Abstract:
Transonic and supersonic aerofoil inverse design for different flight conditions is carried
out using Deep Neural Networks (DNN). DNN are combined with a comprehensive and
complete database of aerodynamic data and aerofoil geometry parameters to form the pillars of a surrogate inverse aerodynamic design tool. The framework of this research starts
with the aerofoil parameterisation. The Class/Shape Transformation functions (CST) was
selected for the parameterisation process due to its high accuracy and flexibility when
describing complex shapes. An automated mesh technique is created and implemented
to discretise the flow domain. The aerodynamic computations are performed for 395
aerofoils. Spatial discretisation is accomplished with the Jameson-Schmidt-Turkel (JST)
scheme and convergence is reached by the backward Euler implicit numerical scheme.
Data are collected and managed with the CST parameters for all aerofoils and their respective aerodynamic characteristics from the CFD solver. The Deep Neural Network
is then trained, validated using cross-validation and evaluated against CFD data. An extensive investigation of the effect from different DNN configurations takes place in this
research. Within this thesis, different case studies are presented for different numbers of
design objectives. For the inverse design process the NACA 66-206 aerofoil was selected
as the baseline aerofoil, to reduce the aerodynamic drag coefficient while maintaining or
improving the lift coefficient, to obtain a superior lift/drag ratio compared with the baseline aerofoil. The framework of this thesis have proved to output aerofoil designs with an
improved lift/drag ratio in comparison with the baseline aerofoil.