Documenti di Didattica
Documenti di Professioni
Documenti di Cultura
Abstract
A historical perspective of computational fluid dynamics (CFD) in aerospace in the last 30 years is firstly given. It is
shown that there still remain a number of problems that are geometrically simple but difficult to simulate even after many
simulations were conducted over complex body configurations. The fact indicates that CFD research is now in the ‘‘specific
phase’’ and requires some innovation.
The innovation includes ‘‘evolutionary effort’’ and ‘‘revolutionary effort’’. As an example of evolutionary effort, large
eddy simulations/ Reynolds-averaged Navier–Stokes simulations (LES/RANS) hybrid method and its application
examples are presented. A shift from RANS to LES/RANS hybrid method occurs not because of the advancement of
computers but because of our recognition that separated flows are inherently unsteady and successful simulations require
LES-like computations.
Comment is given that there may be other types of research necessary to make CFD a real useful tool for a design in
addition to simply showing CFD capability for complex body configurations. As one of the examples, construction of a
CFD database is presented. Another issue is to make CFD infrastructures so that people outside CFD community may use
CFD as a tool to formulate or refine their ideas.
To find out revolutionary effort, the message given by Prof. Dean Chapman in 1977 is referred. Observation of current
CFD research reveals that evaluation methods of ‘‘scale effect’’ that were believed to be the most important benefit of CFD
have not yet been established. Such establishment is the key for the revolution of CFD and researchers need to focus their
effort on the development of technologies to evaluate scale effect. Only with such new CFD technologies can ‘‘conceptual
design with CFD’’ become feasible.
r 2005 Elsevier Ltd. All rights reserved.
Contents
1. Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 456
2. Historical perspective of CFD in aerospace . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 457
3. CFD in the specific phase—current status . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 459
3.1. CFD vs. EFD . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 459
3.2. Simple problems still remain . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 459
0376-0421/$ - see front matter r 2005 Elsevier Ltd. All rights reserved.
doi:10.1016/j.paerosci.2005.09.001
ARTICLE IN PRESS
456 K. Fujii / Progress in Aerospace Sciences 41 (2005) 455–470
instance, it is questionable if we can evaluate scale that the messages and the examples in this manu-
effect (Reynolds number effect) so long as we use script only reflect the author’s narrow experience
Reynolds-averaged Navier–Stokes (RANS) equa- and limited exposure to the field of CFD.
tion simulations with turbulence models. Progress
of computer speeds has not yet solved and will not 2. Historical perspective of CFD in aerospace
solve this problem.
In this paper, the development of CFD in the last Practical flow simulations in aerospace using the
30 years and some of the remaining problems are compressible Navier–Stokes equations first ap-
presented from a historical perspective. Based on peared in 1985. As shown in the overview article
these observations, the future direction of CFD is in Aerospace America in 1986 [3], transonic flow
then discussed. Before closing the Introduction, the simulations over a commercial-type wing and a
author would like to leave the following message. wing-fuselage were carried out in Japan almost at
Remember that the CFD history considered here is the same time as those for fighter aircraft at the
limited to practical applications in aerospace. NASA Ames Research Center. The results [4,5] are
Everybody knows that there has been a much shown in Figs. 2(a) and (b). These were the first
earlier CFD effort in the 1900s. In addition, note example of the Japanese GFLOPS supercomputer,
Fig. 2. Three-dimensional Navier–Stokes simulations in 1986: (a) practical wing and (b) practical wing-fuselage.
