Abstract
To improve equipment throughput and device yield, faults in plasma equipment should be stringently diagnosed. An ex situ diagnosis technique is presented. This was accomplished by recognizing X-ray photoelectron spectroscopy pattern by using a modular backpropagation neural network (BPNN). Each BPNN comprising a modular network was specific to a variation in a process parameter. For comparison, principal component analysis (PCA) was applied to X-ray photoelectron spectroscopy (XPS) data and with these data another modular network was constructed. A total of 17 XPS patterns were used to construct a modular network. Model performance was evaluated in terms of the recognition and diagnosis accuracies. Model accuracy was also investigated as a function of hidden neuron number or threshold. The optimized model trained with PCA-XPS with 100% data variance demonstrated a smaller prediction error compared to XPS and PCA-XPS with 99% data variance. Meanwhile, the PCA-XPS model with 100% data variance yielded a significant improvement of about 32% in fault diagnosis compared to pure XPS model. The improvement was more pronounced under stricter monitoring conditions.
ACKNOWLEDGMENTS
This study was supported by Grand No. R11-2000-086-0000-0 from the Center of Excellency Program of the KOSEF, MOST.