25-09-2012, 11:54 AM
Evaluation of Multidisciplinary Optimization (MDO) Techniques Applied to a Reusable Launch Vehicle
multi discipline optimisation.pdf (Size: 1.3 MB / Downloads: 87)
Abstract
Optimization of complex engineering systems has always been an integral part of design. Due to the size and complexity of aerospace systems the design of a whole system is broken down into multiple disciplines. Traditionally these disciplines have developed local design tools and computer codes (legacy codes) allowing them to perform optimization with respect to some aspect of their local discipline. Unfortunately, this approach can produce sub-optimal systems as the disciplines are not optimizing with respect to a consistent global objective. Multidisciplinary design optimization (MDO) techniques have been developed to allow for multidisciplinary systems to reach a global optimum. The industry accepted All-at-Once (AAO) technique has practical limitations and is confined to only small, conceptual level problems.
New multi-level MDO techniques have been proposed which may allow for the global optimization of the large, complex systems involved in higher levels of design. Three of the most promising multi-level MDO techniques, Bi-Level Integrated System Synthesis (BLISS), Collaborative Optimization (CO) and Modified Collaborative Optimization (MCO) are applied, evaluated and compared in this study.
Introduction
Optimization of complex engineering systems has always been an integral part of design. Man has never created anything which he then didn’t wonder how he could make better. This is true in the aerospace industry dating back to wind tunnel studies conducted by the Wright bothers to study wing shapes. Originally those that created aerospace vehicles were responsible for every aspect from wing shape to propulsion. As the size and complexity of aerospace systems grew, though, it became apparent that the design of such enormously complex problems would have to be broken down into disciplines with groups concentrating only on their own part of the whole.
While breaking apart the overall problem into different contributing analyses (CA’s) made it humanly possible to design inhumanly complex systems, the ability for all designers to see how their specific changes would affect the overall goodness of the whole was lost. Communicating between disciplines became increasingly difficult and thus each discipline developed metrics to which to optimize their own individual part of the total system or vehicle. Unfortunately, no discipline is an island as every discipline will affect another and vice versa. If they don’t agree as to what the coupling variable values should be across disciplines then the system is not converged or valid.
Benefits of MDO
The first goal is to understand the overall benefit of using an MDO method to optimize the RLV test problem versus a few trials converged using a fixed point iteration process (FPI). For this purpose several FPI models of the next generation RLV were created in order to offer some insight as to what is the best vehicle that could be designed without using MDO. The results achieved using FPI will be compared to the techniques mentioned in order to better understand the benefits derived from MDO.
FPI is not an MDO method it is merely a way to converge a multidisciplinary analysis (MDA) process. It does not perform any global optimization to find the optimum system configuration. In discipline a designer will try several configurations and converges each one using FPI. Then one of the configurations is deemed the best and selected for further study. Thus, FPI usually results in the best configuration for the options tested, but does guarantee that the true optimum will be found.
The argument for using FPI to test a limited number of configurations versus applying an MDO process are practical or “real world” in nature. FPI has been the method of choice in the aerospace industry. Thus there are already legacy tools and design practices developed by many individuals over time. Changing this structure would require a large initial investment. Also, the experts performing each analysis have been trained and are experienced in solving the problem as it is currently formulated for FPI
Authenticity of Test Problem
It is intended that the test problem, the optimization of a next generation RLV, have enough realism that it can add to a growing body of work attempting to evaluate some of the most novel and promising MDO algorithms: CO, MCO and BLISS.5,6,7,8 These MDO techniques were “crafted” as opposed to “rigorously derived.”9 They lack a general mathematical proof showing for which MDO problems they are suited. Thus, the algorithms require that they be validated via a statistically significant number of test cases of realistic, complex system applications.9
To ensure realism in the test problem, legacy tools were employed which are the same or very similar to those used in industry. Also, one of the tools selected, POST, while being the industry standard code is known to be very troublesome to work with. This will add the realism of the problem as in the “real world” one often has to work with the tools available and cannot just select tools, such as using all Microsoft spreadsheets, that are usually well behaved.
Comparison between CO, MCO & BLISS
There are some deficiencies in the currently accepted design methods used in industry, namely discipline optimization with FPI does not do system-level optimization and AAO cannot be applied to large, complex engineering problems. New multi-level MDO techniques, of which CO, MCO and BLISS are three of the most promising, are attempting to overcome some of these deficiencies. The work presented intends to some add insight as to which of the most novel techniques, CO, MCO or BLISS, showed the most promise when applied to the RLV test problem.
It is difficult to draw any conclusions between, CO, MCO and BLISS, using the current literature available.5,6,7,8,11 Test applications that have produced varying degrees of success, but it is difficult to determine if this is due to differences in the algorithms or other factors.
MDO Background
Aerospace system analysis and design is usually broken down into multiple disciplines due to their enormous complexity. This impedes MDO as the local-level disciplines lose knowledge of how their local design affects the system as a whole. AAO can be used to solve the MDO problem, but it takes away all design responsibilities from the discipline experts and AAO does not perform well in higher fidelity applications with a lot of design variables (See 8 AAO: All-at-Once pg 36). Multi-level MDO techniques attempt to correct this problem by allowing some form of local-level design optimization while adding a system-level optimizer guide the process to a globally optimized solution.
One of the greatest impediments in the acceptance of multi-level
MDO techniques is the fact that they have to date been “crafted” as opposed to “rigorously derived.”9 They lack a general mathematical proof showing for which MDO problems they are suitable. Thus, the algorithms require that they be validated via a statistically significant number of test cases of realistic, complex system applications.9 Program managers or designers are reluctant to implement the new techniques to a “real world”, industry application as they cannot be confident that the multi-level MDO algorithm will indeed arrive at the global optimum. This creates a Catch-22 scenario for multi-level MDO techniques as the only way to gain acceptance is to show success in a statistically significant number of “real world”, industry-sized test problems; paradoxically not many application attempts are made as program managers are not willing to risk their projects on an unproven method. This contradiction means that acceptance of any multi-level MDO technique may be a slow one.