You are on page 1of 22

VKI lecture series on Optimization Methods & Tools for Multicriteria/Multidisciplinary Design November 15-19, 2004

AERONAUTICAL APPLICATIONS OF EVOLUTIONARY DESIGN


I. Kroo Stanford University, U.S.A.

1. Summary
This note describes the use of evolutionary design methods in aeronautics, highlighting the types of problems for which these methods are best suited and detailing some algorithms that have been useful in this domain. Three sample problems exemplify the utility of these methods and suggest numerous future applications for evolution-based design.

2. Introduction
Optimization is of great importance in aeronautical design. Since small changes in geometry produce large changes in transonic flow characteristics and reductions in structural weight may lead to significant range increases, for example, it is critical to find a design that is very close to best a design that roughly resembles a good design may be a bad design. Therefore, optimization in some form is ubiquitous in aerospace: from aerodynamic shape optimization to structural design, optimal control, and trajectory optimization. These complex simulations are often time consuming, requiring the solution of nonlinear partial differential equations represented with millions of grid points or structural solutions with hundreds of thousands of degrees of freedom. In some cases, the governing equations may be manipulated to provide gradient information efficiently using the adjoint equations. So what makes gradient-free, population-based design methods useful for aeronautical applications? Genetic algorithms and evolutionary optimization are perhaps intriguing from their simplicity and easy parallelization, but are often much less efficient than gradient methods, even when gradients must be computed with finite differences. This is especially true for high dimensional design problems the type of problem often encountered in aerodynamic shape optimization, structural sizing, or trajectory optimization. The simplicity and novelty of techniques such as simulated annealing, genetic algorithms, and other evolutionary methods has sometimes led to their ill-considered application to aeronautical problems that could be solved more efficiently with conventional numerical optimization methods. Even for problems with multiple local minima, it can often be argued that a gradient descent method started (in parallel perhaps) from several points is a better approach.

VKI lecture series on Optimization Methods & Tools for Multicriteria/Multidisciplinary Design November 15-19, 2004

These notes deal with applications for which the use of evolution-based design methods are justified. Classes of interesting problems include : 1. Objective or constraint functions with multiple local minima either due to the inherent nature of the problem or due to numerical noise or stochastic elements in the problem formulation. 2. Problems that involve searching, not for a single optimal point, but rather a population of good designs. 3. Problems for which the parallel, adaptable evolutionary algorithm may be exploited to do more than just function minimization. Function Smoothness: Although many problems in a specific domain may be sufficiently smooth for gradient-based methods to perform efficiently, a variety of interesting, complex problems, especially those involving large-scale multidisciplinary design, are sufficiently discontinuous that such methods are not practical. This may be a result of coding practices that did not recognize the need for smooth function evaluations. (For example, the panel code A502 writes intermediate results to disk because of the limited core memory available when it was written. This affects the precision of the results and sometimes causes difficulty for optimization.) Function discontinuities may also be created by the discretization of the problem itself, as in a vortex lattice code that requires certain relationships between panel placements on nearby surfaces. Algorithms that compute supersonic wave drag or sonic boom using area-rule methods must create oblique cuts through the geometry and the resolution required to provide smooth finite difference gradients requires at least an order of magnitude more time than required for a very reasonable function value. Other cases are more subtle: some functions are inherently statistical in nature such as simulations that involve probabilistic gust models, or are truly complex dynamical systems for which time-extended simulations produce results that are very sensitive to initial conditions. Other cases may be continuous in value but still cause problems with some optimizers as was recently encountered with a design problem that included a standard atmosphere model (which has discontinuous derivatives at the start of the isothermal region).

VKI lecture series on Optimization Methods & Tools for Multicriteria/Multidisciplinary Design November 15-19, 2004

Figure 1. Small variations in fuselage diameter produce large effects on sonic boom metric, creating problems for conventional optimization. Multiobjective design: When a scalar objective and well-defined constraints are known, conventional, single-objective optimization should be employed to find solutions. But often this is not the case. Especially in the formulative stages of problem definition, one frequently is interested in establishing optimal sensitivities among constraints and objective. This may be done using post optimality gradient information [1], but if the trade-offs are needed over a large range of parameters and the system is nonlinear, this is of limited use. As shown in the subsequent examples, many problems in aeronautical design involve multiple trade-offs that are not naturally posed as a single objective and multiple constraints. Evolutionary methods excel at such problems because the population used to sample information can also represent much of the design space of interest. And in some cases the fact that one is interested in a range of solutions prevents premature stagnation of an evolving population. One may also exploit the efficient multiobjective handling of evolutionary algorithms to improve the performance of multifidelity or multidisciplinary design processes by introducing analysis fidelity or interdisciplinary compatibility as additional objectives. Evolution of complexity: The flexibility of design encoding in an evolutionary algorithm also permits much more general descriptions of designs to be used in the design process. In addition to discrete variables, it is possible to allow the complexity of the design description to evolve with the design itself. Starting with a simple description of a design, an evolutionary method that supports variable complexity modeling can slowly add detail and arrive at designs that are conceptually distinct from those envisioned at the start of the optimization process. This concept has been studied by Gage [2] in some topological aerodynamic and structural design problems, including the discovery of novel wing tip geometries described in later sections of these notes. 3

