Ijtpp 07 00020

International Journal of
Turbomachinery
Propulsion and Power
Article
Optimization of Turbine Blade Aerodynamic Designs Using
CFD and Neural Network Models
Chao Zhang * and Matthew Janeway
Siemens Digital Industries Software, Cypress, CA 90630, USA; matt.janeway@siemens.com

* Correspondence: chao.zhang1@siemens.com
Abstract: Optimization methods have been widely applied to the aerodynamic design of gas turbine
blades. While applying optimization to high-fidelity computational fluid dynamics (CFD) simulations
has proven capable of improving engineering design performance, a challenge has been overcoming
the prolonged run-time due to the computationally expensive CFD runs. Reduced-order models and,
more recently, machine learning methods have been increasingly used in gas turbine studies to predict
performance metrics and operational characteristics, model turbulence, and optimize designs. The
application of machine learning methods allows for utilizing existing knowledge and datasets from
different sources, such as previous experiments, CFD, low-fidelity simulations, 1D or system-level
studies. The present study investigates inserting a machine learning model that utilizes such data
into a high-fidelity CFD driven optimization process, and hence effectively reduces the number of
required evaluations of the CFD model. Artificial Neural Network (ANN) models were trained on
data from over three thousand two-dimensional (2D) CFD analyses of turbine blade cross-sections.
The trained ANN models were then used as surrogates in a nested optimization process alongside
a full three-dimensional Navier–Stokes CFD simulation. The much lower evaluation cost of the
ANN model allows for tens of thousands of design evaluations to guide the search of the best blade
profiles to be used in the more expensive, high-fidelity CFD runs, improving the progress of the
optimization while reducing the required computation time. It is estimated that the current workflow
Citation: Zhang, C.; Janeway, M. achieves a five-fold reduction in computational time in comparison to an optimization process that
Optimization of Turbine Blade is based on three-dimensional (3D) CFD simulations alone. The methodology is demonstrated on
Aerodynamic Designs Using CFD the NASA/General Electric Energy Efficient Engine (E3) high pressure turbine blade and found
and Neural Network Models. Int. J. Pareto front designs with improved blade efficiency and power over the baseline. Quantitative
Turbomach. Propuls. Power 2022, 7, 20. analysis of the optimization data reveals that some design parameters in the present study are more
https://doi.org/10.3390/ijtpp7030020 influential than others, such as the lean angle and tip scaling factor. Examining the optimized designs
Academic Editor: Tom Verstraete also provides insight into the physics, showing that the optimized designs have a lower amount
of pressure drop near the trailing edge, but have an earlier onset of pressure drop on the suction
Received: 25 April 2022
side surface when compared to the baseline design, contributing to the observed improvements in
Accepted: 28 June 2022
efficiency and power.
Published: 30 June 2022
Publisher’s Note: MDPI stays neutral Keywords: CFD; optimization; aerodynamics; gas turbines; machine learning; neural networks
with regard to jurisdictional claims in
published maps and institutional affil-
iations.
1. Introduction
The present study applies machine learning methods in a CFD-based optimization for
Copyright: © 2022 by the authors. turbine blade aerodynamics. Literature on optimization and machine learning methods
Licensee MDPI, Basel, Switzerland. used in gas turbine studies will be reviewed, followed by a summary of the motivation of
This article is an open access article the present work.
distributed under the terms and Numerical optimization has been widely used in the design and analysis of gas tur-
conditions of the Creative Commons bines. In some earlier studies, specific optimization algorithms have been investigated to
Attribution (CC BY-NC-ND) license leverage lower fidelity models to achieve fast optimization time. One study on aerody-
(https://creativecommons.org/ namic wing optimization has used an approximation and model management optimization
licenses/by-nc-nd/4.0/).
Int. J. Turbomach. Propuls. Power 2022, 7, 20. https://doi.org/10.3390/ijtpp7030020 https://www.mdpi.com/journal/ijtpp

