Thomas Duriez, N. Gautier, Bernd R. Noack, Carine Fourment, Steven L. Brunton, Markus Abel, Marc Segond, Jean-Charles Laurentie, Jean-Paul Bonnet, Joel Delville, Michel Stanislas, Laurent Cordier, Vladimir Parezanovic, Jean-Luc Aider, C. Raibaudo, Christophe Cuvier, Institut Pprime (PPRIME), Université de Poitiers-ENSMA-Centre National de la Recherche Scientifique (CNRS), Ambrosys GmbH, Potsdam, Germany, Physique et mécanique des milieux hétérogenes (UMR 7636) (PMMH), Ecole Superieure de Physique et de Chimie Industrielles de la Ville de Paris (ESPCI Paris), Université Paris sciences et lettres (PSL)-Université Paris sciences et lettres (PSL)-Sorbonne Université (SU)-Centre National de la Recherche Scientifique (CNRS)-Université de Paris (UP), Laboratoire de Mécanique de Lille - FRE 3723 (LML), Université de Lille, Sciences et Technologies-Ecole Centrale de Lille-Université de Lille-Centre National de la Recherche Scientifique (CNRS), University of Washington [Seattle], ENSMA-Centre National de la Recherche Scientifique (CNRS)-Université de Poitiers, and Université de Lille, Sciences et Technologies-Centrale Lille-Centre National de la Recherche Scientifique (CNRS)
We propose a novel closed-loop control strategy of turbulent flows using machine learning methods in a model-free manner. This strategy, called Machine Learning Control (MLC), allows – for the first time – to detect and exploit all enabling nonlinear actuation mechanisms in an un-supervised automatic manner. In this communication, we focus on MLC applications for in-time control of experimental shear flows and demonstrate how it outperforms state-of-the-art control. In particular, MLC is applied to three different experimental closed-loop control setups: (1) the TUCOROM mixing layer tunnel, (2) the Gortler PMMH water tunnel with a backward facing step, and (3) the LML Boundary Layer wind tunnel with a separating turbulent boundary layer. In all three cases, MLC finds a control which yields a significantly better performance with respect to the given cost functional as compared to the best previously tested open-loop actuation. We foresee numerous potential applications to most nonlinear multiple-input multiple-output (MIMO) flow control problems, particularly in experiments. In particular, the model-free architecture of MLC enables its application to a large class of complex nonlinear systems in all areas of science.