AIAA JOURNAL

Turbulence Modeling via Data Assimilation and Machine

Learning for Separated Flows over Airfoils

XiangLin Shan,∗ YiLang Liu,† WenBo Cao,‡ XuXiang Sun,§ and WeiWei Zhang¶
Northwestern Polytechnical University, 710072 Xi’an, People’s Republic of China
https://doi.org/10.2514/1.J062711
Reynolds-averaged Navier–Stokes (RANS) models, which are known for their efficiency and robustness, are widely
used in engineering applications. However, RANS models do not provide satisfactory predictive accuracy in many
engineering-relevant flows with separation. Aiming at the difficulties of turbulence modeling for separated flows at
high Reynolds number, this paper constructs turbulence models using data assimilation technique and deep neural
network (DNN). Due to the uncertainty of traditional turbulence models, the parameters of Spalart–Allmaras (SA)
turbulence model are optimized with experimental data to provide high-fidelity flowfields. Then DNN model maps the
mean flow variables to eddy viscosity and replaces the SA model to be embedded within a RANS solver by iterative
mode. Different from many existing studies, this DNN model does not depend on traditional turbulence models during
the simulation process. This approach is applied to turbulent attached and separated flows and can significantly
Downloaded by Tongji University on August 2, 2023 | http://arc.aiaa.org | DOI: 10.2514/1.J062711

improve the accuracy for new flow conditions and airfoil shapes. Results show that the mean relative error of lift
coefficient above the stall decreases by over 57% for all the airfoils.

Nomenclature simulation (LES), are overly expensive and time-consuming. There-

Cd = drag coefficient fore, from the view of affordability, Reynolds-averaged Navier–Stokes
Cl = lift coefficient (RANS) equations with turbulence models (e.g., k-ε [1], k-ω [2],
Cp = pressure coefficient Spalart–Allmaras (SA) [3], Shear Stress Transport (SST) model [4],
and Reynolds-stress models [5,6]) will continue to remain the critical
d = wall distance
approaches for the foreseeable future [7,8]. However, RANS models
k = turbulent kinetic energy, m2 ∕s2
are based on many assumptions, which increase the uncertainty and
p = mean pressure, Pa
R = rotation rate tensor, 1∕s limitation, and do not provide satisfactory predictive accuracy in many
Re = Reynolds number complex flows with streamline curvature, strong adverse pressure
S = strain rate tensor, 1∕s gradients, and separation [9]. Hence, to improve the accuracy of
u i = mean velocity, m/s simulating turbulent flows, it is significant to establish a robust and
V = velocity vector, m/s accurate turbulence model.
xi = Cartesian coordinates, m Recently, data-driven methods have been increasingly applied to
α = angle of attack, deg turbulence modeling with the help of machine learning (ML) and
δij = Kronecker symbol high-fidelity data [10,11]. Generally, aiming to improve the accuracy
of RANS, research on data-driven turbulence modeling can be
ν = kinematic molecular viscosity, m2 ∕s
divided into the following two categories.
νt = kinematic eddy viscosity, m2 ∕s
1) Modify the traditional turbulence models. Cheung et al. [12]
ρ = density, kg∕m3
applied Bayesian inference to quantify the uncertainty of SA turbu-
τij = Reynolds stress, m2 ∕s2
lence model and correct the parameters. Similarly, Ray et al. [13,14]
studied the parameter uncertainties of k-ε models and applied cali-
I. Introduction brated parameters to the jet in the crossflow. Further studies about
calibrating model parameters are concluded in [10]. Insufficiently,
T URBULENCE widely appears in natural and industrial envi-
ronments, and, at the same time, is the most complicated kind of
the fluid motion. Nowadays, computational fluid dynamics (CFD)
the calibrated model parameters are not identical for different flows.
To overcome this defect, Duraisamy et al. [15], Singh et al. [16–18],
simulation and experimental measurement are the two main mea- and Parish and Duraisamy [19] put forward a data-driven modeling
sures for engineering problems involving turbulence. However, both framework of field inversion and machine learning (FIML). They
high-Reynolds-number experiment and high-fidelity numerical sim- modified the RANS model by a spatially varying correction factor
ulation, such as direct numerical simulation (DNS) and large-eddy acquired by the adjoint-based method. Then they used ML to predict
the correction factor and obtained improved predictions in various
similar flows. Yang and Xiao [20] operated regularizing ensemble
Received 7 December 2022; revision received 22 April 2023; accepted for Kalman filtering to solve the inverse problem in FIML and reduced
publication 18 June 2023; published online 28 July 2023. Copyright © 2023
the complexity of code modification. The flow inversion and ML are
by the American Institute of Aeronautics and Astronautics, Inc. All rights
reserved. All requests for copying and permission to reprint should be sub-
successively implemented in the above studies. Differently, Holland
mitted to CCC at www.copyright.com; employ the eISSN 1533-385X to [21] integrated ML in flow inversion and promoted improved gen-
initiate your request. See also AIAA Rights and Permissions www.aiaa.org/ eralization.
randp. 2) Construct data-driven models to replace the traditional turbu-
*Graduate Student, School of Aeronautics, International Joint Institute of lence models. An ML method aiming at general closures must
Artificial Intelligence on Fluid Mechanics; [email protected]. possess the objective principles of physical relations (realizability
†
Associate Professor, School of Aeronautics, International Joint Institute of [22], invariance under rotations or reflections [23], etc.). Researchers
Artificial Intelligence on Fluid Mechanics; [email protected]. mainly adopted the following two strategies to realize these objective
‡
Graduate Student, School of Aeronautics, International Joint Institute of principles.
Artificial Intelligence on Fluid Mechanics; [email protected].
§
Graduate Student, School of Aeronautics, International Joint Institute of a) Represent Reynolds stresses by linear or nonlinear eddy-
Artificial Intelligence on Fluid Mechanics; [email protected]. viscosity hypothesis. For elementary research of high-Reynolds-
¶
Professor, School of Aeronautics, International Joint Institute of Artificial number turbulence models, RANS data are adopted in many
Intelligence on Fluid Mechanics; [email protected] (Corresponding existing studies. Zhu et al. [24,25] and Sun et al. [26] established
Author). a data-driven linear eddy-viscosity model trained by SA data and
Article in Advance / 1
 2 Article in Advance / SHAN ET AL.

adopted an iterative mode between data-driven models and RANS i.e., FIML framework [16–19,45–47]. In this paper, we adopt this
solvers. Follow-up studies illustrated that this coupling mode can idea and employ the DA technique for flow inversion. Meanwhile,
effectively improve the convergence compared with frozen coupling to be independent of traditional models, our data-driven model
mode, particularly at high Reynolds number [27]. Furthermore, a maps the mean flow variables to eddy viscosity and replaces the
typical framework for data-driven nonlinear eddy-viscosity model is traditional turbulence models. Our framework is tested on three
the tensor basis neural network (TBNN). Considering Pope’s work wind-turbine airfoils of different Reynolds numbers and angles of
[28] that Reynolds stresses can be represented as the linear combi- attack and provides improved accuracy and generalization.
nation of 10 basis tensors, Ling et al. [29–31] presented TBNN and The rest of the paper is organized as follows. Section II introduces
injected the predicted Reynolds stress anisotropic tensor into RANS the research method, including the general framework of data-driven
solver to obtain the improved velocity field. Following the works of turbulence modeling, DA technique, and ML method. In Sec. III, this
Ling et al., Jiang et al. [32] introduced a new turbulent timescale to approach is applied to turbulent flows over the wind-turbine airfoils.
prevent nonunique mappings. Recently, Zhang et al. [33] applied an Section IV gives a discussion about the method used in this paper.
ensemble Kalman method to train TBNN by indirect observation Finally, conclusions are drawn in Sec. IV.
data. In addition, Weatheritt and Sandberg [34,35] and Zhao et al.
[36] used gene expression programming to establish the explicit
expression of nonlinear eddy-viscosity model. II. Methodology
b) Represent Reynolds stresses by eigenvalues and eigenvec- A. General Framework of Data-Driven Turbulence Modeling
tors. This idea was applied to quantifying the uncertainties in The incompressible RANS equations are given by
RANS models earlier [37–39]. In recent years, Wang et al. [40]
represented Reynolds-stress discrepancies between RANS and ∂u i
DNS as the discrepancies in six Galilean invariant quantities
0
∂xi
Downloaded by Tongji University on August 2, 2023 | http://arc.aiaa.org | DOI: 10.2514/1.J062711

