Design of closed-loop control strategies for fluid flows using deep neural network surrogate models

View Video Presentation: https://doi.org/10.2514/6.2022-3635.vid

We present results for three different closed-loop control techniques that can be implemented from neural network surrogate models. These differentiable models with control inputs are trained with data obtained from open-loop controlled tests. The control approaches are implemented in order to stabilize three nonlinear dynamical systems: a predator-prey population model, the Lorenz system using parameters that enable chaos, and the flow past an airfoil, simulated by a high-fidelity CFD tool. The first technique presented is model predictive control, as literature supports it as a good approach in different flow control problems. The second one implements a NN model as a controller with nonlinear dynamics to overcome the need of running an optimization problem at every control iteration. Finally, the constructed surrogate models are linearized by computing the Jacobian matrix of the surrogate models to enable the design of linear controllers using well established techniques. All three approaches present good results effectively stabilizing the proposed low dimensional models. The linearized surrogate model is also tested for controlling the airfoil problem and is able to attenuate flow oscillations along the wake.