Multiresolution pattern detector networks for controlling plasma etch reactors

摘要

A critical step in semiconductor wafer fabrication is to halt a plasma reactor as soon as the etched film clears to expose the underlying layer. Typically, changes int he optical emissions spectrum are employed to detect this process endpoint. However, differences in plasma chemistry and reactor chambers, small exposed areas, and variations in wafer patterns complicate this control strategy. Our novel approach uses the characteristic local shape of spectral lines as a guide to finding endpoint. With the Haar wavelet representation, we model shapes over many resolutions. A network of pattern detector 'neutrons' runs real- time to locate these shapes and stop the etch process. This provides robust endpoint detection under widely varying reactor and wafer conditions. Our neural network endpointer has been successfully tested on data gathered for six months at a major wafer fabrication facility.