Machine learning tabulation of thermochemistry: Transfer learning for non-adiabatic flames

Artificial Neural Networks (ANNs) are a powerful tool to accelerate the calculation of the thermochemistry source term in turbulent reacting flow simulations. However, their capacity for transfer learning, i.e. their ability to predict flames different from the problems they were trained on, depends on the way training data are generated. In the Hybrid Flamelet/Random Data (HFRD) approach (Ding et al., Combust. Flame 231:111493, 2021), data are generated with a generic approach that combines canonical laminar flames with randomisation, and the resulting dataset enables training ANNs that can generalise to wide families of turbulent flames. However, in many applications such as gas turbines, non-adiabatic processes such as wall heat loss result in an extended composition space. In the present work, we introduce a temperature-profile-control strategy for data generation that enlarges the HFRD composition space coverage by employing a set of one-dimensional burner-stabilised flames with specified temperature profile at various flow rates and mixture fractions. The resulting ANNs are applied first to a one-dimensional freely-propagating CH 4 /H 2 -air premixed flame, to demonstrate transfer to a laminar flame problem with a different fuel inlet. Subsequently, the overall approach is applied to the simulation of the model gas turbine combustor PRECCINSTA with a Large Eddy Simulation - Probability Density Function (LES-PDF) method. It must be noted that the machine learning approach is applicable to any turbulence-chemistry interaction model involving real-time integration of kinetics. Results from the ANN-accelerated simulations show very good agreement with simulations involving direct integration of the kinetics. • New machine learning method for thermochemistry tabulation in non-adiabatic flows. • Enhanced coverage of composition space is demonstrated. • Training data generation is shown to reduce ANN prediction errors for all species. • Application to PRECCINSTA burner shows very good agreement with experiments.