ARTICLE IN PRESS
458 K. Fujii / Progress in Aerospace Sciences 41 (2005) 455–470
FUJITSU VP400, which appeared in 1985. It took leave than enter the field. Innovation starts to switch
about 2 h of computer time to achieve rough from product to process technologies, i.e. to design,
convergence to the steady state. It takes 3–4 ms per development, manufacturing innovation. As the
grid point per iterations for our CFD program to product features stabilize the specific phase is
simulate 3-D compressible Navier–Stokes equations reached where significant changes in product
on current PCs, whereas it took 7–9 ms on the features are unlikely’’. From Prof. Murman’s talk,
VPP400 in 1985. As shown in this example, steady- the author thought that Utterback’s theory also
state flow simulations with less than one million grid applied to the CFD in aerospace. In early days,
points may be carried out within an hour or so using numerical algorithms were developed to solve the
current PCs. Although simulations of the enormous basic equations, for instance, non-linear potential,
number of cases are required in the design process, Euler or Navier–Stokes equations and correspond-
they can be carried out within a reasonable time ing computer programs were created to conduct
frame. flow simulations. A lot of ideas appeared, such as
Geometry complexity is not the problem to be implicit time-integration schemes; approximate fac-
solved by computer advancement. People know that torization, approximate LU decomposition,
bottleneck in flow simulations now is not computer LU–SGS, TVD-like schemes for space discretiza-
time but the time required for preparing surface and tions; flux vector splitting, flux difference splitting,
volume grid data from a CAD geometry. There still AUSM, etc. With the aid of such algorithm
remain other problems, however, in the current development and appearance of supercomputers, a
CFD to be solved with innovative ideas. lot of researchers entered the area of CFD and its
Earll Murman, Professor at MIT, gave a general technology rapidly progressed. It was a period of
lecture in the ICAS 2000 symposium held in CFD to become matured and people enjoyed
Harrogate, England, and talked about aeronautical visualized images that showed CFD’s capability to
design engineering and manufacturing [6]. One of handle complex body configurations and complex
the figures in his manuscript is replotted in Fig. 3, physics. This period is considered to be the ‘‘fluid
which shows time evolution of a number of major phase’’ where product innovation occurs. There are
US aerospace companies. He mentioned that the still some efforts even now but I would say that the
trend follows a classic pattern of product evolution importance of such effort mainly finished early in
exhibited by many industries as studied and the 1990s, when we obtained basic CFD methods to
reported by Utterback. He said, ‘‘In the early years solve a wide variety of aerodynamic problems.
of a new product, the fluid phase, the basic product (Note that only limited areas of CFD applications
features are evolving and many startup companies are discussed and there are areas in which even the
enter the field. At some point, a dominant design mathematical model has not been well established).
emerges when the basic product features become From the late 1980s to the middle of the 1990s, there
established and a transitional phase is entered. was a discussion on ‘‘overset or patched structured
Many factors come into play to establish the grids’’ or ‘‘unstructured grids’’, which were strate-
dominant design including technology, infrastruc- gies to solve problems. Improvement of the
ture, customer expectations, individual entrepre- efficiency of the CFD solution process using parallel
neurs, etc. At this point, more companies start to computers was another hot topic, and even the
international symposium named ‘‘Parallel CFD’’
Fluid to Transitional to Specific appeared. In this period, small but inevitable effort
25
to use CFD technology for practical problems was
? the main focus. I call it ‘‘transitional phase’’ because
20 process innovation occurred (although the original
number of Firms
transitional phase to the current specific phase, the In wind tunnel experiments, many people do not
number of the researchers entering the area became spend much time on the items described in the table.
less than the number of researchers leaving the area. There are people conducting research for new
In ‘‘specific phase’’, changes in product features are measurement techniques such as PIV or PSP, but
unlikely and some innovation is required. The the majority of people simply operate wind tunnels
message here is not a negative one suggesting that with existing measurement techniques. They devel-
CFD research is finished or diminishing, but a op neither pressure sensors nor force balances for
positive one insisting that we, CFD researchers, their experiments. In other words, the items for
have to look back at what we have done and EFD listed in Table 1 are not the main topics for the
reconsider what we need to do for the future. By experimental research but just tools for their
doing so, CFD technology will continue to extend experiments. On the other hand, when talking about
its use. CFD research, people think about the items shown
in Table 1. Actually, conferences on CFD have
3. CFD in the specific phase—current status focused on these topics. CFD has been focusing too
much on the single aspect as a computational tool
3.1. CFD vs. EFD and emphasis on aerospace engineering has been
somewhat lost in the mind of CFD researchers. We
CFD researchers have been trying to prove the will come back to this point in Section 4.
capability of CFD. In general, our effort mainly
focused on fulfilling the requirement of solving 3.2. Simple problems still remain
practical problems with a certain level of fidelity
within reasonable time. As a result, we have been CFD researchers know that CFD is a powerful
showing simulation examples of the flows for more tool but not a magic tool that solves any problem.
