Detecting measurement outliers: remeasure efficiently

Shrinking structures, advanced optical proximity correction (OPC) and complex measurement strategies continually challenge critical dimension (CD) metrology tools and recipe creation processes. One important quality ensuring task is the control of measurement outlier behavior. Outliers could trigger false positive alarm for specification violations impacting cycle time or potentially yield. Constant high level of outliers not only deteriorates cycle time but also puts unnecessary stress on tool operators leading eventually to human errors. At tool level the sources of outliers are natural variations (e.g. beam current etc.), drifts, contrast conditions, focus determination or pattern recognition issues, etc. Some of these can result from suboptimal or even wrong recipe settings, like focus position or measurement box size. Such outliers, created by an automatic recipe creation process faced with more complicated structures, would manifest itself rather as systematic variation of measurements than the one caused by 'pure' tool variation. I analyzed several statistical methods to detect outliers. These range from classical outlier tests for extrema, robust metrics like interquartile range (IQR) to methods evaluating the distribution of different populations of measurement sites, like the Cochran test. The latter suits especially the detection of systematic effects. The next level of outlier detection entwines additional information about the mask and the manufacturing process with the measurement results. The methods were reviewed for measured variations assumed to be normally distributed with zero mean but also for the presence of a statistically significant spatial process signature. I arrive at the conclusion that intelligent outlier detection can influence the efficiency and cycle time of CD metrology greatly. In combination with process information like target, typical platform variation and signature, one can tailor the detection to the needs of the photomask at hand. By monitoring the outlier behavior carefully, weaknesses of the automatic recipe creation process can be spotted.