acquired by eigendecomposition. Yin et al. [41] employed a

similar method and summarized feature selection criteria. More-
∂u i ∂u i u j 1 ∂p ∂2 u i ∂ −ui0 uj0
over, considering the poor convergence of the Reynolds-stress 
− ν  (1)
model constructed above, Wu et al. [42,43] decomposed Reyn- ∂t ∂xj ρ ∂xi ∂xj ∂xj ∂xj
olds stress into linear term and nonlinear terms and overcame the
ill-conditioning of RANS equations to some degree. On this where  and  0 denote mean and fluctuation, respectively. Reynolds-
basis, McConkey et al. [44] added the non-negativity constraint stress tensor τij
−ui0 uj0 is the unclosed term that needs to be modeled.
of linear terms and promoted numerical stability. The linear eddy-viscosity hypothesis is commonly used to transform
In conclusion, for the first category, data assimilation (DA) methods, the unclosed problem of Reynolds stresses into the unclosed problem of
such as adjoint-based method and ensemble Kalman filtering method, kinematic eddy viscosity νt, as shown in Eq. (2), where 2∕3kδij
are adopted for field inversion to obtain the spatial distribution of denotes the isotropic part of the tensor.
correction factor, and then ML methods, such as random forest and
deep neural network (DNN), are used to construct a mapping function ∂ui ∂uj 2
between flow variables and the correction factor. However, these meth- −ui0 uj0
νt  − kδij (2)
∂xi ∂xi 3
ods usually depend on the traditional turbulence models, which contain
many uncertainties. For the second category, data-driven Reynolds- In this section, we introduce the general data-driven framework for
stress constitutive relation is widely explored. Compared with the RANS closure models. As shown in Fig. 1, the procedure of data-driven
traditional turbulence model, the accuracy of the data-driven model is turbulence modeling is divided into three parts: dataset acquisition,
greatly improved, but there are many bottlenecks. For instance, ML, and iterative calculation. Dataset contains high-fidelity numerical
high-fidelity data of DNS/LES are limited, especially at high Reyn- simulation data (DNS, LES, etc.) and experimental data. Numerical
olds number. Furthermore, convergence and stability of coupling simulation can directly provide the flowfield, including the average
calculation are the major points for the applications of new models, field and fluctuating field, and the Reynolds stresses can be acquired by
particularly for Reynolds-stress models. time averaging. The experimental data only contain limited data and the
Combining the strengths of previous studies, this paper proposes a flowfield data can be obtained by DA, which is used in this paper
method of turbulence modeling, which can augment the traditional (Sec. II.B).
SA turbulence model and improve the accuracy for high-Reynolds- Modeling process mainly includes two parts: constructing flow
number separated flows. In existing studies, for high-Reynolds- features from the mean flow variables and training the ML model.
number turbulent flows, researchers mainly use experimental data, The flow features are taken as the input of ML model and the

Fig. 1 General framework of data-driven turbulence modeling.

 Article in Advance / SHAN ET AL. 3

unclosed turbulent quantities (Reynolds stresses or eddy viscosity)

are taken as the output. The ML model in this paper is DNN and
unclosed turbulent quantity is eddy viscosity (Sec. II.C).
Finally, ML model is embedded in the CFD solver and the coupling
calculation between ML model and RANS equations is carried out.
There are two coupling modes between ML model and RANS equa-
tions: iterative mode and frozen mode. For iterative mode, ML model
plays the role of the traditional algebraic eddy-viscosity model. In
each iteration step, the average flow variables calculated by the RANS
equations are transferred to the ML model, and the turbulent quantities
calculated by the ML model are returned to the RANS equations. The
coupling computation is not over until the given convergence tolerance
is reached. On the other hand, for frozen mode, ML model maps the
mean flow features computed by the baseline RANS model to the high- Fig. 2 Structure of DNN.
fidelity turbulent quantities. While predicting, the baseline RANS
simulation is performed to acquire the baseline mean features. The
model is implemented once, and the predicted quantities correspond to neural layers. The first layer and the last layer are called input layer
the high-fidelity turbulent quantities, which are then substituted into and output layer, respectively, and the middle layers are called hidden
the RANS equations and frozen until convergence. Liu et al. [27] layers.
compared the convergence and numerical stability of these two modes The different layers are connected by the nonlinear transformation
and indicated that the convergence of iterative mode is better for represented by
Downloaded by Tongji University on August 2, 2023 | http://arc.aiaa.org | DOI: 10.2514/1.J062711

separated flow at high Reynolds number. Considering the convergence

problem, we adopt the iterative mode. h
gW ⋅ X  b (4)

B. Data Assimilation where X represents the input vector, and it is also the output vector of
DA technique could integrate the limited measurement data and the previous layer. W and b represent the weight matrix and bias vector,
numerical modeling, augment the turbulence model prediction, and respectively. The nonlinearity is mainly introduced by the activation
recover high-fidelity turbulent flowfield quickly and accurately [48,49]. function g⋅, including sigmoid, tanh, ReLU, and LeakyRelu. The
DA technique applied to turbulence may be classified into two cate- activation function adopted in this work is LeakyRelu:
gories: 1) solving the high-order inverse problems by adjoint-based
method, ensemble Kalman filtering method, etc. [20,33,50–53], and x; x≥0
LeakyReLUx
(5)
2) calibrating the turbulence model parameters [12–14,54–58]. All λx x<0
these methods can help to obtain spatial distribution of high-fidelity
data. The requirement is to acquire the high-fidelity flowfield based on where λ is an adjustable hyperparameter. Furthermore, for enforcing
experimental data. The method of the second category is adopted in this positivity of eddy viscosity, ReLU function is used in the output layer.
paper, and an effective optimization algorithm is employed to calibrate ReLU function is defined by
the parameters of turbulence models. We choose the SA model [3] as a
baseline turbulence model as shown in Appendix A. x; x≥0
The parameters of SA model θ
Ct3 ; Ct4 ; Cb1 ; Cb2 ; σ; Cv1 ; ReLUx
(6)
κ; Cw2 ; Cw3  are taken as optimization variables. For the high-fidelity 0 x<0
measurement data, experimental surface pressure distribution of air-
foils is adopted as high-fidelity data. We aim to minimize the mean The form of ReLU function guarantees that all outputs are positive.
square error of pressure distribution calculated by SA model of new The main hyperparameters of DNN are width (number of neurons
parameters. Therefore, the optimization problem can be expressed as in each layer), depth (number of hidden layers), learning rate, and
follows: weight of L2-norm regularization, i.e., λ2 in Eq. (12). All these
hyperparameters are optimized by Bayesian optimization process
1 n
2 [63]. Bayesian optimization is a method of using Bayesian theorem
min fθ
CipCFD θ − CipEXP (3) to guide the search to find the minimum or maximum value of the
θ n i
1 objective function, i.e., at each iteration, using previously observed
historical information for the next optimization step. Compared with
where n is the number of measurement points on the airfoil surface, many gradient-based optimization techniques, Bayesian optimiza-
CipCFD θ is the pressure coefficients corresponding to a set of SA tion requires relatively fewer model evaluations and is therefore
model parameters, and CipEXP is the experimental pressure coef- advantageous when the computational cost of evaluating the objec-
ficients. To solve this optimization problem, this paper employs tive function is high, such as training a neural network.
teaching–learning-based optimization (TLBO) [59,60]. TLBO is a
population-based meta-heuristic optimization technique that sim- 2. Features Construction
ulates the environment of a classroom to optimize a given objective Since the input features of the ML strongly affect the accuracy and
function. The hyperparameter of TLBO is the size of population, generalization of the model, feature construction is one of the most
i.e., number of calling RANS per generation, and maximum num- important processes of ML. To construct a turbulence closure model,
ber of iterations. Meanwhile, TLBO is a global optimization algo- the input features must completely come from the mean flowfield and
rithm, which is widely used in many engineering optimization reflect the turbulence characteristics. At the same time, the process of
problems [61,62]. After solving the optimization above, the flow- feature construction should also include the physical properties of
field data can be obtained by RANS simulation with the optimized turbulence and ensure invariance properties, i.e., invariant under
SA model, which is called DA model in the following. rotational and reflectional transformations and Galilean transforma-
tion of the coordinate system. As shown in Table 1, there are six input
C. Machine Learning features in this paper.
1. Deep Neural Network Note that q1 is the mixing velocity, umix , which is calculated by the
As one of the most well-known methods in ML, DNN possesses a mixing length theory, i.e.,
strong capability to approximate nonlinear systems, which benefits
from its structure as shown in Fig. 2. DNN is composed of multiple umix
lmix kRk (7)
 4 Article in Advance / SHAN ET AL.

