Computational Optimization of a Two Jet Flow Control System

1
Computational Optimization of a Two Jet Flow
Control System
Liang Huang and Raymond LeBeauDept. of
Mechanical Engineering, University of Kentucky
Increasingly powerful and less expensive
computers combined with improved search
algorithms have opened new horizons for
computational optimization problems. The example
presented is the optimization of a flow control
system for an airfoil consisting of a single
blowing jet and a single suction jet. Flow
control systems like this are designed to reduce
the drag and increase the lift of an airplane
wing. The best jet configuration was determined
through linking a genetic search algorithm with a
computational fluid dynamics solver. After
evaluation of thousands (out of millions) of
possible solutions, an optimal solution was
determined. The results indicate that the
location of the suction jet, followed by the
orientation of the suction jet and location of
the blowing jet, were the most important
parameters for this optimizing this system.
Overall, the best configuration increased lift by
about 4 and reduced drag by about 12. This
project demonstrates the potential applicability
of the techniques employed in this project to
larger-scale, more complex problems. This
combination of search algorithms and
computational fluid dynamics potentially will
train many-jet flow control as well as other
systems regulating fluid flows.
The statistics of the evolution are shown for all
five parameters in this figure. The suction
location is the first to converge, consistent
with this parameter having the strongest effect
on the flow field. The blowing angle and
amplitude do not completely con-verge after 100
generations, which reflects their relatively
insignificant effect on the overall lift and drag
of the final airfoil configuration.
Abstract
Suction Location
Suction Angle
Blowing Angle
Blowing Location
Blowing Amplitude
The optimum configuration results in an increase
in lift of about 4 and a decrease in drag of
about 12. Separating the suction and blowing
jets, pressure distribution (above) shows that
the suction jet primarily increases lift on the
leading half of the airfoil, while the blowing
actually reduces overall lift slightly while
reducing drag on the trailing half. Physically,
both jets weaken the separation bubble (left).
Theoretically with stronger or more jets, it is
possible to nearly eliminate the region of
separation. Future simulations will add more jets
into the system.
Methodology
A genetic search algorithm (GA) searches a
complex parameter space by allowing genomes, in
this case a set possible jet configurations
(strength, location, angle), to evolve over a
series of generations. At each generation, the
likelihood of genome survival to the next
generation is based on the fitness of that
configuration. The fitness is a combination of
the lift and drag over the airfoil, evaluated by
a computational fluid dynamics (CFD) simulation.
Techniques such as crossover and mutation
maintain diversity and prevent convergence to a
local, but not global, optimum. For this study,
the test case was a NACA0012 at an angle of
attack of 18o and a Reynolds number of 500,000.
One hundred generations of 32 genomes each were
evaluated, with each genome consisting of five
parameters (suction jet angle and location,
blowing jet strength, angle, and location), the
suction jet strength being fixed. These types
of large-scale CFD projects, involving thousands
of simulations, have become more practical with
the emergence of commodity clusters that can
provide hundreds of thousands of high-power CPU
hours per year for a relatively low cost. This
project has relied on commodity clusters as a
primary source of computer power. Additional
simulations have been performed on CFD-ME, a new
computer system at the University of Kentucky for
CFD research.
The test case for the presented results is the
NACA0012 airfoil shown above. The figure shows
the parameters of a single jet location,
measured in terms of percent chord (0 is the
leading edge, 1 is the trailing edge), jet
amplitude, measured as jet velocity relative to
the wind speed, and jet angle, where -90o to 0o
indicates suction and 0o to 90o indicates
blowing. For the study presented here, the
suction jet velocity was fixed at 3 of the
freestream wind.
Outcomes

• GA-CFD successfully locates optimal
  configuration of two-jet system.
• Optimal configuration consistent with expected
  flow physics.
• Optimal configuration increases lift by 4,
  decreases drag by 12.
• Suction jet location and angle and blowing jet
  location are the more
• important parameters in this case.
• Improved GA leads to faster convergence in test
  cases.
• The time to complete GA-CFD is currently 45-70
  days on 45-60 processors.
• GA-CFD is readily executable on lower cost
  commodity clusters.
• Mr. Liang Huang has nearly completed his PhD
  based on this research.
• Paper presented at AIAA Aerospace Sciences
  Meeting (Jan, 2004) and to
• be presented the Third International
  Conference on Computational Fluid
• Dynamics (June, 2004).
• Paper will be published in AIAA Journal of
  Aircraft (Accepted, Oct, 2003).
• Ongoing research includes further reducing
  computational time, evaluating
• more complex jet configurations, and
  increasing simulation accuracy.

The plots to the left show a comparison of the
original genetic algorithm (top) and the improved
genetic algorithm (bottom) solving Ackleys test
function. Results with four different sets of
initial genomes are shown, the solid line showing
the mean value at each generation. The improved
algorithm increases diversity early and hastens
convergence later. All cases con-verge to the
test function solution (about 0.005), but the
improved algorithm converges faster, achieving
the solution at about generation 600 as opposed
to 900.
Kentucky Fluid Cluster 2 (KFC2) is a commodity
cluster used in this project. It consists of 48
processors (1.8 GHz Athlons) linked by a four-way
channel-bonded ethernet network. Total cost in
January, 2003 for this system was about 35K.

Thanks to the Kentucky Science and Engineering
Foundation, Kentucky NASA EPSCoR, and the KAOS
Laboratory, UK-CFD, and the CFD-ME project at the
University of Kentucky for their support of this
research.