Reduction of chemical reaction mechanisms has been long studied to minimize the computational cost of reacting flows or the number of parameters of catalytic reaction models for estimation from experimental data. Conventional reduction techniques encompass either a tabulation of the reaction rates of elementary reactions or a reduction of elementary reactions. We introduce a Python-TensorFlow tool to learn chemical kinetics (LearnCK) systematically and automatically from microkinetic models using artificial neural networks (NNs). The approach constructs overall reactions among stable species only and interconversion rates and dramatically reduces the number of species and, thus, of the differential equations (the most expensive aspect in reacting flows). Doing this also removes the stiffness and nearly eliminates the complexity and cost of estimating the entire thermochemistry and kinetic rate expressions for computing reaction rates. Python programming automates training data generation, extracts metadata for fitting the NNs, and deploys the NN model. Since NNs are black boxes, we propose an approach to conserve mass. We demonstrate the method for the ammonia synthesis on Ru and the methane non-oxidative coupling over a single-atom Fe/SiO₂ catalyst. The latter model includes over 500 gas and surface species and a combined 9300 gas and surface reactions. We demonstrate a nearly 1000-fold computational speedup and exceptional predictive accuracy using up to 8 overall reactions. The NN model is embedded in macroscopic reactor flow models to estimate uncertainty.