RESEARCH ARTICLE | JUNE 25 2024

A physics-constrained and data-driven method for modeling

supersonic flow 
Tong Zhao (赵通); Jian An (安健) ; Yuming Xu (许煜明); Guoqiang He (何国强); Fei Qin (秦飞) 

Physics of Fluids 36, 066118 (2024)

https://doi.org/10.1063/5.0206515


View Export
Online Citation

17 July 2024 12:14:16

 Physics of Fluids ARTICLE pubs.aip.org/aip/pof

A physics-constrained and data-driven method

for modeling supersonic flow
Cite as: Phys. Fluids 36, 066118 (2024); doi: 10.1063/5.0206515
Submitted: 2 March 2024 . Accepted: 4 June 2024 .
Published Online: 25 June 2024

Tong Zhao (赵通), Jian An (安健), Yuming Xu (许煜明), Guoqiang He (何国强), and Fei Qin (秦飞)a)

AFFILIATIONS
National Key Laboratory of Solid Rocket Propulsion, Northwestern Polytechnical University, Xi’an 710072, China

a)
Author to whom correspondence should be addressed: [email protected]

ABSTRACT
A fast solution of supersonic flow is one of the crucial challenges in engineering applications of supersonic flight. This article introduces a
deep learning framework, the supersonic physics-constrained network (SPC), for the rapid solution of unsteady supersonic flow problems.
SPC integrates deep convolutional neural networks with physics-constrained methods based on the Euler equation to derive a new loss func-
tion that can accurately calculate the flow fields by considering the spatial and temporal characteristics of the flow fields at the previous
moment. Compared to purely data-driven methods, SPC significantly reduces the dependency on training data volume by incorporating
physical constraints. Additionally, the training process of SPC is more stable than that of data-driven methods. Taking the classic supersonic

17 July 2024 12:14:16

forward step flow as an example, SPC can accurately calculate strong discontinuities in the flow fields, while reducing the data volume by
approximately 60%. In the generalization test experiment for forward step flow and compression ramp flow, SPC also demonstrates good pre-
dictive accuracy and generalization capability under different geometric configurations and inflow conditions.
Published under an exclusive license by AIP Publishing. https://doi.org/10.1063/5.0206515

I. INTRODUCTION involve some form of linearization during implementation,10,11 and

The efficient solution of supersonic flow problems has been a thus generate truncation error. This truncation error may affect the
major challenge in the field of aircraft design, and the conventional application of ROMs in complex design scenarios.
solution is computational fluid dynamics (CFD) simulations. With the In recent years, artificial intelligence technologies, led by artificial
progressive advancement of computer technology, CFD has become neural networks (ANNs), have achieved remarkable advancements,
indispensable in solving aerodynamic challenges. However, as engi- particularly in the fields of natural language processing (NLP) and
neering demands have expanded, the prolonged computational time computer vision (CV).12–14 Due to their ability to extract hidden infor-
and substantial resource consumption associated with CFD have mation from data, artificial neural networks are a strong fit for nonlin-
emerged as critical limitations to its further development. Thus, there ear systems. Traditional fluid mechanics disciplines have also
arises a compelling need to explore alternative methods. embraced machine learning as a prominent frontier in research. In this
Solving the fluid control equations is a high-dimensional nonlin- context, some researchers15–19 have introduced machine learning to
ear problem. To tackle the inherent high-dimensionality, reduced partially replace CFD governing equation computations or to intro-
order models (ROMs) are considered to be an effective method. duce error corrections, with the goal of enhancing the accuracy of
Among these, proper orthogonal decomposition1 (POD) and dynamic equation solving. Tracey et al.,16 in their work, utilized neural networks
mode decomposition2 (DMD) have shown promising performance.3–5 to substitute the calculation of source terms in the Spalart–Allmaras
Thomas et al.6 first extended the POD to flow fields modeling and model. This enhancement improved both the methods and precision
aerodynamic analysis of the AGARD 445.6 airfoil. Baker et al.7 applied of RANS computations and enabled adaptability to a certain range of
the DMD to process Reynolds-averaged Navier–Stokes (RANS) simu- Reynolds numbers. Kurz and Beck20 employed artificial neural net-
lation data and extract flow field structures at specific frequencies. works featuring multi-layer perceptron (MLP) and gated recurrent
Apart from these classical approaches, fitting functions to existing data unit (GRU) architectures to learn the closure terms of filtering opera-
are also a ROMs strategy, polynomial basis functions8 and radial basis tors used in large eddy simulation (LES). These methods still rely on
functions9 are used to establish functional mappings between specific the CFD solution itself and are designed to expedite calculations while
inputs and outputs. However, it is worth noting that ROMs inherently preserving generality.

Phys. Fluids 36, 066118 (2024); doi: 10.1063/5.0206515 36, 066118-1

Published under an exclusive license by AIP Publishing
 Physics of Fluids ARTICLE pubs.aip.org/aip/pof

However, researchers have also considered neural networks as been used to model and predict flow fields at locations such as the
black-box models to replace the complex and time-consuming process inlets and isolator sections of hypersonic vehicles.43
of solving Navier–Stokes equations. Kim and Lee21 combined genera- Physics-driven methods have encountered certain bottlenecks in
tive adversarial networks (GANs) with recurrent neural networks modeling supersonic flows, while data-driven methods have made
(RNNs) to replace the direct numerical simulation (DNS) model for advances in the field. Consequently, researchers have attempted to
simulating turbulent channel flows. By incorporating samples with dif- combine the two approaches. Mao et al.44 achieved favorable results by
ferent Reynolds numbers, this approach exhibited good generalization introducing clustering-based training samples at potential discontinu-
performance. Li et al.22 constructed a deep learning framework using ity locations in solving two-dimensional oblique shocks and shock
convolutional neural networks (CNNs) for blade physical fields and tube problems. Liu et al.45 weakened the weights of PINNs near strong
aerodynamic predictions. Benefiting from CNN’s impressive perfor- discontinuities, allowing PINNs to predominantly learn from the
mance in computer vision, the abundant grid data within flow fields smooth regions of the flow fields. The natural occurrence of disconti-
naturally corresponds to pixels in images. Consequently, numerous nuities within the flow fields, influenced by compressible characteris-
research outcomes have emerged using CNNs and its various variants tics, was thus captured. This method yielded positive outcomes in
for flow modeling, prediction, and reconstruction.23–26 solving shock tube problems, the Lax problem, and two-dimensional
Data-driven flow modeling methods have made significant Riemann problems. Gao et al.46 treated training residuals as posterior
strides. However, these approaches typically require a substantial predictions within PINNs, continuously adjusting the neural network’s
amount of data27 to achieve robust generalization performance. In sit- sampling points in space during training, and achieved satisfactory
uations where obtaining sufficient high-precision data may pose chal- computational results.
lenges, introducing physical constraints can be a valuable strategy. The From the previous research, we can summarize as follows: ‹.
physics-informed neural networks27,28 (PINNs) method, introduced Data-driven methods for flow field modeling face challenges due to
by Raissi, incorporates physical conservation laws into neural networks their substantial data requirements for training, which inevitably intro-
to solve partial differential equations (PDEs). It successfully predicted duce computational costs. The potential of PINNs in modeling com-
the Karman vortex behind a cylinder in the absence of data. Rao plex supersonic flow has not yet been fully exploited. One of the
et al.29 further enhanced Raissi’s work by introducing strong boundary questions that needs addressing is how to effectively integrate the
condition constraints, achieving higher modeling accuracy for flow physics-driven and data-driven components in PINNs to enhance
around a cylinder. Harada et al.30 utilized physics-constrained neural their stability, convergence, and generalization during the training pro-

