Turbulent flow closed-loop control with machine learning – FLOwCON

FlowCon is a project devoted to the development and the experimental demonstration of innovative methods for the closed loop control of turbulent fluid flows using machine learning.

Applications of flow control are ubiquitous, in particular in aerospace, both civilian and military: aerodynamic drag reduction, jets mixing and vectorization, mitigation of noise emission due to impacting vortices, reduction of the vortex-induced vibrations and suppression of the hyper-lift flaps to name only a few examples.

In contrast with most current control strategies, this project aims to develop and offer control methods applicable to operational configurations. In particular, the proposed methods could be applied to systems governed by strongly nonlinear equations. Information on the system to be controlled only comes from a very few, wall-mounted, sensors. The environment of operation is potentially severely noisy and non-stationary (drift over time). Further, controllers must be synthesized and evaluated in real-time (closed-loop) with a very limited computational power, compatible with embedded hardware.

We plan to develop control methods using Machine Learning. The available computational power and the real-time constraint make solving the governing equations of high Reynolds number turbulent flows intractable. Standard control methods rely on the knowledge of the state of the system (state vector). Instead of such a rich information, our approach is to use a statistical description, with no need for a detailed information about the flow, of the map between actions from the actuators and effects onto the flow over time. This class of methods lies in the Machine Learning framework and does not require a model a priori but only information available from sensors (data-driven).

Some members of the present project are experts in flow control and have obtained very encouraging early results with this class of methods (for instance, re-attachment of the turbulent flow behind a descending ramp). However, they have also identified some limitations, such as the time required for learning or the difficulty in guarantying robustness of the control with respect to a drift of the system's dynamics.

While they are key enablers of these early successes, current machine learning methods are not suitable for addressing the identified issues and limitations. The control of a system as complex as a turbulent flow is far from the usual applicability framework of this class of algorithms, which typically includes robotics, language processing, image segmentation, etc. A significant effort, gathering both researchers from the machine learning community, in a wide sense, and experts in flow control, is necessary. As a result, thanks to a close interaction and feedbacks from experiments to the theory, significant innovations are expected for the turbulent flow control community, as well as developments of innovative aspects for the machine learning community.

FlowCon is a highly interdisciplinary project involving, for the first time, researchers from the fluid mechanics community as well as machine learning who will work in close interaction to develop innovative solutions allowing to reach the ambitious goals we have set. These developments will be tested on an experimental open cavity flow experiment of moderate complexity (low Reynolds number) and next on a demonstrator consisting of two configurations as realistic as possible: a turbulent flow behind a descending ramp and a NACA profile with a varying angle of attack at high Reynolds number.

This project is supported by Dassault Aviation (see enclosed letter).