0603704 - ÚT 2025 eng A - Abstract
Kubíčková, Lucie - Gebouský, Ondřej - Haidl, Jan - Isoz, Martin
Accelerating shape optimization by adaptively updated deep neural networks.
[OpenFOAM Workshop /19./. 25.06.2024-28.06.2024, Peking]
R&D Projects: GA TA ČR(CZ) TN02000069
Institutional support: RVO:61388998
Keywords : deep neural networks * shape optimization * multiobjective evolutionary algorithms * computational fluid dynamics
OECD category: Applied mathematics
Result website:
https://file.bagevent.com/resource/20241009/1353234153802240.pdf

Modern trend in component design is to tailor the component geometry to a specific application. This stimulates the development of automatic shape optimization methodologies. In this work, we present a hybrid and adaptive shape optimization algorithm (CFDNNetAdapt) that combines deep neural networks (DNNs), computational fluid dynamics (CFD) and multi-objective evolutionary algorithms (MOEAs). The algorithm switches between purely CFD-based optimization and training of DNNs on acquired data. The trained DNNs are subjected to the MOEA, and the DNN-optimization results are checked by CFD. If the DNN error is small enough, the similarity of the CFD- and DNN-results is investigated. If not, simple heuristics is used to adaptively change the tested DNN architectures, and CFD-based optimization is continued. To illustrate the algorithm capabilities, a shape optimization of a single-phase ejector was conducted. Eventually, CFDNNetAdapt was able to outperform the purely CFD-based optimization, especially by speeding up the mid-to-late MOEA iterations.
Permanent Link: https://hdl.handle.net/11104/0361730