Deep learning strategies applied to closed-loop control of unsteady flows

Abstract

The present proposal describes the research plan associated to the research internship (BEPE) of PhD student Tarcísio Costa Déda Oliveira. Tarcísio will develop his research in the Mechanical, Materials, and Aerospace Engineering Department at the Illinois Institute of Technology, Chicago, Illinois, USA. During this period, he will work under the supervision of Prof. Scott Dawson, who has extensive expertise in the fields of fluid mechanics, dynamical systems, control theory, and data science, which are topics related to the current research proposal. Our goal is to apply deep learning techniques to model complex nonlinear flow systems as well as to develop control strategies assisted by deep neural network models. Surrogate models can enable the design of control techniques that, in turn, can be tested in high-fidelity numerical environments for validation. The applications of closed-loop flow control encompass different objectives, such as lift enhancement, drag reduction, acoustic noise reduction and transition delay. In this study, we will apply neural networks to be trained as dynamic surrogate models of fluid flow systems with control inputs. Data from a high-fidelity solver is obtained to train reduced order models that can enable quick control tests and allow for model-based design of controllers. In this work, the control objective will consist in the attenuation of unsteady characteristics of the studied flows either by stabilizing systems that present oscillatory behavior due to nonlinearities leading to limit cycles; or by counteracting perturbations that grow in the flow field. (AU)