Table 1 Input features of DNN number is Rex
Rec x∕c. The concept mixing length was pro-
posed by Prandtl and is similar to the molecular mean free path in gas
Feature Description Sign
dynamics [64]. Mixing velocity can be used to evaluate the velocity
q1 Mixing velocity umix scale in turbulent boundary layer, and we treat mixing velocity as an
q2 Rotation rate kRk input feature to approximate wall turbulence effect. Note that q2 and
q3 Strain rate kSk q3 are rotation rate and strain rate, respectively, which are always
q4 Distance transformation d2 kRk1 − tanhd used in traditional turbulence models, such as SA model; q4 and q5
of rotation rate are distance transformation of rotation rate and distance transforma-
q5 Distance transformation d2 kSk1 − tanhd tion of strain rate, respectively. Here we use an empirical function of
of strain rate wall distance, d2 1 − tanhd, illustrated in [26], to approximate the
q6 Pressure gradient ∂p ∂p
∂xi ∂xi distribution trend of eddy viscosity. Note that q6 is pressure gradient
and used to represent the effect of inverse pressure gradient on flow
Note that k ⋅ k and j ⋅ j indicate matrix Frobenius norm and vector separation. Contour plots of these input features are shown in Fig. 3.
2-norm, respectively, and γ is 1.4. As for the output of DNN model, obviously, eddy viscosity is
invariant under rotational and reflectional transformations and Gal-
ilean transformation. However, because the boundary layer is thin at
where lmix is the mixing length, calculated by high Reynolds number, eddy viscosity near the wall spans various
orders of magnitude and it is difficult to model the eddy viscosity
 ∕A
lmix
min κy 1 − ey ; κδ (8) directly. Therefore, we adopt a function referring to the SA model to
transform the eddy viscosity near the wall as follows:
Downloaded by Tongji University on August 2, 2023 | http://arc.aiaa.org | DOI: 10.2514/1.J062711

0; x<0 χ3 ν~
νt
fv ν~ ; fv
; χ
(10)
χ3  c3v1 ν
0.37x
; 0≤x≤c
δ
Re1∕5
x (9) where cv1
7.1. For any νt, ν~ can be obtained by solving a quartic
0.37c equation and reserving its unique positive real number solution. In the
; x≥c near-wall region, the magnitude of eddy viscosity is small, and the
Re1∕5
c approximate relationship between νt and ν~ is ν~ ∼ νt 1∕4 . Therefore,
the transformed eddy viscosity ν~ is output of DNN model and can
where c is the airfoil chord, and κ
0.41 and A
26. The Reyn- effectively reduce the impact of excessive magnitude change on the
olds number based on airfoil chord is Rec , and the local Reynolds accuracy of DNN model.

Fig. 3 Contour plots of the input features of S809 airfoil at Re
2 × 106 and α
14.2°.
 Article in Advance / SHAN ET AL. 5

3. Training Processing After DA, high-fidelity flowfield data can be obtained by operating
We apply PyTorch [65] to implement the training process of DNN CFD with optimized SA parameters. Then DA data for separated flow
model. Before training, the value of input and output features are and SA data for attached flow are used for ML, as shown in the orange
normalized using their own minimum and maximum by box in Fig. 4. Input and output features described in Sec. II.C.2 can be
constructed from the dataset. Next, Bayesian optimization process
x − xmin is performed to optimize the hyperparameters of the DNN model.
x^
(11)
xmax − xmin Finally, with optimized hyperparameters, the DNN model is trained
and used for the coupling calculation.
This approach can easily scale all the data to [−1; 1] linearly and does
not change the distribution of the data. III. Results
The loss function is defined by A. Dataset and CFD Solver
Turbulence modeling approach is tested by the turbulent flow over
1 N
2 three wind-turbine airfoils at multiple Reynolds numbers, and the
Loss
λ1 y^ i − yi   λ2 kWk (12)
N i
1
airfoil shapes are displayed in Fig. 5. The lift coefficient and surface
pressure distribution of S809/S805/S814 airfoil can be obtained from
where N is the number of training data, y^ i is the predicted value, yi is the existing experimental data [66–68]. Experimental lift coefficients
the real value, and W is the weight in DNN. The multiplier λ1 is set as in this paper are in a fixed transition condition. Because engineering
1.0 and λ2 is a hyperparameter to be optimized. The first term of the demands are to acquire as much information as possible through
a small number of experiments, the pressure distribution of a few
loss function is the mean square error (MSE), and the second term is
angles of attack is used for DA in this work.
L2-norm regularization of the weight in the DNN model, which can
Downloaded by Tongji University on August 2, 2023 | http://arc.aiaa.org | DOI: 10.2514/1.J062711

Numerical cases are simulated by an in-house hybrid unstructured

improve the model robustness and prevent overfitting. In addition, we
CFD code that solves the Reynolds-averaged Navier–Stokes (RANS)
use the AdamWoptimizer, and set the batch size of training data as 64, equations using a second-order cell-centered finite volume method.
which defines the number of training examples in one forward/ The SA model in this code has been verified in Refs. [24,25]. The
backward pass. inviscid fluxes are discretized by the Roe scheme, while the viscous
fluxes are calculated by a second-order central scheme, and the time-
D. Detailed Workflow of Data Assimilation and Machine Leaning marching method is an implicit Gauss–Seidel method. Moreover,
We conclude the main process of DA and ML in Fig. 4. First, DA is due to the possible difficult convergence of the coupling calculation
operated to generate high-fidelity flowfield data, and the detailed between the ML model and RANS solver, our CFD code integrated a
process is shown inside the blue box, as described in the following. novel convergence enhancement method, online dimensional reduc-
Step 1—Initializing/updating: If the current iteration step is one, tion optimization [69], to guarantee the convergence of all the test
the initial population is generated by fluctuating the initial SA para- cases.
meters (baseline), and each member of the population corresponds to
a set of SA parameters. There are nine adjustable SA parameters; i.e.,
the dimension of each member is nine. If the current iteration step is
greater than one, based on errors of each member of the population,
the TLBO algorithm is operated to update the population, i.e., SA
parameters of each member.
Step 2—CFD: Calculate each member by adjusting SA parameters
in the RANS solver.
Step 3—Evaluation: Evaluate the fitness value of each member of
the population, and fitness values are defined by the errors of pressure
distribution of airfoil surface between the experiment and CFD.
Step 4—Repeat steps 1–3 until TLBO converges.

Data Assimilation
Initialize/Update population
by TLBO
Fig. 5 Shape of three wind-turbine airfoils.
CFD

Evaluate errors between No

experiment and CFD

Converge
Yes

Machine DA data SA data

Learning
Construct input and output
features
Bayesian optimization
(hyper-parameters of DNN)

Train DNN model

Fig. 4 Process of data assimilation and machine learning. Fig. 6 Near-wall grid of S809.
 6 Article in Advance / SHAN ET AL.

The grid of S809 is displayed in Fig. 6. The total number of grid B. Results of Data Assimilation
elements is 34,205, the surface of the airfoil is arranged with 400 The DA technique is implemented by the method of Sec. II.B. The
points, and the y value is approximately 1. The grids of other airfoils flowfield data are obtained by numerical simulation with the DA
are generated with the same parameters as S809. model. Flow states and airfoils selected as DA cases are illustrated
in Table 2, and these four cases are denoted as DA1 − DA4 . During
optimization, the maximum fluctuation of the SA parameters is
Table 2 Flow conditions
and airfoils of DA cases
limited by 50% of the original parameters, and the size of the popu-
lation, i.e., the number of calling RANS per generation, is 20. The
Case Airfoil Re (×106 ) α, ° optimized parameters of DA models are presented in Table B1 of
Appendix B. Figure 7 shows the error of pressure coefficient versus
DA1 S809 2 12.2
iteration number for the TLBO optimization process of the DA2 case.
DA2 S809 2 14.2
The error of each member is defined by
DA3 S805 1 12.2
DA4 S805 1 15.2
1 n
2
error
CipCFD θ − CipEXP (13)
n i−1

where the variables are the same as in Eq. (3). From Fig. 7, the error of
the 10th generation decreases 90% compared with the error of the
initial SA model and the optimization converges at the 15th gen-
Downloaded by Tongji University on August 2, 2023 | http://arc.aiaa.org | DOI: 10.2514/1.J062711

eration.
Figure 8 compares pressure distribution calculated by the DA
model and the SA model. It can be seen that the DA model can
predict the pressure distribution more accurately than the SA model.
Table 3 gives the lift coefficient and relative error. It can be seen that
the error of the lift coefficient calculated by the DA model is greatly
reduced. Comparing the area of the pressure distribution curve, we
can see (Fig. 8) that the area of the SA model is larger than that of the
experiment, and the area of the DA model is slightly smaller than that
of the experiment, which is consistent with the trend of the lift
coefficient.
Fig. 7 The convergence history of TLBO optimization process. The From the perspective of flowfield, Figs. 9 and 10 compare the
horizontal axis represents the number of optimized generations, and the contour before and after DA. It can be seen that the eddy viscosity
vertical axis represents the lowest error of 20 members per generation. calculated by the DA model is larger in the wake area, and the

Fig. 8 Pressure distribution of four DA cases.

