A Review of Physics-Informed Machine Learning in F
A Review of Physics-Informed Machine Learning in F
Review
A Review of Physics-Informed Machine Learning in
Fluid Mechanics
Pushan Sharma 1 , Wai Tong Chung 1 , Bassem Akoush 1 and Matthias Ihme 1,2, *
Keywords: physics-informed machine learning; PDE-preserved learning; deep neural network; fluid
mechanics; Navier–Stokes
1. Introduction
1.1. Background and Motivation
In the last few decades, computational fluid dynamics (CFD) of compressible and
Citation: Sharma, P.; Chung, WT.; incompressible fluid flows has progressed significantly through finite difference, finite
Akoush, B.; Ihme, M. A Review of volume, finite elements and spectral methods. With the increasing availability of high-
Physics-Informed Machine Learning performance computational resources, we can now simulate complex turbulent flows
in Fluid Mechanics. Energies 2023, 16, at increasingly higher spatial and temporal resolutions. Despite this progress, several
2343. https://doi.org/10.3390/ challenges still persist for conventional numerical and analytic approaches. For exam-
en16052343 ple, solving inverse problems (e.g., for unknown boundary conditions or experimental
Academic Editors: Michał Jasiński,
parameters) is still prohibitively expensive. More importantly, numerical approaches typi-
Zbigniew Leonowicz and Surender cally simulate configurations under idealized conditions that do not account for realistic
Reddy Salkuti processes such as missing or noisy boundary conditions [1]. Moreover, direct numerical
simulations (DNS) of many practical turbulent systems are still unfeasible due to the com-
Received: 12 January 2023 putational complexity of resolving all spatial scales in multi-physical processes. This is
Revised: 7 February 2023
especially true for complex flows involving phase transitions or chemical reactions, which
Accepted: 24 February 2023
may require solving conservation equations for hundreds of species, introducing additional
Published: 28 February 2023
complexity challenges that arise from dimensionality and stiffness.
Machine learning (ML) and data-driven techniques have been increasingly popular in
scientific and engineering fields, offering a paradigm shift that can help address these chal-
Copyright: © 2023 by the authors.
lenges. Recently, with improved hardware, such as graphic processing units (GPUs) and
Licensee MDPI, Basel, Switzerland. tensor processing units (TPUs), data storage technologies, and access to a plethora of meth-
This article is an open access article ods through open-source libraries, ML offers new opportunities for investigating modeling
distributed under the terms and and predicting fluid flows. ML can be employed to supplement incomplete domain-specific
conditions of the Creative Commons knowledge in conventional experimental or numerical configuration by (i) exploring large
Attribution (CC BY) license (https:// design spaces, (ii) identifying hidden patterns, features or multi-dimensional correlations,
creativecommons.org/licenses/by/ and (iii) managing ill-posed problems. As such, ML has become a popular tool for studying
4.0/). fluid flows, especially when combined with the existing domain knowledge and intellectual
traditions that arise from the study of physical systems. To this end, this paper aims to
review and discuss the intersection of existing physical knowledge with new methods
offered by ML toward solving various problems tied to fluid flows.
Machine Learning
Among the most popular algorithms in supervised learning is the neural network
(NN), which is also the centerpiece of many deep learning techniques. Given the over-
whelming focus across the numerous fields on NNs and deep learning techniques, this
review will focus largely on physics-informed supervised and deep learning techniques.
In its most basic form, NNs are multi-layer perceptrons (MLPs) consisting of multiple
layers, and they generate predictions by forward propagating the outputs of each layer to the
next layer. For example, a one-layer MLP makes a prediction ŷ by first evaluating a linear
function z from inputs X, weights W, and biases b:
challenging problems. Secondly, a deep learning model with 152 layers was demon-
strated by He et al. [9] to outperform human performance in the ImageNet [10] image
recognition challenge. This work demonstrated that the combination of deep multi-layer
architectures can be combined with powerful GPUs and massive datasets to generate highly
accurate predictions.
With the increasing interest in ML methods, data-driven techniques have been applied
extensively to a wide range of fluid mechanics and turbulence problems. For example,
early works [11,12] focused on examining the feasibility of replacing conventional algebraic
turbulence models with ML models. While vanilla (canonical or baseline) ML models
possess the necessary expressiveness to model nonlinear turbulent behavior, these studies
identified the difficulty in accessing sufficient data as a major bottleneck for ML models to
generalize well outside of the training data. As such, many vanilla ML models, which rely
solely on the abundance of data, are currently not suited for many tasks in fluid mechanics
problems. Hence, the research community has developed interest in an alternate paradigm,
which combines existing domain knowledge with ML techniques, which is known as
physics-informed ML (PIML).
1.4. Outline
In this work, we review PIML applied to fluid flows. Given the popularity of ML in
contemporary research, numerous reviews on applying ML to science and engineering
Energies 2023, 16, 2343 4 of 21
can be found from various perspectives including fluid dynamics [45,57–60], combus-
tion [61–63], environmental engineering [64,65], ordinary/partial-differential equations
(ODEs/PDEs) [66] and specific PIML approaches [1,67]. In order to distinguish this review
from existing literature, we aim to provide information and perspectives, specifically for
addressing unique challenges and applications offered by employing PIML methods to
fluid mechanics problems. After this section, we review and discuss previous PIML studies
in Section 2. This is followed by a case study comparing PIML with vanilla ML models in
Section 3. The open questions and opportunities in developing PIML models for fluid flow
problems are addressed in Section 4 before presenting the concluding remarks in Section 5.
Physics-informed data
Physics-informed
Physics-informed loss architecture
Predictions
Figure 2. Different methods for embedding physics domain knowledge into a supervised learning
framework.
the use of the divergence of the Reynolds stress tensor, i.e., the Reynolds stress vector, as
a label, instead of the Reynolds stress tensor. This is based on the observation that errors
in the mean velocity arise when inserting the Reynolds stress tensor (extracted from DNS
databases) into the mean momentum balance equations. As shown in Figure 3, when
demonstrated on a turbulent (Re = 3200) square duct flow, mean velocity fields constructed
from the Reynolds stress vector were found to be significantly more accurate than fields
constructed from the Reynolds stress tensor.
Reynolds Reynolds
stress vector stress tensor
RANS DNS approach approach
Figure 3. Physics -informed labels matter: comparison of mean velocity flow fields from two different
ML labeling approaches, with RANS and DNS. Adapted from [76]. Copyright 2019 with permission
from Elsevier.
When dealing with ML features, two competing paradigms can be considered: (i) learn-
ing with hand-engineered features, and (ii) representation learning [77], which seeks to
develop ML models that can perform well with raw unprocessed data. Physics-informed
features belong to the former category. Early ML work on fluid mechanics utilized con-
ventional feature selection techniques without applying domain knowledge. For example,
Duraisamy et al. [12] employed the hill-climbing algorithm [78], which works by increasing
the size of the feature set until the ML accuracy stops increasing, in an inverse modeling
problem for identifying the analytic forms of closure models. Dimensionality reduction
techniques such as proper orthogonal decomposition (POD) [79], a popular method in
analyzing turbulent flows [80], have also been applied as feature extraction techniques for
ML models. For example, Lui and Wolf [81] demonstrated the use of POD for reducing
the dimensionality of turbulent airfoil data before training an NN for a reduced-order
modeling application.
