Professional Documents
Culture Documents
Computational
Fluid Dynamics 2006
Proceedings of the
Fourth International Conference
on Computational Fluid Dynamics, ICCFD4,
Ghent, Belgium, 10–14 July 2006
123
Editors
Prof. Dr. Herman Deconinck Prof. Dr. Erik Dick
von Karman Institute Ghent University
for Fluid Dynamics Fac. Engineering
Waterloosesteenweg 72 Dept. Flow, Heat &
1640 Sint-Genesius-Rode Combustion Mechanics
Belgium Sint-Pietersnieuwstraat 41
deconinck@vki.ac.be 9000 Gent
Belgium
Part II Schemes
Invited Speakers
“This page left intentionally blank.”
Two new techniques for generating exactly
incompressible approximate velocities
Bernardo Cockburn
1 Introduction
We consider the problem of generating exactly incompressible finite element
approximations of the velocity, that is, of approximations which are pointwise
divergence-free inside each element and whose normal component is continu-
ous across inter-element boundaries. This problem has remained open since
the beginning of the finite element simulation of incompressible fluid flow and
has only been recently solved by means of two techniques: The first is a post-
processing technique and the second a hybridization procedure. This paper is
a short review of these techniques.
The post-processing technique is devised specifically for discontinuous
Galerkin (DG) methods and is based in the fact, typical of these methods, that
the divergence-free constraint is imposed through a Galerkin weak formulation
on each element. The post-processing, which can be defined in an element-
by-element fashion, provides an exactly divergence-free approximate velocity
even though the method uses completely discontinuous elements. The tech-
nique can be applied in two or three space dimensions and works for velocity
approximations of any degree. This post-processing technique was introduced
in [2] in the framework of fluid flow in porous medium and, independently,
in [10] in the framework of the incompressible Navier-Stokes equations. The
application of this approach to incompressible elastic materials was carried
out in [11].
The hybridization technique is devised as an implementation tool for finite
element methods (including classical mixed and DG methods) which employ
divergence-free finite dimensional subspaces. It is well known that the main
difficulty with this type of methods is the construction of finite dimensional
spaces of divergence-free functions. Indeed, the construction is, in practice,
impossible due to the many inter-element continuity constraints that need to
be imposed. By using the hybridization technique, those constraints are im-
posed only on the approximate velocity, not on the test functions, by means
of a suitably chosen set of equations. Moreover, the test functions are taken
−∆u + ∇p =f on Ω, (1a)
∇ · u =0 on Ω, (1b)
u =g on ∂Ω. (1c)
2 Post-processing of DG methods
In this section, we describe the post-processing technique for obtaining exactly
divergence-free approximate velocities from the approximation provided by a
local DG method. The material of this section is contained in [10].
σ = ν∇u in Ω, (2a)
−∇ · σ + ∇p = f in Ω, (2b)
∇·u=0 in Ω, (2c)
u=g on ∂Ω. (2d)
Σh = { τ ∈ L2 (Ω)3×3 : τ |K ∈ Pk (K)3×3 , K ∈ Th },
Vh = { v ∈ L2 (Ω)3 : v|K ∈ Pk (K)3 , K ∈ Th ; },
Qh = { q ∈ L2 (Ω) : q|K ∈ Pk−1 (K), K ∈ Th ; (q, 1)Ω = 0 },
Exactly incompressible approximate velocities 5
uσh , τ · ni∂Ωh ,
(σh , τ )Ωh = −ν (uh , ∇ · τ )Ωh + ν hb (3a)
(σh , ∇v)Ωh − hb σh , v ⊗ ni∂Ωh
− (ph , ∇ · v)Ωh + hb
ph , v · ni∂Ωh = (f , v)Ωh , (3b)
p
− (uh , ∇q)Ωh + hb uh · n , qi∂Ωh = 0, (3c)
where
Φk (K) = {φ ∈ Pk (K)3 : ∇ · φ = 0, φ · nK |∂K = 0}.
Notice that if in the conditions (4a), we replace ub p by u, the function uh is
nothing but the well known BDM projection of u; see [3]. As a consequence,
the function uh is well defined. Let us show that it is exactly divergence-free
provided the triangulation Th is conforming, that is, provided if two tetrahedra
have an intersection whose area is different from zero, such intersection is a
full face of each of them.
6 Bernardo Cockburn
Since the triangulation is conforming and since, by the conditions (4a), the
normal component across inter-element boundaries of uh is uniquely defined,
the velocity uh belongs to H(div, Ω) and we can write
(∇ · uh , q)Ω = (∇ · uh , q)Ωh = −(uh , ∇q)Ωh + huh · n , qi∂Ωh ,
and, if q is in the space Qh ,
uph · n , qi∂Ωh
(∇ · uh , q)Ω = − (uh , ∇q)Ωh + hb by (4a) and (4b),
= 0,
since uh satisfies equation (3c). This implies that ∇ · uh ≡ 0 on Ω, as claimed.
This shows that it is possible to obtain a exactly divergence-free velocity uh
from the completely discontinuous velocity uh and the numerical trace u b ph in
an element-by-element fashion.
A natural question is if the new velocity uh converges to the exact solution
with at least the same order of convergence than that of the original velocity
uh ? It is not difficult to see that their difference e := uh − uh depends only
on the normal component of the difference between the uh and the numerical
trace ub ph . Indeed, from the equations (4), we see that
he · nK , ϕiF = h(bup − uh ) · nK , ϕiF ∀ϕ ∈ Pk (F ), for any face F ⊂ ∂K,
(e, ∇ϕ)K = 0 ∀ϕ ∈ Pk−1 (K),
(e, φ)K = 0 ∀φ ∈ Φk (K).
For most DG methods, the difference (b up − uh ) · nK on ∂K depends solely on
the jump of the normal component of uh and the jump of ph . If the numerical
traces of the DG method are suitably chosen, one expects the jumps to be
small, we might actually have that uh and uh do converge with the same
b p at the interior of the domain was
order. In [10], where the numerical trace u
chosen to be the arithmetic average of the two traces of uh , it was proven
that this is indeed the case. Moreover, numerical experiments showed that uh
provided a better approximation than uh since its L2 -error was smaller by a
fixed constant. In a forthcoming paper, we show that this result also holds
when the local space for the pressure on the element K is taken to be Pk (K)
instead of Pk−1 (K) and the numerical trace u b p involves jumps of the pressure
ph .
The results in this section can be extended in a number of ways. First,
notice that any other DG approximation of the equations (3a) and (3b) could
have been used; see the discussion in [9]. Notice also that the finite element
spaces can be different, provided that a suitable post-processing is available.
For example, we could use cubes instead of tetrahedra, or a Raviart-Thomas
projection instead of the BDM projection; see [3].
2.3 Applications
A simple application arises in the passive transport of a given substance by a
fluid might be modeled by the equation
Exactly incompressible approximate velocities 7
∂
c + ∇ · ( u c ) = 0,
∂t
where c is the concentration of the substance and u is the velocity given by
the Stokes system (1). When discretizing this equation, it is very convenient
to use uh since it is very well known that using an exactly divergence-free
velocity approximation gives better results.
A less evident application concerns the discretization of the steady-state
Navier-Stokes equations
−∆u + u · ∇u + ∇p =f on Ω,
∇ · u =0 on Ω,
u =g on ∂Ω.
In [10] it was pointed out that DG methods for the the Navier-Stokes equa-
tions that are energy-stable are not locally conservative and, vice versa, that
DG methods that are locally conservative are not energy-stable unless the
approximate velocity is exactly divergence-free. This difficulty was overcome
without having to resort to working with finite dimensional spaces of exactly
divergence-free velocities as follows.
It is known that the solution of the Navier-Stokes equations can be ob-
tained by solving a sequence of Oseen problems
−∆u + w · ∇u + ∇p =f on Ω,
∇ · u =0 on Ω,
u =g on ∂Ω.
Indeed, if we define Φ(w) as the exact solution u of the above Oseen equations,
it can be proven that the sequence {u(i) }i≥0 , where u(0) is arbitrary and
u(i+1) = Φ(u(i) ), converges to the exact solution of the Navier-Stokes equation
provided f /ν 2 is small enough. A discrete version of this argument provides
the construction of a DG method that is locally conservative and energy-
stable at the same time. Indeed, let us define Φh (w) as the post-processed
DG approximate velocity uh of the above Oseen equations. Then the sequence
(i) (0) (i+1) (i)
{uh }i≥0 , where uh is arbitrary and uh = Φh (uh ) does converges to
an approximation to the exact solution of the Navier-Stokes equation. Such
approximation has optimal approximation properties, as proven in [10]. There
are no other DG methods for the Navier-Stokes equation that are locally
conservative and energy-stable.
The mixed method we consider has been studied in [12, 14]. To define it, we
begin by rewriting the equations (1) in terms of the vorticity ω = ∇ × u.
Since
−∆u = ∇ × ∇ × u − ∇∇ · u,
the Stokes system (1) can be expressed as
ω−∇×u=0 on Ω, (5a)
∇ × ω + ∇p = f on Ω, (5b)
∇·u=0 on Ω, (5c)
u| = g| on ∂Ω, (5d)
u·n=g·n on ∂Ω, (5e)
W =H(curl, Ω),
V =H(div, Ω),
V(b) ={v ∈ V : ∇ · v = 0 and v · n|∂Ω = b}.
for any b ∈ H −1/2 (∂Ω), then (ω, u) is the only element of W × V(g · n)
satisfying
This completes the definition of the mixed method. Notice that, since the
velocities in the space Vh are divergence-free, the pressure does not appear in
its formulation.
Next, we show how to avoid having to construct the spaces Vh by using a
hybridization procedure.
Notice that, by the definition of the space of pressures Qh and the definition of
the normal component of the Dirichlet data gn,h , the last equation implies that
uh belongs to the space Vh (gn,h ). Moreover, since Vh ⊂ Vh (0), we see that
10 Bernardo Cockburn
The method (7) has a unique solution, as shown in [7], and, as the reader
might be expecting, the function (ωh , uh ) coincides with the approximate
solution (ωh , uh ) of the original method (6). (This is why this method is
called the hybridized version of the method (6).) Let us prove this claim.
Notice that, by equation (7c), the definition of gn,h and the definition of the
pressure space Qh , uh belongs to the space Vh (gn,h ). Similarly, by equation
(7d) and the definition of the tangential velocity space Mh , ωh belongs to
Wh . Finally, if we now take the test functions appearing in the equations (7a)
and (7b) in Wh × Vh (0) ⊂ Wh × Vh , they become identical to the equations
(6). This proves the claim.
Working with the hybridized version of the mixed method allows us to
deal only with vorticity and velocity test functions in the spaces Wh and Vh ,
respectively, which are certainly very easy to construct. Moreover, the fact
that these two spaces are sets of completely discontinuous functions, suggests
that it is reasonable to expect that the original variables ωh and uh could
be eliminated from the equations, as wanted. The resulting equations would
then characterize the two Lagrange multipliers, namely, the pressure ph on
the faces and the tangential velocity λh on the interior faces. Next, we give
such a characterization.
To state the main result of this section, we need to introduce some notation.
We define (w(λ), u(λ)) ∈ Wh × Vh and (W(p), u(p)) ∈ Wh × Vh element by
element as follows:
Moreover, the approximate solution (ωh , uh ) of the original mixed method (6)
can be determined locally as follows:
4 Hybridization of DG methods
We have seen how to post-process a DG method which uses completely dis-
continuous approximate velocities to obtain an exactly divergence-free ve-
locity. We have also shown how to hybridize a mixed method using exactly
divergence-free velocity spaces to bypass the impossible task of constructing
those spaces. One then wonders if hybridization could also be used on DG
methods that use exactly divergence-free velocity spaces [10, 9] and, in the
affirmative case, if a result similar to the theorem of section 3 holds?
In [4], it was shown how to hybridize a DG method based on a vorticity-
velocity formulation. However, unlike the hybridization carried out for the
mixed method, only the Lagrange multiplier of the pressure was introduced
and only the vorticity could be eliminated from the equations. A similar hy-
bridization procedure could be applied to DG methods not based in such
Exactly incompressible approximate velocities 13
formulation, like the ones briefly considered in [10] and later further explored
in [9]. However, also for these DG methods, a result similar to the theorem of
section 3 seems difficult to obtain.
On the other hand, there is a wide class of DG methods than can be hy-
bridized just as the mixed method considered in section 3 was. This constitutes
the subject of a forthcoming paper.
References
1. D. N. Arnold and F. Brezzi, Mixed and nonconforming finite element meth-
ods: implementation, postprocessing and error estimates, RAIRO Modél. Math.
Anal. Numér., 19 (1985), pp. 7–32.
2. P. Bastian and B. Rivière, Superconvergence and H(div) projection for dis-
continuous Galerkin methods, Internat. J. Numer. Methods Fluids, 42 (2003),
pp. 1043–1057.
3. F. Brezzi and M. Fortin, Mixed and Hybrid finite element methods, Springer
Verlag, 1991.
4. J. Carrero, B. Cockburn, and D. Schötzau, Hybridized, globally
divergence-free LDG methods. Part I: The Stokes problem, Math. Comp., 75
(2006), pp. 533–563.
5. B. Cockburn and J. Gopalakrishnan, A characterization of hybridized mixed
methods for second order elliptic problems, SIAM J. Numer. Anal., 42 (2004),
pp. 283–301.
6. , Incompressible finite elements via hybridization. Part I: The Stokes sys-
tem in two space dimensions, SIAM J. Numer. Anal., 43 (2005), pp. 1627–1650.
7. , Incompressible finite elements via hybridization. Part II: The Stokes sys-
tem in three space dimensions, SIAM J. Numer. Anal., 43 (2005), pp. 1651–1672.
8. , New hybridization techniques, GAMM-Mitt., 2 (2005), pp. 154–183.
9. B. Cockburn, G. Kanschat, and D. Schötzau, A note on discontinuous
Galerkin divergence-free solutions of the Navier-Stokes equations. Submitted.
10. , A locally conservative LDG method for the incompressible Navier-Stokes
equations, Math. Comp., 74 (2005), pp. 1067–1095.
11. B. Cockburn, D. Schötzau, and J. Wang, Discontinuous Galerkin methods
for incompressible elastic materials, Comput. Methods Appl. Mech. Engrg., 195
(2006), pp. 3184–3204. C. Dawson, Ed.
12. V. Girault and P. A. Raviart, Finite element approximations of the Navier-
Stokes equations, Springer-Verlag, New York, 1986.
13. J.-C. Nédélec, Mixed finite elements in R3 , Numer. Math., 35 (1980), pp. 315–
341.
14. , Éléments finis mixtes incompressibles pour l’équation de Stokes dans R3 ,
Numer. Math., 39 (1982), pp. 97–112.
Role of High-End Computing in Meeting
NASA’s Science and Engineering Challenges
Abstract
High-end computing (HEC) has always played a major role in meeting the
modeling and simulation needs of various NASA missions. Two years ago,
NASA was on the verge of dramatically enhancing its HEC capability and
capacity by significantly increasing its computational and storage resources.
With the 10,240-processor Columbia supercomputer in production since Oc-
tober 2004, HEC is having an even greater impact within the Agency and
beyond. Advanced science and engineering simulations in space exploration,
Shuttle operations, Earth sciences, and fundamental aeronautics research are
occurring on Columbia, demonstrating its ability to accelerate NASA’s ex-
ploration vision. This paper describes how the integrated production environ-
ment fostered at the NASA Advanced Supercomputing (NAS) facility at Ames
Research Center is reducing design cycle times, accelerating scientific discov-
ery, achieving rapid parametric analyses of multiple scenarios, and enhancing
safety for several NASA missions. We focus on Columbia’s impact on two key
engineering and science disciplines: aerospace, and climate/weather. We also
discuss future mission challenges and plans for NASA’s next-generation HEC
environment.
1 Introduction
Over the years, high-end computing (HEC) has played a major role in meet-
ing the modeling and simulation needs of various NASA missions. Two years
ago, having projected its near-term and future high-fidelity computational
requirements, NASA was on the verge of dramatically increasing its HEC
capability and capacity [4]. With NASA’s 10,240-processor supercomputer,
Columbia, in production since October 2004, HEC is having an even greater
impact within the Agency and extending to partner institutions. Significant
cutting-edge science and engineering simulations in the areas of space ex-
ploration, Shuttle operations, Earth sciences, and fundamental aeronautics
2 Aerospace Applications
2.1 Space Shuttle Launch Vehicle Debris Transport Analysis
After the STS-107 incident in February 2003, the Columbia Accident Inves-
tigation Board (CAIB) requested that CFD researchers conduct a detailed
debris transport analysis to provide insight into the actual mechanism of de-
bris shedding from the bi-pod ramp region of the SSLV. The analysis would
16 Biswas, Tu, and Van Dalsem
also furnish input for foam velocity and density for impact testing to deter-
mine damage to the orbiter’s wing leading edge [18].
Subsequently, researchers at NASA Ames developed a CFD process for
determining the aerodynamic characteristics of debris shedding during the
SSLV ascent phase. In 2005, the role of Columbia’s predecessor system in
the accident investigation [4] was reported. Since that time, a complete de-
bris scenario has been conducted on Columbia, which focused on predicting
the aerodynamic characteristics of potential debris sources, such as insulating
foam and ice. This computational analysis was critical to NASA’s Return-to-
Flight (RTF) effort, where scientists performed six-degree-of-freedom (6-DOF)
foam debris transport analyses and visualization to forecast Shuttle damage,
and for damage assessment and repair recommendations during the successful
Discovery flight in July 2005. HEC and CFD will continue to play a key role
for remaining Shuttle flights; high-fidelity simulations being integral to launch
risk analysis and component redesign.
For future flights, debris analysis has been directed to an assortment of
shapes (typically thin and conical) that can potentially be shed from the
external tank (ET) foam. The debris sources and their aerodynamic charac-
teristics are put into the debris transport code, which calculates trajectory
information to assess the potential damage or risk by a specific debris source
to a specific structural component, such as the ET foam impacting the or-
biter wing. For this interactive process to be effective, the debris transport
analysis must be done rapidly. A single trajectory calculation requires 30–60
CPU-hours on the Columbia system, which provides enough throughput to
rapidly and efficiently run hundreds of trajectories in a day, using only a frac-
tion of the computational resources. A system with the power of Columbia is
absolutely essential to run the typically hundreds of thousands of trajectories
analyzed over the entire vehicle for each iteration. Results in Fig. 1 show that
the average drag for the oscillating trajectory of an idealized frustum and the
tumbling trajectory of a highly asymmetric debris piece are similar. Also note
that farther downstream the debris travels before impact, the greater is the
impact kinetic energy, as the aerodynamic drag is constantly increasing the
relative velocity between the debris and the orbiter.
This is not the case when considering the crossrange behavior (see Fig. 2).
The dynamically stable oscillating frustum generates almost no crossrange,
as the lift force oscillates first in one direction and then in the other, with
little net effect. In order to provide a representative distribution, researchers
used several shapes to develop the crossrange constraints. These include real
digitized shapes, idealized frustums, ideal frustums with the center of mass
offset, and slightly asymmetric shapes such as elliptical frustums with the
small diameter slightly offset from the larger. The crossrange envelopes show
a zero-lift trajectory emanating from the ET flange region. The intersection
of this cone with the SSLV indicates that the fuselage and wing of the orbiter
have potential for debris impacts from this flange location, along with regions
of the left solid rocket booster.
High-End Computing at NASA 17
Fig. 1. Comparison of drag (left) and kinetic energy (right) for various debris shapes
released at Mach 2.5. Symbols show unconstrained 6-DOF trajectories compared
with trajectories using a nominal drag model based on the ensemble-averages.
In May 2002, numerous cracks were found in the SSME #1 flowliner; specifi-
cally, at the gimbal joint in the liquid hydrogen (LH2 ) feedline flowliner. Since
18 Biswas, Tu, and Van Dalsem
then, repairs have been made to existing cracks on all orbiters. Long-term sci-
entific investigations continue, because the root cause of the original cracks
was not conclusively established and remaining Shuttle flights are involved
until 2010.
High-fidelity computations have been conducted on the Columbia super-
computer to analyze the SSME LH2 feedline flowliner [12]. Numerous com-
putational models were used to characterize the unsteady flow features in the
turbopump, including the Low-Pressure-Fuel-Turbopump (LPFTP) inducer,
the orbiter manifold, and an experimental hot fire test article representing
the manifold. Findings show that unsteady flow stemming from the LPFTP
inducer is one of the major contributors to high-frequency cyclic loading that
results in fatigue damage to the flowliners.
The flow fields for the orbiter manifold and the hot fire test article were
computed and analyzed on Columbia using the INS3D incompressible Navier-
Stokes flow solver [10, 11, 13] The first computational model included only
the LPFTP inducer; by studying it, scientists were able to compare unsteady
pressure values against existing data. To resolve the complex geometry in
relative motion, an overset grid methodology [8] was employed, containing
57 overlapping zones with 26.1 million grid points. The second computational
grid system, consisting of 264 overset grids with 65.9 million grid points, added
the flowliner geometry and is shown in Fig. 3. The flowliner component alone
contained 212 grids and 41 million points.
Fig. 3. Computational model for LPFTP inducer and the LH2 flowliner: Grid (left),
and computed results illustrating unsteady interaction of the flow in the bellows
cavity and the back flow from the inducer (right).
applied to launch abort using the Apollo configuration [14]. A PRA identifies
the best level of fidelity for modeling critical failure modes associated with
launch abort. Columbia is then used to conduct higher-fidelity modeling on
specific failure modes. Two failure modes examined so far include booster
explosion and those caused by re-contact with the booster during separation.
Analysis of the booster failure mode using Apollo data showed a possible
catastrophic failure, leading to detonation of the propellant and creating blast
wave overpressures that could fatally damage the LAS (see Fig. 5). As the risk
model was being developed, it became apparent that the booster type and the
nature of the failure it was likely to encounter, determined the environments
Fig. 5. Surface pressures (left) and flowfield (right) for blast wave propagating
through wake of maneuvering LAS at t = 41.6, 72.0, and 85.5 msec (top to bottom).
High-End Computing at NASA 21
under which the LAS must operate to ensure a successful abort. The process
for characterizing this interaction must be carefully modeled and simulated.
One particular weakness found in an engineering-level model was the effect
of headwind as the CEV ascends. To account for these effects in the risk
analysis, high-fidelity blast wave models were built and simulated on Columbia
using the Overflow Navier-Stokes code [15]. Results showed that headwinds
significantly affect the nature and magnitude of the shock wave as it impacts
an escaping CEV. This means that the warning time required to initiate the
abort sequence is also affected. Additional work in high-fidelity simulations is
being done to help engineers generate requirements for the LAS while taking
headwind into consideration.
Another failure mode dependent on high-fidelity simulation involves the
ability of the LAS to achieve clean separation of the CEV from the booster
stack in the event of impending catastrophic failure. Simply put, the CEV
must not scrape or re-contact the booster stack. This failure mode is par-
ticularly demanding because it involves complex proximity aerodynamics—
modeling transonic flow as well as the complex flow at the small gap between
the CEV and the booster stack at separation. Both Navier-Stokes simulations
using Overflow, and Euler simulations using FlowCart [1], were conducted,
and their results validated against transonic wind tunnel and abort flight test
data from the Apollo era [4].
All these cases are computationally expensive to simulate. The complexity
of the geometry and the flow-field required about 30 million grid points, which
enabled good scalable performance up to 250 processors. About 20 cases were
computed using Overflow at various ascent trajectories and separation thrust
levels. Each case required approximately 20,000 CPU-hours on Columbia, in-
cluding the computation of the initial steady-state solution. All failure modes
benefited immensely from the HEC resources at the NAS facility. These tools
and processes will likely be applied to analyze the actual LAS design, and to
further understand the CEV failure modes and their impact on the vehicle’s
survivability.
To increase understanding and predictive capability for the ocean’s role in fu-
ture climate change scenarios, NASA has initiated a project called Estimating
the Circulation and Climate of the Ocean, Phase II (ECCO2): High-resolution
global-ocean and sea-ice data synthesis [17]. The goal is to produce increas-
ingly accurate combinations of all available global-scale ocean and sea-ice data
at resolutions that begin to resolve ocean eddies and other narrow current sys-
tems, which transport heat, carbon, and other properties in the ocean. These
data syntheses are used to quantify the role of the oceans in the Earth’s
22 Biswas, Tu, and Van Dalsem
carbon cycle, understand recent changes in the polar oceans, and monitor
time-evolving term balances within and between different components of the
Earth system. This work aims to harness NASA’s computational resources
such as Columbia, advances in CFD and software engineering, and the ability
to solve massive control problems.
The most challenging numerical experiment undertaken to date is a near-
global simulation with 1/16 ◦ horizontal grid spacing (approximately 6 km at
the Equator and 1 km at high latitudes). The number of surface grid cells
is about 25 million and the configuration has 50 vertical levels, bringing the
total number of cells to just over 1.25 billion. Each of the 3D fields that
describe the simulation domain and its time-evolving state requires 10 GB
of storage. This configuration has been integrated on the 2,048-CPU sub-
cluster of Columbia [9] (see Sec. 4). This workload is distributed evenly over
1,920 processors, so that each CPU is responsible for simulating about 586,000
grid cells (equivalent to a surface region roughly 210×210 km2 ). Decomposing
the workload over this many processors yields a setup that, with extensive
diagnostics and analysis options included, uses about 870 MB of main memory
per processor. With a timestep of two minutes, this performance allows a year
of simulation to be completed in less than ten days.
To investigate solution convergence as horizontal resolution is increased,
ECCO2 researchers are conducting a series of numerical simulations at 1/4 ◦ ,
1/8 ◦ , and 1/16 ◦ resolutions. Figure 6 shows significant changes in solution
with varying resolution. Each plot captures the change in simulated sea-
surface height due to eddy activity over a single month. Variation with reso-
lution occur in regions where eddies are prevalent (such as the Gulf Stream,
the Kuroshio, the Agulhas, the Drake Passage, and the Antarctic Circumpo-
lar Current). For example, in the Gulf Stream, the area where the sea-surface
height changes vigorously increases with higher resolution. Key behaviors,
such as how tightly waters stick to the coast, or how far energetic eddies
penetrate the ocean interior, also change significantly between resolutions.
Fig. 6. Gulf Stream region sea-surface height difference plots for one month at
resolutions of 1/4 ◦ , 1/8 ◦ , and 1/16 ◦ (left to right). Color scale -0.125m to 0.125m.
High-End Computing at NASA 23
During the 2004 hurricane season, the 1/4 ◦ model, which doubled the
resolution adopted by most global models in operational NWP centers at that
time, was running in real-time and providing remarkable landfall predictions
up to five days in advance for major hurricane such as Charley, Frances, Ivan,
and Jeanne [2]. Moreover, the model proved capable of resolving problems such
as erratic track, abrupt recurvature, and intense extratropical transition. In
the 2005 hurricane season, new research focused on the validation of the 1/8 ◦
fvGCM performance on hurricane forecasts, while the real-time 1/4 ◦ forecasts
provided a baseline for comparisons. Being a global mesoscale-resolving model,
the 1/8 ◦ resolution was the first to simulate mesoscale vortices (such as the
Catalina Eddy and the Hawaiian Lee Vortex), which were generated by the
interaction of the large-scale flows with better resolved surface forcing.
The 2005 Atlantic hurricane season was the most active in recorded history.
There were 28 tropical storms and 15 hurricanes, four of which were Category
5. Accurate forecasts of these storms was a significant challenge to global
and mesoscale modelers. It is well known that GCMs’ insufficient resolutions
undermine intensity predictions. Using the power of Columbia, NASA re-
searchers demonstrated that this limitation could be overcome by performing
six 5-day forecasts of hurricane Katrina [21] with the 1/8 ◦ fvGCM, showing
promising intensity forecasts with small errors in center pressure of only ±12
hPa. Notable improvement in Katrina’s intensity forecasts occurred when the
grid spacing decreased from 1/4 ◦ to 1/8 ◦ , at which the near-eye wind distri-
bution and the radius of maximum wind could be resolved more realistically.
While the mesoscale-resolving fvGCM has produced very promising results
for the past two years, significant potential for further advancement is still
ahead.
4 Columbia Description
bandwidth of 3.2 GB/s, while the Bx2 have NUMAlink4 where the bandwidth
is doubled to 6.4 GB/s.
The 20 Columbia nodes are connected by Voltaire InfiniBand fabric, as
well as via 10- and 1-gigabit Ethernet connections. Four of the Bx2 nodes
are tightly linked with NUMAlink4 (as well as the other fabrics) to form a
2,048-processor shared memory environment. Each processor in the 2,048-
CPU subsystem is a 64-bit Intel Itanium2, running at 1.6 GHz, with 9 MB of
level 3 cache and a peak performance of 6.4 GFlop/s. One other Bx2 node is
equipped with these same processors. The remaining 15 nodes have Itanium2
processors running at 1.5 GHz, with 6 MB of level 3 cache, and a peak per-
formance of 6.0 GFlop/s. Columbia is attached to more than 1 PB of online
RAID storage through a Fibre Channel switch.
Each 512-processor Altix node has several salient features that make it
particularly well-suited for executing large-scale compute and data-intensive
applications that are interesting to NASA. For example, its less than 1 mi-
crosecond latency to memory significantly reduces the communication over-
head. Typical problems are physics-based simulations involving a discretized
grid of the physical domain that is partitioned across multiple processors ex-
changing boundary data at every time step. Columbia also was the first system
(in November 2004) to exceed 1 TB/s memory bisection bandwidth on the
STREAM benchmark [22]. With global shared memory and cache coherency,
the nodes enable application programmers to use simpler and more efficient
programming models than message passing. Problems requiring adaptive grid-
ding and dynamic load balancing are much easier to program, control, and
solve on Columbia, leveraging shared memory paradigms such as OpenMP
and MLP [23]. In addition, the large shared memory of 1 TB per node allows
bigger problems to remain resident on the system.
The development and operating environment on Columbia features a 64-
processor SGI Altix front end, a Linux-based operating system, Altair PBS
Professional job scheduler, Intel Fortran/C/C++ compiler, and SGI ProPack
software.
Over the last few decades, simulation methodologies have generally advanced
along with computational technologies. Advanced tools have been developed
to the point that many daily engineering and science problems can now be
routinely computed; however, this is still done mostly using geometrically and
or physically simplified or truncated models. Some of the physical models, such
as those for turbulence and transition, and for high-temperature real gas, have
not been advanced much more than what was available in the 1970s or early
’80s.
To realize the full benefit of HEC, more inclusive modeling of geometry
and physics is needed. Attempts to solve these problems have been made with
26 Biswas, Tu, and Van Dalsem
References
1. M.J. Aftosmis, M.J. Berger, and G. Adomavicius, A parallel multilevel method
for adaptively refined Cartesian grids with embedded boundaries, 38th AIAA
Aerospace Sciences Mtg., AIAA Paper 2000-0808 (Reno, NV, 2000).
2. R. Atlas, O. Reale, B.-W. Shen, S.-J. Lin, J.-D. Chern, W. Putman, T. Lee,
K.-S. Yeh, M. Bosilovich, and J. Radakovich, Hurricane forecasting with the
high-resolution NASA finite volume general circulation model, Geophys. Res.
Lett. 32(3), 2005, L03807, doi:10.1029/2004GL021513.
3. F.R. Bailey, High-end computing challenges in aerospace design and engineering,
3rd Intl. Conf. on CFD (Toronto, Canada, 2004), pp. 13–26.
4. B.L. Barrier and O.C. Pendergraft Jr., Transonic aerodynamic characteristics
of a powered wind-tunnel model of the Apollo Launch Escape Vehicle during
separation, NASA TM-X 1336, 1967.
5. R. Biswas, M.J. Djomehri, R. Hood, H. Jin, C. Kiris, and S. Saini, An
application-based performance characterization of the Columbia supercluster,
SC05 (Seattle, WA, 2005).
6. R. Biswas, D. Kwak, C. Kiris, and S. Lawrence, Impact of the Columbia super-
computer on NASA space and exploration missions, 2nd Intl. Conf. on Space
Mission Challenges for Information Technology (Pasadena, CA, 2006), pp. 51–
58.
7. W. Brooks, M.J. Aftosmis, B. Biegel, R. Biswas, R. Ciotti, K. Freeman, C.
Henze, T. Hinke, H. Jin, and W. Thigpen, Impact of the Columbia super-
computer on NASA science and engineering applications, 7th Intl. Wkshp. on
Distributed Computing (Kharagpur, India, 2005), Springer LNCS 3741, 2005,
pp. 293-305.
8. P.G. Buning, D.C. Jespersen, T.H. Pulliam, W.M. Chan, J.P. Slotnick, S.E.
Krist, and K.J. Renze, Overflow user’s manual—version 1.8g, Technical Report,
NASA Langley Research Center (Hampton, VA, 1999).
28 Biswas, Tu, and Van Dalsem
1 Introduction
is the straight line, while the grid points move along each line. If the upstream
departure point T is given as (ξ, η)=(xi − u∆t, yj − v∆t), at first one pair of
lines satisfying xi1 < ξ < xi1+1 is searched. We should remind that xi can be
far from xi1 and xi1+1 so that large CFL(u∆t/∆x, v∆t/∆y) computation is
possible.
i1 i1+1
j1+1 j2+1
T( , )
T
A(i1, )
x j1
A(i1+1, )
x j2
x
Fig. 1. (Left) Soroban grid arrangement. (Right) Close view on a two-dimensional
plane.
Fig. 2. Soroban grid around a circle. Finite element mesh is used just for visualiza-
tion of Soroban grid
Next, two pairs of points satisfying yj1 < η < yj1+1 and yj2 < η < yj2+1
are searched along two lines at x = xi1 and x = xi1+1 , respectively. Therefore,
Multi-Phase Flow Computation with Soroban grid CIP Method 33
Here, we shall discuss the possibility to use the Soroban grid as for the adaptive
grid to the moving body. We must remind that all the discussion can be
straightforwardly applied to three dimensions. At first, let us consider one-
dimensional case. For the simplest choice of the monitoring function to the
variation, we use the following quantity:
∂f 2 12 ∂2f
M (x, t) ≡ 1 + α +β (4)
∂x ∂x2
where two parameters α and β can be chosen depending on problems we treat.
Therefore monitoring function M becomes large for larger gradient region.
Since the Soroban grid is straight in one-direction, it is much easier to generate
the adaptive grid points along the line. The reorganization of the grid point
34 Takashi Yabe, Youichi Ogata and Kenji Takizawa
Fig. 3. Accumulated monitoring function I (x) is divided into equal pieces. The x
boundary of each piece gives the grid point.
move by line
move by point
Fig. 4. The straight lines can move in horizontal direction and grid points move
along the straight lines.
itoring function to the variation, and local mesh refinement can be achieved.
The drawback was that the Soroban grid thus generated would not be smooth
because the grid size would be automatically determined by the monitoring
function and thus the ratio of grid size would become extremely large. One way
to avoid such drawback is to smooth the spatial derivatives in the monitoring
function, or to use distance function to a body surface.
The CIP method is suitable for this mesh system because it uses only two
stencils to make up cubic-interpolation function, and the calculation of large
CFL(> 10) at locally refined mesh is easily performed. Mesh generation and
searching of upstream departure point are very simple and almost mesh-free
treatment is possible.
The underlying physics included in the above equations for continuum dy-
namics is complex and may include processes that have different time scales
of variation. It is expedient to separate the solution procedure into several frac-
tional steps. For example, Eq.(5) can be separated into the advection phase,
∂f n ∗
+ (u · ∇) f = 0 (f , ∂χ f ) → (f , ∂χ f ) (6)
∂t
and the non-advection phase.
∂f
f ∗ → f n+1
=S (7)
∂t
The value and the spatial gradient in the advection phase are solved by the
CIP method. The time evolution of the spatial gradient in the non-advection
36 Takashi Yabe, Youichi Ogata and Kenji Takizawa
Eq.(7)
∂ (∂χ f )
= ∂χ S (8)
∂t
Usually, it is not easy to get a finite-difference form of ∂χ S. However, using
Eq.(7), this term can be estimated as
n+1 ∗ n+1 ∗
(∂χ f ) − (∂χ f ) (δχ f ) − (δχ f )
= (9)
∆t ∆t
where δχ f represents a centered finite-difference form of ∂χ f in the χ direction.
Thus the time evolution of ∂χ f is estimated by the time evolution of f already
given by Eq.(7).
The solution of Eq.(7) can be obtained by a simple finite difference like
f n+1 = f ∗ + S∆t. However, we have to use a more sophisticated method in
order to deal with compressible and incompressible fluids, or multi-phase flow
simulutaneously. Next section, the method and its application to Soroban grid
will be discussed.
Yabe and Wang[7] adopted the primitive Euler form instead of the conserva-
tive form to construct the pressure equation.
The original CCUP method[7] was proposed only for a special equation of
state, but here we use
∂p ∂p
∆p = ∆ρ + ∆T (10)
∂ρ T ∂T ρ
where ∆p means the pressure change pn+1 − p∗ during one time step and * is
the profile after advection. From this relation, once ∆ρ and ∆T are predicted,
∆p can be predicted based on Eq.(10); ∂p/∂ρ, ∂p/∂T are given by the EOS.
Since the non-advection terms are separated from the advection terms by
the CIP, we can concentrate on the non-advection terms related to sound
waves, which are the primary cause of the difficulty posed by the large sound
speed of liquid, and hence ρ, T are simply given by
∆ρ = −ρ∗ ∇ · un+1 ∆t ρ∗ Cv ∆T = −PT H ∇ · un+1 ∆t (11)
un+1 in this equation is given by an equation of motion as
∇pn+1
∆u = − ∆t (12)
ρ∗
Multi-Phase Flow Computation with Soroban grid CIP Method 37
n+1
Then substituting the given p into Eq.(12), we obtain the velocity un+1
n+1
and then the density ρ from Eq.(11). From this pressure, density can be
solved in terms of pressure. Equation (13) has many important features. This
equation shows that, at sharp discontinuities, n · (∇p/ρ) is continuous. Since
∇p/ρ is the acceleration, it is essential that this term be continuous since the
density changes by several orders of magnitude at the boundary between liquid
and gas. In this case, the denominator of ∇p/ρ changes by several orders,
and the pressure gradient must be calculated accurately enough to ensure
continuous change of acceleration. It has been proved in many applications of
Eq.(13) that it works robustly even with a density ratio larger than 1000 and
enables us to treat both compressible and incompressible fluids.
xin 1 xin
N
y j +1
dS n
yj
xi 1 dS w xi dS e xi +1
dV
dS s
y j 1
xis 1 S xis
Fig. 5. C-CUP method in Soroban grid
When Eq.(13) is solved with Soroban grid, the adequate formulation can
be derived by the integration of Eq.(13) in the shaded square (≡ Ω) in Fig.5
as
pn+1 − p∗ ∇ · u∗
Z Z Z
1 n+1
∇· ∗
∇p dV = 2
dV + dV (14)
Ω ρ P
Ω ∆t2 ρ∗ C 2 + T H Ω ∆t
s ρCv T
where dV = (xi+1 − xi−1 ) (yi+1 − yi−1 ) /4, dSe = −dSw = (yi+1 − yi−1 ) /2,
dSn = −dSs = (xi+1 − xi−1 ) /2, and ∂Ω means the integration along the cell
boundary in two dimension. It is not to mention that three dimension is also
straightforward using Eq.(15)
Although density at points (N, S), and velocity at cell boundaries (n, s, w, e)
in Fig.5 are needed in the formulation of Eq.(15) with Soroban grid, they can
be interpolated by the CIP method. The methods to solve this poisson equa-
tion for pressure have been established like SOR, Bi-CGSTAB and so on, thus,
CCUP method can be used even with Soroban grid as well[14].
Karmann Vortex
Skimmer
Container
The last application is the interaction between a ship (a container) and stokes
waves. The length between profile of a container is set to be 283.8[m]. The
wavelenghth λ is also 283.8[m], which is the same as a container length, and
wave amplitude is 8.0[m]. The calculation domain is 800[m]×400[m]×200[m],
and 10 [m/sec] fixed surge motion is set. The number of Soroban grid points
is changed at each time step from 100, 000 to 250, 000.
Figure 8 shows the motion of a container by stoke waves, and Figure 9
shows the bottom of a container with Soroban grid.
40 Takashi Yabe, Youichi Ogata and Kenji Takizawa
Fig. 7. Grid points used in the skimmer simulation. Vertical plane shows the grid ar-
rangement. For visualization, grid points in three dimensions are depicted by spheres
whose size depends on the density.
5 Conclusion
We have reviewed various families of the CIP method to solve the problem
of multi-phase flow, and proposed recent advance of multi-phase flow com-
putation with Soroban grid CIP method. Although the conventional C-CUP
method in Cartesian grid has already been established and applied to a lot of
multi-phase flow simulations like laser-processing, skimmer, milkcrown and so
on that include solid, liquid and gas as well as phase transition, Soroban grid
CIP method is able to open new avenue for significantly improved scheme to
solve multi-phase flow simulations. We confirmed that the CIP can realize the
third-order accuracy both in time and space even in such arbitrary mesh. The
merit of the Soroban grid can be summerized as
1. It is easy to generate, and load to generate grid is very low.
2. Searching the upstream point is easy and can be done by one step by
refering to the index.
3. The original high accuracy of CIP can be kept.
4. CFL can be large.
Soroban grid is being applied to other fields like plasma simulation
with Vlasov-Boltzmann equations and the propagation of electromagnetic
wave with Maxwell’s equations[17] and so on, and more applications to the
hyperbolic-type equations will be discussed in future papers.
Multi-Phase Flow Computation with Soroban grid CIP Method 41
References
1. Harlow,F.H. and Amsden,A.A: Numerical Simulation of Almost Incompressible
Flow, J.Comput.Phys., 3, 80–93 (1968)
2. Takewaki, H., Nishiguchi,A. and Yabe, T: The Cubic-Interpolated
Pseudo-Particle (CIP) Method for Solving Hyperbolic-Type Equations,
J.Comput.Phys., 61, 261–268 (1985)
3. Takewaki, H. and Yabe, T: Cubic-Interpolated Pseudo Particle (CIP) Method
Application to Nonlinear or Multi-Dimensional Problems, J.Comput.Phys., 70,
355–372(1987)
4. Yabe, T., and Aoki, T.: A Universal Solver for Hyperbolic Equations by Cubic-
Polynomial Interpolation, Comput. Phys. Commun., 66, 219–232 (1991)
5. Yabe, T., Ishikawa, T., Wang, P. Y., Aoki, T., Kadota, Y., and Ikeda, F.: A
Universal Solver for Hyperbolic Equations by Cubic-Polynomial Interpolation
II. Two- and Three- Dimensional Solvers, Comput. Phys. Commun., 66, 233–
242 (1991)
6. Yabe, T., Mizoe, H., Takizawa, K., Moriki, H., Im, H. N., and Ogata, Y.:
Higher-Order Schemes with CIP Method and Adaptive Soroban Grid Towards
Mesh-Free Scheme, J.Comput.Phys., 194, 57–77 (2004)
7. Yabe, T., and Wang, P. Y.: Unified Numerical Procedure for Compressible and
Incompressible Fluid, J.Phys.Soc.Japan., 60, 2105–2108 (1991)
8. Yabe, T., Tanaka, R., Nakamura, T., and Xiao, F.: An Exactly Conservative
Semi-Lagrangian Scheme (CIP-CSL) in One Dimension, Mon.Wea.Rev., 129,
332–344 (2001)
9. Nakamura, T., Tanaka, R., Yabe, T., and Takizawa, K.: Exactly Conservative
Semi-Lagrangian Scheme for Multi-Dimensional Hyperbolic Equations with Di-
rectional Splitting Technique, J.Comput.Phys., 174, 171–207 (2001)
Multi-Phase Flow Computation with Soroban grid CIP Method 43
10. Takizawa, K., Yabe, T., Chino, M., Kawai, T, Wataji, K, Hoshino, H., and
Watanabe, T.: Simulation and Experiment on Swimming Fish and Skimmer by
CIP Method, Computers & Structures, 83, 397–408 (2005)
11. Yabe, T., Xiao, F., and Utsumi, T.: Constrained Interpolation Profile Method
for Multiphase Analysis, J.Comput.Phys., 169, 556–593(2001)
12. Wang, P. Y., T, Yabe, T, Aoki.: A General Hyperbolic Solver - the CIP Method
- Applied to Curvilinear Coordinate, J.Phys.Soc.Jpn., 62, 1865–1871(1993)
13. Aoki, T.: Multi-dimensional advection of CIP(cubic-interpolate propagation),
CFD.J, 4, 279–291(1995)
14. Takizawa, K., Yabe, T., Tsugawa, Y., Tezduyar, T.E., Mizoe, H. : Computation
of free-surface flows and fluid-object interactions with the CIP method based
on adaptive meshless soroban grids , Comput.Mech, in print(2006)
15. Yabe, T., Ogata, Y., Takizawa, K., Kawai, T., Segawa, A., Sakurai, K. : The
next generation CIP as a conservative semi-Lagrangian solver for solid, liquid
and gas, J.Comput.Appl.Math. 149, 267–277 (2002)
16. Yabe, T., Takizawa, K., Chino, M., Imai, M., Chu, C.C.,: Challenge of CIP as
a Universal Solver for Solid, Liquid and Gas, Int. J. Numer. Meth. Fluids. 47,
655–676 (2005)
17. Ogata, Y.,Yabe, T.,Odagaki, K.: An Accurate Numerical Scheme for Maxwell
Equation with CIP-Method of Characteristics, Comm.Comp.Phys, 1, 308–
331(2006)
“This page left intentionally blank.”
Part II
Schemes
“This page left intentionally blank.”
On the Computation of Steady-State
Compressible Flows Using a DG Method
1 Introduction
2 Numerical Method
points, the numerical fluxes are computed exactly as usual straight-sided ele-
ments. Using only straight-sided elements instead of boundary fitted elements
represents a huge simplification of the code implementation and tremendous
saving in both storage requirements and computing costs. In our implementa-
tion, the normals at the boundary integration points are computed using the
local true surface normal based on the analytically defined boundary geome-
tries, which are handily available in our geometry definition file.
It has been well known that just like DGM are more sensitive to the treatment
and implementation of slip boundary conditions at curved boundaries than
those obtained with FVM of the same order of accuracy, DGM are also much
more sensitive to the treatment and implementation of the slope limiters than
their FV counterparts. Slope limiters frequently identify regions near smooth
extrema as requiring limiting, and this typically results in a reduction of the
optimal high-order convergence rate and degenerates solution accuracy. To ad-
dress this concern, the limiters are applied only where they are really needed.
This is accomplished using the so-called discontinuity detector [4], which is
helpful to distinguish regions where solutions are smooth and discontinuous.
Then, the limiting is only used near discontinuities and high-order accuracy
can be preserved in smooth regions. Unfortunately, almost all discontinuity
detectors, will mistakenly identify local high gradient smooth extrema such
as leading edge regions as shock regions due to the high comparable gradient
there. For aerodynamic applications, the active limiters at the leading edge of
an airfoil will pollute the solution in the flow field and ultimately destroy the
high order accuracy solution. To address this concern, a physics-based shock
detector [5] is introduced by taking into consideration the two facts that the
normal flow before a shock is supersonic and normal flow after the shock is
subsonic and a shock wave can only be a compressive wave. Our experience in-
dicates that this shock detector can effectively eliminate high gradient smooth
regions, and performs better than alternative indicators for the steady-state
flow problems considered here.
3 Computational Results
A. Subsonic Flow past a sphere
The first example is a subsonic flow past a sphere at a Mach number of
M∞ =0.5. This test case is chosen to verify the implementation of boundary
conditions for curved geometries for the DG method, and assess the order
of accuracy of the discontinuous Galerkin method and performance of the
p-multigrid. The computational experiments have been performed on three
successively refined grids in order to obtain quantitative measurement of the
order of accuracy for different order DG methods, Fig.1 shows the medium
50 Hong Luo, Joseph D. Baum and Rainald Löhner
-1.5 -1.5 0 0
-0.5 -0.5
-2 -2
-1 -1
-2.5 -2.5 -1.5 -1.5
Log(Error-L2)
-2 -2
-3 -3
Log(resi)
-2.5 -2.5
-3.5 -3.5
-3 -3
-4 -4 -3.5 -3.5
-4 -4
-4.5 DG(P0) (Slope=0.67) -4.5
DG(P1) (Slope=1.82) -4.5 TVDRK(P2) -4.5
DG(P2) (Slope=2.71) p-Multigrid(P2)
-5 -5 -5 -5
-0.7 -0.6 -0.5 -0.4 -0.3 -0.2 -0.1 0 0 100 200 300 400 500 600 700 800 900 1000
Log(cell-size) CPU (Second)
Fig. 1. The medium unstructured surface meshes used for computing subsonic flow
past a sphere at M∞ = 0.5, computed Mach number contours obtained by DG(P2)
on the medium mesh, accuracy summary for DG(P0), DG(P1) and DG(P2) methods,
and comparison of convergence history versus CPU time between TVBRK and p-
multigrid methods.
mesh used in the computation, the computed Mach number contours in the
flow field obtained by DG(P2) solution on the medium mesh, the spatial accu-
racy details of each DG method for this numerical experiment, and the com-
parison of convergence histories versus CPU time for the DG(P2) solution on
the coarse mesh between 3-stage TVBRK method and p-multigrid method.
The numerical results indicate that the discontinuous Galerkin method ex-
hibits a O(hp+1 ) order of convergence on smooth solutions, and show that
our simplified implementation of boundary conditions for curved geometries
well conserves the formal order of the DG method. The p-multigrid method
is orders of magnitude faster than its explicit TVDRK counterpart.
0.02 0.02
Mach Number
0.8 0.8
0.015 0.015
4 Conclusions
Computation of compressible steady state flows using a high-order discontin-
uous Galerkin finite element method is presented in this paper. All three areas
required in the practical application of DG methods: efficient flow solver, im-
plementation of boundary conditions for curved geometries, and limiting for
discontinuities have been discussed. The developed DG method provides a vi-
able, attractive, and competitive alternative to the finite volume method for
CFD applications. Future work will explore application of this method for the
solution of the Navier-Stokes equations.
References
1. Luo, H., Baum, J.D., Löhner, R.: A p-multigrid Discontinuous Galerkin Method
for the Euler Equations on Unstructured Grids. Journal of Computational
Physics, Vol. 211, No. 2, pp. 767-783 (2006)
2. Bassi, F. and Rebay, S.: High-Order Accurate Discontinuous Finite Element
Solution of the 2D Euler Equations. Journal of Computational Physics, Vol.
138, pp. 251-285 (1997)
3. Krivodonova, L. and Berger, M.: High-order Implementation of Solid Wall
Boundary Conditions in Curved Geometries. Journal of Computational Physics,
Vol. 211, No. 2, pp. 492-512 (2006)
4. Krivodonova, L., Xin, J., Remacle, J.F., Chevaugeon, N., and Flaherty, J.E.:
Shock Detection and Limiting with Discontinuous Galerkin Methods for Hyper-
bolic Conservation Laws. Applied Numerical Mathematics, Vol. 48, pp. 323-338
(2004)
5. Luo, H., Baum, J.D., Löhner, R.: A physics-based shock detector for Discontin-
uous Galerkin Method. In preparation (2006)
Space-Time Discontinuous Galerkin Method
for Large Amplitude Nonlinear Water Waves
Summary.
A space-time discontinuous Galerkin (DG) finite element method for nonlinear
water waves in an inviscid and incompressible fluid is presented. The space-time DG
method results in a conservative numerical discretization on time dependent deform-
ing meshes which follow the free surface evolution. The dispersion and dissipation
errors of the scheme are investigated and the algorithm is demonstrated with the
simulation of waves generated by a wave maker.
Key words: space-time discontinuous Galerkin method, nonlinear water waves
1 Introduction
Large amplitude nonlinear water waves are frequently modeled by considering
the fluid as inviscid, incompressible and irrotational. This makes it possible to
describe the flow field with a potential function φ, which satisfies the Laplace
equation. In addition, the potential function must satisfy a kinematic and
dynamic boundary condition at the free surface:
∂ζ ∂φ
+ ∇ s φ · ∇s ζ − =0 (1)
∂t ∂z
∂φ 1
+ ∇φ · ∇φ + z = 0, (2)
∂t 2
∂ ∂ ∂ T ∂ ∂ T
with ζ the wave height, t time, and ∇ = ( ∂x , ∂y , ∂z ) , ∇s = ( ∂x , ∂y )
defined with respect to a Cartesian coordinate system (x, y, z), where z is the
coordinate direction pointing upward from the flat water surface. At a wave
maker a prescribed normal velocity is imposed and at other solid surfaces a
zero normal velocity. The equations are made dimensionless with an average
water depth and the gravitational constant.
The numerical solution of large amplitude nonlinear water waves is non-
trivial since the free surface boundary conditions must be imposed at the
actual free surface, which must be determined as part of the solution process.
The free surface experiences large deformations during the wave evolution, in
particular due to wave interactions. An additional complication is that the
mathematical model has no inherent dissipation, which makes it difficult to
ensure sufficient stability and robustness of the numerical scheme.
In this paper we will present a new space-time discontinuous Galerkin
finite element algorithm to compute large amplitude nonlinear waves. The al-
gorithm uses basis functions which are discontinuous both in space and time,
and discretizes the equations directly in four dimensional space, with time as
the fourth dimension. This approach combines the benefits of discontinuous
Galerkin methods, for instance their suitability for hp-mesh adaptation and
parallel computing, with an accurate representation of the time-dependent
boundary using deforming elements. The space-time DG algorithm is closely
related to the so called arbitrary Lagrangian Eulerian technique, which pro-
vides more flexibility in the mesh deformation algorithm, see e.g. [2], but the
space-time DG method ensures that the discretization remains conservative on
deforming meshes. The space-time discontinuous Galerkin method discussed
in this paper is an extension to nonlinear waves of the spatial discretization
discussed in [1] for linear waves using the space-time techniques presented in
[2].
with L2 (Enh ) the space of Lebesgue square integrable functions on Enh and
Pp (K) the space of polynomials of degree p on K. The finite element space
Wpn associated with the free surface is defined as
1
hvi := (vL + vR ) , [[v ]] := vL n̄L + vR n̄R , (3)
2
with vL := v|∂KL and vR := v|∂KR , and KL , KR the elements connected to
the face F ∈ FIn with outward space normal vectors n̄L and n̄R , respectively.
For q ∈ Σpn we similarly define qL , qR , hqi and [[q ]].
The novel ingredient in the finite element formulation discussed in this
paper is the incorporation of the kinematic condition at the free surface (1)
as a natural boundary condition in the finite element formulation. The space-
time finite element discretization then will automatically account for the mesh
movement necessary to follow the free surface waves. In order to accomplish
this we need to establish a relation between the function f = ζ − z, describing
the free surface, the wave height ζ and the space-time normal vector n. A
straightforward calculation using (1) shows that
∇f ( ∂f , ∇f )T 1 ∂ζ ∂ζ ∂ζ
n = (nt , n̄)T = = ∂t = ( , , , −1)T , (4)
|∇f | |∇f | |∇(ζ − z)| ∂t ∂x ∂y
∂ ∂ ∂ ∂ T
with ∇ the space-time nabla operator defined as ∇ = ( ∂t , ∂x , ∂y , ∂z ) . We
can rewrite the kinematic condition (1) as:
1 ∂f ∇f
+ ∇φ · = 0, (5)
|∇f | ∂t |∇f |
which implies using (4) that the space component of the normal velocity at
the free surface ΓS is equal to
−1 ∂ζ
n̄ · ∇φ = . (6)
|∇(ζ − z)| ∂t
This relation can be used directly in the DG formulation of the Laplace equa-
tion, which is defined as:
Find a (φh , ζh ) ∈ Vpn × Wpn , such that for all (v1 , v2 ) ∈ Vpn × Wpn ,
∂ζh
ζh+ − ζh−
Bh (φh , v1 ) + ( ∂t
, v1 )ΓSn + ( , v+ ) + = Lh (v1 ), (7)
|∇h (ζh − z)| |∇h (ζh+ − z)| 1 ΓS (tn )
∂φh − +
( , v2 )ΓSn + (ζh , v2 )ΓSn + (φ+h − φh , v2 )ΓS (t+
n)
(8)
∂t
1 X X
+ ( (∇h φh − RF ([[φh ]])) · (∇h φh − RF ([[φh ]])), v2 )ΓSn = 0,
2
F∈FI
n nF∈FI
± ±
where ΓS (t+
n ) = lim ΓS (tn + ), φh = lim φh (tn ± ), ζh = lim ζh (tn ± ), with
↓0 ↓0 ↓0
φ− −
h and ζh the known potential and wave height at t = tn . The operators
Bh : Vp × Vpn → R and Lh : Vpn → R are defined as:
n
Z Z
Bh (u, v) = ∇h u · ∇h vdx − [[u]] · ∇h v + [[v ]] · ∇h u ds
En
h Γ0n
56 Yan Xu and Jaap J.W. van der Vegt
X Z
+ (ηF + nf ) RF ([[u ]]) · RF ([[v ]])dx, (9)
F∈FIn En
h
Z
Lh (v) = vgN ds,
n
ΓN
3
with the local lifting operator RF : L2 (F) → Σpn :
Z Z
RF (q) · σdx = q · hσi ds, ∀σ ∈ Σpn . (10)
En
h F
The resulting nonlinear equations are solved with a Newton method with
special attention given to the discretization of the free surface terms. During
each time step and Newton iteration the mesh is adjusted to accommodate for
the free surface motion. For more details about the scheme and the numerical
implementation, we refer the reader to [3].
∂ζh
Bh (φh , v1 ) + ( , v1 )ΓSn + (ζh+ − ζh− , v1+ )ΓS (t+
n)
= Lh (v1 ), (11)
∂t
∂φh − +
( , v2 )ΓSn + (φ+
h − φh , v2 )ΓS (t+
n)
+ (ζh , v2 t)ΓSn = 0. (12)
∂t
We use the Fourier ansatz for the coefficients φbj and ζbj in the space-time DG
discretization at time level tn = n∆t:
-1.7E
0.8 7 0.8 -0 6
E-0 5.0E-07
5.0
5.0E-07
6
0.6
E-0 0.6
2.0 -5.6E-07
5.0E-07
! "#
! "#
6 -5.6E-07
E-0
0.4 2.0 0.4
-5.6E-07
0.05 0.1 0.15 0.2 0.25 0.05 0.1 0.15 0.2 0.25
h h
Fig. 1. Dispersion (left) and dissipation (right) for the space-time DG method with
quadratic polynomial basis functions.
Np (Nz −1) bF
with φbF
D ∈R , φS ∈ RNp , ζbF ∈ RNs , where Np = (p + 1)3 , Ns = (p +
2
1) and p the order of the polynomials. The suffix S refers to the coefficients in
elements connected to the free surface and D to elements not connected to the
free surface. The matrices P, Y ∈ RNp ×Ns , H, V ∈ RNs ×Np and S ∈ RNs ×Ns
are related to the free surface. The matrix MF ∈ RNp Nz ×Np Nz is a Hermitian
positive definite block-tridiagonal matrix.
The eigenvalue λ is computed with MATLAB for a wide range of ∆t values
and k ∈ (0, 2π]. For all cases, the modulus of λ is always less than or equal to
one, hence the numerical discretization is unconditionally stable.
We also use Fourier analysis to compute the dispersion and dissipation
error of the numerical scheme by comparing the frequency and dissipation of
the discrete modes with the exact harmonic wave solution. The results for
quadratic polynomial basis functions are shown in Fig. 1.
4 Numerical example
0.05 0.1
-0.05 -0.05
0 1 2 3 4 0 1 2 3 4
x x
Fig. 2. Wave profile at T = 7.5 (left) and T = 20 (right) generated by a wave maker
at x = 0 for polynomial basis functions of degree p = 1 and 2 using the space-time
DG method.
5 Concluding remarks
References
1. J.J.W. van der Vegt, S.K. Tomar, Discontinuous Galerkin method for linear
free-surface gravity waves, J. Sci. Comput. 22-23 (2005) 531-567.
2. J.J.W. van der Vegt and H. van der Ven, Space-time discontinuous Galerkin
finite element method with dynamic grid motion for inviscid compressible flows,
Part I. General formulation, J. Comput. Phys., 182 (2002), 546-585.
3. J.J.W. van der Vegt and Y. Xu, Space-time discontinuous Galerkin method for
nonlinear water waves, preprint for J. Comput. Phys., 2006.
A discontinuous Galerkin method with
Hancock-type time integration for hyperbolic
systems with stiff relaxation source terms
1 Introduction
In recent years, the need for robust high-order (more than second-order ac-
curate) discretizations for high-fidelity CFD on unstructured grids has been
widely recognized. To be attractive, a high-order method must not only be ac-
curate, but efficient as well, thus reducing the total CPU time needed to yield
a given accuracy. In this paper we propose a combination of two approaches
toward robust, accurate, and efficient schemes for advection-dominated flows
on unstructured grids, one at the PDE level and the other at the discretization
level.
The first approach is replacing everybody’s favorite Navier-Stokes (NS)
equations by a larger set of first-order hyperbolic-relaxation PDEs which con-
tains the NS equations. (N.B.: here “first-order” refers to the order of the
PDEs.) In comparison with second-order PDE models such as the NS equa-
tions, first-order PDE models offer many numerical advantages. For example,
the latter require smaller discrete stencils, reduce communications in par-
allel processing, can replace global stiffness from diffusion terms with local
stiffness from source terms, and yield the best accuracy on non-smooth, adap-
tively refined grids. We will therefore consider the one-dimensional hyperbolic-
relaxation form
1
∂t u + ∂x f (u) = s(u), (1)
where u is the vector of conserved variables, f is the flux of u, s is the source
term, and is the relaxation time. This is the form of moment closures of the
Boltzmann equation, where the source term describes departure from local
thermodynamic equilibrium. In the near equilibrium ( 1), that is, when
the system is stiff, the system formally reduces to a system of second-order
equations,
∂t U + ∂x F(U) = ∂x (D∂x U), (2)
where D is a tensor of diffusion coefficients, with eigenvalues proportional to
.
The second approach is to adopt a discretization method which can pre-
serve compactness in high-order methods. In standard finite-volume methods,
higher-order accuracy relies on piecewise-polynomial reconstruction, which
requires extended stencils. For instance, stencils for the quadratic reconstruc-
tion (third-order) on tetrahedral grids include 50 to 70 cells [2]. Discontinuous
Galerkin (DG) methods overcome the issue of reconstruction by using extra
equations for updating the polynomial representation of state variables. For
a comprehensive literatures review, see Cockburn and Shu [1]. Currently, the
most successful DG methods are semi-discretizations combined with TVD
Runge-Kutta (RK) ODE solvers [1], denoted as RKm-DG(k) where m is the
order of the RK method and k is the degree of polynomial of basis function.
The method proposed here is based on the “upwind moment scheme”
recently developed by Huynh for hyperbolic conservation laws [4]. The solu-
tion representation is only piecewise linear; higher-order extensions will be
presented elsewhere. The two key characteristics of this method are: 1) cell
variables are updated over a half time step without any interactions with
neighboring cells (Hancock’s observation [8]); 2) the gradient of each flow
variable evolves by an independent equation (DG representation). The result-
ing scheme looks promising in comparison to the currently popular methods
based on a semi-discretization with method-of-line (M.O.L.) time integration.
It is a fully-discrete, one step method with one intermediate update step
needed for computing the volume integral of the fluxes. It requires solving
a Riemann problem twice at each cell interface but achieves third-order ac-
curacy in time and space. Interestingly, the upwind moment scheme for a
linear advection equation reduces to Van Leer’s “scheme III” [7], which is a
DG spatial discretization with an exact shift operator for time evolution. It
was shown that the method is linearly stable up to CFL number 1 with an
upwind flux, whereas for the RK2DG(1) (second-order) the limit is 1/3 and
for the RK3DG(2) (third-order) it is 1/5 [1]. Here we extend the upwind mo-
ment method to a simple hyperbolic-relaxation system, apply a Von Neumann
analysis to it, and confirm its results by numerical experiments.
A discontinuous Galerkin method with Hancock-type time integration 61
Z tn+1 Z xj+1/2
u(x, t)v(x)dx
=− f (u(x, t))v(x) dt
Ij tn Tn xj−1/2
ZZ ZZ
1
+ f (u(x, t))vx (x)dxdt + s(u(x, t))v(x)dxdt. (3)
Ij ×T n Ij ×T n
Note that Eq.(3) is still exact. To approximate the boundary and volume
integrals on the right-hand side, we use Legendre polynomials for both test
and basis functions. Thus, the basis of v(x) be {φ0 (x), φ1 (x)} where φ0 (x) =
1, Rφ1 (x) = (x − xj )/h, and u(x, t) ≈ ū(t) + ∆u(t)(x R − xj )/h where ū(t) ≡
1 12
h Ij u(x, t)dx is the cell average and ∆u(t) ≡ h2 Ij (x − xj )u(x, t)dx is the
undivided gradient of u. At the boundary, the midpoint rule is used for the
time integration of the interface flux, thus the upwind flux f ∗ (uL , uR ) ≡
A+ uL + A− uR with A(u) ≡ ∂f /∂u is evaluated based on the state variables
at the half-time level,
Z tn+k
1 n+k/2 ∗ n+k/2 n+k/2
f (u(xj+1/2 , t))dt ≈ fj+1/2 ≡ fj+1/2 (uj+1/2,L , uj+1/2,R ), (4)
k∆t tn
n+k/2
where ∆t = tn+1 − tn . The cell-interface values, uj+1/2,L/R are obtained by
a Taylor-series expansion of u(x, t) in space and time using the Lax-Wendroff
procedure (replacing time derivative by spatial derivative); the final forms are
62 Yoshifumi Suzuki and Bram van Leer
n+k/2 1 k∆t n k∆t n+k/2
uj+1/2,L = ūnj + I− A(ūnj ) ∆uj + s(uj+1/2,L ),
2 h 2
(5)
n+k/2 1 k∆t n k∆t n+k/2
uj+1/2,R = ūnj+1 − I+ A(ūnj+1 ) ∆uj+1 + s(uj+1/2,R ).
2 h 2
Note that implicit character is caused by the source terms. The first update
formulas to compute the intermediate values at the Radau point (k = 1/3)
are found, after some algebra, to be
fluxes respectively. The unity value in the Rusanov flux comes from the dimen-
sionless frozen wave speed. For DG methods, the amplification factor becomes
a 2 × 2 matrix even though a scalar equation is considered. Thus two eigen-
values exist, but only the one accurate for low frequencies is described here.
Recalling that the eigenvalue of the exact advection operator is −irk, where
k is the wave number, we only list the numerical error part of the eigenvalues
in the long wave-length limit:
1 hq i
λRK2HR2 = c3 (rν)2 + k 3 + 9c4 − (rν)3 k 4 + . . . ,
2 r
r
λRK2DG(1) = c3 (rν)2 k 3 + c4 − 9(rν)3 k 4 + . . . , (8)
q
r
λRK3DG(1) = c4 + 3(rν)3 k 4 + . . . ,
q
[(rν)2 − 1]2
λHaDG(1) = c4 3rν(qν − 1) + k4 + . . . ,
q/r − rν
where c3 = −irh2 /6, c4 = −rh3 /72. The CFL number ν is defined by
the dimensionless frozen wave speed instead of the advection speed r, thus
ν ≡ 1∆t/h. These results show that the leading error of both HaDG(1) and
RK3DG(1) is fourth-order whereas the other two methods are third-order.
Thus DG(1) with M.O.L. can be a third-order method if third-order RK is
used, however, it requires three flux calculations at each cell face whereas
HaDG(1) requires two to achieve the same order. Note the vanishing of the
error in HaDG(1) for qν = rν = 1. Regarding stability, it was shown that
both RK2DG(1) and RK3DG(2) with the LxF flux are linearly unstable [6].
The stability limit of each method with upwind and Rusanov fluxes is as fol-
lows: RK2HR2, qν ≤ 1; RK2DG(1), qν ≤ 1/3; RK3DG(1), qν ≤ 0.409 [1].
For HaDG(1) we heuristically found that the linear stability limit is different
for different fluxes: HaDG(1) with upwind, rν ≤ 1; HaDG(1) with Rusanov,
ν ≤ 1/3.
Next, consider a 2 × 2 linear system of hyperbolic-relaxation equations;
u = [u, v]T , f = [v, u]T , and s = [0, ru − v]T in Eq.(1). This system has
“frozen” wave speeds ±1 when relaxation is weak ( 1); when the relaxation
dominates ( 1), it reduces to advection-diffusion equation, ∂t u + r∂x u ≈
(1 − r2 )∂xx u, with an “equilibrium” wave speed of r. For stability, |r| ≤ 1.
Previously, it was shown that the upwind flux applied to the model equations
reduces to the Rusanov flux when 1 [3]. A Von Neumann analysis in the
long wave-length limit shows the dominant dispersion/dissipation errors are
given by Eq.(8) with q = 1 where λexact = −irk − (1 − r2 )k 2 in this case. To
confirm the analysis, consider an initial-value problem on a periodic domain
with a harmonic initial condition u0 = v0 ≡ cos(2πx), x ∈ [0, 1]; the other
parameters r, are chosen so that the reduced equation becomes an advection-
dominated advection-diffusion equation (r = 1/2, = 10−5 ). The new method
is compared with an HR2 and a DG(1) method, both incorporating the IMEX-
64 Yoshifumi Suzuki and Bram van Leer
RK method [5]. The CFL conditions based on the unit wave speed are set as
νHaDG(1) , νRK2DG(1) = 0.3, and νRK2HR2 = 0.8. Figure 1 shows that the new
method is the least dissipative and dispersive of all, whereas the RK2HR2
method produces a completely inaccurate solution. Figure 2 shows a grid
convergence study of the solution at the final time tend = 300. Third-order
convergence is observed for the new method whereas RK2HR2 and RK2DG(1)
show second-order convergence in the L2 -norm.
0
10
10
0.6 :Exact Solution 2
0.4
−2
0.2 10
2
0
−0.2
−3
:HaDG(1)
−0.4 10 :RK2DG(1)
:RK2HR2
−0.6 3
−0.8 −4
10 1 2 3
−1 10 10 10
0 0.2 0.4 0.6 0.8 1
X Degrees of freedom
References
1. B. Cockburn and C. -W. Shu. Runge-Kutta discontinuous Galerkin methods for
convection-dominated problems. J. Sci. Comput., 16(3):173–261, 2001.
2. M. Delanaye. Quadratic reconstruction finite volume schemes on 3D arbitrary
unstructured polyhedral grids. AIAA Paper 1999-3259, 1999.
3. J.A.F. Hittinger, Y. Suzuki, and B. van Leer. Investigation of the discontinuous
Galerkin method for first-order PDE approaches to CFD. AIAA Paper 2005-
4989, 2005.
4. H. T. Huynh. An upwind moment scheme for conservation laws. In ICCFD 3,
Toronto, Canada, 2004.
5. L. Pareschi1 and G. Russo. Implicit-explicit Runge-Kutta schemes and appli-
cations to hyperbolic systems with relaxation. J. Sci. Comput., 25(1):129–155,
2005.
6. W. J. Rider and R. B. Lowrie. The use of classical Lax-Friedrichs Riemann
solvers with discontinuous Galerkin methods. Int. J. Numer. Meth. Fluids,
40:479–486, 2002.
7. B. van Leer. Towards the ultimate conservative difference scheme. IV. A new
approach to numerical convection. J. Comput. Phys., 23:276–299, 1977.
8. B. van Leer. Upwind and high-resolution methods for compressible flow: From
donor cell to residual-distribution schemes. Commum. Comput. Phys., 1(2):192–
206, 2006.
Very High Order, Non-Oscillatory Fluctuation
Distribution Schemes
1 Introduction
Fluctuation distribution schemes for approximating multidimensional systems
of conservation laws have developed to a stage where they can be used reliably
to produce accurate simulations of complex steady state fluid flow phenomena
using unstructured meshes [6]. These methods are often required to avoid
producing unphysical, numerically induced, oscillations in the flow field, and
this has so far restricted them to second order accuracy.
More recent research has led to higher order methods (for both steady state
and time-dependent flows) which use the fact that a k th order method can be
derived by evaluating the fluctuation exactly with respect to a (k −1)th degree
polynomial representation of the dependent variable and then distributing it
in a linearity preserving manner [4]. So far, two successful approaches have
been proposed for constructing this high order interpolant within each mesh
cell [4, 5]. A third alternative will be presented here.
As they stand, none of these three approaches can guarantee the absence
of spurious oscillations from the flow without the application of an additional
smoothing stage. This paper will briefly describe a technique which can be
combined with any of the above procedures to provide schemes which are both
higher than second order accurate and free of spurious oscillations. Brief re-
sults will be shown to demonstrate its effectiveness in approximating the scalar
advection equation on two-dimensional, unstructured, triangular meshes.
2 Fluctuation Splitting
Consider the two-dimensional scalar conservation law given by
ut + fx + gy = 0 or ut + λ · ∇u = 0 (1)
on a domain Ω, with u(x, y, t) = g(x, y, t) imposed on the inflow part of the
T
∂f ∂g
boundary ∂Ω. λ = ∂u , ∂u defines the advection velocity associated with
the conservation law (1). This equation has an associated fluctuation, assumed
here to be calculated over a triangular mesh cell 4 and given by
ZZ I
φ = − λ · ∇u dΩ = u λ · dn , (2)
4 ∂4
where ∆t is the time-step, Si is the area of the median dual cell corresponding
to node i, αij is the distribution coefficient which indicates the appropriate
proportion of the fluctuation φj to be sent from cell j to node i, and ∪4i
represents the
P set of cells with vertices at node i. Conservation is assured
as long as i∈4j αij = 1 , ∀j, where 4j represents the set of nodes at the
vertices of cell j.
[(αij )N ]+
(φji )P SI = P j N +
φj = (αij )P SI φj . (8)
k∈4j [(αk ) ]
This scheme can easily be shown to be conservative, positive for the time-step
given by (7), and linearity preserving, i.e. (αij )P SI is bounded so the order
of accuracy of the steady state scheme is one higher than the degree of the
polynomial used to represent u (in this case second order) [1, 3].
1
0.8
0.9
0.6 0.8
0.7
0.4
0.6
0.2 0.5
0.4
0 0.3
−1
−0.8 1
−0.6 0.2
−0.4
−0.2
0 0.5 0.1
0.2
0.4
0.6
0.8 0 0
1 −1 −0.8 −0.6 −0.4 −0.2 0 0.2 0.4 0.6 0.8 1
1
0.8
0.9
0.6 0.8
0.7
0.4
0.6
0.2 0.5
0.4
0 0.3
−1
−0.8 1
−0.6 0.2
−0.4
−0.2
0 0.5 0.1
0.2
0.4
0.6
0.8 0 0
1 −1 −0.8 −0.6 −0.4 −0.2 0 0.2 0.4 0.6 0.8 1
Fig. 1. The new scheme applied to circular advection of a square wave (top) and a
cosine-squared profile (bottom).
∗
∗ [(αij )N ]+ ∗
(φji )P SI = P j N∗ +
φLIM
j = (αij )P SI φLIM
j . (14)
k∈4j [(αk ) ]
As with the PSI scheme, the limiting procedure will never increase the magni-
tude of the distribution coefficients, so the positivity condition for the scheme
(13) is actually stronger than necessary.
4 Results
The two-dimensional scalar advection equation (1) is approximated, over the
domain [−1, 1] × [0, 1] with λ = (y, −x)T and u(x, y, 0) = g(x) for x ∈ [−1, 0]
and y = 0 (where g(x) is chosen to take a variety of forms to demonstrate the
properties of the schemes) [8].
Figure 1 shows the results obtained from the positive, high order scheme
derived from extending the stencil. Replacing g(x) with much smoother func-
tion, and carrying out a series of experiments on successively refined meshes
suggests an order of accuracy of 2.36 in the L1 norm and 2.23 in the L∞ norm.
Very similar results are obtained by applying the same limiting procedure to
the submesh reconstruction and gradient recovery schemes [8]. In all cases the
oscillations are removed completely and the results are significantly better
than those of the PSI scheme. All of the experiments have converged to their
steady states to machine accuracy.
70 M.E.Hubbard and N.Z.Mebrate
5 Conclusions
A new fluctuation distribution scheme has been presented which is demon-
strated to give higher than second order accuracy at the steady state for
the scalar advection equation without introducing any spurious oscillations.
The procedure used to impose positivity on the high order scheme is gener-
ally applicable, but is here combined with an approach which reconstructs a
quadratic interpolant within each mesh cell by extending the stencil to include
the neighbouring cells’ nodes, averaging across cell edges to give continuity
(and hence conservation). This approach has also been successfully applied to
the two-dimensional inviscid Burgers’ equation, but it remains to extend it to
three-dimensional problems, higher than third order accuracy (both concep-
tually straightforward) and nonlinear systems of equations (not so). Ongoing
research has shown that this limiting procedure can also be used to con-
struct a fully consistent, positive, high order fluctuation distribution scheme
for time-dependent situations.
References
1. Abgrall, R.: Toward the ultimate conservative scheme: Following the quest. J.
Comput. Phys., 167, 277–315 (2001)
2. Abgrall, R., Barth, T.J.: Residual distribution schemes for conservation laws via
adaptive quadrature. SIAM J. Sci. Comput., 24, 732–769 (2002)
3. Abgrall, R., Mezine, M.: Construction of second order accurate monotone and
stable residual distributive schemes: the unsteady case. J. Comput. Phys., 188,
16–55 (2003)
4. Abgrall, R., Roe, P.L.: High order fluctuation schemes on triangular meshes. J.
Sci. Comput., 19, 3–36 (2003)
5. Caraeni, D., Fuchs, L.: Compact third-order multidimensional upwind scheme
for Navier-Stokes simulations. Theor. Comp. Fluid Dyn., 15, 373–401 (2002)
6. Deconinck, H., Sermeus, K., Abgrall, R.: Status of multidimensional upwind
residual distribution schemes and applications in aeronautics. In: AIAA paper
2000–2328 (2000)
7. Deconinck, H., Struijs, R., Bourgois, G., Roe, P.L.: High resolution shock cap-
turing cell vertex advection schemes for unstructured grids. In: Computational
Fluid Dynamics, VKI-LS 1994-05. Von Karman Institute, Brussels (1994)
8. Hubbard, M.E.: Non-oscillatory third order fluctuation splitting schemes for
steady scalar conservation laws. Submitted to J. Comput. Phys.
High-order residual distribution : discontinuity
capturing crosswind dissipation and diffusion
g h
a b c
k 1
Fig. 1. P Lagrangian triangles with P conformal sub-triangulation
where ui = u(xi , yi ) and ψi the i-th basis function. We call ϕi the piecewise
linear basis function associated to i ∈ T defined on the P 1 conformal sub-
triangulation (see figure 1). Clearly, in the P 1 case ϕi = ψi .
The paper is divided in two parts. In the first one we show the general
construction of k + 1-th order schemes for hyperbolic scalar conservation laws.
The extension to (1) is considered in the second part.
Since βiLDA is uniformly bounded (w.r.t. mesh size h and solution uh ), the
LDA scheme respects the accuracy condition (5)
High-order monotone RDS and extension to advection-diffusion 73
The N scheme is monotone (in the sense described in section 2) and first
order. On can easily show that the N is obtained by adding to the LDA
scheme a crosswind (shock capturing) dissipation term [1] :
X −1 X
ΦN
i = Φi
LDA
+ dNi , dN
i = kj+ ki+ kj+ (ui − uj ) (9)
j∈T j∈T
Due to (9), this is equivalent to add to the LDA scheme a residual shock
capturing crosswind dissipation term :
ΦB
i = Φi
LDA
+ θdN i (11)
Limiting the distribution coefficient of the N scheme
X N,+(PSI scheme) :
N,+
ΦTi s = ΦPSI
i = βi
PSI Ts
Φ , β P SI
i = βi / βj (12)
j∈Ts
Ts
with βjN = ΦNi /Φi . The PSI scheme verifies both the monotonicity re-
quirement (ΦTi s × ΦN PSI
i ε0), and the accuracy condition (5) (βi bounded).
0
PSI(P2)
PSI (P3)
1.2 Exact
PSI (P1)
1 −5 b
0.8 1
0.6
−10
0.4
ln(error(u))
0.2
u
0 −15 a
-0.2
c
-0.4 1
−20
d
-0.6
e
-0.8
f
-1 −25
−5.5 −5 −4.5 −4 −3.5 −3 −2.5 −2
ln(h)
0 0.2 0.4 0.6 0.8
x
Fig. 2. Rotation of sin(16πx) and grid convergence (constant advection of cos(πx))
PSI (P3) schemes, while on the right a cut at y = 0.75 of the PSI (P2) and
PSI (P3) solutions. Some oscillations are present in very few mesh points after
the shock, however, their amplitude is small (below 15% local value of u).
1
p
1.6
0.9
1.4
PSI (P3)
0.8 Max : 1.75 1.2 PSI (P2)
Min : -0.76
0.7 1
0.6 0.8
0.5 0.6
u
y
0.4
0.4
0.2
0.3
0
0.2
-0.2
0.1
-0.4
0
0 0.5 1 0.7 0.8 0.9 1
x x
Fig. 3. Burgers equation : PSI (P3) solution contours and cut at y = 0.75 of the
PSI (P2) ans PSI (P3) solutions
High-order monotone RDS and extension to advection-diffusion 75
hand, Galerkin and stabilized Galerkin schemes perform very well. The prob-
lem is to build Petrov-Galerkin (PG) discretizations consistent with a given
RD scheme, and to combine the two discretizations to obtain uniformly (w.r.t.
h and P e) accurate approximations.
The solution proposed in [1] is the following. Given a RD scheme with dis-
tribution coefficients βjTs , such that ΦTj s = βjTs ΦTj s , build continuous piecewise
polynomialZ test functions respecting the consistency conditions
1
ωj dx dy = βjTs ∀ j ∈ T and ∀Ts ⊂ T (13)
|Ts |
Ts
In the last definition, the first line represents the RD scheme3 , while the second
line contains the PG discretization of the diffusive terms. As pointed out in [1],
(15) does not introduce any coupling between the discrete advection operator
and diffusion operators. Defining the following discrete Peclet number
|ΦTs |
P eh = P R
| ν∇uh · ∇ϕi dx dy|
j∈Ts Ts
Ts
Z
ξ(P e h
)(βiTs ΦTs − ΦC
i ) + ξ(P e h
)(3βjTs − 1) ν∇uh · ∇S Ts dx dy (16)
Ts
with ξ(·) continuous, and such that limx→0 ξ(x) = 0 and limx→∞ ξ(x) = 1.
Scheme (16) reduces to a Galerkin type high order approximation in diffusion
3
the central finite element contribution ΦC
i is added and subtracted for reason
which will be soon clear
76 N. Villedieu-Ligout, M. Ricchiuto, and H. Deconinck
−6 −8
−10
−8
b
−12
−10
aa
c
−14
bb b
ln(error(u))
−12 cc
d
−16 a
−14
aa b
−18
−16
−20
h
q ln(error(u) = 3.18 ln(h) − 0.63
−18
Numerical error of u with LDA* −22 i
ln(error(u)) = 2.13 ln(h) −2.99
k
Numerical error of u with LDA
−20 −24
−5.5 −5 −4.5 −4 −3.5 −3 −2.5 −2 −6 −5.5 −5 −4.5 −4 −3.5 −3 −2.5 −2
Pe 0.3125 ln(h) 5 e f ln(error(u)) g
References
1. Ricchiuto, M., Villedieu, N., Abgrall, R., Deconinck, H.: High-order residual
distribution schemes : discontinuity capturing crosswind dissipation and exten-
sion to advection-diffusion, VKI LS on Higher Order Discretization Methods for
Computational Physics, Von Karman Institute for Fluid Dynamics (2005)
2. Abgrall, R. : Very high order residual distribution methods, VKI LS-CFD, (2005)
3. De Palma, P. : Implicit third-order-accurate residal distribution schemes for
unsteady hyperbolic problems, VKI LS -CFD (2005)
4. Nishikawa, H. : Higher order discretization of diffusion terms in residual distri-
bution methods, VKI LS-CFD, (2005)
5. Abgrall, R., Roe, P.L.: High order fluctuation schemes on triangular meshes. J.
Sci. Comput., 19(3), 3–36 (2003)
High-Order Fluctuation-Splitting Schemes for
Advection-Diffusion Equations
1 Introduction
2 Fluctuation-Splitting Schemes
Consider solving conservation laws of the form
ut + fx + gy = 0 (2)
∗
This work was supported under NASA URETI award NCC 3989 funded by NASA
Marshall and Glenn centers and by the USAF.
where {Tj } is a set of triangles that share node j, Mj is the median dual
cell area, and φTj = βjT φT : βjT is the distribution coefficient that assigns the
fraction of the fluctuation sent to node j in triangle T . If βjT is bounded, zero
fluctuation implies no updates in the solution: the scheme preserves polyno-
mial solution ( depends on the quadrature used) on arbitrary unstructured
grids; this is called the residual property and is part of the reason for the
reduced mesh sensitivity of FS schemes. In this work, we are interested only
in steady-state solutions and so we march in time simply using the forward
Euler time integration until solutions do not change.
For the advection-diffusion equation where (f, g) = (au − νux , bu − νuy ),
the fluctuation φT is the sum of two parts, advection φTa and diffusion φTd .
The former should be distributed with upwinding while the latter should be
distributed isotropically. But they should not be distributed separately, or
the residual property is lost and more importantly the accuracy of the scheme
reduces to 1st-order [2]. In this work, we therefore consider only the total
fluctuation, and distribute it with a combined upwind-isotropic distribution
coefficient proposed in [2] which becomes upwind (the LDA scheme) in the
advection limit and isotropic in the diffusion limit. This distribution coefficient
is bounded, and therefore the residual property is guaranteed: the accuracy
of the FS scheme is now determined by the accuracy of the fluctuation [4].
3 Second-Order Fluctuations
For conservation
Z laws, the fluctuation
I can be evaluated as a contour integral
− (fx + gy ) dxdy = − (f dy − g dx) (5)
T ∂T
This is convenient because only the solution variation along the element
boundary is relevant to the accuracy of Rthe fluctuation. Consider a single
2
term over an edge, say from node 1 to 2, 1 f dy. For second-order accuracy,
it suffices to use the trapezoidal rule
Z 2
f1 + f2
f dy = (y2 − y1 ) + O(h3 ) (6)
1 2
Fluctuation-Splitting Schemes for Advection-Diffusion 79
(u x) 2 , (u y) 2
2
um
T m
T
1
ui , pi , qi 3 (u x) 1 , (u y) 1
This formula is 2nd-order accurate provided the nodal fluxes f1 and f2 are
also second-order accurate. Note that for the advection-diffusion equation the
flux now involves the solution gradient ux which is not available at nodes and
must be evaluated with sufficient accuracy. In other words, we must follow
the principle that the quadrature formula and the fluxes at the quadrature
points must be of comparable accuracy, in the same spirit of other high-order
methods [5, 6, 7]. This has not been stressed so much in FS methods for
hyperbolic problems, where the solution values are stored at the quadrature
points, and the condition is automatically met, but becomes important for
problems such as advection-diffusion. The task becomes one of evaluating the
diffusive fluxes (gradients) with sufficient accuracy at the quadrature points.
In this paper we compare, theoretically and experimentally, two approaches.
One of these is interpolation, and the other is solution of a first-order system
having the gradient values as unknowns. The first of these is explicit, but the
second is implicit.
One approach is to directly obtain the gradients by reconstruction, for
example by a least-squares(LS) quadratic reconstruction which is 2nd-order
on general triangular grids. Then, the fluctuation will have a residual property,
preserving exact linear solutions. Note that although the diffusion part is exact
for quadratic functions but it is only 2nd-order because both the quadrature
and the recovered gradients are 2nd-order (being exact for quadratic does not
imply 3rd-order accuracy for terms involving 2nd-order derivatives!). This is
probably the simplest approach to achieve 2nd-order accuracy, but requires
the stencil to be enlarged (Figure 1), even more so for higher-order as at least
cubic reconstruction is needed beyond O(h2 ). In the rest of the paper, we refer
to this approach as the direct gradient evaluation(DGE) approach.
An alternative is the FOS approach in which we compute the nodal gra-
dients as additional unknowns, solving the equivalent first-order system
ut + fx + gy = 0, p − ux = 0, q − uy = 0 (7)
where (f, g) = (au − νp, bu − νq). Fluctuations are defined for the addi-
tional equations with piecewise linear variation of u, p, q, and then distributed
isotropically to update p, q; this minimizes those fluctuations in an L2 norm.
Values of p and q computed this way are used to evaluate the diffusive part of
the flux. This allows us to keep the stencil compact (Figure 2). The resulting
80 Hiroaki Nishikawa and Philip Roe
scheme has the residual property that exact linear solutions are preserved.
Observe that this FOS approach does not require any solution reconstruction
and the scheme remains compact.
Incidentally, it is possible and interesting to interpret the FOS approach
as a reconstruction method. In fact, the method is solving a globally coupled
system for p and q in an iterative manner. Interestingly enough, if we invoke
‘mass-lumping’ to remove the coupling, we end up with the familiar Green-
Gauss formula at every node. A similar observation is made in [8] where a
global system is derived by the Galerkin method. Certainly, the mass-lumping
yields a simple explicit formula for gradients, but in FS schemes we iterate
toward a steady-state anyway, and so we may as well iterate for gradients
along the way. Naturally, without mass-lumping, gradients are more accurate.
4 High-Order Fluctuations
For higher-order accuracy, we introduce a virtual node, m, at the midpoint
over the edge, and evaluate the fluctuation by Simpson’s rule (Figure 3),
Z 2
f1 + 4fm + f2
f dy = (y2 − y1 ) + O(h5 ) (8)
1 6
This integrates cubic polynomials exactly and therefore can be 4th-order accu-
rate. To match the accuracy, we now need 4th-order accurate fluxes at nodes
including the midpoint. This requires not only gradients at all points for the
diffusive flux but also the midpoint value um for the advective part of fm ; all
with 4th or at least 3rd-order accuracy to give a high-order scheme.
To estimate um with high-order, we follow Caraeni and Fuchs[9]. First
recover the gradients ((ux )i , (uy )i ) at nodes, and then evaluate um by the
Hermite cubic interpolation along the edge (Figure 3).
u1 + u2 1
um = − {(us )2 − (us )1 } (9)
2 8
where (us )i ≡ (ux )i (x2 − x1 ) + (uy )i (y2 − y1 ). This formula is 3rd-order with
quadratic gradient reconstruction and 4th-order with cubic reconstruction.
This completes the advective part of the high-order fluctuation.
For the diffusive part, the DGE approach requires cubic/quartic recon-
struction to ensure 3rd/4th-order accuracy. The reconstruction stencil will
then become very large, but can be reduced somewhat by utilizing the mid-
point values um already recovered for the advective part. For example, a node
shared by 4 triangles has now 12 neighbor nodes (instead of 4) which enables
cubic reconstruction. But this means that we must prepare two reconstruction
algorithms: one for estimating um in the advective flux; the other for evalu-
ating the diffusive flux. Also, we would need to store the midpoint values.
In contrast, the FOS approach can be extended to high-order without
such elaboration. All we need are high-order accurate gradient variables (p, q)
at nodes including the midpoints. The midpoint values can be estimated in
exactly the same way as um . Then, with midpoint values available, we can
Fluctuation-Splitting Schemes for Advection-Diffusion 81
define high-order fluctuations for the slope equations with piecewise quadratic
variation of u, p, q, and distribute these to compute p, q at the original
nodes [3]. Clearly, this scheme has the residual property: preserving exact
quadratic/cubic solutions with quadratic/cubic gradient recovery. Observe
now that essentially we need only a single algorithm that estimates midpoint
values via gradient recovery and the Hermite interpolation over edges, and we
only need to call this for 3 times to estimate um , pm and qm . We point out
also that this scheme can be made even simpler by substituting (9) into (8) to
eliminate the midpoint values, resulting an element-wise high-order correction
scheme [3], i.e. no need to store the midpoint values in practice.
5 Results
We show results for the advection-diffusion
√ equation with the exact solution
u = − cos(πη)exp(0.5ξ(1 − 1 + 4π 2 ν 2 )/ν) where ξ = ax + by, η = bx − ay,
ν = 0.1 and (a, b) = (7, 4) in the square domain [0, 1] × [0, 1]. This is a
case where the advective and the dffusive terms are equally important and a
scheme obtained by adding the Galerkin scheme to an advection scheme indeed
loses its accuracy. We tested the DGE schemes with quadratic/cubic/quartic
LS gradient reconstruction, and the FOS schemes with no reconstruction
and quadratic/cubic LS reconstruction, using the inverse distance-squared
weighting for all reconstructions[10]. We determined the order of convergence
through a series of computations using three different unstructured grids hav-
ing 441, 1681, and 6561 nodes. Figures 4 and 5 show the results (◦: 2nd; ∗: 3rd;
: 4th). First, 2nd, 3rd, 4th-order accuracy for both types of schemes were
confirmed. Second, the convergence of the 3rd-order schemes looks somewhat
better than expected. This could be because of the combination of the 4th-
order quadrature formula and 3rd-order fluxes. Finally, and remarkably, the
FOS schemes generate much lower error levels than the DGE schemes (note
that the figures are equally scaled); this may be because the gradients in the
flux are more accurate in the FOS approach as mentioned in Secion 2.
6 Concluding Remarks
The design principle of high-order methods equally applies to FS schemes: the
accuracy is determined by the quadrature formula and the fluxes at quadra-
ture points, and this is particularly important for the diffusive flux. With this
in mind, we compared two different strategies (FOS and DGE) for solving the
advection-diffusion equation, and conclude that the FOS approach demon-
strate several advantages over the DGE approach. First, the FOS schemes are
much more compact and accurate than the DGE schemes: no gradient recovery
required for 2nd-order, and quadratic/cubic reconstruction for 3rd/4th-order.
Second, high-order FOS schemes require only a single additional algorithm
that estimates midpoint values via gradient recovery (and this can be done
without storing the midpoint values[3]) whereas high-order DGE schemes add
82 Hiroaki Nishikawa and Philip Roe
-2 -2
2.0
-3 -3
Log(error)
Log(error) 1.9
-4 3.4 -4
3.3
-5 -5
3.9 3.8
-2 -1 0 -2 -1 0
Log(h) Log(h)
much more complications to coding. Also, the FOS approach provides a solid
base for developing P2 -schemes (midpint values are computed as unknowns);
preliminary studies are reported in [3]. Finally, we remark that in the FOS
approach, only the gradients in the diffusive flux need to be stored (not gradi-
ents of all variables). In the Navier-Stokes equations, this means that we need
to store 3 viscous stresses and 2 heat fluxes in 2D, and 6 viscous stresses and
3 heat fluxes in 3D.
References
1. High Energy Flow Solver Synthesis (LAURA Benchmarks: Cylinder) at
http://hefss.larc.nasa.gov/
2. H. Nishikawa and P. L. Roe, On High-Order Fluctuation-Splitting Schemes for
Navier-Stokes Equations, in Third International Conference in CFD, Toronto,
2004.
3. H. Nishikawa, Higher-Order Discretization of Diffusion Terms in Residual-
Distribution Methods, in 34th VKI CFD Lecture Series Very-High Order Dis-
cretization Methods, 2005-01.
4. R. Abgrall, JCP, 167, pp. 277-315, 2001.
5. T. J. Barth and P. O. Frederickson, AIAA Paper 90-0013, 1990.
6. B. Cockburn and C. W. Shu, JCP, 141, pp. 199-224, 1998.
7. Y. Liu and M. Vinokur and Z. J. Wang, IJNMF, 216, pp.780-801, 2006.
8. R. Löhner, CFD Via Unstructured Grids: Trends and Applications, in Frontiers
of Computational Fluid Dynamics, John Wiley & Son Ltd., pp. 117-133, 1994.
9. D. Caraeni and L. Fuchs, Theoretical and Computational Fluid Dynamics, 15,
pp.373-401, 2002.
10. T. J. Barth, Aspects of Unstructured Grids and Finite-Volume Solvers for
the Euler and Navier-Stokes Equations, in Unstructured Grid Methods for
Advection-Dominated Flows, pp. 6-1-6-11, Neuilly Sur Seine France, 1992.
Construction of Higher Order Residual
Distribution Schemes
In all what follows, we assume that the solution is represented by a pth order
Lagrange continuous interpolant : ΦE is evaluated by an accurate enough
quadrature formula.
ΦE → 0 ⇒ ΦE
i → 0.
Now, the idea is to find a mapping F that transform a low order distribution
{βiL }i∈E into a high order one {βiH }i∈E (we forget the super script E since
there is no ambiguity).
max(0, βiL )
βiH = F(βiL ) = X
max(0, βiL )
i
to solve (5), we cannot reach a steady state in general, and this condition is
essential for accuracy.
We notice that we have this problem for non upwind schemes only. In fact,
the structure of the matrix A is not good when there exists a node i that does
not receive any part of the residual of the surrounding cells. In others words,
the scheme can locally be downwind, which is known to be unstable. The fact
that it does not blow up is related to the positivity of the scheme. Details can
be found in [1].
The solution presented here is an extension of what has been proposed in
[1]. The idea is to modify the scheme with a SUPG like term. Denoting by
βiH Φ the basic high order scheme, and by Φ∗i the stabilized one, we set:
Z
Φ∗i = βiH Φ + hθ (λ∇ψi )(λ∇u)
E
This additional term has the following properties (k being the order of the
polynomial approximation):
• it is streamline dissipative,
• it keeps the conservation property,
• it is O(hk+1 ),
The sensor may depend on the structure of the solution (the ideal case would
be θ ' 1 in the smooth part and θ ' 0 at the discontinuities. The quality of the
results is in practice unsensitive to the actual θ. The theoretical drawback of
this proecedure is that we lose the theorical positivity of the scheme. However,
in practical applications, no oscillation are observed.
3 Numerical Results
In this section, we present three different computations: two for scalar prob-
lems, one for the Euler equations. All the computations are done on triangular
unstructured grids. We denote by LF the low order Lax-Friedrichs scheme,
LLF the limited LF and LsLF the limited-stabilised LF .
• Figure 1: the scalar problem is the pure advection of a sin squared. On
the left, one can see that the LLF is unable to converge to a steady state.
With the LsLF we can reach the zero machine. The effect on the isolines
of the solution can be seen on the middle (LLF ) and the right (LsLF ).
• Figure 2: shows the advantage of going higher order for convection prob-
lems. The problem is a solid rotation of a sin. We compare here the use of
P 1 (left) and P 2 (right) interpolant (with the same number of points). We
can see the smooth and oscillatory signal is significally better represented
on the right.
Construction of Higher Order RD Schemes 87
-2
-4
-6
-8
residual
-10
-12
-14
-16
stabilised scheme
without stabilisation
-18
0 2000 4000 6000 8000 10000 12000 14000 16000 18000 20000
iterations
LFLS_P1.desc LFLS_P2.desc
Fig. 2. Results on solid rotation: 2nd order scheme (left), 3nd order scheme
(right)
4 Future Work
The results of the LsLF scheme are very satisfactory. Without any tuning
parameters, it is able to handle difficult problems remaining stable and con-
vergent. Future extensions are planned: unsteady problems, 3D computations
and viscous flows.
88 Rémi Abgrall and Cédric Tavé
choc_2D_O1.desc choc_2D_O2.desc
Fig. 3. Axisymmetric shock at Mach 2: Mach number, 1st order scheme (left),
2nd order scheme (right)
Fig. 4. 4-states Riemann problem: Density, 40x40 grid (left) and 200x200 grid
(right)
References
1. R. Abgrall: Essentially non oscillatory Residual Distribution schemes for hyper-
bolic problems, J. Comput. Phys., 2006, vol. 214
2. R. Abgrall, M. Mezine: Construction of second-order accurate monotone and sta-
ble residual distribution schemes for steady problems, JCP 195 (2004) 474-507
3. M. Ricchiuto, N. Villedieu, R. Abgrall and H. Deconinck: High-order residual
ditribution schemes: discontinuity capturing crosswind dissipation and extension
to advection diffusion, J. Comput. Applied math, 2006, accepted
4. R. Struijs: A multi-dimensional upwind discretization method for the Euler equa-
tions on unstructured grids, PhD, Von Karman Institute for Fluid Dynamics,
1994
5. A. Kurganov and E. Tadmor, Numer Meth in PDEs, vol 18, 2002.
6. R.Abgrall and P.L. Roe, High-order fluctuation schemes on triangular meshes, J.
Sci. Comput, vol 19, 2003.
Stable and convergent residual distribution for
time-dependent conservation laws
1 Generalities
We consider the discretization of the time dependent hyperbolic problem
∂u
+ ∇ · F(u) = 0 on Ω × [0, tf ] ⊂ R2 × R+ (1)
∂t
on unstructured grids. We present residual distribution (RD) schemes which
(i) give non-oscillatory solutions, (ii) are second order accurate by construc-
tion, and (iii) lead to well-posed algebraic problems, that is, they ultimately
lead to linear systems A x = y, with A invertible. How to construct nonlinear
RD satisfying (i) and (ii) is known for some time [3]. However, it is the satis-
faction of (iii) that ensures that a (unique) discrete solution exists, and that
second order of accuracy is actually obtained in practice (convergence).
Due to boundedness of the βj∗ s defined by (4), and due to the properties of
the mapping, the resulting scheme is formally second order accurate, and it
does have a strong monotonicity preserving (viz. L∞ -stable) character [2, 3].
The whole procedure leads to a nonlinear algebraic system of type (3).
Let us suppose to be approximating a smooth solution, such that we can
linearize both the nonlinear part of (1) (or equivalently (2)), and also (3).
To this linearized version of (3) we associate the matrix engendered by the
component of the algebraic problem corresponding to the (linearized) operator
M(·) in (2). Let us denote by M∗h this matrix. Our (linearized) system reads
M∗h un+1 = Bh∗ (5)
where un+1 contains the unknown nodal values, and in Bh∗ we have dumped
whatever does not depend on un+1 .
3 T
for clarity, we drop the super-script
Stable and convergent RD for unsteady problems 91
It is even possible that λii = 0, for some i : the mapping, in general, weak-
ens the diagonally dominant character of Mh , eventually leading to a ill-
conditioned system matrix M∗h .
Another way to see it is the following. The whole construction is based on
∗
the constraint φM j × βj φε0. Upwinding is not included in the process : it is
likely that the application of the mapping might lead, locally in an element,
to a down-wind discretization, known to have poor stability.
Even so, the code never blows-up due to the L∞ −stable character of the
scheme. However, if a numerical output is obtained, the iterative convergence
in the solution of (3) is often poor, and the result might be affected by spurious
modes not identified/dumped (lack of dissipation/lack of uniqueness).
Let us now go back to the full nonlinear case. As mentioned before, the
convergence problems are relevant mainly when approximating smooth solu-
tions. When dealing with discontinuities the method has enough numerical
viscosity to converge relatively well, both in the inner iterations and with
mesh refinement. An heuristic justification of this fact is the following [2].
In elements containing a singularity, the element residual φT and the first
order monotone residuals φM j scale according to the same power of the mesh
size h. In particular, in two dimensions, simple arguments lead to
φT ≈ hk∆FkT and φM
j ≈ hk∆ukT
being k∆FkT and k∆ukT reference values for the norms of flux and variable
differences over T . The two scaling are easily obtained from the definition of
φT , and from the one of positive first order dissipation terms. If T contains a
singularity, what we can say is that k∆FkT and k∆ukT are bounded, which
j = O(h) and φ = O(h). Since φj = O(φ ), then φj /φj = O(1).
∗ ∗
leads to φM T T M
This means that, across a discontinuity, the mapping is likely to preserve more
the algebraic structure of the system obtained with the low order method.
92 Mario Ricchiuto and Rémi Abgrall,
4
even though, to be rigorous, we should evaluate the extra term in entropy variables
Stable and convergent RD for unsteady problems 93
c c c
a a a
b b b
line01 line01
101.662 101.662
a c
a
99.2033 99.2033
96.7448 96.7448
p
p
94.2863 94.2863
b b
a
91.8278 91.8278
0 0.25 0.5 0.75 1 0 0.25 0.5 0.75 1
b
x x
Fig. 1. Vortex advection. Top (left to right): Exact, LLF and LLFs (h = 1/40).
Bottom (left to right): LLFs (h = 1/80), cut along centerline for LLF and LLFs
4 Computational results
We discuss some results obtained solving the Euler equations with the nonlin-
ear discretization constructed starting from the Lax-Frederich’s scheme com-
bined with Crank-Nicholson time integration :
|T | δun+1 ψin + ψin+1
Z
1 X
φLF
i = + , ψ i = ∇·F dx+αT (ui −uj )(10)
3 ∆tn+1 2 3
T j∈T
We call LLF scheme the one obtained by applying (4) to (10). The scheme
obtained by adding the stabilization term (8) is referred to as LLFs.
We first consider a smooth problem, consisting of the advection of a vortex
in inviscid flow (see [6] for details). On fig. 1 we report, on the top row, the
exact solution (pressure contours), the solution of the LLF scheme and of the
LLFs scheme on a coarse grid (h = 1/40). On the bottom row we report
the solution of the LLFs scheme on the finer mesh (h = 1/80), and the 1D
pressure distributions through the vortex core. The top row pictures clearly
show the spurious modes not dumped by the LLF scheme, and the effectiveness
of stabilization in suppressing these modes. The contour plot on the bottom
row, as well as the 1D line plots, confirm this observation and demonstrate
the truly second order of convergence of the stabilized scheme (as shown by
the reduction in the L∞ pressure error).
Then on fig. 2 we report the results on the well known problem of the
Mach 3 wind tunnel with a forward facing step [10]. The result show the
monotonicity preserving character of the scheme, and the effectiveness of the
definition of the solution monitor (9).
94 Mario Ricchiuto and Rémi Abgrall,
5.50353
4.30635
3.10917
1.91199
0.714813
0 0.75 1.5 2.25 3
Fig. 2. Mach 3 forward facing step. Density contours (left) and distribution at
y = 0.5 (right). Results at time t = 4 computed with the LLFs scheme
5 Conclusions
We discussed the construction of non-oscillatory RD schemes which are stable
and give genuine second order of accuracy in practical applications. The ap-
proach proposed opens the way to a new class of schemes which need very few
matrix operations. Hence, they are more efficient, while retaining the advan-
tages of the RD approach. In particular, the extension to arbitrary accuracy
is quite natural and it is under way.
References
1. Abgrall, R., Andrianov, N., Mezine, M.: Towards very high-order accurate
schemes for unsteady convection problems on unstructured meshes. Int. J. Nu-
mer. Meth. Fluids, 47, 679–691 (2005)
2. Abgrall, R.: Essentially non-oscillatory Residual Distribution schemes for hy-
perbolic problems. J. Comput. Phys., 214, 773–808 (2006)
3. Abgrall, R., Mezine, M.: Construction of second-order accurate monotone and
stable Residual Distribution schemes for steady flow problems. J. Comput.
Phys., 188, 16–55 (2003)
4. Bochev, P.B., Gunzburger, M.D., Shadid, J.N.: Stability of the SUPG finite
element method for transient advection-diffusion problems. Comp. Meth. Appl.
Mech. Engrg., 193, 2301–2323 (2004)
5. Berman, A., Plemmons, R.J.: Nonnegative Matrices in the Mathematical Sci-
ences. Academic Press, 1979
6. Dobevs, J., Deconinck, H.: Second Order Blended Multidimensional Residual
Distribution Scheme for Steady and Unsteady Computations. J. Comput. Appl.
Math., accepted (2006)
7. Ricchiuto, M., Abgrall, R., Deconinck, H.: Construction of very high order
Residual Distribution schemes for unsteady advection: preliminary results. VKI
LS 2003-05 (2005)
8. Ricchiuto, M., Abgrall, R., Deconinck, H.: Application of conservative
residual distribution schemes to the solution of the shallow water equa-
tions on unstructured meshes. J. Comput. Phys., (2006). Available online.
doi:10.1016/j.jcp.2006.06.024.
9. Ricchiuto, M., Csı́k, Á., Deconinck, H.: Residual Distribution for general time
dependent conservation laws. J. Comput. Phys., 209, 249–289 (2005)
10. Woodward, P.R., Colella, P.: The numerical simulation of two-dimensional flows
with strong shocks. J. Comput. Phys., 54, 115–173 (1984)
An ALE Formulation of the Multidimensional
Residual Distribution Scheme for
Computations on Moving Meshes∗
1 Introduction
The Residual Distribution (RD) schemes for unsteady computations are sub-
ject of on-going intensive research [1]. Until recently, the RD schemes for
unsteady transonic flow simulations had modest success due to difficulties to
deal with large time-steps. As the solution, a new Bx scheme was introduced
in [3]. The scheme is based on a blend of the non-oscillatory first order lumped
mass matrix formulation of the N scheme and the second order full mass ma-
trix formulation of the LDA scheme using a simple pressure switch. In smooth
parts of the flow linear stability is sufficient and the LDA scheme is active. In
shocks, the N scheme is activated to prevent oscillations in the solution. The
blending parameter is smooth, giving superior iterative convergence proper-
ties and scaled such that second order of accuracy is achieved. The topic of
this paper is the extension of the Bx scheme for the computation on moving
meshes with application to transonic flutter computations.
2 ALE Formulation
We define the ALE mapping which for each t ∈ I associates a point Y of ref-
erence configuration Ω0 to a point x on the current domain configuration Ωt ,
At : Ω0 ⊂ Rd → Ωt ⊂ Rd , x(Y, t) = At (Y). The ALE mapping At is chosen
sufficiently smooth and invertible with nonzero determinant of Jacobian JAt .
A domain velocity w(x, t) is defined as the time derivative of x for constant
Y. We start from the conservative ALE formulation of the Euler equations in
d spatial dimensions
∗
This research was partially supported by Research plan of MSMT no. 6840770003.
3 Numerical Scheme
The problem is solved on mesh τ h consisting of simplex elements {E}. The
class of unsteady RD schemes can be formulated as a relaxation procedure in
dual time. The solution at time level n + 1 is obtained for m → ∞
X
un+1,m+1
i = un+1,m
i − σi φE
i , (3)
E∈Di
αn+1 µ(E n+1 )un+1 + αn µ(E n )uni + αn−1 µ(E n−1 )un−1
φE,N
i = i i
+
d+1
n+1 αn+1 µ(E n+1 ) + αn µ(E n ) + αn−1 µ(E n−1 )
∆t ki+ (un+1 − ūn+1
i − uin ) ,
d+1
(5)
with coefficients
1 + 2τ τ2 tn+1 − tn
αn+1 = , αn = −1 − τ, αn−1 = , τ= n . (7)
1+τ 1+τ t − tn−1
Each of the terms of (5) corresponds to the respective term given in (4). This
scheme is non-oscillatory, but only first order accurate.
We next consider the linear second order LDA scheme, derived from the
analogy with the Petrov-Galerkin (PG) finite element method (FEM) for
scalar problems. The solution and the mesh velocity are approximated by the
linear Galerkin shape functions ψi from the current domain configuration [4].
The equation is multiplied by the PG test function ϕE i = ψi +βi
LDA
−1/(d+1)
from the current domain configuration, with distribution coefficient βiLDA =
ki+ N , then it is integrated over the domain. The time derivative is replaced
by the 3BDF scheme. The scheme is again written as (3), with the element
contribution
X
φE,LDA
i = (αn+1 µ(E n+1 )un+1
j + αn µ(E n )unj + αn−1 µ(E n−1 )ujn−1 )m̃E
ij +
j∈E
n+1
X X
βiLDA kj uj −(αn+1 µ(E n+1 )+αn µ(E n )+αn−1 µ(E n−1 )) un+1
j m̃E
ij ,
j j∈E
(8)
where m̃E
R
ij = E ϕi ψj dx/µ(E) is the element contribution to the mass matrix
divided by the volume of E. The scheme for scalar problems is extended to
the system of equations simply by replacing the distribution coefficient by
the distribution matrix βjLDA with formally the same definition. The element
contribution to the mass matrix becomes
E 1 1 + δij 1 LDA
m̃ij = − I + βj . (9)
d+1 d+2 d+1
The upwind matrices ki are evaluated in a suitably chosen mean state to
assure conservativity of the scheme [7]. Note, that unlike some other ALE FE
schemes, this scheme does not contain a time derivative of the mass matrix.
Both N and LDA schemes respect the discrete Geometric Conservation Law
by construction.
Finally, the nonlinear Bx scheme takes advantage of the monotone behav-
ior of the N scheme in the shocks and the high accuracy of the LDA with
the mass matrix in the smooth regions of flow. The nodal contribution of the
Bx scheme is given by φBx N
i = θφi + (1 − θ)φi
LDA
. The element-wise blending
coefficient is defined as [3] θ = min(1, sc h), sc = ( ∂p
2 +
∂t + ∇x p · v) /δpv , where
v is the velocity vector of the flow, p is the static pressure, h the diameter of
the element and δpv is a product of the characteristic pressure and velocity
in the domain. The blending coefficient θ is O(h) in the smooth parts of the
flow, giving second order of accuracy, and O(1) in shocks, effectively switching
back to the non-oscillatory N scheme.
98 Jirı́ Dobes and Herman Deconinck
4 Numerical Results
For the comparison, the cell centered finite volume (FV) scheme with linear
reconstruction, Roe’s Riemann solver and Barth’s limiter is considered [2].
The time is discretized with the 3BDF scheme and the mesh movement is
treated using the Scheme A [6].
The first test case involves the well known test case AGARD CT 5 of the
NACA 0012 airfoil, which is sinusoidally pitching around its a quarter chord.
The free stream Mach number is 0.755. The airfoil performs a pitching motion
α = 2.51◦ sin(2kt) + 0.016◦ , k = ωc/2u∞ = 0.0814, where c is the chord, u∞
is the free-stream velocity and ω the frequency. The problem is solved on
an unstructured mesh consisting of 5711 nodes and 11153 elements with 206
nodes around the airfoil.
0.4
0.3
cl
0.2
0.1
! -2 -1 1
-0.1
-0.2
-0.3
-0 4
Fig. 1. NACA 0012 test case. Continuous line is the Bx scheme, discountinous line
is the FV scheme. Left and middle: isolines of the pressure at time t = 116. Right:
dependence of the lift coefficient on the angle of incidence.
LDA FV
1.5
FV fine Bx
0.5 FV
0.008
!z’
0.006
0.004 0
0.002
0 -0.5
!y
-0.002
-0.004
-1
-0.006 N
LDA
-0.008
0 0.02 0.04 0.06 0.08 0.1
-0.01 t
0 0.05 0.1 0.15 0.2 0.25 0.3
t
assumption was used. The flow conditions are given by p∞ = 25714 Pa and
ρ∞ = 0.4 kg/m3 . The critical Mach number Ma cr ∞ that is, the lowest free
stream Mach number for which an unstable aero-elastic mode of the panel
appears, is given in reference [8]. Using a theoretical method the authors
get Ma cr cr
∞ ≈ 2.27 and using their numerical scheme Ma ∞ ≈ 2.23, what they
consider an “excellent agreement”.
The elastic panel is discretized with 60 × 2 elements. The computational
domain is formed by a half-circle of radius R = 5. The mesh consists of
3451 nodes and 6722 triangular elements, giving 50 elements along the panel.
Reference FV computation is performed on a regular quadrilateral mesh of
300 × 100 elements with 100 elements along the panel (referred as “fine”).
The integral of the deflection of the panel for Ma = 2.2 is plotted in figure 2.
The neutral response was correctly reproduced. Although the LDA scheme is
only linearly stable, it is able to capture weak shock waves in a non-oscillating
manner. The nonlinear Bx scheme gives similar results as the LDA scheme,
which are very different from the first order N scheme. Finally, the transonic
flutter of the AGARD 455.6 wing (“solid model”) is considered [5]. The elastic
wing was discretized using 350 tri-quadratic elements. The CFD mesh consists
of 22k nodes and 118k tetrahedral elements. The neutral response regime was
chosen, which is characterized by the flutter speed index of 0.5214 and the free
stream Mach number Ma ∞ = 0.92 with flow medium Freon-12. One period
was divided in 120 time-steps. The integral of the wing velocity is plotted in
100 Jirı́ Dobes and Herman Deconinck
5 Conclusions
A novel ALE formulation of the residual distribution schemes for the compu-
tation on moving meshes is presented. The nonlinear Bx scheme consists of a
blend of the LDA and N schemes. The LDA scheme is formulated with finite
element type mass matrix. The N scheme is a simple extension of the steady
version of the scheme. Results are compared with a FV method. The method
is validated on forced oscillating airfoil, where it shows clearly better accuracy
than the FV scheme, and on two typical cases of aeroelastic problems.
References
1. Abgrall R., Mezine M.: Construction of second order accurate monotone and
stable residual distribution schemes for unsteady flow problems. JCP, 188:16–
55 (2003).
2. Barth T.J.: Aspects of unstructured grids and finite-volume solvers for the Euler
and Navier-Stokes equations. In: 25th CFD VKI LS (1994).
3. Dobevs J., Deconinck H.: Second order blended multidimensional residual dis-
tribution scheme for steady and unsteady computations. J. of Comp. and Appl.
Math. (JCAM) (2005), accepted for publication.
4. Dobevs J., Deconinck H.: Extension of linear second order residual distribution
schemes for computations on moving meshes. In: J. Pvrı́hoda, K. Kozel (editors),
Proceedings of “Topical Problems of Fluid Mechanics 2006”, 27–30, IT CAS CZ
(2006), ISBN 80-85918-98-6.
5. E. Carson Yates J.: AGARD standard aeroelastic configurations for dynamic
response. Technical Memorandum 100492, NASA (1987).
6. Koobus B., Farhat C.: Second-order time-accurate and geometrically conserva-
tive implicit schemes for flow computations on unstructured dynamic meshes.
Comput. Methods Appl. Mech. Engrg., 170(1–2):103–129 (1999).
7. Michler C., Sterck H.D., Deconinck H.: An arbitrary Lagrangian Eulerian formu-
lation for residual distribution schemes on moving grids. Computers and Fluids,
32(1):59–71 (2003).
8. S. Piperno, Farhat C.: Partitioned procedures for the transient solution of cou-
pled aeroelastic problems part II: Energy transfer analysis and three-dimensional
applications. Comput. Meths. Appl. Mech. Engrg., 190(24):3147–3170 (2001).
Solution of the steady Euler equations using
Fluctuation Splitting schemes on quadrilateral
elements
1 Introduction
Fluctuation Splitting (FS) schemes, also known as Residual Distribution
schemes, are a well-established technique for solving hyperbolic problems on
triangular meshes, see [1] and the references therein. As such, they are not
adequate for solving high Reynolds number viscous flows, where an accu-
rate resolution of the boundary layer requires high aspect ratio quadrilateral
cells. Therefore, there has been a renewed interest in the application of this
methodology on quadrilateral meshes. In [2] the authors presented a thor-
ough analysis of the properties of FS schemes when applied to quadrilateral
elements, with reference to the solution of the linear advection and advection-
diffusion equations. Using a multidimensional Fourier analysis, it is shown
that linearity preserving (LP) schemes lack dissipation for the high wavenum-
ber modes, which may hamper convergence to the steady-state solution, or
even lead to instability problems. A remedy is employed, especially needed
for convection dominated problems, namely, adding an SUPG based artificial
dissipation term to the LP signals, which is effective in stabilizing high wave-
number modes, even if it does not ensure monotonicity. This work extends the
approach of [2] to the solution of the Euler equations and applies it with suc-
cess to well documented subsonic, transonic, and supersonic two-dimensional
inviscid flows.
2 Numerical method
The Euler equations in quasi-linear form are written as:
∂Q ∂Q ∂Q ∂Q
+ A · ∇Q = +A +B = 0, (1)
∂t ∂t ∂x ∂y
K+ = R Λ+ L , K− = R Λ− L , (4)
+ −
where Λ and Λ contain the positive and negative eigenvalues, respectively,
Λ± = 1/2 (Λ ± |Λ|). The fluctuation over an element is computed as:
Z Z I
E
Φ = ∇· F dΩ = A · ∇Q dΩ = F · n d` . (5)
ΩE ΩE ∂ΩE
ΦE
Z
∂Ni ∂Ni
=h A +B dΩ
ΩE ∂x ∂y |ΩE |
Z Z
h ∂Ni ∂Ni
= Â dΩ + B̂ dΩ ΦE . (13)
|ΩE | ΩE ∂x ΩE ∂y
3 Results
The first test-case considered is the well-known supersonic jet interaction test-
case. It consists in the interaction between two horizontal, supersonic jets,
which are suddenly brought into contact. The upper stream is a high Mach
number, low pressure flow, while the lower stream is a lower Mach number and
higher pressure flow. The problem is solved on the square domain [0, 1]×[0, 1],
with the prescribed inlet boundary conditions:
M = 4, v = 0 , p = 0.25 , ρ = γ/2 for x = 0, y ∈ [1/2, 1] ,
M = 2.4 , v = 0 , p = 1 , ρ=γ for x = 0, y ∈ [0, 1/2] .
Computations have been performed on an 80 × 80 Cartesian grid and on an
unstructured grid of triangles with spacing h = 1/64 on boundaries. Figure 1
shows the density iso-lines obtained using the monotone N scheme on the
quadrilateral grid and the limited LDA scheme on both the quadrilateral grid
and the triangular one. The results show that the limited scheme resolves all
waves, namely, the shock, contact wave and expansion fan, much better than
the monotone N scheme; however, the residual is not fully converged because
of the limiter.
Then, the steady subsonic flow inside a channel with a cosine-shaped bump
on the bottom wall has been considered [7]. The channel has dimensions
[0, 4] × [0, 1] with a straight upper wall and a cosine shaped lower wall with
coordinates:
0 ≤ x ≤ 1 y = 0,
1 ≤ x ≤ 3 y = (1 − cos[(x − 1) π])/10 ,
3 ≤ x ≤ 4 y = 0.
