Generative Model Based Adaptive Importance Sampling for Flux Calculations in Process TCAD

Abstract

A key part of advanced three-dimensional feature scale etching and deposition simulations is calculating the particle flux distributions. The most commonly applied flux calculation approach is top-down Monte Carlo which, however, introduces numerical noise. In principal, this noise can be reduced by increasing the number of simulated particles but doing so also increases the overall running time. For complex geometries, especially high aspect ratio structures, which are very prominent in state of the art three-dimensional electronic device designs, increasing the number of samples is not a viable approach: Only a very small subset of simulated particles contributes to reducing the noise in remote and obscured surface regions. We thus propose an adaptive importance sampling approach based on a generative model to more efficiently focus the sampling on those surface regions with high noise levels. We show that, for a constant number of simulated particles, our approach reduces the noise levels in the calculated flux by about 33% for a representative high aspect ratio test structure.