One of the first works on embedding physical knowledge into ML involves a feature
engineering technique by Ling et al. [47]. In this work, the authors embedded rotational
invariance and non-dimensionality into features for predicting Reynolds stress anisotropy.
This approach showed improvements in predictive accuracy in random forests after fea-
ture processing, while the use of untransformed data in NNs resulted in more accurate
predictions. Due to the mixed effectiveness these techniques, physics-informed feature
Turbulence modelling using deep neural networks 159
Energies 2023, 16, 2343 6 of 21
(b)
techniques are typically applied on a case-by-case basis, and they can benefit from bench-
marking, as will be discussed in Section 4. Other concepts can also be embedded into the
ML model via feature transformation. For example, Xie et al. [82] transformed the spatial
(a) coordinates of features and labels to match coordinates orthogonal to curved surfaces, and
they demonstrated an 18% reduction in error when modeling turbulent closure.
Any tensor b which satisfies this condition will automatically satisfy Galilean
invariance. There are only a finite number of tensors because by the Caley–Hamilton
2 0. 62 . :2
theory, higher order products of these two tensors can be reduced to a linear
combination of this tensor basis. The five tensor invariants 1 , . . . , 5 are known
scalar functions of the elements of S and R . Pope (1975) gave a detailed derivation
of these 10 tensors, T (1) , . . . , T (10) and 5 invariants 1 , . . . , 5 , which are listed below:
5
9
2:1. 76 26.
2
T (1) = S T (6) = R 2 S + SR 2 3
I · Tr(SR 2 ) >
>
>
T = SR RS (2)
T = RSR R SR (7) 2 2 >
>
er =
Hidden layers Output layer
T (3)
= S 2 1
I · Tr( S 2
) T = 2
Hidden layers RS
(8)
SRS S 2
(2.3)
3 >
2 1 2 2 2 2 2 2 2 2 >
>
T =R
(4)
3
I · Tr(R ) T = R S + S R
(9)
3
I · Tr(S R )>>
2 2 2 2 2 2
;
T (5) = RS S R T (10)
= RS R R S R
GURE 1. (Colour online) Schematic of neural
Figure 4. Tensor network architectures.
Basis neural network architecture. Adapted from [48]. Copyright 2016 with
/5
Frezat et al. [86] enforced as well as Galilean, rotational and translational invari-
1
refore, if the input to the neural network is a rotated velocity gradient The resulting ML model demonstrated
ances, through modifications to a CNN architecture.
output should be the corresponding lowerrotated anisotropy
errors than tensor.
vanilla CNN This Wang
models. can beet al. [87] proposed the equivariant NN that
y constructing an integrity basis introduced
of the input tensors.into
symmetries Further
the MLexplanation
model predictions. Liu et al. [36] introduced a PDE-
bases for isotropic functions canpreserving
be found NNin Smitharchitecture
(PPNN) (1965). that demonstrated excellent stability for spatiotemporal
pecific case of interest here, withpredictions.
input tensors S andworks
The PPNN R , Pope (1975)residual
by forming has connections between input velocities
downsampled on the right-hand side (RHS) of the Navier–Stokes equations to result in
derived the relevant integrity basis. Pope proved that in the most general
more stable predictions than a vanilla CNN model. This approach will also be demon-
ble case, an eddy viscosity model that is a function of only S and R can
strated in a case study in Section 3. Wang et al. [88] proposed a TurbulentFlowNet (TFNet)
d as a linear combination of 10 architecture
isotropic basis tensors:
consisting of three encoder layers that received three differently filtered input
flow fields to account for the multi-scale nature of turbulent flows. The resulting archi-
10
X tecture outperformed standard CNN architectures in a 2D turbulent flow configuration.
b= g(n) ( 1 , .Many
. . , 5of (n)
)Tthese . architectures have been tested(2.2)on different configurations. Thus, while
n=1 these studies do demonstrate that PIML techniques can outperform vanilla ML methods,
it is still difficult to ascertain the benefits of one PIML approach over another without
b which satisfies this condition will automatically
benchmarking, satisfyin Section
as will be discussed Galilean4.
There are only a finite number of tensors because by the Caley–Hamilton
her order products of these two tensors can be reduced to a linear
n of this tensor basis. The five tensor invariants 1 , . . . , 5 are known
ions of the elements of S and R . Pope (1975) gave a detailed derivation
(1) (10)
Energies 2023, 16, 2343 7 of 21
Fourier Inverse
Filter
Transform Fourier
Output
X W + 𝝈 𝑦"
(a) The full architecture
Input of neural operator: start from input a. 1. Lift to a higher Activation
Weights dimension channel space
by a neural network P . 2. Apply four layers of integral operators and activation functions. Function3. Project back to
the target dimension by a neural network Q. Output u. (b) Fourier layers: Start from input v. On top: apply
the Fourier transform F ; a linearbehind
Figure 5. Mechanism transform R on Neural
the Fourier the lower Fourier
Operator. modesfrom
Adapted and [89].
filters out the higher modes;
then apply the inverse Fourier transform
2.3. Physics-Informed Loss Functions F 1
. On the bottom: apply a local linear transform W .
top: development
FigureA2:major The architecture of the
in PIML, neural
which operators;
deserves bottom:
extensive Fourier
discussion layer.
due to its prolif-
eration and influence, is the Physics-Informed Neural Network (PINNs), which was first
applied to solving a 1D Burgers equation and later applied to solving the Navier–Stokes
(PINNs) (Raissi et al.,in2019)
equations aimover
a 2D flow at the latter and
a cylinder [25].can
For therefore
the problembeinvolving
computationally
the Burgersexpensive.
equation, This
the PINN could generate spatiotemporal predictions with a O(10−4 ) mean-squared-error
makes them impractical for applications where a solution to the PDE is required for many different
solely with initial and boundary condition data. As with vanilla neural networks, the
instances of the parameter. On the other hand, our approach directly approximates the operator and
predictive accuracy of PINN was still found to be dependent on the model’s hyperparam-
is therefore much cheaper and faster, offering tremendous computational savings when compared to
eters, and it increased with the amount of training data provided. In the 2D flow over
traditional solvers.
cylinderFor an example
problem application
[25], PINNs to Bayesiantoinverse
were demonstrated problems,
predict the pressuresee
andSection
viscosity5.5.
of
the Navier–Stokes equations in an inverse modeling problem. This work differed from the
Discretization. Since
previous our (which
studies data ajfocused
and uj onare, in general,
model functions,
input, output, to work with
and architecture) bythem numerically,
introducing
we assume access only toinpoint-wise
soft-constraints evaluations.
the loss function L: Let Dj = {x1 , . . . , xn } ⇢ D be a n-point
discretization of the domain D and assume we N have observations aj |Dj 2 R
n⇥da
, uj |Dj 2 Rn⇥dv ,
for a finite collection of input-outputarg
pairs
minindexed
∑ (Ldataby To
(yi )j.+ µ Lkbe
(yi ,discretization-invariant,
θ )), the
(2) neural
operator can produce an answer u(x) for any x 2 D, potentially x 2
θ i = 1 / Dj . Such a property is highly
desirable as it allows a transfer of solutions between different grid geometries and discretizations.
where N is the number of training samples, µ is a tunable weighting parameter, and L k
contains the k-th constraint, which is related to the partial differential equations (LPDE ), initial
3 N EURAL O PERATOR
conditions, and boundary conditions. For an example involving the 1D Burgers equation:
∂u ∂u ∂2 u
The neural operator, proposed in (Li et al., +
2020b), =
∂xis formulated
as an iterative architecture
u ν (3) v 7!