The isentropic exit Mach number is equal to M∞ = 0.5. At inlet points, the
total pressure, total enthalpy and flow angle are imposed, whereas the static
FS on quadrilateral elements 105
to validate the method, the transonic inviscid flow through a cascade of gas
turbine rotor blades (VKI LS–59) has been considered [8]. The blade profile
and the grid used are shown in figure 2a, together with a close-up view of the
leading and trailing edges. The grid is an hybrid multiblock one. An “O” type
structured block (561 × 17 elements) surrounds the blade profile and an un-
structured block made up of 5929 triangles covers the outer domain. Thus, the
References
1. H. Deconinck and R. Abgrall: VKI 34th CFD Lecture Series, Higher order dis-
cretization methods (2006).
2. P. De Palma, G. Pascazio, D. T. Rubino, M. Napolitano: J. Comput. Phys., in
press, (2006).
3. A. Csı́k, M. Ricchiuto, H. Deconinck: J. Comput. Phys., 179, 286 (2002).
4. R. Struijs, H. Deconinck, P.L. Roe: VKI LS 1991-01 (1991).
5. P. De Palma, G. Pascazio, G. Rossiello, M. Napolitano: J. Comput. Phys., 208,
1 (2005).
6. R. Abgrall and F. Marpeau: J. Scient. Comput., in press (2006).
7. M. Napolitano, A. Bonfiglioli, P. Cinnella, P. De Palma, G. Pascazio, in Frontiers
of computational fluid dynamics 2002, World Scientific, pp. 113-127 (2002).
8. R. Kiock, F. Lethaus, N. C. Baines, C. H. Sieverding: ASME J. Eng. Gas Tur-
bines Power, 108, 277 (1986).
A Residual-Based Compact Scheme
for All-Speed Flows on Unstructured Grids
2 Scheme formulation
2.1 Space discretization
where w = (ρ, ρu, ρv, ρw, ρE)T is the vector of conserved variables and F is
the physical flux tensor. The finite volume approximation of (1) on a general
unstructured grid reads :
dwi 1 X
=− Hi,k |Γi,k | (2)
dt |Ωi |
k
with wi the cell-averaged state in cell i of volume |Ωi | identified here with the
cell-center value (since only second-order accurate methods will be considered)
and Hi,k the numerical flux approximating F · n through face k of cell i, with
area |Γi,k |.
Inviscid case
Let us first assume F = F E (w) where F E contains the physical fluxes defining
the Euler equations. The numerical flux of a conventional upwind scheme will
typically read Hi,k = HE (wi,k
L R
, wi,k ; ni,k ) where HE depends on the specific
scheme retained for approximating F E (Roe, AUSM+, HLLC . . .) and wi,k L
,
R
wi,k denote the reconstructed states on the left and right sides of the interface
L/R
Γi,k , with unit normal vector ni,k pointing outward cell i. States wi,k are
computed as [6] :
L R
wi,k = wi + φi ∇wi · rik , wi,k = wo(i,k) + φo(i,k) ∇wo(i,k) · ro(i,k)k
where index o(i, k) refers to the cell sharing the interface Γi,k with cell i, rik
(resp. ro(i,k)k ) denotes the vector extending from the cell center i (resp. o(i, k))
to the center of the interface Γi,k . The gradient ∇wi is evaluated at the cell
center i using a least-square formula applied on a prescribed spatial support
and φi is a slope limiter, computed according to Barth and Jespersen’s formula
as modified by Venkatakrishnan [7]. Alternatively, the numerical flux of the
second-order RBC scheme reads :
1
Hi,k = (F E (win ) · ni,k + F E (wo(i,k) n
) · ni,k ) − di,k (3)
2
The dissipation flux vector di,k added to the centered evaluation of the normal
flux is based on a discrete integral form of the steady residual r = ∇ · F E
computed at the center of face Γi,k :
−−−−−−→
||Ci Co(i,k) ||
Z
1
di,k = Φi,k r dΩ (4)
2 |Ωi,k | Ωi,k
where Ci and Co(i,k) denote the cell centers on each side of the interface Γi,k
and Φi,k is a matrix coefficient of order O(1) designed so as to ensure the
dissipation of the scheme in the multi-D case (see [1] for more details). For a
2D problem, let us introduce f E , g E the physical fluxes associated with the
first and second space direction and AE , B E their respective Jacobian matrix.
Let us denote :
A RBC Scheme for All-Speed Flows on Unstructured Grids 109
(l) −1
(J⊥ )i,k = AE E
i,k (nx )i,k + Bi,k (ny )i,k = (T⊥ )i,k · Diag((λ⊥ )i,k ) · (T⊥ )i,k
(l) −1
(Jk )i,k = AE E
i,k (tx )i,k + Bi,k (ty )i,k = (Tk )i,k · Diag((λk )i,k ) · (Tk )i,k
where t is a unit vector tangent to face Γi,k . Matrix Φ has the same eigenvec-
(l)
tors as J⊥ , that is Φi,k = (T⊥ )i,k · Diag(φi,k ) · (T⊥−1 )i,k and its eigenvalues are
defined from the eigenvalues of J⊥ and Jk by :
(l)
(l) (l) |Γi,k | |(λ⊥ )i,k |
φi,k = sgn((λ⊥ )i,k ) min(1, −−−−−−→ (m(J )) ) (5)
||Ci Co(i,k) || k i,k
(l)
where m(Jk ) = minl (|λk |). The eigenvalues and eigenvectors at face Γi,k
are typically evaluated using a Roe-averaged state. In practice, the volume
integral over the cell Ωi,k enclosing interface Γi,k , which is used to compute
the residual at the face, is expressed as a surface integral over ∂Ωi,k . The
cell Ωi,k is built on the cell centers Ci , Co(i,k) and the vertices of face Γi,k ,
1 2
denoted Ni,k , Ni,k , so that the integral on ∂Ωi,k is easily computed applying
the trapezoidal rule on each segment of this closed contour and using on one
hand directly available cell-center states and on the other hand node values
estimated from area-weighted averages over the neighboring cell centers. Note
that, since this integral is multiplied by the distance between the cell-centers
located on each side of the face, such a first-order evaluation of the node values
is sufficient to ensure a second-order dissipation.
Viscous case
When solving the Navier-Stokes equations, F in (1) takes the form F E (w) −
F V (w, ∇w) where F V contains the viscous fluxes. The extension of a standard
upwind scheme to the Navier-Stokes equations is straightforward since the
inviscid and viscous numerical fluxes remain decoupled; the full numerical
flux takes the form :
1
Hi,k = HE (wi,k
L R
, wi,k ; ni,k ) − (F⊥V (wNi,k
1 , ∇wN 1 ) + F
V
2 , ∇wN 2 ))
⊥ (wNi,k
2 i,k i,k
where F⊥V = F V ·n. The node values are simply evaluated using area-weighted
averages over the neighboring cell centers while the node gradients are com-
puted with a least-square approach. The RBC numerical flux retains an ex-
pression similar to (3), with a purely centered approximation for both the
inviscid and viscous contribution completed by a residual-based dissipation
now built on a residual that includes the inviscid and viscous flux balance :
1 1
Hi,k = ((F⊥E )ni + (F⊥E )no(i,k) ) − ((F⊥V )Ni,k
1 + (F⊥V )Ni,k
2 ) − di,k (6)
2 2
where F⊥E = F E · n and di,k is still given by (4) with Φ unchanged w.r.t. the
inviscid case but with r = ∇ · (F E − F V ). Since the residual contains both
inviscid and viscous fluxes, evaluating the dissipation flux requires to compute
viscous fluxes and hence gradients of w not only at the grid nodes, which is
110 Christophe Corre, Alberto Beccantini, and Thibaud Kloczko
needed anyway for the consistent centered flux, but also at cell-centers, which
is done using again a least-square formula.
where the numerical flux formula H(i) retained in the implicit stage is differ-
ent from the one used in R since it does not affect the accuracy of the method
provided convergence on m is reached. In the present study, the implicit treat-
ment is the matrix-free method described in [10], which preserves a reduced
cost per iteration even when low-Mach matrix preconditioning is used.
3 Numerical results
A few results demonstrating the RB approach offers an interesting alternative
to existing treatments on unstructured grids are now provided. When coupled
with the proposed matrix-free implicit stage and applied to the transonic flow
over a staggered biplane, the second-order RBC scheme converges to steady-
state faster than the second-order limited AUSM+ scheme and yields a sharp
capturing of the shock appearing in the divergent of the nozzle geometry
created by the two airfoils (see Fig. 1). The RBC results compare also well
with those given by the low-Mach HLLC scheme when computing the low-
subsonic flow over a multi-element airfoil (see Fig. 2). When applied to the
computation of the unsteady low-subsonic laminar flow over the NACA0015
airfoil at high incidence, the RBC scheme yields average values of 0.70 and
0.322 respectively for lift and drag coefficients (see Fig. 3), in good agreement
with C̄L = 0.67, C̄D = 0.304 found in [11] using a structured incompressible
flow solver. Current developments of the RBC approach on unstructured grid
include extension to 3D flows as well as to general equations of state, allowing
to deal for instance with dense gas flows [12].
0
10
10-1
-2
10 AUSM+
RBC
-3
10
Residual
-4
10
-5
10
10-6
-7
10
10-8
1000 2000 3000 4000
Iterations
Fig. 1. Flow at M∞ = 0.7 over a staggered biplane. Convergence history for AUSM+
and RBC (CF L = 106 ); RBC pressure contours on a 17170-elements triangular grid.
112 Christophe Corre, Alberto Beccantini, and Thibaud Kloczko
5
HLLC (slat)
HLLC (main body)
4 HLLC (flap)
RBC
3
-Cp
2
-1
Fig. 3. Flow at M∞ = 0.2, Re∞,c = 1000, α = 16◦ over a NACA0015 airfoil. Lift
and drag evolution using RBC in a 13966-elements triangular grid; instantaneous
streamlines at t = 50.
References
1. Lerat, A., Corre, C.: Comp. & Fluids, 31,639–661 (2002)
2. Lerat, A., Corre, C.: J. Comput. Phys., 170, 642–675 (2001)
3. Corre, C., Hanss, G., Lerat, A.: Comp. & Fluids, 34, 561–580 (2005)
4. Lerat A., Corre, C.:34th VKI Lectures Series on CFD (2005)
5. Corre, C., Lerat, A.: In CFD 2004, Groth & Zingg (eds), Springer (2006)
6. Barth, T.J., Jespersen, D.C.: AIAA Paper 89-0366 (1989)
7. Venkatakrishnan, V.: Comp. & Fluids, 18, 191–204 (1990)
8. Turkel, E. : Applied Numerical Mathematics, 12, 257–284 (1993)
9. Luo, H. Baum, J.D.,Löhner, R. : AIAA J., 43, 1160–1166 (2005)
10. Beccantini, A. et al : In CFD 2004, Groth & Zingg (eds), Springer (2006)
11. Fang, K.C. : PhD thesis, U. of Cincinnati (2000)
12. Congedo, P.M., Cinnella, P., Corre, C. : AIAA J. (2006), accepted.
Vorticity Preserving Scheme
for Unsteady Compressible Flows
1 Introduction
ω̃ n+1 − ω̃ n
δ1 µ2 δ2 µ1 n
+ u0 . + v0 . µ1 µ2 ω̃ =
∆t δx δy j+ 12 ,k+ 12
(7)
δ 2 µ2 δ 2 µ2
∆t δ1 µ1 δ2 µ2
u20 1 22 . + 2u0 v0 . + v02 2 21 . ω̃ n
2 δx δxδy δy j+ 1 ,k+ 1
2 2
Similarly to the LWN scheme for the state vector w, this discrete equation
for ω̃ includes some dissipation, so that the LWN is not vorticity preserving,
except for pure acoustics (u0 = v0 = 0).
vanishes and the scheme is vorticity preserving. Scheme (9) will be refered to as
the RBV scheme. Note that the improvement in vorticity transport provided
by RBV over LWN is mainly due to the use of the complete residual (i. e.
including wt ) in the numerical dissipation of the scheme.
4 Numerical applications
Scheme (9) has been formulated on curvilinear meshes in a finite-volume
framework and applied to various compressible flows problems using the non
linear Euler equations on smooth and irregular meshes. For applications on
non uniform grids, a tenth-order filter [4] is added to the dual-time subitera-
tion scheme.
Fig. 1. Vorticity field at t = 100 on an irregular Cartesian mesh. (a) Exact solution,
(b) RBV, (c) RBV with filtering
Vorticity Preserving Scheme for Unsteady Compressible Flows 117
The RBV scheme has has been applied to the case of unsteady head-on inter-
action between a Scully vortex and a NACA 0012 airfoil, at M∞ = 0.5 and
zero incidence, for which experimental results are available [5]. This model
problem is representative of the parallel interaction between a helicopter ro-
tor blade and the vortex emitted at the tip of a preceding blade, which is at
the origin of the blade vortex interaction (BVI) noise. Figure 2 shows that
the use of the RBV scheme leads to grid-convergence for this complex highly
unsteady problem. The RBV scheme preserves the vortex in its displacement
on a long distance before and during the interaction with the profile (see
Fig. 3). Note that supersonic conditions are reached locally when the vortex
impinges the airfoil. Numerical tests have shown that the RBV scheme gives a
converged solution on a coarser grid than with classical schemes of same order
of accuracy. The vortex advection with very little dissipation is necessary to
capture the acoustics in the near field, as illustrated on Fig. 4.
2 2
1
0
1
-1
CP
CP
-2
-3
0
-4
-5
-1 -6
3.3 3.4 3.5 3.6 3.7 3.8 3.3 3.4 3.5 3.6 3.7 3.8
t t
References
1. K. W. Morton and P. L. Roe: Vorticity-Preserving Lax-Wendroff-Type Scheme
for the System Wave Equation, SIAM J. Sci. Comput., 23:170–192, 2001.
2. R. H. Ni: A Multiple-Grid Scheme for Solving the Euler Equations, AIAA Jour-
nal, 20:1565–1571, 1982.
3. A. Lerat and C. Corre: A Residual-Based Compact Scheme for the Compressible
Navier-Stokes Equations. J. Comput. Phys., 170:642–675, 2001.
4. M. R. Visbal and D. V. Gaitonde: On the Use of High-Order Finite-Difference
Schemes on Curvilinear and Deforming Meshes. J. Comput. Phys., 181:155–185,
2002.
5. S. Lee and D. Bershader: Head-On Parallel Blade-Vortex Interaction, AIAA
Journal, 32:16–22, 1994.
118 Falissard, Lerat and Sidès
(a)
(b)
(c)
Fig. 3. Grid convergence of the vorticity field at successive times. Points in vortex
core : (a) 4 pts, (b) 8 pts, (c) 16 pts.
Summary. The Spectral Difference (SD) method has been developed recently by
[1] for the wave equations on unstructured triangular grids and further developed
by [2] for 2D Euler equations. In this paper, the SD method is extended to solve vis-
cous flow governed by compressible Navier-Stokes equations in both 2D and 3D, for
unstructured triangular and hexahedral grids, respectively. Some numerical results
are presented to demonstrate its capability.
1 Introduction
Two sets of nodal points within the computational element or cell are defined.
One set is named the solution points where the solution unknowns are placed.
The other set is called the flux points, where fluxes in all coordinate directions
are located. The solution and flux points for triangular and quadrilateral cells
†
Correspondence can be addressed to the first author. Tel: +1 515 294 1614; Fax:
+1 515 294 3262; Email: zjw@iastate.edu
are illustrated in Figure 1. The placement of solution and flux points within
a hexahedral cell is very similar to Figure 1a. If the solution points support a
Lagrangian interpolation of degree p, the flux points usually support a degree
p+1 Lagrangian polynomial.
Fig. 1. Placement of solution () and flux (◦) points for triangular and quadrilateral
elements
3 Numerical Results
3.1 Accuracy Study
Unsteady viscous flow over a circular cylinder was selected to demonstrate the
SD method for time-accurate viscous flow computations. Both a triangular
122 Z. J. Wang, Yuzhi Sun, C. Liang and Yen, Liu
grid and a hexahedral grid were used. Since the flow is two-dimensional, the
hex grid only has one cell in the span-wise direction. Both grids are shown in
Figure 2. The triangular grid has 3,516 cells, while the hex grid has 590 cells.
For both grids, the cylinder wall is approximated with piecewise quadratic
curves or patches. The inflow Mach number is 0.2.
The cl and cd histories computed with the triangular grid are shown in
Figure 3a. Based on the history, the vortex shedding frequency can be com-
puted, and then the Strouhal number. The computed Strouhal number from
the 2nd order SD simulation is 0.156, comparing with the experimentally
measured 0.148. The entropy contours are shown in Figure 3b. The Karman
vortex street is captured well. On the hex grid, both the 3rd and 4th order
Fig. 3. Flow over a cylinder using the second order SD scheme on the triangular
grid
Fig. 4. Vorticity magnitude contours for flow over a cylinder computed with the
4th order SD scheme on the hex grid
124 Z. J. Wang, Yuzhi Sun, C. Liang and Yen, Liu
4 Conclusions
In this paper, the spectral difference method is extended to the Navier-Stokes
equations on both triangular and hex grids. Two approaches were tested in
the computation of viscous fluxes, and both give similar results. Although the
2nd-order SD scheme is convergent, higher order SD schemes appear to have a
weak instability, and further investigations will be carried out to find the cause
and a remedy. The extension to viscous flow on hex grids is fully satisfactory.
High-order accuracy was numerically demonstrated using the Couette flow.
The SD schemes are then used to compute unsteady flow over a circular
cylinder, and the von Kamnan vortex street was captured, and the computed
Strouhal number agrees well with experimentally measured value. Currently
we are developing an efficient implicit/multigrid solution algorithm.
5 Acknowledgements
References
1. Y. Liu, M. Vinokur, M., and Z.J. Wang: Proceedings of the ICCFD-3, July 12-16,
2004, Toronto, Canada, Springer, 2004.
2. Z.J. Wang and Y. Liu: AIAA paper 2005-5112, 2005.
3. B. Cockburn and C.-W. Shu: J. Comput. Phys. 141, 199 - 224 (1998).
4. D.A. Kopriva: J. of Comput. Phys. 143, 125-158 (1998).
5. Z.J. Wang, L. Zhang and Y. Liu: J. Comput. Phys. 194, 716-741 (2004).
6. G. May and A. Jameson: AIAA Paper No. 2006-304, 2006.
7. B. Cockburn and C. -W. Shu: SIAM J. Numer. Anal. 35, 2440-2463 (1998).
8. C.H.K. Williamson: J. Fluid Mech. 206, 579 (1989).
Strictly Stable High Order Difference Methods
for the Compressible Euler and Navier-Stokes
Equations
Bernhard Müller
1 Introduction
High order finite difference methods have been constructed to be strictly stable
for linear hyperbolic and parabolic problems. The methods employ high order
central approximations in the interior and special boundary stencils to satisfy
the summation by parts (SBP) property leading to discrete energy estimates
[1][2][3][4][5].
Instead of standard central high order difference approximations, disper-
sion relation preserving schemes can be used in the interior to yield SBP
operators [6]. With entropy splitting, i.e. splitting of the nonlinear Euler and
Navier-Stokes equations into a conservative portion and a symmetric non-
conservative portion, generalized energy estimates can be derived to yield
strict stability [7][8].
Here, the perturbation form of the nonlinear Euler and Navier-Stokes equa-
tions has been employed to minimize cancellation errors at low Mach numbers
and to reduce to the linearized equations, if the nonlinear terms are neglected
[9][10][11]. The SBP operator with sixth order in the interior and third order
near the boundaries by B. Strand [2] is employed to approximate the spatial
first derivatives in the nonlinear Euler and Navier-Stokes equations. Second
derivatives are computed by applying the first order difference operator twice.
The explicit fourth order Runge-Kutta method is used for time integration.
High wave number oscillations are damped by a sixth order explicit filter.
Sound propagation of the Kirchhoff vortex, of rocket-launch noise before lift-
off and of point sources in the atmosphere has been simulated by solving the
nonlinear and linearized Euler equations [12][13][14][15][11][16][6]. The numer-
ical approach has recently been extended to the Navier-Stokes equations [17].
High accuracy is required for correctly computing not only sound propagation
over long distances and times but also sound generation due to viscous and
nonlinear effects.
The energy method shows us how to construct difference operators such that
high order and strict stability including the boundary conditions are guaran-
teed. As a model equation, we consider the convection diffusion equation with
homogeneous boundary conditions
ut + aux = buxx , 0 ≤ x ≤ 1, (2)
u(x, 0) = f (x) , 0 ≤ x ≤ 1, (3)
u(0, t) = u(1, t) = 0 , 0 ≤ t. (4)
where a and b are assumed to be constant and positive.
Strictly Stable High Order Difference Methods 127
Let vj = vj (t) denote the approximation of the exact solution of the con-
vection diffusion equation at grid point xj = jh with the grid spacing h = N1 .
We use the notation v = [v0 , v1 , ..., vN ]T . We define the discrete scalar product
and norm
(u, v)h = huT Hv , ||u||2h = (u, u)h , (5)
where H is a diagonal and positive definite matrix.
A difference operator Q satisfies the SBP property, if [2]
(u, Qv)h = uN vN − u0 v0 − (Qu, v)h . (6)
Thus, the discrete energy estimate for vt + aQv = bQQv becomes [17]
d
||v(t)||2h ≤ av02 + 2b[vN (Qv)N − v0 (Qv)0 ] . (7)
dt
The boundary conditions can be weakly imposed by the simultaneous approx-
imation term (SAT) [3] to get strict stability, i.e. continuous dependence on
the initial and boundary data and up to O(∆x) the same growth rate as the
continuous convection diffusion equation [5][18].
6
−h2 D (!)
0
0 0.5 1 1.5 2 2.5 3
! = 2" # h
Fig. 1. Damping of modes with nondimensional wave number ξ = k∆x of the exact
second derivative (-), the standard central fourth order difference operator for the
first derivative D1 applied twice (- -), the standard central fourth order difference
operator for the second derivative D2 (-·), the fourth order difference operator Q
derived for the self-adjoint form (·) and the fourth order operator L based on finite
elements with mass lumping (x) [18].
4 Results
propagation [17]. With the present approach, the direct numerical simulation
of aeolian tones at higher Reynolds numbers can be attacked.
! p = [−0.1:0.005:0.1]*M2.5 −4 ! p = [−0.1:0.005:0.1]*M2.5 −4
x 10 x 10
Acoustic pressure w.r.t. pmean, Re=150, M=0.1, 513 x 294 grid Acoustic pressure w.r.t. pmean, Re=150, M=0.2, 513 x 294 grid 6
100 1 100
4
80 80
60 0.5 60
2
40 40
0 0
20 20
0 0
y
y
−2
−0.5
−20 −20
−4
−40 −40
−1
−60 −60 −6
−80 −80
−1.5 −8
−100 −100
−100 −50 0 50 100 −100 −50 0 50 100
x x
p̃0
Fig. 2. Nondimensional instantaneous fluctuation pressure ρ∞ c 2
= [−0.1 : 0.005 :
∞
2.5
0.1]M∞ for M∞ = 0.1 (left) and M∞ = 0.2 (right) at Re∞ = 150.
5 Conclusions
Since strict stability is taken into account in the construction of summation
by parts (SBP) difference operators, those strictly stable high order difference
methods for the Euler and Navier-Stokes equations have been providing reli-
able numerical tools for sound generation and sound propagation.
Acknowledgements
The present work was performed while the author was associated with the
Division of Scientific Computing, Department of Information Technology, Up-
psala University, Sweden. The parallelization of the Navier-Stokes code by my
former colleague Jarmo Rantakokko is gratefully acknowledged.
References
1. B. Gustafsson, H.-O. Kreiss, J. Oliger: Time Dependent Problems and Differ-
ence Methods. John Wiley & Sons, New York (1995).
2. B. Strand: Summation by Parts for Finite Difference Approximations for d/dx.
J. Comput. Physics 110, 47–67 (1994).
3. M.H. Carpenter , D. Gottlieb, S. Abarbanel: Time-Stable Boundary Conditions
for Finite-Difference Schemes Solving Hyperbolic Systems: Methodology and
Application to High-Order Compact Schemes. J. Comput. Physics 111, 220-
236 (1994).
130 Bernhard Müller
1 Introduction
Recently, interests of many scientists and engineers are shifting from Reynolds-
averaged Navier-Stokes simulations to direct numerical simulations, large-
eddy simulations and aero-acoustic simulations. A major problem for these
simulations is that they are computationally very expensive because they re-
quire very high spatial resolution to guarantee accuracy of their results.
An approach to solve this problem is to use high-order schemes with less
grid points. According to Balsara[1], resolution of eighth-order scheme is four
times as high as second-order schemes in each dimension while it costs only
three times as much as second-order one. Thus, for three dimensional turbu-
lence problem, computational costs can be saved by factor of 44 /3(=83) with
eighth order scheme. Moreover, according to Shu[2], when high-order scheme
is implemented in multi-dimensional problem, finite difference schemes, that
can be constructed by dimension by dimension procedure, costs ten times as
little as finite element methods or finite volume methods which need multi
dimensional reconstruction.
Therefore various high order finite difference schemes were recently proposed.
These schemes include compact scheme proposed by Lele[3] and adapted to
generalized coordinate by Gaitonde and Visbal[4], weighted essential no os-
cillatory (WENO) scheme proposed by Jiang et al[5], and weighted compact
non-linear schemes (WCNS) proposed by Deng and Zhang[6, 7, 8]. WENO or
WCNS scheme has been developed to solve flow-fields including discontinuity
such as shock wave without numerical oscillation.
When finite difference schemes are adapted to generalized coordinate form,
capability of preserving uniform flow is important because noise from un-
preserved uniform flow hides very small oscillation, such as turbulent flow
structure or aero-acoustic wave. So far, for two-dimensional problem, preser-
vation of uniform flow has been achieved by using same stencils for both flow
WENO schemes use rth-order numerical flux f̃j+ 12 defined implicitly as fol-
lowing expression.
∂f 1
= f̃j+ 12 − f̃j− 12 + O(∆xr ) (1)
∂x j ∆x
It should be noted that this numerical flux is different from the physical
flux f̂j+1/2 . The numerical flux is constructed by weighted function of fluxes
+
of computational nodes. Moreover numerical flux is divided into f̂j+1/2 and
−
f̂j+1/2 . Weighted upwind-stencil is used for each divided flux. Thus in this
formulation, it is difficult to use same stencil for both flow computation and
metrics evaluation. Detailed procedure is explained in the reference of Deng
et al[6].
In this study, an explicit version of WCNS is investigated because it is more
efficient than the compact version[6]. This WCNS scheme uses rth-order phys-
ical flux defined explicitly by the following expression.
f̂j+ 12 = f (u(xj+ 12 )) + O(∆xr−1 ) (2)
This physical flux can be constructed by interpolation of flow variables such
as MUSCL procedure instead of finite difference scheme[6]. However, as is
physical flux, simple difference scheme achieves only second order as follows.
Uniform Flow Preserving Property of WENO and WCNS 133
∂f 1
= f̂j+k+ 12 − f̂j−k− 21 + O(∆x2 ) (3)
∂x j ∆x
Where the above expression is same as second order central difference on
twice finer computational grid. Thus to achieve higher order accuracy, sth-
order staggered finite difference schemes written as following expression is
necessary.
∂f 1 X
= bk f̂j+k+ 12 − f̂j−k− 21 + O(∆xr ) + O(∆xs ) (4)
∂x j ∆x
Since this high-order staggered finite difference scheme is linear scheme, it
seems that idea of Gaitonde and Visbal[4] can be used to preserve uniform
flow. Actual metric treatment to preserve uniform flow is explained as fol-
lows. In WCNS, cell-interface metric and high-order staggered finite differ-
ence schemes are used. This cell-interface metric is high-order-interpolated
from cell-node metric. If conservative form of metric terms is computed by
same cell-interface interpolation and same high-order staggered finite differ-
ence schemes, uniform flow seems to be preserved.
Table 1. Results of uniform flow preserving test.(n is number of the grid points)
P 2 1P 2
max(v) max(w) n1 w n w
WENO-G 8.75E-02 0.474989 9.98E-04 3.08E-02
WENO-C 6.82E-15 7.46E-15 3.92E-30 4.04E-30
WCNS-G1 1.75E-14 1.75E-14 8.00E-30 8.00E-30
WCNS-G2 1.13E-03 1.13E-03 7.49E-08 7.49E-08
5 Conclusions
Uniform flow and vortex preserving properties on wavy grids of four high-
order sschemes have been investigated while idea of Gaitonde and Visbal is
applied to WCNS scheme. The present results show that WCNS computed
in generalized coordinate system using conservative form of metric evaluated
by the same interpolation and the same finite difference schemes as flow com-
putation, has very good properties for both uniform flow and vortex flow
preservation.
Uniform Flow Preserving Property of WENO and WCNS 135
Fig. 1. Grid for unform flow preserving test (21x21x21, every 2 line is shown)
Fig. 2. Grid for vortex preserving test (41x41x21, every 2 line is shown)
References
1. Balsara, D. and Shu, C.W.: Journal of Computational Physics, 160, pp. 405-452
(2000)
2. Shu, C.W.: International Journal of Computational Fluid Dynamics (2003)
3. Lele, S. K.: , Journal of Computational Physics, 103, pp.16-44 (1992)
4. Gaitonde, D. V. and Visbal, M. R.: AIAA Paper 99-0557 (1999)
5. Jiang, G.-S. and Shu, C.-W.: Journal of Computational Physics, 126, pp. 200-
212 (1996)
6. Deng, X. G. and Zhang, H.: , Journal of Computational Physics, 165, pp. 22-44
(2000)
7. Deng, X. G. and Mao, M.: , Computational Fluid Dybnamics Journal, 13, pp.
173-180 (2004)
8. Deng, X. G., Liu, X. and Zhang, H.: , AIAA paper, 2005-5246, (2005)
9. Shu C.-W. and Osher, S.: , Journal of Computational Physics, 77, pp. 439-471
(1988)
10. Shu C.-W. and Osher, S.: , Journal of Computational Physics, 83, pp.32-78
(1989)
Implementation of an Enhanced Flux
Formulation for Unsteady Navier-Stokes
Solutions
1 Introduction
2 Computational Formulation
2.1 Preconditioned Equations
For unsteady flows, the governing equations are written in the so-called dual-
time form [1].
∂Qp ∂Q ∂Ei ∂Vi
Γp + + =H+ (1)
∂τ ∂t ∂xi ∂xi
where the τ represents the pseudo-time derivative, introduced purely for nu-
merical purposes. The terms comprising this pseudo-time term are:
ρ0p
p 0 ρT
Qp = uj Γp = uj ρ0p ρδjk u j ρT
T h0 ρ0p − (1 − ρhp ) ρuk ρhT + h0 ρT
In the above matrix, h0 is the total enthalpy and the term ρ0p represents
the scaled pseudo-property that is responsible for proper conditioning of the
system. The definition of this term is given by:
1 ρT (1 − ρhp )
ρ0p = 2
− (2)
Vp ρhT
where:
k1 V
Vp = min max V, , k2 V · Str , c (3)
Remin
Remin is the minimum cell-Reynolds number, Str is the physical Strouhal
number and k1 and k2 are scaling constants.
The enhanced flux formulation consists in replacing the offending matrix dis-
sipation term with one that possesses the correct scaling properties. The cor-
responding flux becomes:
1 α
Ẽ = (EL + ER ) − Γp |Λ|(QpR − QpL ) (5)
2 2
where Λ = diag(σ[Γp−1 Ap ], uδkk , u). In other words, the spectral radius of
the preconditioned Jacobian is used only in the dissipation terms of the con-
tinuity equation, while the standard convective velocity scale is used in the
momentum and energy equations. Thus, the offending dissipation terms in
the momentum and energy equations are effectively removed in this formula-
tion, while the proper definition is retained in the continuity equation. The
enhanced formulation thus insures optimal accuracy for steady, viscous or
unsteady computations.
3 Results
3.1 Taylor Problem
To test the artificial dissipation control, we show results from the numerical
simulation of the Taylor problem. The Taylor problem represents an analytical
solution of the incompressible Navier-Stokes equations for an infinite array of
counter-rotating vortices (see Fig. 1), whose amplitudes decay with time:
1
u = − cos x·sin y·e−2νt , v = sin x·cos y·e−2νt , p = − (cos 2x+cos 2y)·e−4νt (6)
4
Figure 2 shows the pseudo-time convergence for several sequential physical
time-steps using different schemes. It is evident that the scheme without any
preconditioning scaling or only inviscid scaling performs relatively poorly. In
particular, the inviscid choice appears to be very stiff for this choice of time-
scale (CF Lu = 0.1) and performs worse when the time-scale is made smaller.
The standard matrix dissipation scheme and the modified dissipation scheme
140 G. Xia, S. Sardeshmukh, V. Sankaran and C. L. Merkle
the best comparison with the exact solution and out-performs the traditional
matrix dissipation scheme by a significant margin. It is also noteworthy that
the no-preconditioning as well as the inviscid-preconditioning cases are both
very inaccurate, in part because of convergence issues and in part because of
the implications on the sizing of the dissipation terms.
Fig. 4. Comparison of DES predictions with experiments. Left side shows results
for different numerical schemes, right-side shows results for different values of CDES .
4 Summary
References
1. S. Venkateswaran and C. L. Merkle, ’Efficiency and Accuracy Issues in Navier-
Stokes Computations’, AIAA Paper 2000-2251, 2000.
2. Sankaran, V. and Merkle, C. L., ’Artificial Dissipation Control for Unsteady
Computations’, AIAA Paper 2003-3695, 2003.
3. Comte-Bellot, G., and Corrsin, S., ’Simple Eulerian Time Correlation of Full- and
Narrow-Band Velocity Signals in Grid-Generated Isotropic Turbulence’, Journal
of Fluid Mechanics, Vol. 48, pp. 273-337, 1971.
4. Strelets, M., ’Detached Eddy Simulation of Massively Separated Flows’, AIAA
Paper 2001-0879, 2001.
5. Xia, G., Sankaran, V., Li, D., and Merkle, C. L., ’Modeling of Turbulent Mixing
Layer dynamics in Ultra-High Pressure Flows’, AIAA Paper 2006-3729, 2006.
Computation of Eigenspaces of Hyperbolic
Systems
Key words: symmetric hyperbolic systems, method for calculating invariant sub-
spaces, calculation of eigenvalues and eigenfunctions
1 Introduction
2 Problem Statement
For the Maxwell and the acoustics equations, the relations on the vertical
characteristics have, respectively, the form
∂ ∂ ∂
div(E) = 0, div(µH) = 0 and rot U = 0.
∂t ∂t ∂t
To avoid the difficulties related to the null space of the operator D, the
system (1) is often reduced to the second-order equations, which have no ver-
tical characteristic. For instance, the acoustics equations (5) can be reduced
to the wave equation. However, this transition may cause difficulties in set-
ting up boundary conditions, writing a scheme, etc., in the case of variable
coefficients describing the medium. Note that hyperbolic formulation is more
natural for application of theory and numerical methods and is of great inter-
est for problems in magnetohydrodynamics.
3 Computations
The problem of calculating invariant subspaces for differential operators is
not widespread in computational practice, especially for nonself-adjoint oper-
ators. The results presented below are obtained in numerical experiments. We
try to start directly from the first-order system (1) and develop a technique
that makes it possible to eliminate the influence of the zero frequency on the
numerical process. For simplicity we consider a model problem for the system
(5), though the numerical method discussed below is also applicable to the
more general cases. The model problem is the acoustics equations (5) with
the constant coefficients c0 = 1 and ρ0 = 1 in Ω = (0, π)2 with the condition
p = 0 imposed on the boundary ∂Ω.
For the differential operator D corresponding to this boundary value prob-
lem and acting on a function u = (u, v, p) according to the rule Du =
−(px , py , ux + uy ), (x, y) ∈ Ω, we find a grid approximation of the invariant
subspace SD associated with the eigenvalues in the interval 0 < |λ| ≤ 4. From
this problem it is easy to find analytically all eigenvalues and eigenfunctions
[1], and one may verify dim(SD ) = 16.
To construct discrete approximations, we introduce a square grid with
a mesh size h. For simplicity of the presentation the approximations of D
are described for a strictly interior grid cell. We introduce the difference op-
erators D1 and D2 . The operator D2 is obtained by approximating all the
spatial derivatives at each cell by central divided differences. On the grids in-
troduced above, D2 inherits skew-symmetry inherent in D and approximates
D to O(h2 ) accuracy. The operator D1 is the upwind difference operator that
approximate D to O(h) accuracy. Both operators are supplemented by con-
servative boundary conditions.
Let us briefly describe the spectra of D, D2 , D1 for the model problem. In
the complex plane, the spectrum of D consists of the point (0, 0) of infinite
multiplicity and points on the √ imaginary axis. For this example the exact
eigenvalues λ are simply ±i m2 + n2 , 0 < m, n ∈ N and among them
146 S.K. Godunov, O.B. Feodoritova and V.T. Zhukov
exactly 16 values belong to the interval 0 < |λ| ≤ 4, but only 10 of them are
distinct. Remark also the zero eigenvalue is separated from the other ones.
The spectrum of D2 consists of the point (0, 0) of high multiplicity and
points on the imaginary axis. On the 32×32 grid, the spectrum consists of 3072
eigenvalues (counting multiplicities). There are 1122 eigenvalues in 0 < |λ| ≤ 4
of which only 96 are nonzero. The distinct eigenvalues form 9 conjugate pairs,
which are given in Table 1. The table also lists their multiplicities and the
number of pairs that approximate the eigenvalues of D.
The remaining eigenvalues are of grid origin. They are associated with os-
cillating grid modes. Moreover, such modes are also present in the eigenspaces
corresponding to the eigenvalues of D2 in 0 < |λ| ≤ 4. It is practically im-
possible to distinguish the desired approximations of the smooth part of the
spectrum due to presence of such spurious solutions among the eigenfunctions
of D2 . These spurious modes have no relation to the differential problem; actu-
ally, the presence of such modes speaks against D2 . However, it is the operator
D2 that inherits the skew-symmetry of D.
The spectrum of D1 has a rather complex structure. Figure 2a shows the
entire spectrum of D1 for the 32 × 32 grid. On the whole, the spectrum of D1
does not approximate that of D. In contrast to D, the operator D1 has a lot
of real eigenvalues ( O(Nh2 ) ) and part of then are located in the vicinity of
the origin. However, the spectrum of D1 includes some points that correspond
to the eigenvalues of D in 0 < |λ| ≤ 4. It is seen from Figure 2b that each
point in the spectrum of D marked by a circle is associated with a point in
the spectrum of D1 that has the same ordinate up to several digits. When the
grid is refined, these points in the spectrum of D1 approach the corresponding
points in the spectrum of D.
Let us also remark that the operator D1 can be used in an explicit time-
integration scheme with stability restriction τ = O(h) but D2 is an unpractical
operator in such computations.
Computation of Eigenspaces of Hyperbolic Systems 147
15 a 4 b
x
3 x
10
x
2 x
5 x
1
0 0 xxxxxxxxxxxxxxx x x xxxxxx x x
-1
-5 x
-2
x
-10 x
-3 x
x
-15 -4
-40 -30 -20 -10 0 -0.4 -0.2 0
first
-1
X second
10 Eigenpairs:
X third
-2 X
fourth
10
fifth
X
2
-3
10 X
1
X
-4
10
References
1. Godunov, S.K., Feodoritova, O.B. and Zhukov, V.T.: A method for computing
invariant subspaces of symmetric hyperbolic systems. Comput. Maths and Math.
Phys., 46(6), 971–982 (2006), Pleiades Publishing Inc., 2006
2. Godunov, S.K. and Selivanova, S.V.: Experiments Involving Resonance for Spec-
tral Analysis of Skew-Symmetric Operators. Sib. Zh. Vychisl. Mat., 9(2), 123–
136 (2006)
A Proposed Cure to the Carbuncle
Phenomenon
1 Introduction
Numerical shock prediction is a very important aspect of computing aerody-
namic flows, and shock capturing finite-volume methods are commonly used
to predict shocks in various situations with considerable success. However, it
seems that most shock capturing methods fall short in predicting very strong
shocks, which is a crucial element in designing hypersonic vehicles. Except for
a few notoriously diffusive schemes ([21] and the simplest version of Harten-
Lax-van Leer ([7]), most schemes4 exhibit some form of anomaly when pre-
dicting strong shocks. The commonest of these is the carbuncle phenomenon,
produced when computing a hypersonic flow past a blunt body such as a cir-
cular cylinder. Instead of having a smooth bow shock profile upstream of the
cylinder, the solution features a pair of oblique shocks ahead of the stagna-
tion region (Fig 2). Such a solution is actually a true solution of the Euler
equations, and can even be produced experimentally [2]. Many have proposed
cures to the carbuncle problem [15], [14], [12], [13], [11], [3], [20] but none are
universally accepted, and most papers begin by criticizing previous work.
The most common view follows Quirk [16] in supposing that the carbuncle
manifests a two-dimensional numerical instability of the Euler equations, often
4
There are proposals to adopt a hybrid of very dissipative and less dissipative
fluxes, deploying the former near the shock and the latter away from shock but
the basis of the switch is somewhat ad hoc. Furthermore, it is not clear how any
switch would work for complex problems like shock-boundary layer interactions
or shock-contact interactions.
2
3
Mach
20
2
18
1
16
1
14
YCoord
YCoord
12
0 0
10
8
-1
6
-1
4
-2
2
-3 0
-2
-4
-4 -3 -2 -1 0 1 2 3 4 -3 -2 -1 0 1 2
XCoord XCoord
Fig. 1. Quadrilateral grids on a circular Fig. 2. Results of the original Roe flux
cylinder with 80 x 160 cells. with M=20. Contours of Mach number
and velocity vectors.
2
Mach
20 0.75
18
1
16
14
YCoord
12
0.5
0
P
10
8
6 0.25
-1
4
0
-2 0
4 Numerical Results
The test case is the steady-state flow past a two dimensional cylinder. Various
grids and Mach numbers have been employed; a typical structured quadrilat-
eral grid is shown in Figure 1. In all cases reported here, we used a first
order explicit method with ν = 0.2 and the code was run until the residual
is of O(10)−8 . Second-order results were reported in [9]. All of our results for
structured grids indicate that results from the entropy-stable flux (for exam-
ple Figure 3) are quite free of the carbuncle phenomenon. However, the shock
profile is slightly broader with the introduction of a few intermediate cells
(Figure 4) and this may be the price we have to pay.
Previous studies [3] suggest that the carbuncle can be made to appear if the
cell aspect is increased, with the short dimension parallel to the shock. This
is consistent with a view that one needs numerical damping in the tangential
direction. However, our entropy-stable computations were not sensitive to this
aspect of the grid (compare Figures 5 and 6 with Figure 3).
Also, the profile around the shock exhibits spurious oscillations. This is
because our flux function guarantees only the correct sign of entropy pro-
duction but not necessarily the correct amount. The actual production is of
order δ 2 whereas the required production is of order δ 3 . To achieve mono-
tonicity, a flux function must generate ‘enough’ entropy production across
a shock. Flux functions that have this property are called entropy-consistent
(EC) fluxes [18]. We therefore modify the eigenvalues of equation (4) such that
|Λ̂| = |Λ̂| + α|[Λ]| [10]. For the choice of α = 0.2, the entropy-consistent flux
A Proposed Cure to the Carbuncle Phenomenon 153
0.6 0.6
0.4 0.4
0.2 0.2
YCoord
YCoord
0 0
-0.2 -0.2
-0.4 -0.4
-0.6 -0.6
Fig. 5. ES flux with 20 x 200 cells. Fig. 6. ES flux with 20 x 600 cells.
2
Mach
20 0.75
18
1
16
14
YCoord
12
0.5
0
P
10
6 0.25
-1
4
2
0
-2 0
References
1. Barth, T. In Num. Meth. for Gasdyn. Syst. On Unstructured Meshes. 1999.
2. Bogdonoff, S., and Vas, I. J. Aero. Sci 26 (1959), 584.
3. Dumbser, M., Moschetta, J., and Gressier, J. JCP 197 (2004), 647.
4. Edwards, J. In CFD Conference (2001), no. A01-31131, AIAA Conference.
5. Gnoffo, P. Ann. Rev. Fluid Mech. 31 (1999), 459–494.
6. Gnoffo, P., and White, A. No. 2004-2371, AIAA Thermophysics Conference.
7. Harten, A., Lax, P.D. and van Leer, B. SIAM Review (1983)
8. Hughes, T. Comp. Meth. App. Mech. Eng. 54 (1986).
9. Ismail, F. PhD thesis, The University of Michigan, 2006.
10. Ismail, F., and Roe, P. in preparation for JCP .
11. Kim, S.-s., Kim, C., Rho, O.-H., and Hong, S. JCP 186 (2003), 342.
12. Lin, H.-C. JCP 117 (1995), 20–27.
13. Liou, M. JCP 160 (2000), 623–648.
14. Pandolfi, M., and d’Ambrosio. JCP 166 (2001), 271–301.
15. Peery, K., and Imlay, S. No. 88-2924, AIAA Conference.
16. Quirk, J. Int. J. Num. Meth. Fluids 18 (1994), 555–574.
17. Roe, Nishikawa, Ismail, and Scalabrin. No. 2005-4872, AIAA Conference.
18. Roe, P. to be published in JCP .
19. Roe, P. J. Comp. Physics 43 (1981), 357–372.
20. Sanders, R., Morano, E., and Druguet, M. JCP 54 (1994).
21. Steger, J., and Warming, R. JCP 40 (1981), 263–293.
22. Tadmor, E. In Acta Numerica. 2002.
The High Order WLSQR Scheme and its
Applications in Turbomachinery
Jivrı́ Fürst
1 Introduction
This article describes the development of a high order numerical method for
the solution of compressible transonic flows. The discretisation in space is
based on the standard finite volume method of Godunov’s type. A higher
order of accuracy is achieved by the piecewise polynomial interpolation similar
to the ENO or weighted ENO method (see e.g. [8]).
The here mentioned method is developed with the aim to simplify the
implementation of the reconstruction procedure especially for the case of un-
structured meshes. The reconstruction procedure uses single stencil and com-
putes the interpolation polynomial by minimizing the weighted interpolation
error over the cells in this stencil.
The complete finite volume scheme equipped with the piecewise linear
reconstruction has been successfully used for the solution many transonic flow
problems (see e.g. [4, 3]). This article presents the extension of the method to
piecewise parabolic case with the application for the solution of flows through
turbine cascades.
The flow is described by the set of Euler or Navier-Stokes equations in
conservative form
Wt + F (W )x + G(W )y = F v (W )x + Gv (W )y + S(W ), (1)
T
where W = [ρ, ρu, ρv, e] is the vector of conservative variables, F (W ) and
G(W ) are the inviscid fluxes, F v (W ) and Gv (W ) are the viscous fluxes (F v =
Gv = 0 for the case of Euler equations) and S(W ) is a source term [2].
The equations equipped with proper boundary conditions are solved nu-
merically using an unstructured mesh and a finite volume scheme with all
unknowns located at cell centers. The fluxes through the cell interfaces are
approximated by the Gauss quadrature with the physical fluxes replaced by
the numerical ones
L/R
Here Wijq denotes the values of the vector of unknowns interpolated to the
Gauss point q of the interface Ci ∩Cj from the left cell (superscript L) or from
the right cell (superscript R). The resulting finite volume scheme for inviscid
case can be then written in semi-discrete form as
J
dWi XX
|Ci | =− ωq F AU SM P W + (Wijq
L R
, Wijq , Sijq ). (3)
dt
j∈Ni q=1
The weights wij are chosen in such a way, that w is big whenever the solution
is smooth and w is small when the solution is discontinuous. The single stencil
Mi is selected according to the order of the polynomial P .
1
Herefrom comes the name of the method - the Weighted Least Square Recon-
struction.
The WLSQR Scheme 157
The formally second order scheme can be obtained by using linear polynomials
Pi . For this case, the choice of Mi := M1i = {j : Ci ∩Cj 6= ∅} (i.e. cells touching
Ci at least by a vertex) has been tested together with
v
u h−r
wij = t , (6)
u
φi −φj p
h + hq
with h being the distance between cell centers of Ci and Cj and p = 4, q = −3,
and r = 3. The analysis of simplified cases has been carried out in [4] showing
the stability of WLSQR interpolation for special discontinuous data.
Entropy
Exact solution
-0.3356
-0.3358
-0.336
-0.3362
-0.3364
0 1 2 3
CoordinateY
Fig. 1. Isolines of Mach number and distribution of entropy along the left wall for
Ringleb’s flow problem.
It follows from this test case, that for the scheme with a piecewise quadratic
reconstruction the boundary has to be approximated also with higher order
of accuracy, otherwise the numerical results can be worse than the results of
low-cost second order method.
3 Applications in Turbomachinery
The above mentioned numerical method has been applied for the solution
of transonic flows in 2D turbine cascades. The compressible viscous flow is
described by the set of Favre averaged Navier-Stokes equations (RANS) cou-
pled with the TNT k − ω model of turbulence (see [9]). The transonic flow
through a 2D turbine cascade was solved using a hybrid mesh with quadri-
laterals around the profile, in the mixing region behind the outlet edge and
at the outlet boundary. The remaining part of the domain was filled up with
triangles. The total number of elements was 24087 with the y1+ < 1.
Figure 2 shows the isolines of the Mach number the detail of isolines of
entropy near the outlet edge obtained with the help of second and third order
method for the flow characterized by the outlet Mach number M2i = 0.906 and
Reynolds number Re = 848000. The isolines of entropy document clearly the
difference between those two results - the second order scheme gives stationary
solution whereas the wake is unsteady for the third order solution.
The proper choice and implementation of the outlet boundary condition plays
an important role in the calculation of transonic flows in cascades. It is obvi-
ous, that uniform outlet pressure gives the non-physical reflection of the shock
waves at the boundary. The implementation of the so-called non-reflecting
boundary conditions can be complicated for the case of unstructured meshes.
Therefore a simpler outlet condition with given average static pressure at the
The WLSQR Scheme 159
Fig. 2. Isolines of Mach number (above) and entropy (below) in 2D turbine cascade,
second and third order solution.
outlet was chosen in this case. Although this boundary condition admits non-
unique solution, it was used successfully in many calculations with structured
meshes (see e.g. [6]). On the other hand, this boundary condition sometimes
fails for the case of unstructured meshes with triangles at the outlet boundary.
See fig. 3 for the non-physical reflection at the outlet. However, the numerical
experiments show that the non-reflective behavior can be recovered by using
one or two layers of quadrilaterals at the outlet.
Fig. 3. Non-physical shock reflection and non-reflective behavior at the outlet con-
dition, (M2i = 1.162).
160 Jivrı́ Fürst
Conclusion
The article describes briefly the weighted least-square reconstruction proce-
dure and extends the previous work by the polynomial approximation of the
boundary. The resulting is successfully applied for the solution of turbulent
transonic flow through turbine cascades.
Problems with the outlet condition with given mean pressure are described
and a temporary solution is proposed.
References
1. T. J. Barth and D. C. Jesperson: The design and application of upwind schemes
on unstructured meshes. AIAA Paper 89–0366, AIAA, Jan 1989.
2. M. Feistauer, J. Felcman, and I. Stravskraba: Mathematical and Computational
Methods for Compressible Flow. Numerical Mathematics and Scientific Compu-
tation. Oxford University Press, 2003. ISBN 0-19-850588-4.
3. Jivrı́ Fürst and Karel Kozel. Second and third order weighted ENO scheme
on unstructured meshes. In F. Benkhaldoun and R. Vilsmeier, editors, Finite
Volumes for Complex Applications. Problems and Perspectives. Hermes, July
2002.
4. Jivrı́ Fürst. A finite volume scheme with weighted least square reconstruction.
In S. Raghay F. Benkhaldoun, D. Ouazar, editor, Finite Volumes for Complex
Applications IV, pages 345–354. Hermes Science, 2005. ISBN 1-905209-48-7.
5. G. Giocchia. Exact solutions to transonic and supersonic flows. Technical Report
AR-211, AGARD, 1985.
6. Jan Halama, Tony Arts, and Jaroslav Fovrt. Numerical solution of steady and
unsteady transonic flow in turbine cascades and stages. Computers and Fluids,
33:729–740, 2004.
7. Ami Harten. High resolution schemes for hyperbolic conservation laws. Journal
of Computational Physics, 49:357–393, 1983.
8. Changquink Hu and Chi-Wang Shu. Weighted essentially non-oscillatory
schemes on triangular meshes. Journal of Computational Physics, 150:97–127,
1999.
9. J. C. Kok: Resolving the dependence on free stream values for the k-omega
turbulence model. Technical Report NLR-TP-99295, NLR, 1999.
Building Better (Weighted) ENO Methods
William J. Rider1
Applied Physics Division, Los Alamos National Laboratory, Los Alamos NM 87545
rider@lanl.gov
the order of the norm for measuring error. In particular, this relation implies
first-order accuracy at extrema. With ENO’s basic success came a number
of pathologies that came with the freedom associated with completely adap-
tive stencils. Weighted ENO (WENO) methods were introduced to eliminate
these problems by retaining the basic stencil adaptivity that makes ENO well-
behaved at discontinuities, but providing much better stability, efficiency and
accuracy where the solution is smooth [6, 5]. This is achieved by combining
stencils in a weighted way to achieve better stability along with higher ac-
curacy. The key to the method’s accuracy and self-adaptive nature are the
smoothness detectors that provide the stencil selection. Each of the stencils
that could result from the original ENO stencil selection is used in the final
method through a weighting procedure where the weights are selected to pro-
vide higher accuracy (2m − 1 where m is the order the individual stencils).
The well-developed method used a smoothness detector based on the integral
of the total variation for the stencil polynomial’s derivatives. This method has
achieved great success in a broad spectrum of applications. This method is
extended to high-order methods up to 11th order [1], and I have implemented
a 13th order version. A major problem with the method is the substantial
complexity and cost of the smoothness detectors as the order of the method
increases.
3 Implementation
To implement these methods one needs a set of functions that will allow us to
choose the stencil closest to the TVD comparison method. These are based on
the standard functions such as the minmod function that returns the argument
with the smallest magnitude unless the arguments differ in sign,
1
minmod (a, b) = (sign (a) + sign (b)) (|a + b| − |a − b|) , (1)
4
which returns a, b or 0. The second function is used with the hierarchical ENO
methods that returns the argument with the minimum magnitude,
1
xminmod (a, b) = (sign (a + b)) (|a + b| − |a − b|) , (2)
2
Building Better (Weighted) ENO Methods 163
which returns either a or b. These two functions can define two functions that
are used to provide a bounding functionality needed for comparison ENO
methods, the median function that returns the argument that is bounded by
the other two,
median (a, b, c) = a + minmod (b − a, c − a) , (3)
which returns a, b or c. If one argument is TVD, it will return the TVD
argument or that closest to it. The other option is to return the argument
that is closest to another
xmedian (a, b, c) = a + xminmod (b − a, c − a) , (4)
in this case the function returns either b or c depending which is closest to a.
These functions can be used to orchestrate a “playoff system” of sorts
where comparison with the TVD stencil is used to determine stencil fitness.
Take the example of a fourth-order stencil with four different alternative fourth
order stencils Sk , and a TVD stencil for comparison, ST V D ,
Sa = median (ST V D , S1 , S2 ) , Sb = median (ST V D , S1 , S2 ) ,
S = median (ST V D , Sa , Sb )
A weighted ENO scheme would proceed in a similar fashion although all
the stencil will contribute to the final one. The smoothness detectors will us
a difference between the TVD and high-order stencil that then will be used
to compute the weights just as the standard WENO methods.
1 w̄k
ISk = 2 ; wk = ,
(Sk − ST V D ) IS 2 +
where w̄k are the weights that would produce 2m − 1 stencil with m being
the order of the candidate stencil Sk . The weights are then renormalized
so their sum is one. Truncation error analysis confirms that this technique
produces methods with the same accuracy as the original WENO methods.
The difference in the efficiency is easy to see through the complexity and
number of operations to implement the smoothness detectors.
is(0) = 22658.D0*f(ic-d)**2 - 208501.D0*f(ic-d)*f(ic-c)
& + 482963.D0*f(ic-c)**2 + 364863.D0*f(ic-d)*f(ic-b)
& - 1704396.D0*f(ic-c)*f(ic-b) + 1521393.D0*f(ic-b)**2
& - 288007.D0*f(ic-d)*f(ic-a)
& + 1358458.D0*f(ic-c)*f(ic-a)
& - 2462076.D0*f(ic-b)*f(ic-a) + 1020563.D0*f(ic-a)**2
& + (86329.D0*f(ic-d) - 411487.D0*f(ic-c)
& + 758823.D0*f(ic-b) - 649501.D0*f(ic-a))*f(ic)
& + 107918.D0*f(ic)**2
This is the ninth-order WENO smoothness detector for a single (the most
upwind) stencil in Fortran [1]. On the other hand, consider the new WENO
method’s smoothness detector for all five stencils.
164 William J. Rider
4 Analysis
The analysis of the methods insofar as accuracy is concerned is simple. All
the methods trivially produce the order of accuracy of the stencils supplied.
In particular, the median function will produce the order of accuracy of two
of its arguments irregardless of the nature of the third argument. Thus if two
of the arguments to the median function are of a certain order, the median
function will return a result of at least that order. The xmedian function will
also provide the order of accuracy necessary with ENO schemes. The difficult
and unsolved task is to determine whether the method produces a result that
is T V D + O hm+1 . The original ENO methods did not necessarily provide
proven results of this nature. This has led to the conjecture that ENO schemes
are total variation bounded instead. I believe that the comparison framework
of ENO methods leads to a form that can be more straightforwardly provide
the structure for a subsequent proof. This is because of the strong connection
of the method to a method that is provably TVD. I will focus on the semi-
discrete three-point case for a positive velocity then,
∂uj
= −Cj−1/2 (uj − uj−1 ) , (5)
∂t
where Cj−1/2 ε0 and is usually a nonlinear function for high-order TVD meth-
ods (Cj−1/2 = 1/∆x for first-order upwind) [4]. An example of high-order
TVD method would produce Cj−1/2 = 1/∆x + Qj+1/2 − Qj−1/2 where the
TVD requirement would place restrictions on the form Qj−1/2 might take.
The comparison based ENO schemes would then be built upon these schemes.
It seems plausible that the choice of the stencil that is closest to the TVD
scheme would produce a total variation behavior desired with ENO schemes,
T V D + O hm+1 .
5 Results
Space is limited, but a simple explanatory result can be offered. I will solve
the motion of a complex waveform containing both smooth and discontinuous
data using a simple scalar wave equation [7] with a standard and comparison-
based ENO methods. We will compare both the results and the total variation
behavior using 200 cells.
Building Better (Weighted) ENO Methods 165
1.1
1 ENO
ENO!C
Exact
0
!0.1
!0.995 0 0.995
x/t
Fig. 1. Computation of a waveform for one period using a 6th order standard
ENO methods and the 6th order comparison ENO scheme. The standard ENO
(ENO) produces an error of 1.70 × 10−2 , the comparison ENO (ENO-C) produces
an error of 1.21 × 10−2 .
I can also examine the L1 error in the solution for this case. The error with
the comparison method is smaller than the classic ENO method and its total
variation behavior is better. Furthermore, the total variation behavior indi-
C
8
ENO
ENO!C
TV
7
0 1 2
time
Fig. 2. Computation of a waveform for one period using a 6th order standard
ENO methods and the 6th order comparison ENO scheme. The total variation is
compared. The total variation change is smaller with the ENO-C method, and does
not have the large excursion late in the calculation.
166 William J. Rider
cates that the method is acting just as the classic ENO method. Although we
do not present the weighted version of the ENO method here similar results
are obtained. While the total variation behavior of these weighted schemes is
worse than the classic WENO method, the errors are smaller and the oscilla-
tions produced are less severe.
Conclusion
I have introduced a methodology to develop ENO method that is different than
the previously available ENO techniques. The approach employs a comparison
principle where the high-order stencil; is selected on the basis of its “closeness”
to a TVD (or monotone) stencil. This approach can lead to a large degree of
flexibility in the choice of method properties. The results produced by these
methods are promising and generally less oscillatory and more accurate than
the standard ENO methods.
References
1. D. Balsara and C.-W. Shu. Monotonicity preserving weighted essentially non-
oscillatory schemes with increasingly high order of accuracy. Journal of Compu-
tational Physics, 160(2):405–452, 2000.
2. A. Harten. High resolution schemes for hyperbolic conservation laws. Journal
of Computational Physics, 49:357–393, 1983.
3. A. Harten, B. Engquist, S. Osher, and S. Chakravarthy. Uniformly high or-
der accurate essentially non-oscillatory schemes, III. Journal of Computational
Physics, 71:231–303, 1987.
4. A. Jameson and P. D. Lax. Conditions for the construction of multipoint to-
tal variation diminishing schemes. Applied Numerical Mathematics, 2:335–345.,
1986.
5. G.-S. Jiang and C.-W. Shu. Efficient implementation of weighted ENO schemes.
Journal of Computational Physics, 126:202–228, 1996.
6. X.-D. Liu, S. Osher, and T. Chan. Weighted essentially non-oscillatory schemes.
Journal of Computational Physics, 115:200–212, 1994.
7. A. Suresh and H. T. Huynh. Accurate monotonicity-preserving schemes with
Runge-Kutta time stepping. Journal of Computational Physics, 136:83–99, 1997.
8. P. K. Sweby. High resolution schemes using flux limiters for hyperbolic conser-
vation laws. SIAM Journal on Numerical Analysis, 21:995–1011, 1984.
Discontinuity Diagnosis Essentially
Non-Oscillatory Schemes
1 Introduction
(1) Compute the divided difference table of f (u) up to rth desirable order.
(2) Construct the base stencil for each node. Choose the smoothest 3-point
stencil for each node as the base stencil by comaparing the divided differences
like ENO-3 schemes. Let us denote the base stencil at node j by S(j):
j j j
S(j) = (kmin , kmin + 1, kmax ), (3)
j j
where kmin and kmax are the left and right nodes of S(j). The base stencil is
chosen for each node, not for each cell boundary.
(3) Start ENO-3 schemes to approximate the numerical flux fˆj+1/2 at
the cell boundary of xj+1/2 . Let us denote the ENO-3 stencil at xj+1/2 by
S(j + 1/2):
S(j + 1/2) = (kmin , kmin + 1, kmax ). (4)
The discontinuity diagnosis mechanism is introduced into ENO-3 schemes
here. Firstly, DDENO schemes analyzes the information coming from the left
neighboring base stencil of node kmin . It compares the present stencil S(j +
1/2) with the base stencil S(kmin ) of node kmin . Secondly, it analyzes the
information coming from the base stencil S(kmax ) of the right adjacent node
kmax .
Discontinuity Diagnosis Essentially Non-Oscillatory Schemes 169
(i) If
S(kmin ) = S(j + 1/2) = (kmin , kmin + 1, kmax ), (5)
it means that there is a discontinuity between the nodes kmin − 1 and kmin .
Otherwise, it means that the left region of the present stencil S(j + 1/2) is
smooth and the present stencil of ENO-3 can extend at least one point to the
left.
(ii) If
S(kmax ) = S(j + 1/2) = (kmin , kmin + 1, kmax ), (6)
it means that there is a discontinuity between the nodes kmax and kmax + 1.
Otherwise, it means that the right region of S(j + 1/2) is smooth and it can
add at least one point to the right.
(iii) If
S(kmin ) = S(kmax ) = S(j + 1/2) = (kmin , kmin + 1, kmax ), (7)
it means that the present stencil S(j + 1/2) is surrounded by two discontinu-
ities at both sides and DDENO ceases to increase points and just uses ENO-3
to approximate the numerical flux at xj+1/2 .
(iv) Otherwise, if
kmin (S(kmin )) < kmin and kmax (S(kmax )) > kmax , (8)
DDENO extend ENO-3 to ENO-5.
(4) Inductively, DDENO repeats step (3) to increase the ENO-3 to higher
order accuracy in smooth regions.
We remark that in the logical operation (iv), we use the logical opera-
tor ”and”. It is a very strong restriction and we do not need to know which
nodes are chosen by S(j + 1/2). If we use the less strict logical operator ”or”,
we should care about the condition of the present stencil. In some severe
positions, the numerical flux is one-sidedly interpolated or extrapolated and
high-order ENO schemes will induce oscillation. Therefore, the logical opera-
tor ”or” should be carefully used. We do not suggest use higher order one-side
extrapolation to approximate the numerical fluxes near discontinuities.
3 Numerical results
1 1
0.8 0.8
0.6 0.6
0.4 0.4
0.2 0.2
0 0
0 0.2 0.4 0.6 0.8 1 0 0.2 0.4 0.6 0.8 1
Fig. 1. DDENO-3-9 and ENO-3, Sod’s problem, 100 points, t=0.2 (solid lines are
DDENO-3-9 and dashed lines are ENO-3): (left) pressure; (right) density.
1.4
1.2
0.8
0.6
0.4
Fig. 2. DDENO-3-9 and ENO-3, Lax’s problem, density, 100 points, t=0.13 (solid
line is DDENO-3-9 and dashed line is ENO-3).
0 2 4 6 8 10
Fig. 3. DDENO-5-9 and ENO-5, Shu’s problem, density, 200 points, t=1.8 (solid
line is DDENO-5-9 and dashed line is ENO-5).
(3.857, 2.629, 10.333), x < 1.0
(ρ, u, P ) = (10)
(1 + ε sin(5x), 0, 1), xε1.0
where ε = 0.2. The problem is run with 200 points in the interval [0, 10]. The
results are plotted in Fig. 3. We mention above that the logical operator ”or”
in logical equation (8) should be used carefully. This problem looks severe
but it is not difficult for ENO schemes to deal with because sine waves are
periodically continuous functions. In this example, we use ENO-5 as the base
stencil because the fluxes are one-side smooth. Then we use logical operator
”and” to extend ENO-5 to ENO-9 at more smooth regions. Thus fifth-order
one-sided interpolation or extrapolation is performed in this example. There-
fore, the accuracy of the results in this example is fifth-order or ninth-order.
From Fig. 3 we can see the expected improvements in resolution with higher
orders.
172 Yun-Feng Liu and Jian-Ping Wang
4 Concluding remarks
In this paper, we introduce a discontinuity diagnosis mechanism into the flux-
version ENO schemes and propose discontinuity diagnosis essentially non-
oscillatory (DDENO) schemes. DDENO schemes use 3-point smoothest ENO
stencil as the base stencil and detect discontinuities by comparing the present
base stencil with neighboring base stencils. This mechanism prevents high-
order ENO schemes from choosing the weaker discontinuities when it encoun-
ters two or more than two discontinuities at both sides. DDENO schemes
are higher order accurate at smooth regions and three-order (the accuracy of
the ENO base stencil) at discontinuities. Numerical experiments demonstrate
that DDENO schemes work well in the examples and achieve higher order
accuracy at smooth regions.
Acknowledgments
The authors gratefully acknowledge Professor Chi-Wang Shu for helpful dis-
cussions and suggestions about flux-version ENO schemes. We also thank Dr.
Hong-Wei Liu at The Hong Kong University of Science and Technology for
helpful discussions.
References
1. Harten, A., Engquist, B., Osher, S., Chakravarthy, S.J.: Uniformly high order
essentially non-oscillatory schemes, III, J.Comput. Phys. 71, 231, (1987)
2. Shu, C.W., Osher, S.: Efficient implementation of essentially non-oscillatory
shock capturing schemes, J.Comput. Phys. 77, 439, (1988)
3. Shu, C.W., Osher, S.: Efficient implementation of essentially non-oscillatory
shock capturing schemes II, J.Comput. Phys. 83, 32, (1989)
4. Jiang, G.S., Shu, C.W.: Efficient implementation of weighted ENO schemes,
J.Comput. Phys. 126, 202, (1996)
5. Steger, J.L., Warming, R.F: Flux vector splitting of the inviscid gasdynamic
equations with application to finite-difference methods, J.Comput. Phys. 40,
263, (1981)
6. Sod, G.A.: A survey of finite difference methods for systems of nonlinear hyper-
bolic conservation laws, J.Comput. Phys. 27, 1, (1978)
7. Lax, P.D.: Weak solutions of nonlinear hyperbolic equations and their numerical
computation, Comm. Pure Appl. Math. 7, 159 (1954)
Third Order Reconstruction on Unstructured
Highly Irregular 3D Meshes
1 Introduction
A
!h
B C
2 Third–order reconstruction
For third–order reconstruction within the cell Ωh the information is collected
from all neighbouring cells (see Fig. 1 for 2D illustration). The function values
ϕp = ϕ(rp ) are located at cell centres rp (p = 0, . . . , m).
We seek the reconstruction function in the form:
1 T
ϕ (r) = ϕ0 + GT (r − r0 ) + (r − r0 ) H (r − r0 ) (1)
2
where G and H stand for approximated gradient vector and Hessian matrix
respectively (at the point r0 ). To obtain third order accuracy, G and H have
to be estimated with second and first order accuracy, respectively.
To achieve this goal both G and H are expressed as as linear combinations
of the function increments:
m
X m
X
G= Gp wp (ϕp − ϕ0 ) , H= Hp wp (ϕp − ϕ0 ) (2)
p=1 p=1
Table 1. Percentage of mesh cells for which gradient and Hessian could not be
calculated using the standard procedure of Section 2
m m
1
rpc · rTpc ,
P P
M = M [r0 , . . . , rm ] = rpc = rp − rc , rc = m+1 rp (5)
p=0 p=0
The new cloud becomes fully isotropic since the new matrix:
M? = M [r?0 , . . . , r?m ] = I (7)
Third Order Reconstruction on Unstructured Highly Irregular 3D Meshes 177
On this transformed cloud both the gradient G? and the Hessian H? can
be evaluated using the procedures introduced in Section 2.
The original gradient and Hessian can be recovered now as:
4 Numerical results
Fig. 3. The grid used in numerical test with a local stretch ratio η ≈ 24.
0.62
0.62
0.72 0.55
0.53
0.82
0.50 0.88 0.50
0.
0.45 42
Fig. 4. Mach isolines in the 3D channel. Inlet Mach number 0.5 (dashed lines -
isotropic mesh, solid lines - stretched mesh of Fig 3).
178 J. Rokicki and R. Wieteska
The performance of the full method was investigated for the compressible
flow in 3D channel with sinusoidal bump (see Fig. 3). The mesh consisted of
21915 isotropic tetrahedral cells. To illustrate the robustness of the method
presented in both Sections 2 and 3, the fully subsonic flow (Ma=0.5 at in-
let) was calculated on the original mesh as well as on the mesh with highly
stretched cells (stretch ratio reaching 24).
The comparison of Mach isolines is presented in Fig. 4 - showing almost
perfect coincidence of both simulations. On the other hand the method of
Section 1 alone is unable for the stretched mesh to deliver reasonable accuracy.
5 Conclusions
References
1. Carl F., Ollivier-Gooch: Quasi-ENO Schemes for Unstructured Meshes Based on
Unlimited Data-Dependent Least-Squares Reconstruction. Journal of Computa-
tional Physics, 133, 6–17, (1997)
2. Shi J., Hu Ch., Shu Ch-W.: A Technique of Treating Negative Weights in WENO
Schemes. Journal of Computational Physics, 175, 108–127, (2002).
3. Hu Ch., Shu Ch-W.: Weighted Essentially Non-oscillatory Schemes on Triangular
Meshes. Journal of Computational Physics, 150, 108–127, (2002).
4. Henrick A.K., Aslam T.D., Powers J.M.: Mapped weighted essentially non-
oscillatory schemes: Achieving optimal order near critical points. Journal of
Computational Physics, 207, 542–567, (2005).
5. Schatzman M.: Numerical Analysis - a mathematical introduction. Clarendon
Press - Oxford, (2002).
6. Wieteska R., Majewski J., Rokicki J.: Investigation of WENO Schemes For 3D
Unstructured Grids. Computer Assisted Mechanics and Engineering Sciences,
(2006), (in print).
An Intrinsically Multi-Dimensional
Acoustics Convection Upstream Resolution
Algorithm for the Euler Equations
Joe Iannelli1
1 Introduction
∂fjq ∂fjp
" !#
∂fj ∂ N N N
∂q
− εψ c αai aj + α ai aj + ai + ai δ = 0 (3)
∂xj ∂xi ∂xj ∂xj ∂xj
5 Computational Results
solid wall specification only extends to x = 0.4; beyond this station, the upper
and lower boundaries correspond to outlets, to allow the reflected shocks to
cross the computational boundaries with minimal further reflection. The total
outlet remains supersonic. At the solid upper and lower walls, the inviscid-
flow wall-tangency boundary condition is enforced via surface integrals [1, 3].
The final steady state is computationally achieved with ψ = 0.4 by advancing
the solution in time with a Courant number equal to 100.
The wall-tangency boundary conditions on the upper and lower boundary
walls induce emergence of two separate oblique shocks of different strengths,
shocks that propagate toward each other, interact, and reflect away from each
other, towards the outlet as displayed in Figure 1
M3 = 2.28
M2 = 2.20
In comparison to the solution in previous articles [2, 3], this solution re-
mains essentially non-oscillatory even at the compression corners, which are
two geometric singularities. This performance is achieved by simply increas-
ing ψ to 2, but only locally over the four elements closest to each singularity.
Overall, the Mach-number distribution presents an essentially non-oscillatory
solution with crisply calculated incident and reflected shocks. In particular,
the algorithm allows the reflected shocks to cross the outflow boundary essen-
tially unperturbed. Significantly, this computational solution mirrors the avail-
able exact solution, with four juxtaposed plateaus connected by four oblique
shocks. The calculated Mach numbers in the plateaus downstream of the four
shocks are M2 = 2.20, M3 = 2.28, M4 ' M5 = 2.08; the shock inclination
angles are θ2 = 28.53o , θ3 = −26.90o , θ4 = 29.52 and θ5 = −30.15, computed
results that coincide with the corresponding exact values.
The ACG-GMRES algorithm has rapidly computed the solution at each
time station and quickly achieved a steady state, with a reduction of the
equation-system residual norm to 1 × 10−14 , hence machine zero, achieved in
about 45 IRK cycles “C” at a constant maximum Courant number of 100.
The recorded essentially monotone decrease of the residual corresponds to an
exponential convergence rate.
184 Joe Iannelli
6 Concluding Remarks
The Acoustics-Convection Galerkin formulation induces a consistent multi-
dimensional upstream-bias directly in the continuum, at the differential equa-
tion level, before the spatial discretization. The discrete equations are obtained
by way of a classical Galerkin finite element discretization of the formulation
characteristics-bias system.
These time-dependent discrete equations are then integrated in time via
an implicit Runge-Kutta procedure coupled with GMRES, with analytically
determined Jacobians. This procedure remains non-linearly stable and affords
rapid convergence to steady states, with continuously updated upstream di-
rections and maximum Courant number of order 100.
References
1. Iannelli J., Characteristics Finite Element Methods in Computational Fluid Dy-
namics, Springer Verlag, 2006.
2. Iannelli J., “SUPG Formulation of the Acoustics-Convection Upstream Resolu-
tion Algorithm for Compressible Flows”, 44th Aerospace Sciences Meeting and
Exhibit, 2006, Reno, NV; AIAA 2006-1087.
3. Iannelli J., “Finite element and implicit Runge-Kutta implementation of an
acoustics-convection upstream resolution algorithm for the time-dependent two-
dimensional Euler equations”, Int. J. Numer. Meth. Fluids 2005; 49:1261-1286.
4. Iannelli J., “Derivation and characteristics analysis of an acoustics-convection
uptream resolution algorithm for the two-dimensional Euler and Navier-Stokes
equations”, Int. J. Numer. Meth. Fluids 2005; 49:1233-1260.
5. Hughes T.J.R., Recent progress in the development and understanding of SUPG
methods with special reference to the compressible Euler and Navier-Stokes
equations. Int. J. Numer. Meth. Fluids 1987; 7:11-21.
6. Aliabadi S.K., Tezduyar T.E., Parallel Fluid Dynamics Copmutations in
Aerospace Applications. Int. J. Numer. Meth. Engr. 1995; 21:783-805.
7. Saad Y., Schultz M.H., “GMRES: A Generalized Minimal Residual Algorithm
for Solving Nonsymmetric Linear Systems”, SIAM J. Sci. Stat. Comput. 1986;
7(3):856-869.
Multi-dimensional Limiting Process for Two-
and Three-dimensional Flow Physics Analyses
1 Introduction
Since the late 1970s, numerous ways to control oscillations have been studied
and several limiting concepts have been proposed. Most representatives would
be TVB, ENO/WENO and TVD[1]. However, most oscillation-free schemes
have been based on the mathematical analysis of one-dimensional convection
equation. Although they may work very well in many cases, it is often insuf-
ficient or almost impossible to control oscillations near shock discontinuity in
multiple dimension. For that reason, it is obvious that more practical oscilla-
tion control method for multi-dimensional applications is needed.
Kim and Kim[2] extended the one-dimensional monotonic condition to
two-dimensional flow and presented the two-dimensional limiting condition
successfully. With the limiting condition, a multi-dimensional limiting process
(MLP) is proposed which gives more accurate results for the two-dimensional
Euler and Navier-Stokes equations. It is this approach which prompts the
work of the present research. As a continuous study on MLP, we extend the
idea of MLP to three-dimensional space and also consider three-dimensional
MLP approach with high-order polynomial interpolation.
The present research essentially extends the MUSCL approach[3] with TVD
limiter. One-dimensional limiting condition using TVD constraint can be writ-
ten as follows.[1]
0 ≤ φ(r) ≤ min(2r, 2). (1)
The extension of Eq.(1) in a dimensional splitting manner may be in-
sufficient to prevent oscillations in multi-dimensional flow. The dimensional
splitting extension does not possess any information on property distribution
at cell vertex points, which would be essential when property gradient is not
aligned with grid lines. Thus, as an extended condition including the missing
information, Kim and Kim[2] proposed multi-dimensional limiting condition
for two-dimensional flow.
In order to realize the limiting condition in three-dimensional situation,
the values at vertex points are required to satisfy the following condition.
Φ̄min max
p,q,r ≤ Φi+p/2,j+q/2,k+r/2 ≤ Φ̄p,q,r , (2)
where Φi+p/2,j+q/2,k+r/2 is a vertex point value and Φ̄min max
p,q,r and Φ̄p,q,r are the
minimum and maximum cell-averaged values among neighboring candidates,
respectively. To derive the multi-dimensional limiting function from Eq.(2),
we need to express the vertex point values in terms of variations at the cell-
interfaces.
After that, we determine each variation to satisfy the multi-dimensional
limiting condition using the limiting coefficient α. The coefficient α possesses
the information of multi-dimensionally distributed physical property. Thus,
with the help of the coefficient α, we can formulate the multi-dimensional
limiting function. Here, we will give only the final formulation of MLP for
three-dimensional flow and refer to [4] for most details of derivation procedure.
αL,R and βL,R in Eq.(3) and Eq.(4) can be summarized as follows. Along the
ξ-direction, if ∆Φpξ ≥ 0,
ξ max
2max 1, r L,i,j,k Φp,q,r − Φi,j,k
αL = g , (5)
∆Φqη ∆Φrζ
1 + ∆Φp + ∆Φp ∆Φi+1/2,j,k
ξ ξ i,j,k
ξ min
2max 1, 1/r R,i+1,j,k Φp,q,r − Φi+1,j,k
αR = g , (6)
∆Φqη ∆Φrζ
1 + ∆Φp + ∆Φp ∆Φi+3/2,j,k
ξ ξ i+1,j,k
ξ ∆Φ
i+1/2,j,k ξ ∆Φ
i+1/2,j,k
where rL,i,j,k = ∆Φi−1/2,j,k , rR,i+1,j,k = ∆Φi+3/2,j,k , g(x) = max(1, min(2, x)).
Along the η- and ζ-direction, the left and right values at the cell-interfaces
can be formulated in the same way.
Combining Eq.(5) and Eq.(6) with β in the form of third order and fifth
order polynomial, we finally obtain three-dimensional versions of MLP3 and
MLP5, respectively. For more detailed explanation, see Ref.[2].
3 Numerical Results
3.1 Three-dimensional Supersonic Corner Flow
We also verify the complex shock structure capturing ability of MLP5 with
various Mach number and inclination angle of wedge.
Figure 2 is the numerical schlieren image using MLP5. The formulation
of the triple-point and accurate prediction of reflected shock wave, slip line
and spiral curl are important from the standpoint of numerical analysis. The
numerical results indicate that MLP is one of the sound numerical method
for complex shock structure capturing in multiple dimension.
Moreover, the curls interact with a pillar-vortex near the corner of the
walls, which leads to much more complicated flow structure. In figure 4, we
can see the interaction of curls and pillar-vortex tip, and the profile is very
sensitive to numerical scheme. The result of M-AUSMPW+ with MLP5 is
clearer than that of AUSMPW+ with van leer limiter, which means that we
can achieve accuracy enhancement by using M-AUSMPW+ with MLP5.
190 Sung-Hwan Yoon, Chongam Kim and Kyu-Hong Kim
4 Concluding Remarks
References
1. P. K. Sweby :High Resolution Schemes Using Flux Limiters for Hyperbolic Con-
servation Laws. SIAM J. on Num. Analysis, Vol. 21(5), 995–1011 (1984)
2. K. H. Kim and C. Kim : Accurate, Efficient and Monotonic Numerical Methods
for Multi-dimensional Compressible Flows, Part II Multi-dimensional Limiting
Process. J. of Comput. Physics, Vol. 208(2), 570–615 (2005)
3. B. van Leer : Toward the ultimate conservative difference scheme. J. of Comput.
Physics, Vol. 32, 101–136 (1979)
4. S. Yoon, C. Kim, and K. Kim : Multi-Dimensional Limiting Process for Three
Dimensional Compressible Flows. 17th AIAA Computational Fluid Dynamics
Conference, Toronto, Ontario, June 6-9 (2005)
5. V. Daru, C. Tenaud : Evaluation of TVD high resolution schemes for unsteady
viscous shocked flows. Computers and Fluids, 30 89–113 (2001)
A Multidimensional Kinetic Upwind Method
for Euler Equations
Y’ Y !
Y
l^ C1
X’
^
l
+ s
Pi s C1 P0 ! X
P0 ! X
Pi
Fig. 1.
In the previous section we had assumed that ζ(θ) is given. Note that ζ(θ)
ˆ of component of flux vector Q along `.
is a derivative along any direction `, ˆ
Since the Euler equations are hyperbolic in nature we need to evaluate ζ(θ) in
an upwind way along any direction θ. This has been achieved by developing
special flux vector splitting(SFVS) [3, 2] method which splits the special flux
E into E + and E − and using the least squares princilple to evaluate derivative.
Note that the special flux E,
E = Q `ˆ = GX cos θ + GY sin θ (8)
contains the flux components GX and GY which are along x and y directions
while we need to split the flux along `.ˆ Hence the name special flux vector
splitting (SFVS) is used. The [3] can be referred for the detailed derivation
and the special split flux expressions. Using SFVS the directional derivative
becomes,
∂E ∂E + ∂E −
ζ(θ) = = + (9)
∂s ∂s ∂s
To compute the derivatives of split fluxes in the above equation we can again
use the least squares approximation of derivatives, as in standard LSKUM,
along with special split fluxes. The upwinding is achieved by stencil division,
as in LSKUM, along each direction (Fig.(1.b)). The required least squares
formulae for spatial derivatives in eqn.(9) are given in [3]. It is thus clear
that the split derivatives in the expression for the directional derivative, ζ(θj )
can be determined in an upwind way by using the above formula along each
direction θj , for j = 1, . . . . . . , m, m being number of directions which can be
194 Keshav S Malagi, P S Kulkarni, and S M Deshpande
+ −
arbitrary. It is important to note that ζ(θj ) = ∂E ∂E
∂s + ∂s is determined for
each `ˆ or equivalently each θj . Full stencil into two halves is required but the
substencils keep changing. The union of two substencils is always the same
full stencil. It is interesting to observe that we are using least squares twice.
First it is used in generating the data and then in fitting Fourier series to the
data so generated.
3 q-KUMARI
Even though KUMARI gives good accuracy, it is only first order accurate.
It is highly desirable to get more accurate results by reducing the dissipa-
tion further by making it higher order accurate. It can be made higher order
accurate by evaluating ζ(θ) to a higher order which will result in less dissipa-
tion and more accurate results. This is achieved by using defect correction [1]
along with q-variables [7] (also called as entropy variables). The higher order
accurate formula for ζ(θ) is given by,
P P P P
2
∆Ẽi+ ∆x0i ∆y 0 i − ∆Ẽi+ ∆yi0 ∆x0 i ∆y 0 i
ζ (2) (θj ) = 2
∆x0 2i ∆y 0 2i − ( ∆x0 i ∆y 0 i )
P P P
Pi C1+
P P P P
2
∆Ẽi− ∆x0i ∆y 0 i − ∆Ẽi− ∆yi0 ∆x0 i ∆y 0 i
+ 2
∆x0 2i ∆y 0 2i − ( ∆x0 i ∆y 0 i )
P P P
Pi C1−
The method has been successfully tested on a standard shock reflection prob-
lem with M = 2.9, θ = 110 over a domain with [0 ≤ x ≤ 3] and [0 ≤ y ≤ 1].
We can see from the plot of pressure Fig.(2.a) at y=0.35 along x-axis that the
shock is captured crisply for KUMARI method as compared to LSKUM. The
better residue fall Fig.(2.b) is obtained for KUMARI method, signifying bet-
ter convergence. Studies on sensitivity of the method to varying aspect ratio
AR, have been carried out and is found to be less sensitive and are reported
in [2, 5]. The q-KUMARI has been applied to standard test cases of flow
past NACA 0012 airfoil for supersonic, subsonic and transonic flow regimes.
A Multidimensional Kinetic Upwind Method for Euler Equations 195
References
1. S M Deshpande: ”Meshless method, accuracy, symmetry breaking, upwinding
and LSKUM”, Fluid mechanics report Report No 2003 FM 1, Department of
Aerospce engineering, Indian Institute of Science, Bangalore.
2. Keshav S Malagi, P S. Kulkarni, S M Deshpande: ”Rotationally invariant kinetic
upwind method(KUMARI)”, Fluid mechanics report Report No 2004 FM 23,
Department of Aerospce engineering, Indian Institute of Science, Bangalore.
3. Keshav Shrinivas Malagi: ”Rotationally invariant kinetic upwind
method(KUMARI)”, MSc thesis submitted, Department of Aerospce en-
gineering, Indian Institute of Science, Bangalore.
4. Mandal J C and S M. Deshpande: ”Kinetic Flux Vector Splitting for Euler
Equations”, Computers and Fluids, Vol 23. No. 2, pp.447-478, 1994.
5. Keshav S. Malagi, P S. Kulkarni, S M. Deshpande: ”Rotationally invariant ki-
netic upwind method (KUMARI)”, Proceedings of ”8th Annual AeSI CFD Sym-
posium”, Aug-2005, NAL, Bangalore India.
6. P. D. Lax and X D. Liu: ”Solution of two dimensional Riemann problem of
gas Dynamics by positive schemes”, SIAM J Sci. Comput. Vol. 19, No.2, pp.
319-340,1998.
7. S M. Deshpande, K Anandhanarayanan, C. Praveen and V. Ramesh: ”Theory
and applications of 3-D LSKUM based on entropy variables”, Int. J for Num.
Methods in Fluids: vol40, Nos 1-2, pp.47-62, 2002.
196 Keshav S Malagi, P S Kulkarni, and S M Deshpande
0
4
KUMARI
3.5
KUMARI LSKUM
LSKUM -5
3
RESIDUE
pressure
EXACT SOLUTION
2.5
-10
2
1.5
-15
1
0 1 2 3 0 5000 10000 15000
X N
0.8
0.6
0.4
0.2
0
0 0.2 0.4 0.6 0.8 1
(g)Pressure contours for blast wave (h)Pressure contours for blast wave
test case with q-KUMARI test case with q-LSKUM
Fig. 2.
High Resolution Quantum Kinetic Beam
Schemes and Its Applications to Ideal
Quantum Gas Dynamical Flows
∂n(x, t)
+ ∇x · j(x, t) = 0 (2)
∂t Z
∂mj(x, t) dp p
+ ∇x · p f (p, x, t) = 0 (3)
∂t h3 m
dp p p2
Z
∂(x, t)
+ ∇x · f (p, x, t) = 0. (4)
∂t h3 m 2m
Here the number density, number density flux, and energy density are given,
respectively, by
dp p2
Z Z Z
dp dp p
n(x, t) = f, j(x, t) = f, (x, t) = f. (5)
h3 h3 m h3 2m
One can obtain these macroscopic quantities in closed form in terms of the
Bose or Fermi functions in the local equilibrium limit. The effect of the ex-
ternally applied field U (x, t) is not considered. We first consider the following
local equilibrium quantum distribution in two space dimensions,
2
(0) +(py −muy )2 ]/2mkB T (x,y,t)
f2 (px , py , x, y, t) = [z −1 e[(px −mux ) + θ]−1 (6)
where θ = +1 denotes the Fermi-Dirac statistics and θ = −1 the Bose-
Einstein statistics, z(x, y, t) = eµ(x,y,t)/kB T (x,y,t) is the fugacity, ux (x, y, t),
and uy (x, y, t) are the mean velocity components. Then the number density
n(x, y, t) is given by
Z ∞
dpx dpy (0) h1 (z)
n(x, y, t) = 2
f2 (px , py , x, y, t) = (7)
−∞ h λ2
the momentum j(x, y, t),
Z ∞
dpx dpy px (0)
jx (x, y, t) = f (px , py , x, y, t) = n(x, y, t)ux (x, y, t) (8)
−∞ h2 m 2
Z ∞
dpx dpy py (0)
jy (x, y, t) = f (px , py , x, y, t) = n(x, y, t)uy (x, y, t) (9)
−∞ h2 m 2
and the energy density (x, y, t),
Z ∞
dpx dpy (p2x + p2y ) (0) h2 (z) 1
(x, y, t) = 2
f2 = + n(u2x + u2y ) (10)
−∞ h 2m βλ 2
q
βh2
where λ = 2πm is the thermal wavelength and β = 1/kB T (x, y, t).
In the above, the function hν (z) is equal to gν (z) which denotes the Bose
function of order ν when we consider the boson gas. The Bose function is
defined by
Z ∞ ∞
1 xν−1 X zl
gν (z) ≡ dx −1 x = (11)
Γ (ν) 0 z e −1 lν
l=0
And in the case of Fermi gas,hν (z) is equal to fν (z),the Fermi function, defined
by
High Resolution Quantum Kinetic Beam Schemes 199
∞ ∞
xν−1 X (−z)l
Z
1
fν (z) ≡ dx = − ν . (12)
Γ (ν) 0 z −1 ex + 1 l
l=0
In Bose-Einstein gas,the fugacity could not greater than one or less than zero,
but in Fermi-Dirac gas we do not have such limits.
The main idea of the beam scheme is to approximate the local equilibrium dis-
tribution function into several delta functions. The distribution in two spatial
(0)
dimensions, f2 , is approximate by
f2 (p, x, t) ∼
(0)
= qi,j (px , py )
= ai,j δ(px − px0 , py − py0 )
−
+ bi,j δ(px − p+ x0 , py − py0 ) + bi,j δ(px − px0 , py − py0 )
−
+ ci,j δ(px − px0 , py − p+ y0 ) + ci,j δ(px − px0 , py − py0 ) (13)
where p±
x0 = px0 ± ∆px and p±
= py0 ± ∆py . The unknown parameters
y0
a, b, c, px0 , py0 , ∆px , and ∆py in the beam approximation can be found in
[15,16] and they are given by
2πm 4 g22 (z) 1 2πm g22 (z)
a=( )[g1 (z) − ], b = c = ( )
β 3 g3 (z) 3 β g3 (z)
s
3m g3 (z)
∆px = ∆py = px = mux , pyo = muy
2β g2 (z) o
The conservative quantities carried by each beam in cell (i, j) are Qσ,i,j =
(Rσ , Mσ , Nσ , Eσ )i,j , with
Z
dpx dpy
Rσ,i,j = cσ,i,j δ(px − p̄x,σ , py − p̄y,σ ) (14)
h2
Z
dpx dpy px
Mσ,i,j = cσ,i,j δ(px − p̄x,σ , py − p0y ) (15)
h2 m
Z
dpx dpy py
Nσ,i,j = 2
cσ,i,j δ(px − p0x , py − p̄y,σ ) (16)
h m
p2x + p2y
Z
dpx dpy
Eσ,i = cσ,i,j δ(px − p̄x,σ , py − p̄y,σ ) (17)
h2 2m
where p̄x,σ = p0x , for σ = 1, 4, 5, p̄x,2 = p0x −∆px , and p̄x,3 = p0x +∆px ; p̄y,σ =
p0y , for σ = 1, 2, 3, p̄y,4 = p0y − ∆py , and p̄y,5 = p0y + ∆py ; and cσ,i,j = ai,j ,
for σ = 1, and cσ,i,j = bi,j , if σ = 2, 3, 4, 5. The overall conservative quantities
P5
of gases in cell (i, j) are the summation of every beam ie. Qi,j = σ=1 Qσ,i,j .
The final 2D governing equations in generalized coordinates (ξ, η) based on
the beam splitting method can be expressed as:
200 Y. H. Shi and J. Y. Yang
Fig. 1. Number density isolines of quantum gas with the same fugacity in state 1
Fig. 1. Number density isolines of quantum gas with the same fugacity in state 1:
:Bose-Einstein gas(left),
Bose-Einstein Fermi-Dirac
gas(left), Fermi-Diracgas(right)
gas(right)
n
if ∆+ Qnσ,i 6=
[Chou97] Chou,
δσ,i+ 12 S.
σ̄(ā Baganoff,
= Y. & σ,i+
1 )(dσ,i+1
2
D.−(1997)
dσ,i )/∆Kinetic
+ Qσ,i , flux-vector 0
splitting for
(31)the
Navier-Stokes equations, 0, J. Comput. Phys. 130, pp.otherwise 217-230.
[Des86] Deshpande,
∆± Qnσ,i = ±(Q S. (1986)
n NASA-TM
σ,i±1 − Qσ,i )
n 01234. (32)
[Har83] Harten, A. (1983) High resolution schemes for hyperbolic conservation
where Comput.
laws,J. is a small positive
Phys., 49, real
357. number. The setting of the parameters ω and
ϑ inHarten,
[Har87] the TVD A., and ENO schemes
Engquist, B., Osher,areS. & Charkravathy, S. R. (1987) Uniformly
high order
ω =accurate
0, ϑ = essentially
0 → T Vnon-oscillatory
D2 schemes, III, J. Comput. Phys.,
71, 231.ω = 0, ϑ = 0.5 → EN O2 (33)
ω =C.
[Hir88] Hirsch, 1, (1988)
ϑ = 0 Numerical → EN O3 Computation of Internal and External Flows,
Vol. I & II, Wiley.
[KB62] Kadanoff, L. P. & Baym, G. (1962) Quantum Statistical Mechanics, Chap.
6.4Benjamin,
Numerical New York.
Examples and Discussions
[Ohw02] Ohwada, T. (2002) On the construction of kinetic schemes, J. Comput.
Phys. 177, pp. 156-175.
The numerical examples are performed in two dimension. The unsteady quan-
[Pul80]
tumPullin, D. I.wave
gas shock (1980)pass J. byComput. Phys.
a circular 34, 231-244.
cylinder. Initially, the shock with Mach
[PX93] Prendergast,
number Ms = 2.0 K.isH.ahead& Xu, K.(1993)
of the circularNumerical
cylinder. Ahydrodynamics
simple cylindrical from gas-
grid
kinetic
systemtheory,
of 361J.×Comput.
241 was Phys. used, 109, 53. of 361 rays around the cylinder
consisting
[Rei81]
andReitz, R. D.(1981)
241 circles between J. Comput. Phys.
the cylinder 42, 108.
surface and outer boundary which is
[SP74]slightly
Sanders, R. H. &
stretched withPrendergast,
∆rmin = .007. K. H.The 1974 The possible
diameter of the relation
cylinder ofis the 1.0 3-
kiloparsec arm to explosions
and the distance between the in the galactic
origin of thenucleus,
cylinderAstrophys. J. 188,
and the outer 489.
bound-
[Tor99]
aryToro,
is 7.0.E.The
F. (1999)
fugacity Riemann Solvers and Numerical
of the Bose-Einstein gas in theMethods for Fluid
left part(state Dy-
1) is
namics, Springer.
z1 = 0.8,and in right part (state 2) is 0.744. The corresponding macroscopic
[UU33] Uehling,are
quantities E. A.(ρ1&, uUhlenbeck,
1 , 1 , T1 ) =G. E. (1933)
(1.205, Transport
0.000, phenomena
0.603, 0.749), (ρ2 , u2 , in
2 , Einstein-
T2 ) =
Bose and1.000,
(2.410, Fermi-Dirac
4.214, 1.763).gases.It I,isPhys. Rev.that
observed 43, the
552.complicated flow interac-
[Xu13]tion
Xu, K. & Prendergast,
resulting in Mach shocks, K. H.second
(1994)contact
Numerical Navier-Stokes
discontinuities, andsolutions
triple point from
gas-kinetic theory, J.For
were well captured. Comput.
the case Phys. 114, 9. gas, The initial conditions are
of Fermi-Dirac
[YS06] Yang, J. Y. & Shi, Y. H. (2006) A kinetic Beam Scheme for Ideal Quantum
Gas Dynamics, Proc. Roy. Soc. A 462, 1553-1572.
202 Y. H. Shi and J. Y. Yang
set as Ms = 2.0 and fugacity z1 = 0.8 and the flow conditions (ρ1 , u1 , 1 , T1 ) =
(0.254, 0.000, 0.127, 0.432), (ρ2 , u2 , 2 , T2 ) = (0.508, 1.000, 0.889, 1.116) The fu-
gacity in state 2 is 0.576.
References
1. Chou, S. Y. & Baganoff, D. (1997) Kinetic flux-vector splitting for the Navier-
Stokes equations, J. Comput. Phys. 130, pp. 217-230.
2. Deshpande, S. (1986) NASA-TM 01234.
3. Harten, A. (1983) High resolution schemes for hyperbolic conservation laws,J.
Comput. Phys., 49, 357.
4. Harten, A., Engquist, B., Osher, S. & Charkravathy, S. R. (1987) Uniformly
high order accurate essentially non-oscillatory schemes, III, J. Comput. Phys.,
71, 231.
5. Hirsch, C. (1988) Numerical Computation of Internal and External Flows, Vol.
I & II, Wiley.
6. Kadanoff, L. P. & Baym, G. (1962) Quantum Statistical Mechanics, Chap. 6.
Benjamin, New York.
7. Ohwada, T. (2002) On the construction of kinetic schemes, J. Comput. Phys.
177, pp. 156-175.
8. Pullin, D. I. (1980) J. Comput. Phys. 34, 231-244.
9. Prendergast, K. H. & Xu, K.(1993) Numerical hydrodynamics from gas-kinetic
theory, J. Comput. Phys. 109, 53.
10. Reitz, R. D.(1981) J. Comput. Phys. 42, 108.
11. Sanders, R. H. & Prendergast, K. H. 1974 The possible relation of the 3-
kiloparsec arm to explosions in the galactic nucleus, Astrophys. J. 188, 489.
12. Toro, E. F. (1999) Riemann Solvers and Numerical Methods for Fluid Dynam-
ics, Springer.
13. Uehling, E. A. & Uhlenbeck, G. E. (1933) Transport phenomena in Einstein-
Bose and Fermi-Dirac gases. I, Phys. Rev. 43, 552.
14. Xu, K. & Prendergast, K. H. (1994) Numerical Navier-Stokes solutions from
gas-kinetic theory, J. Comput. Phys. 114, 9.
15. Yang, J. Y. & Shi, Y. H. (2006) A kinetic Beam Scheme for Ideal Quantum
Gas Dynamics, Proc. Roy. Soc. A 462, 1553-1572.
Semi-GLS Stabilization of FEM Applied to
Incompressible Flows with Higher Reynolds
Numbers
1 Introduction
Reliable modelling of flow for high Reynolds numbers, which appear in en-
gineering practice, is still not satisfactorilly solved. In the paper we employ
stabilization technique derived from Galerkin Least Squares (GLS) method.
In our paper [3] we modified the GLS method introduced in [7]. Our semi-GLS
algorithm alows to solve flows with markably higher Reynolds numbers than
without stabilization. Unfortunately the stabilization process leads to the loss
of accuracy, as was shown in [3]. In this paper we make use of a posteriori error
estimates of the finite element solution in order to trace the loss of accuracy of
the stabilized algorithm. We present some results of numerical experiments to
show the impact of this stabilization on the solution, namely on the accuracy
of the solution.
2 Model Problem
Let Ω be an open bounded domain in R2 filled with an incompressible viscous
fluid, and let Γ be its boundary. Isothermal flow of such fluid is governed by
the following Navier-Stokes system
∂u
+ (u · ∇)u − ν∆u + ∇p = f in Ω × [0, T ] (1)
∂t
∇ · u = 0 in Ω × [0, T ] (2)
u = g on Γg × [0, T ] (3)
−ν(∇u)n + pn = 0 on Γh × [0, T ] (4)
u = u0 in Ω, t = 0 (5)
where
• t denotes time variable
• u = (u1 , u2 )T denotes the vector of flow velocity
• p denotes the pressure divided by the density
• ν denotes the kinematic viscosity of the fluid supposed to be constant
• f denotes the density of volume forces per mass unit
• Γg and Γh are two subsets of Γ satisfying Γ = Γ g ∪ Γ h , µR1 (Γg ∩ Γh ) = 0
• n denotes an outer normal vector R to the boundary Γ with unit lenght
• g is a given function satisfying Γ g · ndΓ = 0 in the case of Γ = Γg
• u0 is a given flow field satisfying ∇ · u0 = 0
3 Stabilized Formulation
Following ideas of Hughes et al. [7], we apply the GLS stabilizing technique
with some modifications. Applying stabilization to the momentum equation
(1) and adding the continuity equation (2), we introduced in [3] the stabilized
algorithm that we call semi-GLS:
Find uh (t) ∈ Vgh , t ∈ [0, T ] and ph (t) ∈ Qh , t ∈ [0, T ] satisfying for any
t ∈ [0, T ]
BsGLS (uh , ph ; vh , ψh ) = LsGLS (vh , ψh ), ∀vh ∈ Vh , ∀ψh ∈ Qh (6)
uh − ugh ∈ Vh (7)
where
Z Z
∂uh
BsGLS (uh , ph ; vh , ψh ) ≡ · vh dΩ + (uh · ∇)uh · vh dΩ +
∂t
Z Z Ω Z Ω
ν ∇uh : ∇vh dΩ − ph ∇ · vh dΩ + ψh ∇ · uh dΩ +
Ω Ω Ω
N Z
X ∂uh
+ (uh · ∇)uh − ν∆uh + ∇ph · τ [(uh · ∇)vh − ν∆vh + ∇ψh ] dΩ ,
TK ∂t
K=1
Semi-GLS Stabilization of FEM for Flows with Higher Reynolds Numbers 205
Z N Z
X
LsGLS (vh , ψh ) ≡ f · vh dΩ + f · τ [(uh · ∇)vh − ν∆vh + ∇ψh ] dΩ .
Ω K=1 TK
4 Measuring of Error
1
E2 (u1h , u2h , ph , TK )
|TK |
to the solution norm on the whole domain Ω related to unit area of Ω
1
U2 (u1h , u2h , ph , Ω)
|Ω|
i.e.
5 Numerical Results
We presented a series of results for steady case of the problem of ‘lid driven
cavity’ in the paper [3]. Here we show the errors on elements measured by
means of a posteriori error estimates explained in Section 4. On Fig. 1 we
show the streamlines and pressure contours obtained by semi-GLS algorithm
on the mesh 128 × 128 elements for the cavity problem with Re = 10 000.
On Fig. 2 we present, for the same data, the errors on elements done by
a posteriori estimates. We can observe that the accuracy is violated not only
at the upper corners (where singularities arise) but also at other unexpected
places.
Unsteady flow past NACA 0012 airfoil was investigated in [3] as a practical
application. In [3] the results of Guermond and Quartapelle [5] were com-
pared to ours obtained by the semiGLS algorithm. To show the power of the
algorithm, we present in Fig. 3 the streamlines and pressure contours for the
problem with Reynolds number 100 000.
6 Conclusion
The developed algorithm is suitable to solve flow problems for higher Reynolds
numbers, approximately ‘double’ than the method without stabilization. In
[3], [8] a lot of numerical tests are given. They show that the stabilization has
also some negative effects and we have to be aware of them. Important tool
for testing the reliability of the numerical solution are the a posteriori error
estimates.
Acknowledgement: This work has been supported by the grant No.
1ET400760509 of the Information Society Program of the Czech Acad. Sci.
and by the State Research Project No. MSM 684 0770010.
Semi-GLS Stabilization of FEM for Flows with Higher Reynolds Numbers 207
Fig. 1. Streamlines (left) and pressure contours (right) by the semi-GLS algorithm,
Re = 10 000, mesh 128 × 128
0.8
0.6
AEE
6
5
4
0.4 3
2
1
0
0.2
0
0 0.2 0.4 0.6 0.8 1
Fig. 2. A posteriori estimates error: cavity, semi-GLS, Re= 10 000, mesh 128 × 128
208 Pavel Burda, Jaroslav Novotný, and Jakub vSı́stek
Fig. 3. Streamlines (left) and pressure contours (right) by the semi-GLS algorithm,
t = 3.6s, Re = 100 000
References
1. Burda, P., Novotný, J., Sousedı́k, B.: A posteriori error estimates applied to
flow in a channel with corners. Mathematics and Computers in Simulation, 61,
375-383 (2003)
2. Burda, P., Novotný, J., vSı́stek, J.: Precise FEM solution of a corner singularity
using an adjusted mesh. Int. J. Numer. Meth. Fluids, 47, 1285–1292 (2005)
3. Burda, P., Novotný, J., vSı́stek, J.: On a modification of GLS stabilized FEM for
solving incompressible viscous flows. Int. J. Numer. Meth. Fluids, 51, 1001–1016
(2006)
4. Franca, L.P., Madureira, A.L.: Element diameter free stability parameters for
stabilized methods applied to fluids. Comput. Methods Appl. Mech. Engrg.,
105, 395–403 (1993)
5. Guermond, J.L., Quartapelle, L.: Calculation of viscous incompressible viscous
flow by an unconditionally stable projection FEM. J. Comp. Phys., 132, 12–33
(1997)
6. Hughes, T.J.R., Franca, L.P., Balestra, M.: A new finite element formulation for
computational fluid dynamics: V. Circumventing the Babuvska-Brezzi condition:
A stable Petrov-Galerkin formulation of the Stokes problem accommodating
equal-order interpolations. Comput. Methods Appl. Mech. Engrg., 59, 85-99
(1986)
7. Hughes, T.J.R., Franca, L.P., Hulbert, G.M.: A new finite element formulation
for computational fluid dynamics: VIII. The Galerkin/least-squares method for
advective-diffusive equations. Comput. Methods Appl. Mech. Engrg., 73, 173-
189 (1989)
8. vSı́stek, J.: Stabilization of finite element method for solving incompressible
viscous flows. MSc. Thesis, Czech University of Technology, Praha (2004)
Finite volume box scheme for a certain class of
nonlinear conservation laws in mixed form
1 Introduction
∀e ∈ Eb ∩ ∂Ω2 , uh · ne = 0} ,
where RT 0 (T ) = [P 0 (T )]2 +(x, y)t P 0 (T ), and that the non-conforming Crouzeix-
Raviart finite element space is defined as
Z
Pnc,∂Ω1 = ph ∈ L (Ω); ∀T ∈ Th , ph |T ∈ P (T ); ∀e ∈ Ei , [ph ]e = 0;
1 2 1
e
Z
∀e ∈ Eb ∩ ∂Ω1 , ph = 0 ,
e
pnh ∈ Pnc,∂Ω
1
Seek
Z 1
such that
−1
δt Πh ψ(ph ) − ψ(pn−1
n
h ) qh +
Ω Z Z (5)
X
n n n
Πh (f n )qh , 1
[A(p )∇p + B(p )] · ∇q = ∀qh ∈ Pnc,∂Ω .
h
h h h 1
T ∈Th T Ω
Statement 3 in the previous theorem means that starting with the discrete
primal variable solving (5), the local flux reconstruction formula (6) yields a
locally conservative flux. Observe that a locally conservative reconstruction of
the discrete flux is, in general, not possible when working with more standard
P 1 conforming finite elements to approximate the primal variable.
Error L2L2
-1
-Z (m)
0.1
-1.2
-1.4
-1.6 0.01
-4.5 -4 -3.5 -3 -2.5 -2 -1.5 -1 -0.5 0 10 20 30 40 50 60 70 80
Pressure head (m) Number of nodes
Fig. 1.
0 0 0
Fig. 3.
0.45 10 100
outflow box scheme
inflow 1 conformal scheme
0.4 8
0.01
6
maxdiv
0.35 1e-04
theta
Flux
Fig. 4.
Conclusions
Keeping the mesh fixed, the finite volume box scheme is more expensive than
the P 1 conforming finite element scheme. However the finite volume box
scheme uses the degrees of freedom more efficiently and ensures local mass
conservation. Further work will deal with the following question: how much
is a conservative flow field needed for accurate, low-order pollutant transport
simulations?
References
1. Courbet, B. and Croisille, J.-P.: Finite volume box schemes on triangular meshes,
Math. Mod. Numer. Ana., 32(5), 631-649, June (1998)
2. Croisille, J.-P.: Finite volume box schemes and mixed methods, Math. Mod.
Numer. Ana., 31(5), 1087–1106, (2000)
3. Marini, L.D.:An inexpensive method for the evaluation of the solution of the
lowest order Raviart-Thomas mixed method, SIAM J. Num. An., 22, 493-496
(1985)
4. Ginzburg, I., Carlier, J.-P. and Kao, C.: Lattice Boltzmann approach to
Richards’ equation, CMWR Proceedings, North Carolina, USA, (2004)
5. Beaugendre, H. , Ern, A., Esclaffer, T., Gaume, E., Ginzburg, I. and Kao, C.: A
seepage face model for the interaction of shallow water tables with the ground
surface: Application of the obstacle-type method, J. of Hydrology, April (2006)
Numerical study of the Colocated Clustered
Finite Volume Scheme
Summary. The classical formulation of the finite volume scheme for incompressible
flows in simple geometries is based on orthogonal grids with staggered meshes [1]
to avoid the growth of unphysical oscillations, also called ”spurious checkerboards
modes”. For more complex domains, the use of unstructured meshes seems to be
more judicious but it raises the problem of the definition of the staggered grids. In
case of unstructured meshes, a colocated approach is preferred. This arrangement has
become popular since the works of Rhie and Chow [3] that formulated a numerical
method ”Momentum interpolation” to avoid the apparition of unphysical pressure
oscillations. Numerous finite volume schemes have been performed with the Rhie and
Chow interpolation or its variants, as for example the CFD code Fluent [4]. This
paper presents a colocated and clustered finite volume method for unstructured grids
and some numerical tests.
Let us consider an ideal gas in an enclosure with impervious walls. The di-
mensionless low Mach number equations:
∂
ρ + ∇.(ρv) = 0 (1)
∂t
∂ RaP r
(ρv) + ∇.(ρv ⊗ v) = −∇p − ρj + P r∇.τ̄¯ (2)
∂t εT
∂ γ − 1 dP
(ρθ) + ∇.(ρθv) = + ∇.(k∇θ) (3)
∂t εT γ dt
Z Z
P (t)
dx = m(Ω), p(x, t)dx = 0 (4)
Ω 1 + εT θ Ω
2 Numerical scheme
The discrete equations are obtained by integrating
the continuous equations over each control volume
K and by applying the Gauss theorem. To make
easier the presentation of the numerical scheme
and to take into account the steadiness of the
solutions presented in this paper, only the space
discretization for two-dimensional domains is pre-
sented hereafter (for more details, refer to paper
[5],[6]).
One of the main objectives conducing to elabo-
rate the numerical scheme was the simplicity. The Fig. 1. Notations
simplicity is first due to the choice of a colocated
scheme: the discrete unknowns are evaluated at
the same point xK , representative of the control volume K and usually called
centre (see Fig.1). Moreover, the approximation of the flux of a scalar vari-
able, normal to the edge of a control volume σKL , only depends on the two
cells K and L having this edge in common. This particular choice of normal
flux induces two requirements:
• each interface between two adjacent control volumes is orthogonal to the
line joining their centres (σKL ⊥(xK xL )),
• xK xL · nKL > 0 with nKL being the unit vector normal to the edge σKL .
These conditions are suitable for the equations of transport-diffusion such
as the energy equation but they are not sufficient to ensure a satisfactory
convergence for the Stokes problem. To obtain a good coupling between the
velocity and the pressure field, the straight line (xK xL ) must intersect at the
centre of the edge σKL at the point xKL . Therefore, triangles can be chosen as
control volumes and the circumcenters as centre of the cells. The requirement
xK xL · nKL > 0 is automatically verified once the triangulation satisfies the
Delaunay criterion.
Numerical study of the Colocated Clustered Finite Volume Scheme 217
To avoid the work of the pressure forces, the discrete pressure gradient
must be the transpose of the divergence term in discrete mass equation.
The coefficients bKL and bLK being defined by the relations 5, the expres-
sion bKL pL + bLK pK is obviously not equal to the pressure on the edge,
except when bKL = 0.5.
Non-linear inertial force: TKL = σKL (ρv(x) · nKL ) φ ds(x)
R
φK + φL
TKL = MKL
2
To satisfy the kinetic energy balance the two following conditions are
required:
1. the edge velocity must include the stabilization term ρKL m
dKL (pL −
KL
pK ),
218 O. Touazi, E. Chénier, and R. Eymard
3 Numerical study
Spatial accuracy. The spatial accuracy of the numerical method is checked
by simulating the natural convective flow, under the Boussinesq approx-
imation, in a [0, 1]-square enclosure with isothermal rigid walls and for
P r = 1, Ra = 1 and λ = 10−3 . The chosen analytical solution is: u(x, y) =
2π cos(πy) sin(πy) sin2 (πx), v(x, y) = −2π cos(πx) sin(πx) sin2 (πy), θ(x, y) =
sin2 (πx) sin2 (πy) and p(x, y) = sin2 (πx) sin2 (πy).
The accuracy of the numerical method presented in Fig.3 is measured by the
L∞ norm (kφK − φ(xK )k∞ = maxK |φK − φ(xK )| with φ ∈ {v (j) , p}) and H1
Numerical study of the Colocated Clustered Finite Volume Scheme 219
Fig. 2. Meshes, (a) Squares, (b) Regular triangles (c) Random triangles
(j)
norm (kvK − v (j) (xK )k2H1 = √K mK i (Gji (xK ) − Gji (xK ))2 ) as a function
P P
of the average mesh size h = 1/ n and for three kinds of meshes consisting of
n squares (Fig. 2(a)) or regular triangles (Fig. 2(b)) or random triangles (Fig.
2(c)). For the random grid and a fixed number of cells, a series of 50 computa-
tions were performed to obtain an average values for each norm. As foreseen,
the numerical results deteriorate with the deconstruction of the mesh. Notice
that about 2% and 50% of the cell centres are located outside of their con-
trol volumes for the regular and random triangles respectively. However, the
orders of convergence for the velocity and pressure in L∞ and H1 norms are
quite independent of the cell-shape.
Low mach number comparison. The differentially heated cavity is simu-
lated and results are compared to a benchmark (Tab. 1). The physical param-
eters are described in [10]. The distribution of the cell sizes of the rectangular
grid is sinusoidal with finer mesh in the boundary layer. The triangular mesh
consists of triangles quite regular. This mesh is therefore fewer appropriate
log10(||e(p)||)max
-0.5
1
-2
log10(||u||)H
-0.5
Square (1)
-2.5 -1 -1 Regular triangles (1)
-1.5 Random triangles (0.9)
-3
-2 -1.5
-3.5
-2.5
-2
-4 -3
-4.5 -3.5 -2.5
-2.8 -2.6 -2.4 -2.2 -2 -1.8 -1.6 -3 -2.5 -2 -1.5 -1 -0.5 -2.8 -2.6 -2.4 -2.2 -2 -1.8 -1.6
log10(h) log10(h) log10(h)
1 (1)
kvK − v 1 (xK )k∞ kvK − v (1) (xK )k2H1 kpK − p(xK )k∞
than rectangles to represent the solutions in the boundary layers where the
longitudinal and transversal characteristics of the flow vary on very different
scales. This remark explains the less accuracy obtained with triangles than
rectangles. It can be however noticed than the relative discrepencies to the
Vierendels results are good and less than 1.1%.
220 O. Touazi, E. Chénier, and R. Eymard
References
1. Harlow, F., Welch J.: Numerical Calculation of Time-Dependent Viscous In-
compressible Flow of Fluid with a Free Surface. Phys. Fluids, 8, 2182–2189
(1965)
2. Patankar, S.V : Numerical Heat Transfer and Fluid Flow. MacGraw-Hill, New-
York, (1980)
3. Rhie, C.M, Chow, W.L.: Numerical Study of the Turbulent Flow Past an Airfoil
with Trailing Edge Separation. AIAA J., 21, 1523–1532 (1983)
4. Mathur, S.R, Murthy, J.Y.: A Pressure-Based Method for Unstructured Meshes.
Num. heat Transfer, part B, 31, 195–215 (1997)
5. Chénier, E., Eymard, R., Touazi, O.: Numerical results using a Colocated Finite
Volume Scheme on Unstructured Grids for Incompressible Fluid Flows. Num.
Heat Transfer, part B, 49, 259–276, (2006)
6. Touazi, O., Chénier, E., Eymard, R.: Simulation of natural convection with the
Clustered Finite Volume Scheme. Computers & Fluids, submitted (2006)
7. Brezzi, F., Pitkaranta, J.: On the Stabilization of Finite Element Approxi-
mations of the Stokes Equations. Notes Numer. Fluid Mech., Vieweg, Braun-
schweig, 10 (1984)
8. Kechkar, N., Silvester, D.: Analysis of Locally Stabilized Mixed Finite Element
Methods for the Stokes Problem. Mathematics of Computation, 58(197), 1–10,
(1992)
9. Eymard, R., Gallouët, T., Herbin, R.: A cell-centred finite volume approxima-
tion for anisotropic diffusion operators on unstructured meshes in any space
dimension. IMAJNA, accepted for publication (2005)
10. Paillère, H., Le Quéré, P., Weisman, C., Vierendeels, J., Dick, E., Braack, M.,
Dabbene, F., Beccantini, A., Studer, E., Kloczko, T., Corre, C., Heuveline,
V., Darbandi, M., Hosseinizadeh, S.F.: Modelling of natural convection flows
with large temperature differences: A benchmark problem for low Mach number
solvers.ESAIM, Math. Model. Numer. Anal., 617–621 (2005)
Arbitrary High Order Finite Volume Schemes
on Unstructured Meshes
1 Introduction
The spatial and temporal accuracy of the final scheme depends merely on
the order of the argument of the numerical flux used in (2). The key point
is the reconstruction operator presented in [2] that makes use of techniques
developed originally in the discontinuous Galerkin finite element framework.
First, it uses a hierarchical orthogonal basis to perform reconstruction. Second,
reconstruction is not done in physical coordinates, but in a reference coordi-
nate system which eliminates scaling effects and thus avoids ill-conditioned
reconstruction matrices. Third, the proposed algorithm does not reconstruct
point-values but entire polynomials which can be easily evaluated and differ-
entiated at any point, which is crucial for an efficient implementation of the
ADER approach that essentially relies on spatial derivative information. To
achieve non-oscillatory properties for nonlinear hyperbolic systems, we extend
the component-wise WENO reconstruction presented in [2] to characteristic
reconstruction. Numerical fluxes at the element interfaces are then computed
solving generalized Riemann problems (GRPs), according to the ADER ap-
proach of Toro et al., [9, 10, 8]. The resulting one-step ADER finite volume
scheme obtained in this way performs only one nonlinear WENO reconstruc-
tion per element and time step and can thus be implemented very efficiently
even for unstructured grids in three space dimensions. We emphasize that the
method is of uniform high order of accuracy in space and time. The surface
integral appearing in (2) is approximated using Gaussian quadrature formulae
of suitable order of accuracy.
The linear weights are chosen according to [2], putting a large linear weight
λ1 >> 1 on the central stencil and λi = 1 else. Finally, we obtain for each
characteristic variable for each characteristic direction nj a reconstruction
polynomial of degree M that is obtained from a nonlinear combination of the
reconstruction polynomials on each stencil i:
ns
(WENO,nj ) (i,nj )
X
ĉpl = ωp(i,nj ) ĉpl , ∀0 ≤ j ≤ NE . (8)
i=1
Since in our approach the final result of the reconstruction operator is one
single reconstruction polynomial defined in the whole element, we still have
to perform a selection among the various characteristic reconstruction poly-
nomials. Therefore, we transform all characteristic WENO polynomials back
to conservative variables, i.e.
(WENO,nj ) nj (WENO,nj )
ŵpl = Rpq · ĉql , ∀0 ≤ j ≤ NE , (9)
224 Michael Dumbser and Claus-Dieter Munz
and then choose for each conservative variable the smoothest polynomial ob-
tained from all the different characteristic reconstructions. First, the oscilla-
tion indicators for the polynomials (9) are computed in (10),
(WENO,nj )
σpnj = ŵpl (WENO,nj )
· Σlm · ŵpm , (10)
then the smoothest one is selected according to an ENO-type algorithm:
( nN
n
nj 1 if σp j = min σpn0 , σpn1 , ..., σp E ,
ωp = (11)
0 else.
