Machine learning control and modeling -How taming turbulence can be made easy, efficient, fast and fun!
Résumé
Closed-loop turbulence control has current and future engineering applications of truly epic proportions, including cars, trains, airplanes, jet noise, air conditioning, medical applications, wind turbines, combustors, and energy systems, i.e., well-known topics in the GDR 2502. A key feature, opportunity and technical challenge is the inherent nonlinearity of the actuation response [1]. For instance, excitation at a given frequency will affect also other frequencies. This frequency cross-talk is not accessible in any linear control framework. Recently, Artificial Intelligence (AI) / Machine Learning (ML) has opened game-changing new avenues [2]: the automated model-free discovery and exploitation of unknown nonlinear actuation mechanisms directly in the plant and the automated reduced-order modeling from these data. In this talk, we review recent successes on these avenues for broadband frequency turbulence with distributed actuators. Methodological advances include (1) a ML response model predicting performance increases by actuation [3]. (2) the cluster-based network model for automated robust identification of coherent-structure dynamics [4], (3) the explorative gradient method for actuation optimization with the convergence rate of a gradient method and an exploration of global minima [6], and, last but not least, (4) a novel fast-learning gradient-enriched machine learning control which optimizes MIMO feedback laws [5]. Thus, we achieve: (1) 31% drag reduction of a turbulent boundary layer with spanwise traveling surface waves [3, 4], (2) 17% drag reduction of a slanted Ahmed body with 5 groups of orientable actuation jets [6], (3) a significant increase of mixing of a turbulent jet with a novel distributed unsteady actuation [7] and (4) an understanding of the coherent structure dynamics. Nan Deng and Guy Cornejo Maceda will elaborate recent advances for the fluidic pinball during in this meeting.