt

Fig. 3. Computational meshes for FVM. Note that the BCs impact on all the four steps: they are used to define the physical boundaries of the problem domain in step
(i), they influence the definition of the integral equation in step (ii), e.g., by specifying no-slip conditions on the faces of control volumes placed near the channel or
piece border, and they are used in step (iii) to compute the fluxes in control volumes on domain borders. Both BC and IC are used in step (iv): BC are continuously
applied to correctly constrain the solution; IC are used to initialize the system variables and kick-start the iterative process.

6
T. Botarelli et al. Engineering Applications of Arti cial Intelligence 148 (2025) 110347

weights and bias optimization are carried out. minimize the PDE residual and train the NN.
To evaluate the applicability of PINN in solving the problems defined
5.2. Implementation details for N-S equation resolution in Section 3, different neural architectures with different activation
functions have been implemented. More specifically, we initially
In the case of N-S equations ((1) and (2) or (8) and (9) for their a- employed a PINN architecture based on a simple feed-forward network,
dimensional representation), the right parts of equations are moved on i.e., a multilayer perceptron (MLP). Then, the so-called Modified Fourier
the left part to have all equations equals to zero. Then, with the forward Network (MFN) (MFN Lib. and NVIDIA Corporation) has been used. It is
pass of N , output estimates of velocity and pressure for a set of points a modified version of network exploiting Fourier features (Tancik et al.,
{ }
2020). Such architecture has been devised to overcome any spectral bias
c ∈ C are obtained, i.e., ucx , ucy , pc . AD is used to compute the partial
(i.e., the difficulty in learning high-frequency functions) that typically
derivatives (e.g., ∂ux /∂x and ∂uy /∂y for the continuity equation (1)) to afflicts MLP (Rahaman et al., 2019). A feature encoding layer is used to
evaluate the PDE residuals. Briefly, the system evaluates how much the project the input feature into a higher dimensional space using
derivatives of output velocity and pressure estimates combined ac­ high-frequency functions. In the Modulus library (NVIDIA Modulus lib.)
cording to (1) and (2) (i.e., (8) and (9)) differ from zero, thus producing used in our solution, such encoding is learned during training and in the
an indication on how close the net is to correctly estimate velocity and MFN two additional transformation layers are used to further transform
pressure. Therefore, the first terms in the loss function L are the PDE the Fourier features into another feature space to improve convergence
y
residuals for the N-S momentum (L xM , L M ) and continuity (L C ) equa­ in training and final accuracies. Further details on the MLP and MFN
tions. In addition, IC and BC are also enforced into the loss function to architectures are reported in Appendix A.
constrain the net in learning the correct solution. IC and BC are evalu­ Regarding activation functions, a number of solutions have been
ated on specific points lying on the inlet (Γinlet ), outlet (Γoutlet ), channel assessed, namely: the Hyperbolic Tangent (Tanh), the Swish function (or
(Γchan ) and piece (Γpiece ) walls, i.e.: SILU, Sigmoid Linear Unit) (Dung et al., 2023), the Gaussian Error
Ni Ni
Linear Unit (GELU) (Hendrycks and Gimpel, 2016) and considering
x 1 ∑ y 1 ∑ ⃒ ⃒
non-adaptive and adaptive activations (Jagtap et al., 2020). Optimiza­
L = |ux (i, t)− 1|, L = ⃒uy (i, t)⃒ ∀ i ∈ Γinlet ∧ t ∈ [0, 3]
in
Ni i=1 in
Ni i=1 tions have been carried out using ADAM (adaptive moment estimation),
(16) with an initial learning rate (lrstart ) of 0.001 and an exponential decay
with γ = 0.999947 (Wang et al., 2023a), such that
No
1 ∑
L out = |ux (i, t)− 1| ∀ o ∈ Γoutlet ∧ t ∈ [0, 3] (17) lrstep = lrstart • γ step (21)
No o=1
Since our setups consider a temporal evolution with t ∈ [0,3], solving
1 ∑
Nc w
1 ∑
Nc w
⃒ ⃒ the problems on the whole-time window at once would lead to poor
L xcw = |ux (cw , t)|, L y
cw =
⃒uy (cw , t)⃒ ∀ cw ∈ Γchan ∧ t
results. This has been assessed in some preliminary tests where a single
Ncw cw =1 Ncw cw =1
time window has been used, confirming the findings presented in the
∈ [0, 3] (18) literature (Wight and Zhao, 2020; Krishnapriyan et al., 2021). See Ap­
pendix B for further discussion on single vs multiple time windows and
Np w Np w
x 1 ∑ y 1 ∑ ⃒ ⃒ an experiment. Such disappointing results are mainly due to an insuf­
L = |ux (pw , t)|, L = ⃒uy (pw , t)⃒ ∀ pw ∈ Γpiece ∧ t
pw
Npw pw =1 pw
Npw pw =1 ficient number of temporal samples. To avoid the use of a fixed sam­
pling, that can be used to ensure a sufficient time coverage and would
∈ [0, 3]
lead to inaccuracies in not considered locations, we defined smaller time
(19) windows in which space-time points are randomly sampled with a suf­
y ficient density on both time and space. Moreover, in this way each time
where: L xin , L L out are related to the IC on inlet and outlet, L xcw , L ycw
in , window works on a normalized time span in [0,1] that helps to better
and L xpw , L ypw to the BC on the channel and piece walls, and Ni , No , Ncw , deal with numerical stability. Thus, an approach with a set of sequential
and Npw indicate the total number of points used in the inlet, outlet, moving time windows has been used to reduce the problem complexity,
channel and piece wall respectively. The final multi-objective loss splitting the time range in multiple T windows (with T = 3 in our cases).
function L is defined as follows: For each time window a network with the same architecture and
y hyperparameters is trained anew, using as initial conditions for the
L = wxM L x y x
M + wM L M + wC L C + win L in
x
+ wyin L y
in + wout L out + wxcw L x
cw
y x y
sampling points at t = 1 and t = 2 the final estimates obtained in the
+ wcw L cw + wpw L pw + wpw L pw
y x y
previous time step. The maximum number of training steps for each time
(20) window has been set to 100000.
The number of sampling/collocation points per step on which to
y y y y
where: wxM , wM , wC , wxin , win , wout , wxcw , wcw and wxpw , wpw , are the weights evaluate the losses has been variated in different cases to assess the
used to obtain L by linear scalarization. Higher the weight, greater is relation between the achievable accuracy and the training times: indeed,
the impact of the specific loss/constraint. However, in order to obtain a higher number of sampling points would lead to better accuracies,
valid results all the constraints must be considered, and giving to some since more points are constrained to follow the PDEs and the IC and BC.
of the sub-losses a higher importance/weight lead to unsatisfactory Conversely, having to compute the loss on a huge number of points
outcomes. In addition, the sub-losses can have different scales that, if not would increase the computational times. Details and results on such a
considered, could result in an unbalanced loss. Moreover, the losses can trade-off are reported in the next sections.
have different scales, not easy to know before training, that if not
considered could lead to an unbalanced loss. Therefore, manual tuning 6. Validation of PINN results against CFD/FVM simulations
of the weights is very difficult. For this reason, we have used the
ReLoBraLo (Relative Loss Balancing with Random Lookback) balancing In this section, the results of the experimental analysis are reported.
schema (Bischof and Kraus, 2021), with parameters α, E[ρ] = β, and T In order to assess the velocity and pressure estimate accuracy, results
set to 0.95, 0.99, 1, respectively. When using ReLoBraLo, weights are produced by the PINN have been compared with the outputs of the FVM
iteratively updated according to the single sub-loss history to balance solution described in Section 4. The metrics for the assessment are re­
contributions. Finally, Stochastic Gradient Descent (SGD), or other ported in Section 6.1.
optimization algorithms, are performed through backpropagation to

7
T. Botarelli et al. Engineering Applications of Arti cial Intelligence 148 (2025) 110347

The results on the usage of different architectures are provided in window during training. Three consecutive time windows are used.
Section 6.2. The outcomes obtained by using different activation func­ In Fig. 4, the steady state (after the transient) results of velocity
tions are reported in Section 6.3, while by using different number of magnitude are shown, while in Fig. 5, error maps for the MLP-PINN and
neurons, different number of layers are discussed in Section 6.4. These MFN-PINN with respect to OpenFOAM are reported. To appreciate the
cases have been carried out using the setup with a single box piece in the transient evolution videos of the renderings are accessible on
channel/pipe with small inlets/outlets, represented in Fig. 2(a). Using https://www.snap4city.org/1010.
the best architecture and activation, results of new trainings with the In Fig. 6, the time trend of the MAE (tʹ) on the transient phase
configurations shown in Fig. 2(b) and (c) are reported in Section 6.5. computed between the OpenFOAM solution and each of the PINN ar­
Assessments on using a different number of points are reported in Sec­ chitectures is provided for both velocity magnitude and pressure. In
tion 6.6 for all the piece geometries to show the achievable trade-off Table 2, the MAEavg and MAEst for the velocity magnitude, ux and uy
between accuracy and training times. velocity component, and pressure are reported together with the
Note that PINN inference times, i.e., the times required to obtain a training times. The MFN-PINN architecture outperforms the MLP-PINN,
prediction using a trained model are on average less than 6 min for the that shows a diverging error as time increases. Differently, the MFN-
considered time span, including also the time required to load the PINN quickly converged to low MAE (around 0.02) in particular for
sampling points in memory and save the results. Therefore, in the the second and last time windows (i.e., t’ > 1). This result demonstrates
following, we focused on the training times that undoubtfully dominate the efficacy of the feature transformation applied by the MFN that can
the computational times required to get CFD simulations with PINN. better learn the N-S solution. In the above comparison of MLP and MFN
For completeness, in addition to the results presented in the architectures, the same number of points has been used. On the other
following sections, additional video renderings of the experiment results hand, a relevant increment of points for the MLP does not allow to reach
are reported in a web page accessible at the following link https://www. similar results as those obtained by using MFN. Therefore, even if the
snap4city.org/drupal/node/1010. training times for the MFN are longer with respect to those required by
All computations have been carried out exploiting a workstation the MLP, the differences in performance, the MFN has been adopted for
equipped with an Intel Intel(R) Xeon(R) CPU E5-2680 v2 @ 2.80 GHz, all the subsequent cases.
with 24 GB or RAM and an NVIDIA GTX4090 GPU.