VKI lecture series on Optimization Methods & Tools for Multicriteria/Multidisciplinary Design November 15-19, 2004

3. Evolutionary Design Methods


A large number of evolutionary algorithms have been proposed for use in a wide range of problem domains. From evolutionary programming [3] to evolution strategies [4], genetic algorithms [5], and genetic programming [6], population-based design and optimization methods have become popular for applications ranging from classifiers to engineering design. The present notes do not attempt to describe the basic operations of such systems (see [7]) and are not intended as a tutorial on evolutionary methods, but rather highlight some interesting applications of these methods in aeronautical design and suggest features of some of these methods that may be important in this domain. 3.1 Fundamentals Many evolutionary design approaches involve the same fundamental elements, starting with a concept for encoding the design information. Although binary and floating point descriptions are commonly employed, the details of the chosen encoding (whether binary or floating point) are of great, and often underappreciated, significance to the performance of the method. Indeed, it has been noted that once the design has been parameterized, much of the important design work is done and that the search for an optimum in this space is of lesser interest. Binary encoding allows design with specified precision (determined by the number of bits in the genotype string associated with a particular design variable) and permits optimization with discrete choices, which could include parameters such as number of engines or seats, or could allow only finite changes in wing span, perhaps aligned with panel width in a potential flow analysis. The cross-over operation permits replication of certain advantageous aspects of the current generation in subsequent generations. The most common single point cross-over operation between two binary encoded designs leads to offspring that differ from the parents in one coordinate direction (design variable) and can differ from the parents by no more than a factor of 2. This type of parallel search that requires improvement with changes in a single design variable is the simplest, but perhaps not the most efficient search heuristic. One might like to move in a direction that more likely leads to an improved design and this may require changing more than one variable at a time. This requires multi-point crossover, and has been investigated by several researchers [8]. Mutation is also a common evolutionary operator, playing the major role in design perturbation for some methods such as evolutionary programming (EP), while playing a secondary role in others, reducing the likelihood of stagnation in genetic algorithms (GAs) and developing adaptively in some evolution strategy (ES) approaches (e.g. [9]). The relative importance of these two operators in various evolutionary algorithms is discussed in some detail in [10].

VKI lecture series on Optimization Methods & Tools for Multicriteria/Multidisciplinary Design November 15-19, 2004

Similarly, many different approaches to selection have been studied, with perhaps the most important distinction related to whether the algorithm preserves an optimum that has been found (elitist strategy). Again, the number of different algorithms that may be formed even with common types of encoding, recombination, mutation, and selection strategies is large and while some algorithms are more commonly used, little consensus exists that a particular approach is generally to be preferred. Indeed the No Free Lunch theorems of Wolpert and Macready [11] suggest that the performance of optimization algorithms is tied closely to the problem domain. Thus, rather than evaluating the plethora of potential approaches, we focus on some specific implementations that have interesting properties or have been used effectively in aeronautical design problems. 3.2 A Simple Real-Encoded Genetic Algorithm One of these algorithms (termed PCGA) that the author has found to be very effective is closely associated with evolution strategies. The method involves a real encoding of the design variables and a recombination algorithm that simulates multi-point cross-over. It also shares some of the features of the Nelder-Meade nonlinear simplex algorithm, a gradient-free optimization method that is perhaps the most widely used optimization method (e.g. fminsearch in MatLab). In PCGA, two offspring are created as linear combinations of the parents, effectively interpolating between the two previous designs or extrapolating by a bounded distance. The best child replaces the worst parent, preserving the best design in the population and requiring only local comparisons, which can be easily parallelized. The system mimics some of the behavior of the Nelder-Meade algorithm, interpolating or extrapolating, then replacing the worst design by an improved result. Because the recombination algorithm is sensitive to the distance between parents, tight convergence is possible, although the space-spanning requirements of the simplex or other parallel direction search methods [12] requires a minimum population size, related to the dimension of the search space or explicit enforcement of linear independence in the population. Figure 3 shows the performance of this simple algorithm on a 12dimensional algebraic test problem with different population sizes; convergence is achieved after approximately 50-100 generations. A gradient-based search method might be expected to require an order of magnitude fewer function evaluations, but as illustrated in subsequent examples, this population-based method exhibits several features of use in aeronautical design problems. 3.3 Multiobjective Design The previous algorithm is easily modified to maintain a population of designs selected on the basis of multiple design criteria. In this case the metric used to evaluate an individuals relative fitness is a ranking based on dominance [13]. The rank, ri, of the ith design is given by: ri =1+ni where ni is the number of other designs in the population for which all objective values exceed those of design i. At each iteration, the non-dominated solutions are therefore assigned a rank of 1, and form the Pareto set of that generation. Selection is based on the overall ranking of the individuals. Through the generations, the algorithm drives the population toward better solutions, eventually leading to a Pareto set that contains the optimal solutions as illustrated in figure 2. 5