∂t ∂x2 0
v1 7! . . . 7! vT where vj for j = 0, 1, . . . , T 1 is a sequence of functions each taking values in
Rdv . As shown in Figure 2 (a), the input a 2 A is first lifted to a higher dimensional representation
v0 (x) = P (a(x)) by the local transformation P which is usually parameterized by a shallow fully-
connected neural network. Then we apply several iterations of updates vt 7! vt+1 (defined below).
The output u(x) = Q(vT (x)) is the projection of vT by the local transformation Q : Rdv ! Rdu .
In each iteration, the update vt 7! vt+1 is defined as the composition of a non-local integral operator
K and a local, nonlinear activation function .
the loss function in Equation (2) for predicting velocity u with N number of training samples
(and viscosity ν) may take the form:
1 N ∂û ∂û ∂2 û
N ∑ kû − uk2 +µ
| {z } ∂t
+ û − ν 2
∂x ∂x 2
(4)
i =1 | {z }
Ldata
LPDE
where û is the predicted velocity from PINN, the first term of Equation (4) is the velocity
mean-squared error (MSE), while the second term represents a residual term that constrains
the optimization problem in order to produce a solution that matches (3). The schematic of
the PINN architecture for the 1D Burgers equation is depicted in Figure 6.
Figure 6. Schematic of PINN algorithm. Reprinted from [1]. Copyright 2021 with permission from
Springer Nature Limited.
The PINN framework was later extended to specifically tackle several laminar and in-
compressible fluid flow problems. This included simulating vortex-induced vibrations [22]
and tackling ill-posed inverse fluid mechanics problems through a so-called hidden fluid
mechanics framework [26]. Later, the PINN framework was demonstrated to solve the incom-
pressible Navier–Stokes equations in turbulent flow conditions [27]. In this work, two PINN
loss functions were suggested based on two forms of the incompressible Navier–Stokes
equations: (i) one based on the velocity–pressure (VP) formulation and (ii) another based
on the velocity–vorticity (VV) formulation. Both PINN approaches were tested on a range
of laminar canonical cases, as well as a turbulent channel flow, with Re = O(104 ), from the
Johns Hopkins Turbulent Database [92]. Both forms of the PINN were found to be effective
in simulating laminar cases, with less than 1% velocity MSE. However, a ∼10% velocity
MSE was obtained when simulating the turbulent channel DNS with the VP-PINN, while
the VV-PINN failed to converge for the turbulent cases. Similar accuracies in both laminar
and turbulent cases have been reported in another study involving PINNs with RANS-
specific loss functions [29]. While PINNs currently struggle to learn predicting turbulent
configurations, Jin et al. [27] noted that PINNs were particularly useful for applications
outside typical capabilities of traditional CFD solvers, such as inverse modeling, where the
authors could predict the flow Reynolds number from a few samples of velocity data.
PINNs have been employed for inverse modeling in experimental fluid mechanics.
Cai et al. [28] proposed a PINN-based method to infer the full continuous 3D pressure
and velocity fields on an espresso cup from the snapshots of temperature fields obtained
from tomographic background-oriented Schlieren (Tomo-BOS) imaging. On the other
Energies 2023, 16, 2343 9 of 21
hand, Eivazi and Vinuesa [93] used noisy experimental measurements to obtain super-
resolution of flow field data in time and space using PINN.
PINNs have now influenced approaches in which loss functions are formulated in
the following non-exhaustive list of flow problems. Bode et al. [56] modeled the subgrid-
scale stresses by adding a continuity residual term g = k∇ · uk2 to the loss function
of a generative adversarial network (GAN) model in a turbulent reacting flow problem.
This model, known as physics-informed enhanced super-resolution GAN (PIESR-GAN),
optimizes the physics-informed loss, which ensures continuity, along with the adversarial
and accuracy loss terms. Jiang et al. [94] also relied on PDE residuals to accurately perform
super-resolution on a turbulent Rayleigh–Bernard problem. Sun et al. [95] added a loss term
for the subgrid-scale stresses to an MSE loss function for the eddy viscosity in a turbulent
airfoil problem. An entropy term was added by Guan et al. [96] in modeling turbulence
to 2D homogeneous isotropic turbulence configurations. Laubscher [35] demonstrated
significant improvements when predicting temperature, velocity, and species mass fractions
when using segregated-network PINN instead of a vanilla PINN on simulations of water
vapor in dry air flowing in a 2D rectangular duct.
Here, we use the PPNN approach to reconstruct the 2D flow fields of a lid-driven cavity
problem. Our objective is to assess the advantages of using PIML models over baseline
vanilla ML models in terms of long-term prediction and generalizability, especially for
unseen parameter spaces.
∂u 1 2
+ u · ∇u = −∇ p + ∇ u
∂t Re (5)
∇·u = 0
where u = [u( x, y, t), v( x, y, t)] T is the velocity vector, and p( x, y, t) is the pressure. The
length of the computational domain is unity in both directions. The velocity of the lid is
considered to be utop = 1. No-slip condition is enforced at the other three boundaries.
Neumann boundary conditions are applied for the pressure at all boundaries. Four different
Reynolds numbers are considered for this study: 400, 500, 600 and 750. In all cases, the
velocity and pressure fields are generated using the FD solver for 5000 numerical steps
with a time step of ∆t = 0.01. A uniform Cartesian mesh with 100 × 100 is used in the
numerical solver.
The training dataset involves 100 snapshots from only two Re, i.e., 500 and 750. These
snapshots are collected every 20 numerical steps to form the training trajectory, which
means the learning step is set as τ = 20∆t = 0.2. Thus, the training set covers a total time
of T = 100τ. The trained model is then used to predict the flow field for the other Re,
i.e., 400 (extrapolation) and 600 (interpolation), which are not part of the training set. The
PPNN-specific training and prediction parameters are described in the following sections.
Figure 7. (a) The conceptual schematic of PPNN algorithm. (b) The detailed implementation of
PPNN. Reprinted from [36].
where ut (λi ) is the velocity magnitude of the reference solution at time t for the physical
parameter, λi (which corresponds to the i-th Re in this case), and f θ represents the trained
NN update with weights θ̃. The predicted solution by PPNN at time t − ∆t is denoted by
ût−∆t and evaluated as
where n is the number of evolving steps starting from the initial high-resolution solution
field, u0 . The relative testing errors of the magnitude of velocity corresponding to the
prediction of PPNN, PPNN-partial and baseline ConvResNet model compared to the
ground truth reference are shown in Figure 8. Clearly, the prediction of the baseline
model suffers from error accumulation, which is a well-established problem within auto-
regressive models. However, the physics-informed ML model, PPNN, suppresses the
error accumulation and outperforms the baseline model by providing accurate predictions
of the velocity field even beyond the training period (t > T) for Re = 500 and 750. In
case of unseen conditions (Re = 400 and 600), PPNN shows very low roll-out errors
compared to the baseline model. Even for PPNN-partial cases for both training and test Re,
Energies 2023, 16, 2343 12 of 21
the preservation of partially known PDE structures outperforms the vanilla ConvResNet
model. Although there is an error for Re = 400 and 600, PPNN-partial exhibits a larger
error than fully known PDE preserved PPNN, the error never accumulates, and it provides
a stable and reasonably accurate solution.
