Automation enhancements in multidisciplinary design optimization.

摘要

The process of designing complex systems has necessarily evolved into one which includes the contributions and interactions of multiple disciplines. To date, the Multidisciplinary Design Optimization (MDO) process has been addressed mainly from the standpoint of algorithm development, with the primary concerns being effective and efficient coordination of disciplinary activities, modification of conventional optimization methods, and the utility of approximation techniques toward this goal. The focus of this dissertation is on improving the efficiency of MDO algorithms through the automation of common procedures and the development of improved methods to carry out these procedures.; In this research, automation enhancements are made to the MDO process in three different areas: execution, sensitivity analysis and utility, and design variable move-limit management. A framework is developed along with a graphical user interface called NDOPT to automate the setup and execution of MDO algorithms in a research environment. The technology of automatic differentiation (AD) is utilized within various modules of MDO algorithms for fast and accurate sensitivity calculation, allowing for the frequent use of updated sensitivity information. With the use of AD, efficiency improvements are observed in the convergence of system analyses and in certain optimization procedures since gradient-based methods, traditionally considered cost-prohibitive, can be employed at a more reasonable expense. Finally, a method is developed to automatically monitor and adjust design variable move-limits for the approximate optimization process commonly used in MDO algorithms. With its basis in the well established and probably convergent trust region approach, the Trust region Ratio Approximation method (TRAM) developed in this research accounts for approximation accuracy and the sensitivity of the model error to the design space in providing a flexible move-limit adjustment factor. Favorable results are obtained using the TRAM strategy in comparison to existing move-limit strategies.