A Comparison of Methodologies for Generating Design-Requirement-Driven Conceptual Aircraft Designs

Traditional aircraft design techniques leveraging high-fidelity analyses are often limited by the prohibitive cost of design space exploration. Therefore, in the context of large design spaces and unconventional aircraft concepts, the use of machine learning to propose suitable designs represents an attractive solution for conceptual-stage exploration, a process termed generative design. In this paper, we construct a large dataset of narrow-body commercial airliner configurations, which we use to compare several generative machine learning techniques - including variational autoencoders (VAEs), normalizing flows, and diffusion models - to assess their ability to generate diverse, feasible, and realistic conceptual aircraft designs in a fraction of the time required to perform a full conceptual design. Preliminary results indicate that normalizing flows and diffusion models generate the more representative and diverse designs than VAEs. This study highlights how machine learning can support conceptual aircraft design, with relevance to unconventional configurations and sustainable aviation.