Transonic Inverse Airfoil Optimisation Using Autoencoders

This work shows a methodology able to perform inverse airfoil optimisation using neural networks based on autoencoder architectures. An autoencoder neural network is built to codify surface pressure coefficients into a low-dimensional latent space. The latent space is represented by a set of latent variables that can be decoded into the surface pressure coefficients. The autoencoder is trained on a range of random airfoil shapes obtained through a suitable parameterization. This allows the set up of an optimiser based on the latent variables to obtain a particular surface pressure distribution. To recover the corresponding airfoil shape, a mapping between the parametric variables describing the airfoil shapes and the latent variables is also built. The autoencoder ability to recover the correct surface pressures is first demonstrated; this is followed by an inverse design optimisation problem. Results show the viability of the proposed approach to produce feasible designs at a fraction of the cost of higher fidelity methods.