Int. J. Turbomach. Propuls. Power 2022, 7, 20 2 of 19
method to incorporate low fidelity, computationally cheaper models with occasional re-
course to higher fidelity, more expensive models, resulting in threefold saving in optimiza-
tion time [1]. Another study adopted a similar concept, employing a trust-region approach
to interleave the exact models with cheaper surrogate models during optimization iter-
ations [2]. These methods demonstrate the possibility of obtaining optimized solutions
on a limited computational budget by incorporating lower-fidelity surrogate models. In
more recent years, an increasing number of optimization studies have relied on using
parametric CFD models. In optimization of the coolant flow passage of the NASA C3X
vane, different designs were evaluated repeatedly through CFD runs [3]. In the study
of a marine high-pressure turbine [4], ten design parameters controlling multi-row film
cooling designs were built into the CFD model and optimized based on a non-dominated
sorting genetic algorithm. Multiple studies were also conducted on ultra-super-critical
steam turbines [5,6], in which the blade aerodynamic efficiency was optimized using 2D
and 3D CFD simulations driven by Siemens’ Simcenter HEEDS commercial Sherpa opti-
mization algorithm. A CFD-based co-optimization strategy was presented in [7], which
demonstrates a workflow for coupling different disciplines into a nested optimization loop
to conduct parallel blade aerodynamic and thermal optimizations. In addition to improving
turbine blade designs, optimization has also been applied to improving the operations
of gas turbine engines, for instance, to find the best valve setup parameters that reduce
fuel consumption [8].
With the advancements in computer science and data storage, an increase in interest
in application of machine learning methods to gas turbine designs has been observed.
One area of such applications seeing increased interest has been the prediction of key
performance metrics using models trained on input data gained through past simulations
or experiments. In an earlier study [9], the outlet temperature and fuel mass flow rate
at different operating conditions for a 255 MW single-shaft gas turbine were predicted
by building a three-layer neural network model. Another application of the Artificial
Neural Network (ANN) model has been seen in a turbine film cooling study to predict the
instantaneous temperature distributions along the blade surface as well as the cooling effec-
tiveness [10]. In another study of a jet engine power plant [11], a machine learning method
combining physics-based and measurement-driven modeling was developed and used to
conduct preventive maintenance and diagnose faults. Machine learning methods were also
applied to a Viper 632-43 military turbojet engine to predict the exhaust temperature using
models trained on data collected from a gas turbine simulation program [12]. Extending
from predicting individual engineering metrics, machine learning has also been used to
predict field quantities representing more complicated underlying physics. A method using
gradient boosted trees was used to develop models of aerodynamic loads on vibrating
turbine blades and demonstrated to have good agreement with detailed CFD results [13].
In another study, the turbine surface pressure distribution was predicted using transfer
learning models, which transfer knowledge from a large-scale but low-fidelity dataset to a
small-scale but high-fidelity dataset, shown to have a low prediction error with reduced
cost [14]. Machine learning has also been applied to the prediction of operating charac-
teristics of gas turbine engines, using real-time data of power plants to develop neural
networks [15,16]. In addition to the above applications, machine learning has also been
studied to develop turbulence closure models. In one study for wake mixing, a machine
learning model was demonstrated to be robust across several different operating conditions
when integrated into a RANS CFD model of a low-pressure turbine [17]. A review article
on machine learning methods for science and engineering particularly highlighted the need
for interpretable, generalizable, expandable, and certifiable machine learning techniques
for safety-critical applications [18].
Recent works have also focused on using machine learning embedded into design opti-
mization procedures. In the optimization of a centrifugal compressor impeller [19], an ANN
model was first developed using CFD and FEA data from a Design of Experiment (DOE)
study. Then, the ANN model was applied in an optimization procedure, which resulted in a
1% increase in isentropic efficiency and 10% reduction in the blade stress. In another study
investigating a carved blade tip [20], 55 CFD runs were conducted to generate ANN meta
models, which were then used in a genetic algorithm routine to optimize the blade tip shape.
In a missile control surface optimization study [21], machine learning, reinforcement learn-
ing, and transfer learning were integrated into the optimization procedure and leveraged
CFD in the evaluation iterations. In another study of 2D airfoil optimization [22], a deep
convolutional generative adversarial network was trained and embedded as a surrogate
model in an optimization framework. In still another study of a compact turbine rotor [23],
machine learning models were trained and used to optimize the efficiency and torque based
on a gradient-based multi-objective optimization algorithm. In addition to using machine
learning models in optimization, several CFD and optimization studies have compared
machine learning models with response surface models (RSM). In a study of aircyclone
optimization [24], it is concluded that ANN offers an alternative and powerful approach to
response surface methods for modeling the cyclone pressure drop, benchmarked against
experimental data. In another study of modeling and optimizing a perforated baffle used
for turbine blade passage cooling [25], both ANN and RSM methods were found capable of
predicting friction factor and Nusselt number values, although the RSM method performed
slightly better than ANN in that study. A more recent study on cyclone optimization also
tested a RSM and several machine learning models. A GMDH-neural network model was
found to be superior and chosen for the optimization process [26].
As discussed in the above literature, optimization that leverages high-fidelity CFD
simulations can provide accurate and realistic optimal designs in general. While leveraging
a neural network as a surrogate to replace the CFD evaluations in the optimization loop
can significantly improve the computational time, the fundamental challenge is that a
neural network model is a lower fidelity model compared to a CFD simulation, and
therefore, it may not be as accurate as CFD in predicting certain design variants required
by the optimization process. Further, neural network models may also be trained based on
previously available datasets that come from different studies, such as 1D and 2D simulation
data, and experimental data, all of which will result in the neural network being a reduced
order model compared to 3D CFD. As such, relying solely on the neural network when the
predictive accuracy is necessary can lead to errant results. The present study presents a
nested optimization workflow that can leverage both neural networks and high-fidelity,
3D CFD simulations at the same time to ensure every “best” design is studied in detail by
the CFD tool. The introduction of a neural network into the optimization allows for over
70% reduction in the number of CFD evaluations and thus makes significant reductions in
computational cost compared to a process relying exclusively on CFD simulations. The
methodology is demonstrated through an aerodynamic optimization using a rebuilt model
of the NASA/General Electric E3 high pressure turbine blade [27–29].
2. Methodology
Artificial Neural Network models are used alongside 3D CFD simulations in a nu-
merical optimization procedure to improve the aerodynamic performance of a turbine
blade. ANN models are typically trained using large datasets obtained from previous
experimental or numerical studies. In companies/organizations that conduct R&D on engi-
neering designs, these datasets are usually available from previous studies. The present
study proposes a methodology of using such existing knowledge from previous studies to
train ANN models and then use the ANN models in an CFD-based design optimization
process. A dataset from a previously published work is obtained to conduct the ANN
modeling training in the present study. The overall framework of the research methodology
is shown in Figure 1, which illustrates how different analysis models/tools required by
the optimization process are created. First, data were obtained from a previous study [7]
containing 3204 CFD design evaluations of different 2D turbine blade profiles. Performance
metrics including efficiency and power were extracted as targets, with the blade geometric
parameters serving as input features, to train ANN models. The ANN hyperparameters
[7] containing 3204 CFD design evaluations of different 2D turbine blade profiles. Perfor-
mance metrics including efficiency and power were extracted as targets, with the
Int. J. Turbomach. Propuls. Power 2022, 7, 20
blade
4 of 19
geometric parameters serving as input features, to train ANN models. The ANN hyperpa-
rameters were subsequently optimized. The blade cross-sectional profiles were parame-
terized using a Class-Shape Transformation (CST) method [30]. Using the 2D blade cross-
were subsequently optimized. The blade cross-sectional profiles were parameterized using
sectional profile,
a Class-Shape the 3D blade(CST)
Transformation CADmethod
geometry [30].was created
Using the 2Dwith
bladeadditional parameteriza-
cross-sectional profile,
tion allowing for the variation of further design transformation,
the 3D blade CAD geometry was created with additional parameterization allowing such as the thickness,
for
twist, and lean.ofThe
the variation blade
further CADtransformation,
design geometry wassuch thenasused to build a twist,
the thickness, 3D Reynolds
and lean.Averag-
The
ing Navier–Stokes
blade CAD geometry (RANS) CFD used
was then model, whicha solves
to build for theAveraging
3D Reynolds flow field Navier–Stokes
in a single blade
passage.
(RANS)Once CFD the ANN
model, models
which andfor
solves thethe
3Dflow
CFDfield
simulation areblade
in a single independently
passage. Oncedeveloped,
the
ANN
they aremodels and the
integrated 3DaCFD
into simulation
nested are independently
optimization procedure developed,
for repeatedthey are integrated
evaluations of dif-
into ablade
ferent nested optimization
design variants procedure
to optimize forthe
repeated
design evaluations
parameters.ofIndifferent
the nested blade design
optimization
variantsan
process, to inner
optimize the design loop
optimization parameters. In theusing
is executed nestedthe optimization
ANN models process, an inner the
to improve
optimization loop is executed using the ANN models to improve the blade
blade cross-sectional profile. In each iteration of the global optimization, the best blade cross-sectional
profile.
profile is In each iteration
selected from theof the global
inner optimization,
optimization stepthe
to best blade
create a 3Dprofile
CFD is selected
blade modelfrom with
the inner optimization step to create a 3D CFD blade model with additional parameters
additional parameters such as the scaling factors, twist, and lean angles. The optimization
such as the scaling factors, twist, and lean angles. The optimization targets improving the
targets improving the efficiency and power of the blade. Details of the optimization pro-
efficiency and power of the blade. Details of the optimization process will be discussed in a
cess
laterwill be discussed
session in a later session in this chapter.
in this chapter.
Figure
Figure1.1.The
Theoverall
overallframework
framework of
of the
the research methodology.(2D
research methodology. (2DCFD
CFDdata
datawas
was collected
collected from
from a a
previous study [7].)
previous study [7].)
2.1.Artificial
2.1. ArtificialNeural
NeuralNetwork
Network
A dataset representing different design variants of 2D aerodynamic CFD results was
A dataset representing different design variants of 2D aerodynamic CFD results was
obtained from a previous study [7] and used to train ANN models. The dataset features
obtained from a previous study [7] and used to train ANN models. The dataset features
blade design and performance parameters for 3204 designs. The input parameters for the
blade
ANNdesign
models and performance
include 20 bladeparameters for 3204 These
design parameters. designs. The input
parameters areparameters
the weighting for the
ANN models include 20 blade design parameters. These parameters
factors used by the CST method [30] to control variations in the blade profile. Under the are the weighting
factors used bythese
CST method, the CST method
weighting [30] are
factors to control variations
multiplied in thepolynomials
by Bernstein blade profile. Undera the
to define
CST method,
shape function,these and weighting factors aremultiplied
then subsequently multiplied byby Bernstein
a class polynomials
function to define the to define
aero- a
shape function, and then subsequently multiplied by a class function
dynamic blade profile. An order-9 Bernstein polynomial was deemed sufficiently flexible to define the aero-
dynamic blade profile.
for the purposes of theAn order-9
study Bernstein
[7] and polynomial
that choice is repeatedwas deemed
here, sufficiently
resulting in 10 design flexible
parameters
for the purposesfor eachof the of the suction
study and pressure
[7] and sides
that choice is of the blade:
repeated {wu,1
here, , wu,2 , . .in
resulting wu,10
. , 10 }
design
parameters for each of the suction and pressure sides of the blade: {𝑤 , , 𝑤 , , … , 𝑤 , }
and w l,1 , w l,2 , . . . , w l,10 . These 20 parameters are used as input features in the ANN
models
and {𝑤 , and
, 𝑤 , optimization
, … , 𝑤 , }. These parameters in the subsequent
20 parameters are used as blade
inputoptimization
features inprocess.
the ANN Two
mod-
performance metrics in the dataset were used as output parameters,
els and optimization parameters in the subsequent blade optimization process. Two per- isentropic efficiency,
formance metrics in the dataset were used as output parameters, isentropic efficiency, and
and power output, for which two ANN models were developed, respectively. The efficiency
and power are defined based on enthalpy quantities of the flow, as follows:
hi − h o
η= , (1)
hi − ho,isen.
Pow = ∆h = hi − ho (2)
The Keras API for TensorFlow [31] was used to construct the ANN models. In consid-
eration of the different distributions of the efficiency and power data and the robustness
of the model, a separate ANN model was constructed for each of these two quantities.
After some preliminary tests on ANN models of 5, 6, and 7 layers, 7 and 6 layers were
constructed, respectively, into the ANN models representing efficiency and power. The
choice of the number of layers in ANN usually varies in different problems. It will be
shown that the prediction errors are within acceptable range for the chosen number of
layers, and that the errors will be further reduced by optimizing the hyperparameters of
the models. The first layer of each ANN model has 20 inputs representing the blade input
features (CST weights) and the last layer has one output, which is either efficiency or power.
The number of neurons in the hidden layers of each ANN, along with other topological and
training parameters, were optimized in a separate hyperparameter optimization process.
As a starting point for that process, the hyperparameters for each ANN model were set
using the values provided in the first two columns of Table 1.
Table 1. Hyperparameters of the initial and optimized ANN models.
Initial ANN Initial ANN Optimized ANN Optimized ANN

Hyperparameters
for Efficiency for Power for Efficiency for Power
Number of Neurons-Layer 2 20 20 5 23
Number of Neurons-Layer 6 5 – 62 –
Activation Function-Layer 2 relu relu relu selu
Activation Function-Layer 3 relu relu sigmoid softmax
Activation Function-Layer 4 relu relu elu elu
Activation Function-Layer 5 relu relu elu softmax
Activation Function-Layer 6 relu – softmax –
Learning Rate 0.001 0.001 6.4895 × 10−5 1.47285 × 10−3
Number of Epochs 200 200 6123 4000
To improve the predictive performance of the ANN, the hyperparameters were opti-
mized using Siemens commercial optimization tool HEEDS and its Sherpa algorithm [32].
A diagram of the optimization process is provided in Figure 2. The optimizer repeatedly
tunes all the hyperparameters for the ANN models listed in the index column of Table 1
with an objective to minimize the model test error on a held-out testing dataset. In each
optimization cycle, the dataset is randomly split into a training and testing set with a ratio
of 95/5. The training set is used the train the ANN model. The training process involves a
typical backpropagation procedure with an 80/20 split of the training set, over multiple
epochs with TensorFlow’s ADAM optimizer. After the training step is completed, the
model’s performance is then evaluated against the held-out test set. The evaluation error is
passed back to the optimization solver to tune the hyperparameters for the next cycle. A
total number of 240 iterations were executed during the optimization of each of the two
ANNs. The evaluation number, 240, was based on the Sherpa algorithm best practice [32],
which recommends the number of evaluations being about 10 times the number of input
variables for single objective optimization problems, due to the hybrid and adaptive meth-
ods employed by the algorithm. In the ANN optimization, 12 hyperparameters must be
t. J. Turbomach. Propuls. Power 2022, 7, x FOR PEER REVIEW 6 of

