2D flame temperature and soot concentration reconstruction from partial discrete data via machine learning: A case study

Abstract

Flame reconstruction provides a valuable tool for understanding the thermodynamic properties of combustion. Based on the interrelation between different flame locations, this study proposed a new machine learning-based approach to reconstruct the overall 2D temperature and soot concentration from partial discrete data. Six cases were used to evaluate the reconstruction performance of different models (ANN, SVR, and RF), with accuracy assessed through visual observation, scatter plots, and error statistics. Results indicated that the machine learning models could well reconstruct the overall flame field from partial data, with their performance ranked as follows: RF > ANN > SVR. Moreover, the prediction accuracies of ANN, SVR, and RF for flame temperature were all superior to those for soot concentration. For the reconstruction of the temperature fields in Cases 1–3, the predicted values from the optimal RF model closely matched the measurement. For the reconstruction of soot concentration information in Cases 4–6, the predicted values from the RF also showed a similarity to the measurement. This methodology advanced flame reconstruction by leveraging the interrelation between data from different flame locations, enhancing the utilization efficiency of limited data and reducing the reliance on measurement resources.