Prediction of reaction rate coefficients using neural networks

Detailed kinetic models are essential to study, interpret and optimize the plasma reactors used in a variety of applications. However, their predictive power is often hindered by uncertainties in several reaction rate coefficients. In this work, we use artificial neural networks (ANN) to predict a selected subset of these rate coefficients from the knowledge of the species densities in various experimental conditions. The ANN is trained on data generated by numerical simulations using an established kinetic scheme where the reaction rate coefficients are sampled around their established valued. It then receives the steady-state densities of several species and for various experimental conditions as input and learns to inversely map the simulations to determine the corresponding reaction rate coefficients. Moreover, as the measurement of some species is experimentally challenging, we assess which species are essential for accurate predictions and which can be omitted with minimal loss of performance. To this purpose, we use the Morris sensitivity analysis method to hierarchize the influence of each species on the model's output and subsequently evaluate how the performance degrades as the number of inputs is reduced. This approach guides both the simplification of the ANN and the need for specific experimental measurements to improve the reaction mechanism.