and more complex body configurations. CFD Even though there appear a lot of examples of
research is not complete and it is true that such simulations for complex body configurations, there
effort should be continued, but at the same time, the are left many physical problems that look simple
need of users from outside the CFD community but difficult to simulate. Computations may be easy
should be considered. There may be other types of but simulations (meaning with satisfaction to the
research necessary for CFD to become a really required accuracy) are difficult. Here, three exam-
useful tool for design purpose. ples are shown for the future discussion of the
Table 1 shows the comparison of CFD with present paper.
experimental fluid dynamics (EFD) with a wind
tunnel experiment being a representative of EFD. 3.2.1. Thin-airfoil stall characteristics
This comparison may be personally biased and Precise estimation of maximum lift with a stall
there may exist a much different way of thinking, angle of a wing is an important issue for the
but this can be taken as one approach. The last aerodynamic design of aircraft. It is important and
column illustrates disadvantages of each apparatus. necessary to develop a prediction method of such
unsteady flows at high Reynolds numbers with
practical computational costs. Paul Rubbert, the
Table 1 leader of the CFD group at Boeing Commercial
Wind tunnel experiment and computational fluid dynamics Company from the 1980s until he retired, said to us,
‘‘Flow simulations at cruise condition can be done
EFD CFD
by non-linear potential equations and boundary
Wind tunnels Computers layer theory. Solutions of Navier–Stokes equations
Measurement techniques Numerical algorithms are needed for the simulation under buffet or stall
Manufacturing techniques Programming techniques
conditions,’’ when he visited Japan in 1985. CFD
(parallel language, etc.)
Model manufacturing CAD interface, grid has not yet answered his comments. It is still
generation difficult to simulate massively separated unsteady
Data acquisition Post-processing turbulent flows near stall conditions even though
Data handling Visualization software the conventional CFD technology has enabled
Reynolds number effect Discretization error,
precise numerical analysis of attached flows at
turbulence model, etc.
relatively low angles of attack. In 2000, there was
ARTICLE IN PRESS
460 K. Fujii / Progress in Aerospace Sciences 41 (2005) 455–470
0.065 0.98
Ps
0.86
0.035
0.82
0.025
0.78
0.015 0.74
0 0.2 0.4 0.6 0.8 1 0 0.2 0.4 0.6 0.8 1
(a) Chordwise location (b) Chordwise location
Fig. 7. Effect of grid resolutions: (a) location of vortex cores and (b) strength of vortex.
aerospike nozzles and other configurations having dy and steady recirculating region may be the
the base region were simulated by the group of result of an average of strongly unsteady flows.
present author using RANS model. However, Capturing such unsteady flow behavior is inevitable
pressure distributions over the base area were not for the flow analysis and eventual control of the
well captured in any simulation. A similar report flows. Without LES that captures unsteady beha-
was published in Europe. vior of the flows, accurate result may not be
obtained even as an average as the following
3.2.4. Observations from the examples examples will show.
From the three examples shown above, we notice LES has been applied to the flows of some airfoils
that there still remain problems, which are geome- near stall at high Reynolds numbers [15–17]. Mellen
trically simple but difficult to simulate by CFD. We et al. [15] suggested that LES can successfully
can easily obtain numerical solutions, but the resolve the turbulent transition directly and predict
reliability stands on a fragile base. flow behaviors including separation and reattach-
Even with less accurate solutions, we may find ment, if the mesh is adequately fine near the walls to
important physics or some data useful for design resolve near-wall turbulent structures while the
and analysis, as reliability really depends on the results with coarse mesh resolution are generally
requirements by the people who use CFD. How- disappointing. However, the mesh requirements for
ever, insufficient solutions (like improper modeling LES become enormous (from our estimation, it
or insufficient grid resolution) are sometimes used would require about 500 times more grid points
for discussions. Validation of the physical and necessary than those used in the computation shown
mathematical models and verification of the numer- below). In addition, fine mesh resolution near wall
ical solutions have been discussed in many CFD limits the time step size for the computation.
workshops. They have shown some guidelines for Therefore, it still remains difficult to apply LES to
specific applications especially from the viewpoint complex flows at high Reynolds numbers as seen in
of computational mechanics, but most of them are many engineering problems under the current
not discussed from real engineering viewpoints. computer environment.