17 July 2024 12:14:16

networks to solve Euler’s equations and reinforced the Navier–Stokes cess. ›. In the domain of subsonic flows, substantial progress has been
equation residuals through transfer learning. This approach led to the achieved, particularly in addressing aerodynamic problems related to
swift reconstruction of the RAE2822 airfoil flow fields. Combining the aircraft wings. These methods have demonstrated positive results and
excellent fitting capabilities of CNNs for flow field data, Gao et al.31 successful generalizations. However, research in applying intelligent
established physics-informed CNNs for modeling and predicting com- methods to model supersonic flow has not been as extensive as sub-
plex geometric flow fields. In this domain, Wang et al.32 proposed a sonic flow, and there is still much potential for progress. Tackling these
physics-informed machine learning (PIML) framework for solving tur- challenges and addressing these gaps in research is pivotal for advanc-
bulent models. It leverages observational data to create a machine ing the application of physical law and data in fluid dynamics model-
learning model mapping flow fields features to Reynolds tensors, ing, particularly in scenarios involving supersonic flow conditions.
replacing the need for expert experience. Moreover, a comprehensive By combining the strengths of the physics-constrained neural
review in Ref. 33 provides a detailed overview of the current applica- network and CNNs, this paper aims to advance the field of modeling
tions of PINNs in the field of fluid mechanics. In many previous stud- supersonic compressible Euler equations. The main contents of this
ies,34–37 physics-informed CNNs show great potential for physics paper are as follows: Sec. II briefly describes the methodology used in
problems. this study, and focuses on the network structure and the novel loss
In the aforementioned studies, researchers have made substantial
function proposed. Section III will introduce the cases used in this
progress in studying both compressible and incompressible flows, pre-
paper, as well as the evaluation methods for network hyperparameters,
dominantly focusing on subsonic and transonic regimes. For com-
data volume, and model generalization. The fourth section is the dis-
pressible flows involving supersonic conditions, the presence of strong
play and analysis of the calculation results of the third part. The latest
discontinuities in the flow fields has posed challenges for both data-
section is the conclusion of the full text.
driven methods and PINNs. Addressing the issue of separation and
reattachment in supersonic wall flows, Zangeneh et al.17 employed
DNS data to train a random forest model, enhancing the accuracy of II. METHODOLOGY
wall-modeled large eddy simulations for supersonic flows. Zanjani A. Overview of the method
et al.38 predicted the flow fields of a supersonic nozzle through CNNs 1. Convolutional neural networks
and optimized the nozzle shape using ANNs with genetic algorithms.
Wu et al.39 also achieved high accuracy prediction of supersonic flow CNNs automatically learn features of images at various channels
fields around airfoils using sparse convolutional neural networks. With through convolution and pooling operations, which align with human
hypersonic vehicles being a major research focus in the realm of super- understanding of image processing. As a result, CNNs have been
sonic flows, researchers have employed various methods, including widely applied in the field of computer vision. Flow fields, to a certain
multilayer perceptron networks,40 single-path CNNs,41 multi-path extent, exhibit hierarchical features similar to images; the application
CNNs,41,42 and multi-branch fusion CNNs.24 These methods have of CNNs for modeling flow fields has garnered significant attention.

Phys. Fluids 36, 066118 (2024); doi: 10.1063/5.0206515 36, 066118-2

Published under an exclusive license by AIP Publishing
 Physics of Fluids ARTICLE pubs.aip.org/aip/pof

2. PINNs the spatial discretization using the specified weight convolution kernel,
and the last part is the loss function of the coupled physical constraints
PINNs leverage the automatic differentiation algorithm of neural proposed in this paper.
networks to iteratively train and approximate specific solutions to dif- First, compared to subsonic flow, one of the most distinctive char-
ferential equations.47 acteristics of supersonic flow is its sensitivity to fluid compressibility.
The specific set T is composed of two subsets, Tf and Tb , repre- Supersonic flow fields exhibit strong discontinuities such as shock waves.
senting the sampling points within the computational domain and on Since there is a huge difference in the physical properties before and
the computational domain boundary, respectively. Tf and Tb together after the shock wave, we regard these discontinuities in the flow fields as
are referred to as the collection of sampling points. To constrain the an edge feature in the field of computer vision. As a variant of CNNs,
neural network’s output u ^ , the loss function is defined as a weighted U-net has excellent performance in semantic segmentation and edge
sum of the L2 norm of the residual equations and boundary recognition; it was initially built by Ronneberger et al.48 in 2015. In neu-
conditions, ral networks, deeper architectures enhance the capability to extract
Lðh; TÞ ¼ wf Lðh; Tf Þ þ wb Lb ðh; Tb Þ; (1) intrinsic data features, but the edge features in images tend to weaken as
the network layers deepen. U-Net addresses this issue by extracting

 1 X @u @ 2 u @u image features through down-sampling while retaining more edge tex-
L h; Tf ¼ f xi ; ;      ; yi ;      ;
Tf x2T @xi @xi2 @yi tures through the concatenation operation during up-sampling. The
f
2 principle of the SPC can be simply expressed as UNN tþ1
¼ NNðU t ; X; hÞ,
@u where NN represents network, U t is the tensor of flow fields at time t,
ti ;      ; h ; (2)
@ti 2 and UNN tþ1
is the flow field at the next time step computed by SPC. For
1 X unsteady flow simulations, the predicted flow fields at t by the network
Lðh; Tb Þ ¼ Bðu^ ; xÞ22 : (3) are used as inputs for predicting the flow fields at t þ 1.
Tf x2T
b
However, SPC is developed based on CNNs. Therefore, for the
In the above-mentioned equation, wf and wb represent weights, computation of the differential terms in Eq. (8), SPC does not use the
while the other terms involve derivatives, partial derivatives, normal traditional approach of iterating through the grid to implement finite
derivatives, and similar differential terms. Unlike conventional numer- differences. Instead, it utilizes convolution kernels with fixed weights
ical methods that rely on finite differences or similar techniques to for the rapid computation of tensors. This computational method can
be expressed as

17 July 2024 12:14:16

compute differential terms, in this context, these differential terms can
be efficiently obtained through automatic differentiation operations.  
@UNN 1
¼ ðUNN  W@x Þi;j ; W@x ¼ ½1; 0; 1: (4)
@x i;j Dx
B. Supersonic physics-constrained (SPC) model
1. Model structure
2. Loss function
Figure 1 illustrates the overall structure of SPC, which currently
enables fast computation of compressible supersonic inviscid transient The compressible supersonic inviscid flow is described by Euler
flow fields, and the framework consists of three parts. The first part is equations. The Euler equation in a two-dimensional rectangular coor-
the transient flow fields generator based on U-net, the second part is dinate system can be expressed as follows:

FIG. 1. Overall network architecture of SPC.

