Apprentissage artificiel de modèles réduits physiques pour la prévision rapide de la durée de vie d'aubes de turbine avec quantification d'incertitudes

par Thomas Daniel

Projet de thèse en Mécanique

Sous la direction de David Ryckelynck.

Thèses en préparation à Paris Sciences et Lettres , dans le cadre de Ingénierie des Systèmes, Matériaux, Mécanique, Énergétique , en partenariat avec ENSMP MAT. Centre des matériaux (Evry, Essonne) (laboratoire) , MAT- Simulation des matériaux et des structures - SIMS (equipe de recherche) et de École nationale supérieure des mines (Paris) (établissement de préparation de la thèse) depuis le 01-12-2018 .


  • Résumé

    As in the vast majority of industrial domains, the numerical simulation is a tool used in many stages of Safran's activities. The complexity of the models leads to computation times of several hours (or even days) for a single run, although optimization and uncertainty quantifications require many runs. Hence, new strategies must be found. In this thesis, we are interested in the lifetime prediction of high-pressure turbine blades in aircraft engines. The computation chain contains a coupled aerothermal fluid-solid computation and an ElastoViscoPlastic (EVP) cyclic computation of which the lifetime calculation is a post-treatment (see Figure 1). Boundary conditions for the EVP part are provided by the aerothermal computation. The objective is to increase the speed of the EVP cyclic part by constructing a dictionary of reduced order models using machine learning tools. The idea is to use neural networks in order to identify a hyper-reduced model which is adapted to the boundary conditions. To do so, a clustering algorithm will be applied to a collection of numerous low-fidelity computations so as to build a dictionary of possible models for lifetime predictions. Then, an efficient criterion based on classification methods taken from data science will be chosen to decide in real time which hyper-reduced model is the best for fast and accurate predictions. Some theoretical questions will also be addressed, such as the unicity of the stabilized cycle, the conditions for the reduced models to converge towards the stabilized cycle, etc. Finally, the methodology will be applied to an uncertainty quantification study on the lifetime computation of high-pressure turbine blades, for which loading cycles are not accurately known despite having an important influence.

  • Titre traduit

    MACHINE LEARNING OF PHYSICAL REDUCED ORDER MODELS FOR FAST LIFETIME COMPUTATION OF TURBINE BLADES WITH UNCERTAINTY QUANTIFICATION


  • Résumé

    As in the vast majority of industrial domains, the numerical simulation is a tool used in many stages of Safran's activities. The complexity of the models leads to computation times of several hours (or even days) for a single run, although optimization and uncertainty quantifications require many runs. Hence, new strategies must be found. In this thesis, we are interested in the lifetime prediction of high-pressure turbine blades in aircraft engines. The computation chain contains a coupled aerothermal fluid-solid computation and an ElastoViscoPlastic (EVP) cyclic computation of which the lifetime calculation is a post-treatment (see Figure 1). Boundary conditions for the EVP part are provided by the aerothermal computation. The objective is to increase the speed of the EVP cyclic part by constructing a dictionary of reduced order models using machine learning tools. The idea is to use neural networks in order to identify a hyper-reduced model which is adapted to the boundary conditions. To do so, a clustering algorithm will be applied to a collection of numerous low-fidelity computations so as to build a dictionary of possible models for lifetime predictions. Then, an efficient criterion based on classification methods taken from data science will be chosen to decide in real time which hyper-reduced model is the best for fast and accurate predictions. Some theoretical questions will also be addressed, such as the unicity of the stabilized cycle, the conditions for the reduced models to converge towards the stabilized cycle, etc. Finally, the methodology will be applied to an uncertainty quantification study on the lifetime computation of high-pressure turbine blades, for which loading cycles are not accurately known despite having an important influence.