Rotor37 Aerodynamic Optimization: A Machine Learning Approach

Abstract The present paper presents an enhanced method for aerodynamic optimization of NASA Rotor37 row based on Machine Learning (ML) algorithms. An aerodynamic database has been developed using a commercial 3D (three dimensional) computational fluid dynamics (CFD) solver; a RANS (Reynold Averaged Navier Stokes) steady approach with a two-equation SST (Shear Stress Transport) model has been adopted for the aerodynamic computations. The database geometries have been parametrized through autoencoders with the aim of automatically extracting characteristic geometric features and obtaining extensive parameterization of the blade. Autoencoders are a specific type of Neural Network which allow to feed non-parametric CAD designs into ML models. The autoencoder latent parameters describe the blade 3D geometry and can be used as an alternative to the standard geometric parameters in describing the shape of each sample. The main advantage is that autoencoders enable an automatic parameterization of 3D geometries, thus overcoming the limits imposed by manual parameterization. A Neural Network has been used in order to predict global performance (e.g., pressure ratio, efficiency) and 3D field quantities (pressure and temperature distribution). The optimization through reinforcement learning algorithms has been carried out in order to maximize the blade efficiency; geometries have been generated by exploiting the latent parameterization obtained by autoencoder. The accuracy of the ML algorithm forecast has been evaluated through CFD simulations carried out on the optimal sample. The results related to the optimized sample have been presented and highlight all the benefits of the proposed approach.