Deep Reinforcement Learning in Fluid Mechanics: application to Active Flow Control
Tid: To 2020-01-30 kl 10.30 - 11.30
Plats: Faxen, FPL, Teknikringen 8
Medverkande: Jean Rabault (Oslo University)
Abstract. Fluid Mechanics features many challenging problems, due to the combination of non-linearity and high dimensionality that arises from the Navier Stokes equations. Such problems include, to name but a few, Active Flow Control (AFC) and Optimal Design (OD). While both topics have been the focus of much research over the last 40 years, traditional methods based on linearization of the system and gradient descent face many challenges when applied to Fluid Mechanics problems. Therefore, new tools may be required to provide further improvements on those topics.
In this talk, we will offer a brief introduction to the main ingredients behind Artificial Neural Networks (ANNs) and Deep Reinforcement Learning (DRL), and give a high-level overview of how such methods work. Following this, we will present recent results of DRL in AFC. In particular, we will discuss why recent milestones achieved in AFC are promising for real-world applications, and how these compare to more traditional methods such as the adjoint method. Finally, we will provide a brief overview of available resources, including Open Source code packages and example code showcasing recent applications, that may help people new to the field to start working on such topics.