Modeling of plasma process data using a multi-parameterized generalized regression neural network

  • Authors:

  • Byungwhan Kim;Minji Kwon;Sang Hee Kwon

  • Affiliations:

  • Department of Electronic Engineering, Sejong University, 98 Kunja-Dong, Kwangjin-Ku, Seoul 143-747, Republic of Korea;Department of Electronic Engineering, Sejong University, 98 Kunja-Dong, Kwangjin-Ku, Seoul 143-747, Republic of Korea;Department of Electronic Engineering, Sejong University, 98 Kunja-Dong, Kwangjin-Ku, Seoul 143-747, Republic of Korea

  • Venue:
  • Microelectronic Engineering
  • Year:
  • 2009

Quantified Score

Hi-index 2.88

Visualization

Abstract

For characterization or optimization process, a computer prediction model is in demand. A new technique, to improve the prediction performance of conventional generalized regression neural network (GRNN) of plasma process data was presented. Genetic algorithm (GA) was applied to optimize multi-parameterized training factors of GRNN. To evaluate the technique, two data sets were collected from the etchings of silica and silicon carbide (SiC) thin films in inductively coupled plasmas. Both data sets called Data I and Data II were statistically characterized by means of 2^3 and 2^4 full factorial experiment plus one center point. The GRNN models trained with these data were tested with additional six and 16 experiments. A total of eight etch outputs were modeled and compared with conventional GRNN and statistical regression models. The five etch outputs comprising Data I include silica etch rate, aluminum (Al) etch rate, Al selectivity, profile angle, and DC bias. Data II consisted of three etch outputs, including SiC etch rate, surface roughness, and profile angle. Compared to GRNN models, GA-GRNN models yielded more than 40% and 15% improvements for all etch outputs comprising Data I and Data II, respectively. Similar improvements were also demonstrated with respect to statistical regression models. All these results reveal that a multi-parameterization of training factors and GA optimization is an effective technique to considerably improve the prediction performance of conventional GRNN model.