Professional Documents
Culture Documents
Aerospace
Aerospace
Aerospace
a r t i c l e i n f o a b s t r a c t
Article history: The design for the conflicting requirements of low drag and low Radar Cross Section (RCS) causes
Received 7 April 2021 difficulty in achieving an aerodynamically superior stealth aircraft. To cater the issue, a multidisciplinary
Received in revised form 14 November 2022 design exploration and optimization framework is proposed. A shared parameterized aircraft geometry
Accepted 3 January 2023
is used for high fidelity aero-stealth analysis using Computational Fluid Dynamics (CFD) and Shooting
Available online 9 January 2023
Communicated by Nikolaos Gatsonis
and Bouncing Rays (SBR) techniques. A surrogate model using a machine learning approach involving
Gaussian Process (GP) modelling is generated for efficient and rapid design space exploration. A gradient-
Keywords: free metaheuristic optimization scheme using multi-objective Genetic Algorithm (GA) is employed to
CFD minimize drag and RCS. The results show that the proposed framework provides a reliable environment
RCS for the multidisciplinary design exploration of an aerial vehicle.
Shooting and bouncing rays © 2023 Elsevier Masson SAS. All rights reserved.
Gaussian process
Multidisciplinary optimization
Multi-objective genetic algorithm
https://doi.org/10.1016/j.ast.2023.108114
1270-9638/© 2023 Elsevier Masson SAS. All rights reserved.
Z.U.D. Taj, A. Bilal, M. Awais et al. Aerospace Science and Technology 133 (2023) 108114
RCS response with five geometric parameters as predictors. Hence, When the incident wave strikes the target, it behaves like an
GP based surrogate modelling using space filling LHS plan can be antenna and extracts power by using its capture area (σ ). The ex-
employed for efficient and rapid design space exploration. tracted power ( P ) is given by:
Much research has been employed to find the optimized geom-
etry in a modelling and simulation environment using numerous σ Y 0 | E i |2
P = σ Wi = (2)
optimization schemes [16,17]. These schemes are generally divided 2
into two main categories – gradient and non-gradient based opti- The scattered wave power density (W s ) at an observed distance
mization algorithms [18]. Gradient-based algorithms have the lim- (R) is given by Equation (3).
itation of getting stuck in local optimum solutions but are fast and
efficient. However, non-gradient based algorithms, like metaheuris- P σ Y 0 | E i |2
Ws = = (3)
tic optimization schemes, are used to explore the global optimum 4π R2 8π R 2
solution of a large design space with discrete variables but re- However, the scattered wave power density can also be ex-
quire excessive function calls [19]. Many metaheuristic schemes
pressed as:
have been applied by the researchers like Genetic Algorithm (GA),
simulated annealing optimization and particle swarm algorithms Es Hs Y 0 | E s |2
[20,21]. For example, Jun-xue Leng, et al. [22] used GA to improve Ws = = (4)
2 2
the aerodynamic performance of reusable launch vehicles based
Comparing Equations (3)–(4) yields the capture area or radar
on reverse design approach. Similarly, Tian-tian Zhang, et al. [23]
cross section of the target given by Equation (5).
optimized the aerodynamics of hypersonic-gliding vehicle config-
urations during conceptual design phase using multi-objective GA | E s |2
and presented the solutions using Pareto optimal front. In addi- σ = lim 4π R 2 (5)
R →∞ | E i |2
tion, Lyu, et al. [24] in a wing shape optimization study found
that the number of iterations for gradient free schemes increase There are numerous Computational Electromagnetic (CEM) al-
in a quadratic and cubic manner with increasing dimensionality as gorithms to simulate RCS of target objects, which involve solving
compared to gradient based schemes, which show linear behavior. differential or integral forms of Maxwell’s equations [27,28]. Phys-
However, Yang Shen, et al. [25] used GA optimization scheme cou- ical Optics (PO), Method of Moments (MOM) and Shooting and
pled with Kriging surrogate model for rapid function evaluations Bouncing Rays (SBR) are the commonly used CEM algorithms. PO
of increased number of iterations. Therefore, GA optimization cou- method is fast but has very low accuracy due to the inability to
pled with GP surrogate model can be employed in our proposed capture edge diffraction effects and multiple bounces reflection
environment. effects. Whereas MOM method is a high-fidelity method but re-
The present work is focused on proposing a multi-disciplinary quires high computational time. To overcome these limitations,
design exploration and optimization framework. The environ- SBR method is used to compute RCS of electrically large struc-
ment is used for optimizing the fighter aircrafts for high aero- tures yielding fast and reasonably accurate results [29–31]. The
dynamic performance and low radar signature using high fidelity technique efficiently caters edge diffraction, creeping waves and
aerodynamic-stealth analysis and a single parameterized geometry. multiple bounce effects. Ahmad, et al. [32] benchmarked the ac-
In order reduce the calculations cost, GP based surrogate model curacy of SBR technique by comparison with MOM and measured
is generated on which gradient free GA optimization algorithm is results.
