Projet de thèse en Mécanique des fluides
Sous la direction de Luc Pastur et de Bernd Noack.
Thèses en préparation à université Paris-Saclay , dans le cadre de École doctorale Sciences mécaniques et énergétiques, matériaux et géosciences (Gif-sur-Yvette, Essonne ; 2015-....) , en partenariat avec IMSIA - Institut des Sciences de la Mécanique et Applications Industrielles (laboratoire) et de école nationale supérieure de techniques avancées (établissement de préparation de la thèse) depuis le 01-10-2018 .
The starting point of this project are recent successes in machine learning control (MLC) applied to experiments of closed-loop turbulence control in mixing enhancement, reduction of circulation zones, control of wakes behind staggered cylinders, force control of a car model and strongly nonlinear dynamical systems featuring aspects of turbulence. In all cases, a simple genetic programming algorithm has learned the optimal control for the given cost function and out-performed existing open- and closed-loop approaches after few hundred test runs. Yet, there are still numerous opportunities to reduce the learning time by avoiding the testing of similar control laws, to improve the performance measure by generalizing the considered control laws and to design reduced-order models for universal classes of instabilities encountered in fluid flows that could help for effectively designing efficient controllers. The focus of this PhD project is a two-dimensional Navier-Stokes simulation on the newly introduced fluidic pinball, namely a uniform flow around 3 cylinders, as shown in figure 1. The cylinders can be rotated around their axis, providing 3 control inputs. The flow can also be sensed with several downstream velocity sensors, making this flow configuration a non-trivial multiple-input/multiple-output (MIMO) system. A typical goal for the control law is to stabilize the wake and thus reduce the drag. Several strategies were proved to be highly effective in reducing the drag of cylinder wakes flows, namely phasor control, base-bleed, boat-tailing, high-frequency forcing, low-frequency forcing, among others. Most of them involve nonlinear frequency cross-talk, something hardly manœuvrable with linear control strategies. Yet, with its geometric configuration, all these nonlinear control strategies can be applied to the fluidic pinball. Henceforth, although this configuration be geometrically simple to allow testing of thousands of control laws within minutes on a Laptop, it is yet physically complex enough to host a wide spectrum of dynamical regimes and is therefore representative for a wide range of flow instabilities -- beyond the simple fluidic pinball configuration chosen in this project. In addition, the MIMO configuration makes it a challenging configuration for control design in general, and for machine learning control in particular. The fluidic pinball is therefore proposed as a microcosmos of nonlinear dynamics and nonlinear control.
Controlling the fluidic Pinball with machine and human learning
The starting point of this project are recent successes in machine learning control (MLC) applied to experiments of closed-loop turbulence control in mixing enhancement, reduction of circulation zones, control of wakes behind staggered cylinders, force control of a car model and strongly nonlinear dynamical systems featuring aspects of turbulence. In all cases, a simple genetic programming algorithm has learned the optimal control for the given cost function and out-performed existing open- and closed-loop approaches after few hundred test runs. Yet, there are still numerous opportunities to reduce the learning time by avoiding the testing of similar control laws, to improve the performance measure by generalizing the considered control laws and to design reduced-order models for universal classes of instabilities encountered in fluid flows that could help for effectively designing efficient controllers. The focus of this PhD project is a two-dimensional Navier-Stokes simulation on the newly introduced fluidic pinball, namely a uniform flow around 3 cylinders, as shown in figure 1. The cylinders can be rotated around their axis, providing 3 control inputs. The flow can also be sensed with several downstream velocity sensors, making this flow configuration a non-trivial multiple-input/multiple-output (MIMO) system. A typical goal for the control law is to stabilize the wake and thus reduce the drag. Several strategies were proved to be highly effective in reducing the drag of cylinder wakes flows, namely phasor control, base-bleed, boat-tailing, high-frequency forcing, low-frequency forcing, among others. Most of them involve nonlinear frequency cross-talk, something hardly manœuvrable with linear control strategies. Yet, with its geometric configuration, all these nonlinear control strategies can be applied to the fluidic pinball. Henceforth, although this configuration be geometrically simple to allow testing of thousands of control laws within minutes on a Laptop, it is yet physically complex enough to host a wide spectrum of dynamical regimes and is therefore representative for a wide range of flow instabilities -- beyond the simple fluidic pinball configuration chosen in this project. In addition, the MIMO configuration makes it a challenging configuration for control design in general, and for machine learning control in particular. The fluidic pinball is therefore proposed as a microcosmos of nonlinear dynamics and nonlinear control.