Closed- loop control of the wake of a road vehicle – COWAVE

In this project, we will realize experiments in wind and water tunnels, numerical simulations and develop control strategies based on Machine Learning. Two cars’ models will be used: the square back bluff body (Ahmed) and a reduced-scale car model. The first will be essentially used in the water tunnel and the second in the wind tunnel. Control strategies will be tested in both configurations by combining fixed or moving flaps and micro jet actuators. Closed-loop control will also be developed in these situations. The most efficient control strategies will be tested on a real car. Experiments will be done in PRISME and at the Pprime institute. LHEEA will take in charge the numerical simulations. Finally, PSA will provide the vision of an automobile manufacturer and help on the use for a real car of the developed control strategies.

Measurements done in a wind tunnel on an Ahmed body and high-performance computations showed promising results. A detailed wake analysis of the basis model allowed an in-depth characterisation of the transfer mechanisms between the external flow and the recirculation bubble behind the body. This mechanism is known to be responsible of a large part of the drag. By comparing the numerical and experimental results, we determined the turbulence model (DDES), hybrid RANS-LES modelisation of the turbulence that leads to the best agreement. An experimental parametric study of the influence of the length and inclination of horizontal and vertical flaps was then performed. Following this parametric study, two configurations were chosen for further analysis using numerical simulations. These configurations were chosen by taking as a compromise the efficiency of the control, on the one hand, and the size of the flaps, on the other hand. We have thus chosen flaps of moderate length that maintain a good efficiency. We have also started experimental and numerical studies on a reduced-size size model of real car (C4). In parallel, work has been done on closed-loop control strategies. After testing several Machine Learning approaches, we choose to use Reinforcement Learning. For the numerical developments, a generic case of wake flow at low Reynolds number was selected: the case of the fluidic pinball. Very encouraging results have been obtained (65% drag reduction from three sensors positioned in the wake).

To characterize the flow wake and determine aerodynamic performances, we will perform experimental wind tunnel studies on the small-scale model and numerical simulations. Experimental studies in an hydraulic tunnel will start soon at the Pprime Institute. They will allow to enlarge the range of Reynolds number studied in this project and to get rid of the scaling problem. The Reinforcement Learning control strategy will be tested in the hydraulic tunnel. The best performing strategies will be applied in the wind tunnel on the Ahmed's body and also on the small-scale model. If we have time, tests will be carried out at the end of the program on a real vehicle.

Publications
1. Bucci M.A., Semeraro O., Allauzen A., Wisniewski G., Cordier L. & Mathelin L., Control of chaotic systems by deep reinforcement learning, submitted. 2019. Accepted in PRSA. Available on arXiv 1906.07672.
2. R. K. Niven, L. Cordier, E. Kaiser, M. Schlegel, B. R. Noack. «Rethinking the Reynolds Transport Theorem, Liouville Equation, and Perron-Frobenius and Koopman Operators ». Submitted to Journal of Physics A: Mathematical and Theoretical

Communications
1. Bucci M.A., Semeraro O., Allauzen A., Cordier L., Wisniewski G. & Mathelin L., Control of a chaotic dynamical system with a deep reinforcement learning approach, 90th GAMM Annual Meeting, Vienna, Austria, February 18 - 22, 2019.
2. Bucci M.A., Semeraro O., Allauzen A., Cordier L., Wisniewski G. & Mathelin L., Control-oriented model learning with a recurrent neural network, APS-DFD: Annual Meeting of the American Physical Society, Atlanta, GA, USA, November 18 - 20, 2018.
3. N. Kumar, F. Kerhervé, L. Cordier, Data Assimilation for control of a plane mixing layer, GDR Contrôle des Décollements, IMFT Toulouse, 8-9 Novembre 2018
4. N. Kumar, F. Kerhervé, L. Cordier, Dynamic reconstruction of a numerical 2D cylinder wake flow using Data Assimilation, 5th Symposium on Fluid-Structure-Sound Interactions and Control (FSSIC2019), Crete , Greece, 27-30 August 2019
5. R. K. Niven, A. Mohammad-Djafari, L. Cordier, M. Abel and M. Quade, Bayesian Identification of Dynamical Systems, Workshop on Bayesian Inference and Maximum Entropy Methods in Science and Engineering (MaxEnt 2019), Garching/Munich, GERMANY, Jun 30-Jul 5, 2019
6. W. ZEIDAN, N. MAZELLIER, E. GUILMINEAU, A. KOURTA, Aerodynamic drag reduction and flow control of a simplified road vehicle, 24ème Congrès Français de Mécanique, 26-30 Août 2019, Brest, France.

Three-dimensional bluff-body wakes are of key importance due to their relevance to the automotive industry. Such wakes contribute to consumption and greenhouse gas emissions. Drastic European Union limitations concerning these two mechanisms conduct the car industry to think about efficient vehicles. In this project, we propose robust drag and fuel reduction solutions for road-vehicles by closed-loop control of turbulent flows working efficiently for a range of operating conditions including changing oncoming velocity and transient side winds. To achieve this goal, we combine passive, active control and closed-loop strategies by using compliant deflectors, unsteady micro-jets and Machine Learning techniques. This project aims to prove a feasibility of the control from laboratory scale up to a full-scale industrial demonstrator. The main repercussions of the project will be on the reduction of the environmental impacts of transport industry and the gain of competitiveness and employment.
This project consists on experiments in wind and water tunnels, numerical simulations and control strategies. Two models will be used, the square back bluff body and a reduced scale car model. The latter is representative of SUV and is inspired from the model used in collaborative work between POAES and PRISME. Control strategies will be tested in both configurations by combining passive and active actuation i.e. fixed or moving flaps and micro jet actuators. Closed-loop control will also be developed in these situations. Control strategies will be mainly developed by PPRIME. Experiments will be done in PRISME and PPRIME. LHEEA will take in charge numerical simulations and optimization. Finally, PSA will provide the vision of an automobile manufacturer on the industrial feasibility of the developed control strategies.