Deep learning model for ion sputtering dynamics with molecular dynamics simulation

Atomistic simulations have emerged as an effective means for elucidating physical mechanisms of atomic-scale manufacturing. In particular, molecular dynamics (MD) simulations have been utilized for modeling plasma-surface interactions. This work presents a deep neural network based reduced order modeling combined with intensive MD simulations for revealing the fundamental nature of argon ion sputtering on copper substrate. Using MD simulations, sputtering yields, angular and energy distributions of both sputtered atoms and scattered ions are examined for varying characteristics of incident ions and the target surface. Then, beta variational autoencoder is used to reduce the high dimension of distributional outputs to a lower dimensional latent vector through an encoder part and reconstruct it to the original dimension through a decoder part. With the inputs and the corresponding reduced outputs, a regression model consisting of fully connected layers was trained in a supervised manner to approximate the nonlinear relations between the inputs and outputs. The present model enables us to handle the surface dynamics for sputtered and scattered species in an efficient and effective fashion, and can be further extended various other surface communications at the atomistic level.