par D'Alessio, Giuseppe ;Aversano, Gianmarco ;Zdybal, Kamila ;Cuoci, Alberto ;Parente, Alessandro
Référence 17th International Conference on Numerical Combustion(6-8 May 2019: Aachen, Germany)
Publication Non publié, 2019-05-07
Abstract de conférence
Résumé : The high dimensional, non-linear nature of combustion problems modeled by detailed kinetic schemes makes the interpreta- tion of combustion data very challenging. Machine learning methodologies allow for a lower-dimensional representation of high-dimensional data. Thus, they represent an essential tool in supporting data interpretation and visualization. Principal Component Analysis (PCA), Independent Component Analysis (ICA), Dynamic Mode Decomposition (DMD) and auto-encoders are the most employed techniques in data analysis. PCA finds a new, reduced set of uncorrelated variables, which are a linear combination of the original ones, that account for most of the original data variance. ICA finds a new set of variables, linear combination of the original ones, that maximize a certain inde- pendence measure. DMD finds a set of modes each of which is associated with a fixed oscillation frequency and growth rate. Auto-encoders find a non-linear, lower dimensional manifold that approximates the original data with low errors. These methods can be trained on cheap data (1D flames) for subsequent analysis of more complex data-sets (3D turbulent flames). The potential of these methodologies has been investigated on a DNS simulation of a co-flow syngas turbulent jet. Different regions of the flame could be identified in an unsupervised fashion, characterized by different main chemical species and processes. The original data reconstruction from their low-dimensional representation has been used as quality measure for the feature extraction process.