The final reconstruction polynomial for the conservative variables valid in the
whole element T (m) is thus given by
NE
(WENO,nj )
X
WENO
ŵpl = ωpnj ŵpl . (12)
j=0
Z Z Z
Y X Y X Y X
0 0 0
z
z
-0.5 0.5 -0.5 0.5 -0.5 0.5
0.5 0.5 0.5
0 0 0
0 x 0 x 0 x
y y y
-0.5 -0.5 -0.5 -0.5 -0.5 -0.5
To assess the accuracy of the proposed method, we solve the three di-
mensional system of Euler equations with source term in the computational
3
domain Ω = − 21 ; 12 with six periodic boundary conditions using a sixth
NG L∞ L1 L2 O L∞ O L1 O L2
16 1.0136E-04 1.0987E-05 1.9043E-05
20 3.2029E-05 4.1034E-06 7.1727E-06 5.2 4.4 4.4
24 1.1199E-05 1.3482E-06 2.3640E-06 5.8 6.1 6.1
32 1.7327E-06 2.0130E-07 3.5644E-07 6.5 6.6 6.6
226 Michael Dumbser and Claus-Dieter Munz
Fig. 2. Results for the double Mach reflection problem obtained with a third order
1 1
ADER-FV scheme on different grids: h = 100 (top), h = 400 bottom.
this rolling process due to a lack of physical viscosity, this test case is com-
monly used to assess the numerical viscosity of a scheme. The more rolling the
less the numerical viscosity. We compute this test case on differently refined
meshes, where the mesh spacing h is the edge length at the wall boundary. At
the boundaries, we use purely one-sided reconstruction stencils and then solve
inverse generalized Riemann problems to prescribe the correct flux at the wall,
see [4]. This is done by symmetry considerations, where point-symmetry is ap-
plied for the normal velocity and axial-symmetry for all other components.
Comparing our results with those presented in [5] we note that our third
order solution shows more rolling on the finest grid, i.e. less numerical dissi-
pation. This is probably due to the following facts: in [5] a third order TVD
Runge-Kutta time integration scheme was used which implies three nonlinear
Arbitrary High Order Finite Volume Schemes on Unstructured Meshes 227
WENO reconstructions per time step with their associated limiting, whereas
in the ADER approach only one WENO reconstruction is performed per time
step. Furthermore, the third order method presented in [5] relies on an optimal
pointwise WENO reconstruction based on the combination of first order poly-
nomials. Our third order reconstruction, however, is based on a combination
of second order polynomials.
References
1. D. Balsara and C.W. Shu. Monotonicity preserving weighted essentially non-
oscillatory schemes with increasingly high order of accuracy. Journal of Com-
putational Physics, 160:405–452, 2000.
2. M. Dumbser and M. Käser. Arbitrary high order non-oscillatory finite vol-
ume schemes on unstructured meshes for linear hyperbolic systems. Journal of
Computational Physics, 2006. in press.
3. M. Dumbser and C.D. Munz. Building blocks for arbitrary high order discon-
tinuous Galerkin schemes. Journal of Scientific Computing, 27:215–230, 2006.
4. M. Dumbser. Arbitrary High Order Schemes for the Solution of Hyperbolic
Conservation Laws in Complex Domains. Shaker Verlag, Aachen, 2005.
5. C. Hu and C.W. Shu. Weighted essentially non-oscillatory schemes on triangu-
lar meshes. Journal of Computational Physics, 150:97 – 127, 1999.
6. G.-S. Jiang and C.W. Shu. Efficient implementation of weighted ENO schemes.
Journal of Computational Physics, 126:202–228, 1996.
7. M. Käser and A. Iske. ADER schemes on adaptive triangular meshes for scalar
conservation laws. Journal of Computational Physics, 205:486–508, 2005.
8. T. Schwartzkopff, M. Dumbser, and C.D. Munz. Fast high order ADER schemes
for linear hyperbolic equations. Journal of Computational Physics, 197:532–539,
2004.
9. E.F. Toro and V. A. Titarev. Solution of the generalized Riemann problem for
advection-reaction equations. Proc. Roy. Soc. London, pages 271–281, 2002.
10. V.A. Titarev and E.F. Toro. ADER schemes for three-dimensional nonlinear
hyperbolic systems. Journal of Computational Physics, 204:715–736, 2005.
“This page left intentionally blank.”
Part III
Algorithms
“This page left intentionally blank.”
A high scalability parallel algebraic multigrid
solver
1 Introduction
To accelerate the convergence of the outer loop effort was expanded to de-
velop more efficient algorithms [3] to treat the inter-equation coupling of the
momentum and pressure equations either in a segregated manner [4] or more
implicitly [5]. Effort to accelerate the inner loop, have led to the development
of geometric multigrid methods in the 1970’s and later in the 1980’s to alge-
braic multigrid methods (AMG) [6] that extended the main ideas of geometric
multigrid methods to purely algebraic settings, yielding more robustness and
algorithmic simplicity. This dual approach sequential has been quite success-
ful in tackling increasingly larger cfd simulations, but the reliance of many
industries on using CFD to simulate large scale engineering problems has led
to a more drastic approach to decrease simulation time, i.e parallel processing,
which enables the scaling of the CFD problem through the decomposition of
the problem into as many sub-problems as there are processors. The Paral-
lelization of CFD codes can be achieved in several ways however when the
target of the parallelization effort is clusters of interconnected processors [7]
then the Domain Decomposition approach is remains the most effective. In
Domain Decomposition the parallelization is enforced by dividing (partition-
ing) the domain of interest into a number of sub-domains or partitions (usu-
ally one for each processor). In this case inter-domain synchronization where
neighboring sub-domains swap solution information is essential in ensuring
a consistent solution. The synchronization can be performed on one extreme
only at the finest mesh, with the multigrid solver basically playing the role of
its sequential counterpart over the sub-domain or partition, or on the other
extreme the synchronization can be performed at each multigrid level, with
the multigrid solver then playing a additional role of smoother across parti-
tions but at the expense of additional communication cost between partitions
[8]. The subject of this paper is the development of a highly scalable Alge-
braic Multigrid solver. After a brief review of the different synchronization
strategies possible across partitions results for the performance of the pAMG
solver for different grid sizes and partition numbers are presented.
4 Test Problems
∇ · (Γ ∇φ) = 0 (1)
Γ = φ0.1 (2)
where the various boundary conditions are shown in Figure 1. Four un-
structured meshes were used for the geometric discretization; 99,454 - 199,222
- 298,154 - 397,393 elements. The runs were performed on a cluster of ten G5
dual processor machines running at 2.0 GHz each with a memory of 512 MB
per node. The interconnection network is a 100-MBits Ethernet switch. The
runs were made until the residual was reduced to a value of 10−4 .
234 T. Saad and M. Darwish
4.1 speedup
The speedup, for a given number of processors P, is defined as the ratio of the
solution time of the sequential algorithm to the solution time on P processors.
Fig. 2. Speedup
Figure 2 shows the speedup obtained for the four meshes using up to 16
processors. For the 99,454 elements mesh, the scalability starts deteriorating
A high scalability parallel algebraic multigrid solver 235
The communication and computational costs are directly related to the num-
ber of shadow elements and the number of core elements since the computation
takes place over the core elements and the communication takes place over
the shadow elements respectively.
Figure 3 shows the efficiency versus the maximum shadow-to-core ra-
tio.This graph basically characterizes the relation between the maximum
shadow-to-core ratioand the efficiency of the parallel solver. For the 99,454
elements mesh, as the shadow-to-core ratio exceeds 0.008, the efficiency starts
to decrease. For the 199,222 elements mesh, the efficiency starts to degenerate
at a shadow-to-core ratio exceeding 0.02. The same happens for the other two
meshes. In general, it is noticed that when the shadow-to-core exceeds - on
average - a value of 0.02, the efficiency starts to deteriorate.
5 Conclusion
6 Acknowledgment
The authors wish to acknowledge the generous support of the Lebanese Na-
tional Council for Scientific Research (LNCSR).
236 T. Saad and M. Darwish
References
1. Bucker H.M.: Iteratively solving large sparse linear systems on parallel com-
puters. NIC Serices, John Von Neumann Institute for Computing, Julich, 10,
521-548 (2002)
2. Patankar S.V.: Numerical heat transfer and fluid flow. Washington; New York:
Hemisphere Pub. Corp.; McGraw-Hill, 1980
3. Moukalled F., Darwish M.S.: A Unified Formulation of the Segregated Class
of Algorithms for Fluid Flow at All Speed, Num. Heat Transfer, 37, 103-139
(2000)
4. Issa, R.I.: Solution of the Implicit Discretized Fluid Flow Equations by Operator
Splitting, Mechanical Engineering Report, FS/82/15, Imperial College, London,
(1982)
5. Webster R.: Iteratively solving large sparse linear systems on parallel comput-
ers. Int. J. Numer Meth. Fluids, 36, 743-773 (2001)
6. Brandt A. ’Algebraic Multigrid Therory: THe symmetric Case’ Appl. Math.
COmput, vol 19, pp.24-56, 1986.
7. V. Dolean and S. Lanteri, ”Parallel multigrid methods for the calculation of
unsteady flows on unstructured grids: algorithmic aspects and parallel perfor-
mances on clusters of PCs,” Parallel Computing, vol. 30, pp. 503-525, 4. 2004
8. Mitchell W. ’A parallel multigrid methoid using the full domain partition’,
Electron. Trans. Numer. Anal., vol. 6, pp. 224-233, 1998
9. A. Krechel and K. Stüben. Parallel algebraic multigrid based on subdomain
blocking. Parallel Computing, 27:1009-1031, 2001
10. Gropp W. D., ’Parallel computing and domain decomposition’, in T. F. Chan,
D. E. Keyes, G. A. Meurant, J. S. Scroggs, and R. G. Voigt, eds., Fifth Confer-
ence on Domain Decomposition Methods for Partial Differentia equations, pp.
349-362, SIAM, 1992.
Jacobian-Free Newton-Krylov Methods:
Issues and Solutions
1 Introduction
The present algorithm originated with the work of Pueyo and Zingg [6], who
developed a Newton-Krylov algorithm applicable to the computation of aero-
dynamic flows on single-block structured meshes using scalar artificial dissi-
pation and the Baldwin-Lomax turbulence model. The algorithm has since
been extended to incorporate matrix dissipation, the Spalart-Allmaras one-
equation turbulence model, including trip terms for laminar-turbulent transi-
tion, and multi-block meshes.
Our algorithm is based on a Jacobian-free inexact-Newton strategy using
ILU-preconditioned GMRES, as described above. There are numerous param-
eters involved, such as the tolerance to which the inner iterations are solved,
the amount of fill permitted in the incomplete factorization, and a parameter
used in forming the approximate Jacobian upon which the ILU factorization
is based. These have been carefully chosen based on numerous experiments to
maximize efficiency while maintaining robustness, such that a fixed set of pa-
rameters can be used for a broad range of flow conditions, including complex
flows over multi-element geometries. Considerable effort has been put into op-
timizing the pseudo-transient phase. In the original algorithm of Pueyo and
Zingg [6], an approximate factorization algorithm was used initially. This has
been replaced by a Newton-Krylov pseudo-transient scheme with gradually
increasing spatially-varying time steps. In particular, it is critical to maintain
positivity of the Spalart-Allmaras turbulence model variable, since the model
can become unstable when negative values occur. The technique suggested
by Spalart and Allmaras [17] to maintain positivity involving M-matrices is
incompatible with Newton-like convergence. We use an alternative approach
based on a separate local time step for the turbulence model designed to main-
tain positivity during the pseudo-transient phase. Full details of the present
algorithm are deferred to a forthcoming paper and thesis [18].
Two examples are shown in Fig. 1. Convergence of the error in the lift co-
efficient3 is plotted as a function of both CPU time (top axis label) and equiv-
alent function evaluations, i.e. CPU time divided by the time required for one
right-hand side evaluation (bottom axis label). The present results are labelled
“NK” and, in order to provide a well-known reference, are contrasted with
the convergence of an approximate-factorization algorithm (labelled “AF”)
3
Defined as the difference between the current lift coefficient and the fully con-
verged lift coefficient.
240 David W. Zingg and Todd T. Chisholm
0.01 0.01
0.001 0.001
Cl error
Cl error
1e-04 1e-04
1e-05 1e-05
1e-06 1e-06
1e-07 1e-07
1e-08 1e-08
0 200 400 600 800 1000 1200 1400 0 2000 4000 6000 8000 10000
Equivalent RHS evaluations Equivalent RHS evaluations
180
160
140
of zero indicates nonconvergence. The figure shows that root nodes upstream
of the airfoil can lead to varied performance of the algorithm, while root nodes
located above and below often lead to nonconvergence. It is clear that it is
best to select a root node directly downstream of the airfoil. This choice has
been used for the results shown previously.
Another problem arises if the magnitude of the residual of one equation (e.g.
the turbulence model) is much larger than that of the others. The exit toler-
ance for the linear solver can then be satisfied if the larger residuals are re-
duced, even though the other equations have not converged sufficiently, thus
preventing convergence of the Newton iterations. This problem can be di-
agnosed by examining the residual norms of each equation separately. The
equations can be suitably scaled such that the residuals are of the same order.
We also perform a minimum of five inner iterations per outer iteration.
The choice of the parameter in (1) is a balance between round-off and trun-
cation errors. Many authors use
√
m
= (2)
||v||2
where m is the value of machine zero. Problems arise when there exist large
differences in the magnitudes of the entries in Q but not in v or vice versa.
The former occurs when the variables are poorly scaled, the latter when the
residual equations are poorly scaled. Poor scaling of the variables can occur
for various reasons. In our case, the nondimensionalization of the turbulent
242 David W. Zingg and Todd T. Chisholm
viscosity using the actual viscosity leads to values of the turbulence parameter
that greatly exceed the nondimensional mean-flow variables. It can then be
difficult or impossible to find a suitable value of that leads to sufficiently low
truncation error and round-off error. This difficulty is particularly insidious
because the linear iterations appear to converge. However, since the matrix-
vector products are not accurate, GMRES is converging to the solution of
a different linear problem. The problem can be diagnosed by recalculating
the linear residual after the GMRES iterations have terminated. Note that
GMRES does not directly evaluate the residual at each iteration, but instead
finds it indirectly, assuming that the matrix-vector products are accurate. If
the residual calculated subsequently differs from that reported by GMRES,
then the Jacobian-free matrix-vector products are inaccurate. The solution
is to rescale both the variables and the residual equations on the fly so that
both are well scaled. In addition, we use a value of that is two orders of
magnitude greater than that given by (2) to reduce round-off error.
5 Conclusions
A number of subtle problems that can arise when Newton-Krylov algorithms
are applied to the Reynolds-averaged Navier-Stoes equations have been dis-
cussed. Strategies have been presented for identifying and addressing these
issues.
References
1. Knoll, D.A., and Keyes, D.E., J. Comp. Phys., 193:357-397, 2004.
2. Venkatakrishnan, V., and Mavriplis, D.J., J. Comp. Phys., 105:83-91, 1993.
3. Degrez, G., and Issman, E., VKI Lecture Notes 1994-05, 1994.
4. Barth, T.J., and Linton, S.W., AIAA Paper 95-0221, 1995.
5. Nielsen, E.J., Anderson, W.K., Walters, R.W., and Keyes, D.E., AIAA Paper
95-1733, 1995.
6. Pueyo, A., and Zingg, D.W., AIAA J., 36:1991-1997, 1998.
7. Geuzaine, P., Lepot, I., Meers, F., and Essers, J.A., AIAA Paper 99-3341,1999.
8. Saad, Y., and Schultz, M.H., SIAM J. Sci. Stat. Comput., 7:856-869, 1986.
9. Nemec, M., and Zingg, D.W., AIAA J., 40:1146-1154, 2002.
10. Chisholm, T.T., and Zingg, D.W., AIAA Paper 2003-3708, 2003.
11. Nemec, M., Zingg, D.W., and Pulliam, T.H., AIAA J., 42:1057-1065, 2004.
12. Isono, S., and Zingg, D.W., AIAA Paper 2004-0433, 2004.
13. Smith, T.M., Hooper, R.W., Ober, C.C., Lorber, A.A., and Shadid, J.N., AIAA
Paper 2004-0743, 2004.
14. Luo, H., Baum, J.D., and Lohner, R., AIAA Paper 2004-1103, 2004.
15. Nichols, J.C., and Zingg, D.W., AIAA Paper 2005-5230, 2005.
16. Mulder, W.A., and van Leer, B., J. Comp. Phys., 59:232-246, 1985.
17. Spalart, P., and Allmaras, S., AIAA Paper 92-0439, 1992.
18. Chisholm, T.T., Ph.D. Thesis, University of Toronto Institute for Aerospace
Studies, 2006.
Non-stationary two-stage relaxation based on
the principle of aggregation multi-grid
1 Introduction
Huge algebraic systems of equations are routinely solved by iterative meth-
ods. Some of these use a relaxation paramater (SOR, damped Jacobi,...),
the choice of which is sometimes awkward. Often these methods are used as a
smoother for a multi-grid method. In [1] it was shown that a V-cycle multi-grid
method using the Galerkin construction will converge whenever the smoother
converges, which depends on the relaxation parameter. We propose a proce-
dure to automatize the choice of the parameter and limit ourselves initially
to (semi-iterative) damped Jacobi and related methods as smoothers for a
multi-grid method. The idea is to add a small number of lines to an existing
code as a patch. Emphasis is put on simplicity and low cost and applicability
to general cases of matrices to guarantee convergence independently of the
input relaxation parameters (or at least to reduce its influence). We compare
different variations of the idea on the advection-diffusion equation.
2 The algorithm
Vectors will be written in boldface and matrices in capital letters. We are
interested in solving linear algebraic equations of the type
Au = f (1)
We assume that A ∈ IRn×n is non-singular; u, f ∈ IRn×1 . We will solve this
system with an iterative procedure (also called relaxation scheme). The latter
typically splits the matrix A in two parts: A = M − N , where M is non-
singular. The solution is found by iteratively applying
M um+1 = N um + f (2)
starting from an initial guess uo . This will also be symbolically written as
um+1 = S(um ) (3)
2.3 Cost
The dominant cost of an iterative scheme is the matrix-vector (M-V) multi-
plication. We will allocate one workunit (WU) to a M-V multiplication and
neglect the cost of the scalar products. We can quantify the cost of a damped
relaxation as 2 WU (predictor step). The corrector step (κ-modification) also
requires two M-V multiplications. The cost of the multi-grid defect correction
is more difficult to quantify, depending on the real cost of the transfer opera-
tors. As we are only interested in the robustness (convergence) of the scheme,
and all variants use the same two-grid defect correction, the exact cost is a
moot point and can be estimated as 3 WU.
3 Results
The equations under considerations are the two dimensional advection-diffusion
equations:
cost, we lost a lot of the robustness. The modified MG7 (called MG7e), being
independent of the input parameters, gave very good results when we fixed
κ after 4 iterations, being better than the baseline, except for a very narrow
range of parameters where it was roughly equivalent. We show the results for
ω1 = ω2 in figure 3 As the scheme is applicable to any damped smoother
we also try it on damped Symmetric Gauss-Seidel instead of Jacobi. The
schemes with the most stable output were MG3,5 and 7. Again they would
need to have the computation of κ frozen after some iterations to make them
computational competitive with MG1 using the optimal combinations of the
relaxation parameters.
4 Conclusions
We used a simple aggregation procedure using a single aggregate to automatize
the choice of relaxation parameters. We applied the scheme to a damped
Jacobi and damped Gauss-Seidel iterative method as a smoother for multi-
grid, allowing for two different relaxation parameters during the iteration in
a periodic fashion, thus creating a semi-iterative relaxation. We found that
by varying the exact application of the method we could make the scheme
(almost) independent of the input parameters.
References
1. K. Stüben. A review of algebraic MG. Tech. rep. 69, GMD, Bonn, 1999
2. J.M. Ortega, W.C. Rheinboldt. Iterative Solution of Nonlinear Eq. in Several
Variables. Classics in Appl.Math. 30, SIAM, ISBN 0-89871-461-3, 2000
3. R. Haelterman. Aggregation Multi-grid for Residual Distribution Schemes. PR
2005-14, 2005, Von Karman Institute for FD
248 R. Haelterman, J. Vierendeels and D. Van Heule
Table 1. Optimal results for the different multi-grid schemes on the Poisson prob-
lem.
MG1 MG2 MG3 MG4 MG5 MG6 MG7
minimal WU 112 90 154 108 99 99 143
optimal ω1 0.50 0.55 N/A 0.50 N/A 0.90 N/A
optimal ω2 1.35 N/A N/A 1.60 0.95 N/A N/A
Table 2. Optimal results for the different multi-grid schemes on the advection-
diffusion equation with = 1.
Table 3. Optimal results for the different multi-grid schemes on the advection-
diffusion equation with = 100.
Table 4. Optimal results for the different multi-grid schemes on the advection-
diffusion equation with = 1000.
1 Introduction
Governing equations featuring strongly discontinuous coefficients arise in
many physical and engineering applications. A case in point concerns flow
in porous media which in a general sense, is anisotropic and heterogeneous
in nature, and the use of simple homogeneous isotropic assumptions is in-
adequate [1, 2]. Indeed, accurate numerical solutions can only be achieved if
the discretisation employed encompasses its surrounding neighbouring cells to
take account of the full hydraulic conductivity tensors [3]. The sharp discon-
tinuous characteristics observed in these tensorial coefficients require careful
attention.
Accordingly, a multigrid strategy stands out as a clear candidate for for-
mulating an effective, fast and robust numerical solver [4]. We note however,
that the use of uniformly-spaced grids would be highly inefficient for porous
media flows in which discontinuities in hydraulic conductivity are often lo-
calised or preferentially stratified. Hence, this paper describes in addition the
implementation of automatic grid refinement as a means of further improv-
ing the computational efficiency of such a solver, without incurring loss in
accuracy.
The multigrid scheme employed over the solution domain Ω uses a sequence
of progressively finer grids, Gk , where k = 0, 1, 2, . . . , kmax correspond to de-
creasing mesh size, ∆k = 0, 1, 2, . . . , kmax . On grid level k, ∆k is defined as
the number of mesh nodes in each co-ordinate direction, nk = 2k+c+1 , where c
is a constant defining the coarsest grid level. In addition, both cell- and patch-
wise adaptive grid refinement strategies are explored. The former consists of
locally refined arbitrary regions comprised of individually marked grid cells;
the latter is made up of coherent uniformly patched regions on a collection
of marked grid cells which intersect or overlaps one another. The τ -indicator
[4] is the natural choice for quantifying errors since this information is readily
available on each multigrid level. The criterion for grid refinement, τkk−1 > ε,
is defined as a user specified tolerance.
3 Results
0 0.5 1
1 1
1.0
0.0 0 0
0 0.5 1
0 0.5 1 1 1
1 1
0.5 0.5
0.5 0.5
0 0
0 0
(c) 0 0.5 1
(d) 0 0.5 1
Fig. 1. Discontinuous two region anisotropy: (a) problem definition; (b) patch-
wise adaptive grid structure; (c) cell-wise adaptive grid structure; (d) pressure iso-
contours.
mesh refinement demonstrates significant savings in CPU time and thus re-
sources. For example, when k = 8, there is as much as a factor of four savings
in CPU time to be achieved compared to the time expended non-adaptively.
Modified Stone’s problem [7]. Here we examine the effectiveness of
adaptive local mesh refinement for a problem exhibiting multiple sources
and sinks in a domain with anisotropic and heterogeneous structures. The
problem is depicted in Figure 3(a) for Neumann boundary conditions with
source and sink singularities, 1.0, 0.5, 0.6, −1.83, −0.27, located at co-ordinates
(3, 3),(3, 27),(23, 4),(14, 15),(27, 27) respectively, and hydraulic conductivities
given by:
10 1 1 100 0 00
KA = , KB = , KC = , KD = . (4)
01 1 100 0 1 00
Automatic adaptive grid refinement takes place on grid levels, 6 ≤ k ≤ 7, with
adaptive tolerance, ε = 0.1τkk−1 . Figures 3(b) and 3(c) shows that both adap-
tive schemes effectively identify regions requiring further refinement. Here,
the differences in mesh adaptation approach are more clearly depicted. No-
tably, cell-wise adaptive is particularly efficient at determining areas of rapid
change in which further refinement is required. Nonetheless, both approaches
252 Y.C. Lee and P.H. Gaskell
-2 3
10 WA 10 Non-adaptive, WA
CV Non-adaptive, CV
-3 Patch-wise, WA
-4
10 Cell-wise, CV
-5
10 1
10
-6
10
-7 0
10 4 5 6 7 10 4 5 6 7
Grid level, k Grid level, k
(a) (b)
Fig. 2. Error and CPU times associated with (a) non-adaptive and (b) adaptive
solutions (CV and WA) for the problem specified in Figure 1(a).
provide, in general, good solution accuracy; Figure 3(d) shows the resultant
iso-contour pressure profile for this problem.
0 0.5 1
1 1
28
D
21 A
0.5 0.5
16
12
C B
5
0 0 0
0 0.5 1
0 0.5 1 1 1
1 1
0.5 0.5
0.5 0.5
0 0
(c) 0
0 0.5 1
0
(d) 0 0.5 1
Fig. 3. Generalised Stone’s problem: (a) problem definition; (b) patch-wise adaptive
grid structure; (c) cell-wise adaptive grid structure; (d) pressure iso-contours.
Solution of porous media flow problems with discontinuous coefficients 253
y
0 0.5 1
1 1
0.5 0.5
A
1/8
0 x 0 0
0 0.5 1 0 0.5 1
1 1 1 1
0 0 0 0
(c) 0 0.5 1
(d) 0 0.5 1
Fig. 4. Diagonally layered anisotropy: (a) problem definition; (b) cell-wise adaptive
grid structure; (c) pressure iso-contours; (d) streamline contours.
150
100
50
4 5 6 7 8
Grid level, k
Fig. 5. Diagonally layered anisotropy: CPU times associated with the problem
specified in Figure 4(a).
times. Once again, CPU time savings becomes more significant with adaptivity
as finer grid levels of increasingly smaller mesh size are used.
4 Conclusion
Both the WA and CV discretisation methods allow problems exhibiting non-
diagonal permeability or hydraulic conductivity to be solved accurately, while
the associated adaptive multigrid solver offers significant savings in terms of
CPU time and resources. Both patch- and cell-wise adaptivity is shown to be
an effective means of solution which, when coupled with an efficient multigrid
strategy, leads to an accurate, fast and robust solver that can be readily
extended to a range of flow applications, especially where localised features
are important.
References
1. Molenaar, J.: A simple cell-centred multigrid method for 3D interface problems.
Comput. Math. Applic., 31(9), 25–33 (1996).
2. Calvo, P., Lisbona, F.:A finite volume multigrid method for flow simulation on
stratified porous media on curvilinear co-ordinate systems. Int. J. Numer. Meth.
Fluids, 37, 375–397 (2001).
3. Edwards, M.G.: Simulation with a full-tensor coefficient velocity field recovered
from a diagonal-tensor solution. SPE J. 5(4), 387–393 (2000).
4. Brandt, A.: Multigrid Techniques: 1984 guide with applications to fluid dynam-
ics. GMD-Studie Nr. 85, Sankt Augustin, West Germany (1984).
5. Crumpton, P.I., Shaw, G.J., Ware, A.F.: Discretisation and multigrid solution
of elliptic equations with mixed derivative terms and strongly discontinuous
coefficients. J. Comp. Phys., 116, 343–358 (1995).
6. Durlofsky, L.J.: Numerical calculation of equivalent grid block permeability ten-
sors for heterogeneous porous media. Water Resources Research, 27(5), 699–708
(1991).
7. Stone, H.L.: Iterative solution of implicit approximations of multidimensional
partial differential equations. SIAM J. Numer. Anal., 5, 530–558 (1968).
Stability of Pressure-Correction Algorithms for
Low-Speed Reacting and Non-Reacting Flow
Simulations
1 Introduction
A promising technique for reliable and accurate simulations of turbulent non-
premixed flames is large-eddy simulation (LES). With this technique time-
accurate flow field predictions are imperative and efforts for improvement of
algorithms in terms of computational cost are of great value. One way of
reducing computing time is to solve the Navier-Stokes equations in their low-
Mach number formulation, and to apply a segregated solution method to the
set of equations. Such methods are called pressure-correction methods, since
the kinematic part of the pressure in low-Mach number flows acts as a variable
whose magnitude is determined by imposing a constraint on the velocity field
in a corrector step. In flows where the density remains constant in time and
space, these methods have been extensively used [1] and do not suffer from
inexpected stability problems. However, when these methods are applied to
variable density flows in the context of non-premixed combustion simulations,
instabilities arise.
2 Governing Equations
Since we are dealing with low-Mach number flows, the Navier-stokes equa-
tionspare rewritten by expanding every variable in power series of M̃∞ =
u∞ / p∞ /ρ∞ , and taking the asymptotic limit for M̃∞ going to zero1 . For
every variable the lowest order term remains in the equations, except for the
pressure, which has now two parts: a thermodynamic part p0 and a kinematic
1
The subscript ∞ denotes reference values
2
part p2 . The pressure is p = p0 + M̃∞ p2 and the low-Mach number equations
read [2]:
p0 = p0 (t), (1)
∂ρ ∂ρui
+ = 0, (2)
∂t ∂xi
∂ρuj ∂ρui uj ∂p2 1 ∂τij
+ =− + , (3)
∂t ∂xi ∂xj Re∞ ∂xi
∂T ∂T 1 ∂ ∂T γ − 1 dp0
ρ + ui = κ + , (4)
∂t ∂xi Re∞ P r∞ ∂xi ∂xi γ dt
∂u
with shear stress tensor τij = µ ∂xj ∂ui
+ ∂xji − 23 µ ∂u
∂xk δij . ρ denotes the lo-
k
3 Different Formulations
3.2 Improvements
Several arguments form the basis for improved schemes: (a) the scheme must
remain stable, even in regions with high density jumps; (b) mass must be con-
served; (c) a pressure Poisson-equation with constant coefficients is desirable;
(d) in statistically stationary low-speed non-premixed combustion applica-
tions under adiabatic circumstances, no energy-equation is solved when unity
Lewis number is assumed for all species since the static enthalpy can then
immediately be recovered from the mixture fraction. Therefore an algorithm
which uses a constraint solely based on the energy equation, is not a viable
approach.
258 Pieter Rauwoens, Krista Nerinckx, Jan Vierendeels and Bart Merci
The equation for temperature (7) is rewritten as an equation for the density:
∂ρ ∂ρ
+ ui = 0, (15)
∂t ∂xi
from which the density at the new time level ρn+1 immediately follows. The
mass conservation equation (11) is replaced by
†
ρn+1 − ρn
∂ρui
=− . (16)
∂xi ∆t
†
The dagger-symbol indicates a partly-implicit value given by (ρu) = ρn u∗ +
0
(ρu) . These definitions give rise to a pressure Poisson equation of the form
∂ρn u∗i ∂ 2 p0 ρn+1 − ρn
− ∆t 2 = − . (17)
∂xi ∂xi ∆t
The advantage of the latter equation over (12) is that, for a case with pure
convective transport of density in a region with constant velocity, (17) gives
a constant pressure field, which is physically correct.
The density field at the new time is calculated by using the continuity equation
(5), or the equation of temperature in the form (15). The constraint on the
velocity field follows from the combination of these two equations. Eliminating
the time derivative of the density, we get:
∂ρui ∂ρ
= ui (18)
∂xi ∂xi
If we evaluate the latter equation at time level n + 1, we can derive the
Poisson-like equation
∗
∂ 2 p0 ∂ρn+1 ∗ ∆t ∂p0
∂ρui
− ∆t 2 = ui − n+1 . (19)
∂xi ∂xi ∂xi ρ ∂xi
r is given a value of 20. The interval of the jump is [20, 40]. The equations are
solved using a 1D finite volume method on a collocated mesh. The pressure
term in the momentum equation is discretized centrally. First order upwinding
is used for the convective terms. All calculations are done with a CFL-number
of 0.9.
2.5 2.5
density/20 density/20
velocity velocity
2 2
1.5 1.5
1 1
0.5 0.5
0 0
−0.5 −0.5
10 15 20 25 30 35 40 45 50 10 15 20 25 30 35 40 45 50
2.5 2.5
density/20 density/20
velocity velocity
2 2
1.5 1.5
1 1
0.5 0.5
0 0
−0.5 −0.5
10 15 20 25 30 35 40 45 50 10 15 20 25 30 35 40 45 50
Fig. 1. Convection of a density jump with factor 20 in a straight channel after one
time step with schemes 1 to 4; 1: top left, 2: top right, 3: bottom left; 4: bottom
right.
2 x
3 x x x
4 x x x
260 Pieter Rauwoens, Krista Nerinckx, Jan Vierendeels and Bart Merci
Results for the different schemes for the convection of a density jump after
one time step are given in Fig. 1. The properties of the schemes are summerized
in Table 1. The major requirement for the algorithm is to be stable. From the
study, it follows that only algorithm 4 renders stable results for high density
ratios, and is applicable in combustion simulations, making use of the mixture
fraction approach, under the conditions as stated in section 3.2. Furthermore
mass is conserved with this scheme. It has to be noted that an extra cost is
involved in this case, since the elliptic equation for the pressure, that follows
from the constraint on the velocity, does not have constant coefficients, and
therefore needs recalculation at each timestep. Scheme 3 does not suffer from
this shortcoming, but, although stable in the test case of a density jump, does
not guarantee stability in a more general case, e.g. when conductive effects
are taken into account (proof not given in this paper).
5 Conclusions
References
1. Chorin, A.J.: A numerical method for solving incompressible viscous flow prob-
lems. J. Comp. Phys. 2, 12–26 (1976)
2. Müller, B.. Low Mach number asymptotics of the Navier-Stokes equations. J.
Eng. Math., 34, 97–109 (1998)
A simple hybrid well-balanced method for a
2D viscous shallow water model
F. Marche
1 Introduction
Viscous shallow water models are extensively used in the field of coastal engi-
neering. Recent studies by Brocchini et al. [2] have shown that computation of
2D bore propagation over a complex topography is a difficult task. Dissipation
in 2D bores crossing breakwaters leads to the generation of large scale vortic-
ity. Numerical investigations revealed that such study demands a high order
accuracy well-balanced method. Another difficulty concerns the treatment of
the shoreline. Flow properties change rapidly as the water depth vanishes, and
the location of the shoreline is always evolving. Various methods have been
proposed to cope with this, often leading to complex algorithms [9].
We introduce here an hybrid Finite-Volumes (FV) Finite-Differences (FD)
method for the solution of a 2D viscous shallow water model derived by
Marche in [7] :
∂t h + div (hu) = 0 ,
∂t (hu) + div (hu ⊗ u) + g h∇h = −g h∇d − a0 (h) u (1)
−a1 (h) h |u| u + 2µ div hD(u) + 2µ∇(h div u),
water depth positivity for interface values. This property appears to be fun-
damental for simulations involving moving shorelines. This solver is then com-
bined with the hydrostatic reconstruction [1]. This raises a new second order
accuracy well-balanced model which can handle bore propagation over strong
varying topography. In addition, an efficient algorithm is proposed to accu-
rately simulate wetting and drying phenomena.
In a second step, a FD discretization of the remaining source terms is per-
formed. This step enables to take into account dissipation processes, through
friction at the bottom and diffusion, without disturbing the well-suited prop-
erties inherited from the previous step.
A new model ’SU RF− W B’ based on this method is applied to waves propa-
gation situations. First, the ability of the well-balanced scheme to accurately
handle moving shorelines is highlighted through comparisons with analytical
solutions. Then, we propose a comparison between the new viscous formu-
lation and a classical formulation. Some realistic applications will be also
introduced during the communication.
2 Numerical method
2.1 Convective step
with left and right numerical fluxes through the mesh interfaces :
Fn,−
i+ 1
= Fni+ 1 + Sni+ 1 − , Fn,+
i+ 1
= Fni+ 1 + Sni+ 1 + , (5)
2 2 2 2 2 2
The interface value Ui+ 12 (0, Ui+ 12 − , Ui+ 12 + ) between cell i and i + 1 is ob-
tained from a VFRoe-ncv linearized Riemann solver. This solver relies on a
symmetrizing change of variable [5] which preserves the water depth posi-
tivity at least for interface values. The key point is that this interface value
is computed from values Ui+ 12 − and Ui+ 12 + issued from a combined limited
linear/hydrostatic reconstruction. More precisely, considering the cell i, we
compute first linear reconstructions Ui,r and Ui,l respectively at i + 12 − and
A simple hybrid well-balanced method for viscous flows 263
In a second step we take into account the dissipative processes through the
following system :
∂t h = 0,
∂t (hu) = −a0 (h) u − a1 (h) h |u| u + 2µ div hD(u) + 2µ∇(h div u).
Considering the values (h∗ , (hu)∗ ) issued from the previous step, this system
is solved relying on the following semi-implicit discretization :
hn+1 = h∗ ,
∗ n+1 ∗
h u − (hu)
= −a0 (h∗ ) un+1 − a1 (h∗ ) h∗ |u∗ | un+1 (10)
∆t ∗ ∗
∗ ∗
+2µ div h D(u ) + 2µ∇(h div u ).
Using a second order accuracy centered discretization of the viscous terms and
avoiding any division by h∗ , this step can clearly handle the occurrences of dry
areas which may arise through the convective step and does not modify the
steady states at rest. Finally, these two steps are combined in a Runge-kutta
time marching algorithm to achieve second order accuracy in time. The time
step ∆t for both steps is defined with ∆t = min(∆tc , ∆ts ) where ∆tc and ∆td
are respectively obtained from stability analysis for each step [8].
264 F. Marche
h* hs*
-0.15
-0.05
-0.15
-0.05
0.05
0.15
0.05
0.15
-0.1
-0.1
0.1
0.1
0
0
-1
2
Analytical solutions
Numerical results
-0.8
4
-0.6
6
-0.4
x*
t*
8
-0.2
0
10
0.2
12
Fig. 1. The Carrier and Greenspan solution. Comparison between numerical and
analytical results. Left : surface elevation versus the onshore coordinate for different
values of time. Right : shoreline motion during three periods
3 Validation
We propose here some validations relying on analytical solutions or compar-
isons with more classical models.
We compare here numerical results (solid lines) with the Carrier and Greenspan
1D analytical solutions [3] (in dots). This solution represents the motion of a
periodic wave traveling shoreward and being reflected out to sea, generating
a standing wave which is let run-up and run-down on a plane beach. Fig.1
highlights the ability of the model in the computation of oscillating shore-
lines. The analytical solution is used as a left inlet boundary condition and
we use ∆x = 0.01 and CF L = 0.7. The shoreline oscillations are accurately
computed even after a large number of period. We stress out that many other
well-balanced methods, highly validated in hydraulic situations, are unable to
provide such results [7].
Fig. 2. Thacker’s solution. Contour plots for h+d (a): exact solution, (b): numerical
results, at t = 2T .
even after 2 periods. These results are more accurate than those presented by
Hubbard and Dodd [6] with their adaptive refinement method.
We study the evolution of a mound of water over a flat bottom which is re-
leased, generating a bore-like wave such that the front of the water propagates
with an inclination of 45◦ with respect to the boundaries of the domain. Flow
properties computed on the central cross-section orthogonal
to the propa-
gating front with the new viscous term 2µ div hD(u) + 2µ∇(h div u) are
compared to the results obtained with the classical term 4µ div hD(u) , as-
suming that the boundaries induced effects can be neglected for this section
and for a small time. Values of ∆x = ∆y = 0.01 and CF L = 0.7 are used.
h Q
0.5
0.6
0.7
0.8
0.9
0.32
0.34
0.36
0.38
0.42
0.44
0.4
1
-0.4
-0.4
-0.2
-0.2
a)
r
b)
0
r
0
0.2
0.2
0.4
0.4
References
1. Audusse, E., Bouchut, F., Bristeau, M.O., Klein, R., Perthame, B.: A fast and
stable well-balanced scheme with hydrostatic reconstruction for shallow water
flows. SIAM J.Sci.Comp., 25(6), 2050–2065 (2004)
2. Brocchini, M., Mancinelli, A., Soldini, L., Bernetti, R.: Structure-generated
macrovortices and their evolution in very shallow depths. Proc. Coastal Eng.,
772–783 (2002)
3. Carrier G.F., Greenspan H.P.: Water waves of finite amplitude on a sloping
beach, J. Fluid Mech., 4, 97–109 (1958)
4. Dressler R.F.: Comparison of theories and experiments for the hydraulic dam-
break wave, Int. Assoc. Sci. Hydraul., 38, 319–328, (1954)
5. Gallouet, T., Herard, J.M., Seguin, N.: Some approximate Godunov schemes
to compute shallow-water equations with topography. Computers and Fluids,
32, 479–513 (2003)
6. Hubbard M.E., Dodd N.: A 2D numerical model of wave run-up and overtop-
ping, Coastal Engineering, 47, 1–26 (2002)
7. Marche, F.: Derivation of a new two-dimensional viscous shallow water model
with varying topography, bottom friction and capillary effects. To appear in
European. J. Mech. B/Fluids.
8. Marche, F., Bonneton, P., Fabrie, P., Seguin, N.: Evaluation of well-balanced
bore-capturing schemes for 2D wetting and drying procedure. To appear in
Int. J. Num. Meth. Fluids.
9. Prasad, R.S., Svendsen, I.A.: Moving shoreline boundary condition for
nearshore models, Coastal Engineering, 49, 239–261 (2003)
10. Thacker WC.: Some exact solutions to the nonlinear shallow-water wave equa-
tions, J. Fluid Mech., 107, 499–508 (1981)
A kinetic energy-preserving P1 iso P2/P1
finite-element method for computing unsteady
incompressible flows
1 Introduction
Numerical experience with Direct Numerical and Large Eddy Simulations
(DNS, LES) of incompressible turbulenth flows consistently shows that cen-
tral schemes are to be preferred to stabilized methods [1, 2, 3, 4]. The turbu-
lence community has therefore put forth considerable effort to develop cen-
tral schemes that preserve the discrete kinetic energy (KE) in the inviscid
limit, which guarantees numerical stability. Designing KE preserving finite-
difference and finite-volume (FD,FV) schemes is not trivial, in particular for
higher orders and on irregular meshes [5, 6].
This contribution confirms earlier results of Sung et al. [7] who, to our best
knowledge, first demonstrated that skew-symmetric Galerkin FE methods nat-
urally preserve the KE. They thus provide a straightforward and systematic
method of creating KE preserving schemes of any order accuracy, even on
highly distorted unstructured meshes. Numerical evidence of KE preservation
is provided by implementing a P1 iso P2/P1 FE method. We also investi-
gate the behavior of schemes that do not exactly preserve the discrete kinetic
energy, namely (a) when non skew-symmetric formulations of the convective
terms are used and (b) when numerical stabilization is added.
1
Cc = ∇ · (uu) ; Ca = (u · ∇)u ; Cs = 2 (cc + ca ) . (2)
Although these three forms are analytically identical, they are not equivalent
when solved numerically.
We seek to obtain a FE approximation to the exact solution of Eqs. (1) in
compatible velocity and pressure spaces:
n o
Uh = span {Ni ; i = 1, . . . , n} ; Ph = span Ñj ; j = 1, . . . , m . (3)
Herein, Ni and Ñj represent basis functions with compact support. The
Galerkin FE discretization of (1) follows from the variational statement [8]
Z
qRCont (u) dΩ = 0 ∀q ∈ P, u ∈ U, (4)
Ω
Z
v · RM om (u, p) dΩ = 0 ∀v, u ∈ U; q, p ∈ P. (5)
Ω
This is, of course, valid when (v, q) are equal to the discrete solution fields (u,
p). For a skew-symmetric formulation of the advective terms, one can easily
verify that this leads to the following KE preservation statement:
Z
REn dΩ = −V ≤ 0, (6)
Ω
where
∂u2 /2 1 1
REn = + ∇· [uu2 /2] + ∇· [up] + ∇· [ (u · ∇)u] ; V = ∇u ⊗ ∇u.
∂t Re Re
c
b
1
a
2
Fig. 1. Illustration of the P1 iso P2/P1 stable linear element pair. Pressure degrees
of freedom are only stored in vertices marked by circles.
A kinetic energy-preserving P1 iso P2/P1 finite-element method 269
4 Results
π m/s π m/s
(periodic) (periodic)
π/10 m/s
(periodic)
6.28 6.28
3.14
y
3.14
y
6.28 6.28
3.14 3.14
y
6.28 6.28
3.14 3.14
y
Fig. 4. Comparison of u−velocity profiles after 2 time units: (a) baseline, (b) base-
line with momentum dissipation.
A kinetic energy-preserving P1 iso P2/P1 finite-element method 271
at different instances in time. Due to dispersion errors and the nonlinear na-
ture of the equations, the solution is severely distorted after only one time
unit. Shortly thereafter, the solution becomes fully chaotic. Nevertheless, the
numerical method remains stable. Fig. 4 compares solutions obtained after
2 time units. The baseline scheme rapidly distributes kinetic energy over all
available wave numbers. Adding a small amount of artificial momentum dis-
sipation suffices to maintain the flow structures.
Fig. 5(a) shows that for ∆t = 0.001 (CFL = 0.013), the time integration
errors become so small that the time integration can be considered ‘exact’
for the 100 time units considered. In this case, KE is preserved up to four
significant digits after 100 time units. As shown in Fig. 5(b), the use of non
skew-symmetric nonlinear terms leads to nonlinear instabilities independent
of the time step size. Figs. 5(c–d) show that the addition of higher-order dissi-
pation, either to the momentum or continuity equations, leads to a damping of
the flow structures. The PSPG stabilization damps out even the bulk velocity
once chaos appears.
10 8
∆t = 0.01
∆t = 0.001 7
9
8 6
〈KE〉
〈KE〉
7 5
6 4 Baseline, ∆t = 0.001
Advective, ∆t = 0.001
Advective, ∆t = 0.0005
5 3
Conservative, ∆t = 0.001
Conservative, ∆t = 0.0005
4 2
0 20 40 60 80 100 0 5 10 15 20 25
Time Time
(a) (b)
6
5
5
4
4
〈KE〉
〈KE〉
3
3
2 Baseline
2 Baseline JST momentum diss.
JST momentum diss. PSPG pressure diss.
1 PSPG pressure diss. 1
Pressure Laplacian
Pressure Laplacian
0
0 20 40 60 80 100 0 20 40 60 80 100
Time Time
(c) (d)
Fig. 5. Time evolution of the mean kinetic energy: (a) baseline formulation, in-
fluence of the time step; (b) influence of the formulation of the nonlinear terms;
(c) influence of adding stabilization on the resolved mesh; (d) influence of adding
stabilization on the marginally resolved mesh.
272 A kinetic energy-preserving P1 iso P2/P1 finite-element method
5 Conclusions
The shown results clearly support our claim that Galerkin FE methods are
well-suited for making kinetic KE preserving numerical methods. Although
stable, KE preserving central schemes are found to generate rapidly generate
chaotic solutions. Whether this is desirable, or not, remains to be established.
A stability analysis for the considered flow would be most useful for this.
References
1. S. Ghosal. An analysis of numerical errors in large-eddy simulations of turbu-
lence. Journal of Computational Physics, vol. 125, pp. 187:206 (1996).
2. A. G. Kravchenko and P. Moin. On the effect of numerical errors in large eddy
simulations of turbulent flows. Journal of Computational Physics, vol. 322, pp.
131:310 (1997).
3. N. Park and J.-Y. Yoo and H. Choi. Discretization errors in large eddy sim-
ulation: on the suitability of centered and upwind-biased compact difference
schemes. Journal of Computational Physics, vol. 198, pp. 580:616 (2004).
4. L. Georges, G. Winckelmans and Ph. Geuzaine. Improving shock-free com-
pressible RANS solvers for LES on unstructured meshes. JCAM, accepted for
publications, 2006.
5. Y. Morishini and T. Lund and O. Vasilyev and P. Moin. Fully conservative
higher-order finite-difference schemes for incompressible flow. Journal of Com-
putational Physics, vol. 143, pp. 90:124 (1998).
6. O. Vasilyev. High-order finite-difference schemes non non-uniform meshes with
good conservation properties. Journal of Computational Physics, vol. 157(2),
pp. 746:761 (2000).
7. J. Sung, H.G. Choi, J.Y. Yoo. Time-accurate computation of unsteady free
surface flows using an ALE-segregated equal-order FEM. Comput. Methods
Appl. Mech. Engrg., vol. 190, pp. 1425:1440 (2000).
8. A. Ern and J.L. Guermond. Theory and Practice of Finite Elements. Applied
Mathematical Series, Vol. 159. Springer-Verlag, New York, 2004.
9. D. Vanden Abeele and G. Degrez and D.O. Snyder. Parallel turbulent flow
computations using a hybrid spectral/finite-element method on Beowulf clus-
ters. Proceedings of the Third International Conference on Computational Fluid
Dynamics (ICCFD3), Springer Verlag, 2004.
10. R.J. McDermott. Toward one-dimensional turbulence subgrid closure for large
eddy simulation. PhD Thesis, University of Utah, 2005.
Study on the segregation algorithms of the
incompressible Navier-Stokes equations using
P1P1/P2P1 finite element formulation
1 Introduction
Since velocity and pressure fields are obtained at each separate step in a
segregated formulation, it produces a small and simple matrix at each step
than in an integrated formulation. And the memory and cost requirements of
computations are significantly reduced because the pressure equation for the
mass conservation of the Navier-Stokes equations is constructed only once if
the mesh is fixed. However, segregation formulation solves Poisson equation
of elliptic type so that it always needs a pressure boundary condition along
a boundary even when no physical information on pressure is given. In gen-
eral, open boundary problems sometimes have no information on both the
pressure and pressure gradient at the exit because velocities on downstream
boundary are unknown. Therefore, the problem of proper imposition of out-
flow boundary condition is not clearly stated, which is an important issue in
the segregation approach [1, 2].
On the other hand, an integrated finite element formulation in which the
governing equations are simultaneously treated has an advantage over a seg-
regated formulation in the sense that it can give a more robust convergence
behavior because all variables are implicitly combined especially when the flow
fields are complicated, as in strong swirl flows and impinging jet. Further an
integrated formulation needs a very small number of iterations to achieve con-
vergence. However, the application of an ILU type preconditioner to a global
matrix resulting from an integrated formulation is not so straightforward as
that to a small and simple matrix resulting from a segregated formulation [3].
Thus, the saddle-point-type matrix (SPTM) in the integrated formulation is
assembled and preconditioned every time step, so that it needs a large mem-
2 Theoretical background
2.1 P1P1 segregation formulation
The P2P1 semi-segregation formulation (P2P1) used in the present study can
be written as follows:
ûi − uni 1 1 1
+ (uˆj uˆi,j + unj uni,j ) = − pn,i + n
(τ̂ij + τij ),j (7)
∆t 2 ρ 2ρ
u∗i − ûi 1
= pn,i (8)
∆t ρ
( un+1 − u∗ 1
i i
= − pn+1
∆t ρ ,i (9)
un+1
i,i = 0
In the first and second fractional step it solves the same equations as P1P1 seg-
regated formulation to satisfy the momentum equation. For time integration,
the Crank-Nicolson scheme, which is second-order accurate and uncondition-
ally stable, is to be used. In order to utilize the fact that a separate pressure
equation, which is assembled and preconditioned only once, is to be solved in
the P1P1 fractional step method, a fixed symmetric SPTM has been obtained
for the continuity constraint of the present P2P1 finite element formulation.
3 Numerical examples
Nozzle
B=1
H=2
y
x u = 0, v = 0
Impingement plate
W=8
Fig. 1. The flow regions considered for the impinging jet problem
The results show that the P2P1 is about two times faster than the P1P1. In
particular, the third step of P2P1 is faster than the third and fourth step of
P1P1. Fig. 2 compares the eigenvalue distribution of the ILU (or AILU) pre-
conditioned matrix to that without the preconditioning in both the P2P1 and
P1P1. A more clustered eigenvalue distribution is obtained with precondition-
ing than without preconditioning. Fig. 3 compares the local Nusselt number
(Nu) obtained from the present computations at ReB = 450 and the result
of Cziesla et al. [4]. Pressure boundary condition of an elliptic type pressure
equation in the P1P1 should be given even when the physical information of
the pressure boundary condition is not provided. Therefore, we can confirm
that the pressure boundary conditions have an effect on the accuracy and
convergence of the solutions.
In the second example we solved the flow of a viscous fluid past a stationary
isolated sphere at the Reynolds number 300 (Re is based on inlet fluid velocity
and diameter of the sphere). ICEM-CFD package is used for unstructured
mesh generation. Total numbers of nodes and elements in the P2P1 mesh are
173584 and 128211, respectively. In order to find the same accuracy as P2P1,
P1P1 needs much more nodes and elements than P2P1. Total numbers of
nodes and elements used in the P1P1 are 493320 and 2910779, respectively.
Time increment ∆t = 0.2 is used. Time history of the pressure is shown in
Fig. 4. It is clear that a periodic flow pattern appears behind the sphere. The
corresponding Strouhal number of the present is 0.139. This also agrees well
Study on the segregation algorithms using P1P1/P2P1 element 277
1 1
Imaginary
Imaginary
0.75
0.5 0.5
0.25
0 0
-0.25
-0.5 -0.5
-0.75
-1 -1
4 8 12 16 20 24 28 32 36 0 0.25 0.5 0.75 1 1.25 1.5 1.75 2
Real Real
1 1
Imaginary
Imaginary
0.5 0.5
0 0
-0.5 -0.5
-1 -1
0 4 8 12 16 20 24 28 32 0 0.25 0.5 0.75 1 1.25 1.5 1.75 2
Real Real
Fig. 3. Nu distribution for uniform and parabolic inlet velocity profiles at the nozzle
exit (ReB = 450)
with the value of 0.137 found by the numerical experiments of Johnson and
Patel [5].
4 Conclusion
The present P2P1 semi-segregation method has the ability to give a more
accurate solution than the existing P1P1 segregation method. In terms of
the pressure boundary condition, the P2P1 semi-segregation method is more
flexible to use because the P1P1 segregation method always needs a pressure
278 Myung H. Cho, Hyoung G. Choi and Jung Y. Yoo
0.08
0.04
Present study, P1P1
Present study, P2P1
0
Pressure
-0.04
-0.08
-0.12
-0.16
-0.2
3100 3200 3300 3400 3500
Time
References
1. Choi, H.G., Choi, H., Yoo, J.Y.: A fractional of the unsteady incompressible
Navier-Stokes equations using SUPG and linear equal-order element methods.
Comput. Methods in Appl. Mech. Engrg., 143, 333–348 1997
2. Ramaswamy, B. and Jue, T.C.: Some recent trends and developments in finite
element analysis for incompressible thermal flows. Int. J. Numer. Meth. Engrg.,
35, 671–707 1992
3. Nam, Y.S., Choi, H.G., Yoo, J.Y.: AILU preconditioning for the finite element
formulation of the incompressible Navier-Stokes equations. Comput. Methods in
Appl. Mech. Engrg., 191, 4323–4339 2002
4. Cziesla, T., Tandogan, E., Mitra, N.K.: Large-eddy simulation of heat transfer
from impinging slot jets. Numer. Heat. Trans. Part A, 32, 1–17 (1997)
5. Johnson, T.A. and Patel, V.C.: Flow past a sphere up to a Reynolds number of
300. J. Fluid. Mech., 378, 19–70 (1999)
A Mach-uniform algorithm: coupled versus
segregated approach
1 Introduction
Historically, algorithms for CFD have been grouped into two classes: coupled
density based methods and segregated pressure based methods. Both types of
algorithms have been adapted to make them Mach-uniform, i.e. applicable for
all speeds (see f.e. the references in [1]). In this paper, we present the idea
that only a classification with regard to the solution technique (coupled versus
segregated) is valuable: three different types of algorithms are constructed,
each of them based on the same idea to reach Mach-uniformity, but applying
a different solution technique.
2 Semi-implicit system
Mach-uniform efficiency implies a good convergence rate for any level of the
Mach number. In [1] we explain how Mach-uniformity can be reached by treat-
ing the acoustic and diffusive (friction, heat conduction) terms implicitly. The
convective terms, on the other hand, can be treated explicitly. A finite volume
method for the two-dimensional Navier-Stokes equations therefore results in
the following semi-implicit system
h i
(m+1) [m]
X [m] (m+1)
ρi − ρi Vi /∆tI + ρf wf dsf = 0, (1)
h i
(m+1) [m] [m] (m+1)
X X
(ρv)i − (ρv)i Vi /∆tI + (ρvw)f dsf + pf dxf =
(m+1)
X
(τ .1̄y .n̄)f dsf − ρi gVi , similar for u, (2)
h i
(m+1) [m] [m] (m+1)
X
(ρE)i − (ρE)i Vi /∆tI + (ρH)f wf dsf =
(m+1) (m+1)
X X
(κ∇T.n̄)f dsf + (τ .w̄)f .n̄dsf − ρi vi gVi , (3)
with ρ the density, p the pressure, T the temperature, E the total energy, H
the total enthalpy, ḡ the gravitational acceleration vector, τ the viscous stress
tensor and κ the heat conduction coefficient. The summation runs over all
faces f of the control volume around node i. dsf is the length of the face f
and wf represents the projection of the velocity vector w̄ = u1̄x + v 1̄y on the
outward normal n̄ of the face f , i.e. wds = udy + vdx.
In the convective fluxes, the transported quantities u, v, H are upwinded
using a higher order method. Like in [2], we use for the mass flux and the
pressure at a face a blending of the high speed AUSM+ flux and a low-speed
central flux. A pressure dissipation term is added to the mass flux to prevent
pressure-velocity decoupling at low Mach numbers.
A multistage time stepping with k stages is used,
[m]
Q[0] = Qn ; ...; Q[m+1] = Q[0] + αm+1 ∆tQt ; ...; Qn+1 = Q[k] . (4)
n and n + 1 represent respectively the old and the new time level, while [m]
and [m + 1] represent stage levels. Q is the state vector and ∆t = tn+1 − tn
[m]
the time step. The slope Qt = ∂Q ∂t is determined as Qt = ∆Q[m] /∆tI =
(m+1) [m] I I
(Q − Q )/∆t . ∆t a chosen internal time step. The unknown state
Q(m+1) is determined from the semi-implicit system, and then inserted into the
multistage time stepping, Q[m+1] = Q[0] +αm+1 ω∆Q[m] , with ω = ∆t/∆tI . In
the following we will use a Fourier stability analysis to analyse the algorithms.
Notice that ∆tI determines the shape of the Fourier symbol, while ω acts as
a scaling factor for this curve.
The way in which this semi-implicit system is solved for the updated values
(m + 1) can now be chosen: three different solution techniques are considered,
varying between a fully coupled and a fully segregated approach. In a fully
coupled method, the semi-implicit (4Nx4N)-system (1)-(3) is solved in a cou-
pled way, in each step of the multistage time stepping. As a second solution
technique, we consider the coupled pressure and temperature algorithm from
[1]. This algorithm finds its place in between the fully coupled and the fully
segregated approach. An explicit convective predictor step is followed by an
acoustic / diffusive corrector step in which a (2Nx2N)-system of correction
equations has to be solved. The third algorithm is a fully segregated pressure-
correction algorithm with pressure corrections determined from the energy
equation. As explained in [1], this fully segregated algorithm can be applied
for a perfect gas flow without heat transfer. Here, the dimension of the system
to be solved in each multistage step is (NxN) only.
A Mach-uniform algorithm: coupled versus segregated approach 281
The invicid flow of a perfect gas is considered. First, we focus on low speed
flow. In the stability analysis we consider a flow in the x-direction aligned
with the grid, at M = 0.001. We use the Van Leer-κ scheme for the trans-
ported quantities (with κ = 1/3), and the multistage time stepping with four
stages and standard coefficients. The time step ∆tI is calculated from a chosen
convective CFL number, ∆tI = cf lI ∆x/u and the scaling factor ω is chosen.
Figure 1 shows the Fourier symbol. For the fully coupled algorithm
(COUP), cf lI = has to be decreased down to 0.5 to obtain a good shaped
curve. Next, the curve can be scaled up with ω = 2.7 before it crosses the
time stepping curve.
In the predictor-corrector approach, we consider three algorithms: a pres-
sure-correction algorithm with pressure corrections form the energy equation
(PE), one with pressure corrections from the continuity equation (classical
approach) (PC), and the coupled pressure and temperature correction algo-
rithm (PT). Explicit convective terms are used. Apparently, the PE algorithm
can use the highest global CFL number (i.e. cf lI ∗ ω). For the PT algorithm
the global CFL number is somewhat lower. For the PC algorithm the global
CFL number is the lowest. This confirms that, for the case of a perfect gas
without heat transfer, the corrections should by determined from the energy
equation.
3 3
2 2
1
1
–2
3 3
2
2
1
1
–2
–1
Fig. 1. Fourier symbol, M = 0.001. Top left: fully coupled algorithm, cf lI = 0.5,
ω = 2.7. Top right: PE algorithm, cf lI = 0.6, ω = 2.7. Bottom left: PT algorithm,
cf lI = 0.6, ω = 2.4. Bottom right: PC algorithm, cf lI = 0.2, ω = 2.7.
282 Krista Nerinckx, Jan Vierendeels, and Erik Dick
As a numerical test case, we take the inviscid flow past a bump in a channel,
with an inlet Mach number of 10−5 . Figure 2 compares the convergence plots.
First, the maximum value for cf lI was determined to keep the computation
stable with ω = 1. Next, this maximum value for cf lI was kept constant and
ω was increased as much as possible. For the predictor-corrector algorithms,
the PE algorithm converges the fastest. The PC algorithm needs the most
time to converge, due to the use of the continuity equation to determine the
pressure (see [1]). The PT algorithm has a convergence rate in between the
PE and the PC algorithm. Indeed, due to the influence of the continuity
equation in the corrector step, the convergence slows down compared to the
PE algorithm. The PE algorithm has the same convergence behavior as the
fully coupled algorithm. However, the cost to solve one time step is lower
for the PE algorithm: only a (NxN)-system has to be solved per time step,
opposed to a (4Nx4N)-system for the fully coupled method. Therefore, the
more segregation is introduced, the cheaper the simulation can be done.
-11.5
-12
log ( |Res Mom| )
-12.5
-13
-13.5
-14
PC
-14.5 PE
-15 COUP PT
-15.5
100 200 300 400 500 600
Number of time steps
Fig. 2. Test case bump, Min = 10−5 . ρv-momentum residual as a function of the
number of time steps. COUP: cf lI = 0.7, ω = 2.2. PE: cf lI = 0.7, ω = 2.9. PT:
cf lI = 0.7, ω = 1.9. PC: cf lI = 0.5, ω = 1.8.
Similar tests were done for high Mach number flow. The same conclusions
hold.
In [1] it was shown that the PE algorithm suffers from a diffusive time step
limit due to the conductive terms. For the coupled pressure and temperature
correction algorithm, however, such a limit is absent. The stability analysis
now confirms this. The cf lI number is chosen as the maximum value for the
Euler case and is kept constant. When the value of κ is increased, the PE
algorithm becomes unstable if κ is taken too high. For the PT algorithm,
however, κ can be taken infinitely high.
Thus, the PT algorithm is advantageous for those cases where the heat
conduction is important. What’s more, for such a case the behavior of the PT
algorithm improves. Indeed, in case of Euler equations, the convective CFL
A Mach-uniform algorithm: coupled versus segregated approach 283
number can be taken at most 1.1 when implicit convective terms are used.
However, if the diffusive terms dominate (i.e. a cell Peclet number P ec <
O(1)), the convective CFL restriction is weakened, and even can disappear
completely. This is confirmed by the Fourier stability analysis, where the CFL
number can be taken higher as the value of P ec decreases.
The gravity terms in the y-momentum and the energy equation act as source
terms. Therefore, they may introduce an additional time step restriction. To
analyse this, we add gravity terms to the Euler equations, and perform the
Fourier stability analysis for a vertical flow at M = 0.001.
In the fully coupled algorithm, the gravity terms can be treated implicitly
without difficulties. For the Euler equations, i.e. g = 0, the maximum value of
the convective CFL number cf lv is 0.6 when a first order upwind discretisation
is used for the transported quantities. We take cf lv = 0.6 and increase the
value of g (which represents some non-dimensional number).
Figure 3 compares the eigenvalue plots for the algorithm without gravity
(g = 0) and with an extremely high value for g (g = 106 ). We conclude that
the fully coupled algorithm remains stable, no matter how high g is taken. No
additional time step restriction due to the gravity terms is introduced.
In a segregated predictor-corrector algorithm, however, it is no longer pos-
sible to treat the gravity terms in a fully implicit way. To get insight in the
effect of the gravity term, we analyse the eigenvalues of the predictor amplifi-
cation matrix. We find that the eigenvalues are completely determined by two
1 1
–2 0
–2 0
1 4
0 –2 0
–2
–2
–4
Fig. 3. Fourier symbol. Top: fully coupled algorithm with implicit gravity terms,
left: g = 0, right: g = 106 . Bottom: PT algorithm, left: g = 10−7 , cf lg = 0.071,
right: g = 10−5 , cf lg = 7.1.
284 Krista Nerinckx, Jan Vierendeels, and Erik Dick
5 Conclusion
References
1. K. Nerinckx, J. Vierendeels, and E. Dick. Mach-uniformity through the cou-
pled pressure and temperature correction algorithm. Journal of Computational
Physics, 206:597–623, 2005.
2. J. Vierendeels, B. Merci, and E. Dick. Blended AUSM+ method for all speeds
and all grid aspect ratios. AIAA Journal, 39(12):2278–2282, 2001.
Crank-Nicolson Scheme for Solving Low Mach
Number Unsteady Viscous Flows Using an
Implicit Preconditioned Dual Time Stepping
Technique
1 Introduction
The aim of this paper is to simulate low Mach number unsteady viscous flows
using a density-based finite volumes solver. This kind of solver is known to
encounter some difficulties to simulate low Mach number flows in which the
density is almost constant. Convergence fails or accuracy decreases when the
speed of sound is much greater than the fluid velocity. Local preconditioning
methods intend to solve this problem by altering the time derivative terms
of the Navier-Stokes equations in order to artificially modify the speed of
sound, improving the convergence and accuracy in the case of low Mach num-
ber steady flows [1], [4]. Is it still the case for unsteady flows for which a
dual time stepping approach is used ? Does the condition number depend on
the time step ? Does the artificial dissipation terms of numerical advective
flux functions should be adequately modified ? In this paper, a semi-implicit
Crank-Nicolson time integration scheme is used while the advective terms up-
winding is performed by the AUSM+up scheme designed by Liou [2]. At each
time step the solution is found by means of a fully implicit Newton-GMRes
dual time stepping (pseudo-transient) technique. An eigenvalues analysis will
answer to the previous questions for this special case. It will show that the
preconditioning technique designed for steady flows leads to an ill-conditioned
system for small CF Lc numbers1 . Therefore, a new unsteady preconditioning
technique will be introduced and the AUSM+up artificial dissipation func-
tion will also be modified to achieve robustness and accuracy for all Mach
and CF Lc numbers.
1
CF Lc = c ∆
∆t
must be close to unity or to the inverse of the Mach number M −1
when the time step is dictated by the speed of sound or by the fluid velocity (that
is when CF Lu = kuk∆∆t ) respectively.
3 Local preconditioning
7
10
No Precond. 10
0
10
6 Steady Precond.
Unsteady Precond.
5
10
-1
4
10
10
CN
fc
103
-2
M = 1
102 10 M = 10
-1
M = 10 -2
M = 10 -3
101 -4
M = 10
-3
100 10
-1 0 1 2 3 4
10 10 10 10 10 10 10 0 10 1 10 2 10 3
CFL c CFL c
The eigenvalues analysis of the pseudo-transient system (2) is made on its non
spatially discretized version. If the source term g wl is assumed to have no
effect on the dispersion equation and if the diffusive terms are supposed to be
negligible the system canbe written in its quasi-linearform as
∂w Str ∂w ∂w ∂w
P + S w + θ Ax + Ay + Az =0 (5)
∂τ ∆t ∂x ∂y ∂z
where the matrices S, Ax , Ay and Az are the jacobian matrices of time deriva-
tive and advective terms
of f . By introducing a Fourier mode w (τ, x, y, z) =
w0 exp kT x − i ω τ into the system, this last equation reduces to a disper-
sion equation for ω. Due to the finite number of mesh nodes, the wavelengths
are limited by the grid spacing. To take this filtering into account, one can set
kkk = φ/∆ where ∆ is a measure of the grid spacing, which is evaluated as the
minimal distance between the node and its neighbours, and φ ∈ [0, π]. Defin-
T
ing successively the wave speed λ = ω/kkk and uk = ukkkk , the eigenvalues are
found to be such that
λ1,2,3 = θ uk 1 − i CF L−1 and λ4,5 = λ± = θ uk 1 − i CF L−1
u u T± (6)
The CF Lu number based on the local velocity and the CF Lc number based
on the speed of sound are defined through the cell Strouhal number Str by
Str = Str ∆/ ∆t φ uk = θ CF L−1 −1
u and M Str = θ CF Lc . It can be seen
that the first three eigenvalues have a real part corresponding to the advective
velocity and an imaginary part corresponding to the damping brought by the
time derivatives term of the implicit Crank-Nicolson scheme. Nevertheless this
decoupling is no longer valid for the two other eigenvalues since the factors
T± depend on M , M∗ = c0 /c (where the artificial speed of sound is defined as
c02 = ρ hT / ρ hT ρ0p + ρT (1 − ρ hp ) ) and CF Lc .
288 Scheme for Solving Low Mach Number Unsteady Viscous Flows
s
1 1 2 4 M∗2
T± = 1 + M∗2 ± (1 − M∗2 ) − 2 (7)
2 2 i M + CF L−1
c
This expression reduces to the results found be Weiss and Smith [5] in the
case of steady Euler flows for which CF Lc tends to infinity.
10 3
10 0
10 -1
-2
10
-3
10
10 -4
-5
10
-6
10
0 15 30 45 60 75
Newton Iterations
Fig. 2. Laminar flow past a cylinder at Re = 100 and M0 = 10−3 . Left : Streamlines
at different times. Right : Dual time stepping convergence curves (CF Lc = 103 ).
4 Results
A time periodic laminar flow past a cylinder at Re = 100 and M = 10−3 is
solved using the three different preconditioning techniques with a time step
corresponding to CF Lc = M −1 = 103 (based on the smallest grid spacing). It
can be seen in Figure 2 that the new unsteady and the steady preconditioning
exhibit equivalent convergence rates while the unpreconditioned method does
not converge as previously stated by theory for such CF Lc numbers (see
Figure 1). The vortex shedding Strouhal number is found to be Strvs =
0.166. The drag coefficient CD has a mean value of 1.32 and oscillates with
an amplitude of 0.0085 whereas the lift coefficient CL oscillates around zero
with an amplitude of 0.315. These values are in agreement with numerical and
experimental values found in the literature.
The 3D laminar lid driven cavity flow test case at Re = 1000 and M0 =
10−3 is solved. Streamlines and convergence curves are shown in Figure 3.
Tests on convergence have been made by integrating the equations with a time
step that corresponds to CF Lc = 102 . As predicted by theory (see Figure 1),
the unsteady preconditioning has a better convergence rate. Furthermore, less
GMRes iterations are performed to solve the linear systems when CF LnP tends
to infinity that is when the effect of the preconditioning matrix P becomes
negligible. That illustrates the role of the scaling function fc on convergence.
Conclusion
A new unsteady preconditioning technique, taking the CF Lc number into ac-
count, has been designed for low Mach unsteady flows solved by a fully implicit
dual time stepping scheme. The scaling function of the artificial dissipation
290 Scheme for Solving Low Mach Number Unsteady Viscous Flows
0.4 0.4
0.2 0.2
0 0
y
y
-0.2 -0.2
-0.4 -0.4
No Prec.
GMRes Iterations
10 4 Steady Prec.
50
Unsteady Prec.
10 2
40
0
10
10 -2 30
10 -4 20
-6
10 No Prec.
10 Steady Prec.
-8
10 Unsteady Prec.
0
0 5 10 15 20 25 0 5 10
Newton Iterations Newton Iterations
Fig. 3. 3D lid driven cavity flow (M0 = 10−3 and Re = 1000). Up left : streamlines
at t = 2. Up Right : streamlines at t = 4. Down Left : Dual time stepping convergence
curves (CF Lc = 103 ). Down Right : GMRes convergence (CF Lc = 103 ) curves.
terms in the AUSM+up scheme has been consequently modified. Simple test
cases have illustrated the improvement of convergence predicted by theory. A
better study of time and space accuracy should be performed in future works.
References
1. Y.-H. Choi and C. L. Merkle. The application of preconditioning in viscous
flows. Journal of Computational Physics, 105:207–223, 1993.
2. M.-S. Liou. A further development of the ausm+ scheme towards robust and
accurate solutions for all speeds. June 2003. AIAA Paper 2003-4116.
3. E. Turkel and Verr N. Vasta. Local preconditioners for steady and unsteady flow
applications. ESAIM: Mathematical Modelling and Numerical Analysis, 39:515–
535, 2005.
4. D. Vigneron, J.-M. Vaassen, and J.-A. Essers. An implicit finite volume method
for the solution of 3d low mach number viscous flows using a local preconditioning
technique. In Acoment 2005 Conference, Ghent, Belgium., May-June 2005.
5. J. M. Weiss and W. A. Smith. Preconditioning applied to variable and constant
density flows. AIAA Journal, 33(11):2050–2057, November 1995.
Heated Wake by Deferred Corrected ULTRA
Summary. The instability problem associated with DWF formulation (an implicit
implementation of ULTRA) is removed by a deferred correction procedure. With
this deferred correction, ULTRA becomes one of the most accurate flux-limiting
strategies for unsteady flow simulations. Unlike some TVD schemes, ULTRA exerts
control only occasionally when the DWF value exceeds its bounds. This deferred
corrected DWF strategy is applied to a series of numerical simulations conducted
for the advection of heat in the wake of a heated cylinder for two Reynolds numbers,
Re = 200 and 1000.
1 Introduction
A large number of flux limiters have been developed to control unphysical
oscillations associated with advection schemes ( Van Leer [5]; Gaskell and
Lau [3]; Leonard and Mokhtari [6]; Leonard and Drummond [4]). These lim-
iters may alter the accuracy of the base schemes. Therefore, their application
should be minimized to only those instances when intervention by the lim-
iters is absolutely necessary. In this paper, computations were conducted for
a realistic flow of the wake behind a heated circular cylinder. The goal was
to study the performance of a number of flux limiters on a practical problem.
Unlike some TVD (Total-Variation Diminution) schemes, the method of UL-
TRA (Universal Limiter for Tight Resolution and Accuracy) by Leonard and
Mokhtari [7] is most accurate due to its ability to optimize for the highest
order of accuracy during the strategy’s intervention. In an implicit implemen-
tation, Leonard and Mokhtari [7] introduced the DWF formulation. However,
the original formulation is numerically unstable. In the present investigation,
this numerical instability is removed by a deferred correction procedure to en-
sure diagonal dominance of the solution matrix. Calculations were carried out
using a 2D (two-dimensional) model without sub-grid viscosity3 , for Reynolds
3
The effect of the sub-grid viscosity was considered and reported in the Ph.D.
thesis by Bouhairie [2].
FOD n
1
a b
0.8
K
IC
QU
0.6 f g CD h
0.4
m
DWF 0.2
0 c d
S OU FOU
-0.2 e
-0.4
Fig. 1. DWF values as a function of the normalized variable φ̃C for various finite-
difference schemes. The constraints imposed by ULTRA are along the line e-c for
the SOU (second order upwind) scheme, between the line a-b for the FOD (first
order downwind) scheme and the line c-d for the FOU (first order upwind) scheme,
and along the line g-h for the CD (central difference) scheme.
number, Re = 200 and 1000. The advective process at these Reynolds num-
bers is sufficiently dominant to provide a rigorous assessment of the strategy’s
performance.
Fig. 2. Heat wake at Re = 1000 obtained using CD (central difference) scheme (a)
without and (b) with the flux limiter ULTRA. The negative temperature associated
with the spurious oscillations (blue in part (a) of the figure) is as high as -6.58o C.
20 20
Largest Tma x over time = 14.94 Largest Tma x over time = 10.36
(a) CD Smallest Tm in over time = -6.575 (b) QUICK Smallest Tmin over time = -0.728
15
14.94 C o
Average Tma x over time = 6.63294 15 Average Tma x over time = 3.75699
Average Tm in over time = -2.04733 o Tmin over time = -0.190863
Average
10.36 C
10 10
T 5 T 5
0 0
-0.73oC
-5 -5
-6.58oC
20 20
60 70 80 90 100 60 70 80 90 100
Largest Tm ax over time = 7.79 Largest Tma x over time = 8.874
time
Smallest Tmin over time = -0.0001092 time
Smallest T over time = -0.0001337
15 (c) ULTRA-CD Average Tma x over time = 3.15418 15 (d) ULTRA-QUICK min
Average T over time = 3.43111
ma x
Average Tm in over time = 0 Average Tm in over time = 0
10
7.79oC 10 8.87oC
T 5 T 5
0 0
-5 -5
20 20
60 70 80 90 100 60 70 80 90 100
Largest Tma x over time = 6.422 Largest Tma x over time = 6.257
time
Smallest Tmin over time = 0 Smallest Tmtime
Dimensionless in over time time
=0
15 Dimensionless time
Average Tma x over time = 2.84479 15 Average Tma x over time = 2.77966
(e) TVD-CD (f) TVD-QUICK
Fig. 3. Temperature maximum and minimum at a cross section 20 radii downstream
of the heated cylinder.
3 Results
f CD g CD h
c d
SOU FOU
e
IC K g CD h
QU
m
c d
SOU FOU
e
Fig. 4. DWF values and the areas where the order of accuracy are reduced in the
wake. Points in (a) and (c) mark the intervention by the flux limiter ULTRA. White
and pink in (b) and (d) denote the areas where the order of accuracy is first and
second order, respectively. Green denotes the default order not altered by the flux
limiter.
4 Conclusion
Stability in the solution process is achieved in the present calculations of the
unsteady flow by deferred correction and by application of a flux limiter to
suppress oscillations. This combination ensures the robustness of the algo-
rithm, which is essential in the numerical simulations of advection-dominant
processes at high Reynolds number. Numerical oscillations in highly unsta-
ble base schemes, such as the CD scheme, is managed by the flux limiter to
produce acceptable performance. The advantage of ULTRA and SMART is
the minimal intervention on their base schemes, unlike the TVDs examined in
this paper. As shown in Figure 4, the order of the accuracy of the base scheme
is compromised only sparingly in small and isolated areas in the wakes.
References
1. S. Bouhairie and V. H. Chu: Two-dimensional simulation of unsteady heat
transfer from a circular cylinder in cross flow. J. Fluid Mech., (in press), 2006.
2. S. Bouhairie: Computational methods for calculating heat transfer from a circu-
lar cylinder in a cross flow. PhD thesis, McGill University, Montreal, Quebec,
Canada, 2005.
3. P. H. Gaskell and A. K. C. Lau: Curvature compensated convective transport:
SMART, a new boundedness preserving transport algorithm. Int. J. Num. Meth.
Fluids, 8:617–641, 1988.
4. B. P. Leonard and J. E. Drummond. Why you should not use ‘HYBRID’,
‘POWER-LAW’ or related exponential schemes for convective modelling - there
are much better alternatives. Int. J. Num. Meth. Fluids, 20:421–442, 1995.
5. B. Van Leer. Towards the ultimate conservative difference scheme. II. mono-
tonicity and conservation combined in a second-order scheme. J. Comp. Phys.,
14:361–370, 1974.
6. B. P. Leonard and S. Mokhtari. Beyond first-order upwinding: the ULTRA-
SHARP alternative for non-oscillatory steady-state simulation of convection. Int.
J. Num. Meth. Engng., 30:729–766, 1990.
7. B. P. Leonard and S. Mokhtari. ULTRA-SHARP nonoscillatory convection
schemes for high-speed steady multidimensional flow. Technical Report NASA
TM 102568 (ICOMP-90-12), NASA-Lewis Research Center, 1990.
8. P. L. Roe. Large scale computations in fluid mechanics, pt. 2. In S. Osher
E. Engquist and R. J. C. Sommerville, editors, Lectures in Applied Mathematics,
volume 22, pages 163–193. Amer. Mathematical Soc., Providence, RI, 1985.
Convergence Acceleration for Euler Equation
based on SPR
1 Introduction
Wavelet coefficients of each odd numbered positions are defined as the differ-
ence between real values and approximated values.
In threshold process, the new threshold method which consists of the
0.004
0.0035
Conventional
0.003
Previous threshold
0.0025 Modified threshold
0.002
0.0015
Accuracy
0.001
0.0005
cases, ∆x2 is smaller than . From Fig. 2, the modified threshold value main-
tains the 2nd order of spatial accuracy by changing the threshold value from
to ∆x2 . Secondly, we test the stabilization technique in constructing SPR
dataset throughout NACA0012 airfoil calculations with 201 × 97 grid points.
The 2nd level SPR method with = 10−5 is applied. The angle of attack is
5◦ and the MAch number is 0.8. Remaining grid points in SPR dataset are
shown in Fig. 3. If the stabilization technique is not used, unnecessary grid
points are remaining in a SPR dataset. The remaining grid points in a SPR
dataset without the stabilization technique are about 60% of total number
of grid points. However, when the stabilization technique is used, grid points
are extracted effectively according to the flow solution and the remaining grid
points in a SPR dataset 37% of total number of grid points. Consequently,
the compression ratio of SPR method is improved about 170%.
Thirdly, we test the restriction technique of meaningless variations of
flow variables. The residual histories are presented in Fig. 4 for subsonic and
transonic cases. In the subsonic case (a), the characteristic of convergence is
Convergence Acceleration for Euler Equation based on SPR 301
10-1
-2 Conventional Conventional
10 Restriction-Off Restriction-Off
Restriction-On Restriction-On
10-2
Residual History
Residual History
10-3
-3
10
-4
10 -4
10
-5
10 10-5
50 100 150 200 50 100 150 200
Time Time
substantially enhanced and the total computing time decreases by half when
tiny variations are restricted. In the transonic case (b), even though comput-
ing time reduces substantially, the improvement of computing efficiency does
not reach the level of that of subsonic case. It is because the large amount
of error is generated inevitably nearby shock discontinuity, flows downstream
and contaminates a substantial portion of computational domain, which is be-
yond the order affected by restriction technique. If some way of treating these
errors induced by shock discontinuity is introduced additionally, the compu-
tational efficiency may be improved more even in transonic case.
Combining all of newly introduced methods, the modified thresholds
value, the stabilization technique and restriction technique, we check the over-
all enhancement in computing efficiency with AUSMPW+ scheme for spatial
discretization and 4th order Runge-Kutta explicit methods for time integra-
tion. The total iteration number, CPU time gain and difference of lift co-
efficient according to each case are summarized in table 1. From table 1,
Table 1. Test cases and results for the flow at NACA0012 airfoil
NACA0012 Time SPR Iteration CPU Time Lift Difference
Airfoil Integration Level Number Time Ratio Coefficient (%)
Dense Grid Conventional 0.5497
Subsonic Conventional 14912 1952.22 1.00 0.5337 2.911
(Ma=0.3, Runge-Kutta Level1 7853 793.72 2.46 0.5327 3.093
AOA=5) (201×97) Level2 7227 569.42 3.43 0.5349 2.692
Level3 8018 652.58 2.99 0.5318 3.256
Dense Grid Conventional 1.0629
Subsonic Conventional 12013 1532.98 1.00 1.0497 1.242
(Ma=0.8, Runge-Kutta Level1 8054 893.58 1.72 1.0491 1.298
AOA=5) (201×97) Level2 7737 764.08 2.01 1.0453 1.656
Level3 7861 786.34 1.95 1.0465 1.543
302 Dohyung Lee, Hyungmin Kang and Dongho Lee
differences of lift coefficients have little relation with the number of multires-
olution level and change within 2nd order of accuracy. Comparing with time
ratio of 4th order explicit Runge-Kutta method, the overall computing time
substantially decreases and about 300% enhancement of efficiency is achieved
for subsonic flow case and 190% is for transonic case.
4 Concluding Remarks
The modified SPR method is developed to improve computing efficiency of
conventional Euler solver while conserving the spatial accuracy of the solver.
The threshold value is modified so as to consider grid spacing and as a result,
it is realized that SPR method maintains the spatial accuracy of conventional
solver. Also, by the stabilization technique, unnecessary grid points are ef-
fectively cut off in a SPR dataset and the compression ratio is enhanced. In
time integration, if variations of flow variables are smaller than the threshold
value, ´, they are restricted by adopting a weighting factor and convergence
of conventional solver is improved.
The modified SPR method can be easily generalized to three-dimensional
Euler equation and more complicated Navier-Stokes equation. Also, it is com-
pletely independent of other convergence acceleration methods such as the
multigrid method. Therefore the modified SPR method combined with multi-
grid method can elevate the computing efficiency even more.
References
1. Bacry, E., Mallat, S., Papanicolaou, G.: A wavelet based space-time adaptive
numerical method for partial differential equation. Math. Model. Numer. Anal.,
26, 793–834 (1992)
2. Holmström, M., Walden, J.: Adaptive wavelet method for hyperbolic PDEs. J.
Sci. Comput., 13, 19–49 (1998)
3. Holmström, M.: Solving hyperbolic PDEs using interpolation wavelets. SIAM J.
Sci., Comput., 21, 405–420 (1999)
4. Sjögreen, B.: Numerical experiments with the multiresolution scheme for the
compressible Euler equations. J. Comput. Phys., 117, 251–261 (1995)
Acceleration of Unsteady Incompressible Flow
Calculation Using Extrapolation Methods
1 Introduction
2 Extrapolation methods
Use of an implicit time-marching method is assumed. An intermediate solution
vector is identified by two indices, the time step and the number of iterations.
For example, xj,k represents the intermediate solution vector obtained after
k iterations at the j-th time step.
MPE was proposed by Leland and Rollett[2] for linear systems. Its origin is
found in Cabay and Jackson[3] and Mevsina[4].
When a conventional iterative method based on matrix splitting is used,
the intermediate solution vector xj,k is expressed in the form
N
X
xj,k = x˜j + ai λki φi (1)
i=1
where N stands for the degree of freedom, x˜j is the converged solution at the
j-th time step, and λi and φi are the i-th eigenpair of the iteration matrix.
Although eigenpairs depend on the time step, the subscript j is omitted in
Eq.(1) for simplicity. Suppose that the eigenvalues are arranged in order of
magnitude, namely 1 ≥ |λ1 | ≥ · · · ≥ |λN |. It is assumed in MPE that the first
p eigenvalues λ1 , · · · , λp are dominant and responsible for slow convergence:
p
X
xj,k = x˜j + ai λki φi (p N ). (2)
i=1
From Eq.(2), the shift vector sj,k ≡ xj,k − xj,k−1 is expressed in the similar
form
p
X
sj,k = bi λki φi . (3)
i=1
The slowly decaying errors are removed by linear extrapolation and the con-
verged solution x˜j can be obtained from
(c0 + c1 + · · · + cp−1 )x˜j = c0 xj,k + c1 xj,k+1 + · · · + cp−1 xj,k+p−1 (4)
where {c0 , · · · , cp−1 } is a set of coefficients such that
c0 sj,k + c1 sj,k+1 + · · · + cp−1 sj,k+p−1 = 0. (5)
In practice, the coefficients {c0 , · · · , cp−1 } are calculated from the least square
problem derived from Eq.(5) because Eq.(5) itself is not determinate. As most
eigenvalues are neglected, MPE is not effective unless the slowly decaying
errors are dominant and should not therefore be applied at an early stage of
iterations.
MPE explained above was originally developed for linear systems. The
applicability of MPE to non-linear systems was theoretically given by Shimano
Acceleration of Flow Calculation Using Extrapolation Method 305
TVE predicts the converged solution at the present time step from a series of
converged solutions at previous time steps.
First, we consider a linear system. In the time-marching, an increment of
the converged solution, ∆x̃j (≡ x̃j − x̃j−1 ), is subject to sequential linear map-
ping. The behaviour of the incremental vector is determined by eigenvalues of
the mapping. When the system contains periodic fluctuations, eigenvalues of
which absolute value is equal to unity play an important role. If the number
of such eigenvalues is M , one component of the incremental vector is written
in the following form:
M
X
∆x̃j = {am sin(j∆tωm ) + bm cos(j∆tωm )} (6)
m=1
The coefficients {c1 , · · · , cq } are calculated by solving the least square problem
derived from Eq.(9) with all components of ∆x̃j taken into account.
TVE is applicable not only to linear systems but also to nonlinear ones because
terms generated by nonlinearity can be expressed in the same form as Eq.(6).
306 Kenjiro Shimano, Shun Yonezu, and Yoshiteru Enomoto
Re=100 Re=500
200 × 120 400 × 240 200 × 120 400 × 240
MPE only 25 25 25 25
TVE+MPE 15 25 15 25
Acceleration of Flow Calculation Using Extrapolation Method 307
the solution converge faster than TVE alone. Despite a jump-up in the residual
caused in the x-momentum equation, decay of the residual was considerably
quick after the MPE application.
Remarkably, the rapid residual reduction by MPE lasted longer when TVE
and MPE were jointly used. TVE worked as a preconditioner which strength-
ened the effect of MPE. This can be also observed in Fig.3 where averaged
speed-up factors are compared. At each calculation except Re = 100 with
400 × 120 grid points, the speed-up achieved by TVE+MPE is higher than
the product of those achieved by TVE and MPE separately. The speed-up by
TVE+MPE ranged approximately from 2.0 to 4.4. In calculation at Re = 100
with 400 × 120 grid points, the solution did not converge when only TVE was
applied. Further study is necessary to investigate the reason.
10-1 10-1
No extrapolation No extrapolation
10-2 TVE 10-2 TVE
10-3 10-3
Residual
Residual
10-4 10-4
10-5 10-5
10-6 10-6
10-7 10-7
20 40 60 20 40 60
Iterations Iterations
(a) x-momentum equation (b) Continuity equation
10-1 10-1
MPE MPE
10-2 TVE+MPE 10-2 TVE+MPE
10-3 10-3
Residual
Residual
10-4 10-4
10-5 10-5
10-6 10-6
10-7 10-7
20 40 60 20 40 60
Iterations Iterations
(a) x-momentum equation (b) Continuity equation
5
TVE 4.41
4 MPE
TVE+MPE
Speed-up
3
2.59 2.65
2.15
1.98 1.92
2 1.76
1.40 1.50
1.35
1.19
1
0
Re=100 Re=100 Re=500 Re=500
200x120 400x240 200x120 400x240
Fig. 3. Speed-up
Conclusions
A new extrapolation method, TVE, was proposed for unsteady flow analysis.
A joint effect of TVE and an existing extrapolation method,MPE, was inves-
tigated by applying them to an unsteady N-S solver. TVE provided a good
initial guess for the iterative solver. An acceleration effect of MPE was im-
proved by joint application of TVE because TVE worked as a preconditioner
for MPE. As a result, the speed-up achieved by the combination of TVE and
MPE ranged approximately from 2.0 to 4.4.
References
1. Shimano, K., Arakawa, C.: Incompressible Navier-Stokes Solver Using Extrap-
olation Method Suitable for Massively Parallel Computing, Comp. Mech., 23,
172–181 (1999)
2. Leland, R.W., Rollett,J.S. : A comparison of parallel extrapolation methods
In: Hackbusch, W. (ed) Parallel Algorithms for Partial Differential Equations.
Viewveg (1991)
3. Cabay, S., Jackson, W.: A Polynomial Extrapolation Method for Finding Limits
and Antilimits of Vector Sequences, SIAM J. Numer. Anal., 13-5, 734–752 (1976)
4. Mevsina, M.: Convergence Acceleration for the Iterative Solution of the Equa-
tions X=AX+f, Comp. Meth. in Appl. Mech. and Eng., 10, 165–173 (1977)
Improved Numerical Simulations of
Incompressible Flows Based on
Viscous/Inviscid Interaction Procedures
1 Introduction
0.5
p
−C
−0.5
−1
3 Numerical Results
Algebraic grids are first generated around NACA 0012. The governing equa-
tions are discretized using finite volumes. Line relaxations are adopted to
solve the potential equation iteratively together with tri-diagonal solvers for
the momentum equations, marching in the flow direction.
Improvement of the numerical method in the present work includes the
solution of potential equation in a mapped plane using ADI method and using
a third order upwinding scheme for the convection terms in the momentum
equations. Details of the Numerical Methods are given in ref [7].
In figures 1 - 2 , the results for steady flow calculations at Re = 500
and α = 10◦ are shown. In figures 1(a), 1(b), the u-velocity contours are
plotted based on the present formulation and based on the method of ref.
[1] where the normal momentum equation is solved for the pressure. The
Simulations of Incompressible Flows Based on Viscous/Inviscid Interaction 313
4 Concluding Remarks
For external aerodynamic problems, it is not necessary to solve Navier Stokes
equations everywhere in the field. Viscous / Inviscid interaction procedures
with potential flow in the outer region is an alternative. Using Helmholtz-
type velocity decomposition and a modified Bernoulli’s law for the pressure
in terms of the potential, leads to an attractive formulation. Preliminary nu-
merical results are encouraging for both steady and unsteady laminar incom-
pressible flow simulations. Unlike the stream function / vorticity formulation,
the present method can be extended to three dimensions in a straight forward
manner.
References
1. Hafez, M. , Shatalov A., Wahba, E.: Numerical Simulations of Incompress-
ible Aerodynamic flows using Viscous / Inviscid Interaction Procedures. Comp.
Methods in Appl. Mech. and Engrg., Vol 195, pp. 3110-3127 (2006)
2. Shatalov, A., Nakajima, M., Hafez, M.: Numerical Simulations of Unsteady Lam-
inar Incompressible Flows using Viscous / Inviscid Interaction Procedures, CFD
Journal, Vol 15, pp. 246-254 (2005)
3. Cousteix, J., Mauss, J.: Approximations of the Navier-Stokes equations for high
Reynolds number flows past a solid wall, Journal of Computational and Applied
Mathematics, Vol 166, pp. 101-122 (2004)
4. Gordnier, R.E., Rubin, S.G.: Flow Simulations using a composite velocity pro-
cedures for potential, Euler and RANS equations, Computer and Fluids, Vol 17,
pp. 85-98 (1989)
5. Quarteroni, A.: Domain Decomposition Methods for PDE’s, Oxford University
Press (1999)
6. Amsden, A., Harlow, F.: A Simplified MAC Technique for Incompressible Fluid
Flow Calcualtions, JCP, Vol 6, pp. 322-325 (1970)
7. Shatalov, A.: Ph.D. Dissertation, Univ. of Calfornia, Davis (2006)
314 M. Hafez, A. Shatalov, and M. Nakajima
time=5 time=5
2 2
1.5 1.5
1 1
y
y
0.5 0.5
0 0
−0.5 −0.5
−0.5 0 0.5 1 1.5 2 2.5 3 −0.5 0 0.5 1 1.5 2 2.5 3
x x
time=5
time=5
SMAC
1 Present Method (2nd order upwinding)
1
Present Method (3rd order upwinding)
0.8
0.8
0.6 0.6
v
0.4 0.4
0.2 0.2
0 0
−0.2
−0.2 0 5 10 15 20 25 30
0 2 4 6 8 10 12 14 16 18 20
time (a) time
1 Introduction
2 Numerical Method
2.1 Discretization of Governing Equations
For overset mesh methods, a search procedure is needed for identifying the lo-
cation of the nodes on opposite overlapping blocks. For unstructured meshes,
the search should be performed for all nodes of all mesh blocks, because the
nodes and the cells are randomly distributed. Once the search process is com-
pleted for all nodes, the information is used for clarifying the node type, either
active or non-active, and for determining the weighting factors for interpola-
tion. To overcome the large overhead involved in the search process, a fast
and robust neighbor-to-neighbor search technique was implemented by uti-
lizing the property of the linear shape functions. The geometric relationship
between the beeline connecting the starting position and the target point and
the faces of the cells the beeline crosses are shown in Fig.1 (a).
Under parallel environment, this point search may develop across the subdo-
main boundary, and thus the number of data information in each processor
changes at every iteration. To handle this varying number of data efficiently, a
new data structure for parallel distributed memory machines was introduced
Parallel Unstructured Overset Mesh Technique 319
as shown in Fig.1 (b). Each node is tagged as either active or non-active based
on the distance-to-wall technique [2]. Determination of cell type, active, in-
terpolation, or non-active, for hole-cutting was made based on the number of
active nodes assigned to each tetrahedral cell.
Since transfer of the flow variables between mesh blocks was made through
interpolation, the results obtained from overset mesh methods do not satisfy
the conservation property. The local solution accuracy degrades furthermore
when the characteristic cell size changes drastically from one mesh block to
another. In order to reduce this error, the interpolation was made by consid-
ering not only the cell itself containing the interpolation receiver but also the
neighbor cells enclosing the cell.
To treat overlapped bodies in relative motion, those cells with all nodes lo-
cated underneath the solid boundary of other objects were excluded from the
calculation, while the cells with nodes partially immersed inside the body were
treated as active cells for computation. The interpolation of the flow variables
for these cells was made such that the flow tangency condition was satisfied
at the solid surface as shown in Fig.1 (c). For unsteady flow calculations, the
flow variables were assigned also for non-active cells by interpolation and ex-
trapolation from surrounding cells so that these variables can be used in the
subsequent time steps as they become active.
For the validation of the present method, single/triple ejector rack cases were
simulated by modelling each individual body with a separate mesh block.
The main block covers the entire flow field including wing/pylon, and the
other blocks represent the stores. The stationary main block was composed
of 404,738 tetrahedral cells and each subgrid block was composed of 249,536
tetrahedral cells. The calculations were made at a freestream Mach number
of 0.95 and an angle of attack of 0 degree. The six degree-of-freedom equa-
tions of motion were coupled to the flow equations for the prediction of the
time-accurate store trajectories. The results in Figs.2 (a)–(d) show that the
predicted trajectories compare well with the experimental data [5] and other
predictions [6] for all stores. In Figs.2 (e) and (f), the separation sequence of
single and three stores are presented at every 0.1 sec.
The second validation of the present method was made for a generic missile
with deploying fins. The fins are made of a NACA0018 airfoil section and
are located at 80% from the nose of the body. The calculation was made at
320 Mun Seung Jung and Oh Joon Kwon
4 Concluding Remarks
A parallel unstructured overset mesh method has been developed for the sim-
ulation of unsteady flow fields. Applications were made to the single/triple
ejector rack, a generic missile with deploying fins, and a rotor-fuselage config-
uration. The predicted results showed reasonable agreement with experiment
and other predictions within the accuracy of the present inviscid calculation.
It was demonstrated that the present method is efficient and robust for sim-
ulating unsteady flows involving multiple bodies in relative motion.
Parallel Unstructured Overset Mesh Technique 5
Parallel Unstructured Overset Mesh Technique
321
Parallel Unstructured Overset Mesh Technique 5
Parallel Unstructured Overset Mesh Technique 5
Fig. 1. Node
Fig. search,
1. Node parallelization
search, parallelization and interpolation
and interpolation inside
inside the solid
the solid surface
surface
Fig. 1. Node search, parallelization and interpolation inside the solid surface
Fig. 1. Node search, parallelization and interpolation inside the solid surface
Fig. 3. Pressure contour and hole-cutted meshes for a missile with deploying fins
at selected time levels
Fig. 3. Pressure contour and hole-cutted meshes for a missile with deploying fins
at
3.selected
Fig.Fig. Pressuretimecontour
3. Pressure levels
contourand
andhole-cutted meshesforfor
hole-cutted meshes a missile
a missile withwith deploying
deploying fins fins
at selected time
at selected levels
time levels
322
6 Mun
Mun Seung
Seung Jung and Oh
Jung and Oh Joon
Joon Kwon
Kwon
Fig. 4.
Fig. 4. Rotor-fuselage
Rotor-fuselage configuration
configuration and
andtime-averaged
time-averagedvelocity
velocitydistributions
distributions
References
[TJ95] Thoms, R.D., Jordan, J.K.: Investigation of Multiple-Body Trajectory
References
Prediction Using Time-Accurate Computational Fluid Dynamics. AIAA
paper 95-1970 (1995)
[NT00]
1. Thoms,Nakahashi, K.,Togashi,
R. D., Jordan, F., Sharov, of
J. K. : Investigation D.:Multiple-Body
Intergrid-Boundary Definition
Trajectory Predic-
tion Using Time-Accurate Computational Fluid Dynamics. AIAA paper 38,11,
Method for Overset Unstructured Grid Approach. AIAA Journal, 95-1970
(1995) 2077–2084 (2000)
[TI03]
2. Togashi,
Nakahashi, F., Ito, Y.,
K., Togashi, F.,Nakahashi,
Sharov, D. K., Obayashi, S.: Overset
: Intergrid-Boundary Unstructured
Definition Method
for Overset Unstructured Grid Approach. AIAA Journal, 38, 11,2003-3405
Grid Method for Viscous Flow Computations. AIAA paper 2077–2084
(2000) (2003)
[PK04] Park, Y.M., Kwon, O.J.: Simulation of Unsteady Rotor Flow Fields Using
3. Togashi, F., Ito, Y., NaNakahashi, K., Ohayashi, S. : Overset Unstructured Grid
Unstructured Sliding Meshes. Journal of the American Helicopter Soci-
Method for Viscous Flow Computations. AIAA paper 2003-3405 (2003)
ety,49,4,391–400 (2004)
4. Park, Y. M., Kwon, O. J. : Simulation of Unsteady Rotor Flow Fields Using
[LS95] Lijewski, L.E., Suhs, N.E.: Time Accurate Computational Fluid Dynamics
Unstructured Sliding Meshes. Journal of the American Helicopter Society, 49,
Approach to Transonic Store Separation Trajectory Prediction. Journal
4, 391–400 (2004)
of Aircraft, 31,4,886–891 (1995)
5. Lijewski,
[LP00] L. E.,Park,
Lee, S., Suhs,M.,
N. E. : Time
Cho, K.W.,Accurate Computational
Kwon. J.H.: New Fully Fluid Dynamics
Automated Ap-
Proce-
proach dure
to Transonic Store Separation
for the Prediction Trajectory Journal
of Store Trajecotry. Prediction. Journal37,6,1038–
of Aircraft, of Aircraft,
31, 4, 886–891 (1995)
1049 (2000)
6. Lee,
[LK89] S.,Liou, S.G.,Cho,
Park, M., K. W., Kwon,
Komerath, J. H. : New H.M.:
N.M., McMahon, Fully Automated Procedure for
Velocity Measurements
the Prediction of Store
of Airframe Trajectory.
Effects on a Rotor Journal of Aircraft,
in Low-Speed 37, 6,Flight.
Forward 1038–1049 (2000)
Journal of
7. Liou, S.Aircraft,
G. Komerath, N. M.,
24,4,340–348 McMahon, H. M. : Veloctity Measurements of
(1989)
Airframe
[ZR97] Effects
Zori, on a
L.A.J., Rotor in Low-Speed
Rajagopalan, Forward Flight.
R.G.: Navier-Stokes Journal of Aircraft,
Calculations Rotor-
24, 4, 340–348
Airframe(1989)
Interaction in Forward Flight. Journal of the American Heli-
8. Zori, L.copter
A. J., Rajagopalan, R. G. : Navier-Stokes
Society, 42,3,235–243 (1997) Calculations of Rotor-Airframe
Interaction in Forward Flight. Journal of the American Helicopter Society, 42,
3, 235–243 (1997)
A Parallel Overset Adaptive Cartesian/Prism
Grid Method for Moving Boundary Flows
Summary. The use of overset grids in CFD started more than two decades ago,
and has achieved tremendous success in handling complex geometries[6]. In particu-
lar, overset grids have the advantage of avoiding grid remeshing when dealing with
moving boundary flow problems. In this paper, two particular unstructured grids
are advocated for moving boundary flow simulation, i.e., the use of overset adap-
tive Cartesian/prism grids. Semi-structured prism grids are generated around solid
walls. These prism grids then overlap a single adaptive Cartesian background grid.
With the adaptive Cartesian grid, the mesh resolution of the prism grid near the
outer boundary can easily match that of the oversetting Cartesian grid cells. In ad-
dition, the tree-based data structure of the Cartesian grid can be used efficiently in
hole-cutting and donor cell identification. In order to expedite the simulations, the
entire solver was made capable of running in parallel. The overset adaptive Carte-
sian/prism grid method is tested for both steady and unsteady flow computation. It
is demonstrated that moving boundary flow computations can be automated with
minimum user interferences.
1 Introduction
grids are used to generate holes in the adaptive Cartesian grid to facilitate
data communications. If the bodies move, the prism grids move with the bod-
ies, while the Cartesian grid remains stationary. After a few (tens of) time
steps, new holes are cut out of the Cartesian grids, and new donor cells are
also identified. Solution fields are interpolated from the old Cartesian grid to
the new grid using cell-wise linear reconstruction.
The paper is organized as follows. The numerical method is reviewed in the
next section. Solutions for some problems(using the current approach) are
presented in Section 3. Conclusions are summarized in Section 4.
2 Numerical Method
2.1 Generation of Overset Cartesian/Prism Mesh
Kallinderis[1] defined the term node-manifold as the list of faces confining the
node to be marched. Common sense tells us that the marching vector at any
node should not make an angle greater than 90◦ with the face normals of its
manifold. If the above criterion is violated, the tip of the marching vector
is not visible from all the faces of the manifold. So the paramount objective
here is to ensure that the marching vectors satisfy the visibility criterion. The
secondary objective is to impose orthogonality. An angle based weighting is
used to obtain the initial guess for the marching vector. The marching vector
is then refined locally in order to satisfy the secondary objective. An inverse
distance based smoothing was used to further smooth the marching vectors.
The marching step was obtained based on the local curvature. The nodes in the
current layer were smoothened. The intersections (if any) were obtained and
were corrected on a local basis. After the prism grid generation, an adaptive
Cartesian grid was generated automatically matching the grid resolution near
the outer boundaries of the prismatic grids
Vin+1 − Vin
= Σf vgn dSf (3)
∆t
The above equation is employed to obtain vgn . In addition, the volume swept
by a cell is the sum of volumes swept by its faces, i.e.
Vin+1 − Vin = Σf ∆Vf (4)
Comparing equations 3 and 4, we arrive at the following equation:
∆Vf
vgn = (5)
∆tdSf
For a steady inviscid flow, the velocity components at a ghost cell for a solid
wall boundary are computed as
ughost = u − 2nx vn , vghost = v − 2ny vn (6)
where vn = unx +vny . For unsteady moving boundary problems, vn is modified
to unx + vny − vgn . For an unsteady viscous problem, the velocity components
at a ghost cell are computed as
ughost = −u + 2nx vgn , vghost = −v + 2ny vgn (7)
In the far field, a characteristic analysis based on Reimann invariants is used
to determine the values of the flow variables on the outer ghost cells. This
acts as a non-reflecting boundary for the unsteady applications. The solution
at the face center of the hole/interpolation boundary faces are obtained by
extrapolating the solution from the corresponding donor cell
3 Numerical Results
3.1 Problem 1: Turbulent flow over a moving sphere
This case was selected to validate the moving grid flow solver. The sphere
moves from right to left in quiescent air with a Mach number of 0.2. If the
reference frame is fixed on the moving sphere, the flow field should reach a
steady state after the initial transients propagate out of the solution domain.
The moving grid flow solver was first verified that the GCL was satisfied.
Then it was used to solve the moving sphere problem. The Reynolds number
of the flow was 1.1e6. The SA method was employed to close the turbulence
equations. Roughly 6.0e5 cells were used for the current simulation.
The density field created the moving sphere after a long time was compared
with that created by turbulent flow over a stationary sphere in figure 1. It is
observed that the density fields are similar. The Cp and the skin friction coeffi-
cients are plotted in figure 2. The Cp values are in good accord with the exper-
imental results[2] and the simulation results obtained by other researchers[3].
326 Ravishekar Kannan and Z.J. Wang
A good match could not be obtained for the skin friction coefficient. This
discrepancy was also mentioned by other researchers. This discrepancy is in
all probability due to the fact that the turbulence model enforces the effect of
turbulence throughout the computational domain. In reality the effect of tur-
bulence become predominant only after some distance. Hence the mismatch.
The Cd was in good agreement with the experimental values and is presented
in the below table
Table 1. Data on Drag; Experimental results from Aachanbach[2] and DES results
from Constantinescu[3]
Current Simulation 0.1
DES 0.084
Experiment 0.12
Fig. 1. The density distribution for moving sphere in quiescent air and flow around
a stationary sphere
1 8
Experiment
Experiment
DES 6 DES
Moving Grid Moving Grid
0
cf * sqrt(Re)
4
Cp
2
−1
0
−2 −2
0 60 120 180 0 60 120 180
Theta in degrees Theta in Degrees
4 Conclusions
The moving boundary method was tested on both turbulent and unsteady
problems. It produced good results for both the cases. This method combines
328 Ravishekar Kannan and Z.J. Wang
Fig. 3. Vortices shed by the airfoil during the upstroke and downstroke positions
5 Acknowledgments
The authors gratefully acknowledge the support from the Air Force Office
of Scientific Research (Grant Number FA9550-04-1-0053), with Dr. Fariba
Fahroo as the Technical Monitor.
References
1. Kallinderis, Y., and Ward, S., ”Prismatic Grid Generation for 3-D Complex
Geometries,” AIAA Journal, Vol 31, No. 10, 1993, pp. 1850-1856.
2. Achenbach, E., 1974, ”Vortex shedding from the spheres”, J. Fluid Mech., 62,
pp. 209-221
3. Constantinescu, G.S., Pacheco, R. and Squires, K.D., 2002, ”Detached-Eddy
Simulation of Flow over a Sphere,” AIAA Paper 2002-0425
4. Von Karman, T., and Burgers, J.M., “General Aerodynamic Theory-Perfect Flu-
ids,” Aerodynamic Theory, Vol. 2. Julius-Springer, Berlin, 1943, p. 308.
5. Anderson, J.M, Streitlien, K., Barrett D.S, and Triantafyllou, M.S, ”Oscillating
foils of high propulsive efficiency,” JFM, Vol. 360, 1998, pp. 41-72
6. Togashi, F., Nakahashi K., Ito Y., Iwamiya T. and Shimbo Y.,”Flow Sim-
ulation of NAL Experimental Supersonic Airplane/Booster Using Overset
Grids”,Computers and Fluids, Vol 30, pp. 673-688,July 2001
Navier-Stokes Simulation of Local Winds Over
the Earth’s Topography
Abstract
A numerical approach is described that simplifies and automates the CFD
solution process so that Earth scientists can utilize high-resolution Navier-
Stokes flow solvers as a research tool to investigate wind events on the Earth’s
surface. The current approach utilizes the OVERFLOW-2 structured overset
RANS code. A genetic algorithm is used to obtain an optimal multi-zone
overset grid system that reduces the grid size and simulation time by main-
taining high resolution over high-gradient land regions and lower resolution
over low-gradient water regions. Flow simulations are presented that include
flow separation and reattachment over mountainous terrain for coastal islands
in Alaska (USA) and British Columbia (Canada).
1 Introduction
Earth scientists often rely on local meteorological measurements and simple
linear airflow models to study how winds impact the Earth’s surface. Surface
winds directly impact the Earth’s surface by eroding and depositing dust and
soil. High speed surface winds can cause catastrophic damage to forest ecosys-
tems throughout the world . In high latitude coastal regions of North Amer-
ica (southeastern Alaska) and Europe maritime cyclonic windstorms disturb
forested ecosystems by snapping and uprooting trees and causing landslides.
[1]
Reynolds-averaged Navier-Stokes (RANS) simulations have not been read-
ily used to predict winds with micro-scale spatial resolution (grid resolutions
under 1 Kft) because these high-resolution simulations require large compu-
tational resources, and advanced Computational Fluid Dynamic (CFD) codes
often require expert knowledge to effectively use them. However, large super-
computing centers are becoming more readily available at government labs
and universities. For example, NASA’s Advanced Supercomputing (NAS) Di-
vision at Ames Research Center makes available to scientists 10,240 1.5 GHz
2 Solution Procedure
The OVERFLOW-2 CFD RANS code [2] is used to simulate high-resolution
viscous flow over large coastal regions of Alaska and British Columbia. This
code uses overset structured grids, has a variety of steady-state and time-
accurate numerical integration algorithms, and several turbulence models. A
low Mach number preconditioner maintains solution accuracy and code effi-
ciency for low-speed flow simulations. Much of the solution process has been
automated in order to simplify code execution and provide a flow simulation
tool for scientists that are not expert users of CFD codes.
An Automated Planetary Grid (APG) code has been developed to au-
tomatically generate overset grids and OVERFLOW-2 input files. The grid
generation process begins by reading in a Digital Elevation Model (DEM)
of the Earth’s surface. The DEM is an ascii file that defines the terrain ele-
vation (z) on a uniformly spaced xy-grid. The micro-scale spatial resolution
is often on the order of tens to hundreds of feet. Since these coastal regions
involve islands surrounded by large bodies of water, improved computational
efficiency is obtained by keeping land regions at the DEM resolution and re-
ducing the resolution of water regions by a factor of two. These water regions
are flat and have small flow gradients. This is accomplished by partitioning
the terrain surface into structured overset water and land zones.
The zonal partition is established by first subdividing the terrain into small
water and land sub-grids, see Fig. 1. Water sub-grids are regions that only have
zero (sea level) elevations, while land sub-grids are regions that have at least
one non-zero elevation. The process of reassembling these sub-grids into an
optimal water and land zonal topology is based on a Genetic Algorithm (GA)
reported by Holst and Pulliam.[3] Certain rules and constraints are included in
the optimization process. For example, grid zones obtained by reassembling
sub-grids must be rectangular, a requirement of the OVERFLOW-2 code.
Moreover, when forming rectangular zones, water sub-grids can be converted
into higher resolution land sub-grids, but not visa versa. This is done to ensure
DEM resolution is maintained over all land regions. Additional constraints
include a minimum number of sub-grids in a zone, and a maximum zonal
aspect ratio to reduce the GA optimization time and avoid very small isolated
zones.
A two-objective GA optimization is performed, subject to the rules and
constraints mentioned above, that simultaneously minimizes the total number
of land zones and the total number of grid cells, including the overset zonal
Navier-Stokes Simulation of Local Winds Over the Earth’s Topography 331
overlap. A “pareto front” is the result of the optimization process, see the
example in Fig. 2. This pareto front represents the minimum number of grid
cells for a specific number of land zones. Any other combination of land and
water zones will have more surface cells and lie above the pareto front. In the
example shown in Fig. 2, the number of land zones varies from one to five
along the pareto front. Any combination of water and land zones that has
more than five land zones will result in a higher number of surface cells. The
choice of which combination of water and lands zones to use on the pareto
front depends on the requirements of the flow simulation. For example, one
may wish to have fewer land zones to simplify flow visualization or reduce the
number of zonal interfaces. In the cases presented in this study, the number
of zones that results in the least number of grid cells is chosen. This would
correspond to five land zones in the Fig. 2 example.
Once the surface zonal topology is determined, the APG code generates
volume grids using a hyperbolic grid generation method. User input controls
the grid attributes, e.g., initial body-normal spacing, stretching factors, and
far-field grid boundary height. The APG code also automatically generates all
OVERFLOW-2 input files and force/moment input files to monitor solution
convergence. Post-processing software is provided to extract the flow solution
at a fixed distance Above Ground Level (AGL) for data analysis and flow
visualization. The web-based AeroDB framework [4] is used to run the flow
solver and simplify the job submission and solution monitoring process.
3 Numerical Results
The OVERFLOW-2 RANS code [2] uses a Message-Passing Interface (MPI)
for parallel processing. The Pulliam-Chaussee diagonal central-difference al-
gorithm [5], together with matrix artificial dissipation, is used to integrate
the RANS equations. Steady solutions are obtained using a multi-grid tech-
nique with local time-stepping based on a constant Courant-Friedrichs-Levy
(CFL) condition. The Menter SST turbulence model [6] is used to model fully
turbulent flow.
332 Chaderjian, Ahmad, Kramer and Holst
Figure 3 shows the result of the APG zoning algorithm for the Queen
Charlotte Islands, British Columbia, Canada. This is an archipelago of more
than 150 islands with an overall length of 170 miles and width of 83 miles. The
surface grid resolution for land zones is 328 ft while the water zone resolution
is 656 ft. Throughout this paper, the water grid resolution is always twice
as coarse as land grids. The GA optimization selected 21 land grids and 15
water grids, resulting in a 25% reduction in the total grid size. Four slender
water grids 8 miles wide (using exponential stretching) are added around the
perimeter of the grid system to allow the turbulent boundary layer to grow to
sufficient height. Typical Earth boundary layers vary from 1000 ft to over 3000
ft, depending on the time of day and local weather conditions. This 40 zone
grid system consists of 262 million grid points and covers over 14,000 sq. mi.
The serial (1 CPU) APG code took 30 seconds for the GA zonal optimization,
and 50 minutes to generate all 40 volume grids on the NAS Columbia system.
(a) Wind speed shaded surface. (b) Streamlines and wind speed
Dark regions (low speed); light re- shaded surface. Dark regions (low
gions (high speed). speed); light regions (high speed).
Fig. 6. Wind speed shaded surface and streamlines at 27 ft AGL. Mitkof Island,
Alaska. Dark regions (low speed); light regions (high speed).
Fig. 6 shows the wind speed and flow streamlines on a surface 27 ft AGL
for Mitkof Island, Alaska, USA. The free stream conditions are identical to
the Queen Charlotte Islands, except this high-wind event comes directly from
the South. There are 4 land zones at 200 ft resolution and 3 water zones
at 400 ft resolution. As before, water zones are added around the perimeter
334 Chaderjian, Ahmad, Kramer and Holst
R. Keppens1,2,3 , Z. Meliani2,4
1
Centre for Plasma Astrophysics, K.U.Leuven
2
FOM-Institute for Plasma Physics Rijnhuizen, Nieuwegein
3
Sterrenkundig Instituut, Utrecht
4
Max-Planck-Institut fur Astrophysik, Garching
Bj Bi
2 j i j i j i ij
+∂i (ξ + B )v v − − (v · B)(B v + v B ) + δ ptot = 0, (2)
Γ2
∂t τ + ∂i τ v i + ptot v i − (v · B)B i = 0,
(3)
j j i j i i
∂t B + ∂i B v − v B = 0, ∂i B = 0. (4)
In these equations, we introduced the total pressure ptot , containing both
kinetic and magnetic energy density, and we write the energy density τ in the
pre-chosen frame as
(v · B)2 B2
ptot = p + + , (5)
2 2Γ 2
B2 1
τ =ξ+ + (v 2 B 2 − (v · B)2 ) − p − D. (6)
2 2
In the case of an ideal equation of state with constant adiabatic index γ, the
still unspecified variable ξ becomes
γp
ξ = Γ2 ρ + ≡ Γ 2 w. (7)
γ−1
The quantity w (expression inside the brackets) represents the relativistic fluid
enthalpy which includes a rest mass energy contribution. Note in particular
that the magnetic field induction equation (4), and its vanishing spatial di-
vergence, are identical in form to their non-relativistic ideal MHD variants,
Grid-adaptive Simulations of Relativistic Flows 337
and that the relativistic hydrodynamic equations are contained as the limit of
vanishing magnetic field. We further adopted conventional scaling where the
light speed c = 1.
reproduces the main morphological features found in these jets. This picture
can be compared to Fig. 8 of Mignone et al. [10]: one finds good agreement on
the internal shocks in the jet beam, with qualitative reproduction of the fine
structure in the jet cocoon produced by shear flow instabilities at the contact
interface between jet and swept-up shocked ambient material.
Fig. 2. Resulting density structure of a Γ ' 7, beam Mach number vz /cs ' 11
relativistic hydro jet after 80 time units.
Fig. 3. Toroidally magnetized jets, with high (left) and low (right) Poynting flux
at the inlet.
Lorentz factors of varying plasma beta. We will apply both relativistic hy-
dro and magnetohydro simulations to the modeling of the enigmatic Gamma
Ray Burst phenomenon, where current models favor extremely relativistic
flow conditions. To this end, we will demonstrate robustness at high Lorentz
factors (order Γ ' 100) in multi-dimensional simulations.
Acknowledgement. We acknowledge financial support from the Netherlands Orga-
nization for Scientific Research, NWO grant 614.000.421, initial coding efforts by J.
Bergmans, and stimulating discussions with A. Achterberg.
References
1. D. Balsara, ApJSS 132 83 (2001).
2. J. Bergmans, et al., Lect. Not. Comp. Sc. Eng. 41 223 (2005).
3. L. Del Zanna, N. Bucciantini, and P. Londrillo, A&A 400 397 (2003).
4. B. Giacomazzo and L. Rezzolla, astro-ph gr-qc/0507102v1.
5. R. Keppens, et al., Comp. Phys. Comm. 153 317 (2003).
6. A.V. Koldoba, O.A. Kuznetsov, and G.V. Ustyugova, MNRAS 333 932 (2002).
7. S.S. Komissarov, MNRAS 308 1069 (1999).
8. T. Leismann, et al., A&A 436 503 (2005).
9. J.M. Martı́, et al., ApJ 479 151 (1997).
10. A. Mignone, T. Plewa, and G. Bodo, ApJSS 160 199 (2005).
Solution of Laminar Combusting Flows Using a
Parallel Implicit Adaptive Mesh Refinement
Algorithm
1 Introduction
Numerical methods have become an essential tool for investigating combus-
tion processes. In spite of the advances in solution algorithms and computer
hardware, the solution of combusting flows can still place severe demands on
available computational resources. In the previous work, Northrup and Groth
[1] and Gao and Groth [2] have developed a parallel adaptive mesh refinement
(AMR) algorithm that both reduces the overall problem size and the time to
calculate a solution for non-premixed laminar and turbulent combusting flows.
