Conditional scalar dissipation rate modeling for turbulent spray flames using artificial neural networks

Abstract Machine learning techniques have been used for the closure of the conditional scalar dissipation rate in turbulent spray flames. Statistical data are extracted from carrier-phase direct numerical simulation (CP-DNS) results for the generation of artificial neural networks that are trained to predict the conditional scalar dissipation rate such that they could serve as a sub-grid model for large eddy simulations of spray combustion. A quantitative error assessment suggests that predictions of the conditionally averaged dissipation rate are in excellent agreement with CP-DNS data. Further comparison with commonly used models for the unconditionally filtered dissipation rate promises significant improvements if ANNs are used for closure. The artificial neural networks also help to identify the important features that affect the local dissipation rates. The results suggest that few - mostly droplet related - parameters suffice as input features for accurate ANNs. This is in contradiction to standard modeling techniques that are solely based on gas phase properties and highlights the need to revisit scalar dissipation rate modeling for spray flames if analytical expressions are to be used.