Session: 36-09 Machine Learning & Artificial Intelligence Methods - Part 4

Submission Number: 177674

Assessment of a Full-Field Neural Surrogate (MARIO) for Turbomachinery Flow Applications 

Learning full flow fields is a very current challenge in surrogate modelling applications of neural networks for physics applications. Yet, the advatanges of methods capable of predicting entire fields at a glance are not clear.
When a desired output is known (which is often in the form of a scalar or 1D distribution), there is not necessarily a need to predict the entire field to obtain it. Generally, simpler and much smaller Deep Learning (DL) models such as Multilayer Perceptrons (MLP) would suffice. As a result, application cases for models capable of predicting entire fields are not always evident. 
The general belief is that the capacity to predict full fields would trigger a need for such models. 
In this context, the present study aims at evaluating the capacity of a state-of-the-art DL model, namely MARIO (Modulated Aerodynamic Resolution Invariant Operator), to learn realistic flow fields in the context of turbomachinery. To achieve this, two data-sets are used: VKI-LS59—composed of 2D RANS computations of a turbine cascade with geometric variations and various operating points—and faNN380—composed of 3D RANS computations of a compressor rotor at various operating points. 
An inherent strength of implicit neural representations (INR, the category of networks MARIO falls in) is demonstated by tackling a full scale data-set of 3D RANS simulations with a total of 6.4M points in the learned blocks, and a training time of one hour on a single GPU. 
The model is evaluated following two perspectives: a deep learning perspective, with metrics such as the mean squared error, and an engineering perspective, through the post-processing of relevant metrics on the predicted flow fields. It is then compared to smaller models in the form of simple MLPs on the direct prediction of post-processed data of interest. 
It is discussed that although the field predictions may not seem satisfying according to deep learning metrics, many relevant quantities can be accurately derived from the predicted fields. This includes 0D metrics such as the isentropic efficiency, 1D metrics such as the isentropic Mach number on the blade surface and 2D metrics such as the shock position along the blade. Furthermore, the accuracy in learning the flow topology leads to the posibility of using the surrogate model as a mean to qualitatively explore the flow field and identify regions of interest in the design space.
As a result, a light is shed on the potential of this full-field model in the context of turbomachinery for real world applications.
 

Presenting Author: Jean Fesquet ISAE-Supaero

Presenting Author Biography: Jean is a PhD student at ISAE-Supaero

Authors:

Jean Fesquet ISAE-Supaero
Michael Bauerheim ISAE-Supaero
Nicolas Binder ISAE-Supaero