3.0 3.0
2.5 Training Forecasting 2.5 Forecasting
(Unseen dataset)
2.0 Re = 500 (Baseline) 2.0 Re = 400 (Baseline)
Re = 500 (PPNN) Re = 400 (PPNN)
Re = 500 (PPNN-partial) Re = 400 (PPNN-partial)
1.5 Re = 750 (Baseline) 1.5 Re = 600 (Baseline)
t
t
To emphasize the advantage of PPNN even further, the contours of the velocity
field prediction from PPNN, PPNN-partial and the baseline model along with ground
truth are shown in Figure 9 for Re = 400. Both PPNN predictions emulate the ground
truth accurately which is consistent with the error plot (Figure 8). In contrast, despite
accurate predictions from the baseline model during the few initial timesteps, error rapidly
accumulates to provide random, noisy and unphysical solution field. Overall, PPNN (an
example of a PIML approach) demonstrates significantly better generalizability, stability
and robustness in terms of long-term prediction than the baseline (vanilla) ML model.
0.8
Baseline
0.6
0.4
PPNN
0.2
(partial)
PPNN
0.0
Figure 9. Predicted snapshots of magnitude of velocity obtained from the finite-difference solution
(Truth), baseline ConvResNet architecture (Baseline), the PPNN with fully known Navier–Stokes
equation, and the PPNN with partially known Navier–Stokes equation at Re = 400, which is not part
of the training set.
Energies 2023, 16, 2343 13 of 21
5. Conclusions
In this article, we have reviewed the physics-informed ML techniques for applications
within fluid flow problems. The general idea behind PIML models is to integrate domain
knowledge and/or the information about the governing PDEs with the deep NN models.
Incorporating these aspects with traditional ML algorithms have resulted in better and
more stable prediction accuracy, faster training and improved generalizability. Thus, the
successful embedding of physical knowledge, either as input features/labels or in the form
of a modified loss function or by preserving the PDE structures in the model architecture
itself, holds exciting promise for the advancement of solving fluid flows, especially complex
Energies 2023, 16, 2343 16 of 21
turbulent flows. Here, we have highlighted the applications of PIML in the field of fluid
mechanics, ranging from knowledge discovery, prediction of spatiotemporal dynamics in
different flow configurations, as well as super-resolution and turbulent closure modeling.
We also demonstrate the utility of a PIML algorithm through a standard fluid flow problem.
Fluid mechanics is a data-rich field that is founded on physical concepts encapsulating
conservation relations. Relying solely on the data-driven methods may omit these concepts,
as data-driven methods could ignore certain physical constraints or conservation principles.
Therefore, fluid mechanics problems can be an excellent testbed for developing novel PIML
algorithms. While PIML approaches have demonstrated significant progress in fluid flow
problems, several open questions and opportunities still exist. As each PIML model has
only been demonstrated on particular problem configurations and datasets, benchmarking
existing PIML methods is still an open challenge, which will need to be addressed as ML
within fluid mechanics matures. It is also important to keep track of the fast-moving and
mercurial developments of ML outside fluid mechanics. In these domains, new innovations
are introduced at a rapid pace, leaving many existing methods obsolete. Researchers should
take inspiration from these developments either to improve specfic components of existing
PIML models (such as optimization routines and loss functions) or to develop entirely new
algorithms for resolving existing challenges within flow problems. As PIML development
continues to progress, this set of approaches is expected to play a significant role in the
offering computationally efficient predictive models to fluid mechanics that could be more
reliable than their vanilla counterparts, especially in complex flow configurations.
Author Contributions: Conceptualization, P.S., W.T.C. and M.I.; software, P.S. and B.A.; resources,
M.I.; data curation, P.S. and B.A.; writing—original draft preparation, P.S. and W.T.C.; writing—review
and editing, M.I.; visualization, P.S., W.T.C., and B.A.; supervision, M.I.; project administration, M.I.;
funding acquisition, M.I. All authors have read and agreed to the published version of the manuscript.
Funding: This research was funded by DOE Basic Energy Science: DE-SC0022222 and NASA Early
Stage Innovation Program: 80NSSC22K0257. W.T.C is grateful for partial financial support from the
Stanford Institute for Human-centered Artificial Intelligence Graduate Fellowship.
Data Availability Statement: The data supporting this review are available from the corresponding
authors upon request.
Conflicts of Interest: The authors declare no conflict of interest.
Abbreviations
The following abbreviations are used in this manuscript:
ML Machine learning
PIML Physics-informed machine learning
NN Neural network
CFD Computational fluid dynamics
DNS Direct numerical simulation
GPU Graphics processing unit
TPU Tensor processing unit
MLP Multi-layer perceptron
CNN Convolutional neural network
RL Reinforcement learning
Re Reynolds number
PINN Physics-informed neural network
PPNN PDE-preserved neural network
LES Large-eddy simulation
RANS Reynolds average Navier-Stokes
ODE Ordinary differential equation
PDE Partial differential equation
Energies 2023, 16, 2343 17 of 21
References
1. Karniadakis, G.E.; Kevrekidis, I.G.; Lu, L.; Perdikaris, P.; Wang, S.; Yang, L. Physics-informed machine learning. Nat. Rev. Phys.
2021, 3, 422–440. [CrossRef]
2. Bishop, C.M. Pattern Recognition and Machine Learning; Springer: Berlin, Germany, 2006.
3. Nair, V.; Hinton, G.E. Rectified linear units improve restricted Boltzmann machines. Proc. Int. Conf. Mach. Learn. 2010,
27, 807–814.
4. Amari, S.i. Backpropagation and stochastic gradient descent method. Neurocomputing 1993, 5, 185–196. [CrossRef]
5. Rumelhart, D.E.; Hinton, G.E.; Williams, R.J. Learning representations by back-propagating errors. Nature 1986, 323, 533–536.
[CrossRef]
6. LeCun, Y.; Bottou, L.; Bengio, Y.; Haffner, P. Gradient-based learning applied to document recognition. Proc. IEEE 1998,
86, 2278–2324. [CrossRef]
7. Mnih, V.; Kavukcuoglu, K.; Silver, D.; Graves, A.; Antonoglou, I.; Wierstra, D.; Riedmiller, M. Playing Atari with deep
reinforcement learning. arXiv 2013, arXiv:1312.5602.
8. Mnih, V.; Kavukcuoglu, K.; Silver, D.; Rusu, A.A.; Veness, J.; Bellemare, M.G.; Graves, A.; Riedmiller, M.; Fidjeland, A.K.;
Ostrovski, G.; et al. Human-level control through deep reinforcement learning. Nature 2015, 518, 529–533.
9. He, K.; Zhang, X.; Ren, S.; Sun, J. Deep residual learning for image recognition. In Proceedings of the IEEE Conference on
Computer Vision and Pattern Recognition (CVPR), Las Vegas, NV, USA, 27–30 June 2016; pp. 770–778.