Phys. Fluids 36, 066118 (2024); doi: 10.1063/5.0206515 36, 066118-3

Published under an exclusive license by AIP Publishing
 Physics of Fluids ARTICLE pubs.aip.org/aip/pof

@U @FðUÞ @GðUÞ methods within a two-dimensional plane. Spatial discretization can

þ þ ¼ 0; (5)
@t @x @y be achieved using the second-order central difference scheme, while
time discretization can be performed using the first-order implicit
and Euler method. The discretized Euler equations can be represented as
3
2 2 3 2 3 follows:
q qu qv
6 qu 7 6 qu þ p 7
2 6 quv 7
U¼6 7 FðUÞ ¼ 6 7 GðUÞ ¼ 6 7 Þ GðUjþ1 Þ  FðUj1 Þ
tþ1 tþ1
4 qv 5; 4 quv 5; 4 qv2 þ p 5: U tþ1  U t FðUiþ1 Þ  FðUt1
tþ1 tþ1
þ þ ¼ 0: (8)
E uðE þ pÞ vðE þ pÞ Dt 2Dx 2Dy
(6) The loss function constructed based on the discretized Euler
equations can be expressed as follows:
The Euler equations consist of a set of nonlinear partial differen-
 tþ1
tial equations formed by the conservation of mass, momentum, and 1 X  UNN  UNN
t FðUNN tþ1
Þ  FðUNN
tþ1
Þ
Losspdes ¼ þ
iþ1 i1
energy. In the given equations, q, u, v, p, and E, respectively, represent Tf x2T  Dt 2Dx
density, velocity components in the x and y directions, pressure, and f

energy per unit volume. To complete the formulation of the equations, GðUNN tþ1
Þ  GðUNN
tþ1
Þ 2
þ
jþ1 j1
: (9)
an ideal gas equation of state is necessary to describe the relationship
2Dy 
between pressure and energy,
  In order to tackle the strongly coupled and nonlinear nature of
1 the two-dimensional Euler equations, this study utilizes mathematical
p ¼ ðc  1Þqe ¼ ðc  1Þ E  pðu þ v Þ :2 2
(7)
2 manipulation to completely decouple the nonlinear Euler equations
Indeed, in many cases, the ratio of specific heats (c) can be into four independent solvable wave equations,
assumed to be a constant value. A commonly used value for c is 1.4.
Using the Euler equations as the basis for the loss function, dis- @U @U @U
þA þB ¼ 0; (10)
cretization can be carried out on these equations using finite difference @t @x @y

2 3

17 July 2024 12:14:16

0 1 0 0
6 7
6 ðc  3Þu2 þ ðc  1Þv2 7
6 ð3  cÞu ðc  1Þv c  17
6 2 7
A¼6
6
7;
6 uv v u 0 7
7
6 7
4 ðc  2Þðu3 þ uv2 Þ a2 u ð3  2cÞu2 þ v2 a2 5
 þ ðc  1Þuv cu
2 c1 2 c1

2 3
0 0 1 0
6 7
6 uv v u 0 7
6 7
6 ðc  1Þu2 þ ðc  3Þv2 7
B¼6
6 ðc  1Þu ð3  cÞv c177;
6 2 7
6 7
4 ðc  2Þðv3 þ u2 vÞ a2 v ð3  2cÞv2 þ u2 a2 5
 ðc  1Þuv þ cv
2 c1 2 c1

t
where a represents the local speed of sound. Additionally, to enhance where UCFD is obtained by solving the Euler equation. Furthermore, by
the temporal constraint of the neural network model and accomplish incorporating constraints for boundary conditions, SPC’s loss function
unsteady computations, a combined approach of data-driven and for the supersonic flow model is as follows:
physics-constrained modeling is employed. This entails integrating
CFD data during the time progression. Ultimately, the derived loss Loss ¼ x1 Losspdes þ x2 LossBC : (12)
function for the physical equations is given as
III. EXPERIMENTAL SETUP
 tþ1
1 X  UNN  UCFD
t tþ1
UNN  UNN
tþ1
A. Experimental configurations
¼ þ
iþ1 i1
Losspdes A
Tf x2T  Dt Dx The forward step flow is a classic case of supersonic flow, and its
f

tþ1 2 geometric structure is depicted in Fig. 2. All physical quantities in

tþ1
UNN  UNN 
þB
jþ1 j1
; (11) Fig. 2 are expressed in a dimensionless manner. The computational
Dy 
domain spans the range x 2 ½0; 4, y 2 ½0; 1. A solid step of height 0.2

Phys. Fluids 36, 066118 (2024); doi: 10.1063/5.0206515 36, 066118-4

Published under an exclusive license by AIP Publishing
 Physics of Fluids ARTICLE pubs.aip.org/aip/pof

TABLE I. Neural network model parameter selection experiment.

Activation Convolutional Kernel Learning

No. function kernels size rate

1 Tanh 16 33 0.005

2 Tanh 24 55 0.01
3 Tanh 32 73 0.05
FIG. 2. The basic configuration of supersonic forward step flow problem. 4 Relu 16 73 0.01
5 Relu 24 33 0.05
and length 3.0 is located at the lower boundary of the computational 6 Relu 32 55 0.005
domain, while the remaining regions constitute the fluid domain. The 7 LeakyRelu 16 55 0.05
right boundary of the computational domain is designated as an out- 8 LeakyRelu 24 73 0.005
flow boundary, and the left boundary represents an inflow boundary 9 LeakyRelu 32 33 0.01
with a supersonic inflow. CFD simulation data are obtained using a
second-order accurate finite difference method, with a spatial step size
of Dx ¼ Dy ¼ 0:01 and a time step size of Dt ¼ 0:015. The validation of supersonic flow, obtaining high-fidelity numerical results requires
of the computational methods and grid independence is shown in substantial computational resources. Additionally, acquiring experi-
Appendix A.49 mental data can be even more costly. Therefore, the key focus of
research in flow modeling intelligence is how to use fewer training
samples while maintaining the computational accuracy of deep
B. Effect of hyperparameters
learning-based flow modeling. Consequently, we verified the applica-
In order to determine the optimal network hyperparameters and tion of purely data-driven neural networks in modeling supersonic
evaluate the influence of different hyperparameters on prediction flows. The loss function for the purely data-driven neural networks is
results, this study conducts an orthogonal experiment focusing on four as follows:
key hyperparameters commonly emphasized in convolutional neural
networks: activation function, number of convolutional kernels, learn- 1X
Loss ¼ Lossdata ¼ jðUNN Þi  ðUCFD Þi j2 :

17 July 2024 12:14:16

(13)
ing rate, and kernel size. n
The training of the network utilizes the Adam optimizer. Discrete We aim to determine that the SPC with physics constrain has a
the total computation time of 3 s into 200 time steps. The details of the reduced demand for data volume. This study compares the prediction
orthogonal experiment design are provided in Table I. results of both the data-driven neural network and SPC under different
data distribution scenarios. The network structure and hyperpara-
C. Loss function and dataset volume meters were chosen based on the previously determined optimal
The quantity of samples in the training set has a significant parameter combination. The blue dots in Fig. 3 correspond to the posi-
impact on the results of deep learning model. However, in the context tions of the corners. Each subplot represents the distribution of the

FIG. 3. Dataset distribution. The volume of data used for training is (a) 3, (b) 4, (c) 6, and (d) 9.

