Surrogate models of capacitively-coupled plasmas by machine learning

Numerical simulations of radio-frequency (RF) driven capacitively-coupled plasmas (CCP) were performed under various conditions and the obtained phase-averaged profiles of their physical quantities such as the electron density and electrostatic potential were used to construct surrogate models of the plasma discharge. While the first-principles-based plasma simulation requires some time to predict the steady-state plasma profiles correctly under the given discharge conditions, the surrogate model can predict such profiles instantaneously because it interpolates the stored data.[1] Such surrogate models, if constructed for realistic plasmas with complex gas chemistry, could be used for the real-time control of plasma processing. In this study, we used one-dimensional fluid-model (FM) and particle-in-cell/Monte Carlo collision (PIC/MCC) simulations of argon CCP to obtain profile data. The neutral-network-based model that we constructed was found to be the best among our other models based on different regression techniques. Because PIC/MCC simulations are more time-consuming than FM simulations, we applied a transfer learning with extensive FM simulation data and a relatively small amount of PIC/MCC simulation data.