10. Russakovsky, O.; Deng, J.; Su, H.; Krause, J.; Satheesh, S.; Ma, S.; Huang, Z.; Karpathy, A.; Khosla, A.; Bernstein, M.; et al.
ImageNet large scale visual recognition challenge. Int. J. Comput. Vis. 2015, 115, 211–252.
11. Tracey, B.D.; Duraisamy, K.; Alonso, J.J. A machine learning strategy to assist turbulence model development. In Proceedings of
the 53rd AIAA Aerospace Sciences Meeting, Kissimmee, FL, USA, 5–9 January 2015; p. 1287.
12. Duraisamy, K.; Zhang, Z.J.; Singh, A.P. New approaches in turbulence and transition modeling using data-driven techniques. In
Proceedings of the 53rd AIAA Aerospace Sciences Meeting, Kissimmee, FL, USA, 5–9 January 2015; p. 1284.
13. Jolliffe, I.T.; Cadima, J. Principal component analysis: A review and recent developments. Philos. Trans. Royal Soc. A 2016,
374, 20150202. [CrossRef] [PubMed]
14. Jain, A.K. Data clustering: 50 years beyond K-means. Pattern Recognit. Lett. 2010, 31, 651–666. [CrossRef]
15. Reddy, G.; Celani, A.; Sejnowski, T.J.; Vergassola, M. Learning to soar in turbulent environments. Proc. Natl. Acad. Sci. USA 2016,
113, E4877–E4884. [CrossRef]
16. Novati, G.; de Laroussilhe, H.L.; Koumoutsakos, P. Automating turbulence modelling by multi-agent reinforcement learning.
Nat. Mach. Intell. 2021, 3, 87–96. [CrossRef]
17. Fukami, K.; Fukagata, K.; Taira, K. Super-resolution reconstruction of turbulent flows with machine learning. J. Fluid Mech. 2019,
870, 106–120. [CrossRef]
18. Wang, Z.; Chen, J.; Hoi, S.C. Deep learning for image super-resolution: A survey. IEEE Trans. Pattern Anal. Mach. Intell. 2020,
43, 3365–3387. [CrossRef] [PubMed]
19. Maulik, R.; San, O.; Jacob, J.D.; Crick, C. Sub-grid scale model classification and blending through deep learning. J. Fluid Mech.
2019, 870, 784–812. [CrossRef]
20. Chung, W.T.; Mishra, A.A.; Perakis, N.; Ihme, M. Data-assisted combustion simulations with dynamic submodel assignment
using random forests. Combust. Flame 2021, 227, 172–185. [CrossRef]
21. Li, B.; Yang, Z.; Zhang, X.; He, G.; Deng, B.Q.; Shen, L. Using machine learning to detect the turbulent region in flow past a
circular cylinder. J. Fluid Mech. 2020, 905, A10. [CrossRef]
22. Raissi, M.; Wang, Z.; Triantafyllou, M.S.; Karniadakis, G.E. Deep learning of vortex-induced vibrations. J. Fluid Mech. 2019,
861, 119–137. [CrossRef]
Energies 2023, 16, 2343 18 of 21
23. Callaham, J.L.; Rigas, G.; Loiseau, J.C.; Brunton, S.L. An empirical mean-field model of symmetry-breaking in a turbulent wake.
Sci. Adv. 2022, 8, eabm4786. [CrossRef]
24. Chung, W.T.; Mishra, A.A.; Ihme, M. Interpretable data-driven methods for subgrid-scale closure in LES for transcritical
LOX/GCH4 combustion. Combust. Flame 2022, 239, 111758. [CrossRef]
25. Raissi, M.; Perdikaris, P.; Karniadakis, G. Physics-informed neural networks: A deep learning framework for solving forward
and inverse problems involving nonlinear partial differential equations. J. Comp. Phys. 2019, 378, 686–707. [CrossRef]
26. Raissi, M.; Yazdani, A.; Karniadakis, G.E. Hidden fluid mechanics: Learning velocity and pressure fields from flow visualizations.
Science 2020, 367, 1026–1030. [CrossRef]
27. Jin, X.; Cai, S.; Li, H.; Karniadakis, G.E. NSFnets (Navier-Stokes flow nets): Physics-informed neural networks for the incompress-
ible Navier-Stokes equations. J. Comp. Phys. 2021, 426, 109951. [CrossRef]
28. Cai, S.; Wang, Z.; Fuest, F.; Jeon, Y.J.; Gray, C.; Karniadakis, G.E. Flow over an espresso cup: Inferring 3-D velocity and pressure
fields from tomographic background oriented Schlieren via physics-informed neural networks. J. Fluid Mech. 2021, 915, A102.
[CrossRef]
29. Eivazi, H.; Tahani, M.; Schlatter, P.; Vinuesa, R. Physics-informed neural networks for solving Reynolds-averaged Navier–Stokes
equations. Phys. Fluids 2022, 34, 075117. [CrossRef]
30. Wang, H.; Liu, Y.; Wang, S. Dense velocity reconstruction from particle image velocimetry/particle tracking velocimetry using a
physics-informed neural network. Phys. Fluids 2022, 34, 017116. [CrossRef]
31. Qiu, R.; Huang, R.; Xiao, Y.; Wang, J.; Zhang, Z.; Yue, J.; Zeng, Z.; Wang, Y. Physics-informed neural networks for phase-field
method in two-phase flow. Phys. Fluids 2022, 34, 052109. [CrossRef]
32. Aliakbari, M.; Mahmoudi, M.; Vadasz, P.; Arzani, A. Predicting high-fidelity multiphysics data from low-fidelity fluid flow and
transport solvers using physics-informed neural networks. Int. J. Heat Fluid Flow 2022, 96, 109002. [CrossRef]
33. Ji, W.; Qiu, W.; Shi, Z.; Pan, S.; Deng, S. Stiff-PINN: Physics-informed neural network for stiff chemical kinetics. J. Phys. Chem. A
2021, 125, 8098–8106. [CrossRef] [PubMed]
34. Weng, Y.; Zhou, D. Multiscale physics-informed neural networks for stiff chemical kinetics. J. Phys. Chem. A 2022, 126, 8534–8543.
[CrossRef]
35. Laubscher, R. Simulation of multi-species flow and heat transfer using physics-informed neural networks. Phys. Fluids 2021,
33, 087101. [CrossRef]
36. Liu, X.Y.; Sun, H.; Zhu, M.; Lu, L.; Wang, J.X. Predicting parametric spatiotemporal dynamics by multi-resolution PDE
structure-preserved deep learning. arXiv 2022, arXiv:2205.03990.
37. Ren, P.; Rao, C.; Liu, Y.; Wang, J.X.; Sun, H. PhyCRNet: Physics-informed convolutional-recurrent network for solving
spatiotemporal PDEs. Comput. Methods Appl. Mech. Eng. 2022, 389, 114399. [CrossRef]
38. Kochkov, D.; Smith, J.A.; Alieva, A.; Wang, Q.; Brenner, M.P.; Hoyer, S. Machine learning-accelerated computational fluid
dynamics. Proc. Natl. Acad. Sci. USA 2021, 118, e2101784118. [CrossRef] [PubMed]
39. De Avila Belbute-Peres, F.; Economon, T.; Kolter, Z. Combining differentiable PDE solvers and graph neural networks for fluid
flow prediction. In Proceedings of the 37th International Conference on Machine Learning, Virtual, 13–18 July 2020; pp. 2402–2411.
