{Seminar} @ CDS: 20th March: “Machine learning techniques for extreme events detection/prediction in fluid mechanics”

Extreme events appear in many fluids mechanical systems, such as in atmospheric flows, oceanography, or wind turbines. These extreme events are sudden, unsteady, transient large nonlinear deviation of the flow away from its mean state. All these events are generally accompanied by detrimental and potentially catastrophic consequences. Therefore, the ability to predict such events is of the utmost importance. However, such a task is extremely challenging because of (i) their unpredictability that stems from the complex nonlinear interactions existing within the flows; (ii) their chaotic nature where any infinitesimal perturbation will lead to substantially different evolutions (the so-called butterfly effect) and (iii) their high-dimensionality which makes any data-processing techniques challenging. Despite these difficulties, recent advances in deep learning techniques have enabled advances in the understanding and predictions of such extreme events in chaotic systems and turbulent flows.

 

In this talk, we will present works related to the development of such deep learning-based techniques. Specifically, we will tackle three different aspects. First, we will introduce a clustering-based approach to identify pathway to extreme events in chaotic systems enabling the identification of critical system states. Secondly, we will present nonlinear surrogate modelling techniques applied to turbulent systems that can provide accurate reduced-order models able to predict the onset of extreme events in either thermoacoustic systems or turbulent flows. Finally, the scalability of such techniques to three-dimensional turbulent flows with extreme events will be presented with a proposed multiscale convolutional autoencoder echo state network (CAE-ESN). We will show that it is able to reproduce the chaotic dynamics of a turbulent channel flow, including the statistical occurrence of extreme events and we will introduce some approaches which can support the interpretability of the proposed CAE-ESN.

 

BIOGRAPHY

Dr. (Nguyen) Anh Khoa Doan is an Assistant Professor in AI for Fluid Mechanics at the faculty of Aerospace Engineering at TU Delft. He obtained a dual Master degree in Aerospace Engineering at the Free University of Brussels (ULB) and the Institut Supérieur de l’Aéronautique et de l’Espace (ISAE-SUPAERO) in 2012. In parallel, he also received a Master of Research from the Université de Toulouse III Paul Sabatier. Subsequently, he obtained his PhD in Engineering at the University of Cambridge in 2018 with his research covering the direct numerical simulations of turbulent (reacting) flows and of a novel low-emission combustion concept called MILD combustion. During this period, he also covered research topics related to hydrogen combustion and turbulence/combustion modelling and spent some time as a visiting researcher at the Sandia National Laboratories, USA. After this, he worked as a postdoctoral fellow at the Technical University of Munich at the Institute for Advanced Study and the Mechanical Engineering department from 2018 focusing on the development of Machine Learning and AI-based tools for the prediction and control of extreme events in turbulence and turbulent reacting flows, before joining TU Delft in March 2021.