A preconditioned nonlinear multigrid algorithm with multi-stage semi-implicit
time marching scheme as a smoother was used to integrate the governing
partial differential equations. Although accurate solutions were obtained, for
viscous reacting flows, the approach is non-optimal and in many cases a large
number of multigrid cycles and solution residual evaluations were required to
obtain steady state solutions. A fully implicit treatment would seem neces-
sary to deal with the numerical stiffness of the reactive flow problems and
reduce the time to achieve a solution. However, the parallel implementation
of implicit methods with AMR can be challenging.
Recently, a parallel implicit AMR scheme has been proposed that combines
a block-based AMR approach with Newton’s method to solve the equations
governing two-dimensional inviscid flows of compressible gas dynamics [3]. A
Jacobian-free inexact Newton method is used to solve the system of nonlinear
equations arising from a finite-volume spatial discretization procedure and
a preconditioned generalized minimal residual (GMRES) method is used to
solve the resulting system of linear equations at each Newton step. An addi-
tive Schwarz preconditioner is used in combination with block-fill incomplete
lower-upper (BFILU) preconditioning to improve performance of the linear
iterative solver. The Schwarz preconditioning and block-based AMR readily
allow efficient scalable parallel implementations of the approach.
This work considers the extension and application of the preceding implicit
AMR scheme to the prediction of non-premixed laminar diffusion flames. De-
tails are provided concerning the mathematical models used to describe the
laminar flow combustion processes and a description is given of the parallel
implicit AMR finite-volume scheme, including details of the inexact Newton
and GMRES iterative methods, AMR strategy, and parallel implementation.
Numerical predictions are described for methane-air laminar diffusion flames.
The results demonstrate the validity of the parallel implicit AMR approach
and the efficiency of the method for resolving laminar flame structure.
2 Mathematical Modelling
The Navier-Stokes equations for a thermally perfect mixture, given by
∂ ∂
(ρ) + ∇ · (ρu) = 0 , (ρu) + ∇ · (ρuu + pI) = ∇ · τ − ρg , (1)
∂t ∂t
∂ p
[ρE] + ∇ · ρu(E + ) = ∇ · (u · τ − q) − ρg · u , (2)
∂t ρ
∂
(ρcs ) + ∇ · (ρcs u) = ∇ · (ρDs ∇cs ) + ρω˙s , (3)
∂t
are used herein to model the combustion of gaseous fuels and oxidizers. Equa-
tions (1)–(2) reflect the conservation of mass, momentum, and energy for the
reactive mixture, ρ is the mixture mass density, u is the mixture velocity, E
is the total specific energy of the mixture given by E = e + 12 |u|2 , e is the
specific internal energy, p is the mixture pressure, τ is the fluid stress tensor
for the mixture, q is the heat flux vector, and g is the acceleration due to
gravitational forces. Equation (3) is the species concentration equation for
species s, where cs is the species mass fraction, Ds is the diffusion coefficient,
and ω˙s is the time rate of change of the species concentration due to finite-
rate chemistry. It followsP from the caloric equation of state for a thermally
N
perfect mixture that e = s=1 cs hs − ρp , where hs is the species enthalpy, N
is the
PN number of species, and the ideal gas law for the mixture is given by
p = s=1 ρcs Rs T , where Rs is the species gas constant and T is the mixture
temperature.
Expressions for species and mixture thermodynamic and transport proper-
ties are required to complete the system of partial differential equations for the
reactive mixture. The empirical expressions complied by Gordon and McBride
[4] are used to specify the hs and the species specific heat, cps , entropy, ∆ss ,
viscosity, µs , and thermal conductivity, κs , as functions of temperature. Per-
fect mixture rules are used to determine the thermodynamic properties of the
reactive mixture and Wilke’s [5] and Mason and Saxena’s [6] mixture rules
are used to evaluate the mixture viscosity, µ, and thermal conductivity, κ, re-
spectively. The species diffusion coefficients, Ds are calculated by additionally
specifying a Schmidt number, Scs = µ/ρDs for each species.
Simplified one- and two-step reduced chemical reaction mechanisms are
used to model the chemical kinetics of methane-air combustion [7]. In the
one-step scheme, methane oxidation is represented as follows:
Parallel Implicit AMR for Laminar Combusting Flows 343
4 Numerical Results
The proposed parallel implicit AMR algorithm is now applied to the solution
of an axisymmetric co-flow methane-air diffusion flame. In particular, the
solution of the steady laminar flame studied by Mohammed et al. [19] and
Day and Bell [20] is considered. The flame boundary and initial conditions
are the same as those used in the previous studies. The Mach and Reynolds
number based on the diluted methane fuel flow in the inlet are M = 0.0016
Parallel Implicit AMR for Laminar Combusting Flows 345
0.04 0.04 0.04 T (K)
2200
2160
2120
2080
2040
2000
1960
1920
1880
1840
1800
1760
1720
1680
1640
1600
1560
1520
1480
1440
1400
0.02 0.02 0.02 1360
1320
1280
1240
1200
1160
z (m)
z (m)
z (m)
1120
1080
1040
1000
960
920
880
840
800
760
720
680
640
600
560
520
480
0 0 0 440
0 0.01 0 0.01 0 0.01 400
Initial Level - 96 Blocks(4x8) 3rd Level - 195 Blocks (4x8) 5th Level - 396 Blocks (4x8)
3072 Cells 6240 Cells 12672 Cells
and Re = 169. Additional details concerning this case are given elsewhere
[19, 20, 1].
The predicted solution for the laminar diffusion flame obtained using the
parallel implicit AMR method is shown in Figure 1. The figure shows the
computed isotherms and flame structure obtained using a sequence of adap-
tively refined grids starting from the initial mesh (96 solution blocks with
3,072 cells) and proceeding to the final mesh after five levels of refinement
(396 blocks with 12,672 cells). A comparison of the results of Figure 1 with
those given in the previous studies [19, 20] reveals, that in spite of the inherent
simplifications used in the two-step reaction mechanism, the predicted flame
structure agrees very well with the previous work. The “wishbone” structure
of the high-temperature region is present and the computed lift-off and flame
heights are 0.05 cm and 3.3 cm, respectively, with a maximum centre-line
temperature of 2080 K. All of these values agree reasonably well with the
previously published results. The predicted value of the CO mass fraction
concentration at z = 3 cm along the centerline is cCO = 0.026 and, consid-
ering the limitations of the reduced chemistry mechanism used herein, is in
good agreement with the value of Mohammed et al. [19], who report a mass
fraction of cCO = 0.03 at this location.
Estimates of the parallel performance and scalability of the proposed al-
gorithm for the laminar diffusion flame problem are shown in Figure 2. The
figure depicts both the parallel speed-up, Sp = (t1 /tp )p, and parallel efficiency,
Ep = Sp /p, for a fixed-size problem as a function of the number of processors,
p, where tp is the processor time required to solve the problem using p proces-
sors and t1 is the time required to solve the problem using a single processor.
The performance curves are shown and compared to their idealized values for
up to 48 processors for the diffusion flame calculation on a mesh consisting of
96 4×8 cell solution blocks (3,072 cells). Although there are some inefficiencies
in the Schwarz preconditioning, it is quite evident that the speedup is nearly
linear and the efficiency remains above about 90% for up to 48 processors.
346 Scott A. Northrup and Clinton P. T. Groth
1
40
30
0.6
20
0.4
Relative Speedup
10 Relative Efficiency 0.2
Ideal Speedup
Ideal Efficiency
0
10 20 30 40
Number of Processors (p)
Fig. 2. Parallel performance of the Newton algorithm for laminar diffusion flame
calculation on a mesh consisting of 96 4×8 cell solution blocks (3,072 cells) showing
the parallel speed-up, Sp = (t1 /tp )p, and efficiency, Ep .
5 Conclusions
A parallel implicit AMR scheme has been developed for solving laminar com-
busting flows. The combination of finite-volume discretization procedure, par-
allel block-based AMR strategy, low-Mach-number preconditioning, and New-
ton method solution procedure has resulted in a powerful computational tool
for predicting a wide range of steady laminar reactive flows.
References
1. Northrup, S. A. and Groth, C. P. T., Paper 2005–0547, AIAA, January 2005.
2. Gao, X. and Groth, C. P. T., Paper 2006-1448, AIAA, January 2006.
3. Groth, C. P. T. and Northrup, S. A., Paper 2005-5333, AIAA, June 2005.
4. Gordon, S. and McBride, B. J., Reference Publication 1311, NASA, 1994.
5. Wilke, C. R., J. Chem. Phys., Vol. 18, 1950, pp. 517–519.
6. Dixon-Lewis, G., Combustion Chemistry, edited by W. C. Gardiner, Springer-
Verlag, New York, 1984, pp. 21–126.
7. Westbrook, C. K. and Dryer, F. L., Combust. Sci. Tech., Vol. 27, 1981, pp. 31.
8. Weiss, J. M. and Smith, W. A., AIAA J., Vol. 33, No. 11, 1995, pp. 2050–2057.
9. Barth, T. J., Paper 93-0668, AIAA, January 1993.
10. Roe, P. L., J. Comput. Phys., Vol. 43, 1981, pp. 357–372.
11. Coirier, W. J. and Powell, K. G., AIAA J., Vol. 34, No. 5, May 1996, pp. 938.
12. Sachdev, J. S., Groth, C. P. T., and Gottlieb, J. J., Int. J. Comput. Fluid Dyn.,
Vol. 19, No. 2, 2005, pp. 157–175.
13. Saad, Y., Iterative Methods for Sparse Linear Systems, PWS, Boston, 1996.
14. Dembo, R. S., Eisenstat, S. C., and Steihaug, T., SIAM J. Numer. Anal.,
Vol. 19, No. 2, 1982, pp. 400–408.
15. Gropp, W. D., Kaushik, D. K., Keyes, D. E., and Smith, B. F., Parallel Com-
puting, Vol. 27, 2001, pp. 337–362.
16. Knoll, D. A. and Keyes, D. E., J. Comput. Phys., Vol. 193, 2004, pp. 357–397.
17. Mulder, W. A. and van Leer, B., J. Comput. Phys., Vol. 59, 1985, pp. 232–246.
18. Aftosmis, M. J., Berger, M. J., Murman, S. M., Paper 2004-1232, AIAA, 2004.
19. Mohammed, R. K., Tanoff, M. A., Smooke, M. D., A. M. Schaffer, A. M., and
Long, M. B., 27th Symposium on Combustion, Combustion Institute, Pitts-
burgh, 1998, pp. 693–702.
20. Day, M. S. and Bell, J. B., Combust. Theory Modelling, Vol. 4, No. 4, 2000,
pp. 535–556.
Towards Direct Numerical Simulation of a
Diffusion Flame-Shock Interaction with an
AMR Algorithm
1 Introduction
The simulation of unsteady reactive flows where very stiff phenomena occur
requires algorithms with specific properties. In particular, strong and rapid
fluctuations have to be captured precisely. It is therefore necessary to use a
method that is able to describe on a large range of Mach number the physical
and chemical events with good accuracy and stability if the wavelike oscil-
latory behavior is to be taken into account. In this paper we present three
simulations realized with the reactive solver MAJIC [1]. It is a second order
accuracy solver based on time splitting, MUSCL approximation and one step
time scheme on uniform Cartesian meshes.
The hyperbolic terms of the Navier-Stokes equations are generally the most
difficult to solve because of the presence of nonlinear terms. For these terms,
adaptive limiters have been proposed in [1] [2] [3] associated with AUSM+
flux splitting [4]. This approach is compared to ENO and Jameson’s schemes
in the case of freely decaying isotropic turbulence.
The parabolic terms have also to be discretized correctly. A centered
second-order scheme in space is used. In order to study the parabolic op-
erator, first the diffusion of two species and generation of acoustic waves after
one time step is studied and compared with the analytical solution. Then as
we proceed in time, a quasi-incompressible convection-diffusion flow which
couples the hyperbolic and parabolic operators is simulated.
Finally, in order to show that the code is able to work also in the do-
main of compressible reactive flows, we associate a source operator with the
hyperbolic-parabolic operators. A reactive supersonic flow where a H2 -air cir-
cular diffusion flame interacts with a shock is studied. For this last case,
MAJIC has been coupled with an adaptive mesh refinement (AMR) plat-
form developed at ONERA ([5],[6]). The AMR method is the one presented
by Berger, Oliger, Colella, Quirk and many other authors (see the website
http://seesar.lbl.gov/AMR/) and adapted to finite differences (see [7]).
constant pressure, hi the enthalpy per unit mass of the ith species and e the
internal energy per unit mass. If we assume
CP = CP1 Y1 + CP2 Y2 and λ = λ1 X1 + λ2 X2 , (Xi is the mole fraction of
the ith species) with
CP i = ani T n + ... + a0i , λi = bm i T
m
+ ... + b0i (n, m) ∈ N,
we are able to obtain the pressure space variation at time t1 = t0 + δt from
the space variations of the mass fraction of H2 at time t0 :
2
(W2 −W1 )
p(x, t1 ) = p(t0 ) + (t1 − t0 ) CP [Yc1 (W 2 −W1 )+W1 ]
G(x, t0 ) with
h i
W1 W2 (λ1 −λ2 ) λ
G(x, t0 ) = λY1xx + [Y1 (W2 −W1 )+W1 ]2 − CP (CP1 − CP2 ) Y1x Y1x .
c represents the speed of sound and Wi the molecular weight of the ith species.
When the molecular weights are equal, the diffusive interface does not pro-
duce acoustic waves, but as soon as the molecular weights are different, two
acoustic waves set up and propagate in the opposite directions. At t0 , we take
Y1 = 21 (1 + tanh C(xc − |x − x0 |)), x0 = 5mm, xc = 2.5mm, T (x) = 1000K,
p(x) = 1atm., δt = 10−9 s, ∆x = 25µm. At t1 , for two initial thicknesses
δY ≈ 100µm (C = 500) and δY ≈ 1mm (C = 50) of the diffusion layer de-
fined as δY = |xY1 =0.001 − xY1 =0.999 |, we plot p(x, t1 ) − p(t0 ) and the acoustic
pressure obtained by the simulation. Fig. 2 shows the analytical and numerical
curves. When the diffusion layer is spread off on only few cells, the maximum
numerical values are perceptibly weaker. As soon as the thickness encloses a
sufficiently number of cells (δY ≥ 10∆x), the curves are close and the code
reproduces well the phenomenon. The log-log plot of the error versus grid size
350 G. Billet, J. Ryan and M. Borrel
dp (exact)
dp (num)
3 dp (exact)
300
dp (num)
C = 500
C = 50
200 2
1
100
dp (pa)
dp (pa)
0
0
-1
-100
-2
-200
-3
0.0024 0.0025 0.0026 0.0015 0.002 0.0025 0.003
x (m) x (m)
for the parabolic operator is shown on Fig. 2. For this last result, the thick-
ness remains constant (δY ≈ 1mm) and the grid size varies. The second-order
(s ≈ 2) is obtained for log10 ∆x ≤ −2.6 (δY ≥ 40∆x).
y
0.004 0.004 0.004
0 0 0
0 0.005 0.01 0 0.005 0.01 0 0.005 0.01
x x x
Fig. 3. Acoustic field at three different times (left: t = 10−6 s, center: t = 5 10−6 s,
right: t = 6 10−5 s).
4.2 Results
For this AMR computation, the finest mesh size is ∆x = ∆y = 25µm. Two
levels of embedded grids have been used (of ratio 2). Patches are defined
using a sensor based on the gradient of all primitive variables. During the
computation, the finest time step is 10−11 s ≤ δt ≤ 5 10−10 s. The results are
presented at a time when half of the bubble has crossed the shock. The density
distribution is shown on Fig. 4 where a typical shape of the shock deformed
by the flame is visible. At the same time, the distribution of some species are
plotted on Fig. 4 (for y ≥ 0). The shock has not a great influence on YH2O , but
the distribution of the other presented species changes strongly, particularly
for YOH , YHO2 , YH2O2 . These variations are due to the heat supply of the
shock that develops temporarily an endothermal process in the flame and
modifies the kinetic of the reactions and particularly the reactions that have
the shortest characteristic times.
5 Conclusion
The MAJIC solver effectiveness has been shown with different flows. For the
first flow, the results are compared with those of high order methods. For the
second flow, numerical and analytical solutions have been compared. Circular
permutation of the split-ups of the hyperbolic operator has shown to be effec-
tive in keeping the flow symmetry. The structure of this solver is well adapted
to an AMR strategy. MAJIC integration within the AMR platform has been
validated. The domain of research is thus enlarged since we may simulate
detailed reactive phenomena where high speed flows keep close to low speed
flows. Even if some work remains to be done to optimize the coding, for exam-
ple concerning high performance parallel computation, perspectives for com-
puting complex 3D flows are quite good. The physical models can be adapted
with the level of refinement : Euler/Navier-Stokes, mono/multi-species, empir-
ical laws/detailed transport mechanisms, global/detailed chemical scheme,...
352 G. Billet, J. Ryan and M. Borrel
y
4.0E-03 0.038 0.005
0.0015 0.0015
0.0005 0.0005
0.003 0.004 0.005 0.006 0.003 0.004 0.005 0.006 0.003 0.004 0.005 0.006
x x x
y
9.0E-04 3.4E-05 6.2E-06
0.0015 0.0015 2.0E-05 0.0015 3.0E-06
0.003 0.004 0.005 0.006 0.003 0.004 0.005 0.006 0.003 0.004 0.005 0.006
x x x
Fig. 4. Density field and mass fraction zooms of H2O, OH, H, HO2 and H2O2
species when the flame crosses the shock.
References
1 Introduction
The behaviour of thin liquid films whether forced to spread or deposited as a
distinct pattern on the surface of a substrate, is of enormous significance to
many manufacturing and biological processes. The topic of the present study is
the flow of continuous thin liquid films over surfaces containing topographical
features. In the electronics sector (displays, printed circuits, micro-devices,
sensors etc), for example, the industrial goal is often to minimise free surface
deviations from planarity either for aesthetic reasons or to ensure predictable
product properties [1].
These three-dimensional flows present extremely challenging practical de-
sign problems since free surface disturbances caused by even small-scale to-
pography can persist over length scales several orders of magnitude greater
than the actual size of the topography. The focus here is on the numerical
modelling of such flows, an area which is still in its infancy. The majority of
previous numerical studies have modelled the flow using the long wave, lu-
brication approximation which reduces the three-dimensional Navier-Stokes
equations to more tractable two-dimensional fourth order partial differential
equations for film thickness, pressure and (for evaporative flow) solvent con-
centration. Most have solved the resultant time-dependent lubrication equa-
tions with alternating-direction implicit (ADI) algorithms using alternating
sweeps in each direction so that only a banded system of equations needs to
be solved at each time step [2].
Recently, however, a Multigrid approach with adaptive time-stepping has
been developed as a more efficient alternative to ADI schemes, having already
been applied successfully to droplet spreading flows [3] and continuous film
flow with [4] or without evaporation [5]. This paper highlights the additional
benefits of adopting an automatic local grid refinement procedure within the
Multigrid algorithm which allows fine grids to be used only where they are
needed, i.e. near the topographies themselves, and much coarser grids to be
used over the rest of the substrate. Several previous numerical studies have
Free surface
z
H(X,Y)
y
RT Inclined plane
S0
LP
WP
!
density ρ, viscosity µ and surface tension σ, and its motion governed by the
Navier-Stokes and continuity equations, viz:
∂U
ρ + U .∇U = − ∇P + µ ∇2 U + ρ g, (1)
∂T
∇.U = 0, (2)
where U = (U, V, W ) and P are the fluid velocity and pressure respectively,
T is time and g = g(sin θ, 0, − cos θ) is the acceleration due to gravity.
Assuming that = H0 /L0 is small, where H0 and L0 are the character-
istic film thickness and in-plane length scale respectively, yields the following
lubrication equation for non-dimensional film thickness h,
∂ h3 ∂p ∂ h3 ∂p
∂h
= −2 + , (3)
∂t ∂x 3 ∂x ∂y 3 ∂y
and pressure, p:
6 2
p = − 3 ∇2 (h + s) + 61/3 N (h + s) , (4)
β β
Adaptive Multigrid Solutions of Thin Film Flows over Topography 355
1/3
σH0
where L0 = β 3ρg sin θ and N measures the influence of gravity on free
surface shape. Topographies are defined via arctangent functions which enable
the steepness of their sides to be controlled easily and the boundary conditions
are that the flow is fully developed upstream and downstream. Further details
are given in [4].
3 Numerical Method
3.1 Spatial Discretisation
The lubrication equations (3) and (4) are solved on a square computational
domain, (x, y) ∈ Ω = (0, 1) × (0, 1), with equal, uniform grid spacings in the
x and y directions, ∆ say, leading to the following discretised equations:
1 h3 h3
∂hi,j
= 2 |i+ 12 ,j (pi+1,j − pi,j ) − |i− 12 ,j (pi,j − pi−1,j ) +
∂t ∆ 3 3
3 3
h h
| 1 (pi,j+1 − pi,j ) − | 1 (pi,j − pi,j−1 ) −
3 i,j+ 2 3 i,j− 2
3
h3
2 h
|i+ 12 ,j − |i− 12 ,j , (5)
∆ 3 3
6
pi,j + (hi+1,j + si+1,j ) + (hi−1,j + si−1,j ) + (hi,j+1 + si,j+1 ) +
∆2
√
3
(hi,j−1 + si,j−1 ) − 4(hi,j + si,j ) − 2 6N (hi,j + si,j ) = 0, (6)
3 3
for each, (i, j), in the computational domain. The terms, h3 |i± 12 ,j , h3 |i,j± 21
are the pre-factors obtained from linear interpolation between neighbouring
vertices. Time integration is performed using the standard, second-order ac-
curate Crank-Nicholson method and writing the right hand side of equation
(5) as F (hi,j , pi,j , hi±1,j , pi±1,j , hi,j±1 , pi,j±1 ) leads to the equation
∆tn+1
hn+1
i,j − F (hn+1 n+1 n+1 n+1 n+1
i,j , pi,j , hi±1,j , pi±1,j , hi,j±1 , pi,j±1 )
n+1
2
∆tn+1
= hni,j + F (hni,j , pni,j , hni±1,j , pni±1,j , hni,j±1 , pni,j±1 ), (7)
2
for which ∆tn+1 = tn+1 − tn , and the right hand sides are given in terms of
known values at the end if the nth time step, t = tn .
356 Y.C. Lee, H.M. Thompson, and P.H. Gaskell
Equations (6) and (7) and are solved using a Full Approximation Storage
Multigrid algorithm within a Full Multigrid Cycle. Error reduction is per-
formed using a fixed number of pre- and post- Red-Black Gauss-Seidel New-
ton relaxations. On the coarsest grid level, the discretised equations are solved
using Newton iteration. Adaptive local mesh refinement is implemented us-
ing the Multi-Level Adaptive Technique (MLAT) first proposed by Brandt [8].
The τ -indicator is used to quantify errors since information is readily available
from the different grid levels. The difference in truncation error on successive
grids Gk and Gk−1 is approximated via a relative truncation error quantity,
τkk−1 , with large values of τkk−1 indicating regions of significant error between
successive grid levels and where correspondingly further grid refinement is
necessary.
A general approach to the discretisation at local refinement interfaces is
to conserve numerical flux at both the coarse and fine locally refined regions,
where the numerical flux across a control volume is defined via;
Z Z
F = u · n dS . (8)
S
For equation (7)
3 3
h ∂p h ∂p
u = uh = −2 , , (9)
3 ∂x 3 ∂y
and for equation (6)
6
u = up = − ∇(h + s) . (10)
β3
Further details of the adaptive approach used here are given in [9].
4 Results
The cases considered here are for the flow of thin water films of asymptotic
film thickness H0 = 100µm, viscosity 0.001Pa s, density ρ = 1000kg m−3
and surface tension σ = 0.07N m−1 down a substrate inclined at 30◦ to the
horizontal and with a constant inlet flow rate Q0 = 1.635 × 10−6 m2 s−1 [1].
These parameters yield a Capillary length Lc = 0.78mm and N = 0.122,
the latter value indicating that gravity has little influence on the free surface
shape. All results are obtained using an FMG V(4,2) cycle with a coarse grid
with 9x9 nodes in each direction and finest grid levels up to 513x513 (k = 6).
Figure 2 illustrates the flexibility of the adaptive approach by solving flow
over a square 39mm x 39mm domain containing three trenches each of char-
acteristic length 3.9mm, depth 10µm and with square, circular and a diamond
cross-sectional shapes indicated on the right hand plan view of the computa-
tional grids. The left hand side of the Figure shows the resultant free surface
Adaptive Multigrid Solutions of Thin Film Flows over Topography 357
3
10
2
10
adaptive
non-adaptive
1
10 2 3 4 5 6
grid level, k
Fig. 3. Comparison between CPU times for adaptive and non-adaptive solution of
flow past a square trench.
358 Y.C. Lee, H.M. Thompson, and P.H. Gaskell
than the non-adaptive ones. For the case of a 513x513 grid (k = 6), for exam-
ple, the adaptive solution calculates a solution of equal accuracy in less than
10% of the CPU time.
5 Conclusion
References
1. Decre, M.J., Baret, J,C.: Gravity driven flows of viscous liquids over two-
dimensional topographies. J. Fluid Mech., 487, 147–166 (2003).
2. Schwartz, L.W., Eley, R.R.: Simulation of droplet motion on low energy and
heterogeneous surfaces. J. Coll. Int. Sci., 202, 173–188 (1998).
3. Gaskell, P.H., Jimack, P.K., Sellier, M., Thompson, H.M.: Efficient and accurate
time adaptive multigrid simulations of droplet spreading. Int J. Numer. Meth.
Fluids, 45, 1161–1186 (2004).
4. Gaskell, P.H., Jimack, P.K., Sellier, M., Thompson, H.M., Wilson, M.C.T.:
Gravity-driven flow of continuous thin liquid films on non-porous substrates
with topography. J. Fluid Mech., 509, 253–280 (2004).
5. Gaskell, P.H., Jimack, P.K., Sellier, M., Thompson, H.M.: Flow of evaporat-
ing, gravity-driven thin liquid films over topography. Phys. Fluids, 18, 031601
(2006).
6. Bai, D., Brandt, A.: Local Mesh Refinement Multilevel Techniques. SIAM J.
Sci. Stat. Comput., 8(2), 109–134 (1987).
7. Trangenstein, J.A., Kim, C.: Operator Splitting and Adaptive Mesh Refinement
for the Luo-Rudy I Model. J. Comput. Phy., 196, 645–79 (2004).
8. Brandt, A.: Multi-Level Adaptive Solutions to Boundary-Value Problems. Math.
Comp., 31, 333–90 (1977).
9. Lee, Y.C., Thompson, H.M., Gaskell, P.H.: An efficient adaptive multigrid algo-
rithm for predicting thin film flow on surfaces containing localised topographic
features. Computer and Fluids, to appear (2006).
A meshless solver for computing viscous flows
on Cartesian like grids
1 Introduction
Cartesian mesh calculations provide great potential for future industrial cal-
culations. While the unstructured data based finite volume method has had
success in terms of solving inviscid flows on such grids, the performance of
the new class of meshless solvers (in particular LSFD-U), is yet to be tested.
It is needless to mention that the Cartesian grids provide a most natural way
of obtaining point distribution for meshless solvers. The present attempt is
to evaluate the performance of the meshless solver LSFD-U on a point dis-
tribution obtained from a Cartesian grid. It is remarked that the LSFD-U
procedure has already been successfully used for solving inviscid flows [1, 2].
Also, certain aspects of its viscous implementation along with positivity anal-
ysis are presented in reference [3]. Similar to Cartesian mesh calculations, one
of major hurdles in the extension of meshless solver for viscous flows, is the
non-positivity of the viscous discretization procedure.
Inspired by the success of the structured grid in computing viscous flows,
a novel grid stitching strategy has been proposed [4] for generating what we
refer to as “Cartesian like grids”. The details of this procedure are presented in
the next section. These grids have been successfully used for the computation
of laminar viscous flows [5, 6] within the framework of unstructured data
based cell center finite volume procedure. In this work, the point distribution
obtained from Cartesian like grids is used in conjunction with the LSFD-U
solver.
3 LSFD-U
. NW
x
NE
x
. g .j NW NE
N
x x x .
Wx O
xE
f
.
J . x
SW
x
S
x
SE
. o NW N NE
x x
. Wx O
. xE
x
Wx x
. O
. x x x
. SW S SE
x
SW
x
S
x
SE
(a) Cluster for LSFD-U (b) Cartesian type (c) Hanging node type
1.14653 0.959
1.08556 0.908
1.02458 0.857
0.963612 0.806
0.902639 0.755
0.841666 0.704
0.780693 0.653
0.71972 0.602
0.658747 0.551
0.597774 0.5
0.536801 0.449
0.475828 0.398
0.414855 0.347
0.353882 0.296
0.292909 0.245
0.231936 0.194
0.170963 0.143
0.10999 0.092
0.0490175 0.041
(a) Point distribution (b) Equi φ contours: Variant 1a (c) variant 1a+limiter
The LSFD-U solver along with point distribution obtained from Cartesian
like grids, is employed to solve inviscid and viscous flows. Numerical results
are compared in Figure 3 with available experimental or standard numerical
values. Also, the solutions are compared with the results obtained from un-
structured data based cell center finite volume solver, HIFUN-2D on the same
grids. van Leer and Roe schemes are used for interfacial convective fluxes for
inviscid and viscous flows, respectively. SGS implicit relaxation procedure [2]
is employed for convergence acceleration. Strong boundary treatment involv-
ing no slip condition and Riemann boundary condition are implemented at
the wall and far field points, respectively. Algebraic Baldwin-Lomax turbu-
lence model is implemented to compute eddy viscosity. Details of initial and
solution adaptive Cartesian like grids are shown in table 1. Details of the test
cases are described in table 2.
The finite volume solutions compare well with benchmark results on the ini-
tial grids. Whereas the LSFD-U solver requires finer grids. Finer grids are
obtained using solution adaptive gridding strategy. The surface coefficient dis-
tribution obtained on adapted grids using LSFD-U solver compare well with
the standard results. It can be observed for laminar viscous flow solutions,
the oscillations in the wall data are reduced in case of finer grids. However,
unacceptable oscillations are seen in turbulent flow test case for both pressure
and skin friction coefficient distribution. The loss in robustness of the flow
solver on highly stretched turbulent flow grids is primarily attributed to non-
positive viscous discretization procedure and ill-conditioning of the geometric
matrix associated with the LSFD-U procedure.
5 Conclusions
The point distribution obtained from Cartesian like grids in conjunction with
meshless solver LSFD-U, is employed for simulating inviscid and viscous flows.
Partial success is achieved for turbulent flow computations involving highly
stretched viscous grids. In our view, the work presented has provided useful
direction to further research in meshless solvers as applied to viscous flows.
The present effort is towards evolving suitable grid generation and discretiza-
tion strategies for solving turbulent flows within LSFD-U framework.
p
−C
3
−1
−2
−0.5 0 0.5 1 1.5
x/c
(a) Grid1 (b) Grid in the slotted region (c) −Cp distribution
Inviscid flow: NLR 7301, M∞ =0.185, α=6o
1.5 1
cart−lsfdu−level1
cart−lsfdu−level0
1 Fortunato
cart−fv−level0
0.5 0.5
−Cp
Cf
−0.5
0
cart−lsfdu−level1
cart−lsfdu−level0
Fortunato
−1
cart−fv−level0
−1.5
0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 1 −0.5
0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 1
x/c x/c
0.4 cart−lsfdu−level0
0.1
cart−lsfdu−level1
venkatakrishnan
0.2
cart−fv−level0
−0.2
−Cp
Cf
0
−0.4
−0.6
cart−lsfdu−level0
cart−lsfdu−level1
−0.8 venkatakrishnan
cart−fv−level0
−0.1
−1
−1.2
0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 1
0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 1
x/c x/c
cart−lsfdu−level1
5 cart−lsfdu−level0
1
Exeriment
0.5
3
−Cp
0
Cf
1
−0.5
cart−lsfdu−level1
cart−lsfdu−level0
0
Exeriment
−1 cart−fv−level0
−1
−1.5 −2
0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 1 0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 1
x/c x/c
1 Introduction
The Immersed Boundary (IB) method simplifies the grid generation process
for the simulation of flows with complex and/or moving solid boundaries by
avoiding the need for a body-fitted mesh. The IB technique was originally
developed for incompressible flows [1, 2, 3] using Cartesian grids. Recently,
some of the authors have extended the IB technique to compressible flows [4]
using the preconditioned Navier–Stokes equations, which allow one to provide
accurate and efficient solutions for a wide range of the Mach number. To date,
IB methods employ structured grids, which allow only limited control on the
distribution of the grid points in the computational domain; in fact, clustering
of grid points is needed close to solid boundaries in order to describe its ge-
ometry accurately and, since mesh lines run through the entire computational
domain, a high concentration of grid points is obtained also in regions away
from the solid walls, where flow gradients are usually small.
In order to cope with this problem, a flexible local grid refinement technique
is to be employed, increasing the mesh resolution near the body. This work
extends the method of [4] to the solution of three-dimensional flows using a
semi-structured grid obtained by coarsening a uniformly fine mesh far from
high-gradient flow regions, such as boundary layers and shocks.
Local grid refinement allows for efficient clustering of cells close to the im-
mersed boundary. The basic idea was recently introduced by Durbin and Iac-
carino [9] for a finite difference discretization and extended to a finite volume
formulation by Iaccarino et al. [10]. The following description of the algorithm
refers to two dimensions, the extension to three-dimensions being straightfor-
ward. An auxiliary structured grid is employed to handle the data structure
of the semi-structured locally refined grid (shown in figure 1). The auxiliary
grid covers the whole computational domain employing the finest mesh size
on the semi-structured grid. Therefore, each cell of the semi-structured grid is
bounded by the lines passing through the vertices (i, j) and (i + ∆i, j + ∆j),
see figure 1, where the indeces i = 0, . . . , Ni and j = 0, . . . , Nj refer to the
auxiliary (finest) grid and ∆iε1, ∆jε1 depend on (i, j). Therefore, having an
auxiliary grid with Ni × Nj cells, the N < Ni × Nj cells belonging to the
semi-structured grid are defined using the two couples of indices (i, j) and
(∆i, ∆j), with a total memory requirement of 4 N integers. In addition, an
array of integers, ID(i, j), is needed to store the correspondence between the
cells of the auxiliary and semi-structured grids. All of the cells of the for-
mer grid not employed in the latter one, namely, those included in the range
[i : i + ∆i − 1] and [j : j + ∆j − 1], are tagged using the same cell number. The
total storage required for allocating ID(i, j) is, therefore, Ni × Nj integers.
The connectivity information for each cell is retrieved by querying the array
ID(i, j).
Immersed boundary compressible flow simulations 367
N1
j+!j=const
(i+2,j+2)
W1
P E1
W2
(i,j)
j=const
S1 S2
i=const i+!i=const
where φj,wall is the value of the flow variable to be imposed at the immersed
surface, αi = 1/di and βj = 1/dj , di and dj being the distances between the
surrounding cell centers and the interface cell center and between the wall
intersections and the interface cell center, respectively, and
N
Xnbr Nib
X
q= αi + βj . (3)
i j
It can be shown that in the one-dimensional case, this procedure coincides with
the linear interpolation scheme used in [2, 4]. The pressure gradient along the
normal to the immersed surface is set to zero by assigning the corresponding
flow field value at the interface cell. In case of an adiabatic surface, also the
temperature gradient is set to zero.
3 Results
0.19
0.18
0.17
0.16
St
0.15
0.14
T*=1.0 exp.
T*=1.1 exp.
0.13 T*=1.5 exp.
T*=1.8 exp.
T*=1.0 present
T*=1.1 present
0.12 T=*1.5 present
T*=1.8 present
0.11
60 80 100 120 140 160
Re
Fig. 2. Strouhal number vs Reynolds number for the unsteady flow past a heated
circular cylinder.
Immersed boundary compressible flow simulations 369
4
3
3
2
2
1
1
Y
0
Y
0
-1
-1
-2
-2
-3
-3
-4
-5 -4 -3 -2 -1 0 1 2 3 4 5
X
-4
-5 -4 -3 -2 -1 0 1 2 3 4 5
The steady turbulent supersonic flow past a circular cylinder has been con-
sidered as a suitable test case to validate the method for compressible flows
with shocks. The flow with M∞ = 1.7 and Re∞ = 2 × 105 has been computed,
the inlet values of the turbulence kinetic energy and specific dissipation rate
2
being k/U∞ = 0.0009 and ωD/U∞ = 40, respectively. For the considered val-
ues of M∞ , a bow shock is formed upstream of the cylinder; the subsonic flow
behind the shock close to the cylinder accelerates along its surface forming
a supersonic-flow region, enveloping the subsonic recirculation region behind
370 de Tullio et al.
the cylinder, and two symmetric tail shocks are formed at the end of the sep-
aration region. Results have been obtained using a rectangular computational
domain with dimensions [−10 D; 15 D] × [−10 D; 10 D], D being the diameter
of the cylinder centred at the origin. Standard characteristic boundary condi-
tions have been imposed at the inlet and outlet surfaces and free-shear wall
boundary conditions are imposed at the far-field boundaries. Numerical results
are obtained using the third-order-accurate upwind scheme. Simulations on a
coarse grid have been performed at first to locate the position of the shocks,
approximatively. Based on such a solution, a semi-structured grid with 75556
cells and 545699 faces has been obtained starting from a very fine auxiliary
grid with 1805 × 2159 cells and removing cells away from the shock regions. A
local view of such semi-structured grid is shown in figure 3. The Mach number
contours are given in figure 4, showing that a clear description of the shocks
and of the wake are obtained, thanks to the local grid refinement in those
regions. The computed separation angle, measured clockwise from the leading
edge, is equal to 113◦ and agrees well with the corresponding experimental
datum, 112◦ [13]. Moreover, the computed and experimental drag coefficients
are equal to 1.41 and 1.43, respectively. All numerical results agree very well
with the numerical solutions obtained using the same numerical method and
a body-fitted grid [4] and reasonably well with the experimental ones.
References
1. C. S. Peskin, J. Comput. Phys. 10 (1972) 252.
2. E. A. Fadlun, R. Verzicco, P. Orlandi, J. Mohd-Yosuf, J. Comput. Phys. 161
(2000) 35.
3. R. Mittal, G. Iaccarino, Annu. Rev. Fluid Mech. 37 (2005) 239.
4. P. D. Palma, M. de Tullio, G. Pascazio, M. Napolitano, Comput. Fluids 35
(2006) 693.
5. D. C. Wilcox, Turbulence models for CFD, 2nd Edition, DCW Industries, Inc.,
1998.
6. C. L. Merkle, in: M. Hafez, K. Oshima (Eds.), Computational Fluid Dynamics
Review 1995, John Wiley & Sons, pp. 419–436 (1995).
7. T. H. Pulliam, D. S. Chaussee, J. Comput. Phys. 39 (1981) 347.
8. H. van der Vorst, SIAM J.Sci.Statist.Comput. 13 (1992) 361.
9. P. A. Durbin, G. Iaccarino, J. Comput. Phys. 128 (2002) 110.
10. G. Iaccarino, G. Kalitzin, P. Moin, B. Khalighi, Paper AIAA-2004-0586.
11. A.-B. Wang, Z. Trávnı́ĉek, K.-C. Chia, Phys. Fluids 12 (2000) 1401.
12. M. Sabanca, F. Durst, Phys. Fluids 15 (2003) 1821.
13. V. A. Bashkin, A. V. Vaganov, I. V. Egorov, D. V. Ivanov, G. A. Ignatova,
Fluid Dynamics 37 (2002) 473.
Incompressible Flow Simulations Using Virtual
Boundary Method with New Direct Forcing
Terms Estimation
In this paper, a new external forcing terms estimation in the virtual boundary
method is proposed. In the usual virtual boundary method with feedback and
direct forcing terms estimations, the unphysical oscillations near the boundary
appear in the pressure field. In order to remove these unphysical oscillations,
the new direct forcing terms are added not only near the boundary but also
inside the boundary. In the incompressible flows past a circular cylinder and
a sphere, the pressure oscillations near the boundary can be removed and
the flow characteristics on the boundary are in very good agreement with the
reference results.
1 Introduction
For the practical flow simulations, the boundary fitted coordinates (BFC) is
usually adopted. This BFC approach has the high adaptability to the body
configuration. However, in the complicated flow geometry, it is difficult to
generate the computational grid. And in the BFC approach, it is necessary
that the governing equations are transformed from the physical plane to the
computational plane, so that the transformed governing equations have more
terms than the original equations in Cartesian coordinates. Therefore, the
computational effort is larger than the Cartesian coordinates. Then, in recent
years, Cartesian grid approach is highlighted again for the numerical flow
simulations. There are mainly three ways for Cartesian grid approach, i.e.,
immersed boundary method, cut cell method, and building cube method.
In this work, we consider the virtual boundary method [1]-[4] that belongs
to the immersed boundary method. In order to satisfy the velocity condi-
tions at the (virtual) boundary points, the virtual boundary method requires
the external forcing terms added to the momentum equations. These forcing
terms are usually estimated by two ways, that is, feedback [1] and direct [3]
procedures. Also, in both forcing terms estimations the grid points added the
forcing terms are restricted only near the boundary. Therefore, the unphysical
oscillations near the boundary appear in the pressure field. These unphysical
oscillations give the difficulty of pressure prediction on the (virtual) boundary.
In the present work, we propose the new direct forcing terms estimation in
order to remove these unphysical oscillations.
In order to validate the present new direct forcing terms estimation, the
direct numerical simulations (DNSs) of incompressible flow past a circular
cylinder and a sphere are considered.
U is the specified velocity and U = 0 in solid media. Then, the type of pres-
sure equation can be changed automatically from Poisson equation in the fluid
region to Laplace equation in the solid region.
In order to calculate the additional forcing terms, there are two ways in
the fractional step approach. One is to calculate in Eq.(3) and another is to
calculate in Eq.(4). The present forcing terms estimation gives the reasonable
solutions in both calculations. However, in the usual direct forcing estimations,
the reasonable solutions cannot be obtained for calculating forcing terms in
Eq.(4). Then, in this work, we adopt the calculation of forcing terms in Eq.(3).
(a) Usual feedback forcing (b) Usual direct forcing (c) Present direct forcing
(a) Usual feedback forcing (b) Usual direct forcing (c) Present direct forcing
Fig. 7. Iso-surface of ∇2 p
4 Concluding Remarks
In this paper, a new direct forcing terms estimation is proposed in the virtual
boundary method. The present approach is applied to the numerical simula-
tions of flow past a circular cylinder and a sphere. In comparison with the
results obtained by usual feedback and direct forcing terms, the present ones
show that the smooth pressure field can be obtained so that the character-
istic quantities are in more excellent agreement with the reference solutions.
Also, since the pressure in solid media can be obtained precisely, the present
approach is very fruitful for fluid-structure interaction analysis.
This work was supported in part by a Grant-in-Aid for Scientific Research
(16560146) from the Japan Society for the Promotion of Science.
References
1. Goldstein, D., Handler, R., Sirovich, L.: Modeling a No-slip Flow Boundary with
an External Force Field. J. Comput. Phys., 105, 354–366 (1993)
2. Saiki, E.M., Biringen, S.: Numerical Simulation of a Cylinder in Uniform Flow:
Application of a Virtual Boundary Method. J. Comput. Phys., 123, 456–465
(1996)
3. Fadlun, E.A., Verzicco, R., Orlandi, P., Mohd-Yusof, J.: Combined Immersed-
Boundary Finite-Difference Methods for Three-Dimensional Complex Flow Sim-
ulations. J. Comput. Phys., 161, 35–60 (2000)
4. Nishida, H.: Cartesian Grid Approach with Virtual Boundary Method and Its
Applications. Notes on Numerical Fluid Mechanics, Springer, 78, 17–30 (2001)
5. Nishida, H., Satofuka, N.: A Variable Order Method of Lines: Accuracy, Conserva-
tion and Applications. Lecture Notes in Computational Science and Engineering,
Springer, 21, 167–174 (2002)
6. Fornberg, B.: A Numerical Study of Steady Viscous Flow Past a Circular Cylin-
der. J. Fluid Mech., 98-4, 819–869 (1980)
7. Dennis, S.C.R., Chang, G.Z.: Numerical Solution for Steady Flow Past a Circular
Cylinder at Reynolds Number up to 100. J. Fluid Mech., 42-3, 471–489 (1970)
8. Rosenfeld, M.: Grid Refinement Test of Time-Periodic Flows over Bluff Bodies.
Computers Fluids, 23-5, 693–709 (1994)
A Hybrid Building-Block and Gridless Method
for Computing Shock Waves
1 Introduction
The numerical methods for the solution of the compressible Euler and Navier-
Stokes equations can be classified by the mesh they use as structured grid
methods, unstructured grid methods, Cartesian grid methods, and gridless
methods. Each of these methods, advocated, promoted, developed, and used
by their respective supporters, has its own advantages and disadvantages. The
structured grid methods have a disadvantage in mesh generation for complex
geometries. The main advantage of the unstructured grid methods is the ease
of grid generation for complex configurations. However, the computational
costs and memory requirements are generally higher than their structured
grid counterparts. The advantages of the Cartesian grid methods include ease
of grid generation, lower computational storage requirements, and significantly
less operational count per cell. However, the main challenge in using Cartesian
methods is how to deal with arbitrary boundaries, as the grids are not body-
aligned and the cells of a Cartesian mesh near the body can extend through
surfaces of boundaries. Lately, gridless methods came into focus. This class
of methods is essentially propelled by the fact that there exists a perceived
difficulty in generating volume filling grids for complex geometries in spite
of significant progress in the theory and practice of mesh generation over the
last decade. However, global conservation of mass, momentum, and energy for
these methods is not necessarily assured. Furthermore, these gridless meth-
ods are generally slower than their mesh-based counterparts. Recently, a new
approach [1] named building-cube method has been developed for large-scale
high resolution flow computations around complex geometries. In this method,
a flow field is divided into a number of cubes (squares in 2D) of various sizes.
Each cube is a computational sub-domain with a Cartesian mesh of equal
spacing and equal number of nodes. In this approach, the wall boundary is
defined by a staircase representation in order to keep the simplicity of the
algorithm and to minimize the memory requirement per node. Unfortunately,
2 Numerical Method
The basic idea behind the present hybrid method is to combine a building-
block method and a gridless method in an attempt to develop a fast, low
storage method for time-accurate computation of the unsteady Euler equa-
tions for complex geometries. In this method, body geometries are first defined
and the desired grid sizes in the flow field are specified via a background mesh
and/or sources. A baseline block (root block) covering the computational do-
main is then refined to create eight blocks of equal size. In turn, these blocks
are refined again until a suitable block size is achieved, which is determined by
the specified number of cells in x-, y-, and z-directions in a block and the pre-
scribed grid size. Each block is a sub-domain of the flow computation, where
a uniform-spacing Cartesian mesh is used. All blocks have the same number
of Cartesian cells in all three directions, so that the local computational res-
olution is determined by the block size. In order to minimize the numerical
errors between blocks, a smoothing of the size differences among blocks is
performed, only allowing the size difference between two adjacent blocks to
two. An unstructured data structure is used to manage the blocks and com-
municate with each other. Meanwhile, a triangulation of the body geometries
is performed using the advancing front method based on the representation
of the body geometries and the prescribed grid size. The center points of tri-
angles are chosen as a set of boundary points, which will be updated using
the gridless method. Note that such chosen points have a well-defined nor-
mal direction, therefore avoiding ambiguity of normal definitions for surface
singularities such as the tip of an airfoil trailing edge. Since a majority of
computational cells is solved using the building-block Cartesian method, the
A Hybrid Building-Block and Gridless Method 379
3 Computational Results
A. Steady two-dimensional oblique shock wave
A 2D shock wave reflecting from a rigid surface ((ρ, u, v, p)(0,y,t) =
(1, 2.9, 0, 1/1.4), (ρ, u, v, p)(x,1,t) = (1.69997, 2.61934, −0.50632, 1.52819)) is
considered to validate and verify the baseline building-block Cartesian grid
method, which is the foundation of the proposed hybrid method. The com-
putational domain is a rectangle of length 4.1 and height 1. Fig. 1 shows
the unstructured boundary surface mesh generated using a specified uniform
grid size of 0.05, the computed density contours in the flow field, and the
comparison of the computed density distribution with the exact solution at
y = 0.5. When the numbers of cells in x-, y-, and z-directions are specified as
82x20x20, a single block will be generated automatically. When these numbers
are halved, eight blocks will be created. When the numbers of cells in all three
directions are set to one, 32,800 (82x20x20) blocks will then be generated.
The CPU cost between the building-block Cartesian grid method and a
unstructured grid method [4] for 500 time steps is compared in Table 1, where
the number of faces for the building-block Cartesian grid and the number of
edges for the unstructured grid are also given. The unstructured tetrahedral
grid consists of 283,359 elements, 52,148 points, and 8,276 boundary points.
Note that computing cost for the building-block Cartesian grid method and
the unstructured grid method is proportional to the number of faces and edges,
respectively, where Riemann fluxes need to be evaluated for each method
and represent the most dominate CPU consuming operations. On the one
hand, when a single block is used, the building-block Cartesian grid method
380 Hong Luo, Joseph D. Baum and Rainald Löhner
2.8 2.8
2.6 Exact solution 2.6
Computation
2.4 2.4
2.2 2.2
2 2
Density
1.8 1.8
1.6 1.6
1.4 1.4
1.2 1.2
1 1
0.8 0.8
0 0.5 1 1.5 2 2.5 3 3.5 4 4.5
X-coordinates
Fig. 1. Unstructured boundary surface mesh used for the representation of gridless
points for stationary shock reflection problem (nface=16,548,nboun=8,276), com-
puted density contours in the flow field for stationary shock reflection problem, and
comparison of the computed density profile with the exact solution at y=0.5 for
stationary shock reflection problem.
Table 1. Computing cost analysis for the building-block Cartesian grid and
unstructured-grid methods.
Method FEFLO B-Block(32,800) B-Block(8) B-Block(1)
CPU (Second) 956 197 135 132
No of faces (edges) 343,780 102,080 102,080 102,080
Speed-up (Fluxes) 1 3.37 3.37 3.37
Speed-up (Others) 1 1.44 2.10 2.15
Speed-up (Total) 1 4.85 7.08 7.24
4 Conclusions
A hybrid Building-block Cartesian grid and gridless method has been devel-
oped for solving the unsteady compressible Euler equations. The developed
method combines the efficiency of a Cartesian method and the flexibility of a
gridless method for the complex geometries. The numerical results obtained
indicate that the use of this hybrid method leads to a significant improvement
in performance over its unstructured grid counterpart without compromising
solution accuracy, demonstrating the great potential and benefits of this hy-
brid method for the simulation of transient shock interaction problems. An
overall speed-up factor from six to twelve and a saving in storage requirements
up to one order of magnitude for a typical 3d simulation in comparison with
the unstructured grid method are obtained.
References
1. Nakahashi, K., and Kim, L.S.: Building-Cube Method for Large-scale, High Res-
olution Flow Computations. AIAA Paper 2004-0423 (2004)
2. Kirshman, J., and Liu, F.: Gridless Boundary Condition Treatment for a Non-
Body-Conforming Mesh. Journal of Computational Physics, Vol. 201, No. 1, pp.
119-147 (2004)
3. Luo, H., Baum, J.D., Löhner, R.: A Hybrid Cartesian Grid and Gridless Method
for Compressible Flows. Journal of Computational Physics, Vol. 214, pp. 618-632
(2006)
4. Luo, H., Baum, J.D., Löhner, R.: A Fast, Matrix-free Implicit Method for Com-
pressible Flows on Unstructured Grids. Journal of Computational Physics, Vol.
146, pp. 664-690 (1998)
382 Hong Luo, Joseph D. Baum and Rainald Löhner
Summary. Conventional error indicators can be effectively employed for grid re-
finement, but provide no information on error levels in the domain, and consequently
no “termination criterion”. Error estimators, on the other hand give reasonable es-
timate of error distribution and serve to decide suitable “termination criterion”.
In this paper, we discuss the design of a new residual–based error referred to as
“<–parameter”. The relation between the estimated error and the global errors are
investigated. Consequently, the <–parameter is used to design an effective AMR
strategy, which is applied successfully to inviscid flows in two and three dimensions.
1 Introduction
The ultimate aim of any grid adaptation strategy is to minimise the global
error. Therefore, some of the earlier works deal with estimating the global
error itself for the purpose of adaptation. On the other hand, in this work,
we estimate the local truncation error, in the form of the <–parameter. The
global error (GE) can be shown to be composed of two parts: Transmitted
error, which is error convected by the flow from “error sources” and Cell error,
which is error generated in each cell. The <–parameter is a cell–level imbalance
and can detect “error source”, which the global error cannot. Consequently,
an adaptive strategy based on the residual would be preferred as it results
in a faster decrease in error levels, and better error control. These arguments
are verified on the isentropic flow (M∞ = 0.6, AoA= 0o ) flow past a NACA
0012 airfoil. Starting from the same initial grid, a fixed percentage of cells
are refined at each level using the global error and the <–parameter. The
final grids resulting from this selective refinement are shown in Fig.2(a) and
Fig.2(b). Error levels resulting from residual based refinement are seen to
be lower than those from resulting from the global error based refinement
(Fig.2(c)).
(a) Final grid from (b) Final grid from < (c) Global error plot
GE based refinement based refinement
Fig. 2. Global and estimator error studies on isentropic flow past airfoil
386 Ganesh N., Nikhil V Shende, and Balakrishnan N.
5 Numerical Simulations
The proposed adaptive strategy is applied to three different test cases, rep-
resentative of different flow physics. All simulations are performed using the
in-house code HIFUN–2D/HIFUN–3D, which are based on cell–centered finite
volume approach. Time integration is achieved using an implicit Symmetric
Gauss Seidel (SGS) procedure [4] and Venkatakrishnan’s limiter [7] is used to
enforce monotonicity. VanLeer’s scheme [6] is used to calculate the interfacial
fluxes and the divergence of velocity is employed as an error indicator, unless
otherwise mentioned.
5.1 Inviscid flow past double ellipse (M∞ =8.15, AoA= 30o )
Fig. 3. Initial, intermediate and final meshes for flow past double ellipse
6 Conclusion
A new residual–based error estimator referred to as the <–parameter has
been developed from from first principles. The consistency of the estimator
and its relation with the global error are investigated. It is shown that the
388 Ganesh N., Nikhil V Shende, and Balakrishnan N.
Fig. 5. Initial, intermediate and final meshes for flow past ONERA M6 wing
References
1. Bassi F., Rebay S., JCP, Vol.138, pp 251-285, 1997
2. Desideri. et.al.(Ed), Hypersonic flow for reentry problems, Vol.2, 1991
3. Ganesh N., Nikhil V Shende, Balakrishnan N., Proc. 7th AeSI CFD Symposium,
2004
4. Nikhil V Shende, Balakrishnan N., AIAA Jl., 42(9), Sep. 2004
5. Sun M., Takayama K., Error localization in solution-adaptive grid, JCP,
Vol.190, pp 346-350, 2003
6. van Leer B.,Report No. 82-30, NASA Langley Research Centre, Virginia.
7. Venkatakrishnan, V., “Convergence to Steady State Solutions of the Euler
Equations on Unstructured Grids with Limiters”, JCP, Vol. 118, 1995, pp 120–
130.
Anisotropic solution-adaptive technique
applied to simulations of steady and unsteady
compressible flows
1 Introduction
Grid adaptation is a very powerful tool for optimizing CFD calculations. Un-
fortunately the typical isotropic adaptation used for 3D flows may result in
excessive number of elements. The reason for this is that during refinement of
the grid around, e.g. the shock wave the grid cells becomes smaller uniformly
in all directions. It means that after splitting N cells in the direction normal
to the shock wave the number of the cells will be increased to N K 3 (K de-
scribes how many new edges are created after splitting an old one - if an edge
of a cell is split into two edges then K = 2). When the anisotropic adaptation
is used the grid is refined only in the direction normal to the shock wave and
the number of the cells would be increased to N K.
2 Anisotropic Adaptation
The adaptation algorithm presented in this paper relies on grid remeshing.
The remeshing is based on the generation of a new grid in the Riemann space
in which uniform spacing of the cells is assumed. The final spacing of the new
grid is a function of the metric field used for the definition of the Riemann
space. Detailed description of the anisotropic grid generation algorithms can
be found in [4] and [2].
The metric used for the definition of the Riemann space is obtained from
the solution field by means of the presented error estimator. The estimator
provides data in a form of a metric tensor which describes the directionality of
flow features. Several possibilities exist with regard to the choice of the error
estimator. In the present case the flow solver is using linear reconstruction
within each cell. Therefore it is convenient to use an estimate of the interpo-
lation error as the error indicator ([3], [4]).
Calculation of the metric. Assume that E denotes a grid cell inside which a
function u is being interpolated and xc is the center of E. Then after dropping
terms of higher order, the interpolation error for E can be estimated [4] as:
εE ≤ max x − xc )T |H|(x − xc
(1)
x∈E
4 Numerical Results
Adaptation of the MHD steady flow in a sinusoidal 3D channel. The
first testcase is the steady MHD flow through a 3D channel with a sinusoidal
bump, with supersonic inlet flow. The wall is fully conducting, hence van-
ishing normal B-field is imposed. Inside the channel the pattern of oblique
shock waves should be resolved. The initial grid and the grid after the 3rd
adaptation are shown on the Figure 1. The corresponding field of magnitude
of magnetic induction vector is shown on Figure 2. It is worth noting that sig-
nificant improvement of the solution was obtained while the number of points
was increased by a factor less than 1.6.
a)
b)
Fig. 1. Grids used for calculation of the MHD problem inside a channel with sinu-
soidal bump: (a) initial grid (23117 nodes), (b) grid after 3rd adaptation loop (36361
nodes)
Anisotropic solution-adaptive technique 393
4.30776
4.01185
3.71595
3.42004
3.12413
2.82822
2.53231
2.23641
1.9405
1.64459
1.34868
1.05277
a)
4.30776
4.01185
3.71595
3.42004
3.12413
2.82822
2.53231
2.23641
1.9405
1.64459
1.34868
1.05277
b)
Fig. 2. The field of the magnitude of the magnetic induction vector: (a) calculated
on the initial grid (b) calculated on the grid after 3rd adaptation loop
5 Conclusions
The presented anisotropic adaptation algorithm has been successfully applied
to steady and unsteady inviscid flows. The adaptation has improved the qual-
ity of the solution and recovered details of flow which were missing in the
solution based on uniform grid. The presented algorithm is capable to locally
predict the grid areas that should be adapted.
394 Jerzy Majewski and Aristotelis N. Athanasiadis
a)
6.2
5.8
5.4
4.9
4.5
4.1
3.6
3.2
2.8
2.3
1.9
1.5
1.0
0.6
0.2
b)
Fig. 3. Results after third adaptation for t = 3.0: (a) grid (42193 nodes), (b) density
field
Acknowledgements The work was done with cooperation with Von Karman
Institute and was supported by Belgian Federal Science Policy Office.
References
1. Alauzet, F., George, P.L.,Mohammadi, B., Frey, P., Borouchaki, H.: Transient
fixed point-based unstructured mesh adaptation, Int. J. Numer. Meth. Fluids,
43, 729–745 (2003)
2. Athanasiadis, A.N.: Three-Dimensional Hybrid Grid Generation with Applica-
tion to High Reynolds Number Viscous Flows, Ph.D. thesis, Von Karman Insti-
tute, Universit´e Libre de Bruxelles (2004)
3. Castro-Dı́az, M.J., Hecht, F., Mohammadi, B., Pironneau, O.: Anisotropic Un-
structured Mesh Adaptation for Flow Simulation, International Journal for Nu-
merical Methods in Fluids, 25, 4, 475–491 (1997)
4. George, P.L., Borouchaki,H.: Delaunay Triangulation and Meshing - Application
to Finite Element, Hermes, ISBN 2-86601-692-0, Paris (1998)
5. Formaggia, L., Perotto, S.: Anisotropic Error Estimation for Finite Element
Methods, 31st Computational Fluid Dynamics, VKI Lecture Series 2000-05
(2000)
6. Majewski, J.: An anisotropic adaptation for simulation of compressible flows,
Mathematical Modelling and Analysis, 7, 127–134 (2002)
7. Woodward, P.R., Colella, P.: The numerical simulation of two-dimensional flows
with strong shocks, J. Comput. Phys., 54, 115–173 1984
Simulation of Flow around Wing Sections by
Building-Cube Method
1 Introduction
Flow around an airplane includes various scales; from micro-scale near the
surface to macro-scale around the airplane. The macro-scale flow basically
determines the surface pressure distribution around an airplane. The micro-
scale flow dynamics also has a significant contribution to such as the skin
friction, the flow separations and the aerodynamic noise generation. Current
CFD has become a powerful tool to accurately simulate macro-scale flows for
designing airplanes. The micro-scale flow simulation, however, is still difficult
due to the extremely large difference in the flow characteristic length.
In order to simulate the micro-scale flows in the macro-scale, an approach
based on Cartesian grid named ’Building-Cube Method (BCM)’ [1] was devel-
oped. The main strategy of the method is to use very fine and orderly grid in
the boundary layer regions in order to resolve micro-scale flows. The objective
of this study is to apply the BCM to simulate flows around airfoils/wing in
order to evaluate the capability of the method.
2 Building-Cube Method
Fig. 1. Two-dimensional cube frame (left) around A-airfoil and an enlarged view
(right) of the Cartesian grid around a tripping wire.
3 Computed Results
3.1 NACA0012 Airfoil with Trailing Edge Modifications
Two-dimensional flows around NACA0012 airfoil at relatively low Reynolds
number of 5 × 104 based on the chord length were simulated to observe the
398 Kazuhiro Nakahashi and Yota Sakurai
effect of small trailing edge modification to the flows. The free stream Mach
number is 0.5 and the angle of attack is three degrees. With this Reynolds
number, the boundary layer on the airfoil is essentially laminar [5] so that the
2D computations were performed without using turbulence models.
The number of cubes is 522 and each cube has 64x64 Cartesian mesh in
it. The total number of computational cells is 522 × 64 × 64 = 2, 138, 112. The
minimum spacing in the near-wall cube is 4.58 × 10−4 to the chord length.
Outer boundary is located at 60 chord length.
Figure 3(a) shows the pressure distribution around the original airfoil.
The laminar boundary layer on the upper surface separates at mid-chord and
generates vortices that flows down along the airfoil surface. When a vortex
reaches at the trailing edge, it generate a pressure wave that propagates in all
directions from the trailing edge.
Fig. 3. Pressure distributions around NACA0012 airfoils; (a) original, (b) with a
slit, (c) with a downward plane.
In order to investigate the cause of the pressure wave generation, the trail-
ing edge was modified as shown in Fig.3(b) and (c). The modification shown
in Fig.3(b) is to add a slit of 0.01c width at 0.9 chord length. The second one
shown in Fig.3 (c) is to add a small downward plate of 0.02 chord height at
the trailing edge. These geometrical modifications are very easy because of
the Cartesian grid in each Cube.
With these small modifications near the trailing edge, the frequencies of
the vortex shedding were significantly affected as shown in Fig.3(b) and (c).
These results indicate an existence of a feedback loop between the pressure
wave at the trailing edge and the vortex shedding on the upper surface due
to the laminar separation.
Simulation of Flow around Wing Sections by Building-Cube Method 399
The Aerospatiale A-airfoil at an angle of attack =13.3 deg, the maximum lift
configuration, and at Re = 2.1 × 106 and M∞ = 0.15 was chosen for the
3D computation because of the availability of experimental data as well as
computational results using RANS and LES [6]. The outer boundary of the
computational region locates at 25 chord length far from the airfoil. Total
number of cubes is 569. The minimum spacing near the airfoil surface is
1.9 × 10−4 with respect to the chord length if each cube employs 128x128
Cartesian mesh and 0.95 × 10−4 for 256x256 mesh. The tripping wire used
in the experiment at x/c=0.3 on the lower side was geometrically included in
the grid as shown in Fig. 1.
For 3D computations, 3D zone is selected similar to Mary and Sagaut
computation [7]. That is, the airfoil boundary layer on the upper side is covered
by 3D cubes but the one on the lower side is treated as 2D as shown in Fig. 2.
With this case, the number of 3D cubes is 141 and the number of 2D cubes is
428. The spanwise extent is 0.012 to the chord length and a periodic condition
was used in the spanwize direction. The computation was started by the 2D
flow field with 64x64 grid in all cubes. Then, the grid in cube is doubled to
setup the initial flowfield. The 3D field is computed by using the 2D result
as the initial condition. All 2D and 3D computations are carried out without
using any turbulence models.
The 2D and 3D results are shown in Fig. 4 for instantaneous density distri-
bution near the leading edge. In the 2D result, small vortices are generated in
a short distance from the leading edge and flows downward. The 3D result also
shows the generation of small vortices. However, those vortices disappear in
a short distance and the flow becomes fully three-dimensional. This indicates
a transition process; the development of 2D waves, followed by the formation
of 3D flow structures, as was also shown by Mary and Sagaut [7].
400 Kazuhiro Nakahashi and Yota Sakurai
The pressure plateau near the leading edge and the boundary layer separa-
tion on the upper surface near the trailing edge observed by experiment were
well reproduced by the present computations of 2D of 128x128 and 256x256
as well as 3D with 128x16x128 grid. In the 3D computations, however, a large
vortical structure appeared when the spanwise resolution was increased with
the same spanwize extent. The scale of the vortical structure is much larger
than the spanwise extent. More investigation is required especially for the
spanwize extent in 3D computations.
4 Conclusions
References
1. Nakahashi, K.: High-Density Mesh Flow Computations with Pre-/Post-Data
Compressions, AIAA 2005-4876, 17th AIAA CFD Conf. (2005).
2. Obayashi, S., Guruswamy, G. P.: Convergence Acceleration of a Navier-Stokes
Solver for Efficient Static Aeroelastic Computations, AIAA J., 33, 6, 1134–1141
(1995).
3. Yamamoto, S., Daiguji, H.: Higher-Order-Accurate Upwind Schemes for Solving
the Compressible Euler and Navier-Stokes Equations, Computer and Fluids, 22,
2/3, 259–270 (1993).
4. Matsuno, K.: Improvement and Assessment of an Arbitrary-Higher-Order Time-
Accurate Algorithm, Computer and Fluids, 22, 2/3, 311–322 (1993).
5. Elimelech,Y., et. al.: On the Onset of Transition at Low Reynolds Number Flow
Over Airfoils, AIAA 2005-5311 (2005).
6. Mellen, C. P., Frohlich, J., Rodi, W.: Lessons from LESFOIL Project on Large-
Eddy-Simulation of Flow Around an Airfoil, AIAA J., 41, 4, 573–581 (2003).
7. Mary, I., Sagaut, P.: Large Eddy Simulation of Flow Around an Airfoil, AIAA
2001-2559, 15h AIAA CFD Conf. (2001).
Robust Mesh Deformation using the Linear
Elasticity Equations
Richard P. Dwight
1 Introduction
The ability to deform meshes according to changes in the computational do-
main is a fundamental tool in CFD, and a foundation on which many other
processes are built. It is for example essential in aerodynamic optimization,
where remeshing the geometry after each design step would cause the change
in the quantity of interest to be swamped by the discretization error [2].
Achieving high-order accuracy in time is much easier when the mesh topology
remains unchanged between time steps [4], but grid deformation is commonly
used in a wide variety of other applications simply because it tends to be
considerably cheaper and more convenient than remeshing; an example is
aero-elastic coupling.
However in practice most mesh deformation algorithms suffer from ro-
bustness problems when faced with large deformations (in particular those
involving large relative motion of bodies), but poor quality input meshes
cause difficulties for small deformations too. Even worse many important
problems involve these two aspects combined; a particularly acute example
is the movement of the flaps and slats of 3d high-lift configurations, for which
few algorithms perform satisfactorily.
One method that shows considerable promise models the mesh as an elas-
tic solid using the equations of linear elasticity. It was applied in [1], together
with refinement and derefinement for time-dependent problems undergoing
substantial deformation. Stein et al. [3] have applied it with an elastic stiff-
ness varying in inverse proportion to cell volume thereby preserving quality
2 Governing Equations
The equations of linear elasticity govern small displacements u(x) = (u, v, w)
of an elastic solid subject to body forces and surface tractions. Using the
summation convention they may be written
∇·σ =f on Ω, σ = λTr()I + 2µ, (1)
where f is some body force, Ω is the computational domain, and σ is the
stress tensor, given in terms of the strain tensor by the constitutive relation,
where Tr is the trace, λ and µ are the Lamé constants, and are a property
of the elastic material. It is convenient to express these in terms of Young’s
modulus E and Poisson’s ratio ν as
νE E
λ= , µ= . (2)
(1 + ν)(1 − 2ν) 2(1 + ν)
E > 0 may be thought of as the stiffness of the material, where large E
indicates rigidity. Poisson’s ratio ν is a measure of how much the material
shrinks in the lateral direction as it extends in the axial direction; for physical
materials −1 < ν < 12 . The linear kinematic law = 12 ∇u + ∇uT , quantifies
the change in length and orientation of a material fibre in the elastic body.
Boundary conditions are Dirichlet, u = g on ∂Ω, completing the system.
A significant advantage of the application of these equations for mesh
deformation is that diverse features required in practice can be readily and
simply realized in a manner consistent with the model of the mesh as an
elastic body. For example for a symmetry plane it is convenient that mesh
points move, but remain within the plane, which may be implemented by
the specification of only one coordinate of the displacement. For adaptive
methods it may be desirable to deform the mesh such as to cluster mesh
points in regions of interest, which is readily possible by applying a non-zero
body force f .
U h = uh |uh ∈ H h (Ω)n , uh = g h on ∂Ω ,
Φh = φh |φh ∈ H h (Ω)n , φh = 0 on ∂Ω ,
160
Std Elastic - NACA
Mod Elastic - NACA
140 Std Elastic - Highlift
Mod Elastic - Highlift
Max deflection (Degrees)
120
100
80
60
40
20
0
-1 -0.5 0 0.5
Poisson Ratio !
Fig. 1. Maximum possible deflection for a single step of the standard and modi-
fied elastic models for the NACA0012 and high-lift configurations. Initial and 51◦
defected high-lift grids.
Increment 2°
101 Increment 10°
Increment 30°
Smallest angle in triangles (Degrees)
Increment 90°
100
10-1
0 1 2 3 4
Rotations (Degrees/360)
5 Numerical Results
Two two-dimensional test cases are considered: a Euler grid for a NACA0012
single element aerofoil with a coarse fully unstructured Euler grid, a circular
farfield at a distance of 40 chord lengths, and the aerofoil is rotated about
Robust Mesh Deformation 405
its nose while the farfield is held stationary, see e.g. Figure 2. Also a NS grid
about a three-element high-lift configuration where the slat is deflected and
the remaining elements held fixed, see Figure 1. The robustness of the modified
equations is compared with that of the standard equations for −1 < ν < 0.5,
by determining the maximum deflection that still results in a valid grid, with
no negative volumes. In all cases E is proportional to the inverse of the cell
volume. The results are shown on the left of Figure 1; the modified constitutive
law allows a 70% greater rotation of the NACA, and a 40% greater deflection
of the flap. The grid for the greatest valid flap deflection is shown on the right
of Figure 1. Near the body the point distributions are almost unmodified due
to the much greater elastic stiffness there, a desirable property given that the
initial meshes are likely to resolve surface regions well.
Provided intermediate surface definitions exist, exceptionally large defor-
mations are possible if several steps are taken. For example for the NACA0012
by first calculating the grid for a 90◦ turn an additional 90◦ rotation was pos-
sible. Figure 2 shows the NACA rotated completely 4 times, the left-hand plot
shows the rotation angle against the smallest angle in the grid (a measure of
grid quality) for several increment sizes. The same may be done for the high-
lift case, resulting in a valid grid for a flap deflection of 155◦ , the tip of the
flap almost touching the main element, and an entire grid block compressed
into the intervening space, Figure 3. Of course a flow solution on this grid
reveals defects in the solution, notably the black triangle of low pressure at
the trailing edge of the main element. Finally Figure 4 shows the deformation
of a 3d turbulent wing-body configuration mesh, whereby the calculation was
performed in two steps, a deflection first to 45◦ and then to 90◦ . The mesh
close to the wing is almost unchanged; because of the relatively rapid grid
coarsening with distance from the surface, the algorithm is able to put the
weight of the deformation onto exclusively large cells.
Y
X
Smallest angle in quads (Degrees)
0
10
Y
X
-1
10
Z
Increment 1° Y
X
Increment 2°
Increment 5°
Increment 10°
Increment 30°
0 50 100 150
Flap Deflection (Degrees)
and mesh refinement could be performed based on this indicator [1]. Finally
a demanding practical test case should be considered.
References
1. T. Baker and P. Cavallo. Dynamic adaption for deforming tetrahedral meshes.
37th AIAA Aerospace Sciences Meeting, Reno, NV. AIAA-1999-3253, 1999.
2. R.P. Dwight and J. Brezillon. Effect of various approximations of the discrete
adjoint on gradient-based optimization. Proceedings of the 44th AIAA Aerospace
Sciences Meeting and Exhibit, Reno NV. AIAA-2006-0690, 2006.
3. K. Stein, T. Tezduyar, and R. Benney. Mesh moving techniques for fluid-
structure interactions with large displacements. J. Appl. Mech., 70:58–63, 2003.
4. Z. Yang and D. Mavriplis. Unstructured dynamic meshes with higher-order
time integration schemes for the unstready Navier-Stokes equations. 43th AIAA
Aerospace Sciences Meeting, Jan. 10-13, Reno, NV. AIAA-2005-1222, 2005.
Part V
where K = ρkΦU
µH is the non dimensional coefficient of permeability for k
the intrinsic permeability and Φ the porosity of the medium without gravity
(in practise we set K = 1020 in the fluid and K = 10−8 inside the obstacles),
Re = ρUµH is the non dimensional Reynolds number and (Uref , pref ) is a
reference flow. A Poiseuille flow is given upstream and a no-slip boundary
condition is set on the walls of the channel, the boundary condition on ΓN
allows to convey properly the vortices downstream [2].
The equations are solved in primitive variables on Cartesian staggered grids
with a multigrid algorithm coupled to a cell-by-cell relaxation procedure..
The discretization is done by a global second order scheme in time and space
including an accurate third order approximation of convection terms [4]. A
view of the flow behind the arrays of cylinders is given on figure 1.
The theory concerning the wavelet packets is detailed in [6] and will not be
reminded here. The same Daubechies type wavelets are used in the current
paper to build the packets array, and the entropy criterion is used to select
the best basis. In [6], few tests have been performed in order to get the best
wavelet mother, and to determine the number of scales necessary for an effi-
cient representation of the flow.It is shown that it is necessary to go over at
least 5 scales from the finest scale corresponding to a 640 × 2560 grid to the
coarsest scales as long as one is concerned with smoothing the discontinuities.
Using an appropriate threshold as in [5] for an academic test it is possible
to separate the structures of the flow. The overall filtering process can be
summarised as follows:
1. Computation of the wavelet packets decomposition of the two components
of the velocity U = (u, v) over 5 scales.
2. Separation of the velocity fields into two sub fields: one sub field, Us =
(us , vs ), where the wavelet packet coefficients have their modulus larger
Role of coherent structures in 2D turbulence 411
than a given threshold and another one, If = (uf , vf ) where the wavelet
packet coefficients are smaller than .
3. Construction of the corresponding vorticity fields, ωs and ωf . The filtered
field ωs is then essentially formed of the solid rotation part of the vortices,
and the filtered field ωf is formed of the vorticity filaments that roll up in
spirals inside the vortices as can be seen on figure 2 on a window of the
flow behind the cylinders.
4. Computations of the physical data: energy, enstrophy, spectra and fluxes.
100
200
300
400
500
600
100
200
300
400
500
600
Fig. 2. Snapshots at the end of the channel (kinj ≈ 20). Whole pressure and
vorticity fields (top) and filtered vorticity sub fields (bottom).
412 Ch.-H. Bruneau, P. Fischer and H. Kellay
Using this efficient filtering, it is possible to separate the whole flow into two
orthogonal coherent sub flows that have their own dynamics as it is shown
below. The flow is not splitted into coherent vortices and background or noise
but into solid rotation part of the vortices and filaments that are also present
inside the vortices. The first sub field contains more than 95% of the energy
and 70% of the enstrophy while the second subfield contains less than 5% of
the energy but 30% of the enstrophy. That is why the dynamics of the flow
starting from the first subfield deviates from the dynamics of the whole flow
as an important part of the enstrophy is missing. The first sub field is strongly
related to the pressure field and can isolate the vortices while the second sub
field is a continuous field that operates the necessary link between the vortices
to yield the whole dynamics.
For the last step of the filtering process a collection of 25 windowed snapshots
thanks to a Tukey window with a small parameter is used to compute the
energy spectra and the fluxes. The time between two snapshots is long enough
for describing a significant evolution of the flow. Indeed between the 1st and
the 25th snapshots, the vortices cross several times the observation box.
The global spectrum shows clearly on figure 3 that two ranges coexist on both
sides of the injection scale which is k = 20 in this simulation and the sum of
the spectra of the two orthogonal sub fields corresponds to the spectrum of
the whole flow.
To confirm the coexistence of both cascades the energy flux ΠE and the
0
10
−5
10
−10
10
Total energy E
Filtered energy Es
Filtered energy Ef
Slope k−2
Slope k−5.5
−15
10
0 1 2
10 10 10
Fig. 3. Original and filtered (with 5 scales packets) energy spectra (kinj ≈ 20).
enstrophy flux ΠZ are computed. We recall that the energy flux is computed
from the non linear term in the Navier-Stokes equation written in Fourier
space:
Role of coherent structures in 2D turbulence 413
Z +∞
ΠE (k) = TE (k 0 )dk 0 (2)
k
0.2
!E
!s " s
E
0.1 !f " f
E
−0.1
−0.2
−0.3
−0.4
−0.5
0 1 2
10 10 10
Fig. 4. Energy fluxes of the whole flow and of the separate structures (kinj ≈ 20).
side of the injection scale and thus an inverse energy cascade stands from this
injection scale to larger scales and the flux vanishes for larger scales. A main
positive part of the enstrophy flux is obtained on the right hand side of the
injection scale and thus a direct enstrophy cascade stands from this scale to the
smaller scales. Let us note also a negative part of the flux before the injection
scale showing that some vorticity is brought by the inverse energy cascade.
These fluxes confirm what can be seen on the spectra. The whole spectrum
is given before the injection scale by the solid rotation field and beyond the
injection scale by the filaments field, the other field being negligible in both
cases.
414 Ch.-H. Bruneau, P. Fischer and H. Kellay
1000
!
Z
!s " s
Z
f"f
!Z
500
−500
0 1 2
10 10 10
Fig. 5. Enstrophy fluxes of the whole flow and of the separate structures (kinj ≈ 20).
3 Conclusions
References
1. Angot, Ph., Bruneau, C.H., Fabrie, P.: A penalization method to take
into account obstacles in an incompressible flow. Num. Math., 81, 497–
520 (1999)
2. Bruneau, C.H.: Boundary conditions on artificial frontiers for incompress-
ible and compressible Navier-Stokes equations. M2AN, 34, 303–314 (2000)
3. Bruneau, C.H., Kellay, H.: Coexistence of two inertial ranges in two-
dimensional turbulence. Phys. Rev. E, 71, 046305 (2005)
4. Bruneau, C.H., Saad, M.: The 2D lid-driven cavity problem revisited.
Computers & Fluids, 35, 326–348 (2006)
5. Farge M., Kevlahan N.: Vorticity filaments in two-dimensional turbulence:
creation, stability and effect. J. Fluid Mech., 346, 49-76 (1997)
6. Fischer, P.: Multiresolution analysis for two-dimensional turbulence. Part
1: Wavelets vs Cosine packets, a comparative study. Discr. Cont. Dynam-
ical Syst. B, 5, 659–686 (2005)
Temporal Evolution of Dominant Flow
Structures in Turbulent Channel Flow
1 Introduction
A valuable approach for the numerical investigation of turbulent flows is the
Direct Numerical Simulation (DNS ) according to which the system of the
Navier-Stokes equations is integrated numerically with no modifications of any
kind. The critical aspect in following this approach is the accuracy of the calcu-
lations, that has to be sufficiently high to resolve the essential turbulent scales.
In turbulent shear flows a case that has been studied by some authors is that
of the plane channel. DNS calculations of turbulent channel flow have been
carried out by Kim et al. [1], Lyons et al. [2], Kasagi et al. [3], Rutledge and
Sleicher [4], Moser et al. [5]. In these works the Navier-Stokes equations in ro-
tational form are integrated by means of the fully spectral Fourier-Chebychev
numerical technique introduced in [1]. In the present work a mixed spectral-
finite difference technique for the numerical integration of the Navier-Stokes
equations is introduced and the issue is addressed of the coherent structures
of turbulence in the wall region of turbulent channel flow. The calculations
are performed at nominal friction Reynolds number Reτ = 180. The coherent
turbulent structures are educed with the technique of the Proper Orthogonal
Decomposition (POD) and the evolution of the dominant flow structures near
the wall is followed in time.
∂ui
=0 (1)
∂xi
∂ui ∂ ∂p 1 ∂ 2 ui
+ (ui uj ) = − + (2)
∂t ∂xj ∂xi Reτ ∂xj ∂xj
where ui (u, v, w) are the velocity components in the cartesian coordinate
system xi (x, y, z). Variables and operators are nondimensionalized√ by the
channel half-height δ for lengths, wall shear velocity uτ = τw /ρ for ve-
locities, ρu2τ for pressure and δ/uτ for time, being Reτ = uτ δ/ν the friction
Reynolds number, ρ the fluid density and ν the fluid kinematic viscosity. The
fields are admitted to be periodic in the streamwise (x ) and spanwise (z )
directions and equations (1-2) are Fourier transformed accordingly. The non-
linear terms in the momentum equation are evaluated pseudospectrally, by
anti-transforming the velocities back to physical space to perform the prod-
ucts (FFTs are used). A dealiasing procedure is applied to avoid errors in
transforming the results back to Fourier space. In order to have a better spa-
tial resolution near the walls, a grid-stretching law of hyperbolic-tangent type
is introduced for the grid points along y, the direction orthogonal to the solid
walls. For time advancement a third-order Runge-Kutta algorithm is imple-
mented and time marching is accomplished with the fractional-step method
([1]).
By recalling the wall formalism, one has: x+ + 2
i = xi uτ /ν = xi /δτ , t = tuτ /ν =
+ + +
tuτ /δτ , δ = δ/δτ , u = u/uτ , Reτ = uτ δ/ν = δ/δτ = δ , where u de-
notes a mean x -velocity averaged on a x-z plane and time, δτ = ν/uτ is
the viscous length and δ/uτ the viscous time unit. The characteristic param-
eters of the simulations are the following. Computing domain: Lx = 2πδ,
Ly = 2δ, Lz = πδ; L+ + +
x = 1131, Ly = 360, Lz = 565. Computational grid:
+
Nx = 96, Ny = 129, Nz = 64. Grid spacing: ∆x+ = 11.8, ∆ycenter = 4.4,
+
∆ywall = 0.87, ∆z + = 8.8. It can be verified that there are 6 grid points in
the y direction within the viscous sublayer (y + ≤ 5). The Kolmogorov spa-
tial microscale, estimated using the criterion of the average dissipation rate
per unit mass across the width of the channel, results η + ≈ 1.8. After the
Evolution of Dominant Structures in Turbulent Channel Flow 417
domain D. In the present work the POD is used for the analysis of the fluctuat-
ing portion of the velocity field. The two homogeneous directions are handled
in Fourier space, so that the optimal representation of the velocity field in the
statistical sense outlined above is sought in the direction normal to the walls.
3 Results
Fig. 3. Surfaces of constant streamwise fluctuation reconstructed from the first five
most energetic POD modes at t+ = 136.8, t+ = 165.6, t+ = 172.8 and t+ = 190.8
(left to right)
420 Giancarlo Alfonsi and Leonardo Primavera
the detachment of the extended surface at one end of the negative streamwise-
elongated structure occurs and, at the other end, the first manifestation of the
structure’s break-up becomes evident. At t+ = 172.8 and t+ = 190.8 the pro-
cess of disruption of the originary structure becomes more evident.
4 Concluding Remarks
The analysis of the first 5 most energetic eigenfunctions of the decomposition
has shown the existence of two kinds of different structures, i.e. elongated
streamwise-independent modes aligned in the x direction and bean-shaped
streamwise-dependent modes. The analysis of the evolution in time of the
flow field produced by the sum of the first 5 most energetic POD modes
has revealed the existence of dominant structures elongated in the stream-
wise direction whose shape is altered due to the process of interaction with
the bean-shaped, quasi-streamwise flow modes. This process of interaction is
responsible for the gradual break-up of the x -elongated structures.
References
1. Kim, J., Moin, P., Moser, R.D.: Turbulence statistics in fully developed channel
flow at low Reynolds number. J. Fluid Mech., 177, 133–166 (1987)
2. Lyons, S.L., Hanratty, T.J., McLaughlin, J.B.:Large-scale computer simulation
of fully developed turbulent channel flow with heat transfer. Int. J. Numer.
Meth. Fluids, 13, 999–1028 (1991)
3. Kasagi, N., Tomita, Y., Kuroda, A.: Direct numerical simulation of passive
scalar field in a turbulent channel flow. ASME J. Heat Transf., 114, 598–606
(1992)
4. Rutledge, J., Sleicher, C.A.: Direct simulation of turbulent flow and heat trans-
fer in a channel. Part I: smooth walls. Int. J. Numer. Meth. Fluids, 16, 1051–
1078 (1993)
5. Moser, R.D., Kim, J., Mansour, N.N.: Direct numerical simulation of turbulent
channel flow up to Re = 590. Phys. Fluids, 11, 943–945 (1999)
6. Dean, R.B.: Reynolds number dependence of skin friction and other bulk flow
variables in two-dimensional rectangular duct flow. ASME J. Fluids Eng., 100,
215–223 (1978)
7. Lumley, J.L.: Stochastic tools in turbulence. Academic Press, (1971)
8. Sirovich, L.: Turbulence and the dynamics of coherent structures. Part I: co-
herent structures. Part II: symmetries and transformations. Part III: dynamics
and scaling. Quart. Appl. Math., 45, 561–590 (1987)
9. Berkooz, G., Holmes, P., Lumley, J.L.: The Proper Orthogonal Decomposition
in the analysis of turbulent flows. Ann. Rev. Fluid. Mech., 25, 539–575 (1993)
LES of passive scalar in compressible mixing
layers
C. Le Ribault
1 Introduction
ibility of a mixing layer. Mixing layers at the convective Mach numbers of 0.16,
0.5 and 1.1 are simulated. At the first convective Mach number, Mc = 0.16,
the compressibility effects are negligible. The convective Mach numbers of 0.5
has been chosen to correspond to the experiments of Elliot and Samimy [2].
At this convective Mach number, the compressibility effetcs are moderate. At
the convective Mach number of 1.1, important compressibility effects are ex-
pected. The Reynolds number, based on the inflow momentum thickness δθ0
and the velocity difference, is equal to Re = ρδθ0 ∆U/µ = 300. The grid sizes
are also non-dimensionalized by the inflow momentum thickness.
Table 1 summarizes the dimensions of the computational domains and the
number of grid points for the different cases.
Only one filter size ∆ = 2h, where h is the local grid spacing, is used.
4 Momentum mixing
An experimental curve, called “Langley curve” is obtained from the compi-
lation of results corresponding to experiments with air-air shear layers. This
LES of passive scalar in compressible mixing layers 423
curve gives the shear momentum thickness non dimensionalized by its incom-
pressible counterpart in function of the convective Mach number. Shear layer
thickness growth rates predicted by the Langley curve has been plotted on
the figure 1 along with LES results. The spreading rate of the mixing layers
decreases with the convective Mach number in agreement with experimental
results.
1.0
0.8
(d!"/dx)/(d!"/dx)inc
0.6
0.4
Langley experimental curve
LES Smag. dyn.
LES dyn. mixed
0.2
0.0 0.5 1.0 1.5
Mc
The figure 2 presents the dependence of the peaks streamwise and trans-
verse velocities on Mc . The results of the two turbulence models are com-
pared to experimental results. The difference between the experimental re-
sults are very important on those quantities. All the Reynolds stress compo-
nents slightly decrease with the convective Mach number. The dynamic mixed
model predicts values slightly closer to experimental results than the dynamic
Smagorinsky model but the difference between the two models predictions is
weak.
5 Scalar mixing
The figure 3 compares the similarity profiles of passive scalar turbulence in-
tensities obtained by the two subgrid models at the convective Mach numbers
of 0.16 and 1.1. The self-similarity shape is predicted by the two models. The
difference between the prediction of the two models is small. The peak is
smaller at the convective Mach number of 1.1.
The peak values of the rms of the scalar predicted by the two subgrid
models are plotted together with the experimental results of Clemens [1] in
424 C. Le Ribault
0.20 0.20
rms(u)/Δu
rms(v)/!u
0.10 Goebels & Dutton (1991) 0.10
0.00 0.00
0.00 0.50 1.00 1.50 0.0 0.5 1.0 1.5
Mc Mc
Fig. 2. Dependence of peak rms streamwise (left) and vertical (right) velocity
0.3 0.3
0.2 0.2
rms(")
rms(")
0.1 0.1
0 0
−1.5 −0.5 0.5 1.5 −1.5 −0.5 0.5 1.5
y/! y/!
Fig. 3. Comparison of the RMS of the passive scalar obtained with the two subgrid
models Mc = 0.16 (left) and Mc = 1.1 (right)
function of the convective Mach number (figure 4). The peak value decreases
with increasing compressibility, even at the convective Mach number of 0.5. It
is in agreement with the DNS results of Freund [2] for the annular mixing layer.
Clemens [1] experimental scalar fluctuation statistics show that fluctuations
are smaller at Mc = 0.82 than at Mc = 0.35.
0.25
0.20
0.15
rms(!)
0.10
Dyn. Smag
0.05 dyn. mixed
Clemens and Paul (1995)
0.00
0.00 0.50 1.00 1.50
Mc
4
η=0.29
3.5 "=0.29 η=0.18
"=0.18 4 η=0
3 "=0.0 η=−0.16
η=−0.29
"=#0.16
2.5 "=#0.29
rms(ξ)
PDF(!)
2
1.5
0.5
0 0
0 0.5 1 0 0.5 1
! ξ
Figure 6 shows the PDF of the mixture fraction obtained with the two
subgrid models for Mc = 0.16 and for Mc = 1.1 at one section η = 0. There
is a weak trend of narrower and taller PDFs with increasing convective Mach
number. This is consistent with smaller RMS fluctuations. Large eddies will
broaden the PDFs since fluid can be transported from regions far from the
measurement location. On the contrary, a layer that is dominated by small
eddies will produce concentration PDFs that are more narrow. This argu-
ment suggests that scalar transport in compressible layer may occur through
smaller-scale eddies.
7 Isovalues
PDF(ξ) 2 2
PDF(ξ)
1 1
0 0
0 0.5 1 0 0.5 1
ξ ξ
8 Conclusion
LES of spatially developing mixing layers have been performed at three dif-
ferent convective Mach numbers. Two-subgrid models are compared: the dy-
namic Smagorinsky model and the dynamic mixed model. The principal com-
pressiblity effects such as the decrease of the RMS statistics and qualitative
changes in the PDFs and the structures shape are predicted by both models.
References
1. Clemens N.T., Paul P.H.: Scalar measurements in compressible mixing layers,
Physics of Fluids, 7, no 5 (1995)
2. Freund J.B., Moin P., Lele S.: Compressibility effects in a turbulent annular
mixing layer. Part 2. Mixing of a passive scalar , 421 (2000)
3. Le Ribault C.:Large eddy simulation of compressible mixing layers, International
Journal of Dynamics of Fluid , 1, no 1 (2005)
4. Le Ribault C., Sarkar S., Stanley S.: Large-eddy simulation of evolution of a
passive scalar in a plane jet, AIAA Jounal, 39, no 8, (2001)
5. Sankaran V., Menon S.: LES of scalar mixing in supersonic mixing layers , Pro-
ceedings of the combustion Institute, 30, (2005)
Buoyancy Effect on Turbulence using Blocks
Summary. Buoyancy effect on turbulence was simulated by LES and K- models
using blocks as computational elements. Three values of the Smagorinsky model
coefficient were used in the LES: Cs = 0.0375, 0.075 and 0.15. The fourfold varia-
tion in the coefficient (sixteenfold in subgrid viscosity) had produced essentially the
same results. The insensitivity of the LES results to model coefficient were in sharp
contrast to the simulations of the same flow using the K- model. The buoyancy-
extended coefficient c3 in the K- model was not a constant but dependent on
buoyancy stratification. The optimal values of the coefficient were c3 = 1.0 for
stably stratified flow and c3 ' 0.34 ∼ 0.45 for unstably stratified flow. A slight
modification to the K- model coefficient can led to significantly different prediction
by the model.
1 Introduction
Controlling unphysical oscillations is a problem that has initiated the de-
velopment of many computational strategies (Van Leer [7]; Gaskell and Lau
[4]; Leonard and Mokhtari [9]; Leonard [8]) However, few implementations of
these strategies were carried out for realistic flow computations. In this paper,
the LBM (Lagrangian block method) of Chu and Altai [1] was employed to
simulate the buoyancy effect on turbulence for two flows with opposite buoy-
ancy stratification. In one flow, turbulence was produced by Kelvin-Helmholtz
(KH) instabilities initiated by shear across a stably stratified interface (Figure
1a). In the other case, turbulence was produced by the Rayleigh-Taylor (RT)
instabilities due to unstable stratification (Figure 1b). In the case of stably
stratified flow, the mixing across the interface is bounded by the stablizing
effect of the buoyancy force. In unstably stratified flow, the mixing across the
unstable interface is unbounded as the turbulent motion is reinforced by the
positive work of the buoyancy force The LBM is employed for the sinulations
because the method is free of unphysical oscillations and relatively not diffu-
sive. It is a numerical scheme that can resolve accurately the sharp gradients
across laminar and turbulent interface [2] (see also Figure 2).
The LES (Large Eddy Simulation) of the turbulent flows were conducted using
following sub-grid scale viscosity of the Smagorinsky model:
∂ ũ 2 ∂ṽ ∂ṽ ∂ ũ 2
vSG = Cs2 ∆2 [2( ) + 2( )2 + ( + ) ] (1)
∂x ∂y ∂x ∂y
in which (ũ, ṽ) are filtered components of simulated velocity, and ∆ = [(∆x)2 +
1
(∆y)2 ] 2 is the mesh size of the computation grid. A Smagorinsky coefficient
of Cs = 0.075 was selected for the simulations shown in Figure 1. The results
obtained for the stably stratified flow have produced a Richardson number for
the asymptotic state in close agreement with the laboratory observation by
Thorpe [10]. The simulations of the RT instabilities for the unstably stratified
flow produced the following formulae for the depth of the mixing layer:
σ t2 σ t
= 0.015 2 if t < 6 ts ; = 0.090 if t > 6 ts (2)
h1 ts h1 ts
which agree with the experimental observation by Li and Chu [6]. In these
formulae for the RT instabilities, the lengthp scale and time scale are: h1 =
initial depth of the upper dense layer, ts = h1 /g 0 = time scale, and g 0 = re-
duced gravity. The depth of the mixing region σ is determined by the following
integrals of the buoyancy profile G(z):
R∞ R∞
2 (z − z)2 (dG/dz)dz
o R
z(dG/dz)dz
σ = ∞ , z = Ro ∞ (3)
o
(dG/dz)dz o
(dG/dz)dz
Buoyancy Effect on Turbulence using Blocks 429
Fig. 2. One and four rotations of an 8x8 block in a velocity field produced by solid
body rotation. The numerical diffusion associated with LBM is almost negligible.
Fig. 4. Depth of the mixing region in the unstably stratified flow produced by the
RT instability. The dash line delineates the results obtained by LES and laboratory
experiments of Li and Chu [6]. The thin solid line delineates the results obtained by
K- model using the model coefficient c3 = 0.4, while the thick solid line the results
obtained using c3 = 0.445 if t < 4.3 ts and c3 = 0.34 if t > 4.3 ts
Buoyancy Effect on Turbulence using Blocks 431
1.6
1.2
0.8
0.4
0.0
0 4 8 12 16
flow was more difficulty because the simulation results were sensitive to the
values of the model coefficient c3 . Signficantly different results were obtained
if the value of c3 were only slightly deviated from the optimal. To fit the
simulation with the experimental data, the optimal values are c3 = 0.445 if
t < 4.3 ts and c3 = 0.34 if t > 4.3 ts (see Figure 4).
4 Conclusion
The LES of the turbulent flow was successfully conducted using one Smagorin-
sky model coefficient, Cs = 0.075, for the entire range of Richardson number.
However, the K- model simulations were not as successful. The buoyancy-
extended coefficients in the K- model was not a constant but dependent
on buoyancy stratification. To reproduce the experimental observation, the
buoyancy-extend coefficient was taken to be c3 = 1.0 for stably stratified
flows, and c3 ' 0.34 to 0.455 for unstably stratified flows. The K- model
results are sensitive to slight changes in the values of the model coefficients.
In contrast, the LES were not as sensitive to the model coefficient. As shown
in Figure 5, with the fourfold variation in the Smagorinsky model coefficient,
Cs = 0.0375, 0.075 and 0.15, the LES produced essentially the same results.
References
1. V. H. Chu and W. Altai. Simulation of shallow transverse shear flow by gener-
alized second moment method. J. Hydraulic Research, 39(6):575–582, 2001.
432 Vincent H. Chu, Wihel Altai, and Camilo E. Pinilla
1 Introduction
2 Numerical aspects
We advance the incompressible Navier-Stokes equations in time in a 2nd order
accurate manner, by setting the following continuity and momentum residuals
to zero at each new time step n + 1:
RCn+1 = ∇· (u n+1 ) = 0, (1)
n+1 1
RM = (u n+1 − u n )/∆t + ∇p n+1/2 − ∇2 (u n+1 + u n ) (2)
2Re
+ (3/2)∇· (u n u n ) − (1/2)∇· (u n−1 u n−1 ) = 0. (3)
Herein, the pressure and viscous terms are treated in an implicit manner
using the Crank-Nicolson scheme, whereas the convective terms are treated
explicitly using the 2nd order Adams-Bashforth method.
2.1 Spatial representation
Both azimuthal and axial directions are periodic for the considered geometry.
It is then natural to represent any flow unknown q (where q represents either
the pressure p or any component of the velocity vector u) by means of a trun-
cated Fourier series in at least one of these directions. In cross planes formed
by the two remaining directions, we introduce a finite element representation
on an unstructured triangular mesh (P1 elements). In the present study, we
have performed all simulations using two different hybrid spectral/FE ap-
proaches:
Planar model: We use a cartesian coordinate system (x, y, z) and discretize
in the axial direction using a spectral expansion:
Nz /2 NNod
1 X X
q n (x = rcosθ, y = rsinθ, z) = Qnk,j Nj (x, y)eIk2πz/Lz , (4)
Nz
k=−Nz /2 j=1
The discrete equations for the unknowns Qnk,i are obtained via the following
weighted residual formulation:
Z Z
φk Ni RCn+1 dV + τpspg φk ∇N˜ i · R n+1 dV = 0, (6)
M
Z
n+1
φk Ni RM dV = 0. (7)
It can be shown that the above approach allows to decouple the 3D linear
problem (Eq. 3) to be solved at each time step into a series of decoupled
2D linear problems, one for each Fourier mode. Coupling still occurs through
the (nonlinear) convective terms, which are however treated explicitly and
therefore do not complicate the (costly) linear solution phase. For the complete
technical details of the discretization procedure outlined above, the reader is
referred to [1].
The discretized equations can be easily computed directly in terms of
the Fourier components, except for the convective terms, which couple each
Fourier mode to all others. The convective terms are therefore computed in
physical space and then transformed to Fourier space using the Fast Fourier
Transform (FFT) algorithm. The basic idea behind this pseudo-spectral solu-
tion procedure is as follows :
(1) Evaluate the convective terms in physical space, using the previously cal-
culated u n and u n−1 ;
(2) transform these terms to Fourier space using the FFT algorithm;
(3) set up the linear problems corresponding to each Fourier mode;
(4) apply appropriate boundary conditions for each Fourier mode;
(5) solve the linear systems to obtain the updated solution unknowns Qn+1 k,j
in Fourier space;
(6) apply the inverse Fourier transform to obtain the updated unknowns q n+1
in physical space.
436 Detandt et al.
3 Results
We use the same geometry for each computation (η = rroi = 0.875; roλ−r z
i
=
2.5). While our code is able to use fully unstructured meshes, we have choose
to use a structured mesh for this simple geometry (grid spacing is constant
in each direction). For Reynolds number 140, we use a 323 mesh (Nz = Nr =
Nθ = 32) for both versions, while for the turbulent test case (Re=3000), we
use 643 mesh in order to resolve all the turbulent length scales. Five horizontal
ribs have also been placed on the inner cylinder to evaluate the influence of
these on the mixing properties around the inner cylinder. The rib size is 1/15th
of the gap and corresponds to the experiments described in [2].
For the laminar test case at Re=140, the following integral flow quantities
Z 2π Z h Z ro
1
E= |u|2 r dr dz dθ, (9)
2hs2 (ωri )2 0 0 ri
Z 2π Z h Z ro
1
L= uθ r2 dr dz dθ, (10)
2hs3 ωri 0 0 ri
Z 2π Z h Z ro
1
E= |∇ × u|2 r dr dz dθ, (11)
2h(ωri )2 0 0 ri
are compared to literature results by Marcus [6] and Manna et al. [4] to assess
the quality of our results.
For the turbulent simulations at Re=3000, the Kolmogorov length has also
been computed to assess whether the mesh used is sufficiently fine or not :
s
ν
lKolm = p . (12)
2Sij Sij
∂ui ∂uj
Herein Sij is the strain tensor defined as Sij = 21 ∂x j
+ ∂xi
. Defining4 a
cell size as the radial distance between mesh nodes, a simulation is resolved
if lCell ≤ lKolm . In the present case, the Kolmogorov length scale is averaged
over two directions of periodicity (< lKolm >2 = 1 R 2π R Lνz √ ).
2πh 0 0
2Sij Sij
3.1 Reynolds 140
This low value of the Reynolds number, for which the flow is steady and
axisymmetric, is used to assess the performance of both versions of the solver.
The integral flow quantities are compared to numerical results presented in
Marcus [5] and in Manna et al. [4].
The Kolmogorov length scale is 3 times greater than the cell size which
indicates that our simulations are resolved on this mesh.
4
Different definition of a cell size lCell are acceptable. This explains why the cri-
terium consider acceptable ratio llKolm
Cell
should be of the order of 1 and not exactly.
DNS of Taylor-Couette flows 437
(a) Planar version (b) Cylindrical version (c) Cylindrical version (with ribs)
Fig. 1. Vz instantaneous contours and streamstraces at Re=3000
4 Conclusion
In the present paper we have demonstrated the performance of two different
combinations of Spectral and Finite Elements discretisations of Navier-Stokes
438 Detandt et al.
equations. Good agreement was found between both versions and with liter-
ature.
The main objective of the present paper was to investigate the influence
of putting riblets on the inner cylinder. The next step in this investigation
will be to test the influence of adding vertical riblets. Further analysis on the
transport of a passive scalar is needed to verify the most effective configuration
of riblets for electro-chemical applications.
Acknowledgments
The research was supported by the Institute for the Promotion of Innovation
by Science and Technology in Flanders (IWT, MuTEch project, SBO contract
040092) at VKI and KUL, and by the Higher Education and Scientific Re-
search division of the French Community Government at ULB (ARC project
02/07-283). Yves Detandt was supported by a FRIA fellowship. Computa-
tions are performed on the linux cluster ANIC4 financed by several contracts
(see http://anic4.ulb.ac.be).
References
1. David Vanden Abeele, Gérard Degrez, and Snyder Deryl Owen. A combined
spectral/finite elements method for the direct and large eddy simulation of tur-
bulent flows in complex, two dimensional geometries. In Proceeding CMFF 03,
pages 783–790, 2003.
2. O. Cadot, Y. Couder, A. Daerr, S. Douady, and A. Tsinober. Energy injection in
closed turbulent flows : Stirring through boundary layers versus inertial stirring.
Physical Review E, 56(1), 1997.
3. Vanden Abeele David, Degrez Gérard, and Snyder Deryl Owen. Parallel turbu-
lent flow computations using a hybrid spectral/finite element method on beowulf
clusters. In Proceeding ICCFD 3, 2004.
4. M. Manna and A. Vacca. An efficient method for the solution of the incompress-
ible Navier-Stokes equations in cylindirical geometries. Journal of Computational
Physics, 151:563–584, 1999.
5. Marcus P.S. Simulation of Taylor-Couette flow. Part 1. Numerical methods and
comparison with experiment. Jounral of Fluid Mechanics, 146:45–64, 1984.
6. Marcus P.S. Simulation of Taylor-Couette flow. Part 2. Numerical results for
wavy-vortex flow with one travelling wave. Journal of Fluid Mechanics, 146:65–
113, 1984.
Flux Limiting Schemes for Implicit Large Eddy
Simulation of Synthetic Jets
Summary. The evolution of synthetic jets into quiescent air has been simulated
by performing three-dimensional numerical investigations using various flux limiting
high-resolution schemes in the context of Implicit Large Eddy Simulation (ILES).
The results compare well with existing experimental data from NASA Langley Re-
search Center.
1 Introduction
A synthetic jet refers to a flow created with zero net mass flux. The synthetic
jet is generated using a device consisting of an oscillating diaphragm, where
fluid is alternately pushed in and out of an orifice. The exiting fluid separates
at the edges of the orifice and rolls up to form a vortex ring which propagates
away from the orifice due to its own self induced velocity. As the diaphragm
moves away from the orifice fluid is drawn back into the the cavity which
results in an overall zero net mass flux across the orifice exit. Hence, synthetic
jets are synthesized from the working fluid and do not require complex pip-
ing in order to generate the flow. We examine the evolution of synthetic jets
into quiescent air by performing three-dimensional numerical investigations
using various flux limiting high-resolution schemes for solving the compress-
ible Navier-Stokes equations in the context of Implicit Large Eddy Simulation
(ILES) [1]. Various limiting approaches have been considered including Min-
bee; Superbee; van Albada [2] and the limiter developed by Drikakis [3]. The
computational results have been validated against experimental data obtained
from the NASA Langley workshop on CFD Validation of Synthetic Jets and
Turbulent Separation Control. Results of phase-averaged velocity fields close
to the orifice have shown a good comparison between the computational and
experimental results.
2 Numerical Method
A three-dimensional parallel compressible Navier-Stokes flow solver has been
used for the simulations. This section will outline the various limiting ap-
proaches considered in the context of high-resolution methods. High-resolution
methods are nonlinear, nonoscillatory methods which use the local solution
to select a technique for approximating the solution. This suggests that high-
resolution methods tend to adapt themselves to the particular circumstances
so that the solution obtained is accurate and has some physical meaning. It has
been shown by Drikakis [1] and Margolin and Rider [4] that high-resolution
methods can be used in (under-resolved with respect to grid resolution) turbu-
lent flow computations without the need to resort to a turbulence model. Lim-
iters are the general nonlinear mechanism that distinguishes modern methods
from classical linear schemes. Their role is to act as a nonlinear switch between
more than one underlying linear method thus adapting the choice of numeri-
cal method based upon the behaviour of the local solution. Limiters result in
nonlinear methods even for linear equations in order to achieve second-order
accuracy simultaneously with monotonicity. Limiters can act like dynamic,
self-adjusting models, modifying the numerical viscosity to produce a nonlin-
ear eddy viscosity [1, 4].
The high-resolution scheme employed here is the Godunov-type, characteristics-
based scheme by Eberle [5]. It is a Godunov-type method that defines the con-
servative variables along the characteristics as function of their characteristic
values. The scheme will not be presented here in detail and the author would
refer the reader to the paper by Eberle [5] for further details. The cell centered
data is interpolated to the cell faces using a MUSCL-type high-order interpo-
lation which incorporates the various flux limiters studied in this paper. This
method can be expressed as;
L 1
Ui+1/2 = Ui + [(1 − k)(Ui − Ui−1 ) + (1 + k)(Ui+1 − Ui )]. (1)
4
R 1
Ui+1/2 = Ui+1 − [(1 − k)(Ui+2 − Ui+1 ) + (1 + k)(Ui+1 − Ui )]. (2)
4
This form of the extrapolation is symmetric about the interface i + 1/2, and k
is a free parameter between -1 and 1. The MUSCL scheme as presented above
is not stable on it’s own and hence a limiting method must be incorporated.
Limiters are generally defined using the parameter r where;
Ui+1 − Ui Ui+1 − Ui
rL = , rR = . (3)
Ui − Ui−1 Ui+2 − Ui
A limiting function φ(r) is defined giving a symmetric limited MUSCL scheme
in the form;
L 1 L 1
Ui+1/2 = Ui + [(1 − k)φ(r )(Ui − Ui−1 ) + (1 + k)φ (Ui+1 − Ui )].(4)
4 rL
Flux Limiting Schemes for ILES of Synthetic Jets 441
R 1 1
Ui+1/2 = Ui+1 − [(1−k)φ(rR )(Ui+2 −Ui+1 )+(1+k)φ (Ui+1 −Ui )].(5)
4 rR
The four limiters investigated in this paper are the van Albada (VA), Minbee
(MB), Superbee (SB) and the limiter developed by Drikakis (DD) [3]. They
can be expressed as;
r(1 + r)
φV A = (6)
1 + r2
φM B = min(1, r) (7)
φSB = max(0, min(1, 2r), min(r, 2)) (8)
N
2N r 2r
φDD = 1 − 1 + 1− (9)
1 + r2 1 + r2
In all the above cases we have used k = 1/3 which gives third order accuracy
for the MUSCL reconstruction.
the flow having to travel through the cavity which has not been modelled in
the computational case. A phase lag in the computations can be seen on this
figure with the experimental results reaching a peak velocity earlier than the
computations. At a distance of 4mm from the orifice exit the peak velocity
in the computations compares well with the hot wire data but the PIV mea-
surements show a large increase in velocity. The various limiters compare well
especially in the suction phase, with Minbee, van Albada and the DD limiter
giving the best results in comparison to the PIV data. Again, the phase lag
in the computational results is especially noticeable in the expulsion phase.
Phase averaged results of u- and v-velocity profiles along the horizontal line
2mm above the orifice exit at phase angles of 90◦ and 270◦ are shown in figure
30 PIV data
HW data
LDV data 30
DD limiter
20 van Albada limiter DD limiter
Superbee limiter van Albada limiter
25
Minbee limiter Superbee limiter
Minbee limiter
10 PIV data
Velocity m/s
Velocity m/s
20
HW data
0 15
10
-10
5
-20
0
0 100 200 300 0 50 100 150 200 250 300 350
Phase deg Phase deg
Fig. 2. Phase-averaged u-velocity at 0.1mm (a) and 4mm (b) above the center of
the orifice exit.
Flux Limiting Schemes for ILES of Synthetic Jets 443
6
DD limiter
van Albada limiter DD limiter
25 van Albada limiter
Superbee limiter 4
Minbee limiter Superbee limiter
PIV data Minbee limiter
20 PIV data
2
Velocity m/s
Velocity m/s
15
0
10
-2
5
-4
0
-6
-2 -1 0 1 2 3 4 -2 -1 0 1 2 3 4
Position across slot mm Position across slot mm
DD limiter
-1 van Albada limiter
Superbee limiter 3
Minbee limiter
-1.5 PIV data DD limiter
2 van Albada limiter
Superbee limiter
-2 Minbee limiter
1
Velocity m/s
PIV data
Velocity m/s
-2.5
0
-3
-1
-3.5 -2
-4 -3
-2 -1 0 1 2 3 4 -2 -1 0 1 2 3 4
Position across slot mm Position across slot mm
3. The various limiters tend to over predict the peak u-velocity at phase of
90◦ with Minbee and van Albada giving the best results away from the slot
edges. In contrast Minbee, van Albada and the DD limiter under predict the
maximum and minimum velocities in the cross-stream direction whereas Su-
perbee gives an extremely close match especially in the region of the orifice
edges. At maximum suction (270◦ ) all four limiters give different results in
comparison with the PIV u-velocity data. All limiters over predict the mini-
mum velocity with van Albada and the DD limiter showing the best results
with respect to the velocity profile shape. The cross streamwise profile shows
a steep gradient across the orifice exit in the computational results whereas
the PIV shows a more smoother transition across the orifice. A comparison
of phase averaged centerline streamwise velocity at maximum expulsion and
suction are shown in figure 4. Both plots show that the velocity at the orifice
exit is over predicted by all the different limiters. No limiter manages to cap-
ture the position and magnitude of the maximum expulsion velocity. The van
444 Sanjay Patel and Dimitris Drikakis
DD limiter
van Albada limiter
30 Superbee limiter 10
Minbee limiter
PIV data
25
Velocity m/s
0
Velocity m/s
20
DD limiter
15 -10 van Albada limiter
Superbee limiter
Minbee limiter
10 PIV data
-20
0 2 4 6 8 0 2 4 6 8
Streamwise position mm Streamwise position
◦ ◦
(a) Phase 90 (b) Phase 270
4 Conclusions
The evolution of a three-dimensional synthetic jet in quiescent air has been
simulated using various limiting schemes. The results obtained have compared
relatively well with the experimental data from NASA Langley. It was found
that the various limiters investigated differed with regards to the amount
of dissipation provided by the scheme. All limiting schemes gave reasonable
results close to the orifice exit but further downstream the difference in the
schemes was much higher, with the van Albada limiter and the limiter of
DD performing best. The study of various limiters is complex with limiters
behaving differently depending on the characteristics of the flow situation. It is
clear from these results that the cavity section with the oscillating diaphragm
needs to be modelled in the computations. This should provide a much more
realistic velocity profile at the orifice exit and hence give more accurate results.
References
1. Drikakis, D: Advances in turbulent flow computations using high-resolution
methods. Progress in Aerospace Sciences. 39, 405–424 (2003).
2. Toro, E: Riemann solvers and numerical methods for fluid dynamics. Springer-
Verlag (1997)
3. Zóltak, J. and Drikakis, D: Hybrid upwind methods for the simulation of un-
steady shock-wave diffraction over a cylinder. Computer Methods in Applied
Mechanics and Engineering. 162, 165–185 (1998).
4. Margolin, L.G. and Rider, W.J: A rationale for implicit turbulence modeling.
International Journal for Numerical Methods in Fluids. 39, 821–841 (2002).
5. Eberle, A: Characteristic flux averaging approach to the solution of Euler’s equa-
tions. Computational Fluid Dynamics, VKI Lecture Series. (1987).
6. Yao, C., Chen, F.J., Neuhart, D. and Harris, J: Synthetic jet flow field database
for CFD validation. 2nd AIAA Flow Control Conference. 2218, (2004).
Implicit Large Eddy Simulation of a Flow
around a Cylindrical Body
Summary. Numerical simulation was carried out for a flow around a cylindrical
body. The incompressible Navier-Stokes equations were solved by the multidirec-
tional finite-difference method. For high-Reynolds-number flows, a third-order up-
wind scheme was utilized for the convective terms to stabilize the computation. As
a result of the computation, drag variation caused by different section shape of body
was captured well.
1 Introduction
Many simulations of a flow past a streamlined body have been performed.
But those of bluff body have not been done much. To resolve the boundary
layer, we use finite difference approach. Flow around a car is one of the most
important applications. But flow around a car is very complicated, grid gen-
eration is not easy. To understand the most important physical features of
these flows, we compute a flow around a cylindrical body.
2 Computational Method
The governing equations are the 3-d incompressible Navier-Stokes equations
and the equation of continuity as follows:
divu = 0, (1)
∂u 1
+ u · gradu = −gradp + 4u. (2)
∂t Re
where u, p, t and Re denote the velocity vector, pressure, time and the
Reynolds number respectively. For high-Reynolds-number flows, time-dependent
computations are required owing to the strong unsteadiness.
P2 P2 P5
Px2 Px1
Px2 P2 Px1
P3 P0 P1 P3 P0 P1
P3 P0
h P1
P4 Px3 P4 Px4
Px3 P4 Px4
P6
r
d
has been found to be the most suitable for high-Reynolds number flow com-
putations. The second-order Crank-Nicolson implicit scheme is used for time
integration. The equations are iteratively solved at each time step by SOR
method. For Poisson’s equation, a multi-grid method is utilized.
In the computations, the flow around a square cylinder with rounded cor-
ners is solved. For a section of the cylinder we define d as its side length and
r as the radius of rounded corners (Fig.2). Computations are carried out for
several cases of r/d. An O-type grid is employed, and the grid points can be
concentrated near the body surface. The number of grid points is 257×129×81,
and the aspect ratio is one. In the spanwise direction, a periodic condition is
employed. Figure 3 shows body-fitted grid used in this computation. Global
view and local view near a corner are displayed in Fig.3 (a) and (b) respec-
tively. Fully developed 2-D flow is used as an initial condition for the 3-D
computation to save computation time.
3 Results
flow field visualized by stream lines, pressure contour lines and shading of
lateral component of vorticity.
Also, other computations at Re = 103 to 105 were carried out at Re = 103
to 105 , where aspect ratio based on the chord length is three and the number
of grid points is 129 × 65 × 193. These results are shown in Figures 7 and
8 presenting the instantaneous flow field at r/d = 0 and 0.333 respectively.
The Reynolds number is 105 . Contour surfaces of longitudinal component of
vorticity and pressure contour lines are visualized.
4 Conclusion
Simulation of flows around a cylindrical body was carried out by using
multidirectional-finite-difference method with O-grid. As a result of the com-
putation, flow fields were clarified in each case, and the drag variation caused
by different section shape of the body was captured well. Differences of flow
characteristics by r/d were elucidated by visualizing suitably.
References
1. K. Kuwahara and S. Komurasaki, ”Simulation of High Reynolds Number Flows
Using Multidirectional Upwind Scheme”, AIAA Paper 2002-0133 (2002)
2. S. Komurasaki and K. Kuwahara, ”Implicit Large Eddy Simulation of a Subsonic
Flow around NACA0012 Airfoil”, AIAA Paper 2004-0594 (2004)
3. T. Kawamura and K. Kuwahara, ”Computation of high Reynolds number flow
around a circular cylinder with surface roughness”, AIAA Paper 84-0340 (1984)
4. H. Suito, K. Ishii and K. Kuwahara, ”Simulation of Dynamic Stall by Multi-
Directional Finite-Difference Method”, AIAA paper 95-2264 (1995)
5. T. Igarashi, ”Characteristics of the Flow around Rectangular Cylinders (The
Case of the Angle of Attack 0 Deg)”, J. of JSME(B), 50-460, 3185–3192 (1984)
(in Japanese)
Implicit Large Eddy Simulation of a Flow around a Cylindrical Body 449
Fig.8. Flow around a rectangular cylinder with r/d = 0.333 at Re = 105 , contour
surfaces of longitudinal component of vorticity, pressure contour lines.
LES study of the impact of the wake structures
on the aerodynamics of a simplified ICE2 train
subjected to a side wind
1 Introduction
=0.580
=0.03
=0.07
=0.14
=0.44
=0.75
End car dummy
Support in the
Leading car End car experiment
Veiw from
front
Inner gap
Bogies
Spoiler in the
experiment Bogies in the
a) Ground board b) experiment
Lateral wall
Sli
Exit
pw
all
a) b)
Lateral wall
2 Results
Numerical accuracy is investigated by performing two simulations with dif-
ferent number of nodes: 6 and 12 million nodes. The spatial resolutions of
the first cell layer at the model and floor are shown in Table 1 for the two
meshes, where u∗ is the friction velocity, n is the distance between the first
node and the train surface in the wall normal direction, ∆s is the cell width in
the streamwise direction and ∆l is the cell width in the span-wise direction.
Table 1. Spatial resolutions for the coarse and the fine mesh simulations.
# #
" "
! !
!" !"
! "!! #!! $!! ! "!! #!! $!!
Angle ( ) Angle ( )
$ $
# #
" "
! !
!" !"
! "!! #!! $!! ! "!! #!! $!!
Angle ( ) Angle ( )
$ $
# #
" "
! !
!" !"
! "!! #!! $!! ! "!! #!! $!!
Angle ( ) Angle ( )
Fig. 3. Time-averaged surface pressure: (solid) fine mesh, (dashed) coarse mesh,
(symbols) experimental data.
car. In general, although the LES Reynolds number is seven times lower than
the experimental one, the LES results are in fairly good agreement with the
measurements.
Figure 4 shows the LES wake structures by means of isosurface of the in-
stantaneous pressure. The flow separates from the top-side face at the wind-
ward edge. It reattaches to the surface forming small separation bubble on
the top side face. On the other side, the flow remains attached to the bottom-
side face before it enters the wake. Both of the flows on the bottom-side and
top-side faces separate from the surface when they reach the lee-side face to
form line vortices emerging from the separation lines. These line vortices are
forming the wake flow patterns. The upper vortices are stronger and more
steady than the lower-side ones. Figure 4 shows that the first upper vortex
stretches in the wake flow all the way from the nose tip to the end of the
leading car. It grows in size but remains attached to the train surface. It is
steady for a length of about 5D from its onset at the front of the train before
it becomes unsteady and breaks up. The lower vortices are highly unsteady.
