Adaptive Surrogate-Based Multi-Disciplinary Optimization for Vane Clusters

Abstract

The presented work is part of a research project aimed towards multi-disciplinary robust shape optimization of low pressure turbine (LPT) vane clusters. Multi-disciplinary analysis for vane cluster optimization is used to evaluate design constraints, involving 3D aerodynamic Navier-Stokes simulation, transient thermal analysis, structural analysis and life prediction. The expense of these simulations combined with high-dimensional design space, makes the application of gradient-based or stochastic optimizers inefficient. To overcome these issues, a surrogate-based optimization approach is proposed here. High quality surrogate models are required for accurate description of the constraints with life prediction. Adaptive Global Surrogate-Based Optimizer, based on Gaussian-Process (GP) surrogate models and Expected Improvement infill criteria is employed, which allows to efficiently increase the surrogate quality while approaching the optimal solution at the same time. Additional techniques are introduced to deal with the geometry rebuild failure, as some combinations of the design parameters may produce infeasible geometry. The adaptive optimization method is successfully applied to the multi-disciplinary problem for the vane cluster shape optimization. The comparison of the method performance with a gradient-based optimizer indicates that a much lower number of true simulations is needed by the proposed method to find an optimal design. Successful optimization results shows the ability of the method to handle simulation crashes, caused by geometry rebuild failure.