Fault Detection via Optical Filtering and Statistical Correlation

Present inspection techniques in the semiconductor industry are generally serial in nature. As a result of this serial nature, certain types of tests are prohibitively expensive or even impossible. For example, to completely test a 64K random access semiconductor memory can require up to N3 ([65,536]3) or about 2.8 x 10 14 individual tests. At 10 7 tests per second, testing 24 hours per day, one memory chip would take 326 days to test! The solution to this problem is simple--the chips are not fully tested. A very real problem, however, is the inspection of the optical masks used to make semiconductors. Todays inspection techniques, which are generally visual and serial in nature, can take days to weeks of inspection time. Needless to say, only the master set of masks is inspected and then used to make sub-master sets. Another approach to visual inspection is based on the Fourier transform properties of optical lenses. In this method, information processing is accomplished in the two-dimensional Fourier domain using photographic techniques. The limitations of this method are those imposed by the photographic process--lens aberrations, non-linearities, film granularity, etc. If some or all of these limitations could be overcome, the parallel information processing afforded by the Fourier domain technique would make full inspection practical or economical for many applications where it cannot now be considered. The purpose of this dissertation is to address and solve the problem presented by certain of these photographic limitations. A technique of spatial modulation along with a technique of two dimensional statistical correlation are used to determine the existence of faults in a given image contaminated by photographic process noise. These techniques are developed analytically, and a digital simulation is performed to verify the analysis