Data-driven subfilter modelling of thermo-diffusively unstable hydrogen–air premixed flames

Abstract

This article is dedicated to Moshe Matalon on the occasion of his 70th birthday, for his numerous contributions to the field of combustion and, in particular, to the rich and varied topic of premixed flame stability. Here, we follow in his footsteps and propose a subfilter modelling framework for thermo-diffusively unstable premixed flames, such as lean hydrogen-air flames. Performing an optimal estimator analysis for the unfiltered and filtered heat release rate of the lean premixed hydrogen-air flames, the latter is found to require at least two scalars for an appropriate representation while for large filter sizes, the heat release appears to require only one scalar for parametrisation. As a result, we develop a modelling strategy based on the construction of thermochemical tables for each unclosed term as a function of two variables as well as the filter size. The framework is based on the filtered tabulated chemistry approach, where, in lieu of a one-dimensional unstretched flame, we adopt a data-driven paradigm and filter fully resolved two-dimensional simulations of variable size. Models originating from small-and medium-sized simulations are tested a-priori on a large-size simulation, thus highlighting the role of the lateral domain in the dataset used for tabulation. The concept of a minimum domain size is thus discussed, leading to a dataset exhibiting the minimal properties for sufficiently accurate thermochemical tables. The strategy is shown to be more accurate than a classical one-dimensional filtered tabulated chemistry approach and shows promise in future LES modelling of laboratory and industrial scale hydrogen flames.

Co-optimized machine-learned manifold models for large eddy simulation of turbulent combustion

“…Additionally, Yellapantula et al demonstrated that an ANN of equivalent form with ξ i = (Z, C) could be fit to filtered data from a DNS of a stratified turbulent premixed flame to develop a model for filtered reaction rates in that flame [48] . Lapenna et al showed a similar result for thermodiffusively unstable flames, using an optimal estimator analysis to choose manifold parameters and replacing the variance with the filter width [60] . Therefore, the CMLM approach uses this form for subfilter closure as well.…”

“…In essence, optimization of the neural network identifies manifold parameterizations that are tuned to not only provide low error predictions of the outputs, but also to allow effective subfilter closure based on the variances of the manifold parameterizing variables. Conceptually, the result is equivalent to basing subfilter closure on filtered one-dimensional flames, as in the M-FFGM or filtered tabulated chemistry for large eddy simulation (F-TACLES) approaches [60] , but it can be applied to more complex data sets, including those requiring additional parameters beyond those typically used for one-dimensional flames.…”

Comparison of Flame Propagation Statistics Based on Direct Numerical Simulation of Simple and Detailed Chemistry. Part 2: Influence of Choice of Reaction Progress Variable

“…The ever-growing computing power is an enabler for two trends in recent combustion DNS research: Firstly (not discussed here but underlining the importance of combustion DNS), the availability of high-fidelity datasets has given a significant boost to the development of turbulence and combustion models using machine learning methods [5] and secondly detailed chemical kinetic mod-els for larger and larger hydrocarbon and oxygenated hydrocarbon molecules are being produced [6]. While practical fuels for transportation often contain thousands of distinct chemical compounds [7], there is also an increasing interest in simple fuels such as hydrogen [8,9] or syngas [10,11] that could play an important role during the transition of the fuel landscape. Their main components such as H 2 , CO and CH 4 are conceptually at the base of detailed chemical reaction mechanisms that describe much more complex hydrocarbon combustion [12].…”

Gas Turbine Combustion Technologies for Hydrogen Blends

“…The strategies employed to avoid or reduce its effects are substantially based on empirical methods. In addition, lean H 2 flames are highly thermo-diffusively unstable, which means that wrinkles in the flame will tend to become more pronounced, thus justifying the higher turbulent flame speed of lean H 2 flames compared to other lean mixtures [48].…”