40. Um, K.; Brand, R.; Fei, Y.R.; Holl, P.; Thuerey, N. Solver-in-the-loop: Learning from differentiable physics to interact with iterative
PDE-solvers. Adv. Neural Inf. Process. Syst. 2020, 33, 6111–6122.
41. Gao, H.; Sun, L.; Wang, J.X. PhyGeoNet: Physics-informed geometry-adaptive convolutional neural networks for solving
parameterized steady-state PDEs on irregular domain. J. Comp. Phys. 2021, 428, 110079. [CrossRef]
42. Geneva, N.; Zabaras, N. Modeling the dynamics of PDE systems with physics-constrained deep auto-regressive networks.
J. Comp. Phys. 2020, 403, 109056. [CrossRef]
43. Ranade, R.; Hill, C.; Pathak, J. DiscretizationNet: A machine-learning based solver for Navier–Stokes equations using finite
volume discretization. Comput. Methods Appl. Mech. Eng. 2021, 378, 113722. [CrossRef]
44. Rao, C.; Sun, H.; Liu, Y. Embedding physics to learn spatiotemporal dynamics from sparse data. arXiv 2021, arXiv:2106.04781.
45. Duraisamy, K.; Iaccarino, G.; Xiao, H. Turbulence Modeling in the Age of Data. Annu. Rev. Fluid Mech. 2019, 51, 357–377.
[CrossRef]
46. Ling, J.; Templeton, J. Evaluation of machine learning algorithms for prediction of regions of high Reynolds averaged Navier-
Stokes uncertainty. Phys. Fluids 2015, 27, 085103. [CrossRef]
47. Ling, J.; Jones, R.; Templeton, J. Machine learning strategies for systems with invariance properties. J. Comp. Phys. 2016, 318, 22–35.
[CrossRef]
48. Ling, J.; Kurzawski, A.; Templeton, J. Reynolds averaged turbulence modelling using deep neural networks with embedded
invariance. J. Fluid Mech. 2016, 807, 155–166. [CrossRef]
49. Parish, E.J.; Duraisamy, K. A paradigm for data-driven predictive modeling using field inversion and machine learning. J. Comp.
Phys. 2016, 305, 758–774. [CrossRef]
50. Xiao, H.; Wu, J.L.; Wang, J.X.; Sun, R.; Roy, C. Quantifying and reducing model-form uncertainties in Reynolds-averaged
Navier-Stokes simulations: A data-driven, physics-informed Bayesian approach. J. Comp. Phys. 2016, 324, 115–136. [CrossRef]
51. Singh, A.P.; Medida, S.; Duraisamy, K. Machine-learning-augmented predictive modeling of turbulent separated flows over
airfoils. AIAA J. 2017, 55, 2215–2227. [CrossRef]
Energies 2023, 16, 2343 19 of 21
52. Wang, J.X.; Wu, J.L.; Xiao, H. Physics-informed machine learning approach for reconstructing Reynolds stress modeling
discrepancies based on DNS data. Phys. Rev. Fluids 2017, 2, 034603. [CrossRef]
53. Lapeyre, C.J.; Misdariis, A.; Cazard, N.; Veynante, D.; Poinsot, T. Training convolutional neural networks to estimate turbulent
sub-grid scale reaction rates. Combust. Flame 2019, 203, 255–264. [CrossRef]
54. Beck, A.; Flad, D.; Munz, C.D. Deep neural networks for data-driven LES closure models. J. Comp. Phys. 2019, 398, 108910.
[CrossRef]
55. Maulik, R.; San, O. A neural network approach for the blind deconvolution of turbulent flows. J. Fluid Mech. 2017, 831, 151–181.
[CrossRef]
56. Bode, M.; Gauding, M.; Lian, Z.; Denker, D.; Davidovic, M.; Kleinheinz, K.; Jitsev, J.; Pitsch, H. Using physics-informed enhanced
super-resolution generative adversarial networks for subfilter modeling in turbulent reactive flows. Proc. Combust. Inst. 2021,
38, 2617–2625. [CrossRef]
57. Kutz, J.N. Deep learning in fluid dynamics. J. Fluid Mech. 2017, 814, 1–4. [CrossRef]
58. Brunton, S.L.; Noack, B.R.; Koumoutsakos, P. Machine learning for fluid mechanics. Annu. Rev. Fluid Mech. 2020, 52, 477–508.
[CrossRef]
59. Brunton, S.L. Applying machine learning to study fluid mechanics. Acta Mech. Sin. 2022, 37, 1–9. [CrossRef]
60. Zhu, L.T.; Chen, X.Z.; Ouyang, B.; Yan, W.C.; Lei, H.; Chen, Z.; Luo, Z.H. Review of machine learning for hydrodynamics,
transport, and reactions in multiphase flows and reactors. Ind. Eng. Chem. Res. 2022, 61, 9901–9949. [CrossRef]
61. Ihme, M.; Chung, W.T.; Mishra, A.A. Combustion machine learning: Principles, progress and prospects. Prog. Energy Combust.
Sci. 2022, 91, 101010. [CrossRef]
62. Zhou, L.; Song, Y.; Ji, W.; Wei, H. Machine learning for combustion. Energy AI 2022, 7, 100128. [CrossRef]
63. Echekki, T.; Farooq, A.; Ihme, M.; Sarathy, S.M. Machine Learning for Combustion Chemistry. In Machine Learning and Its
Application to Reacting Flows: ML and Combustion; Springer International Publishing: Cham, Switzerland, 2023; pp. 117–147.
64. Zhong, S.; Zhang, K.; Bagheri, M.; Burken, J.G.; Gu, A.; Li, B.; Ma, X.; Marrone, B.L.; Ren, Z.J.; Schrier, J.; et al. Machine learning:
New ideas and tools in environmental science and engineering. Environ. Sci. Technol. 2021, 55, 12741–12754. [CrossRef]
65. Willard, J.; Jia, X.; Xu, S.; Steinbach, M.; Kumar, V. Integrating scientific knowledge with machine learning for engineering and
environmental systems. ACM Comput. Surv. 2022, 55, 1–37. [CrossRef]
66. Hao, Z.; Liu, S.; Zhang, Y.; Ying, C.; Feng, Y.; Su, H.; Zhu, J. Physics-informed machine learning: A survey on problems, methods
and applications. arXiv 2022, arXiv:2211.08064.
67. Cai, S.; Mao, Z.; Wang, Z.; Yin, M.; Karniadakis, G.E. Physics-informed neural networks (PINNs) for fluid mechanics: A review.
Acta Mech. Sin. 2022, 37, 1–12. [CrossRef]
68. Wu, J.L.; Wang, J.X.; Xiao, H.; Ling, J. A priori assessment of prediction confidence for data-driven turbulence modeling. Flow
Turbul. Combust. 2017, 99, 25–46. [CrossRef]
69. Li, J.; Cheng, K.; Wang, S.; Morstatter, F.; Trevino, R.P.; Tang, J.; Liu, H. Feature selection: A data perspective. ACM Comput. Surv.
