E-Thesis 136 views 28 downloads
The Behaviour of Evolutionary Algorithms for the CFD-Driven Design Optimisation of Aerofoils / HANNAH DITCHBURN
Swansea University Author: HANNAH DITCHBURN
PDF | E-Thesis – open access
Copyright: The Author, Hannah R. Ditchburn, 2023.Download (4.3MB)
This thesis utilises aerodynamic shape optimisation software AerOpt and FLITE2D, to explore the behaviour of three Evolutionary Algorithms, Differential Evolution (DE), Modified Cuckoo Search (MCS), and Particle Swarm Optimisation (PSO), to optimise a 2D nonsymmetric aerofoil, providing an evaluatio...
Swansea, Wales, UK
|Degree level:||Master of Research|
|Degree name:||MSc by Research|
|Supervisor:||Evans, Ben J. and Walton, S.|
No Tags, Be the first to tag this record!
This thesis utilises aerodynamic shape optimisation software AerOpt and FLITE2D, to explore the behaviour of three Evolutionary Algorithms, Differential Evolution (DE), Modified Cuckoo Search (MCS), and Particle Swarm Optimisation (PSO), to optimise a 2D nonsymmetric aerofoil, providing an evaluation of their aerodynamic optimising capabilities.The aerofoil used in test cases is the NACA21120, where a variation of control node approaches are utilised to alter the aerofoil’s geometry. In the first set of test cases, a control node is placed on the upper surface to allow the thickness to be altered, and in the second set of cases, six control nodes are arranged along the boundary of the aerofoil, to examine the overall shape change.A mesh convergence study helped to determine the best mesh settings for the given problem. Each algorithm is tested in a subsonic, transonic, and supersonic flow regime to ensure the test cases fulfil the CFD aspect of the research. All flow regimes were treated as viscous with the relevant Reynolds number applied. To provide an analysis on how tuning the input parameters affects the algorithm’s behaviour, the number of agents were inputted were varied from 10 to 50 to 99. The generations number was set to 99, and the fitness objective was to optimise for the lift-drag ratio (L/D), throughout all optimisations.The first set of results (one control node) found that fitness improvements were largest in the transonic cases, increasing the L/D by an average percentage of 213%. The aerofoil’s L/D at Mach 0.5 was improved by an average of 80%, and Mach 1.5 by 33%. Each algorithm showed a similar trend in which the control node was positioned at the final generation in the design space, this varied depending on the Mach number being optimised for, either resulting in an increase or decrease in the aerofoil thickness. Varying the number of agents inputted, had a more significant effect on MCS, whereas DE and PSO showed more consistent results regardless of the number of inputted agents. Generally, PSO displayed fastest convergence of all the agents, shortly followed by DE, followed by MCS.The second set of results (six control nodes) were optimised for identical input parameters but for simplicity, at a single flow regime, Mach 1.5. Differing from the first set of results showing similar control node placement within the design space, the second set of results showed the algorithm’s position some of the control nodes in different locations within the design space. Despite the similar fitness improvement values seen between DE and PSO, the final geometries were observed to be somewhat varied, where DE reduced the thickness of the trailing edge, but PSO increased it. MCS displayed similar geometry change to PSO but with more conservative control node movement.
Aerofoil, Agent, Computational Fluid Dynamics, Control Node, Evolutionary Algorithm, Fitness, Generation, High Performance Computing
Faculty of Science and Engineering