|
ASMC `08 annotations: Where defects, systematic and random, get their due |
|
|
|
May 05, 2008 at 01:55 PM |
For those who thrive on the nitty-gritty of semiconductor defect inspection, review, and characterization, the Advanced Semiconductor Manufacturing Conference (ASMC) provides enough source material--or at least info on defect source mechanisms--to satiate most appetites.
This year's symposium is no exception, with a couple dozen oral presentations and poster papers either directly or indirectly addressing issues of defectivity in the process flows of the world's fabs.
In their love/hate relationship with initialisms and acronyms, defect and yield engineers are no different from their colleagues working in other parts of chip manufacturing. What's another helping of alphabet soup to categorize and memorize, when your head is already swimming with letters and jargon? KLA-Tencor's Chris Young, presenting a paper jointly written by his customers at UMC and his fellow KLA-T'ers, added another one to the ever-growing list: DBB. No, it's not chipspeak for "defects big and bad" or "designed by blindmen," but stands for "design-based binning."
DBB is one of the growing number of tools in the design for manufacturability kitbag, specifically targeted to deal with systematic defect excursion monitoring in the 65-nm device generation and beyond. The technique falls under the chapter titled "When Good Patterns Go Bad" in the ever-growing process interaction and variation catalog.
As Young pointed out, sometimes a marginal design can pass lithographic qualification but fail when certain process variations kick in. Marginal pattern failures can occur at random locations, and typical die-to-die repeater analyses often fail to pick up these yield-detracting aberrations, which are structurally systematic but spatially random. A clever solution has emerged, where design and defect inspection have been joined at the hip, using design data to identify systematic defects from the inspection results and binning the results for future identification and categorization. So-called "design clips" are employed to group "structural repeaters" based on their pattern background.
Young went through a case study during his morning talk, which described a broken poly structure classified as a "defocus excursion" during a 65-nm ramp. The problem showed up intermittently, and could be traced back to all the litho scanner tools but didn't correlate to the same location in the design every time. The team broke the puzzle down via an odor-free fishbone analysis summary, renaming the poly defocus as a "poly hammerhead" because of its unique design features (though he made no mention whether defects are always renamed with a piscine title when analyzed with the fishbone chart).
The hammerhead was classified as public enemy number one, achieving top billing in the systematic defect pareto. A commonality analysis was performed, which showed that design/OPC was off the hook as anything more than a marginal root cause. Turns out the bigger fish to fry was a problem with scanner aberrations and stage tilt issues revealed as the process drift-causing bad actors, root causes traced back to the shakin' goin' on during one of those all-too-frequent Taiwan earthquakes. (Hmm, maybe those initial UMC--and TSMC for that matter--reports of limited production impacts in the wake of the quakes should get another look?)
Once corrective action was taken, Young said that the problem has not reoccurred. Preventive maintenance cured the discrepancies, and continued monitoring has verified that the marginal, defect-prone patterns have not returned. Users can define their libraries and keep a growing file of these pattern failures, to identify their occurrence as needed. During the Q&A after his talk, he explained that once the pattern library is established, a software program takes over, which means, if I'm not mistaken, that DBB borrows a page from the automated defect classification (ADC) playbook, which has been an important weapon in the battle against process defectivity for several years.
Young made a solid argument on the compelling, if still somewhat limited, value of DBB for effectively controlling excursions during a production ramp and the technique's broader role in connecting the D and the M in DFM. For the data-hungry defect and yield contingent at ASMC, it was also something they could sink their teeth into.
|
|
