This paper draws motivations from the fact that current efforts of engineering optimization process were mostly carried out from scratch; i.e., tabula rasa-style design. In contrast however, the humans routinely take advantage of the knowledge that drawn from past experiences whenever a new task is met. Such transfer learning process by leveraging knowledge from completed tasks can be promising to significantly improve the performance over the current state-of-the-art algorithms, particularly in solving expensive black-box problems. In light of the above, we propose a cokriging based transfer optimization framework for the design of turbomachinerey cascades, which was demonstrated in redesign optimization of the GE-E^3^ first-stage vane. Specifically, when building cokriging surrogate in such transfer optimization context, the samples from already completed task (also called source task) are treated as low-fidelity (LF) data. Meanwhile, since these LF samples act as “past experiences”, they remained unchanged during the whole optimization process, i.e., we only query points of the high-fidelity (HF) model in iteration process. The acquisition function of expected improvement is adopted to guide the search process in such transfer optimization framework. On the other hand, the cokriging requires the initial HF samples be a subset of LF dataset, therefore, one of our efforts were drawn to designing selection strategies of initial HF samples, in order to make full use of the “past experiences”. In addition, we also need to take care of the bad effects of “negative transfer”, i.e., sometimes the “past experiences” may do harm to the optimization of target task. Three strategies for the selection of initial HF samples were proposed, their performance were compared in different scenarios as (1) the optimization of GE-E^3^ blade in different working conditions and (2) the optimization of blade section contours at different spans of the vane cascade (e.g., at the hub and pitch span, respectively ). In order to analyze the performance of initial selection strategies in avoiding “negative transfer”, the correlation coefficients between source and target tasks in each optimization process were calculated. The results showed that, by learning from the past experiences, the transfer optimization framework can reduce the computational cost as much as 50%. More importantly, our proposed transfer learning strategies were shown can effectively avoid “negative transfer” and thus making our algorithm always achieve better solutions than the compared state-of-the-art algorithms.


Transfer Optimization in Accelerating the Design of Turbomachinery Cascades

Division / Sponsor:
International Gas Turbine Institute (IGTI)

Paper Proceedings: https://asmedigitalcollection.asme.org/GT/GT2020/volume/84058

Turbo Expo - Virtual Conference

Supporting the need among researchers worldwide for additional online and multimedia content, ASME now offers the ASME Conference Video Collection. This platform will collect and provide long-term access to conference presentation videos, encompassing all technology areas within ASME's technical conference and events program. About 30,000-40,000 engineers across all mechanical engineering and related disciplines participate in ASME conferences each year. This includes authors of approximately 6,000 to 8,000 technical papers published by ASME, the majority of whom also submit video presentations for our virtual conferences. The video collection will be available as an add-on to existing product packages or as a separate subscription.