Machine learning-based prediction of the electron energy distribution function and electron density of argon plasma from the optical emission spectra

Abstract

Optical emission spectroscopy (OES) is a highly valuable tool for plasma characterization due to its nonintrusive and versatile nature. The intensities of the emission lines contain information about the parameters of the underlying plasma–electron density ne and temperature or, more generally, the electron energy distribution function (EEDF). This study aims to obtain the EEDF and ne from the OES data of argon plasma with machine learning (ML) techniques. Two different models, i.e., the Kernel Regression for Functional Data (KRFD) and an artificial neural network (ANN), are used to predict the normalized EEDF and Random Forest (RF) regression is used to predict ne. The ML models are trained with computed plasma data obtained from Particle-in-Cell/Monte Carlo Collision simulations coupled with a collisional–radiative model. All three ML models developed in this study are found to predict with high accuracy what they are trained to predict when the simulated test OES data are used as the input data. When the experimentally measured OES data are used as the input data, the ANN-based model predicts the normalized EEDF with reasonable accuracy under the discharge conditions where the simulation data are known to agree well with the corresponding experimental data. However, the capabilities of the KRFD and RF models to predict the EEDF and ne from experimental OES data are found to be rather limited, reflecting the need for further improvement of the robustness of these models.