Machine learning-based prediction of species mass fraction and flame characteristics in partially premixed turbulent jet flame

This study uses machine learning (ML) and large eddy simulation (LES) to predict the mass fractions of species and flame characteristics in partially premixed turbulent jet flames. A flamelet-based chemistry technique was used to perform high-fidelity LES of Sandia Flame D. The resulting dataset was used to train three ML models—neural networks (NNs), linear regression, and decision tree regression—for surrogate prediction. Among them, the NN model achieved the highest accuracy, with R-squared values exceeding 0.9998 and mean absolute error values below 1.0 × 10−4 across all species. Sensitivity analysis identified temperature and progress variable as dominant input features. Uncertainty quantification confirmed high model confidence in stable regions, while elevated uncertainty was observed for the hydroxyl radical due to its short-lived, highly reactive nature in turbulent flame fronts. The NN surrogate was also used for targeted optimization, enabling to find ten combustion states with species compositions that were consistent with experimental data within the reported range of uncertainty. Four fitness functions—Euclidean distance, Manhattan distance, collinearity coefficient, and amplitude correlation coefficient—were applied to guide the optimization process. Manhattan distance consistently demonstrated the lowest absolute errors for key species, such as CH4(0.0014), OH(0.0002), and O2(0.0072), indicating its superior accuracy and compatibility with the LES benchmark data. Additionally, the ML surrogate achieved a 17.25× reduction in total computing time compared to the LES solver, promising efficient parametric exploration and rapid predictive capability. These findings demonstrate the potential of ML-based surrogates to support real-time combustion diagnostics, optimization, and design.