 Article in Advance / SHAN ET AL. 7

Table 3 Lift coefficient and relative error of SA model flow condition, which are not optimal in other flow conditions. To
and DA model compared with experimental data further test the applicability of the DA model, the lift curves of S809
at Re
1 × 106 and Re
2 × 106 calculated by the DA1 model are
α
12.2° α
14.2° α
12.2° α
15.2°
shown in Fig. 11 and compared with the SA model and the experi-
Parameter (S809) (S809) (S805) (S805)
ment. These results indicate that the DA model cannot accurately
Exp, CL 0.996 1.061 1.164 1.098 predict lift coefficients for either attached flows or separated flows.
SA, CL 1.278 1.264 1.352 1.370 To measure the mean relative error of the lift curves between
(relative error) (28.3%) (19.1%) (16.2%) (24.8%) different models, we define the following formula:
DA, CL 0.933 1.069 1.080 1.100
(relative error) (6.3%) (0.7%) (7.2%) (0.2%)
1 n
1 n
eRCFD
jCilEXP − CilCFD j CilEXP (14)
Exp, experimental data. n i
1
n i
1

where n is the number of calculation states, and eR and CilCFD

separation vortex is larger, which is finally reflected in the improve- correspond to the mean relative error of different models and lift
ment of the accuracy of pressure distribution and lift coefficient. coefficients of different models, respectively.
For the S809 at Re
1 × 106 (Fig. 11a), eR
0.1200 (SA
C. Applicability of DA Model for Different Flow Conditions model), eR
0.1225 (DA1 model), the mean relative error of the
Before training DNN, the applicability of the DA model is checked DA1 model is larger than the SA model. For the S809 at Re
2 ×
for a comparative analysis. As can be seen from Table B1, the result of 106 (Fig. 11b), eR
0.1341 (SA model), eR
0.0835 (DA1 model),
DA is a set of optimized turbulence model parameters under a single the mean relative error of the DA1 model is lower than the SA model,
Downloaded by Tongji University on August 2, 2023 | http://arc.aiaa.org | DOI: 10.2514/1.J062711

Fig. 9 Eddy viscosity field and streamline of S809 at Re
2 × 106 and α
14.2°.

Fig. 10 Streamwise velocity field of S809 at Re
2 × 106 and α
14.2°.

Fig. 11 Lift curves of S809 calculated by the DA1 model.

 8 Article in Advance / SHAN ET AL.

but the DA1 model cannot provide satisfactory accuracy for the sample selection method [24,70] is operated to select 5–30% data of
attached flows. Therefore, the DA1 model only provides high accu- the flowfield at each angle of attack (AOA). The ratio of DA data to
racy at Re
2 × 106 and α
14.2°, and cannot be generalized to SA data (η) is adjusted by different sampling rates, and the number of
other flow conditions. To construct a generalized turbulence model, samples is kept as constant at each AOA. The mean relative error of
DNN is applied to model the eddy viscosity, and the approach is the DNN model varies with η, as shown in Fig. 13. When η is too
tested on the following two examples. small, the error is large because of too few DA data, which leads to
large deviation of the separated flow. When η is too large, the error is
D. Results of Hyperparameter Optimization also large due to too few SA data, and the accuracy of the attached
flow decreases. The error is the smallest when η
1.62; hence η is
Using the training data in Table 4, we operate a Bayesian optimi- taken as 1.62 in the following cases. Furthermore, it can be seen that
zation process. The test set is randomly selected from 20% of all data when η is close to 0, the DNN error is close to the SA error (0.1341),
of α
3.1∕8.2∕12.2∕14.2 of the S809 airfoil. The rest data construct and when η is greater than 2.74, the DNN error is close to the DA1
training set and validation set, whose ratio is 8∶2. Before optimizing, error (0.0835). Then, the DNN model is trained with η
1.62, and
we set the optimization range of the hyperparameters; i.e., the depth the loss function value versus iteration step is shown in Fig. 14.
of DNN ranges from 1 to 8, the width of DNN ranges from 1 to 512, The lift curve calculated by DNN is displayed in Fig. 15a, compar-
the learning rate ranges from 10−5 to 0.01, and the weight of L2-norm ing with the experiment, SA, and DA. The errors of several angles of
regularization ranges from 0 to 0.001. The results of the Bayesian attack are presented in Table 5. As can be seen, for separated flows,
optimization process are depicted in Fig. 12. We mainly show R2 and the DNN model maintains the accuracy of DA, while for the attached
MSE versus iteration, and R2 is defined by flow, the DNN model maintains the accuracy of SA (different from
the result of the DA model in Sec. III.B). Figure 15b shows the drag
N
y^ i − yi 2 coefficient curve, and results from Singh et al. [18] are also used for
Downloaded by Tongji University on August 2, 2023 | http://arc.aiaa.org | DOI: 10.2514/1.J062711

R2
1 − i
1
(15)
N
i
1 y i − yi 2 comparison (the training data of Singh et al. are slightly different
from this paper).
As can be seen, the model at iteration
84 is the best model on the
test set. The final model structure at iteration
84 is (6, 265, 362,
362, 362, 362, 362, 362, 469, 1), indicating that the depth is 8, and the
width is nonequivalent in each hidden layer. The optimized learning
rate is 0.00013, and the weight of L2-norm regularization λ2 is 0,
indicating that the overfitting problem is not significant in this
problem. All the DNN models in the following are trained by these
hyperparameters.

E. Case 1: Generalization About Angle of Attack

Fig. 13 Lift curve’s eR of S809 calculated by the DNN model versus η at
Table 4 presents the training set and test set of case 1; flow data of Re
2 × 106 .
four angles of attack of S809 are selected as training data. To predict
the lift curve, the training data must include the flow conditions of
attached flow and separated flow. Flowfield data at α
3.1 and 8.2°
are calculated by the SA model, and flowfield data at α
12.2 and
14.2° are calculated by the DA model.
There are two sources of the training samples in this paper, i.e., SA
data and DA data. The SA data mainly affect the accuracy of the
attached flow at low angles of attack, and the DA data mainly affect
the accuracy of the separated flow. Therefore, the ratio of these two
sources will affect the overall accuracy of the model. In this paper, the

Table 4 Training set and test set of Case 1

Set Re (×106 ) α, ° Airfoil

Training 2 3.1/8.2/12.2/14.2 S809
Test 2 0–15.2 S809
Fig. 14 Loss function value versus iteration step.

Fig. 12 Bayesian optimization process curves.

 Article in Advance / SHAN ET AL. 9

Fig. 15 Lift and drag coefficient of the DNN and SA models compared with experimental data at Re
2 × 106 (for drag coefficient in the fixed transition
condition, experimental data only give data of α ≤ 10.2°).

Table 5 Lift coefficient and relative error of SA model and DNN model compared with experimental data at Re
2 × 106
Downloaded by Tongji University on August 2, 2023 | http://arc.aiaa.org | DOI: 10.2514/1.J062711

Parameter α
8.2° α
10.2° α
12.2° α
13.2° α
14.2° α
15.2

Exp, Cl 0.932 0.981 0.995 1.039 1.061 1.076
SA, Cl (relative error) 0.972 (4.3%) 1.160 (18.2%) 1.278 (28.4%) 1.282 (23.4%) 1.264 (19.1%) 1.255 (16.6%)
DNN, Cl (relative error) 0.959 (2.9%) 0.969 (1.2%) 0.940 (5.5%) 0.962 (7.4%) 0.994 (6.3%) 1.008 (6.3%)

Exp, experimental data.

