POD-Driven Adaptive Sampling for Efficient Surrogate Modeling and its Application to Supersonic Turbine Optimization

Abstract

A new approach for adaptively sampling a design parameter space using an error estimate through the reconstruction of flow field by a combination of proper orthogonal decomposition (POD) and radial basis function network (RBFN) is presented. It differs from other similar approaches in that it does not use the reconstructed flow field by POD for the evaluation of objective functions, and thus it can be a subset of the flow field. Advantages of this approach include the ease of constructing a chain of simulation codes as well as the flexibility of choosing where and what to reconstruct within the solution domain. An improvement in achieving a good prediction quality, with respect to other adaptive sampling methods, has been demonstrated using supersonic impulse turbine optimization as the test case. A posteriori validation of the surrogate models were also carried out using a set of separately-evaluated samples, which showed a similar trend as the Leave-One-Out (LOO) cross-validation. The progressively enriched surrogate model was then used to achieve the more uniformly populated Pareto front with fewer number of function evaluations.