Laser Induced Breakdown Spectroscopy Study of Non-premixed Flames with Machine Learning

Abstract

In-situ non-premixed flames were studied by the combination of laser induced breakdown spectroscopy (LIBS) and machine learning to obtain an accurate depiction of the local structures of diffusion flames. Six candles were utilized as representative sources of the non-premixed flames. The use of the intensity ratio of H/O provided the point information about the distribution of the local fuel to the oxidizer without disturbing the flame. The H/O ratio tends to decrease from the flame centerline to the border. The concentrations of H and C at the flame centerline are high due to the ionization/fragmentation of candle wax molecules. To discriminate the non-premixed flames with the candle as a representative, twenty-nine machine learning classification models of the classification learner app were implemented on LIBS data. The quadratic discriminant analysis and neural network models provided unprecedented optimum results for training and for testing in the case of distinguishing all candles. The cubic support vector machines (SVM) and Neural Network gave the best results for both training (validation)and predicting the spatial positions in a candle flame. The reported procedure can be potentially applied to the aviation, space, and engine manufacturer industries to improve the efficiency of combustion and reduce pollutant emissions.