The cores of these vortices oscillate in a horizontal plan with low frequencies,
LES of side-wind flow over ICE2 train 455
where they attach and detach to and from the surface of the train with the
same frequency. The vortex shedding in the wake is mainly from the lower
vortices. Attachment and detachment of the wake vortices to the train leave
disturbances on its surface pressure which affect the aerodynamic coefficients.
Our LES resulted in wake structures are in agreement with those obtained
Upper vortex
Detached Attached
vortex vortex
Lower vortex
a) b)
Incident wind
a) b)
Fig. 5. (a) Experimental view (taken from [2]). (b) DES results (taken from [3]).
0.46 0.2
0.44 0.15
0.42
0.1
0.4
0.38 0.05
0.36 0
0 10 20 30 40 0 10 20 30 40
a) t b) t
3 15
2 10
1 5
0 0
0 0.1 0.2 0.3 0.4 0.5 0.6 0 0.1 0.2 0.3 0.4 0.5 0.6
c) d)
Fig. 6. (a) and (b) are time history of the side force and the lift force coefficients,
Cl and Cs , respectively. (c) and (d) are the corresponding autopower spectra.
Acknowledgments
This work was supported by the Swedish Agency for Innovation System (VIN-
NOVA), Bombardier Transportation, and Scania. Computer time on HPC2N
(High Performance Computing Center North), NSC (National Supercomputer
Center in Sweden), HIVE supercomputer at SWEGRID, and the HELIOS su-
percomputer provided by UNICC at Chalmers are gratefully acknowledged.
References
1. T. W. Chiu and L. C. Squire. An experimental study of the flow over a train
in a crosswind at large yaw angles up to 90o . Journal of Wind Engineering and
Industrial Aerodynamics, 45:47–74, 1992.
2. J. M. Copley. The three-dimensional flow around railway trains. Journal of Wind
Engineering and Industrial Aerodynamics, 26:21–52, 1987.
3. D. Wu. Predictive prospects of unsteady detached-eddy simulations in industrial
external aerodynamic flow simulations. Diploma thesis. matriculation number:
219949, Lehrstuhl fur Strömungslehre und Aerodynamishes Institute, Aachen,
Germany, 2004.
Use of Immersed Boundary Technique in a
Cartesian LES solver to study wake flows
Key words: LES, Immersed Boundary Technique, Wake Flow, Parallelism, Finite
Differences
Introduction
U˚i − U i
n
∂Pn n n
=− + Hj (U i , U j ) (3)
∆t ∂xi
1
Free/Libre and Open Source Software
2
Wall Adapting Local Eddy viscosity
Immersed Boundary Technique 459
− U˚i
n+1
Ui ∂dP
=− (4)
∆t ∂xi
1 ∂ U˚i
= ∇2 dP (5)
∆t ∂xi
Due to (4), the boundary conditions for the modified pressure dP are homo-
geneous Neumann on all the non immersed boundaries where the predicted
velocity is assumed the same as the corrected one. After (5), the modified
pressure is updated: Pn+1 = Pn + dP.
In MiOma the linear system associated to (5) is built on a cell-by-cell
basis dynamically at the initialization phase of the solver[2]: the construction
is done in a typical direction splitting approach. This characteristic will be
exploited in the future when the IBT implemented in MiOma will be extended
to pressure equation as well.
One interesting feature of the IBT is the direct link between CAD3 and the
CFD environments; indeed, the solid can be directly created in a CAD soft-
ware and read by the CFD software. To take into account solid location, the
pre-processor analyzes the CAD generated geometry file and distinguishes
cells as belonging to the solid, the fluid, or the boundary in between. This
procedure can be performed at the pre-processor stage in case of absence of
relative motion of the body with respect to the reference frame: for a moving
body the procedure hereafter described should be moved inside the solver.
There exist three ways to treat a 2D closed surface in a three dimensional en-
vironment, such as analytical functions, NURBS 4 , or discrete representation,
through a discrete set of points and a surface triangulation.
Based on the suggestions of previous studies[6, 7, 4], the last method was
chosen, through the use of the STL file format: being a de facto standard, the
format is almost always available when exporting geometries from any CAD
software. The STL or file format is an ASCII or binary file format, containing
a list of the triangular surfaces that describe a solid model plus some auxiliary
data. The main advantages to deal with such a triangulated geometry is the
capability to build general procedures.
3
Computer Aided Design
4
Non Uniform Rational B-Splines
460 J. Bodart, R. Giammanco, P. Rambaud and C. Benocci
In(3) Out(2)
Not modified
Corrected
In(1)
Set to 0
Fluid cell
Solid cell
Boundary cell
(a) Cell attributes and related IBT approach to (b) Ray-surface: number of
the velocity components according to the stag- intersection’s parity deter-
gered arrangement mine in or out position
Fig. 1.
5
Free and Open Source Software
Immersed Boundary Technique 461
4.5D
uniform area
14D 3.5D
5D 17D
6
Pentium IV , 3.4GHz
462 J. Bodart, R. Giammanco, P. Rambaud and C. Benocci
2 0.4
Spectral Simulation (Mittal & Balachandar) Spectral Simulation (Mittal & Balachandar)
MIOMA simulation (with IB) MIOMA simulation (with IB)
X/D =1.2
1 0.2
X/D=1.2
X/D =1.5 0
0
X/D=1.5
2
u’ /U∞
u/U∞
2
-1
X/D=2.0
X/D =2.5 -0.4
-2
X/D=2.5
-0.6
X/D =3.0
-3
X/D=3.0
-0.8
-2 -1.5 -1 -0.5 0 0.5 1 1.5 2 0
Z/D (normal direction) Z/D (normal position)
(a) Mean streamwise velocity in the (b) Time averaged streamwise velocity
wake fluctuations in the wake
Conclusion
A promising method has been introduced to study flows around bluff bodies
using IBT. A succesful validation at low Reynolds number around a circu-
lar cylinder is obtained. It has to be underlined how easy and fast are the
CAD and meshing stages. Future implementation will include moving and/or
deforming object, as well as mesh refinement localised near the body.
References
1. GTS development team. Gts web page. http://gts.sourceforge.net/, 2006.
2. R. Giammanco. Numerical Study of Coherent Structures within a legacy LES
code and development of a new parallel Frame Work for their computation. PhD
thesis, von Kármán Institute for Fluid Dynamics, 2005.
3. F. H. Harlow and J. E. Welsh. Numerical calculation of time-dependent vis-
cous incompressible flow of fluid with free surface. The Physics of Fluids, 8, N.
10:2182–2189, 1965.
4. G. Kalitzin and G. Iaccarino. Toward immersed boundary simulation of high
reynolds number flows. CTR-Annual Research Briefs, pages 369–379, 2003.
5. R. Mittal and S. Balachandar. On the inclusion of thre dimensional effects in
simulations of two-dimensional bluff body wake flows. Proceedings of ASME
Fluids engineering division Summer Meeting, jun 22-26 1997.
6. R. Verzicco and G. Iaccarino. Immersed boundary technique for turbulent flow
simulations. Applied mechanics review, 56:331–341, 2003.
7. R. Verzicco and G. Iaccarino. Immersed boundary technique for large eddy sim-
ulation. In Von Karman Institute, editor, VKi Lecture series on Large Eddy
Simulation and related techniques: Theory and Applications, march 2006.
8. C.H.K. Williamson. Vortex dynamics in the cylinder wake. Annual review of
Fluid Mechanics, 28:477–539, 1996.
Stochastic generation of velocity fluctuation for
turbulent inflow and initial condition
1 Introduction
It is well known that besides the mathematical model and numerical scheme,
proper specification of boundary and initial conditions is of great importance
in the representation of a physical system. From a theoretical point of view
an adequate set of boundary and initial conditions is necessary to achieve a
unique solution of the set of partial differential equations to be solved. This
issue, is even more critical in turbulent flow simulations, using DNS or Large
Eddy Simulation (LES) approaches. Since in these approaches more flow field
structures are captured, all the resolved scales must be quantified at the inflow
or initial conditions. Moreover, due to the chaotic character of the Navier-
Stokes equations, turbulent flow field simulations are strongly influenced by
their inflow or initial conditions [5]. Therefore, any uncertainty in these data
In this section, the new methodology for generating artificial turbulent veloc-
ity fields is explained. To improve the readability of the paper, mathematical
derivations are discussed for a two-dimensional velocity field which is consis-
tent with the considered test case. In the new method, each velocity com-
ponent u and v is formed by a linear combination of different uncorrelated
random fields fij with zero mean, i.e.,
u = a11 f11 + a12 f12 ,
v = a21 f12 + a22 f22 . (1)
field, fij in equation (1), can be generated by applying a gaussian filter, with
appropriate filter width, to an arbitrary white noise field ξij . Moreover, it is
assumed that in equation (1) the cross-term coefficients aij and aji are es-
tablished with a common structure fij . However, the sign of these cross-term
coefficients depend on the sign of the respective Reynolds-stress. This leads
to:
aij = sign(hui uj i).aji = ±aji . (3)
With these assumptions the following set of equations for the generated ve-
locity field is obtained:
r2 r2
u = a11 . exp(− 2 ) ◦ ξ11 + a12 . exp(− 2 ) ◦ ξ12 ,
σ11 σ12
r2 r2
v = ±a12 . exp(− 2 ) ◦ ξ12 + a22 . exp(− 2 ) ◦ ξ22 , (4)
σ12 σ22
where ◦ denotes convolution. Correlation functions for the generated velocity
field expressed by equation (4) are:
r2 r2
r r
2 π 2 π
Ruu = (a11 σ11 ). exp(− 2 ) + (a12 σ12 ). exp(− 2 ),
2 2σ11 2 2σ12
r2 r2
r r
π π
Rvv = (a212 σ12 ). exp(− 2 ) + (a222 σ22 ). exp(− 2 ), (5)
2 2σ12 2 2σ22
r2
r
2 π
Ruv = sign(huvi)(a12 σ12 ). exp(− 2 ).
2 2σ12
Therefore, Reynolds-stresses and integral length scales can be easily calculated
by respectively setting r = 0 and integrating with respect to r in the expression
for the correlation functions (5). These provide six linear algebraic equations
for six unknown filter widths σij and coefficients aij in equation (4).
Table 1. Properties of different turbulent initial conditions used for different simula-
tions. The imposed statistical quantities for each initial turbulent field are specified.
prescribed turbulent kinetic energy components huui, hvvi and hwwi. Initial
condition IC2 has complete Reynolds-stresses information generated based on
Lund’s transformation [3]. Initial condition IC3 has prescribed turbulent ki-
netic energy components huui, hvvi,hwwi and normal integral length scales
luu , lvv and lww , generated by applying appropriate gaussian filters on the
velocity field of IC1. Finally, initial condition IC4 is generated based on the
new approach and satisfies both Reynolds-stresses and integral length scales
lij , including cross terms.
In the considered temporal homogeneous turbulent shear flow, the im-
posed mean flow in streamwise direction x is a linear function of the normal
coordinate z:
U (z) = Sz, (6)
where S is the mean shear rate. For all simulations the kinematic viscosity is
ν = 0.001 and S = 10.0. All the flow field parameters are in arbitrary units
[4]. The fully compressible Navier-Stokes equations are solved (DNS), only for
the fluctuating part of the velocity field −→
u (−
→x , t),
→−
−
U (→x , t) = U (z)−→
ex + −
→
u (−
→
x , t), (7)
→−
−
where U (→ x , t) is the complete velocity field and −→
ex is the unit vector in
streamwise direction x. The set of Navier-Stokes equations are solved based on
a cell-averaged finite-volume approach. Convective and viscous fluxes are dis-
1
cretized on a uniform mesh with 1003 grid cells, with grid spacing 4x/(ν/S) 2
equal to 6, using a second-order discretization. For time integration an explicit
four-stage compact-storage Runge–Kutta scheme of second-order accuracy is
used.
Figure 1 shows the evolution of the normalized Reynolds-stress compo-
nents for these simulations. Comparing the evolution of these quantities in
figure 1 reveals that the flow field started from initial condition with pre-
scribed integral length scales, i.e. IC4 and IC3 are much closer to the reference
simulation. Moreover, using the initial velocity field IC4, which is generated
based on the new procedure, results in a more accurate flow field evolution
compared to the simulation with only prescribed normal integral length scale
IC3. The difference between the latter two cases can be better observed by
Stochastic generation of turbulence 467
considering their respective deviation from the reference simulation. The error
is determined as:
Z τ =end q
2
errtot (φ) = φ(τ ) − φref (τ ) dτ, (8)
τ =τ ∗
0.48 0.33
0.46 0.32
0.31
0.44
<vv>/q
<uu>/q
0.3
0.42
0.29
0.4
0.28
0.38 0.27
0.36 0.26
5 10 15 20 5 10 15 20
St (a) St (b)
0.32 0.6
0.31 0.5
0.3
<−uw>/u’w’
0.4
<ww>/q
0.29
0.3
0.28
0.2
0.27
0.26 0.1
0.25 0
5 10 15 20 5 10 15 20
St
(c) St (d)
Fig. 1. Flow field evolution of homogeneous turbulent shear flow, initialized with
different random p (b):hvvi/q; (c):hwwi/q; (d):−huwi/u0 w0 where
fields (a):huui/q; p
0 0
q = hui ui i, u = huui and w = hwwi. (—): reference simulation, (◦): IC4 or
new approach, (square): IC3, (.): IC2, (4): IC1.
468 M. Fathali, M. Klein, T. Broeckhoven, C. Lacor, and M. Baelmans
results are obtained from initial condition where all the Reynolds-stresses and
integral length scales are prescribed, i.e. IC4. Listed in table 2 are total errors
for all simulations computed based on the equation (8) and normalized with
the total error of IC4.
4 Conclusion
References
1. Klein, M., Sadiki, A., and Janicka, J., “A digital filter based generation of inflow
data for spatially developing direct numerical or large eddy simulations,” Journal
of Computational Physics, Vol. 186, 2003, pp. 652–665.
2. Kempf, A., Klein, M., and Janicka, J., “Efficient Generation of Initial- and In-
flow Conditions for Transient Turbulent Flows in Arbitrary Geometries,” Flow,
Turbulence and combustion, Vol. 74, No. 1, 2005, pp. 67–84.
3. Lund, T. S., Wu, X., and Squires, K. D., “Generation of turbulent inflow data
for spatially-developing boundary layer simulations,” Journal of Computational
Physics, Vol. 140, 1998, pp. 233–258.
4. Rogers, M. M. and Moin, P., “The Structure of the Vorticity Field in Homoge-
neous Turbulent Flows,” Journal of Fluid Machanics, Vol. 176, 1985, pp. 33–66.
5. Dziomba, B. and Fiedler, H. E., “Effect of initial conditions on two-dimensional
free shear layers,” Journal of Fluid Mechanics, Vol. 152, 1985, pp. 419–442.
Study on Numerical and Modelling Error in
LES of a Channel Flow Using Explicit Filtering
Tellervo T. Brandt
1 Introduction
It is well recognised that the numerical error involved in a large eddy simu-
lation (LES) applying low-order finite-difference schemes is large when com-
pared to the effect of sub-grid scale (SGS) model [1]. However, depending on
the applied SGS model and the choice of the numerical method, the modelling
error can be of the same size as the numerical error and the two error compo-
nents may interact with each other [2, 3]. In this paper, the error components
are studied a posteriori in a fully developed turbulent channel flow between
two infinite parallel walls using explicit filtering. In Reference [4], this case was
studied a posteriori using an approach based on implicit filtering [3], and here
the two approaches are compared. Previously, explicit filtering with smooth
filters was noticed to have a large effect on the simulations [5], and here the
effect is studied a posteriori to exclude the effect of numerics.
Table 1. Applied grid resolution in wall units, ∆+ , and extent of the computational
domain, l, in streamwise (x), spanwise (y) and wall-normal (z) directions
grid ∆x+ ∆z + ∆y + lx /h lz /h ly /h
LES 1 44 23 1.0, . . . , 20 6 3 1
fine-grid LES 2 22 12 0.5, . . . , 10 6 3 1
fine-grid LES 3 15 8 0.5, . . . , 10 3 1.6 1
wall units: x+ = Reτ x, where x is scaled by the channel half-height
4 A Posteriori Tests
On the left-hand side of Fig. 1, we have the mean-velocity profile obtained from
simulation on grid 1 and the corresponding finer-grid cases, and in addition,
the DNS from [6] and the grid 3 case with filtering but no SGS model, which
approximates a filtered DNS, are included. The difference between grid 2 and
grid 3 results is small and thus, grid 3 result is a fair approximation to grid-
independent LES. The error components (1) evaluated using these results
are depicted on the right-hand side of the figure. In all the simulations with
filtering, the shape of the velocity profile is almost the same, and it differs from
the DNS result. Since the change is visible already in the filtered DNS, the
shape of the profile is due to filtering and not due to modelling or numerics.
The large effect of filtering is visible also on the right-hand side of the figure
where the effect of filtering is larger than the total simulation error. The
difference between grid 3 case and the filtered DNS is small which indicates
that the effect of SGS modelling is small and this is visible also in the small
modelling error. Both modelling and numerical error are visible mainly in the
mean bulk velocity, and the error components are thus nearly constants in the
logarithmic layer. In addition, the numerical and modelling error are of the
opposite sign and they partially cancel out each other.
1.5
20
1
18
0.5
16
0
! U+
14
U+
12 !0.5
LES, grid 1 !1
10 fine!grid LES, grid 2 total
fine!grid LES, grid 3 filter
8 filtered DNS !1.5 modelling
DNS numerical
6 !2
10 100 10 100
y+ y+
In the earlier study using implicit filtering, similar counteraction of the nu-
merical and modelling error was noticed, and there the numerical and mod-
elling error were approximately of the same size [4]. In the present results,
both the error components are smaller, which is probably partly due to ex-
plicit filtering, since it reduces the effect of the Smagorinsky model [5], and
partly due to different resolutions in the wall-normal direction.
In Figure 2, the corresponding plots are given for the deviatoric streamwise
Reynolds stress. Here, the effect of modelling is even smaller than in the mean-
velocity profile, and most part of the error is formed by explicit filtering.
The numerical error is again of the opposite sign to the modelling error and
472 Tellervo T. Brandt
thus, they partially cancel out each other. The error components are similar
for the other diagonal Reynolds stresses. In the earlier study with implicit
filtering, the obtained modelling error was larger than here, the distribution
of the numerical error somewhat different, and counteraction of the error
components was not visible there.
0 2
!1
1
!2
!3 0
! u’u’*
!u’u’*
!4
!1
!5
LES, grid 1
!6 fine!grid LES, grid 2 !2 total
fine!grid LES, grid 3 filter
!7 filtered DNS modelling
!3
DNS numerical
!8
0 0.05 0.1 0.15 0.2 0 0.05 0.1 0.15 0.2 0.25
y/h y/h
1.5
20
1
18
0.5
16
0
! U+
14
U+
12 !0.5
LES, grid 1 !1
10 fine−grid LES, grid 2 total
fine−grid LES, grid 3 filter
8 filtered DNS !1.5 modelling
DNS numerical
6 !2
10 100 10 100
y+ y+
implicit filtering [4]. Increasing the model length scale improves the prediction
of the peak value of the Reynolds stress, but at the same time, it widens the
distribution. This is visible also in the total error which decreases in the near-
wall region but increases after the peak of the Reynolds stress when compared
to the case with lower model constant in Fig. 2. However, changes in the total
error are mainly due to the counteraction of the numerical and modelling
error. Increasing the model length scale increases the effect of modelling and
the modelling error, and now, modelling error is larger than the numerical
error, but they are still of different sign. When compared to the case with
the smaller model length scale, the numerical error clearly increases, and the
increase is stronger than in the mean-velocity profile. Similar behaviour was
noticed in the study with implicit filtering [4]. However, there the numerical
error of the Reynolds stress increased, as it should, also when the model length
scale was reduced to 0.5∆.
0 2
!1
1
!2
!3 0
! u’u’*
!u’u’*
!4
!1
!5
LES, grid 1
!6 fine!grid LES, grid 2 !2 total
fine!grid LES, grid 3 filter
!7 filtered DNS modelling
!3
DNS numerical
!8
0 0.05 0.1 0.15 0.2 0 0.05 0.1 0.15 0.2 0.25
y/h y/h
5 Conclusions
In this paper, a posteriori tests were performed using explicit filtering in the
channel flow. The numerical error related to the second-order scheme and the
modelling error of the standard Smagorinsky model were approximately of
the same size and of different sign. When compared to results obtained with
implicit filtering [4], the results were similar except for the effect of explicit
filtering itself. Filtering reduced both error components. In the present results,
the large effect of filtering with a smooth filter was demonstrated at fine grid
resolution, and this verifies that the strong effect of filtering obtained in [5]
was not due to numerics.
The a posteriori tests were repeated with a larger model length scale. This
showed that the slight improvement of simulation results with this model
length scale that was noticed in [5] was due to interaction of the error com-
ponents. The increased modelling error cancelled out the numerical error.
In the study with implicit filtering, some increase of numerical error with
increasing model length scale was visible in the mean-velocity profile and in
the Reynolds stresses, and this trend was noticed also here. This behaviour
suggests that in these a posteriori studies, part of the effect of modelling is
included in the numerical error as the model length scale is increased.
References
1. S. Ghosal. An analysis of numerical errors in large-eddy simulations of turbulence.
Journal of Computational Physics, 125:187–206, 1996.
2. B. Vreman, B. Geurts, and H. Kuerten. Comparison of numerical schemes in
large-eddy simulation of the temporal mixing layer. International Journal for
Numerical Methods in Fluids, 22:297, 1996.
3. B. Geurts and J. Fröhlich. A framework for predicting accuracy limitations in
large-eddy simulation. Physics of Fluids, 14(6):L41–L44, June 2002.
4. T. Brandt. A posteriori study on modelling and numerical error in LES applying
the Smagorinsky model. In Complex Effects in LES, Lecture Notes in Computa-
tional Science and Engineering. Springer, 2006. Accepted for publication.
5. T. Brandt. Study of large eddy simulation and Smagorinsky model using explicit
filtering. In 36th AIAA Fluid Dynamics Conference and Exhibit, San Francisco,
California, June 5-8 2006. AIAA-2006-3541.
6. T. S. Lund. On the use of discrete filters for large eddy simulation. Center
for Turbulence Research, Annual Research Briefs, pages 83–95, 1997. Stanford
University.
7. O. V. Vasilyev, T. S. Lund, and P. Moin. A general class of commutative filters for
LES in complex geometries. Journal of Computational Physics, 146(1):82–104,
October 1998.
8. R. D. Moser, J. Kim, and N. N. Mansour. Direct numerical simulation of tur-
bulent channel flow up to Reτ = 590. Physics of Fluids, 11(4):943–945, April
1999.
Wall Boundary Conditions for Variational
Multiscale Large-Eddy Simulations
1 Introduction
momentum and energy equations on the wall to zero, and explicitly constrain-
ing the velocity and temperature variables. The value of density can be left
free to be determined by the numerical solution.
As described in reference [4], weak solid-wall conditions can be based on
the adjoint-consistent form of the penalty method introduced by Nitsche (see
[5] for details). As the near-wall flows considered here are essentially incom-
pressible, such conditions can be constructed by adding the following two
terms to the momentum components of (2):
− (2µ∗ ∇s W · n, U − g)Pn + (Wαµ∗ /h, U − g)Pn (9)
∗
Where g is the desired boundary state, µ = µ + µm is the sum of the physical
and model viscous coefficients, h is a measure of the element size and and α is
a positive parameter. In addition we explicitly constrain the normal velocity
to zero and the wall temperature to its specified value, and leave the value of
density free as in the hard condition procedure.
We first consider a low Reynolds number Reτ = 180 case with a standard
level of LES resolution. For the spatial discretisation, a continuous hierarchical
basis with linear and quadratic element modes was used on a 16 × 16 × 16
mesh. The element geometries were stretched in the wall-normal direction
using a hyperbolic tangent function. The flow conditions for this case and all
subsequent cases have been scaled so that the maximum Mach number is less
than 0.3.
Figure 1 compares RMS velocity profiles from a single-scale method with
with van Driest damping (SvD) and a variational-multiscale method (VM)
with data obtained from direct numerical simulation (DNS) [7]. Results for
both hard (H) and and weak (W) solid-wall conditions are shown. For the
single-scale method, van Driest damping is required to achieve acceptable re-
sults on the current mesh. In contrast, the VM produces good results with only
a constant-coefficient model. As both results are relatively resolved, however,
the impact of weak boundary condition is minimal. In this case the first inter-
element boundary is located at y + = 5.9, which provides sufficient resolution
of the near-wall gradient when quadratic interpolation is used.
We now consider a higher Reynolds number, Reτ = 590, with a coarser LES
resolution. The same quadratic basis as used previously is employed on a
8 × 8 × 8 mesh. No stretching is used, so that the first intra-element interface
is located at y + = 74.
Both the SvD and VM methods with hard conditions underpredict the wall
velocity gradient (figure 2, left). In the case of the VM, this results in a higher
Wall Boundary Conditions for Multiscale LES 479
3.5
SvD H 1.4 SvD H
SvD W SvD W
3 VM H VM H
VM W 1.2 VM W
DNS DNS
2.5
1
2
0.8
wrms
urms
1.5
0.6
1 0.4
0.5 0.2
0 0
0 50 100 150 200 0 50 100 150 200
+
y y+
30 3.5
VM H
VM W
3 DNS
25
2.5
20
Mean Velocity
2
urms
15
1.5
10
1
SvD H
5 SvD W
VM H 0.5
VM W
DNS
0 0
0 100 200 300 400 500 600 0 100 200 300 400 500 600
+ +
y y
mean velocity. In the case of SvD, the coarse mesh appears to accentuate
the effect of the dissipative model on the large scales, resulting in a reduced
velocity profile. As in the previous case, the use of weak boundary conditions
does not significantly benefit the SvD results. As can be seen in figures 2 and
3, however, the VM results are strongly improved. The mean flow prediction
is far better, particularly at large y + , as is required for a good wall model.
The fluctuating velocity components are also improved at large y + .
480 S. J. Hulshoff and E. A. Munts
2.5
1.4 VM H VM H
VM W VM W
DNS DNS
1.2 2
1
1.5
0.8
wrms
vrms
0.6 1
0.4
0.5
0.2
0 0
0 100 200 300 400 500 600 0 100 200 300 400 500 600
+
y y+
Concluding remarks
References
1. Hughes, T. J. R., Mazzei, L. and Jansen, K. E., Large Eddy Simulation and the
Variational Multiscale Method. Comput.Visual Sci. 3 47–59 (2000)
2. Collis, S. S., The DG/VMS Method for Unified Turbulence Simulation. AIAA
Paper No. 2002-3124 (2002)
3. Munts, E., Space-time multiscale methods for large-eddy simulation. Ph.D. The-
sis, Faculty of Aerospace Engineering, Delft University of Technology (2006)
4. Bazilevs Y., and Hughes, T. J. R., Weak Imposition of Dirichlet Boundary Con-
ditions in Fluid Mechanics, ICES Report 05-25, University of Texas at Austin
(2005)
5. Arnold, D. N., Brezzi, F., Cockburn, B, and Marini, L. D., Unified Analysis of
Discontinuous Galerkin Methods for Elliptic Problems. SIAM J. Numer. Anal.,
Vol 39, No 5, pp 1749-1779 (2002)
6. Schumann, U., Subgrid Scale Model for Finite Difference Simulations of Turbu-
lent Flows in Plane Channels and Annuli. J. Comp. Physics, 18 376–404 (1975)
7. Moin, P., and Kim, J., Numerical investigation of turbulent channel flow. Journal
of Fluid Mechanics, 118, 341–371 (1982)
The Sampling Based Dynamic Procedure for
Numerical Discretization Enhancement
1 Introduction
Recently, a new sampling formalism for Large Eddy simulation was proposed
by Winckelmans et al. [1] and Knaepen et al. [2], which is a projection method
for Navier-Stokes equations from continuum space to a discrete space, using
a sampling operator instead of a filter operator. Since the sampling operator
is not commutative with spatial derivatives, a closure term appears which
represents the loss of information due to the projection on a discrete mesh. In
e.g. [2] a Smagorinsky model was proposed that, by relying on a generalized
dynamic procedure, succeeded in accounting for the subgrid scales. In this
paper, we investigate the ability of this sampling based dynamic procedure, in
combination with an appropriate model for the truncation error, to obtain
higher-order numerical accuracy. Two such possible models are presented.
Further, we show that Richardson extrapolation is a simplified formulation of
this procedure.
We define the sampling operator S∆1 , which operates between the continuum
space Ω ⊂ IRn and the discrete space Ω ∆1 with N number of grid points,
and spacing ∆1 . This sampling operator S∆1 is idempotent, and commutative
with the product of the non-linear terms. However, S∆1 does not commute
with spatial derivatives. We use the notation S∆1 ◦ ui = ui and S∆1 ◦ ∂ = δ.
Applying S∆1 to the continuity equation and the momentum equations gives
δui
= Π ∆1 (1)
δxi
∂ui δui δp δ 2 ui
+ uj =− + ν 2 + Σi∆1 (2)
∂t δxj δxi δxj
The truncation errors are due to the non-commutativity of the operator S∆1
with the spatial derivatives, and have the basic form
δui ∂ui
Π ∆1 = − (3)
δxi ∂xi
!
δ 2 ui ∂ 2 ui
δui ∂ui δp ∂p
Σi∆1 = uj − + − −ν − (4)
δxj ∂xj δxi ∂xi δx2j ∂x2j
C∆2 and Ni∆2 are called here the continuity and Navier-Stokes operators,
respectively. Applying S∆2 to the continuous set equations gives
δe
ui
0 = C∆2 (e
u i ) + Π ∆2 = − + Π ∆2 (7)
δxi
∂e
ui δe
ui δep δ2 u
ei
= Ni∆2 (e
ui ) + Σi∆2 = −e
uj − + ν 2 + Σi∆2 (8)
∂t δxj δxi δxj
Ideally, the latter set should also be obtained by applying the sampling oper-
ator S∆2 to the first set of equations (5)-(6) and because of S∆2 ◦ S∆1 = S∆2 ,
following relation is imposed:
S∆2 ◦ C∆1 (ui ) − C∆2 (eu i ) = Π ∆ 2 − S∆ 2 ◦ Π ∆ 1 (9)
S∆2 ◦ Ni∆1 (ui ) − Ni∆2 (e
ui ) = Σi∆2 − S∆2 ◦ Σi∆1 (10)
These explicitly express the commutation errors made by the projection
Ω ∆1 → Ω ∆2 . This expression can be determined in terms of the resolved
velocity ui since S∆2 ◦ ui = u
ei = uei . Suppose models are adopted for both
truncation errors Π ∆1 and Σi∆1 with the basic forms
Π ∆1 = C π mπ,∆1 , Σi∆1 = Ciσ mσ,∆
i
1
(11)
π
and analogously for the test-level ∆2 . Then the fields C and can be ob- Ciσ
tained from relation (9)-(10).We suggest to use a least square method with
The Sampling Based Dynamic Procedure for Numerical Improvement 483
δ n u ∆
Consider the Taylor series expansion of the nth discrete derivative δxn
on grid with spacing ∆, for a k th -order central discretization scheme
∆
∂nu δ n u k∂
k+n
u
+ O ∆k+2
(x) = + ck ∆ (12)
∂xn δxn ∂xk+n
2∆
∂nu δn u k ∂
k+n
u
+ O ∆k+2
e
n
(x) = n
+ ck (2∆) k+n
(13)
∂x δx ∂x
We assume now, that the leading order truncation term is an adequate model
for the complete truncation error. Discretization of this term and applying
the generalized dynamic procedure leads to a coefficient ck . Substitution in
(12) results finally in
n ∆ k+n 2∆
2∆ k+n ∆
δn u
k δ u δ u e δ u
−
e
2
∂nu δxn δxk+n δxn δxk+n
(x) ≈ (14)
k+n 2∆
∂xn ∆
δ k+n u
kδ u
−
e
2
δxk+n δxk+n
This expression is closely related to Richardson extrapolation. The Richardson
extrapolation formula is obtained for
2∆ ∆
δ k+n u
e δ k+n u ∂ k+n u
≡ ≡ (15)
δxk+n δxk+n ∂xk+n
th
which gives a (k + 2) -order accurate central scheme for the nth derivative.
For further analysis of expression (14), and the implications of approximation
(15), a Fourier analysis for a single wave is performed for the 2nd -order ac-
curate gradient. A blending factor f is introduced in order to switch between
the dynamic procedure (f = 1) and Richardson extrapolation (f = 0), and
to investigate intermediate behaviour. Modified wavenumbers of the resulting
expression are given in figure 1 for a 2nd , 4th (equals Richardson Extrapola-
tion f = 0) and 6th order central scheme, and also for the dynamic procedure
formula at values of f = 1, 12 , 13 , 14 , 15 , 10
1
. We noticed from a semilog error plot
1
that f = 5 is the value for which the dynamic procedure obtains best results.
For the lower wavenumbers the accuracy is slightly better than 6th -order, for
the higher wavenumbers it is slightly less.
484 Dieter Fauconnier, Chris Delanghe, Erik Dick
1
0.9
0.8
0.7 f=1
f=1/2
0.6
f=1/3
f=1/4
!"/!max
0.5
f=1/5
f=1/10
0.4
0.3
0.2
0.1
0
0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 1
!/!
max
5 Results
The testcase is a laminar flow in a 2D driven cavity at Reynolds number
Re = 400. A pseudo-compressible code is used with a 3rd -order Runge-Kutta
method for stepping in pseudo time. Spatial discretization is 2nd -order central.
No artificial stabilisation is used in the continuity equation. Therefore, the
fields contain a minimal spurious pressure mode, that however does not affect
the velocity results. The exact truncation errors are (16)-(17). We use a 6th -
order central scheme on a 180 × 180 mesh, to generate the reference solution.
The different simulations are given in Table 1.
Using an exact truncation model in combination with the dynamic pro-
cedure, a global least-square averaging was chosen for C π en C σ . For the
486 Dieter Fauconnier, Chris Delanghe, Erik Dick
! !
u v
0.025
0.02
0.02
0.015
0.015
0.01
0.01
0.005
0.005
0
0
−0.005
−0.005
−0.01
−0.01
−0.015
−0.015
−0.02
−0.02
−0.025
−0.025
−0.03
References
1. G. Winckelmans, L. Bricteux, and L. Georges. The “sampling-based dynamic
procedure” for LES without filtering: validation using finite differences. In
E. Lamballais et al., editors, Ercoftac Workshop - Direct and Large-Eddy Sim-
ulations 6, pages 140–141, Poitiers, September 2005. Université de Poitiers. IN
PRESS.
2. B. Knaepen, O. Debliquy, and D. Carati. Large-eddy simulation without filter.
J. Comp. Phys., 205:98–107, 2005.
Application of a Unique Eddy–Viscosity Model
for a Hybrid LES–RANS Method
1 Introduction
Although LES is a highly promising simulation technique, it still suffers from
extremely large resources required for the resolution of the near–wall region,
especially for high–Re flows. That is the main motivation for setting up hy-
brid LES–RANS methods. Whereas RANS suits reasonably well to attached
boundary layers, requiring much less CPU–time and memory than LES, the
latter is recommended for complex large–scale flow phenomena, which RANS
often fails to predict correctly. Both characteristics are combined in hybrid
LES–RANS methods to obtain an optimal solution at low cost. Meanwhile a
variety of different hybrid concepts were proposed including DES.
In the present study a non–zonal approach based on two unique models is
preferred. The predefinition of RANS and LES regions is avoided and a grad-
ual transition between both methods takes place which weakens the problem
of setting up an appropriate coupling strategy. The new hybrid LES–RANS
approach relies on a one–equation model for the turbulent kinetic energy in
both modes. In addition to a linear eddy–viscosity model (LEVM), an explicit
algebraic Reynolds stress model (EARSM) is applied in order to account for
the Reynolds stress anisotropy. Both model variants have been tested on the
basis of the standard plane channel flow. Additionally, for the LEVM the flow
over a periodic arrangement of hills was considered.
is the main target for RANS, the equation for the length scale can be simpli-
fied to an analytical relation leading to a one–equation model for both zones.
Presently, the near–wall model by Rodi et al. [1] is applied in the viscosity–
affected near–wall layer and combined with a one–equation SGS model in the
outer LES region. The resulting unique model consists of a transport equation
for the modeled turbulent kinetic energy kmod in RANS mode and the sub-
grid scale turbulent kinetic energy ksgs in LES mode, respectively. For LES
the unknown dissipation correlation is modeled by ε = Cd k 3/2 /∆ leading to
νt = Cµ k 1/2 ∆ (Cµ = 0.05, Cd = 1.0).
As shown by Rodi et al. [1], for the near–wall region (RANS mode) the ap-
plication of the wall–normal velocity fluctuations (v 0 2 )1/2 as velocity scale is
better suited to characterize the turbulent motion than k. The length scale
lµ,v appearing in the eddy–viscosity relation νt = (v 02 )1/2 lµ,v now scales lin-
early with the wall–distance y and needs no damping function anymore, i.e.,
lµ,v = Cl,µ y. Indeed, the eddy–viscosity reduction in the vicinity of the wall is
an effect of the decreasing v 0 fluctuations rather than a viscous one and hence
naturally taken into account. The dissipation length lε,v used to define the
dissipation rate ε = (v 02 )1/2 k/lε,v also scales linearly near the wall. Only in
the immediate vicinity of the wall the distribution must be modified to yield
the correct behavior of ε ∼ y 0 as y goes to zero. In order to apply (v 0 2 )1/2 as
the velocity scale, an equation relating the wall–normal velocity fluctuations
to k = kmod was introduced [1] so that the transport equation does not have
to be modified. However, some adjustments on the model constants have been
done in order to fit more closely to high–Re flows [2].
In addition, an explicit algebraic Reynolds stress model (EARSM) de-
signed by Wallin and Johansson [3] for pure RANS applications has been
implemented in complement to the LEVM. This special EARSM was cho-
sen because of its near–wall treatment ensuring realizability of the individ-
ual stresses. The objective is to investigate whether a more advanced for-
mulation, which avoids the limitations of LEVM such as the poor predic-
tion of the normal Reynolds stresses, could be beneficial to the approach
suggested. The EARSM provides an algebraic relation for the Reynolds
stress tensor [3], which can be introduced in the momentum equation as
(ex) (ex)
u0i u0j = 23 k δij − 2 νt S ij + k aij . Here aij represents an extra anisotropy
tensor which is computed explicitly based on the normalized mean strain and
rotation tensors. Hence in this formulation the transport equation for k and
the relation for determining νt are still needed but the additional term takes
the anisotropy of the stresses appropriately into account. A further conse-
quence of applying an EARSM is on the production term in the transport
equation for k, which does not require any modeling anymore. Concerning
the interface between RANS and LES, an approach involving an interface
treatment without synthetic or reconstructed turbulence is preferred. Since
the criterion used by Rodi et al. [1] in their two–layer approach was found to
be unsuitable in the hybrid LES–RANS technique, it has been abandoned [2].
Unique Eddy–Viscosity Model for a Hybrid LES–RANS Method 489
1/2
As a remedy, the dynamic criterion y ∗ = kmod y/ν has been chosen to define
the LES–RANS interface [2]. y ∗ ≤ 60, which is also the validity restriction of
the relation given for (v 02 ) in the LEVM, means RANS mode, whereas y ∗ > 60
sets the simulation in LES mode. Alternatives are discussed in [2].
3 Results
Numerical Methodology and Test Cases
All simulations were performed with the LES code LESOCC, used for the solu-
tion of the filtered Navier–Stokes equations. This is a 3–D finite–volume solver
for arbitrary non–orthogonal and non–staggered (block–structured) grids [4].
The discretization is second–order accurate in space and time.
Two different test cases are considered. The first is the classical plane
channel flow at Reτ = 590 [6] performed on two grids. The dimensions of the
computational domain are: 2π (streamwise)× π (spanwise) × 2 (wall–normal
direction). The first grid A consists of 128 × 128 × 128 control volumes (CVs),
which represents a resolution of ∆x+ = O(30) (streamwise) and ∆z + = O(15)
(spanwise). The first grid point (half first cell) is located at y + = 0.68. The
second grid B contains 64×64×64 CVs leading to a resolution of ∆x+ = O(60)
and ∆z + = O(30). Here the first grid point is located at y + = 1.46.
A more challenging test case is the flow over periodic hills at Reb =
10, 595 [5]. The grid consists of 160×100×60 CVs in streamwise, wall–normal
(∆ycrest /h = 0.005, 1st CV height) and spanwise direction, respectively. This
grid has been originally designed for DES. The number of grid points in the
spanwise direction is moderate so that the RANS region becomes larger.
The results of the EARSM hybrid method are shown and compared with
the LEVM hybrid formulation which does not account for the anisotropy. Since
the approaches using the EARSM in both RANS and LES modes (E–RL) or
in RANS only (E–R) provide similar results, only the latter is presented in
Fig. 1. The performance of the LEVM is not studied here for the channel
flow but for the hill flow (c.f. [2]). The EARSM hybrid method is first studied
without and finally with the non–modeled production term in k–equation.
25 3
L 1.4 L
E–R E–R
20 2.5 1.2
E–Rnp E–Rnp
DNS DNS
RMS(u′ u′ )tot
RMS(v ′ v ′ )tot
2 1
15
+ 1.5 L 0.8
U
10 E–R 0.6
1 E–Rnp
DNS 0.4
5 0.5 0.2
0 0 0
1 10 100 0.1 1 10 100 0.1 1 10 100
y+ y+ y+
Fig. 1. Channel flow at Reτ = 590 on coarse grid B. Mean streamwise velocity
U + (left). Total normal Reynolds stresses (u0 u0 )tot (center) and (v 0 v 0 )tot (right). L:
LEVM hybrid method. E–R: hybrid method with EARSM in RANS mode.
y/h
y/h
1.5 1.5 1.5
1 1 1
0.5 0.5 0.5
0 0 0
-0.2 0 0.2 0.4 0.6 0.8 1 1.2 0.02 0.04 0.06 0.08 0.1 -0.04 -0.03 -0.02 -0.01 0
U/Ub ufuftot ufvftot
3 3
2 2
y
y
1 1
0 0
0 2 4 6 8 0 2 4 6 8
x x
Fig. 2. Periodic hill flow at Reb = 10, 595. Mean velocity U/Ub at x/h = 0.5, 2 and
6 (left), u0 u0 tot (middle) and u0 v 0 tot (right) at x/h = 2. Streamlines of the averaged
flow of the reference solution [5] (left) and the hybrid approach (right).
obvious when the value of uτ used for the normalization of each computation
is set to the DNS value (not shown here). The consequence is interesting and
this result could provide a good indication on the manner how to reduce this
unphysical step specific to the hybrid technique and also observed by other
groups. Regarding the prediction of the normal Reynolds stresses (Fig. 1, cen-
ter and right), the modeled part is reduced in comparison with the previous
hybrid EARSM (E–R). This reduction is compensated by an increase of the
resolved contribution in the RANS region for (u0 u0 ) but not for (v 0 v 0 ) and
(w0 w0 ), which leads to an under–estimation of these stresses. The Reynolds
shear stress estimation improves with the non–modeled production term. The
region close to the wall is now perfectly predicted and the magnitude of the
kink observed around the interface location is weakened. Furthermore, the
curves at this kink location are smoothened. kmod is naturally that quantity
which is mostly affected by the implementation of the production term. Here
the drop of kmod is significant. The peak value of kmod is almost halved for
E–Rnp compared to E–R. This behavior is not desired since the objective is
to obtain a mainly modeled contribution in RANS mode rather than an in-
creased resolved one. However, the compensation by kres is not large enough,
which provokes the present under–estimation of the total kinetic energy ktot .
the DES and the highly–resolved reference LES data [5]. The mean velocity
U/Ub is plotted at the positions x/h = 0.5, 2 and 6, which represent the
beginning and the center of the recirculation region and the flow after reat-
tachment, respectively. The mean velocity U/Ub is well reproduced by the
hybrid model at each position. However, concerning the wall shear stress dis-
tribution τw , the hybrid simulation shows at the hill crest larger deviations
(also noticed in RANS simulations) than the DES, which is in good agreement
with the reference LES. Regarding the recirculation region the LES applying
the Smagorinsky model predicts the best separation point (x/h = 0.201) with
respect to the reference LES (x/h = 0.19). Whereas the DES under–predicts
this location (x/h = 0.173) but stays in close agreement with the reference,
the hybrid simulation (x/h = 0.254) delays this point marginally. The hybrid
version (x/h = 4.751) and the DES (x/h = 5.197) show a slightly delayed
reattachment compared to the reference (x/h = 4.694). It is interesting that
the hybrid approach shows a better prediction of the reattachment location
than the DES and even the LES (x/h = 4.547). In comparison with LES and
DES, which show deviations of −3.1% and +10.7%, respectively, the hybrid
approach leads to a deviation of +1.2% only. Fig. 2 depicts plots of the aver-
aged streamlines for the hybrid technique and the reference LES. The center
of the recirculation zone is predicted similarly by the hybrid version and the
DES, but in comparison with the reference data, this location is slightly shifted
backward. However, both hybrid and DES still offer satisfactory results.
In general, the Reynolds stresses are well predicted by the hybrid approach
(Fig. 2). The profiles are overall recovered with respect to the reference LES.
Regarding the intensity, only slight under– or over–estimations are observed.
However, at the position x/h = 0.5 the hybrid technique shows an under–
estimation in the prediction of the peaks located at the vicinity of the lower
wall for v 0 v 0 tot and u0 v 0 tot . Other simulations, not mentioned here, show that
these under–estimations result from a lack of resolved scales in this region. In
conclusion, the hybrid simulation gives encouraging statistical results similar
(or in some cases better) than DES. Nevertheless, the hybrid method sug-
gested can be seen as superior in the prediction of the recirculation region.
In the next step of this study, the hill flow test case has to be considered in
combination with the EARSM formulation taking the anisotropy into account.
References
1. Rodi, W., Mansour, N.N., Michelassi, V. (1993). J. Fluids Eng. 115: 196–205.
2. Breuer, M., Jaffrézic, B., Arora, K. (2005). EUROMECH Colloquium 469, LES
of Complex Flows, TU Dresden, Germany, Oct. 6–8, 2005.
3. Wallin, S., Johansson, A.V. (2000). J. Fluid Mech. 403: 89–132.
4. Breuer, M. (1998). Int. J. for Num. Methods in Fluids 28: 1281–1302.
5. Breuer, M. (2005). personal comm., http://www.hy.bv.tum.de/DFG-CNRS/
6. Moser, R.D., Kim, J., Mansour N.N. (1999). Phys. Fluids 11: 943–945.
A locally superconvergent scheme for the
simulation of turbulent flows in complex
geometries
1 Introduction
The present work is part of a research project aimed at developing and validat-
ing numerical codes for the simulation of complex turbulent flows of industrial
or environmental interest. In particular, we focus on such classes of flows for
which the most classically used turbulence approach in engineering applica-
tions, i.e. the Reynolds-averaged Navier-Stokes (RANS) approach, encounters
difficulties in giving accurate predictions. This is the case for flows character-
ized by massive separated regions or by significant unsteady phenomena, as,
for instance, in separated wakes behind bluff bodies. For such classes of flows
a more promising turbulence approach, which has received large attention in
the last decade, is large-eddy simulation (LES).
In our research project we mutually adapted an industrial CFD method-
ology and a LES approach [1] [2]. Our starting point was a numerical solver
for the Navier-Stokes equations in the case of compressible flows and perfect
Newtonian gases, based on a mixed finite-element/finite-volume scheme for-
mulated for unstructured grids made of tetrahedral elements. Finite elements
(P1 type) and finite volumes are used to treat the diffusive and convective
fluxes, respectively. This numerical technology has been extensively used for
RANS simulations of many different flows (see, for instance, [3]).
A significant part of our work concerned the adaption of the numerics
to the LES approach. Indeed, some numerical features which are positive in
a RANS simulation, such as, for instance, numerical viscosity, may become
dangerous in LES. In our opinion, the most critical point is indeed numerical
viscosity and we proposed [2] a MUSCL scheme in which stabilization is ob-
tained through numerical diffusion based on sixth-order derivatives, which has
a dissipative effect much more localized on high frequencies than the one of
stabilizations based on second-order derivatives. In this way we can reduce the
interaction between, from one hand, numerical dissipation, which damps in
priority the highest frequencies, in particular those for which the phase error
is too large and can produce oscillations, and from the other hand, the LES
closure model, which should reproduce the effects of unresolved frequencies on
the resolved ones. Moreover, a key coefficient (γ) permits to tune numerical
dissipation to the smallest amount required to stabilize the simulation.
In the present study, we push forward this concept and we investigate
the application of a locally superconvergent scheme, that is a scheme which is
second-order accurate on the whole unstructured mesh but enjoys accuracy up
to 5th order in regions where the grid is cartesian. This is achieved by applying
the linear reconstruction to flux functions rather than to flow variables.
2 Numerical method
6
5
5
4
4
3
Z
Z
3 2
2 1
6
1 6 0
4
6
0 4
Y
4
0 2 Y
2 2 X
2
X 4
0
6 0 0
(a) (b)
Fig. 1. New finite-volume cells in 3D: (a) division in tetrahedrons, (b) trace of NGC
on a tetrahedron resulting from the previous division.
(among those of the four tetrahedorn faces), which is closest to the center of
the circumscribed sphere. Although the NGC, as well as the MC, can be built
starting from a generic tetrahedrization, it is interesting to consider the case
of a Cartesian mesh, thus, made of rectangle parallelepipeds (thereafter called
bricks), which are cut in a particular way in tetrahedrons, following [5]. This
division splits each brick in six identical tetrahedra, each being the mirror
image of its neighbors (see Fig. 1(a)). Starting from such a tetrahedization,
the NGC cells are bricks, centered around the vertices of the mesh, as can be
seen in Fig. 1b, in which the trace of the division of an element into NGC is
shown.
The convective fluxes are then discretized in terms of fluxes through the
common boundaries shared by neighboring cells. The Roe scheme (with Turkel
preconditioning) represents the basic upwind component for the numerical
evaluation of the convective fluxes F:
F(Wi ) + F(Wj ) 1
ΦR (Wi , Wj , ) = · nij − γs P −1 |P R| (Wj − Wi )
2 2
in which ΦR (Wi , Wj ) is the numerical approximation of the flux between
the i-th and the j-th cells, Wi is the solution vector at node i, nij is the
integrated outward normal to the cell boundary and R(Wi , Wj , nij ) is the
Roe matrix. The matrix P (Wi , Wj ) is the Turkel-type preconditioning term,
introduced to avoid accuracy problems at low Mach numbers [7]. Finally, the
parameter γs , which multiplies the upwind part of the scheme, permits a
direct control of the numerical viscosity, leading to a full upwind scheme for
γs = 1 and to a centered scheme when γs = 0. As well known, the spatial
accuracy of this scheme is only first order. In our previous works, the MUSCL
linear reconstruction method (“Monotone Upwind Schemes for Conservation
Laws”, [8]) was employed to increase the order of accuracy of the Roe scheme.
The basic idea of this approach is to use in the Roe flux computation the
extrapolated values of the flow variables at the cell interface. In [2], a linear
extrapolation was proposed, based on a combination of different approximate
gradients of the variables, which leads to a numerical viscosity proportional
496 M.V. Salvetti, S. Camarri, B. Koobus, A. Dervieux
45&26&78 34%15%67
!# 45698 34587
!"
45&2698 34%1587
456&78 345%67
:;,!,!<=,
2 ,3++-+
1#+2**,*
!%
#
!" !$
!"
:;,!,%%,
!$
!" # #
!" !"
&'()*+,-.,/-0*1 %&'()*+,-+.,/)0
Fig. 2. L2 norm of the error at (a) x = 0.82L; (b) x = 0.2L and AR ' 33. AR is
the maximum aspect ratio of the grid elements.
Table 1. Flow around a square cylinder. Main features of the simulations and
related values of the flow bulk parameters. C d is the mean drag coefficient,Cl0 the
r.m.s. of the lift coefficient and lr the length of the mean recirculation bubble.
tive fluxes are small. Note, however, that, as previously, this effect might be
masked by the influence of the discretization error of diffusive terms, and, in
this case, also by the subgrid scale model.
References
1. Camarri, S., Salvetti, M.V., Koobus, B. and Dervieux, A.: Large-eddy simulation
of a bluff-body flow on unstructured grids. Int. J. Num. Meth. Fluids, 40, 1431–
1460 (2002).
2. Camarri, S., Salvetti, M.V., Koobus, B. and Dervieux, A.: A low-diffusion
MUSCL scheme for LES on unstructured grids. Computers and Fluids, 33, 1101–
1129 (2004).
3. Koobus, B., Farhat, C. and Tran, H.: Computation of unsteady viscous flows
around moving bodies using the k-epsilon turbulence model on unstructured
dynamic grids. Comp. Meth. Appl. Mech., 190(11-12), 1441–1466 (2000).
4. Dervieux, A. and Désidéri, J.A.: Compressible flow solvers using unstructured
grids. INRIA Report, RR-1731 (1992).
5. Gourvitch, N., Rogé, G., Abalakin, I., Dervieux, A. and Kozubskaya, T.: A
tetrahedral-based superconvergent scheme for aeroacoustics. INRIA Report, RR-
5212 (2004).
6. Rodi, W., Ferziger, J.H., Breuer, M. and Pourquie, M.: Status of large eddy
simulation: results of a workshop. ASME J. Fluids Eng., 119, 248–262 (1997).
7. Turkel, E.: Review of preconditioning methods of fluid dynamics. Appl. Num.
Math., 12, 257-284 (1993).
8. van Leer, B.: Towards the ultimate conservative difference scheme, V. A second-
order sequence to Godunov’s method. J. Comp. Physics, 32, 101-136 (1979).
URANS Analysis of Flow-Induced Cavity
Resonances
1 Introduction
The physics of cavity flow oscillation has been the subject of research for
many decades. Rossiter was the first researcher who described an acoustic
feedback mechanism [3]. The flow forms a shear layer which amplifies flow
disturbances and the subsequent scattering of these disturbances into acoustic
waves at the downstream corner. These acoustic waves propagate upstream
and excite further disturbances in the shear layer, creating a feedback loop,
which often leads to self-sustained oscillations at discrete resonant frequencies.
He also introduced a semi-empirical formula for the dimensionless frequencies
(Strouhal number, Str) of this periodic phenomenon:
fL n−γ
Str = = , n = 1, 2, ... (1)
U∞ M + 1/k
with f the frequency, L the length of the cavity, U∞ the free-stream velocity,
n the mode number, M the Mach number, and constants k and γ are 0.57
and 0.25, respectively.
Numerical and experimental studies reveal two distinct types of flow os-
cillations inside the cavity: a shear-layer and a wake mode. Shear-layer mode
2 Computational setup
1
Fluent6.2 , Fluent Inc, 2005
URANS Analysis of Flow-Induced Cavity Resonances 501
a laminar flat-plate Blasius boundary layer along the wall and spanning the
cavity. The ratio L/θ0 is constant and equals 102. The parameter θ0 corre-
sponds to the momentum thickness of the boundary layer at the upstream
edge for the initial condition. Furthermore, the URANS simulations are per-
formed using the simplistic Cartesian multi-block grid. The mesh consists of
two blocks and the grid is stretched around the edges. The cavity block con-
sists of 101 × 51 mesh points and the block above the cavity has 301 × 131
mesh points.
3 Flow results
In order to study the physical structure of flow inside the cavity, an average
over 10 sampled data points in one period of oscillation is performed. Figures 2
(a) and (b) illustrate the time-averaged streamlines of the cavity with M=0.7
using κ-ω and RSM simulations, respectively. The flow pattern displayed in
Figure 2 (a) shows the presence of two vortices in the cavity, though one is
predominant. The streamlines above the cavity are nearly horizontal along
the mouth of the cavity. The presented flow structure for shear-layer mode is
similar to the observations which have been reported by other investigators
[7, 1, 2]. The time-averaged streamlines obtained from RSM simulations are
illustrated in Figure 2 (b) where only one large vortex in the cavity can be
seen. The mean flow pattern is similar to the case when flow converges to
the steady state due to the high dissipation [5]. It can be concluded from
comparison between Figures 2 (a) and (b) that RSM is more dissipative than
κ-ω model. Therefore, it is expected that the fluctuations captured in shear
layer by RSM have smaller amplitudes than when the κ-ω model is used.
4 Frequency analysis
The time signals of primitive variables for each simulation are recorded at
the observation point which is located at the mouth of cavity at x = 0.78L.
502 M. Mesbah, W. Desmet, and M. Baelmans
(a)
(b)
Fig. 2. Mean streamlines of cavity flow with L/D = 4 and M=0.7. (a) κ-ω simula-
tion, (b) RSM simulation,
Fourier analysis of the recorded signals provides the frequency spectrum where
several tonal frequencies and their harmonics become evident.
Figures 3 (a) and (b) show the spectrum of the recorded pressure for M=0.5
using RSM and k − ω models respectively. Oscillation modes corresponding
to the first and second Rossiter modes can be observed. These figures reveal
that the simulation using k − ω model is able to capture the first and second
Rossiter’s modes with the similar amplitudes, while RSM simulation captures
only the second mode.
170 170
Mode 1 Mode 2
160 160
Mode 2
140
SPL (dB)
140
SPL (dB)
130 130
120 120
110 110
100 100
0 0.2 0.4 0.6 0.8 1 1.2 1.4 1.6 1.8 2 0 0.2 0.4 0.6 0.8 1 1.2 1.4 1.6 1.8 2
Str Str
(a) (b)
Fig. 3. Power spectra of pressure at the point located on top of the cavity for
M=0.5. (a)κ-ω model (b) RSM model
The frequencies of the two most dominant peaks in the spectra for all
simulations, as well as predictions from Rossiter’s model [3], LES [1] and DNS
[2] are presented in Figure 4. The obtained results show a good agreement with
Rossiter’s predictions which are based on experiments.
URANS Analysis of Flow-Induced Cavity Resonances 503
0.7
Str= f.L/U
0.6
0.5
0.4
0.3
0.2
0.1
0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9
Mach
An overview of the oscillation modes that are obtained using different ap-
proaches is presented in Table 1. DNS simulations show the wake mode for
M ≥ 0.4. In this oscillatory mode, the frequency of oscillation is not changed
by increase of Mach number. LES simulations illustrate almost the same re-
sults but at the M = 0.4 both shear-layer and wake modes are observed. In
simulations using URANS, the RSM model does not reveal any oscillation
for M < 0.5. For the simulations with M ≥ 0.5 only shear-layer mode is ob-
served. The k − ω model provides an oscillatory flow field for M ≥ 0.4 and like
the RSM model the flow oscillates in the shear-layer mode but with higher
amplitude.
References
1. Rubio, G., De Roeck, W., Baelmans, M., Desmet,W.: Numerical identification of
flow-induced oscillation in Rectangular Cavities using Large Eddy Simulation.
Accepted for publication in the International Journal for Numerical Methods in
Fluids (2006)
2. Rowley, C. W.: Modeling simulation and control of cavity flow oscillations., PhD
Thesis, California Institute of Technology (2002)
3. Rossiter, J. E.: Wind tunnel experiments on the flow over rectangular cavities at
subsonice and transonic speeds. Royal Aircraft Establishment, Technical Report
64037 (1964).
4. Gharib, M., Roshko, A.: The effect of flow oscillation of on cavity drag. JFM
(1987)
5. Sinha, N., Dash, S. M., Chidambaram, N., Findlay, D.: A perspective on the
simulation of cavity aeroacoustics. AIAA paper, 98-0286 (1998)
6. Henderson, J., Badcock, K., Richards, B. E.: Subsonic and transonic transitional
cavity flows. AIAA paper, 2000-1966 (2000)
7. Shieh, C. M., Morris, P.: Parallel computational aeroacoustic simulation of tur-
bulent subsonic cavity flow. AIAA paper, 2000-1914 (2000)
8. Cain, A. B., Rubio, A. D., Bortz, D. M., Banks, H. T., Smith, R.C.: Optimizing
control of open bay acoustics. AIAA paper, 2000-1928 (2000)
Revisiting URANS computations of the flow
behind a backward-facing step using
second moment closures
1 Introduction
&#!$.
,%
'()*+ /0+)*+
&#!$1
"#!$!- ! & &#!$%
"
"#!$% "#!$-%
Mesh 1 2 3 4 5 6 Mesh 7 8 9 10 11 12
fr 1 1.5 2 2.5 3 4 fr 1 1.2 1.4 1.6 1.8 2
Ncell 1396 3396 5970 9842 14234 25154 Ncell 3954 5936 7920 10816 13838 17256
y1+ 150 100 75 60 50 37 y1+ 3 2.5 2.1 1.9 1.6 1.5
Table 1. Mesh characteristics. Left: high-Reynolds models (k − ε and LRR). Right:
low-Reynolds model (EB-RSM).
based on an implicit Euler scheme, and time-step convergence has been care-
fully checked. Space derivatives are approximated by a second-order central-
difference scheme (CDS) and first-order upwind-difference scheme (UDS) for
comparison. Six other meshes, numbered from 1 to 6, are used for the high-
Reynolds number models (k − ε and LRR), mesh 1 being the reference mesh
(see Tab. 1, left). Six meshes, numbered from 7 to 12, are also used for the
low-Reynolds number model (EB-RSM), mesh 7 being the reference mesh (see
Tab. 1, right). fr is the factor of refinement in each direction compared to the
reference mesh. Ncell is the number of cells and y1+ is the distance of the first
point to the wall in wall units. Boundary layer profiles are extracted from
separate calculations and applied at the inlet of the domain in order to match
the boundary layer thickness 4h upstream of the step corner observed in the
experiments.
3 Results
The Elliptic Blending Model significantly improves the skin friction coefficient
Cf (Fig. 2) and the reattachment length lr (Fig. 3) compared to the standard
k − ε and LRR models with wall functions. In Fig. 2, the friction coefficients
obtained with the EB-RSM model with the different meshes are superimposed,
showing that the results are not sensitive to a reasonable coarsening of the
near-wall mesh, which is very interesting for industrial applications. The EB-
RSM mean velocity profiles (Fig. 2) show that the boundary layer downstream
of reattachment recovers more slowly than the experimental data. This feature
is common to most of the turbulence models, SMCs as well as eddy-viscosity
models. All the computations with the k − ε model give steady solutions,
Fig. 2. Left: mean skin friction coefficient. Right: mean velocity profile. Symbols:
experiment [4]; lines: EB-RSM calculation.
508 A. Fadai-Ghotbi, R. Manceau, and J. Borée
Fig. 3. Top left: mean reattachment length. Bottom left: isocontours of vertical re-
solved velocity with mesh 10 and EB-RSM model (dashed lines are negative values).
Right: maximum value of the ratio resolved energy/total energy as a function of the
streamwise location (EB-RSM calculation).
whatever the mesh or the space discretization used. On the contrary, the
behaviour of the SMCs is strongly dependent on numerics: using a UDS,
whatever the meshes and the model, the solution is steady; on the contrary,
using a CDS and meshes similar to those used by Lasher & Taulbee [12], the
solution is unsteady, exhibiting a single frequency. This behaviour shows the
significant influence of numerical diffusion in unsteady flows, consistent with
the conclusions drawn by Lasher & Taulbee. The correct order of magnitude
(St = 0.2) of the Strouhal number is found with both SMCs: St = 0.21 (EB-
RSM) and St = 0.16 (LRR). This frequency is associated with the vortex
shedding which can be visualized in Fig. 3. These large-scale structures are
convected downstream at 0.5U0 as reported in the experiment. The capture of
these structures by the recirculation zone, observed in the experiments, is not
reproduced: consequently, the flapping motion, which is linked to the capture
and release of the shed vortices [11] is missed.
Moreover, the resolved, large-scale structures contain a very limited part of
the total fluctuating energy: indeed, the ratio kres /ktot is less than 5%, where
kres is the kinetic energy of the resolved structures and ktot = kres + kmod is
the sum of the resolved and the modelled kinetic energies.
One could expect that refining the mesh would intensify the resolved en-
ergy, by reducing the numerical errors. On the contrary, what is obtained is a
drastic reduction of the resolved energy and eventually a steady solution for
the finest mesh (mesh 12), as seen in Fig. 3. This figure shows the stream-
wise evolution of M (x) for each mesh used in the EB-RSM calculation, where
M (x) = M axy kres (x, y)/ktot (x, y) . (The results with the LRR model, not
shown here, are qualitatively similar).
The origin of this behaviour can be traced to the excitation of the most
unstable mode of the shear layer by numerical errors at the step corner.
Indeed, it is well known, at least for a 1D convection/diffusion equation,
that central differencing can generate numerical oscillations if the Peclet
number exceeds 2 [8]. The Peclet number compares the relative magni-
tudes of convection and diffusion: in the boundary layer just upstream the
step corner, the convection and diffusion can be considered as essentially
Revisiting URANS computations of the backstep flow using SMCs 509
Fig. 4. Left: streamwise evolution of the local Peclet number at y + = 15. Right:
streamwise evolution of ∂x U at y + = 15.
Conclusion
The Elliptic Blending Model, a recent near-wall second-moment closure, has
been tested on the backstep flow at Re = 37500. The skin friction and the
reattachment length are highly improved compared to the standard k − ε
or LRR models with wall-functions, and are not sensitive to a reasonable
coarsening of the near-wall mesh. The EB-RSM is much more robust than
previous models based on Elliptic Relaxation method ([7],[15]) and can be
510 A. Fadai-Ghotbi, R. Manceau, and J. Borée
References
1. F. Archambeau, N. Méchitoua, and M. Sakiz. Code Saturne: A finite volume
code for the computation of turbulent incompressible flows - Industrial applica-
tions. Int. J. on Finite Volume, Electronical edition: http://averoes.math.univ-
paris13.fr/html, ISSN 1634(0655), 2004.
2. G. Bosch and W. Rodi. Simulation of vortex shedding past a square cylinder
with different turbulence models. Int. J. Num. Meth. Fluids, 28:601–616, 1998.
3. S. Carpy and R. Manceau. Turbulence modelling of statistically periodic flows
: the case of the synthetic jet. In Proc. 6th Int. Symp. Engng. Turb. Modelling
and Measurements, Sardinia, Italy, 2005.
4. D. Driver and H.L. Seegmiller. Features of a reattaching turbulent shear layer
in divergent channel flow. AIAA J., 23:163–171, 1985.
5. D.M. Driver, H.L. Seegmiller, and J.G. Marvin. Time dependent behavior of a
reattaching shear layer. AIAA J., 25:914–919, 1987.
6. P. A. Durbin. Near-wall turbulence closure modeling without “damping func-
tions”. Theoret. Comput. Fluid Dynamics, 3:1–13, 1991.
7. P. A. Durbin. A Reynolds stress model for near-wall turbulence. J. Fluid Mech.,
249:465–498, 1993.
8. J. H. Ferziger and M. Perić. Comp. meth. for fluid dynamics. Springer, 1996.
9. G. Iaccarino, A. Ooi, P. Durbin, and M. Behnia. Reynolds averaged simulation
of unsteady separated flow. Int. J. Heat & Fl. Flow, 24:147–156, 2003.
10. S. Johansson, L. Davidson, and E. Olsson. Numerical simulation of vortex shed-
ding past triangular cylinders at high Reynolds number using a k−ε turbulence
model. Int. J. Num. Meth. Fluids, 16(10):859–878, 1993.
11. M. Kiya and K. Sasaki. Structure of a turbulent seperation bubble. J. Fluid
Mech., 137:83–113, 1983.
12. W.C. Lasher and D.B. Taulbee. On the computation of turbulent backstep
flow. Int. J. Heat and Fluid Flow, 13:30–40, 1992.
13. B. E. Launder, G. J. Reece, and W. Rodi. Progress in the development of a
Reynolds-stress turbulence closure. J. Fluid Mech., 68(3):537–566, 1975.
14. R. Manceau. An improved version of the Elliptic Blending Model. Application
to non-rotating and rotating channel flows. In In Proc. 4th Int. Symp. Turb.
Shear Flow Phenomena, Williamsburg, VA, USA, 2005.
15. R. Manceau and K. Hanjalić. Elliptic Blending Model: a new near-wall
Reynolds-stress turbulence closure. Phys. Fluids, 14(2):744–754, 2002.
Contribution of Turbulence Equation Terms to
the Shear Stress Balance
1 Introduction
From the modelling point of view there are mainly two benefits from Menter’s
Shear Stress Transport (SST) turbulence model [1] as compared to other two-
equation models: blending from ω- to ε-equation while moving away from the
wall and the application of the Bradshaw assumption in boundary layers. The
former reduces the dependency on the prescribed free stream turbulence val-
ues while the latter prevents the overprediction of eddy viscosity in adverse
pressure gradient flows by assuming that the shear stress does not increase
more rapidly than the turbulence kinetic energy. On the other hand, both fea-
tures strongly rely on the wall distance for determining their exertion regions.
This can pose a problem in complex 3D boundary layers of large elongation.
For example, the above mentioned regions only partly cover the turbomachin-
ery side wall boundary layers, as will be shown in section 3. Hence a criterion
is proposed which locally determines the blending functions by balancing the
terms of the k-equation. This is considered to be a more physical criterion as
compared to the wall distance, since it represents a charcteristic condition of
the wall bounded layers. Consequently, an improved balancing of shear stress
can be expected. Below the new approach is denoted as the BSS model, which
is an abbreviation for Balance of Shear Stress.
2 Modelling Approach
The equations for turbulent kinetic energy k and dissipation rate ω of the BSS
model are formally the same as the SST model by Menter [1]:
∂ (%k) ∂ (%Ui k) ∗ ∂ ∂k
+ = P̃k − β %kω + (µ + σk µT )
| ∂t
{z } | ∂x {zi }
| {z } ∂xi ∂xi (1)
|{z}
| {z }
klocal kconv kprod kdest kdif f
∂ (%ω) ∂ (%Ui ω) 2 ∂ ∂ω
+ = 2α%Sij Sij − β%ω + (µ + σω µT ) +
∂t ∂xi ∂xi ∂xi (2)
1 ∂k ∂ω
2 (1 − F1 ) %σw2
ω ∂xi ∂xi
To avoid anomalous turbulence production in stagnation regions, the produc-
tion term of k-equation is limited:
∂Ui ∂Ui ∂Uj
P̃k = min µT + , 10β ∗ %kω (3)
∂xj ∂xj ∂xi
Furthermore, the eddy viscosity µT , containing the Bradshaw assumption, is
defined as follows:
a1 %k
µT = p with: a1 = 0.31 (4)
max a1 ω, F2 2Sij Sij
The constants are blended from the corresponding constants from the k-ε and
k-ω model, e.g. α = α1 F1 +α2 (1−F1 ). Following constants have been applied:
α1 = 5/9, β1 = 3/40, σk1 = 0.85, σω1 = 0.5, α2 = 0.44, β2 = 0.0828, σk2 = 1,
σω2 = 0.856 and β ∗ = 0.09.
Instead of using the wall distance as proposed by Menter [1] the blending
functions F1 and F2 in Eqs. (2) and (4) are determined by weighting the terms
of the k–equation (1). The following instructions are applied subsequently.
First of all, F1 is determined:
QS1 − 0.5 |kprod |
F1 = with: QS1 = max min , 4.0 , 0.5 (5)
3.5 |kconv | + |kdif f |
This approach is based on the idea that the source terms dominate near the
wall and transport terms prevail far away from wall.
Secondly, F2 is the maximum value of F1 and a function which depends
on the ratio of turbulence production to destruction:
−0.9 QS2 −0.2
F2 = max F1 , QP0.2 · 0.3
h i
|k |
with: QP = max min |kprod dest | , 1.1 , 0.9 (6)
h i
|k |
and: QS2 = max min |kconv prod |+|kdif f | , 0.5 , 0.2
QP typically exceeds 1 in the wake region of the boundary layer, and assures
F2 = 1. This is usually the region where the Bradshaw assumption is effective
and prevents the overproduction of eddy viscosity. Furthermore, the term
Contribution of Turbulence Equation Terms to the Shear Stress Balance 513
√ p
10(r∗ − 1)κ Ui Ui
∗ 2Sij Sij
F2 = F2 · 1 − max ,0 for: r = p > 1(8)
βω ω 2Ωij Ωij
According to Kovzulović and Röber [3] the regions of curvature-induced turbu-
lence amplification can be identified by r∗ > 1, using the following definitions
of strain Sij and vorticity Ωij :
1 ∂Ui ∂Uj 1 ∂Ui ∂Uj
Sij = + and: Ωij = − (9)
2 ∂xj ∂xi 2 ∂xj ∂xi
Furthermore, the streamline curvature κ is determined by deriving the tangent
unit vector T in flow direction s:
s 2 2 2
dT ∂Tx ∂Ty ∂Tz
κ= = + + (10)
ds ∂s ∂s ∂s
Following transformation yields dT/ds:
∂T ∂T ∂T
x x x
∂x ∂y ∂z Tx
dT ∂T y ∂Ty ∂Ty
U
= ∂x ∂y ∂z Ty with: T= (11)
ds ∂Tz ∂Tz ∂Tz |U|
Tz
∂x ∂y ∂z
3 3
y/! a) k-" y/! b) BSS
"FS =108 Hz "FS =108 Hz
2 "FS =102 Hz 2 "FS =102 Hz
1 1
0 0
0 20 40 60 µ /µ 80 0 20 40 60 µ /µ 80
T T
0.10
y [m] Exp.
SST
0.08 BSS
k-!
0.06
T5 T6
0.04
T1 T3
0.02
0.00
0.000 0.002 0.004 0.006 0.008 0.010
2
-u’v’/Ue [-]
1.8 0.90
" [-] !is [-]
1.7 0.85
!is
1.6 0.80
1.5 0.75
"
Exp.
1.4 SST 0.70
BSS
1.3 0.65
14.0 14.5 15.0 15.5 16.0 16.5 17.0
m’ [kg/s]
SST BSS
blade blade
µT/µ: 0 200 400 600 800 1000 µT/µ: 0 200 400 600 800 1000
SST BSS
lines = const. µT
lines = const. µT (cf. upper plot)
(cf. upper plot)
Fig. 4. Eddy viscosity µT and blending function F1 in the casing boundary layer
of the rotor, xax /lax = const. = 0.7, ṁ = 16 kg/s
of 240 000 nodes and the dimensionless wall distance of the wall-adjacent
nodes is y + ≈ 50. Overall, the predicted pressure ratios and efficiencies for
SST and BSS model agree very well with the measurement at higher mass
flow rates. However, the SST model does not yield converged solutions when
moving towards the lower mass flow range. In contrast, the BSS model is
able to predict stable operating points closer to the stall line, although the
gap to the experimentally acquired stall line is still large. The authors hope to
considerably reduce this gap by including the effects of streamline curvature [3]
to the BSS model. Investigation of Fig. 4 reveals that the blending function F1
of the SST model only partly covers the tip vortex region, which is denoted by
high values of eddy viscosity. Hence, the switching between ω- and ε-equation
starts already in the vortex core region. On the other hand, the BSS model
adopts the blending to the turbulence structure itself, which is the reason
for slightly better predictions of the BSS model at lower mass flow rates.
516 Dragan Kovzulović and Thomas Röber
Besides the side wall region, the boundary layer at the blade surface also
shows reasonable blending.
4 Conclusions
The terms of the k-equation are used for the blending of the SST turbulence
model, leading to improved balancing of shear stresses in complex 3D bound-
ary layers, as demonstrated on a simulation of a transonic compressor. In
addition, the computation of the wall distance is not required. The BSS ap-
proach shows much less dependency on the free stream values and improved
turbulence predictions in adverse pressure gradient flows as compared to the
k-ω model. Future work will focus on the refinement of blending functions,
on application to multistage turbomachines and on the interaction with the
effects of streamline curvature.
References
1. Menter, F.R., Kuntz, M., Langtry, R.: Ten years of industrial experience with the
SST turbulence model. In: Hanjalić, K. (ed) Proc. 4th. Int. Symp. on Turbulence,
Heat and Mass Transfer. Begell House, New York (2003)
2. Wilcox, D.C.: Reassessment of the scale-determining equation for advanced tur-
bulence models. AIAA Journal, 26, 1299–1310 (1988)
3. Kovzulović, D., Röber, T.: Modelling the Streamline Curvature Effects in Tur-
bomachinery Flows. Proc. of ASME Turbo Expo, Paper No. ASME–GT2006–
90265, Barcelona, Spain (2006)
4. Yang, H., Nürnberger, D., Kersken, H.-P.: Toward excellence in turbomachin-
ery computational fluid dynamics: A hybrid structured-unstructured Reynolds-
averaged Navier-Stokes solver. J. Turbomach., 128, 390–402 (2006)
5. Samuel, A.E., Joubert, P.N.: A boundary layer developing in an increasingly
adverse pressure gradient. J. Fluid Mech., 66, 481–505 (1974)
6. Müller, M.: Private Communication. Technical University Darmstadt, Depart-
ment of Gas Turbines and Aerospace Propulsion, Darmstadt (2006)
7. Blaha, C., Hennecke, D.K., Fritsch, G., Hoeger, M., Beversdorf, M.: Laser-2-
Focus Measurements in a Transonic Compressor Blisk-Rotor and Comparison
with 3D Numerical Simulations. Proc. 13th Int. Symp. on Air Breathing Engines,
Paper No. ISABE 97–7069, Chattanooga, USA (1997)
Part VI
tation into equation (1) yields the unsteady residual, Rˆ∗ k = i kV ŵk + R̂k . The
nonlinearity of the unsteady residual stems from the spatial operator which
can be computed using a Fast Fourier Transform (FFT) algorithm. Consistent
with the time accurate approach, a pseudo-time derivative can be added, and
a time-stepping scheme can be employed to numerically integrate the resulting
equations; V ∂∂τ ŵk
+ Rˆ∗ k = 0. In the NLFD case, an unsteady residual exists
for each wavenumber used in the solution and the pseudo-time derivative acts
as a gradient to drive the absolute value of all of these components to zero
simultaneously.
The NLFD discrete adjoint equation is developed using a pseudo-spectral
approach similar to the one applied to the flow equations, which can be as
expressed as V ∂ψ ∂t + R(ψ) = 0, where R(ψ) is the sum of all the spatial
operators, both convective and dissipative, used in the discretized adjoint
equations. Next, we assume that the adjoint variable and spatial operator can
be expressed as a Fourier series. The derivation of the NLFD adjoint then
follows that of the NLFD flow equations. The NLFD adjoint equations are
∗ ∗
expressed as, V ∂∂τ ψ̂k [ = 0, where R(ψ)
+ R(ψ) k
[ = i kV ψ̂k + R(ψ)
k
[ .
k
2 Results
log(Residual)
log(Residual)
0
10
−2
10
−1
10
−3
−2
10
10
−3
10 −4
10
−4
10
200 400 600 800 1000 1200 1400 1600 1800 2000 2200 2400 50 100 150 200 250 300 350 400
Multigrid Cycles Multigrid Cycles
Fig. 1. Viscous NLFD Flow Solver Fig. 2. Viscous NLFD Adjoint Solver
Convergence; M∞ = 0.822, ωr = 0.102 Convergence; M∞ = 0.822, ωr = 0.102
This section documents the results of the redesign of the LANN Wing to re-
duce the time-averaged drag coefficient for a fixed time-averaged lift coefficient
and wing thickness ratio. The simulation was performed at a Mach number,
M∞ = 0.82, reduced frequency, ωr = 0.102, and deflection of ±0.25◦ about
the zero angle of attack. As established in the previous section, five time steps
per period will be used for the redesign of the LANN Wing. During the ini-
tial stage to compute the flow solution, the time-averaged lift and drag are
computed and used as the target lift and objective function.
Figure 7 illustrates the wing surface pressure contour and the initial and
final pressure distributions at three span locations at the 0◦ phase after 50
design cycles. The pressure contour illustrates the severe weakening of the
522 S. Nadarajah and A. Jameson
λ-shock system and this is further validated in the three pressure coefficient
plots at span stations 6.2%, 49.2%, and 92.3% (initial pressure-dotted line;
final design-solid). The mean angle of attack was perturbed from the initial
zero degrees to 0.344◦ to maintain the time-averaged lift coefficient at 0.348.
The time-averaged drag coefficient reduced by 5.65% from 0.0115 to 0.01085
within 50 design cycles.
Figure 8 illustrates the gradients of the objective function for various tem-
poral modes and demonstrates the need for at least five time steps per period
to capture accurate gradients. Lastly, Fig. 9 presents the convergence of the
time-averaged drag coefficient from 115 drag counts to 108.5.
3 Conclusion
Five time steps per period is sufficient to resolve the unsteady viscous tran-
sonic flow field via the NLFD approach. A redesign of the LANN Wing has
been demonstrated with a reduction of the time-averaged drag coefficient by
5.65% while maintaning the time-averaged lift coefficient constant. Accurate
gradients are achieved with five time steps per period as well. These results fur-
ther demonstrate the potential of the method to provide significant improve-
ments to more realistic problems such as helicopter rotors, turbomachinery,
and other unsteady devices operating in the transonic regime.
References
1. A. Jameson. Computational aerodynamics for aircraft design. Science, 245:361–
371, 1989.
2. S. Nadarajah and A. Jameson. Optimal control of unsteady flows using a time
accurate method. AIAA paper 2002-5436, 9th. AIAA/ISSMO Symposium on
Multidisciplinary Analysis and Optimization Conference, Atlanta, GA, Septem-
ber 2002.
3. S. Nadarajah, M. McMullen, and A. Jameson. Optimum shape design for un-
steady flows using time accurate and non-linear frequency domain methods.
AIAA paper 2003-3875, 16th. Computational Fluid Dynamics Conference, Or-
lando, FL, June 2003.
4. J. Reuther, A. Jameson, J. Farmer, L. Martinelli, and D. Saunders. Aerody-
namic shape optimization of complex aircraft configurations via an adjoint for-
mulation. AIAA paper 96-0094, 34th Aerospace Sciences Meeting and Exhibit,
Reno, Nevada, January 1996.
5. R. J. Zwaan. Data set 9, LANN wing. Pitching oscillation. Technical report.
Agard-R-702 Addendum no. 1., AGARD, Amsterdam, Netherlands, 1985.
Optimum Shape Design for Unsteady Three-Dimensional Viscous Flows 523
0.44 0.0155
NLFD 1−Harmonic
NLFD 2−Harmonic 0.015
0.43
NLFD 3−Harmonic
0.0145
0.42
Drag Coefficient
Lift Coefficient
0.014
0.41 0.0135
0.013
0.4
0.0125
0.39
0.012
0.0115
0.38
0.35 0.4 0.45 0.5 0.55 0.6 0.65 0.7 0.75 0.8 0.35 0.4 0.45 0.5 0.55 0.6 0.65 0.7 0.75 0.8
Angle of Attack Angle of Attack
−1
−1 NLFD 3 Viscous NLFD 3 Viscous
NLFD 5 Viscous NLFD 5 Viscous
NLFD 7 Viscous −0.8 NLFD 7 Viscous
−0.8 NLFD 3 NLFD 3
NLFD 11 NLFD 11
Exp Upper −0.6 Exp Upper
−0.6 Exp Lower Exp Lower
−0.4
−0.4
Pressure Coefficient, Cp
Pressure Coefficient, Cp
−0.2
−0.2
0
0
0.2 0.2
0.4 0.4
0.6 0.6
0.8 0.8
0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 1 0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 1
x/c x/c
◦ ◦
(a) η = 20%, α = αmean = 0.59 (b) η = 65%, α = αmean = 0.59
0 0
NLFD 3 Viscous
0.02 0.05 NLFD 3 Viscous
NLFD 5 Viscous
NLFD 5 Viscous
NLFD 7 Viscous
0.04
0.1 NLFD 7 Viscous
imag(Cp )
real(Cp )
0.15
0.06
0.2
0.08
0.25
0.1
0.3
0.12
0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 1 0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 1
x/c x/c
Fig. 6. Comparison of the Real and Imaginary Components of the First Harmonic
Pressure Coefficient for Various Modes
524 S. Nadarajah and A. Jameson
Cp = -2.0
NLR LANN WING
Mach: 0.822 Alpha: 0.344 IT: 1
CL: 0.332 CD: 0.01512 CM:-0.3034
TIMEAVE CL: 0.348 TIMEAVE CD: 0.01085
Design: 50 Residual: 0.1135E+01 Grid: 257X 65X 49
Cp = -2.0 Cp = -2.0
Fig. 7. Initial and Final Pressure Distribution for Various Span Locations at Phase
= 0 deg
−3
x 10
2 0.0115
1
0.0114
0
−1 0.0113
Gradient Value
−2
Drag Coefficient
0.0112
−3
−4
0.0111
−5
−6 0.011
−7
0.0109
−8
0.0108
20 40 60 80 100 120 140 160 0 5 10 15 20 25 30 35 40 45 50
Grid Point Design Cycles
2 Numerical results
Numerical results for constrained shape optimization are presented. The pur-
pose is to minimize the drag coefficient while keeping the lift coefficient of
the airfoil constant by means of an equality constraint. Also, lower inequal-
ity constraints are enforced on the relative maximum thickness, on the nose
radius and on the trailing edge angle. Unconstrained optimization algorithms
that include the constraints as penalty terms in the objective could be used.
However, their accuracy is known to be poor and moreover they can be ill–
conditioned. Therefore, optimization algorithms capable of handling directly
equality and inequality constraints are necessary. Two algorithms have been
employed: a widely used sequential quadratic programming (SQP) algorithm
and a sequential linear programming (SLP) algorithm known as the method
of centers [8].
1.4
1.4
0.06
1.2 EXACT, [ 49/19 ]
1.2
0.04 1 APPROX 1, [ 44/17 ]
1
0.06
0.04
0.02
-0.02
-0.04
-0.06 EXACT
APPROX 1
APPROX 2
-0.08
0 0.2 0.4 0.6 0.8 1
The flow and the adjoint equations are solved using an implicit pseudo–
time stepping scheme. During the optimization, the two solvers are converged
until the residual norm is reduced about 6 orders of magnitude. A spring
analogy is used to deform the mesh and automatic differentiation in forward
mode has been used to derive the geometric sensitivities, i.e., the derivatives
with respect to the shape parameters. The shape of the airfoil is parameterized
by means of Chebyschev polynomials. Further details can be found in [4, 5].
The computations presented below have been performed on an unstruc-
tured mesh of triangles, see Fig. 1.a, with 12161 nodes, 500 of which are on the
airfoil. The mesh has been deformed in order to be used for all airfoils. In or-
der to make sure that the mesh was capable of capturing weak shocks, a mesh
of 30092 nodes, 700 of which are on the airfoil, has been used to check the
optimization results. The shock–free pressure distributions obtained on the
first mesh after optimization have been re–obtained on the second mesh. The
SQP algorithm from the Matlab Optimization Toolbox4 has been used for the
NACA64A410 and RAE2822 cases whereas the SLP algorithm [5] has been
used for the NACA0012 case. For the two algorithms a tolerance of 10−5 for
both the objective function and the constraints has been adopted as criterion
to stop the optimization.
NACA64A410, M∞ = 0.75 and α = 0 deg angle of attack. The pressure con-
tours for the original airfoil are shown in Fig. 1.b. The airfoils optimized using
the exact and the first two approximations are shock–free, see Figs. 1.c, 1.d
4
Matlab
1984–2004,
c The Mathworks, Inc.
Comparison of Exact and Approximate Discrete Adjoint 529
and 1.e. As can be seen from Fig. 1.g the three airfoils are different. Com-
pared to the airfoil obtained from the exact adjoint, the other two airfoils
show a maximum difference in y–coordinates of 5.6 × 10−3 and 2 × 10−3 re-
spectively. However, they all satisfy the design problem accurately. The airfoil
obtained using the third approximation shows three weak shocks on the upper
side because its optimization has stalled, see Fig. 1.f. In Fig. 1.i the objec-
tive function, which is the drag divided by its initial value, is shown at each
gradient iteration and, in the legend, the ratio of function/flow evaluations to
gradient/adjoint evaluations is shown. The ratio is around 2.6 except for the
third approximation. The latter has exceeded the maximum number, 2000, of
allowed function calls.
RAE2822, M∞ = 0.75 and α = 2 deg angle of attack. The pressure contours
for the original airfoil are shown in Fig. 2.a. As before, the airfoils obtained
from the exact and the first two approximate adjoint codes are shock–free, see
Figs. 2.b, 2.c and 2.d. The geometry of the three airfoils is different, especially
for the first 40% of the chord on the lower side, see Fig. 2.f. Maximum differ-
ence in y–coordinates are of the order of 10−3 . The optimization stalled when
the third approximation was used. The airfoil still exhibits a weak shock, see
Fig. 2.e.
3 Conclusions
It has been shown that the first two approximate adjoint codes considered
in this work can be effective for shape optimization purposes. However, the
SQP algorithm has shown to converge to different airfoils, i.e., different local
minima, for each approximation. Nevertheless, each of these airfoils satisfied
the design problem. The SLP algorithm appeared to be insensitive to the two
approximations and converged to a single airfoil.
The third approximation, which neglects the reconstruction operator, has
shown to be ineffective. For the third approximation, both the SQP and the
SLP algorithms have stalled during the optimization process, failing to con-
verge to a shock–free airfoil.
Acknowledgments: This research was supported by the Dutch Technology
Foundation STW, applied science division of NWO and the technology pro-
gram of the Dutch Ministry of Economic Affairs.
References
1. Barth, T.J.:Analysis of Implicit Local Linearization Techniques for Upwind and
TVD Algorithms. AIAA Paper 87–0595 (1987).
2. Barth, B.J.: Aspects of Unstructured Grids and Finite–Volume Solvers for the
Euler and Navier–Stokes Equations. VKI Lecture Series, 1991-06 (1991).
3. Barth, T.J., Linton S.W.: An Unstructured Mesh Newton Solver for Compress-
ible Fluid Flow and its Parallel Implementation. AIAA Paper 95–0221 (1995).
4. Carpentieri, G., van Tooren, M.J.L., Koren, B.: Improving the Efficiency of Aero-
dynamic Shape Optimization on Unstructured Meshes. AIAA Paper 2006-298
(2006).
5. Carpentieri, G., van Tooren, M.J.L., Koren, B.: Aerodynamic Shape Optimiza-
tion by Means of Sequential Linear Programming Techniques. ECCOMAS CFD
2006 Paper, to appear (2006).
6. Dwight, R.P., Brezillon, J.: Effect of Various Approximations of the Discrete
Adjoint on Gradient–Based Optimization. AIAA Paper 2006-690 (2006).
7. Giles, M.B., Pierce, N.A.: An Introduction to the Adjoint Approach to Design.
Flow, Turbulence and Combustion, 65, 393–415 (2000)
8. Vanderplaats, G.N.: Numerical Optimization Techniques for Engineering Design.
3th Edition, Vanderplaats Reasearch & Development (2001)
Adjoint Sensitivity Computations for an
Embedded-Boundary Cartesian Mesh Method
and CAD Geometry
1 Introduction
Complex geometry remains a challenging issue facing the application of ad-
joint and flow-sensitivity methods in practical engineering design. Among the
most promising approaches for complex-geometry problems is the embedded-
boundary Cartesian mesh method [1]. In this approach, the discretization of
the surface geometry is decoupled from that of the volume mesh, which enables
rapid and robust mesh generation, and ultimately, an automatic analysis of
aerodynamic performance. The purpose of this work is to extend the automa-
tion and efficiency of Cartesian methods to the computation of aerodynamic
sensitivities for shape optimization problems.
The most common way to account for the effect of boundary shape per-
turbations in the adjoint and flow-sensitivity formulations is via domain map-
ping approaches. In implementations on body-fitted structured or unstructured
meshes, this involves the use of mesh-perturbation schemes. For local shape
deformations, the extent of the mesh perturbations can be limited to just
the boundary cells [2, 3, 4]. The approach we propose here for non-body-
fitted Cartesian meshes is similar, but the boundary faces of the volume mesh
maintain their Cartesian orientation as the surface evolves. This approach
permits the computation of mesh sensitivities via a direct linearization of the
boundary-cell geometric constructors of the mesh generator.
3 Shape Sensitivities
The flow equations are discretized on a multilevel Cartesian mesh. The mesh
consists of regular Cartesian hexahedra everywhere, except for a layer of body-
intersecting cells, or cut-cells, adjacent to the boundaries. An infinitesimal
perturbation of the boundary shape affects only the cut-cells. Consequently,
the mesh-sensitivity term ∂M/∂T, which contains the linearization of the
Cartesian-face areas and centroids, volume centroids and the wall normals
and areas with respect to the surface triangulation, is non-zero in these cells.
The crux in the evaluation of ∂M/∂T is the linearization of the geometric con-
structors that define the intersection points between the surface triangulation
and the Cartesian hexahedra.
We explain the salient steps of the linearization using the example shown
in Fig. 1, where a Cartesian hexahedron is split into two cut-cells by the
surface triangulation. We require the linearization of the intersection points
that lie on Cartesian edges, e.g., point A, and also those that lie on triangle
edges, e.g., point B. Focusing on point B, its location along the triangle edge
V0 V1 is given by
Adjoint Sensitivity Computations 533
ro
Inflow
Mi
ri r
!
Outflow
B = V0 + s(V1 − V0 ) (5)
where s denotes the distance fraction of the face location relative to the ver-
tices V0 and V1 . The linearization of this geometric constructor is given by
4 Verification Studies
Supersonic Vortex Problem. We investigate the error convergence rate of
a representative objective function and its gradient on a model problem with
a known analytic solution [6]. The problem involves isentropic flow between
concentric circular arcs at supersonic conditions, as shown in Fig. 2. The ob-
jective function is the integral of pressure along the outer arc, which is similar
to the lift and drag objectives used in aerodynamic design. We compute the
534 M. Nemec and M. J. Aftosmis
-1
ρ Objective (J)
-2 dρ/dMi dJ/dMi
-2
log10(L1 Error)
log10(Error)
-3 -3
-4
Slope=2.04
-4 Slope=2.05
-5
1 1.25 1.5 1.75 2 1 1.25 1.5 1.75 2
0.5*log10(Cells) 0.5*log10(Cells)
(a) Density and density gradient (b) Objective function and gradient
Fig. 3. Error convergence (slopes are computed using the three finest meshes)
gradient and the sensitivities of the flow solution, Eq. 3, with respect to the
inlet Mach number, Mi . The problem is solved on a sequence of five nested
Cartesian meshes. Figure 3 summarizes the results. Figure 3(a) shows the er-
ror convergence rates in the L1 norm of density and its sensitivity to variations
in Mi . The error convergence rate of the objective function and its gradient
is shown in Fig. 3(b). The asymptotic convergence rate of all errors, which is
measured over the three finest meshes, is just slightly over 2. These results
verify the accuracy of the linearization and the convergence of these methods
to the continuous problem.
Shape sensitivities for the NACA 0012 airfoil. The goal of this verifica-
tion test is to study the convergence of gradients for design variables that alter
the shape of the boundary. We consider a subsonic flow over the NACA 0012
airfoil at M∞ = 0.5 and α = 1 deg. We investigate the sensitivity of lift to
the angle of attack using two equivalent approaches. First, we consider the
influence of the angle of attack via a change in the farfield boundary condi-
tions, which does not alter the relationship between the mesh and the airfoil.
We contrast this with an angle of attack change implemented via a rigid-body
rotation of the airfoil about its trailing edge within a fixed mesh. The mesh-
refinement study is performed on a sequence of five nested Cartesian meshes
for each airfoil orientation.
The results are summarized in Fig. 4. Referring to Fig. 4(b), note that the
differences in the gradients between the farfield and rigid-body rotation cases
are decreasing as the mesh is refined. Additional regression analysis of this
data indicates that the rate of convergence is first-order. This is a consequence
of the fact that the mesh perturbations are confined to only the cut-cells.
Unlike the supersonic vortex problem, a perturbation of the angle of attack
via rigid-body rotation modifies the cut-cell boundary and introduces an error
in the objective function proportional to second-order spatial discretization.
Nevertheless, the gradient values for the rigid-body rotation case, even on the
coarsest mesh of roughly 3, 200 cells, are within 1% of the fine-mesh values.
Adjoint Sensitivity Computations 535
6 Conclusions
We have presented an approach for the computation of aerodynamic shape
sensitivities using a discrete formulation on Cartesian meshes with cut-cells at
the wall boundaries. The verification studies show that the convergence rate
of gradients is second-order for design variables that do not alter the bound-
ary shape, and is reduced to first-order for shape design variables. This is a
consequence of confining the mesh sensitivities to the cut-cells. The design
0.142
0.142
Farfield
Shape
0.14 0.14
dCL/dα
Farfield
CL
Shape
0.138 0.138
0.136
0.136
Fig. 4. Convergence of lift and its gradient with respect to angle of attack: imple-
mented via the farfield boundary (“Farfield”) and rigid-body rotation (“Shape”)
536 M. Nemec and M. J. Aftosmis
1e-02
1e-04
1e-06
Objective Function
L2 Gradient
1e-08
1e-10
0 1 2 3 4 5
Design Iterations
1 Introduction
in which {Sq = (xq , yq )}q=0,n and {Bnq (t)}q=0,n corresponds to the Bézier
control points and the Bernstein polynomials of degree n respectively.
Let us consider γ = (y0Γ , · · · , ym
Γ T
) , the ordinates of the shape grid-points,
as control variables and m + 1 parameters tk given (t0 = 0 < t1 < · · · < tm =
1). Then, it clearly appears that the ordinates of the Bézier control points,
Multilevel Bézier gradient method 539
T
i.e. α = (y0 , · · · , yn ) , are directly related with the control variables applying
(2) with y(tk ) = ykΓ for k = 0, · · · , m. More precisely, the following linear
operator from α to γ can be considered:
P : IRn+1 −→ IRm+1
(3)
α 7−→ γ = P (α) = M α in which Mij = Bnj (ti ).
Due to the linearity of P , it seems then possible to define a multilevel approach
in a similar way as done in [1]. Indeed, let us consider the following descent
direction at iteration r and level l:
j (l) (l)
d(l) (l) (l) T
r = M (M ) gr with Mij = Bnl (ti ). (4)
The only difference with respect to the algorithm defined in [1] is the choice of
the matrix M (l) , i.e. the choice of the specific subspace used in the multilevel
process. Nevertheless, it is not so obvious to describe the different levels since
the subparametrisation definition occurs in a less natural way. Indeed, the
abscissae of shape grid-points ({xΓk }k=0,m ) being given, at each level, X (l) =
(l) (l) (l) (l)
(x0 , · · · , xnl )T with nl > nl−1 and T (l) = (t0 , · · · , tm )T should be defined
such that (2) be verified, i.e.:
nl
(l) (l)
X
xΓk = x(tk ) = Bnq l (tk )x(l)
q for k = 0, · · · , m (5)
q=0
Let us also consider the operator P (L) , defined by eq. (3) with n = nL ,
which relates the finest level of Bézier control points and the shape grid-points
ordinates. Then, at a particular level l, one can directly relate the current set
of Bézier control points with the shape grid-points as done in (3) with n = nl ,
or alternatively, apply successively the operators PlL and P (L) . The two ways
should be equivalent from a theoretical view point, even if they can differ in
the implementation. Thus, the following algorithm can be used instead of eqs.
(1) and (4):
γr+1 = γr − ωr M (L) DL L T
l (Dl ) (M
(L) T
) gr (7)
In (7), even if the degree elevation explicitly appears through DL
l ,
the control
variables are still the ordinates of the shape grid-points. Nevertheless, due
to the linearity of the degree elevation procedure, it can be also envisaged
to directly use the Bézier control points as design variables in a multilevel
gradient-based approach. More precisely:
αr+1 = αr − ωr DL
l (Dl ) Gr
L T
(8)
in which Gr is the gradient with respect to α = (y0 , · · · , ynL )T at iteration r.
As soon as αr+1 is obtained, the shape grid-points can be also updated
by γr+1 = P (L) αr+1 . Then, starting from (8) and using the linearity of the
operator P (L) , the following expression is achieved for γr+1 :
γr+1 = γr − ωr M (L) DL
l (Dl ) Gr
L T
with Gr ∈ IRnL +1 (9)
But, since Gr can be related with the gradient with respect to γ = (y0Γ , · · · , ym
Γ T
)
by Gr = (M ) gr , the algorithm defined by (9) coincides with the one de-
(L) T
fined in (7).
Multilevel Bézier gradient method 541
4 Numerical experiments
A first validation of the present optimisation procedure has been done on a 2D
convergent-divergent nozzle inverse problem for inviscid subsonic flows (the
flow is modelled by the Euler equations). The corresponding cost functional is
expressed as a boundary integral which involves the flow variables through the
pressure field. This test-case is a classical one [4] already used for the multi-
level approach associated to grid-point coordinates parametrisation [1, 3]. An
initial constant-section nozzle and a target sine shape are considered while
the computational mesh is composed here of 1900 nodes with, in particular,
63 points on the upper boundary of the throat nozzle, i.e. where the shape
should be optimised.
Fig. 1 summarizes the convergence behaviour for different gradient pre-
conditioning. A reference run is obtained considering as control variables, the
ordinates of 31 shape grid-points among the 63 available. This one level strat-
egy yields a very slow convergence since the cost functional is diminished of
only about one order of magnitude. Indeed, for the same computational cost,
with a multilevel approach based on shape grid-points and a V-cycle strategy
on 4 levels (3, 7, 15 and 31 parameters), a plateau is attained after a decrease
of nine orders of magnitude. Concerning the preconditioning based on Bézier
control points as described in Sec. 3, both one level approach with 10 and 15
parameters respectively and a multilevel approach with a V-cycle strategy on
3 levels (5, 10 and 15 parameters) are considered. The convergence behaviour
of the three approaches is very similar for the decrease of the first five orders
of magnitude. For the one level approach, surprisingly the convergence curve
is better with less parameters but this behaviour inverts asymptotically. The
multilevel approach seems more advantagous only for very small values of the
cost functional with, in particular, an overlap of the plateau obtained with
the shape grid-points multilevel approach.
5 Conclusion
-1
-2 31 param.
shape grid-points
3-7-15-31 param.
log(cost functional)
-3
10 param.
Bezier control points 15 param.
-4
5-10-15 param.
-5
-6
-7
-8
-9
0 5000 10000 15000 20000 25000 30000
iterations
on the one hand the multilevel is particularly well-adapted for shape grid-
points representation since, in this case, the presence of non-smooth shapes is
avoided, and on the other hand, the present test-case probably suffers from its
simplicity. Thus, the validation should be also carried out on other test-cases
in which the Bézier representation could furnish major contribution. Note
that, since the Bézier curves act as a basic tool for polynomial shape repre-
sentation, one can also envisage to extend the formulation to more complex
curves as B-splines, and also, to 3D case through, for instance, tensorial Bézier
parametrisation. Furthermore, since the present approach connects multilevel
gradient-based methods with Bézier hierarchical parametrisation originally
defined for gradient-free approaches [2], another possible development could
be to define a hybrid method, which combines favourable features of different
kinds of optimisation algorithm.
References
1. Beux, F., Dervieux, A.: A hierarchical approach for shape optimization. Engi-
neering Computations, 11, 25–48 (1994)
2. Désidéri, J.-A.: Hierarchical optimum-shape algorithms using embedded Bézier
parameterizations. Numerical Methods for Scientific Computing, Variational
Problems and Applications, Y. Kuznetsov (ed), CIMNE, Barcelona (2003)
3. Held, C., Dervieux, A.: One-shot airfoil optimisation without adjoint. Computers
and Fluids, 31/8, 1015–1049 (2002)
4. Beux F.: Shape optimization of an Euler flow in a nozzle. Notes on Numerical
Fluid Mechanics, 55, Périaux et al. eds, Vieweg Publishing, 115-131 (1994)
High-Fidelity Multi-criteria Aero-structural
Optimisation using Hierarchical Parallel
Evolutionary Algorithms
The wing geometry is represented by three aerofoil sections with sixteen vari-
ables for each section, eight variables that describe the wing planform and
eleven variables to describe the internal structure (Figure 1a and 1b). The
aerofoil geometry is represented by the combination of a mean line and thick-
ness distribution. Both lines are represented by Bézier curves with leading and
trailing edge points fixed at (0.0,0.0) and (1.0,0.0) respectively, and a variable
number of intermediate control points whose x-positions are fixed in advance
and whose y-heights form the problem unknowns. In this case 6 free control
points on the mean line and 10 free control points on the thickness distribu-
tion are taken. The wing planform design variables are indicated in Figures
1a and 1b. The optimisation is subject to various design constraints which are
imposed through several penalties details of which are omitted here.
The two fitness functions to be optimised are defined in Eqn 1 and Eqn 2:
f1 = min(1/(L/D) + penalty) (1)
f2 = min(W ingW eight + penalty) (2)
High-Fidelity Multi-criteria Aero-structural Optimisation using HAPEA 545
An aero structural solver was developed. This solver integrates the two analy-
sis tools for FEA and CFD namely- MSC. NASTRAN and the high order panel
method- PANAIR. The entire aero-structural program is controlled through a
MatlabÂvZ script file. This allows for an easy coupling of the different required
programs as one continuous MatlabÂvZ data structure could be utilized to
define all the information passed between programs.
3.4 Implementation
The present optimisation has been carried out using two approaches; the first
one uses a traditional evolutionary method with a single population model
based upon a fine mesh aerodynamic model of each candidate wing. This
model uses between 1904 and 2992 aerodynamic panels. The second method
uses a hierarchical topology of mesh resolutions comprising three levels. The
same aerodynamic resolution as was used in the single population experiments
was used for the top level; a coarser mesh model used in an intermediate
aerodynamic level (between 1428 and 2244 aerodynamic panels), and an even
coarser mesh (between 952 and 1496 aerodynamic panels) used in the lowest
model to explore the design search space. Three P4, 2.4 GHz machines were
used in the calculation and the population size for this problem was set to 30
in all levels.
546 L. F. González, L. Damp, J. Périaux, and K. Srinivas
3.5 Results
The algorithm was run for 500 function evaluations and took 36 hours to com-
pute. Figure 2a shows the Pareto fronts obtained by using the two approaches.
The Pareto members selected for further investigation are also indicated. By
inspecting Figure 2a and b) it can be seen that the use of a multi-fidelity (hier-
archical) approach gives an overall lower front as compared to a single model
approach. Both approaches give results which lie to the left of the benchmark
value for the MALE wing. The algorithm progress for the hierarchical ap-
proach is shown in Figure 2b. The fitness for all levels decreases with time
and it can be seen that after 350 evaluations, the fitness of each level reaches
a steady state value indicating an optimum solution.
Three wings from the hierarchical optimisation strategy results are taken
for further evaluation (Pareto members 3, 8 and 23) in order to illustrate the
two objective extremes and a compromise between the two results. Pareto
3 has the best aerodynamics, Pareto 23 the best structural response to the
aerodynamic loading and Pareto 8 lies between the two solutions.
Figure 3a shows a top view of the wings selected for further analysis and
Figure 3b compares the baseline aerofoils and the aerofoil sections of each
Pareto member at the root, crank and tip.
A Mach number sweep was performed at a constant angle of attack for
the selected Pareto members and the benchmark. Over the range of Mach
numbers investigated, the Pareto members have lower drag coefficient than
the benchmark wing (results not shown here). The Pareto members three
and eight operate at an almost constant drag value over the range of Mach
numbers. When the coefficient of lift is plotted versus the coefficient of drag
for the selected Pareto members and the benchmark MALE wing in Figure
4a, one can see that the performance of all the wings belonging to the Pareto
front is higher than that of the benchmark wing. The required coefficient of
lift of 0.64 is met at a lower coefficient of drag with all the selected Pareto
High-Fidelity Multi-criteria Aero-structural Optimisation using HAPEA 547
Fig. 3. a) Top views of selected Pareto members b) Selected MALE Pareto member
Root aerofoil sections
Fig. 4. a) Coefficient of Lift vs. Coefficient of Drag for selected MALE Pareto
members and Benchmark. b) MALE Pareto 3 Displacement
member wings. Figure 4b shows one example of the deflected Pareto wings.
The original wing position and internal structure is shown as a wire frame
for comparison. All the MALE wings deflect nonlinearly. A nonlinear analysis
was necessary as the deflections encountered are too large for a Linear Static
analysis and if used, would have yielded incorrect results.
A summary of each of the Pareto and benchmark wing overall character-
istics is shown in Table 3 for comparison. It can be noted that all Pareto
member wings out-performed the benchmark wing on L/D ratio with Pareto
members 8 and 23 had weights which were lower than that of the benchmark
wing while keeping the deflection below 20 percent.
4 Conclusions
Table 1. Summary of Displacement, Strain and Mass for Selected MALE Pareto
members and the Benchmark
Description Pareto 3 Pareto 8 Pareto 23 Benchmark
Deflection % 16 17.3 13.8 12.5
Maximum Strain 0.000396 0.00052 0.000485 0.000403
Mass (kg) 233.735 217.873 185.1 220.9
L/D 43.367 44.4 44.25 23.9
Acknowledgements
References
1. K. Deb. Multi-objective genetic algorithms: Problem difficulties and construction
of test problems. Evolutionary Computation, 7(3):205–230, 1999.
2. N. Hansen and A. Ostermeier. Completely derandomised self-adaption in evo-
lution strategies. In Evolutionary Computation, volume 9, pages 159–195. MIT
Press, 2001.
3. M. Sefrioui and J. Périaux. A hierarchical genetic algorithm using multiple mod-
els for optimization. In M. Schoenauer, K. Deb, G. Rudolph, X. Yao, E. Lutton,
J.J. Merelo, and H-P. Schwefel, editors, Parallel Problem Solving from Nature,
PPSN VI, pages 879–888. Springer, 2000.
4. J. Sobieske and R.T. Haftka. Multidisciplinary aerospace design optimization:
survey of recent developments. In Structural Optimization, v14, pp. 1-23. 1997.
5. E. J. Whitney, M. Sefrioui, K. Srinivas, and J. PuAvSriaux. Advances in hier-
archical, parallel evolutionary algorithms for aerodynamic shape optimisation.
JSME (Japan Society of Mechanical Engineers) International Journal, 45(1),
2002.
Flow Control Optimization Using Neural
Networks and Genetic Algorithms
Evolutionary algorithms have now been used as tool to optimize complex de-
sign spaces in aerospace applications, notably in the areas of Multidisciplinary
Design Optimization (MDO) [4, 2] and flow control [9]. However, in the latter
area a limiting factor has been the cost of evaluating the performance of each
tested flow control configuration. This process is conventionally accomplished
through computational fluid dynamics (CFD) simulation of the design, but
in cases that require a full, viscous Navier-Stokes solver, the computational
demands still can be high even for simple flow control applications like steady
suction or blowing. More challenging flow control applications typically re-
quire even more expensive simulations.
A possible means for accelerating the evaluation process within a search
algorithm is to replace CFD computations with an artificial neural network
(NN). The NN initially requires training and testing for the given problem,
the data for which is provided by CFD simulations. Since neural networks do
not actually solve the Navier-Stokes equations, the proposed optimum per-
formance regions determined by the GA-NN system need to be confirmed.
However, as proposed, the NN approach can replace nearly all of the CFD
computations for a majority of the generations of the GA, dramatically re-
ducing the computational cost. The NN can also be used to test GA methods
with a realistic, but far less costly fitness evaluation, leading to improved GA
design for both NN and CFD evaluations.
1 Numerical Methodology
1.1 Geometry
presented in previous work [8] and are generally satisfactory for the given grid.
For the jet array GA-CFD computations, the total number of grid points is
between 140,000 and 150,000. The computations are done on the structured
two-dimensional code GHOST. GHOST is an in-house, finite volume CFD
code developed at University of Kentucky by P. G. Huang. This code has
been tested extensively and is routinely used for turbulence model validation
and 2D flow control analysis [9, 8]. The baseline case under consideration is
both fully turbulent and at a large angle of attack producing strong separa-
tion if flow control is not applied. As such, these simulations employ Menter’s
two-equation SST turbulence model.
Two real-coded genetic algorithm approaches have been applied in this project
as an objective is to determine the GA with a minimum of fitness evaluations
for cases when relatively expensive CFD computations are required to eval-
uate the fitness functions. The first GA is a Continuous Genetic Algorithm
[5] with elitism. The second GA is the EARND GA with diversity control,
which is detailed in [6, 8]. For the EARND GA, additional modifications have
been made to a standard GA to accelerate convergence: a normal distribu-
tion scheme [11] and dynamic updating of the parameter bounds to eliminate
unnecessary evaluations (computations) in un-promising areas. Finally, diver-
sity control isa dded to the selection function during the initial part (typically
20%) of the evolution to reduce local early convergence.
Fitness evaluation remains the primary cost of both these optimization ap-
proaches, as each evaluation of an individual requires a very costly CFD sim-
ulation. While the computations considered in this project are on the order
of CPU hours, not days, this is still quite expensive when thousands of simu-
lations are required. An alternative to the CFD computation of every fitness
is to use interpolation schemes; however, these must be applied with care as
a GA approach fundamentally assumes a multi-dimensional solution space
with multiple local maximum and minimum that need to be maintained, not
blurred out by poor interpolation. One option is use a non-linear approach
such as a NN. Generally, a NN consists of layers of interconnected nodes, each
node producing a non-linear function of its input [10]. Our neural network ap-
proach is based on the Pyro [1] library implemented in Python. Currently, we
are using one input layer with the optimization variables like suction strength,
suction angle, etc. Then a hidden layer connects this input layer to an output
layer of two nodes representing lift and drag. There exist no general rules on
the structure of the network and we train several different networks with an
accuaracy of three percent. Next, we select the best network with the least
amount of nodes in the hidden layer for the NN-GA approach.
Flow Control Optimization 551
Three parameters for each jet are selected for the search investigation, namely
jet location Lj (measured in percent chord), suction/blowing amplitude A
(measured in term of orifice velocity relative to the inflow freestream velocity),
and suction/blowing angle θ (Fig. 1). The jet width for both suction and
blowing is fixed at 2.5% chord length based on a study by Dannenberg [3].
Note that a negative θ corresponds to suction while positive indicates blowing.
Thus, perpendicular suction is −90o and perpendicular blowing is 90o . The
previous single jet studies [7] have demonstrated that the potential flow control
due to suction tends to dominate that of blowing in this configuration if they
are of comparable magnitudes. To prevent the suction jet from dominating
the genetic search (and thereby significantly reducing the complexity of the
parameter space), the total suction amplitude is limited to be no more than
0.03, or 3% of the inlet velocity for the two jet case, while the net blowing
velocity can reach 0.2. These values correspond to jet momentum coefficients
(Cµ ) of 2.25x10−5 and 0.001, respectively, if applied to a single jet.
0
18
Fig. 1. Jet parameters and flow conditions for the multiple jet CFD simulations
vary in the same ranges as the two jet case. The amplitude of one suction jet
is fixed as before, with the others allowed to vary. The GA used in the four
jet case is the Continuous GA with 48 individuals per generation.
2 Results
While the two jet GA-CFD results are fairly definitive, the four jet results
should be considered tenative as the Continuous GA currently has too low
a rate of mutation, resulting in too many similar values in some parameters,
and only 15 generations have been included, short of a full evolution. Still, the
results suggest that the four jet case actually is favoring placing both suction
jets and one blowing jet near the leading edge in a series, while the remaining
blowing is again located downstream.
The results for the best individual for the two and four jet cases are given
in Table 1. The optimal drag seems to be predicted very well by the NN,
whereas there are differences in the lift prediction.
To illustrate the behavior of the GA-CFD and GA-NN system better, the
fitness for the best 2,500 individuals is compared in Fig. 2.
Fig. 2. Comparison of the fitness function between the CFD-GA approach and the
NN-GA approach for the two and four jet cases
Fig. 3. Comparison of the lift prediction between the CFD-GA approach and the
NN-GA approach for the two and four jet cases
Fig. 4. Comparison of the drag prediction between the CFD-GA approach and the
NN-GA approach for the two and four jet cases
554 Raymond P. LeBeau, Jr., Narendra K. Beliganur and Thomas Hauser
References
1. D.S. Blank, D. Kumar, L. Meeden, and H. Yanco. Pyro: A python-based ver-
satile programming environment for teaching robotics. Journal of Educational
Resources in Computing, 2004.
2. S. Chen, F. Zhang, and M. Khalid. Aerodynamic optimization for a high-lift
airfoil/wing configuration. In 22nd AIAA Applied Aerodynamics Conference,
Providence, RI, number 2004-5078, August 2004.
3. R.E. Danneberg and J.A. Weiberg. Section characteristics of a 10.5-percent
thick aifoil with area suction as affected by chordwise distribution of perme-
ability. NASA Technical Note 2847, Ames Aeronautical Laboratory, Moffet
Field, CA, 1952.
4. L.F. Gonzalez, E.J. Whitney, K. Srinivas, and J. Periaux. Optimum multidis-
ciplinary and multi-objective wing design in cfd using evolutionary techniques.
In International Conference on Computational Fluid Dynamics 3, Toronto,
Canada, July 2004.
5. R.L. Haupt and S.E. Haupt. Practical Genetic Algorithms. John Wiley and
Sons, 2004.
6. L. Huang, , R.P. LeBeau, and T. Hauser. Application of genentic algorithm
to two-jet control system on naca 0012 airfoil. In International Conference on
Computational Fluid Dyanmics 3, Toronto, Canada, July 2004.
7. L. Huang, P.G. Huang, R.P. LeBeau, and T. Hauser. Numerical study of blow-
ing and suction control mechanism on naca0012 airfoil. AIAA Journal of Air-
craft, 41:1005–1013, 2004.
8. L. Huang, P.G. Huang, R.P. LeBeau, and T. Hauser. Optimization of blowing
and suction control on naca 0012 airfoil using a genentic algorithm. In 42nd
AIAA Aerospace Sciences Meeting and Exhibit, Reno, NV, number 2004-0423,
January 2004.
9. R.P. LeBeau, L. Huang, and Th. Hauser. Application of evolutionary algo-
rithms to small jet arrays for flow control. In 17th AIAA Computational Fluid
Dynamics Conference, number 2005-4859, June 2005.
10. D. Michie, D.J. Speigelhalter, and C.C. Taylor. Machine learning, neural and
statistical classification, February 17 1994.
11. A. Oyama, S Obayashi, and K. Nakahashi. Real-coded adaptive range genetic
algorithm and its application to aerodynamic design. JSME International Jour-
nal Series A, 43:124–129, 2000.
Shape Optimization for Dense Gas Flows in
Turbine Cascades
1 Introduction
In recent years, great attention has been paid to a class of fluids of the retro-
grade type (i.e. fluids that superheat when expanded), known as the Bethe–
Zel’dovich–Thompson (BZT) fluids, which exhibit in the vapor phase, above
the upper saturation curve, a region of negative
values of the Fundamental
ρ ∂a
Derivative of Gasdynamics Γ := 1 + a ∂ρ , with ρ the fluid density, a
s
the sound speed, and s the entropy. In the transonic and supersonic regimes,
this leads to nonclassical gasdynamic behaviors, such as expansion shocks and
mixed waves. Moreover, flow discontinuities with jump conditions in the vicin-
ity of the Γ = 0 contour have necessarily limited strength, producing losses
(entropy rise) one order of magnitude lower than usual [2]. The interested
reader may refer to [3] for a review of the complex dynamics of BZT fluids.
An appealing application of BZT fluids is efficiency enhancement for Or-
ganic Rankine Cycles (ORCs). ORCs’ working fluids are heavy organic com-
pounds with large heat capacities: interestingly, several of these fluids possess
BZT properties. One major source of losses in ORC turbines is wave drag,
since they usually operate in the transonic/supersonic regime: the use of a
BZT fluid could avoid shock formation and, ideally, allow isentropic turbine
expansion. Unfortunately, simply utilizing a BZT working fluid is not suffi-
cient to maximize the reduction in losses. Operating the turbine cascade at a
pressure and temperature near the thermodynamic region where BZT effects
appear is also necessary. On the other hand, this region, called the inversion
zone, has a quite limited extent. As a consequence, a reduction in the cascade
pressure ratio is required to operate the turbine entirely within this zone. Now,
it is known from thermodynamic theory that a too small pressure ratio leads
to poor global thermal cycle efficiency. Thus, the development of BZT ORCs
needs finding a reasonable tradeoff between two opposite requirements: on the
one hand, turbine expansion must happen as close as possible to the inversion
zone, on the other, turbine pressure ratio must be sufficiently high for achiev-
ing high global cycle efficiency and power output. Previous studies about
dense gas flows in turbine cascades [4] have shown that, for moderately high
turbine pressure ratios, the use of a BZT working fluid may allow efficiency
improvements up to 3% over air for properly chosen operating conditions. For
higher pressure ratios, efficiency improvements with respect to a perfect gas
progressively disappear, whereas BZT fluids remain advantageous compared
with other working fluids typically employed in Rankine cycles, such as steam
or toluene. Increasing the inlet pressure and Fundamental Derivative along
an isentrope crossing the inversion zone, efficiency tends initially to increase
and reach an optimum. At optimum conditions, the average value of Γ is less
than 1 at the wall, and the sound speed increases with dropping pressure.
This limits Mach number growth during flow expansion and reduces shock
strength. For higher inlet values of Γ , Γinl , efficiency drops again. The higher
the cascade pressure ratio, the higher the Γinl value corresponding to peak
efficiency, and the quicker efficiency drops when moving away from optimum
conditions. Note however, that efficiency improvements observed in previous
studies were simply due to the special nature of the working fluids, since the
blade shapes considered were typical gas turbine blade sections, not specifi-
cally adapted for dense gas flows. The objective of the present study is finding
optimal blade shapes for BZT ORC turbines, providing high efficiency over a
large range of operating conditions and working with high cascade pressure
ratios.