employed. In order to simulate the incident electromagnetic (EM) wave,
The research is presented in following sections. Section 2 in- SBR algorithm uses a sequence of optical ray tubes. The ray tube
cludes a brief background of the employed high fidelity RCS com- obeys geometric optics principles during propagation. For a homo-
putation, aerodynamic evaluation, GP regression modelling and geneous isotropic medium, in geometric optics, Fermat’s principle
GA optimization techniques. The validation of electromagnetic and states a ray tube propagates on the straight line. Thus, its path of
aerodynamic solvers is done in Section 3. Section 4 outlines the propagation can be written as [33]:
proposed multidisciplinary design exploration and optimization
environment. Section 5 includes the results and discussion of the r (x, y , z) = r0 (x0 , y 0 , z0 ) + s sx , s y , s z t (6)
obtained optimized aircraft geometries using GA based optimiza-
Here r0 represents the starting point, s represents the direction
tion scheme. Section 6 concludes the research.
and t depicts time for point r0 on the path followed by ray. If the
direction of incident ray is (θi , φi ), s can be written as follows:
2. Background
sx = − sin θi cos φi , s y = − sin θi sin φi , s z = − cos θi (7)
2.1. Stealth analysis
The ray will refract or reflect at the target’s surface when it in-
RCS of an aircraft is the measure of its stealth. Lower the radar tersects the target. The direction of refraction and reflection can be
signature means better the stealth, smaller the detection range and found using Snell’s law. If n̂ represents the normal at the point of
lesser the response time. RCS of a target is the projected area of an intersection, and î denotes the incident direction, then the reflec-
object, which would scatter the same power in the same direction tion direction ŝ is shown by:
as a metallic sphere [26]. It is measured in m2 or dBsm, and shows
ŝ = î − 2(î · n̂)n̂ (8)
the comparison between the incident wave power density at the
target and the scattered wave power density received at the radar. The above steps are repeated for a ray until it departs the tar-
The incident wave power density (W i ) is given by Equation (1). get and then shoot towards infinity to acquire a ray path. Then to
simulate multiple reflections of a ray tube on the surface of a tar-
Ei Hi Y 0 | E i |2
Wi = = (1) get, geometrical optics is utilized. During propagation, the phase
2 2 and field strength also change in the tube. Upon i + 1th intersec-
Where, E i = incident electric field strength, H i = incident mag- tion with target (r i +1 ), the field strength can be calculated from
netic field strength and Y 0 = admittance of free space. the last intersection i.e., the i th intersection (ri ):
2
Z.U.D. Taj, A. Bilal, M. Awais et al. Aerospace Science and Technology 133 (2023) 108114
ri +1 = (DF )i · G
E · ri · e − j (phase) (9) In order to model the turbulence, numerous turbulence mod-
i
els are used including Direct Numerical Simulation (DNS) model,
Where phase = k r i +1 − ri represents the change in phase. Large Eddy Simulation (LES) model and Reynolds-Averaged Navier-
G denotes the reflection coefficient, (DF )i represents the diver- Stokes (RANS) model. In this research, RANS model is used due
i to its inexpensive nature of obtaining good quality results with
gence factor at ri , and Ei (ri ) indicates the electric field incident at
sufficient accuracy. It solves the entire Navier-Stokes equations by
ri .
performing time averaging to extract mean flow properties from
The scattering of the object in far field is given by accretion of
the instantaneous ones. Among numerous types of RANS models,
the individual scattered fields of the tube, which result from re-
Shear Stress Transport (SST) k − ω turbulence model is used in
flections off the surface of target. Finally, after multiple reflections
this work as shown by Equation (16)–(17) [35]. It uses k − ε in the
the ray tube shoots at infinity and the method utilizes PO princi-
freestream and k − ω in the near wall region to capture the viscous
ples on the ray tube to accomplish the far field integration. Using
effects.
