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

Paper Number: 121752

121752 - Data-Driven Inverse Design Method for Turbomachinery 

Inverse design methods have been proposed for turbomachinery design for a few decades, where various approaches such as analytical correlations are used to design the turbomachinery component given the prescribed output performance. In the wake of the advance of numerical methods and computation resources, 3-dimensional (3D) simulations have become a standard for turbomachinery designs, where hundreds and thousands of simulations are run throughout the design cycle of a component. The wealth of accurate simulation results, in the era of big data, opens the door to design methods, where a machine learning data-driven model learns to generate new designs able to meet the user prescribed conditions. 

 

In this paper, a data-driven inverse design method based on Artificial Neural Networks (ANN) is proposed for turbomachinery. Using simulated data, we train a multilayer perceptron with design parameters as input, and performance attributes, e.g., efficiency obtained from 3D simulations, as output. Now, target performance can be prescribed and fed to the trained model in order to automatically generate diverse designs. As validated in the use case of turbocharger turbine stage, the true efficiencies of the generated designs are in good agreement with the prescribed target efficiencies. Furthermore, the trained model can generate accurate designs significantly outside of the trained data set distribution, indicating good generalization properties.  

 

The proposed ANN-based inverse design method is a novel approach based on big data. Our method is widely applicable to numerous other tasks since the model optimization is purely data-driven, and non-intrusive to the simulation or data-generation processes. Combining the data from different processes, multi-component and multi-physics aspects can easily be considered and designs can be readily generated. The method can trivially be scaled up to cater for huge amount of data provided the availability of computer resources. The method also lays down the foundation for further developments adapting advanced generative artificial intelligence techniques. This paper illustrates the interdisciplinary outcome of turbomachinery and computer sciences and is highly relevant for turbomachinery and in general engineering designs. 

Presenting Author: Kwok Kai So Turbo Systems Switzerland Ltd

Presenting Author Biography: Dr. So has completed his Bachelor of Mechanical Engineering at The University of Hong Kong, his Diplome d'Ingenieur at the École nationale supérieure d'arts et métiers in France, his Dr. Ing at the Chair of Aerodynamics and Fluid Mechanics at the Technische Universität München in Germany, and his Diploma of Advanced Studies in Data Science at ETH Zurich, Switzerland. Since 2012, Dr. So has been working at Turbo Systems Switzerland Ltd on CFD, optimization and machine learning topics.

Authors:

Kwok Kai So Turbo Systems Switzerland Ltd
Luis Salamanca Swiss Data Science Center, ETH Zurich and EPFL, Switzerland
Firat Ozdemir Swiss Data Science Center, ETH Zurich and EPFL, Switzerland
Fernando Perez-Cruz Swiss Data Science Center, ETH Zurich and EPFL, Switzerland / Computer Science Department, ETH Zurich, Switzerland