Phys. Fluids 36, 066118 (2024); doi: 10.1063/5.0206515 36, 066118-5

Published under an exclusive license by AIP Publishing
 Physics of Fluids ARTICLE pubs.aip.org/aip/pof

FIG. 4. Distribution of datasets with changes in configuration (a) forward step. (b) three-wedge ramp.

dataset, and the volume of data used for training represents 3, 4, 6, and The computation using the trained network consumes only
9, respectively. 0.012 s, which has a high acceleration effect. The loss values during
training are analyzed separately in Sec. IV B and Fig. 7 of this paper.
D. Generalization ability in different boundary Other details about training and prediction are also described in more
conditions detail below.
In this section, the generalization ability of SPC will be verified by
changing the geometric boundary and inflow conditions. In the vari- A. Effect of network hyperparameters
ous geometric experiments, the flow velocity is fixed at u ¼ 3, and the
In the study of supersonic flows, wave structures are critical focal
coordinates x and y of the step corner are used to characterize the step.
points. To precisely evaluate the accuracy of neural network predic-
As shown in Fig. 4, the blue dots represent the distribution of the train-
tions, this paper intends to employ numerical error and shock wave
ing set, and the red dot represents the position of the testing set. The
prediction result within the black dashed lines signifies interpolation position error as evaluation metrics,
generalization ability, while outside the black dashed lines represents  t 
 xpred  xCFD
t

extrapolation generalization ability. In the experiments involving ERRORp ¼  t
  100%;
 (14)
xCFD

17 July 2024 12:14:16

changes in flow velocity, a fixed geometric of forward step with coordi-
nates x ¼ 1.0 and y ¼ 0.2 is used. The incoming velocity is treated as ERRORN ¼ jutpred  utCFD j; (15)
the variable. The training set includes inflow velocities ranging from t t
2.4 to 2.6, with a spacing of 0.2 between each training sample. where ERRORp represents the shock position error, xpred and xCFD are
On the other hand, we employ a three-wedge ramp as another the coordinates of the shock positions in the neural network predic-
case to verify the generalization performance. For experiments involv- tions and CFD results, respectively. A smaller value of ERRORp indi-
ing geometric variations, this paper uses the dashed region in Fig. 4(b) cates a higher degree of overlap between the predicted shock position
as the sampling range. The coordinate y of each control point can be by the neural network and the actual shock position from CFD.
randomly selected within the upper and lower intervals. Ten samples However, ERRORN represents the numerical error, utpred and utCFD are
are generated for training. To ensure the rationality of the turning predicted by the neural network and obtained from CFD, respectively.
angle range, the range of values for each control point pi is listed in Numerical error is helpful as we observe the error distribution from
Table II. For experiments involving variations in inflow velocity, we fix the perspective of the entire flow fields, while shock wave position
the coordinates of control points pi as {p0 ¼ 0, p1 ¼ 0.15, p2 ¼ 0.50, error will focus on the accuracy of the intermittent predictions.
p3 ¼ 0.61} and uniformly sample seven samples within the interval The full results of the orthogonal experiments are presented in
u ¼ [2.2, 3.4] for training. Appendix B. It is evident that the training results for the first and third
parameter combinations have failed. The range analysis of the calcula-
IV. RESULT AND DISCUSSION
tion results of all 9 test cases in this study is carried out, Table IV
The training of SPC was done on RTX 3090, based on the presents the sum of experimental results corresponding to each column’s
Pytorch framework, while the necessary CFD data were computed on influencing factor as Ki, and the average value of Ki as Li. The magnitude
an Intel Xeon Gold 6226R central processing unit (CPU), and the
other necessary hardware and software environments as well as the TABLE III. Software and hardware configuration.
time consumed for the model training are displayed in Table III.
TABLE II. The range of control points in Fig. 4(b). Name Version/time

CPU Intel Xeon Gold 6226R 2.9 GHz

pi Minimum Maximum
Graphics processing unit (GPU) NVIDIA GeForce RTX3090
0 0.00 0.00 Pytorch 1.12 þ cu11.6
1 0.08 0.24 CUDA 12.1
2 0.37 0.60 CFD calculation time 161.91 s
3 0.41 0.66 SPC train time 2 h 18 min

Phys. Fluids 36, 066118 (2024); doi: 10.1063/5.0206515 36, 066118-6

Published under an exclusive license by AIP Publishing
 Physics of Fluids ARTICLE pubs.aip.org/aip/pof

TABLE IV. Range analysis of the effect of network hyperparameters on the results. We analyze the reasons for the influence of the above hyperpara-
meters. According to Fig. 5, in the conventional image recognition
Hyperparameters problem, Tanh is widely used as an activation function. With the
advantage of fast training speed, it can meet the needs of high precision
Activation Convolutional Learning Kernel
in learning results. However, solving the physical problem is different
function kernels rate size ERRORp
from the image recognition. The flow process is continuous and more
1 Tanh 16 0.005 33 sensitive to the change in boundary conditions. Therefore, Tanh makes
2 Tanh 24 0.01 55 10.63 the neural network show the same output under different boundary
3 Tanh 32 0.05 73 conditions, because it is easy to be saturated. The input and output of
the Relu are linear in the domain greater than zero, so it does not have
4 Relu 16 0.01 73 8.48
the disadvantage of saturation. However, in the definition domain less
5 Relu 24 0.05 33 13.7
than zero, the output of the Relu is all zero, which makes the neuron
6 Relu 32 0.005 55 10.22 no longer responsive to any input after completing a large gradient
7 LeakyRelu 16 0.05 55 14.45 operation, which makes the network fail to update the gradient during
8 LeakyRelu 24 0.005 73 10.88 the training process. The LeakyRelu improves the disadvantage of the
9 LeakyRelu 32 0.01 33 1.43 Relu without negative input by giving a small linear component to a
K1 210.63 122.93 121.10 115.13 negative input.50 It shows better stability and higher accuracy than
K2 32.40 35.21 20.54 35.30 others. The learning rate is a secondary factor affecting the learning
K3 26.76 111.65 128.15 119.36 results. The smaller learning rate converges slowly, but the larger learn-
L1 70.21 40.98 40.37 38.38 ing rate will overfit the model. After comparison, 0.01 is the optimal
L2 10.80 11.74 6.85 11.77 learning rate. In addition, the more number of convolution kernels is
L3 8.92 37.22 42.72 39.79 not better. More convolution kernels can extract more features, but at
the same time, it will cause too many network parameters and take a
Rj 61.29 29.24 35.87 28.02 long time for network training. The size of the convolution kernel has
the least influence on the results, the training speed of the small convo-
Priority Activation function, Learning rate, lution kernel is faster. The large convolution kernel can capture a larger
Convolutional kernels, Kernel size flow fields space and consider the spatial characteristics of the flow

17 July 2024 12:14:16

more comprehensively, but it has little effect on this study.
Optimal LeakyRelu24 0.01 55
It can be concluded that the optimal level for activation function
is LeakyRelu. Similarly, the optimal levels for convolutional kernels,
learning rate, and Kernel size can be determined as 24, 0.01, and 5  5,
of Li determines the optimal level for that parameter, and consequently,
respectively. Based on the Rj, the sequence of the primary and second-
the optimal combination of various hyperparameters can be derived. Rj
ary effects of the hyperparameters on the calculation accuracy is activa-
represents the range of the corresponding column’s influencing factor,
tion function, learning rate, convolutional kernels and kernel size.
which is the difference between the maximum and minimum values of
the average indicator values at various levels. Rj reflects the extent of
experimental results’ variation when the influencing factor of that col- B. Analysis of data volume
umn changes. The larger Rj is, the greater the impact of that parameter The purpose of PINNs is to provide a novel approach for solving
on the results, making it more crucial. Therefore, the magnitude of Rj can partial differential equations. However, as more complex flow prob-
determine the priority order of various hyperparameters. lems are tackled, the mathematical equations that describe the flow

FIG. 5. Activation function (a) Tanh, (b) Relu, and (c) LeakyRelu.