and adaptive methods employed by the algorithm. In the ANN optimization, 12 hyperp
rameters must be optimized, and thus it requires at least 120 evaluations according to th
best optimized,
practice. In
andconsideration
thus it requiresofatthe
leastrelatively fast run
120 evaluations time for
according thebest
to the ANN training
practice. In proce
in comparison toofthe
consideration the3D CFD run
relatively time,
fast run that
time forevaluation number
the ANN training is further
process doubled to 2
in comparison
to the 3D CFD run time,
to provide better accuracy. that evaluation number is further doubled to 240 to provide
better accuracy.
Figure 2. ANN
Figure model
2. ANN optimization
model diagram.
optimization diagram.
The optimized hyperparameters are obtained and provided in the last two columns of
The optimized hyperparameters are obtained and provided in the last two column
Table 1. After optimization, the evaluation errors for efficiency and power were reduced,
of Table 1. After
respectively, fromoptimization,
0.33% to 0.10%,theandevaluation
from 0.43% toerrors
0.10%. for
Theefficiency
performanceand power
of the initial were r
duced,
andrespectively,
optimized ANNfrom 0.33%
models to examined
are also 0.10%, and from 0.43%
by comparing to 0.10%.values
the predicted The vs.performance
true
values from
the initial and the test set, shown
optimized ANN in models
Figure 3. It
areis observed that the optimized
also examined ANN models
by comparing the predicte
result in better alignment between the predicted values and the true values, as well as
values vs. true values from the test set, shown in Figure 3. It is observed that the optimize
narrower bandwidths of the prediction errors on the histogram plots. From Figure 3, it
ANN is models resultthat
also observed in better
the biasalignment between
of the optimized ANN theforpredicted values
efficiency was andreduced,
greatly the true value
as well asdata
as the narrower
points are bandwidths of theon
evenly distributed prediction
both sides of errors onofthe
the line histogram
unity slope, shown plots. Fro
Figure 3, itsecond
in the is alsosubplot.
observed Thisthat the biaswith
is consistent of the
theoptimized
histograms of ANN for efficiency
the prediction errorswas
in great
Figureas4,the
reduced, which
data shows
pointsthatare
both the bias
evenly and bandwidth
distributed of the
on both errors
sides of were improved
the line of unity slop
in the optimized ANN models. The optimized ANN models are used in the subsequent
shown in the second subplot. This is consistent with the histograms of the prediction e
blade optimization process in this study. The Python scripts for training the ANN models
rors with
in Figure 4, which
the optimized shows that both
hyperparameters the biasin and
are provided [33]. bandwidth
The CPU timeof forthe errors
training an were im
proved in the ANN
individual optimized
largely ANN
depends models. The optimized
on the number of Epochs. ANN models
In the present are the
study, usedCPU in the su
sequent
timesblade optimization
for training processrepresenting
the ANN models in this study. The Python
efficiency and power scripts for training
with their respectivethe AN
optimized
models with the hyperparameters were 14.3 min andare
optimized hyperparameters 8.5 min, respectively.
provided in [33].The
The total
CPU compute
time for trai
times for optimizing the hyperparameters of these ANN models were, respectively, 13.07 h
ing an individual ANN largely depends on the number of Epochs. In the present stud
and 11.65 h. This optimization was carried out based on parallel execution of 10 training
the CPU
modelstimes
at thefor training
same the ANN
time using a 6-coremodels
CPU. representing efficiency and power with the
respective optimized hyperparameters were 14.3 min and 8.5 min, respectively. The tot
compute times for optimizing the hyperparameters of these ANN models were, respe
tively, 13.07 h and 11.65 h. This optimization was carried out based on parallel executio
of 10 training models at the same time using a 6-core CPU.
Int.
Int.J. J.Turbomach.
Turbomach.Propuls.
Propuls.Power 2022,7,7,20
Power2022, x FOR PEER REVIEW 7 of2319
7 of
Int. J. Turbomach. Propuls. Power 2022, 7, x FOR PEER REVIEW 8 of 23
Figure3.3. Predicted
Figure Predicted vs.
vs. true
true values
values of
of the
the initial
initial and
and optimized
optimized ANN
ANNmodels.
models.
Figure 4.
Figure 4. Histogram
Histogram plots
plots of
of the
the prediction
prediction errors
errorsof
of the
theinitial
initialand
andoptimized
optimizedANN
ANNmodels.
models.
2.2.
2.2. Parametric
Parametric BladeBlade CADCAD and and CFD
CFD Model
Model
An initial turbine
An initial turbine blade bladeCAD CADmodel modelwas was built
built byby adapting
adapting thethe E3 high
E3 high pressure
pressure tur-
turbine
bine bladeblade profiles.Using
profiles. Using thethefirst
firststage
stagerotor
rotorblade
blade coordinates
coordinates provided
provided at at three
three
spanwise
spanwiselocations
locations[27]—hub
[27]—hub(12.693in),
(12.693in),pitch pitch(13.571in),
(13.571in),and
andtiptip
(14.41in)—and
(14.41in)—and accounting
account-
for
ing for the incoming flow angle [28,29], a 3D CAD model was constructed by lofting these
the incoming flow angle [28,29], a 3D CAD model was constructed by lofting these
cross-sectional
cross-sectional profiles,
profiles, as as shown
shown in in Figure
Figure 5a, 5a, to
to provide
provide aabaseline
baseline case.
case. Applying
Applying the the
same
same philosophy
philosophy of of lofting
lofting cross cross sections
sections and and atat the
the same
same time
time utilizing
utilizing the
the CST
CST method
method
for the cross-sectional
for the cross-sectionalprofile profiledefinition
definition[27], [27],a parameterized
a parameterized blade
blade CADCAD model
model waswas cre-
created for the design optimization study. First, the CST method
ated for the design optimization study. First, the CST method creates a base cross-sectional creates a base cross-
sectional profile,
profile, which which
is used as is used
the pitchasprofile.
the pitch Theprofile. The CST
CST method method
allows allows
varying thevarying the
base profile
base profile in design optimization through manipulating
in design optimization through manipulating the 20 weighting parameters, the 20 weighting parameters,
{{𝑤
wu,1,, 𝑤wu,2, …
, .,.𝑤. , wu,10 } and wl,1 , wl,2 , . . . , wl,10 . Then, the base profile is scaled to create
, , , } and {𝑤 , , 𝑤 , , … , 𝑤 , }. Then, the base profile is scaled to create the
the hub and tip profiles. In the scaling process, the twist angles and chord lengths of the
hub and tip profiles. In the scaling process, the twist angles and chord lengths of the hub
hub and tip profiles are matched to their counterparts of the original E3 blade; 4 additional
and tip profilesnare matched to their counterparts o of the original E3 blade; 4 additional
scaling
scaling factors,
factors, {𝜉 ξ hub,l, ,, 𝜉ξ hub,u , ξ , , 𝜉, ξ tip,u
, , 𝜉 tip,l
, were applied to the lower and upper profiles
, }, were applied to the lower and upper profiles of
the hub and tip sections, respectively, to allow for small variations of the thicknesses of
the hub and tip sections so that they can be optimized. The ranges of these variations are
defined conservatively out of practical considerations of the original E3 engine blade
shape. In general, the scaling allows the hub profile to be thicker and the tip profile to be
thinner. Finally, a tilt angle, 𝜃, and a lean angle, 𝛽, were also built into the CAD model to
of the hub and tip sections, respectively, to allow for small variations of the thicknesses
of the hub and tip sections so that they can be optimized. The ranges of these variations
are defined conservatively out of practical considerations of the original E3 engine blade
shape. In general, the scaling allows the hub profile to be thicker and the tip profile to be
thinner. Finally, a tilt angle, θ, and a lean angle, β, were also built into the CAD model to
be optimized later. Based on a reverse engineering study analyzing different engine blade
Int. J. Turbomach. Propuls. Power 2022, 7,geometries [34], the tilt angle can vary from 0 to 0.164◦ , and the lean angle can vary
x FOR PEER REVIEW 9 offrom
23
◦ ◦
−0.086 to 0.086 in the optimization study. A schematic plot of these angles is shown
in Figure 5b.
(a) (b)
Figure 5. Blade geometry schematics: (a) blade geometry adapted from E3 profile coordinates; (b)
Figure 5. Blade geometry schematics: (a) blade geometry adapted from E3 profile coordinates;
blade angles schematic.
(b) blade angles schematic.
Using
Using the 3D3D
the blade,
blade, a single-blade
a single-blade passage
passage CFDCFD model
model waswasdeveloped
developed using
using Siemens
Siemens
multi-physics
multi-physics package Simcenter STAR-CCM+. The continuity, momentum, and en-
package Simcenter STAR-CCM+. The continuity, momentum, and fluid fluid
ergy transport
energy transportequations
equations are are
solved
solvedusing a coupled
using a coupled solver
solver(density-based)
(density-based) following
following aa
finite volume approach with a 2nd order upwind discretization
finite volume approach with a 2nd order upwind discretization scheme on a polyhedral scheme on a polyhedral
grid. Menter’s
grid. Menter’s SSTSSTK-Omega
K-Omega model
model [35] with
[35] with allall
y+y+ treatment
treatment is is
used
used asasa closure
a closure totothethe
turbulence
turbulence model.
model. The
The allall
y+y+wall
walltreatment
treatment adjusts
adjusts the application
the application ofof a turbulence
a turbulence wall
wall
function
function based
based ononthe
thelocal
local y+y+
value
value ofofthethenear
near wall
wallmesh
mesh cell. The
cell. The CFDCFD simulations
simulations were
were
performed using the test conditions. Following E3 engine rotor
performed using the test conditions. Following E3 engine rotor testing conditions and testing conditions and 2D2D
CFD practices from NASA studies [28,29], a total pressure of 344,777
CFD practices from NASA studies [28,29], a total pressure of 344,777 Pa and a total temper- Pa and a total tem-
perature
ature ofof709709KK were
were usedusedforfor
thethe inlet
inlet boundary
boundary condition,
condition, while
while an atmospheric
an atmospheric pres-of
pressure
sure of 101,325
101,325 Pa wasPa was defined
defined at the The
at the outlet. outlet.sameThe same boundary
boundary conditions conditions
representing representing
the engine
the engine
testing testing conditions
conditions were alsowere alsoin
applied applied in the previous
the previous 2D CFD simulations
2D CFD simulations in [7].2D
in [7]. (These
(These 2D CFD were
CFD datasets datasets
used were used toANN
to develop developmodelsANNinmodels
the presentin the present
work.) work.)
A grid A grid
sequencing
sequencing
method was method
used towas used toinitial
provide provide initial conditions,
conditions, by solvingby solvingflow
inviscid inviscid flow equa-
equations repeat-
tions
edlyrepeatedly
on a set ofon a set of gradually
gradually refined meshrefined mesh
grids. grids. Automatic
Automatic CFL number CFL number
control control
was also
was also applied
applied to adjusttotheadjust
CFL the CFL in
number number
response in to
response to linear
linear solver solver convergence
convergence behavior duringbe-
havior during the
the algebraic algebraic
multigrid multigridTo
procedure. procedure. To ensure
ensure overall overall
solution solution convergence
convergence is reached, in
is addition
reached, to in addition
monitoring to monitoring
the residuals theofresiduals of the governing
the governing fluid equations,
fluid equations, stop-
stopping criteria
were
ping also set
criteria based
were alsoonsetthe asymptotic
based values of engineering
on the asymptotic quantities, such
values of engineering as isentropic
quantities, such
efficiency, power output, maximum surface temperature, and
as isentropic efficiency, power output, maximum surface temperature, and total pressure total pressure difference.
The CFD methodology has been applied in a previous study based on a 2D version
difference.
of the
TheE3CFDblade and validated
methodology hasagainst other experimental
been applied in a previousand study numerical
based onresults of the E3
a 2D version
ofblade
the E3[7]. In the
blade and present
validatedstudy, a gridother
against independence
experimental investigation
and numerical has also been of
results performed
the E3
for the
blade [7].3D CFD.
In the Three study,
present sets ofaCFDgridgrids with different
independence resolutions
investigation haswere
alsotested. The chosen
been performed
forgrid
the features
3D CFD.12 million
Three setspolyhedral
of CFD grids cells
withwith near-wall
different y+ values
resolutions wereall tested.
below The 1.4, based
chosenon
which the solution calculates a baseline efficiency of 96.0% and a power of 5.638 × 10on 4 W.
grid features 12 million polyhedral cells with near-wall y+ values all below 1.4, based
(Notethe
which power is calculated
solution calculates based on the efficiency
a baseline single blade of CFD
96.0%model and ainpower
the current
of 5.638study,
× 10which
4 W.
has apower
pie sector angle of 4.737 ◦
(Note is calculated based).on The therelative
single difference
blade CFDbetween model in thethechosen gridstudy,
current and its
next-level
which has arefined gridangle
pie sector is 0.1% of for efficiency
4.737°). and 0.02%
The relative for power.
difference between the chosen grid
and its next-level refined grid is 0.1% for efficiency and 0.02% for power.
2.3. Optimization Strategy