6.3. Impact of different activation functions

6.1. Assessment metrics
In this section, the results obtained by using different activation
To measure the differences between the FVM and PINN results, the functions for the proposed PINN solution are discussed. Solutions have
Mean Absolute Error at time t (MAE(t’)) is computed according to (22) been carried out by using the MFN-PINN architecture with Tanh acti­
taking into account the N points uniformly distributed in channel C used vations. Additional cases have been conducted by using the SILU and
in the FVM simulation (different for each case and with an order of GELU functions. All activation functions have been evaluated also in the
magnitude of 104 ). They are used in the inference phase of the trained adaptive activation mode. Errors have been computed comparing our
PINN (i.e., for the evaluation PINN and FVM work on the same points). MFN-PINN-based results with respect to the OpenFOAM solution. As
MAE(t’) is defined as: shown with previous cases, in Table 3, MAEavg , MAEst , and training times
N
are reported, and in Fig. 7 MAE(tʹ) for t’ ∈ [0, 3] velocity and pression are
1 ∑ shown. Results demonstrate that there are no strong differences among
MAE(tʹ) = |OF(c, tʹ) − PINN(c, tʹ)| (22)
N c=1 the activation functions, with the Adaptive GELU obtaining the best
average scores for velocity estimates. SILU achieves the best results for
where: OF(c, tʹ) and PINN(c, tʹ) are the results obtained with FVM the pressure p. Tanh seems to obtain on average the worst results:
OpenFOAM and PINN simulations respectively, evaluated in the points indeed, Tanh has upper and lower bounds, while GELU and SILU are
c ∈ C in each a-dimensional time t’ ∈ {0, T}, with T = 3. Note that: unbounded, and this can mitigate vanishing gradient problems and help
OF(c, tʹ) and PINN(c, tʹ) can be computed on specific variables: velocity to learn representations with large dynamics. Note that we did not
magnitude (‖u‖), horizontal (ux ) and vertical (uy ) components of the include in this test the ReLU (Rectified Linear Unit) activation. Even if
velocity, and pressure (p). In this way, it is possible to observe the ReLU does not provide an upper bound, like SILU and GELU, it is a non-
temporal evolution of the errors at each time instant. To assess the error smooth piecewise linear function that is not differentiable at zero and
for the whole timespan, the mean MAE over time is used and defined as: can lead to inaccurate or suboptimal solutions (Dung et al., 2023).
T Finally, considering the training times, all the adaptive activations
1∑
MAEavg = MAE(tʹ) (23) required longer times due to the additional parameters to be learnt.
T tʹ=0
According to these results, the best activation seems to be the adaptive
On the other hand, to estimate the error at steady state (when the versions of the SILU and GELU, and the adaptive GELU has provided a
transient is finished) we defined the MAEst as: faster training time. Therefore, in the following, the GELU Adaptive
activation function has been used.
T
1 ∑
MAEst = MAE(tʹ) (24)
(T − k) tʹ=k
6.4. Impact of different layer width and network depth

where: k has been set to 2.5, since we observed that for tʹ ≥ 2.5 the
In order to identify the better ranked architecture an analysis of the
velocity flow fields stable.
impact of the number of neurons of each network layer has been per­
formed. Using the MFN-PINN and the GELU adaptive activation,
6.2. Impact of different network architectures different layers’ widths with 64, 128, 256, and 512 neurons have been
assessed. Results are shown in Fig. 8 and Table 4, where transient errors
Results obtained with the different PINN architectures – the Multi- for the velocity magnitude and pressure, and average errors with the
Layer Perceptron (MLP-PINN), and the Modified Fourier Network training times are reported, respectively.
(MFN-PINN) – are hereafter discussed. For both MLP-PINN, MFN-PINN 6 As a general consideration, the increment of the number of neurons
layers with 512 neurons each have been used. The Tanh activation generally leads to lower errors. Configurations with 64 or 128 appeared
function has been used, and 4850 points have been adopted in each time to be unable to learn the N-S equations, since they resulted in high er

8
T. Botarelli et al. Engineering Applications of Arti cial Intelligence 148 (2025) 110347

Fig. 4. Steady-state renderings of the velocity magnitude estimates ‖u‖ obtained with (a) OpenFOAM, (b) MLP-PINN, and (c) MFN-PINN at time 2.5 for case study of
Section 6.2 on neural architecture evaluation. While OpenFOAM and MFN-PINN renderings are practically identical, the MLP-PINN is clearly unable to estimate the
correct flows. Please observe Fig. 5 for the related error maps. (Best viewed in color). (For interpretation of the references to color in this figure legend, the reader is
referred to the Web version of this article.)

Fig. 5. Absolute error map for the steady-state results at time 2.5 of the velocity magnitude ‖u‖ for cases of Section 6.2 where different neural architectures are
evaluated. In (a) the error map of the estimates obtained with the MLP-PINN; in (b) error map obtained with the MFN-PINN. Please note that the color-scale ranges of
the error maps are different from the scale of the flows of Fig. 4. When using the MLP-PINN, high errors with respect to the OpenFOAM estimates are observed, while
the MFN-PINN is able to obtain very small differences. (Best viewed in color). (For interpretation of the references to color in this figure legend, the reader is referred
to the Web version of this article.)

Fig. 6. Mean absolute error, MAE(tʹ), plots for MLP-PINN (in purple) and MFN-PINN (in red) of: (a) velocity magnitude ‖u‖, and (b) pressure p. MAE(tʹ) is computed
according to eq. (22) exploiting the OpenFOAM estimates. MFN-PINN clearly outperform the MLP-PINN that archived higher errors with and evident growth as time
proceeds. (Best viewed in color). (For interpretation of the references to color in this figure legend, the reader is referred to the Web version of this article.)

Table 2
MAEavg and MAEst considering the velocity magnitude (‖u‖), horizontal (ux ) and vertical (uy ) components of the velocity, and pressure (p). The last column reports the
training times for the different PINN architectures. Best values (i.e., lower errors/lower times) are highlighted in bold. Even if MLP-PINN, being a simpler architecture,
can be trained in shorter times it achieved errors an order of magnitude larger compared to the MFN-PINN that is therefore to be preferred.
Architecture MAEavg MAEst Training Times (hh:mm)

‖u‖ ux uy p ‖u‖ ux uy p

MLP-PINN 0.121 0.111 0.053 1.489 0.130 0.119 0.054 1.526 03:32
MFN-PINN 0.017 0.017 0.009 0.179 0.011 0.010 0.005 0.171 04:55

rors. Conversely, layers with 256 and 512 neurons obtained very similar ported cases, a layer width of 256 neurons has been used, since it can
errors. In particular, when the system reaches the steady state (tʹ > achieve low errors with reduced training times.
2.5). However, having wider layers impacts negatively on the execution A similar test has been conducted to assess the impact of different
time required for the network training passing from 04h:25m (for 256 network depths by evaluating results using networks composed by 3, 6,
neurons) to 05h:49m (for 512). For these reasons, in the following re­ and 9 hidden layers. Fig. 9 and Table 5 show the transient errors for both

9
T. Botarelli et al. Engineering Applications of Arti cial Intelligence 148 (2025) 110347

Table 3
MAEavg and MAEst on the estimates obtained with the MNF-PINN using different activation functions. Errors are reported for velocity magnitude (‖u‖), horizontal (ux )
and vertical (uy ) components of the velocity, and pressure (p). In the last column the training times required for each case are shown. Best values are highlighted in
bold. Note that, the results for Tanh are the same as those reported in Table 2 for the MFN-PINN. SILU and GELU (both the adaptive and non-adaptive versions) have
obtained very close results. However, the GELU-Adaptive appears to have achieved the best trade-off between high accuracy (low errors) and faster training times, thus
it has been selected for the subsequent tests.
Activations MAEavg MAEst Training Times (hh:mm)

‖u‖ ux uy p ‖u‖ ux uy p

Tanh 0.017 0.017 0.009 0.179 0.011 0.010 0.005 0.171 04:55
Tanh Adaptive 0.018 0.017 0.011 0.158 0.010 0.010 0.005 0.119 05:32
SILU 0.014 0.013 0.007 0.107 0.009 0.008 0.005 0.093 05:47
SILU Adaptive 0.013 0.013 0.007 0.126 0.008 0.007 0.004 0.138 06:10
GELU 0.013 0.013 0.007 0.166 0.009 0.008 0.005 0.174 05:26
GELU Adaptive 0.013 0.012 0.007 0.145 0.008 0.008 0.004 0.138 05:49

Fig. 7. Plots of MAE(tʹ), computed according to eq. (22), considering in (a) the velocity magnitude ‖u‖ and in (b) the pressure p, for the whole-time span (t ∈ [0, 3])
computed with respect to the OpenFOAM solution for the different activation functions. While both versions of Tanh (red and orange plots) show higher ‖u‖ errors,
SILU and GELU share very close results (see Table 3 to observe more detailed error results). (Best viewed in color). (For interpretation of the references to color in this
figure legend, the reader is referred to the Web version of this article.)

