Physics-Informed Deep Learning for Turbulent Reacting Flows

by Anh Khoa Doan, Wolfgang Polifke and Luca Magri

Motivation

Air transportation is anticipated to double over the next couple of decades, which calls for novel methods to cut back on up to 80% in oxides of nitrogen (NOx) and up to 50% in noise, as set by the Advisory Council for Aerospace Research. To develop new clean aircraft engines, gas turbines are designed to burn in a lean regime to reduce NOx emissions. The downside is that lean flames burn very unsteadily due to their sensitivity to the turbulent environment of the combustion chamber. In this complex multi-physical environment, three main physical subsystems can be identified: (i) acoustics, (ii) aerodynamics, and (iii) flame dynamics (chemical reaction). These subsystems interact with each other contributing differently, but fundamentally to engine noise; thermo-acoustic instabilities; and rare and extreme events such as flashback or blowoff. If uncontrolled, these instabilities can lead to the catastrophic failure of the engine. It becomes, therefore, crucial to predict and control those extreme events.

This project explores the fundamental interactions between these phenomena using innovative machine-learning based algorithms and aims to develop a new artificial intelligence framework which synergetically combines physical first principles and data-driven approaches for the prediction and control of these instabilities.

Objectives and Strategy

This project uses data-driven methods, such as neural networks and reservoir computing, enhanced with physical knowledge of the systems, such as conservation laws, for the prediction of these extreme events in turbulent reacting flows. In particular, dataset of simulations of turbulent flows and gas turbines obtained with conventional Computational Fluid Dynamics (CFD) and experimental measurements are used to train these deep learning networks which are then also physics-constrained to ensure the physicality of their predictions. More specifically, the project has two main components:

Modelling Turbulent Flows using Physics-Informed Deep Learning: The goal is to develop physics-informed neural networks which can act as digital twin of the real turbulent systems. To ensure the accuracy of these networks and the physicality of their predictions, they are trained by combining existing simulations and the physical knowledge, in the form of conservation equations that the system must respect. The available dataset provides a general information on the dynamics of the combustors while the physical information enhances the training information used by the network, allowing it to infer behaviours outside the regime present in the training dataset. This will allow the physics-informed neural network to accurately reproduce the long- and short-term statistics of the combustors.

Deduce extreme events and sensitivity characteristics from the physics-informed data-driven model: From the robust model obtained at the previous stage, the goal is to deduce the behaviour of the gas turbine to external perturbations and the probability of extreme events (thermoacoustics instabilities, flashback, blow-off, ...) to occur. This can be obtained by deducing the sensitivity from our physics-informed machine-learning model using adjoint methods. Ultimately, the objective is for these physics-informed networks to infer the behaviour of gas turbines for some given operating conditions or under some prescribed perturbations and to identify potential critical situations. Using such a network will permit to avoid the expensive experimental campaigns normally used to assess the stability of gas turbines.

Acknowledgement