The present study adopts a nested optimization process. The nested optimization is
ower 2022, 7, x FOR PEER REVIEW 10 of 23

for a 3D CFD run to complete. For each 3D CFD simulation run, 250 evaluations of differ-
ent blade cross-sectional profiles were completed in the inner optimization loop using the
ANN model. Second,2.3. theOptimization Strategy
inner, ANN-driven optimization passes the best blade profile of
The present study adopts a nested optimization process. The nested optimization
the present optimization cycle to the 3D CFD run. This architecture effectively allows us-
is a slight variation of a previous co-optimization strategy introduced in [7]. The nested
ing the ANN-driven optimization to guidetwo
optimization accomplishes theessential
3D CFD search,
tasks. and thus
First, because an ANNreduces modelthe num-
calculation
consumes significantly less computing resources
ber of 3D CFD runs compared to an optimization process that is solely based on evaluat- and time than a 3D CFD run, nesting an
inner optimization based on the ANN model can effectively utilize the time while waiting
ing 3D CFD models. for a 3D CFD run to complete. For each 3D CFD simulation run, 250 evaluations of different
The goal of the optimization
blade cross-sectional is to maximize
profiles were completedbladeinefficiency,
the inner optimizationη, and loop power usingoutput,
the ANN
model. that
Pow. The design variables Second,arethe inner, explored
being ANN-driven inoptimization
optimization passes arethethebestones
blade discussed
profile of the
present optimization cycle to the 3D CFD run. This architecture effectively allows using the
in the previous session, 𝑤 , , 𝑤optimization
ANN-driven , , … , 𝑤 , to, guide 𝑤 , ,the
𝑤 ,3D , …CFD
, 𝑤 search,
, , 𝜉 and, thus , 𝜉 reduces
, , 𝜉 the , , 𝜉number , , 𝜃 of
and 𝛽. The optimization
3D CFD procedure
runs compared repeatedly executes
to an optimization process thethat ANN and
is solely CFD
based models to
on evaluating 3D
CFD design
evaluate different blade models. variants, while extracting performance metrics to search
The goal of the optimization is to maximize blade efficiency, η, and power output, Pow.
for better designs. AsThe
thedesign
ANN model is much less computationally expensive compared
variables that are being explored in optimization are the ones discussed in the
to CFD, it is leveraged in an session,
previous optimization
wu,1 , wu,2 ,loop
. . . , wby
u,10 ,itself
wl,1 , wto
l,2 , yield optimized
. . . , wl,10 cross-sectional
, ξ hub,l , ξ hub,u , ξ tip,l , ξ tip,u , θ and
Thesubsequently
blade profiles, which β.are optimization procedureused to repeatedly
construct executes
3D the ANN and
shapes for CFDCFDmodels to evaluate
simulations.
different blade design variants, while extracting performance metrics to search for better
Using the ANN embedded optimization to guide the search for the best blade profiles
designs. As the ANN model is much less computationally expensive compared to CFD, it is
allows for reducing the number
leveraged in anof expensive
optimization loop CFD evaluations.
by itself to yield optimized As such, a nested
cross-sectional blade optimi-
profiles,
which are subsequently used to construct 3D shapes
zation workflow is constructed, as illustrated in Figure 6. The workflow is realized using for CFD simulations. Using the ANN
embedded optimization to guide the search for the best blade profiles allows for reducing
commercial optimization software Simcenter HEEDS and its Sherpa optimization algo-
the number of expensive CFD evaluations. As such, a nested optimization workflow
rithm [30] is used in both Optimizer
is constructed, 1 and Optimizer
as illustrated in Figure 6. 2.The The Sherpais algorithm
workflow realized using constantly
commercial
evaluates the characteristics
optimization of the problem
software Simcenterby using
HEEDS a and
hybrid combination
its Sherpa optimization of algorithm
search strat- [30] is
used in both Optimizer 1 and Optimizer 2. The Sherpa algorithm constantly evaluates
egies at each stage ofthethe optimization process to best traverse the design landscape. In
characteristics of the problem by using a hybrid combination of search strategies at
addition, it also adapts
eachthestage
tuning
of theparameters
optimization process of thetosearch strategies
best traverse the designto the specific
landscape. In region
addition,
of the search. The HEEDS’ Sherpa algorithm has been effectively applied to optimizeofgas
it also adapts the tuning parameters of the search strategies to the specific region the
search. The HEEDS’ Sherpa algorithm has been effectively applied to optimize gas turbine
turbine applications, applications,
as shown in studies [5–7].
as shown in studies [5–7].
Figure 6. Optimization Figure

iteration hierarchyiteration
6. Optimization and data flow.and data flow.
hierarchy
In the global optimization loop (Optimizer 1 as shown in Figure 6), a multi-objective

In the global optimization loopis(Optimizer
trade-off problem 1 asmaximizing
solved targeting shown inbothFigure 6), aand
efficiency multi-objective
power. Two con-
trade-off problem is straints
solvedaretargeting maximizing
set: efficiency both
must be greater thanefficiency andmust
95% and power power.
also beTwo con-
greater than
a baseline power value. These constraints are included in the optimization performance
straints are set: efficiency must be greater than 95% and power must also be greater than
function, based on considerations of practical gas turbine design metrics and general
a baseline power value. These constraints are included in the optimization performance
function, based on considerations of practical gas turbine design metrics and general per-
formance characters of the E3 engine. The global optimization solver, Optimizer 1, affects
the blade shape change by directly optimizing the section scaling factors and blade angles
performance characters of the E3 engine. The global optimization solver, Optimizer 1,

affects the blade shape change by directly optimizing the section scaling factors and blade
angles in the CAD. In addition, Optimizer 1 also affects an inner optimization process,
which is represented by Optimizer 2. Optimization 1 controls the selection of the CST
weights used as input parameters for Optimizer 2, through manipulating a random seed,
R, which sets the initial state for the global search strategies employed by the Sherpa
algorithm. Although Optimizer 1 does not directly manipulate the CST weights, because R
controls their selection, it acts as a 1D encoding of the entire 20-dimensional design space
for Optimizer 1 to explore. Optimizer 1 repeatedly evaluates and optimizes different design
variables during the global optimization cycles; in each evaluation cycle, Optimizer 2
optimizes the CST weights using the ANN models, with an objective to maximize the
efficiency and a constraint on minimum power value. Overall, 250 blade cross-sectional
profiles were evaluated in each execution of Optimizer 2. It directly explores the shape
profiles through manipulating the weighting parameters used in the CST method. In each
evaluation cycle of Optimizer 1, the best blade profile found by Optimizer 2 is applied in
the more expensive and high-fidelity CFD simulation. The performance function in each
optimization is defined as the sum of normalized objectives minus the sum of normalized
constraints. In a generic form, the performance function is formulated as follows:
i = Nobj Obji j= N 104 × ConsViol 2

