Deep learning dynamical latencies for the analysis and reduction of combustion chemistry kinetics

The modeling of chemical kinetics holds many challenges, as well as a necessity for more efficient modeling techniques, together with dimensionality reduction techniques. This work studies the application of time-lag auto-encoders for the analysis of combustion chemistry kinetics. Such a technique allows a better reconstruction of the thermochemical temporal advancement in relation to traditional reduction techniques (principal component analysis) while applying a potential denoising operation. Moreover, the reduced manifolds or latencies are provided with physical meaning, which further analysis gives insight into key chemical reactions and interactions between chemical species, allowing for a deeper understanding of the chemical mechanism itself.