Automatically Analyzing Effects of Process Equipment on Yield

Laura Peters, Senior Editor -- Semiconductor International, 8/1/1999

   Part 7 of Series

Part 7 in this series of articles on Integrated Yield Management by Nick Atchison and Ron Ross of Silicon Systems (Santa Cruz, Calif.) demonstrates automated numerical analysis of the complex effects wafer processing equipment has on yield. Typical yield analysis techniques using wafer averages minimally detect 3% yield differences between equipment with 95% confidence levels, while this new technique expands the confidence level and detects mean probability differences of l 0.2% with a sample size of only 2250 wafers. The method uses tool-to-tool comparative analysis on wafer geometric zones to identify low-level yield differences between different process tools that, when combined, significantly degrade overall yield. This and other articles in this series are presented at www.Semiconductor.Net.

The equipment yield loss method (Y_E) determines random defect (Y_D) and systematic (Y_S) components by performing wafer zone analysis on each wafer lot for each piece of equipment at each process step, calculating the mean, median and standard deviation of yield. Zonal-based analysis prevents obscuring of regional effects typical of wafer-average yield analyses. One chooses zones to maximize the probability of independent yield variation. After identifying zones (Figure shows zones of a 150 mm wafer), the user evaluates the data to determine the lowest-yielding process tools. It also improves data resolution by removing low-use equipment from the evaluation. Later, low-use equipment data is checked for normal distribution and to see if the mean matches that of high-use tools, differences that may indicate statistically significant problems. Finally, the user performs in-depth graphical analysis on the worst-case process equipment and processes.

Fig 1 Wafer zone analysis prevents obscuring of regional effects typical of wafer-average yield analyses.

The method arranges analysis zones to confine regionally distinct yield loss propensities to their own domain. The computer program's calculation of the mean, median and standard deviation yield for each zone, wafer lot and tool is so computationally intensive, Atchison and Ross suggest using 'C' code programs optimized for high-speed analysis of wafer data. For example, an analysis taking six days using RS1 code can require only six minutes using optimized C code.

Once lowest-yielding wafer zones are located and correlated with a process tool and step, the method's 'quad' graphical analysis compares: analyzed wafers versus the rest of the population; process tools with identical performance; two (or more) tools showing different yields in a zone; and tools demonstrating identical performance over most of the probability plot yet different l ow-yielding regions. The user repeats quad analysis on the chosen wafers to reduce yield distribution standard deviations and enable 0.2% equipment yield difference measurements with >95% confidence.

Once tools with yield issues are isolated, the user performs any of the integrated yield management methods presented in previous articles including cluster analysis, parametric yield limit analysis, design-of-experiments or evaluation of KLA inspection using yield data for each equipment set at a given process stage. Y_E mean difference, probability plots, box plots, trend plots and zone histograms help isolate causes of yield differences from one piece of equipment to another. Such tools illustrate data spreads, confidence intervals and yield distributions.

The Y_E method successfully identified a consistent, 2% lower yield in zone A1 for a laser scriber -- yield loss remedied by adjusting the scriber's vent flow. In another case, zone 2 had a skewed distribution on the yield histogram where superimposed curves on the main curve corresponded to the number of times the wafers were reworked. Rework elimination and closer clean monitoring restored yield curves to normal.