Figure 16 shows the comparison of the pressure distribution on three test sets, and the DNN model can also be improved by
the training set. It is emphasized again that the training data of 8.2° high accuracy under AOA that does not appear in the training set,
are SA data, and the training data of 12.2 and 14.2° are DA data. which verifies the generalization of the DNN model with respect to
Figure 16a illustrates that the results of DNN are in good agree- AOA. Figure 18 shows the coefficients of skin friction of different
ment with those of SA and experiments. From Figs. 16b and 16c, models at three angles of attack. The location of Cf
0 represents
we can see that the results of DNN are close to those of DA, and the separated point. Obviously, compared with the SA model,
the error is low. Figure 17 compares the pressure distribution on there are two prominent changes in the flowfield of the DNN

Fig. 16 Pressure distribution of DNN model, DA model, SA model, and experiment at Re
2 × 106 on training set.

Fig. 17 Pressure distribution of DNN model, SA model, and experiment at Re
2 × 106 on test set.
 10 Article in Advance / SHAN ET AL.

Fig. 18 Coefficients of skin friction plots of DNN model, DA model, and SA model at Re
2 × 106 .

model. The first is that the flow separates earlier, and the second F. Case 2: Generalization About AOA, Airfoil, and Re
is that the separation bubble is enlarged. Both of these changes In this case, flow data of two airfoils are selected as training data, as
help reduce the lift coefficient and improve the accuracy for lift shown in Table 6. For the flow conditions of α < 8°, the SA model is
curve as shown in Fig. 15 and Table 5. Meanwhile, Fig. 18b shows used to obtain the training data, and for other flow conditions, DA
Downloaded by Tongji University on August 2, 2023 | http://arc.aiaa.org | DOI: 10.2514/1.J062711

that the coefficients of skin friction of the DNN model are in good models are applied.
agreement with those of the DA model on the upper surface, but The comparison of the lift coefficients for the three airfoils at
tend to those of the SA model from x∕c
0.2 to x∕c
0.5 on different Reynolds numbers is sketched in Figs. 13–15. It can be
the lower surface. Figures 19 and 20 compare near-wall eddy-
viscosity profiles and near-wall velocity profiles of the SA model
and the DNN model at Re
2 × 106 and α
13.2°. Eddy viscos- Table 6 Training set and test set of Case 2
ity of the DNN model is larger than that of the SA model, espe-
cially the maximum of a profile. Figure 20b shows that the flow Set Re (×106 ) α, ° Airfoil
predicted by the DNN model begins to separate at x∕c
0.5, Training 2 3.1/8.2/12.2/14.2 S809
but the velocity profile of the SA model still has a large kinetic 1 3.1/8.2/12.2/15.2 S805
energy and similar to the profile of the DNN model at x∕c
0.2.
Test 1/1.5/2/2.5/3 0–20.2 S809
Figure 20c indicates that the height of the recirculation zone of the 0.5/0.7/1/1.5/2 0–20.1 S805
DNN model is higher than that of the SA model by comparing the 0.7/1/1.5/2/3 0–20.2 S814
location of u∕U0
0.

Fig. 19 Near-wall eddy-viscosity profiles in semilog plots of DNN model and SA model at Re
2 × 106 and α
13.2°. The vertical axis represents the
distance nondimensionalized by airfoil chord, and horizontal axis represents the eddy viscosity nondimensionalized by freestream laminar viscosity.

Fig. 20 Near-wall velocity profiles in semilog plots of DNN model and SA model at Re
2 × 106 and α
13.2°. The vertical axis represents the distance
nondimensionalized by airfoil chord, and horizontal axis represents the streamwise velocity nondimensionalized by freestream velocity.
 Article in Advance / SHAN ET AL. 11

seen that the DNN model possesses higher accuracy for the lift Re
1 × 106 . For higher Re, although the stall point can be accu-
coefficient above the stall. Figures 21 and 22 display the test cases rately predicted, the steep drop of lift coefficient above the stall
of the same airfoil (S805, S809) as the training set, but with different cannot be accurately predicted. Even so, the error of the DNN model
Re and AOA. Lift coefficients of the DNN model are more accurate is still lower than that of the SA model.
than that of the SA model. Figure 23 shows the lift curve of the new Figure 24 presents the mean relative error of lift coefficients in stall
airfoil (S814), which does not appear in the training set. The DNN zone (α > 8°), and compares results from the SA model and the DNN
model can obviously provide higher accuracy at Re
0.7 × 106 and model for the three airfoils at five different Re. Since the error of the
Downloaded by Tongji University on August 2, 2023 | http://arc.aiaa.org | DOI: 10.2514/1.J062711

Fig. 21 Lift curve of S805 airfoil at different Reynolds numbers. The training set is marked by the green circles.

Fig. 22 Lift curve of S809 airfoil at different Reynolds numbers. The training set is marked by the green circles.
 12 Article in Advance / SHAN ET AL.
Downloaded by Tongji University on August 2, 2023 | http://arc.aiaa.org | DOI: 10.2514/1.J062711

Fig. 23 Lift curve of S814 airfoil at different Reynolds numbers.

Fig. 24 Lift coefficients’ eR in stall zone (α > 8°) of three airfoils at different Reynolds numbers.

linear segment of the lift curve is small, only results of the stall zone distribution of the three airfoils at different Re. It can be seen that
are selected to calculate eR . It can be seen that the errors of all three the accuracy of the DNN model is significantly higher than that of
airfoils reduce by over 60%. the SA model. Figure 29 shows the coefficients of skin friction
Figure 25 displays the diagram of drag coefficient versus lift plots of the S805 airfoil for three angles of attack at Re
1 × 106,
coefficient of three airfoils at several Reynolds numbers. Due to and illustrates that the location of separation on the upper surface
the improvement in the accuracy of the lift coefficient, DNN can of the DNN model is fronter than the SA model, responding to the
predict a more accurate stall point (the drag begins to increase reason why the lift coefficient predicted that the DNN model is
faster) than the SA model. Figures 26–28 compares the pressure lower.
 Article in Advance / SHAN ET AL. 13

Fig. 25 Drag coefficients versus lift coefficients of three airfoil at three Reynolds numbers.
Downloaded by Tongji University on August 2, 2023 | http://arc.aiaa.org | DOI: 10.2514/1.J062711

Fig. 26 Pressure distribution of S805 airfoil for three angles of attack at Re
1 × 106 .

Fig. 27 Pressure distribution of S809 airfoil for three angles of attack at Re
2 × 106 .

Fig. 28 Pressure distribution of S814 airfoil for three angles of attack at Re
1.5 × 106 .
 14 Article in Advance / SHAN ET AL.

Fig. 29 Coefficients of skin friction plots of S805 airfoil for three angles of attack at Re
1 × 106 .

To sum up, the DNN model possesses good generalization ability by physical feature construction, transformed output, and designed
for airfoil shapes and flow conditions. Concerning the generalization loss function. Considering the flow characteristics at a high Reyn-
of the airfoil shapes, the DNN model is tested on the new airfoil olds number, a nonlinear wall-distance transformation of eddy
(S814) with a larger thickness, and the mean relative error of lift viscosity is applied to the output of the DNN model and can
Downloaded by Tongji University on August 2, 2023 | http://arc.aiaa.org | DOI: 10.2514/1.J062711