Unfortunately, the threshold has not been estab- To overcome these difficulties, LES/RANS hy-
lished in many flow physics, and the decision brid methodology was proposed in recent years
depends on a researchers’ experience. We always [18,19]. The hybrid method used by the present
have to foresee the insufficiency behind CFD author is different from detached eddy simulations
simulations. (DES) which is a modification of the Spalart–All-
maras turbulence model [19]. The term ‘‘LES/
4. Paradigm change for the future prospect RANS hybrid’’ used here, however, includes many
of the similar approaches and the difference of the
4.1. Evolutionary effort—1: RANS to LES/RANS formulations does not influence the present discus-
hybrid methods sion. The hybrid approach is a relatively new
method in which the RANS formulation is applied
There is an obvious shift in the CFD research near the solid surface, while the LES formulation is
from RANS simulations (with turbulence model) to applied to massively separated flow regions. The
large eddy simulations (LES) due to rapid computer hybrid methodology is considered to require less
advancement. As Phillip Spalart suggested in Ref. computational cost than LES as it alleviates the
[14], emerging technology produces hybrid strate- required mesh resolution near walls and the
gies retaining conventional turbulence modeling in resultant time step limitation. The following simula-
attached region while adopting LESs in the region tion examples that the author’s group has con-
of separated flows. The examples in Section 3.2 all ducted show the effectiveness of LES/RANS hybrid
include large recirculating flow regions and are good methodology for solving the problems shown in
test cases for such hybrid strategies. The shift is Sections 3.2.1 and 3.2.3.
supported by the rapid progress of computer
performance, but more importantly, we start to 4.1.1. Thin-airfoil stall characteristics—RANS/LES
recognize that the nature of flow physics, even from hybrid simulations
the engineering viewpoint, requires unsteady flow As presented in Section 3.2, the estimation of
simulations. Separated flows are inherently unstea- thin-airfoil stall characteristics using any RANS
ARTICLE IN PRESS
K. Fujii / Progress in Aerospace Sciences 41 (2005) 455–470 463
Fig. 10. Simulation result by the LES/RANS hybrid method with compact differencing: (a) instantaneous and (b) time averaged.
Detailed discussions of the flow structures as well tions using DES (a kind of LES/RANS hybrid
as the computational method are presented in the methods) and showed that both steady and un-
original paper [23]. steady sub-structures exist in the shear layer
depending on the grid resolution [26].
4.1.2. High a flows over a delta and double-delta Visbal and Gordnier computed the flow at a low
wings—not yet solved Reynolds number using a compact difference
It is quite natural to hypothesize that the scheme [27] and studied the vortex structure in the
discrepancy between the computational and experi- shear layer. As these results suggested and the
mental results presented in Section 3.2.2 may come author pointed out at the beginning of this section,
from insufficient grid resolution. As presented in we may need to resolve the structure of the rolling-
Ref. [10], four million background grid points with up shear layer for accurate prediction of the leading-
another four million grid points locally adapted to edge separation vortex flows over a delta wing. The
the vortical flow region were used for the simula- approach should use a LES or LES/RANS hybrid
tion. The result obviously showed some improve- method and such effort is underway. The leading-
ment but was still not satisfactory. Then, we applied edge separation vortex flow structure is much more
a compact difference scheme (as the simulation of complicated than has been expected and accurate
thin-airfoil characteristics shown above) to improve simulation is still very difficult even for such simple
the spatial resolution with a limited number of grid body configurations.
points [24].