Stratton-Chu integral formula, the generic equation for scattered
field of target surface is written as [33]: ∂ ∂ ∂ ∂ω
(ρω) + (ρω U i ) = (μ + 0.5μt ) + P ω − Dω
. ∂t ∂ xi ∂ xi ∂ xi
jk e −jkr
Es r , ω = r̂ × r , ω
Z 0 r̂ × n̂ × H 0.856ρ ∂ k ∂ ω
4π r + 2(1 − F ) (16)
s ω ∂ xi ∂ xi
∂ ∂ ∂ ∂k
+ r , ω
E × n̂ e jkr̂ ·r (10) (ρ k) + (ρ kU i ) = (μ + μt ) + P k − Dk (17)
∂t ∂ xi ∂ xi ∂ xi
Here r̂ and r denote unit and position vector of point of obser-
Where, P w = rate of production of ω , D w = rate of dissipation
vation, n̂ represents unit normal vector of bin on target surface, r
of ω , P k = and rate of production of k, D k = rate of dissipation of
shows position vector
of any surface,
Z0 represents the free space k, F = blending function and μ = molecular viscosity coefficient.
r , ω
impedance, and H r , ω
and E represent total magnetic
and electric scattered fields respectively. 2.3. GP regression
2.2. Aerodynamic analysis GP regression models use supervised machine learning tech-
nique to estimate the response based on the training data. How-
An aircraft design is an iterative process. The flight-testing and ever, the generated surrogate trained model acts as a black box
wind tunnel experiments are the industry standards for exact eval- and does not take a functional form hence it is non-parametric in
uation of aerodynamic performance. However, the associated high nature. In addition, the surrogate model is not deterministic but
cost and time makes it unfeasible for their integration in this iter- stochastic in nature i.e., it returns a value with a probability dis-
ative design cycle. As a result, CFD has gained confidence for the tribution instead of a fixed value every time. For a given data set
incorporation in aircraft design processes. In CFD, Navier-Stokes D of xn variables and f n responses, as shown in Equation (18), a
equations are solved which are a coupled system of nonlinear par- Gaussian process is defined as a stochastic process in which each
tial differential equations that represent the unsteady, compress- finite set of random variables in D is normally distributed or has a
ible, three-dimensional viscous flows. Equations (11)–(15) repre- multivariate Gaussian distribution [36], as shown by Equation (19).
sent complete set of Navier-Stokes equations including continuity,
momentum, and energy equations [34]. D = {(x1 , f 1 ) , (x2 , f 2 ) , . . . , (xn , f n )} (18)
Dρ f ∼ N (μ, K ) (19)
+ ρ∇ · V = 0 (11)
Dt
Du ∂P ∂ τxx ∂ τyx ∂ τzx Where, f is the prior and has a Gaussian distribution with
ρ + =− + + + ρ fx (12) mean μ and covariance K . For two variables, it is expressed as:
Dt ∂x ∂x ∂y ∂z
Dv ∂P ∂ τxy ∂ τyy ∂ τzy
ρ =− + + + + ρ fy (13) f1 μ1 K 11 K 12
Dt ∂y ∂x ∂y ∂z ∼N , (20)
f2 μ2 K 21 K 22
Dw ∂P ∂ τxz ∂ τyz ∂ τzz
ρ =− + + + + ρ fz (14) The covariance defines the resemblance between the values of
Dt ∂z ∂x ∂y ∂z
the data set as shown in Equation (21).
