Multifidelity Approaches for Parallel Multidisciplinary Optimization
Andrew March, Karen E. Willcox · 12th AIAA Aviation Technology, Integration, and Operations (ATIO) Conference and 14th AIAA/ISSMO Multidisciplinary Analysis and Optimization Conference · 2012
Optimization of systems is often plagued by computationally expensive simulations and the need for iterative analysis to resolve coupling among subsystems. These challenges are typically severe enough as to prohibit formal system optimization. This paper formulates two new methods to parallelize the optimization of multidisciplinary systems. The rst method decomposes the system optimization problem into multiple subsystem optimiza- tions that are solved in parallel. The second method generates a list of designs at which computationally expensive simulations should be run, evaluates those designs in parallel, and then solves an inexpensive surrogate-based optimization problem. Both methods en- able the use of multidelity optimization to nd an optimal solution with respect to the highest-delity models available. In addition, both methods exploit high-delity sensitivity information if available, but do not require gradients of the high-delity models. Thus, the methods are applicable to problems with black-box codes and/or noisy function evalua- tions. The two methods are demonstrated on three analytical optimization problems, and a multidisciplinary, multidelity aerostructural design optimization problem.