Results of simulations over a simple delta wing
and a double-delta wing are shown in Figs. 13(a) 4.1.3. Supersonic base flows—RANS/LES hybrid
and (b) as spanwise surface pressure distributions at simulations
certain chordwise locations over the wing surface. As pointed out in Section 3.2, the base pressure of
Both results show another improvement, but the a blunt body at supersonic speeds has not been
location of the main vortex still exhibits discrepancy predicted well. There are various techniques for
from the experiment. The computed locus of the numerically simulating such complicated turbulent
strake and wing vortices and resultant spanwise flows, but LES/RANS hybrid method may be a
pressure distributions in another chordwise stations good choice.
showed differences. Note that we have used a The experiment [28] by Herrin and Dutton for
RANS model and/or laminar flow simulations so axisymmetric base flow is taken here as an example
far. With improvements in the measurement tech- for the simulation. A free-stream Mach number of
niques, good experimental data became available. M ¼ 2.46 and a unit Reynolds number of 45 million
For a delta wing, Mitchell et al. [25] showed the –per meter are imposed as an inflow. With the base
presence of stationary sub-structures in the rolling- radius of 31.75 mm, the resulting Reynolds number
up shear layer as a leading-edge separation vortex in based on the diameter is 2.858 million. The
mean flow measurement at near one million computational grid used in this study is roughly
Reynolds numbers. They also carried out computa- three million [29].
1.6 1.6
Compact (6th order)
1.4 1.4 Roe (3rd order)
Roe (Fine grid+Zonal)
1.2 1.2 Experiment
1 1
−CP
0.8
−CP
0.8
0.6 0.6
0.4 0.4
Compact (6th order)
0.2 Roe (3rd order) 0.2
Experiment
0 0
0 0.2 0.4 0.6 0.8 1 0 0.2 0.4 0.6 0.8 1
(a) Spanwise Location (b) Spanwise Location
Fig. 13. Spanwise CP distributions: (a) delta wing and (b) double-delta wing.
ARTICLE IN PRESS
K. Fujii / Progress in Aerospace Sciences 41 (2005) 455–470 465
CA
over aircraft was discussed frequently and, e.g., it -0.50
was shown that the estimation error of transonic M=0.70 CFD Grid 01
M=0.70 Experiment
drag of wings by the Navier–Stokes simulations is -1.0 M=0.90 CFD Grid 01
M=0.90 Experiment
less than certain drag counts [30]. For the body M=1.20 CFD Grid 01
configurations of space transportation systems, we -1.5 M=1.20 Experiment
M=2.00 CFD Grid 01
cannot clearly tell how accurate we can predict the M=2.00 Experiment
geometries and flow configurations by simply use the rather basic disciplines for the rather simple
changing input data. In reality, CFD is used in the body configurations. The second reason is a critical
preliminary design process of aircraft wings or other reason; for any CFD simulations, we have to choose
elements in aerospace applications. However, such the right modeling and use the right equations based
CFD use is limited to the problems where a on our knowledge of fluid dynamics and aerospace
‘‘concept’’ is fixed, and CFD has been only used engineering. Even for preliminary conceptual de-
for the configuration update to increase aerody- sign, we need to use CFD to capture flow domains
namic performance where the basic concept has with certain confidence. The methodology and
been determined. Suppose we live in the period of software tool have not been well established.
the 15th century when Leonardo Da Vinci was alive Everybody knows that preprocessing of CFD to
and think of flying in air without knowing create surface and volume grids over body geome-
aerodynamic theory, we would need to consider tries requires both a grid generation specialist and
many possibilities of ‘‘wing’’ shapes. We might large amounts of time and human effort. We need
consider a flapping wing or a rotating wing. With to develop a software tool where we can carry out
CFD technology, we would be able to try many flow simulations without much effort. A tool with
possibilities without making real physical models, which we can simply assemble geometrical elements
which would really help the design process. That is by dragging a PC mouse may be a good example. It
the advantage of CFD. We have seen a lot of CFD also requires a database of geometrical elements for
results for conventional wings and body configura- aircraft and spacecraft.
tions, but we have rarely seen examples of a new Since the bottleneck of the time required for the
‘‘concept’’ in aerospace applications. CFD is a good flow simulation is the heavy requirement for
and easy tool when working on a new concept. preparing surface and volume grid data from the
There may be a possibility of new concepts even for CAD geometry, a revolution may occur here.