Phys. Fluids 36, 066118 (2024); doi: 10.1063/5.0206515 36, 066118-7

Published under an exclusive license by AIP Publishing
 Physics of Fluids ARTICLE pubs.aip.org/aip/pof

FIG. 6. (a) Results of the CFD at t ¼ 1 s, t ¼ 2 s, and t ¼ 3 s. (b) Results of the physics-driven neural network at t ¼ 1 s, t ¼ 2 s, and t ¼ 3 s. (c) Results of the SPC at t ¼ 1 s,
t ¼ 2 s, and t ¼ 3 s.

phenomena become more intricate. Solely relying on data-free PINNs differential information from the physical equations in training neural
presents certain challenges in solving these complex equations. For the networks, SPC can help reduce the demand for training samples.
problem studied in this paper, if the data-driven approach is not intro- In order to explore the reasons for the above-mentioned conclu-

17 July 2024 12:14:16

duced, the neural network loss function is given by Eq. (9). The net- sions, we conducted an analysis of the physical conservation during
work training results are shown in Fig. 6(b). network training by calculating the loss values for various flow param-
It can be clearly found that the physics-driven neural network eters within the flow fields. It can be observed that both the data-
lacks the constraints in time dimension and is unable to solve the driven neural network and SPC exhibit an overall decreasing trend in
unsteady flow. However, when we use Eq. (12) as the loss function, the losses during training. However, the loss values for the data-driven
computational results after adding CFD data as training samples are neural network show oscillations as they fit the training data.
shown in Fig. 6(c). This demonstrated that SPC effectively enhances Comparing the analysis regarding mass and energy conservation
the convergence difficulties encountered by PINNs when solving high- shown in Figs. 7(c) and 7(d), it is evident that data-driven training
dimensional nonlinear problems. exhibits nonphysical phenomena of mass and energy non-
The introduction of a data-driven approach helps in the accurate conservation during the training process. However, after incorporating
prediction of SPC, so it is necessary to explore the demand for the vol- the physics-constrained network, SPC’s loss values rapidly converge to
ume of data for SPC. The results of the above study are shown in a lower level. Especially when comparing the statistics on conservation
Appendix C. Table V shows the error of different cases. Comparative with the purely data-driven approach, SPC demonstrates a much
analysis of the four different training sets described above shows that stronger physical interpretability throughout the training process. As a
the network model including the physics-constrained model requires
result, SPC for flow fields modeling not only captures the parameter
only 3 training samples, while the data-driven deep learning method
variations in unsteady flow processes but also accelerates the conver-
requires 9 samples, provided that the shock wave position error is less
gence of training and learns the underlying physical laws. This
than 5%. This observation highlights the supportive role of high-order
approach produces results that are more compelling and interpretable.

C. Generalization ability
TABLE V. Error results for each of the data-driven neural network and SPC.
1. Forward step flow
Data-driven SPC
Volume of First, we verify the prediction results for the forward step prob-
datasets ERRORN ERRORP (%) ERRORN ERRORP (%) lem. The results in Fig. 8 correspond to five sampling points from the
test dataset with different geometries, none of which were included in
1 3 0.281 100 0.022 2.78 the training samples. In Fig. 8(a), the coordinates of the corner point
2 4 0.108 44.35 0.021 2.41 are x ¼ 0.5 and y ¼ 0.2. In this case, due to the proximity of the step to
3 6 0.059 15.77 0.016 1.81 the inlet, a third reflected shock wave appears at the upper boundary
4 9 0.045 4.94 0.012 1.60 of the flow fields, a special case not present in the training set. This
indicates that the SPC can effectively learn the physical characteristics

Phys. Fluids 36, 066118 (2024); doi: 10.1063/5.0206515 36, 066118-8

Published under an exclusive license by AIP Publishing
 Physics of Fluids ARTICLE pubs.aip.org/aip/pof

17 July 2024 12:14:16

FIG. 7. Conservation analysis: (a) Loss values of data-driven neural network, (b) loss values of SPC, (c) mass conservation comparison, and (d) energy conservation
comparison.

of the flow fields, with a numerical error of 0.023 and a shock posi- error in the position of the third reflected shock wave in the upper
tion error of 0.020. In Fig. 8(b), the coordinates of the corner point right corner of the computational domain. This discrepancy might be
are x ¼ 1.0 and y ¼ 0.2. Only two reflected shock waves appear on due to the lower incoming flow speed, resulting in a weaker intensity
the upper wall in this case, with an average numerical error of 0.012 of the reflected shock wave and a thicker shock wave surface in the cal-
and a shock position error of 3.82%. The error is mainly concen- culations, leading to distortion in the shock wave position calculation.
trated near the last reflected shock wave. As observed in Fig. 8(b), For this case, the numerical error of the fluid fields is 0.022, and the
the predicted shock waves by SPC are closer to the upstream region error in shock wave position is 4.82%. Figure 9(b) displays a compari-
compared to the results from CFD calculations. In Fig. 8(c), repre- son and error analysis of the computational results for the case with
senting the corner point coordinates of x ¼ 1.3 and y ¼ 0.3, due to u ¼ 4. In this case, the left region of the computational domain is
an increase in step height, the bow shock at the upper boundary of slightly slower. This might be attributed to the neural network model
the flow fields transitions to a normal shock wave. SPC also cap- not capturing the incoming flow boundary accurately, the affected area
tures this change, with a numerical error of 0.018 and a shock posi- is only a small portion of the entire computational domain. However,
tion error of 3.73%. In Fig. 8(d), the corner point coordinates are the shock wave position calculation is fairly accurate, the numerical
x ¼ 1.6 and y ¼ 0.4, and in Fig. 8(e), the corner point coordinates error of the fluid fields is 0.053, and the error in shock wave position is
are x ¼ 2.0 and y ¼ 0.4. In these two cases, the original bow shock 2.54%.
wave completely transforms into a normal shock wave by further Regarding the error in mentioned above, it should be noted that
elevation of the step height. SPC predicts the shock wave fronts SPC is still a model based on data. Physics-constrained networks are
clearly and accurately, with high precision in other areas of the flow introduced to minimize the demand for high-resolution supersonic
fields. The average numerical errors for the flow fields are 0.018 and flow data. At the same time, physics-constrained networks make
0.013, and the shock position errors are 2.06% and 2.00%, model training easier to converge and more physically interpretable,
respectively. but this does not mean that SPC can realize the modeling of supersonic
Second, we further validate the generalization of SPC to changes flow entirely through physics. The complexity of the flow field charac-
in inflow conditions. Figure 9(a) shows the computational results and teristics is still an important factor affecting the accuracy. This will be
error of CFD and SPC for the step problem when u ¼ 2, the predicted more intuitively reflected in the experiment on the compression ramp
shock wave surface by SPC is relatively clear. However, there is a small in Sec. IV C 2.

Phys. Fluids 36, 066118 (2024); doi: 10.1063/5.0206515 36, 066118-9

Published under an exclusive license by AIP Publishing
 Physics of Fluids ARTICLE pubs.aip.org/aip/pof

17 July 2024 12:14:16

FIG. 8. CFD and SPC results for different geometrical configurations of the forward step, (a) x ¼ 0.5 and y ¼ 0.2; (b) x ¼ 1.0 and y ¼ 0.2; (c) x ¼ 1.3 and y ¼ 0.3; (d) x ¼ 1.6
and y ¼ 0.4; and (e) x ¼ 2.0 and y ¼ 0.4.

