AI meets turbulence: Lagrangian and Eulerian data-driven tools for optimal navigation and data-assimilation
| Quand ? |
Le 04/04/2022, de 11:00 à 12:00 |
|---|---|
| Où ? | Salle Condorcet (1 place de l'Ecole) |
| Participants |
Luca Biferale |
| Ajouter un événement au calendrier |
vCal
iCal
|
We examine the applicability of Artificial Intelligence tools to different open problems in fluid dynamics, from the search for an optimal navigation strategy in complex environments to data reconstruction from partial measurements of turbulent flows. To solve navigation problems we follow a Reinforcement Learning (RL) approach. Here, we will focus on the problem of finding the path that minimizes the navigation time between two given points in a fluid flow. I will show how RL is able to take advantage of the flow properties in order to reach its target, providing stable solutions with respect to perturbations on the initial conditions and to addiction of external noise. These results illustrate the potential of RL algorithms to model adaptive behavior in real/complex flows and pave the way towards the engineering of smart unmanned autonomous vehicles. The search for optimal navigation strategies is key in several applications, with a potential breakthrough in the open challenge of Lagrangian data assimilation (DA). In the DA direction, we also explore the capability of Generative Adversarial Network (GAN) to generate missing data. In this direction, I will present a quantitative investigation of their potential in reconstructing 2d damaged snapshots extracted from a large numerical database of 3d turbulence in the presence of rotation. I will briefly compare GAN with different, well-known, data assimilation tools, such as Nudging, an equation-informed protocol, or Gappy POD, developed in the context of image reconstruction. I will discuss how one can use DA tools with a reverse engineering approach, to investigate theoretical questions like which features of the input flow data are required/"more important" in order to obtain a better full-field reconstruction.