Eventos Anais de eventos
EPTT 2024
14th Spring School on Transition and Turbulence
Neural Networks as Flow Controllers: a Study on Robustness
Submission Author:
Tarcísio Déda , SE , Brazil
Co-Authors:
Tarcísio Déda, William Wolf, Scott Dawson
Presenter: Tarcísio Déda
doi://10.26678/ABCM.EPTT2024.EPT24-0052
Abstract
We explore the application of neural networks for developing surrogate models and controllers aiming for closed-loop stabilization. Leveraging a grid of velocity sensors in a confined cylinder wake, convergence to equilibrium is demonstrated. The methodology, {developed in previous work,} involves training neural network surrogate models (NNSMs) that are leveraged for designing neural network controllers (NNCs). During the training of NNSMs, sensor selection minimizes the number of probes required by implementing a sparsity layer {within the neural network architecture}. Equilibrium estimation, required for providing a control setpoint, is achieved through the Newton method applied to the NNSMs. The networks are trained iteratively, with progressive improvements as more data near equilibrium becomes available. {The present work focuses on analyses} conducted to assess the robustness of these trained NNCs. First, we show the main results involving a series of tests conducted with the Lorenz system, in which we verify the behavior of the closed-loop system subject to phenomena such as plant variations, measurement noise and disturbances. A controller is then trained to stabilize a confined cylinder flow at Reynolds number $\mathrm{Re} = 150$. We subsequently demonstrate that the controller is able to stabilize unstable flows at lower and higher Reynolds numbers than that for which the NNC was trained. {As the Reynolds number increases, we find that reducing the timestep at which control is applied can be required to achieve complete stabilization.
Keywords
flow control, machine learning, Nonlinear control