2. Compression ramp flow highly accurate predictions of shock wave behavior. For the extrapola-
tion case, the coordinates of point p are [0, 0.16, 0.61, 0.72]. In this
This section is a verification of the generalization of SPC on the case, the angles of wedges are larger, resulting in a more significant
three-wedge compression ramp flow. Among the experiments of compression effect. As depicted in Fig. 10(b), the error is primarily
geometries change, an interpolated working condition and an exter- concentrated near the fourth wedge angle. Notably, SPC predicts shock
nally interpolated working condition are chosen in this paper. In the wave positions closer to the upstream region compared to CFD calcu-
interpolation case, the coordinates of point p are [0, 0.12, 0.47, 0.50], lations. The average flow field error for this case is 0.010, with a shock
the results predicted by SPC are illustrated in Fig. 10(a). For the inter- wave position error of 2.63%, SPC demonstrates significant precision
polation case, the average numerical error is only 0.008 with a shock in predicting the behavior of the ramp.
wave position error of 2.25%. The error between the predictions and Figure 11(a) shows the computational and predicted results of the
the true value is mainly in the wave back of the reflected shock wave ramp at u ¼ 2, with a good level of agreement between the predicted
near the upper wall, and the reflection angle is slightly larger in the results and the true values. Due to the slower inflow velocity, the oblique
prediction result of SPC. However, overall, SPC provides clear and shock waves generated by the first and second ramps do not reflect at the

FIG. 9. CFD and SPC results for different inflows of the forward step. (a) u ¼ 2.0 and (b) u ¼ 4.0.

Phys. Fluids 36, 066118 (2024); doi: 10.1063/5.0206515 36, 066118-10

Published under an exclusive license by AIP Publishing
 Physics of Fluids ARTICLE pubs.aip.org/aip/pof

FIG. 10. CFD and SPC for different geometrical configurations of a three-wedge compression ramp. (a) pi ¼ [0, 0.12, 0.47, 0.50]. (b) pi ¼ [0, 0.16, 0.61, 0.72].

FIG. 11. CFD and SPC results for different inflows of a three-wedge compression ramp. (a) u ¼ 2.0 and (b) u ¼ 3.6.

17 July 2024 12:14:16

upper boundary. The numerical error in the flow fields is 0.009, with a framework is analyzed in two cases, forward step flow and compres-
shock wave position error of 2.92%, primarily concentrated near the sion ramp. The study arrives at the following conclusions:
reflected shock waves on the lower wall. In Fig. 11(b), at u ¼ 3.6, the
(1) SPC incorporates U-Net as its core architecture. By leveraging the
shock waves generated by each ramp converge and reflect on the upper
exceptional edge detection capabilities inherent to U-shaped net-
wall, the reflection angles predicted by SPC are slightly larger than the
works, the framework demonstrates remarkable proficiency in delin-
CFD calculation results. The numerical error in the flow fields is 0.010,
eating the discontinuous features prevalent in supersonic flow fields.
with a shock wave position error of 3.56%. SPC demonstrates excellent
(2) In response to the unique characteristics of supersonic flows, we
generalization to variations in incoming flow, further confirming its
have devised an innovative loss function that synergistically integra-
applicability in modeling hypersonic flow. tes physics-constrained with data-driven methodologies. This
SPC exhibits an excellent generalization capability. For the forward
approach substantially diminishes the reliance on extensive datasets
step problem, although the training samples are only distributed over a for training deep neural networks in flow field analysis, facilitating
space of 1 unit in length, SPC can expand the computational domain to more efficient convergence and enhancing physical interpretability.
1.5 units in length, SPC performance remains excellent in changing (3) Demonstrating robust generalization capabilities, SPC excels in
incoming flow conditions. However, SPC is capable of accurately scenarios requiring extrapolation. The case studies, involving
describing the evolution process from bow shocks to normal shocks, varied configurations and inflow conditions, indicate that SPC
especially the interaction of the shock wave with the solid boundary. In can effectively extend the computational domain beyond half
the case of compression ramps, we observed an intriguing phenomenon, the span of the original data distribution.
SPC demonstrates higher prediction compared to forward steps. We
hypothesize that this is attributed to the geometric configuration of the In conclusion, this research provides a viable alternative for rap-
ramp, which tends to feature oblique shock waves in the flow fields idly acquiring high-fidelity supersonic flow field data. Moving forward,
rather than bow shocks prevalent in step problems. The flow fields our focus will shift toward improving computational accuracy and
behind ramps tend to be more uniform, facilitating SPC in learning dis- applying SPC to more complex scenarios in supersonic flow computa-
continuous features more effectively. This is one more way to indicate tion. This will drive the progress of numerical simulations in super-
that data quality is important for the accurate modeling of SPC. sonic fluid dynamics.

V. CONCLUSION ACKNOWLEDGMENTS
In this study, a deep learning framework, SPC, for unsteady The authors would like to express their thanks for the support
supersonic inviscid flows is proposed, and the applicability of the from the National Natural Science Foundation of China (Grant No.

Phys. Fluids 36, 066118 (2024); doi: 10.1063/5.0206515 36, 066118-11

Published under an exclusive license by AIP Publishing
 Physics of Fluids ARTICLE pubs.aip.org/aip/pof

52106165). Also, the authors thank the reviewers for their DATA AVAILABILITY
recommendations. The data that support the findings of this study are available from
the corresponding author upon reasonable request.
AUTHOR DECLARATIONS
Conflict of Interest APPENDIX A: NUMERICAL CALCULATION METHODS
The authors have no conflicts to disclose. The data used in this paper are obtained by solving the Euler
equations based on the rhoCentralFoam solver of OpenFOAM,
which is a widely accepted solver in the field of supersonic flow.
Author Contributions
The second-order total variation diminishing (TVD) scheme is used
Tong Zhao: Conceptualization (equal); Data curation (equal); and a minmod limiter is added to alleviate the oscillation near the
Methodology (equal); Software (equal); Validation (equal); Writing – discontinuity. The courant number (CFL) number is limited to 0.1.
original draft (equal). Jian An: Data curation (equal); Formal analysis The Sod shock tube and the Lax shock tube are the most widely
(equal); Methodology (equal); Validation (equal); Writing – review & used cases for verifying the accuracy of the solver. Both of them are
editing (equal). Yuming Xu: Conceptualization (equal); Data curation 1 D Riemann problems, which have analytical solutions under fixed
(equal); Investigation (equal); Methodology (equal); Writing – review boundary conditions, described by Euler equations as
& editing (equal). Guoqiang He: Data curation (equal); Formal analy- 0 1 0 1
q qu
sis (equal); Validation (equal). Fei Qin: Funding acquisition (equal); @@ A @ @ 2 q 1
qu þ qu þ p A ¼ 0; E ¼ þ qu2 : (A1)
Project administration (equal); Visualization (equal); Writing – review @t @x c1 2
E uðE þ pÞ
& editing (equal).

17 July 2024 12:14:16

FIG. 12. The verification of solver. (a) Lax
problem. (b) Sod problem.

FIG. 13. Comparison of three grids. (a)

Forward step flow and (b) three-wedge
compression ramp flow.

Phys. Fluids 36, 066118 (2024); doi: 10.1063/5.0206515 36, 066118-12

Published under an exclusive license by AIP Publishing
 Physics of Fluids ARTICLE pubs.aip.org/aip/pof

TABLE VI. Effect of neural network hyperparameter selection on training results.