coefficients reduces by over 60% compared with the SA model. As effectively reduce the difficulty of training. Meanwhile, we add
for the generalization of flow conditions, only two Reynolds numbers the non-negativity constraint term of eddy viscosity to the loss
(1 × 106 and 2 × 106 ) with four angles of attack each are used to train function to promote numerical stability.
the DNN model, and the accuracy of the lift coefficients of Re ranging Because of the complexity of turbulence, high-fidelity data and
from 0.5 × 106 to 3 × 106 has improved a lot. physics-informed ML framework are required to construct a general
turbulence model. This paper verifies that, for the same type of flow
(turbulent flows over airfoil), effective generalization can be achieved
IV. Discussion using a small amount of data. An extended work is to overcome more
Based on the DA and DNN models, we have obtained better results, types of model uncertainties and build a general model for different
such as the more accurate stall curve than the SA model. However, types of complex engineering flow, which need to be further inves-
from the perspective of model architecture and source of uncertainty, tigated in future.
the method in this paper still does not break through some assumptions
(linear eddy-viscosity hypothesis, SA model framework, etc.). In
general, four layers of simplifications are typically required to formu- Appendix A: Spalart–Allmaras Model
late a RANS closure [10]. These simplifications introduce four types The one-equation Spalart–Allmaras turbulence model [3] is
of uncertainties, i.e., uncertainties introduced by ensemble averaging,
uncertainties in the functional and operational representation of Reyn- ∂~ν ∂~ν C ν~ 2
 uj
Cb1 1 − ft2 S^ ν~ − Cw1 fw − b1 ft2
olds stress, uncertainties in functional forms within a model, and ∂t ∂xj κ2 d
uncertainties in the coefficients within a model. Since the training data
1 ∂ ∂~ν ∂~ν ∂~ν
in this paper came from the solution of the SA turbulence model with  ν  ν~   Cb2 (A1)
calibrated parameters, only the discrepancies within the modeling σ ∂xj ∂xj ∂xj ∂xj
coefficients in the SA model were considered and reduced. The other
three types of uncertainties are not eliminated from the methodological where ν~ is the solution variable, which is used to compute the
level in our paper. On the other hand, some other methods, such as kinematic eddy viscosity by
Reynolds stress models, may show potential. Reynolds stress models
at least overcome the uncertainties introduced by the linear eddy- χ3 ν~
viscosity hypothesis, and Reynolds stress models combined with a νt
ν~ fν1 ; fν1
; χ
(A2)
χ 3  C3ν1 ν
data-driven approach will be explored in our future work.
The expressions of related variables are displayed below:
V. Conclusions
υ χ
This paper developed a turbulence modeling approach combin- S^
Ω  2 2 fν2 ; fν1
1 − (A3)
ing DA and deep neural network to solve the problems of data κ d 1  χfν1
acquiring and turbulence modeling for high-Re separated flow.
Different from modeling the correction factor in the traditional 1 ∂ui ∂uj
models, this paper directly constructs a model for eddy viscosity Ω
2Ωij Ωij ; Ωij
− (A4)
2 ∂xj ∂xi
to replace the traditional models. The DA method can acquire
optimized parameters of the SA model based on the experimental
pressure distribution data and then obtain the high-fidelity flow- ft2
Ct3 exp−Ct4 χ 2  (A5)
field for separated flows. Calibrated SA models with optimized
parameters can provide accurate pressure distribution and lift 1  C6w3 1∕6 υ~
coefficient, but are case-by-case models. Therefore, trained with fw
g ; g
r  Cw2 r6 − r; r
(A6)
high-fidelity separated flowfield data obtained by the DA method g6  C6w3 Sκ 2 d2
and attached flowfield data acquired by the SA model, the DNN
model can be competent in both the attached flows and separated where the initial parameters in the above equations are
flows. This approach is tested at different Reynolds numbers over
three wind-turbine airfoils, can greatly improve the accuracy 2 C 1  Cb2
σ
; Cb1
0.1355; Cb2
0.622; Cw1
b1  ;
of numerical simulation, and can possess good generalization 3 κ2 σ
ability. Cw2
0.3; Cw3
2; κ
0.41; Cv1
7.1; Ct3
1.2; Ct4
0.5
In addition, an important point is how to design a physics-
informed DNN model. In this paper, the DNN model is augmented (A7)
 Article in Advance / SHAN ET AL. 15

Table B1 Optimized parameters of DA models

Model Ct3 Ct4 Cb1 Cb2 σ Cv1 κ Cw2 Cw3

SA 1.200 0.500 0.136 0.622 0.667 7.100 0.410 0.300 2.000
DA1 1.159 0.375 0.153 0.311 0.334 3.550 0.481 0.309 1.500
DA2 1.204 0.544 0.129 0.923 0.635 4.608 0.607 0.150 3.000
DA3 1.415 0.395 0.088 0.311 0.835 7.295 0.396 0.230 1.007
DA4 0.600 0.250 0.139 0.311 0.478 3.550 0.615 0.150 3.000

Appendix B: Results of Optimized Parameters pp. 4893–4909.

for DA Models https://doi.org/10.2514/1.J057204
[14] Ray, J., Lefantzi, S., Arunajatesan, S., and Dechant, L., “Bayesian Param-
Using the data assimilation method, we can obtain the optimized eter Estimation of a k-Epsilon Model for Accurate Jet-in-Crossflow
parameters of the SA model, which are shown in Table B1. Mean- Simulations,” AIAA Journal, Vol. 54, No. 8, 2016, pp. 2432–2448.
while, initial parameters of the SA model are also displayed for https://doi.org/10.2514/1.J054758
comparison. [15] Duraisamy, K., Zhang, Z. J., and Singh, A. P., “New Approaches in
Turbulence and Transition Modeling Using Data-Driven Techniques,”
53rd AIAA Aerospace Sciences Meeting, AIAA Paper 2015-1284, 2015.
Acknowledgments [16] Singh, A. P., Duraisamy, K., and Zhang, Z. J., “Augmentation of
This work was supported by National Natural Science Foundation Turbulence Models Using Field Inversion and Machine Learning,”
Downloaded by Tongji University on August 2, 2023 | http://arc.aiaa.org | DOI: 10.2514/1.J062711

