Surrogate Models for Low Temperature Plasma Simulations with Deep Learning

Advances in machine learning algorithms have recently generated considerable interest for constructing computationally efficient counterparts of complex dynamical systems such as low temperature plasmas. Accurate multiphysics/multiscale plasma simulations are often slow to execute for real world problems e.g. plasma reactors used for etching and deposition processes in semiconductor manufacturing, that limits their applicability for design and optimization purposes. A promising route to accelerate simulations by building fast and accurate surrogates with machine learning provides high computational gains for real-time prediction and multiscale simulations. In this work, we present a case study of development of a parametric emulation framework based on non-intrusive, data-driven methods using deep neural networks. We developed surrogate models for periodic steady-state radio frequency powered capacitively coupled plasmas where deep learning is used for parametric interpolation of reduced spatiotemporal modes and thus can be considered as equation-free approach for latent-space representation of plasmas. We assess the viability of such method to allow design parameter exploration and to enable new, previously unfeasible computational research.