Neuronové sítě pro aerodynamickou optimalizaci založenou na metamodelech

Artificial Neural Networks for Surrogate-based Optimization in Preliminary Aerodynamic Design

článek ve sborníku ve WoS nebo Scopus

en

A preliminary aerodynamic design often imposes requirements on global optimum search within a large, highly multimodal design space. Tools typically deployed to evaluate individual design candidates are very computationally expensive, being part of the finite volume computational fluid dynamics class. This virtually prevents deployment of traditional stochastic global optimization approaches, such as evolutionary algorithms. Hence, there has been a growing interest in metamodelling techniques, providing a reliable surrogate of the simulator response to an optimization algorithm. Efficient deployment of such techniques within preliminary aerodynamic design is of interest to Garteur Action Group 52 members. The present paper describes the involvement of Brno University of Technology, Institute of Aerospace Engineering in the AG52. The considered test case is based on the RAE2822 aerofoil constrained multipoint optimization problem. The overall problem setup is given along with selected surrogate modelling and optimization techniques. The presented approach featuring artificial neural networks is able to produce highly reliable metamodels with cutting-edge performance as documented by the AG52 performance metrics comparison.

Předběžný aerodynamický návrh často klade požadavky na hledání globálního optima v rámci velkého, vysoce multimodálního návrhového prostoru.Nástroje typicky užívané pro výpočet jednotlivých návrhových bodů jsou velmi výpočetně náročné. Většinou se jedná o implementaci metody konečných objemů. Tento fakt v podstatě vylučuje užití tradičních stochastických optimalizačních metod jako jsou evoluční algoritmy. Z tohoto důvodu roste zájem o optimalizaci založenou na metamodelech, které poskytují spolehlivou náhradu skutečné odezvy při výrazně nižší výpočetní náročnosti. Tento přístup je zkoumán v rámci pracovní skupiny Garteur AG 52. Článek popisuje přínos VUT v Brně k aktivitám AG52. Vyvinutý metamodel vykazuje špičkové parametry a překonává náhradní modely ostatních partnerů ve všech sledovaných metrikách.

A preliminary aerodynamic design often imposes requirements on global optimum search within a large, highly multimodal design space. Tools typically deployed to evaluate individual design candidates are very computationally expensive, being part of the finite volume computational fluid dynamics class. This virtually prevents deployment of traditional stochastic global optimization approaches, such as evolutionary algorithms. Hence, there has been a growing interest in metamodelling techniques, providing a reliable surrogate of the simulator response to an optimization algorithm. Efficient deployment of such techniques within preliminary aerodynamic design is of interest to Garteur Action Group 52 members. The present paper describes the involvement of Brno University of Technology, Institute of Aerospace Engineering in the AG52. The considered test case is based on the RAE2822 aerofoil constrained multipoint optimization problem. The overall problem setup is given along with selected surrogate modelling and optimization techniques. The presented approach featuring artificial neural networks is able to produce highly reliable metamodels with cutting-edge performance as documented by the AG52 performance metrics comparison.

náhradní modelování, metamodel, vícebodová optimalizace, Garteur AG52, RAE2822

surrogate modelling, metamodel, multi-objective optimization, multi-point, Garteur AG52, RAE2822

2015

01.09.2015

University of Strathclyde

Glasgow, UK

9788890632310

Eurogen 2015 Extended Abstracts Book

1

28–34

7