Predictive Data-Driven Model Based on Generative Adversarial Network for Premixed Turbulence-Combustion Regimes

ABSTRACT

Premixed flames exhibit different asymptotic regimes of interaction between heat release and turbulence depending on their respective length scales. At high Karlovitz number, the dilatation caused by heat release does not have any relevant effect on turbulent kinetic energy with respect to non-reacting flow, while at low Karlovitz number, the mean shear is a sink of turbulent kinetic energy, and counter-gradient transport is observed. This latter phenomenon is not well captured by closure models commonly used in Large Eddy Simulations that are based on gradient diffusion. The massive amount of data available from Direct Numerical Simulation (DNS) opens the possibility to develop data-driven models able to represent physical mechanisms and non-linear features present in both these regimes. In this work, the databases are formed by DNSs of two planar hydrogen/air flames at different Karlovitz numbers corresponding to the two asymptotic regimes. In this context, the Generative Adversarial Network (GAN) gives the possibility to successfully recognize and reconstruct both gradient and counter-gradient phenomena if trained with databases where both regimes are included. Two GAN models were first trained each for a specific Karlovitz number and tested using the same dataset in order to verify the capability of the models to learn the features of a single asymptotic regime and assess its accuracy. In both cases, the GAN models were able to reconstruct the Reynolds stress subfilter scales accurately. Later, the GAN was trained with a mixture of both datasets to create a model containing physical knowledge of both combustion regimes. This model was able to reconstruct the subfilter scales for both cases capturing the interaction between heat release and turbulence closely to the DNS as shown from the turbulent kinetic budget and barycentric maps.

KEYWORDS:

• Premixed flames
• turbulent combustion modeling
• generative adversarial network
• machine learning

Acknowledgments

The research leading to these results has received funding from the European Union’s Horizon 2020 research and innovation program under the Center of Excellence in Combustion (CoEC) project, grant agreement no. 952181.

Simulations were performed with computing resources granted by RWTH Aachen University under project rwth0658. The authors gratefully acknowledge the computing resources from the DEEP-EST project, which received funding from the European Union’s Horizon 2020 research and innovation program under the grant agreement no. 754304. We thank Mr Sedona for the support in the porting of the application to DEEP-EST.

Disclosure statement

No potential conflict of interest was reported by the author(s).

Additional information

Funding

The research leading to these results has received funding from the European Union’s Horizon 2020 research and innovation program under the Center of Excellence in Combustion (CoEC) project, grant agreement No 952181.