P= ∑ i =1 Normi
− ∑ j=1 cons
Norm j 2
, (3)
In the above formula, Obji represents the objective quantity, such as η and Pow;
ConsViol represents the amount of violation resulted from each constraint definition
(η > 0.95, Pow > Powbaseline ); the normalization factors Norm are selected using the respec-
tive η and Pow values of the baseline design.
3. Results
3.1. Optimization Results
In the present optimization study, a total of 163 blade designs have been evalu-
ated in 3D CFD, with 40,750 2D blade profiles being evaluated by the ANNs. An effi-
ciency vs. power plot showing all 3D blade design points is provided in Figure 7. A Pareto
front is formed in the upper right region of Figure 7, which contains optimal designs with
the best trade-off relations between efficiency and power. Among the optimal designs, the
maximum increase in efficiency is over 3%, and the maximum increase in power is around
8%. To obtain insight into the design variable distributions of the optimal designs, two
groups of optimal designs are selected—a group of top 10 high-efficiency designs, and
another group of top 10 high power designs. The responses and selected design input
variables (which include the two blade angles and four scaling factors) for these two groups
of designs are highlighted on a parallel plot in Figure 8. The parallel plot reveals some qual-
itative trends in the design parameters: (1) both the high-efficiency and high power designs
favor large lean angles, β; (2) the high-efficiency designs also favor a large scaling factor for
the tip upper surface, ξ tip,u , while the high power designs favor a small-to-medium value
for ξ tip,u ; (3) a weaker influence of the scaling factor for the hub lower surface, ξ hub,u , is
also observed. Note that the lean angle was constrained conservatively within a smaller
range in consideration of commercial turbine geometric characteristics provided in [27].
tive trends in the design parameters: (1) both the high-efficiency and high power designs
favor large lean angles, 𝛽; (2) the high-efficiency designs also favor a large scaling factor
for the tip upper surface, 𝜉 , , while the high power designs favor a small-to-medium
value for 𝜉 , ; (3) a weaker influence of the scaling factor for the hub lower surface,
𝜉 20 , , is also observed. Note that the lean angle was constrained conservatively within
Int. J. Turbomach. Propuls. Power 2022, 7, a
11 of 19
smaller range in consideration of commercial turbine geometric characteristics provided
in [27].
Figure 7. 7.Efficiency
Figure Efficiencyvs.
vs. power forall
power for alldesigns
designspoints
pointsin in
thethe optimization
optimization study.
study.
Figure 8. 8.Parallel
Figure Parallelplot
plothighlighting selecteddesign
highlighting selected design parameters
parameters of different
of different groups
groups of designs.
of designs.
AAquantitative
quantitative analysis hasalso
analysis has alsobeen
beenconducted.
conducted.TheThe Pearson
Pearson correlation
correlation coefficients
coefficients
of of
thethe dataare
data arecalculated,
calculated, based
basedon
onthe
thecovariance
covarianceand standard
and deviations.
standard The definition
deviations. The definition
is givenasasfollows:
is given follows:
n
∑ i =1 ( x i − x ) Y i − Y

r= q ∑ q (𝓍 ̅ )(𝓎 𝓎) 2, , (4)
𝔯=
∑in=1 ( xi∑− x(𝓍 )2 ∑ n
Y − Y (4)
i
̅ ) i =∑1 (𝓎 𝓎 )
n o
the above formula, x, Y ∈ η, Pow, β, θ, ξ hub,l , ξ hub,u , ξ tip,l , ξ tip,u , ( x 6= Y ).
In the In above formula, 𝓍, 𝓎 ∈ 𝜂, 𝑃𝑜𝑤, 𝛽, 𝜃, 𝜉 , , 𝜉 , , 𝜉 , , 𝜉 , , ( 𝓍 ≠ 𝓎). 𝓍 and 𝓎
x i and Y i represent
represent sample values sample values corresponding
corresponding to each blade to eachdesign,
blade design,
and n isand
then total
is the number
total of
number of designs, which is 163 in the present study. The Pearson correlation coefficients,
designs, which is 163 in the present study. The Pearson correlation coefficients, the histo-
the histogram distributions of selected parameters, and the data sample plots are shown on
gram distributions
a correlation plot inofFigure
selected parameters,onand
9, respectively, thethe upper data sample
right plots
region, alongarethe
shown on a cor-
diagonal
relation
line joining the upper left and lower right corners, and on the lower left region. Large line
plot in Figure 9, respectively, on the upper right region, along the diagonal
joining the upper
magnitudes of theleft and lower
Pearson right corners,
correlation coefficients andfor on the
ξ tip,ulower leftrow,
, η (1st region.
last Large
column) magni-
tudes
and {ofβ, thePowPearson
} (2nd row, correlation
3rd column) coefficients
indicate strong 𝜉 , , 𝜂 (1stinrow,
for relationships theselast
two column)
pairs of and
{𝛽,variables.
𝑃𝑜𝑤} (2nd Therow,
large3rd
coefficient
column) values are confirmed
indicate by their respective
strong relationships datatwo
in these sample
pairsplots
of varia-
of ξThe
bles. tip,u vs. η (last
large row, 1st column)
coefficient values are β vs. Pow (3rd
andconfirmed row, 2nd
by their column),
respective in which
data sample most
plots of
of the design points follow the linear regression lines closely. In addition, the correlation
𝜉 , vs. 𝜂 (last row, 1st column) and 𝛽 vs. 𝑃𝑜𝑤 (3rd row, 2nd column), in which most
plot also shows that the effect of the scaling factor for the hub lower surface, ξ hub,u , is
of rather
the design
mild.
points follow the linear regression lines closely. In addition, the correlation
plot also shows that the effect of the scaling factor for the hub lower surface, 𝜉 , , is
rather mild.
Figure 9. Correlation
Figure plot
9. Correlation of selected
plot design
of selected parameters
design parametersofof
allall
designs.
designs.(The two
(The most
two mostinfluential
influential
numbers
numbersand their
and corresponding
their correspondingplots
plotsare
arehighlighted
highlighted using the dashed
using the dashedarrows.)
arrows.)
The Theefficiency
efficiencyand
andpower
powerare areplotted
plottedversus
versus the
the two parameters
parameters with withthethestrongest
strongest
correlations,
correlations, leanangle
lean angle(𝛽),
(β),and
and tip upper
uppersurface
surfacescaling
scalingfactor (ξ tip,u
factor (𝜉 ),, shown
), shownin Figure 10.
in Figure
It is observed that (1) most of the improved designs favor larger lean angles,
10. It is observed that (1) most of the improved designs favor larger lean angles, and (2) and (2) among
the
Int. J. Turbomach. Propuls. Power 2022,
among7, x improved
FOR
thePEER designs,
REVIEW
improved a smaller
designs, ξ tip,u leads
a smaller 𝜉 , toleads
greater
to power
greaterimprovement,
power improvement,while a14larger
of 23
while
leads to high-efficiency gain. To analyze the effects of the blade
a larger 𝜉 , leads to high-efficiency gain. To analyze the effects of the blade shape and
ξ tip,u shape and the fluid
dynamics, two designs are selected for further investigation—a high-efficiency design and
the fluid dynamics, two designs are selected for further investigation—a high-efficiency
a high power design, which will be referred to as “Design A” and “Design B,”, respectively.
design and a high power design, which will be referred to as “Design A” and “Design B,”,
These two designs are also marked in Figure 10 and will be discussed in the next session.
respectively. These two designs are also marked in Figure 10 and will be discussed in the
next session.
(a) (b)
Figure 10.
Figure 10. Two
Twoviews
viewsofofaa3D
3Dplot
plotshowing
showingefficiency,
efficiency, power,
power, lean
lean angle,
angle, andand
tip tip upper
upper surface
surface scaling
scaling factor of all designs: (a) a normal view; and (b) a view obtained by rotating the left-side
factor of all designs: (a) a normal view; and (b) a view obtained by rotating the left-side plot around
plot around the 𝜂 axis by 180°.
the η axis by 180◦ .
The total run time for the optimization process was around 30 h (with six CFD eval-
uations running in parallel, each consuming 160 compute cores). There are 27 input vari-
ables in the global optimization problem. As illustrated in Figure 6, the integration of the
inner optimization loop allows for optimizing the 20 blade profile parameters using the
The total run time for the optimization process was around 30 h (with six CFD evalua-
tions running in parallel, each consuming 160 compute cores). There are 27 input variables
in the global optimization problem. As illustrated in Figure 6, the integration of the inner
optimization loop allows for optimizing the 20 blade profile parameters using the inexpen-
sive ANN models. Throughout the optimization process, the ANN models were evaluated
40,750 times; in each optimization cycle, the best blade cross-sectional profile obtained
from the inner ANN optimization was passed on to the 3D CFD for evaluation. As a result,
the number of design variables being directly exposed in the 3D CFD runs is reduced to
six. Since the required number of CFD evaluations largely depends on the number of
input variables in an optimization problem, the nested optimization strategy only requires
163 evaluations of 3D CFD simulations. In comparison, if an optimization study of the
same problem does not adopt the nested optimization strategy and is solely based on CFD,
then all 27 design variables (rather than 6) will be evaluated based on 3D CFD runs. Based
on a quick estimation, this will result in 4.5 times (=27/6) the number of CFD runs, which
is 733 CFD runs, and will cost roughly 135 h of optimization run time if the same compute
resource settings were used.
3.2. Fluid Dynamics Analysis