Fig. 8. Plots of MAE(tʹ), computed according to eq. (22), of the velocity magnitude ‖u‖ (a) and pressure p (b) computed w.r.t. the OpenFOAM solution for the
different layer widths, i.e., 64 (purple), 128 (cyan), 256 (yellow), and 512 (red) neurons. Plots are related to the results presented in Table 4. Except for the first-time
window (tʹ ∈ [0, 1]) the network using layers of 256 neurons achieves very close ‖u‖ errors if compared to those of the network with 512 neurons per layer. Due to this
small difference and to its faster training times (see Table 4), the solution with 256 has been preferred. (Best viewed in color). (For interpretation of the references to
color in this figure legend, the reader is referred to the Web version of this article.)

velocity magnitude and pressure, and the average errors together with 6.5. Results on different piece geometries
their respective training times for the three network depth configura­
tions. The architecture using 3 hidden layers requires an inferior In this section, results on the circular cylinder (see Fig. 2(b)) and the
training time and obtains the worst performances, in particular in the three-boxes (see Fig. 2(c)) piece geometries are presented. We used the
first two time windows. Solutions using 6 and 9 hidden layers achieve MFN-PINN architecture, with 6 layers of 256 neurons each, and GELU
comparable results, with the latter obtaining slightly lower errors in the adaptive activation function. The networks have been trained from
transient phase. However, solution with 9 layers increases the training scratch for each experiment. Fig. 10 shows the rendering of the steady
time to 05h:18m with respect to the 04h:25m required by using 6 hidden state velocity magnitude obtained with OpenFOAM and PINN and the
layers. Therefore, in the following tests the architecture with 6 hidden respective error maps. In Table 6, MAEavg and MAEst are reported.
layers has been used since it provides better compromise of errors and Finally, in Fig. 11, the transient error plots of the velocity magnitude and
training time. pressure are presented. For comparison purposes, we reported also the
results obtained with the single-box piece used in the previous sections.
Results are favorable for all the different piece geometries, meaning that

10
T. Botarelli et al. Engineering Applications of Arti cial Intelligence 148 (2025) 110347

Table 4
MAEavg and MAEst considering estimates for the velocity magnitude (‖u‖), horizontal (ux ) and vertical (uy ) components of the velocity, and pressure (p) obtained with
the MFN-PINN, using the GELU-Adaptive activations for the different layer widths (from 64 to 512 neurons for each hidden layer). In the last column training times are
reported. In bold, the best values are reported. Note that, the results for 512 are the same as those reported in Table 3 for the GELU Adaptive activation. As expected, the
greater the layer width is, the lower are the errors (except for the pressure, as explained in the text) and the longer the training times. Considering the accuracy-time
trade-off, we selected the architecture with 256 neurons per layer, achieving low errors (very close to those obtained using 512 neurons) with a substantial saving in the
required training times (01h:24m less than the time used by the 512-test case).
Layer width MAEavg MAEst Training Times (hh:mm)

‖u‖ ux uy p ‖u‖ ux uy p

64 0.068 0.064 0.032 0.092 0.037 0.034 0.018 0.094 03:24

128 0.040 0.038 0.020 0.187 0.018 0.017 0.009 0.223 03:38
256 0.019 0.019 0.010 0.091 0.010 0.010 0.005 0.116 04:25
512 0.013 0.012 0.007 0.145 0.008 0.008 0.004 0.138 05:49

Fig. 9. Plots of MAE(tʹ) of both velocity magnitude ‖u‖ (a) and pressure p (b) computed w.r.t. the OpenFOAM solution for the different network depths, i.e., 3
(purple), 6 (cyan), and 9 (red) hidden layers. Using 6 or 9 layers produced very similar results with slightly higher errors for the 6-layer case for t ∈ [0,1]. However, as
reported in Table 5, a sensible reduction of training times can be achieved using 6 hidden layers. (Best viewed in color). (For interpretation of the references to color
in this figure legend, the reader is referred to the Web version of this article.)

Table 5
MAEavg and MAEst considering both velocity magnitude (‖u‖), horizontal (ux ) and vertical (uy ) components of the velocity, and pressure (p) for the different network
depths (3, 6, and 9 hidden layers). In the last column training times are reported. Best values are reported in bold. Note that, the results using 6 layers are the same as
those reported in Table 4 for the case with 256 neurons. Using 3 layers resulted in higher errors. Conversely, few differences are appreciable using 6 or 9 layers, and 9
layers require longer training. Thus, the architecture with 6 hidden layers has been selected.
Hidden layer number MAEavg MAEst Training Times (hh:mm)

‖u‖ ux uy p ‖u‖ ux uy p

3 0.033 0.031 0.017 0.144 0.016 0.015 0.007 0.166 03:22

6 0.019 0.019 0.010 0.091 0.010 0.010 0.005 0.116 04:25
9 0.018 0.017 0.010 0.185 0.010 0.010 0.005 0.234 05:18

the MFN-PINN solution has achieved low errors in particular for the reported together with the measured training times, while in Fig. 12 the
steady state. Among the geometries, the case with three-boxes shows the error evolution for all the time instants is presented, as to both velocity
higher velocity errors, followed by the case with single box. This could and pressure. It can be observed that, as in the previous cases, the cir­
be due to the sharp borders of those geometries, thus leading to greater cular cylinder appears to be the easiest to solve, achieving the lowest
difficulties in learning the air flow dynamic, when impacting on the velocity errors. Low errors are anyhow obtained also in the other two
pieces. Indeed, in Fig. 10(c) and (i) the more evident errors are accu­ geometries, and for all cases velocity errors are reduced, as the number
mulated near the front corners of the pieces. On the other hand, the of points increases. Pressure results show fewer regular trends, with
smoother shape of the cylinder leads to better estimates of the flow lower values not always obtained using a higher number of collocation
velocity. Pressure seems to obtain less accurate predictions, however, points. This behavior is due to the fact that we did not impose any
analyzing the OpenFOAM results, pressure varies between [− 8.816, pressure constraints on IC and BC.
1.390], a wider range with respect to the velocity covering the range When using a smaller number of points, training times are reduced,
[0.000, 1.996]. Therefore, such errors are acceptable, taking also into passing from an average of 07h:24m for the High case, to 04h:24m for
account that no specific constraints have been defined for the pressure in the Mid case, and to 02h:47m for the Low case. However, MAEavg and
the PINN loss. MAEst provided relatively small differences between Mid and High cases.
These results suggested that a careful selection of the number of points
produces a significant reduction in training times without relevant drops
6.6. Impact of different number of points on accuracy and time
in performance, and for this reason in the following a novel training
strategy exploiting such a consideration is proposed (see Section 7.2).
In this section, the trade-off between accuracy and the number of
points used is assessed, also considering the impact on the training
7. Training strategies for faster learning with PINN
times. Table 7 shows the different number of points used as High, Mid,
and Low numbers/cases. The Mid number of points has been adopted for
As reported in the previous section, using adequate architecture and
the results of previous sections. In Table 8, MAEavg and MAEst are

11
T. Botarelli et al. Engineering Applications of Arti cial Intelligence 148 (2025) 110347

Fig. 10. Rendering of the velocity magnitude ‖u‖ results regarding the tests using different pieces inside the channel, as described in Section 6.5. PINN results are
obtained using the MFN-PINN with GELU adaptive activation, 6 hidden layers, and 256 neurons for each layer. In the first row the renderings of the single-box case
are reported; in the second row the results using the circular cylinder; in the third row those considering the three-boxes. Column wise, the first column (a, d, g)
reports the OpenFOAM results, while the second column (b, e, h) shows the MFN-PINN estimates. The PINN estimated flows are very close to those obtained with the
OpenFOAM FVM for all the considered piece geometries. To better highlight any differences, in the third column (c, f, i) the absolute error maps, i.e., the pointwise
absolute difference of the PINN and OpenFOAM ‖u‖ estimates, are reported. Please note that in the error maps a smaller range of values ([0.0, 0.1]) is used in the
colormaps to better show the slight differences. Observing the renderings, most of the errors seems to accumulate near the piece border, with the circular cylinder
exhibiting the lowest errors due to its more aerodynamic shape compared with the geometries of the single box and three boxes. (Best viewed in color). (For
interpretation of the references to color in this figure legend, the reader is referred to the Web version of this article.)

