JPResUnet: A Joint Probability Density Function Translation Model in Partially Premixed Flames

Abstract

Machine learning (ML) models are often constrained by their limitations in extrapolation, which restricts their applicability in engineering contexts. Conversely, while exhibiting broad generality, many established scientific models seem to lack the necessary accuracy. This study addresses these challenges by introducing JPResUnet (Joint PDF Residual U-net), a novel model that integrates the strengths of both ML and traditional scientific approaches to predict sub-grid joint probability density functions (PDFs) in partially premixed flames. JPResUnet employs a residual U-Net architecture to translate classic $\beta$-PDFs to sub-grid PDFs. The model is trained using direct numerical simulation (DNS) data from methane-air moderate or intense low-oxygen dilution (MILD) combustion and is initially tested through a priori assessments on out-of-sample data. Comparative analyses against an artificial neural network (ANN) and the $\beta$-PDF approach demonstrate that JPResUnet consistently outperforms these methods in capturing complex sub-grid features with greater accuracy and robustness for both box and Gaussian kernels of varying widths, and for more extrapolated cases. Subsequent a posteriori assessment involves two versions of JPResUnet with different output PDF resolutions, which are deployed for large eddy simulation (LES) of a multi-regime burner through the look-up table (LUT) approach. The higher resolution model yields improvements in temperature estimates compared to the conventional LUT method. This highlights the potential of the JPResUnet model for robust and accurate LES of reacting flows with ML.