White Papers
Alternate Sampling Methods for use with Multidisciplinary Design Optimization in a High Performance Computing Environment
Srinivas Kodiyalam and Jaroslaw Sobieszczanski-Sobieski
Abstract
Multidisciplinary Design Optimization (MDO) embodies a set of methodologies which provide a means of coordinating efforts and possibly conflicting recommendations of various disciplinary design teams with well-established analytical tools and expertise. MDO involves multiple disciplines, engineering, business and program management, often with multiple, competing objectives. These disciplines may just be analysis codes, which contain a body of physical principles, or, in addition, they may possess some intelligent decision-making capabilities. In an attempt to address the issues involved with the MDO process, formal methods have been derived, making use of consistent mathematical concepts, unique data structures, and alternative system representation techniques.
Simulation based detailed design of complex systems, more specifically, aerospace and automotive systems, is increasingly becoming a distributed design activity involving multiple decision teams each with very high fidelity models and analysis tools as well as heterogeneous computing environments.
Introduction
Another aspect of the simulation based design process is that it relies heavily on complex computer analysis codes and simulations (e.g., FEA and CFD) to improve the product design. These time consuming and expensive analyses are repeatedly invoked during optimization making the design exploration and multidisciplinary design optimization time very long, if not prohibitive. Two solutions are possible to make these problem solution times tractable:
(i) Use of Approximation Models (also referred to as Surrogate Models) for the design objectives and constraints in conjunction with the numerical search process. Since these approximate models are inexpensive to evaluate for a new set of data or values assigned to design variables, we can afford to evaluate approximate responses many more times without having to worry about the computational resources. Consequently, various different types of studies including design optimization using high fidelity analysis become possible. The purpose of these studies must be extraction of data that are directly useful for the design decisions.
(ii) Use of High Performance Computing (HPC) servers with a large number of processors to enable multiple levels of parallelism (coarse and fine grained parallelism) for higher throughput computing and faster solution turn around times.
In this work, both of these above solutions are investigated. The novel aspect of this work is an investigation of alternate methods for sampling the design space in conjunction with concurrent processing on a 256 processor, SGI HPC server, Origin 2000. These new methods are variation of DOE methods with the principal aim being to maximize the information gathered about the design space of interest. The data gathered from the analysis of the sampled design space if further used for construction of metamodels (approximations) and optimization. The application problems considered in this work include composites lay-up design optimization of an Air Borne Laser (ABL) Optical Bench that accounts for structural, thermal and optical line of sight constraints, as well as a conceptual design of a supersonic business aircraft involving aerodynamics, structures and propulsion disciplines.
Download The Complete Paper
(A one-time registration will be requested)
