Deep Learning for the Identification of Key Dynamical Chemical Variables

What is it about?

The modeling of combustion chemistry kinetics is a very challenging task, principally due to the high dimensionality (many chemical species to solve), which can be also accompanied by other closure physical models, such as turbulence, acoustics, and radiation, to name a few. Many attempts have been made to simplify the problem of chemistry kinetics modeling. However, none of them presents itself as a general solution to all the possible cases. Nowadays, with the increasing popularity of Machine Learning (ML) techniques, there have been many developments for their application to the problem of combustion modeling and not only, but also for the task of dimensionality reduction, or in other words, to simplify the model to be applied, as well as explore spaces that hold less number of variables (i.e. number of chemical species).  In this work, a time-lag autoencoder (TAE) neural network is applied for the analysis of combustion chemistry kinetics, such a model consists of a neural network that associates inputs and outputs vectors, which represent a thermochemical state vector (Temperature and chemical species mass fractions values) while keeping a constant time shift between the associated input and output at every time sample; at the same time, this model applies a dimensionality reduction. It is from this dimensionality reduction that is possible to obtain what is usually called latencies, which are the main study object of this paper.

Featured Image

Photo by Patrick Hendry on Unsplash

Photo by Patrick Hendry on Unsplash

Why is it important?

Machine Learning (ML) techniques for dimensionality reduction can be quite effective in reducing the number of variables, however, most of them lack a physical interpretation of the obtained variables. In this work, the temporal characterization of the resulting models has been exploited to interpret the obtained latencies (or reduced variables), giving place to the identification of key chemical species, which hold the statistical variance of the chemical mechanism (set of chemical reactions that lead from the reactants to the products). Such chemical species have been denominated as  “chemical carriers”, analyzing such chemical species interactions, can potentially lead to the finding of key chemical reactions involved in the chemical mechanism, which could help in the development of reduced models. Additionally, the qualities of the time-lag autoencoder (TAE) for the preserving of information have been tested, making a comparison with a Principal Component Analysis technique (PCA); this PCA technique holds great popularity due to its simplicity for calculation. However, as is observed in this paper, it has some difficulties with the correct representation of the combustion non-linearities, which are better represented by the TAE network. Lastly, the TAE technique also allows for a temporal advancement of the thermochemical state vector.