A Hybrid Machine Learning approach for Flamelet Modelling of Turbulent Reacting flows

Simulation of turbulent reacting flows requires accurate modeling of complex chemical reactions and their interactions with the turbulent flow field. One of the common approaches to reduce the computational cost of reacting flow simulations is the use multidimensional lookup tables. One such approach is the Unsteady Flamelet Progress Variable (UFPV) approach. While these UFPV tables have demonstrated effectiveness in accurately representing turbulent reacting flows under certain regimes, they are computationally expensive due to the necessity of storing large, multidimensional datasets and performing complex interpolation schemes. This leads to significant memory demands and limitations imposed by the curse of dimensionality, impacting their efficiency and scalability. In response to this limitation, we propose a hybrid machine learning (ML) approach to replace the use of multidimensional UFPV tables, thereby reducing memory requirements and eliminating the need for interpolation schemes. The proposed ML model is validated through a Large Eddy Simulation (LES) of turbulent CH4/air reacting flow, confirming the model’s accuracy. By using this hybrid ML approach, we not only reduce the memory footprint associated with large multidimensional tables but also enhance the model’s scalability and flexibility in handling complex chemical kinetics.