Artificial Neural Networks for Reaction Rate Prediction in ArO and UO Plasma Chemistry

Both experimental and computational approaches to the characterization of plasma-chemical reaction networks are faced with combinatorial complexity. Plasma discharges in complex chemistries (e.g. atmospheric discharges, complex precursors) may have too many species to simulate spatially, and lifetimes for these species may be very short with signatures that are difficult to discern from each other. Other information, principally rate expressions, may be challenging to obtain experimentally. Application of artificial neural networks (ANNs) has shown promise towards solving these problems.

In this work we show progress towards rate expression prediction in 1) a UO laser-ablation discharge over varying laser intensities, and 2) an argon parallel-plate discharge over varying pressures and voltages. For each reaction in each chemical network, a small 2-layer 128-neuron neural network is trained to predict rate expressions from chemical species densities. Mean square percent errors in predicting out-of-training-set data were found to be below 1% for most reactions. Further, we show that ANN predictions remain robust when network inputs are restricted to 10-20% of the full species list.