Chemical dynamics constitute a bottleneck in the acceleration of reaction flow simulation. This is, in large part, due to the stiffness of the associated ODEs, i.e. the simultaneous presence of slow and fast timescales. We present several results aiming to replace stiff ODE solvers with Machine Learning approaches. Our studies underscore the importance of dataset balance, the necessity to preprocess simulation data and the potential of dimensionality reduction. We illustrate our results by showing significant increase in computational efficiency while simulating the behaviour of 0D reactors with a large number (52) of chemical species.