of China (No. 92152301) and the foundation of National Key Labo- 55th AIAA Aerospace Sciences Meeting, AIAA Paper 2017-0993, 2017.
[17] Singh, A. P., and Duraisamy, K., “Using Field Inversion to Quantify
ratory of Science and Technology on Aerodynamic Design and Functional Errors in Turbulence Closures,” Physics of Fluids, Vol. 28,
Research (No. 614220121010115). No. 4, 2016, Paper 045110.
https://doi.org/10.1063/1.4947045
References [18] Singh, A. P., Medida, S., and Duraisamy, K., “Machine-Learning-Aug-
mented Predictive Modeling of Turbulent Separated Flows over Air-
[1] Launder, B. E., and Sharma, B. I., “Application of the Energy- foils,” AIAA Journal, Vol. 55, No. 7, 2017, pp. 2215–2227.
Dissipation Model of Turbulence to the Calculation of Flow Near a https://doi.org/10.2514/1.J055595
Spinning Disc,” Letters in Heat and Mass Transfer, Vol. 1, No. 2, 1974, [19] Parish, E. J., and Duraisamy, K., “A Paradigm for Data-Driven Predic-
pp. 131–137. tive Modeling Using Field Inversion and Machine Learning,” Journal of
https://doi.org/10.1016/0094-4548(74)90150-7 Computational Physics, Vol. 305, Jan. 2016, pp. 758–774.
[2] Wilcox, D. C., “Reassessment of the Scale-Determining Equation for https://doi.org/10.1016/j.jcp.2015.11.012
Advanced Turbulence Models,” AIAA Journal, Vol. 26, No. 11, 1988, [20] Yang, M., and Xiao, Z., “Improving the k-Omega-Gamma-A(r) Tran-
pp. 1299–1310. sition Model by the Field Inversion and Machine Learning Framework,”
https://doi.org/10.2514/3.10041 Physics of Fluids, Vol. 32, No. 6, 2020, Paper 064101.
[3] Spalart, P., and Allmaras, S., “A One-Equation Turbulence Model for https://doi.org/10.1063/5.0008493
Aerodynamic Flows,” 30th Aerospace Sciences Meeting and Exhibit, [21] Holland, J. R., Baeder, J. D., and Duraisamy, K., “Field Inversion and
AIAA Paper 1992-0439, 1992. Machine Learning with Embedded Neural Networks: Physics-
[4] Menter, F. R., “2-Equation Eddy-Viscosity Turbulence Models for Consistent Neural Network Training,” AIAA Aviation 2019 Forum,
Engineering Applications,” AIAA Journal, Vol. 32, No. 8, 1994, AIAA Paper 2019-3200, 2019.
pp. 1598–1605. [22] Schumann, U., “Realizability of Reynolds-Stress Turbulence Models,”
https://doi.org/10.2514/3.12149 Physics of Fluids, Vol. 20, No. 5, 1977, pp. 721–725.
[5] Launder, B. E., Reece, G. J., and Rodi, W., “Progress in Development of https://doi.org/10.1063/1.861942
a Reynolds-Stress Turbulence Closure,” Journal of Fluid Mechanics, [23] Lumley, J. L., “Computational Modeling of Turbulent Flows,” Advances
Vol. 68, No. 3, 1975, pp. 537–566. in applied mechanics, Vol. 18, Jan. 1979, pp. 123–176.
https://doi.org/10.1017/s0022112075001814 https://doi.org/10.1016/S0065-2156(08)70266-7
[6] Wallin, S., and Johansson, A. V., “An Explicit Algebraic Reynolds [24] Zhu, L., Zhang, W., Kou, J., and Liu, Y., “Machine Learning Methods
Stress Model for Incompressible and Compressible Turbulent Flows,” for Turbulence Modeling in Subsonic Flows Around Airfoils,” Physics
Journal of Fluid Mechanics, Vol. 403, Jan. 2000, pp. 89–132. of Fluids, Vol. 31, No. 1, 2019, Paper 015105.
https://doi.org/10.1017/s0022112099007004 https://doi.org/10.1063/1.5061693
[7] Slotnick, J. P., Khodadoust, A., Alonso, J., Darmofal, D., Gropp, W., Lurie, [25] Zhu, L., Zhang, W., Sun, X., Liu, Y., and Yuan, X., “Turbulence Closure
E., and Mavriplis, D. J., “CFD Vision 2030 Study: A Path to Revolutionary for High Reynolds Number Airfoil Flows by Deep Neural Networks,”
Computational Aerosciences,” NASA CR-2014-218178, 2014. Aerospace Science and Technology, Vol. 110, March 2021, Paper
[8] Durbin, P. A., “Some Recent Developments in Turbulence Closure 106452.
Modeling,” Annual Review of Fluid Mechanics, Vol. 50, Jan. 2018, https://doi.org/10.1016/j.ast.2020.106452
pp. 77–103. [26] Sun, X., Cao, W., Liu, Y., Zhu, L., and Zhang, W., “High Reynolds
[9] Craft, T. J., Launder, B. E., and Suga, K., “Development and Application Number Airfoil Turbulence Modeling Method Based on Machine
of a Cubic Eddy-Viscosity Model of Turbulence,” International Journal Learning Technique,” Computers & Fluids, Vol. 236, March 2022,
of Heat and Fluid Flow, Vol. 17, No. 2, 1996, pp. 108–115. Paper 105298.
https://doi.org/10.1016/0142-727x(95)00079-6 https://doi.org/10.1016/j.compfluid.2021.105298
[10] Duraisamy, K., Iaccarino, G., and Xiao, H., “Turbulence Modeling in [27] Liu, Y., Cao, W., Zhang, W., and Xia, Z., “Analysis on Numerical
the Age of Data,” Annual Review of Fluid Mechanics, Vol. 51, Jan. 2019, Stability and Convergence of Reynolds Averaged Navier–Stokes Sim-
pp. 357–377. ulations from the Perspective of Coupling Modes,” Physics of Fluids,
[11] Duraisamy, K., “Perspectives on Machine Learning-Augmented Vol. 34, No. 1, 2022, Paper 015120.
Reynolds-Averaged and Large Eddy Simulation Models of Turbu- https://doi.org/10.1063/5.0076273
lence,” Physical Review Fluids, Vol. 6, No. 5, 2021, Paper 050504. [28] Pope, S., “A More General Effective-Viscosity Hypothesis,” Journal of
https://doi.org/10.1103/PhysRevFluids.6.050504 Fluid Mechanics, Vol. 72, No. 2, 1975, pp. 331–340.
[12] Cheung, S. H., Oliver, T. A., Prudencio, E. E., Prudhomme, S., and https://doi.org/10.1017/S0022112075003382
Moser, R. D., “Bayesian Uncertainty Analysis with Applications to [29] Ling, J., Kurzawski, A., and Templeton, J., “Reynolds Averaged Turbu-
Turbulence Modeling,” Reliability Engineering & System Safety, lence Modelling Using Deep Neural Networks with Embedded Invari-
Vol. 96, No. 9, 2011, pp. 1137–1149. ance,” Journal of Fluid Mechanics, Vol. 807, Nov. 2016, pp. 155–166.
https://doi.org/10.1016/j.ress.2010.09.013 https://doi.org/10.1017/jfm.2016.615
[13] Ray, J., Dechant, L., Lefantzi, S., Ling, J., and Arunajatesan, S., “Robust [30] Ling, J., and Templeton, J., “Evaluation of Machine Learning Algo-
Bayesian Calibration of a k-Epsilon Model for Compressible Jet-in- rithms for Prediction of Regions of High Reynolds Averaged Navier
Crossflow Simulations,” AIAA Journal, Vol. 56, No. 12, 2018, Stokes Uncertainty,” Physics of Fluids, Vol. 27, No. 8, 2015, Paper
 16 Article in Advance / SHAN ET AL.

085103. Aerospace Science and Technology, Vol. 126, July 2022, Paper 107328.
https://doi.org/10.1063/1.4927765 https://doi.org/10.1016/j.ast.2022.107328
[31] Ling, J., Jones, R., and Templeton, J., “Machine Learning Strategies for [49] Stuart, A., and Zygalakis, K., Data Assimilation: A Mathematical
Systems with Invariance Properties,” Journal of Computational Phys- Introduction, Oak Ridge National Lab., Oak Ridge, TN, 2015, Chap. 1.
ics, Vol. 318, Aug. 2016, pp. 22–35. [50] He, C., Wang, P., and Liu, Y., “Data Assimilation for Turbulent Mean
https://doi.org/10.1016/j.jcp.2016.05.003 Flow and Scalar Fields with Anisotropic Formulation,” Experiments in
[32] Jiang, C., Mi, J., Laima, S., and Li, H., “A Novel Algebraic Stress Model Fluids, Vol. 62, No. 5, 2021, p. 117.
with Machine-Learning-Assisted Parameterization,” Energies, Vol. 13, https://doi.org/10.1007/s00348-021-03213-8
No. 1, 2020, p. 258. [51] Foures, D. P., Dovetta, N., Sipp, D., and Schmid, P. J., “A Data-
https://doi.org/10.3390/en13010258 Assimilation Method for Reynolds-Averaged Navier–Stokes-Driven
[33] Zhang, X.-L., Xiao, H., Luo, X., and He, G., “Ensemble Kalman Method Mean Flow Reconstruction,” Journal of Fluid Mechanics, Vol. 759,
for Learning Turbulence Models from Indirect Observation Data,” Nov. 2014, pp. 404–431.
Journal of Fluid Mechanics, Vol. 949, Oct. 2022, p. A26. https://doi.org/10.1017/jfm.2014.566
https://doi.org/10.1017/jfm.2022.744 [52] Kato, H., Yoshizawa, A., Ueno, G., and Obayashi, S., “A Data Assimi-
[34] Weatheritt, J., and Sandberg, R., “A Novel Evolutionary Algorithm lation Methodology for Reconstructing Turbulent Flows Around Air-
Applied to Algebraic Modifications of the RANS Stress-Strain Rela- craft,” Journal of Computational Physics, Vol. 283, Feb. 2015,
tionship,” Journal of Computational Physics, Vol. 325, Nov. 2016, pp. 559–581.
pp. 22–37. https://doi.org/10.1016/j.jcp.2014.12.013
https://doi.org/10.1016/j.jcp.2016.08.015 [53] Zhang, X.-L., Michelén-Ströfer, C., and Xiao, H., “Regularized Ensem-
[35] Weatheritt, J., and Sandberg, R. D., “The Development of Algebraic ble Kalman Methods for Inverse Problems,” Journal of Computational
Stress Models Using a Novel Evolutionary Algorithm,” International Physics, Vol. 416, Sept. 2020, Paper 109517.
Journal of Heat and Fluid Flow, Vol. 68, Dec. 2017, pp. 298–318. https://doi.org/10.1016/j.jcp.2020.109517
https://doi.org/10.1016/j.ijheatfluidflow.2017.09.017 [54] Deng, Z., He, C., Wen, X., and Liu, Y., “Recovering Turbulent Flow
Downloaded by Tongji University on August 2, 2023 | http://arc.aiaa.org | DOI: 10.2514/1.J062711