D V2 ∂ ∂T ∂ ∂T ∂ ∂T ⎡ ⎤
e+ = ρ q̇ + k k + + k k (x1 , x1 ) k (x1 , x2 ) k (x1 , xn )
Dt 2 ∂x ∂x ∂y ∂y ∂z ∂z
⎢ k (x2 , x1 ) ···
⎢ k (x2 , x2 ) k (x2 , xn ) ⎥
⎥
∂ (u τxx ) ∂ u τyx ∂ (u τzx ) K =⎢ .. .. .. ⎥ (21)
− ∇ · (P V ) + + + ⎣ . . . ⎦
∂x ∂y ∂z
k (xn , x1 ) k (xn , x2 ) ··· k (xn , xn )
∂ v τxy ∂ v τyy ∂ v τzy ∂ ( w τxz )
+ + + + The covariance is modelled using a kernel function that is cho-
∂x ∂y ∂z ∂x
sen based on the periodic or aperiodic nature of the model re-
∂ w τyz ∂ ( w τzz ) sponse. RCS response evaluated for varying geometric parameters
+ + +ρ f ·V (15)
∂y ∂z is found to be aperiodic [15] and thus Matérn 5/2 covariance func-
Where, ρ = density of the fluid, V = velocity, P = pressure, tion, given by Equation (22), is used to model the covariance of the
τi j = normal stress for i = j and shear stress for i = j, f = body data.
force, e = internal energy, k = thermal conductivity, q̇ = volumet- √ √ 5
ric heat addition per unit mass and T = temperature. k x, x∗ = σ f2 e − 5r
(1 + 5r + r2) (22)
3
3
Z.U.D. Taj, A. Bilal, M. Awais et al. Aerospace Science and Technology 133 (2023) 108114
Here σ f2 is the maximum allowable covariance and r is the dis- 2.4. GA metaheuristic optimization scheme
tance given by:
GA metaheuristic optimization scheme is based on the process
D of natural selection and mimics natural genetics. It uses a popula-
(xk − x∗ )2
r= k
(23) tion consisting of a set of design points and thus are less prone to
k =1
τk2 be trapped in local solutions. The variables are considered as chro-
mosomes or individuals that are represented by binary string of
Here τk is the kernel length scale for each predictor, which digits where each digit is considered as a gene or bit. Whereas a
is used to gauge the distance between the actual and predicted generation or population is the set of these chromosomes whose
values. In order to predict a function value f ∗ for a data point fitness values are evaluated at the end of each iteration. Fig. 1
x∗ , we assume that this new data has also a Gaussian distribu- shows the flow chart explaining the working and steps of the al-
tion with a mean μ∗ and covariance K ∗ as shown by Equation gorithm.
The algorithm is initiated by generation of μ individuals pop-
(24).
ulation and their fitness values are obtained using the objective
functions. The number of individuals is then selected for reproduc-
f ∗ ∼ N (μ∗ , K ∗ ) (24) tion based on the tournaments to be carried out between them.
The process works on the principle of random selection of μ indi-
In the matrix form, it is expressed as:
viduals, known as tournament size, and competing them for high-
⎡ ⎤ ⎛⎡ ⎤ ⎡ ⎤⎞ est fitness function values.
f1 μ1 K 11 K 12 K 1∗ Next step is the reproduction which is further divided in two
⎣ f 2 ⎦ ∼ N ⎝⎣ μ2 ⎦ , ⎣ K 21 K 22 K 2∗ ⎦⎠ (25) steps – crossover and mutation. In crossover, the genetic shuffling
f∗ μ∗ K 1∗ T K 2∗ T K ∗∗ of μ parent individuals into λ offspring occurs, thus allowing GA
to search for global optimum solution. Every bit of a parent chro-
Where, K ∗ is the similarity between new data point with the mosome is assigned a probability to decide the swap with the cor-
remaining points, given by Equation (26), and K ∗∗ is the similarity responding bit of the other parent chromosome. In mutation, the
of new data point with itself or simply 1. genetic diversity of the process is preserved by randomly swap-
ping certain bits in λ offspring. The fitness of λ offspring is then
K ∗ = k (x∗ , x1 ) k (x∗ , x2 ) . . . k (x∗ , xn ) (26) evaluated and compared with μ parents. The highest fit μ individ-
uals are finally replaced to pass on to next generation and entire
To check the accuracy of the surrogate models, Root Mean process is repeated in next iterations until the required global op-
Square Relative Error (RMSRE) is calculated using Equation (27) timum solution is not achieved.
[37].
