Genetic Algorithm (real-coded ARGA) is used for design optimization.

To represent flow fields accurately and produce reliable designs, threedimensional Navier-Stokes computation is used for aerodynamic analysis. To reduce turn-around time, the computation is parallelized on the SGI ORIGIN2000 cluster at the Institute of Fluid Science, Tohoku University of Japan, by distributing flow analyses of design candidates to 64 processing elements. The present method is applied to aerodynamic redesign of NASA rotor67 20. THREE-DIMENSIONAL NAVIER-STOKES SOLVER FOR CASCADE FLOW Flow field inside high-speed axial-flow turbomachinery is highly three-dimensional and involves significant viscous effects, such as boundary-layer separations and shock wave/boundary layer interactions. Therefore three-dimensional Navier-Stokes computations are essential for blade shape optimization because further improvement in the aerodynamic performance requires detailed knowledge of the flow structure such as secondary flows and tip clearance flow. In this study, the three-dimensional NavierStokes code TRAF3D21,22 is used for aerodynamic analysis of blade designs. Capability of the present code has been validated by comparing the computed results to some experiments such as the Goldman annular vane with and without end wall contouring, the low speed Langston linear cascade21 as well as the NASA rotor67 22. The present code solves the three-dimensional full Reynolds-averaged Navier-Stokes equations. The present code uses a central-differencing scheme including artificial dissipation terms introduced by Jameson, Schmidt, and Turkel23 to maintain stability and to prevent oscillations near shocks or stagnation points. In order to minimize the amount of artificial diffusion inside the shear layer, the eigenvalues scaling of Martinelli24 and Swanson and Turkel25 are used. The two-layer eddy-viscosity model of Baldwin and Lomax is used for the turbulence closure. The system of the differential equations is advanced in time using an explicit four-stage Runge-Kutta scheme. In order to accelerate convergence of calculations, local time-stepping, implicit residual smoothing26, and the Full Approximation Storage (FAS) multigrid technique27 are used. At the subsonic axial inlet, the flow angles, total pressure and total enthalpy are specified according to the theory of characteristics while the outgoing Riemann invariant is taken from the interior. At the subsonic axial outlet, the average value of the static pressure at the hub is prescribed and the density and components of velocity are extrapolated together

with the circumferential distribution of pressure. The radial equilibrium equation is used to determine the spanwise distribution of the static pressure. On sidewalls, the momentum equation, the no-slip condition, and the temperature condition are used to compute pressure and density. For the calculations presented in this paper, all the walls have been assumed to be adiabatic. The periodicity from blade passage to blade passage is imposed by setting periodic phantom cell values. At the wake, where the grid is not periodic, the phantom cells overlap the real ones. Linear interpolations are then used to compute the value of the dependent variables in phantom cell. The three-dimensional grids are obtained by stacking two-dimensional grids generated on the blade-to-blade surface. These two-dimensional grids are of C-type and are elliptically generated, with controlled grid spacing and orientation at the wall. The problem of grid skewness due to high stagger or large camber is addressed by allowing the grid to be non-periodic on the wake28. By adding lines near the wall, viscous grids are obtained from the inviscid grids. The wall normal spacing scaled with the axial chord is 10-4. In the spanwise direction a standard Htype structure has been adopted. Near the hub and tip walls geometric stretching is used for a specified number of grid points, after which the spanwise spacing remains constant. The number of the grid points is 201 chordwise x 53 tangential x 57 spanwise. Among the 201 chordwise grid points, 149 grid points are distributed along the blade shape. The computational grid for NASA rotor67 is shown in Fig. 1. BLADE SHAPE PARAMETERIZATION Here a rotor blade shape is represented by four blade profiles, respectively at 0%, 31%, 62%, and 100% spanwise stations (all spanwise locations discussed here are measured from the hub) and linearly interpolated. Each of these sectional profiles can be uniquely defined by using a mean camber line and a thickness distribution and they are parameterized by the third-order B-Spline curves. Parameterization using B-Spline curves is one of the most popular approaches for airfoil designs. When B-Spline curves are used for shape parameterization, positions of control points of the BSpline curves are often considered as the design parameters. Here, five control points are used for the mean camber line as illustrated in Fig 2. For the thickness distribution, two control points are added at the leading edge and the trailing edge so that these points represent leading edge and trailing edge radii, respectively. Chordwise locations of the control points at leading edge and trailing edge are frozen to

