LES models for turbulent hydrogen flames with convolutional neural networks

Attili, A., Jansen, M.G. D, Sorace, Nicola, Bruce, M., Grenga, Temistocle, Nista, Ludovico, Berger, Lukas and Pitsch, Heinz (2023) LES models for turbulent hydrogen flames with convolutional neural networks. In Proceedings of the 45th Meeting of the Italian Section of the The Combustion Institute. Associazione Sezione Italiana del Combustion Institute. pp. 189-194 .

Abstract

Lean hydrogen flames are prone to thermodiffusive instabilities, which have a strong effect on the structure and dynamics of the flame and can enhance flame speed by several times. Conventional combustion models perform poorly for unstable hydrogen flames and often fail in capturing the effects of thermodiffusive instabilities. In this work, the capability of Convolutional Neural Networks (CNN) to model the unclosed reaction rate term in Large Eddy Simulations (LES) of turbulent hydrogen premixed flames is investigated using high-fidelity data from a large-scale Direct Numerical Simulation (DNS). It is shown that the CNN model can accurately reproduce the filtered reaction rate over a large range of filter sizes. Traditional models usually require at least two scalars, e.g., two progress variables or a progress variable and a mixture fraction, to capture the local fluctuations of equivalence ratio caused by thermodiffusive effects; remarkably, the CNN-based model requires only a single progress variable, due to its ability to consider the topology of the three-dimensional progress variable field, which embeds the information regarding the fluctuations of equivalence ratio. Finally, the capability of the CNN to generalize to different filter sizes and filter kernels is investigated.

This record has no associated files available for download.

Contributors

Author: Temistocle Grenga

Download statistics