Une nouvelle méthode multi-élément ...
Document type :
Communication dans un congrès avec actes
Title :
Une nouvelle méthode multi-élément non-intrusive basée sur la clusterisation agglorémative capable d'approcher les systèmes dynamiques bifurcatifs
Author(s) :
Vauchel, Nicolas [Auteur]
DAAA, ONERA [Lille]
Gomez, Thomas [Auteur]
Laboratoire de Mécanique des Fluides de Lille - Kampé de Fériet [LMFL]
DAAA, ONERA [Lille]
Gomez, Thomas [Auteur]
Laboratoire de Mécanique des Fluides de Lille - Kampé de Fériet [LMFL]
Conference title :
CFM 2019
City :
BREST
Country :
France
Start date of the conference :
2019-08-26
Keyword(s) :
QUANTIFICATION
INCERTITUDE
MULTI-ÉLÉMENT
CHAOS POLYNOMIAL
META-MODÈLES
INCERTITUDE
MULTI-ÉLÉMENT
CHAOS POLYNOMIAL
META-MODÈLES
English keyword(s) :
UNCERTAINTY QUANTIFICATION
MULTI-ELEMENT
POLYNOMIAL CHAOS
AGGLOMERATIVE CLUSTERING
MACHINE LEARNING
NEURAL NETWORK
MULTI-ELEMENT
POLYNOMIAL CHAOS
AGGLOMERATIVE CLUSTERING
MACHINE LEARNING
NEURAL NETWORK
HAL domain(s) :
Sciences de l'ingénieur [physics]
Physique [physics]
Physique [physics]
English abstract : [en]
A multi-element non-intrusive generalised polynomial chaos method is developed to approximate state variable of a bifurcative dynamical system. It rely on a partitioning of the stochastic space along discon-tinuities of ...
Show more >A multi-element non-intrusive generalised polynomial chaos method is developed to approximate state variable of a bifurcative dynamical system. It rely on a partitioning of the stochastic space along discon-tinuities of any shapes, efficient even when the discontinuities are running partially, and allowing the use of a low degree of approximation. An agglomerative clustering method is implemented in a space defined by the values of the quantity of interest (QOI) and of its derivatives with respect to the stochastic parameters. The stochastic space is first properly sampled. Then, at every instant, several partitions of the stochastic space are tested and the best one is selected as the one minimizing the maximum of the cross-validation errors of the local surrogate models. Once the local models are obtained and the sto-chastic space is correctly sliced, a neural network classifier is learned to determine in which element a sample is for further evaluations at the considered instant. The method uses the same samples for the space partitioning and for the training of the local models with regression thus reducing the computational cost. The long-time integration problem is countered and the stochastic space can be split according to the discontinuity induced by the potential bifurcation. The results of the method are advantageously compared to those obtained with the direct use of a Gradient Tree Boosting algorithm at every instant and the classical Pseudo projection method for a system leading to a bifurcation.Show less >
Show more >A multi-element non-intrusive generalised polynomial chaos method is developed to approximate state variable of a bifurcative dynamical system. It rely on a partitioning of the stochastic space along discon-tinuities of any shapes, efficient even when the discontinuities are running partially, and allowing the use of a low degree of approximation. An agglomerative clustering method is implemented in a space defined by the values of the quantity of interest (QOI) and of its derivatives with respect to the stochastic parameters. The stochastic space is first properly sampled. Then, at every instant, several partitions of the stochastic space are tested and the best one is selected as the one minimizing the maximum of the cross-validation errors of the local surrogate models. Once the local models are obtained and the sto-chastic space is correctly sliced, a neural network classifier is learned to determine in which element a sample is for further evaluations at the considered instant. The method uses the same samples for the space partitioning and for the training of the local models with regression thus reducing the computational cost. The long-time integration problem is countered and the stochastic space can be split according to the discontinuity induced by the potential bifurcation. The results of the method are advantageously compared to those obtained with the direct use of a Gradient Tree Boosting algorithm at every instant and the classical Pseudo projection method for a system leading to a bifurcation.Show less >
Language :
Anglais
Peer reviewed article :
Oui
Audience :
Internationale
Popular science :
Non
Source :
Files
- https://hal.archives-ouvertes.fr/hal-02364686/document
- Open access
- Access the document
- https://hal.archives-ouvertes.fr/hal-02364686/document
- Open access
- Access the document
- https://hal.archives-ouvertes.fr/hal-02364686/document
- Open access
- Access the document
- https://hal.archives-ouvertes.fr/hal-02364686/document
- Open access
- Access the document
- document
- Open access
- Access the document
- DAAA19059.1554299257.pdf
- Open access
- Access the document
- document
- Open access
- Access the document
- DAAA19059.1554299257.pdf
- Open access
- Access the document