[36] Zhao, Y. M., Akolekar, H. D., Weatheritt, J., Michelassi, V., and Field from Local Quantity Measurement: Turbulence Modeling Using
Sandberg, R. D., “RANS Turbulence Model Development Using Ensemble-Kalman-Filter-Based Data Assimilation,” Journal of Visuali-
CFD-Driven Machine Learning,” Journal of Computational Physics, zation, Vol. 21, No. 6, 2018, pp. 1043–1063.
Vol. 411, June 2020, Paper 109413. https://doi.org/10.1007/s12650-018-0508-0
https://doi.org/10.1016/j.jcp.2020.109413 [55] Kato, H., and Obayashi, S., “Approach for Uncertainty of Turbulence
[37] Emory, M., Larsson, J., and Iaccarino, G., “Modeling of Structural Modeling Based on Data Assimilation Technique,” Computers & Flu-
Uncertainties in Reynolds-Averaged Navier-Stokes Closures,” Physics ids, Vol. 85, Oct. 2013, pp. 2–7.
of Fluids, Vol. 25, No. 11, 2013, Paper 110822. https://doi.org/10.1016/j.compfluid.2012.09.002
https://doi.org/10.1063/1.4824659 [56] Zeng, F., Zhang, W., Li, J., Zhang, T., and Yan, C., “Adaptive Model
[38] Wu, J.-L., Wang, J.-X., and Xiao, H., “A Bayesian Calibration- Refinement Approach for Bayesian Uncertainty Quantification in Tur-
Prediction Method for Reducing Model-Form Uncertainties with Appli- bulence Model,” AIAA Journal, Vol. 60, No. 6, 2022, pp. 3502–3516.
cation in RANS Simulations,” Flow Turbulence and Combustion, https://doi.org/10.2514/1.J060889
Vol. 97, No. 3, 2016, pp. 761–786. [57] Subbian, G., Botelho e Souza, A. C., Radespiel, R., Zander, E., Mosha-
https://doi.org/10.1007/s10494-016-9725-6 gen, T., Friedman, N., Moioli, M., Breitsamter, C., and Sorensen, K.,
[39] Tracey, B., Duraisamy, K., and Alonso, J., “Application of Supervised
“Calibration of an Extended Eddy Viscosity Turbulence Model Using
Learning to Quantify Uncertainties in Turbulence and Combustion
Bayesian Update,” AIAA Journal, Vol. 60, No. 6, 2022, pp. 3487–3501.
Modeling,” 51st AIAA Aerospace Sciences Meeting Including the New
https://doi.org/10.2514/1.J060527
Horizons Forum and Aerospace Exposition, AIAA Paper 2013-0259,
[58] Doronina, O. A., Murman, S. M., and Hamlington, P. E., “Parameter
2013.
Estimation for Reynolds-Averaged Navier-Stokes Models Using
[40] Wang, J.-X., Wu, J.-L., and Xiao, H., “Physics-Informed Machine
Approximate Bayesian Computation,” AIAA Journal, Vol. 59, No. 11,
Learning Approach for Reconstructing Reynolds Stress Modeling Dis-
2021, pp. 4703–4718.
crepancies Based on DNS Data,” Physical Review Fluids, Vol. 2, No. 3,
2017, Paper 034603. https://doi.org/10.2514/1.J060308
https://doi.org/10.1103/PhysRevFluids.2.034603 [59] Rao, R. V., Savsani, V. J., and Vakharia, D. P., “Teaching-Learning-
[41] Yin, Y., Yang, P., Zhang, Y., Chen, H., and Fu, S., “Feature Selection and Based Optimization: An Optimization Method for Continuous Non-
Processing of Turbulence Modeling Based on an Artificial Neural Net- Linear Large Scale Problems,” Information Sciences, Vol. 183, No. 1,
work,” Physics of Fluids, Vol. 32, No. 10, 2020, Paper 105117. 2012, pp. 1–15.
https://doi.org/10.1063/5.0022561 https://doi.org/10.1016/j.ins.2011.08.006
[42] Wu, J.-L., Xiao, H., and Paterson, E., “Physics-Informed Machine [60] Rao, R. V., Savsani, V. J., and Vakharia, D. P., “Teaching-Learning-
Learning Approach for Augmenting Turbulence Models: A Compre- Based Optimization: A Novel Method for Constrained Mechanical
hensive Framework,” Physical Review Fluids, Vol. 3, No. 7, 2018, Paper Design Optimization Problems,” Computer-Aided Design, Vol. 43,
074602. No. 3, 2011, pp. 303–315.
https://doi.org/10.1103/PhysRevFluids.3.074602 https://doi.org/10.1016/j.cad.2010.12.015
[43] Wu, J., Xiao, H., Sun, R., and Wang, Q., “Reynolds-Averaged Navier– [61] Khatir, S., Wahab, M. A., Boutchicha, D., and Khatir, T., “Structural
Stokes Equations with Explicit Data-Driven Reynolds Stress Closure Health Monitoring Using Modal Strain Energy Damage Indicator
Can Be Ill-Conditioned,” Journal of Fluid Mechanics, Vol. 869, Coupled with Teaching-Learning-Based Optimization Algorithm and
June 2019, pp. 553–586. Isogoemetric Analysis,” Journal of Sound and Vibration, Vol. 448,
https://doi.org/10.1017/jfm.2019.205 May 2019, pp. 230–246.
[44] McConkey, R., Yee, E., and Lien, F. S., “Deep Structured Neural Net- https://doi.org/10.1016/j.jsv.2019.02.017
works for Turbulence Closure Modeling,” Physics of Fluids, Vol. 34, [62] Shukla, A. K., Singh, P., and Vardhan, M., “An Adaptive Inertia Weight
No. 3, 2022, Paper 035110. Teaching-Learning-Based Optimization Algorithm and its Applica-
https://doi.org/10.1063/5.0083074 tions,” Applied Mathematical Modelling, Vol. 77, Jan. 2020, pp. 309–
[45] Yan, C., Li, H., Zhang, Y., and Chen, H., “Data-Driven Turbulence 326.
Modeling in Separated Flows Considering Physical Mechanism Analy- https://doi.org/10.1016/j.apm.2019.07.046
sis,” International Journal of Heat and Fluid Flow, Vol. 96, Aug. 2022, [63] Snoek, J., Larochelle, H., and Adams, R. P., “Practical Bayesian
Paper 109004. Optimization of Machine Learning Algorithms,” Proceedings of the
https://doi.org/10.1016/j.ijheatfluidflow.2022.109004 25th International Conference on Neural Information Processing Sys-
[46] Yan, C., Zhang, Y., and Chen, H., “Data Augmented Turbulence Mod- tems, Vol. 2, Curran Associates Inc., Lake Tahoe, NV, 2012, pp. 2951–
eling for Three-Dimensional Separation Flows,” Physics of Fluids, 2959.
Vol. 34, No. 7, 2022, Paper 075101. [64] Erhard, P., Etling, D., Muller, U., Riedel, U., Sreenivasan, K., and
https://doi.org/10.1063/5.0097438 Warnatz, J., Prandtl-Essentials of Fluid Mechanics, Springer Science
[47] Rumsey, C. L., Coleman, G. N., and Wang, L., “In Search of Data- & Business Media, New York, 2010, Chap. 6.
Driven Improvements to RANS Models Applied to Separated Flows,” [65] Paszke, A., Gross, S., Chintala, S., and Chanan, G., “Pytorch: Tensors
AIAA Scitech 2022 Forum, AIAA Paper 2022-0937, 2022. and Dynamic Neural Networks in Python with Strong GPU Acceler-
[48] Liu, Y., Zhang, W., and Xia, Z., “A New Data Assimilation Method of ation,” PyTorch: Tensors and Dynamic Neural Networks in Python with
Recovering Turbulent Mean Flow Field at High Reynolds Numbers,” Strong GPU Acceleration, Vol. 6, No. 3, 2017, p. 67.
 Article in Advance / SHAN ET AL. 17

[66] Somers, D. M., “Design and Experimental Results for the S809 Airfoil,” Optimization,” arXiv:2212.03183, 2022.
National Renewable Energy Lab. TR NREL/SR-440-6918, Golden, https://doi.org/10.48550/arXiv.2212.03183
CO, 1997. [70] Zhu, L., Wang, T., Fan, Z., Xiang, X., Zhang, W., and Yuan, X.,
[67] Somers, D. M., “Design and Experimental Results for the S805 Airfoil,” “Physics-Assisted Recursive Method for Sample Selection from Wall-
National Renewable Energy Lab. TR NREL/SR-440-6917, Golden, Bounded Turbulence Data,” Physics of Fluids, Vol. 34, No. 8, 2022,
CO, 1997. Paper 085132.
[68] Somers, D. M., “Design and Experimental Results for the S814 Airfoil,” https://doi.org/10.1063/5.0101008
National Renewable Energy Lab. TR NREL/SR-440-6919, Golden,
CO, 1997. P. Durbin
[69] Cao, W., Liu, Y., Shan, X., Gao, C., and Zhang, W., “A Novel Con- Associate Editor
vergence Enhancement Method Based on Online Dimension Reduction
Downloaded by Tongji University on August 2, 2023 | http://arc.aiaa.org | DOI: 10.2514/1.J062711