3. Validation
1 f p − f a 2
RMSRE =
f
(27) To validate the electromagnetic solver, SBR results are bench-
N a marked using MOM and experimental measurement of an aircraft
N
geometry [32]. The frequency is taken as 1 GHz and the polariza-
Where, f a = actual function response, f p = predicted function tion mode is vertical polarization. The comparison of SBR, MOM
response and N = number of runs. and experimental RCS shows a good agreement as shown in Fig. 2.
4
Z.U.D. Taj, A. Bilal, M. Awais et al. Aerospace Science and Technology 133 (2023) 108114
Fig. 2. Comparison of RCS between simulated and experimental data [32] at 1 GHz frequency and VV polarization.
Fig. 3. Boxplots of relative error between simulated and experimental data (a) Electromagnetic solver (b) Aerodynamic solver.
Fig. 4. Comparison between CFD and experimental data [39] of generic UCAV configuration SACCON (a) Drag coefficient (b) Lift coefficient.
The boxplots of relative error between them are shown in Fig. 3(a). measured in anechoic chamber [30]. However, the same method-
Furthermore, the RMSRE between SBR and MOM is 1.47 and SBR ology can be extended to design and optimize a new and complex
and experimental result is 2.16, thus confirming that SBR technique fighter aircraft geometry with more stealth features like twin ver-
can be efficiently employed for RCS simulations. tical fins, variable wing sweep, etc. Table 1 shows geometric vari-
For validating the aerodynamics solver, simulation results are ables of the baseline aircraft geometry.
benchmarked using experimental data of generic UCAV configura- The parameterized geometry is then shared for the coupled
tion SACCON [39]. The flow conditions include Mach number of stealth and aerodynamic analysis to be carried out using SBR
0.149 and angle of attack from 0◦ to 18◦ at Reynold’s number of technique and CFD respectively. The analysis is performed in
1.6 million with reference to mean aerodynamic chord. The com- three steps: pre-processing, simulation, and post-processing. In
parison of CFD and experimental lift and drag coefficients shows pre-processing, geometry cleaning along with the formulation of
an excellent correlation as shown in Fig. 4. The boxplots of rela- physical problem, computational domain of interest and automatic
tive error between them are shown in Fig. 3(b). Furthermore, the scripted grid generation is done. In simulation, meshed geome-
RMSRE between experimental and CFD data of drag coefficient is try is used for the analysis by defining the boundary conditions
0.27 and lift coefficient is 0.21, thus confirming that CFD can be for the problem. The simulation parameters are listed in Table 2.
efficiently employed for aerodynamic simulations. The stealth analysis is performed only on one frequency and polar-
ization to keep the number of simulation variables tractable. The
4. Methodology frequency of 1 GHz is chosen because it is the typical operating
frequency of ground-based radars [15].
An automated multi-disciplinary design exploration and opti- The mesh settings for CFD analysis are shown in Table 3. Fig. 7
mization framework is proposed in this research, which is shown shows the grid of the employed aircraft model. The drag and lift
in Fig. 5. Firstly, a geometry of a fighter aircraft including eight were taken as converged to a tolerance of 1e-04 for 10 successive
parametric variables is created as shown in Fig. 6. The geometry iterations. In post-processing, the results in the form of RCS, drag
selection is driven by the availability of aircraft model and data and lift coefficients are extracted and visualized.
5
Z.U.D. Taj, A. Bilal, M. Awais et al. Aerospace Science and Technology 133 (2023) 108114
Table 1
Baseline geometric parameters.
Table 2
Simulation parameters.
6
Z.U.D. Taj, A. Bilal, M. Awais et al. Aerospace Science and Technology 133 (2023) 108114
7
Z.U.D. Taj, A. Bilal, M. Awais et al. Aerospace Science and Technology 133 (2023) 108114
Table 4
Grid independence study.
Mesh level Mesh size CD Percentage error C D from M 0 CL Percentage error C L from M 0
M0 7,153,002 0.050838 – 0.35685 –
M1 4,502,804 0.050543 0.6% 0.35470 0.6%
M2 3,619,205 0.052635 3.0% 0.37906 6.2%
Fig. 10. Pareto optimal front. Fig. 12. Aerodynamics optimized geometry.