current commercial or fighter aircraft. Fig. 19 shows Currently, many people try to use Cartesian-type
an example of the Mars airplane. This configuration grids to avoid lengthy and tough body-fitted grid
was not designed with the sophisticated CFD generation. The difficulty exists in capturing proper
software, but can be a good representative of the viscous layers over the body surface; but if this is to
target of advanced CFD technologies. be solved, this will totally change CFD simulations
There may be two reasons why CFD has not been in aerospace [35–39].
used for the development of new concepts. One Obviously, CFD will not replace the wind tunnel
simple reason is that we have little effort. We have experiment. CFD and the wind tunnel experiment
been trying to develop software for more and more will go along side by side. We need to use CFD for
sophisticated disciplines (equations and discretiza- the tasks that are too difficult for the experiments.
tion) and apply them to more and more complex That is the key issue for the future prospect of CFD.
body configurations but have made little effort to
5. Conclusions
[21] McCollough GB, Gault DE. Examples of three representa- [31] Fujimoto K, Fujii K. Assessment of CFD estimation of
tive types of airfoil-section stall at low speed. NA- aerodynamic characteristics of basic reusable rocket config-
CATN2502; 1951. urations. Trans Jpn Soc Aeronaut Space Sci 2005;48(159).
[22] McCollough GB, Gault DE. Boundary-layer and stalling [32] Fujimoto K, Fujii K. Computational analysis of the
characteristics of the NACA64A006 Airfoil [27] Section. aerodynamic characteristics of capsule configurations to-
NACA TNI923; 1949. ward the development of reusable rockets. J Spacecr Rockets
[23] Kawai S, Fujii K. Analysis and prediction of thin-airfoil stall 2005; to appear.
phenomena with hybrid turbulence methodology. AIAA J [33] Tsukada H, Fujimoto K, Nonomura T, Miyaji K, Fujii K.
2005;43(5):953–61. Numerical analysis of the aerodynamic characteristics of
[24] Arasawa T, Fujii K, Miyaji K. High-order compact SSTO configurations with an aerospike nozzle. AIAA paper
difference scheme applied to double-delta wing vortical no. 2005-1043; 2005.
flows. J Aircr 2004;41(4):953–7. [34] Fujii K, Miyaji K. WEB-CFD and beyond-CFD for non-
[25] Mitchell A, Molton P, Barberis D, Delery D. Characteriza- CFD researchers. AIAA paper no. 2002-0753; 2002.
tion of vortex breakdown by flow field and surface [35] Aftosmis MJ, Berger MJ, Melton JE. Robust and efficient
measurement. AIAA paper no. 2000-0788; 2000. Cartesian mesh generation for component-based geometry.
[26] Mitchell A, Morton S, Forsythe J. Analysis of delta wing AIAA paper no. 97-0196; 1997.
vortical substructures using detached-eddy simulation. [36] Wang ZJ, Chen RF. Anisotropic solution-adaptive viscous
AIAA paper no. 2002-2968; 2002. Cartesian grid method for turbulent flow simulation. AIAA
[27] Visbal MR, Gordnier GE. On the structure of the shear- J 2002;40(10):1969–78.
layer emanating from a swept leading edge at angle of [37] Rogers SE, Aftosmis MJ, Pandya SA, Chaderjan NM, Tejnil E,
attack. AIAA paper no. 2003-4016; 2003. Ahmad JU. Automated CFD parameter studies on distributed
[28] Herrin JL, Dutton JC. Supersonic base flow experiments parallel computers.AIAA paper no. 2003-4229; 2003.
in the near-wake of a cylindrical afterbody. AIAA J [38] Murman SM, Aftosmis MJ, Berger MJ. Implicit approaches
1994;32(1). for moving boundaries in a 3-D Cartesian method. AIAA
[29] Kawai S, Fujii K. Computational study of a supersonic base paper no. 2003-1119; 2003.
flow using hybrid turbulence methodology. AIAA J 2005; [39] Chaderjian NM, Rogers SE, Aftosmis MJ, Pandya SA,
43(6):1265–75. Ahmad JU, Tejnil E. Automated CFD database generation
[30] Warfield MJ. The year in review, applied aerodynamics. for a 2nd generation glide-back booster. AIAA paper no.
Aerosp Am 2001;December:12–3. 2003-3788; 2003.