Knowledge transfer accelerated turbine blade optimization via an sample-weighted variational autoencoder

Abstract

The aerodynamic shape optimization (ASO) of turbine blades is a typical expensive black-box problem. With the ever increasingly strict design requirements, high-fidelity simulations are demanded, and the related simulation cost increases sharply, which means fewer simulation calls are allowed within a design optimization cycle. Then, the problem as “how to achieve better solution with fewer simulations” becomes rather crucial. Inspired by the concept of transfer learning, we propose to leverage the samples of related completed tasks (also known as source samples) to address the above problem. Specifically, an sample-weighted variational autoencoder (VAE) is proposed, which includes the source samples as part of the training dataset to boost the proportion of high-scoring solution candidates in the learnt design space of target problem, and thus increasing the probability of attaining substantially better solutions. The including of source samples in VAE training also helps to increase the accuracy of reparameterizing the source samples in target design space. Furthermore, a multi-fidelity surrogate is used to guide the optimization search, which extracts information from the source samples to help the algorithm to more quickly arrive at the small neighborhood of the true optimal of the target problem. With the above, a knowledge transfer accelerated aerodynamic shape optimization procedure is proposed for the turbine blade design, labeled as KT-ASO. Through test on a highly-loaded low-pressure turbine blade and compared against the ASO procedures without knowledge transfer, the proposed KT-ASO is shown to reduce at least 50% function calls when achieving similar optimal solutions. In the meantime, under the same sample budget, our proposed KT-ASO is shown to attain much better solutions than the compared ASO procedures.