VKI lecture series on Optimization Methods & Tools for Multicriteria/Multidisciplinary Design November 15-19, 2004

In order to maintain diversity among non-dominated solutions, niching is used. The concept is to enforce a minimum distance between solutions of same rank to avoid clustering around local extrema and ensure that the solution population fills the available design space. A normalized distance between two solutions, i and j, of same rank, dij, is used as metric for this purpose. It is defined as: dij2 = ! (fik - fjk )2 / (fmax k - fmin k)2 where fmax k and fmin k are the maximum and minimum objective function values of the kth objective, respectively. The niche count is calculated by summing sharing function values across all solutions of rank ri: nci = ! S(dij ), with S(dij) = 1 - dij ", if dij ! "; and S(dij) = 0, otherwise. The parameter " is set by the user (typical values are 0.2-0.6). Finally, the shared fitness F of solution i is computed and substituted for raw fitness, F: Fi = Fi nci Multiobjective genetic algorithms generally lack a true convergence criteria. The Pareto set indicates the best solutions of a given generation and the population average rank is commonly used to gauge progress towards the optimal population an average population rank of 1 indicates that all population members are non-dominated and lie on the Pareto front.

Figure 2. Simple 2 variable, 2 objective test problem. The solid line is the Pareto front, generated using conventional optimization with a variable weighting between objectives. The points show the multiobjective GA population after the 1st and 10th generations. Note that the multiobjective front requires about the same number of generations to converge as a single objective GA.

VKI lecture series on Optimization Methods & Tools for Multicriteria/Multidisciplinary Design November 15-19, 2004

Figure 3. Convergence history for a 12-variable problem using a single-objective GA with various population sizes (left) and Pareto front for a 12-variable, 2-objective problem after 40 generations (right). 3.4. Constraints Constraints are an integral part of many engineering problems. In most genetic algorithms constraints are accommodated by one of the following techniques: rejecting population members that violate one or more constraints; penalizing infeasible designs by decreasing their fitness, therefore increasing rank; or introducing a modification (repair) of the design that enforces feasibility. Although the first approach ensures that the population remains feasible, it may be difficult to generate a reasonable initial population of feasible designs. Also by rejecting many infeasible designs, especially in early generations, it is easy to lose genetic diversity that may be required to achieve the best designs subsequently. The third approach may be most efficient, but is often difficult to implement, and so penalty methods are commonly used. The lack of dependence on gradient and Hessian information means that the penalty function need not be smooth or even continuous and L1 norms are commonly used (in contrast to L2 measures of constraint violation in many gradient-based methods). This avoids the problem of slightly infeasible solutions with quadratic exterior penalties and requires only that the penalty weight exceed the value of the Lagrange multiplier that would be computed by a constrained optimization method. Improved performance has also been noted in [2] with the use of an extended encoding of the design variables whose expression is governed by constraint activity. This allows designs to adapt to changes in the active set of constraints that may occur during the evolutionary process or in response to a dynamic environment. 3.5. Unconventional Uses of Evolutionary Methods Although evolutionary algorithms are rarely used when the underlying function evaluations are expensive, this is not always true. The increasing popularity of surrogate methods for optimization of complex functions, may lead to a role for EAs even in this domain. Response surfaces are sometimes constructed over a certain domain using

VKI lecture series on Optimization Methods & Tools for Multicriteria/Multidisciplinary Design November 15-19, 2004

