Modélisation du plissement de sous-maille par apprentissage automatique pour les simulations aux grandes échelles de la combustion turbulente prémélangé

Large Eddy Simulation (LES) is increasingly being used for the design of new combustion systems, particularly decarbonized ones. In the case of premixed combustion, part of the flame wrinkling is not resolved on the LES grid and requires specific models. These models are traditionally algebraic, based on analytical closures, and dynamic models, based on filtering operations of the LES field. However, each of these models has its limitations: algebraic models are not very generalizable, and dynamic models do not allow predictions at large filter sizes. An emerging alternative is the use of Machine Learning-based models, particularly neural networks. Built from DNS data, these models have significant predictive power. Many studies have demonstrated the performance of convolutional neural networks in predicting filtered reaction rates. However, the models developed so far are designed for a fixed and relatively small filter size, in order to construct a DNS database in a reasonable amount of time. This significantly limits their application to industrial cases, which typically operate at variable filter sizes. The objective of this thesis is to propose a methodology for generating neural networks that can operate at variable and potentially large filter sizes. To achieve this, it is proposed to build a model at a fixed filter size and to develop an analytical scaling model for larger filter sizes. This method will first be tested on simple canonical cases, such as flames evolving in homogeneous turbulence, and then applied at the end of the thesis to a more complex case.