Table 6
MAEavg and MAEst for the test cases using different piece geometries (Section 6.5). Errors are related to both velocity magnitude (‖u‖), horizontal (ux ) and vertical (uy )
components of the velocity, and pressure (p). As illustrated also in Figs. 10 and 11, low errors are achieved for all the geometries with the circular cylinder obtaining the
lowest errors.
Piece MAEavg MAEst

‖u‖ ux uy p ‖u‖ ux uy p

Single box 0.019 0.019 0.010 0.091 0.010 0.010 0.005 0.116
Circular cylinder 0.011 0.010 0.006 0.065 0.007 0.006 0.004 0.040
Three boxes 0.024 0.023 0.014 0.048 0.013 0.012 0.006 0.021

activation functions, the proposed PINN results on the three test cases model trained on the single box to assess the possibility of using a
are very satisfactory, showing low differences with respect to the results pre-trained network to solve different problems and reduce training
obtained by OpenFOAM as CFD simulations, both in the transient phase times. Moreover, results on fine-tuning between consecutive time
and at steady state. One of the main limitations in the usage of PINN in windows for the single box geometry are reported.
real scenarios seems to be the training times (see Tables 4, 5 and 8). • Case B) Multi-resolution training: in Section 7.2 we proposed a
Therefore, in this section, alternative training techniques with the training solution named multi-resolution in which the number of
objective of reducing the time required to train neural networks are points used varies in the time windows. Under the hypothesis of
proposed. being more interested in simulating the steady state of the solutions,
we trained the first two time-windows with fewer points, rising their
• Case A) Fine-tuning: in Section 7.1 the circular cylinder and the numbers in the final window. Note that, this is not possible in CFD
three-boxes geometries have been addressed by fine-tuning of the

12
T. Botarelli et al. Engineering Applications of Arti cial Intelligence 148 (2025) 110347

Fig. 11. Plots of MAE(tʹ) for the test cases considering different piece geometries: circular cylinder (purple), three-boxes (cyan) and single box (red). In (a) plots of
velocity magnitude ‖u‖ errors; in (b) those considering the pressure p. Plots are related to renderings shown in Fig. 10, and errors reported in Table 6. Observing the
errors during the time span (tʹ ∈ [0, 3]), the circular cylinder is confirmed to be the easiest test case with the lowest errors. The single box and the three-boxes, even
though showing higher errors due to their more complex geometries, still obtain very good accuracies. (Best viewed in color). (For interpretation of the references to
color in this figure legend, the reader is referred to the Web version of this article.)

)
same used in the previous version (lr = 10− 3 and one with an order of
Table 7 − 2
magnitude larger (lr = 10 ). To achieve this strategy, the sequential
Different numbers of collocation points used in the tests are described in Section
6.6. In the table, the number of points used for each specific area of the
solver of Modulus (SeqSolver lib.) has been modified to introduce the
considered setup are reported (inlet, outlet, channel walls, piece walls, and in possibility to re-load weights from a previously trained model. This
the interior part of the channel where the fluid flow moves). In the last row, the experiment is particularly useful to assess the applicability of PINN in
total number of points is reported. real scenario, where a single well-trained model could be deployed in
Area High Mid Low
different domains with lower effort, thus reducing time and simulation
costs.
Inlet 500 250 100
MAEavg and MAEst errors are reported in Table 9, and transient errors
Outlet 500 250 100
Channel walls 2500 1250 500 in Fig. 13. We reported errors at different number of epochs to highlight
Piece walls 1250 600 250 possible gains in training times compared with the measured errors. For
Channel interior 5000 2500 1000 the circular cylinder at lr = 10− 3 , starting from 40000 epochs the
Total 9750 4850 1950
trained model already achieves velocity errors very similar to the model
trained from scratch (discussed in Section 6.5 and Table 6), and after
solvers, where the mesh grid must be kept fixed for the whole reaching 60000 epochs, errors are equals or even lower using fine-
simulation. tuning. Note that by increasing the learning rate a slight precision
• Case C) Parametrized training: in Section 7.3, we proposed to use a improvement of velocity estimation is obtained with more than 60000
parametrized representation of the piece geometry. In this way it is epochs. A similar behavior can be observed for the three-boxes geome­
possible to train a PINN model to solve a larger number of setups try, with comparable or lower errors starting from 40000 epochs.
with pieces of different dimensions. Pressure shows a more erratic behavior as above discussed.
With the approach proposed, training times have been significantly
reduced, passing from 04h:24m for training the model from scratch, to
02h:01m when fine-tuning for 40000 epochs, to 03h:00m for 60000
7.1. Case A) fine-tuning results epochs. Note that, a small overhead in the measured times has been
obtained, since to get all the results at different epochs reported we
In this case, the results on the circular cylinder and three-boxes test incrementally augmented the number of epochs: for example, the model
cases have been obtained by fine-tuning the network model trained for trained till 40000 epochs has been saved and then reloaded to be used as
computing the solution of the single box case. The analysis estimated the a starting point to perform additional training steps up to 60000 epochs
errors at different numbers of epochs, assessing any possible reduction of causing an overhead due to saving and loading. When training in a single
the training times. Different learning rates have been evaluated, the

Table 8
MAEavg and MAEst for the different geometries and different numbers of collocation points (see Section 6.6). Errors consider both velocity magnitude (‖u‖), horizontal
(ux ) and vertical (uy ) components of the velocity, and pressure (p). In the last column training times (in hours and minutes) are reported. Best values for each different
geometry are reported in bold. Increasing the number of sampling points helps to obtain more accurate estimates at the expense of higher training times, since the loss
function is evaluated on a greater number of points. These results highlight the relevance of choosing an adequate number of points, particularly for the more complex
geometries like the single box and the three-boxes cases.
Geometry Collocation points MAEavg MAEst Training times (hh:mm)

‖u‖ ux uy p ‖u‖ ux uy p

Single box Low 0.024 0.023 0.013 0.071 0.013 0.013 0.006 0.048 02:47
Mid 0.019 0.019 0.010 0.091 0.010 0.010 0.005 0.116 04:25
High 0.017 0.016 0.009 0.089 0.009 0.008 0.005 0.068 07:14
Circular cylinder Low 0.012 0.012 0.007 0.128 0.007 0.006 0.004 0.166 02:47
Mid 0.011 0.010 0.006 0.065 0.007 0.006 0.004 0.040 04:24
High 0.010 0.010 0.006 0.219 0.007 0.006 0.004 0.226 07:11
Three boxes Low 0.028 0.026 0.016 0.059 0.015 0.014 0.008 0.036 02:48
Mid 0.024 0.023 0.014 0.048 0.013 0.012 0.006 0.021 04:24
High 0.021 0.019 0.012 0.112 0.010 0.009 0.005 0.126 07:47

13
T. Botarelli et al. Engineering Applications of Arti cial Intelligence 148 (2025) 110347

Fig. 12. Plots of MAE(tʹ) for results using different number of collocation points described in Section 6.6, and Table 8. In (a) and (b) plots of velocity magnitude ‖u‖
and pressure p error, respectively, for the single box geometry. Similarly, plots for the circular cylinder are shown in (c) and (d), while for the three-boxes are
depicted in (e) and (f). All the plots have confirmed that the PINN can achieve accurate estimates, in line to those obtained with OpenFOAM FVM. It can be noted that
the circular cylinder is less affected by the different number of points. Conversely, slight improvements are achieved in the single box and three-boxes test case, when
using a higher number of points. (Best viewed in color). (For interpretation of the references to color in this figure legend, the reader is referred to the Web version of
this article.)

run up to a given number of epochs, by skipping intermediate savings of the network from scratch, both by moving to different geometry
and loadings, the total training times are lower. setups or between successive time windows.
A second application of fine-tuning is to exploit the model trained for
a time-window and fine-tune it in the successive time window. Let us
7.2. Case B) multi-resolution training
now use the single-box geometry. Let M 0 be the trained model for the
first time-window (obtained with the solution described in Section 6.5).
In this section, results using the proposed multi-resolution training
Then, model M 1 of the second time window has been obtained by fine-
approach are presented. In this case, the goal has been to identify a
tuning of M 0 . Similarly, the third time window model M 2 , has been solution to accurately simulate the FD behavior in the last time instants
obtained by fine-tuning M 1 . (less importance is given to the first steps of the transient phase). This
In Table 10 MAEavg , MAEst , and training times (considering the time approach can be an interesting feature of proposed PINN which could
required to train the 2nd and the 3rd windows) are reported. In Fig. 14, shorten the time needed to produce the results of classic CFD simulation,
transient plots are shown. Results Identical to those obtained by training in which the mesh grid must remain fixed and require to estimate ve­
each time window from scratch have been achieved by fine-tuning for locity and pressure for each time instant in order to estimate/propagate
100000 epochs (except for the pressure that improved in the fine-tuning the data to the successive one.
case), and by fine-tuning for 50000 epochs, errors are very close to those In this case, the first two time-windows have been trained, by using a
obtained in the training from scratch case. Moreover, by using 50000 reduced number of points. On this view, two multiresolution cases have
fine-tuning epochs, training times halve from 03h:17m to 01h:38m, with been analyzed, namely: (a) using the Low number of points (i.e., 1950);
a speed-up of 2. (b) reducing the number of points to 1000 (Inlet:75, Outlet: 75, Channel
In conclusion, fine-tuning is a successful approach to reduce training Walls: 250, Piece Walls: 100, Channel Interior: 500). Please note that in
times and to obtain similar or better results with respect to any training both cases, in the third final window the points are those of the High