sampling techniques based on results from design of experiments. If the response surface models of objectives or constraints are very simple (e.g. quadratic models), many different optimization algorithms are suitable to find the optimum in the model domain. The difficulty with model updating and restriction of very simple fits to small regions of the design space has led to the use of models based on Kriging, neural networks, and other interpolatory methods that may pose difficulties for some optimization methods (e.g. introducing discontinuities or local minima). In this case evolutionary design methods work well with the surrogate models, finding optimal points (or even fronts) in a robust manner, and since evaluation of the surrogate model is very fast, the penalties associated with large numbers of (now cheap) function evaluations may be insignificant (see [14]). Similarly, GAs have been combined with neural networks, to find weights that lead to desired behavior even when direct training is not possible. Fogels use of this technique in his intriguing Checkers player [15] and the success of the method in coordinated Mars rover search strategies [16], suggest the range of applications for this approach. Similarly, GAs may be used with multi-fidelity simulation methods, with a (sub)population of higher fidelity analyses used to calibrate or correct a lower fidelity simulation that provides fitness values for an evolutionary design method. Multiobjective evolutionary design in which simulation fidelity constitutes one of the objectives may also provide a direct role for evolutionary methods in multi-fidelity optimization. Complex multidisciplinary optimization problems are sometimes decomposed by replacing interdisciplinary coupling with an extra design parameter and an associated compatibility constraint [17]. This increases the dimensionality of the optimization problem, but simplifies the problem in other ways that may be desirable. Gradient-based, constrained optimization is generally used to efficiently incorporate the new design variable and constraint into the system, but this is less easily accomplished when an evolutionary method is used for design. However, by incorporating compatibility as an explicit part of a multiobjective evolutionary design procedure and including computed values of interdisciplinary coupling parameters directly in the population, even highdimensionality coupling can be accommodated efficiently in a population based design strategy. The final unconventional application of evolutionary design involves the use of these methods for searching through a design space of variable dimension. This concept is suggested by the success of natural evolution in producing very efficient complex design solutions. Indeed natural evolution is quite different from conventional numerical optimization, primarily in that the parameterization, and the complexity, of the problem evolves with the solution. Consider the evolution of the human eye as a parametric optimization problem. To completely describe the eyes components, from lens to retinal cone cells with a parameterization that would lend itself to conventional optimization is not possible. But even if it were, and our initial design resembled a microscopic hydra, it is inconceivable that the extremely high dimensional design space could be searched in such a way that light sensitivity in hydra cells would lead to a fully functional optical system through the usual kind of hill-climbing function maximization. Instead, evolution 8

VKI lecture series on Optimization Methods & Tools for Multicriteria/Multidisciplinary Design November 15-19, 2004

appears to involve the evolution of design complexity rather than just the search for the best values of an a priori parameterization. Although it may be possible to construct a more conventional optimization method that admits a variable complexity design description, this is readily achieved with population-based evolutionary methods. By creating a flexible design language and creating a cross-over algorithm that can accommodate variable string length representations of the design, an evolutionary design method can create designs with an adaptive parameterization that is not only capable of evolving complexity from simple designs, but leads to optimal solutions much more efficiently than conventional methods. This is demonstrated in the final examples of the next section.

4. Applications
This section highlights some recent applications of evolutionary methods in aeronautical design to illustrate the growing utility of such methods with increasingly available clusters of computers. 4.1. Supersonic Design for Low Boom and Efficient Cruise Supersonic aircraft are restricted from overland operation in much of the world because of the noise created by shockwaves that propagate to the ground even from altitudes in excess of 50,000 ft (15 km). Several recent design studies have explored the possibility of small supersonic aircraft that might be capable of efficient flight with shaped signatures that produce much lower noise levels on the ground. By tailoring the distribution of lift and aircraft volume properly, the conventional N-shaped boom signature may be transformed into a signature with several smaller shocks or even a smoother ramp type signature with much lower noise. Design techniques for achieving these modified boom signatures were developed in the 1970s and 1980s and employed linear, axisymmetric aerodynamic theory to solve the inverse problem of sonic boom propagation [18]. One could specify a desired ramp or flat-top boom signature and compute the distribution of lift and volume required to achieve this. Unfortunately, this distribution was not necessarily a low drag distribution and resulting configurations pay a performance penalty to achieve the low boom signature. An alternative approach using numerical optimization has been developed recently [19 21] to permit optimal tradeoffs between boom and drag, and to allow the use of higher fidelity aerodynamic models in this process. Evolutionary design methods are used for three reasons: 1. When boom loudness is measured with typical noise metrics (such as dBA or PNL) the response is not smooth (as shock coalescence during propagation introduces large sensitivities in the results see figure 1). 2. Since the required boom loudness has not yet been established, we are interested in the trade space in which low drag and low boom are desirable features. Multiobjective GAs provide a convenient method of creating this family of designs. 9

VKI lecture series on Optimization Methods & Tools for Multicriteria/Multidisciplinary Design November 15-19, 2004

3. The parallel nature of evolutionary methods and modern cluster-based computing may be exploited to improve computational performance. To accomplish this improved-fidelity multi-objective design, a parametric description of a small supersonic aircraft (70,000 lbs at Mach 1.5) was developed that could generate a complete aircraft geometry from a set of 10 to 30 parameters. From this geometric description an automated CFD meshing tool generated the flow field grid or panel discretization required to obtain a solution from an Euler solver or surface panel code (figures 4, 5). These solutions were combined with an acoustic propagation code to obtain the acoustic signature of the aircraft on the ground.

Figure 4. Local pressure distribution from Euler-based flow solver on a supersonic business jet described by a simple parametric model.

Figure 5: Surface panel representation (A502) of a generic supersonic business jet with nacelles modeled as equivalent bodies.

