Application of Deep Learning for Fan Rotor Blade Performance Prediction in Turbomachinery

Abstract

Turbomachinery fan optimization is a complex multidisciplinary process, which forces engineers to rely on strong theoretical assumptions and/or can be very computationally expensive. Additionally, with new constraints arising, such as distorted inflow in boundary layer ingestion cases, it is essential to find surrogate models able to account for the requirements and produce satisfying results, while capitalizing on the computational and experimental data already produced on other (e.g., previously developed) configurations. Toward this objective, the present study aimed at predicting the performance of the rotor of a turbomachine fan stage using deep learning (DL) techniques. These approaches have been showing increasingly convincing results in recent times, yet usually applied to toy problems or simplified configurations. Thus, this work evaluates the feasibility of applying DL models to optimize the shape of realistic fan rotor blades. To that end, a pipeline is presented to generate and mesh new geometries, run simulations, and finally train deep neural networks to be used as surrogates for performance prediction. In this framework, a u-net-type deep neural network was used to predict 2D wake flow fields of entropy and two 0D metrics, efficiency, and pressure ratio from the geometry of the blade and its operating conditions. To reduce the complexity of the predictive tasks, a transformative approach is used, by opposition to a fully generative one. For model testing and training, 75 geometries were built through interpolation of pre-existing, parametrized rotor blades. In turn, Reynolds-averaged Navier–Stokes (RANS) computations at various operating points were performed. The model was compared to proper orthogonal decomposition (POD)-Kriging techniques. Results showed that the neural network was only a slight improvement on an isogeometry dataset, but widely outperformed the POD-Kriging model on the multigeometry dataset. As a conclusion, it provided a good proof of concept to learn flow field views and global performance metrics on realistic, 3D, fan rotor geometries to be later used for optimization.