Skip to main content
To KTH's start page To KTH's start page

Data-driven methods for analysis and control of fluid flows: Deep learning and reinforcement learning strategies

Time: Thu 2020-05-14 10.30 - 11.30

Location: Faxén, FPL, Teknikringen 8

Participating: Onofrio Semeraro

Export to calendar

Abstract. In recent years, statistical learning by use of data-science tools has been boosted by the increased computational power and the availability of large amount of high quality data. In this talk, we propose a brief overview on how these technique can be integrated within fluid mechanics by introducing some of the basic ideas behind the machine learning. The aim is provide elements for a demystifying perspective on the properties of these tools and the extent to which data-enhanced analysis can be effective in exploring and controlling complex dynamics with respect of the current state-of-art.

As reference case, we consider a 1-D equation, the Kuramoto-Sivashinsky equation, in the chaotic regime. This nonlinear system is well documented and historically used for the analysis of front flames; from the dynamical point of view, it shares some features that are typical of turbulence and as such is a rather convenient test case before attacking full-scale fluid mechanics problems. During the talk, we will focus on the flow control problem, a rather long standing research area in fluid mechanics. Within the Reinforcement learning framework, we will show how is possible to approximate the nonlinear optimal control problem associated to the Hamilton-Jacobi-Bellman equation. The possibility of applying these data-driven algorithms by solely relying on local measurements for the control via localised actuation will be discussed.