Applying machine learning for the prediction of electron fluxes in capacitively coupled plasma

Fluid models of low temperature plasmas are often used to investigate application-relevant timescales, dimensionality, and length scales compared to fully kinetic models. This work seeks to explore an application of machine learning to improve the fluid closures in low temperature plasmas, which is demonstrated by predicting electron fluxes in capacitively coupled plasmas. Using a multi-layer perceptron with training data taken from particle-in-cell (PIC) simulations, electron fluxes are more accurate than the commonly used local mean energy approximation (LMEA). These network inputs are the same required to evaluate the fluxes for the LMEA using the drift-diffusion model (electron density and gradient, electric field, electron mean energy), displaying the strength of deep learning relative to traditional methods. Evaluation of the network is significantly faster than running a PIC code. At low pressures where the LMEA is known to fail, nonlocal effects become significant, so a ‘neighborhood’ of information is passed on to the network. Each individual point is given to the model with the state of several surrounding cells included, allowing the model to predict the flux at a location while consulting nonlocal data, exceeding the accuracy of the LMEA. Further examination of these nonlocal cases, as well as transferability of the model between various pressures is to be explored.