In the present study, dense gas flows are modeled by the compressible Euler
equations for single-phase, non-reacting flows, completed by Martin-Hou real-
istic equation of state [5]. The governing equations are discretized by a finite
volume method for structured multiblock grids, using a third-order accurate
centered spatial approximation [6]. The solution is advanced in time using a
four-stage Runge-Kutta scheme with local time-stepping and implicit residual
smoothing. The solver has been previously validated for a variety of perfect
and dense gas flow simulations.
The flow solver is coupled with a multi-objective genetic algorithm (MOGA).
Genetic algorithms (GA) have proved their interest with respect to gradient-
based methods because of their high flexibility and their ability to find global
optima of multi-modal problem. The MOGA applied in this study is the Non-
Dominated Sorting Algorithm proposed by Srinivas and Deb [7]. For single
objective problems, the algorithm uses an elitism strategy to ensure that the
best individuals survive when algorithm evolves. For multi-objective problems,
a Pareto-based genetic algorithm is applied. The MOGA has been previously
validated for the optimization of airfoil shapes in dense gas flows [8]. In order
to select a proper starting population for the genetic algorithm, the repre-
Shape Optimization for Dense Gas Flows in Turbine Cascades 557
3 Results
The MOGA optimizer, coupled with the flow solver and a structured mesh
generator is applied to shape optimization of turbine cascades. Blade shape
is parametrized using Bezier polynomials, starting from the baseline profile
VKI LS-59 [9]. Sixteen control points are imposed to parametrize the profile.
Globally the optimization problem depends on twenty-five variables. Geomet-
ric constraints are imposed on the maximum blade thickness and on trailing
edge thickness (normalized with the blade chord) that may vary within ±10%
of the corresponding non-dimensional thickness for the baseline geometry. The
turbine cascade is operated with an inlet flow angle of 30◦ and a pressure ra-
tio of 1.82 corresponding, for a perfect diatomic gas, to sonic isentropic exit
conditions. For dense gas flows, inlet thermodynamic conditions, i.e. the ther-
modynamic operation point, should be also specified. All the computations
presented in the following are performed on C-grids composed by 192 × 16
cells, with average non-dimensional height of the closest cells to the wall ap-
proximately equal to 5 × 10−3 . This mesh refinement provides a reasonable
tradeoff between accuracy and computational cost. Since the objective func-
tion chosen for optimization runs is cascade efficiency (defined as real-to-ideal
static enthalpy drop), preliminary computations have been performed, for the
baseline configuration, to investigate the sensitivity of this parameter to mesh
refinement. Results computed on finer grids composed by 382 × 32 (first cell
height 10−3 ) at several flow conditions show that the magnitude of cascade
efficiency tend to increase of about 2∇ · 3% when refining the grid, but trends
of behavior (e.g. efficiency dependency on inlet conditions) are well conserved.
Firstly, a single-objective shape optimization for a diatomic perfect gas
(PFG) has been performed in order to maximize turbine efficiency. The ini-
tial population of the genetic algorithm is selected after a preliminary DOE
over 100 individuals. After about 20 generations the mean and maximum
value of the objective function in the population reach approximately the
same value, indicating that the GA has converged to a population of almost
identical (optimal) individuals. The solution for the baseline blade is char-
acterized by an oblique shock at about 70% the chord and a second shock
attached to the trailing edge (Fig. 2a); computed efficiency is 93.2%. Figure
2b shows the solution after shape optimization: the oblique shock is almost
suppressed, the trailing edge shock is weaker with respect to the reference
case and efficiency grows to 96.2%. Figures 2c,d display Mach number and
pressure (normalized by inlet stagnation conditions) distributions at the wall.
For the optimal configuration the maximum Mach number at the wall and,
consequently, shock strength, is noticeably reduced compared to the baseline
configuration; consequently, wave drag lowers and cascade efficiency improves.
558 Pietro Marco Congedo, Paola Cinnella, and Christophe Corre
0.4 0.4
0.2 0.2
0 0
y
-0.2 -0.2
Fig. 1. Optimal shapes for perfect (a) and dense (b) gas flow.
The optimal individual (Fig. 1a) has a thicker leading edge, thinner trailing
edge, and greater camber than the baseline configuration.
Then, computations are performed with the BZT fluorocarbon PP10 as
the working fluid. Three optimization runs are undertaken. In the first and
second run, a single objective function is maximized, i.e. cascade efficiency
at fixed operating conditions pinl /pc = 1.00, ρinl /ρc = 0.752, Γinl = 0.416
(optimization point OPT1) and pinl /pc = 1.10, ρinl /ρc = 1.09, Γinl = 1.91
(optimization point OPT2), respectively. Subscript c indicates critical point
values. Both operation points lay on an isentrope crossing the inversion zone,
respectively at lower and higher pressure than peak efficiency conditions. In
the third run, both objectives (i.e. efficiencies at OPT1 and at OPT2) are
simultaneously maximized. Here again, a DOE is preliminarily run in order
0.5
0.5
1
1
1.1 1.1
0.5
0.5 0.5
0.5 1.2
1
1.1 1.1
0.5
0.5 1.2
(a) (b)
1
1.4
Optimal
1.2 Baseline 0.8 Pressure
Adimensional Pressure
1
Mach number
0.6
0.8
Suction
Suction
0.6 0.4
0.4
0.2
0.2
Pressure Optimal
Baseline
0 0
0 0.2 0.4 0.6 0.8 1 0 0.2 0.4 0.6 0.8 1
x/c (c) x/c (d)
Fig. 2. PFG flow, isoMach lines: a) baseline; b) optimized. c,d) Wall distributions.
Shape Optimization for Dense Gas Flows in Turbine Cascades 559
0.8 0.8
1.3
0.8
0.75 1.3 0.8
1
1
1.2
1.3
0.8
0.8
1.3 1.2
(a) (b)
1.5 2.5
Baseline Baseline
Pareto 1 Pareto 1
2
Best OPT1 Best OPT1
Fundamental Derivative
Best OPT2 Best OPT2
1 1.5
Mach number
Pressure
Suction
1
0.5 0.5
Suction
Pressure 0
0 -0.5
0 0.2 0.4 0.6 0.8 1 0 0.2 0.4 0.6 0.8 1
x/c (c) x/c (d)
to properly initialize the population of the GA. Figure 1b shows optimal indi-
viduals for operating conditions OPT1 and OPT2 and an individual selected
on the Pareto front resulting from the two-point optimization. The baseline
profile is also represented. For each configuration, a parametric study of cas-
cade efficiency at off-design conditions has been performed. In addition to
operation points OPT1 and OPT2, two test points corresponding to operat-
ing conditions pinl /pc = 1.02, ρinl /ρc = 0.813, Γinl = 0.886 (TEST1) and
pinl /pc = 1.05, ρinl /ρc = 0.944, Γinl = 1.62 (TEST2) are considered. Results
are summarized in Table 1. Note that optimal individuals provide an improve-
ment of 1∇ · 2% only over the baseline at optimization conditions. However,
improvements up to 7% are obtained at conditions TEST1 and TEST2, which
lay close to peak efficiency conditions for the baseline cascade. The best overall
results are obtained for the individual derived from the two-point optimization
(”Pareto 1”). Note that its shape is quite similar to that of the optimal indi-
vidual for perfect gas flow. Figures 3a,b show iso-Mach lines for the baseline
configuration and for individual ”Pareto 1” at operating condition TEST2.
Figures 3c,d show the Mach number and Γ distributions at the wall for the
best individuals from single-objective optimizations, the ”Pareto 1” individ-
ual and the baseline blade. Here again the mechanism leading to efficiency
improvement is a reduction in the maximum Mach number of the flow and
hence of shock strength: major gains come from a significant weakening of the
trailing edge shock for the optimized configuration. This is related to lower
560 Pietro Marco Congedo, Paola Cinnella, and Christophe Corre
values of Γ at the rear part of the suction side, causing slower growth of the
Mach number when the flow re-expands downstream of the first shock.
Table 1. Efficiencies (%) for optimal blade shapes at several operating conditions.
4 Final remarks
Shape optimization for flows of perfect and dense gases in turbine cascades
has been achieved by means of a multi-objective genetic algorithm. For perfect
gas flows, shape optimization allows efficiency improvements of about 3% over
the baseline configuration. For dense gas flows, proper optimization starting
from the same baseline geometry leads to efficiency gains up to 7%. Multipoint
optimization allows improving performance over a large range of thermody-
namic operating conditions. For the high cascade pressure ratio considered in
this study, BZT effects play a minor role in efficiency improvement. The use
of properly designed turbine cascades working with somewhat lower pressure
ratios could allow higher efficiency improvements due to BZT effects, opening
the door to the development of BZT turbines for Organic Rankine Cycles.
References
1. P.A. Thompson: Phys. Fluids 14, 1843 (1971)
2. M.S. Cramer, A. Kluwick: J. Fluid Mech. 142, 9 (1984)
3. M.S. Cramer: Nonclassical dynamics of classical gases. In: Nonclassical waves in
real fluids, ed by A. Kluwick (Springer, Berlin, 1991) pp 91–145
4. P. Cinnella, P.M. Congedo, D. Laforgia: Transonic flows of BZT fluids through
turbine cascades. In: Computational Fluid Dynamics 2004, ed by C. Groth and
D.W. Zingg (Springer, Berlin, 2006). In press.
5. J.J. Martin, Y.C. Hou: A.I.Ch.E. J., 1, 142 (1955)
6. P. Cinnella, P.M. Congedo: AIAA J. 43, 2458 (2005)
7. N. Srinivas, K. Deb: Evol. Comput. 2, 221 (1995)
8. P.M. Congedo, C. Corre, P. Cinnella: Airfoil shape optimization for transonic
flows of Bethe-Zel’dovich-Thompson fluids. AIAA J. (2006), accepted.
9. R. Kiock et al.: ASME J. Eng. Gas Turb. Power 108 : 810 (1996)
Numerical Simulation of Control of Flow Past
a Stalled NACA0015 Airfoil with Plasma-based
Body Forces
1 Introduction
2 Numerical method
The flow around the airfoil is assumed to be two-dimensional.The fluid dynam-
ics is described by the full RANS equations with the k-g turbulence model,
augmented by terms representing the local forcing of the DBD device. They
may be written in non-dimensional form:
∂Q ∂FI ∂GI ∂FV ∂GV
+ + = + +S (1)
∂t ∂x ∂y ∂x ∂y
where Q = {ρ ρu ρv ρe}T , S = {0 FT avex FT avey 0}T is plasma-
based body force, and is obtained from the relation[6]:
∗
FT∗ ave = cα ϑ∗ (ρc ec ) ∆∗T e E∗ δ (2)
where ϑ∗ = 4.2 kHz is the frequency of the applied voltage, ∆∗T e = 67 µs is
an effective duty cycle to recover an effective mean force, cα = 1 is a factor
for collision efficiency, ρc = 1011 cm−3 is the charge density and E∗ is the
electric field. δ is 1 when the magnitude of the electric field is at or above
breakdown level (|E∗ b | = 3 kV mm−1 for air, the plasma exists in this region)
and zero otherwise. Here superscript * denotes dimensional quantity.
Neglecting the influence of charges on the electric potential ϕ∗ , the electric
field is obtained by the Gauss’ law for the electric potential,
∆ϕ∗ = 0 , E∗ = −∇ϕ∗ (3)
The non-dimensionalization of the FT∗ ave is accomplished through the fol-
lowing relation:
FT∗ ave Lref
FT ave = 2 (4)
ρref Uref
where Lref , ρref and Uref are dimensional reference quantities, and set to
chord length, density and velocity of incoming flow respectively.
To simplify the numerical procedure, the influence of the flow on the
plasma is neglected assuming that the low speed flow changes the thermody-
namic state quantities of the fluid trivially and the fluid moves little distance
during a cycle. Equation (3) is solved by implicit central scheme with Gauss-
Sedial technique. The convergence criteria is
∗(n)
i , j , k
X ϕ
max ϕ∗(n+1) − ϕ∗(n) ≤ 10−8 ϕ∗(n) , ϕ∗(n) = (5)
av av Ncell
i,j ,k
method for convective terms, a central scheme for viscous terms, and first-
order backward difference for time derivative term. The implicit discretization
of the spatial derivative terms is adopted to increase the stability, and then
LU–SGS technique is used to solve the algebraic equations.
3 Results
The problem considered is the flow past a NACA0015 airfoil at angles of attack
from 2◦ to 10◦ and unit Reynolds number is 1.9512 × 105 m−1 . The sizes of
the actuators and their locations on the airfoil, and the size of the airfoil are
from reference [8].
Fig. 1. the computational domain (left), mesh (center ) and contour of the electric
field intensity(right)
564 Meiliang Mao, etc..
A O-type mesh is employed. The enlarged meshes are shown in Fig. 2, and
the small panes labeled with number from 1 to 8 denote the locations of eight
actuators.
The simulated cases are listed as following: (1) the influence on the flow
and aerodynamics at different angles of attack caused by the plasma due
to the actuator 1; (2) the influences of flow and aerodynamics at angle of
attack 10◦ with plasmas of 0, 2, 4, 6 and 8 actuators at different locations;
(3) the influence of a single actuator at different location on the flow and
aerodynamics at angle of attack 4◦ . Limited by the length of the paper, only
a few typical results are presented here.
Figure 3 shows that the plasma at the location of pane 1 can damp the sep-
aration at leeside of the airfoil obviously. The separation at 2◦ angle-of-attack
is eliminated, the sizes of separation are reduced and the separation pattern
of the multi-vortex are eliminated or delayed at higher angles of attack.
Figure 4 presents the curve of the normal force vs iteration steps at various
angle-of-attack. The plasma at pane 1 changes the normal force from the
constant oscillation to a certain value at 4◦ angle-of-attack, and from irregular
oscillation to a constant one at 10◦ angle-of-attack.
Figure 5 presents the curves of the peak and the valley of the oscillat-
ing aerodynamics on the airfoil vs angle-of-attack with/without the plasma
at pane 1. It shows that the plasma can reduce the amplitude, and reduce
chordwise force, while increase normal force.
Figure 6 presents the flow around the airfoil at 10◦ angle-of-attack with
the plasma of six actuators at different locations. It shows that the plasma
on the adjacence of the leading edge can damp the separation more efficiently
than the trailing edge.
4 Concluding Remarks
Neglecting the influence of net charge density, the spatial distributions of the
electric field around the actuators have been obtained by solving the potential
equation with several types of far field boundary conditions. The numerical
results show that the influences of the interaction between actuators on the
electric field in the plasma region are trivial and neglected, the plasma-based
body forces acting on the fluid in all cases with various number of actuators
can be obtained simply and conveniently.
566 Meiliang Mao, etc..
The numerical results agree with Roth’s experimental observation that the
plasmas locating on the leading part of the airfoil are more efficient in control
of the airfoil stall than on the trailing part. The oscillation amplitude of aero-
dynamics due to the flow separation over the airfoil is damped or eliminated,
the normal force and the stability increase while the chordwise force decreases
as plasma-based body force increases.
References
1. Roth J.R., etc.: Flow RE-attachment and Acceleration by Paraelectric and
Peristaltic Electrohydrodynamic (EHD) Effects. 41st Aerospace Science Meet-
ing and Exhibit, Reno, Nevada (2003)
2. Post M., etc.: Separation Control on High Angle of Attack Airfoil Using Plasma
Actuators. AIAA Journal, Vol.42, No.11, 2177(2004)
3. G.I. Font, etc.: Plasma Discharges in Atmospheric Pressure Oxygen for Bound-
ary Layer Separation Control. 36th AIAA Plasmadynamics and Lasers Confer-
ence, Toronto, Ontario, Canada (2005)
4. T.J. Sommerer, etc.: Numerical Investigation of the Kinetics and Chemistry of
RF Glow Discharge Plasma Sustained in He, N2, O2, He/N2/O2, He/CF4/O2,
and SiH4/NH3 using a Monte Carlo-fluid hybrid Model. J. App. Phys, 71(4),
1654(1992)
5. Roth, J.R.: Aerodynamics Flow Acceleration Using Paraelectric and Peristaltic
Electrohydrodynamic Effects of a One Atmosphere Uniform Glow Discharge
Plasma. Phys. Plasmas, Vol.10, 2117(2003)
6. W.Shyy, etc.: Modeling of glow discharge-induced fluid dynamics. J. Applied
Phys., Vol.92, No.11, 6434(2002)
7. Y.B.Suzen, etc.: Numerical Simulation of Plasma Based Flow Control Applica-
toins. 35th Fluid Dynamics conference and Exhibit, Toronto, Ontario, Canada
(2005)
8. Roth, J.R.: Subsonic Plasma Aerodynamics Using Paraelectric and peristaltic
electrohydrodynamic Effects. 29th IEEE International Conference on Plasma
Science, Banff, Alberta, Canada (2002)
9. D.V. Gaitonde, etc.: Control of Flow Past a Wing Section with Plasma-based
Body Forces. 36th AIAA Plasmadynamics and Lasers Conference, Toronto,
Ontario, Canada (2005)
Active control of a transitional flow over a
backward-facing step
Introduction
The backward-facing step case study, exhibiting all the essential features of
separated flows, is one of the most studied benchmarks in internal channel
flows. Research on control of the flow in such a configuration is recently de-
velopped. In severeal cases the control is performed using flow through slots
implemented in the step or the lower wall of the channel. The aim of the
present work is to develop a control, based on pulsed inlet velocities on a
two-dimensional backward-facing step for an incompressible transitional flow.
This work follows a previous study [3] on the numerical investigation of the
recirculation areas obtained by an inlet pulsed velocity in a laminar flow.
Here, the impact of the control on several significant characteristics of the
flow are carefully analysed. The numerical simulations are performed using a
Vortex-In-Cell method [1]. VIC methods for viscous flows may be veiwed as an
appealing alternative to pure grid-free vortex methods in simple geometries.
In this case, the use of fast Poisson solvers enables fast velocity evaluations.
Compared with pure finite-difference methods, VIC methods offer the advan-
tage of a robust and accurate treatment of the convective part of equations
with time steps not constrained by convective CFL conditions [2]. Moreover,
the diffusion part of equations are solved by the Random Walk Algorithm
that is well adaptable to the Lagrangian particle transport [5], [6]. In this
paper, after a brief description of the flow configuration and the numerical
scheme’s behaviour and validation, the control procedure and devices are pre-
sented. Then, open-loop and closed-loop control results are studied in order to
achieve an efficient control strategy, both with intrusive (located inside flow)
and non intrusive (wall based) sensors.
Lup Ldown
The last step of the algorithm is to generate new vortex elements on the wall
to cancel the slip velocity and to satisfy the no-slip boundary condition.The
circulation of each vortex is finally distributed on the nodes of the mesh imbed-
ded on rotationnal zones, which allows to recover the vorticity field ω n+1 (i, j)
on the grid, and to repeat the whole procedure to get the variables at the
following time steps. The boundary conditions in vortex methods are based
on stream function Ψ definition related to the vorticity. On the bottom part
of Γwall the no-slip boundary condittion yields to Ψ = 0. On Γin the uniform
velocity is defined as ∂Ψ/∂y = 1. For the top Γwall , Ψ = Ψ (B) = Hup − Hstep
and for Γout , ∂ 2 Ψ/∂x2 = 0.
2 Control approaches
Sensor signal
Energy
0
1
−50
−100 0.5
−150
−200 0
100 200 300 400 500 600 700 0 0.1 0.2 0.3 0.4 0.5
Time Frequency
d2udy2 at sensor located at x=7, y=0 5 Frequency analysis
x 10
150 2
100
50 1.5
Sensor signal
Energy
0
1
−50
−100 0.5
−150
−200 0
100 200 300 400 500 600 700 0 0.1 0.2 0.3 0.4 0.5
Time Frequency
d2udy2 at sensor located at x=8, y=0 5 Frequency analysis
x 10
150 2
100
50 1.5
Sensor signal
Energy
0
1
−50
−100 0.5
−150
−200 0
100 200 300 400 500 600 700 0 0.1 0.2 0.3 0.4 0.5
Time Frequency
2 2
Fig. 2. Signal of ∂ u/∂y (t) at non intrusive sensors on the bottom wall at x = 6
(top), x = 7 (middle) and x = 8 (bottom) and spectral analysis.
ZZ
dx dy
- The fluctuation coefficient, defined as J3(t) = (u0 (t))2 +(v 0 (t))2
;
B 1+ 2
where B is the box [0; 6] × [0; 0.5], and where u0 (t) (resp. v 0 (t)) is the variation
of u(t) (resp. v(t)) around its mean value.
The time-average functionals related to above parameters for a simulation
between t = 320 and t = 600 are noted by Lr , J1, J2 and J3. The objective of
this work is to control these time-average functionals. The main control device
is to replace the uniform inlet velocity by a pulsing inlet velocity from t = 300
to t = 600. The Ψ boundary conditions are fitted to this new inlet condition.
The pulsing frequency and amplitude choice will be defined in the next section.
Lr J1
5 70
65
4
60
3
55
2 50
0 0.1 0.2 0.3 0.4 0 0.1 0.2 0.3 0.4
f f
J2 J3
200 3.2
180 3.18
3.16
160
3.14
140
3.12
120
3.1
0 0.1 0.2 0.3 0.4 0 0.1 0.2 0.3 0.4
f f
Fig. 3. Open-loop control : Functionals versus frequency. Dashed line : uncontrolled
simulation.
the efficiency of the control almost linearly increases with its value.
where M = 2.5E − 03 is chosen to make vary uin (t) in the range of [0.8; 1.2] as
in the previous section, and ∂ 2 u/∂y 2 (t) Sn is the value of ∂ 2 u/∂y 2 (t) related
to a non intrusive sensor Sn located at xSn = 8 and ySn = 0.
where the intrusive sensor Si is located at x = 6 and y = 0.33, and where uSi ,
umin
Si and umax
Si are respectively the average, the minimum and the maximum
values of the uncontrolled velocity uSi (t) computed for the same time and
space location. We set A = 0.2, leading to an inlet velocity uin oscillating
between 0.8 and 1.2. Values of the functionals are reported on Table 1.
Lr J1 J2 J3
Uncontrolled case 4.52 63.53 122.09 3.190
Best value, open-loop 2.09 55.28 186.71 3.137
Closed-loop and non intrusive sensor 2.94 57.41 141.83 3.162
Closed-loop and intrusive sensor 2.38 52.64 159.95 3.150
the intrusive control sensor is located in the transport dominated part of the
flow and then is not perturbated by the near-wall local effects. However the
optimal open-loop control results remain the best (except for J1) compared
to the closed-loop control values. It seems to be due to the noise free property
of uin (t) in the open-loop control.
Conclusion
In this work, both open and closed loop control methods were applied to a
backward-facing step flow, pulsing the inlet velocity conditions. The achieved
results showed that the natural fundamental frequency of the flow offers the
characteristic value to tune the pulsing frequency. The control should there-
fore be based on this value, whatever the used devices. We also observed the
efficiency of implementing the intrusive sensors. This approach improves the
closed-loop control results, reducing the near-wall noisy flow fluctuations.
References
1. Christiansen, J.P. : Numerical Simulation of Hydrodynamics by the Method of
Point Vortices, J. Comp. Phys., 13, 363–379 (1971).
2. Cottet, G.H. and Koumoutsakos, P. : Vortex methods : Theory and Practise,
Cambridge University Press (2000).
3. Creusé, E., Mortazavi, I.: Simulation of Low Reynolds Number Flow Control
over a Backward-Facing Step Using Pulsed Inlet Velocities, AMRX, 4, 133–152
(2004).
4. Ghoniem, A. F. and Gagnon, Y. : Vortex Simulation of Laminar Recirculating
Flow, J. Comp. Phys., 68, 346–377 (1987).
5. Mortazavi, I., Micheau, P. and Giovannini, A. : Etude de la convergence
numérique d’une méthode vortex pour un écoulement à grand nombre de
Reynolds dans un mélangeur, C.R. Mécanique, 330, 409–416 (2002).
6. Sethian, J.A., Ghoniem, A.F. : Validation Study of Vortex Methods, J. Comp.
Phys., 74, 283–317 (1988).
7. Wee, D., Yi, T., Annaswamy, A. and Ghoniem, A.F. : Self-sustained oscilla-
tions and vortex shedding in backward-facing step flows : Simulation and linear
instability analysis, Physics of Fluids, 16, 9, 1–12 (2004).
Synthetic Jet Actuator Modeling for Flow
Control Applications
DIASP - Politecnico di Torino, Corso Duca degli Abruzzi,24, 10129 Torino, Italy,
michele.ferlauto@polito.it, roberto.marsilio@polito.it
Summary. A numerical model of synthetic jet actuator, suitable for flow control
simulations, is presented. The controlled flowfield is simulated by a standard CFD
method for compressible RANS equations, while flow inside the actuator is reduced
to a one-dimensional piston flow. The nonlinear matching between the two systems
ensures conservation of the mass, momentum and energy.
1 Introduction
Recent advances in flow control deals with Synthetic Jet (SJ) manipulators.
This technology has widespread potential applications in aeronautics, with
the aim of enhancing performances and safety margins against flow separa-
tion and stall. With little costs one can, for instance, modify the pressure
distribution over an airfoil at high angle of attack and enhance remarkably
the stall characteristics. The application of such devices has been investigated
for separation control [1], noise or drag reduction and lift enhancement [2],
flow vectoring [3] and more. The related results reveal the great potentials of
flow manipulation via synthetic jets, as well as the need of tools to investigate
and assess adequate control strategies.
From the CFD point of view, one has to deal with unsteady, turbulent
flows in time varying domains. Therefore, the simulation of flowfields actively
controlled via synthetic jets requires accurate computational schemes in both
time and space. The full simulation of the flowfield inside the actuator and
of the main flow is feasible in simple cases and for short integration time. As
alternative, the followed approach is to simulate the presence of actuators by a
boundary condition (BC) that imposes a periodically oscillating velocity pro-
file at the actuator orifice [4]. This modelisation has difficulties in evaluating
the correct pressure level into the actuator since it does not ensure the con-
servation of mass, momentum and energy [5], which is crucial in the suction
phase of the cycle, where pressure is the variable that determines the influx
of mass. More complicated models, based on linear governing equations, also
Main Flow
Piston
fail in capturing flow resonances and nonlinearities [6][7]. For the region far
from the junction between actuator and main flow, a reduced order model has
been suggested in [5] . The approach retains most of the features of the full
simulation with significant CPU savings but a large number of computational
points are still needed to resolve accurately the junction region. This fact rep-
resents a limitation for the application of the model to flow control problems,
involving long time unsteady simulations with one or multiple jet actuators.
In the present paper a nonlinear model of synthetic jet actuator is pro-
posed, based on the domain decomposition approach. The model is a reduction
of [5] suitable for flow control applications. The main flowfield is simulated
using standard CFD technique for compressible, turbulent flows. The flow into
the actuator is modeled as the one-dimensional piston flow, based on the Euler
equation. The interaction at the interface of the two flowfields is evaluated by
solving a Riemann problem. With little CPU expense compared to the BC ap-
proach, the proposed approach ensures the conservation of mass, momentum
and energy across the two systems. Simulations of some flow manipulations
are discussed and compared with available data from literature.
2 Mathematical Model
The main flowfield is simulated using a finite volume discretization of the com-
pressible Reynolds Averaged Navier-Stokes equations, second order in time
and space by using Flux-Difference Splitting and Essentially Non-Oscillatory
schemes. The one-equation model of Spalart-Allmaras is used for the turbu-
lence modeling. The computational domain is truncated to a finite extent in
Synthetic Jet Actuator Modeling for Flow Control Applications 575
Consider the synthetic jet device in Fig. 1 attached, for instance, beneath a
flat plate on which develops a turbulent boundary layer. The synthetic jet
is created at the slot by solving the flow generated by an oscillation piston
moving inside the duct. The motion of the piston is given by Y (t) = A0 sin(ωt)
where ω is the angular frequency of the piston and A0 is the amplitude. The
duct is characterized by is length L and width h. For flow control applications,
the details of the flow inside the actuator are unessential, while the interaction
with the external flow is of relevance. The flow into the actuator is then
modeled as the piston flow, based on the one-dimensional Euler equation. At
the interface the flow state between the two fields is evaluated by solving a
Riemann problem. Within this model, a profile varying along the actuator
orifice is regarded as a sequence of pistons moving synchronously. Similarly,
a vibrating membrane can be simulated by a series of pistons moving with
imposed amplitudes but same frequency (see Fig.1). Viscous effects inside the
piston duct can be accounted for by a distributed loss model.
The numerical solver adopted for the flow in the duct integrates the one-
dimensional, compressible, unsteady Euler equations in time varying domains.
The following transformation of the independent variables is performed
x − b(t)
ξ= , τ =t (1)
c(t) − b(t)
where b(t) and c(t) are the time varying position of the left and right boundary,
respectively. The flow governing equations are rewritten as
∂ U ∂ U ∂ξ F (U ) ∂ξ
+ + =0 (2)
∂τ J ∂ξ J ∂t J ∂x
where
ρ ρu
∂ξ
U = ρu F (U ) = p + ρu2 , J= (3)
∂x
e (e + p)u
and are integrated using a finite volume method fully compliant to that used
for the external flow and having the same accuracy properties. Conventionally,
the left boundary represents the piston wall, moving with velocity b(t) = Y (t).
The right boundary (c(t) = 0) is assumed as the synthetic jet slot, where a
Riemann problem between the internal and the external flow is computed.
576 M. Ferlauto and R. Marsilio
1 1.1
1
0.9 Freq = 20Hz
Freq = 20Hz 0.9 THE ORY
0.8 THEORY Freq = 40 Hz
Freq = 40 Hz 0.8
0.7
0.7
Ucl/Ucl-max
0.6
U/Ucl
0.6
0.5
0.4 0.5
0.3
0.4
0.2
0.1
0 0.3 0 1 2
-2 -1 0 1 2 10 10 10
Y/b X/h
T T T
1.00172 1.00172 1.00172
1.00154 1.00154 1.00154
1.00137 1.00137 1.00137
1.00119 1.00119 1.00119
1.00102 1.00102 1.00102
1.00084 1.00084 1.00084
1.00067 1.00067 1.00067
1.00049 1.00049 1.00049
1.00031 1.00031 1.00031
1.00014 1.00014 1.00014
0.999964 0.999964 0.999964
d e f
T T T
1.00172 1.00172 1.00172
1.00154 1.00154 1.00154
1.00137 1.00137 1.00137
1.00119 1.00119 1.00119
1.00102 1.00102 1.00102
1.00084 1.00084 1.00084
1.00067 1.00067 1.00067
1.00049 1.00049 1.00049
1.00031 1.00031 1.00031
1.00014 1.00014 1.00014
0.999964 0.999964 0.999964
3 Numerical Results
In this section two case of flow manipulation using synthetic jet actuator
are described: the generation of a synthetic jet in a quiescent air and the
interaction of a flat plate boundary layer and the synthetic jet.
The synthetic jet produced in a quiescent external flow has been extensively
studied ([8, 9]). The observation that the far field effects of a turbulent jet
can be reproduced by this type of actuators is the starting point of the whole
research on synthetic jets. At moderate frequencies, during the ejection phase,
the piston pushes the flow out of the actuator and a pair of counter rotating
Synthetic Jet Actuator Modeling for Flow Control Applications 577
0.03
0.025
0.02
0.015
Y
0.01
0.005
-0.005
1 1.1 1.2
X
0.03
0.025
0.02
0.015
Y
0.01
0.005
-0.005
1 1.1 1.2
X
vortices form at the manipulator orifice (Fig. 3). This vortex pair travels in
the external flow by its own induced velocity. When the piston reverts its
motion, it entrains external fluid through the slot. As long as the vortices
have already travel away from the orifice, they are not effected by the motion
of the entrained fluid. During the subsequent ejection phase, a new dipole
is formed into the fluid. The self-similar characteristics of the flow are show
in Fig. (Fig2) where the transverse velocity profiles at different stations, ob-
tained numerically for different frequencies (f1 = 20Hz and f2 = 40Hz) are
compared with the theoretical solution of a turbulent two-dimensional jet.
The decay of the centerline jet velocity also reach the theoretical rate at a
distance of about 10h .
The interaction of a synthetic jet with a flat plate boundary layer is stud-
ied numerically. Due to the lack of availability of experimental data in tran-
sonic, high-Reynolds-number flow regime, the simulations we made are not
suitable to validate rigorously the actuator model. The test-case, fully three-
dimensional, proposed in [7] has been reduced to a two-dimensional configura-
tion while maintaining the main flow parameters. In Fig. 4 two flow patterns
are shown, the upper concerns the blowing phase, the lower the suction phase.
Although obvious discrepancies between the 3D and 2D flow configurations,
the main flow feature on which the test-case focused, that is the time evolution
of the velocity at the actuator orifice centerline Ucl , was recovered (see Fig. 5).
The observed behaviour is due to the nonlinearities. The replacemente of the
actuator by a periodic boundary condition would lead instead to a sinusoidal
evolution in time.
578 M. Ferlauto and R. Marsilio
4 Conclusions
The main outlines of a numerical model of synthetic jet actuator has been
presented. The model aims to overcome drawbacks related to the lack of con-
servativeness of other approaches suitable for flow control simulations. The
significant increase of computational costs of more sophisticate modelisations
is avoided. A preliminary validation has been performed. Rigorous validation
and fine tuning are expected during the experimental testings of the under-
going research project.
Acknowledgments
This work was supported by Regione Piemonte through the grant E40-2004.
References
1. Seifert, A., Darabi, A., Wygnanski, I.: Delay of airfoil stall by periodic excitation.
AIAA J., 33(4), 691-707 (1996).
2. Smith, D., Amitay, M., Kibens, V., Parekh, D., Glezer, A.: Modification of lifting
body aerodynamics using synthetic jet actuators. AIAA Paper 98-0209 (1998).
3. Guo, D., Gary, A.W.: Vectoring Control of a Primary Jet with Synthetic Jets.
AIAA paper 2001–0738 (2001).
4. Donovan, J., Kral, L., Cary, A.: Active flow control applied to an airfoil. AIAA
Paper 98-0210 (1998).
5. Yamaleev, N.K., Carpenter, M.H.: A Reduced-Order Model for Efficient Simu-
lation Synthetic Jets. NASA TM 2003-212664 (2003).
6. Gallas, Q. et al.: Lumped element modeling of piezoelectric-driven synthetic jet
actuators. AIAA paper 2002-0125 (2002).
7. AA.VV. : CFD Validation of Synthetic Jets and Turbulent Separation Control.
Langley Research Center Workshop, Virginia, USA (2004).
8. James, R.D., Jacobs, J.W., Glezer, A.: A round turbulent jet produced by an
oscillating diaphram. Phys. Fluids, 9(8) (1996).
9. Rizzetta, D.P., Visbal, M.R., Stanek, M.J.: Numerical investigation of synthetic
jet flowfields. AIAA Paper 982910 (1998).
Numerical Study of Transonic Drag Reduction
for Flow Past Airfoils Using Active Flow
Control
Summary. The objective of this paper is to evaluate, via numerical simulation, the
feasibility of weakening the shock wave(s) and reducing the pockets of supersonic
regions on airfoils/wings in transonic flight using active flow control (AFC) devices
such as synthetic jet actuators (SJA), and thereby achieving the desired changes
in aerodynamic forces and moments. Most importantly, it is shown that both the
enhancement in lift and reduction in drag can be achieved by using multiple synthetic
jet actuators suitably placed on the upper surface of the airfoil.
1 Introduction
In recent years, there has been a great deal of emphasis toward the develop-
ment of advanced aerodynamic technologies based on the fluidic modification
of aerodynamic and propulsive flow fields that can cover multiple flight condi-
tions without the need for conventional control surfaces such as flaps, spoilers
and variable wing sweep. The fluidic modification (or flow control) can be
accomplished by employing micro-surface effectors or actuators operated by
an intelligent control system. One of the promising device for AFC that has
been explored extensively is the zero net-mass-flux synthetic jet (or oscilla-
tory jet) actuator. With synthetic jet actuator, a variety of impressive flow
control results have been achieved experimentally including the vectoring of
conventional propulsive jets, modification of aerodynamic characteristics of
bluff bodies, control of lift and drag of airfoils, enhanced mixing in circular
jets, reduction of skin-friction of a flat plate boundary layer etc. For subsonic
flow past airfoils at low angles of attack, recently it was shown experimentally
by Amitay et al [1] and computationally by the authors of this paper [2] that
the pressure drag of an airfoil can be significantly reduced with a minimum
change in lift by modification of the apparent aerodynamic shape of the airfoil.
This virtual aerodynamic shape modification can be achieved by creating a
small domain adjacent to the upper airfoil surface (downstream of the leading
edge) using a synthetic jet actuator which displaces the local streamlines suffi-
ciently to modify the local pressure distribution. These results have been very
2 Solution methodology
Table 1. Comparison of the computed lift and drag coefficients for the flow past a
NACA64A010 airfoil at M∞ = 0.8, Rec = 5.67e + 6 and α = 2.8◦
Table 2. Comparison of the computed lift and drag coefficients for the flow past a
NACA64A010 airfoil at M∞ = 0.9, Rec = 5.67e + 6 and α = 2.8◦
number Re = 5.67 10+6, with and without AFC using three synthetic jets
on the upper surface of the airfoil; experiments for this configuration were
performed at Boeing. The three synthetic jets were installed on the upper
surface of the airfoil at x/c = 0.1, 0.53 and 0.68, c being the airfoil chord. The
three jets were identical in width b/c =0.00084 with frequency f = 220 Hz and
velocity amplitude Mj = 0.05, and operated in phase. The grid requirements
were very sensitive to the flow conditions. To maintain y + < 2, a different
computational grid was required for each flow condition M∞ = 0.8 and 0.9.
Figures 4(a) and 4(b) show the close-up view of the two grids used for M∞
= 0.8 and M∞ = 0.9 respectively. Both the grids have a C-H topology with
296 x 61 nodes; the main difference between the two grids is in the cluster-
ing near the airfoil surface. The boundary conditions for the three synthetic
jets are imposed at the airfoil surface without accounting for the cavity. It
is important however that in addition to sinusoidal velocity of the synthetic
jet, the pressure and temperature at the three jet locations (x/c = 0.1, 0.53,
and 0.68) are correctly specified. We employ the pressure and temperature
at these locations calculated for the baseline airfoil (without synthetic jets).
Figures 5(a) and 5(b) show the computed pressure coefficient Cp at M∞ =
0.8 and M∞ = 0.9 respectively. Figures 6(a) and 6(b) respectively show the
pressure drag coefficient without and with AFC, and the pressure lift coeffi-
cient without and with AFC for M∞ = 0.8, while Figures 7(a) and 7(b) show
these quantities for M∞ = 0.9. These figures demonstrate the effectiveness
of AFC using suitably placed multiple synthetic jets, in both reducing the
drag and enhancing the lift. Tables 1 and 2 respectively show the reduction
in drag and increase in lift at Mach 0.8 and 0.9 using AFC. These results are
very close to the experimental values obtained in the experiments performed
at Boeing. These calculations, for the first time, demonstrate the potential of
AFC in transonic drag reduction.
Transonic drag reduction using active control 583
Fig. 1. Numerical results for a NACA 64A010 airfoil: M∞ = 0.8, Rec = 5.67e+6,
α = 0, 1 and 2◦ , b/c = 0.0044, f = 150Hz, Mj = 0.3, xj /c= 0.53.
Fig. 2. Numerical results for a NACA 64A010 airfoil: M∞ = 0.8, Rec = 5.67e+6,
α = 1◦ , b/c = 0.0035, 0.0044, and 0.0055, f = 150Hz, Mj = 0.3, xj /c= 0.53
Fig. 3. Numerical results for a NACA 64A010 airfoil: M∞ = 0.8, Rec = 5.67e+6,
α = 1◦ , b/c = 0.0044, f = 150Hz, Mj = 0.1, 0.2 and 0.3, xj /c= 0.53
584 Jose Vadillo and Ramesh K. Agarwal
Fig. 4. Computational domain and grids used in the simulations for flow past a
NACA 64A010 airfoil with three synthetic jet actuators.
Conclusions
Fig. 6. Comparison of the computed pressure drag coefficient and pressure lift
coefficient, without and with AFC for flow past a NACA 64A010 airfoil at M∞ =
0.8, Rec = 5.67e+6, α = 2.8◦ , b/c = 0.00084, f = 220 Hz, and Mj = 0.05 using three
synthetic jet actuators located at x/c = 0.1, 0.53, and 0.68 on the upper surface of
the airfoil.
Fig. 7. Comparison of the computed pressure drag coefficient and pressure lift
coefficient, without and with AFC for flow past a NACA 64A010 airfoil at M∞ =
0.9, Rec = 5.67e+6, α = 2.8◦ , b/c = 0.00084, f = 220 Hz, and Mj = 0.05 using three
synthetic jet actuators located at x/c = 0.1, 0.53, and 0.68 on the upper surface of
the airfoil.
References
1. Amitay, M., Horvath, M., Michaux, M. and Glezer, A., ”Virtual Aerodynamic
Shape Modification at Low Angles of Attack Using Synthetic Jet Actuators”,
AIAA Paper 01-2975, 2001
2. Vadillo, J.L., Agarwal, R.K., Cary A.W. and Bower, W.W., ”Numerical Study
of Virtual Aerodynamic Shape Modification of an Airfoil Using a Synthetic Jet
Actuator”, AIAA Paper 03-4158, 2003
3. Vadillo, J.L., Agarwal, R.K. and Hassan, A.A., ”Modulation of Aerodynamic
Forces and Moments at Transonic Speeds Using Active Flow Control”, AIAA
Paper 05-0486, 2005
4. Bush, R.H., ”The Production Flow Solver of the NPARC Alliance”, AIAA Paper
88-0935, 1988.
“This page left intentionally blank.”
Part VII
1 Introduction
2 Numerical methods
Given the initial distribution of the level set function and velocity field, the
following incompressible Navier-Stokes equations are to be solved to obtain
an updated velocity and pressure field using the P1P1 splitting finite element
formulation:
Du
ρ(φ) = ∇p + ∇ · (2µ(φ)D) − σκ(φ)δ(φ)∇φ + ρ(φ)g (1)
Dt
Level Set based FEM of Bubble-in-Liquid Simulation 591
where, φ is the level set function and fluid density and viscosity, curvature
of an interface and is represented by a function of φ. The surface tension
effect has been evaluated by using the CSF(continuum surface force) method
proposed by Brackbill et al.
3 Results
3.1 Tank sloshing problem
Tank sloshing problem at a low filling level has been solved using the present
level set FEM and compared with the existing results. Fig. 1 shows the pres-
sure history at a point of the bottom of the tank with time. The present result
agrees well with the experiment.
For the validation of the present FEM code for the bubble-in-liquid simulation,
a rising bubble problem in a quiescent liquid has been solved. The dimension-
less parameters introduced in the present simulation are density
√ ratio(ρg /ρl =
0.001), viscosity ratio(µg /µl = 0.01), Reynolds number(ρl R (gR)/µl ) and
Bond number(ρl R2 /σ). Non-dimensionalized terminal velocities calculated in
1.4
1.2
U
1
Case 2
0.8
0.6
0.4
0.2
0
2 4
t
the present simulation for various Reynolds and Bond numbers are compared
with those obtained by Ryskin and Leal[2], who used a body-fitted coordinate
method in order to find a shape at the equilibrium state from the stress bal-
ance. Their calculation can be considered to be the most accurate since the
curvature along the interface can be estimated most accurately by the body-
fitted coordinate. The results by Son[3] have been obtained by combining the
level set method with VOF. Fig. 2 shows the time history of the terminal
velocity for the case II. It is shown that the simulation reaches a terminal ve-
locity as time goes on. It needs to be noted that the present bubble-in-liquid
simulation method based on FEM and level set method alone gives a satisfac-
tory solution, particularly in terms of mass conservation. It should be noted
that Sussman and Puckett[7] and Son have used a CLSVOF(combined level
set set VOF) in order to circumvent volume conservation problem of level set
method based on finite difference method.
4 Conclusion
(1) The most satisfactory result has been obtained with the Taylor-Galerkin
discretization of the advection and redistancing equations compared to the
SUPG type discretization.
(2) The P1P1 splitting finite element method combined with the Taylor-
Galerkin discretization of the hyperbolic type equations solves some bench-
mark problems well and it is directly applicable to a problem with unstruc-
tured mesh.
(3) The present FEM code based on level set method gives a reasonable so-
lution for the benchmark problem of a rising bubble in a quiescent liquid
compared with the solution of Ryskin and Leal, who used a body-fitted coor-
dinate method.
594 Hyoung G. Choi and Jung Y. Yoo
References
1. Sussman, M., Smereka, P. and Osher, S.: A Level Set Approach for Computing
Solutions to Incompressible Two-Phase Flow. Comput. J. Comput. Phys.,114,
146–159 1994
2. Ryskin, G. and Leal, L.G.: Numerical solution of free-boundary problems in fluid
mechanics. J. Fluid. Mech., 148, 1–17 1984
3. Son, G.: Efficient implementation of a coupled level-set and volume-of-fluid
method for three dimensional incompressible two-phase flows. Numer. Heat.
Transfer. Part B., 43, 549–565 2003
4. Pillapakkam, S.B. and Singh P.: A level-set method for computing solutions to
viscoelastic two-phase flow. J. Comput. Phys., 174, 552–578 (2001)
5. Lin, C., Lee, H., Lee, T. and Weber, L.J.: A level set characteristic Galerkin nite
element method for free surface flows. Int J. Numer. Meth. Fluids, 49, 521–547
(2005)
6. Sung, J.Y., Choi, H.G. and Yoo, J.Y.: Time-accurate computation of unsteady
free surface flows using an ALE segregated equal-order FEM. Comput. Methods.
Appl. Engrg., 190, 1425–1440 (2000)
7. Sussman, M. and Puckett, E.G.: A coupled level set and volume-of-fluid method
for computing 3D and axisymmetric incompressible two-phase flows. J. Comput.
Phys., 162, 301–337 (2000)
Level Set method for Curvature-driven Flows
in Microfluidics
Paul Vigneaux
Summary. A Level Set method for bifluid flows in Microfluidics is presented. After
a stability analysis which conducts to a restrictive stability condition on the compu-
tational time step due to surface tension, we introduce a splitting to shorten simula-
tion times. Numerical results of these curvature-driven flows are in good agreement
with physical experiments and are used to explore mixing processes inside micro-
droplets.
1 Introduction
Over the last decade, manipulations of tiny volumes of fluid have dramati-
cally improved. For instance, microchannels, with a section in the order of 100
square micrometers, are now casted in polymer or etched in silicium. Volume
of fluids are transported inside these channels. This miniaturization allows to
increase experiments in field such as biology, chemistry or pharmacology (viz.
genome sequencing, pharmacological molecules targeting). The Laboratory of
the Future (LOF) - a joint laboratory between Rhodia, Bordeaux 1 Univer-
sity and the CNRS (France) - develops experiments in microchannels in order
to analyse flows of polymers, concentrated surface-actives or emulsions. LOF
also pay a lot of attention to the mixing of these complex fluids. As far as the
Applied Mathematics Laboratory of Bordeaux (MAB) is concerned, we are in-
terested in the numerical dimension of these experiments through simulation.
This allows us in turn to compare the results with the physical studies of LOF.
2 Physical Problem
We are concerned by the simulation of immiscible bifluid flows in microchan-
nels. In such systems, the Reynolds number is lower than one. As a con-
sequence, we can use a quasi-stationary Stokes equation to model the flow.
Furthermore, we can assume that fluids are incompressible. The fact that the
Reynolds number is low could let us to believe that flows are laminar and that
the shape of the structures in the flow (e.g. droplets) is a simple one ; indeed,
this is confirmed by LOF experiments. Nevertheless, these flows are such that
surface tension forces are much stronger than inertia forces which can induce
vortices inside the structures. In addition, the geometries of our applications
are such that we have to deal with moving interfaces with topological changes.
By the way, we want to use a Level Set method [1] [3] which, in this case, is
particularly suited to easily follow interfaces.
3 Numerical Approach
It is the Eulerian nature of the Level Set method which allows to easily handle
topological changes of the interface. This method has been introduced by
Osher and Sethian in [1], for fronts propagating with curvature dependent
speed. It was later used by Sussman et al. [2] for incompressible two-phase
flows. Let us recall that the Level Set method consists in representing the
interface, Γ ⊂ Rd , as the zero level set of a continuous level set function,
φ : Ω ⊂ Rd → R. The function φ is defined everywhere in the domain Ω
whereas Γ = {x ∈ Ω / φ(x) = 0}.
The motion of the interface is obtained by solving the transport equation with
velocity field v :
φt + v.∇φ = 0 (1)
Actually, v is the desired velocity of the underlying problem on the interface
and is arbitrary elsewhere. However, it was showed that a smooth extension
of the interface velocity off the front gives better results.
div(u) = 0 (3)
φt + u.∇φ = 0 (4)
Level Set method for Curvature-driven Flows in Microfluidics 597
The normal and the curvature are simply computed thanks to the level set
function :
∇φ
n= and κ = div(n) (7)
|∇φ|
It is also known, that one has to periodically reinitialize φ to a signed
distance function since, along the computation, if v is not the extension of
the interface velocity, φ does not remain a distance function. This can leads
to severe loss of mass. One way to reinitialize a level set function φ0 to a
signed distance function is to solve the so-called reinitialization equation :
φt + sign(φ0 )(|∇φ| − 1) = 0 with initial condition φ(., t = 0) = φ0 . Since
all the information propagates from the interface, no boundary condition is
needed. Level Set method well lends itself to Microfluidics because interfaces
have a near circular shape which sprawls all over the width of the channel.
By the way, reinitialization and mass correction with translation of φ can be
legitimately used and achieve good mass conservation and accurate curvature
computation.
Geometry and Boundary Conditions
All our numerical simulations are realized in the 2D case. The domain geom-
etry is composed of channels which cross each other with injections of various
fluids at various extremities, or exits of fluids, depending on the physical ex-
periments. (see Figure 1)
In addition, we must impose sliping boundary conditions (BC) on the walls
of the channels :
u.τ = αus (η) + βLs (η) ∂(u.τ )
∂nw (8)
u.nw = 0
where nw and τ are respectively the normal and the tangent to the boundary,
us is the slip velocity and Ls is the so called “slip length” coefficient both
depending a priori on η since sliping is different from one fluid to another. We
suggest the hybrid condition (8) to take into account whether the interface
touches the wall (α 6= 0 and β = 0 : non homogeneous Dirichlet BC) or not
(α = 0 and β 6= 0 : Navier BC). Classical boundary conditions are used at
the inlets and outlets.
598 Paul Vigneaux
In order to solve the Stokes equation (2), we use a finite volume discretization
on staggered cartesian grid and an Augmented Lagrangian method to take
into account the incompressibility constraint (3). If there are solid obstacles
in the flow, we use a penalization method of the obstacles [4] in order to keep
the cartesian grid. The transport equation (4) is solved through the fifth order
WENO schemes [5].
The time step ∆t must obey stability conditions due to convective and
surface tension terms. The time step associated to the convection is classicaly
∆tc = ∆x/max(v). Using an approach that allows to retrieve the stability
condition of Brackbill et al. [6] for Navier-Stokes equation, we find that the
stability constraint associated to the surface tension term in Stokes equation
is ∆tσ = ση ∆x. For our microfluidic applications, ση ∼ 1 and max(v) ∼ 10−2
which means that ∆tσ is 100 times smaller than ∆tc . Thus, since the smaller
time step is taken, the stability constraint is very restrictive. In the following,
we propose an approach to relax this stability constraint.
black one is the interface of the droplet which is, in this case, stationary, the
red curve is a perfect circle superimposed on the droplet in order to show
the difference with the asymptotic shape obtained when there is no injection
(static case). Figure 3 shows, for the same simulation, the velocity field in
the droplet’s frame of reference. We can see that the asymptotic shape is
obtained since velocity is tangent at the interface, thus there is no deformation
of the droplet. A bigger droplet, with the same physical properties, is shown
in Figure 4 with the velocity field in the droplet’s frame of reference : we
observe additional vortices compared to the small droplet. These simulations
allows to study the mixing inside droplets for different flow conditions. Figure
5 shows a train of droplets which flow upward and exit the two horizontal
channels. In Figure 6, we see two droplets, arriving by the two horizontal
channels, colliding and flowing downward. These two cases of figures 5 and 6
shows the ability of the code to handle topological changes of the interface.
5 Conclusion
We have presented a Level Set method for the simulation of flows in Microflu-
idics, which appears to be well adapted to deal with such curvature-driven
flows. A stability analysis shows that surface tension induces a restrictive
condition on the computational time step. We thus introduced a splitting
method to relax this constraint and shorten simulation times. Finally we give
numerical results which show the ability of the method to handle moving in-
terfaces in Microfluidics ; the agreement with physical experiments leads us
to use this code with our collegues to explore the mixing inside microdroplets.
References
1. Osher, S., Sethian, J.: Fronts propagating with curvature-dependent speed: al-
gorithms based on Hamilton-Jacobi formulations. J. Comput. Phys., 79, 12–49
(1988)
2. Sussman, M., Smereka, P., Osher, S.: A level set approach for computing solu-
tions to incompressible two-phase flow. J. Comput. Phys., 114, 146–159 (1994)
3. Osher, S., Fedkiw, R.: Level Set Methods and Dynamic Implicit Interfaces.
Springer, New York (2003)
4. Angot, P., Bruneau, Ch.-H., Fabrie, P.: A penalization method to take into
account obstacles in incompressible viscous flows. Numer. Math., 81, 497–520
(1999)
5. Jiang, G-S., Peng, D.: Weighted ENO schemes for Hamilton-Jacobi equations.
SIAM J. Sci. Comput., 21, 2126–2143 (2000)
6. Brackbill, J. U., Kothe, D. B., Zemach, C.: A continuum method for modeling
surface tension. J. Comput. Phys., 100, 335–354 (1992)
600 Paul Vigneaux
inlet
u u
wall outlet
1.1 Methodology
Porosity is defined as the ratio of the volume of the total pore space filled by
the fluid to the total volume of the porous medium, all within the domain of
interest. Since mesh cells in the domain have been tagged as solid or fluid, the
volume of total pore space can easily be obtained numerically by finding the
sum of volumes of the cells tagged as fluid. Similarly, the volume of the solid
matrix is the sum of volumes of the cells tagged as solid. Finally, the total
medium volume is simply the sum of the volumes of all cells. Figures 1.a − 1.c
are three representative porous media obtained numerically. These three plots
are used to demonstrate that various solid microstructures can be obtained
from triangular, rectangular, and hybrid meshes respectively. They exhibit
different characteristics determined by inputs provided in the processes of
seeding and growing inside unstructured mesh beds. The blue and red colors
represent fluid and solid elements, respectively.
Fig. 1. Randomly formed solid matrix representing porous medium in 2-D domain
(a) fibrous shape formed with triangular mesh; (b) small circular shape solids; (c)
large solids in circular shape with hybrid mesh.
Mesh-Based Microstructure Representation Algorithm 603
2 Sensitivity test
Fig. 2. Sensitivity test of pressure drop versus initial seed used in random number
generator.
604 May-Fun Liou and Issac Greber
at exit which are shown in black line with square symbol; the effect from the
porous micro-structure persists.
2.2 Geometric types
The solid structure is approximated by two types of geometric representations,
sphere-like and fiber-like. Table 1 lists structures of different porosities. In the
case of sphere-like porous medium, d is the radius of the sphere, while d is the
thickness of fiber for fibrous solid phases. The fibrous structure is composed
of slender and long fiber in sheets in which the form drag is much smaller
than the circular cylinder tubes structure. Figure 4 shows the effects of solid
shape and permeability on the sphere-like and fibrous structures listed above.
2
The non-dimensional pressure gradient parameter is defined by P ∗ = L∆pµ dV ,
Fig. 4. Sensitivity test of pressure drop versus Reynolds number in various types
of solid structure.
Mesh-Based Microstructure Representation Algorithm 605
where Red is the Reynolds number based on the pore scale ρ Vµ d . It is evident
that the sphere-like structure causes higher pressure drop than the fibrous
structure.
2.3 Grid Independence Study
It is expected that the more the mesh is refined, the more closely the micro-
scale can be simulated. In the vicinity of the solid-fluid interface, there exist
large velocity and temperature gradients, and an adequate number of mesh
points is necessary to accurately resolve these variations. Also, the use of fine
grids in the flow direction may be critical when the effects of axial conduc-
tion and convection are of concern. However, the desire for finer meshes is
limited by available computing resources. Hence, grid-independence tests are
conducted before this methodology could be used for further computations.
These tests were done by varying mesh sizes while holding porosity fixed.
4 Conclusions
A robust modeling technique for porous medium transport has been proposed
for detailed quantification of how microscopic properties affect macroscopic
transport phenomena. The pore density is currently controlled by random
number generator; it can be extended to conform specified geometries. To meet
the requirement of sufficient resolution around solid structures, the adaptive
meshing technique can be incorporated to locally refine meshes in regions
that require it. Replacement of the non-slip condition with a Maxwellian slip
condition at the interface or wall may be added if the local characteristic
length (finest mesh scale) is comparable to the mean free path of the fluid [3].
606 May-Fun Liou and Issac Greber
6 May-Fun Liou and Issac Greber
Fig. 6. (a) Nusselt number distribution along the axial direction; (b) Convergence
history of residuals from solving Navier-Stokes equations for fluid and solving heat
conduction equations for solid.
Fig. 6. (a) Nusselt number distribution along the axial direction; (b) Convergence
References
history of residuals from solving Navier-Stokes equations for fluid and solving heat
conduction equations for solid.
1. Matsumoto, M. and Kurita, Y., ”Twisted GFSR generators II,” ACM Trans. on
Modeling and Computer Simulation,” 4, 254–266, (1994).
2. Chen, K.-H., Norris, A. T., Quealy A., and Liu, N.-S., ” Benchmark Test Cases
References
for the National Combustion Code,” AIAA Paper 98-3855, (1998).
3. Barber, R. W. and Emerson, D. R., ”The Influence of Knudsen Number on
[1] Matsumoto, M. andDevelopment
the Hydrodynamic Kurita, Y., ”Twisted GFSR Parallel
Length within generators II,”Micro-Channels,”
Plate ACM Trans. on
Modeling and Computer Simulation,” 4, 254–266, (1994).
Advances on Fluid Mechanics IV, Rahman et. al. (Eds.), WIT press, (2002).
[2] Chen, K.-H., Norris, A. T., Quealy A., and Liu, N.-S., ” Benchmark Test Cases
for the National Combustion Code,” AIAA Paper 98-3855, (1998).
[3] Barber, R. W. and Emerson, D. R., ”The Influence of Knudsen Number on
the Hydrodynamic Development Length within Parallel Plate Micro-Channels,”
Advances on Fluid Mechanics IV, Rahman et. al. (Eds.), WIT press, (2002).
The Numerical Simulation of Liquid Sloshing
in Microgravity
University of Groningen, IWI, P.O. Box 800, 9700 AV Groningen, The Netherlands
r.luppes@math.rug.nl
1 Introduction
The description of the interaction between the sloshing fluid inside Sloshsat
and the motion of Sloshsat itself requires a coupling between the mathemat-
ical models for fluid dynamics and solid-body dynamics. These models are
extensively described in [12, 13, 14].
The fluid flow inside Sloshsat can be described by means of the Navier-
Stokes equations for incompressible flow in terms of the fluid velocity u, the
pressure p and the fluid density ρ and viscosity µ. The influence of the solid-
body motion on the fluid is described through a virtual body force, which
depends on the angular velocity ω and acceleration ω̇ and the linear acceler-
ation q̇ of Sloshsat. These quantities follow from the equations for solid-body
dynamics, describing conservation of linear and angular momentum of ’dry’
Sloshsat. The equations contain additional terms representing the force and
torque that the fluid via pressure and viscous effects exerts on the tank wall,
as well as an additional external force describing the thruster induced forces
on the satellite. The equations for solid-body dynamics are rewritten by using
the total (liquid+solid) mass and moment of inertia of the coupled system. In
this way, the system can be solved iteratively in a stable manner [9, 12, 13].
The Numerical Simulation of Liquid Sloshing in Microgravity 609
continuity of
θ normal and
static tangential
contact angle stress
conservation
no slip of mass and
momentum
Fig. 2. Schematic overview of the liquid model. Dark and light shading represent
solid body and liquid, respectively.
For the spatial derivatives of the flow model, finite-volume discretisation is ap-
plied, making use of geometry apertures describing the parts of the cell faces
that are open for flow [11, 12, 14]. To enhance the stability of the numeri-
cal scheme, interpolations are chosen carefully to ensure that the discretised
equations have the same symmetry properties as the continuous equations
[16]. For the temporal derivatives in the flow model the forward Euler method
is used, leading to a Poisson equation for the pressure that is solved iteratively
using SOR with automatically adjusted relaxation for optimal convergence be-
haviour [17]. The temporal derivatives of the solid-body model are integrated
in time using a 4th order Runge-Kutta method. This leads to a linear system
of 6 six unknowns (3 components of both q̇ and ω̇), which is solved by means
of Gaussian elimination during each Runge-Kutta step [12, 13].
When the pressure at a time level is determined, the velocity field is easily
obtained and the free surface can be displaced. This is done using an adapted
version of the volume-of-fluid (VOF) method introduced in [18]. An important
610 R. Luppes, J.A. Helder and A.E.P. Veldman
drawback of the original method is the creation of unrealistic small drops that
get separated from the bulk fluid. Moreover, free-surface apertures may be
outside their natural range. Therefore, a correction is used in terms of a local
height function that is based on the orientation of the free surface [12, 11]. An
alternative method, which is however far more difficult to implement, has been
proposed in [19]. Both methods have been compared in [11], leading to the
conclusion that for certain flow applications the second method yields slightly
better results. However, for the Sloshsat simulations in the present study, the
VOF method with local height function leads to satisfactory results.
The simulation of a Sloshsat manoeuvre is started with a lasting period
of constant rotation ω around the stable axis of maximum moment of inertia
(MOI) and q = 0, without any thruster action. In this initial period, the water
inside the tank can settle towards a certain stable configuration. Only the
discretised flow model is solved, as ω and q are constant in this period. The
final state after the initial period is referred to as t = 0. At t = 0, the manoeuvre
is started and all discretised equations are solved in sequential order.
During a period of 8 days, several experiments have been carried out with
Sloshsat in an orbit around earth. These experiments were controlled by means
of 12 thrusters that could be fired with a frequency of 30Hz. The status
of each thruster (whether or not a thruster has been fired) was recorded.
The thruster forces that operate on Sloshsat can be determined subsequently
from the measured pressure inside the fuel tank. The various experiments and
measurement equipment are extensively described in [5, 6, 7].
In Fig. 3 (left column), measurements of a spin-up around the axis of maxi-
mum MOI are compared with the results of a simulation. The rotational veloc-
ity changes from ω =(-0.004;0.104;-0.007)T towards a rotation with ωy ≈ 0.179
and nutating ωx and ωz components. The increase of ωy is almost perfect.
However, the results of ωx and ωz show a clear difference. Although the fre-
quency of the nutation in ωx and ωz is comparable, in the simulation the
amplitudes are too small. A proper calibration of the measured data may
yield a favourable correction of the small measured values of ωx and ωz . The
decrease of the amplitude is too strong in the simulation, probably as a result
of the use of first-order spatial discretisation to enhance stability. This gen-
erally results in additional non-physical (numerical) damping, depending on
the magnitude of fluid velocity. Capillary forces are of less importance for the
observed damping, as ω is relatively large in this experiment.
The results for a nutation-avoidance manoeuvre (NAM) are depicted in
Fig. 3 (middle column), where the objective of the experiment is to move
ω without incurring too much nutation. Initially, the fluid is adapted to
ω =(0.003;-0.101;0.007)T . In the ωx component (top figure), a series of thruster
pulses can be distinguished at t = 0, 70 and 140s, resulting in a rapid change of
The Numerical Simulation of Liquid Sloshing in Microgravity 611
5 Conclusions
The influence of sloshing liquid on board spacecraft has been investigated by
carrying out experiments with the satellite ’Sloshsat FLEVO’. These experi-
ments were supported by a computational model based on the Navier-Stokes
equations for 3D incompressible free-surface flow with capillary surface physics
and coupled solid-liquid interaction dynamics.
Measurements of the rotational velocity ω of Sloshsat have been com-
pared with numerical simulations. The obtained frequencies are reasonably
comparable at various satellite manoeuvres. Hence, in the numerical model,
the sloshing frequency of the water inside the tank is adopted correctly by the
satellite motion. The nutation damping in the simulations is too large in case
of large rotational velocities and large-scale liquid motion, because of addi-
tional numerical diffusion in the computational model. The use of higher-order
discretisation to reduce the level of numerical diffusion is recommended.
At low rotational rates and small-scale liquid motion, capillary effects are
important for the damping of manoeuvre-induced oscillations. A dynamic
contact-angle model probably yields better agreement between simulations
and experiments in these cases. A proper calibration of the small components
of ω may yield a favourable correction of the measured values.
References
1. Strikwerda, T.E., Ray, J.C., Haley, D.R. and O’Shaughnessy, D.J. In Guidance
and Control 2001, volume 107, pages 597–614. Am. Astron. Soc., 2001.
2. http://near.jhuapl.edu.
3. http://www.history.nasa.gov/alsj/a11/a11.landing.html.
4. http://www.nlr.nl/public/publications/pdf/f183-02.pdf.
5. Vreeburg, J.P.B. In 56th Int. Astronautical Congr. IAF, 2005. IAF-05-C1.2.09.
6. Vreeburg, J.P.B. Technical Report CR 96755 L, NLR, 1996.
7. Vreeburg, J.P.B. In 50th Int. Astronautical Congr. IAF, 1999. IAF-99-J.2.04.
8. Prins, J.J.M. In 51st Int. Astronautical Congr. IAF, 2000. IAF-00-J.2.05.
9. Vreeburg, J.P.B. and Veldman, A.E.P. In R. Monti, editor, Physics of Fluids in
Microgravity, pages 293–321. Gordon and Breach Acad. Publ., 2002.
10. Veldman, A.E.P and Vogels, M.E.S. Acta Astronautica, 11:641–649, 1984.
11. Kleefsman, K.M.T., Fekken, G., Veldman, A.E.P., Iwanowski, B. and Buchner,
B. J. Comp. Phys., 206:363–393, 2005.
12. Gerrits, J. Dynamics of Liquid-Filled Spacecraft. PhD thesis, University of
Groningen, The Netherlands, 2001.
13. Gerrits, J. and Veldman, A.E.P. J. Eng. Math., 45:21–38, 2003.
14. Luppes, R., Helder, J.A. and Veldman, A.E.P. In 56th Int. Astronautical Congr.
IAF, 2005. IAF-05-A2.2.07.
15. Van Mourik, S., Veldman, A.E.P and Dreyer, M.E. Microgr. Sci. Techn., 17-
3:91–98, 2005.
16. Verstappen, R.W.C.P. and Veldman, A.E.P. J. Comp. Phys., 187:343–368, 2003.
17. Botta, E.F.F. and Ellenbroek, M.H.M. J. Comp. Phys., 60:119–134, 1985.
18. Hirt, C.R. and Nichols, B.D. J. Comp. Phys., 39:201–225, 1981.
19. Youngs, D.L. Technical Report AWRE/44/92/35, AWRE, 1987.
Numerical Prediction of Interfacial Instability
1 Introduction
4
Some rare and tentative attempts for a sharp treatment are noted below. The
elegant and essentially exact work with the boundary integral method is limited
to the ideal limits of inviscid [5, 15], or Stokes [14, 16] flows.
Fig. 1. Spurious velocity and grid convergence. a) Spurious velocity for Rayleigh-
Taylor instability. b) Spurious velocity for a static drop. c) Grid convergence
of velocity, pressure and stresses for discontinuous flow test, defined as fol-
lows. A bubble of radius R = 0.4 is suspended in the center of the com-
putational domain h1×1. The i pressure
h and
−velocity
inside and outside of the
i
bubble are set to P, Vξ , Vη (r) = P0 + PI − P0 Rr 3 ; VξI Rr 2 ; VηI Rr 3 and
h − + + i
P − σ
R
R
r ; VξI 1 − 2 µ− r
R 1− r
R ; Vη 1 − 3 µ− 1 − r
R , respectively; where
I µ I µ
− µ+
P0 = 20, P
I
= 10, Vξ
I
= 1, Vη
I
= 1, σ = 1 and µ−
= 100. Pressure and velocity
fields are shown in d).
Concerns about the fidelity of DIMs have been raised previously by Kang
et al. [6] and Nourgaliev et al. [10] but could not be resolved, while Ye et al.
[24, 21], and Lee and LeVeque [7] offered sharp interface treatments but not
in the interfacial instability context considered here and in Theofanous et al.
[19, 20].
(SAMR) computational environment [1] that includes a Level Set (LS) treat-
ment [9, 11]. The C2 -grid is dynamically generated from the LS function in
a narrow band around the interface in order to allow conservative treatment
of the fluid dynamics, and to accurately piecewise-linearly represent the in-
terface, while avoiding small-cell issues. This complex data structure will be
denoted hereafter as adaptive Cartesian Cut-Cell Mesh Refinement (aC3 MR).
This quasi-Lagrangian-type treatment of the interface is complemented, dur-
ing fluid dynamics updates of the cut-cells, by using dynamic flux treatment
at interfacial edges, taking into account the momenta of inertia of irregular
cells, producing up to the third-order accuracy, with appropriate mapping of
cell-average to cell-center variables and communication between irregular C2
and structured C1 grids. The cost of supporting the C2 -grid is minimal as it
is generated in a narrow band around the interface, which pertains only to
the finest level of SAMR hierarchy.
The dynamic coupling between the two fluids on either side of the in-
terface is accomplished through a special discretization of fluid dynamics at
the centers of interfacial edges of the cut-cells. For this we have developed a
weighted least-squares, multi-dimensional, Taylor-expansion-based reconstruc-
616 Robert Nourgaliev, Meng-Sing Liou, and Theo Theofanous
The code, MuSiC-SIDS, has been built on the basis of previous versions,
MuSiC-ua2 LS, and -aCBM, it has been run on a 40-node Xeon-based clus-
ter, and shown to exhibit the good scalability needed for massively parallel
simulations.
Acknowledgments. This work was carried out for the Joint Science and
Technology Office, Defense Threat Reduction Agency, with the support and
cooperation of Mr. C. Fromer and Dr. J. Hannan. We are also grateful to Drs.
G. Nakafuji, F. Handler (LLNL), and R. Babarsky (NGIC) for continuing,
long term support and collaboration.
References
1. Berger, M.J., and Colella, P.: Local Adaptive Mesh Refinement for Shock Hydrodynamics.
J. Comput. Phys., 82, 67–84 (1989)
2. Brackbill, J.U., Kothe, D.B., and Zemach, C.: A continuum method for modeling surface-
tension. J. Comp. Phys., 100(2), 335–354 (1992)
3. Chandrasekhar, S.: Hydrodynamic and Hydromagnetic Stability. Dover Publications Inc.,
New York (1981)
4. Coward, A.V., Renardy, Y.Y., Renardy, M., and Richards, J.R.: Temporal Evolution of
Periodic Disturbances in Two-Layer Couette Flow. J. Comp. Phys., 132, 346–361 (1997)
5. Hou, T.Y., and Lowengrub, J.S., and Shelley, M.J.: The long-time motion of vortex sheets
with surface tension, Phys. Fluids, 9(7), 1933 (1997)
6. Kang, M., Fedkiw, R.P., and Liu, X.-D.: A Boundary Condition Capturing Method for
Multiphase Incompressible Flow. SIAM J. Sci. Comp., 15(2), 323–360 (2000)
7. Lee, L., LeVeque, R.J.: An immersed interface method for incompressible Navier-Stokes
equations. SIAM J. Sci. Comp., 25(3), 832–856 (2003)
8. Nourgaliev, R.R., Dinh, T.N., and Theofanous, T.G.: A Pseudo-Compressibility Method
for the Numerical Simulation of Incompressible Multifluid Flows. Intern. J. Mult. Flow,
30(7-8), 901–937 (2004)
9. Nourgaliev, R.R., Dinh, T.N., and Theofanous, T.G.: Adaptive Characteristics-Based
Matching for Compressible Multifluid Dynamics, J. Comp. Phys., 213(2), 500–529 (2006)
10. Nourgaliev, R.R., Dinh, T.N., and Theofanous, T.G.: Sharp Treatment of Surface Tension
and Viscous Stresses in Multifluid Dynamics, AIAA 2005-5349, 17th AIAA CFD Conference,
Toronto (2005)
11. Nourgaliev, R.R., and Theofanous, T.G.: High-Fidelity Interface Tracking in Compressible
Flows: Unlimited Anchored Adaptive Level Set, J. Comp. Phys, under review (2006)
12. Nourgaliev, R.R., Liou, M.-S., and Theofanous, T.G.: Sharp Treatment of Interface Jump
Conditions in Multifluid Dynamics. J. Comp. Phys., to be submitted (2006)
13. Peskin, C.: Flow patterns around heart valves: a digital computer method for solving the
equations of motion. PhD Thesis, Physiol., Albert Einstein Coll. Med., Univ. Microfilms.
378:72-30 (1972)
14. Pozrikidis, C.: Instability of multi-layer channel and film flows, In: Advances in Applied
Mech., Ed. H. Aref and E. van der Giessen, Elsevier, (2004)
15. Prosperetti, A.: Boundary integral methods, In: Course on Drop-Surface Interactions, Ed.
M. Rein, Springer, (2002)
16. Ryskin, G., and Leal, G.: Numerical Solution of Free Boundary Problems in Fluid Mechan-
ics. Part 1. The Numerical Technique, J. Fluid Mech. 148(1) (1984)
17. Scardovelli, R., and Zaleski, S.: Direct numerical simulations of free-surface and interfacial
flow. Annu. Rev. Fluid Mech., 31, 567–603 (1999)
18. Tauber, W., Unverdi, S.O., and Tryggvason, G., The nonlinear behavior of a sheared im-
miscible fluid interface, Phys. Fluids, 14, 8 (2002)
19. Theofanous, T.G., Nourgaliev, R.R., Li, G.J., and Dinh, T.N.: Compressible Multi-
Hydrodynamics (CMH): Breakup, Mixing, and Dispersal, of Liquids/Solids in High Speed
Flows, In: IUTAM Symp. on Computational Approaches to Disperse Multiphase Flow, Ed.
A. Prosperetti and S. Balachandar, Springer, (2006)
20. Theofanous, T.G., Nourgaliev, R.R., Wiri, S., and Liou, M-.S.: Can Diffuse Methods be
Tolerated in Numerical Predictions of Interfacial Instability? Phys. Fluids, under review
(2006)
21. Shyy, W.: Multiphase computations using sharp and continuous interface techniques for
micro-gravity applications. Comptes Rendus Mechanique, 332, 375–386 (2004)
22. Unverdi, S.O., and Tryggvason, G.: A Front-Tracking Method for Viscous, Incompressible,
Multi-Fluid Flows. J. Comp. Phys., 100, 25 (1992)
23. Yiantsios, S.G., and Higgins, B.G.: Linear-stability of plane poiseuille flow of 2 superposed
fluids, Phys. Fluids, 31, 11 (1988)
24. Ye, T., Shyy, W., and Chung, J.N.: A fixed-grid sharp-interface method for bubble dynamics
and phase change. J. Comp. Phys., 174, 781–815 (2001)
A Lattice Boltzmann based Single-Phase
Model: Surface Tension and Wetting
Summary. A 2D single-phase free surface tracking model using the Lattice Boltz-
mann method (LBM) is developed in this study. In contrast to the conventional
multi-phase models, it is not necessary to simulate the motion of the gas phase
using this LBM single-phase algorithm, and thus improves the computational effi-
ciency without sacrificing the underlying physics. Based on Gunstensen0 s immiscible
Lattice Boltzmann algorithm, a perturbation is added to the distribution functions
of the interface cells for introducing the surface tension and wetting into the LBM
single-phase model. Implementations of the model are demonstrated for simulations
of droplet deformation under the surface tension effect and the droplet spreading on
solid surface with various wetting properties. Simulation results show that the pro-
posed model can generate a surface tension that is isotropic, and the model allows
for continuous control of contact angles by varying the mass gradient at the wall
boundary cells.
1 Introduction
The simulation of flows with free surfaces has a variety of technical appli-
cations such as flow through porous media, boiling dynamics and etching
processes. Difficulties in the simulation of these types of problems lie in mod-
eling interface dynamics and dealing with the complex boundaries. Several
LBM models to simulate the interfacial dynamics have been developed in the
past decade. Gunstensen et al.[1] proposed a color model for simulating im-
miscible binary fluids based on Lattice Gas model of Rothman and Keller[2].
D0 Ortona et al.[3] studied the surface tension and wetting properties by mod-
ifying the color redistribution step and obtained good agreement with theory.
Shan and Chen[4, 5] and Shan and Doolen[6] considered a fluid with S differ-
ent components on a regular lattice. In order to include the surface tension,
they modified the collision operator by adding the force on the kth phase ow-
ing to a pairwise interaction between different phases. Orlandini et al.[7] and
Swift et al.[8, 9] proposed a free energy approach for simulating binary fluids
and liquid-gas fluids. The surface tension is considered by incorporating a free
The test cases chosen for the application of the single-phase model concerns
surface tension and wetting. The model will be applied to two cases, droplet
deformation under the effect of surface tension, and droplet spreading on a
solid wall with different wetting properties. In our simulation for surface ten-
sion and wetting, gravity is not included as the dimensionless Bond number
Bo is less than 1.0 indicting negligible effect of the gravitational force com-
pared with the surface tension force.
Test case 1: Droplet deformation due to surface tension. From a thermody-
namics viewpoint, in the absence of external forces a two-phase system, under
surface tension effect, tends to minimize its interface area. This is caught by
the LBM single-phase model developed in this study. The simulation was car-
ried out in a computational domain with 40 × 40 lattices filled with fluid. The
relaxation time τ was set as 2.0. Here the droplet initial shape is assumed as
a square, and then it undergoes deformation under the effect of surface ten-
sion, as shown in Fig.1. It can be seen that the interface is quite smooth. In
calculation, surface tension control parameter is β = 0.002, corresponding to
the lattice surface tension of 0.37 ± 0.01 . The effective radius of the circular
blob as a function of the angular direction (−180◦ to 180◦ ) was checked. The
average value of the radius is equal to 22.5 cell units, and small fluctuation is
observed, of about 0.4 cell unit, which demonstrates that our model generates
an isotropic surface tension. The effect of surface tension on droplet shape is
also studied, and the results are shown in Fig.2. When the surface tension
control parameter β is set as 0.001, it can be seen that the final shape is not
an exact circle. While, when β is set as 0.002 or 0.003, the final shape of the
622 X.Q. Xing et al
Fig. 1. Time evolution of a droplet changing from square shape to circular shape
droplet is a circle. The simulation results are in accordance with the surface
physics. The higher the surface tension forces, the faster the deformation. The
computational cost is also reduced significantly when the surface tension con-
trol parameter β increases.
Test case 2: Droplet wetting. Wetting involves interactions of liquids with a
solid surface. In simulation, collision parameter was chosen as τ = 2.0 , surface
tension control parameter β = 0.002. Initial droplet shape is assumed as a cir-
cle with a radius of 15 lattice units. Then the droplet undergoes deformation
on the solid surface due to surface tension effect, and finally reaches equi-
librium shape, Fig.3 shows the final equilibrium shape of droplets in contact
with a solid surface with different wetting properties. The different wetting
results are obtained by assigning different values of the mass gradient ∂m ∂n at
the wall cell. For Figs.3(a) to (i), the corresponding mass gradients are 1.0,
0.9, 0.8, 0.7, 0.6, 0.5, 0.4, 0.2 and 0.0 respectively. Contact angles correspond-
ing to Figs.3(a) to (i) are 180◦ , 160◦ , 142◦ , 125◦ , 105◦ , 90◦ , 75◦ , 40◦ , and 0.0◦
respectively. Computational costs are in the order of 104 time steps.
In Fig.4, the symbols are the contact angles obtained from the final equilib-
rium shape of droplets simulated by the model proposed in this study. The
solid line is obtained from a simple model, namely a zero mass gradient corre-
sponds to a zero contact angle and a mass gradient equal to the liquid density
corresponds to a 180◦ contact angle, and intermediate situations should lead
to an angle between 0◦ and 180◦ according to a linear rule. It can be seen
that contact angle varies almost linearly with the mass gradient of the wall
cells, which is in good agreement with the simple model.
4 Conclusion
A LBM based single-phase free surface tracking algorithm has been developed.
The proposed model has been implemented to simulate the deformation of a
droplet due to surface tension effect, the time evolution of a droplet spreading
Single-Phase Model: Surface Tension and Wetting 623
Fig. 3. Final equilibrium shape of droplets on solid surfaces with different wetting
properties
180 x
x
x simulation results
150
theoretical value x
x
Contact Angle
120
x
90 x
x
60
x
30
0x
0 0.2 0.4 0.6 0.8 1
Mass gradient
References
1. Gunstensen,A.K., Rothman,D.H.: Lattice Boltzmann Model of Immiscible Flu-
ids, Phys. Rev. A 43, 4320–4327 (1991).
624 X.Q. Xing et al
1 Introduction
In flow situations where multiple material phases are found at scales that can-
not be resolved by direct numerical simulation, the dynamics can be described
by effective-field, or coarse-grain models [1-3]. Here we are concerned with the
flow involving two compressible fluids (at least one of them moving at high
Mach number), the basic formulation of the so-called two-fluid model, and
the numerical issues that arise from the well-known non-hyperbolic and non-
conservative characters of this model. Of particular interest is the application
of this model to flows with steep gradients in flow quantities and material
densities, as occurring in high-speed, disperse flows [4]. These, in conjunction
with the numerical issues just noted, can cause numerical difficulties, such
as spurious oscillations and non-convergence of solutions. The “remedy” of
incorporating adequate amounts of numerical diffusion is not an option.
Other “remedies” found in past work, all for low Mach number condi-
tions, are based on the recognition that certain inviscid interactions between
the phases have been lost in the coarse-graining procedure (space-, time-, or
ensemble-averaging) and that they must be recovered through modeling. In
particular we note the concepts of virtual mass [5-7] and interfacial pressure
[8]. The latter is to represent a local difference between the pressures in the
two phases; but its origin and role remain obscure. From our present stand
point we need to address the well-posedness of the two-fluid models and how
much the solution changes by the change of parameters that control the “de-
gree” of hyperbolicity. In particular, we will examine how hyperbolicity affects
the numerical solution in the presence of shocks and contact discontinuities,
under general flow conditions (in speeds, thermodynamic and material states).
∂Uk ∂Fk
+ = Pk + Pint vm v
k + Fk + Fk , k = gas or liquid phase (1)
∂t ∂x
where Uk = αk ρk (1, uk , Ek )T denote the conservative variables, Fk = αk ρk uk
(1, uk , Hk )T + (0, αk p, 0)T the inviscid fluxes, Pk = p(0, ∂αk /∂x, −∂αk /∂t)T
the pressure interaction between phases, and Fvk the viscous diffusion terms.
The inviscid interfacial transfer terms include effects of interfacial pressure
(Pint
k ) and virtual
vm
P mass (Fk ). αk is the volume fraction of phase k, satisfying
the constraint k αk = 1. The other variables are standard.
3 Hyperbolicity
As the interfacial terms are typically expressed as first derivatives in space
and time, they can exert a significant influence on the eigenvalues, hence the
hyperbolicity of the system. Studies of hyperbolicity of two fluid models have
been reported in the literature for both concepts (see for example [7,8]), but
they were all made for incompressible flows under the assumptions w2 << a2g ,
where w is the relative velocity and ag the speed of sound in gas phase. the
result is a simple, explicit hyperbolicity condition, however, this is not the case
in general. In what follows we will establish the conditions under which the
system becomes hyperbolic for all flow conditions and demonstrate the effects
of hyperbolicity on the numerical solutions, especially in terms of convergence
and accuracy.
Following Drew et al. [5], we express the virtual mass term by Fvm
d =
fdvm (0, 1, uint )T , with
∂(uc − ud ) ∂uc ∂ud
fdvm = Cvm αd ρc [ + uvm
c − uvm
d ], (4)
∂t ∂x ∂x
Hyperbolicity, Discontinuities, and Numerics of Two-Fluid Models 627
where subscripts “c”, and “d” refer to the continuous and disperse phase
respectively, and superscript “vm” is to indicate the velocities appropriate to
the substantial derivatives involved in the virtual mass formulation [5]. Also,
uint is the interfacial velocity, and Cvm is an appropriate coefficient which
we will use as a parameter in investigating the hyperbolicity of the system.
We solve for the boundary of the hyperbolic zone using the full characteristic
equation of the system (no assumptions on the magnitude of the relative
velocities), and we use Cp∗ and Cvm as parameters alone or in combination.
In Fig. 2(a) we show minimum values of Cvm required for hyperbolicity as
a function of the value of Cp∗ covering the whole space of void fractions and
correspondingly of disperse flow regimes, from bubbly to droplet, but under
the assumption w2 << a2g . Note that in certain combinations the magnitude of
the virtual mass needed for hyperbolicity appears to somewhat exceed what
is known to represent physical reality—in particular the low void fraction
(bubbly) flows. In Fig. 2(b) we show the regularizing potency of the virtual
mass term alone, as a function of Mach number referred to the relative velocity
(w). In contrast to the case just shown, the very high void fraction (droplet)
flows appear to need somewhat excessive values of the coefficient, while the
need for regularization of bubbly flows diminishes with increasing relative
velocity, to a vanishingly small quantity beyond Mach number of about 0.3.
Also it should be clear that in Fig. 2 the limit w → 0 is approached correctly
in that the virtual mass term needed for regularization also goes to zero, as
appropriate to reflect the well-posed homogeneous flow model, even though
the coefficient itself appears to remain finite.
Fig. 2. The hyperbolicity boundary on Cvm for the virtual mass in [5].
Fig. 3. Faucet problem. (a): effect of hyperbolicity, and (b): grid convergence (note
the expanded scale in x to view the solution near the volume fraction front.)
Fig. 4. Shock tube problem [11]. (a): with interfacial pressure alone and (b): with
virtual mass alone [5].
630 Liou, M.-S. et al.
4 Concluding Remarks
We showed the effects of interfacial pressure and virtual mass on hyperbolicity
of the two-fluid equations, providing a complete map in an all-speed range.
In turn, the effects of hyperbolicity on stability and convergence of numerical
solutions were investigated using an accurate method. We concluded that
the concern due to the con-conservative form did not impact convergence
or uniqueness, as long as the system was conditioned adequately. On the
other hand our numerical results indicate that placing modestly inside the
hyperbolic zone may not be always sufficient for numerical regularization,
and this requires further investigation. Also more work is needed, especially on
the interfacial pressure term, to clarify whether the amount of regularization
indicated from the hyperbolicity analysis is quite consistent with physical
reality.
Acknowledgments. This work was carried out for the Joint Science and Technol-
ogy Office, Defense Threat Reduction Agency, with the support and cooperation
of Mr. C. Fromer and Dr. J. Hannan. We are also grateful to Drs. G. Nakafuji, F.
Handler (LLNL), and R. Babarsky (NGIC) for continuing, long term support and
collaboration.
References
1. Drew, D., Passman, S.L.: Theory of Multicomponent Fluids. Springer-Verlag
NY (1999)
2. Ishii, M.: Thermo-Fluid Dynamic Theory of Two-Phase Flow. Eyrolles, Paris
(1975)
3. Zhang, D.Z., Prosperetti, A.: Physics of Fluids, 6, 2956–2970 (1994)
4. Theofanous T. et al: Computational Approaches to Multiphase Flow (Ed. Bal-
achander, S. and Prosperetti, A.), Springer, 353–369 (2006)
5. Drew D. et al.: Int. J. Multiphase Flow, 5, 233–242 (1979)
6. Drew D., Lahey R.T.: Int. J. Multiphase Flow, 13, 113–121 (1987)
7. Watanabe, T., Kukita, Y.: Nuclear Eng. and Design, 135, 327–340 (1992)
8. Stuhmiller, J.H.: Int. J. Multiphase Flow, 3, 551–560 (1977)
9. Liou, M.-S.: J. Comput. Phys. 214, 137–170 (2006)
10. Ransom, V.H.: Multiphase Science and Technology, 3 (1987)
11. Toumi, I.: Nuclear Eng. and Design, 123, 147–168 (1996)
Computation of Multiphase Mixture Flows
using RoeM and AUSMPW+ Schemes
Seung-Won Ihm, Kyung Rok Lee, Chongam Kim∗ , and Kyu Hong Kim
Summary. RoeM[1] and AUSMPW+[2] schemes, originally developed for the ro-
bust and accurate simulation of single phase gas dynamics, are extended to the
compressible gas-liquid two-phase flows. We adopted homogeneous equilibrium
model(HEM) using mass fraction to describe different two phases. In the Eulerian-
Eulerian framework, HEM assumes dynamic and thermal equilibrium of the two
phases in the same computational mesh. We derived new pressure weighting func-
tion from the mixture equation of state(EOS) to construct the new shock disconti-
nuity sensing term(SDST), which is commonly used in RoeM and AUSMPW+ for
the stable numerical flux calculation. The developed two-phase versions of RoeM
and AUSMPW+ are applied on several air-water two-phase test problems. In spite
of the large discrepancy of material properties such as density, enthalpy, and speed
of sound, the numerical results show that both schemes provide very satisfactory
solutions.
1 Governing Equations
HEM[3] with mass fraction is adopted to describe two-phase flows. As we
assume fully compressible flows, the governing equations are composed of
mixture mass, momentum, and energy conservation laws together with one
phase mass conservation law. The two-dimensional Euler equations are as
follows:
∂Q ∂E ∂F
+ + =0, (1)
∂t ∂x ∂y
T
Q = [ρm , ρm u, ρm v, ρm E, ρm Y1 ] ,
T
E = [ρm U, ρm uU + nx p, ρm vU + ny p, ρm HU, ρm Y1 U ] ,
where ρm is a mixture density, and Y1 stands for the mass fraction of gas
phase.
And the following definition of mixture density combined with each phase’s
EOS plays a role of mixture EOS[4]:
1 1
ρm = P Yi
=P Yi
. (2)
ρ̂i (p,T ) ρ̂i (p,hi )
In Eq. (2), ρ̂i means density defined on the occupied volume of i-th fluid,
while ρm is defined on the each whole computational mesh. We adopted the
stiffened fluid EOS for liquid phase, and the ideal gas EOS for gas phase[5].
Finally, the system is closed by the following assumption of dynamic and
thermal equilibrium:
p = pl = pg , T = T l = T g . (3)
Then, the density ratio can be changed into the ratio of p̄.
ρm,L p̄(pL ) p̄L
≈ = (6)
ρm,R p̄(pR ) p̄R
Now we can use min p̄p̄RL , p̄p̄RL as a SDST. For one dimensional normal
shock condition of M=1.5 and 2, the proposed SDST has the value of 1/O(1)
even for the near liquid phase.
RoeM and AUSMPW+ for Two-phase Flows 633
1 , û2 + v̂ 2 = 0
f= ,
|M̂ |h , û2 + v̂ 2 6= 0
( p̄ ,j p̄
1−min( p̄ i , i+1,j
p̄i ,j )
g= |M̂ | i+1,j , M̂ 6= 0 .
1 , M̂ = 0
It is noted that Mach number-based functions f and g are introduced to
balance damping and feeding rates, which leads to a shock-stable Roe scheme.
single phase can be extended. In the following, the AUSMPW+ scheme for
the two-phase flow is given.
E1/2 = M̂L+ c1/2 φL + M̂R− c1/2 φR + (PL+ pL + PR− pR ) , (10)
1) M1/2 ≥ 0 ,
M̄L+ = ML+ + MR− [(1 − ω)(1 − fR ) − fL ] ,
M̄R− = MR− ω(1 + fR ) ,
2) M1/2 < 0 ,
M̄L+ = ML+ ω(1 + fL ) ,
M̄R− = MR− + ML+ [(1 − ω)(1 − fL ) − fR ] ,
where, ω = max(ω1 , ω2 ) ,
3
ω1 = 1 − min p̄p̄RL , p̄p̄RL ,
min(p̄ ,p̄ ,p̄ ,p̄ )
ω2 = 1 − max(p̄1,L 1,R 2,L 2,R
,p̄1,R ,p̄2,L ,p̄2,R ) ,
1,L
p̄ ρm,L ρm,R
fL,R = L,R p̄s − 1 × (1 − ω2 ) × min ρm,R , ρm,L ,
T
φ = [ρm , ρm u, ρm v, ρm H, ρm Y1 ] .
Our modifications are
Note that the speed of sound at phase interface is lower than that of either
phases, because HEM allows mixture region. From the mixture EOS, typical
averaging of mass fraction at phase interface can show this characteristic, but
any averaging of left and right speed of sound cannot predict low speed of
sound. It is also noted that the function f is modified to stabilized the scheme
near the phase interface of large density ratio.
5 Numerical Results
In order to test the shock stability of the schemes, we solved two-phase mix-
ture moving shock problem on a perturbed grid. Figure 1 shows the results of
each scheme for Y1 = 0.1 mixture flow. While Roe scheme destroys the nor-
mal shock structure, both RoeM and AUSMPW+ clearly capture the shock,
showing their robust and stable behavior even in the two-phase flow region.
RoeM and AUSMPW+ for Two-phase Flows 635
In this challenging problem, the liquid shock of M=1.7 encounters with the
phase interface of about 1:1000 density ratio. Initial right-going liquid shock
is located at x=0.1, and phase interface is at x=0.5. Results at t=0.3ms are
shown in Fig. 2. Well known physics including shock transmission/reflection
pattern, acceleration after the shock-phase interface interaction are shown.
6 Conclusion
Numerical methods for simulating compressible two-phase flows with large
density ratio are presented. We extend the RoeM and AUSMPW+, which are
originally developed for single-phase gas dynamics, to two-phase flows. For
the two-phase flows using homogeneous equilibrium model(HEM) with mass
fraction, new shock discontinuity sensing term(SDST) is derived from mixture
EOS. The developed two-phase versions of RoeM and AUSMPW+ are tested
on several air-water two-phase problems. In spite of the large discrepancy of
material properties, the numerical results show a good performance of the
developed schemes. We expect the proposed SDST can be used for other nu-
merical schemes for two-phase flows. And let us mention that both the SDST
and the developed schemes have a consistent form for isothermal compressible
flows.
References
1. S. S. Kim, C. Kim, O. H. Rho, S. K. Hong : Cures for the shock instability:
Development of a shock-stable Roe scheme, J. Computational Physics, 185,
342–374 (2003)
2. K. H. Kim, C. Kim, and O. H. Rho : Methods for the Accurate Computations of
Hypersonic Flows, Part I: AUSMPW+ Scheme, J. Computational Physics, 174,
38–80 (2001)
3. S. Venkateswaran, J. W. Lindau, R. F. Kunz, C. L. Merkle : Computation of Mul-
tiphase Mixture Flows with Compressibility Effects, J. Computational Physics,
180, 54–77 (2002)
4. Robin K. S. Hankin : The Euler Equations for Multiphase Compressible Flow in
Conservation Form, J. Computational Physics, 172, 808–826 (2001)
5. H. Luo, J. D. Baum, R. Lohner : On the computation of multi-material flows
using ALE formulation, J. Computational Physics, 194, 304–328 (2004)
6. K. H. Kim, J. H. Lee, O. H. Rho : An Improvement of AUSM Schemes by In-
troducing the Pressure-Based Weight Functions, Computers and Fluids, 27(3),
311–346 (1998)
A relaxation method for the Kapila model
1 Mathematical model
We are concerned with the numerical approximation of a two phase compress-
ible flow. We assume the phases non-miscible. For the sake of simplicity, the
viscosity effects are neglected. We consider the physical model, derived in [1],
defined as:
∂t W + ∂xF + B ∂x U = 0 (1)
with:
α1 0 u00β 0 α1
α1 ρ1 α1 ρ1 u 0 0 0 0 0 ρ1
α2 ρ2 , F = α22ρ2 u , B = 0 0 0 0
W= 0 , U = ρ2 (2)
ρu ρu + p 0 0 0 0 0 u
E ρHu 0000 0 ε
where α1 ∈ (0, 1) is the volume fraction for the first phase and α2 = 1 − α1
is recovered from the saturation constraint. For each phase k, we note ρk > 0
the partial density. ε > 0 is the fluid internal energy, u the velocity and p the
pressure. The mixture density, total energy and enthalpy are:
Finally, the internal energy is given by the Stiffened gas equation of state:
p + γP ∞ 1 X αk γP ∞ X αk γk P ∞
k
ρε = where = and = (5)
γ−1 γ−1 γk − 1 γ−1 γk − 1
k k
Let note that in the general case of a non-conservative system, the fluctuations
φ ij and φ ji are not equal. For conservative systems, the fluctuations balance
and the scheme becomes conservative.
If we assume that we can approximate the solution of the Riemann prob-
lem (6) by constants discontinuous waves, our formulation meets the wave
A relaxation method for the Kapila model 639
propagation method [2]. Let suppose that the approximate solution satisfies
the consistency condition:
Xm
λl δ Wl = ∆FF + B̃
B ∆U (8)
l=1
where λl denotes the speed of the discontinuity, δ Wl the jump of the solution
and m the number of discontinuities. B̃B is a constant matrix standing for the
non-conservative terms.
Then the numerical scheme (7) can be reformulated as follow:
ai X
Wn+1 − Wni + C NC
i Ψij · ηij + Ψij · ηij = 0 (9)
∆t
j∈ν(i)
where:
C 1
NC 1
Ψij = F j + F i − |Ã
Aij | (∆W)ij , Ψij = B̃B ij (∆U)ij (10)
2 2
and the viscosity matrix is given by the relation:
Xm
|Ã
A|ij (∆W)ij = |λl |δ Wl (11)
l=1
NC
Ψij stands for the non-conservative contribution. For the conservative part,
C
we have the flux Ψij with its centered and dissipative terms put in evidence.
An additional difficulty for the Kapila system is the term β which expression
depends on the closure (5). We want our numerical scheme free of thermody-
namical closure. In this way, we introduce the term β̃ for which we demand
two properties: it must be a Riemann invariant across shock and rarefaction
waves, the volume fraction saturation constraint must be recovered. We pro-
pose β̃ = α1 − y1 a21 /a2 where yk = αk ρk /ρ is the mass fraction of the fluid k
2
1 X yk
and the positive reals ak are such that: 2 =
a a2k
k=1
Then we rewrite the relaxation system in the form:
1 ∂FFR UR
∂U
∂t WR + A R ∂x WR = R with A R = + BR (14)
λ ∂ WR WR
∂W
With λ set to infinity, the system is unconditionally hyperbolic and linearly
degenerated. We illustrate the Riemann invariants associated to each wave
in the table (1). As the relaxation system is linearly degenerated , it is easy
−1
to determine the viscosity matrix |Ã AR | = P̃P R |Λ̃
ΛR |P̃
P R . Actually, the matrix
ΛR must contain the eigenvalues of the relaxation system and the matrix
Λ̃
P R must contains non nul vectors proportionnal to the solution jumps. After
P̃
straightforward calculations, we find: L
L ∗ ∗ R
λ = λu−a/ρ
|Λ̃
Λ|R = Diag |λ |, |u |, ..., |u |, |λ |
u∗ = λu
λR = λu+a/ρ
L L
β /ρ 1 0 0 0 0 0 β R /ρR
L
∗a
y1L 0 1 0 0 0 0 y1R L
J = H L
− u
ρL
y2L 0 0 1 0 0 0 y2R
R
where ∗a (15)
y2L 0 1 0 0 0 0 y2R J R
= H R
+ u
PR =
P̃ R
λL 0 u ∗ u ∗ 0 0 0 λR
ρ
aL
JL 0 0 0 0 0 0 JR L ∗ R
∆ = (u − λ )
ρL
∆L 0 0 0 0 0 0 ∆R
aR
0 0 0 0 001 0 ∆R = (λR − u∗ ) R
ρ
4 Numerical results
The first numerical test (figure 1) is proposed in [6]. We ran this experiment
with the relaxation method and the acoustic solver [7] which is based on a
A relaxation method for the Kapila model 641
5 Conclusion
A relaxation system for the Kapila model has been devised. The numerical
implementation is based on Godunov type methods feeting the propagation
waves methods. Two experiments have been run proving the accuracy and re-
liability of the method. The future work is devoted to the implicit formulation
and the addition of viscosity effects.
References
1. A.K. Kapila, R. Menikoff, D.S. Stewart: Two-Phase Modeling of Deflagration
to Detonation Transition in Granular Materials: Reduced Equations. Physics of
Fluids, 13, 3002–3024, (2001).
2. LeVeque, Randall, Shyue, Keh-Ming: Two-dimensional front tracking based on
high resolution wave propagation methods. Journal of Computational Physics,
123, Issue 2, 354–368 (1996)
3. S. Jin, Z. P. Xin: The relaxation schemes for systems of conservation laws in
arbitrary space dimensions. Comm. Pure Appl. Math. 48, 235–278, (1995).
4. T.P. Liu: Hyperbolic conservation laws with relaxation. Comm. Math.
Phys.,108, 153–175, (1987).
5. J. Whitham Linear and Nonlinear Waves. Wiley, New-York, (1974).
6. X. Y. Hu, B. C. Khoo: An interface interaction method for compressible multi-
fluids. Journal of Heat and Mass Transfert, 45, No 6 (2001), 1287–1307
7. H. Guillard, A. Murrone: A five equation reducezd model for compressible two
phase flow problems. Technival Report RR-4778, INRIA, 2003.
8. J.U. Brackbill, D.B. Kothe, C. Zemach: A Continuum Method for Modeling
Surface Tension. Journal of Computational Physics, 100, 335–354, (1992)
Direct numerical simulation of bubbly
Taylor-Couette flow
1 Introduction
Taylor-Couette flow is a typical research target for investigating the fluid me-
chanical phenomena in shear flows. It features the flow of an incompressible,
viscous fluid contained in the gap between two concentric cylinders at low
Reynolds numbers (see references [1] and [2]). The Couette apparatus has
originally been developed for measuring the viscosity of a fluid at small im-
posed angular velocities of the cylinders. A simple sketch of the test case
geometry is shown in figure 1.
In the present study, we will concentrate on the case where the outer
cylinder (of radius ro ) is immovable and the inner cylinder (of radius ri ) is
rotating. It will be assumed that there is no axial flow. The shear stresses
on the surface of the inner rotating cylinder entrain the fluid to rotate. Cen-
trifugal forces then lead to vortical structures, their complexity depends on
the rotational speed, domain geometry and fluid properties. A wide variety of
2 Numerical Method
The velocity field u of the primary phase flow in the reactor is obtained by
DNS of the incompressible Navier-Stokes equations. For the present simula-
tions, we write them in a cylindrical coordinate system, which is the most
natural for flow problems in axisymmetric geometries.
∂rvr ∂vθ ∂vz
+ +r = 0, (1)
∂r ∂θ ∂z
∂rvz r ∂p ∂ ∂vz 1 ∂ 2 vz ∂ 2 vz
+ − ν( (r )+ + r )
∂t ρc ∂z ∂r ∂r r ∂θ2 ∂z 2
DNS of bubbly Taylor-Couette flow 645
∂vz ∂vz ∂vz
= − rvr + vθ + rvz , (2)
∂r ∂θ ∂z
∂v 3 CD ρc 1 ρc Du dv
= |u − v| (u − v) + −
∂t 4 d ρd 2 ρd Dt dt
ρc Du ρc
+ + CL (u − v) × ω, (5)
ρd Dt ρd
646 Thomas Nierhaus et al.
where ρd is the density of the dispersed phase. The terms on the right
hand side are, in order of appearance, the Stokes drag force, the virtual mass
force, the force due to the fluid pressure gradient and viscous stresses and the
slip-shear lift force. In the lift force term, ω = ∇ × u denotes the vorticity
of the fluid. We neglect gravitational effects on the bubbles, because from [8]
we see that they are not playing a major role in the flow domain of interest.
Following empirical correlations for a fluid bubble, the drag coeffcient CD is
calculated according to [6]. The lift coefficient is set to CL = 0.5 throughout
our investigations.
Equation 5 and the relation v = dx/dt form a system of six ordinary dif-
ferential equations in three space dimensions, which is solved for each bubble
using the Crank-Nicholson scheme. Sequential tracking of all bubbles in the
system is performed at each time step of the Navier-Stokes solver. Momen-
tum back-coupling from the dispersed to the continuous phase is neglected,
because the ratio of momentum between the phases is very small [7]. Also no
back-coupling through the continuity equation has been included.
The Lagrangian solver adapts the geometical partitioning strategy of the
Navier-Stokes solver (see section 2.1), to which it is coupled via standardized
data interfaces. This allows to distribute the tracking of the bubbles to the
parallel partitions according to the partitioning of the computational domain.
The parallel architecture of the Lagrangian solver is realized through MPI.
3 Results
The test case geometry and flow conditions are chosen to match the experi-
mental setup from [8], where glycerol bubbles of ρd = 4.12kg/m3 were pro-
duced by cavitation in a mixture of water-glycerol of ρc = 1088kg/m3 and
ν = 1.7 · 10−5 m2 /s. The inner and outer cylinder radii are ri = 30mm and
ro = 35mm while the reactor hight is h = 110mm. There is a fixed non-slip
wall at the bottom of the reactor while the top is formed by a free surface
where the pressure is fixed. The computational grid consists of 80×220 points
in the radial and axial directions, while 64 Fourier modes are used in the
azimuthal direction. The grid is refined near the reactor walls.
At the start of the simulation, the fluid is at rest and the inner cylinder is
put in rotation. Due to shear stress, the azimuthal velocity propagates to the
outer cylinder. Counter-rotating vortices propagate from the top and bottom
to the half-height of the cylinder. Since the reactor oprates at Re > Rec3 ,
these Taylor vortices become wavy. This trend is more pronounced near the
upper free surface, because the bottom wall limits the axial flow modulations
while the upper free surface does not limit them. An average in time and over
the azimuthal angle direction of the flow field is presented in figure 2. The
wavy Taylor vortices are clearly visible in figures 2a and 2b. Once the Taylor
cells are established, 70 000 bubbles of diameter d = 0.5mm are injected into
the flow at random initial positions. The bubbles start to follow the flow
DNS of bubbly Taylor-Couette flow 647
in the azimuthal direction. Due to centrifugal and lift forces, they migrate
toward the center of the Taylor vortices. After 100 inner cylinder rotations,
the averaged volume fraction in time and azimuthal direction shown in figure
2c is obtained. The bubbles are accumulated inside the vortices as expected.
The concentration ragions are broader with increasing axial direction, because
the wavy vortices are located in the upper part of the reactor.
Comparing our results with the experiments described in [8], we see from
figures 3a, 3b and 3c that the bubbles naturally collapse to form bubble rings
around the cylinder. These rings strongly modify the flow field by reducing
the number of Taylor cells. It is obvious that this phenomenon cannot be
reproduced numerically by Lagrangian tracking of the bubbles (see result
in figure 3d), since it fails the dispersed entities hypothesis. Phase coupling
through the continuity equation, as it is applied in [9] and has been neglected
in the present work, could lead to a better prediction of the flow field in this
case.
Conclusion
DNS of Taylor-Couette flow laden with gas bubbles has been applied. The
main point to retain from this numerical test case is that the forces acting
on the bubbles are modeled effectively and experimental observations can
be qualitatively reproduced. The accumulation of gas in the center of the
Taylor vortices modifies the flow field by reducing the number of Taylor cells,
a phenomenon which cannot be predicted with the numerical method used.
Phase coupling through the continuity equation could lead to better results.
Acknowledgements
References
1. Taylor, G.I.: Stability of a viscous liquid contained between two rotating cylin-
ders. Philos. Trans. R. Soc. London, A223, 289–343 (1923)
2. Andereck, C.D., Liu, S.S., Swinney, H.L.: Flow regimes in a circlular couette sys-
tem with independently rotating cylinders. J. Fluid Mech., 164, 155–183 (1986)
3. Krivilyov, M., Detandt, Y., Vanden Abeele, D., Degrez, G., Fransaer, J.: Direct
numerical simulations of Taylor-Couette flow using a hybrid spectral/finite ele-
ment approach. In: 7th National Congress on theoretical and applied Mechanics,
Mons, Belgium (2006)
4. Vanden Abeele, D., Degrez, G., Snyder, D.O.: Parallel turbulent flow computa-
tions using a hybrid spectral/finite element method on Beowulf clusters. In: 3rd
International Conference on Computational Fluid Dynamics, Toronto, Canada
(2004).
5. Maxey, M.R., Riley, J.J.: Equation of motion for a small rigid sphere in a nonuni-
form flow. Phys. Fluids, 26, 883–889 (1983).
6. Lain, S., Bröder, D., Sommerfeld, M., Göz, M.F.: Modelling hydrodynamics and
turbulence in a bubble column using the Euler-Lagrange procedure. Int. J. Mul-
tiphase Flow, 28, 1381–1407 (2002).
7. Nierhaus, T., Vanden Abeele, D., Deconinck, H., Planquart, P.: Modelling of tur-
bulent bubble-laden flow in a parallel plate electrochemical reactor. In: 13th In-
ternational Conference on Fluid Flow Technologies, Budapest, Hungary (2006).
8. Djeridi, H., Gabillet, C., Billard, J.Y.: Two-phase Couette-Taylor flow: Arrange-
ment of the dispersed phase and effects on the flow structures. Phys. Fluids,
16(1), 128–139 (2004).
9. Ferrante, A., Elghobashi, S.: On the physical mechanisms of drag reduction in a
spatially developing turbulent boundary layer laden with microbubbles. J. Fluid
Mech., 503, 345–355 (2004).
Numerical simulation of unsteady flow inside
an impulsively started liquid drop
1 Introduction
In this work, the transient flow around a liquid drop after injection in a qui-
escent fluid is calculated. This kind of fluid flow occurs in many natural phe-
nomena and technological processes. Numerical simulation of this problem is
important for industrial applications like atomization, ink-jet printing, powder
metallurgy, electrochemistry and the behavior of emulsions. The description
of such flows is essential if fluid flow is coupled with phase transformations
as is the case during the rapid solidification of atomized metal drops. The
scientific background on this subject goes back to the works of Stokes, and
Hadamard and Rybczynski and to the more recent works of [1, 3, 2]. However,
the majority of this work is devoted to the study of the steady flow around a
drop. The first goal of this article is to model and analyze the transient flow
inside a drop. The second goal is to find a formula for the characteristic time
of the internal transient flow useful for practical purposes.
2 Model
Suppose a fluid of viscosity µext and density ρext is moving unsteadily around
a spherical drop, having viscosity µint and density ρint . Then the governing
equations of the model are the Navier-Stokes equation (NSE) and the equa-
tion of motion. The dimensionless time-dependent NSE written in terms of a
stream function Ψ
E2Ψ E2Ψ
4 1 2 ∂Ψ ∂ ∂Ψ ∂ ∂
E Ψ + sin θ Re 2 2 − 2 2 = (E 2 Ψ ),(1)
2 ∂r ∂µ r sin θ ∂µ ∂r r sin θ ∂t
subject to
1 U 2
{e
vr , veθ } < M at r = 0; Ψ →− r (1 − µ2 ) as r → ∞; (2)
2 Uss
describes the problem of unsteady fluid flow around a liquid drop. The super-
script ˜ is used for functions inside a drop and there is no superscript for the
outer flow. Here t is time, r and θ are the radial and angular coordinates, U is
∂2 1−µ2 ∂ 2
the instantaneous drop velocity, µ = cos θ, E 2 = ∂r 2 + r 2 ∂µ2 is the Stokes
operator, Reext = 2aUss /νext and Reint = 2aUss /νint are the external and in-
ternal Reynolds numbers (a – drop radius, Uss – steady-state drop velocity,
ν – kinematic viscosity), vr = r12 ∂Ψ∂µ and vθ =
√ 1 2 ∂Ψ
∂r are the radial and
r 1−µ
tangential velocities, τrθ is the shear stress. The stream function is scaled by
Uss a2 , velocity by Uss , the radial coordinate by a and time by a2 /ν.
The instantaneous velocity of the drop is found from the Newton’s equa-
dU
tion of motion: (mint + madd ) = Fgrav + Fbuoy + Fdrag + Fhist , where mint
dt
and madd are the drop and added masses, F are the forces acted on the drop.
Incorporating the history force into the generalized drag force and express-
ing the forces through parameters of the drop, the dimensionless equation of
motion in scalar form becomes:
dÛ 1 ρ̂ − 1 Reext 3 1
= − Reext Û 2 CD , (3)
dt 2 ρ̂ + 12 Fr 16 ρ̂ + 12
2
where the superscript ˆ denotes dimensionless parameters, Fr = Uss /ag is the
Froude number, CD is the drag coefficient, ρ̂ = ρint /ρext is the density ratio.
(3) is subject to Û (t = 0) = Û0 where Û0 is the dimensionless injection
velocity.
For numerical modeling, the stream function is expanded through Leg-
endre polynomials and (1) is solved using the Van Dyke’s series-truncation
technique by the method of weighted residuals. Hermite polynomials are used
as the interpolations functions and the Crank-Nicolson scheme as the integra-
tion method [2]. The nonlinear system of equations is solved by the Newton-
Raphson method. The equation of motion is integrated by a second order
Runge-Kutta algorithm. The initial regime of flow immediately after injec-
tion is characterized by a thin boundary layer near the drop surface. In order
to describe this initial phase correctly, a boundary layer transformation of
space and time coordinates is implemented in the model [1]. Such coordinate
transformation is important for drops injected with a high initial velocity U0
and can be skipped for problems with small U0 .
3 Results
Fig. 1. Flow patterns at fixed Reext = 100 and different values of µ̂ and ρ̂.
driven by friction acting from the free-stream flow on the drop. The second
vortex is located at the back of the drop and is driven by the return flow in the
wake behind the drop. As the flow inside the drop accelerates with time due
to action of shear stresses on the surface, the ratio of the inertial to viscous
forces grows, which leads to the decay of the small vortex, Fig. 2e-f.
Fig. 2. Transient flow around a liquid drop at different dimensionless times t after
injection. The simulation is performed at Re = 100, µ̂ = 240 and ρ̂ = 2.4 × 105 .
4 Analysis
The transition from the two-vortices to single-vortex regime occurs if two con-
ditions are fulfilled. Firstly, the Reynolds number should be Re > Rec in order
to achieve separation of the external flow. Secondly, the viscosity of the drop
should be µ̂ 1. Under these conditions, the external flow reaches its steady
regime in time τext ∼ 1. The total time τint required for the internal flow to
reach its steady regime is much bigger: τint τext . The internal flow is better
described by the reduced internal Reynolds number Re∗int = Reint (vint max
/Uss ),
max
where vint is the actual (time-dependent) velocity maximum inside the drop.
Then the transient two-vortices regime is observed if Re∗int . 1. Increase of
Re∗int with time leads to the decay of the small vortex. The transition from
the two-vortices to a single vortex regime can be characterized in terms of
Unsteady flow inside a liquid drop 653
1.2
0.8
int / f(µ)
0.6
vavg
0.4
0.2
Gas bubble
Oil drop
Metal drop
0
0 0.05 0.1 0.15 0.2 0.25 0.3 0.35 0.4
(! / µ) t
avg
Fig. 3. Volume-averaged dimensionless internal velocity vint as a function of di-
mensionless time t, where f (µ) = µ if µ > 1; f (µ) = 1 if µ ≤ 1.
the stream function: its first derivative (∂Ψ/∂r)/(1 − µ2 )1/2 becomes nonneg-
ative at r = 1. This implies that the tangential velocity becomes vθ > 0 over
the drop surface. The results reported early in [3] are not fully confirmed in
the present study, i.e. we did not observe the internal flow break-up in the
steady-state regime.
The development of the internal flow is evaluated in terms of the bulk
avg
average internal velocity vint as a function of time. Results for all test cases
are presented in Fig. 3. The curve for the gas bubble shows a smooth increase
avg
of vint with time and the transitional regime takes τint = 0.1 (ρ̂/µ̂) dimen-
sionless time units. Transition times for the oil and metal drops are about
0.2 (ρ̂/µ̂) and 0.35 (ρ̂/µ̂). Hence, the dimensional characteristic time τint can
be approximated by the formula
ρ̂ a2
τint ∼ M , M = 0.1..0.5. (4)
µ̂ νext
The correspondence between the calculated transition time and this formula
is reasonable, taking into account that τint differs by 5 order of magnitude for
drops with different µ̂ and ρ̂. The dimensional average internal velocity in the
steady state is estimated by
Uss , if µ̂ > 1,
avg
vint = µ̂ (5)
Uss , if µ̂ < 1.
654 M. Krivilyov and J. Fransaer
Conclusions
The transient flow inside an impulsively started liquid drop is analyzed by
means of a numerical model. The steady flow inside a drop is always a single
Hill vortex. In case of a highly viscous drop, a transitional flow regime with
two vortices is predicted, where the smaller vortex decays with time. This
effect can be important in the rapid solidification of metal droplets. A formula
was found for the transition time of the internal flow that gives reasonable
approximation to the calculated ones.
This work was supported by the European Space Agency under the ESTEC
Contract No. 14980/01/NL/SH. We thank Dr. T. Volkmann, Dr. M. Kolbe
and Prof. D. Herlach from DLR for helpful discussions.
References
1. S.C.R. Dennis and J.D.A. Walker. Numerical solution for time-dependent flow
past an impulsively started sphere. Phys. of Fluids., 15(4):517–525, 1972.
2. D.L.R. Oliver and J.N. Chung. Flow about a fluid sphere at low to moderate
Reynolds numbers. J. Fluid Mech., 177(4):1–18, 1987.
3. V.Ya. Rivkind and G.M. Ryskin. Flow structure in motion of a spherical drop in
a fluid medium at intermediate Reynolds numbers. Fluid Mechanics, 11(1):5–12,
1976. Translated from Russian.
4. F. Sy and E.N. Lightfoot. Transient creeping flow around fluid spheres. AIChE
Journal, 17:177–191, 1971.
Accurate and Efficient Solution of 2D Steady
Water Flows with Surface Waves and
Turbulence
Summary. A surface capturing method is developed for steady water–air flow with
gravity. Second-order accuracy is obtained with flux limiting and turbulence is mod-
eled with Menter’s model. The equations are solved efficiently with a combination
of multigrid and defect correction. Results for two test cases confirm the efficiency
and accuracy of the method.
1 Introduction
Numerical simulation of steady water flow plays a continuously increasing role
in the development of ships and offshore structures. Accurate predictions of a
ship’s wave pattern, for example, can help in reducing the ship’s drag during
the design, and thus minimise operating costs [3].
A key choice in the development of a numerical model for water flow is the
representation of the water surface. The most widely used approach is surface
fitting: the computational mesh is deformed during the computation, such that
it coincides with the water surface. Another approach is the surface capturing
technique (the volume of fluid and level set methods are well-known examples)
where the grid is fixed, but the location of the water surface on the grid can
change. These methods are more flexible than surface fitting methods; they
allow steeper waves and more complex object geometries. A disadvantage is
that, for steady flow problems, the equations resulting from classical surface
capturing methods are hard to solve efficiently. Currently, they are usually
solved by time-marching the unsteady flow equations to convergence, which
is a time-consuming process.
tion laws only, which is solved with classical multigrid. Very fast convergence
of the solution process was obtained. But it also became clear that the first-
order, laminar flow model was not sufficient to accurately compute realistic
water flow.
Therefore, the method is improved by making it second-order accurate
and by adding a turbulence model. The second-order model is solved with a
defect-correction procedure, that uses the efficiency of the first-order multigrid
method. Turbulence is modeled with Menter’s model [5]: a simple yet accurate
one-equation turbulence model that has already been applied in some ship-
flow solvers.
This paper briefly describes the first-order accurate discretisation and focuses
on the extension to second-order accuracy and on the inclusion of the tur-
bulence model. The flow equations are given in Sect. 2, the discretisation of
the different terms in Sect. 3. Section 4 describes the defect correction and
the special multigrid method that are used to solve the difficult two-fluid flow
equations with turbulence. Finally, the practical application of the method
is illustrated in Sect. 5 with a test case: the water flow in a channel with a
bottom bump.
2 Flow Equations
The flow equations used here are based on the Reynolds–averaged Navier–
Stokes (RANS) equations. The distinction between the water w and air a is
made by adding a mass conservation equation for the water. Turbulence is
modeled with Menter’s model. The resulting system has conservation laws
only, which makes it suitable for multigrid solution.
∂ ∂ ∂
p + %u2 +
(%uv) = ((µ + µT ) 2ux ) +
∂x ∂y ∂x
∂
((µ + µT ) (uy + vx )) (x–mom.),
∂y
∂ ∂ ∂
p + %v 2 =
(%uv) + ((µ + µT ) (uy + vx )) +
∂x ∂y ∂x
∂
((µ + µT ) 2vy ) − %g (y–mom.),
∂y
(1)
∂ ∂
(u) + (v) = 0 (tot. mass),
∂x ∂y
∂ ∂
(uα) + (vα) = 0 (water mass),
∂x ∂y
∂ ∂ ∂
(ν̃T u) + (ν̃T v) = ((ν + ν̃T ) (ν̃T )x ) +
∂x ∂y ∂x
∂
((ν + ν̃T ) (ν̃T )y ) + P − D (Menter),
∂y
Solution of Steady Water Flows with Waves and Turbulence 657
where α is the volume fraction of water, ν = µ/% and ν̃T is a scaled version
of νT [5]. P and D are turbulence production and dissipation terms, based on
velocity derivatives.
3 Discretisation
The system (1) is discretised with a cell-centered finite-volume technique
on structured, curvilinear grids. The convective and diffusive fluxes and the
source terms are discretised separately, to independently control the stability
of each part. Both first- and second-order accurate fluxes are used. The deriva-
tion of the first-order fluxes is given in [7]. These fluxes are briefly described
here, together with a description of the second-order accurate fluxes and the
turbulence source terms.
