Session: 36-03 Neural-Network based approaches (2)

Paper Number: 127348

127348 - Application of Deep Learning for Fan Rotor Blade Performance Prediction in Turbomachinery 

Turbomachinery fan optimisation 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. Towards this objective, the present study aims at predicting the performance of the rotor of a fan stage of turbomachine using Deep Learning (DL) techniques. These approaches have been showing increasingly convincing results in recent times, yet usually applied on toy models or simplified configurations. Thus, this work evaluates the feasibility of applying DL models to optimise the shape of realistic fan rotor blades. To that end, a pipeline is presented to automatically generate new geometries, construct the associated mesh, run simulations, and finally train deep neural networks to be used as surrogates for optimization. In this framework, Deep neural networks are used to predict axial cuts of the flow-field and various metrics, such as efficiency, from the shape of the rotor and the targeted operating point. From a methodology viewpoint, generative models are often used to tackle problems of the sort, that is predicting high-dimensional outputs (e.g. 2D images) from low-dimensional inputs (e.g. the parametrisation of a blade shape). This leads to complex tasks which render the model training process arduous. To reduce the complexity of the predictive tasks, as well as to increase the size of the dataset artificially, a transformative approach is used here, by opposition to the generative one. It is based (in this case) on a UNet which transforms the output of a sample at a specified operating point and for a given geometry, into the output of another arbitrary operating point on an arbitrary geometry included in the design space. To train the models, a dataset of periodic RANS results of over 80 geometries at various operating points has been built. The generation of the considerable set of geometries is a two-step process. Diverse fan rotor geometries from both the industry and the academia are parametrised using an open-source library. In turn, new blade shapes are created through weighted interpolation of the pre-existing geometry set. This yields a mean to greatly increase the number of shapes in the dataset, while keeping the dimensionality of the design space relatively low. Furthermore, owing to the pipeline developed for this study and the small amount of time required per calculation, a large number of operating points are assessed employing CFD for the various geometries. For the purely convolutional model, axial cuts of the wake flow-field are estimated with good accuracy. A second model capable of outputting performance scalars, namely the isentropic efficiency, the flow coefficient and the stage loading, in addition to axial cuts is trained with the idea of conducting shape and performance optimisation tasks. The models are compared to POD- Kriging techniques on metrics such as sample efficiency and prediction accuracy in both interpolation and extrapolation regimes. Results show that the proposed pipeline based on a transformative UNet outperforms POD, even for small training datasets. As a conclusion, it provides a good proof of concept to learn performance maps and flow field views on realistic 3D geometries of a fan rotor, to be later used for optimisation.

Presenting Author: Jean Fesquet ISAE-Supaero

Presenting Author Biography: Jean Fesquet is an engineer working in the department of aerodynamics, energetics and propulsion at ISAE-Supaero. He focuses currently on deep learning applied to blade performance optimisation in turbomachinery.

Authors:

Jean Fesquet ISAE-Supaero
Michael Bauerheim ISAE-Supaero
Ludovic Rojda ISAE-Supaero
Yannick Bousquet ISAE-Supaero
Nicolas Binder ISAE-Supaero