14
T. Botarelli et al. Engineering Applications of Arti cial Intelligence 148 (2025) 110347

Table 9
MAEavg and MAEst for fine-tuning starting from the model trained on the single box geometry and described in Section 7.1. In the top part the results for circular
cylinder, in the bottom part those for three-boxes. In both cases different learning rates (lr) have been considered. Errors include both velocity magnitude (‖u‖),
horizontal (ux ) and vertical (uy ) components of the velocity, and pressure (p). Training times, for different number of fine-tuning epochs, are shown in the last column.
In bold best values are described, according to the different kind of transfer learning. Low errors are achieved in both cases starting from 40000 epochs with a relevant
reduction of training times if compared with those required to train the model from scratch. With additional epochs, using transfer-learning, the solution shows errors
lower than those yielded with a training from scratch.
Geometry Training type Training epochs MAEavg MAEst Training Times (hh:mm)

‖u‖ ux uy p ‖u‖ ux uy p

Single Box Fine-tuning (lr = 10− 3 ) 5000 0.067 0.065 0.041 0.188 0.064 0.062 0.039 0.240 00:13
to 10000 0.042 0.040 0.025 0.209 0.035 0.033 0.022 0.270 00:29
Circular cylinder 20000 0.023 0.022 0.014 0.229 0.018 0.017 0.011 0.310 00:59
40000 0.012 0.012 0.007 0.248 0.008 0.007 0.005 0.342 02:01
60000 0.010 0.010 0.006 0.252 0.007 0.006 0.004 0.350 03:00
80000 0.010 0.009 0.005 0.248 0.007 0.006 0.004 0.346 04:01
100000 0.009 0.009 0.005 0.247 0.007 0.006 0.004 0.343 05:01
Fine-tuning (lr = 10− 2 ) 5000 0.070 0.067 0.037 0.205 0.074 0.071 0.039 0.222 00:13
10000 0.045 0.042 0.025 0.313 0.046 0.044 0.027 0.380 00:29
20000 0.023 0.021 0.014 0.457 0.022 0.021 0.013 0.585 00:59
40000 0.014 0.013 0.008 0.561 0.012 0.011 0.007 0.722 02:01
60000 0.009 0.009 0.005 0.568 0.007 0.006 0.004 0.705 03:01
80000 0.008 0.008 0.004 0.548 0.007 0.006 0.004 0.651 04:01
100000 0.008 0.008 0.004 0.531 0.007 0.006 0.004 0.609 05:02
Training from scratch 100000 0.011 0.010 0.006 0.065 0.007 0.006 0.004 0.040 04:24
Single Box Fine-tuning (lr = 10− 3 ) 5000 0.091 0.083 0.042 0.088 0.087 0.078 0.04 0.103 00:13
to 10000 0.069 0.064 0.033 0.096 0.063 0.058 0.026 0.116 00:29
Three boxes 20000 0.046 0.043 0.023 0.135 0.037 0.034 0.017 0.165 00:59
40000 0.027 0.025 0.015 0.168 0.019 0.018 0.009 0.214 02:01
60000 0.020 0.018 0.011 0.181 0.012 0.012 0.006 0.233 03:01
80000 0.017 0.016 0.010 0.186 0.010 0.009 0.005 0.24 04:01
100000 0.016 0.015 0.010 0.187 0.009 0.008 0.005 0.242 05:02
Fine-tuning (lr = 10− 2 ) 5000 0.089 0.083 0.038 0.117 0.078 0.072 0.033 0.091 00:13
10000 0.080 0.074 0.036 0.084 0.076 0.069 0.034 0.059 00:29
20000 0.044 0.041 0.022 0.202 0.040 0.038 0.017 0.157 00:59
40000 0.027 0.025 0.014 0.235 0.021 0.02 0.009 0.182 02:01
60000 0.016 0.015 0.009 0.244 0.012 0.011 0.005 0.197 03:01
80000 0.012 0.011 0.007 0.236 0.008 0.007 0.004 0.192 04:01
100000 0.011 0.011 0.006 0.231 0.007 0.006 0.004 0.190 05:02
Training from scratch 100000 0.024 0.023 0.014 0.048 0.013 0.012 0.006 0.021 04:24

Fig. 13. Plots of MAE(tʹ) for the fine-tuning tests starting from a model trained on the single box geometry (Section 7.1, Table 8). In (a) and (b) velocity magnitude
‖u‖ and pressure p errors for the circular cylinder using lr = 10− 3 ; in (c) and (d) the same test using lr = 10− 2 . Similarly, in (e) and (f) velocity magnitude ‖u‖ and
pressure p errors for the three-boxes using lr = 10− 3 ; in (g) and (h) using lr = 10− 2 . In all the plots errors related to the estimates obtained with a model trained from
scratch are reported in red. By augmenting the number of fine-tuning epochs more accurate results are achieved in all the cases. Moreover, 40000 epochs are
sufficient to obtain pretty accurate results, while after 60000 epochs results are practically equals to those obtained by training the model from scratch. (Best viewed
in color). (For interpretation of the references to color in this figure legend, the reader is referred to the Web version of this article.)

15
T. Botarelli et al. Engineering Applications of Arti cial Intelligence 148 (2025) 110347

Table 10
MAEavg and MAEst for the fine-tuning between time windows experiment described in Section 7.1. The table shows errors for both velocity magnitude (‖u‖), horizontal
(ux ) and vertical (uy ) components of the velocity, and pressure (p). In the last column the training times for the different number of epochs used to train the 2nd and 3rd
time windows are reported. In the last row, the errors obtained by training from scratch are reported for comparative ends. Best values are shown in bold. Note that, at
the expense of a slight reduction in accuracy, using 50000 epochs to fine-tune the 2nd and 3rd time windows, training times are halved, thus obtaining a speed-up of 2.
Training epochs MAEavg MAEst Training Times for 2nd and 3rd windows (hh:mm)

‖u‖ ux uy p ‖u‖ ux uy p

50,000 0.012 0.011 0.007 0.066 0.008 0.007 0.005 0.044 01:38
100,000 0.011 0.011 0.006 0.055 0.007 0.006 0.004 0.022 03:17
Training from scratch 0.011 0.010 0.006 0.065 0.007 0.006 0.004 0.040 03:17

Fig. 14. Plots of MAE(tʹ) for the fine-tuning between successive time windows experiment (see Section 7.1, Table 10). Results are obtained considering the single box
geometry. In (a) plot of the velocity magnitude ‖u‖ errors, in (b) the errors on pressure p with different plots according to the used number of epochs (violet for
50000, cyan for 100000). In red errors of the estimates from the model trained from scratch. Note that, using 50000 fine tuning epochs accurate results are achieved
(with slight difference with respect to. 100000 epochs and the training from scratch). (Best viewed in color). (For interpretation of the references to color in this
figure legend, the reader is referred to the Web version of this article.)

level (i.e., 9750). approach opens the possibility to train a single PINN model capable of
Table 11 reports the achieved MAEavg and MAEst together with the providing a range of solutions and setups. To assess this capability, the
number of points used in each time windows and the related training single box case has been adopted defining the box dimensions with the
times (that for the multi-resolution cases are divided in the times following ranges instead of the fixed values described in Section 3. The
required for the first two windows, and the times employed to train the horizontal dimension varies in the range [0.4, 0.6] and the vertical
last window). Transient error plots are shown in Fig. 15. As expected, dimension in the range [0.15, 0.25]. On this basis, a new MFN-PINN
transient errors are higher with respect to the results obtained by model has been trained with the hyperparameters identified, using the
training the model from scratch since less collocation points have been adaptive GELU activation, and the High number of points. Differently
used in the first time-windows. Conversely, steady state results (MAEst ) from the previous cases, the maximum training epochs has been
are practically identical to the case with 3 consecutive trainings, increased to 150000 rising the training times to 18h:23m. However,
demonstrating the validity of the multi-resolution approach for short­ considering the wide range of possible setups that can be solved, such an
ening the execution time and obtaining equivalent results. Therefore, a increase in computational times is acceptable. For example, suppose that
reduction of training time has been achieved passing from the 04h:25m in a production line an autoclave is used to cure 10 pieces with a similar
to 03h:50m in case (b). shape and different dimensions. According to the regular approach a
different PINN should be trained for each different piece requiring
04h:24m. Then the whole set of 10 would require a total training time of
7.3. Case C) parametrized learning
about 44 h. On the contrary, using the proposed parametrized approach
for training, a speed-up greater than 2 would be achieved.
Since in PINNs the grid of points/mesh do not need to be defined, an
To evaluate the achieved results, a set of new OpenFOAM simula­
additional advantage can be recovered from the possibility of providing
tions have been carried out changing the dimensions of the piece inside
the geometry in a parametrized manner. For example, by defining
the channel, named Small and Big, with dimensions [0.4,0.15] and [0.6,
ranges for the element dimensions instead of fixed numbers. This