10

VKI lecture series on Optimization Methods & Tools for Multicriteria/Multidisciplinary Design November 15-19, 2004

The simplest description of the design involved the following 11 design variables: x1x6 : fuselage diameter at 6 stations along the length x7 : horizontal tail x-location x8 : horizontal tail incidence x9 : aircraft angle of attack x10 : wing twist at break point x11 : wing twist at tip Lower bounds for x3 and x4 were set by cabin size requirements, while the wing planform was fixed based on previous design studies that suggested that the optimal planform depended strongly on structural properties and low speed characteristics. These considerations were added in a later study [22], but many interesting results are apparent in the simpler optimization study. The objective was to minimize drag and sonic boom intensity: Minimize Subject to: J1 = !Pmax - !Pmin J2 = Drag Lift = Weight xmin < x < xmax
!Pmax !P

!Pmin

Figure 6: Simple sonic boom criteria. Although the results shown here use ground signature overpressure as some measure of boom strength, other objectives, such as the sonic boom dBA loudness, were also used and were part of the motivation for using evolutionary design methods in this work (see figure 1). Figure 7 shows the initial random population of 204 designs and the population after 40 generations in terms of inviscid lift to drag ratio computed by the surface panel code and boom pressure difference. After 40 generations, the Pareto front is already well defined, providing valuable insight on how the drag varies with sonic boom. Figure 8 shows the resulting Pareto optimal solutions for the fuselage radius and the boom signatures corresponding to the lowest drag and lowest boom members of the population. The optimal fuselage shape for sonic boom corresponds approximately to the upper bound of the Pareto set, illustrating how the low-boom and low-drag designs differ

11

VKI lecture series on Optimization Methods & Tools for Multicriteria/Multidisciplinary Design November 15-19, 2004

in fuselage nose shape, with the low-boom design incorporating a blunt nose. The wing tip twists are also quite different, with greater induced drag for the low-boom design. For these reasons, the lowest-boom design incurred a large drag penalty to achieve a sonic boom overpressure of 65% of the lowest-drag design.

Figure 7. Sonic boom magnitudes (pressure differences) and inviscid lift-to-drag ratio for initial population (left) and population after 40 generations (right).

Figure 8. Optimal designs from Pareto front after 40 generations. Fuselage radius shows nose blunting for reduced boom (left). Boom signature for lowest drag and lowest boom designs in the population. A related study [23] utilized a similar design parameterization and objectives, but replaced the surface panel aerodynamic analysis with an Euler solver. The increased computational time required for the flow solution made direct use of the Euler results

12

VKI lecture series on Optimization Methods & Tools for Multicriteria/Multidisciplinary Design November 15-19, 2004

impractical with the available computer resources. Instead, the multiobjective evolutionary design was based on results from a Kriging representation of the design space, which was iteratively refined. Figure 9 summarizes some of the results of this work, including planform variations that emphasize the importance of additional disciplines, such as structures and controls, in the complete vehicle design process. A more complete supersonic vehicle design study was subsequently undertaken with results presented in [22]. The resulting configuration is represented in figure 4.

Figure 9. Related multiobjective design based on Kriging fit of Euler solutions. 4.2. Aircraft Design for Improved Environmental Performance A second recent example of the use of evolutionary design methods is described in [24]. This research explored the feasibility of including environmental performance as an optimization objective in aircraft conceptual design, allowing a quantitative analysis of the trade-offs between environmental performance and operating cost. A program for aircraft design and optimization was developed, using a multiobjective genetic algorithm to determine optimal aircraft configurations and to estimate the sensitivities between the

13

VKI lecture series on Optimization Methods & Tools for Multicriteria/Multidisciplinary Design November 15-19, 2004

conflicting objectives of low noise, low emissions, and operating costs. The design tool was based on a new application integration framework incorporating a detailed noise prediction code, engine simulator, and aircraft analysis and optimization modules. Results from this study illustrate environmental performance trades and address the feasibility of very low-noise and low-emissions designs that could dramatically decrease the environmental impact of commercial aviation. A library of integrated analysis routines were used to compute many aspects of aircraft design and performance. NASA Langleys Aircraft Noise Prediction Program (ANOPP [25]) was used for jet, fan, and airframe noise modeling while engine performance was estimated using NASA Glenns Engine Performance code (NEPP [26]). The design framework integrates these programs with other disciplinary analyses ranging from component weights to stability and control and mission performance and with evolutionary design methods. The approximate methods are particularly well-suited for evolutionary optimization due to their rapid execution and robustness. In addition to traditional performance constraints such as range and field performance, maximum cumulative certification noise (the sum of the noise at each certification location) was included as an objective. This approach allowed an increase in environmental acceptability to be explicitly specified: from slight improvements in noise to silent aircraft. NOx emissions were constrained through the emission index of the engine and CO2 emissions were modeled based on fuel burned. Design variables included parameters from the aircraft configuration, propulsion system, and flight profile. The design framework permits custom optimization methods to be incorporated directly and a multiobjective optimizer was developed for use in this application. Although gradient-based nonlinear optimization methods provide efficient design tools, they require smooth objective and constraint functions. Especially when one is employing file-based legacy codes that may not have been written with this requirement in mind, discontinuous and noisy function values may make the use of gradient methods difficult. This is the case for the design problems considered here. Figure 10 shows how the predicted noise (from the NASA ANOPP code) changes with engine bypass ratio. Note that the scale is expanded greatly to illustrate the small numerical noise, but the large local gradients call for unconventional design tools. In this research the multiobjective, floating point genetic algorithm, described previously, was used to allow tradeoffs between competing objectives. As in the previous example, the methods used a penalty function approach to constraints, dominance-based ranking, and solution niching.

