UC San Diego

Multidisciplinary design optimization (MDO) is an approach that usesoptimization methods to design complex engineering systems involving multiple disciplines simultaneously. The coupled nature of multidisciplinary systems and the large number of design variables involved in complex systems present unique challenges to solving MDO problems. One of these challenges is the implementation of software necessary to evaluate multidisciplinary models within an optimization framework. When gradient-based optimization approaches are used, efficient and accurate derivatives must be computed for each model evaluation. This dissertation presents an approach that significantly reduces the manual effort required to implement computational models for use within a gradient-based MDO framework, especially in large-scale problems. This dissertation introduces a novel approach to address these challenges and automate the process, enabling accurate and efficient adjoint-based sensitivity analysis for gradient-based MDO in particular. To address these challenges, a three-stage compiler methodology is proposed. The methodology centers around a graph representation that provides a foundation for automating sensitivity analysis in MDO. In addition, the Computational System Design Language (CSDL) is introduced, which allows for a concise description of the physical system. The adoption of CSDL demonstrates a twofold reduction in code complexity for engineering models, significantly reducing the barrier to entry for MDO practitioners. The three-stage compiler also provdes users of CSDL with the ability to measure the effect of the model structure on run-time performance and memory complexity using the graph representation. Finally, the methodology developed in this dissertation is applied to the design of a space-based virtual telescope comprised of two spacecraft flying in formation. A reformulation of the orbit dynamics of the spacecraft is found to avoid the introduction of truncation errors due to tight formation constraints that render solving the optimization problem impossible. A sequential approach to applying MDO to the design of a space-based virtual telescope is also found to be more robust than solving the MDO problem where all disicplines are considered simultaneously.