3 American Institute of Aeronautics and Astronautics

zero and one, respectively. As a result, 14 design parameters are required to represent a sectional shape. Each blade shape is then represented with 56 design parameters. EVOLUTIONARY ALGORITHM EAs mimic mechanism of natural evolution, where a biological population evolves over generations to adapt to an environment by selection according to fitness, recombination and mutation of genes (Fig. 3). In EAs, a design candidate, objective function values, and design variables usually correspond to an individual, fitness, and genes, respectively. Starting with an initial population of design candidates that is often generated by random sampling from the design space, EAs select good design candidates in terms of fitness, which is assigned on the bias of their objective function values. Typically, fitness of a design candidate is its objective function value itself for a single objective problem. Recombination is applied, where new population is generated by exchanging features of the selected designs with the intent of improving the fitness of the next generation. Then, mutation is applied to design parameters of the new population to maintain diversity in the population. One of the key features of EAs is that it searches from multiple points in the design space in contrast to the traditional methods that usually move from a single design point. In addition, EAs use objective function values alone to determine a search direction and do not require gradients of the objective function while the traditional methods use local gradient information of an objective function. These features also lead to advantages such as, 1) Robustness: Deterministic methods, such as the gradient-based methods, typically start with a single design point and use the local gradient information to determine a search direction. As a result, they generally lead to a local, not necessarily a global optimum near the starting point. In contrast to them, EAs determine their search direction globally and probabilistically but efficiently using their unique operators socalled recombination and mutation that give EAs capability of finding global optimums. Compared with other probabilistic methods such as the simulated annealing method29 that is similar to the gradient-based methods but tries a random step according to the so-called Boltzmann probability distribution, EAs are more robust because they maintain a population of design candidates and they don t use function gradients that direct the search toward a local optimum. In addition, EAs have a capability to handle any design problems that may involve

non-differentiable objective function and/or a mix of continuous, discrete, and integer design parameters. 2) Suitability to parallel computing: Because EAs are population-based search algorithms, all design candidates in each generation can be evaluated in parallel by using the simple masterslave concept. Parallel efficiency is extremely high, if objective function evaluations consume most of the computational time. Aerodynamic design optimization is a typical case. 3) Simplicity in coupling evaluation codes: Because EAs use only objective function values of design candidates, EAs do not need substantial modification or sophisticated interface to evaluation codes. If an all-out re-coding were required to every optimization problem, extensive validation of the new code would be necessary every time. EAs can save such troubles. 4) Straightforward application to multiobjective optimization problems: Because EAs maintain multiple designs, EAs can find compromised optimum designs, so-called Pareto-optimal solutions, by introducing Pareto-optimal concept. In the present study, the real-coded AdaptiveRange Genetic Algorithm30 (real-coded ARGA) is used. The real-coded ARGA is an EA that can solve large-scale design optimization problems very efficiently by promoting the population toward promising design regions during the optimization process. To represent design parameters of design candidates, the floating-point representation31 is used where an individual is characterized by a vector of real numbers. It is natural to use the floating-point representation for real parameter optimization problems instead of binary representation, because it is conceptually closest to the real design space, and moreover, the string length is reduced to the number of design variables. The parental selection consists of the stochastic universal sampling32 and the ranking method31. To handle design constraints, the constrained domination approach33 is used. Blended crossover34 (BLX-0.5) is used for recombination. Mutation takes place at a probability of 10% and then adds a random disturbance to the corresponding gene. The present EA adopts the elitist strategy35 where the best and the second best individuals in each generation are transferred into the next generation without any recombination or mutation. Population size is set to 64. The main concern related to the use of a threedimensional Navier-Stokes solver for aerodynamic shape design is the required computational effort.

4 American Institute of Aeronautics and Astronautics

Algorithms, Morgan Kaufmann Publishers, Inc., San Mateo, California, 1987, pp 14-21. 33) Deb, K., Pratap, A. and Moitra, S., Mechanical Component Design for Multiple Objectives Using Elitist Non-Dominated Sorting GA, Lecture Notes in Computer Science 1917 Parallel Problem Solving from Nature PPSN VI, edited by Schoenauer, M., Deb, K., Rudolph, G., Yao, X., Lutton, E., Merelo, J. J., and Schwefel, H.-P., Springer, Berlin, Germany, 2000, pp.859-868. 34) Eshelman, L. J. and Schaffer, J. D., Real-Coded Genetic Algorithms and Interval Schemata, Foundations of Genetic Algorithms.2, Morgan Kaufmann Publishers, Inc., San Mateo, California, 1993, pp 187-202. 35) De Jong, K. A., An Analysis of the Behavior of a Class of Genetic Adaptive Systems, Doctoral Dissertation, University of Michigan, Ann Arbor, 1975.

camber line
0.2 0.15

z/c

0.1
0.25

blade profile
control points B-Spline curves 0.2 0.4 0.6 0.8 1
0.15 0.2

0.05 0 0

x/c
z/c
0.06 0.05 0.04

thickness distribution

0.1 0.05 0 0 0.2 0.4 0.6 0.8 1

z/c

0.03 0.02 0.01 0 0 0.2 control points B-Spline curves 0.4

x/c

x/c

0.6

0.8

Figure 2. B-Spline curves for mean camber line and thickness distribution and the resultant blade profile.

initial population evaluation of designs Next generation selection according to fitness recombination of selected designs

mutation applied to new designs Figure 3. Flowchart of typical evolutionary algorithms.

0.01

entropy production

0.0095 0.009 0.0085 0.008 0.0075 0.007 0 20 40

rotor67 optimized designs

Figure 1. Computational grid over NASA rotor67.

60

80

100

generation
Figure 4. Optimization history in terms of entropy production.

8 American Institute of Aeronautics and Astronautics