14

VKI lecture series on Optimization Methods & Tools for Multicriteria/Multidisciplinary Design November 15-19, 2004

Figure 10. Sensitivity of aircraft noise to small changes in bypass ratio. The design tool allows trade-offs (using the multiobjective genetic algorithm) and case studies to be quickly completed in order to visualize the interrelationships between noise, emissions, and other aircraft performance metrics. A selection of example results is included below. To understand the penalty associated with such design trades, an operating cost metric was minimized together with trip CO2 emissions and cumulative certification noise for a 280-passenger, twin-engine airliner with a 6,000 nm range, and takeoff, cruise, and landing performances in line with industry standards for similarlysized aircraft. The 15 design variables included aircraft geometry, engine parameters, and operational variables. The process of obtaining a low-rank Pareto front can be significantly accelerated by first computing the extreme points of the fronts. This was done by running a single-objective version of the genetic algorithm. These optimal designs were subsequently inserted into the initial population of the multiobjective problems. The resulting Pareto fronts of fuel carried, NOx emissions, and cumulative noise margin vs. cost are shown in Figure 11. The configuration leading to minimum operating cost (Design A) was computed first by running the design tool without specifying any noise or emissions constraints. This aircraft was considered the baseline and is representative of existing aircraft. Reflecting the impact of block time on the cost function, the cruise Mach number is higher than would be required for minimal fuel burn (Design B). Fuel plays an important role in the cost calculation, as illustrated by the similarities in the designs for minimum cost and minimum fuel carried (and therefore, minimum cruise CO2, SO2, and H2O). This tight coupling is also reflected in the relatively small fuel-cost trade space (notice the fuel-cost Pareto front is narrow). Both designs attain high fuel efficiency via large pressure ratios and high turbine inlet temperatures. 15

VKI lecture series on Optimization Methods & Tools for Multicriteria/Multidisciplinary Design November 15-19, 2004

Optimizing the aircraft for lowest fuel yields a 10% decrease in fuel carried (propagating through the design to yield a 5.4% decrease in maximum takeoff weight) but with a cost increase of 2% relative to low-cost Design A. Since the generation of NOx emissions is a strong function of combustor exhaust temperature and compression ratio, Design C compromises fuel efficiency for low NOx emissions by reducing the engine overall pressure ratio and combustor temperature. The resulting 12% reduction in sea-level static thrust relative to the low-cost design mandates that the aircraft fly slower -- Mach 0.67, versus Mach 0.84 for the baseline -- and at lower altitudes (the initial cruise altitude is reduced from approximately 33,000 ft to 28,000 ft). The result is a 53.5% drop in LTO NOx emissions for an 9% and 13% increase in operating cost and fuel consumption, respectively. These divergent requirements for low NOx and low fuel designs are well illustrated by the wide, and very smooth, Pareto front. As a result of the lower cruise Mach number, wing sweep is reduced from 34 to 11 degrees. With significantly reduced available thrust, the wing taper ratio is increased from 0.10 to 0.39 to increase the maximum lift coefficient during for takeoff. For similar reasons, the wing area is enlarged by 12%, contributing to an increase in maximum takeoff weight of 9%. Combined with lower available thrust, the climb performance of the low-NOx design is significantly deteriorated: thrust-toweight ratio drops to 0.296, resulting in the increased cumulative certification noise -over 3.5 dB louder than the baseline Design A.

Figure 11. Pareto fronts of fuel, NOx, and noise vs. relative operating cost.

16

VKI lecture series on Optimization Methods & Tools for Multicriteria/Multidisciplinary Design November 15-19, 2004