Two optimized designs are selected for detailed investigation of the aerodynamic
results: a high-efficiency design, and a high-power design (a.k.a. Designs A and B as
shown in Figure 9). They are compared with the baseline design. The CFD results show the
optimized designs feature less pressure drop on the suction side near the trailing edge. A
pressure coefficient is defined based on the average inlet velocity and density:
p
cp = 1 2
, (5)
2 ρin uin
The pressure coefficient distributions on the hub, mid, and tip sections are compared
in the upper row of plots in Figure 11. The baseline design shows significant low pressure
regions on the suction surface near the trailing edge on all three streamwise sections. In
comparison, the two optimized designs saw increased minimum pressure in those areas at
the hub and mid sections. An enlarged view of the pressure coefficient distribution on the
mid-section near the trailing edge is provided for each design in the lower row of plots in
Figure 11. The baseline design features a shock in this region, highlighted by a relatively
high pressure spot next to the low pressure zone, which is present inside the dotted circle
marked on Figure 11a. This shock feature is absent in the two optimized designs due to the
modifications in the blade shapes in those designs. This observation is consistent with the
Mach number distributions that will be discussed in a later paragraph.
To further investigate the pressure drop on the suction side near the trailing edge,
volumetric renderings of the pressure coefficient scenes are shown in Figure 12, to highlight
the flow in the low-pressure regime by focusing on c p values between −0.35 and −0.1. It is
observed that in the optimized designs, lower pressure coefficients are present near the tip,
but the pressure coefficient has been increased in the mid-session and nearer the hub. The
overall reduction of maximum negative pressure coefficients in the mid and hub regions
offsets the effects of increased tip leakage flow in the optimized designs.
Next, the pressure coefficient distributions on the blade surfaces are investigated.
Scalar scenes of the pressure coefficients on the baseline, Design A, and Design B blades
are shown in Figure 13. A large portion of the suction-side surface in the baseline design
features a significantly negative c p values between −0.35 and −0.23 starting around the
three quarters chord. In the optimized designs, the onset of the negative pressure coefficient
zone moves upstream, nearer the half chord, yet the magnitude of this zone is much less,
with most areas featuring c p values between −0.12 and 0. To achieve a more quantitative
comparison, the pressure coefficient curves along the streamwise direction on the hub, mid,
and tip sections of the three blades are plotted and compared, with the plots of the blade
profiles, shown in Figure 14. The earlier onsets of the lower pressure coefficient region on
comparison, the two optimized designs saw increased minimum pressure in those areas
Int. J. Turbomach. Propuls. Power 2022, 7, 20at the hub and mid sections. An enlarged view of the pressure coefficient distribution on 14 of 19
the mid-section near the trailing edge is provided for each design in the lower row of plots
in Figure 11. The baseline design features a shock in this region, highlighted by a relatively
high pressure spot next to the low pressure zone, which is present inside the dotted circle
the marked
suction-side surfaces
on Figure in the
11a. This twofeature
shock optimized designs
is absent in theare
twodemonstrated by the
optimized designs dueplots.
to The
earlier onset allows in
the modifications forthe
larger
bladeenclosed
shapes inareas
those of Thiscobservation
negative
designs. p curves for
is the optimized
consistent with design,
contributing to the power
the Mach number improvements.
distributions that will be discussed in a later paragraph.
(a) (b) (c)

Figure 11. Static pressure distributions: (a) Baseline design; (b) Design A; and (c) Design B.
To further investigate the pressure drop on the suction side near the trailing edge,
volumetric renderings of the pressure coefficient scenes are shown in Figure 12, to high-
light the flow in the low-pressure regime by focusing on 𝑐 values between −0.35 and
−0.1. It is observed that in the optimized designs, lower pressure coefficients are present
near the tip, but the pressure coefficient has been increased in the mid-session and nearer
the hub. The overall reduction of maximum negative pressure coefficients in the mid and
hub regions offsets the effects of increased tip leakage flow in the optimized designs.
(a) (b) (c)

Next, the pressure coefficient distributions on the blade surfaces are investigated.
Scalar scenes of the pressure coefficients on the baseline, Design A, and Design B blades
are shown in Figure 13. A large portion of the suction-side surface in the baseline design
features a significantly negative 𝑐 values between −0.35 and −0.23 starting around the
three quarters chord. In the optimized designs, the onset of the negative pressure coeffi-
cient zone moves upstream, nearer the half chord, yet the magnitude of this zone is much
less, with most areas featuring 𝑐 values between −0.12 and 0. To achieve a more quanti-
tative comparison, the pressure coefficient curves along the streamwise direction on the
hub, mid, and tip sections of the three blades are plotted and compared, with the plots of
the blade profiles, shown in Figure 14. The earlier onsets of the lower pressure coefficient
region on the suction-side surfaces in the two optimized designs are demonstrated by the
plots. The earlier onset allows for larger enclosed areas of negative 𝑐 curves for the op-
(a) (b) (c)
timized design, contributing to the power improvements.
Figure 13. Pressure coefficient scenes on the surfaces of the blades: (a) Baseline design; (b) Design
Figure Pressure
13. (c)
A; and coefficient scenes on the surfaces of the blades: (a) Baseline design; (b) Design A;
Design B.
and (c) Design B.
To further investigate the differences in solutions, a volume rendering scene is cre-
ated by focusing on visualizing changes in Mach number within the range between 0.8
and 1.2, shown in Figure 15. The goal of this plot is to reveal where sudden changes in
Mach number may occur in the transonic flow regime. A consistent shock structure near
the trailing edge covering the full span length of the blade is observed in the baseline
design, which is consistent with the high pressure spot in the same area highlighted by
the dotted circle in Figure 11a. This shock feature is missing in the two optimized designs
(Designs A and B), due to the weaker gradients in those designs.
(a) (b) (c)

Figure 14. Blade profiles (first row) and pressure coefficients in the streamwise direction (second
Figure 14. Blade profiles (first row) and pressure coefficients in the streamwise direction (second
row): (a) hub sections comparison; (b) mid sections comparison; and (c) tip sections comparison.
row): (a) hub sections comparison; (b) mid sections comparison; and (c) tip sections comparison.
To further investigate the differences in solutions, a volume rendering scene is created

by focusing on visualizing changes in Mach number within the range between 0.8 and
1.2, shown in Figure 15. The goal of this plot is to reveal where sudden changes in Mach
number may occur in the transonic flow regime. A consistent shock structure near the
trailing edge covering the full span length of the blade is observed in the baseline design,
which is consistent with the high pressure spot in the same area highlighted by the dotted
circle in Figure 11a. This shock feature is missing in the two optimized designs (Designs A
and B), due to the weaker gradients in those designs.
(a) (b) (c)

Figure 15. Mach number rendering: (a) Baseline design; (b) Design A; and (c) Design B.
Figure 15. Mach number rendering: (a) Baseline design; (b) Design A; and (c) Design B.
A further comparison is made by plotting the streamlines in the tip leakage regions
A further comparison is made by plotting the streamlines in the tip leakage regions
and showing pressure coefficients on spanwise cross-sections in Figure 16. It appears the
and showing pressure coefficients on spanwise cross-sections in Figure 16. It appears the
two optimized designs feature slightly more chaotic Mach number distributions in the
two optimized designs feature slightly more chaotic Mach number distributions in the
local region near the trailing edge and tip leakage. The leakage flow’s influence on the
local region near the trailing edge and tip leakage. The leakage flow’s influence on the
overall performance is relatively small in comparison to the previously discussed suction-
overall performance is relatively small in comparison to the previously discussed suction-
side effects. It can be observed that the streamlines in the high-efficiency design (Design
side effects. It can
A) are slightly morebe curved,
observed yetthat
thethe streamlines
induced in the high-efficiency
low-pressure design (Design
spots by the secondary flow
A) are slightly
effects of the tipmore curved,
leakage flowsyet
arethe
weakinduced low-pressure
and about spots by the
the same magnitude. secondary
Overall, flow
the dif-
effects
ference caused by the tip leakage flow is small in these three designs. The performance the
of the tip leakage flows are weak and about the same magnitude. Overall,
difference
improvementscausedin by
thethe tip leakage
optimized flow are
designs is small
moreininfluenced
these threebydesigns. Theeffects
other flow performance
dis-
improvements in the optimized designs are more influenced by other flow
cussed earlier, such as the increased minimum pressure on the suction side near the effects discussed
trail-
earlier,
ing edgesuch as the11–13),
(Figures increased
andminimum pressure
the reduction of the on the structure
shock suction side near15).
(Figure the trailing edge
(Figures 11–13), and the reduction of the shock structure (Figure 15).
(a) (b) (c)

