Scale-Bridging Simulations for Plasma-Surface Interactions

Plasma-surface interactions play an integral role for a variety of processes (e.g., sputter deposition, plasma-enhanced catalysis) but combine length and time-scales that differ in orders of magnitude. They are therefore commonly accounted for by oversimplifying (model) assumptions (e.g., lookup tables with rate coefficients), often biased by experience and feasibility. Unbiased data-driven approaches are pursued in this work for the sputtering and the deposition of metals and metal nitrides in Ar and Ar/N₂ plasmas. Obstacles and key perspectives are highlighted for the data generation (i.e., hybrid reactive molecular dynamics / time-stamped force-bias Monte Carlo simulations) as well as for the selection of suitable artificial neural network architectures. The proposed architecture allows to simultaneously include both experimental and simulations data in the training set, despite the different physical properties that are accessible. The role of entrapped working gas atoms (i.e., Ar) for sputtering and thin film deposition is found to be negligible and significant, respectively. Ultimately, the temporal development of sputtering phenomena (e.g., emission of N₂ due to collisions with N split interstitials) and various thin film properties (e.g., composition, density) are predicted with atomic but still high physical fidelity for experimental process times of 45 minutes. Rare events (e.g., impingements of ions whose kinetic energies stem from the tail of the ion energy distribution function) can be of paramount importance to the steady-state and increase the time to reach equilibrium from a few seconds up to 30 minutes. The machine learning predictions took merely 34 GPU hours. In contrast, hybrid reactive molecular dynamics / time-stamped force-bias Monte Carlo simulations would last for more than approximately 8 million CPU years.