Predicting extreme events before they occur with scientific machine learning

Context and overall goal

Climate change and the push to decarbonize society are making extreme events in fluids more common. These events are rare instances where the flow suddenly shifts to extreme states, far from its usual behavior. Such events can occur in various flow systems, including the atmosphere, where atmospheric blocking leads to severe heat waves; the oceans, where rogue waves (extremely large waves) can capsize ships; and in engineering systems such as hydrogen-based clean combustors, where flashback occurs and the flame unexpectedly moves back into the injection system.

At present, accurately predicting these extreme events is challenging due to several obstacles. First, the chaotic nature of turbulent flows makes them difficult to forecast; even the smallest disturbance can cause radically different outcomes (the butterfly effect). Second, extreme events result from complex nonlinear interactions, which differ across systems with varying physical mechanisms, making it hard to apply insights from one system to another. Third, there is a lack of sufficient data on these events.

To revolutionize our approach to extreme events, our Focus Group aims to develop an advanced scientific machine learning framework that synergistically combines deep learning with physics-based methods, thereby enabling the prediction of extreme events in turbulent flows. Such a hybrid physics / machine learning-based framework will overcome the normally data-hungry nature of machine learning tools in the context of the rarity of extreme events in any dataset.

Summary of work and major outcomes

During this TUM-IAS Fellowship, we pioneered the development of physics-informed reservoir computer methods for the time-accurate prediction of extreme events in turbulent flows. We showed that a combination of physical knowledge with a specific type of reservoir computer, called an echo state network, could provide a longer time horizon in the prediction of extreme events in turbulent flows [1]. Furthermore, to handle high-dimensional flows, we developed a new deep learning architecture that combined a convolutional autoencoder with an echo state network to learn and predict the dynamics of turbulent flows (see Fig. 2) [2].

In parallel to those activities dedicated to the prediction of extreme events, we also developed data-driven techniques and applied them to reacting flows (i.e., flows where chemical reactions play an important role, such as in gas turbines). Deep learning techniques were developed to learn the dynamics of reacting flows, in the context of thermoacoustic instabilities prevention. We showed that neural networks could learn the essential dynamics of a flame and be embedded in a reduced-order acoustic model of the combustor to accurately predict the onset of thermoacoustic instabilities when design changes to the combustor are performed [3].Furthermore, we also developed data assimilation techniques to improve the accuracy of coarse simulations, demonstrating how a judicious choice of assimilated quantities from high-fidelity simulations could enable coarse ones to accurately predict ignition kernels.[4]

To provide a platform for the dissemination of our results and create opportunities for research cross-fertilization, our Focus Group organized SoTiC 2021 – Symposium on Thermoacoustics in Combustion: Industry meets Academia in September 2021. This was a five-day symposium, held online due to Covid, with five keynote presentations from both industry and academia and 64 other presentations. Selected papers from the symposium were published in a special issue of the International Journal of Spray and Combustion Dynamics.

Future research

The research achieved during this TUM-IAS Fellowship is paving the way for the future control and prevention of extreme events in turbulent flows in engineering applications. Building on the ability to predict their occurrence, further research will be devoted to developing machine learning-based techniques to identify optimal control strategies that can effectively mitigate these extreme events in practical engineering applications, such as the prevention of flashback in hydrogen-powered gas turbines, and to improve prediction of extreme climate events in weather forecasting. 

In close collaboration with Prof. Thomas F. Sattelmayer (TUM) and Prof. Mirko R. Bothien (Zurich University of Applied Sciences, Alumnus Rudolf Diesel Industry Fellow).

Selected publications

• Doan, N.A.K., Polifke, W. & Magri, L. Short- and long-term predictions of chaotic flows and extreme events: a physics-constrained reservoir computing approach. Proceedings of the Royal Society of London A, 477, 20210135 (2021).
   
• Doan, N.A.K., Polifke, W. & Magri, L. Auto-encoded reservoir computing for turbulence learning. In: Paszynski, M., Kranzlmüller, D., Krzhizhanovskaya, V.V., Dongarra, J.J., Sloot, P.M. (eds) Computational Science - ICCS 2021. ICCS 2021. Lecture Notes in Computer Science, 12746, 344–355. Springer, Cham (2021).
   
• Magri, L., & Doan, N.A.K. Physics-informed data-driven prediction of turbulent reacting flows with lyapunov analysis and sequential data assimilation. In: Pitsch, H., Attili, A. (eds) Data Analysis for Direct Numerical Simulations of Turbulent Combustion, Springer, Cham (2020).
   
• Doan, N.A.K., Polifke, W. & Magri, L. Physics-informed echo state networks. Journal of Computational Science, 47, 101237 (2020).

Full list of publications