Table 11
MAEavg and MAEst for the multi resolution training strategy presented in Section 7.2. Errors are reported for the velocity magnitude (‖u‖), horizontal (ux ) and vertical
(uy ) components of the velocity, and pressure (p). In the last column training times are reported. Note that for the multi-resolution cases, training times are divided in
the times required for the first two windows, plus the times employed to train the last window. In the last row, errors and times related to a model trained from scratch
are shown for comparison. The case (a) and (b) indicate the different number of points used in training the first two windows: 1950 points in (a), 1000 points in (b). Best
values are reported in bold. While the multi-resolution approaches obtain higher MAEavg errors that considers the whole-time span, MAEst errors, limited to the steady
state, are comparable to those obtained by model trained from scratch.
Experiment MAEavg MAEst Training Times (hh:mm)

‖u‖ ux uy p ‖u‖ ux uy p

Multi resolution (a) 0.026 0.024 0.014 0.057 0.011 0.011 0.006 0.026 01:57 + 02:14
Multi resolution (b) 0.030 0.029 0.015 0.110 0.012 0.012 0.006 0.101 01:36 þ 02:14
Training from scratch 0.019 0.019 0.010 0.091 0.010 0.010 0.005 0.116 04:25

16
T. Botarelli et al. Engineering Applications of Arti cial Intelligence 148 (2025) 110347

Fig. 15. Plots of MAE(tʹ). In (a) the velocity magnitude ‖u‖ errors; in (b) the pressure p errors computed with respect to the OpenFOAM solution for different multi
resolution trainings (in purple the (a) case using 1950 points in the first two windows, in cyan the case (b) using 1000 points). In red the reference error plot obtained
from the model trained from scratch. Note that, in particular for ‖u‖, the estimates of the multi-resolution approaches show higher errors in the first-time windows
(tʹ ∈ [0, 2]), while they converge to values comparable to those obtained with the model trained from scratch, achieving practically the same errors, when the system
reaches the steady state (tʹ ≥ 2.5). (Best viewed in color). (For interpretation of the references to color in this figure legend, the reader is referred to the Web version
of this article.)

0.25] respectively, comparing the results obtained with the parame­ adopted in scenarios where similar pieces of different dimensions are
trized approach against the results on the same setup obtained by placed inside the channel.
training the model from scratch (similarly as done in Section 6.5) using
the Mid number of collocation points and 100000 epochs. 8. Conclusions
Transient errors are presented in the plots of Fig. 16, and MAEavg and
MAEst in Table 12. Thus, training from scratch seems to produce better In this study, we explored the application of PINNs to complex fluid
results (with the exception of pressure estimates that as in the previous dynamic scenarios, focusing on self-supervised setups that exploit solely
case tests show a less regular behavior for the reasons previously dis­ physical laws without any data. Our work introduced several novel
cussed). However, there are small differences, in particular when strategies to enhance the efficiency and accuracy of PINNs, particularly
reaching the steady state. Moreover, in the case of the Small rectangle when dealing with the Navier-Stokes equations in various geometries
differences are smaller, and this can give hints on the fact that the less is inspired by industrial autoclave setups. All the PINN results have been
the area of the piece inside the channel, minor is its impact on the flow validated against those obtained by using OpenFOAM, a widely used
dynamics and smaller are the errors achieved in the transient phase. In CFD FVM tool. The comparison included assessing errors in velocity
conclusion, the parametrized training can be a valid solution to be magnitude, horizontal and vertical velocity components, and pressure.

Fig. 16. Plots of the MAE(tʹ) for the parametrized learning tests presented in Section 7.3. In (a) and (b) velocity magnitude ‖u‖ and pressure p errors for the Big
rectangle. In (c) and (d) the same errors for the Small rectangle. In all the plots, results of the parametrized learning are reported in violet, while errors obtained by
training from scratch using the differently sized rectangles are shown in red. It can be noted that, the parametrized training results are compared with the results
obtained by training the network from scratch, in particular when the steady state is reached. (Best viewed in color). (For interpretation of the references to color in
this figure legend, the reader is referred to the Web version of this article.)

17
T. Botarelli et al. Engineering Applications of Arti cial Intelligence 148 (2025) 110347

Table 12
MAEavg and MAEst for the differently sized geometries used in the parametrized learning test case (see Section 7.3 and Fig. 16). For both the Big and Small rectangles,
errors with respect to OpenFOAM are reported considering the parametrized strategy and the training from scratch. Errors consider the velocity magnitude (‖u‖),
horizontal (ux ) and vertical (uy ) components of the velocity, and pressure (p). Lower errors are highlighted in bold for each test case. Results indicate that using
parametrized learning good estimates are achieved comparable with those yielded by a model trained from scratch, in particular for the steady state (MAEst ).
Geometry Training type MAEavg MAEst

‖u‖ ux uy p ‖u‖ ux uy p

Big rectangle Parameterized 0.032 0.030 0.015 0.157 0.010 0.009 0.005 0.163
Training from scratch 0.022 0.021 0.010 0.192 0.013 0.012 0.005 0.137
Small rectangle Parameterized 0.025 0.023 0.017 0.165 0.014 0.013 0.009 0.172
Training from scratch 0.021 0.020 0.014 0.331 0.012 0.011 0.007 0.474

We conducted a large set of experiments to select the optimal neural applications.

architecture and hyperparameters. The study evaluated different neural In summary, this study demonstrates that PINNs are effective tools
network architectures, such as MLP-PINN and MFN-PINN, and activa­ for addressing complex fluid dynamics problems governed by the
tion functions like Tanh, SILU, and GELU. The MFN-PINN architecture Navier-Stokes equations. They offer an effective alternative to tradi­
consistently outperformed the MLP-PINN, demonstrating its efficacy in tional CFD methods, particularly in scenarios requiring rapid simula­
learning the Navier-Stokes solutions. Adaptive GELU activation func­ tions, when a range of pieces have to be used. With ongoing
tions yielded the best average scores for velocity estimates, while SILU advancements and further validation, PINNs have the potential to
excelled in pressure estimates. Note that in future work the adoption of revolutionize the field of fluid dynamics simulation, paving the way for
trainable activation function could be considered (Wang et al., 2023b). broader adoption in practical engineering and industrial applications.
Then, additional piece geometries have been considered to evaluate By embracing these innovative neural network approaches, we can
PINN performance widening the test cases. Using the identified archi­ significantly enhance our capability to simulate and predict fluid
tecture (MFN-PINN with adaptive GELU activation) rapid convergences behavior in various contexts, ultimately leading to more efficient and
to low MAE both during the transient phase and in the steady state phase accurate designs and processes in engineering disciplines. The integra­
have been observed for all the different setups. This further validates tion of PINNs in fluid dynamics simulations presents a transformative
PINN performance that show small differences with respect to estimates approach that combines the strengths of machine learning with the
obtained with CFD methods using the FVM solver. These comparisons foundational principles of physics. This synergy optimizes computa­
established a baseline performance of PINNs in replicating complex fluid tional resources and also opens new avenues for real-time and precise
dynamics scenarios. simulations, making it a valuable asset for modern engineering
Since one of the primary identified limitations of PINN applicability challenges.
in real-world scenarios has been the extensive training time required,
further test and novel training strategies have been proposed. First, the CRediT authorship contribution statement
impact of the number of points used have been assessed, showing that a
careful setup can provide a reduction of training times without strong Tommaso Botarelli: Visualization, Validation, Software, Method­
degradation in the performance. Moreover, we evaluated several alter­ ology, Investigation, Formal analysis, Data curation. Marco Fanfani:
native training strategies. (A) Fine-Tuning: This approach is beneficial Writing – review & editing, Writing – original draft, Visualization, Su­
when a well-trained model is available. It requires about half the time pervision, Software, Investigation, Formal analysis. Paolo Nesi: Writing
needed to train a model from scratch, achieving similar results. (B) – review & editing, Writing – original draft, Validation, Supervision,
Multi-Resolution Learning: This strategy is particularly useful for sce­ Resources, Project administration, Investigation, Funding acquisition,
narios where steady state results are more critical than transient ones. Formal analysis, Conceptualization. Lorenzo Pinelli: Writing – original
Simulation errors for the last time window have been comparable to full draft, Validation, Software, Formal analysis, Data curation.
training from scratch, indicating its effectiveness in reducing training
times. (C) Parameterized Learning: this method proves advantageous
when dealing with multiple similar geometries. A single trained model Declaration of competing interest
can handle a virtually infinite number of setups with comparable ac­
curacies to those obtained from models trained from scratch. Therefore, The authors declare the following financial interests/personal re­
our findings indicate that PINNs can significantly reduce computational lationships which may be considered as potential competing interests:
costs while maintaining accuracy levels comparable to traditional CFD Paolo Nesi reports financial support was provided by University of
methods. This positions PINNs as a viable solution for industrial appli­ Florence.
cations requiring rapid and accurate fluid dynamic simulations.
However, additional studies are necessary to assess the behavior of Acknowledgements
PINNs in three-dimensional geometries and evaluate other training
strategies to further expedite the training process and simulate longer This research has been funded by CAI4DSA actions (Collaborative
transient times. Additionally, different PDEs should be considered, such Explainable neuro-symbolic AI for Decision Support Assistant), of the
as for example heat equations governing the thermal transfer between FAIR national partnership project on foundations of artificial intelli­
the fluid and the pieces inside the channel. Finally, the development of gence, PE 1 PNRR (https://fondazione-fair.it/). Snap4City (https://
PINN-based solutions fully integrated into industrial processes would be www.snap4city.org/) is an open technology and framework of DISIT
required to effectively exploit such novel methods in real-world Lab (https://www.disit.org).