Diffusive fluxes The diffusive fluxes in the momentum equations are mod-
eled with central differences. These are both stable and second-order accu-
rate, therefore they are used for both the first- and the second-order accurate
discretisation. For non-cartesian grids and for the cross-diffusion terms, the
velocity derivatives are computed with a control volume approach.
Source terms Two different source terms appear in the system (1). The
gravity force in the y–momentum equation is discretised as −%i g ∆x ∆y. The
production and dissipation terms in the turbulence equation contain both first-
order and second-order derivatives of the state variables, which are computed
with finite differences in a local cartesian coordinate system. When the grid is
smooth and locally (close to) orthogonal, then this approach is second-order
accurate. It is used for both the first-order and the second-order accurate
schemes.
4 Multigrid
For the solution of nonlinear systems, multigrid is one of the most powerful
techniques known today. Its application to fluid flow solutions is mature and
658 Jeroen Wackers and Barry Koren
well studied [6]. The multigrid principle is to apply a simple solution technique
or smoother to the solution on the finest grid and to copies of this solution
on underlying coarser grids, to eliminate high-frequency and low-frequency
errors, respectively.
nonlinear line smoothing and the computation of the linear operator takes
little extra work, as the linearisation is already needed for the line smoothing
on the finest grid.
Full multigrid As initial condition, we choose ν̃T very close to zero. Then,
when the solution process is started, the turbulence intensity has to grow.
Typically, we see the residual in the turbulence equation increase at first,
because the turbulence intensity increases. Then, when the boundary layers
have more or less developed, it starts to decrease. Experiments show that
multigrid relaxation is of little use in the first stage.
Therefore, full multigrid (FMG) is essential to our method: we find the
initial condition on the finest grid by computing solutions on the coarser grids,
from coarse to fine. Thus, all but the first computations start at the second
stage, with developed boundary layers. The time-consuming development of
turbulence only occurs on the coarsest grid.
5 Results
As a test case, the flow in a channel with a bottom bump is computed. Exper-
imental results for this test are given by Cahouet [1]. The flow has a Froude
number F r = 2.05 and a Reynolds number Re = 1.9 × 105 (based on inflow
water height). The top wall is modeled as a slip wall, the bottom is a no-slip
wall. The bump has a thickness of 44% of the water height. The curvilinear
grid has 128 × 512 cells and is compressed near the boundary layer and the
water surface.
In the velocity plot (Fig. 1a), we see low-velocity regions near the leading
and trailing edge of the bump and high-velocity regions above the bump and
in the air region near the top wall. The turbulent viscosity (Fig. 1b) shows
the boundary layer. Figure 1c gives a comparison of the second-order solution
with Cahouet’s measurements. Overall, the agreement is very good. The slight
difference at the back of the wave may be caused by the more developed
660 Jeroen Wackers and Barry Koren
% # #
$'*
#$* )'(
$')
#$%
# )
"
"
"
#
$'(
&$* $'(
$'%
#
#
&$'
'
&$
& $ $
!% & % ( ) !# $ # % & !# $ # % &
!
a) !
b) c) !
Fig. 1. Cahouet test case, F r = 2.05. Speed (a), turbulent viscosity ν̃T (b), and a
comparison (c) of the second-order α = 0.5 isolines (—) with Cahouet’s experiment
(averaged wave height, - -).
) #
! !
!) !#
)$ (" $# )"%
!" !$
2/34-,156
!( 1.23,+045 !(
!# !'
!' !"
!$ !&
!& !%
! "! #! $! %! ! "! #!! #"! $!!
*+,-.+*/01 )*+,-*)./0
Fig. 2. Cahouet test case, convergence of the sum of the residual for multigrid
(FMG on four grids, left) and line Gauss–Seidel smoothing on a single grid (right).
6 Conclusion
The computational results for the flow in a channel with a bottom bump, show
that the method is efficient. Comparison with experimental results indicates
that it is also accurate.
References
1. Cahouet, J.: Etude numérique et expérimentale du problème bidimensionnel
de la résistance de vagues non-linéaire. Technical report 185, Ecole Nationale
Supérieure de Techniques Avancées, Paris (1984)
2. Dick, E., Linden, J.: A multigrid method for steady incompressible Navier-Stokes
equations based on flux-difference splitting. Int. J. Num. Meth. Fluids 14, 1311–
1323 (1992)
3. Hino, T. (ed) CFD Workshop Tokyo 2005. National Maritime Research Insti-
tute, Tokyo (2005)
4. Koren, B.: A robust upwind discretization method for advection, diffusion and
source terms. In: Vreugdenhil, C.B., Koren, B. (eds) Numerical Methods for
Advection-Diffusion Problems. Vieweg (1993)
5. Menter, F.R.: Eddy viscosity transport equations and their relation to the k − ε
model. J. Fluids Engineering, 119, 876–884 (1997)
6. Trottenberg, U., Oosterlee, C.W., Schüller, A.: Multigrid. Academic Press
(2001)
7. Wackers, J., Koren, B.: A surface capturing method for the efficient computation
of steady water waves. To appear in J. Comp. Appl. Math.
Numerical analysis of bubble migration in
thermocapillary flows of an open cylindrical
container
1 Introduction
2 Numerical methods
2.1 Governing equations
In the present calculations, mass, momentum and energy equation for the
incompressible flow were applied to the thermocapillary flows and the bubble
migrations as shown in Eqs. (1), (2) and (3). These equations were described
in nondimensional forms:
∇ · u = 0, (1)
∂u Pr 2 Pr 1
ρ + (u · ∇)u = −∇p + µ∇ u − ∇k T − κn δε , (2)
∂t Ma Ma We
∂T α 2
+ (u · ∇)T = ∇ T, (3)
∂t Ma
where u, p and T are nondimensional velocity, pressure, temperature and
average curvature, respectively. The above properties of ρ, µ and α are density,
dynamic viscosity and thermal diffusivity, respectively, which are normalized
by liquid properties. The nondimensional numbers of M a, P r and W e are
Marangoni number, Prandtl number and Weber number, respectively. ∇k T
in Eq. (2) is a temperature gradient in the tangential direction on the bubble
surfaces.
Free surface
Heater 1.0
z 0.0
0.1 1.0
r
Fig. 1. 2D mesh arrangement of an open cylindrical container
3 Numerical conditions
In above equations, subscript l means the liquid phase. In the present study,
two kinds of M a number were applied to the steady thermocapillary flows as
shown in Table 1. Case 1 conditions were almost similar to the experiments
and the calculations reported by Kamotani, et al.[4] to verify the present nu-
merical method for thermocapillary flows. Since the oscillation was measured
in Case 1 condition, the smaller M a number of Cases 2 and 3 were applied to
the bubble calculations. We applied two assumptions for the bubble surface
tension; the constant surface tension (Case 2) and the temperature dependent
surface tension (Case 3): We applied Eq. (2) in Case 3 for the bubble migra-
tion, while we neglected the fourth term of the right hand side in Eq. (2) in
Case 2.
The nondimensional temperature of the heater and the side wall were
1.0 and 0.0, respectively. The top surface was assumed to be undeformable in
the present calculations. All the boundary surfaces except for the free surface
were set as no-slip conditions, while the velocities on the free surface were
given in Cartesian forms as shown in Eq.(8);
∂u ∂T ∂v ∂T
=− , =− . (8)
∂z ∂x ∂z ∂y
(a) (b)
Figures 3(a) and 3(b) show the normal and the tangential components
of the bubble surface tension in Case 3 at 386,000 steps after the bubble
calculation was started (point 0 A0 in Fig.4). The tangential components were
calculated in the 3rd term of the right hand side of Eq.(2), while the normal
ones were calculated in the 4th term. From these figures, the forces acted on
the bubble were well calculated by the present LSM model. The gas velocity
profile observed on the bubble at the same step is also shown in Fig. 3(c). The
liquid was observed to flow from the front region of the bubble, while the gas
in the bubble was observed to have two eddy centers and it flowed along the
bubble surface. The above results show the LSM was properly modeled in the
present algorithms.
(a) (b)
(c)
Fig. 3. (a) Normal and (b) tangential components of bubble surface tension and
(c) liquid and gas velocity profile observed on the bubble. All figures were plotted
at 386,000 steps after the bubble calculation was started
Numerical analysis of bubble migration 667
1.0
Non-dimensional height
0.9
A
Case 2
B
0.8
Case 3
Initial position
0.7
(0.7, 0.7)
0.6
0.3 0.4 0.5 0.6 0.7 0.8 0.9 1.0
Non-dimensional radius
The traces of the bubble center in Cases 2 and 3 are shown in Fig.4.
The bubbles flowed along the thermocapillary flow and gradually moved to
the initial eddy center, which described as the point 0 B 0 . The eddy centers
were moved from 0 B 0 to the left hand side region when the bubble migrations
were started. These phenomena were similar in the two cases, however, the
bubble in Case 3 approached closer to the heater surface than that in Case 2
because of the Marangoni effect on the bubble. The bubble velocity in Case
3 became much smaller near the free surface and much larger near the heater
compared with the velocities in Case 2. Finally, the bubbles in the two cases
continued to flow around one flow center. From these results, it became clear
that the bubble migration and the thermocapillary flow affected each other,
and Marangoni effects on the bubbles also affected the bubble migration and
its effects could not be negligible. The movie of the bubble migration in Case
2 was presented in the conference.
Temperature and velocity profiles in the radial direction are shown in
Figs. 5(a) and 5(b), respectively. These profiles were calculated in Case 2 at the
point 0 C 0 in Fig. 4. The temperature around the bubble surface became lower
since the gas temperature of the bubble was lower than the surrounding liquid
1.0 0.10
(a) (b)
Non-dimensional temperature
Non-dimentional velocity
0.8 0.08
0.6 0.06
0.4 0.04
0.2 0.02
0.0 0.00
0.2 0.4 0.6 0.8 1.0 0.2 0.4 0.6 0.8 1.0
Non-dimensional radius Non-dimensional radius
temperature. The temperature gradient near the bubble surface also became
larger than that of the other surface region. The velocity in the region between
the free surface and the bubble surface became larger because of its narrower
flow area. This larger velocity was estimated to cause the bubble acceleration
and the deformation of the thermocapillary flow.
5 Conclusions
In order to evaluate bubble migrations in thermocapillary flows in open
cylindrical containers, a calculation model was developed by applying LSM
to GSMAC-FEM and 2 dimensional calculations were performed in the mi-
crogravity conditions. The following conclusions were obtained from these
calculations:
(1) The bubble and the thermocapillary flow were affected each other and
consequently the bubble migration paths were decided,
(2) The Marangoni forces acted on the bubble surface affected the bubble mi-
gration in the thermocapillary flows and its effects could not be negligible
in the initial bubble migration.
We are also going to calculate 3 dimensional phenomena and evaluate the
interaction between more than two bubbles in the thermocapillary flows.
References
1. Kamotani, Y., Ostrach, S. and Masud, J.: Microgravity experiments and analysis
of oscillatory thermocapillary flows in cylindrical containers. Journal of Fluid
Mechanics 410, 211–233 (2000)
2. Sussman, M., Smereka, P. and Osher, S.: A level set approach for computing
solutions to incompressible two-phase flow. Journal of Computational Physics,
114, 146–159 (1994)
3. Tanahashi, T., Okanaga, H. and Sano, T.: GSMAC finite element method for
unsteady incompressible Navier-Stokes equations at high Reynolds numbers. In-
ternational Journal for Numerical Methods in Fluids, 11, 479–499 (1990)
4. Yabe, T. and Wang, P.: Unified numerical procedure for compressible and in-
compressible fluid. Journal of The Physical Society of Japan, 60, 2105–2108
(1991)
5. Ohira, H., Matsumoto, S., Mashiko, T., Yoda, S. and Kamotani, Y.: Numerical
analysis of bubble migration in thermocapillary flows of an open cylindrical
container. 25th International Symposium on Space Technology and Science, June
4–11, CD-ROM, 2006
Part VIII
Volker Hannemann
Introduction
Target configuration
The generic geometry as depicted in figure 1 is a blunted wedge with a cooled
copper nose part and a radius of 10 mm. On the plain wedge surface the
porous C/C-SiC part is embedded in non permeable C/C-SiC followed by
isolation (KAPYROK). The plain surface has an angle of attack of −30◦ The
free stream flow conditions are: Mach number 7.6, pressure 50 Pa, temperature
491 K, velocity 3730 m/s and a composition of the gas mixture in mass frac-
tions YN 2 = 0.757, YO2 = 0.012, YN O = 0.018, YN ≤ 10−6 , YO = 0.213. These
conditions lead to a laminar and except of the boundary layer almost chemi-
cally frozen flow. Pure nitrogen is used as a coolant. The surface temperature
distribution is of mayor interest, because it is measured in the experiments by
optical techniques. The investigation is focused on the steady state behavior.
Coupling model
For given distributions of the effusion mass flux and the wall temperature
the CFD calculates the wall pressure and heat flux distributions. The CSM
provides the local effusion mass flux and wall temperature for given pressure,
heat flux and known reservoir conditions summarized in table 1. An iterative
exchange between both tools at the surface of the porous medium should
converge to a steady state solution at least with a good starting solution and
sufficient under relaxation.
The flow field analysis was performed with the DLR-TAU code [4] [3],
which is a finite volume solver for the compressible, multi species Navier Stokes
equations. The implementation of the coupling interface in TAU is done inside
the boundary condition for a viscous wall. The values of a conservative variable
on a boundary can either be iterated by closing the respective conservation
equation with an adequate flux or by setting the value directly. The given
mass flux is included in the mass balance and is assumed to enter the flow
field normal to the wall, so the tangential components of the velocity are still
zero at the wall. The momentum of the gas is set according to the mass flux,
in other words the wall normal velocity is set equal to the given mass flow
Numerical investigation of an effusion cooled thermal protection material 673
divided by the iterated local density. The given temperature determines the
inner energy of the gas at the wall. The kinetic energy of the gas due to the
mass flux is determined by one half of the mass flux squared and divided by
the iterated density. For the conditions of interest in this investigation the
kinetic part stays below 0.5% of the inner energy.
Simplified model
The simplified model assumes a given mass flux distribution and iterates an
equilibrated surface temperature. Similar to an adiabatic wall or a wall in
radiative equilibrium the heat flux within the structure is neglected except the
part transfered to the coolant. This approximation is quite useful for free flight
investigations, where no heat sink is within the vehicle, and determines an
upper limit for the predicted peaks of the surface temperature. The heat flow
into the structure (normal component) will diminish the surface temperature,
where as the tangential heat flow spreads the peaks, i.e. decreases the highest
temperatures by increasing the surrounding temperature level. In wind tunnel
test models the internal heat flow can become quite complicated (cooled model
parts, instrumentation) but still neglecting the heat flow within the structure
gives a good first impression.
The mass and momentum conservation equations are closed at the bound-
ary as described in the last section. Without a prescribed temperature the
energy equation has to be closed by an expression for the energy flux across
the wall. The energy flux from the fluid onto the wall consisting of a thermal
and a diffusion part is in this simplified model equalized by the radiation part
and a part due to the coolant flow. The radiation heat flux qrad is modeled for
the simple convex geometry under investigation by the Stefan-Boltzmann law
qrad = σT 4 with the Stefan-Boltzmann constant σ and emissivity factors of
Kapyrok = 0.95 and C/C−SiC = 0.8. At least for low wall temperatures the
4 4
back ground radiation has to be taken into account: qrad = σTwall − σTbg
with a background temperature Tbg usually set to room temperature for wind
tunnel investigations.
The part due to coolant flow can be split into one part which is transported
together with the coolant through the surface (into the adjacent control vol-
ume) and another part is added to account for the cooling of the porous media.
To model the first part it is assumed that the coolant leaves the porous media
with the wall temperature and surface pressure, so the coolant mass flux mc
enters the control volume with the local enthalpy H of the flow field. For the
second part the cooling is set equal to the difference of the coolant energy ec
before and behind the porous media times the mass flux. Together the influ-
ence of the cooling gas is modeled as: qc = mc {H − ec (Twall ) + ec (Tres } where
Tres denotes the temperature of the coolant in the reservoir before entering
the porous medium.
674 Volker Hannemann
Results
To gain a first impression over the parameter space of interest, the investiga-
tion was conducted on an appropriate fixed grid. Except the new parameters
of the simplified model all other parameters were kept constant, which means
a 5 species air model in chemical non equilibrium with reaction rates accord-
ing to Gupta et al. [5] and a thermodynamic model of the species including
coupled electronic excitation, vibration and rotation (the last two only for
molecules). Blottner curve fits are used for the species viscosities and Eu-
cken’s correction for the thermal conductivity. A constant diffusion coefficient
is assumed for all species connected to the viscosity via a constant Schmidt
number of 0.7. The copper nose is assumed to be full catalytic at a constant
temperature of Twall = 500K, where as all other materials are assumed to be
non catalytic at radiative or radiative plus effusion equilibrium temperature.
The mass flux distribution is assumed to be constant due to the homogenous
porous media of constant thickness and the negligible small pressure gradient
along the surface (≈ 103 Pa/m) compared with the strong gradient within the
porous media (≈ 108 Pa/m).
The highest investigated coolant mass flux of mmaxc = 0.354 kg/(sm2 ) is
about 37% of the free stream value or about 8 % of the post shock value out-
side the boundary layer. In the following all effusion mass fluxes are related
to this value. In figure 1 the colors indicate the Mach number contours with-
out effusion mass flux and the black contour lines show the same contours
of the flow field with the maximal effusion mass flux. At this high effusion
mass flux the boundary layer thickens and separates generating a compres-
sion strong enough to change the shape of the leading shock. Figure 2 shows
the temperature distributions along the wedge surface for different effusion
mass flow rates. The dashed lines indicate the changing of the surface ma-
150
100
y [mm]
50
-50
Fig. 1. Mach number contours without effusion mass flux (colors) versus Mach
number contours with effusion mass flux (black contour lines)
Numerical investigation of an effusion cooled thermal protection material 675
1200
1000
T[K]
800
600
400
0.05 0.1 0.15 0.2
x[m]
Fig. 2. Wall temperature along the wedge at different mass flow rates: red/plus
without mass flow, black/cross with 1 mmax
c , green/square with 0.6 mmaxc , or-
ange/triangle with 0.2 mc , violet/diamond with 0.1 mmax
max
c , blue/circle with
0.01 mmax
c
terial. Starting at 500 K on the cooled nose, the temperature rises on the
isolator part and goes to even higher temperatures on the first C/C-SiC part
due to higher emission abilities (radiation cooling) of the isolator. Without ef-
fusion cooling (red line/plus), the temperature stays at a high level but with
a small decrease with increasing distance due to the growing temperature
layer and a small drop when reaching the isolator at the end. At low mass
flow rates of 0.01 mmax
c (blue/circle) the temperature stays nearly at the same
level as without cooling. With a moderate coolant flow 0.1 − 0.2 mmax c (vio-
let/diamond, orange/triangle), the temperature decreases significantly on the
effusion cooled part, then rises again on the non permeable C/C-SiC part and
shows the same small drop on the final isolator part. Between 0.3 − 0.4 mmax c
boundary layer separation starts, so that at the high effusion mass flow rates
0.6 (green/square) and 1 mmax
c (black/cross) the temperature is influenced far
ahead of the effusion cooled zone. In case of the high effusion mass flow rates
the wall temperature on and behind the cooled zone decouples from the flow
field such that the wall temperature becomes dependent only on the coolant
temperature and the (background) radiation. In the temperature distributions
at moderate mass flow rates a small peak is visible at the beginning of the
cooled zone. These peaks as well as the very small cooling effect for the low
mass flow rate originate from an increase of the heat flux due to diffusion.
When blowing air as a coolant into air instead of N2 into air the concentra-
tion gradients diminish and the temperature becomes smaller as can be seen
in picture 3.
676 Volker Hannemann
1280
1260
1240
T[K]
1220
1200
1180
0.05 0.1 0.15 0.2
x[m]
Fig. 3. Closeup view on the wall temperature at different mass flow rates: red/plus
without mass flow, blue/circle with 0.01mmax
c , green/star again 0.01mmax
c but air
instead of N2 as coolant
Concluding Remarks
A simplified one dimensional model for an effusion cooled wall was imple-
mented into the DLR-TAU code. It neglects tangential mass or heat flows
inside the porous medium as well as any heat flux leaving the porous medium
on the back side (adiabatic inner wall). The surface temperature is iterated via
equilibrating the radiation heat flux together with a simple energy gain of the
coolant versus the thermal and diffusion heat fluxes from the flow field. Un-
der these assumptions the model shows the influence of the effusion mass flow
rate on the flow field topology (separation prediction) and the effectiveness of
different coolants due to diffusion.
References
1. Mack, A., Schaefer, R., Esser, B., Guelhan, A.: Fluid Structure Interaction on a
Hypersonic Generic Body-Flap Model. Proceedings of ICCFD3, Toronto (2004)
2. Mack, A., Schaefer, R.: Fluid Structure Interaction on a Generic Body-Flap
Model in Hypersonic Flow. AIAA Journal of Spacecraft and Rockets, Vol. 42,
No.5, (2005)
3. Mack, A.: Analyse von heissen Hyperschallstroemungen um Steuerklappen mit
Fluid-Struktur Wechselwirkung. DLR-FB 2005-23 and PhD, TU Braunschweig
(2005)
4. Mack, A., Hannemann, V.: Validation of the Unstructured DLR TAU-Code for
Hypersonic Flows. AIAA 2002-3111, (2002)
5. Gupta, R.N., Yos, J.M., Thompson, R.A., Lee, K.-P.: A Review of Reaction Rates
and Thermodynamic and Transport Properties for an 11-Species Air Model for
Chemical and Thermal Nonequilibrium Calculations to 30 000 K, NASA Refer-
ence Publication 1232, (1990)
Calculation of Transport Properties
for Entry into the Martian Atmosphere
1 Introduction
The computation of entry problems into a planetary atmosphere strongly re-
lies on the models used to describe the thermal and chemical nonequilibrium
effects in the flow, the radiation phenomenom, and the surface chemistry.
Moreover, the solution of the flow field in the continuum approach also re-
quires an accurate description of the transport properties, the main focus of
this work. The transport properties (diffusion coefficients, viscosity, thermal
conductivity) are described in terms of collision integrals obtained here from
three intermolecular potential models (Tang–Toennies, Born–Mayer and (m,6)
potentials). The Martian atmosphere is essentially made of carbon dioxide
(95.7%), plus some nitrogen (2.7%) and argon (1.6%). A CFD computation is
carried out to investigate the influence of nitrogen and argon on the convective
heat flux.
2 Collision integral
At high temperatures, the macroscopic transport properties cannot be directly
measured. However, they can be derived in kinetic theory from the Boltzmann
equation. They are thus linked to the microscopic scale by collision integrals
(l,s)
Q̄ij that represent energy-averaged cross sections Q(l) of a pair of colliding
particles i and j
Z∞
(l,s) 2 (l + 1) −E (l)
Q̄ij (T ) = h i exp E s+1 Qij dE.(1)
(s + 1)! 2l + 1 − (−1)
l
(kT )s+2 kT
0
where E is the relative kinetic energy, T the gas temperature, k the Boltz-
mann constant, the indices l and s are related to the spectral method used
Z∞
dr
χ (E, b) = π − 2 b p . (3)
r2 1 − ϕ(r)/E − b2 /r2
rmin
where A and b are the parameters of the repulsive part of the potential (short
range) and C2n the dispersion coefficients for the attractive part (long range).
We propose to modify the semi-empirical damping function f2n by adding an
additional parameter z
2n
X (zbr)k
f2n (r) = 1 − e−zbr . (5)
k!
k=0
m!6
60
Born!Mayer
!"2#* (1,1) (10!20 m2) TT polarization
TT original
40 TT modified
20
0
2 3 4
10 10 10
Temperature (K)
Fig. 1. Collision integral for Ar–Ar by use of (m,6) potential from Mourits and
Rummens [6], Born–Mayer potential from Cubley and Mason [3], Tang–Toennies
potential with coefficients obtained from Patil [8], given by Tang and Toennies [9]
and a modified case of them.
6
10 m!6
potential (eV)
3 Born!Mayer
10 TT polarization
TT original
0
10 TT modified
0.01
0.005
0
!0.005
!0.01
1 2 3 4 5 6 7 8 9 10
distance (m!10)
Fig. 2. Intermolecular potential for Ar–Ar based on (m,6) potential from Mourits
and Rummens [6], Born–Mayer potential from Cubley and Mason [3], Tang–Toennies
potential with coefficients obtained from Patil [8], given by Tang and Toennies [9]
and a modified case of them.
from a smooth fit between the values obtained with both types of potential.
Another way to comptute the collision integral is to use a potential valid at
all temperatures, for instance the Tang–Toennies potential with parameters
given in the literature or by estimation of them from the polarization. The
Tang–Toennies potential based on the polarization data does not allow for a
correct description of the collision integrals at all temperatures, since it is just
an estiamtion. The Tang–Toennies potential given by Tang and Toennies [9]
fails at high temperature (more than 2000 K), due to the fact that the repul-
680 Johannes Baumgart et al.
sive part is not well represented. Finally, resonable integrals are obtained, if
the damping function of the Tang–Toennies potential is modified according
to (5). The collision integrals are solved by using a Gauss-quadrature method
and taking into account the singularity of the integrand. Frequently, it is not
possible to retrieve the correct parameters for an intermolecular potential.
For instance, combination rules can be applied for interactions between un-
like particles.
Once the intermolecular potential is known, the collisions integrals can be
solved and the transport properties can be computed. As an example, the
total thermal conductivity is presented in Fig. 3 for different gas mixtures.
The nominal mixture of Mars atmosphere (so-called MARS15) is composed
of CO2 95.7%, N2 2.7% and Ar 1.6%. During an entry, temperature rises,
molecules dissociate, recombine, and molecules and atoms ionize. Our mixture
is composed of 15 species CO2 , N2 , Ar, O2 , C, NO, O, N, CO, Ar+ , C+ ,
N+ , NO+ , O+ , e− . Three other mixtures are also presented: an 8-species
carbondioxide mixture (CO2 , O2 , C, O, CO, C+ , O+ , e− ), a 4-species nitrogen
mixture (N2 , N, N+ , e− ) and a 3-species argon mixture (Ar, Ar+ , e− ). The
Mutation library is used to compute the thermodynamic and the transport
properties based on collision integrals [5]. The thermal conductivity of the four
mixtures are shown in Fig. 3 for a gas in local thermodynamic equilibrium. It
includes three contributions: diffusion, translation, and Eucken’s correction for
the internal energy. The value of the thermal conductivity of carbon dioxide
is close to the one of the nominal mixture.
6
MARS15
5 CO2
N
2
! (W m!1 K!1)
4
Ar
3
0
0 5000 10000 15000
Temperature (K)
Fig. 3. Total thermal conductivity for martian mixture MARS15, pure N2 and Ar
at a pressure of 10 kPa
Transport properties for Martian Atmosphere 681
3 CFD simulation
The influence of the nitrogen and argon compounds on the convective heat
flux is investigated in a simulation of a Martian entry for a gas in chemical
nonequilibrium by means of a finite-volume axisymmetric code solving the
Navier-Stokes equations [1]. In our simulation, ionization is neglected: first a
pure carbon dioxide mixture (C, O, CO, O2 and CO2 ) is chosen, then for the
Martian mixture Ar, N, N2 and NO are added to the previous mixture.
The computations are carried out for the NASA MESUR (Mars Environ-
mental Survey, also called Pathfinder) geometry and flight conditions for the
peak heat load. Chen and Candler [2] performed a Navier–Stokes calculation
with surface catalysis and a CO2 –N2 gas mixture. The boundary conditions
given therein are used. The chemistry model considered in this work is taken
from Park [7]. So far a single temperature model is applied.
The freestream condition of Chen and Candler are used for an altitude
of 41.67 km, where the temperature is 160.9 K and the density 0.2687×10−3
kg/m3 for a mixture of CO2 and N2 . The velocity of the vehicle is 6155 m/s.
At the wall a non catalytic boundary condition is considered for the chemistry.
Radiation of the wall into the farfield is accounted for with an emissitivity co-
efficient of 0.9 assuming that the flow is optically transparent. The radiation of
the wall is computed with the Stefan–Boltzmann law. The velocity boundary
condition is a no-slip wall.
The surface heat flux, here shown in Fig. 4 for the non catalytic wall, is
400
Heat flux (kW m!2)
300
200
MARS
CO2
100 Chen and Candler, Parks kinetic
Chen and Candler, McKenzies kinetic
0
0 0.2 0.4 0.6 0.8 1
Distance from stagnation point (m)
Fig. 4. Surface heat flux for CO2 and MARS mixtures and non catalytic wall
compared to [2] with to different chemistry models
Conclusion
We have proposed a modified potential of Tang–Toennies valid at all tem-
peratures. We have computed the collision integrals to obtain the transport
properties of the Martian atmosphere. The results obtained in local thermo-
dynamic equilibrium or in the CFD computation did not exhibit significant
influence of argon and nitrogen on the convective heat flux. However, it should
be taken into account that some minor species such as CN or C2 may have a
strong influence in the prediction of the radiative heat flux.
References
1. P.F. Barbante. Accurate and efficient modelling of high temperature nonequilib-
rium air flows. PhD thesis, von Karman Institute for Fluid Dynamics, Rhode-
Saint-Genèse, Belgium, 2001.
2. Y.-K. Chen and G. V. Candler. Navier-stokes solutions with surface catalysis
for martian atmospheric entry. Journal of Spacecraft and Rockets, 30(1):32 – 42,
1993.
3. S.J. Cubley and E.A. Mason. Atom-molecule and molecule-molecule potentials
and transport collision integrals for high-temperature air species. The Physics
of Fluids, 18(9):1109 – 1111, 1975.
4. E. Levin and M.J. Wright. Collision integrals for ion-neutral interactions of
nitrogen and oxygen. Journal of Thermophysics and Heat Transfer, 18(1):143–
147, 2004.
5. T. Magin. A Model for Inductive Plasma Wind Tunnels. PhD thesis, von Karman
Institute for Fluid Dynamics, Rhode-Saint-Genèse, Belgium, 2004.
6. F.M. Mourits and F.H.A. Rummens. A critical evaluation of Lennard-Jones
and Stockmayer potential parameters and of some correlation methods. Can. J.
Chem., 55:3007–3020, 1977.
7. C. Park, R.L. Jaffe, and H. Partridge. Chemical-kinetic parameters of hyperbolic
Earth entry. Journal of Thermophysics and Heat transfer, 15(1):76–90, January-
March 2001.
8. S.H. Patil. Thomas-fermi model electron density with correction boundary con-
ditions: applications to atoms and ions. Atomic Data and Nuclear Data Tables,
71:41–68, 1999.
9. K.T. Tang and J.P. Toennies. The van der waals potentials between all the rare
gas atoms from He to Rn. Journal of Chemical Physics, 118:4976–4983, 2003.
Numerical simulation of supersonic flow
around a double ramp configuration and
correlation with experiment
Summary. The aim of this work is to find out optimal geometrical parameters of
a double ramp configuration in a supersonic flow in order to attain the minimum
value of pressure drag. The computations are provided by an URANS-based code
that implements several turbulence closures. The experiments were carried out at
IAG, Germany. The results of the numerical research for Mach numbers M=2.51 and
M=2.995 and for different Reynolds numbers are presented, along with a comparison
of the numerical data with the experimental ones. The obtained information allowed
to disclose three flow regimes, viz., regime with a big separation zone, with a two-
shock flow structure and a transitional regime. An optimal ramp length for each of
the flow cases has been estimated. The verifications of the computations performed
show a reasonable ability of URANS-based applications to predict the structure and
the wall pressure behavior of the detached flow.
1 Introduction
Nowadays the use of high-accuracy models such as DNS, LES and DES for
the simulation of engineering turbulent flows is confronted with high com-
puting costs. That way, advanced RANS-based models seem to be the most
feasible choice to solve the problem. This is one of the reasons for intensive
turbulence model design. Due to the semi-empirical nature of these models,
they demonstrate different levels of obtained accuracy for different applica-
tion areas. Therefore it is believed to be an urgent and important task to
verify the numerical results on a base of experimental data for a wide range
of applications.
The present work is subjected to a supersonic double ramp configuration.
From an engineering point of view the research is conditioned by a growing
interest in the development of scramjet facilities. The double ramp is a compo-
nent of a scramjet inlet used for the so-called external compression upstream
of the intake itself. The objectives of the investigation are experimental and
numerical studies of the aerodynamical characteristics of a scramjet intake
flow.
The problem offers a certain level of complexity since the configuration
has two compression angles resulting in a shock system. Thus, depending
on the geometry of the configuration, separation can occur. In order to get
minimal pressure drag along with the compression needed, conditions for a
short separation zone have to be attained.
Hence this investigation presents two aspects of the problem. From an
engineering point of view the parameters, structure and regimes of the flow
are studied. On a numerical point of view the ability of URANS methods to
produce a correct solution for detached turbulent flows is shown.
Part of this work is carried out in the frame of the Research Training
Center 1095/1 financed by DFG.
A sketch of the investigated problem and the boundary conditions are pre-
sented in Fig. 1. The calculations were conducted jointly with experimental
research [2] for three cases of flow parameters. The flow conditions of the ex-
periments are presented in Table 1. The angles of the first and second ramp
kinks are α1 = 11◦ and α2 = 9◦ , respectively. The distance d between the two
kinks is varied from 0 to 39 mm in calculations as well as in the experiment.
Upper boundary
"Simple wave" boundary conditions
Outlet
boundary
Boundary
layer edge
The numerical investigation is carried out with the original 2D code developed
at ITAM SB RAS, Novosibirsk, Russia [1]. The Favre-averaged compressible
Numerical simulation of supersonic flow around a double ramp 685
Navier-Stokes equations for ideal gas are used. Several turbulence models
for closure are implemented such as the k-ω Wilcox model [5], the BSL and
the SST models by Menter [4]. The numerical technique includes the TVD-
scheme of the Flux Vector Splitting method by van Leer with third order of
accuracy for the approximation of convective terms and second-order central
finite-difference relations for the approximation of viscous terms. The four-step
implicit finite-difference scheme of splitting according to the space directions
realized by scalar sweeps is used for the time approximation.
A regular grid with a toward-the-wall refinement is constructed. Typically
the grid consists of 100∇ · 200 nodes in y-direction and 200∇ · 400 nodes in
x-direction.
All computational results presented below have been obtained with the
aid of k-ω model unless otherwise stated.
3.1 Case 1
Fig. 2. Experimental and calculated wall pressure distributions for the case 1
2 8
1.5
7
1
Cf 10^3
6
3
0.5
4
1
0 5
2
-0.5
-5 0 5 10
x/delta
Fig. 3. Calculated wall skin friction coefficient distributions for the case 1
3.2 Case 2
The results of the computations for Mach number M=2.51 and a lower
Reynolds number Re = 9.82 · 106 , shown in Fig. 4 for the various distances be-
tween the ramp kinks, exibit that similar flow regimes are to be seen. The line
numbers correspond to the following distances: 1 – d = 0 mm, 2 – 9.43 mm,
3 – 12.36 mm, 4 – 18.336 mm, 5 – 24.26 mm, 6 – 30.23 mm, 7 – 34.66 mm, 8 –
39.24 mm. The calculated pressure distributions for this case also agree quite
well with the experimental data. However for the distances corresponding to
the transitional regime a slight data disagreement can be seen. It means that
in the computations the transition to the two-shock regime takes place at a
smaller d. This fact can be explained by an incorrect computational turbulent
viscosity level, which in the k-ω turbulence model is essentially influenced
by a freestream value of the specific turbulence dissipation rate ω. For the
two-shock regime the predicted pressure behind the interaction zone is a bit
higher than in the experiment. A possible reason of this disagreement is an
effect of non-stationarity of the shock wave position, which is not caught by
Numerical simulation of supersonic flow around a double ramp 687
the method. Analysis of the wall skin friction distributions (not shown here)
indicates that the optimal distance between the kinks is 11∇ · 12.4 mm, or,
in terms of d/δ, is 2.13∇ · 2.4.
3.0
3
2 6 7
2.5 8
Pw/P1
4 5
2.0
1
1.5
1.0
-5 0 5 10
x/delta
Fig. 4. Experimental and calculated wall pressure distributions for the case 2
Also calculations of this case for few distances between the kinks have been
performed with BSL and SST turbulence models. The BSL and SST model
formulations allow to avoid the dependency of the results on a freestream value
of turbulence model parameters. The wall pressure distributions and skin
friction distributions indicate an excess turbulent viscosuty level that leads
to a shorter separation length prediction in comparison with both measured
data and k-ω model calculations. Due to the lack of space these data are not
presented.
3.3 Case 3
Finally similar investigations have been performed for the higher Mach num-
ber 3 (Case 3). In this case a good agreement between the experiments and
the numerical computations could also be found as presented in Fig. 5. The
optimal distance between the kinks is estimated to be in the range from 31 to
40 mm.
4
4
3.5 3
1 2
3
Pw/Pw1
2.5
1.5
0.5
-20 0 20 40 60 80
x, mm
Fig. 5. Experimental and calculated wall pressure distributions for the case 3
discovered and the optimal ramp lengths have been estimated. An analysis
shows that with Re number decreasing the transition to the two-shock regime
takes place at a shorter distance between the kinks that agrees with classical
theory of viscous flows.
Future work will be directed toward a non-stationary effect investigation,
experimentally and numerically (with the aid of an URANS algorithm), since
the non-steadiness can essentially influence the flow parameters. [5].
References
1. Borisov, A.V., Fedorova, N.N.: Numerical simulation of turbulent flows near the
forward-facing steps. Thermophysics and Aeromechanics, 4, No. 1, 69–83 (1996)
2. Gaisbauer, U., Knauss, H., Wagner, S.: Experimental Investigation about Exter-
nal Compression of Highly Integrated Airbreathing Propulsion Systems. Basic
Research and Technologies for Two-Stage-to-Orbit Vehicles, WILEY-VCH, 347–
364 (2005)
3. Gaisbauer, U., Knauss, H., Wagner, S., Kharlamova, Yu.V., Fedorova, N.N.:
Experimental Investigation and Numerical Simulation of Supersonic Turbulent
Flows in Vicinity of a Double-Ramp Configuration. ICMAR: Proc. Pt 4 / Ed.
A.M. Kharitonov. -Novosibirsk: Publishing House Nonparel (2004)
4. Menter, F.R.: Two-equation eddy-viscosity turbulence models for engineering
applications. AIAA Journal, 32, No. 8, 1598–1605 (1994)
5. Wilcox, D.C.: Turbulent modeling for CFD. La Canada, California: DCW In-
dustries Inc. (1993)
Effects of Mach number on the combustion
zone length for a RAMAC configuration at
sub-detonative mode
1 Introduction
2.1 Configuration
25
Analytical estimation
Experimental data
20 Numerical prediction
Normalised pressure (P/P0)
15
10
Choke point
0
Projectile
-5
0 0.05 0.1 0.15 0.2
Axial coordinate, m
Fig.
Fig. 2. 2. Normalized
Normalized pressureratio
pressure ratio (P/P
(P/P 0 ) distributions along the projectile surface
0 ) distributions along the projectile surface
and at the centerline behind the projectile up to thermal choking point.
and at the centerline behind the projectile up to thermal choking point.
4
M2.5
3.5 M3.5
M4
2.5
Mach Number
2
L1
L2
1.5 L3
0.5
0
Rear of the projectile
-0.5
0.1 0.12 0.14 0.16 0.18 0.2
Axial coordinate [m]
Fig. 3. Numerical prediction of combustion zone length at three typical Mach num-
bers of 2.5, 3.5 and 4.
The predicted combustion zone lengthes are denoted as L1, L2, and L3, in
correspondence to the incoming Mach numbers of 2.5, 3.5 and 4. The results
(Fig. 3) show that the combustion zone length is inversely proportional to
the Mach number. Taking the combustion zone length at Mach 2 (L1) as a
reference, the combustion zone lengthes at Mach 3.5 (L2) and Mach 4 has
been decreased to about (L1 − L2)/L1 = 33% and (L1 − L3)/L1 = 44%,
respectively. As an example, the combustion zone length (L1) at Mach 2.5 is
about 20% of the control volume length used in the theoretical study. Thus the
predictions from 3D simulation provides at least some qualitative comparisons
with those obtained by Bauer [9].
The pressure contours around the projectile wall and fin surfaces are shown
in Fig. 4a. Because of the high pressure after normal shock, the flow exhibits
quite uniformity between the adjacent fins. Fig. 4b illustrates the pressure
distributions on the tube wall. It shows the zones of interaction of two shock
waves, one from the conical nose and another from the interaction of the
shocks over the fins located in each side.
694 T. Bengherbia et al.
(a) b)
Fig. 4. Contours of pressure distributions at (a) the projectile wall and fin surfaces
and (b) on the tube wall.
4 Conclusions
Numerical simulations have been carried out by solving the three-dimensional
Navier-Stokes equations, in conjunction with the chemical reactive model for
supersonic flows over a Ram Accelerator. Flows at three Mach numbers 2.5, 3.5
and 4 are considered with the sub-detonative propulsion mode applied. The
simulation reveals that the combustion zone length is inversely proportional
to the Mach numbers of 2.5, 3.5 and 4 with a significant reduction about
33% and 44% at two high Mach numbers (3.5 and 4) while referencing to the
prediction length at low Mach number 2.5, qualitatively in agreement with
other investigations.
References
1. Hertzberg A., Bruckner A.P. and Bogdanoff D.W., AIAA Journal, 26(2),
195-203 (1988).
2. Bruckner A.P. et al., J. of Propulsion and Power, 7(5), 828-836 (1991).
3. Bauer P. et al., J. Phys. IV 10, Pr11:59-67 (2000).
4. Bauer P., Knowlen C., Eur. Phys. J. Appl. Phys., 21, 233-238 (2003).
5. Bengherbia T., Yao Y.F. and Bauer P., submitted to Eur. Phys. J. Appl.
Phys. (2006).
6. Bengherbia T., Yao Y.F. and Bauer P., AIAA 2006-0558 (2006).
7. Leblanc J.E., et al., J. Phys. IV 10, Pr11:119-130 (2000).
8. Li C.P., Kailasanath K. and Oran E.S., Combustion and Flame, 108(1-2),
173-186 (1997).
9. Bauer P., Knowlen C. and Bruckner A.P., J. of Propulsion and Power,
21(5), 955-958 (2005).
10. Bauer P., Knowlen C. and Bruckner A.P., Eur. Phys. J. Appl. Phys., 29,
253-258 (2005).
A Finite Element/Finite Volume Mixed Solver
and Applications on Heat Flux Prediction∗
Key words: Discontinuous Galerkin Finite Element Method; Finite Volume Method;
FE/FV Mixed Method
1 Introduction
∗
This work is supported by Chinese National Foundation of Nature Science (No.
10321002 and No. 10472012).
In the local coordinate system, the governing equations can be written as:
∂U ∂ Ẽ ∂ F̃ ∂ G̃ ∂U 1˜
+( + + )/J = + ∇ · H̃ = 0 (1)
∂t ∂ξ ∂η ∂ς ∂t J
where U is the conservative vector, H̃ the flux (including the inviscid and
viscous parts) and J4 the Jacobion. In each element, supposing that:
X
uh (t, x, y, z) = uj (t)ϕj (x, y, z) (2)
j=1
in which the orthogonal basis for cubic cell in local system (Fig. 2) are
ϕ1 = 1, ϕ2 = ξ, ϕ3 = η, ϕ4 = ζ (3)
whereas the orthogonal basis for prismatic cell (Fig. 3) are
ϕ1 = N1 + N2 − N3 , ϕ2 = N2 + N3 − N1 , ϕ3 = N1 + N3 − N2 , ϕ4 = ζ (4)
where N1 , N2 and N3 are the area-coordinates in the bottom triangle.
Substituting relation (2) into equation (1), multiplying by the basis func-
tions, and then integrating in each cell, and setting the weighted-residual equal
to zero,Z we have: Z
d ˜ · H̃ϕj /Jdξdηdς = 0
uh ϕj dξdηdς + ∇ (5)
dt
K K
The integral in equation (5) can be solved with Gauss’s numerical integral
method. For the inviscid flux, Roe’s flux splitting scheme can be used here. To
diminish the unphysical oscillation near shock waves, limiters are introduced
for each type of cells (See previous work [9]). Meanwhile, a LU–SGS implicit
method is adopted to accelerate the convergence history.
On the grids away from boundary, a cell-centered finite volume solver[14]
is adopted here. For the integral form of governing equation, we can rewritten
it as following
∂U X
Vi + H̃ (U )f dSf = 0 (6)
∂t
f
3 Numerical Results
Two hypersonic cases are simulated to validate the mixed solver. The first case
is hypersonic flow over blunt-cone when angle of attack equals to 20◦ . The
computational conditions are same as those in Cleary’s experiment: M∞ =
10.6, Re = 1.1 × 105 , T∞ = 47.3K, Tw = 294.44K.
For this simple model, a structure grid (75×65×61, as shown in Fig. 4(a))is
used, but the data structure is transformed into unstructured format. For the
mixed solver, the DG–FE solver runs only in the inner 20 layers of grid close to
the body surface, whereas the FV solver runs on the other region. Figure 4(b)
shows the pressure contours. It can be seen that the mixed solver captures
the bow shock clearly, and there is no unphysical oscillation in the whole flow
field. In Fig. 4(c), we show the surface flow patterns obtained by the DG–FE
A Finite Element/Finite Volume Mixed Solver 699
solver and the mixed solver. They are almost same as each other, and both
of them capture the second separation clearly. The heat-flux distributions, as
well as experimental data, are shown in Fig. 5 and Fig. 6. It can be seen that
they agree with each other very well. Furthermore, we compare the CPU time
and memory for both solvers, which are listed in Table 1. It can be seen that
about 37% memory and more than 55% CPU time are saved.
Concluding Remarks
A DG–FE/FV mixed solver is presented in this work, and applied successfully
to heat flux prediction of hypersonic viscous flows. The mixed solver is set up
based on unstructured/hybrid grids, so it is suitable for simulations of flows
over complex configurations. Numerical results demonstrate the high efficiency
and accuracy of present mixed solver.
References
1. Hughes T J R.: Recent progress in the development and understanding of SUPG
methods with special reference to the compressible Euler and Navier–Stokes
equations. Int. J. Num. Meth. Fluids, 7:1261–1275(1987)
2. Donea J.: A Taylor–Galerkin method for convective transport problems. Int. J.
Num. Meth. Eng., 20:101–120(1984)
3. Wu W.Y, Cai Q.D: A new NND finite element scheme on unstructured grid,
Acta Aerodynamica Sinica, 16(1):1–13(1998)
4. Duan Z.Y., Tong B.G., Jiang G.Q.: A finite difference/finite element mixed
method and applications on aero-thermal prediction, ACTA Aerodynamica
Sinica, 15(4):1–8(1997)
5. Cockburn B., Hou S., and Shu C.W.: TVD Runge–Kutta local projection dis-
continuous Galerkin finite element method for conservation laws IV: The mul-
tidimensional case. Math. Comp., 54:545–581(1990)
6. Cockburn B., Lin S.Y., and.Shu C.W.: TVD Runge–Kutta local projection
discontinuous Galerkin finite element method for conservation laws III: One
dimensional systems. J. Comput. Phys., 84:90–113(1989)
7. Cockburn B. and Shu C.W: TVD Runge–Kutta local projection discontinuous
Galerkin finite element method for scalar conservation laws II: General frame-
work. Math. Comp., 52:411–435(1989)
8. Cockburn B. and Shu C.W.: The P1 -RKDG method for two-dimensional Euler
equations of gas dynamics. ICASE Report No.91–32(1991)
9. He L.X., Zhang L.P., Zhang H.X.: Discontinuous Galerkin finite element method
on 3D arbitrary elements, Acta Aerodynamica Sinica, accepted(2005)
10. Peter A. Gnoffo and Jeffery A. White.: Computational aerothermodynamic
simulation issues on unstructured grids. AIAA 2004–2371(2004)
11. Weatherill NP.: Unstructured grids: procedures and applications. Handbook of
Grid Generation. CRC Press(1998)[Chapter 26]
12. Pirzadeh S.: Three-dimensional unstructured viscous grids by the advancing-
layers method. AIAA J;34(1):43–9(1996)
13. Kallinderis Y, Khawaja A, McMorris H.: Hybrid prismatic/tetrahedral grid
generation for complex geometries. AIAA J;34:291–8(1996)
14. Zhang LP, Yang YJ, Zhang HX.: Numerical simulations of 3D inviscid/viscous
flow fields on Cartesian/unstructured/prismatic hybrid grids. In: Proceedings
of the 4th Asian CFD Conference, Mianyang, Sichuan, China(2000)
Numerical Analysis for Magnetic Control of
Heat-Transfer and Pressure in Hypersonic
Shock Wave Interference Flows
1 Introduction
Local heat-transfer and pressure increases due to shock wave interference are
critical problems in structural designs for complicated hypersonic vehicles.
Therefore, many experiments and computational simulations have been done
to identify the critical shock interference phenomena [1, 2, 3]. Types of the
shock interference were classified by Edney in Ref. [1]. His classification shows
that some types are severe interference and the others types are not. The most
critical interference is called Type IV.
Conventionally, the thermal protection system and the aerodynamic shape
against the shock interference of the hypersonic vehicles have been designed
based on such experimental and computational data. However, such conven-
tional design approach may increase the weight and lose the aerodynamic char-
acteristics of the vehicles. The actual shock interference phenomena around
the hypersonic vehicle change variously according to its flight conditions and
is not always severe interference. Considering the actual situation, it is bet-
ter to use a dynamic flow control method which can handle various critical
situations during the actual flight.
Recently, some dynamic control methods have been developed. A magnetic
flow control is proposed as one of such control method. This method has the
characteristic that a magnetic field is applied to ionized fluid in hypersonic
flow. Its application to a blunt body in a free stream was computationally
investigated [4, 5]. According to their research, a magnetic field shifts the
bow shock to windward and friction drag and heat-transfer-rate on the body
surface are reduced. Considering that the hypersonic shock interference may
cause ionization, the magnetic flow control can be applied to the hypersonic
2 Computational Methods
Interaction between the ionized fluid flow and the magnetic field induces a
total current. An electromagnetic field exerts the Lorentz force f and does
work P on this current. These values per unit volume are given by
f = ρe E + J × B , P = E · J. (1)
where ρe is the total charge density, J is the total current, E is the electric
field and B is the magnetic field. Neglecting magnetization and polarization
of the fluid, the total current is the sum of the convected charge ρe u and
conduction current j, i.e. J = ρe u + j, where u is the velocity of the fluid.
In general, electromagnetic phenomena are governed by Maxwell’s equa-
tions. However, in many aerodynamic problems, flow time scales are larger
than the reciprocal of the plasma frequency, flow velocities are much less than
the speed of light, and flow length scales are much larger than the Debye
length. Thus simplified Maxwell’s equations can be applied. Under these as-
sumptions, the convected charge ρe u is negligible as compared to the conduc-
tion current j and the first term of the Lorenz force is negligible as compared
to the second term. The conductive current density is assumed to be given by
Ohm’s law
j = σ (E + u × B) , (2)
where σ is the electrical conductivity.
Furthermore, low magnetic Reynolds number approximation is applied.
This approximation is based on the fact that the distortion of the magnetic
field by the flow can be neglected for many aerospace applications. Thus,
the imposed magnetic field is only important. Therefore, the governing two-
dimensional equations in generalized coordinate form are given as
Magnetic Control of Hypersonic Shock Wave Interference Flows 703
∂X ∂Fk ∂Fνk
+ = +S , k = 1, 2, (3)
∂t ∂ξk ∂ξk
where X is the conserved variable vector, Fi are inviscid flux vector and Fνi is
viscous flux vector and all quantities in Eq. (3) are nondimensionalized. The
source term S contains the nondimensional f and P written as
f =Q σ j×B , P =Q σ j·E , (4)
where the quantities in Eqs. (4) are nondimensionalized. The parameter Q in
Eqs. (4) is called interaction parameter and given as
σr B0 U∞ Lr
Q= 2
, (5)
ρ∞ U∞
where the U∞ and ρ∞ represent the free stream velocity and density and Lr
is the blunt diameter. σr and B0 are the reference electric conductivity and
the reference magnetic field magnitude. This parameter Q represents the ratio
between the inertia terms and the electromagnetic terms. In this study, this
parameter denotes a degree of a magnitude of the magnetic field. Then S can
be expressed as S = (0, f1 , f2 , f3 , P )T , where fi is the “ i th ” component of
the nondimensional Lorentz force f .
The computational schemes are the AUSMDV scheme [6] for the convective
terms and the LU-SGS scheme [7] for the time integration. The third order
upwind biased MUSCL TVD method [8] is also employed for the higher-order
approximation of the numerical flux. The magnetic field for the flow control
is two-dimensional magnetic dipole whose origin is set at the center of a blunt
body and whose dipole moment vector is directed to the opposite of free stream
direction. B0 in Eq. (5) is set the magnitude of the magnetic dipole at the top
of th blunt. The electric field is not used. In this paper, the calorically perfect
gas is assumed for qualitative investigation. Therefore, the fluid behind bow
shock wave is given a constant electric conductivity σr .
3 Flow conditions
Flow conditions for this analysis is based on the previous shock interference
study [3]. The Mach number of the free stream is M∞ = 8.03 and the Reynolds
number referring the diameter of the blunt body is Re = 5.15×105 . The angle
of the impinging shock is 18.11 [deg] where the extended line of the impinging
shock impinges at the front edge of the body. This condition cause the Type
IV interference. The Mach number behind impinging shock wave is 5.25. This
impingement is generated at inlet boundary by the oblique shock relation.
The isothermal wall boundary condition is employed on the body surface.
The wall temperature Tw is 294.44 [K] and the free stream temperature T∞
is 111.56 [K]. The simulations are done for Q = 0, 3, 6 and 9.
704 Daisuke Tsubakino, Yoshiteru Tanaka, and Kozo Fujii
4 Results
Figure 1 shows a schematic of Type IV interference. The impinging shock
impinges on the bow shock (B. S.) and generates the triple points (T. P.).
From this tripe point, the share layer (S. L.) and the lambda shock occur.
The lambda shock generates another triple point where another share layer
occurs. Then, the supersonic jet occurs between the two shear layers and
impinges on the body surface. The impingement of jet leads extremely high
heat-transfer-rate and pressure in the localized region.
B. S.
T. P. Body surface
Supersonic jet
M∞ Jet bow shock
The time averaged temperature contours for each Q are shown in Fig.
2. Light color expresses higher value. As shown in the figure, the higher the
magnetic field becomes, the larger the length of shock standoff becomes. Due
to this effect, the triple point made by the interference between the impinging
shock and the bow shock moves downward along the impinging shock. This
means the shear layer generated by this triple point impinges lower part of
the body. As the result, the interference type turns to another type from the
Type IV over Q = 6. This type is called the Type III interference and this
interference type is less severe than Type IV.
The heat-transfer-rate and the pressure distributions on body surface are
shown in Figs. 3,4. The horizontal axis θ means angles of the body surface
measured clockwise from the horizontal direction, and p and q are pressure and
heat-transfer-rate. The subscripts “0” and “w” means the values at a stagna-
tion point without a shock impingement and on the body surface, respectively.
As expected from flow fields, the heat-transfer-rate and the pressure are dras-
tically reduced by the magnetic field. From these results, it is sure that this
control method is effective.
Conclusion
The magnetic control on the shock interference flows was computationally in-
vestigated. The magnetic field can be used to change the interference type to
Magnetic Control of Hypersonic Shock Wave Interference Flows 705
(a). Q = 0 (b). Q = 3
(c). Q = 6 (d). Q = 9
12
Q0
10 Q3
8 Q6
Q9
6
q
0
-80 -60 -40 -20 0 20 40 60 80
theta
12
Q0
10 Q3
8 Q6
Q9
p 6
0
-80 -60 -40 -20 0 20 40 60 80
theta
reduce the local heat-transfer-rate and pressure on the body surface. Further-
more, it is expected that the changes of direction or position of a dipole may
produce more effective control.
References
1. Edney, B. E., “Anomalous Heat Transfer and Pressure Distributions on Blunt
Bodies at Hypersonic Speeds in the Presence of an Impinging Shock,” FFA Report
115, Aeronautical Research Institute of Sweden, Stockholm, Sweden, 1968.
2. Wieting, A. R. and Holden, M. S., “Experimental Shock-Wave Interference Heat-
ing on a Cylinder at Mach 6 and 8,” AIAA Journal , Vol. 27, No. 11, 1989,
pp. 1557–1565.
3. Yamamoto, S. and Kano, S., “Structure of Bow Shock and Compression Wave In-
teractions in Unsteady Hypersonic Shock/Shock Interference Flow,” AIAA Paper
96–2152, 1996.
4. Poggie, J. and Gaitonde, D. V., “Magnetic control of flow past a blunt body:
Numerical validation and exploration,” Physics of Fluid , Vol. 14, No. 5, 2002,
pp. 1720–1731.
5. Fujino, T., Funaki, I., Sugita, H., Mizuno, M., and Ishikawa, M., “Numerical
Analyses on Flow Control around Blunt Body “OREX” by Magnetic Field,”
AIAA Paper 2003–3760, 2003.
6. Wada, Y. and Liou, M. S., “A Flux Splitting Scheme with High-Resolution and
Robustness for Discontinuities,” AIAA Paper 94–0083, 1994.
7. Yoon, S. and Jameson, A., “Lower-Upper Symmetric-Gauss-Seidel Method for
the Euler and Navier-Stokes Equations,” AIAA Journal , Vol. 26, No. 9, 1988,
pp. 1025–1026.
8. Van Leer, B., “Toward the ultimate Conservative Difference Scheme. 4, A New
Approach to Numerical Convection,” Journal of Computational Physics, Vol. 23,
March 1977.
Numerical Simulation of a Flat-Plate
Hypersonic Shock Layer Perturbed
by External Acoustic Waves
1 Introduction
In high-velocity high-altitude flight, the thick boundary layer on the surface
of a flying vehicle merges with the bow shock wave so that, even at a large
distance from a leading edge, they may form the so-called viscous shock layer
(VSL). Like the boundary layer, the laminar VSL is unstable, and flow dis-
turbances arising and developing in this layer induce its transition to the
turbulent flow regime. However, the mechanisms governing their emergence
and development in the hypersonic VSL may differ substantially from those
investigated in supersonic flows with lower Mach numbers [1, 2]. A better
understanding of the receptivity and instability mechanisms is a necessary
condition for the development of efficient methods for controlling the laminar-
turbulent transition in a hypersonic flow around flying vehicles.
The possibilities of experimental modeling of receptivity and evolition of
disturbances in hypersonic wind tunnels are rather limited. The present paper
describes the results of direct numerical simulation of receptivity and evolution
of perturbations in a VSL on a flat plate for a very high Mach number (M∞ =
21) and a moderate Reynolds number (ReL = 1.44 × 105 ). The problem of
interaction of the VSL with free-stream acoustic disturbances of slow and fast
modes and with perturbations introduced in the vicinity of the leading edge
of the plate by means of periodic blowing and suction was studied by solving
the 2D Navier-Stokes equations. The computed results are compared with
characteristics of density fluctuations measured in experiments performed at
the same flow parameters in the T-327A hypersonic nitrogen wind tunnel
based at ITAM.
3 Results
0.2
0.15
0.1
0.05
(b)
0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 1
Fig. 1. Isolines of mean density (a) and density fluctuations (b) at A = 0.0286,
f ∗ = 19.2 kHz
The solid lines in Fig.3 show the variation along the plate of the com-
puted amplitudes of density fluctuations on the SW for slow external acoustic
disturbances of different frequencies f ∗ = 19.2 kHz, 38.4 kHz, 50 kHz, and
80 kHz. Symbols in Fig.3 show the experimentally measured amplitudes at
the same frequencies. The computational and experimental data are in good
agreement. It can also be seen from the figure that an increase in frequency
yields a nonmonotonic variation of the amplitude of density fluctuations on
the SW along the streamwise coordinate. This phenomenon can be explained
by the interaction of external flow perturbations and vortex disturbances in
the VSL, which propagate with different streamwise wavenumbers. This in-
teraction is manifested on the SW as amplitude beatings.
Variation of the angle of incidence of external acoustic waves leads to
changes in both their streamwise phase velocity and their streamwise wavenum-
ber in the external flow. This alters the interference of external acoustic dis-
Hypersonic Shock Layer Receptivity 711
0.2
0.1
(a)
0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 1
0.2
0.1
(b)
0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 1
0.2
0.1
(c)
0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 1
References
1. Ma Y., Zhong, X.: Receptivity of a supersonic boundary layer over a fflat plate.
Part 2. Receptivity to free-stream sound. J. Fluid Mech., 488, 79-121 (2003)
2. Egorov, I.V., Sudakov, V.G., Fedorov, A.V.: Numerical simulation of super-
sonic boundary-layer receptivity to acoustic disturbances. Izvestiya RAN, Mech.
Zhidk. i Gaza, 2, 43-52 (2006)
3. Suresh, A., Huynh , S.T.: Accurate monotonicity-preserving schemes with
Runge-Kutta stepping. J. Comput. Phys., 136, 83-99 (1997)
4. Kudryavtsev, A.N., Mironov, S.G., Poplavskaya, T.V., Tsyryulnikov, I.S.: Exper-
imental investigation and direct numerical simulation of evolution disturbances
in viscous shock layer on a flat plate. J. Appl. Mech. Tech. Phys., 5, 43-52(2006)
5. McKenzie, J.F., Westphal , K.O.: Interaction of linear waves with oblique shock
waves. Phys. Fluids, 11, 2350-2362(1968)
6. Maslov, A.A., Poplavskaya , T.V., Smorodsky, B.V.: Stability of a hypersonic
shock layer on a flat plate. Comptes Rendus Mecanique, 332, 875-880 (2004)
Numerical Simulation of Supersonic Turbulent
Flows over Backward–Facing Steps
Summary. The turbulent flows over 2–D backward–facing steps with different step
inclination angles at Mach numbers M = 2∇ · 5 are considered. The flows are nu-
merically simulated on the basis of the full Favre averaged Navier Stokes equations
complemented with the two–equation turbulence model. The comparison with ex-
perimental data available is performed. On the basis of computations the flow wave
picture has been constructed. Special attention is paid to the investigation of the
temperature factor influence on the structure of recirculation zone.
1 Introduction
P2/P∞
M ∞ , P∞ EF 5
M 4
SW
M 1, P1 3
M 2, P2
2
h
1
0
5 15 25 35
Fig. 1. Scheme of inviscid supersonic Fig. 2. Pressure after the BFS for dif-
flow in vicinity of BFS ferent M
lines stand for non-realizable ”strong” solutions of Euler equations for SW.
Figures shows that the pressure drop occurred in EF is fully compensated
by pressure rise in SW only for small angles α. For large α, pressure after
the SW is essentially lower than that in the incoming flow. In inviscid flow
pressure recovery takes place downstream of BFS in the zone of the EF/SW
interaction resulting in SW degeneration in Mach line.
The wave picture of viscous flow is much more complex because of presence
of boundary layer on the step surface and mixing layer over the separation
region. Reflections and deflections of EF characteristics may occur inside the
BL. As a result, EF becomes non-isentropic and uncentered, the lip shock
[7] and the slip line [4] are formed in the region of EF/BL interaction. The
experiments [6] have shown that lines of constant pressure are not straight,
and the angle of flow deflection in EF is smaller than that calculated by the
inviscid theory. Static pressure after the tail SW recoveries much earlier than
the last EF characteristic meet the tail SW. Pressure distribution along the
surface is non-monotone with pronounced local maximum that was explained
by the lip shock action [5] or the action of expansion fan emanated from the
tripple point of the λ-configuration of shock wave over the separation zone
[9].
The computations were carried out under the conditions of the ITAM wind
blow down tunnel (total temperature T 0 = 300∇ · 500K) and the hot – shot
facility (T 0 = 2000∇ · 3000K). The conditions of the experiments (incoming
Mach number M , unit Reynolds number Re1 , total pressure P 0 and temper-
ature T 0) simulated in the present paper are given in Table 1 .
The calculations have been performed on the basis of the full transient
Favre averaged Navier-Stokes equations and Wilcox two equation turbulence
model [8]. The original numerical algorithm [1] was used with the temporal
approximation based on four step finite-difference scheme of splitting accord-
ing spatial variables. At each fractional step the finite difference scheme was
Backward–Facing Steps 715
M Re1 , 106 /m P 0, KP a T 0, K
2 30 1980 280
2.9 38 418 265
3.75 58 1070 270
5 16.4 800 370
3 36.5 7300 2500
4 77.2 30000 2500
realized by scalar sweeps. The TVD scheme of Flux Vector Splitting by van
Leer [2] and AUSM scheme by Liou [3] of the third order of accuracy has
been used for the approximation of convective terms. The viscous terms have
been approximated with the central finite-difference relations of second order
of accuracy.
The calculation domain does not include the whole model but the part
of it contacted directly to the BFS. From left and right, the computational
domain is bounded with input and output sections. On the top free boundary,
so called ”simple wave” conditions are assigned provided all the disturbances
to come out the computation region. The ”soft” boundary conditions have
been set in the outlet section. In the input section, the incoming flow condi-
tions and the profiles of all gas dynamic and turbulent parameters obtained
from the computation of the turbulent boundary layer on the plate have been
assigned. The no-slip conditions for velocities are specified on the wall. Adia-
batic condition for temperature is used in the computation performed under
the condition of long-duration facility and constant temperature (”cold wall”)
for computations under the condition of hot shot facility. The regular compu-
tational grid condensed toward the solid surfaces has been constructed in the
calculation region. In the most computations, the grid consists of 500 points
in lengthwise direction and 200 points in transverse direction.
3 Computation results
The computations were carried out for α = 8◦ , 15◦ , 25◦ , 45◦ and 90◦ . At
α = 8◦ the flow is attached for all Mach numbers, but because of the inter-
action of the EF characteristics with viscous layer the points when pressure
reach the level determined by inviscid theory go downstream both for expan-
sion and compression corner. At α = 15◦ and 25◦ , a small separation zone
exist in vicinity of internal corner. Fig. 3 presents the computed static pres-
sure distributions (solid lines) for α = 25◦ and α = 45◦ together with inviscid
solutions (dashed lines) and experimental data (symbols). Lines 1 to 4 stand
for M = 2, 2.9, 3.75 and 5, correspondingly. Computed data show good agree-
ment with experimental points for M = 2.9 and 3.75. Figure shows that in
716 N.N. Fedorova, and I.A. Bedarev
Fig. 3. Static pressure distributions for α = 25◦ (a) and α = 45◦ (b) BFS
vicinity of the external corner pressure first drops and then decreases tending
to an inviscid level. A separation zone may be identified by pressure ”plateau”
region. After the reattachment static pressure rises coming to an inviscid so-
lution. But then new compression waves are formed that increase pressure up
to the incoming level. These additional compression waves can clearly be seen
in Fig. 4 where the pressure contours for M = 2.9 are presented. The origin
of these compression waves is the region where the tail shock interacts with
the mixing layer.
Computations have shown that the flow in vicinity of expansion corner is
dividing into two parts. The first, low-speed, low-density near-wall part turns
on angle α and forms the rare base region. The external layer with higher
velocity and density turns on an angle less then α and forms the mixing layer
going from the inclined step face. It was shown, that after crossing the tail
shock the flow still directs to the wall. In the region of mixing layer / tail
shock interaction the secondary compression waves appear that come to the
wall and reflect from it. Further downstream these secondary compression
waves overtake the tail shock and reflect from it as rarefaction wave. The
system of secondary waves exists far downstream from the interaction zone
causing a non-uniform pressure, skin friction and heat transfer behavior.
Next series of computations was done for α = 90◦ at M = 3 and 4.
Two kinds of flow were simulated, a ”cold” flow (T 0 = 300K) and a ”hot”
flow (T 0 = 2500K). For the hot flow, several temperature wall conditions
20
0
0
40 80 120 160
Fig. 4. Static pressure contours for α = 25◦ BFS at M = 2.9
Backward–Facing Steps 717
P/P1
1.2 Cf1.5
*10
3
2
a 1.25 3 2
1 b
1.00 1 3
0.8 1 0.75 4 1
4
0.6 0.50
0.4 0.25
0.00
0.2 2,3,4 -0.25
0 -0.50
-1 1 3 5 7 x/h9 -1 1 3 5 7 x/h
9
Fig. 5. Pressure (a) and skin friction (b) distributions for α = 90◦ BFS at M = 3
were explored, namely adiabatic conditions and ”cold” wall with different
Twall
temperature factors K = Tadiabatic .
In Figure 5 the pressure (a) and skin friction coefficient (b) distributions
along the step surface are presented for M = 3. Here 1 stands for ”cold”
flow with adiabatic temperature conditions on the wall; 2, 3 and 4 denote
numerical solutions for ”hot” flow with adiabatic wall and K = 0.8 and 0.5,
correspondingly. Figure shows that in the case of ”hot” flow the size of sepa-
ration zone is smaller and the base pressure is lower that those in the ”cold”
flow. The temperature factor has no effect on the static pressure distributions
along the wall, but on the skin friction distribution the differences between
the regimes are evident. For M = 4 case there is the same effect.
Analysis of the flow temperature fields has shown that for both consid-
ered Mach numbers the decrease of temperature factor K essentially reduce
the temperature in the recirculation zone. Figure 6 present the streamlines
for ”hot flow” at M = 3 and different wall temperature conditions: adiabatic
temperature (a) and cold wall with K = 0.5 (b). For adiabatic condition the
recirculation zone has typical two-vortex configuration with high–temperature
external vertex and relative cold near–face vortex. Decreasing the wall tem-
perature leads to significant reduction of near-face vortex and lowering the
temperatures of both vortexes. Since the experiments are conducted in the
hot shot facility with run duration of about 100 ms, such temperature condi-
2 2
a b
1.5 1.5
1 1
y/ h
y/ h
0.5 0.5
0
0 x/h 1 2 0
0 x/h 1 2
Fig. 6. Streamlines at vicinity α = 90◦ BFS at M = 3 for adiabatic (a) and cold
wall K = 0.2 (b) temperature condition
718 N.N. Fedorova, and I.A. Bedarev
References
1. Borisov, A.V. and Fedorova, N.N.: Numerical simulation of turbulent flows near
the forward-facing steps. Thermophysics and Aeromechanics, 4, 1, 69–83 (1996)
2. Van Leer, B. Flux–vector splitting for the Euler equation. Lecture Notes in Phys.
170 507–512 (1982)
3. Liou, M.S., Steffen, C.J.Jr A new flux splitting scheme J. Comp. Phys. 107
23–39 (1993)
4. Charwat, A.F. and Yakura, J.K.: An investigations of two-dimensional super-
sonic base pressures. J. Aeronautical Sci. 25, 122-128 (1958)
5. Hama, F.R.: Experimental studies of the lip shock. AIAA Jour., 6, 1086–1088
(1968)
6. Scherberg M.G., Smith H.E.: An experimental study of supersonic flow over a
rearward facing step. AIAA Jour. 5, 51–56 (1967)
7. Weinbaum, S.:Rapid Expansion of a supersonic boundary layer and its applica-
tion to the near wake. AIAA Jour., 4, 35–47 (1966)
8. Wilcox, D.C.: Turbulence modeling for CFD, DCW Ind. Inc., La Canada, Cali-
fornia (1993)
9. Zheltovodov, A.A., Shuelein, E. Kh., Horstman, C.C.: Development of the sepa-
ration uder the interaction of shock and boundary layer pertuberated by expan-
sion fan. Priikl. Mech. Tekn. Fiz.,34, N 3, 58-68 (1993)
Unified Flow Solver Combining Boltzmann and
Continuum Models for Simulations of Gas
Flows for the Entire Range of Knudsen
Numbers
1 Introduction
We have developed a Unified Flow Solver (UFS) for simulations of gas flows
for the entire range of Knudsen numbers from rarefied to continuum regimes
[1, 2]. The UFS methodology is based on the direct numerical solution of the
Boltzmann equation for rarefied flow domains [3, 4] and the kinetic schemes
of gas dynamics for the continuum flow domains [5, 6]. This approach enables
an easy coupling of kinetic and continuum solvers because similar numerical
techniques are used for solving both the Boltzmann and continuum equations.
The UFS can separate rarefied and continuum domains and use appropriate
solvers to combine efficiency of continuum models with accuracy of kinetic
models. Domain decomposition criteria and coupling algorithms are important
part of the UFS. The computational framework of the UFS is provided by
a tree-based data structure of the GERRIS Flow Solver (GFS) [7] enabling
dynamically adaptive Cartesian grid with support of complex boundaries. The
UFS can automatically generate Cartesian mesh around embedded boundaries
defined through standard files, perform dynamic adaptation of the mesh to
the solution and geometry, detect kinetic and continuum domains and select
appropriate solvers based on continuum breakdown criteria. Parallelization
of the UFS has been performed enabling dynamic load balancing for multi-
processor systems. This paper presents solutions of several test problems for
supersonic and subsonic flow regimes to illustrate current capabilities of the
UFS for different applications. UFS extensions to multi-component mixtures
of polyatomic gases are also discussed.
Fig. 1. Gas flow around a cylinder for M = 3 for different Kn numbers (0.5, 0.005).
The density profiles are on the left side, the computational grids with kinetic (red)
and continuum (white) domains are on the right side.
Fig. 2. Left plot: pressure and heat flux on the cylinder surface as a function of angle
for different values of the breakdown parameter S. Right plot: distribution of nor-
mal force over the cylinder surface for different values of the continuum breakdown
parameter S
Fig. 3. Left plot: drag coefficient versus Kn number. Comparison of UFS simulations
with experiment. Solid lines indicate the free molecular flow and continuum limits.
Right plot: pressure on the cylinder surface
Fig. 4. Mass flow rate versus Knudsen number for a channel L/d = 21 (L is length
and d is width of the channel), P0 /P1 = 2.0 (left plot), P0 /P1 = 1.5 (right plot)
case of M = 3, Kn = 0.25. One can see that all curves converge at small S
numbers when the Boltzmann region grows. At the same time, by decreasing
the S number, the computation time increases. Therefore, for quick results
one can use larger S numbers if precision of the order of 10% is satisfactory.
3 Results of simulations
tion is observed in the kinetic domain behind the body. The streamlines are
not closed due to 3D effects.
In Fig. 6 distribution of parameters of different components of air with
chemical reactions for a flow at Mach number M = 2 about the cylinder
along the central streamline are shown. The kinetic-Euler scheme with chemi-
cal reactions is used. It is considered 17 chemical reactions and 5 components,
namely O2 , N2 , O, N and N O.
References
1. Kolobov, V.I., Bayyuk, S.A., Arslanbekov, R.R., Aristov, V.V., Frolova, A.A.,
Zabelok, S.A.: Construction of a Unified Continuum/kinetic Solver for Aerody-
namic Problems, J. Spacecraft and Rockets, 42, 598 (2005)
2. Kolobov, V.I., Arslanbekov, R.R., Aristov, V.V., Frolova, A.A., Zabelok, S.A.:
Unified Flow Solver for aerospace application, AIAA Paper 2006-0988
3. Aristov., V.V.: Direct Methods for Solving the Boltzmann Equation and Study
of Non-Equilibrium Flows, Kluwer Academic Publishers, Dordrecht (2001)
4. Tcheremissine, F.G.: Conservative evaluation of Boltzmann collision integral in
discrete ordinate approximation, Computers Math. App. 35, 215 (1998)
5. Xu, K.: A Gas-kinetic BGK Scheme for the Navier-Stokes Equations and its
Connection with Artificial Dissipation and Godunov Method, J. Comp. Phys.
171, 289 (2001)
6. Ohwada, T., Fukata, S.: Simple Derivation of High-resolution Schemes for Com-
pressible Flows by Kinetic Approach, J.Comp. Phys. 211, 424 (2006)
7. http://gfs.sourceforge.net/
8. Aristov, V.V., Tcheremissine, F.G.: Solving the Euler and Navier-Stokes equa-
tions on the basis of the operator splitting of the kinetic equations, Doklady
USSR Acad. Sci. 272, 555 (1983)
9. Kolobov, V.I., Arslanbekov, R.R., Aristov, V.V., Frolova, A.A., Zabelok, S.A.:
Unified Solver for Rarefied and Continuum Flows with Adaptive Mesh and Al-
gorithm Refinement J.Comput.Phys. (submitted).
10. Maslach, G.J., Schaaf, S.A.: Cylinder Drag in the Transition from Continuum
to Free-Molecule Flow, Phys. Fluids 6, 315 (1963)
11. Lofthouse, A.J., Boyd, I.D., Wright, M.J.: AIAA Paper 2006-993
12. Moss, J.N. et al: Low-Density Aerodynamics of the Inflatable Reentry Vehicle
Experiment (IRVE), AIAA Paper 2006-1189
Adaptive Boltzmann/Navier-Stokes Hybrid
Method for Multi-Scale Gas Flow Simulation
Koji Morinishi1
1 Introduction
Numerical flow simulation about micro devices is one of the new frontiers of
computational fluid dynamics (CFD). It may gives essential knowledge about
gas flows around micro-electro-mechanical systems (MEMS). Since the flows
about the micro devices range from continuum to free molecule flow regimes,
numerical methods designed for automatically simulating the continuum to
free molecule, multi-scale flows are preferable.
Usual continuum methods can not predict flows in transitional and free
molecule regimes. While any kinetic approaches, such as the direct simulation
Monte Carlo (DSMC) method [1], can be theoretically applied to the multi-
scale flows, the approach is computatinally several orders of magnitude more
expensive than continuum methods, especially for flows in the continuum
regime. Thus an adaptive continuum/kinetic hybrid method, in which a kinetic
method is automatically used in the transitional and free molecule regimes
and a continuum method in the rest, is highly preferable for simulating the
multi-scale gas flows.
In the past decade, several studies [2, 3] about continuum/kinetic hybrid
approaches have been reported. Most of them are hybrid approaches using the
Navier-Stokes method and the DSMC method. The DSMC method has widely
been used for simulating high speed rarefied flows. The method, however,
becomes a poor simulation tool for the low speed flows about micro devices,
because huge sample size is required to reduce the inherent statistical scatter
of DSMC method to a level of the small changes of flow quantities in the low
speed flows [4].
In contrast to the DSMC method, a CFD method based on the kinetic
model Boltzmann equation [5, 6] is free from the statistical scatter. The nu-
merical Boltzmann solver is definitely superior to the DSMC method for
simulating the subsonic flows. Moreover a Navier-Stokes/Boltzmann hybrid
method, may be easily constructed, since the same CFD methods can be
adopted for the kinetic model Boltzmann equation as well as the Navier-
Stokes equations. In this paper, a novel continuum/kinetic hybrid solver is
proposed for simulating the continuum to free molecule, multi-scale flows, by
coupling the Navier-Stokes solver and the kinetic model Boltzmann solver.
Z
nu = cf dc
Z (4)
3
nT = (c − u)2 f dc
2
Z
q = (c − u)(c − u)2 f dc
3 Navier-Stokes Equations
γ µ ∂T
qi = − (14)
γ − 1 P r ∂xi
All these quantities in the continuum approach are normalized with a
reference length L, a reference density ρ∞ (= mn∞ ), a reference temperature
T∞ , and a reference velocity U∞ . The reference velocity U∞ is defined as:
p
U∞ = RT∞ (15)
where nondimensional velocities in the continuum
√ approach are greater than
those in the kinetic approach by the ratio of 2.
4 Hybrid Approach
Domain decomposition strategy is adopted for the continuum/kinetic hybrid
approach. A computational domain is decomposed into the continuum regions
and the kinetic regions. The Navier-Stokes solver works on the points in the
continuum regions and the kinetic model Boltzmann solver in the kinetic
regions.
An upwind gridless method [8] with the third order accurate weighted es-
sentially non-oscillatory (WENO) reconstruction is used for discretizing both
the kinetic model Boltzmann equation and the compressible Navier-Stokes
equations. At the interface between the continuum and kinetic regions, infor-
mation exchange is necessary at every time step over the two layers of either
side of the interface, because the stencil of the third order accurate WENO
reconstruction is generally five. At the points incontinuum regions two layer
inside from the interface, Chapman-Enskog distribution functions for the ki-
netic model Boltzman equations are constructed as:
2
1 τij Ci Cj 1 C 5 Ci ∂T
f = fe 1 − + − (16)
ν µ T Pr T 2 T ∂xi
On the other hand, the numerical flux of the Navier-Stokes equations can
be simply evaluated over the interface, since the macroscopic flow quantities
are obtained using the equation (4) at every time step even in the kinetic
regions.
As the continuum/kinetic interface indicator, the gradient-length local
Knudsen number KnGLL
λ
KnGLL = |∇ρ| (17)
ρ
and Garcia’s parameter B [2]
∗
B = max{|τij |, |qi∗ |} (18)
∗
are used. Here τij
and qi∗ are the normalized stress tensor and the normalized
heat flus, respectively.
Adaptive Boltzmann/Navier-Stokes Hybrid Method 729
2.0
1.5
1.0
0.5
2.5
2.5
0.5
1.0
2.0
1.5
Hybrid
2.0
1.5
1.0
0.5
2.5
2.5
0.5
1.0
2.0
1.5
Hybrid
5 Numerical Results
References
1. G.A. Bird. Molecular gas dynamics and the direct simulation of gas flows, Oxford
Science Publications, (1994).
2. A.L. Garcia, J.B. Bell, W.Y. Crutchfield, and B.J. Alder. Adaptive mesh and
algorithm refinement using direct simulation Monte Carlo. Journal of Computa-
tional Physics, 154, 134–155, (1999).
3. Q. Sun, I.D. Boyd, and G.V. Candler. A hybrid continuum/particle approach for
modeling subsonic, rarefied gas flows. Journal of Computational Physics, 194,
256–277, (2004).
4. E.S. Oran, C.K. Oh, and B.Z. Cybyk. Direct simulation monte carlo: Recent
advances and applications. Annual Review of Fluid Mechanics, 30, 403–442,
(1998).
5. K. Morinishi and H. Oguchi. A Computational method and its application to
analyses of rarefied gas flows. Proc. 14th International Symposium on Rarefied
Gas Dynamics, University of Tokyo Press, Vol. 1, 149–158, (1984).
6. K. Morinishi. Numerical simulation for gas microflows using Boltzmann equa-
tion. Computers & Fluids, 35, 978–985, (2006).
7. T. Abe and H. Oguchi. A hierarchy kinetic model and its applications. Proc. 10th
International Symposium on Rarefied Gas Dynamics, Progress in Astronautics
and Aeronautics, AIAA, 58 781–793, (1977).
8. K. Morinishi. Parallel computation of higher order gridless type solver. Parallel
Computational Fluid Dynamics, Elsevier, 427–434, (2003).
Computing Simulation of Hypersonic
Magneto-Fluid-Dynamics Interaction
J.S. Shang∗
1 Introduction
Numerous ideas have been advocated for using electromagnetic force for high-
speed flow control [1-4]. Unfortunately, the inefficient ionization process of
flow medium becomes the major obstacle for applying electromagnetic force
for flow control [5]. In an attempt to accentuate the magneto-fluid-dynamic
(MFD) interaction, some of the research efforts apply an externally magnetic
field to intensify the MFD interaction by invoking the Lorentz force in addi-
tion to Joule heating [1,2,3]. Surzhikov and Shang [6] have also shown that
the Hall current exerts significant influence to the plasma generation via the
electron collision process. All the aforementioned computing simulations are
in a general agreement with experimental observations by Bityurin et al [4].
Therefore, the only viable plasma flow control mechanism may require the
implementation of the plasma actuator as a small perturbation and amplifies
by aerodynamic viscous-inviscid interaction [1,2].
The most effective location of an electromagnetic perturbation is the lead-
ing edge domain where the flow is laminar and the intensity of the perturba-
tion can be relatively weak, but exerts a profound influence on the entire flow
field. The classic hypersonic flow theory by Hayes and Probstein describes the
inviscid-viscous interaction over a sharp leading edge as the pressure inter-
action [7]. The magnitude of the induced pressure is well-known and can be
calculated by a single interaction parameter χ = M 3 (C/Rx)1/2 . The viscous-
inviscid interaction is strongly amplified in hypersonic flows due to the cubic
power dependence on the freestream Mach number. This control mechanism
can be actuated in microseconds to induce a higher surface pressure domain
downstream. Results from both experimental and computational efforts have
shown the chain of events constitute a very effective hypersonic flow control
∗
Emeritus Scientist of US Air Force Research Laboratory and Research Professor
of Wright State university, joseph.shang@wright.edu
2 Analysis
In most aerodynamic application of flow control using plasma actuator, the
Magnetic Reynolds number, the Rem = µm σuL is much less than unity [2].
According to the investigated flow condition, Rem has a value of 8.48 × 10−6 ;
thus the governing equations of the low Magnetic Reynolds number approxi-
mation are fully justified for the present analysis [1,2,6]:
∂ρ
+ ∇ · (ρU ) = 0 (1)
∂t
∂ρU
+ ∇ · ρU U + pI − τ = (J × B) (2)
∂t
∂ρe
+ ∇ · [(ρe + p) U − (U · τ ) − Q] = E · J (3)
∂t
The governing equations system is identical to the Navier-Stokes equations
except the non-zero source terms. While the DCD has the maximum charged
particles number density in the cathode layer is around 8.8 × 1011 /cc, the
mass fraction is less than 10−6 . It is therefore permissible to consider only
the transport properties of the weakly ionized gas and to neglect the effects
of nonequilibrium thermodynamics and chemical kinetics. In the above for-
mulation, the electrical current density, J and electric field intensity,E that
appear in the governing equation are evaluated using a three-components,
two-temperature plasma model by Surzhikov and Shang [6].
The partially weakly ionized gas model is based on the two dominant mech-
anisms in plasma - the drift velocities and diffusion (including the ambipolar
diffusion) of discharged particles. The compatible electrical field intensity, E,
of the discharge domain is obtained by satisfying the charge conservation equa-
tion [1,6]. In globally neutral plasma field, the charge conservation equation
reduces to the Poisson equation of electrical potential, ϕ, E = −∇ϕ:
Hypersonic Magneto-Fluid-Dynamics Interaction 733
e
∇2 ϕ = (ne − ni ) (4)
ε
∂ne
+ ∇ · Γe = α(E, p)|Γe | − βne n+ (5)
∂t
∂n+
+ ∇ · Γ+ = α(E, p)|Γe | − βne n+ (6)
∂t
where the flux vectors of the charged particles densities for electrons and ions
are Γe = −De ∇ne − ne µe E and Γ+ = −D+ ∇n+ + n+ µ+ E. Once the flux
vectors are known, the electrical current density of plasma can be obtained
as J = e(Γ+ − Γe ). α and β are the Townsend discharge and recombination
coefficients. De , D+ are the electron and ion diffusion coefficients, and µe , µ+
are the mobility of the electron and ion respectively [6].
Since the wave speeds of the governing equations span an extraordinary
range from the sonic speed to speed of light, a loosely coupled solving scheme
was developed for the MFD and electro-dynamics equations. Both the govern-
ing equations can be cast into the flux vector form [1,2]. The spatial dsicretiza-
tion is based on a semi-discrete finite-volume scheme [8]. In the flux-difference
splitting procedure, the flux vectors at the control surface are written as an
exact solution to the approximate Riemann Problem.
The initial values and boundary conditions for the governing equation
system are straightforward. For the fluid dynamic variables, the no-slip con-
dition for the velocity components, the vanishing pressure pressure gradient
condition, as well as the prescribed surface temperature are imposed on the
solid surface. For the electromagnetic variables, the full electric potential is
imposed on the anode in reference to the cathode. A key element of bound-
ary condition for the DCD is specified the electron number density on the
cathode for the secondary emission phenomenon. This physical requirement
is met by requiring the normal component of the electron flux at the cathode
to be proportional to the ion counterpart: Γe · n = γΓ+ · n. Where γ is the
secondary emission coefficient which depends on the material of the cathode
and intensity of the electrical field [5]. In the earlier and present investigations,
the value of 0.1 was used [1,2,6].
Taking advantage of the symmetrical property of the rectangular inlet
and the electrodes placement, two relatively sparse mesh systems are used for
numerical resolution assessment. The coarse and fine mesh systems consist of
(85 × 45 × 81) and (105 × 57 × 101) in the ξ, η, and ζ coordinates respectively.
The solutions of the two mesh system differ by a maximum of less than 4% in
the computed Pitot pressure. All the presented numerical results are derived
from the fine-mesh calculations. For convergence acceleration, three-level mesh
sequencing of the multigrid technique is applied [9]. The data processing rate
on a 400 mHz SGI Octane workstation is 61.6 × 10−6 seconds per number
of cells per number of iterations. The convergence criterion of the present
analysis is preset at a value of 3.0 × 10−7 of the normalized global residue for
the finite-volume scheme.
734 J.S. Shang
3 Discussion of Results
The present numerical simulation duplicates the experiment in a Mach 5.15
plasma channel. The flow field is characterized by a static pressure of 78.4 Pa,
a temperature of 43 K, and the streamwise velocity of 675.5 m/s. Under these
condition, the Reynolds number per meter is 2.57×105 . The surface plasma
is generated by a Universal Voltronic DC power supply that can deliver a
current of 800 mA up to 10 kV.
The relatively small rectangular inlet model has the physical dimension of
101.6×38.1×31.1mm (L×W×H). The electrodes are embedded in the phenolic
vertical sidewalls. The cathodes are located upstream of the anodes and at a
distance of 7.9 mm downstream of the sharp leading edge. All electrodes have
the identical dimensions of 31.8×6.4×1.6 mm (length×width×depth).
The DC glow discharge over the vertical side walls is displayed in Figure
1. The surface plasma after the initial breakdown is maintained by an electric
potential of 800 V and a current of 80 mA in the external electric circuit. The
maximum charged particles number density is 8.8×1011 at a total surface DC
discharge power supply of 64 Watts. The electric current density on the anode
is 21.27 mA/cm2 which put the discharge in the abnormal mode. The DCD
in fact exhibits noticeable fluctuations in the Pitot pressure data [2]. The
dominant and visible emission is over the cathode region. This distinctive
feature is the consequence of a layer of positive space charge at the cathode
with a substantial drop of electrical potential.
In the absence of an externally applied transverse magnetic field, the prin-
cipal mechanism of the electromagnetic perturbation to the flowfield is Joule
heating. The contours of the computed Joule heating in the x-z plane is de-
Conclusion
References
1. Shang, J.S. and Surzhikov, S.T., Magneto-aerodynamic Actuator for Hypersonic
Flow Control, AIAA Journal Vol. 43, No. 8, August 2005, pp. 1633-1643.
2. Shang, J.S. Surzhikov, S.T., Kimmel, R. Gaitonde, D.V., Hayes, J.R., and
Menart, J., Mechanism of Plasma Actuators for Hypersonic Flow control,
Progress in Aerospace Sciences, Vol. 41, No. 8, Nov. 2005, pp. 642-668.
3. Macheret, S.O., Shneider, M.N., and Miles, R.B., Magnetohydrodynamic Control
of Hypersonic Flows and Scramjet Inlets Using Electron Beam Ionization, AIAA
J., Vol. 40, January 2001, pp.74-81
Hypersonic Magneto-Fluid-Dynamics Interaction 737
Summary. In the present paper R-M instability problems with high Mach number
and high density ratio are studied numerically. The compressible N-S equations in
cylindrical coordinate are approximated with a new developed sixth order accurate
WGVC scheme (weighted group velocity control). Some results of interaction be-
tween a cylindrical shock and a perturbed cylindrical material interface with shock
Mach numbers Ms = 1.2, 5, 10 and Atwood numbers A = 0.818, 0.905, 0.961,
0.980 (corresponding to ρρout
in
= 10, 20, 50, 100) are presented. The effects of shock
strength, Atwood number and mode competition on characteristics of R-M instabil-
ity are discussed.
1 Introduction
material interface onto a light fluid inside is studied numerically. Some results
for the case of shock Mach numbers M a = 1.2, 5, 10, and Atwood numbers
A = 0.818, 0.905, 0.961, 0.980 are presented. The effects of shock strength,
Atwood numbers and mode competition on R-M instability characteristics are
discussed.
(2k,0) (2k,0)
hj+1/2 − hj+1/2 = ∆x(∂f /∂x)j
+ C∆x2k [∂ 2k f /∂x2k ]j+1/2 − C∆x2k [∂ 2k f /∂x2k ]j−1/2 + O(∆x2k+2 ) (1)
where C is a known constant obtained after Taylor series expansion. With the
enlarged grid points j − k − 1, ..., j, ..., j + k + 1 after discretization of the
terms in [ ] we can get the expression
1
[∂ 2k f /∂x2k ]j+1/2 = ∆x2k [δx2k fj + δx2k fj+1 ].
2
1
where δx2k fj = fj−k − C2k fj−k+1 + · · · + (−1)m C2k
m
fj−k+m + · · · + (−1)2k fj+k .
A higher 2(k + 1)th order accurate symmetrical numerical flux can be written
as
(2k+2,0) (2k,0) C
hj+1/2 = hj+1/2 − [δx2k fj + δx2k fj+1 ]. (2)
2
If the approximation ∆x2k [∂ 2k f /∂x2k ]j+1/2 = δx2k fj−m is used, the following
(2k + 1)th order accurate upwind biased numerical flux can be obtained
(2k+1,+) (2k,0)
hj+1/2 = hj+1/2 − Cδx2k fj−m . (3)
where mε0 for the case C > 0. For the case of large integer m the scheme is
strongly upwind biased, and it may not be stable. The simplest case is m = 0.
740 Fu Dexun, Ma Yanwen, and Liang Xian
|fj+1 − fj |
σ0,j+1/2 = (9)
|fj+1 | + |fj | + ε
Equations (4), (6) and (7) correspond to the steps of extrapolation, group
velocity control and weighting, respectively. The case of k = 3 (6th order) is
used to simulate the R-M instability problems.
is the wave bifurcation stage. When the transmitted shock reaches the center
it reflects, and thus a reflected shock is formed. This reflected wave propagates
outward and interacts with the material interface again. This interaction is
known as reshock. After that a new wave bifurcation is produced. Because the
reflected shock from the center propagates in the light medium towards the
material interface, the reflected wave after collision with the material interface
becomes a new reflected shock. Finally the spike-bubble structure is formed
after multiple reshocks. In DNS we consider the mean radius of the initial
material interface r◦ as the characteristic length, the initial shock speed vs as
the characteristic velocity, and the dimensionless time is defined as t = vs t̄/r◦
The density and the temperature are normalized by the initial density and
temperature inside the material interface.
Generally speaking, when the incident shock is stronger, the speed of the mov-
ing interface, the speed of the transmitted shock and the reflected rarefaction
wave will be faster. In this case the reshock exists in the early stage of strong
shock acceleration. The effect of shock strength is discussed on the basis of
computed results for the case of single mode with a = 0.033, k1 = 12, b = 0,
Re = 50000., Ms = 1.2, 5.0, 10, and Atwood number A = 0.818, 0.905. The
variations of pressure at the cylinder center with time are given in Fig.2 for
different Mach numbers. From the figures it can be seen that the pressure at
the center grows rapidly with increasing Mach number for both cases with
and without the interface perturbation. For example, the maximal pressure
at the center for the case of Ms = 10 is 6000 times larger than the pressure
at t = 0. For the case of Ms = 1.2 the maximal pressure at the center is
only 4 times larger. Also we see that the interface instability restrains the
pressure growth at the center in the later time and the compression degree of
the interior medium C = rmin /r0 increases with increasing Mach number. In
our case C = 0.68, 0.11 and 0.05 for Ms = 1.2, 5 and 10 respectively.
From computed results we can see that the variation of the growth rate
with time is much different for different Mach numbers. Increase of the Mach
number greatly speeds up the non-linear development of the interface insta-
bility and changes the spike-bubble structures.
Fig.3 shows the variation of the growth rate of perturbation with time
for different shock Mach numbers. The growth rate is defined as ∂(rmax −
rmin )/(2∂t) where rmin and rmax correspond to the maximal and minimal
radius of the material interface. It can be seen that the increase of the shock
Mach number changes the characteristics of the growth rate of perturbation.
The density and vorticity contours at different time in the evolution process
of R-M instability are given in Fig.4 and 5 for the cases of Ms = 5 and
Ms = 10 respectively. For clarity the central parts of the contours are amplified
(the radius normalized by the mean radius of the initial material interface is
equal to one). From these figures together with Fig.1 we can see that the spike-
Numerical Simulation of R-M instability at High Mach Numbers 743
bubble structures after multiple reshocks are much different for different shock
Mach numbers. It means that increasing the shock Mach number speeds up
the nonlinear development of R-M instability, and greatly changes the spike-
bubble structures. Fig.6 shows the vorticity contours at different times for the
case of shock Mach numbers Ms = 5. From the figure we can see pairs of
contra-rotating vortices. In the middle stage after several times collisions of
the shocks with the interface, unstable phenomenon occurs at the center.
From spectral analysis it can be seen that with increasing shock Mach num-
ber the Kelvin-Helmoholtz instability resulting from R-M instability develop-
ment speeds up the excitation of small structures with high wave numbers. In
the later stage of development wide range of spectra are found. Finally, the
merging of perturbation waves with different scales leads to breakdown of the
spike-bobble structures, and a central merging region is formed.
The above presented numerical results with high resolution further confirm
the effectiveness of the high order accurate scheme developed in this paper.
The effect of double mode competition on the R-M instability with the
basic wave number of initial perturbation k1 = 12 is also discussed. From the
computed results we see that when the difference between the basic mode and
the subharmonic is large (for example, k1 = 12, k2 = 60) the competition
is weak in the spike-bubble structures, and merging can not be seen in a
short period of time. When the difference between the basic mode and the
g p p p
Conclusions
A high order accurate scheme with weighted group velocity control is devel-
oped to simulate the R-M instability problems with high shock Mach number
and high density ratio. The effects of Atwood number, shock strength and
mode competition on the characteristics of R-M instability are discussed. The
computed results verify the effectiveness of the developed schemes.
References
1. M. Brouillette, Annual Reviews, p445-468, 2002
2. N. J. Zabusky, A. D. Kolnikov, J. Fluid Mech. Vol. 475, p147-162, 2003
3. R. H. Cohen et al., Physics of Fluids, vol. 14, No.10, 2002
4. Q. Zhang and M. J. Granham, Physical Review Letters vol. 79, No.14,1997
5. Q. Zhang and M. J. Granham, Physics of Fluids, vol. 10, No. 4, 1998
6. D. Fu and Y. Ma, JCP 134:1-15, 1997
Part IX
1 Introduction
In this contribution we are concerned with the coupling between a full order
simulation and a reduced order model. The idea is to reduce the extent of the
simulation domain and hence the computational costs.
In a broad sense, there exist many applications where far from the bound-
ary the solution is weakly dependent on the details of the boundary geometry.
In such regions we use a reduced order model based on proper orthogonal de-
composition (POD) [1] to solve the problem. This approach allows a represen-
tation of the solution by a small number of unknowns that are the coefficients
of an appropriate Galerkin expansion. Therefore away from a narrow region
close to the boundary of interest the number of unknowns to be solved for,
is drastically reduced. Like all other approaches based on POD, a solution
database is necessary to build the Galerkin modes, therefore this method will
be useful when many computations for relatively similar cases are to be per-
formed, like for example in shape optimization.
where ||·|| is the norm induced by the canonical l2 scalar product, noted by
(·, ·).
PM
Solution of (3) reduces to the solution of the linear problem i=1 ai (ϕi , ϕj ) =
PM −1
(uΓ , ϕj ), 1 ≤ j ≤ M . Therefore ai = (uΓ , Pi ), where Pi = j=1 [(ϕi , ϕj )] ϕj
is a constant vector computed once for all from the POD modes.
At this point we approximate u2 with û2 and substitute in (1). Since
PM PM
B2 û2 = B2t ū2 + i=1 ai B2t ϕi , we have B2t û2 = B2t ū2 + i=1 B2t ϕi (uΓ , Pi ).
t
PM
Finally, letting Ŝ2 = i=1 B2t ϕi Pi we obtain the approximation of (2)
B1t u1 + (AΓ − Ŝ2 )uΓ = fΓ − B2t ū2 (4)
where B2t ū2 ≡ B2t A−1
2 f2 .
Matrix Ŝ2 is the approximation of the discrete
Steklov-Poincaré operator obtained by the POD expansion. Equations (2)
can of course be solved simultaneously by a standard linear solver.
The models range from Laplace equation with non-linear boundary conditions
to compressible Euler equations. Since this method can be of interest for
optimal design applications, where many different geometries must be tested
to improve performance, in some cases we have explored the idea of simulating
by usual discretization methods just the region where the geometry changes,
modelling the rest by POD.
750 Marcelo Buffoni, Haysam Telib, and Angelo Iollo
A second order finite differences (FD) method coupled to a fix point iteration is
used to solve the Laplace equation inside the square domain shown in fig. 1(a).
The left Dirichlet boundary condition is varied to build the needed database.
In particular, uL = sin (kπy) + y, 1 ≤ k ≤ 49. The boundary conditions on
the other sides are, referred to fig. 1(a), the following: on uU : u = 1, on
∂u
uD : u = 0 and on uR : u4 − u40 + ∂n = 0.
The domain is splitted at d = 1/3. Then, the Galerkin base is generated
over Ω2 using the previously built database. In order to check the accuracy
of the method, a boundary condition which was not included in the database
used to build the POD modes is imposed on the left boundary: uL = y 2 , and
a second order FD method is used to solve the problem in Ω1 , by a fix point
iteration. We use 6 POD modes to recover û2 inside Ω2 . Figure 2 presents the
result of the test by means of the distribution of the relative error between the
FD solution over the entire domain and the approximate Schur complement
approach.
(a) (b)
Fig. 2. Distribution of the relative error between the solution obtained by the
present method and the solution obtained by a second order FD method on the
whole domain. Ω1 (a), Ω2 (b).
(a) (b)
Fig. 3. Percentage error distribution over (a) Ω1 and (b) Ω2 . The reference solution
is obtained by FEM P1 elements on a triangular mesh.
(a) (b)
3.4 Discussion
References
1. J. L. Lumley, “The structure of inhomogeneous turbulent flows,” In Atmospheric
Turbulence and Radio Wave Propagation, edited by A. M. Yaglom and V. L.
Tatarski, Moscow, pp. 166–178, 1967.
2. L. Sirovich, “Turbulence and the dynamics of coherent structures. Parts I,II and
III,” Quarterly of Applied Mathematics, vol. XLV, pp. 561–590, 1987.
3. A. Quarteroni and A. Valli, Domain Decomposition Methods for Partial Dieren-
tial Equations. Oxford Science Publications, 1999.
4. O. Pironneau, F. Hecht, and A. L. Hyaric, “Freefem++, a language implemen-
tation dedicated to the finite element method.” http://www.freefem.org.
5. L. Zannetti and B. Favini, “About the numerical modelling of multidimensional
unsteady compressible flow,” Computers and Fluids, vol. 17, pp. 289–299, 1989.
6. Y. Saad, “Sparsekit: a basic tool kit for sparse matrix computations.” http:
//www-users.cs.umn.edu/~saad/software/SPARSKIT/sparskit.html.
Multi Scale Numerical Simulation of Dispersed
Reacting Flow, with application to Chemical
Vapor Deposition of Alumina
Andrey A. Markov
Summary. The micro equations are coupled with the macro balance equations for
mass, momentum, and energy and written down, using only the Eulerian approach.
The relationship between gas, surface, and solid phases is numerically simulated. The
examples of kinetics include homogeneous and heterogeneous competitive chemical
reactions and chemical condensation in a gas phase, as well as homogeneous reactions
relative to high-temperature kinetics of AlCl3 decomposition in the presence of
additives.
1 Introduction
v,0 v,00 0 00
Here νij , νij and νij , νij are appropriate stoichiometric coefficients for
j’th specie in the i’th chemical process for bulk and surface reaction respec-
vol vol
tively. The values kif , kib are the constant rates for forward and backward
vol
bulk reactions and Q denote the thermal effect. The values kif , kib , Qi are
the constant rates and thermal effects for surface reactions. Let components
Cl , ; l = 1, ..., np change the phase from gas to solid due to chemical
condensation
kl,ph
Cl −→ Pj ; l = 1, ..., np ; np ≤ n
Ql,ph
Including the phase transition, the mass rate production M˙K of specie CK for
an arbitrary set of L simultaneous surface reaction steps is presented in [3].
Let Cj and Mj are the mass fraction and molar mass of j-th specie, ρ is
gas density, JpK is the rate of condensation of specie, M˙iK
vol is the mass flux for
np
1 dρp Vb Vb X
= (1 − ) Ml Jpl (3)
Sb dt V
l=1
Equation (3) represents the mass conservation for a condensed particle that
density and volume equal ρp and Vb = 4πb3 /3 respectively. The subscripts
ex and s are referred to the values near the particle and at the particle sur-
face respectively
The heat transfer and mass transfer coefficients α ,and β ck depend on the
coordinates of a particle. These values are found resolving the flow problem
near the particle in the region b(t) < r < r∞ :
Dρck DρCp T
= ∇r ·Dk (∇r ck ) ; = ∇r · (λ∇r T )
dt dt
where Dk is the diffusivity.
We apply the boundary conditions in the form :
k1f ,Q1
1 kph 2 k2f ,Q2 kph
C3 + C4 ←→ C1 −→ P1 C3 + C4 ←→ C2 −→ P2 (5)
k1b ,−Q1 Qph k2b ,−Q2 Qph
1 2
Here, σ is the energy per unit square for particle surface creation, 0 refers
to dimensional values. The dimensionless similarity parameters that charac-
terize the nonequilibrium rate of chemical-process will be described by the
756 Andrey A. Markov
The equations are numerically solved using the implicit finite–difference ap-
proximations and the splitting of the differential operator as a sum of two
terms LD and LKI . The LD (f ) is the transfer via advection and diffusion
[3]. The high accurate time resolution for stiff chemical system is coupled
Multi Scale Numerical Simulation of Dispersed Reacting Flow 757
Fig. 2. Gas and surface phase values against time at x = 0.5, y = 0.95, T0 =
0.2, b0 = 10−6 , n0 = 5 · 108 a01 = 0.05, a02 = 0.45, c01 = 0.01, c02 = 0.49
with macro equation solver. At the current semi-integer time level we find
n+1/2
the value fs of surface phase of unknown vector f n+1 . The micro–macro
scale interaction can be sketched as shown in table 1.
High-Temperature AlCl3 thermal decomposition in the presence of addi-
tives is simulated. We consider the model of aluminium particle growth, using
the simplified model (4),(5) of surface reactions [1]. Let C1 = Al2 O3 , C2 =
(Al2 O3 )α , C3 = OAlO, C4 = AlO. Some computations are performed for to-
tal set of bulk reactions [2]. The results presented in Fig. 1 are in agreement
with experimental and theoretical ones [2].
The results of computation refer to particle number density n0 = 104 , n0 =
5 · 108 , zkb = 5, b0 = 10−3 , b0 = 10−6 . The reference temperature and molar
0
mass are T0 = 1100K and M00 = M10 , Reynolds and Peclet numbers are equal
to 10−1 .
The parameters that were not varied in simulation are listed below. The
particle parameters, activation energy, and thermal effect are as follows:
0
ρ0p /ρ0 = 10−3 , E1fv
= E1f s s
= 15, E1b s
= 19.2, E2f s
= 22.5, E2b = 28.8, E1ph =
35, E2ph = 50; KJ/mole
Let ξj be the ration of backward to forward heterogeneous reaction rate
0 0
ξj = kjb /kjf = τf s,i /τbs,i
758 Andrey A. Markov
Fig. 2 shows the time history at x = 0.5, Z = 0.95. The temperature, mass
fractions of components in gas and surface phase are presented for low thermal
effect of reactions. The parameters are as follows: T0 = 0.2, n0 = 5 · 108 , b0 =
10−6 ,zkb = 5 initial mass fractions a01 = 0.05, a02 = 0.45, c01 = 0.1, c02 =
−1 −1
0.49, and the rate parameters, ξ1 = 0.91, ξ2 = 0.91, τs,1 = 1012 , τs,2 =
12 −1 14 −1 12
5 · 10 , τp,1 = 10 , τp,2 = 10 . We see the difference between the gas and
surface values when temperature is increasing.
In the following example the initial data at t = 0 is imposed such a way
that the evaporation for B1 and condensation for B2 takes place. However, this
condition is not sufficient in total time interval as it is found in computation.
Particle components B1 (left) and B2 (right) are presented in Fig. 3 at time
t = 2.25, 7 in sections x = 0.4, x = 0.6. the parameters are as follows: b0 =
10−3 , n0 = 5 · 108 ,zkb = 5 for initial mass fractions a01 = 0.12, a02 = 0.4, c01 =
−1
0.08, c02 = 0.4 and the rate parameters, ξ1 = 0.995, ξ2 = 0.091, τs,1 =
12 −1 12 −1 14 −1 12
10 , τs,2 = 5 · 10 , τp,1 = 10 , τp,2 = 10 .
References
1. Grigor’ev, Yu. M., Doronin, S. I., and Filimonov, I. A. Kinetics of phase transition
in cooperative CVD processes. J. Chem. Phys.,18, No. 12, 25–30. 1999.
2. Catoire, L, and Swihart, M. T. High-Temperature Kinetics of AlCl3 Decomposi-
tion in the Presence of Additives for Chemical Vapor Deposition., J. Electrchem.
Society, 149, 261–267,2002.
3. Markov, A. A. Multi Scale Numerical Technique for Strongly Coupled Two-Phase
Flow. In: Proc. NURETH11 Symposium, Avignon, October, 2005.
Computational Analysis of Flow through a
Multiple Nozzle Driven Laser Cavity and
Diffuser
Summary. Computations have been carried out for simulating supersonic flow
through a set of converging-diverging nozzles with their expanding jets forming a
laser cavity and flow patterns through diffusers, past the cavity. A thorough numeri-
cal investigation with 3-D RANS code is carried out to capture the flow distribution
which comprises of shock patterns and multiple supersonic jet interactions. The
analysis of pressure recovery characteristics during the flow through the diffusers is
an important parameter of the simulation and is critical for the performance of the
laser device. The results of the computation have shown a close agreement with the
experimentally measured parameters as well as other established results indicating
that the flow analysis done is found to be satisfactory. The model is extended for
study of different parameters influencing the performance of the system, generating
useful data for an optimised design. The work assumes importance as it validates
the design considerations by the parametric analysis of the geometrical and other
aerodynamic parameters governing the performance of the laser system.
1 Introduction
The configuration selected for the analysis is taken out from a problem re-
alised from the field of work that involves using aerodynamically controlled
laser cavity for generation of high powered laser beams. In this arrangement,
in order to produce an active lasing medium, rapid and deep cooling of a
mixture of gases from a combustion chamber (CO2 − N2 − H2 O) is achieved
by allowing the gases to expand in a set of convergent-divergent nozzles, and
the allowing the supersonic jets to drive the cavity used as the zone for the
laser activity. This cavity would have a distribution of CO2 in non-equilibrium
condition with many quantum levels, providing a favorable condition for the
lasing activity. The bank of nozzles followed with the cavity followed by a
diffuser section for stable operation makes the main parts of the unit, an
aerodynamically controlled laser cavity, as being considered in the present
configuration. Fig 1. shows a 3-D view of the configuration considered.
The simulation of flow in a Gas Dynamics Laser system is done using
an industrially standard CFD tool with the code solving for 3D-RANS. The
process of the modelling is taken up in steps of increasing complexity to
ensure the numerical stability and the computational consistency. The dif-
ferent stages have provided clarity on different parameters that include the
optimal grid distribution and the optimal initial flow profiles to obtain a
stable and consistent solution. The details of the analysis are as follows.
2 Method of Analysis
view of close-up of the mesh generated near in cavity at the upstream of the
diffuser.
The objective of the study is to capture the flow patterns that provide
the details of the shock pattern, the supersonic jet interactions and multiple
reflected shock interactions along with the associated pressure and temper-
ature distributions in the laser cavity and the diffuser. The captured details
of the flow analysis include the pressure recovery patterns and shock interac-
tions in the diffuser, the pressure recovery being an important parameter for
the stable performance of the laser cavity. A study of grid independence was
obtained by a series of test runs with increasing computational nodes. A com-
prehensive study of different turbulent models was carried out to decide the
suitable scheme of turbulence. It was found that k − ωturbulence model was
most appropriate to capture all the features of the flow where as the laminar
model and k − ε turbulence and other models were inadequate to capture the
essential features of the flow.
For the present flow analysis the 3D RANS equations have been considered.
The Reynolds-Averaged Navier-Stokes Equations are solved for steady, single-
phase, compressible and viscous flow.
Mass conservation:
ρ1 U1 = ρ2 U2 = ṁ (1)
Momentum conservation:
p1 + ρ1 U1 2 = p2 + ρ2 U2 2 (2)
Energy conservation:
1 1
Cp T1 + U1 2 + q = Cp T2 + U2 2 (3)
2 2
The following boundary conditions are implemented:
1) No slip, impermeable and adiabatic wall conditions
2)The inlet pressure and temperature of the fluid entering the gas dynamic
Laser system is set as an assigned boundary value
3)The outlet boundary condition is to set with matching with the Ambient
Upwind scheme is used in obtaining the solution. The solution is obtained by
implicit successive iterative scheme. The solution is taken to have converged
with the diminishing RMS residuals reducing to less than of 10-4 on the
variables used.
762 M. A. Sriram, N. K. S. Rajan, and P. S. Kulkarni
5 Concluding Remarks
The results of the computation have captured all the essential features of
the fluid dynamics of the aerodynamically driven laser cavity and the main
thermodynamic parameters are found to have a close match to the experi-
mentally measured parameters indicating the flow modelling is satisfactory.
Optimal mesh generation and grid resolution has been obtained during the
study. The results of the analysis have effectively helped in characterising the
design details and the other aerodynamic parameters that would be useful for
further activities in the area.
1 Introduction
Magnetic Sail is a propulsion system that utilizes the solar wind in the in-
terplanetary space [1]. This propulsion system creates a large magnetic field
around the spacecraft by a super-conducting coil and generates thrust through
the interaction between the solar wind and the magnetic field. The interaction
produces the “magnetosphere” around the spacecraft and the magnetosphere
works as a sail for catching the solar wind. This propulsion system is suitable
for deep space exploration missions because this propulsion system can gen-
erate large thrust and does not need any propellant, and so many researches
about the propulsion system have been conducted recently.
The solar wind jets out of the Sun in all directions. The source of the solar
wind is the Sun’s hot corona. The temperature of the corona is so high that
the Sun’s gravity cannot hold on to it. The solar wind is full-ionized plasma
flow at speeds of about 400km/sec in the interplanetary space, and interacts
with the magnetic field of the Earth and then confines the magnetic field of
the Earth to the magnetosphere.
It is generally known that the solar wind is magnetized because of the
magnetic field of the Sun. The magnetic field in the solar wind is called the
Interplanetary Magnetic Field (IMF ; as shown in Fig.1). The strength and
direction of the IMF is changing in response to activities of the Sun. The IMF
affects the magnetosphere of the Earth and causes magnetic storms and the
aurora.
Similary, it is to be expected that the thrust of Magnetic Sail is affected by
the IMF, because Magnetic Sail generates thrust by its magnetosphere. It is
therefore important to understand how the IMF affects thrust and to estimate
amount of change in thrust. Although many researches about the interaction
between the geomagnetic field and the IMF have been conducted, there is no
reserch on the relation between thrust and the IMF.
In this study, the interaction between the dipole magnetic field by a coil
and magnetized plasma flow is simulated using magnetohydrodynamics, and
the force acting on the coil is examined.
2 Dungey Model
When a dipole magnetic field is confined within plasma flow, magnetic neutral
points exists in the magnetic field. The magnetic field reconnection occurs
at the magnetic neutral point. In the case that the IMF is in the southern
direction, a magnetic field neutral line surrounds the Earth on the equatorial
plane (as shown in Fig. 2). First, a magnetic field reconnection occurs at
a magnetic neutral point on the windward side of the Earth and the IMF is
MHD Analysis of Dipole Field in Magnetized Plasma Flow 767
connected with the geomagnetic field by this magnetic field reconnection. The
connected IMF is transported to the leeward of the Earth by the solar wind
and reconnects at a magnetic neutral point in the geomagnetic tail. Newly-
formed closed magnetic field goes back to the Earth and open magnetic field,
which does not connect the geomagnetic field, is transported to the leeward.
When the IMF is in the northern direction, magnetic neutral points are on the
leeward of the Earth (as shown in Fig. 3). Magnetic reconnections occur at
these magnetic neutral points, however the magentic field reconnections are
weaker than the reconnections in the case of southern IMF. Open magentic
fields formed by the reconnections are transported to the leeward, and closed
magnetic field region is formed around the Earth.
4 Computational Conditions
The interaction between the dipole magnetic field and the magnetized plasma
flow is simulated in two-dimensional space. The simulation box is defined to
be a rectangular coordinate system as shown in Fig. 4. Two rigid cables with
electric currents flowing in the direction perpendicular to the X-Y plane are lo-
cated, and these currents generate a dipole magnetic field in two-dimensional
space. In this simulation, the dipole magnetic moment vector is in the Y direc-
tion. Magnetized plasma flows into the box at the beginning of the simulation.
The magnetic field of the plasma flow is in the Y or minus Y direction. The
plasma flow’s magnetic field in the Y direction corresponds to the southern
IMF in the case of the Earth, and the magnetic field in the minus Y direction
corresponds to the northward IMF.
Computational parameters are in Table. 1.
Simulation area
Y
Cables with currents
5 Simulation Results
The simulation result of flow field and magnetic field is shown in Fig. 5 in the
case of the plasma flow with the southern magnetic field (in the Y direction). It
is observed in Fig. 5 that a bow shock is formed on the windward of the dipole
magnetic field and the magnetosphere is formed around the dipole. The solar
wind flows around the magentosphere without entering into it. The flow field
varies cyclically and At the reconnection point in the magnetotail, a heated
plasma ball, called plasmoid, is ejected to the downstream at regular intervals.
This simulation result is in agreement with the Dungey model denoted in
Subsect. 2.1. Figure 6 shows the time change of the force acting on the
dipole magnetic field (in other words, two cables), here the force and time
are normalized by the characterize force F0 and the characterize time t0 ,
respectively. The force is estimated by calculating the Lorentz force between
the induced current in the magnetosphere and the currents in cables [4]. In
Fig.3, the dashed line shows the force acting on the dipole magnetic field in the
non-magnetized plasma flow for comparison. In the case of non-magnetized
plasma flow, the flow field and the force acting on the dipole are in steady-state
after enough time passing. The force in the magnetized plasma flow changes
in time periodically and is larger than the force in the non-magnetized plasma
flow.
Reconnection Plasmoid
The simulation result of flow field and magnetic field is shown in Fig. 7 in the
case that the plasma flow has the northern magnetic field (in the minus Y
direction). In this case, significant changes in flow field caused by the recon-
nection does not occur, and these simulation results agree with the Dungey
model denoted in Subsect. 2.2. Figure 8 shows the time change of the force
acting on the dipole magnetic field. The force in the plasma flow with northern
magnetic field is smaller than the force in the plasma flow with no magnetic
field and changes in time, but the amplitude of the change is smaller than
that in the case of plasma flow with southern magnetic field.
Reconnection
Conclusion
In this study, the interaction between the dipole magnetic field and the mag-
netized plasma flow was simulated in order to analyze the force acting on
the dipole magnetic field. The flow field and the force change in time. These
changes are caused by the magnetic field reconnection.
References
1. R.Zubrin, D.Andrews, “Magnetic sails and interplanetary space”, Journal of
Spacecraft and Rockets, Vol.28, No.2 (1991).
2. T.Saito, Space Science Rev. 10, 319 (1969).
3. J.Raeder, .et.al., “Global simulation of the Geospace Environment Modeling
substorm challenge event”, Journal of Geophysical Research, Vol.106, No.A1,
381-395 (2001).
4. H.Nishida, .et.al., “Two-dimensional Magnetohydrodynamic Simulation of a
Magnetic Sail”, Journal of Spacecraft and Rockets, 43, 667-672 (2006).
Performance of High Order Filter Methods for
a Richtmyer-Meshkov Instability
Summary. Sixth-order compact and non-compact filter schemes that were de-
signed for multiscale Navier-Stokes, and ideal and non-ideal magnetohydrodynam-
ics (MHD) systems are employed to simulate a 2-D Rightmyer-Meshkov instability
(RMI). The suppression of this RMI in the presence of a magnetic field was investi-
gated by Samtaney (2003) and Wheatley et al. (2005). Numerical results illustrated
here exhibit behavior similar to the work of Samtaney. Due to the different amounts
and different types of numerical dissipations contained in each scheme, the struc-
tures and the growth of eddies for this chaotic-like inviscid gas dynamics RMI case
are highly grid size and scheme dependent, even with many levels of refinement.
spectral element, finite element, discontinuous Galerkin, finite volume and fi-
nite difference spatial base schemes. The type of shock-capturing scheme used
as nonlinear dissipation is very general and can be any dissipative portion of
a high resolution TVD, MUSCL, ENO, or WENO shock-capturing method.
By design, the flow sensors, spatial base schemes and linear and nonlinear dis-
sipation models are stand alone modules, and a whole class of low dissipative
high order schemes can be derived at ease. The objective of this work is to
illustrate the performance of our sixth-order low dissipation filter schemes for
a 2-D inviscid Richtmyer-Meshkov instability (RMI) problem.
RMI occurs when an incident shock accelerates an interface between two fluids
of different densities. This interfacial instability was theoretically predicted by
Richtmyer [3] and experimentally observed by Meshkov [2]. For the present
study, the RMI problem investigated by Samtaney [4] and Wheatley et al.
