Proceedings Article | 2 June 2000
KEYWORDS: Semiconducting wafers, Inspection, Signal processing, Particles, Defect detection, Lithography, Optical inspection, Optical lithography, Visualization, Interference (communication)
Although the subject of frequent concern, criticism, and attention in the modern semiconductor fabrication facility, human after develop inspection (ADI) does not catch the major scrap and yield events early enough, if at all. The overall success of scrap and photo redo reduction programs over past years has resulted in residual problem levels which are difficult to improve upon -- yet still very costly. Detected 'events' are few and far-between, although evidence of their prevalence is frequently seen at subsequent inspections, or finally at probe. In the ASIC fab, they put on-time delivery to customers at risk, because individual wafer lots in an ASIC facility have a designated customer. The sampled area is limited by human throughput to less than 10% of the wafers in a lot. The visual ADI process step is unpopular among manufacturing technicians. It is often a bottleneck in the photo area. Statistically, in a photo area with capacity of 5000 wafer starts per week, only a few wafers processed per day are destined for scrap. Since wafer events occur in sporadic clusters, the photo area experiences only a few significant incidents per month. The typical operator can expect to intercept such an event less than once during several months of otherwise uneventful ADI inspection haystack.' Hence the stubbornness of our residual problem. Going beyond the statistical problem, our current manual macro-inspection equipment is engineered appropriately to ancient IC generations. A collimated, oblique-oriented light was an effective darkfield illumination source, when line widths were much larger than the wavelength of light. When line width is comparable to, or smaller than, the wavelength, the collimated light source produces scintillating diffracted colors on the wafer. Thus diffraction 'noise' significantly buries the defect 'signal' in the typical bright light visual macro inspection. In addition, there is the problem of variability between human inspectors, and the impossibility of accurate classification and recording of defect types, locations, and layer of occurrence. In this paper, we discuss a pilot implementation of an automated macro inspection system at Motorola, Inc., which has enabled the early detection and containment of significant photolithography defects. We show a variety of different types of defects that have been effectively detected and identified by this system during production usage. We introduce a methodology for determining the automated tool's ability to discriminate between the defect signal and process noise. We indicate the potential for defect database analysis, and identification of maverick product. Based upon the pilot experience, we discuss the parameters of a cost/benefit analysis of full implementation. The costs involve tool cost, additional wafer dispositions, and the engineering costs of recipe management. The most tangible measurable benefit is the saved revenue of scrapped wafers. An analysis of risk also shows a major reduction due to improved detection, as well as reduced occurrence because of better containment. This reduction of risk extends both to the customer -- in terms of field failures, OTD, maverick product -- as well as to the production facility -- in terms of major scrap incidents, forced inking at probe, redo, and containment.
