Data driven estimation of fluid flows: long-term prediction of velocity fields using machine learning
Résumé
This paper gives a framework for the data-driven estimation of an unsteady fluid flow field. The strategy combines machine learning tools for the reduction, the reconstruction and the prediction of the considered system. The reduction is performed by linear autoencoding while support vector regression and dynamical mode decomposition are respectively used as reconstruction and prediction models. Starting from an initial condition, reconstructions are frequently assimilated to update erroneous predictions. The procedure is tested on four cases with increasing complexity and robustness is assessed through training and testing errors. Quantitative results suggest that reconstruction and prediction models purely learnt from data can be used for effective data assimilation, hence enabling the long-term prediction of even complex fluid flows.
Origine : Fichiers produits par l'(les) auteur(s)