Appendix A. Multi-Layer Perceptron (MLP) and Modified Fourier Network (MFN) architectures

Hereafter a brief description of the MLP and MFN architectures is provided showing the difference between them. As widely known, a MLP is
composed by an input layer, an output layer, and a variable number of hidden layers. Moving from one layer to the next one, input variables are

18
T. Botarelli et al. Engineering Applications of Arti cial Intelligence 148 (2025) 110347

linearly combined using a weight matrix W and a bias vector b and then non-linearity is introduced by means of some activation function σ , as shown
in the following equation
( )
xk+1 = σ Wk xk + bk (A1)

where xk ∈ Rn+1 and xk+1 ∈ Rm+1 are the input vector of the k-th layer and the relative output vector, respectively, Wk ∈ Rm+1×n+1 is the weight matrix
and bk ∈ Rm+1 the bias vector of the k-th layer. A graphical representation of a generic layer of the MLP is shown in Fig. 17. Hidden layers are stacked
one after the other to obtain the whole network.

[ ]T
Fig. 17. A generic k layer of a MLP architecture. The input vector xk = xk0 , …, xki , …, xkn is multiplied with the weights wMN of the matrix Wk (with N = [0, n] and
[ ]T
M = [0, m]) and the resulting vector is added to the bias vector bk = bk0 , …, bkj , …, bkm . A non-linear activation function σ is used and the output vector xk+1 =
[ ]T
xk+1 k+1
0 , …, xj , …, xk+1
m is obtained.

The MFN architecture extends the MLP by introducing an input encoding layer F, and two transformation layers characterized by specific weight
matrices WT1 and WT2 , plus bias vectors bT1 and bT2 , according to the following equation
( ( )) ( )
xk+1 = 1 − σ Wk xk + bk ⊙ σ(WT1 ϕE + bT1 ) + σ Wk xk + bk ⊙ σ(WT2 ϕE + bT2 ) (A2)

where ⊙ are element-wise multiplications and ϕE is obtained from the encoding layer F as
[ ( ) ( )]T
ϕE = sin 2π f × x0 ; cos 2πf × x0 (A3)
k
In this case, all the Wk and b for each layer plus f ∈ Rnf ×n0 (where nf is the number of frequency sets and n0 the dimension of the input vector x0 ),
{WT1 , bT1 }, and {WT2 , bT2 } are trainable parameters. In Fig. 18, a schematic representation of the MFN is provided.

( )
Fig. 18. Schematic representation of the MFN architecture. The block σ Wk xk +bk is used to represent a MLP layer as shown in Fig. 15. From the input vector x0 the
frequency set ϕE is obtained using the encoding layer F. Then, ϕE is used as input for the two transformation layers and the results are element-wise multiplied with
the outputs of the MLP layers according to equation (A2).

Appendix B. Experimental evidence of the relevance of multiple time windows in PINN training

In this section, we are providing the empirical evidence of the advantages of exploiting multiple time windows in PINN training with respect to the
adoption of a single window for the whole time-span, hereafter experimental results are reported. For this experiment, the single box piece has been
chosen and an MFN with 6 layers of 256 neurons and GELU adaptive activation has been selected, being the configuration that has obtained the best
trade-off between performance and training times (as reported in Sections 6.2, 6.3, and 6.4), excluding the advanced training strategies proposed in
Section 7.
For the single time window test, namely SingleTimeWindow, 300000 epochs have been used for a fair comparison, since it has been the total number
of epochs used in the MultiTimeWindows experiment for the same time span of [0,3]. Results on single window have been compared with respect to the
training reported in Section 6.4 and Tables 4 and 5 using 6 layers with 256 neurons, named hereafter as MultiTimeWindows.

19
T. Botarelli et al. Engineering Applications of Arti cial Intelligence 148 (2025) 110347

As it can be seen in Table 13 and Fig. 19 reporting the error, the MultiTimeWindows produced better results in terms of velocity with respect to the
SingleTimeWindow. The computational effort in both cases depends on the number of points and the number of epochs, which is identical. The slightly
longer times for the MultiTimeWindows is due to the overhead for restarting the model when a new time window is started. According to Fig. 19, the
SingleTimeWindow presents a quite higher error value with respect to the MultiTimeWindows approach. This is mainly due to the fact that (i) the
physical phenomenon presents quite different behaviour in the first part of the time-span with respect to the rest in which the stationary status arrives
(see videos at https://www.snap4city.org/1010 looking at the OpenFoam case of the single box), and (ii) the approach based on SingleTimeWindow
tries to find a unique model for all the time-span, while having three different trainings as in the MultiTimeWindows would produce much more precise
estimations.
Moreover, observing Fig. 19, it can be noted that around t’ = 1s a relevant reduction of the error is present for both solutions. In this case, two
effects have been appreciated. In SingleTimeWindow the effect is only associated with specific flow dynamics, which at that time the flow hitting the
front of the piece inside the channel reaches its maximum speed and then it stabilizes, making easier the fitting with the stationary model which is
dominance in the timespan (see the above-mentioned videos). In the MultiTimeWindows, the size of the first-time window has been experimentally
aligned with the change of flow behaviour (discussed above), and thus the relevant change is also due to the passage to the new time window (as also
occurred at t’ = 2s, for the third window). Then, when the flow reaches a stable condition in all the channel points, the error is continuously reduced
until setting on an almost fixed value after 2.5s (the time we selected to define the steady-state condition, see (24)).
This experimental analysis, better motivate and corroborates the relevance of adopting the MultiTimeWindows solutions in PINN training, con­
firming the literature (Wight and Zhao, 2020; Krishnapriyan et al., 2021).

Table 13
MAEavg and MAEst considering estimates for the velocity magnitude (‖u‖), horizontal (ux ) and vertical (uy ) components of the velocity, obtained for the
MultiTimeWindows, and SingleTimeWindow presented in Appendix B. In the last column, the training times. In bold the best values.

Time Window MAEavg MAEst Training Times (hh:mm)

‖u‖ ux uy ‖u‖ ux uy

MultiTimeWindows 0.019 0.019 0.010 0.010 0.010 0.005 04:25

SingleTimeWindow 0.037 0.035 0.017 0.018 0.017 0.010 04:00

Fig. 19. Plots of MAE(tʹ) of velocity magnitude ‖u‖ computed with respect to the OpenFOAM solution for both the MultiTimeWindows (violet) and SingleTimeWindow
(red) configurations discussed in Appendix B. The MultiTimeWindows approach outperforms the SingleTimeWindow case on estimating ‖u‖ for the whole timespan.
(Best viewed in color).

Data availability Cao, Y., Chandrasekar, A., Radhika, T., Vijayakumar, V., 2024b. Input-to-state stability of
stochastic Markovian jump genetic regulatory networks. Math. Comput. Simulat.
222, 174–187.
No data was used for the research described in the article. Chuang, P.Y., Barba, L.A., 2022. Experience Report of Physics-Informed Neural Networks
in Fluid Simulations: Pitfalls and Frustration arXiv preprint arXiv:2205.14249.
Clyne, T.W., Hull, D., 2019. An Introduction to Composite Materials. Cambridge
References
university press.
De Ryck, T., Jagtap, A.D., Mishra, S., 2024. Error estimates for physics-informed neural
Barzegar Gerdroodbary, M., 2020. Application of neural network on heat transfer networks approximating the Navier–Stokes equations. IMA J. Numer. Anal. 44 (1),
enhancement of magnetohydrodynamic nanofluid. Heat Tran. Res. 49, 197–212. 83–119.
https://doi.org/10.1002/htj.21606. Djeddi, S.R., Masoudi, A., Ghadimi, P., 2013. Numerical simulation of flow around
Baydin, A.G., Pearlmutter, B.A., Radul, A.A., Siskind, J.M., 2018. Automatic diamond-shaped obstacles at low to moderate Reynolds numbers. Am. J. Appl. Math.
differentiation in machine learning: a survey. J. Mach. Learn. Res. 18, 1–43. Stat. 1 (1), 11–20.
Bischof, R., Kraus, M., 2021. Multi-objective loss balancing for physics-informed deep Dung, D.V., Song, N.D., Palar, P.S., Zuhal, L.R., 2023. On the Choice of activation
learning. arXiv preprint arXiv:2110.09813. functions in physics-informed neural network for solving incompressible fluid flows.
Cao, Y., Subhashri, A.R., Chandrasekar, A., Radhika, T., Przybyszewski, K., 2024a. In: AIAA SCITECH 2023 Forum, p. 1803.
Exponential state estimation for delayed competitive neural network via stochastic Eivazi, H., Tahani, M., Schlatter, P., Vinuesa, R., 2022. Physics-informed neural networks
sampled-data control with Markov jump parameters under actuator failure. J. Artif. for solving Reynolds-averaged Navier–Stokes equations. Phys. Fluids 34 (7).
Intell. Soft Comput. Res. 14 (4), 373–385. Faller, William E., Schreck, Scott J., 1997. Unsteady fluid mechanics applications of
neural networks. J. Aircraft 34 (1), 48–55.