Although much of the progress on noise and emissions reduction has been achieved through improvements to aircraft engines, the design tool was also applied to investigate three examples of potential future airframe technologies: laminar flow, advanced structures, and induced drag reduction. In particular the following three assumptions were explored: Increased laminar flow: for wing sweeps less than 20 degrees it is assumed that natural laminar flow is achieved over an average of 60% chord. Advanced structures a factor of 0.8 is applied to the aircraft structural weight to reflect use of new materials or other innovative structural concepts. A 10% reduction in induced drag enabled, perhaps with span extensions and active load alleviation, for example. These technologies, meant to improve aerodynamic efficiency or reduce structural weight, are of interest because they also have significant impact on the environmental performance of the aircraft. Figure 12 illustrates the changes to the aircraft fuel-NOx performance. The advanced technology lowest-NOx aircraft (Design K) produces 34% less NOx per LTO cycle than the conventional low-NOx aircraft (Design C), at 9% lower operating cost. Similarly, the advanced low-fuel candidate (Design J) requires 27% less fuel (and therefore produces 27% fewer fuel-proportional emissions) to complete the mission while releasing 32% fewer NOx emissions than the conventional lowest-fuel aircraft (Design B). Taking advantage of the drag benefits associated with increased laminar flow, the two advanced designs feature wing sweep under 20 degrees while retaining the original cruise Mach number. This is possible with a thinner wing, which is permitted by the reduced fuel capacity requirement. Incorporating these advanced technologies negates some of the adverse effects of optimizing the aircraft for low-noise or low-emissions. Indeed, the advanced low-NOx aircraft (Design K) generates 30% of the NOx emissions of the conventional low-cost design (Design A) at the same operating cost. Because of the reduced structural weight and fuel load (and therefore lower maximum takeoff weight), the two advanced designs require approximately 30% less installed thrust, resulting in a cumulative noise margin of 5.5 EPNdB for Design J and 0.8 EPNdB for Design K.

17

VKI lecture series on Optimization Methods & Tools for Multicriteria/Multidisciplinary Design November 15-19, 2004

Figure 12. Trade-offs between NOx and CO2 (fuel consumption) for conventional aircraft and potential future designs with advanced airframe technologies. 4.3. Variable Complexity Topological Design The final example of evolutionary design applications in aeronautics involves the use of a variable-complexity genetic algorithm for wing design. At the moment, most high fidelity optimization is used to refine a configuration that has been defined using simpler but more comprehensive analyses or design intuition, but as computing power increases, these tools will be used more at the conceptual design phase, leading to the possibility of innovative conceptual solutions that appear from the optimization automatically. This type of emergent innovation is suggested in some recent examples that illustrate how unexpected fundamental concepts can arise from robust configuration optimization. Figure 13 (from [2]) shows how an evolutionary optimization algorithm was used to find the wing geometry that produced minimum total drag, yet fit inside a box of fixed height and span.

18

VKI lecture series on Optimization Methods & Tools for Multicriteria/Multidisciplinary Design November 15-19, 2004

Figure 13. Evolution of C-Wing design from variable complexity nonplanar wing model. Left image shows the half-wing geometries represented in the population. Right image is the design with lowest drag in the given generation. In this case, a variable length encoding was used, describing a wing as a collection of linearly tapered and twisted elements, whose aerodynamic characteristics were computed using a vortex lattice analysis. A population of initially simple designs is shown at the top of figure 13. The optimizer quickly discovers that span reduces drag, and after only 5-6 generations (with population size 500), has found the minimum induced drag for a roughly planar wing. As the span constraint becomes active, the optimizer discovers winglets, adding vertical elements at the wing tips to further reduce vortex drag. Finally, after about 100 generations, the system finds an advantage to adding horizontal tip extensions to the

19

VKI lecture series on Optimization Methods & Tools for Multicriteria/Multidisciplinary Design November 15-19, 2004

winglets, forming a C shape at the tip. This design was investigated further and found to exhibit useful structural and control features in addition to the reduced vortex drag at fixed span. The concept was subsequently patented [27] and is being studied for application to new aircraft concepts at Boeing, NASA and elsewhere (figure 14).

Figure 14. Configurations for a large aircraft and blended-wing-body based on the C-wing concept.

5. Conclusions and Directions for Future Work


Evolutionary methods for aeronautical design are of increasing utility as parallel computation becomes more common and as the complexity of simulated systems increases. The ability of evolutionary methods to handle noisy or probabilistic objectives as well as their natural extension to efficient multicriteria design makes these techniques valuable tools for the design of complex aerospace systems at least in cases with a moderate number of design variables (typically 50 or fewer). The examples shown here illustrate that many interesting design results, that would have been difficult to produce using conventional optimization methods, arise from the application of evolutionary methods. Future development of these methods to improve efficiency is especially important for higher dimensionality, higher fidelity design. However, even minor modifications of current algorithms may provide powerful tools for future problems involving uncertainty, multidisciplinary decomposition, and multifidelity or variable complexity modeling.

6. Acknowledgments
Much of the work summarized here has been accomplished by doctoral students in the Aircraft Aerodynamics and Design Group at Stanford (in particular Drs. Nicolas Antoine, Martin Chan, and Peter Gage). Interested readers should consult their papers and theses listed in the references section for further details. The author would also like to gratefully acknowledge support from the Boeing Company and NASAs Ames, Glenn, and Langley Research Centers.

