Techniques for Improving Neural Network-based Aerodynamics Reduced-order Models

Qingzhao, Wang; Medeiros, Renato; Cesnik, Carlos E. S.; Fidkowski, Krzysztof J.; Brézillon, Joël; Bleecke, Hans

doi:10.2514/6.2019-1849

Cited by 8 publications

(3 citation statements)

References 39 publications

Supporting

Mentioning

Contrasting

Order By: Relevance

“…This approach is called the non-intrusive reduced-order model. The surrogate model can be kriging [375], RBF [376], or ANN [315,[377][378][379]. Another approach, called the intrusive reduced-order model, uses projectionbased reduced model [380,381] to solve reduced-order flow variables; this approach usually requires more computational time and code implementation (such as using automatic differentiation to solve the partial derivatives [382]).…”

Section: Flow Field Modelingmentioning

confidence: 99%

Machine Learning in Aerodynamic Shape Optimization

Li¹,

Martins²

2022

Preprint

View full text Add to dashboard Cite

Large volumes of experimental and simulation aerodynamic data have been rapidly advancing aerodynamic shape optimization (ASO) via machine learning (ML), whose effectiveness has been growing thanks to continued developments in deep learning. In this review, we first introduce the state of the art and the unsolved challenges in ASO. Next, we present a description of ML fundamentals and detail the ML algorithms that have succeeded in ASO. Then we review ML applications contributing to ASO from three fundamental perspectives: compact geometric design space, fast aerodynamic analysis, and efficient optimization architecture. In addition to providing a comprehensive summary of the research, we comment on the practicality and effectiveness of the developed methods. We show how cutting-edge ML approaches can benefit ASO and address challenging demands like interactive design optimization. However, practical large-scale design optimizations remain a challenge due to the costly ML training expense. A deep coupling of ML model construction with ASO prior experience and knowledge, such as taking physics into account, is recommended to train ML models effectively.

show abstract

Section: Flow Field Modelingmentioning

confidence: 99%

Machine Learning in Aerodynamic Shape Optimization

Li¹,

Martins²

2022

Preprint

View full text Add to dashboard Cite

show abstract

“…The 10-fold cross-validation is performed since it is a good compromise between cost and accuracy. 22 Assume that twenty hidden neurons are involved in the first hidden layer. Eight neural networks with sample sizes varying from 100 to 450 are compared for the trial study.…”

Section: Numerical Studiesmentioning

confidence: 99%

Updating multi-fidelity structural dynamic models for flexible wings with feed-forward neural network

Huang

2022

Proceedings of the Institution of Mechanical Engineers, Part G:

View full text Add to dashboard Cite

In multidisciplinary design optimization of aerospace structures (e.g., a flexible wing), it may be convenient and practical to break such a complex problem into multi-fidelity, multi-stage design problems. Structural model updating is needed in multi-fidelity, multi-stage optimizations to ensure the consistency of models with different fidelity. However, due to the inequality in structural parameters, there exists a fundamental difficulty in the model updating from a lower fidelity model to a higher fidelity model. In this paper, a feed-forward neural network is applied to determine the structural dynamic characteristics of a higher fidelity model based upon a lower fidelity model. The feasibility of this approach is demonstrated by updating beam-like wings to a thin shell-based model and a one-cell wing box model, respectively. The quality and accuracy of model updating using the proposed method are also discussed regarding the neural network structure and sample size.

show abstract

“…Study the viability of Machine Learning techniques and Neural Networks algorithms [96] as an alternative/complement to ROM in order to obtain fast approximations of the solution fields of the coupled magnetomechanical presented. In [97], the accuracy of the POD technique is enhanced by making use of neural network in order to interpolate the POD coefficients. Also, data-driven approaches within Machine Learning [98] use numerical data in order to train the model (unsupervised training).…”

Section: Further Research Developmentsmentioning

confidence: 99%

Towards MRI scanner design: the Proper Generalised Decomposition in the context of coupled magneto-mechanical problems

Gassiot¹

View full text Add to dashboard Cite

Latest developments in high-strength Magnetic Resonance Imaging (MRI) scanners, with in-built high resolution, have dramatically enhanced the ability of clinicians to diagnose tumours and rare illnesses. However, their high-strength transient magnetic ﬁelds induce unwanted eddy currents in shielding components, which result in high-frequency vibrations, noise, imaging artefacts and, ultimately, heat dissipation and boiling oﬀ of the helium used to super-cool the magnets. Optimum MRI scanner design requires the capturing of complex electro-magneto-mechanical interactions with high ﬁdelity computational tools. Moreover, manufacturing new MRI scanners still represents a computational challenge to industry due to the large variability in material parameters and geometrical conﬁgurations that need to be tested during the early design phase. This process can be highly optimised through the employment of user-friendly computational metamodels constructed on the basis of Reduced Order Modelling (ROM) techniques, where high-dimensional parametric oﬄine solutions are obtained, stored and assimilated in order to be eﬃciently queried in real time.This thesis presents a novel a priori Proper Generalised Decomposition (PGD) computational framework for the analysis of the electro-magneto-mechanical inter-actions in the context of MRI scanner design to address the urgent need for the development of new cost-eﬀective methods, whereby previously performed compu-tations can be assimilated as training solutions of a surrogate digital twin model to allow for real-time simulations. The PGD methodology is derived for coupled electro-magneto-mechanical problems in an axisymmetric Lagrangian setting, in-cluding the possibility to vary several material and geometrical parameters (as part of the high-dimensional oﬄine solution), that are relevant for the industrial part-ner of the project, Siemens Healthineers. A regularised-adaptive strategy and a staggered PGD approach are proposed in order to enhance the accuracy and robust-ness of the PGD algorithm while preserving its a priori nature. The Lagrangian adaptation of the governing equations will allow for a comparison between staggered and monolithic solvers, where the staggered approach will be shown to enhance the robustness and accuracy of the PGD technique. Moreover, geometric changes in the computational domain will be accounted for in the PGD solution by using a PGD-projection technique that will enable the computation of a separable expression even for geometrical variations, preserving thus the eﬃciency of the online PGD stage. A set of numerical problems will be presented in order to validate the PGD formula-tion, which will be benchmarked against the full order (reference) model. Moreover, a comparison between two families of ROM methods, the a priori PGD and the a posteriori Proper Orthogonal Decomposition (POD), will also be performed in order to assess and compare diﬀerent ROM strategies.

show abstract

Techniques for Improving Neural Network-based Aerodynamics Reduced-order Models

Cited by 8 publications

References 39 publications

Machine Learning in Aerodynamic Shape Optimization

Machine Learning in Aerodynamic Shape Optimization

Updating multi-fidelity structural dynamic models for flexible wings with feed-forward neural network

Towards MRI scanner design: the Proper Generalised Decomposition in the context of coupled magneto-mechanical problems

Contact Info

Product

Resources

About