No. Output ERRORN ERRORP

Network training failure

Network training failure

17 July 2024 12:14:16

5

Phys. Fluids 36, 066118 (2024); doi: 10.1063/5.0206515 36, 066118-13

Published under an exclusive license by AIP Publishing
 Physics of Fluids ARTICLE pubs.aip.org/aip/pof

17 July 2024 12:14:16

FIG. 14. Results for a training data volume of 3. (a) Data driven, (b) SPC, (c) error of data driven, (d) error of SPC, and (e) shock wave position errors.

FIG. 15. Results for a training data volume of 4. (a) Data driven, (b) SPC, (c) error of data driven, (d) error of SPC, and (e) shock wave position errors.

Phys. Fluids 36, 066118 (2024); doi: 10.1063/5.0206515 36, 066118-14

Published under an exclusive license by AIP Publishing
 Physics of Fluids ARTICLE pubs.aip.org/aip/pof

17 July 2024 12:14:16

FIG. 16. Results for a training data volume of 6. (a) Data driven, (b) SPC, (c) error of data driven, (d) error of SPC, and (e) shock wave position errors.

FIG. 17. Results for a training data volume of 9. (a) Data driven, (b) SPC, (c) error of data driven, (d) error of SPC, and (e) shock wave position errors.

Phys. Fluids 36, 066118 (2024); doi: 10.1063/5.0206515 36, 066118-15

Published under an exclusive license by AIP Publishing
 Physics of Fluids ARTICLE pubs.aip.org/aip/pof

The initial conditions of the Lax problem are 6

J. P. Thomas, E. H. Dowell, and K. C. Hall, “Three-dimensional transonic aero-
0 1 0 1 0 1 0 1 elasticity using proper orthogonal decomposition-based reduced-order mod-
q 0:445 q 0:5 els,” J. Aircr. 40(3), 544–551 (2003).
@uA ¼ @ 0:698 A; @ u A ¼ @ 0:0 A; (A2)
7
S. Baker, X. Fang, L. Shen, C. Willman, J. Fernandes, F. Leach, and M. Davy,
“Dynamic mode decomposition for the comparison of engine in-cylinder flow
p x<0:5 3:528 p x>0:5 0:571
fields from particle image velocimetry (PIV) and Reynolds-averaged Navier–
Stokes (RANS) simulations,” Flow, Turbul. Combust. 111, 115 (2023).
and the initial conditions of the Sod problem are 8
W.-X. Ren and H.-B. Chen, “Finite element model updating in structural
0 1 0 1 0 1 0 1 dynamics by using the response surface method,” Eng. Struct. 32(8), 2455–
q 1:0 q 0:125
@uA 2465 (2010).
¼ @ 0:0 A; @ u A ¼ @ 0:0 A: (A3) 9
S. Kazem, M. S. Tameh, and M. M. Rashidi, “An improvement to the unsteady
p x<0:5 1:0 p x>0:5 0:1 MHD rotating flow over a rotating sphere near the equator via two radial basis
function schemes,” Eur. Phys. J. Plus 134(12), 611 (2019).
The verification results of the solver are shown in Fig. 12. It 10
H. Csala, S. T. M. Dawson, and A. Arzani, “Comparing different nonlinear
can be seen from the figure that the solver used can accurately dimensionality reduction techniques for data-driven unsteady fluid flow model-
capture the flow field discontinuity and has high calculation accu- ing,” Phys. Fluids 34(11), 117119 (2022).
racy, which can be used to generate the data required for SPC
11
J. Wen, W. Zhu, X. Jia, F. Ma, and Q. Liu, “Spectral domain graph convolutional
deep neural network for predicting unsteady and nonlinear flows,” Phys. Fluids
training.
35(9), 095107 (2023).
Three grids with 40 000 (grid-1), 48 000 (grid-2), and 57 000 12
S. L. Brunton, “Applying machine learning to study fluid mechanics,” Acta
(grid-3) cells, respectively, were used simulation to verify the grid Mech. Sin. 37(12), 1718–1726 (2021).
resolution for forward step flow. Figure 13(a) displays the computed 13
D. Drikakis and F. Sofos, “Can artificial intelligence accelerate fluid mechanics
results of the velocity near the upper wall on three different grids. research?,” Fluids 8(7), 212 (2023).
There is minimal difference in the results among the grids, indicat-
14
Y. LeCun, Y. Bengio, and G. Hinton, “Deep learning,” Nature 521(7553), 436–
444 (2015).
ing grid independence. However, 35 000 (grid-1), 42 000 (grid-2), 15
K. Duraisamy, Z. J. Zhang, and A. P. Singh, “New approaches in turbulence and
and 49 000 (grid-3) cells are used to verify the grid independence of transition modeling using data-driven techniques,” AIAA Paper No, 2015-1284,
compression ramp flow as shown in Fig. 13(b), and they also dem- 2015.
onstrate good agreement. To conserve computational resources, this
16
B. D. Tracey, K. Duraisamy, and J. J. Alonso, “A machine learning strategy to
study utilizes grid-1 for all computations. assist turbulence model development,” AIAA Paper No. 2015-1287, 2015.
17
R. Zangeneh, “Data-driven model for improving wall-modeled large-eddy simu-

17 July 2024 12:14:16

lation of supersonic turbulent flows with separation,” Phys. Fluids 33(12),
126103 (2021).
APPENDIX B: HYPERPARAMETRIC EXPERIMENTS 18
Z. J. Zhang and K. Duraisamy, “Machine learning methods for data-driven tur-
The setting of hyperparameters is important for the training of bulence modeling,” AIAA Paper No. 2015-2460, 2015.
19
S. Bhushan, G. W. Burgreen, W. Brewer, and I. D. Dettwiller, “Assessment of
neural networks. We conduct orthogonal experiments on four neural network augmented Reynolds averaged Navier Stokes turbulence model
important hyperparameters in convolutional neural networks. A in extrapolation modes,” Phys. Fluids 35(5), 055129 (2023).
total of nine different parameter combinations are configured, and 20
M. Kurz and A. Beck, “A machine learning framework for LES closure terms,”
the results of the network training with each parameter setting are Electron. Trans. Numer. Anal. 56, 117–137 (2022).
displayed in Table VI.
21
J. Kim and C. Lee, “Deep unsupervised learning of turbulence for inflow genera-
tion at various Reynolds numbers,” J. Comput. Phys. 406, 109216 (2020).
22
Y. Li, T. Liu, Y. Wang, and Y. Xie, “Deep learning based real-time energy
extraction system modeling for flapping foil,” Energy 246, 123390 (2022).
APPENDIX C: EFFECT OF DATA VOLUME 23
Z. M. Nikolaou, C. Chrysostomou, L. Vervisch, and S. Cant, “Progress variable
Figures 14–17 verify the prediction ability of the data-driven variance and filtered rate modelling using convolutional neural networks and
flamelet methods,” Flow, Turbul. Combust. 103(2), 485–501 (2019).
model and SPC for supersonic flow field under different data vol- 24
H. Chen, M. Guo, Y. Tian, J. Le, H. Zhang, and F. Zhong, “Intelligent recon-
umes. It can be seen that the addition of physical constraints struction of the flow field in a supersonic combustor based on deep learning,”
reduces the demand for data volume in model training. Phys. Fluids 34(3), 035128 (2022).
25
J. Ren, H. Wang, G. Chen, K. Luo, and J. Fan, “Predictive models for flame evo-
lution using machine learning: A priori assessment in turbulent flames without
REFERENCES
and with mean shear,” Phys. Fluids 33(5), 055113 (2021).
1
J. Lumley, “The structure of inhomogeneous turbulence,” in Proceedings of the 26
A. Seltz, P. Domingo, L. Vervisch, and Z. M. Nikolaou, “Direct mapping from
International Colloquium on the Fine Scale Structure of the Atmosphere and LES resolved scales to filtered-flame generated manifolds using convolutional
Its Influence on Radio Wave Propagation (1967). neural networks,” Combust. Flame 210, 71–82 (2019).
2
P. Schmid and J. Sesterhenn, “Dynamic mode decomposition of numerical and 27
M. Raissi, P. Perdikaris, and G. E. Karniadakis, “Physics-informed neural net-
experimental data,” J. Fluid Mech. 656, 5 (2008). works: A deep learning framework for solving forward and inverse problems
3
Y. Fu, X. Lin, L. Li, Q. Chu, H. Liu, X. Zheng, C.-H. Liu, Z. Chen, C. Lin, T. K. involving nonlinear partial differential equations,” J. Comput. Phys. 378, 686–
T. Tse, and C. Y. Li, “A POD-DMD augmented procedure to isolating domi- 707 (2019).
nant flow field features in a street canyon,” Phys. Fluids 35(2), 025112 (2023). 28
M. Raissi, Z. Wang, M. S. Triantafyllou, and G. E. Karniadakis, “Deep learning
4
B. Zhang, “Nonlinear mode decomposition via physics-assimilated convolu- of vortex-induced vibrations,” J. Fluid Mech. 861, 119–137 (2019).
tional autoencoder for unsteady flows over an airfoil,” Phys. Fluids 35(9), 29
C. Rao, H. Sun, and Y. Liu, “Physics-informed deep learning for computational
095115 (2023). elastodynamics without labeled data,” J. Eng. Mech. 147(8), 04021043 (2021).
5
D. Wang, F. Xie, T. Ji, X. Zhang, Y. Lu, and Y. Zheng, “Prediction of wind shear 30
K. Harada, D. Rajaram, and D. N. Mavris, “Application of multi-fidelity phys-
layer for dynamic soaring by using proper orthogonal decomposition and long ics-informed neural network on transonic airfoil using wind tunnel measure-
short term memory network,” Phys. Fluids 35(8), 085103 (2023). ments,” AIAA Paper No. 2022-0386, 2022.