20

VKI lecture series on Optimization Methods & Tools for Multicriteria/Multidisciplinary Design November 15-19, 2004

7. References
[1] Braun, R.D., Kroo, I., Gage, P., Post-Optimality Analysis in Aerospace Vehicle Design, AIAA-93-3932, Aug. 1993. [2] Gage, P.J., New Approaches to Optimization in Aerospace Conceptual Design, Ph.D. thesis, Stanford University, 1995. [3] Fogel, L.J., Owens, A.J., Walsh, M.J., Artificial Intelligence through Simulated Evolution, New York: Wiley, 1966. [4] Rechenberg, I., Evolutionstrategie: Optimierung technischer Systeme nach Prinzipien der biologischen Evolution, Stuttgart: Fromman-Holzboog, 1973. [5] Holland, J.H., Adaptation in Natural and Artificial systems, Ann Arbor, MI: The University of Michigan Press. 2nd edn. (1992) Boston, MA: MIT Press, 1975. [6] Koza, J.R., Genetic Programming: On the Programming of Computers by Means of Natural Selection, Cambridge, MA: MIT Press, 1992. [7] Goldberg, D.E., Genetic Algorithms in Search, Optimization, and Machine Learning, Addison Wesley, 1989. [8] De Jong, K.A., Spears, W.M., A Formal Analysis of the Role of Multi-Point Crossover in Genetic Algorithms, Annals of Mathematics and Artificial Intelligence, 1992. [9] Schwefel, H.P. Numerical Optimization of Computer Models, John Wiley and Sons, NY, 1981. [10] Spears, W.M., Evolutionary Algorithms: The Role of Mutation and Recombination, Springer, 2000. [11] Wolpert, D.H., Macready, W.G., "No Free Lunch Theorems for Search", IEEE Transactions on Evolutionary Computation, 1, 1997. [12] P. D. Hough, T. G. Kolda, V. J. Torczon. Asynchronous Parallel Pattern Search for Nonlinear Optimization. SIAM J. Scientific Computing, 23(1):134-156, June 2001. [13] Deb, K., Multiobjective Optimization Using Evolutionary Algorithms, Wiley, 2001, pp. 190-196. [14] Chung, H. S., Alonso, J.J., "Multiobjective Optimization Using Gradient Enhanced Genetic Algorithms", 10th AIAA/ISSMO Multidisciplinary Analysis and Optimization Conference, AIAA Paper 2004-4325, Albany, New York, August 30 - September 1, 2004. 21

VKI lecture series on Optimization Methods & Tools for Multicriteria/Multidisciplinary Design November 15-19, 2004

[15] Fogel, D. B., Evolving a checkers player without relying on human experience, intelligence, v.11 n.2, p.20-27, Summer 2000. [16] Agogino, A., Tumer, K., Efficient Evaluation Functions for Multi-Rover Systems, Genetic and Evolutionary Computation Conference, Seattle, WA, June 2004. [17] Cramer, E.J., Dennis, J.E. Jr., Frank, P.D., Lewis, R.M., Shubin, G.R., Problem formulation for multidisciplinary optimization, Technical Report 93-34, Department of Computational and Applied Mathematics, Rice University, Houston, TX 77251, 1993. [18] George, A. R., Seebass, R., Sonic Boom Minimization Including Both Front and Rear Shocks, AIAA Journal, Vol. 9, No. 10, Oct. 1971, pp. 2091-2093. [19] Makino, Y., Aoyama, T., Iwamiya, T.: Numerical Optimization of Fuselage Geometry to Modify Sonic-Boom Signature, Journal of Aircraft, Vol. 36, No. 4, Jul-Aug 1999, pp. 668-674. [20] Alonso, J.J., Kroo, I.M., Jameson, A., Advanced Algorithms for Design and Optimization of Quiet Supersonic Platform, 40th AIAA Aerospace Sciences Meeting & Exhibit, AIAA Paper 2002-0144, Reno, NV, January 2002. [21] Chan, M., Supersonic Aircraft Optimization for Minimizing Drag and Sonic Boom, Stanford University Ph.D. thesis, August 2003. [22] Choi, S., Alonso, J.J., Kroo, I., Wintzer, M., "Multi-fidelity Design Optimization of Low-boom Supersonic Business Jet", 10th AIAA/ISSMO Multidisciplinary Analysis and Optimization Conference, AIAA Paper 2004-4371, Albany, New York, August 30 September 1, 2004. [23] Chung, H., Choi, S., Alonso, J., "Supersonic Business Jet Design Using KnowledgeBased Genetic Algorithm with Adaptive, Unstructured Grid Methodology", 21st AIAA Applied Aerodynamics Conference, AIAA Paper AIAA-2003-3791, Orlando, FL, June 23-26, 2003 [24] Antoine, N.E., Aircraft Optimization for Minimal Environmental Impact, Ph.D. thesis, Stanford University, August 2004. [25] Gillian, R.E., Aircraft Noise Prediction Program Users Manual, NASA Langley Research Center, 1982. [26] Klann, J.L., Snyder, C.A., NEPP Users Manual, NASA Glenn Research Center, March 1997. [27] McMasters, J.H., Kroo, I.M., Advanced Configurations for Very Large Transport Airplanes, Aircraft Design, Vol. 1, No. 4, 1998, pp. 217-42 (see also NASA CR 198351, Oct. 1996. 22

You might also like