A New Paradigm for Large-Scale System Design Optimization

University Of California-San Diego, La Jolla CA

Investigators

Abstract

Design optimization is the computation of design variable values that minimize or maximize an objective subject to constraints, where the objective and constraint functions are the outputs of an engineering model. Applying large-scale optimization - which involves up to thousands of design variables - to engineering design is challenging, because of the conflicting requirements of efficient derivative computation for scalability and coupling multiple disciplines for system-level modeling. However, state-of-the-art gradient-based optimizers, combined with the PI's recent work in developing a unified theory for multidisciplinary derivative computation, have made it feasible to solve large-scale design optimization (LSDO) problems in only hundreds of model evaluations. The objective of this project is to accelerate large-scale system design optimization algorithms by an order of magnitude compared to the state-of-the-art approach. This improvement will be achieved through a paradigm shift enabling a novel optimization algorithm that uses a hybrid of reduced-space and full-space optimization. This project will investigate a new, intrusive paradigm in which the internal components of the model are exposed to the optimizer. An intrusive paradigm enables a novel optimization algorithm that would achieve the robustness of a reduced-space formulation and the efficiency of a full-space formulation if the two formulations can be unified. The difference between the two formulations is that full-space treats the model's states as design variables. This research will result in theoretical and algorithmic contributions to sequential quadratic programming (SQP), which is the most common optimization approach in LSDO. The research project will broaden the unification to general SQP algorithms and leverage adjoint-based error estimation and inexact Newton methods to determine methods for adaptively selecting the hybrid of reduced and full space. The resulting algorithms will be made available through open-source licensing, allowing the efficiency improvements to benefit students, researchers, and practitioners. Moreover, the hybrid algorithm removes the need for practitioners to choose between reduced and full space problem formulation; therefore, less effort and expertise will be required for LSDO. The largest impact will be on industry, where the efficiency and usability improvements will significantly lower the barrier-to-entry for using LSDO to help design complex engineered systems, which will be demonstrated in collaboration with General Atomics Aeronautical Systems. This award reflects NSF's statutory mission and has been deemed worthy of support through evaluation using the Foundation's intellectual merit and broader impacts review criteria.