Figure 16. Tip leakage flow streamlines and pressure coefficients on spanwise cross-sections: (a)
Figure 16. Tip
Baseline leakage
design; flow streamlines
(b) Design and pressure
A; and (c) Design B. coefficients on spanwise cross-sections: (a) Base-
line design; (b) Design A; and (c) Design B.
4. Conclusions
4. Conclusions
The present study has demonstrated an optimization workflow combining the use of
The present study has demonstrated an optimization workflow combining the use of
neural networks and high-fidelity CFD. The neural network models were trained on over
neural networks and high-fidelity CFD. The neural network models were trained on over
three-thousand design data points from a previous publication. The practical implication
three-thousand design data points from a previous publication. The practical implication
of the overall strategy is that in engineering design analysis, existing data sets, which are
of generally
the overall strategy is that in engineering design analysis, existing data sets, which
available from the previous simulation, experimental, or reduced-order studies,
arecan
generally available
be leveraged to build from
neuralthenetwork
previous simulation,
models, which can experimental,
then be usedor in reduced-order
combination
studies, can be leveraged to build neural network
with high-fidelity CFD simulations to guide optimization processes. This models, which can then approach
be used in
combination with high-fidelity CFD simulations to guide optimization
achieves a reduction in the required number of high-fidelity CFD runs, and hence reduces processes. This
approach achieves a reduction in the required number of
the computational cost while maintaining accuracy. The integration of computationallyhigh-fidelity CFD runs, and
hence
inexpensive ANN models, which were evaluated 40,750 times, allows for a relatively of
reduces the computational cost while maintaining accuracy. The integration
computationally
small number (163) inexpensive ANN models,
of CFD evaluations which
in the were
present evaluated 40,750
optimization process,times, allows
resulting in afor
a relatively small
total run time of number
about 30 (163) h (with of6CFD
CFD evaluations in the present
evaluations running optimization
in parallel, process,
each consuming
resulting in a total
160 compute cores).runIt istime of about
estimated that30 h (with
if the nested6 CFD evaluations
optimization running
strategy based in onparallel,
ANN
each
wasconsuming
not used, a160 compute
total number cores).
of 733ItCFDis estimated that will
evaluations if thebenested
requiredoptimization
due to thestrategy
large
based on ANN
number was not
of design used, exposed
variables a total number
in 3D CFD of 733 CFD evaluations
evaluations, resulting will be required
in roughly 135 due
h of to
thethe
large number ofrun
optimization design
time if variables
the sameexposed
computeinresource
3D CFDsettings
evaluations, resulting in roughly
were used.
135 h ofThetheefficacy
optimization
of the run time if theissame
methodology compute resource
demonstrated on a turbine settings
bladewere used.
aerodynamic
The efficacy
problem. of the methodology
ANN models with 7 layers and is demonstrated
6 layers were builton ato turbine
representblade
twoaerodynamic
blade per-
formanceANN
problem. metrics, efficiency,
models withand power,and
7 layers respectively.
6 layers wereThe hyperparameters
built to represent of the
twoANN blade
models were optimized, and the models were used as surrogate
performance metrics, efficiency, and power, respectively. The hyperparameters of the models along with high-
ANN
fidelitywere
models CFDoptimized,
simulations in the
and a nested
models optimization
were usedprocedure
as surrogate to obtain
modelsoptimized
along with blade
high-
designs.
fidelity CFDPareto front designs
simulations in a representing improvedprocedure
nested optimization efficiency and poweroptimized
to obtain were foundblade by
the optimization
designs. Pareto front procedure.
designs A lean angle and
representing a tip scaling
improved factorand
efficiency were shown
power to be
were moreby
found
thefavored by the optimization
optimization procedure. A procedure
lean angle thanandother parameters
a tip in thewere
scaling factor context
shownof thetochosen
be more
blade and
favored analysis
by the methodology
optimization used in
procedure the other
than current study. Examining
parameters the fluid
in the context ofdynamics
the chosen
of the
blade andoptimized
analysisdesigns
methodology vs. the baseline
used in the design reveals
current thatExamining
study. the optimization (1) reduced
the fluid dynamics
the magnitude of the most negative pressure coefficients in the
of the optimized designs vs. the baseline design reveals that the optimization (1) reduced flow on the suction side
the magnitude of the most negative pressure coefficients in the flow on the suctionthe
near the trailing edge, and (2) altered the blade geometry that reduced the shock near side
trailing
near edge. Both
the trailing edge, of and
these(2) aspects
altered have
the effects on improved
blade geometry thatefficiency
reduced and power near
the shock in thethe
optimized
trailing edge.designs.
Both of these aspects have effects on improved efficiency and power in the
As a future extension of this study, the following may be considered. (1) Test the
optimized designs.
performance of other response surface methods and meta models against ANN and apply
As a future extension of this study, the following may be considered. (1) Test the
them in the nested optimization workflow. (2) The present study demonstrates a nested
performance of other response surface methods and meta models against ANN and apply
ANN-CFD optimization methodology, applied to an idealized turbine CFD problem as a
them in the nested optimization workflow. (2) The present study demonstrates a nested
proof-of-concept. In production-level gas turbine designs, more detailed constraints on
ANN-CFD optimization methodology, applied to an idealized turbine CFD problem as a
proof-of-concept. In production-level gas turbine designs, more detailed constraints on
the geometry or loading curve must be incorporated to ensure more realistic optimiza-
tion outcomes. Different power output conditions of the engine may also be explored
in optimization.
Author Contributions: Conceptualization, C.Z.; methodology, C.Z. and M.J.; software, C.Z. and M.J.;
validation, C.Z. and M.J.; formal analysis, C.Z.; investigation, C.Z. and M.J.; resources, C.Z. and M.J.;
data curation, C.Z.; writing—original draft preparation, C.Z.; writing—review and editing, C.Z. and
M.J.; visualization, C.Z.; supervision, M.J.; project administration, C.Z. and M.J.; funding acquisition,
M.J. All authors have read and agreed to the published version of the manuscript.
Funding: This research received no external funding.
Data Availability Statement: Not applicable.
Acknowledgments: This work is a part of the authors’ professional development initiatives spon-
sored by the Simcenter Customer Support organization of Siemens Digital Industries Software. The
authors wish to thank their colleagues at Siemens Digital Industries Software. Additional thanks go
to Yutao He at Jet Propulsion Laboratory, Caltech, for discussions on the research methodology.
Conflicts of Interest: The authors declare no conflict of interest.
Nomenclature
cp pressure coefficient
h fluid enthalpy J/kg
p pressure Pa
P performance function
Pow power W
r radius m
r Pearson coefficient
u velocity m/s
∆h enthalpy change J/kg
w blade profile parameter
η efficiency
ρ density kg/m3
σc centrifugal stress Pa
Ω rotating speed rad/s
Subscripts
h hub
i domain inlet
l lower surface
max maximum quantity
o domain outlet
s solid
t tip
u upper surface
References
1. Alexandrov, N.M.; Lewis, R.M.; Gumbert, G.R.; Green, L.L.; Newman, P.A. Approximation and Model Management in Aerody-
namic Optimization with Variable-Fidelity Models. J. Aircr. 2001, 38, 1093–1101. [CrossRef]
2. Ong, Y.S.; Nair, P.B.; Keane, A.J. Evolutionary Optimization of Computationally Expensive Problems via Surrogate Modeling.
AAIA J. 2003, 41, 687–696. [CrossRef]
3. Mazaheri, K.; Zeinalpour, M.; Bokaei, H.R. Turbine Blade Cooling Passages Optimization Using Reduced Conjugate Heat Transfer
Methodology. Appl. Therm. Eng. 2016, 203, 1228–1236. [CrossRef]
4. Jiang, Y.; Lin, H.; Yue, G.; Zheng, Q.; Xu, X. Aero-Thermal Optimization on Multi-Rows Film Cooling of a Realistic Marine High
Pressure Turbine Vane. Appl. Therm. Eng. 2016, 111, 537–549. [CrossRef]
5. Braining, E.; Weidtmann, K.; Buehler, P.; Lautenschlaeger, A.; Braun, R. Improvement of an Automated Design Process and
Investigation of the Optimization Approach and the Dependencies of the Individual Objectives. In Proceedings of the GPPS
Forum 18, Global Power and Propulsion Society, GPPS-2018-0027, Montreal, QC, Canada, 7–9 May 2018.
6. Braining, E.; Weidtmann, K.; Braun, R.; Rincon DL, H.; Bohn, D. Automated 2D Airfoil Optimization of Intentionally Choked
Blades. In Proceedings of the Global Power and Propulsion Society, GPPS-BJ-2019-0067, Beijing, China, 16–18 September 2019.
7. Zhang, C.; Hodges, J.; Janeway, M. Co-Optimization of Turbine Blade Aero and Thermal Designs Based on Computational Fluid
Dynamics (CFD) Models. In Proceedings of the 5th–6th Thermal and Fluids Engineering Conference, New Orleans, LA, USA,
26–28 May 2021.
8. Pierezan, J.; Maidl, G.; Yamao, E.M.; Coelho, L.S.; Mariani, V.C. Cultrual Coyote Optimization Algorithm Applied to a Heavy
Duty Gas Turbine Operation. Energy Convers. Manag. 2019, 199, 111932. [CrossRef]
9. Bettocchi, R.; Pinelli, M.; Spina, P.R.; Venturini, M. Set up of a Robust Neural Netwok for Gas Turbine Simulation. In Proceedings
of the ASME Turbo Expo 2004, Power for Land, Sea, and Air, Vienna, Austria, Jun 14–17 June 2004.
10. Li, D.; Qiu, L.; Tao, K.; Zhu, J. Artificial Intelligence Aided Design of Film Cooling Scheme on Turbine Guide Vane.
Propuls. Power Res. 2020, 9, 344–354. [CrossRef]
11. Belov, S.; Nikolaev, S.; Uzhinsky, I. Hybrid Data-Driven and Physics-Based Modeling for Gas Turbine Prescriptive Analytics.
Int. J. Turbomach. Propuls. Power 2020, 5, 29. [CrossRef]
12. Giorgi, M.G.D.; Quarta, M. Hybrid MultiGene Genetic Programing—Artificial Neural Networks Approach for Dynamic Perfor-
mance Prediction of an Aeroengine. Aerosp. Sci. Technol. 2020, 103, 105902. [CrossRef]
13. Zhang, M.; Hao, S.; Hou, A. Study on the Intelligent Modeling of the Blade Aerodynamic Force in Compressor Based on Machine
Learning. Mathematics 2021, 9, 476. [CrossRef]
14. Wang, Q.; Yang, L.; Rao, Y. Establishment of a Generalizable Model on a Small-Scale Dataset to Predict the Surface Pressure
Distribution of Gas Turbine Blades. Energy 2021, 214, 118878. [CrossRef]
15. Park, Y.; Choi, M.; Kim, K.; Li, X.; Jung, C.; Na, S.; Choi, G. Prediction of Operating Characteristics for Industrial Gas Turbine
Combustor Using an Optimized Artificial Neural Network. Energy 2020, 213, 118769. [CrossRef]
16. Sun, L.; Liu, T.; Xie, Y.; Zhang, D.; Xia, X. Real-Time Power Prediction Approach for Turbine Using Deep Learning Techniques.
Energy 2021, 233, 121130. [CrossRef]
17. Akolekar, H.; Zhao, Y.; Sandberg, R.D.; Pacciani, R. Integration of Machine Learning and Computational Fluid Dynamics to
Develop Turbulence Models for Improved Turbine Wake Mixing Prediction. In Proceedings of the ASME Turbo Expo 2020:
Turbomachinery Technical Conference and Exposition, London, UK, 22–26 June 2020.
18. Brunton, S.L.; Kutz, J.N.; Manohar, K.; Aravkin, A.Y.; Morgansen, K.; Klemisch, J.; Goebel, N.; Buttrick, J.; Poskin, J.;
Blom-Schieber, A.W.; et al. Data-Driven Aerospace Engineering: Reframing the Industry with Machine Learning. AIAA J.
2021, 59, 2820–2847. [CrossRef]
19. Wang, H.; Xi, G. Optimization of High Speed Centrifugal Compressor for a Micro Gas Turbine Based on CFD and FEM Analysis.
In Proceedings of the ASME Turbo Expo 2010: Power for Land, Sea, and Air, Glasgow, UK, 14–18 June 2010.
20. Maral, H.; Alpman, E.; Kavurmacioglu, L.; Camici, C. A Genetic Algorithm Based Aerothermal Optimization of Tip Carving for
an Axial Turbine Blade. Int. J. Heat Mass Transf. 2019, 143, 11849. [CrossRef]
21. Yan, X.; Zhu, J.; Kuang, M.; Wang, X. Aerodynamic Shape Optimization Using a Novel Optimzer Based on Machine Learning
Techniques. Aerosp. Sci. Technol. 2019, 86, 826–835. [CrossRef]
22. Li, J.; Zhang, M.; Martins, J.R.R.A.; Shu, C. Efficient Aerodynamic Shape Optimization with Deep-Learning-Based Geometric
Filtering. AIAA J. 2020, 58, 4243–4259. [CrossRef]
23. Wang, Y.; Liu, T.; Zhang, D.; Xie, Y. Dual-Convolutional Neural Network Based Aerodynamic Prediction and Multi-Objective
Optimization of a Compact Turbine Rotor. Aerosp. Sci. Technol. 2021, 116, 106869. [CrossRef]
24. Elsayed, K.; Lacor, C. Modeling, Modeling, Analysis and Optimization of Aircyclones Using Artificial Neural Network Response
Surface Methodology and CFD Simulation Approaches. Powder Technol. 2011, 212, 115–133. [CrossRef]
25. Chamoli, S. ANN and RSM Approach for Modeling and Optimization of Designing Parameters for a V down Perforated Baffle
Roughened Rectangular channel. Alex. Eng. J. 2015, 54, 429–446. [CrossRef]
26. Park, D.; Cha, J.; Kim, M.; Go, G.S. Multi-objective Optimization and Comparison of Surrogate Models for Separation Performance
of Cyclone Separator Based on CFD, RSM, GMDH-Neural Network, Back Propagation-ANN and Genetic Algorithm. Eng. Appl.
Comput. Fluid Mech. 2020, 14, 180–201. [CrossRef]
27. Halila, E.E.; Lenahan, D.T.; Thomas, T.T. Energy Efficient Engine High Pressure Turbine Test Hardware Detailed Design
Report; NASA-CR-167955; NASA Lewis Research Center: 1982. Available online: https://ntrs.nasa.gov/citations/19850002687
(accessed on 25 April 2022).
28. Timko, L.P. Energy Efficient Engine High Pressure Component Test Performance Report; NASA-CR-168289; NASA Lewis
Research Center: 1990. Available online: https://ntrs.nasa.gov/citations/19900019237 (accessed on 25 April 2022).
29. Chen, S.-C.S. Axial-Flow Turbine Rotor Discharge-Flow Overexpansion and Limit-Loading Condition, Part 1: Computational
Fluid Dynamics (CFD) Investigation; NASA/TM-2017-219506; Glenn Research Center: 2017. Available online: https://ntrs.nasa.
gov/citations/20170007293 (accessed on 25 April 2022).
30. Kulfan, B. Universal Parametric Geometry Representation Method. J. Aircr. 2008, 45, 142–158. [CrossRef]
31. Abadi, M.; Barham, P.; Chen, J.; Chen, Z.; Davis, A.; Dean, J.; Devin, M.; Ghemawat, S.; Irving, G.; Isard, M.; et al. Tensorflow:
A system for large-scale machine learning. In Proceedings of the 12th USENIX Conference on Operating Systems Design and
Implementation, Savannah, GA, USA, 2–4 November 2016; USENIX Association: Berkeley, CA, USA, 2016; pp. 265–283.
32. Red Cedar Technology. SHERPA—An Efficient and Robust Optimization/Search Algorithm; WP-1023, Rev. 05.08; Red Cedar
Technology: East Lansing, MI, USA, 2021.
33. Zhang, C. A Python Program—Machine Learning Regression for Gas Turbine Blade Efficiency and Power, Version 1.0.0;
Computer software; 2022. Available online: https://github.com/Cool-Chaoz/Machine-Learning-Regression-for-Gas-Turbine-
Blade-Efficiency-and-Power (accessed on 25 April 2022).
34. Mohaghegh, K.; Sadeghi, M.H.; Abdullah, A.; Boutorabi, R. Improvement of Reverse-Engineered Turbine Blades Using Construc-
tion Geometry. Int. J. Manuf. Technol. 2010, 49, 675–687. [CrossRef]
35. Menter, F.R. Two-equation eddy-viscosity turbulence modeling for engineering applications. AIAA J. 1994, 32, 1598–1605.
[CrossRef]

