Modeling of APCVD-doped silicon dioxide deposition process by a modular neural network

This paper describes a methodology based on the combined utilization of bot h a multisensor system and an optimized artificial neural network that has been applied to equipment utilized for the production of doped silicon diox ide films. The model exhibits an average relative error around 1% in predic ting the concentrations of dopants and the thickness of the oxide layer. On e of the major benefits of such a predictor is the ability of providing an on-line estimate of the process yield, thus avoiding off-line testing and g aining a significant reduction of risks of wafer loss. The neural model her e described is currently utilized as a control tool at the Texas Instrument s Avezzano, Italy, plant.