20
T. Botarelli et al. Engineering Applications of Arti cial Intelligence 148 (2025) 110347

Fernlund, G., Mobuchon, C., Zobeiry, N., 2018. 2.3 autoclave processing. In: Carl NVIDIA Modulus lib., NVIDIA Corporation, “NVIDIA Modulus”. Available: https://develo
Zweben, P.B. (Ed.), Comprehensive Composite Materials II, pp. 42–62. https://doi. per.nvidia.com/modulus (Accessed on May 3, 2024).
org/10.1016/b978-0-12-803581-8.09899-4. OpenFoam soft., OpenCFD, “OpenFoam”. [computer software]. Available: https://www.
Ferreira, De Paula, William, Armellini, Fabiano, Santa-Eulalia, Luis Antonio De, 2020. openfoam.com/(Accessed on May 3, 2024).
Simulation in industry 4.0: a state-of-the-art review. Comput. Ind. Eng. 149, 106868. Parish, Eric J., Duraisamy, Karthik, 2016. A paradigm for data-driven predictive
https://doi.org/10.1016/j.cie.2020.106868. ISSN 0360-8352. modeling using field inversion and machine learning. J. Comput. Phys. 305,
Fezai, S., Ben-Cheikh, N., Ben-Beya, B., Lili, T., 2020. Numerical study of obstacle 758–774. https://doi.org/10.1016/j.jcp.2015.11.012. ISSN 0021-9991. https://
geometry effect on the vortex shedding suppression and aerodynamic characteristics. www.sciencedirect.com/science/article/pii/S0021999115007524.
Int. J. Numer. Methods Heat Fluid Flow 30 (2), 469–495. Pinelli, L., Marconcini, M., Pacciani, R., Bake, F., Knobloch, K., Gaetani, P., Persico, G.,
Flores, F., Garreaud, R., Muñoz, R.C., 2013. CFD simulations of turbulent buoyant 2022. Effect of clocking on entropy noise generation within an aeronautical high
atmospheric flows over complex geometry: solver development in OpenFOAM. pressure turbine stage. J. Sound Vib. 529, 116900.
Computers & Fluids 82, 1–13. Rahaman, Nasim, Baratin, Aristide, Arpit, Devansh, Draxler, Felix, Lin, Min,
Ghamlouch, T., Roux, S., Bailleul, J.L., Lefèvre, N., Sobotka, V., 2017. Experiments and Hamprecht, Fred A., Bengio, Yoshua, Courville, Aaron, 2019. On the spectral bias of
numerical simulations of flow field and heat transfer coefficients inside an autoclave neural networks. ICML.
model. In: AIP Conference Proceedings, vol. 1896. AIP Publishing, 1. Raissi, M., Perdikaris, P., Karniadakis, G.E., 2019. Physics-informed neural networks: a
Giaccherini, S., Pinelli, L., Marconcini, M., Pacciani, R., Arnone, A., 2023. Validation of deep learning framework for solving forward and inverse problems involving
an analytical model for the acoustic Impedance Eduction of Multicavity Resonant nonlinear partial differential equations. J. Comput. Phys.
liners by a high-fidelity large Eddy simulation approach. ASME Journal of SeqSolver lib., NVIDIA Corporation. NVIDIA Modulus sequential solver. Available:
Turbomachinery 145 (8), 081002. https://doi.org/10.1115/1.4056984. https://github.
Hendrycks, Dan, Gimpel, Kevin, 2016. Gaussian error linear units (gelus). arXiv Preprint com/NVIDIA/modulus-sym/blob/main/modulus/sym/solver/sequential.py#L30.
arXiv:1606.08415. (Accessed 29 July 2024).
Hirsch, C., 2007. Numerical Computation of Internal and External Flows: the Sun, L., Gao, H., Pan, S., Wang, J.X., 2020. Surrogate modeling for fluid flows based on
Fundamentals of Computational Fluid Dynamics, second ed. Butterworth- physics-constrained deep learning without simulation data. Comput. Methods Appl.
Heinemann. Mech. Eng. 361, 112732.
Hu, B., McDaniel, D., 2023. Applying physics-informed neural networks to solve Szabó, B., Babuška, I., 2021. Finite Element Analysis: Method, Verification and
Navier–Stokes equations for laminar flow around a particle. Math. Comput. Appl. 28 Validation.
(5), 102. Tamil Thendral, M., Ganesh Babu, T.R., Chandrasekar, A., Cao, Y., 2022.
Jagtap, A.D., Kawaguchi, K., Karniadakis, G.E., 2020. Adaptive activation functions Synchronization of Markovian jump neural networks for sampled data control
accelerate convergence in deep and physics-informed neural networks. J. Comput. systems with additive delay components: analysis of image encryption technique.
Phys. 404, 109136. Math. Methods Appl. Sci.
Jin, X., Cai, S., Li, H., Karniadakis, G.E., 2021. NSFnets (Navier-Stokes flow nets): Tancik, M., Srinivasan, P., Mildenhall, B., Fridovich-Keil, S., Raghavan, N., Singhal, U.,
physics-informed neural networks for the incompressible Navier-Stokes equations. Ramamoorthi, R., Barron, J., Ng, R., 2020. Fourier features let networks learn high
J. Comput. Phys. 426, 109951. frequency functions in low dimensional domains. Adv. Neural Inf. Process. Syst. 33,
Krishnapriyan, A., Gholami, A., Zhe, S., Kirby, R., Mahoney, M.W., 2021. Characterizing 7537–7547.
possible failure modes in physics-informed neural networks. Adv. Neural Inf. Wang, J.X., Wu, J.L., Xiao, H., 2017. Physics-informed machine learning approach for
Process. Syst. 34, 26548–26560. reconstructing Reynolds stress modeling discrepancies based on DNS data. Phys.
Lawal, Z.K., Yassin, H., Lai, D.T.C., Che Idris, A., 2022. Physics-informed neural network Rev. Fluids 2, 034603. https://doi.org/10.1103/PhysRevFluids.2.034603.
(PINN) evolution and beyond: a systematic literature review and bibliometric Wang, S., Sankaran, S., Wang, H., Perdikaris, P., 2023. An Expert’s Guide to Training
analysis. Big Data and Cognitive Computing 6 (4), 140. Physics-Informed Neural Networks arXiv preprint arXiv:2308.08468.
Linse, Dennis J., Stengel, Robert F., 1993. Identification of aerodynamic coefficients Wang, H., Lu, L., Song, S., Huang, G., 2023. Learning Specialized Activation Functions
using computational neural networks. J. Guid. Control Dynam. 16 (6), 1018–1025. for Physics-Informed Neural Networks arXiv preprint arXiv:2308.04073.
Meng, X., Li, Z., Zhang, D., Karniadakis, G.E., 2020. PPINN: Parareal physics-informed Wight, C.L., Zhao, J., 2020. Solving Allen-Cahn and Cahn-Hilliard Equations Using the
neural network for time-dependent PDEs. Comput. Methods Appl. Mech. Eng. 370, Adaptive Physics Informed Neural Networks arXiv preprint arXiv:2007.04542.
113250. Xiang, Z., Peng, W., Zheng, X., Zhao, X., Yao, W., 2021. Self-adaptive loss balanced
MFN Lib., NVIDIA Corporation. Modified Fourier network. Available: https://docs. Physics-informed neural networks for the incompressible Navier-Stokes equations.
nvidia.com/deeplearning/modulus/modulus-v2209/user_guide/theory/architecture arXiv preprint arXiv:2104.06217.
s.html#modified-fourier-network. (Accessed 3 May 2024). Xu, S., Sun, Z., Huang, R., Guo, D., Yang, G., Ju, S., 2023. A practical approach to flow
Naderibeni, M., Reinders, M.J., Wu, L., Tax, D.M., 2024. Learning Solutions of field reconstruction with sparse or incomplete data through physics informed neural
Parametric Navier-Stokes with Physics-Informed Neural Networks arXiv preprint network. Acta Mech. Sin. 39 (3), 322302.
arXiv:2402.03153. Zienkiewicz, Olek C., Taylor, Robert L., Zhu, Jian Z., 2005. The Finite Element Method:
its Basis and Fundamentals. Elsevier.

21