Skip to Main content Skip to Navigation

Machine learning for nonlinear model order reduction

Abstract : Uncertainty quantification in computational physics requires running many simulations. For some industrial applications, the complexity of the numerical model is incompatible with the number of simulations to be run. Solving physics equations in a reduced computation time is therefore essential for the design of safe and reliable systems. In this thesis, we propose a new numerical method combining model order reduction and machine learning to compute an approximate solution of a stochastic partial differential equation in a reasonable computation time. With this method, the mechanical behavior of an aircraft engine’s component is predicted 636 times faster than with the original high-fidelity model with less than 3% of errors, which enables quantifying the uncertainties generated by the thermal loading.
Complete list of metadata
Contributor : ABES STAR :  Contact
Submitted on : Friday, January 14, 2022 - 8:15:08 AM
Last modification on : Friday, January 21, 2022 - 10:05:07 AM
Long-term archiving on: : Friday, April 15, 2022 - 6:18:46 PM


Version validated by the jury (STAR)


  • HAL Id : tel-03525700, version 1


Thomas Daniel. Machine learning for nonlinear model order reduction. Mechanics of materials [physics.class-ph]. Université Paris sciences et lettres, 2021. English. ⟨NNT : 2021UPSLM039⟩. ⟨tel-03525700⟩



Record views


Files downloads