Phys. Fluids 36, 066118 (2024); doi: 10.1063/5.0206515 36, 066118-16

Published under an exclusive license by AIP Publishing
 Physics of Fluids ARTICLE pubs.aip.org/aip/pof

31
H. Gao, L. Sun, and J.-X. Wang, “PhyGeoNet: Physics-informed geometry- 41
C. Kong, J. Chang, Z. Wang, Y. Li, and W. Bao, “Data-driven super-resolution
adaptive convolutional neural networks for solving parameterized steady-state reconstruction of supersonic flow field by convolutional neural networks,” AIP
PDEs on irregular domain,” J. Comput. Phys. 428, 110079 (2021). Adv. 11(6), 065321 (2021).
32
J.-X. Wang, J. Wu, J. Ling, G. Iaccarino, and H. Xiao, “A comprehensive 42
C. Kong, C. Zhang, Z. Wang, Y. Li, and J. Chang, “Efficient prediction of super-
physics-informed machine learning framework for predictive turbulence sonic flowfield in an isolator based on pressure sequence,” AIAA J. 60(5),
modeling,” arXiv:1701.07102 (2018). 2826–2835 (2022).
33
S. Cai, Z. Mao, Z. Wang, M. Yin, and G. E. Karniadakis, “Physics-informed 43
D. Meng, M. Shi, Y. Shi, and Y. Zhu, “A machine learning method for transi-
neural networks (PINNs) for fluid mechanics: A review,” Acta Mech. Sin. tion prediction in hypersonic flows over a cone with angles of attack,” AIP
37(12), 1727–1738 (2021). Adv. 12(2), 025116 (2022).
34
P. Ren, C. Rao, Y. Liu, J.-X. Wang, and H. Sun, “PhyCRNet: Physics-informed 44
Z. Mao, A. D. Jagtap, and G. E. Karniadakis, “Physics-informed neural networks
convolutional-recurrent network for solving spatiotemporal PDEs,” Comput. for high-speed flows,” Comput. Methods Appl. Mech. Eng. 360, 112789 (2020).
Methods Appl. Mech. Eng. 389, 114399 (2022). 45
L. Liu, S. Liu, H. Xie, F. Xiong, T. Yu, M. Xiao, L. Liu, and H. Yong,
35
E. McGowan, V. Gawade, and W. (Grace) Guo, “A physics-informed convolu- “Discontinuity computing using physics-informed neural network,” J. Sci.
tional neural network with custom loss functions for porosity prediction in Comput. 98, 22 (2024).
laser metal deposition,” Sensors 22(2), 494 (2022). 46
Z. Gao, L. Yan, and T. Zhou, “Failure-informed adaptive sampling for PINNs,”
36
X. Zhao, Z. Gong, Y. Zhang, W. Yao, and X. Chen, “Physics-informed convolu- SIAM J. Sci. Comput. 45, A1971–A1994 (2023).
tional neural networks for temperature field prediction of heat source layout 47
L. Lu, X. Meng, Z. Mao, and G. E. Karniadakis, “DeepXDE: A deep learning
without labeled data,” Eng. Appl. Artif. Intell. 117, 105516 (2023). library for solving differential equations,” SIAM Rev. 63(1), 208–228 (2021).
37
R. Zhai, D. Yin, and G. Pang, “A deep learning framework for solving forward 48
O. Ronneberger, P. Fischer, and T. Brox, “U-Net: Convolutional networks for bio-
and inverse problems of power-law fluids,” Phys. Fluids 35(9), 093115 (2023). medical image segmentation,” in Medical Image Computing and Computer-Assisted
38
A. Zanjani, A. M. Tahsini, K. Sadafi, and F. Ghavidel Mangodeh, “Shape opti- Intervention – MICCAI 2015, edited by N. Navab, J. Hornegger, W. M. Wells, and
mization and flow analysis of supersonic nozzles using deep learning,” Int. J. A. F. Frangi (Springer International Publishing, Cham, 2015), pp. 234–241.
Comput. Fluid Dyn. 36(10), 875–891 (2022). 49
L. F. G. Marcantoni, J. P. Tamagno, and S. A. Elaskar, “High speed flow simula-
39
M.-Y. Wu, J.-Z. Peng, Z.-M. Qiu, Z.-H. Chen, Y.-B. Li, and W.-T. Wu, tion using openfoam,” Mec. Comput. 31(16), 2939–2959 (2012), available at
“Computationally effective estimation of supersonic flow field around airfoils https://scholar.google.com/scholar_lookup?title=High%20speed%20flow%20simulation
using sparse convolutional neural network,” Fluid Dyn. Res. 55(3), 035504 %20using%20OpenFOAM&author=L.%20Gutierrez%20Marcantoni&author=J.
(2023). %20Tamagno&author=S.%20Elaskar&publication_year=2012&journal=&volume
40
C. Fujio and H. Ogawa, “Deep-learning prediction and uncertainty quantifi- =&pages=2939-2959.
cation for scramjet intake flowfields,” Aerosp. Sci. Technol. 130, 107931 50
D. P. Kingma and J. Ba, “Adam: A method for stochastic optimization,”
(2022). arXiv:1412.6980 (2017).

17 July 2024 12:14:16

Phys. Fluids 36, 066118 (2024); doi: 10.1063/5.0206515 36, 066118-17

Published under an exclusive license by AIP Publishing