Bayesian Optimization for Multi-Objective High-Dimensional Turbine Aero Design
Abstract Industrial design fundamentally relies on high-dimensional multi-objective optimization. Bayesian Optimization (BO) based on Gaussian Processes (GPs) has been shown to be effective for this practice where new designs are picked in each iteration for varying objectives including optimization and model refinement. This paper introduces two industrial applications of BO for turbine aero design. The first application is GE’s Aviation & Power DT4D Turbo Aero Design with 32 design variables. It has a single objective to maximize with 32 input/design variables and thus considered high-dimensional in terms of the input space. BO has significantly succeeded the traditional design schemes. It has been shown that finding the maximum-EI points (inner-loop optimization) could be critical and the influence of inner-loop optimization was evaluated. The second application is for multi-objective optimization. Each simulation run is the aggregate result from multiple CFD runs tuning geometry and took 24 hours to complete. BO has been capable to extend the existing Pareto front with a few additional runs. BO has been searching along the border of the design space and therefore motivate the open-up of design space exploration. For both applications, BO successfully guide the CFD run and allocate design variables more optimum than previous design approaches.