Development of Machine Learning Potentials for Plasma-Surface Interactions

Molecular dynamics (MD) simulations can aid in the understanding of plasma-surface interactions (PSIs) in thin film and surface processing. MD simulations require robust and reliable force fields to accurately model PSIs. Currently, there is no obvious systematic procedure to develop force fields for the many new materials being introduced in plasma processing. Although many empirical force fields for plasma processes have been constructed, these potentials generally result in tradeoffs between accuracy and computational efficiency. Machine learning (ML) techniques can enable the development of ab initio-based models which have the computational efficiency of empirical force fields while maintaining ab initio accuracy. In this work, we have developed a ML model for the interaction between Si and Ar using data from quantum density functional theory calculations to obtain the effective energy hypersurface and associated forces. In order to correct unrealistic behavior at short distances which can be accessed during ion bombardment, we interpolate the ML model with a short-range potential. The results show good agreement with experimental Si etch yields and amorphous layer thicknesses. Further, we use ML potentials to simulate atomic layer etching of Si by Cl₂ molecules and Ar⁺ ions. The etch per cycle as a function of Ar⁺ ion energy shows good agreement with experiment and previous results using empirical force fields. Finally, the distribution of etch products during the ion bombardment is analyzed.