8
Z.U.D. Taj, A. Bilal, M. Awais et al. Aerospace Science and Technology 133 (2023) 108114
Table 5
Geometric parameters of baseline and optimized configurations.
Table 6
Comparison of optimization results between baseline and optimized configurations.
9
Z.U.D. Taj, A. Bilal, M. Awais et al. Aerospace Science and Technology 133 (2023) 108114
[21] Yu Hu, et al., Optimization configuration of gas path sensors using a hybrid [31] Wanli Xu, Lixin Guo, Shuirong Chai, Improved SBR method for backward scat-
method based on tabu search artificial bee colony and improved genetic algo- tering of ship target under shallow sea background, in: 2018 12th International
rithm in turbofan engine, Aerosp. Sci. Technol. 112 (2021) 106642. Symposium on Antennas, Propagation and EM Theory (ISAPE), IEEE, 2018.
[22] Jun-xue Leng, et al., Parameterized modeling and optimization of reusable [32] Ahmad Bilal, et al., Comparison of SBR and MLFMM techniques for the com-
launch vehicles based on reverse design approach, Acta Astronaut. 178 (2021) putation of RCS of a fighter aircraft, IET Radar Sonar Navig. 13 (10) (2019)
36–50. 1805–1810.
[23] Tian-tian Zhang, et al., Parameterization and optimization of hypersonic-gliding [33] Junyang Zhou, Yiping Han, Analyzing the electromagnetic scattering character-
vehicle configurations during conceptual design, Aerosp. Sci. Technol. 58 (2016) istics for plasma targets based on shooting and bouncing ray method, AIP Adv.
225–234. 9 (6) (2019) 065106.
[24] Yin Yu, et al., On the influence of optimization algorithm and initial de- [34] John David Anderson, et al., Computational Fluid Dynamics: An Introduction,
sign on wing aerodynamic shape optimization, Aerosp. Sci. Technol. 75 (2018) Springer Science & Business Media, 2013.
183–199. [35] López Deibi, Diego Domínguez, Jesús Gonzalo, Impact of turbulence modelling
[25] Yang Shen, et al., Constraint-based parameterization using FFD and multi- on external supersonic flow field simulations in rocket aerodynamics, Int. J.
objective design optimization of a hypersonic vehicle, Aerosp. Sci. Technol. 100 Comput. Fluid Dyn. 27.8–10 (2013) 332–341.
(2020) 105788. [36] Eric Schulz, Maarten Speekenbrink, Andreas Krause, A tutorial on Gaussian
[26] Eugene F. Knott, John F. Schaeffer, Michael T. Tulley, Radar Cross Section, process regression: modelling, exploring, and exploiting functions, J. Math. Psy-
SciTech Publishing, 2004. chol. 85 (2018) 1–16.
[27] Z. Lin, et al., Performance of a massively parallel higher-order method of mo- [37] Mustafa Göçken, et al., Integrating metaheuristics and artificial neural networks
ments code using thousands of CPUs, in: 2018 International Conference on for improved stock price prediction, Expert Syst. Appl. 44 (2016) 320–331.
Electromagnetics in Advanced Applications (ICEAA), IEEE, 2018. [38] Johann Dréo, et al., Metaheuristics for Hard Optimization: Methods and Case
[28] Mohammad H. Ahmad, Dayalan Prajith Kasilingam, Spectral domain fast multi- Studies, Springer Science & Business Media, 2006.
pole method for solving integral equations of electromagnetic wave scattering, [39] Thomas Loeser, Dan Vicroy, Andreas Schuette, SACCON static wind tunnel tests
Prog. Electromagn. Res. M 80 (2019) 121–131. at DNW-NWB and 14 x22 NASA LaRC, in: 28th AIAA Applied Aerodynamics
[29] Guocai Zan, et al., Scattering characteristics of the multi-corner reflector based Conference, 2010.
on SBR method, in: 2018 12th International Symposium on Antennas, Propa-
gation and EM Theory (ISAPE), IEEE, 2018.
[30] Ahmad Bilal, et al., ISAR imaging using FFT with polar reformatting of mea-
sured RCS, in: 2020 3rd International Conference on Computing, Mathematics
and Engineering Technologies (iCoMET), IEEE, 2020.
10