Ijtpp 07 00020

Uploaded by

Copyright:

Available Formats

Ijtpp 07 00020

Uploaded by

Document Information

Original Title

Copyright

Available Formats

Share this document

Share or Embed Document

Sharing Options

Did you find this document useful?

Is this content inappropriate?

Copyright:

Available Formats

Ijtpp 07 00020

Uploaded by

Copyright:

Available Formats

International Journal of

Siemens Digital Industries Software, Cypress, CA 90630, USA; matt.janeway@siemens.com

Int. J. Turbomach. Propuls. Power 2022, 7, 20. https://doi.org/10.3390/ijtpp7030020 https://www.mdpi.com/journal/ijtpp

Table 1. Hyperparameters of the initial and optimized ANN models.

Initial ANN Initial ANN Optimized ANN Optimized ANN

Int. J. Turbomach. Propuls. Power 2022, 7, 20 6 of 19

Int. J. Turbomach. Propuls. Power 2022, 7, x FOR PEER REVIEW 8 of 23

2.3. Optimization Strategy

Int. J. Turbomach. Propuls. Power 2022, 7, 20 9 of 19

Figure 6. Optimization Figure

In the global optimization loop (Optimizer 1 as shown in Figure 6), a multi-objective

performance characters of the E3 engine. The global optimization solver, Optimizer 1,

i = Nobj Obji j= N 104 × ConsViol 2

Int. J. Turbomach. Propuls. Power 2022, 7, x FOR PEER REVIEW 12 of 23

Int. J. Turbomach. Propuls. Power 2022, 7, 20 12 of 19

3.2. Fluid Dynamics Analysis

Int. J. Turbomach. Propuls. Power 2022, 7, x FOR PEER REVIEW 16 of 23

(a) (b) (c)

Int. J. Turbomach. Propuls. Power 2022, 7, x FOR PEER REVIEW 17 of 23

(a) (b) (c)

Int. J. Turbomach. Propuls. Power 2022, 7, 20 15 of 19

(a) (b) (c)

To further investigate the differences in solutions, a volume rendering scene is created

(a) (b) (c)

Int. J. Turbomach. Propuls. Power 2022, 7, 20 16 of 19

(a) (b) (c)

You might also like

Pfad - The Proxy pFad of © 2024 Garber Painting. All rights reserved.