Data-driven modeling of the flame composition space of an intrinsically unstable lean premixed ammonia/hydrogen/air flame

This study investigates the capability of different data-driven methodologies of achieving dimensionality reduction of the flame composition space in a large-scale lean premixed, intrinsically unstable ammonia/hydrogen/air flame, while accurately representing key chemical species. Two data-driven techniques, one consisting of a combination of Singular Value Decomposition and Artificial Neural Network (SVD+ANN) and a second consisting of a Linear-Nonlinear Autoencoder (LNAE), were employed to model three relevant species. The models were trained on three different datasets: a small-scale DNS dataset, unstretched 1D flamelets, and a combination of stretched and unstretched flamelets. Error analyses demonstrate that both techniques effectively capture the species behavior using at least two (or three, depending on the behavior of specific species) independent variables, when trained on the small-scale DNS dataset. The models trained on the unstretched flamelet dataset show worse testing performance with both techniques, while the LNAE models trained on the stretched flamelet dataset can accurately reproduce the pollutant concentrations when three or more predictors are used. The results highlight the potential of the LNAE data-driven approach to accurately model key chemical species belonging to a DNS dataset of an intrinsically unstable ammonia/hydrogen flames using strained flamelet data, while highlighting the generalization capabilities of the different training datasets.