A deep learning framework for expert-guided sailplane performance optimisation

Abstract

The aerodynamic optimisation of high-performance aircraft is traditionally hampered by computationally prohibitive simulations, limiting the scope and agility of design space exploration. To overcome this barrier, we introduce a unified, data-driven framework founded on deep neural networks that enables rapid, expertguided optimisation. The framework replaces the conventional simulation pipeline with a suite of bespoke deep learning models: one for the controllable generation of airfoils from intuitive geometric parameters, and another for their instantaneous 2D performance prediction. Applied to the optimisation of a stateof-the-art 15-meter standard class sailplane, the framework achieved a mean lift-to-drag ratio increase of 1.26% over a competitive baseline, while rigorously adhering to all user-defined geometric constraints. The results demonstrate that combining expert-defined geometric constraints with data-driven design exploration can substantially reduce computational overhead while maintaining physical interpretability in preliminary aerodynamic design.