GEC Student Excellence Award Finalist Presentation - Machine-learning-based prediction of plasma characteristics from optical emission spectra

Achieving consistent atomistic control in semiconductor plasma processing calls for real-time monitoring of the plasma conditions. While a simple and noninvasive diagnostic tool such as optical emission spectroscopy (OES) can be employed for such monitoring, interpreting the respective data to determine plasma characteristics is not an easy task. In this study, experimentally obtained OES data and plasma parameters obtained from particle-in-cell/Monte Carlo Collision (PIC/MCC) simulations as well as collisional-radiative model (CRM) simulations, combined with machine learning (ML) techniques were employed to predict plasma characteristics, including electron energy distribution functions (EEDF), from experimentally obtained OES data of radiofrequency-driven argon capacitively-coupled plasmas at various pressures (2-100 Pa). The use of several ML methods, including Multiple Kernel Ridge Regression, Random Forest, and Artificial Neural Networks, was investigated. This study first confirms that OES data of Ar discharges obtained from the coupled PIC/MCC-CRM simulations are in good agreement with measured OES data at low gas pressures. The inverse problem by the ML method also successfully predicts plasma parameters, including the EEDFs, from the measured OES data.