2017, 50, 1–45. [CrossRef]
70. Ding, S.; Zhu, H.; Jia, W.; Su, C. A survey on feature extraction for pattern recognition. Artif. Intell. Rev. 2012, 37, 169–180.
[CrossRef]
71. Germano, M.; Piomelli, U.; Moin, P.; Cabot, W.H. A dynamic subgrid-scale eddy viscosity model. Phys. Fluids A 1991, 3, 1760–1765.
[CrossRef]
72. Vreman, A.W. An eddy-viscosity subgrid-scale model for turbulent shear flow: Algebraic theory and applications. Phys. Fluids
2004, 16, 3670–3681. [CrossRef]
73. Beck, A.; Kurz, M. A perspective on machine learning methods in turbulence modeling. GAMM-Mitteilungen 2021, 44, e202100002.
[CrossRef]
74. Duraisamy, K. Perspectives on machine learning-augmented Reynolds-averaged and large eddy simulation models of turbulence.
Phys. Rev. Fluids 2021, 6, 050504. [CrossRef]
75. Arridge, S.; Maass, P.; Öktem, O.; Schönlieb, C.B. Solving inverse problems using data-driven models. Acta Numer. 2019, 28, 1–174.
[CrossRef]
76. Cruz, M.A.; Thompson, R.L.; Sampaio, L.E.; Bacchi, R.D. The use of the Reynolds force vector in a physics informed machine
learning approach for predictive turbulence modeling. Comput. Fluids 2019, 192, 104258. [CrossRef]
77. Bengio, Y.; Courville, A.; Vincent, P. Representation learning: A review and new perspectives. IEEE Trans. Pattern Anal. Mach.
Intell. 2013, 35, 1798–1828. [CrossRef]
78. Davis, L. Bit-climbing, representational bias, and test suit design. Proc. Int. Conf. Genetic Algorithm 1991, 18–23.
79. Berkooz, G.; Holmes, P.; Lumley, J.L. The proper orthogonal decomposition in the analysis of turbulent flows. Annu. Rev. Fluid
Mech. 1993, 25, 539–575. [CrossRef]
80. Meyer, K.E.; Pedersen, J.M.; Özcan, O. A turbulent jet in crossflow analysed with proper orthogonal decomposition. J. Fluid Mech.
2007, 583, 199–227. [CrossRef]
81. Lui, H.F.S.; Wolf, W.R. Construction of reduced-order models for fluid flows using deep feedforward neural networks. J. Fluid
Mech. 2019, 872, 963–994. [CrossRef]
82. Xie, C.; Xiong, X.; Wang, J. Artificial neural network approach for turbulence models: A local framework. Phys. Rev. Fluids 2021,
6, 084612. [CrossRef]
Energies 2023, 16, 2343 20 of 21
83. Milani, P.M.; Ling, J.; Eaton, J.K. On the generality of tensor basis neural networks for turbulent scalar flux modeling. Int.
Commun. Heat Mass Transf. 2021, 128, 105626. [CrossRef]
84. Fang, R.; Sondak, D.; Protopapas, P.; Succi, S. Neural network models for the anisotropic Reynolds stress tensor in turbulent
channel flow. J. Turbul. 2020, 21, 525–543. [CrossRef]
85. Berrone, S.; Oberto, D. An invariances-preserving vector basis neural network for the closure of Reynolds-averaged Navier–Stokes
equations by the divergence of the Reynolds stress tensor. Phys. Fluids 2022, 34, 095136. [CrossRef]
86. Frezat, H.; Balarac, G.; Le Sommer, J.; Fablet, R.; Lguensat, R. Physical invariance in neural networks for subgrid-scale scalar flux
modeling. Phys. Rev. Fluids 2021, 6, 024607. [CrossRef]
87. Wang, R.; Walters, R.; Yu, R. Incorporating symmetry into deep dynamics models for improved generalization. arXiv 2020,
arXiv:2002.03061.
88. Wang, R.; Kashinath, K.; Mustafa, M.; Albert, A.; Yu, R. Towards physics-informed deep learning for turbulent flow prediction.
Proc. ACM Int. Conf. Knowl Discov Data Min. 2020, 26, 1457–1466.
89. Li, Z.; Kovachki, N.B.; Azizzadenesheli, K.; Liu, B.; Bhattacharya, K.; Stuart, A.; Anandkumar, A. Fourier neural operator for
parametric partial differential equations. Proc. Int. Conf. Learn. Represent. 2021, 9.
90. Ronneberger, O.; Fischer, P.; Brox, T. U-net: Convolutional networks for biomedical image segmentation. In Proceedings of
the Medical Image Computing and Computer-Assisted Intervention–MICCAI 2015: 18th International Conference, Munich,
Germany, 5–9 October, 2015; pp. 234–241.
91. Li, Z.; Peng, W.; Yuan, Z.; Wang, J. Fourier neural operator approach to large eddy simulation of three-dimensional turbulence.
Theor. App. Mech. Lett. 2022, 12, 100389. [CrossRef]
92. Li, Y.; Perlman, E.; Wan, M.; Yang, Y.; Meneveau, C.; Burns, R.; Chen, S.; Szalay, A.; Eyink, G. A public turbulence database cluster
and applications to study Lagrangian evolution of velocity increments in turbulence. J. Turbul. 2008, 9, N31. [CrossRef]
93. Eivazi, H.; Vinuesa, R. Physics-informed deep-learning applications to experimental fluid mechanics. arXiv 2022,
arXiv:2203.15402.
94. Jiang, C.M.; Esmaeilzadeh, S.; Azizzadenesheli, K.; Kashinath, K.; Mustafa, M.; Tchelepi, H.A.; Marcus, P.; Prabhat, M.;
Anandkumar, A. MESHFREEFLOWNET: A physics-constrained deep continuous space-time super-resolution framework. In
Proceedings of the SC20: International Conference for High Performance Computing, Networking, Storage and Analysis, Virtual,
9–19 November 2020; pp. 1–15.
95. Sun, X.; Cao, W.; Liu, Y.; Zhu, L.; Zhang, W. High Reynolds number airfoil turbulence modeling method based on machine
learning technique. Comput. Fluids 2022, 236, 105298. [CrossRef]
96. Guan, Y.; Subel, A.; Chattopadhyay, A.; Hassanzadeh, P. Learning physics-constrained subgrid-scale closures in the small-data
regime for stable and accurate LES. Physica D 2023, 443, 133568. [CrossRef]
97. Xiao, H.; Wu, J.L.; Laizet, S.; Duan, L. Flows over periodic hills of parameterized geometries: A dataset for data-driven turbulence
modeling from direct simulations. Comput. Fluids 2020, 200, 104431. [CrossRef]
98. Rumsey, C.; Smith, B.; Huang, G. Description of a website resource for turbulence modeling verification and validation. Proc.
Fluid Dyn. Conf. Exhib. 2010, 40, 4742.
99. Eckert, M.L.; Um, K.; Thuerey, N. ScalarFlow: A large-scale volumetric data set of real-world scalar transport flows for computer
animation and machine learning. ACM Trans. Graph. 2019, 38, 239. [CrossRef]
100. Bonnet, F.; Mazari, J.A.; Cinnella, P.; Gallinari, P. AirfRANS: High fidelity computational fluid dynamics dataset for approximating
Reynolds-Averaged Navier–Stokes solutions. arXiv 2022, arXiv:2212.07564.
