Reducing the complexity of chemical networks via interpretable autoencoders

In many astrophysical applications, the cost of solving a chemical network represented by system ordinary differential equations (ODEs) grows significantly with size network, and can often represent significant computational bottleneck, particularly in coupled chemo-dynamical models. Although standard numerical techniques complex solutions tailored to thermochemistry somewhat reduce cost, more recently, machine learning algorithms have begun attack this challenge via data-driven dimensional reduction techniques. work, we present new class methods that take advantage data sets (autoencoders), optimization multi-parameter systems (standard backpropagation), robustness well-established ODE solvers explicitly incorporate time-dependence. This method allows us find compressed simplified version large semi-automated fashion be solved solver, while also enabling interpretability compressed, latent network. As proof concept, tested on an astrophysically-relevant 29 species 224 reactions, obtaining reduced but representative only 5 12 x65 speed-up.