An efficient autoencoder neural network plasma-surface interaction model for sputtering processes

Comprehensive low-temperature plasma processing models require an accurate physical description on all time and length scales of the plasma discharge, the plasma transport and the plasma-surface interactions. In this work, we revise a machine learning plasma-surface interface surrogate model for bridging the inherent scales with a high physics-fidelity regression model at modest computational cost, to address the excessive number of considered degrees of freedom. To establish a physical description of mixed material sputtering, the impingement of Ar ions onto Ti-Al composites is studied in this work for varying stochiometries by means of Monte Carlo surface simulations. The obtained data is compressed into a low-dimensional latent space representation by utilizing a variational autoencoder neural network. Thereafter, a comprehensive regression is realized by using a dedicated mapper network and transfer learning of the previously trained decoder network. Excellent generalization is demonstrated, enabling effective denoising of statistically varying data samples and physically sound interpolation even for not previously trained stoichiometries and incident ion energies.