101. Goldbloom, A.; Hamner, B. Kaggle: Your Machine Learning and Data Science Community. 2010. Available online: https:
//www.kaggle.com (accessed on 6 February 2023).
102. Chung, W.T.; Jung, K.S.; Chen, J.H.; Ihme, M. BLASTNet: A call for community-involved big data in combustion machine
learning. Appl. Energy Combust. Sci. 2022, 12, 100087. [CrossRef]
103. Hennigh, O.; Narasimhan, S.; Nabian, M.A.; Subramaniam, A.; Tangsali, K.; Fang, Z.; Rietmann, M.; Byeon, W.; Choudhry, S.
NVIDIA SimNet™: An AI-accelerated multi-physics simulation framework. In Proceedings of the Computational Science–ICCS
2021: 21st International Conference, Krakow, Poland, 16–18 June 2021; pp. 447–461.
104. Wang, Q.; Ihme, M.; Chen, Y.F.; Anderson, J. A tensorflow simulation framework for scientific computing of fluid flows on tensor
processing units. Comput. Phys. Commun. 2022, 274, 108292. [CrossRef]
105. Bezgin, D.A.; Buhendwa, A.B.; Adams, N.A. JAX-Fluids: A fully-differentiable high-order computational fluid dynamics solver
for compressible two-phase flows. Comput. Phys. Commun. 2023, 282, 108527. [CrossRef]
106. Takamoto, M.; Praditia, T.; Leiteritz, R.; MacKinlay, D.; Alesiani, F.; Pflüger, D.; Niepert, M. PDEBench: An extensive benchmark
for scientific machine learning. arXiv 2022, arXiv:2210.07182.
107. Yang, L.; Shami, A. On hyperparameter optimization of machine learning algorithms: Theory and practice. Neurocomputing 2020,
415, 295–316. [CrossRef]
108. Krishnapriyan, A.; Gholami, A.; Zhe, S.; Kirby, R.; Mahoney, M.W. Characterizing possible failure modes in physics-informed
neural networks. Adv. Neural Inf. Process. Syst. 2021, 35, 26548–26560.
109. Hu, W.; Fey, M.; Zitnik, M.; Dong, Y.; Ren, H.; Liu, B.; Catasta, M.; Leskovec, J. Open graph benchmark: Datasets for machine
learning on graphs. Adv. Neural Inf. Process. Syst. 2020, 33, 22118–22133.
Energies 2023, 16, 2343 21 of 21
110. Psichogios, D.C.; Ungar, L.H. A hybrid neural network-first principles approach to process modeling. AIChE J. 1992, 38, 1499–1511.
[CrossRef]
111. Moradi, R.; Berangi, R.; Minaei, B. A survey of regularization strategies for deep models. Artif. Intell. Rev. 2020, 53, 3947–3986.
[CrossRef]
112. Standley, T.; Zamir, A.; Chen, D.; Guibas, L.; Malik, J.; Savarese, S. Which tasks should be learned together in multi-task learning?
In Proceedings of the 37th International Conference on Machine Learning, Virtual, 13–18 July 2020.
113. Sener, O.; Koltun, V. Multi-task learning as multi-objective optimization. Adv. Neural Inf. Process. Syst. 2018, 31, 525–536.
114. Goodfellow, I.; Pouget-Abadie, J.; Mirza, M.; Xu, B.; Warde-Farley, D.; Ozair, S.; Courville, A.; Bengio, Y. Generative adversarial
networks. Commun. ACM 2020, 63, 139–144. [CrossRef]
115. Ho, J.; Jain, A.; Abbeel, P. Denoising diffusion probabilistic models. Adv. Neur. Inf. Process. Syst. 2020, 33, 6840–6851.
116. Rombach, R.; Blattmann, A.; Lorenz, D.; Esser, P.; Ommer, B. High-resolution image synthesis with latent diffusion models. In
Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), New Orleans, LA, USA, 19–24
June 2022; pp. 10684–10695.
117. Baddoo, P.J.; Herrmann, B.; McKeon, B.J.; Kutz, J.N.; Brunton, S.L. Physics-informed dynamic mode decomposition (piDMD).
arXiv 2021, arXiv:2112.04307.
118. Schmid, P.J. Dynamic mode decomposition of numerical and experimental data. J. Fluid Mech. 2010, 656, 5–28. [CrossRef]
119. Liu, X.Y.; Wang, J.X. Physics-informed Dyna-style model-based deep reinforcement learning for dynamic control. Proc. Math.
Phys. Eng. Sci. 2021, 477, 20210618. [CrossRef]
120. Stachenfeld, K.; Fielding, D.B.; Kochkov, D.; Cranmer, M.; Pfaff, T.; Godwin, J.; Cui, C.; Ho, S.; Battaglia, P.; Sanchez-Gonzalez, A.
Learned Simulators for Turbulence. Proc. Int. Conf. Learn. Represent. 2022.
121. Shorten, C.; Khoshgoftaar, T.M. A survey on Image Data Augmentation for Deep Learning. J. Big Data 2019, 6, 60. [CrossRef]
122. Yu, F.; Koltun, V. Multi-scale context aggregation by dilated convolutions. arXiv 2015, arXiv:1511.07122.
123. Zhou, J.; Cui, G.; Hu, S.; Zhang, Z.; Yang, C.; Liu, Z.; Wang, L.; Li, C.; Sun, M. Graph neural networks: A review of methods and
applications. AI Open 2020, 1, 57–81. [CrossRef]
124. Peng, J.Z.; Wang, Y.Z.; Chen, S.; Chen, Z.H.; Wu, W.T.; Aubry, N. Grid adaptive reduced-order model of fluid flow based on
graph convolutional neural network. Phys. Fluids 2022, 34, 087121. [CrossRef]
125. He, X.; Wang, Y.; Li, J. Flow completion network: Inferring the fluid dynamics from incomplete flow information using graph
neural networks. Phys. Fluids 2022, 34, 087114. [CrossRef]
126. Rasmussen, C.E. Gaussian processes in machine learning. In Proceedings of the Summer School on Machine Learning; Springer:
Berlin/Heidelberg, Germany, 2003; pp. 63–71.
127. Hinton, G.E.; van Camp, D. Keeping the neural networks simple by minimizing the description length of the weights. Proc.
Annu. Conf. Comput. Learn. Theory 1993, 6, 5–13.
128. Yang, L.; Meng, X.; Karniadakis, G.E. B-PINNs: Bayesian physics-informed neural networks for forward and inverse PDE
problems with noisy data. J. Comp. Phys. 2021, 425, 109913. [CrossRef]
129. Champion, K.; Lusch, B.; Kutz, J.N.; Brunton, S.L. Data-driven discovery of coordinates and governing equations. Proc. Natl.
Acad. Sci. USA 2019, 116, 22445–22451. [CrossRef] [PubMed]
Disclaimer/Publisher’s Note: The statements, opinions and data contained in all publications are solely those of the individual
author(s) and contributor(s) and not of MDPI and/or the editor(s). MDPI and/or the editor(s) disclaim responsibility for any injury to
people or property resulting from any ideas, methods, instructions or products referred to in the content.