Feature Extraction by means of Unsupervised Learning Techniques on LES Combustion Data

. Principal Component Analysis (PCA) has shown very promising capabilities from the perspective of reduced-order model development as it provides optimal progress variables that can be used for the parametrization of the reacting system as well as to gain insight on the features of combustion processes. Local PCA is able to find lower-dimensional clusters in the data-space, i.e. regions where a subset of the original variables account for most of the original variance, and thus overcomes the limits of global PCA in dealing with the recurrent non-linearities of combustion problems. Self-Organizing Maps (SOMs) are a class of Artificial Neural Network (ANN) which are used to map an ensemble of high dimensional observations onto a non-linear, lower-dimensional grid. Both methodologies are useful in finding structures in the data manifold and in extracting information that can also be used for mechanism reduction by means of an adaptive kinetic scheme. In this work, Local PCA and SOMs are applied to a data-set obtained from a LES simulation of two co-flow jets: the central jet is an equimolar mixture of CH4 and H2; the annulus jet has an oxygen content of 3%. Results show that Local PCA is able to efficiently cluster the data, while also offering a measure for the quality of the clustering process itself. The methodology also extracts information about the dominant variables in each clustered region. On the contrary, SOMs do not provide a measure for the quality of their clustering solution nor do they indicate a subset of dominant variables. Furthermore, the method wrongly detected features in the data manifold.