• Citer

Effects of the training dataset and data preprocessing on adaptive chemistry simulations

par D'Alessio, Giuseppe ;Aversano, Gianmarco ;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

• Détails
• Contenu

Résumé :

Machine learning provides the tools for the mitigation of the high computational costs associated with the numerical simulations of reacting flows with detailed kinetic schemes, while preserving high levels of accuracy. By means of unsupervised learning, the composition space is partitioned into clusters where the full chem- ical mechanism is reduced by Directed Relation Graphs with Error Propagation. Therefore, at each time-step of the numerical simu- lation, all grid points are classified and different reduced schemes are applied locally. Several unsupervised learning techniques for multivariate data partitioning exist in literature, such as Local Principal Component Analysis, Self-Organizing Maps, k-Means, Auto-encoders. The effectiveness of these machine learning models depend on the training data quality and availability, and on how the multivariate dataset is preprocessed. Since a certain manifold in composition space needs to be learned, it is necessary to have sufficient available data coming either from the system of interest itself or from simpler systems that can be representative for it. This implies that also 0D or 1D data can be used to train a machine learning model for a 2D system, as long as the same manifold is represented. The aforementioned techniques and different preprocessing strategies were tested for an adaptive- chemistry simulation of a laminar co-flow methane flame, with the objective to assess the different models robustness and potential in terms of chemistry reduction and simulation reliability.