[6] as indicated in Fig. 1 has been chosen. The mathematical models are
the 2-D Euler gas dynamics equations and the ideal MHD equations. The
computational domain is −2 < x < 6 and 0 < y < 1. A planar shock at
x = −0.2 is moving (left to right) toward the density interface with an incline
angle of θ with the lower end initialized at x = 0. The density ratio across the
interface is denoted by η, and the nondimensional strength of the magnetic
field β = 2p0 /B02 , where the initial pressure in the preshocked regions is p0 = 1,
High Order Filter Methods for RM Instability 773
Fig. 2. Comparison between Euler gas dynamics and MHD for the 6th-order com-
pact spatial base scheme (left) and the 6th-order central (non-compact) spatial base
scheme (right) using a (801 × 101) grid at t = 3.33. MHD solutions shown are mirror
images of the original computations.
and B0 is the initial magnetic field. The initial magnetic field is uniform in the
(x, y) plane and perpendicular to the incident shock front. Numerical results
shown below are for M = 2, θ = 45o , η = 3, β −1 = 0 (Euler gas dynamics) and
β −1 = 0.5 (magnetic field present). The computation stops at an evolution
time t = 3.33. For this set of parameters and all studied numerical schemes,
instability occurs near t = 1.8 for the gas dynamics case but not for the MHD
case for the entire time evolution. Our numerical results exhibit behavior
similar to the study of Samtaney.
Computations by the sixth-order centered spatial compact base scheme
with the compact linear filter [1], in conjunction with a second step nonlin-
ear WENO5 filter (WENOfi) denoted by Comp66+Compfi+WENOfi using a
801 × 101 grid is shown in Fig. 2 (left) for the inviscid gas dynamics (GD)
and the ideal MHD equations. Here WENOfi means the dissipative portion
of the fifth-order WENO scheme (WENO5) in conjunction with our wavelet
flow sensor as the nonlinear filter [10, 12, 13]. The same computation using
the sixth-order central spatial (non-compact) base scheme in conjunction with
WENOfi denoted by CEN66+WENOfi is shown in Fig. 2 (right). The clas-
sical fourth-order Runge-Kutta method is used for the sixth-order compact
and non-compact filter schemes. For this low resolution grid, the accuracy
between the two filter methods is similar. See Fig. 3 for the grid refinement
study below. Computations using Comp66+WENOfi (i.e., without the linear
compact filter step) or Comp66+Compfii (i.e., without the nonlinear WENOfi
filter step) indicate spurious oscillations around shock regions.
The present study arrives at the same conclusion drawn in [7, 12] on the
behavior of compact spatial schemes for problems containing multiscale shock
interaction. High order compact schemes are methods of choice for many in-
compressible and low speed turbulent/acoustic flows due to their advantage
of requiring very low number of grid points per wavelength. In the presence
of multiscale shock interactions, however, this desired property of high order
compact base schemes seems to have diminished in both the gas dynamic and
774 B. Sjögreen and H.C.Yee
MHD test cases that we have studied (compared with the same order of ac-
curacy of non-compact central base schemes). Also the compact spatial base
scheme requires more CPU time per time step and it is less compatible with
parallel computations than the central spatial base scheme. Consequently, the
compact spatial base scheme requires added CPU time in a parallel computer
framework.
Figure 3 shows the inviscid gas dynamics comparison between CEN66 +
WENOfi and a second-order MUSCL for four grids (801 × 101, 1601 × 201,
3201×401, 6401×801). Here, computations using a second-order MUSCL and
a second-order Runge-Kutta method are denoted by MUSCL. Not shown are
the same computation using CEN66 as the base scheme in conjunction with
the dissipative portion of the Harten-Yee scheme and the wavelet flow sensor
as the nonlinear filter (CEN66+HYfi). For similar resolution, MUSCL requires
nearly 3 times finer grid size per spatial direction than CEN66+WENOfi and
CEN66+HYfi. The eddy structures are different among the three filter meth-
ods and they are very different from the Samtaney adaptive mesh refinement
(AMR) simulation with an equivalent uniform grid of 16384 × 2048. Note that
except for WENOfi, the van Leer version of the van Albada limiter is used.
For the second-order MUSCL scheme, the limiter is applied to the primitive
variables. All methods employed the Roe’s approximate Riemann solver for
the gas dynamics case and the Gallice approximate Riemann solver for the
MHD case using our method of solving the conservative MHD system [9].
The present study can serve as a good example of failure of grid refinement
for unsteady chaotic-like inviscid flow. As the grid is refined (in conjunction
with different amounts and types of numerical dissipations contained in each
scheme), smaller and smaller eddies are formed which combine to affect global
flow through the energy cascade effect. Future work, including viscous effects
and extensive comparisons is in progress [14].
3 Concluding Remarks
The efficiency, accuracy and flexibility of the present class of low dissipative
high order filter schemes rest on the fact that the multistep filter can be ap-
plied after a full time step of the Runge-Kutta method, and that a wavelet
flow sensor is employed to control the amount of numerical dissipation where
needed. The major CPU time intensive part of the shock-capturing computa-
tion is the nonlinear filter. In fact, the filter procedure requires slightly more
CPU time than the Harten-Yee and MUSCL schemes. This is due to the fact
that all sixth-order (or any order) filter schemes require only one Riemann
solve per time step per direction (independent of the time discretizations of
the base scheme step) as opposed to two Riemann solves per time step per
direction by the MUSCL and Harten-Yee schemes using a second-order Runge-
Kutta method. WENO5-RK4 requires at least twice the CPU time of all other
methods since four Riemann solves per time step per direction are required by
WENO5-RK4. RK4 stands for the classical 4th-order Runge-Kutta temporal
discretization and WENO5 is the 5th-order WENO spatial scheme.
References
1. D.V. Gaitonde and M.R. Visbal, Further Development of a Navier-Stokes Solu-
tion Procedure Based on Higher-Order Formulas, AIAA Paper 99-0557, 1999.
2. Y.Y. Meshkov, Instability of a Shock Wave Accelerated Interface Between Two
Gases, NASA Tech. Trans. NASA TT F-13074, (1970).
3. R.D. Richtmyer, Taylor Instability in Shock Acceleration of Compressible Fluids,
Commun. Pure Appl. Math. 13, (1960), 297.
4. R. Samtaney, Suppression of the Richtmyer-Meshkov Instability in the Presence
of a Magnetic Field, Physics of Fluids, 15 (2003), L53-L56.
5. B. Sjögreen and H. C. Yee, Multiresolution Wavelet Based Adaptive Numer-
ical Dissipation Control for Shock-Turbulence Computation, RIACS Technical
Report TR01.01, NASA Ames research center (Oct 2000); also, J. Scient. Com-
puting, 20, (2004), 211-255.
6. V. Wheatley, D.I. Pullin and R. Samtaney Regular Shock Refraction at an oblique
Planar Density Interface in Magnetohydrodynamics, J. Fluid Mech. 522, (2005),
179-214.
7. H.C. Yee, N.D. Sandham, N.D., and M.J. Djomehri, Low Dissipative High Order
Shock-Capturing Methods Using Characteristic-Based Filters, J. Comput. Phys.,
150 (1999) 199-238.
776 B. Sjögreen and H.C.Yee
Summary. High fidelity computations were carried out to analyze the orbiter LH2
feedline flowliner. Computations were performed on the Columbia platform which is
a 10,240-processor supercluster consisting of 20 Altix nodes with 512 processors each.
Various computational models were used to characterize the unsteady flow features
in the turbopump, including the orbiter Low-Pressure-Fuel-Turbopump (LPFTP)
inducer. Unsteady flow originating from the orbiter LPFTP inducer is one of the
major contributors to the high frequency cyclic loading that results in high cycle
fatigue damage to the gimbal flowliners just upstream of the LPFTP. An incom-
pressible Navier-Stokes flow solver INS3D, based on the artificial compressibility
method, was used to compute the flow of liquid hydrogen in each test article.
1 Introduction
The incompressible Navier-Stokes flow solver, INS3D [2], [3], [4], based on the
artificial compressibility method [1] was used to compute the flow of liquid
hydrogen in several test geometries including the feed line and an inducer
for pumping. To obtain time-accurate solutions, the equations are iterated
to convergence in pseudo-time for each physical time step until the diver-
gence of the velocity field has been reduced below a specified tolerance value.
The total number of sub-iterations required varies depending on the problem,
the time step size and the artificial compressibility parameter used. Typically
the number ranges from 10 to 30 sub-iterations. The computational require-
ment for the model problem is large, and thus two distinct parallel processing
paradigms have been implemented. These include the Multi-Level Parallel
(MLP) [5] and the MPI/OpenMP [6] hybrid parallel programming models.
Both models contain coarse and fine grain parallelism. Coarse grain parallelism
is achieved through a UNIX fork in MLP and through explicit message pass-
ing in the MPI/OpenMP hybrid code. Fine grain parallelism is achieved us-
ing OpenMP compiler directives in both the MLP and MPI/OpenMP hybrid
codes. Computations were performed to compare the scalability between the
MLP and MPI/OpenMP hybrid versions of the INS3D code on the Columbia
system using the BX2b processors.
All computations included tip leakage effects with a radial tip clearance of
0.006 inches, a pump operating condition of a mass flow rate of 154.7 lbm/sec,
and a rotational speed of 15,761 rpm. A simplified computational model in-
cludes the LPFTP inducer with 4 long and 4 short blades, and a straight
duct, which extends 4 duct diameters upstream of the inducer. The objective
of studying this model is to compare unsteady pressure values against exper-
imental data. To resolve the complex geometry in relative motion, an overset
grid approach is employed. The geometrically complex body is decomposed
into a number of simple grid components. Connectivity between neighboring
grids is established by interpolation at each grid’s outer boundaries. Addition
of new components to the system and simulation of arbitrary relative mo-
tion between multiple bodies are achieved by establishing new connectivity
without disturbing the existing grids. This computational grid has 57 overset
zones with 26.1 million grid points. An extended computational grid system is
based on the first model with the addition of the flowliner geometry. The grid
system includes upstream and downstream slots, the overhang area between
liners and the bellows cavity. This model is very similar to the ground test
article. It consists of 264 overlapped grids with 65.9 million grid points. De-
tails of the grid system are shown in figures 1 and 2. The flowliner component
consists of an axisymmetric chamber around the external wall of the pipe,
Unsteady Flow Simulation of High Speed Turbopumps 779
and two rows of slots in the streamwise direction. Each slot is a rectangular
shaped hole with rounded corners. On the outside wall of the chamber are the
bellows, which are shaped like 10 periods of a sine wave. The bellows cavity is
connected to the duct via the overhang area and the slots. Two-dimensional
overset grids are first created for the bellows, side walls and the overhang area
of the bellows cavity. These are then revolved 360 degrees to form the volume
grids. Each slot consists of a body-conforming grid and a warped Cartesian
core grid in the middle of the hole. The flowliner component alone contains
212 grids and 41 million points.
Fig. 1. Surface grids for LPFTP inducer and the Liquid LH2 flowliner.
3 Computed Example
An initial condition of flow at rest and no inducer rotation is used to start
the computations. Then, the inducer is rotated at full speed as a start-up
procedure. Mass flow is specified at the inflow and characteristic boundary
conditions are used at the outflow. Simulations for 14 inducer rotations were
competed for model I, and 12 inducer rotations were completed for model II.
The time history of non-dimensional pressure difference from INS3D calcula-
tions (model I) at a location where experimental measurements are taken is
plotted in figure 3(a). Even though computed results have not fully converged
to periodicity in time and may still show the evidence of start-up transients,
780 Kiris, Kwak, Chan and Housman
(a) (b)
Fig. 3. Time history of non-dimensional pressure during one inducer rotation (model
I, 14th inducer rotation), and Min/Max values of non-dimensional pressure in hot
fire test.
The region of reverse flow extends far enough upstream to interfere with
both flowliners in the gimbal joint. Figure 4 shows the extent of the backflow
at an instantaneous time. Positive axial velocity values are colored in red and
negative axial velocity values are colored in blue. Axial velocity values are
non-dimensionalized by tip velocity. It is clearly seen that there is a strong
interference between the backflow in the duct and the flow in the bellows
cavity. The backflow velocity magnitude reaches 15-20 % of the tip velocity
magnitude in the overhang area between the liners. It should be noted that
this interaction is unsteady and backflow travels in the circumferential direc-
tion as well. Due to strong interactions in the overhang area, flow is excited
in the bellows area that results in time-dependent recirculation regions. This
Unsteady Flow Simulation of High Speed Turbopumps 781
observation can be seen in figure 5, where velocity vectors are plotted in the
region near the bellows cavity and overhang area. As shown in figure 5, strong
jet flow, with velocities of about 10-15 % of the inducer tip speed, penetrates
directly into the bellows cavity resulting in strong unsteady recirculation re-
gions in the cavity. The time-dependent interaction between the duct and the
bellows cavity can be one of the major contributors for high cycle loading.
Figure 5 also shows that modeling the gap in the overhang area between flow-
liners is very important. Jet-like flow in the overhang area pushes the fluid in
the bellows cavity toward the duct through the slots. Without proper model-
ing of this detailed geometry, one cannot obtain fine scale flow unsteadiness
on the liner. This transient phenomenon creates unsteady pressure loading
spectrum on the flowliner surfaces. Backflow also causes pre-swirl to occur in
the flow approaching the inducer.
Conclusion
High fidelity turbopump and flowliner calculations were performed on the
Columbia supercluster by using both MLP and MPI/OpenMP hybrid imple-
mentations of the INS3D code. CFD results confirmed that there is a backflow
caused by the LPFTP inducer. The region of reverse flow extends far enough
upstream to interfere with both flowliners in the gimbal joint. Computed
results for the straight duct hot fire test configuration have been verified by
correlation with unsteady pressure measurements. CFD results confirmed that
there is a strong unsteady interaction between the backflow regions caused by
the LPFTP inducer and secondary flow regions in the bellows cavity through
the flowliner slots. It is observed that there are more significant unsteady
flow interactions through the downstream slots than those observed in the
upstream slots. Inducer rotations are simulated in order to understand the
root cause of the flowliner crack problem.
Acknowledgments. The authors are grateful to Alex Te for his help in CAD
work, Tim Sandstrom and David Ellsworth for their help in the visualization
work, and to NASA-NESC Flowliner ITA team members for helpful discus-
sions.
References
1. Chorin, A., J., “A Numerical Method for Solving Incompressible Viscous Flow
Problems,” Journal of Computational Physics,’ Vol. 2, pp. 12-26, 1967
2. Kiris, C., and Kwak, D., “Parallel Unsteady Turbopump Simulations for
Reusable Launch vehicle,” Frontiers of Computational Fluid Dynamics,’
Caughey, D.A. and Hafez, M., ed, World Scientific, 2002.
3. Kiris, C., Kwak, D., and Rogers, S., “Incompressible Navier-Stokes Solvers in
Primitive Variables and Their Applications to Steady and Unsteady Flow Simu-
lations,” Numerical Simulations of Incompressible Flows,’ Hafez, M., ed, World
Scientific, 2003.
4. Rogers, S. E., Kwak, D. and Kiris, C., “Numerical Solution of the Incompress-
ible Navier-Stokes Equations for Steady and Time-Dependent Problems,” AIAA
Journal,’ Vol. 29, No. 4, pp. 603-610, 1991.
5. Taft, J. R., “Achieving 60 Gflop/s on the Production CFD Code OVERFLOW-
MLP,” Parallel Computing,’ 27(4):521-536,2001.
6. H. Jin, M. Frumkin and J. Yan, “Automatic Generation of OpenMP Directives
and Its Application to Computational Fluid Dynamics Codes,” in the Proceeding
of the Third International Symposium on High Performance Computing,’ Tokyo,
Japan, Oct. 16-18, 2000.
Use of the Gaussian Moment Closure for the
Modelling of Continuum and Micron-Scale
Flows with Moving Boundaries
1 Introduction
where A is the cell area, w, n̂k and ∆`k are the velocity of, unit normal
to, and length of the k th cell face, while Fk and Uk are the flux dyad and
solution state at the quadrature point on the k th face. Second-order spacial
accuracy is achieved via limited linear reconstruction. Flux evaluations are
carried out using Roe’s approximate Riemann solver [14]. Mesh refinement is
carried out according to physics-based refinement criteria as well as to resolve
embedded/moving boundaries. Refer to the recent paper by McDonald and
Groth [11] for further details of the parallel AMR scheme.
5 Numerical Results
Several flow problems are now considered. The first such problem is subsonic
boundary-layer flow over a flat plate. The goal in this case is to demonstrate
that the embedded boundary treatment and hyperbolic nature of the gov-
erning equations will yield smooth predictions of the frictional forces acting
on the plate; traditional cut-cell-type approaches combined with the Navier-
Stokes equations have been shown to produce large oscillations in viscous drag
predictions [6]. Next, subsonic flow past a circular cylinder is considered. This
case is considered to illustrate the scheme’s applicability to transition-regime
flows. Previous investigation using body-fitted meshes has shown good agree-
ment with experimental results [11] and it will be shown that equally good
results can be achieved with embedded boundary treatment. Finally, a channel
flow with complex moving geometry is considered to illustrate the potential
of the proposed approach for tackling such problems.
Subsonic boundary-layer flow past a flat plate is considered. For the case
of interest, the free-stream Mach number is 0.2 and Reynolds number per unit
length is 2000. For this situation, the Knudsen number was 1.5×10−4 , which
indicates that the flow is in the continuum regime. Two computational meshes
are considered: one aligned with, or at 0◦ to, the plate and a second mesh at
30◦ to the plate. Both meshes initially consisted of one 16 × 16 Cartesian
square block. This block was then subjected to six or eight mesh refinements
such that at each level of refinement, any block crossed by the boundary was
refined. Sections of the resulting meshes are shown in Figure 2.
Numerical predictions of the friction coefficient, Cf , are shown in Figure 3.
In the figure, comparisons are made to the classical boundary layer results of
Blasius. It can be seen that there is good agreement between the computed
results and Blasius’ solution. There are no oscillations present in the predicted
skin friction coefficients, even when the flat plate intersects the grid at an
angle. Furthermore, on the more refined mesh with eight levels of refinement,
the numerical solutions are nearly indistinguishable from the Blasius solution.
As a second case, subsonic flow past an circular cylinder is considered.
Experimental measurements of drag on the cylinder have been obtained by
Use of the Gaussian Moment Closure with Moving Boundaries 787
Fig. 2. (left) Section of 48128-cell mesh with flat plate embedded at 0◦ . (right) Sec-
tion of 65024-cell mesh with flat plate embedded at 30◦ .
Fig. 3. Coefficient of friction calculated for a Cartesian grid with an embedded flat
plate at 0◦ and 30◦ to the grid as compared to Blasius solution.
Coudeville et al. [7]. The Gaussian closure has previously been applied to this
case using body fitted meshes [11] for several speed ratios, S, and a wide
range of Knudsen numbers in the continuum and transition regimes. These
predictions are shown in Figure 4, where numerical results are compared to
experimental data of Coudeville and an analytical expression of Paterson [11].
In order to demonstrate that the present embedded mesh treatment can
recover these previous results with virtually equal accuracy, values for the
coefficient of drag were computed for varying Knudsen numbers for flow with
a speed ratio of 0.107. Predicted results obtained using the adjusted mesh
scheme are compared in Figure 4 with the Gaussian-closure solutions obtained
using a body fitted mesh, as well as to the experimental results. Clearly, the
agreement between body-fitted and non-aligned-mesh results is excellent.
Coirier explored Mach 0.1 flow through a branched channel containing
fourteen pin cooling fins [6]. A similar geometry has been considered here for
both continuum and non-equilibrium flow, except two rows of the pins are now
assumed to oscillate with a prescribed motion. The results of the continuum
situation can be seen in Figure 5.
788 J. G. McDonald, J. S. Sachdev, and C. P. T. Groth
Fig. 4. (left) Coefficients of drag computed using the Gaussian closure with a body-
fitted mesh compared with experimental data and an approximate solution due to
Paterson. (right) Coefficients of drag computed using the Gaussian closure with a
body-fitted mesh and a Cartesian mesh with an embedded boundary at a speed
ratio of 0.107 are compared with experimental data.
Fig. 5. Computational grid and x-direction velocity contours for flow through a
branched channel with moving pins.
Conclusions
The use of the Gaussian moment closures with embedded-mesh treatment for
continuum- and transition-regime flows has been demonstrated. The proposed
algorithm allows for accurate treatment of micron-scale flows with embedded
and possible moving boundaries, not aligned with the mesh.
References
1. M. J. Aftosmis, M. J. Berger, and J. E. Melton. AIAA J., 36(6):952–960, 1998.
2. M. J. Aftosmis, M. J. Berger, S. M. Murman. Paper 2004-1232, AIAA, 2004.
3. S. A. Bayyuk, K. G. Powell, and B. van Leer. Paper 93–3391, AIAA, 1993.
4. M. J. Berger and P. Colella. J. Comput. Phys., 82:67–84, 1989.
5. A. Beskok, G. E. Karniadakis. J. Thermophys. Heat Transfer, 8(4):647, 1994.
6. W. J. Coirier. PhD thesis, University of Michigan, 1994.
7. H. Coudeville, P. Trepaud, E.A. Brun. Proceedings of the Fourth International
Symposium on Rarefied Gas Dynamics, New York, 1965. Academic Press.
8. J. C. Harley, Y. Huang, H. B. Bau, J. N. Zemel. J. Fluid Mech., 284:257, 1995.
9. J. A. Hittinger. PhD thesis, University of Michigan, 2000.
10. C. D. Levermore. J. Stat. Phys., 83:1021–1065, 1996.
11. J. McDonald and C.P.T. Groth. Paper 2005-5035, AIAA, June 2005.
12. S. M. Murman, M. J. Aftosmis, M. J. Berger. Paper 03–1119, AIAA, 2003.
13. S. Osher and J. A. Sethian. J. Comput. Phys., 79:12–49, 1988.
14. P. L. Roe. J. Comput. Phys., 43:357–372, 1981.
15. J. S. Sachdev and C. P. T. Groth. Proceedings of the 3rd International Confer-
ence on Computational Fluid Dynamics, 109–115, 2004.
16. J. S. Sachdev, C. P. T. Groth, and J. J. Gottlieb. Int. J. CFD, 19(2):157, 2005.
Development of “MATIS-SC” for High Speed
Steam Flow with Non-equilibrium
Condensation
1 Introduction
2 Development of “MATIS-SC”
2.1 Calculation Model
There are two main types of steam flow calculation models. One uses ap-
proximate equation of state for steam to calculate state quantities [1, 2, 3]
(named “equation of state model” in this paper) and the other uses
steam table [4, 5, 6, 7] (named “steam table model” in this paper). These
models have both advantages and disadvantages. The equation of state model
Based on Based on
Calculation of Wetness IAPWS- classical classical
IF97 concideration consideration
theory theory
Precisions of State
4
Quantities Calculation
We already have built CFD code “MATIS-C” for compressible fluid and
confirmed in previous researches [9, 10]. “MATIS-C” is 3 dimensional FDM
based code and 2nd order spatial and temporal precision. We incorporate
mixed model into “MATIS-C”, and develop “MATIS-SC” for high speed steam
with non-equilibrium condensation.
The feature of “MATIS-SC” is shown in Table 2. 2nd order TVD scheme
is adopted to avoid the numerical oscillations, and 2nd order spatial precision
and full implicit 2nd order temporal precision with newton iteration. LU-SGS
algorithm is adopted for matrix inversion.
Conclusions
For accurate calculation of a high-speed steam flow with non-equilibrium
condensation in wide range, we make mixed model to combine the advantages
of the existing models and develop “MATIS-SC”. A steam flow in a laval
nozzle is calculated for the benchmark test, and the reproducibility of non-
equilibrium condensation and the validation of the code are confirmed.
References
1. Senoo, S., et al., Proc. of ASME FEDSM, 31191 (2002)
2. Ishizaka, K., et al., Proc. of 6th ISCFD, 1(1995), 479-484
3. Gunter H. Schnerr et al., Proc. of 5th ISCFD, 3 (1993), 87-92
4. Grossman, B. and Walters, R. W., AIAA Journal, 27-5 (1989), 524-531
5. Meng-Sing Liou, Bram van Leer and Jian-Shun Shuen, J. of Computational
Physics, 87(1990), 1-24
6. Senoo, S., et al., Trans. JSME Series B (in Japanese), 671-68 (2002), 2006-2013
7. Gunter H. Schnerr and Ulrich Dohrmann, AIAA Journal, 28-7 (1990), 1187-
1193
8. Release on the IAPWS Industrial Formulation 1997 for the Thermodynamic
Properties of Water and Steam, IAPWS, (1997)
9. Morita, R. et al., 2004 ASME HT/FED Conf., 56017 (2004)
10. Morita, R. et al., Trans. JSME Series B (in Japanese), 696-70 (2004), 1984-1991
11. Yee, H., NASA TM 89464, (1987)
12. Yoon, S. and Jameson, A., AIAA Journal, 26-9 (1988), 1025-1026
13. Binnie, A.M., Green, M.A. et al., Proc. Roy. Soc., A, 181, plate 3, pp. 134-154,
(1943)
Numerical Method for Near-critical Fluids of
Arbitrary Material
1 Introduction
The superscripts ± indicate the sign of characteristic speeds. Fi± and ±i are
the numerical flux vectors and the preconditioned Jacobian matrices divided
by positive or negative characteristic speeds, respectively. The detail expres-
sion in Eq.(2) has been presented in Ref. [6]. This flux-vector splitting form
is applied to the second-order LU-SGS scheme [7]. This form is extended to a
form for supercritical fluids as
Γ D∆Q̂∗m = RHS m + θL ∆tG+ (∆Q̂∗m ) (3)
∆Q̂m = ∆Q̂∗m − Γ −1 D−1 θL ∆tG− (∆Q̂m ) (4)
G+ (∆Q̂∗m ) = (Â+
1 ∆Q̂
∗m
)i−1,j + (Â+
2 ∆Q̂
∗m
)i,j−1 (5)
G− (∆Q̂m ) = (Â− m − m
1 ∆Q̂ )i+1,j + (Â2 ∆Q̂ )i,j+1 (6)
where m is the number of Newton iterations in each time step. The delta-form
fluxes in Eqs.(5) and (6) can be calculated by using Eq.(2). Then, Q̂L(R) is re-
placed by the time derivative δ Q̂L(R) . The script of each time-derivative flux in
the right hand side of Eqs.(5) and (6) denoted by i, j indicates the grid location
of the flux. D is the diagonal matrix approximated by the spectral radius of
the preconditioned Jacobian matrices and modified by the dual-time stepping
as D = Γ −1 Γp + ∆tθL [r(Âi ) + rv (∂Fvi /∂ Q̂)], where r(Âi ) and rv (∂Fvi /∂ Q̂)
indicate the sum of the maximum values of the spectral radius derived from
the Jacobian matrix Âi and ∂Fvi /∂ Q̂. In this study, these values are approx-
imated by r(Âi ) = αc max[λ(Âi )] and rv (∂Fvi /∂ Q̂) = αv (µ + κ)gii /(ρ∆ξi ),
Near-critical Fluids of Arbitrary Material 797
where αc and αv are empirical constants. λ(Âi ) means the eigenvalues of Âi .
µ and κ are the molecular viscosity coefficient and the heat-conductivity co-
efficient. gii = ∇ξi ∇ξi . RHS m is calculated by the following equation with
the Crank-Nicolson method.
RHS m = −(Qm − Qn ) − ∆tL∗ [(Q̂m + Q̂n )/2] (7)
∗
where L means the finite-difference operator of L in Eq.(1). The vector de-
rived from Γ −1 Γp is actually effective only to the continuity equation.
We have ever employed one of cubic equations, that is, the Peng-Robinson
equation of state(P-R EOS) [5], (p + β)(1 − bρ) = ρRT , where β = aρ2 /(1 +
2bρ−b2 ρ2 ). The square of the sound speed is defined by dp/dρ = ρhT /(ρT (1−
ρhp ) + ρρp hT ) , where ρp , ρT , hp and hT are the partial derivatives of the
density and the enthalpy with respect to the pressure and the temperature.
The preconditioning matrix Γ is also modified by such partial derivatives.
These partial derivatives can be derived theoretically from the P-R EOS. The
P-R EOS should be accurate when the bulk conditions in fluids are at super-
critical state. But the P-R EOS may be inaccurate at subcritical conditions.
Unfortunately even if the P-R EOS is used, the coefficients of thermal conduc-
tivity and molecular viscosity should be calculated by any other mathematical
models, because the theoretical derivation from the P-R EOS is impossible.
In such a situation, we know the PROPATH database developed by
Kyushu University. A number of mathematical models for thermal proper-
ties of arbitrary substances are programmed. This database can calculate
about fifty chemical substances at arbitrary temperature and pressure in a
wide range of them, not only supercritical conditions but also atmospheric
and cryogenic conditions. For examples, the EOS for carbon-dioxide is based
on IUPAC [8], and that for water is made from the steam table presented
in IAPWS IF97 [9]. Other properties such as molecular viscosity and heat-
conductivity coefficients are also programmed according to the original paper
where the mathematical model was proposed. The partial derivatives such as
ρp , ρT , hp , and hT are also prepared in the PROPATH.
4 Numerical Examples
(a) t = t1 (b) t = t1 + ∆t
Fig.2 Instantaneous temperature contours(CO2 , Ra=1.3E5)
Next, the above case is calculated with the PROPATH database. The cor-
responding Rayleigh number results in Ra = 1.3 × 105 because of a higher
density near the critical point. The calculated results have a time-dependent
and periodical flow property. The Rayleigh number may be beyond the criti-
cal one. Figure 2(a) and 2(b) show the calculated instantaneous temperature
contours at a different non-dimensional time. The density near the top wall is
relatively higher than that near the bottom wall, because the top wall has a
near-critical temperature. The heaver CO2 moves downward, whereas lighter
CO2 near the bottom wall moves upward due to the buoyancy effect. [t] [t]
The Rayleigh number is increased to Ra = 3×105 by increasing the height
length of the flow field and the same CO2 case except for the height length
Near-critical Fluids of Arbitrary Material 799
(a) t = t2 (b) t = t2 + ∆t
Fig.3 Instantaneous temperature contours(CO2 , Ra=3E5)
(a) t = t3 (b) t = t3 + ∆t
Fig.4 Instantaneous temperature contours(H2 O, Ra=3E5)
of the field is calculated again. Figures 3(a) and 3(b) show the calculated
instantaneous temperature contours at a different non-dimensional time. On
the other hand, the present method is further applied to an R-B convection in
H2 O near critical conditions. The Rayleigh number is fixed to the same value
with the CO2 case, that is, Ra = 3 × 105 . The corresponding bulk pressure is
22.074×106 [Pa] and the top wall temperature Ttop = 650.096[K](∆T = 60[K]).
Figures 4(a) and 4(b) show the calculated instantaneous temperature contours
at a different non-dimensional time. Unsteady flow fields are captured in both
calculated results. Although both cases set the same Rayleigh number, the
wideness of the field at a lower temperature may be different. The field in
H2 O is more widely spread relatively compared with that in CO2 .
This paper focuses on the near-critical fluids. In addition to them, another
distinctive point of the present method can be noticed if the following calcu-
lated results are added here. As a simple flow field, two-dimensional square-
cavity flows are calculated assuming a liquid of water(H2 O) and a vapor of
800 Satoru Yamamoto and Atsushi Ito
Conclusion
References
1. A Program Package for Thermophysical Properties of Fluids(PROPATH),
Ver.12.1, PROPATH GROUP.
2. S. Yamamoto and B.-R. Shin, Proc. of the Second Int. Conf. on Computational
Fluid Dynamics, (2002), 112-117, Springer.
3. J.M. Weiss and W.A. Smith, AIAA J., 33(1995), pp.2050-2056.
4. S. Yamamoto, Proc. of the Third Int. Conf. on Computational Fluid Dynamics,
(2004), Springer, in press.
5. D.-Y. Peng and D.B.Robinson, Ind. Eng. Chem. Fundam, 15(1976), pp.59-64.
6. S.Yamamoto, J. Comp. Phys., 207(2005), pp.240-260.
7. S. Yamamoto, S.Kano, and H. Daiguji, Comp. and Fluids, 27(1998), pp.571-580.
8. S.Angus et al., International Thermodynamic Table of the Fluid State-3 Carbon
Dioxide, IUPAC, Vol.3, (1976).
9. Japan Society of Mechanical Engineers, 1999 JSME Steam Tables, (1999).
Flow Physics and Stokes’ Theorem in
Wind Turbine Aerodynamics
Summary. A viscous lift theorem is derived from a momentum balance and Stokes’
theorem around one section of a wind turbine blade. The theorem is a generaliza-
tion of the classical Kutta-Zhukovsky lift theorem and is validated for 2D attached
and separated flow. The application of the viscous lift theorem within a coupled
Navier-Stokes/Vortex-Panel solver gives insight into the complex 3D aerodynamics
pertinent to wind turbines.
1 Introduction
ison with measured wind tunnel data [5] and a vortex line method (VLM)
[6].
2 Methods
Let us assume 2D, steady, viscous, and incompressible high Reynolds number
flow around an airfoil . A simply connected domain Ω is defined in Figure 1
where (C) is the airfoil surface contour, (Σe) is assumed to be in the inviscid
region, and (Σw) passes through the trailing edge orthogonal to the direction
of U∞ . Applying Gauss’ divergence theorem to the gradient components of a
momentum balance in the normal direction yields
Z Z Z Z
0
L = (ρ∞ U ω) dΩ = −ρ∞ U∞ Γ − ρ∞ (u0 ω) dΩ (1)
Ω Ω
0
for the lift per unit length L after neglecting higher order viscous terms. In
Equation (1), ω is the vorticity component normal to Ω, and u0 is a perturba-
tion velocity with U = U∞ + u0 . Note that the Kutta-Zhukovsky lift theorem
is exactly recovered for irrotational flow and ω = 0. The integrand in Eqn. (1)
reads in conservation form:
∂ 1 ∂
u0 ω = (u0 v 0 ) − u0 2 − v 0 2
(2)
∂x 2 ∂y
Substituting Eqn. (2) into Eqn. (1) and applying Green’s theorem we obtain
L0 = −ρ∞ U∞ Γl = −ρ∞ U∞ (Γ + ∆Γw + ∆Γe ) (3)
Flow Physics and Stokes’ Theorem in Wind Turbine Aerodynamics 803
with
I Z Z
Γ = v ds = ω dΩ (4)
(Σw)+(Σe) Ω
Z
1
∆Γw = (u0 v 0 ) dy (5)
U∞ (Σw)
Z
1 1 02
u − v 0 2 dx + (u0 v 0 ) dy
∆Γe = (6)
U∞ (Σe) 2
Fig. 2. S809 (Re = 106 , α = 4.1◦ ) Fig. 3. S809 (Re = 106 , α = 20◦ )
The PCS method was applied to the parked National Renewable Energy Lab-
oratory (NREL) Phase VI rotor. Results are compared with measured data
obtained in the NASA Ames wind tunnel [5] and the VLM code [6] as a stand-
alone method. The normal force coefficient CN at r/R = 0.47 is shown in Fig.
4 for PCS, VLM, and NREL data as a function of α47 . The blade setting angle
α47 is the geometric incidence of the NREL blade at r/R = 0.47. PCS results
agree very well with NREL data, while VLM shows discrepancies for separated
and stalled flow. Figure 5 shows results obtained for the bound circulation Γ
at a blade pitch angle of α47 = 23.49◦ . The rapid variation in the PCS bound
circulation near r/R = 0.40 is associated with a vortical structure that is
counter-rotating to the neighboring root vortex and attached to a region of
separated flow more outboard. Figure 5 also shows the PCS (K-Z) bound cir-
culation computed by the classical K-Z lift theorem. It is apparent that values
for Γ deviate for r/R ≥ 0.35. The very good agreement of PCS results with
NREL data in Fig. 4 for α47 = 23.49◦ supports the physical correctness of
the solution and computed PCS bound circulation. The Iso-vorticity surface
in Fig. 6 illustrates the vortical structure trailing downstream of the blade.
Such complex 3D flow phenomena cannot be attained with a vortex model
such as VLM that is based on strip theory.
Flow Physics and Stokes’ Theorem in Wind Turbine Aerodynamics 805
Conclusions
A novel method of finding the spanwise bound circulation along a wind tur-
bine blade inside a NS region was proposed from a viscous lift theorem. This
method can serve as an alternative to the classical K-Z lift theorem being
employed in hybrid solvers for wind turbines and rotorcrafts. The bound cir-
culation plays a fundamental role in a coupled Navier-Stokes/Vortex-Panel
solver (PCS) for the computational analysis of the complex flow physics per-
tinent to wind turbines. Some results have been presented and compared with
NREL measurements. PCS is found to give improved performance predictions
compared to a stand-alone vortex method (VLM). The spanwise bound cir-
culation Γ gives insight into the 3D aerodynamics of wind turbines that is
not easily accessible with standard local blade loads such as thrust/torque or
normal/tangential force coefficients.
References
1. Leishman, J. G.: Challenges in Modeling the Unsteady Aerodynamics of Wind
Turbines. AIAA-2002-0037 (2002)
2. Schmitz, S., Chattot, J. J.: A Parallelized Coupled Navier-Stokes/Vortex-Panel
Solver. ASME Journal of Solar Energy Engineering, 127, 475–487 (2005)
806 Sven Schmitz and Jean-Jacques Chattot
3. Xu, G., Sankar, L. N.: Computational Study of Horizontal Axis Wind Turbines.
ASME Journal of Solar Energy Engineering, 122, 35–39 (2000)
4. Bhagwat, M., Moulton, M. A., Caradonna, F. X.: Hybrid CFD for Rotor Hover
Performance Prediction. AIAA-2006-3474 (2006)
5. Hand, M. M., Simms, D. A., Fingersh, L. J., Jager, D. W., Cotrell, J. R., Schreck,
S., Larwood, S. M.: Unsteady Aerodynamics Experiment Phase VI: Wind Tun-
nel Test Configurations and Available Data Campaigns. NREL/TP-500-29955
(2001)
6. Chattot, J. J.: Helicoidal Vortex Model for Steady and Unsteady Flows. Com-
puters and Fluids, 35, 733–741 (2006)
The Role of Separation Bubble on an Airfoil at
Low Reynolds Numbers
Numerical simulations have been made to clarify the relation between the
behavior of separation bubble and aerodynamic characteristics of an air-
foil(NACA0012) at low Reynolds numbers below 105 . Two and three dimen-
sional time-dependent Navier-Stokes equations are solved by finite-difference
approximation without using any turbulence models. And the multi-directional
finite-difference scheme was also utilized to stabilize computations. Syn-
chronous interaction of a separation bubble near the leading edge and vortical
structure generated at the trailing edge is responsible for the aerodynamic
performance of the airfoil.
1 Introduction
2 Computational Method
The governing equations are the unsteady incompressible Navier-Stokes equa-
tions and the equation of continuity:
divBorn − −M ayeru = 0, (1)
∂Born − −M ayeru
+ (Born − −M ayeru · grad) Born − −M ayeru
∂t
1
= −gradp + ∇p, (2)
Re
where Born − −M ayeru, p, t, Re denote the velocity vector, pressure, time,
and Reynolds number.
These equations are solved by the projection method. The pressure field is
obtained by solving Poisson’s equation derived from Navier-Stokes equations.
And we applied SOR method and the multi-grid method to solve the Poisson’s
equation as follows:
Dn
4p = −div (Born − −M ayeru · gradBorn − −M ayeru) + , (3)
δt
D = divBorn − −M ayeru, (4)
where n is the number of time step and δt is the time increment. Dn+1 is
assumed to be zero, but Dn is retained as a corrective term.
A generalized coordinate system with O type grid is employed to avoid
instability induced by concentrated grid points near the centerline of the wake.
All the space derivative terms except the convection terms are represented
by a central difference approximation. The convection terms are represented
by a third-order upwind scheme(K-K scheme¶ ). And for time integration, the
Crank-Nicolson implicit scheme is utilized.
Y Z Z’ Z Z’ Z
Y’ X’ Y’
X’ Y
Y’ Y
X’
X X X Y
X
Fig.2 and 3 show comparison of pressure fieldsk between utilized Multi- di-
rectional finite-difference scheme and without Multi-directional finite-difference
scheme. You can see shakes of pressure in the pressure field where Multi-
directional finite-difference scheme is not utilized, especially near the outer
boundary where grid becomes coarse(Fig.2). But, by utilizing Multi-directional
finite-difference scheme, shakes of pressure are prevented and we could get
convergence well(Fig.3).
3 Computational Results
The numerical simulation was done at the Reynolds number 10000, which is
based on the uniform velocity and the chord length. The angle of attack of
the airfoil is fixed at 10 degrees examining behavior of a separation bubble in
detail.
Fig.4 is the flow field averaged over one period. A pattern of surface oil
traces at the same flow conditions is also obtained experimentally (Fig.5).
A line near the leading edge indicates the position of flow separation and
another line near the trailing edge is so-called a re-attachment line. The region
bordered by the two lines is considered to be a laminar separation bubble. The
agreement between the numerical results (Fig.4) and the experimental one is
satisfactory.
k
Same grid, same caluculation condition
810 Yusuke Nakae et. al.
On the other hand, at the case of minimum lift(ϕ=140◦ ), the detached fluid
lumps with clockwise vorticity just above the trailing edge induces a vortex
with counter-clockwise vortex(colored red) which produces positive pressure
distributions on the surface of suction side.
As time passed, the separated region and the reattachment point moved
toward the trailing edge. The rolling-up lump sweeps away into the wake. And
the lift coefficient gradually increases.
When the long bubble developed enough, reattachment point suddenly
moved toward the leading edge. Then a newly generated separated region
appeared near the leading edge and results in taking a maximum lift coefficient
again.
4 Conclusion
References
1. K. Kuwahara and S. Komurasaki, ”Direct Simulation of a Flow around a Sub-
sonic Airfoil”, AIAA Paper 2001-2545, 2001.
2. K. Kuwahara and S. Komurasaki, ”Simulation of High Reynolds Number Flows
Using Multidirectional Upwind Scheme”, AIAA Paper 2002-0133, 2002.
3. S. Komurasaki and K. Kuwahara, ”Implicit Large Eddy Simulation of a Subsonic
Flow around NACA0012 Airfoil”, AIAA Paper 2004-594, 2004.
4. Thomas J. Mueller and Stephen M. Batill, ”Experimental Studies of Separation
on a Two-Dimensional Airfoil at Low Reynolds Numbers”, AIAA Paper 80-1440,
1980.
5. M. Brendel and Thomas J. Mueller, ”Boundary-Layer Measurements on an Air-
foil at Low Reynolds Numbers”, AIAA Paper 87-0495, 1987.
6. K. Rinoie, ”Measurements of Short Bubble and Long Bubble Formed on
NACA63-009 Airfoil”, Journal of Aeronautical and Space Sciences Japan 38,
249-257, 1990.
7. K. Rinoie and N. Takemura, ”Oscillating Behaviour of Laminar Separation Bub-
ble Formed on an Airfoil near Stall”, CEAS Aerospace Aerodynamics Research
Conference, Cambridge, 21.1-21.11, 2002.
812 Yusuke Nakae et. al.
ϕ = 45◦ ϕ = 270◦
ϕ = 90◦ ϕ = 315◦
Fig. 7. Flow fields at each phase (the blue color corresponds to clockwise vorticity
and the red color corresponds to counter clockwise vorticity)
1 Introduction
Transient natural convection in cubical, rectangular and cylindrical cavities
have attracted numerous researchers due to its implications in a wide set of
technological applications and because of its geometric simplicity. Most of
the work done in the thermal storage area considers water as working fluid.
The usage of the oil - or high Prandtl number fluid - as working fluid in
the area of the thermal storage is rare. To the author’s knowledge only a
few studies have been reported. Lin and Akins [1], studied experimentally
the natural convection in cubical enclosures using different kinds of fluids
(6 ≤ P r ≤ 9000)and different sizes of cubes. These author’s found that the
inclusion of the time and/or Pr number does not improve the correlation
between Nusselt and Rayleigh number and that the use of the conventional
correlations is accurate enough for prediction purposes. Ogawa et al. [2] carried
out three-dimensional steady calculation of natural convection in a fluid with
variable viscosity. They classified the flow patterns into two main regimes
depending on the behaviour of the upper boundary layer; the stagnant lid
regime and the whole layer regime. They concluded that the Nusselt number
of the top surface is highly dependent on the viscosity in the whole layer
regime. Davaille and Jaupart [3] reported experimental results of transient
natural convection at high Ra numbers with large viscosity variation in a
tank with insulated bottom wall and cold top wall. They studied the effect of
the viscosity in the unstable top boundary layer on the onset of instabilities.
They introduced a new viscous temperature scale to compute the heat transfer
rate at the cold top wall of the tank and claimed that this scale is sufficient
to account for the characteristics of convection.
The general objective of the present study is to develop and to validate
a simulation tool able to predict the cooling rates of oil in a tank. This tool
has to take into account the variation of the oil viscosity with temperature,
the influence of the conductivity and the thickness of the walls of the tank to
give detailed information of the cooling rates of the oil under different realistic
thermal boundary conditions, as those present in real systems. As a first step
towards this objective, we considered the transient laminar cooling of a fluid
with P r = 104 in a cubical cavity.
∂T ∗ ∂(u∗j T ∗ ) ∂2T ∗
+ = (4)
∂t∗ ∂x∗j ∂x∗j ∂x∗j
α
The length and velocity scales are L and L , respectively. The non dimensional
∗ T −Tw
temperature is defined as T = Ti −Tw .
The Prandtl and Rayleigh number are defined as:
βg(Ti − Tw )L3 νo
Rao = = 106 , P ro = = 104
νo α α
According to Kays and Crawford [4] , the flow considered can be classified
as laminar. The Prandtl number that appears in the diffusion term in equation
(3) is calculated using equation (1) . The Prandtl and Rayleigh numbers (P ro
and Rao ) in the buoyancy term are based on kinematic viscosity evaluated at
the reference temperature To = (Tw + Ti ) /2.
The associated initial and boundary conditions are:
u∗ = v ∗ = w∗ = Ti∗ = 0 when t = 0
Velocity:
u∗ = v ∗ = w∗ = 0 when t > 0
The different thermal boundary conditions used and the cases considered are
summarised in table (1).
Case (4) considers the heat conduction through the 1cm thick Plexiglass
walls, kkfwall
luid
= 0.47 .
3 Numerical methods
The set of governing equations and boundary conditions have been solved
numerically with the finite voulme, second order accuracy 3DINAMICS code.
816 O. Younis, J. Pallares, and F. X. Grau
This code has been validated and tested by Pallares et al [5] , the complete
mathematical formulation can be found in Cuesta [6] .
The grid and time-step dependence study carried out shows that a mesh of
403 control volumes in the fluid and a time-step of 10−7 are accurate enough.
When the wall conduction is considered, the mesh in the solid walls contains
4 control volumes in the direction perpendicular to the wall, resulting in a
mesh of 483 control volumes.
-0.1
-0.1
-0.2
-0.2
-0.3
-0
.3
7
-0.
-0.4 -0.6
-0.4 -0.5
-0.5
-0.5
-0.6
-0.7 -0.6
-0.8 -0.7
-0.8
-0.9
-0
-1
.9
-1
-0.1
-0.2
-0.3
wall
-0.4
-0.1
-0.5
-0.6
-0.2
-0.7 -0.3
-0.8 -0.4 -0.5
-0
-0.6 -0.8
-1
-0.7 -0.9
.9
-1
VV2 VVWC1
VVWC1
−0.1
−0.2 2
10
−0.3
Nu
−0.4
1
10
−0.5
−0.6
0
−0.7 10
0 500 1000 1500 2000 2 3
10 10
Time (s) Time (s)
∗
(a) Time evolution of T (b) Time evolution of N u
Conclusions
References
1. Lin, Y. S. , Akins, R. G.,ASME conference, July1983.
2. Ogawa, M., Schubert, G. , Zebib, A., J. Fluid Mechanics, 233, 299–328 (1991).
3. Davaille, A. , Jaupart, C. , 253, 141–166 (1993).
4. Kays, W. M. , Crawford ,M. E. , (1993).
5. Pallares, J. , Grau, F. X. , Giralt, F. , Int. J. Heat Mass Transfer, 42, 753–769
(1999).
6. Cuesta, I. ,Ph.D. thesis, Universitat Rovira i Virgili, Tarragona, Spain, (1993).
Aeroacoustic simulation of the flow in a
Helmholtz resonator
1 Introduction
After a lot of work has been performed to reduce the noise generated by vi-
brating structures, aerodynamic noise sources are becoming more and more
important in the transportation industry. Nowadays computational aeroacous-
tics (CAA) emerges as a viable solution to reduce the number of prototypes
and tests, and thus the global design cost and duration. Aeroacoustics is gov-
erned by the compressible Navier-Stokes equations. Since a direct numerical
simulation (DNS) of these equations is impractical for engineering problems,
it thus calls for an alternative approach. The approach followed in this paper
consists in using an acoustic analogy, as first proposed by Lighthill [5]. Acous-
tic analogies rest on the assumption that noise generation and propagation are
decoupled, that is, flow generated noise does not impact the internal dynamics
of the flow. In practice, using an acoustic analogy is a two-step procedure. In
the first step, an unsteady computational fluid dynamics (CFD) analysis is
used to compute aerodynamic sources. The second step consists in computing
the propagation and radiation of these aerodynamic sources.
be used for exterior or interior problems with or without liners, and has been
shown to possess the potential to handle industrial problems, see for example
Ref. [1]. The theory [2] behind the formulation used in Actran/LA is briefly
summarized hereafter.
Starting from the mass and momentum conservation equations, it is possi-
ble to derive Lighthill’s equation without any assumption, as in the beginning
of the original paper [5]. Then, some classical assumptions, valid only in the
case of a low Mach number and a high Reynolds number flow where isentropic
assumptions are reasonable from an acoustic point of view, lead to dramatic
simplifications. The final equation is a true wave equation whose right-hand
side term is the simplified Lighthill’s tensor
∂ 2 ρa ∂ 2 ρa ∂ 2 Tij
2
− a20 = , (1)
∂t ∂xi ∂xi ∂xi ∂xj
with
Tij ' ρ0 vi vj . (2)
In the frequency domain and adapted to a finite element framework, the for-
mulation becomes [2],
Z !
∂ ρ̃a ∂δρa 1 ∂ T̃ij ∂δρa
− k 2 ρ̃a δρa + + 2 dΩ = 0 ∀ δρa . (3)
Ω ∂xi ∂xi a0 ∂xj ∂xi
The only missing quantity is the source term, represented by the divergence
of the Lighthill tensor in the frequency domain. This quantity is computed in
the time domain by the CFD solver and a Fourier transform is then used
before the field is passed to Actran/LA. The Fourier transform is per-
formed using a dedicated package which comes with the standard distribution
of Actran/LA.
The whole computational aeroacoustics process can be summarized as fol-
lows:
1. The user produces a file with the node coordinates of the subset of the
acoustic mesh where the sources must to be accounted for;
2. The CFD solver reads this file; then, during the unsteady CFD simulation,
at each time step, it computes and writes the divergence of the Lighthill
tensor at all these nodes locations;
3. The user does the Fourier transform of the result, using filters if needed;
4. The user launches the Actran/LA simulation, which gives a direct access
to all acoustic fields in the finite and infinite elements, including some
energy indications.
The files produced for Actran/LA by the CFD solver are written in the
Hierarchical Data Format (HDF), an open-source format used to store named
matrices and developed by NCSA. It is an OS-independent binary format.
The Scientific DataSets of HDF version 4 is used.
822 L. Georges, G. Winckelmans, S. Caro, P. Geuzaine
(c) Close-up for the CFD mesh (d) Close-up for the CA mesh
The mesh used for the computational acoustic (CA) simulations is shown in
Figs. 1(b) and 1(d). It contains 79,436 nodes and 472,786 tetrahedra. The data
generated by the CFD computation and used for the acoustic simulation allows
a frequency resolution of 10 Hz. Figure 3 compares the numerical solution with
available experimental data. Both the frequency, computed at 353 Hz, and the
amplitude of the resonance are well reproduced.
70
Numerical simulation
Experiment
60
50
40
Lw (dB)
30
20
10
0
0 200 400 600 800 1000 1200 1400
Frequency (Hz)
Conclusions
A computational aeroacoustics methodology has been presented. It is based
on the Lighthill acoustic analogy and couples a in-house three-dimensional
unstructured flow solver to the Actran/LA commercial finite element solver.
824 L. Georges, G. Winckelmans, S. Caro, P. Geuzaine
Acknowledgment
The first and last authors acknowledge the support by the Walloon Region and
the European funds ERDF and ESF under contract N◦ EP1A122030000102.
References
1. S. Caro, P. Ploumhans, F. Brotz, M. Schrumpf, F. Mendonca, and A. Read.
Aeroacoustic simulation of the noise radiated by an Helmholtz resonator placed
in a duct. AIAA paper 2005-3067, May 2005.
2. S. Caro, P. Ploumhans, and X. Gallez. Implementation of Lighthill’s acoustic
analogy in a finite/infinite elements framework. AIAA paper 2004-2891, May
2004.
3. L. Georges, K. Hillewaert, G. Winckelmans, and P. Geuzaine. Development and
study of discretization schemes on sample unstructured grid topologies for large
eddy simulation. In 36th AIAA Fluid Dynamics Conference and Exhibit, San
Francisco, California, June 2006. AIAA-2006-3059.
4. L. Georges, G. S. Winckelmans, and P. Geuzaine. Improving shock-free com-
pressible RANS solvers for LES on unstructured meshes. Journal of Computa-
tional and Applied Mathematics, In press, 2006.
5. M.J. Lighthill. On sound generated aerodynamically. Proc. Roy. Soc. (London),
A 211, 1952.
Explicit low dispersive and low dissipative
non-centered finite differences and filters
1 Introduction
Since the earliest stages in computational aeroacoustics (CAA), the need for
highly accurate schemes has been recognized [4]. To meet the stringent accu-
racy requirements of CAA, low dispersive, low dissipative and large spectral
bandwidth numerical methods have been designed by optimizing their disper-
sion and dissipation properties in the Fourier space for low wave numbers.
Available optimized finite differences are for instance the explicit Dispersion-
Relation-Preserving scheme of Tam & Webb [3] or the compact schemes of
Lele [2]. Bogey & Bailly [1] also recently developed explicit finite differences
and selective filters accurate for waves down to four points per wavelength.
These examples deal only with centered schemes designed for uniform mesh
grids, but numerical tools with similar properties are also required for some
local specific features of the computational domain.
The implementation of free-field boundary conditions as well as wall
boundary conditions commonly involves small-stencil schemes which locally
lower the order of accuracy. However, to avoid extra damping and loss of ac-
curacy, it may be recommended to use non-centered optimized schemes with
high accuracy properties. The development of appropriate explicit boundary
filters is in particular of crucial importance and must be performed with at-
tention because standard explicit non-centered filters are unstable [5]. The
boundary filters must therefore be optimized in the Fourier space to form
stable algorithms. With this in mind, non-centered seven- and eleven-point
explicit finite differences and selective filters, with properties optimized in the
wave number space, are proposed in the present work. They are designed to
allow the accurate resolution of waves with at least five points per wavelength.
In the present paper, low dispersive and low dissipative seven- and eleven-
point non-centered boundary schemes are designed in section 2. An acoustic
reflection test problem is resolved in section 3 to highlight the properties of
the developed methods with respect to the standard ones. Concluding remarks
are finally drawn in section 4.
where ∆x is the mesh spacing and aj are the stencil coefficients. The scheme
uses P points left to xi and Q points right to xi , and it will referred to as
FDP Q . Applying Fourier transform in space to (1) yields the effective wave
number k ∗ of the scheme
Q
X
k ∗ ∆x = −i aj eijk∆x (2)
j=−P
In the present work, following Tam & Webb [3] for instance, coefficients aj
are determined to minimize the dispersion and the dissipation provided by
the scheme. First, to ensure a minimum order of accuracy, the terms of the
Taylor series of (1) are canceled up to the fourth order. Coefficients aj are
then determined to minimize the integral error defined by
Z π/2 h i d(k∆x)
(1 − α)|k∆x − Re(k ∗ ∆x)| + α|Im(k ∗ ∆x)| (3)
π/16 k∆x
where the parameter α is chosen between 0 and 1 depending on the scheme.
Fourth-order seven- and eleven-point non-centered finite-difference schemes
FDP Q have been designed in this way. Coefficients are given in table 1.
where f d is the filtered function, ∆x is the mesh spacing, dj are the coeffi-
cients of the filter, and σ is a constant taken between 0 and 1 defining the
filtering strength. The algorithm is denoted SFP Q . To design the filter and to
investigate its properties, σ is chosen to be equal to 1, its maximum value.
Explicit non-centered finite differences and filters 827
Consider now the Fourier transform in space of (4) to compute the transfer
function of the filtering:
Q
X
G(k∆x) = 1 − dj eijk∆x (5)
j=−P
Criteria of accuracy based on half a percent error, |k∆x − Re(k ∗ ∆x)|/π <
∗
5 × 10−3 and |1 − eIm(k ∆x) | < 5 × 10−3 , are introduced to illustrate, re-
spectively, the phase and the dissipation errors. The accuracy limits are given
in table 3 for the different eleven-point finite difference schemes. The non-
centered schemes FD46 and FD37 have dispersion limits close to four points per
wavelength. The schemes FD28 , FD19 and FD010 , are more asymmetric and are
consequently more dispersive: their dispersion limits are about five points per
wavelength. As for the dissipation limits, they are about five points per wave-
length for the schemes FD46 , FD37 , FD28 and FD19 , ranging from λ/∆x = 4.30
for the FD37 scheme up to λ/∆x = 6.87 for the FD19 scheme. The disper-
sion limit of the off-centered finite differences FD010 is about ten points per
wavelength. This scheme is nevertheless used at only one mesh point and the
overall accuracy of the eleven-point non-centered schemes is thus set to five
points per wavelength.
The filter properties are also reported in table 3. Limits of accuracy in
amplitude and in phase are estimated using the criteria 1 − |G(k∆x)| < 5 ×
10−3 and |ϕG (k∆x)|/π < 5 × 10−3 . The dissipation limits of the filters SF46 ,
SF37 and SF28 range from four to six points per wavelength. The filter SF15 is
more dissipative with a limit of fourteen points per wavelength, but it is used
828 Berland J., Bogey C. and Bailly C.
only at one mesh point. The accuracy limit in phase of the filter SF46 , whose
stencil is very weakly asymmetric, is close to two points per wavelength. For
the filter SF37 , the limit of accuracy in phase is about about five points per
wavelength, while it is close to seven points per wavelength for the schemes
SF28 and SF15 . The overall accuracy limit may nevertheless be set to five
points per wavelength as the non-centered schemes are implemented on a
small number of mesh points.
3 Test problem
A linear acoustic problem is considered by the solving the dimensionless equa-
tion:
∂u ∂p ∂p ∂u
+ = 0, + =0 (7)
∂t ∂x ∂t ∂x
on a domain extending from x = −50 to x = 50 and discretized by a
uniform grid with ∆x = 1. Wall boundary conditions are enforced at the
boundaries of the computation domain, i.e. ∂x p = 0 at x = ±50. Numer-
ically, ∂x p is set to zero at the boundaries and ∂x u is computed using the
interior points. The initialn disturbance at to= 0 is defined by u = 0 and
2
p = cos [2πx/(a∆x)] exp − ln(2) [x/(b∆x)] , with a = 8 and b = 12. The
perturbation is thus a wave with a dominant component discretized by eight
points per wavelength.
The system of equations (7) is solved up to time t = 200. At this time
position, each part of the initial disturbance, the left-running and the right-
running parts merge after two reflections into a wave identical to the ini-
tial perturbation. Time integration is performed using an optimized six-stage
Runge-Kutta algorithm [1], and the eleven-point optimized finite differences
and selective filtering of Bogey & Bailly [1] are used for centered spatial deriva-
tives. The CFL number is set to 0.8, i.e. ∆t = 0.8, and filtering strength σ is
set to 0.8.
The problem is first solved with centered schemes at the boundaries. The
centered finite differences and selective filters are the standard schemes whose
coefficients are computed by maximizing the order of accuracy. The 11-point
boundary schemes are then implemented to assess their accuracy.
Computed solutions are plotted in Figure 1. Figure 1.a shows the solutions
obtained using only centered finite differences and filters. The waves are seen
to be strongly damped. The solution obtained using the non-centered finite
differences and filters at the boundaries are presented in Figure 1.b. The
computed perturbation is observed to be in good agreement with the exact
solution, and the gain in accuracy is significant compared to the solution
determined with centered schemes at the boundaries.
Explicit non-centered finite differences and filters 829
(a) (b)
1 1
0.5 0.5
0 0
p
p
!0.5 !0.5
!1 !1
!50 !25 0 25 50 !50 !25 0 25 50
x/"x x/"x
Fig. 1. Computed pressure for the reflection test problem. The gray stripes represent
the wall boundaries. Use of: (a) centered finite differences and selective filters. (b)
non-centered optimized schemes. , computed solution; ◦, exact solution.
4 Conclusion
Explicit schemes have been proposed for local features of the computational
domain. In particular, non-centered finite differences and selective filters for
boundary conditions have been designed by minimizing their phase and dis-
sipation errors. These schemes are highly accurate for waves with at least
five points per wavelength. For reflection test problems, they provided results
showing a significant improvement of precision with respect to the centered
schemes.
830 Berland J., Bogey C. and Bailly C.
References
1. Bogey C., Bailly C.: A family of low dispersive and low dissipative explicit
schemes for flow noise and noise computations, J. Comput. Phys., 194, 194–
214 (2003).
2. Lele S.K.: Compact finite difference schemes with spectral-like resolution, J.
Comput. Phys., 103, 16–42 (1992).
3. Tam C.K.W., Webb J.C.: Dispersion-Relation-Preserving finite difference
schemes for computational acoustics, J. Comput. Phys., 107, 262–281 (1993).
4. Tam C.K.W.: Computational aeroacoustics: issues and methods, AIAA J.,
33(10), 1788–1796 (1995).
5. Visbal M.R., Gaitonde D.V.: Very high-order spatially implicit schemes for com-
putational acoustics on curvilinear meshes, J. Comput. Acous., 9(4), 1259–1286
(2001).
A New Principle of Non-Reflecting and its
Application to Hyperbolic Conservation Laws
∂ Ṽ ∂ Ṽ ∂ Ṽ ∂ Ṽ
( + Ã + B̃ + C̃ ) + i(K̃ − ωI)Ṽ = 0, (2)
∂t ∂x ∂y ∂z
where I is the 5 × 5 identity matrix. For any given wave number vector k =
(kx , ky , kz ), the matrix K̃ = kx à + ky B̃ + kz C̃. From the hyperbolicity of (1),
real ωs (eigenvalues) exist such that det|K̃ − ωI| = 0 (dispersion relation),
where the amplitude vector Ṽ = Ṽ(k) may be determined as an eigenvector
of K̃. For each plane wave (Fourier mode), both its amplitude Ṽ and phase
θ remain intact when passing through this interior point O, no distortion or
reflection occurs.
Based on such observations and the theoretical justification, a first princi-
ple of non-reflecting is proposed:
A boundary of the solution domain is said to be non-reflecting if
every point on the boundary surface can be considered as an interior
point.
There are two basic requirements defining an interior point:
1. the point must be physically located in the domain interior;
2. the consistency requirement that the governing PDE system is the only
condition to be satisfied at this point. Other conditions, if any, must com-
ply with the governing PDE system.
In order that all the boundary points be interior points, the solution domain
must be extended to include ghost cells. Then the consistency requirement
leads to:
The continuity criterion - a generic NRBC
Assume that in the domain exterior (i.e. ghost cells), the flow is also gov-
erned by another Euler equation system. If the two Euler equation systems
are identical at all the points on the boundary, the consistency requirement is
satisfied and these boundary points are interior points of the extended domain
(Fig.1a) and hence non-reflective. The continuity criterion leads naturally
to such non-reflecting, it is stated and proved as follows:
At each point on the boundary, if the flow variables V and their
spatial derivatives, namely, Vx , Vy , Vz , are continuous; or, V is
C 1 continuous in space across the boundary, the boundary is non-
reflecting .
Proof : Consider the Euler equations (1) for both domain interior and
exterior. Note that the jacobian matrices Ã, B̃, and C̃ are directly functions
of V (details may be found in [3]). As a result of the C 1 continuity of V, all
the terms
∂V ∂V ∂V ∂V
∂t , ∂x , ∂y , ∂z , Ã, B̃, C̃, and Q̃
in (1) are identical across the boundary surface. Here, continuity of ∂V ∂t is
inferred from (1). Equation (1) now may be considered as a single governing
equation system in the extended domain. The consistency requirement is sat-
isfied at all the boundary points. Hence there is no reflection at the boundary.
Non-reflecting principle 833
a b
C
boundary pt. now, a new
interior pt. F E
R L L
O R
interior pt.
extended domain
!S
L
A boundary point becomes an A R B
interior pt., but consistency boundary
requirement must be observed ghost cell
surface D cntr.
ghost cell
4 Numerical examples
Figure 3 shows propagation of an acoustic pulse together with an entropy
wave and vortex in a square domain. Comparison with exact solution is also
Non-reflecting principle 835
boundary wave No.
surface
x k
x
O
0
boundary
a b surface
element ! S
wave front (characteristic surface) c
Fig. 2. Wave propagation across a boundary; a: normal to the boundary, b: oblique
to the boundary, phase error may occur; c: sketch of x − x0 and k.
t=60
rho contact
shock
contact
t=300 t=500 t=700
Fig. 4. 2-D radial Riemann problem in an oval domain. Density contours show how
a shock is almost perfectly absorbed by the boundary. The mean flow is stationary
(M = 0). A Godunov scheme is used.
836 Ching Y. Loh and Philip C.E. Jorgenson
Fig. 6. 3-D Radial Riemann problem in a sphere, initial high pressure region is a
cube. One million unstructured hexahedral cells with the CE/SE scheme are used.
Concluding Remarks
References
1. R. Courant, and D. Hilbert “ Methods of Mathematical Physics”, Vol.II, John
Wiley & Sons, London, 1962.
2. B. Engquist and A. Majda “Absorbing boundary conditions for the numerical
simulation of waves”, Math. Comp., Vol. 31, pp. 629-651 (1977).
3. C. Hirsch “Numerical Computation of Internal and External Flows” Vol. I and
II, John Wiley & Sons (1993).
4. F.Q. Hu “On Absorbing Boundary Conditions for Linearized Euler Equations
by a Perfectly Matched Layer”, J. Comput. Phys., Vol.129, 201-219 (1996).
5. C.Y. Loh and P.C.E. Jorgenson, AIAA paper 2005-4716, 2005.
Numerical Solution of the Linearized Euler
Equations Using Compact Schemes
Kris Van den Abeele1,3 , Jan Ramboer2 , Ghader Ghorbaniasl1 , Chris Lacor1
1
Vrije Universiteit Brussel, Dept. of Mechanical Engineering, Fluid Dynamics and
Thermodynamics Research Group, Pleinlaan 2, 1050 Brussels, Belgium
kvdabeel@vub.ac.be, ghader@stro.vub.ac.be, chris.lacor@vub.ac.be
2
Cardijnstraat 11A, 1980 Eppegem, Belgium jan.ramboer@gmail.com
3
FWO Research Engineer
1 Numerical Algorithm
The linearized Euler equations (LEE) are given by the following general form:
∂U
+ ∇ · F (U) = 0 (1)
∂t
where the unknowns are U = [ρ 0 , v 0 , p 0 ]T and the fluxes are given by:
ρ0 v 0 + v0 ρ 0
0
F (U) = v0 ⊗ v 0 + pρ0 I (2)
v0 p 0 + γp0 v 0
ρ0 , v0 and p0 are the mass density, velocity and pressure of the mean flow,
while ρ 0 , v 0 and p 0 are small perturbations superimposed on this mean flow.
Rem is the mesh Reynolds number, which determines the amplitude of the
damping term. The coefficients dk = d−k are chosen such that only the high
frequency waves are affected by this term.
U0ij = Unij
Ukij = U0ij + ∆t k−1
N −k+1 Rij 1≤k≤N (6)
Un+1
ij = UNij
In (6), Rkij corresponds to the right hand side of (3), evaluated at stage k. ∆t
is the time step and n is the time index. The N -stage algorithm is N -th order
accurate in time for linear problems.
Solution of the Linearized Euler Equations Using Compact Schemes 839
In (8), r is the distance to the sound source, M0 is the mean flow Mach
number and 1r and 1v are the unit vectors
in the r-direction and the direction
of v0 . Notice the error term O r−5/2 . It shows that the error caused by
the boundary conditions will diminish with increasing distance between the
boundary and the source. The radiation boundary conditions are suitable
for a boundary where only acoustic disturbances exit the domain. If entropy
and vorticity waves also pass through the boundary, an outflow boundary
condition, the formulation of which can be found in [2], must be used. These
boundary conditions will be referred to as the ROBC boundary conditions.
3 Numerical Results
3.1 2D Linearized Euler - Grid Convergence Tests
As a first testcase, the advection of the three characteristic wave solutions of
the 2D LEE’s, namely the entropy, vorticity and acoustic wave, was computed.
840 K. Van den Abeele, J. Ramboer, G. Ghorbaniasl, C. Lacor
L2 -error on the pressure perturbation versus the mesh size at three different
moments in time, for the acoustic pulse. At t = 0.2, the wave has not yet
reached the boundary, at t = 0.5 it is going through the boundary and at
t = 1.0 it has left the domain. From the figure it can be seen that without
ASD the scheme is fourth order accurate when the wave is inside the do-
main. With ASD, this fourth order accuracy is lost. When the wave reaches
the boundary, neither of the boundary conditions preserves the fourth order
accuracy. The error is the largest for the CSBC boundary conditions. It is
caused by the fact that the tangential velocity component is forced to zero,
which in general is not true for the exact solution. For the ROBC boundary
conditions, the error is caused by the term O r−5/2 . However, this term can
be reduced by increasing the distance from the source to the boundary.
The graphs for the entropy and the vorticity wave are not shown here. For
these waves, the advantage of ASD is more apparent, since the error after the
waves have left the domain is smaller because the reflected high frequency
waves are damped. For the entropy wave, both boundary conditions conserve
the fourth order accuracy. For the vortex however, only the ROBC boundary
conditions give good results. It can thus be concluded that the latter are more
suited. For the remaining testcases, ROBC boundary conditions are used.
For these calculations, source terms were added to the right hand side of
(1). The source terms corresponding to dipole and quadrupole sources can be
found in Bailly and Juvé [4]. They result in the sound fields in Figure 2.
The last testcase was the rotor noise testcase as proposed by Tam in [3]. A
non-uniform mesh was used for this, the parameters of which can be found
Solution of the Linearized Euler Equations Using Compact Schemes 841
Fig. 1. Convergence test acoustic pulse with CSBC (left) and ROBC (right).
Fig. 2. Dipole (left) and quadrupole (right) sound pressure field at t = 640∆t.
– – : values from −0.011 to 0.001 with step 0.001. — : values from 0.001 to 0.011
with step 0.001.
plotted. The noise is not radiated outside the duct in the subsonic case, since
the frequency is below the cut-off limit. In the supersonic case, the noise is
radiated outside the duct.
842 K. Van den Abeele, J. Ramboer, G. Ghorbaniasl, C. Lacor
Conclusions
A code for the solution of the 2D LEE’s and the rotor noise equations using a
fourth order accurate compact finite volume scheme was developed. Two dif-
ferent sets of boundary conditions were tested, of which the ROBC’s produced
the best results. The code was tested for the advection of characteristic waves,
for dipole and quadrupole sound sources and for the sound field radiated by
an open and a ducted rotor. Good results were achieved for all testcases.
References
1. Lacor, C., Smirnov, S., Baelmans, M.: A finite volume formulation of compact
central schemes on arbitrary structured grids. J. Comput. Phys., 198, 535–566
(2004)
2. Tam, C. K. W.: Numerical Methods in Computational Aeroacoustics. AIAA
Short Course Computational Aeroacoustics: Methods and Applications, Lahaina,
Hawaii, 10-11 June 2000.
3. Third Computational Aeroacoustics (CAA) Workshop on Benchmark Problems
(2000), NASA CP 2000-209790.
4. Bailly, C., Juvé, D.: Numerical Solution of Acoustic Propagation Problems Using
Linearized Euler Equations AIAA Journal, Vol. 38, No. 1, 2000, pp. 22-29.
Transonic Aeroelastic Computations of a Delta
Wing Configuration with High Fidelity
Equations
RIKEN (The Institute of Physical and Chemical Research), 2-1 Hirosawa, Wako,
Saitama, 351-0198, Japan, htera@riken.jp / keno@riken.jp
1 Introduction
the linear and nonlinear structural models are performed on wing responses
with large deformation. In the comparison, influences of the number of selected
mode in linear structural analysis and assumption of strain-displacement rela-
tionships in nonlinear structural analysis are discussed. Finally, a criterion for
the use and choice of structural models will be set on aeroelastic computations
of a transonic delta wing with large deformations.
Fig. 3. Flow fields and deformed shape using linear and nonlinear structural model
at dynamic pressure of 2.78(psi)
wing responses with large deformation. Although all the computations with
the nonlinear structural models could not capture the experimental trend
accurately, note that the published available data on aeroelastic wing with
large deformation is only this one. For the range of larger dynamic pressure
above 4.0, differences appear apparently between the results with the non-
linear structural models. The reason for these differences has not yet been
understood clearly.
Conclusions
Comparisons of linear and nonlinear structural models coupled with Navier-
Stokes equations have been conducted on aeroelastic computations of a delta
wing with large deformation in the transonic flows. Some criteria on the struc-
tural models are made for the better understanding on the computation of
aeroelastic wing responses. The linear structural models are suitable for the
accurate predictions of transonic flutter boundary. There are little differences
between the linear and nonlinear structural models on the decision of flutter
boundary. Discussions on the wing responses with large deformation after a
flutter onset should be made by the nonlinear structural models. The higher-
order terms in the strain-displacement relationships dose not influence the
wing responses over the present range of deformation.
References
1. Schairer, E., Hand, L.: AIAA 97-2217 (1997)
2. Gordnier, R. E.: Journal of Aircraft., 40, No. 6, 1206–1207 (2003)
3. Attar, P. J., Gordnier, R. E.: AIAA 2005-1915 (2005)
4. Sheta, E. F., Harrand, V. J., Thompson, D. E., Strganac, T. W.: AIAA 2000-1399
(2000)
5. Noguchi, H., Hisada, T.: Computer & Structures, 47, No. 4/5, 699–710 (1993)
6. Goura, G. S. L., Badcock, K. J., Woodgate, M. A., Richards, B. E.: The Aero-
nautical Journal, 105, 215–221 (2001)
7. Bhardwaj, M. K., Kapania, R. K., Eichenbach, E., Guruswamy, G. P.: AIAA
Journal, 36, No. 12, 2179–2186 (1998)
Transonic Buffet over Symmetric Airfoils
Summary. Supercritical airfoils typically have small curvature of the upper surface
as compared with that of a conventional airfoil. At the same time, physical phe-
nomena associated with transonic flow over flattened surfaces have not been well
understood. For instance, the inviscid steady flow over such surfaces may exhibit
an unexpected bifurcation and intricate instability of its entire structure at certain
freestream conditions [1, 2, 3, 4]. In this paper, we examine the viscous transonic
flow. A numerical study confirms the bifurcation and structural instability of the
flow past a flattened airfoil. Apart from the bifurcation, computations demonstrate
periodic oscillations (buffet) caused by the shock-induced boundary layer separation.
with the second-order backward Euler scheme. We used the SST−k − ω tur-
bulence model which predicts the buffet onset quite reasonably [5].
Computations were performed primarily on a mesh of about 2 × 105 grid
points which were clustered in the regions of strong gradients of the solution.
The high accuracy of the method was confirmed by application to the buffet
problem for the 18% circular-arc airfoil and comparison with results available
in the literature [6].
flow [4]. The error in calculation of the endpoints of interval (2) under the
fixed model of turbulence is estimated by ±0.0004 with test computations on
refined meshes.
For the quasi-symmetric flows, the frequency of oscillations drops from 95
to 76 Hz with M∞ increasing from 0.823 to 0.844. For the asymmetric flows
involving three local supersonic regions, the frequency of oscillations decreases
from 89 to 78 Hz with M∞ increasing within the interval (2).
At a double freestream pressure p∞ and corresponding Re = 10.8 × 106 ,
the amplitudes of oscillations increase, whereas the bifurcation interval slightly
shrinks (see Fig. 6). At the end of this interval, an abrupt drop of CL for the
asymmetric flow is caused by instability of the shock foot F residing on the
rapidly thickening boundary layer.
At a halved p∞ and Re = 2.7 × 106 , the oscillations of CL slightly weaken.
The frequency of oscillations also insignificantly depends on Re. This fact,
along with the flow bifurcation, makes the flattened airfoil (1) markedly dif-
ferent from the 18% circular-arc airfoil.
Conclusion
References
1. Jameson, A.: Airfoils admitting non–unique solutions of the Euler equations.
AIAA Paper, No. 91-1625, 1–13 (1991)
2. Hafez, M., Guo, W.: Nonuniqueness of transonic flows. Acta Mechanica, 138,
177–184 (1999)
3. Caughey, D. A.: Stability of unsteady flow past airfoils exhibiting transonic non-
uniqueness. CFD Journal, 13, No. 3 (60), 427 – 438 (2004)
4. Kuz’min, A.: Instability and bifurcation of transonic flow over airfoils. AIAA
Paper, No. 2005-4800, 1–8 (2005)
5. Geissler, W., Koch, S.: Adaptive airfoil. In : Sobieczky, H. (ed) Proc. IUTAM
Symposium Transsonicum IV. Kluwer, Dordrecht, 303–310 (2003)
6. Geissler, W., Ruiz-Calavera, L.P.: Transition and turbulence modelling for dy-
namic stall and buffet. In: 4th Internat. Symp. on Engineering Turbulence Mod-
elling and Measurements, Corsica, France (1999)
7. Hafez, M., Guo, W.: Some anomalies of numerical simulation of shock waves.
Part II: effect of artificial and real viscosity. Computers and Fluids, 28, 721–739
(1999)
852 Alexander Kuz’min and Alexey Shilkin
Fig. 1. Lift coefficient CL versus M∞ for the oscillating transonic flow past airfoil
(1) at α = 0, Re = 5.4 × 106
Fig. 2. Instantaneous Mach number contours in the flow with two pairs of local
supersonic regions at α = 0, Re = 5.4 × 106 , M∞ = 0.832
Transonic Buffet over Symmetric Airfoils 853
Fig. 3. Oscillations of the lift coefficient for the transonic flow shown in Fig. 2
Fig. 4. Instantaneous Mach number contours in the asymmetric flow with three
local supersonic regions at α = 0, Re = 5.4 × 106 , M∞ = 0.832
854 Alexander Kuz’min and Alexey Shilkin
Fig. 6. Lift coefficient CL versus M∞ for the oscillating transonic flow past airfoil
(1) at α = 0, Re = 10.8 × 106 . The lower shaded domain is dropped in order
to avoid excessive complications of the diagram.
Aerodynamic Performance of a Deforming
Elastic Body in Supersonic Flow
1 Introduction
m
k
i
i-1/2
c !l i+1/2
i-1 i+1
The equations of motion based on the wave equation with the term of
damping effect are formulated for the shell made of elastic film. For example,
a balloon that has been filled with fluid. The film is composed of material
particles connected with elastic springs and dampers as shown in Fig. 1, so
that the film can expand and contract with the effect of damped vibration
in response to the fluid dynamic force. In Fig. 1, the segment shown in the
shaded portion of the film corresponds to a material particle. The notations
m, c, k and ∆l are the mass of a material particle, the damping coefficient,
the spring coefficient and the length of segment respectively. In this model, it
is assumed that the grid points are distributed on the circular film at regular
intervals, and the position corresponds to that of a material particle. The
equations of motion in the system of a material particle are given by
d2 xi
m = pi + τi + fi r + fi d + fi p (i = 1, · · · , N ) (1)
dt2
pi = −pout n∆l, τi = Tn∆l (2)
r
fi = k(xi+1 − 2xi + xi−1 ) (3)
fi d = c(ui+1 − 2ui + ui−1 ) (4)
where N is the number of material particles.
The external force consists of the fluid dynamic force, the elastic body force
and the artificial force. In the fluid dynamic force, the pressure pi and shear
stress τi given in Eq. (2) act on the film, where pout is the outside pressure,
T is the viscous stress tensor and n is the unit normal outward vector to the
film. These forces are obtained by solving the Navier-Stokes equations. In the
elastic body force, the restoring force fi r and damping force fi d given in Eq.
(3) and Eq. (4) are defined, where x and u are the position vector and velocity
vector of a material particle respectively.
Aerodynamic Performance of a Deforming Elastic Body 857
Furthermore, suppose that the elastic body deforms under the constraint
for fixed volume and center of gravity. Note that in the case of 2D the volume
is area.
In order to keep the area constant, the following additional force with
adjustable pressure pin is introduced inside the body.
fi p = pin n∆l (5)
The pressure can be determined by solving the following equation through
sub-iterations,
pν+1 ν
in = pin + α(S0 − S), α > 0 (6)
where S0 is the initial area and α is a constant parameter for control of the
area.
In order to fix the center of gravity, the additional force is given to offset
all forces acting on the body as follows:
The equation (1) is rewritten as the general form
d2 xi
m = fi (7)
dt2
by substituting fi as
fi = pi + τi + fi r + fi d + fi p (8)
Consider now summation in Eq. (7),
X d2 xi X
m 2 = fi (9)
dt
the equation of motion for the center of gravity is derived as
d2 x̄
M =F (10)
dt2
with the relations
P
X X mxi
M= m, F = fi , x̄i = P (11)
m
where M , F and x̄ are the total mass, the resultant force and the position
vector at the center of gravity respectively. The restoring force and damping
force automatically satisfy
X X
f r ≡ 0, f d≡0 (12)
By using the resulting Eq. (10) together with Eq. (7), finally the equation of
motion for fixed center of gravity is obtained as
d2 (xi − x̄) F 1
m = fi − m = fi − F (13)
dt2 M N
858 Tomohisa Hashimoto, Koji Morinishi and Nobuyuki Satofuka
3 Numerical Method
Numerical simulations are carried out by coupling the compressible Navier-
Stokes equations together with the equations of motion given in Eq. (13). The
unsteady compressible flow field around a deforming body is solved by using
sub-iterations with pseudo time at every physical time step. In the numerical
method, a central difference scheme with artificial dissipation [2] is used for the
spatial discretization. An iterative scheme in the time integration is the lower-
upper symmetric Gauss-Seidel implicit relaxation method [3]. The overset grid
technique [4] is adopted to remesh the grid easily in treating the deforming
body.
reaches 2.3. The total drag is considered to be the pressure drag because the
friction drag is almost negligible. In the case of k = 50, while the friction drag
slightly decreases, the pressure drag gradually increases until about t = 10
and remains 2.0, so that the total drag increases by 0.6.
Fig. 2. Mach number contours for a circu- Fig. 3. Mach number contours (k = 30)
lar cylinder
1.5 6.0
1.0 5.0
0.5 4.0
-0.5 2.0
Pressure drag
-1.0 1.0
Friction drag
-1.5 0.0
-1.5 -1.0 -0.5 0.0 0.5 1.0 1.5 0 150 300
X Nondimensional time
1.5
1.0
0.5
Y 0.0
-0.5
-1.0
-1.5
-1.5 -1.0 -0.5 0.0 0.5 1.0 1.5
X
Fig. 6. Mach number contours (k = 50) Fig. 7. Histories of body shape (k = 50)
860 Tomohisa Hashimoto, Koji Morinishi and Nobuyuki Satofuka
6.0
5.0
4.0
Cd 3.0
Total drag
2.0
Pressure drag
1.0
Friction drag
0.0
0 25 50
Nondimensional time
Conclusion
The relationship between the shape and its drag force of the elastic body
was investigated. As the elastic body becomes a slender shape, the total drag
increases, where the friction drag is almost negligible in comparison to the
pressure drag. A considerable difference is observed in the time histories of
pressure drag for different spring coefficients. In this paper the spring coeffi-
cient is assumed to be constant over the elastic body, however it is desirable
to control the drag force by changing the spring coefficient from place to place
so that the designated shape can be made under the fluid dynamic force.
References
1. Hashimoto, T., Morinishi, K., Satofuka, N., Numerical simulation of unsteady
flow around deformable elastic body under fluid dynamic force, ECCOMAS 2004,
Paper 364, 2004, pp.1-19
2. Jameson, A. and Baker T.J., Solution of the Euler equations for complex con-
figurations, AIAA Paper 83-1929,1983
3. Yoon, S. and Kwak, D., Multigrid convergence of an LU scheme, Frontiers of
CFD, 1994, pp.319-338
4. Steger, J.L. and Benek, J.A., On the use of composite grid schemes in computa-
tional aerodynamics, Computer Methods in Applied Mechanics and Engineering,
64, 1987, pp.301-320
CAD-Centric Framework for Aero-mechanical
Optimization - Counter-Rotating Fan Design
1 Introduction
Traditionally, the three dimensional blade shape is very often designed by ex-
perienced designers by iteratively and manually modifying the blade shape,
taking into account several results coming from aerodynamic computations
and structural mechanics computations (static and dynamic), among others.
In recent years, progress has been made in the development of automatic
optimization packages capable of optimizing complex shapes using advanced
CFD solver and optimization algorithms [4, 6]. However they are not yet
commonly used in industrial shape optimization. A more intensive use still
requires progress in the field of automatic shape optimization. Some of the
issues are 1) the large number of design variables typical of industrial prob-
lems, 2) the efficient and robust shape parameterization, and 3) the ability to
easily couple multiple simulation tools with minimum labor. The first two are
numerical efficiency issues, while the third one is often overlooked, but just as
important if automatic shape optimization is to be incorporated in a realistic
industrial environment. Emphasized here is the role of CAD systems as an
integral element of optimization framework. A major advantage of modern
CAD systems is that design intent can be incorporated into the CAD model
as a master model, which is a set of parameters that control features within
the model. However the simulation codes often lack access to the feature tree
and parameters defined in the model due to file translations, which results
in loss of information and robustness. In addition, manual intervention that
may be required to heal the model results in a procedure that is not general
enough and automated to be acceptable in industrial cases. It is therefore ad-
vantageous to integrate the CAD-system directly inside the simulation chain,
and an object-oriented framework was developed that can interact with the
underlying CAD systems transparently, based on the CAPRI [2] middleware
which in turn modifies the shape design variables, regenerates the CAD model,
and provides a native geometry representation that can be used for analyses.
2 The Optimizer
The optimization algorithm is based on the use of genetic algorithms because
they provide a very robust method responding to the criteria mentioned above.
One drawback of GAs is that they suffer from a slow convergence because of
its stochastic nature. One approach to accelerate the optimization process is
to use an approximate model. Ref. [3] presents an overview of several similar
optimization techniques. The blade design algorithm is therefore organized
with the following 5 steps :
1. Build a database using a design of experiments procedure (DOE).
2. Build an approximate model using the DOE points, i.e. construct an ana-
lytical relation between the design variables and the simulation responses.
3. Perform a GA optimization to find the optimum using the approximate
model to evaluate the objective functions and constraints.
4. Perform accurate simulations to evaluate and verify the real objective and
constraint values. This new simulation result is added to the database. The
database is therefore always improved with new design points, leading to
an improved approximate model.
5. Go to step 2 until the maximum number of design cycles specified by the
user is reached.
In this research, the radial basis function interpolation technique is used to
build an approximate model, mainly because of its robustness in providing a
more accurate approximate model. Moreover it allows constructing a global
approximate model which is valid for the entire design space. This is an im-
portant aspect for the application of the method to multiple objective op-
timization techniques based on the Pareto front concept, which may need
information from the whole search space. For industrial applications, the com-
putational cost of one optimization iteration mainly depends on the cost of
the simulation employed. In general the cost for building the approximate
model and running the genetic algorithm is from a few seconds to a few
minutes, depending on the number of training examples and the number of
input and output variables. More details about this method can be found in
Refs. [7, 6, 8].
For flow computations the TRAF code developed by the University of Florence
is used in this research [1]. The code can compute the flow in linear and annular
cascades as well as the flow in stationary and rotating blade passages with
tip clearance. The RANS equations are solved using a Runge-Kutta scheme
in conjunction with accelerating techniques : local time stepping, residual
smoothing and Full-Approximation-Storage (FAS) multigrid. These equations
are discretized using finite volumes and a cell-centered scheme with artificial
CAD-Centric Framework for Aero-mechanical Optimization 863
The aim of this optimization is to find the optimal geometry for 3 different
aerodynamic operating conditions : one at choked mass flow, one near effi-
ciency peak and one near the stall flow. The three points are analyzed at the
same blade rotational speed, which is the counter-rotating fan design speed.
These 3 objectives are transformed into a single objective by summing the
864 Hiromasa Kato, Stéphane Pierret, and Rajan Filomeno Coelho
individual penalties. Lower limits on the choked mass flow and the pressure
ratio are imposed as aerodynamic constraints. Structural constraints are also
imposed. An upper limit on the static stress due to centrifugal forces is im-
posed at 450MPa. Then a set of constraints are imposed on the frequencies
of the blade vibration modes. The rotational frequency of the first fan blade
is 72.765 Hz while this frequency is 65Hz for the second fan blade. These fre-
quencies are the first excitation frequencies and are called the 1N excitation
frequencies. The other excitation frequencies are the multiples of these 1N
frequencies. The excitation frequencies controlled during the optimization are
1N, 2N, 3N, 4N, 6N, 8N, and 10N respectively. These frequencies are used in
the constraints as forbidden ranges.
CAD-free Optimization
from 403 kg/s to 406.8 kg/s. The peak adiabatic efficiency point is increased
by 1.5% from 84.7 to 86.7%. The static stresses on the optimized geometries
are 483 MPa for the rotor 1 and 250 MPa for the second rotor.
CAD-Based Optimization
The original counter-rotating fan geometry has been imported into CATIA
V5. For this proof-of-concept test case only the rotor 1 tip section was modi-
fied during the optimization, leading to 7 design variables. Then a DOE with
21 geometries has been run, among which 14 geometries were computed suc-
cessfully. The objective is to maximize the sum of the stage efficiency at 3
operating conditions. A lower bound constraint is imposed on the pressure ra-
tio and choked mass flow. For this test case a constraint on the static stresses is
imposed while no constraint is imposed on the dynamic vibration frequencies.
The convergence history of the optimization process presented in figure
5 shows a rapid convergence to a better geometry in 14 design cycles. The
efficiency operating curve presented in figure 2 shows an efficiency improve-
ment of 0.5% and also a choked mass flow increase. This is a rather important
improvement considering the number of design variables. The geometry mod-
ifications are presented in figure 6.
5 Concluding Remarks
Acknowledgment.
This research was supported by the Walloon Region and the ERDF European
funds under contract no EP1A122030000102.
References
1. A. Arnone, M. Marconcini, and A. Scotti Del Greco. Numerical investigation of
three-dimensional clocking effects in a low pressure turbine. ASME 2003-GT-
38414, 2003.
2. R. Haimes and G. Follen. Computational analysis programming interface. Pro-
ceedings of the 6th International Conference on Numerical Grid Generation in
Computational Field Simulations, University of Greenwich, June 1998.
3. P. Hajela. Nongradient methods in multidisciplinary design optimization - status
and potential. Journal of Aircraft, pages 255–265, January-February 1999.
4. Yongsheng Lian. Multi-objective optimisation using coupled response surface
model and evolutionary algorithm. AIAA paper 2004-4323, 2004.
5. A. Merchant, J L. Kerrebrock, and A H Epstein. Compressors with aspirated
flow control and counter-rotation. AIAA-2004-2514.
6. S. Pierret. Multi-objective and multi-disciplinary optimization of three-
dimensional turbomachinery blades. 6th World Congresses of Structural and
Multidisciplinary Optimization, 30 May - 03 June 2005, Rio de Janeiro, Brazil.
7. S. Pierret. Multi-objective optimization of three-dimensional turbomachinery
blades. Int. Conf. on Computational Methods for Coupled Problems in Science
and Engineering, 25-28 May 2005, Santorini Island, Greece.
8. S. Pierret, H. Kato, R. Coelho, and A. Merchant. Multi-objective and multi-
disciplinary shape optimization. Evolutionary and Deterministic Methods for
Design, Optimization and Control with Applications to Industrial and Societal
Problems 2005.
Aeroelastic Solutions for Viscous Flows using
the Time Accurate and Non-Linear Frequency
Domain Methods
1 Introduction
tation into equation (1) yields the unsteady residual, Rˆ∗ k = i kV ŵk + R̂k . The
nonlinearity of the unsteady residual stems from the spatial operator which
can be computed using a Fast Fourier Transform (FFT) algorithm. Consistent
with the time accurate approach, a pseudo-time derivative can be added, and
a time-stepping scheme can be employed to numerically integrate the resulting
equations; V ∂∂τ ŵk
+ Rˆ∗ k = 0. In the NLFD case, an unsteady residual exists
for each wavenumber used in the solution and the pseudo-time derivative acts
as a gradient to drive the absolute value of all of these components to zero
simultaneously.
The governing coupled second-order structural equations of motion can be
written as
mḧ + Sα α̈ + Kh h = −L
Sα ḧ + Iα α̈ + Kα α = Mea
L is the lift (positive up) and Mea is the moment about the elastic axis (pos-
itive nose up). Sα , Iα , Kh , and Kα are the static imbalance, moment of
inertia, bending stiffness and torsional stiffness respectively. h is the plunging
coordinate (positive down) and α is the angle of attack (clockwise positive).
Given the periodic nature of the problem, both the plunging displacement
h and angle of attack α can be written in spectral forms. After taking advan-
tage of the orthogonality of the Fourier terms and substituting the spectral
representations into our structural equations yields a separate equation for
each wavenumber k in the solution
After solving the simple system above for ĥk and α̂k for each wavenumber we
can use an inverse FFT to determine our displacements at k equally spaced
points in our period. The displacement velocities ḣ and α̇, are computed using
a trigonometric interpolation of the displacements followed by a backward
difference of adjacent points.
Time Accurate and Non-Linear Frequency Domain Methods 869
3 Results
ωc
The reduced frequency and speed index are defined as kc = 2V ∞
and Vf =
V∞
bwα µ respectively. Structural constants used are similar to that of Isogai[4]:
√
approach, most of the computational cost is spent resolving the initial decay
of transients. The NLFD code is able to resolve the solution in as few as five
periods, resulting in a substantial savings. The number of periods needed to
reach a periodic steady state solution may vary for the time accurate method.
However, it remains clear that the NLFD method still offers tremendous sav-
ings due to the reduced number of samples per period needed to resolve the
periodic steady state solution.
An important distinction in the NLFD results (refer to Figs. 10 through
12) is the discontinuity between the first and second periods in the NLFD
solution. It is important for the reader to understand that we are solving for
the periodic steady state solution immediately in the NLFD case, bypass-
ing all the transients present in the time accurate solution. At steady state,
the unsteady motion of the airfoil prescribes non-zero loads at zero angle of
attack which, when input in the structural equations, would yield non-zero
displacements.
Conclusion
The viscous NLFD approach is able to produce accurate lift hysteresis loops
with just three time steps per period or one harmonic mode in addition to the
zeroth mode. Aeroelastic solutions were computed at M∞ = 0.825. Stable,
neutral, and unstable responses were obtained with just three time steps per
period. The viscous NLFD approach provides a cost effective aeroelastic so-
lution when compared to time accurate methods that require at least 24 time
steps per period. Hence, we can conclude that the NLFD method is indeed an
efficient and accurate alternative to the study of unsteady periodic problems.
References
1. J. J. Adamczyk. Model equation for simulating flows in multistage turboma-
chinery. Technical report, NASA Technical Memorandum 86869., Nov 1984.
2. J. J. Alonso and A. Jameson. Fully-implicit time-marching aeroelastic solu-
tions. AIAA paper 94-0056, AIAA 32nd Aerospace Sciences Meeting and Ex-
hibit, Reno, NV, January 1994.
3. K. C. Hall, J. P. Thomas, and W. S. Clark. Computation of unsteady non-
linear flows in cascades using a harmonic balance technique. Technical report,
9th International Symposium on Unsteady Aerodynamics, Aeroacoustics and
Aeroelasticity of Turbomachines, Lyon, France, September 2000.
4. K. Isogai. On the transonic-dip mechanism of flutter of a sweptback wing. AIAA
Journal, 17(7):793–795, July 1979.
5. F. Kachra and S. Nadarajah. Aeroelastic solutions using the time accurate and
non-linear frequency domain methods. AIAA, 44th Aerospace Sc. Meeting, 2006.
6. Matthew McMullen. The Application of Non-Linear Frequency Domain Methods
to the Euler and Navier-Stokes Equations. Ph.d. dissertation, Department of
Aeronautics and Astronautics, Stanford University, Stanford, CA, March 2003.
7. K. Willcox and J. Peraire. Aeroelastic computations in the time domain using
unstructured meshes. Int. J. for Num. Meth. in Engineering, 40:2413–2431,
1997.
Time Accurate and Non-Linear Frequency Domain Methods 871
0.1
NLFD−3 0−Mode
0 0.08
10 NLFD−3 1−Mode
NLFD−5 0−Mode 0.06
−2
10
NLFD−5 1−Mode 0.04
NLFD−5 2−Mode
Lift Coefficient
−4
log(Residual)
10 0.02
NLFD−7 0−Mode
0
10
−6 NLFD−7 1−Mode
Time Accurate, 23 Time Steps
NLFD−7 2−Mode −0.02
NLFD 1−Harmonic Inviscid
−8
10 NLFD−7 3−Mode −0.04
NLFD 2−Harmonic Inviscid
NLFD 3−Harmonic Inviscid
Experimental
−10 −0.06 NLFD 1−Harmonic Viscous
10
NLFD 2−Harmonic Viscous
−0.08 NLFD 3−Harmonic Viscous
−12 NLFD 4−Harmonic Viscous
10
NLFD 5−Harmonic Viscous
−0.1
200 400 600 800 1000 1200 1400 1600 1800 2000 −1 −0.8 −0.6 −0.4 −0.2 0 0.2 0.4 0.6 0.8 1
Multigrid Cycles Angle of Attack
1
Drag Coefficient
Lift Coefficient
8.8
NLFD 1−Harmonic Viscous
8 −2
7.8
−3
7.6
−1 −0.8 −0.6 −0.4 −0.2 0 0.2 0.4 0.6 0.8 1 −1 −0.8 −0.6 −0.4 −0.2 0 0.2 0.4 0.6 0.8 1
Angle of Attack Angle of Attack
8.8
0.005
8.6 0.022
h/b
0
0.021
8.4
−0.005 0.02
8.2 0.019
−0.01
0.018
8
−0.015
0.017
7.8
−0.02 0.016
0.01 0
3
−2
2 −4
0.005 −6
Plung Speed
1 −8
Angle of Attack
−10
0 −12
0
0.0135
22.2 22.4 22.6 22.8
0.013 −1
−0.005 0.0125
−2
0.012
1 Harmonic
0.0115 −3
2 Harmonics
−0.01
0.011 3 Harmonics
−4
20.2 20.4 20.6 20.8 21
−0.015 −5
18 20 22 24 26 28 30 32 18 20 22 24 26 28 30 32
Dimensional Time Dimensional Time
1 Harmonic 0.025
h/b
!
2 Harmonics
4 0.02
3 Harmonics
0.015
AOA and Plung Displacements
2 0.01
Rotational Speed
−4
x 10
20 0.005
0
15 0
10 −0.005
−2
5 −0.01
−0.015
0
−4
−0.02
−5
22 22.5 23
−0.025
−6
16 18 20 22 24 26 28 30 32 34 0 50 100 150 200 250 300
Dimensional Time Dimensional Time
0.03
0.02
0.02
AOA and Plung Displacements
AOA and Plung Displacements
0.01
0.01
0 0
−0.01
−0.01
−0.02
−0.02 h/b
h/b
!
! −0.03
−0.03
0 20 40 60 80 100 120 140 0 20 40 60 80 100
Dimensional Time Dimensional Time
Summary. Fluid-structure interaction is vital for the study of airflow induced vi-
brations, a cause of hard disk drive failure. With the loose fluid-structure coupling,
there is a need for an accurate exchange of information at the fluid-structure in-
terface. This work looks at two variations of the Shepard’s interpolation method to
assess their accuracy in such an exchange and their application in simulating the
airflow induced vibration on the actuator arm of a small one-inch form factor.
1 Introduction
As airflow induced vibrations is one of the primary causes of hard disk drive
(HDD) failure, it is important to understand the airflow characteristics and
fluid-structure interactions for the optimal design of the HDD. While re-
searchers [10] have performed experiments on simple rotating disks within a
casing, it still remains a challenge to perform experiments on the actual hard
disk drives. To this end, simulation is an efficient and economical tool in the
early design stages of the HDD. Coupling a Computational Fluid Dynamics
(CFD) code and a computational structural dynamics (CSD)code to address
the fluid-structure interaction problem, as studied by [8, 2], assumes that the
accuracy of the information exchanged depends on the accuracy of the inter-
polation algorithm used. In the present study, Shepard’s interpolation method
[7] is used in conjunction with several neighbouring point selection techniques
to evaluate the accuracy of these techniques in exchanging the information
for the airflow induced vibration studies for a one-inch HDD.
circumradius, R c
circumcenter, x 0 circumradius, R c
circumcenter, x 0
Z X
Y
Fig. 1. (a) HDD enclosure configuration and construction of the (b) local region of
influence via (c) circumsphere and (d) circumcirle methods.
ard’s interpolation for any node, x, within the radius Rc of each coarse mesh
element can be established as follows:
G : dk ≤ ε
w(x) = (dk )−n : ε < dk ≤ Rc (1)
0 : dk > R c
for some big number G and small tolerance, ε. The first criterion for Equation
1 is selected to avoid discontuinity of w(x) t xk .
The HDD enclosure in Figure 1 consists of a single disk rotating at 3600 rpm,
a latch, and an actuator arm with sliders. The material used for the actuator
arm is aluminium which has a density of 2700 kg/m3, Young’s Modulus of 70
GPa and a Poisson ratio of 0.33.
In order to compare the performance of the various weighting functions,
two test functions have been adapted from Gantovnik et. al. [5], i.e.
f1 (x) = 1 − 2/3(|x1 − 0.001| + |x2 − 0.02| + |x3 − 0.003|) (2)
3
X
2/3(x + x + x ) , xi ≤ 0.03
1 2 3
f2 (x) = i=1 (3)
X3
−2/3(x + x + x ) + 0.05 , xi > 0.03
1 2 3
i=1
Modal analysis has been performed on the actuator arm to compute its
natural frequencies prior to the fluid-structure interaction. Figure 2(a) and
(b) show the first bending and sway modes contributing to the on-track and
876 S. Ali and M. Damodaran
Fig. 2. First (a) bending (mode 1 = 2.066kHz) and (b) sway modes (mode 5 =
17.917kHz) and the initial flow characteristics (c) Gauge static pressure and (d)
Velocity Vectors in the vicinity of the sliders
−9 −10
10 10 −150
on−track
off−track
−10 −11 −160
10 10
−170
−11 −12
10 10
displacement (m)
displacement (m)
circumcircle
Decibel [dB]
circumsphere −180
−12 −13
10 10 Shepard
−190
circumcircle
−13 −14
10 circumsphere 10
Shepard −200
−14 −15
10 10
−210
−15 −16
10 10 −220
1 2 3 4 5 6 7 8 9 10 1 2 3 4 5 6 7 8 9 10 0 5 10 15 20 25
time (s) x 10
−3
time (s) −3 frequency [kHz]
x 10
Based on the displacement amplitudes of Figure 3(a) and (b), the accu-
racy of the interpolation results affects the on-track displacements more than
the off-track displacements. The mean relative errors shown in Table 2 clearly
indicate that the results from the circumcircle technique are relatively more
accurate than those from the circumsphere technique. This difference in accu-
racy could be attributed to the size of the coarser mesh elements in relation
to the thickness of the actuator arm. The inaccuracy of the interpolation re-
sults of circumsphere technique arises from the inclusion of pressure values
from both the upper and lower surfaces of the arm as opposed to those of
the circumcircle technique due to the size of the coarse mesh spanning both
surfaces. However, the mean relative interpolation errors shown in Table 2
for both techniques based on pressures from the nearest neighbouring fluid
nodal points are less than ten percent of the actual pressure values of these
neighbouring points. These figures give some assurance on the accuracy of the
two techniques in capturing the pressure distribution.
Using Discrete Fourier transformation on the displacements determined
from Abaqus for the circumcircle technique, the frequency analysis is plotted
in Figure 3(c). Though offsetted by about 30 dB, the off-track and on-track
power spectrum patterns are similar with several peaks. The first peak could
be found at 2 kHz which is close to the natural frequency of the actuator as
shown in Figure 2(a) and similarly, the peak at 18 kHz is close to the natural
frequency of the fifth mode as shown in Figure 2(b). These close proximities
to the natural frequencies are likely to lead to the problem of vibrational
resonance and affect the integrity of data being written or retrieved by the
actuator arm. However, this analysis is still incomplete due to the lack of
experimental data validation and the fact that the slider to disk gap used for
the present study is around 9 µm which is much larger than the nano-sized
gap of the actual HDD assembly.
Conclusions
The role of interpolation methods in ensuring the accuracy of the information
passed between the CFD and CSD solver is explored in this paper and it is
878 S. Ali and M. Damodaran
found that the displacement amplitudes are sensitive to the difference in in-
terpolation results. The two techniques greatly improve on the interpolation
results of the Shepard’s method by enforcing selection rules for the local sub-
set of the data points with the circumcircle technique observed to be more
accurate. As fluid-structure interactions are applied to smaller structures, the
accuracy of the interpolation methods used in the loose-coupling approach
play an increasingly important role in ensuring the accuracy of the results.
References
1. ABAQUS Inc.: Abaqus documentation, Version 6.5.4 (2005).
2. Damodaran, M., Ali, S. and Dayanandan, S.: Computational Fluid Dynamics in
the Analysis and Design of Engineered Systems. In: Caughey, D.A. and Hafez,
M.A. (ed) Frontiers of Computational Fluid Dynamics 2006, World Scientific,
37-48, (2005).
3. FLUENT Inc.: Flow simulation software user guide, Fluent Version 6.2.16
(2005).
4. Franke, R., and Nielson, G.: Smooth Interpolation of Large Sets of Scattered
Data. International J. Numerical Methods Engineering, 15, 1691-1704 (1980).
5. Gantovnik, V.B., Gurdal, Z., and Watson, L.T.: Linear Shepard Interpolation
for High Dimensional Piecewise Smooth Functions. AIAA 2004-4486, Proceed-
ings of 10th AIAA/ISSMO Multidiscplinary Analysis and Optimization Con-
ference, Albany, New York, 30 Aug - 1 Sep (2004).
6. Goh, H.J., Damodaran, M. and Ng, Q.Y.: Modeling Airflow and Particle Tra-
jectories near the Head/Disk Interface Region of a Small Form Factor Hard Disk
Drive Enclosure. WTC2005-63630, ASME-STLE World Tribology Conference
III, Washington DC, US, 12- 16 Sep (2005).
7. Shepard D.: A Two-Dimensional Interpolation Function for Irregularly Spaced
Data. Proceedings of the 1968 23rd ACM national conference, 517-524 (1968).
8. Tatewaki, M., Tsuda, N., and Maruyama, T.: A Numerical Simulation of Un-
steady Airflow in HDDs. Fujitsu Sci. Tech. Journal, 37, 2, 227-235 (2001).
9. Weisstein. E.W.: Circumsphere. In: MathWorld–A Wolfram Web Resource.,
CRC Press LLC (1999). http://mathworld.wolfram.com/Circumsphere.html
10. Wu, S.C. and Chen, Y.M.: Phase-Averaged Method Applied to Periodic Flow
between Shrouded Co-Rotating Disks. International Journal of Rotating Ma-
chinery, 8, 2, 413-421 (2002).
Computation of Fluid Structure Interaction
with Application to Three-Dimensional
Combustion System
Zinedine Khatir
1 Introduction
The numerical simulation of FSI problems occurs in many engineering and sci-
entific applications, ranging from airfoil oscillations [1] or aero-hydrodynamics,
haemodynamcis [2] to blade flutter analysis in turbomachinery [3] and power
generation when designing gas turbine combustors [4, 5] for instance. Un-
like tightly-coupled approaches where the fluid and structural equations are
solved in a single computational domain, loosely-coupled partitioned proce-
dures [6, 7] use independent techniques for the fluid and structure subdomains
and exchange data along the fluid-structure interface. The loosely-coupled ap-
proach can thus take the full advantage of existing, well developed and tested
codes for fluid and structural analysis.
In this article, we are interested in enforcing the interface boundary con-
dition by transferring the fluid loads to the structure, while displacement is
transferred back to the fluid from the structure [8]. Various methods have been
developed for the problem of data transfer with non-matching meshes [9]. A
detailed review of data transfer methods is given in [10], which can be clas-
sified as methods suitable for the transfer of: (i) load only, (ii) displacement
only and (iii) both load and displacement. The volume spline interpolation
method by Hounjet et al [11] used for the data transfer in the present paper
belongs to the third class. After a brief description of the method, the accuracy
of the interpolation is assessed against simple 2D and 3D test cases. We will
then apply the methodology for the application of FSI in three-dimensional
combustion system, and the numerical results show good agreement with ex-
perimental data.
2 Fluid-Structure Coupling
Principle. The evaluation of the structure load induced by the fluid and
the displacement of the CFD grid induced my the structural motion needs
to satisfy the requirement of conservation of work and accuracy. The princi-
ple of virtual work is usually used to ensure conservation of energy. For this
purpose, a linear transformation is sought. If the displacement ∆xa of the
CFD grid is expressed by means of the structural displacement ∆xs using the
transformation matrix [G] such that:
∆xf = [G] ∆xs (1)
then the requirement for conservation leads to a corresponding matrix for the
transformation of the fluid and structural load forces ff and fs
fsT ∆xs = ffT ∆xf = ffT [G] ∆xs and fs = [G]T ff . (2)
Volume Spline Interpolation. Volume spline interpolation is a very simple
method to interpolate non-uniformly spaced data. The method belongs to the
category of scattered interpolation methods using radial basis functions [12,
13] and is well suited for transferring both load and displacement. Consider
a set of unordered support points xi , i = 1 . . . N with functions values {fi ∈
R}, = 1 . . . N , and the interpolation function defined as:
N
X
f (x) = α0 + αi Ri (x) (3)
i=1
where N a and N s are the number of fluid and structural elements respectively.
The functions E a and E s are given by
s
p
Elm = (xl − xm )2 + (yl − ym )2 + (zl − zm )2
p (5)
a
Elm = (xal − xm )2 + (yla − ym )2 + (zla − zm )2
with (x, y, z) and (xa , y a , z a ) are the coordinates of the structural points and
fluid points respectively. The coupling matrix [G] is then obtained by deleting
the first column of [G]∗ .
3 Numerical Results
2D Simple Test. The interface between CFD and CSD consists of the unit
circle centered at the origin. As seen in Fig. 1-a the two meshes are non-
matching. The aerodynamic input field points is obtained from 30 discretized
points on the unit circle, whereas the structural input field is taken from 23
discretized points on the circle with radius 1.05. After that, the following
deformation is applied to the aerodynamic input points (xi , yi ) to obtain the
new position (x∗i , yi∗ ):
For i=1 to 30:
∗ ∗
θ = 2π
30 × (i − 1); xi = r cos θ = xi ; yi = r sin θ = yi ;
∗ 3
If xi > 0.8: xi = xi − 0.8 ∗ xi
End
The final solution is depicted in Fig. 1-b after interpolation. The deforma-
tion of the structural mesh is of very good quality. The interpolated solution
is compared with the exact solution. The maximum obtained error is around
0.07 which gives second order precision for this test E ' (δx)2 .
3D Simple Test. Here a random pressure field is applied onto the unstruc-
tured CFD mesh as in Fig. 2. This pressure field is then interpolated onto
the structured CSD mesh. As shown in Fig. 2, the input random pressure
field and the interpolated pressure field on the CSD mesh compare well. The
maximum obtained error is E ' 0.008 which leads to a third order precision
with this kind of discretization.
882 Zinedine Khatir
Fig. 1. Two circular concentric non-matching meshes, (dots) CFD mesh and
(square) CSD mesh configuration: (a) Before deformation and (b) After deformation.
(a) (b)
Fig. 2. (a) CFD and (b) CSD non-matching meshes alongside pressure fluctuation.
Fig. 3. (Left) Experimental setup and (Right) Computational domain [5, 17].
3D Fluid Structure Interaction Computation 883
on the liner. Results compare well. The time evolution of the loads is plotted
in Fig. 6. CFD and CSD loads after transfer coincide. Fig. 7 demonstrate
the energy conservation during the simulation. A mean error of 1.26 × 10−5
and 9.01 × 10−4 is found for the time evolution of the 1st and 2nd moment
respectively. This proves again an accurate data transfer between the 2 non-
matching grids as plotted in Fig. 8.
Conclusion
A volume spline method for the data transfer in transient FSI computation
with non-matching meshes has been presented. Results show good agreement
with analytical and experimental results. Further investigations are still being
Fig. 7. Time evolution of 1st (Top) and 2nd (Bottom) Moment (w.r.t axis x=0.895
m) and relative errors.
884 Zinedine Khatir
Fig. 8. Surface plot of CFD (Left) and CSD (Right) loads at time t=0.485s.
Jubaraj Sahu
1 Introduction
2 Numerical Procedure
present numerical study is a big step forward and a direct extension of that
research which now includes numerical simulation of the actual fight paths
of the projectile using coupled CFD/RBD techniques using real-time accu-
rate approach. The complete set of 3-D time-dependent Navier-Stokes equa-
tions is solved in a time-accurate manner for simulations of actual flights. A
commercially available code, CFD++ [6-9], is used for the time-accurate un-
steady CFD simulations. The basic numerical framework in the code contains
unified-grid, unified-physics, and unified-computing features. The user is re-
ferred to these references for details of the basic numerical framework. The
3-D time-dependent Reynolds-averaged Navier-Stokes (RANS) equations are
solved using the finite volume method [7]:
Z I Z
∂
W + [F − G]dA = HdV (1)
∂t V V
where W is the vector of conservative variables, F and G are the inviscid and
viscous flux vectors, respectively, H is the vector of source terms, V is the cell
volume, and A is the surface area of the cell face.
Second-order discretization was used for the flow variables and the tur-
bulent viscosity equation. The turbulence closure is based on topology-
parameter-free formulations. Two- equation [6] and higher order hybrid RANS/
LES [10,11] turbulence models were used for the computation of turbulent
flows. The hybrid RANS/LES approach based on Limited Numerical Scales
(LNS) is well suited to the simulation of unsteady flows and contains no addi-
tional empirical constants beyond those appearing in the original RANS and
LES sub-grid models. With this method a regular RANS-type grid is used
except in isolated flow regions where denser, LES-type mesh is used to re-
solve critical unsteady flow features. The hybrid model transitions smoothly
between an LES calculation and a cubic k − ε model, depending on grid fine-
ness. These models are ideally suited to unstructured book-keeping and mas-
sively parallel processing due to their independence from constraints related
to the placement of boundaries and/or zonal interfaces. For computations of
unsteady flow fields that are of interest here, dual time-stepping as described
below was used to achieve the desired time-accuracy [12]. In addition, the
projectile in the coupled CFD/RBD simulation along with its grid is actually
moved as it flies downrange.
Grid velocity is assigned to each mesh point. This general capability can be
tailored for many specific situations. For example, the grid point velocities can
be specified to correspond to a spinning projectile. In this case, the grid speeds
are assigned as if the grid is attached to the projectile and spinning with it.
Similarly, to account for rigid body dynamics, the grid point velocities can
be set as if the grid is attached to the rigid body with six degrees of freedom
(6 DOF). For the rigid body dynamics, the coupling refers to the interaction
between the aerodynamic forces/moments and the dynamic response of the
projectile/body to these forces and moments. The forces and moments are
computed every CFD time step and transferred to a 6DOF module which
Unsteady Aerodynamics of Projectiles using an Unstructured Technique 889
computes the body’s response to the forces and moments. The response is
converted into translational and rotational accelerations that are integrated
to obtain translational and rotational velocities and integrated once more to
obtain linear position and angular orientation. The 6DOF rigid body dynamics
module uses quaternions to define the angular orientations. However, these
are easily translated into Euler angles. From the dynamic response, the grid
point locations and grid point velocities are set. Typically, we begin with
a computation performed in ”steady state mode” with the grid velocities
prescribed to account only for the translational motion component of the
complete set of initial conditions. At this stage we also impose the angular
orientations from the initial conditions. The complete set of initial conditions
includes both translational and rotational velocity components along with
initial position and angular orientation.
3 Results
Time-accurate unsteady numerical computations were performed using Navier-
Stokes and coupled 6-DOF methods to predict the flow field and aerodynamic
coefficients, and the flight paths of a finned projectile at supersonic speeds. In
all cases, full 3D computations were performed and no symmetry was used.
The supersonic projectile modeled in this study is an ogive-cylinder-finned
configuration (see Figure 1). The length of the projectile is 121 mm and the
diameter is 13mm. The ogive nose is 98.6 mm long and the afterbody has a
22.3 mm, 2.5◦ boat-tail. Four fins are located on the back end of the projectile.
Each fin is 22.3 mm long and 10.16 mm thick. An unstructured computational
mesh was generated for this projectile. In general, most of the grid points are
clustered in the boundary-layer and afterbody fin regions. The total number
of grid points is about 4 million for the full grid.
Here, our primary interest is in the development and application of coupled
CFD and RBD techniques for accurate simulation of the free flight aerody-
namics and flight dynamics of the projectile in supersonic flight. The first
step was to obtain the steady state results for this projectile at a given ini-
tial supersonic velocity. Also imposed were the angular orientations at this
stage. Corresponding converged steady state solution was then used as the
starting condition along with the other initial conditions for the computa-
tion of coupled CFD/RBD runs. Numerical computations have been made for
these cases at an initial velocity of 1034 m/s. The simulations were started
a small distance away from the muzzle. The corresponding initial angle of
attack was, α = 4.9◦ and initial spin rate was 2500 rad/s. Figure 2 shows
Fig. 2. Computed pressure contours Fig. 3. Euler pitch angle vs. x-distance
lations and the 6DOF analysis of the flight results from ARFDAS, software
commonly used for this purpose. Computed results match very well with the
experimental flight test results. The unsteady simulations took thousands of
hours of CPU time on a Xeon PC cluster system running with either 32 to 64
processors.
Another case considered in the study is a complex canard-controlled finned
projectile. Here, the control maneuver is achieved by the two horizontal ca-
nards in the nose section (Figs. 6-8). Unstructured Chimera overlapping grids
were used (see Fig. 6) and solutions have been obtained for several canard
892 Jubaraj Sahu
Conclusions
References
1. Sahu, J., K. R. Heavey, and E. N. Ferry, ”Computational Fluid Dynamics for
Multiple Projectile Configurations.” Proceedings of the 3rd Overset Composite
Grid and Solution Technology Symposium, Los Alamos, NM, October 1996.
2. Sahu, J., K. R. Heavey, and C. J. Nietubicz, ”Time-Dependent Navier-Stokes
Computations for Submunitions in Relative Motion.” 6th International Sym-
posium on Computational Fluid Dynamics, Lake Tahoe, NV, September 1995.
3. Meakin, R. L., ”Computations of the Unsteady Flow About a Generic
Wing/Pylon/Finned-Store Configuration.” AIAA 92-4568-CP, August 1992.
Unsteady Aerodynamics of Projectiles using an Unstructured Technique 893
4. Smith B. L., and A. Glezer, ”The Formation and Evolution of Synthetic Jets.”
Journal of Physics of Fluids, vol. 10, No. 9, September 1998.
5. Davis, S. A. and Glezer, A., ”The Manipulation of Large- and Small-Scales
in Coaxial Jets using Synthetic Jet Actuators”, AIAA Paper No. 2000-0403,
January 2000.
6. Batten, P., U. Goldberg and S. Chakravarthy, ”Sub-grid Turbulence Modeling
for Unsteady Flow with Acoustic Resonance”, AIAA Paper 00-0473, 38th AIAA
Aerospace Sciences Meeting, Reno, NV, January 2000.
7. Peroomian, O., S. Chakravarthy, and U. Goldberg, ”A ’Grid-Transparent’
Methodology for CFD.” AIAA Paper 97-07245, 1997.
8. Peroomian, O., S. Chakravarthy, S. Palaniswamy, and U. Goldberg, ”Conver-
gence Acceleration for Unified-Grid Formulation Using Preconditioned Implicit
Relaxation.” AIAA Paper 98-0116, 1998.
9. Goldberg, U. C., O. Peroomian, and S. Chakravarthy, ”A Wall-Distance-Free
K-E Model With Enhanced Near-Wall Treatment.” ASME Journal of Fluids
Engineering, Vol. 120, pp. 457-462, 1998.
10. Avancha, R. and Pletcher, R. H., ”Large Eddy Simulation of the Turbulent Flow
Past a Backward Facing Step”, AIAA paper No. 2000-0542, January 2000.
11. Arunajatesan, S. and Sinha, N., ”Towards Hybrid LES-RANS Computations
of Cavity Flowfields”, AIAA Paper No. 2000-0401, January 2000.
12. Sahu, J., ”Unsteady CFD Modeling of Aerodynamic Flow Control over a Spin-
ning Body with Synthetic Jet.” AIAA Paper 2004-0747, Reno, NV, 5-8 January
2004.