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

Submission Number: 178867

Deep-Learning Based Implicit Parametrization of Turbomachinery for Optimal Generation of CAE Surrogates 

Computer-Aided Engineering (CAE) is based on the numerical solution of partial differential equations for flow, thermal, and mechanical stresses fields evaluation, which requires significant time and resources. To reduce the associated computational cost, Original Equipment Manufacturers (OEM) have started deploying Artificial Intelligence (AI) surrogate models, trained on datasets of numerical simulations that cover a prescribed design space. Usually, the surrogates use a parametric representation that allows direct control of machine geometries  and  this is the preferred description in industrial design processes. Such representation may not be necessarily optimal in terms of surrogate model performance or of sample efficiency, which is a factor of paramount importance when applying AI surrogate models in actual industrial practice. Also, this makes it harder to train models on geometries described by different parametrizations, sometimes preferred to cover different discipline requirements, and limits the design space which can be explored by an optimizer tied to such a parametrization.  In this work, the use of data-learned latent geometry representations is explored for different turbomachines and tasks. They are compared to parametric description of the same geometry in terms of representation accuracy, downstream surrogate model performances, and sample efficiency. In this study, we consider linear dimensionality reduction methods, morphing approaches, and deep-learning based techniques. The results show interesting connections between the size of the implicit geometry representation, the size of the training dataset, and the type of AI surrogate model used downstream of the geometry encoder. In particular, the deep-learning based geometry representations are at least as efficient as the initial geometry parametrizations, in terms of reconstruction accuracy and impact on the surrogate model performance. This confirms that such methods can be applied to all the cases where a geometry parametrization doesn’t exist or is different for different samples and offer opportunities to speed up multidisciplinary optimization methods. 

Presenting Author: Andrea Panizza Baker Hughes

Presenting Author Biography: Andrea Panizza is Senior Principal AI Specialist in Baker Hughes, mainly working on AI for Engineering Design, Large Language Models and Computer Vision. He has previous experience in Computational Fluid Dynamics and Aerodynamic Design of Centrifugal Compressors. His interests lie in developing and deploying Artificial Intelligence products for real world applications.

Authors:

Andrea Panizza Baker Hughes
Leonardo Pulga Baker Hughes
Laure Barriere Baker Hughes
Alessandro Pela University of Florence
Marco Bicchi Baker Hughes
Andrea Agnolucci Baker Hughes
Federico Funghi Baker Hughes
Vittorio Michelassi Baker Hughes