Bayesian Collaborative Sampling for Multidisciplinary Design
Chung Lee, Dimitri N. Mavris · 12th AIAA Aviation Technology, Integration, and Operations (ATIO) Conference and 14th AIAA/ISSMO Multidisciplinary Analysis and Optimization Conference · 2012
The use of high order analysis codes such as CFD for multi-disciplinary design in early design phases is dicult due to the computational expense of the codes as well as iterative solutions required for inter-disciplinary compatibility. For example, for aero-structural design of a wing, iterative function calls from both disciplines may be needed to ensure that the wing shape used in the aerodynamics discipline is consistent with the deformed shape computed in the structures discipline. Bayesian collaborative sampling (BCS) is an adaptive sampling method that concentrates samples of disciplinary analyses in regions of a design space that are favorable to a system objective and compatible between disciplines. a BCS combines the collaborative optimization (CO) architecture with Bayesian models. In this example, a sparse Bayesian regression model is used to predict expected improvement and probable compatibility discrepancies, which are used as sampling criteria within a bilevel architecture. This technique can be used to explore a multidisciplinary design space when analysis function calls are expensive or limited. A simple glider wing aero-structures design problem is used as a test case. BCS concentrates sample points in regions of high lift-to-drag ratio and inter-disciplinary compatibility.