Characterization and improvement of unpatterned wafer defect review on SEMs Alan S. Parkes *, Zane Marek ** JEOL USA, Inc. 11 Dearborn Road, Peabody, MA 01960 ABSTRACT Defect Scatter Analysis (DSA) provides a fast, accurate method for evaluating the random errors made by defect scanners when a wafer is loaded, aligned, scanned, and unloaded multiple times. The DSA tool includes a 200 mm or 300 mm wafer that has a series of patterned defects; the defects can be located by the inspection scanner and the patterns detected even if there are many other defects present. The DSA software analyses multiple scans, locates the pattern at each site for every scan, and compares the distributions of the pattern at each site to the average reported position, as well as to the known site position. Plots of the predicted positions show the scatter in terms of x, y and rotational errors. The composite plot for all scans and sites represents a figure of merit for the scanner. By itself, DSA is useful for evaluating which scanner makes the smallest set of random errors, and the effects of modifying a scanner to minimize these errors. In addition, the software also generates a defect file in the same format as the input files, showing the average reported positions for the sites of the programmed defects. This file and the wafer provide the best input to the JEOL in-fab microscopes for the LMLS/SSBWA procedures that correct for most of the systematic differences between the predicted positions and the defect positions as relocated on the SEMs 1. The tool consists of one 200mm or 300mm wafer and the analysis software, provided in both UNIX (either 9.05 or 10.2) and Windows 2000. The software will accept defect files in standard KLA, Tencor, and Inspex formats. Keywords: Defect review, Defect relocation, Defect scanner errors, Alignment errors 1. INTRODUCTION Optical defect scanners use various techniques to locate inconsistencies on the surface of wafers, often debris scattered on the surface, imperfections in the surface, missing pattern, or added pattern. Although the scanners can locate very small defects, the available magnification range is not sufficient to enable easy identification of the defect, if at all. If the position of a defect is recorded, the wafer can be loaded in an SEM and moved to the location of that defect. With the available tools on the SEM, including much greater magnification, energy-dispersive x-ray analysis, etc., the identification is much easier. A typical defect review inspection plan would involve loading and aligning the wafer in an inspector so that the stage coordinates of the wafer center and the orientation of the wafer in terms of the position of the notch or flat are determined; then scanning the wafer for defects, converting the stage coordinates at each defect into a coordinate system based on the wafer. This process is reversed in the SEM, again aligning the wafer, converting the predicted coordinates of each defect into SEM stage coordinates, then driving the stage to that position. When the defects are sufficiently large so that a relatively low magnification, with its correspondingly large field of view, will be adequate to see the defect, the accumulated errors in stage positions, wafer alignment, etc., are small enough that the defect will be in the field of view when the SEM is driven to the predicted position. However, as reported defects get smaller and smaller, the necessary increase in magnification reduces the field of view to a point that the defect is often outside of the field. Where an SEM operator is present, various techniques can be used to locate a defect, particularly if there are a few large defects. Once two defects have been correctly located, a transform can be calculated to correct the predicted coordinates into much more accurate ones 2. This requires the presence of an operator, the time to do the searching and checking, and some relatively large defects. Automation of the review process tends to eliminate the operator, and, increasingly, there are no adequately large defects. The best solution to this problem is to reduce the magnitude of the errors in the predicted coordinates of the defects The sources of error must be considered in order to reduce the magnitude of the errors in question. Typically, the scanner will convert the actual stage coordinates into virtual stage coordinates, where the x- and y-axes are straight and perpendicular, and the reported difference between two positions of the stage can be converted exactly into microns.
Even with backlash removed, the reported actual stage coordinates aren t exact, and the factors to convert to a virtual stage are also not exact. The alignment of the wafer depends on locating the positions of several points around the perimeter; there are both setting errors and a possible discrepancy between the position of the alignment microscope and the defect scanner beam. These types of errors are repeated when the wafer is loaded in the SEM, and it is the composite of these errors that must be reduced. This composite error can be separated into two parts, systematic errors and random errors. The conversions between a real stage and a virtual one will introduce systematic errors, errors that will be the same each time a wafer is scanned on a particular scanner, then loaded in a particular SEM. These errors can be studied, and the predicted positions can be corrected before defects are relocated in the SEM. The random errors cannot be corrected for without locating some defects in the SEM, but the major source of random errors is the alignment of the wafer in the defect scanner, and this can be studied to estimate the errors made by a scanner, and determine which scanners will have the smallest random errors in the predicted positions. This study can be done simply by loading, aligning, scanning and unloading a wafer in the defect scanner several times, and comparing the predicted positions of a particular defect in each of the scans 3. However, scanners don t find exactly the same number of defects in different scans of the same wafer, so that making the comparison for several defects across multiple scans is very tedious, and very prone to misidentification errors. 2. METHODOLOGY AND RESULTS To solve this problem, we have designed a special wafer containing a standard pattern of thirteen defects at eighty sites across the wafer. If the scanner detects these defects, a pattern-recognition program can locate the pattern, even among many other defects, and record the position of the center of the pattern at each site. If this wafer is loaded in a defect scanner, aligned, scanned, then unloaded several times, the output files can be analyzed by a computer program that locates the patterns in the first scan, corrects for overall positioning alignment errors, and records the positions of the pattern at every site where it was detected. Using the same alignment error corrections, subsequent scans are read and the patterns located. Once each scan has been examined, the results can be studied at any site with a plot of a scatter diagram, showing the position of the pattern on each scan plotted on a 2-sigma radius circle centered at the average position of the pattern for all scans. Figure 1: Wafer map showing sites where the standard pattern was detected, and a map of defects at one site.
Figure 1 shows a map of the wafer, with the positions of the eighty sites where the standard pattern has been drawn. Those sites where the pattern was detected among the defects in the vicinity of the site are shown in red (black), and the site map shows the defects found at one site, with the center of the standard pattern highlighted. When the first defect file in a series is loaded, the locations of the detected patterns are compared to the known site positions and alignment corrections are calculated to adjust the predicted positions, as shown in the dx:, dy: and dθ: windows. These values are used to correct the reported positions of defects on subsequent scans. The area about each site position can be varied to restrict the number of defects to be searched at each site; in Figure 1 is has been set to 64 sq. mm. The standard pattern consists of eleven defects in a horizontal row, spaced at 100-micron intervals, and two additional defects, located 100 microns above and below the center defect. The pattern recognition algorithm allows for an adjustable error tolerance in the reported positions of the defects; in this case it is set to 9 microns. The algorithm must find at least four consecutive defects in a horizontal row, then find a defect above and below one of these. The alignment of some defect scanners is sufficiently incorrect that the reported defect positions are assigned to the wrong sites. As a check, there are five additional patterns printed on the wafer in the form of a plus, centered at the wafer center. These patterns are turned at right angles to the primary set of patterns. If the scan area is set to the maximum value of 400 sq. mm (or 900 sq. mm, for a 300 mm wafer), these additional patterns will be checked for during the first scan and displayed, as shown in Figure 2. Here the plus is correctly centered on the wafer map. If it isn t, a large initial dx or dy offset can be used to adjust the input defect positions. Figure 2: Wafer map showing the five additional alignment sites. Table 1 summarizes the results of this alignment, after the dx, dy and dθ offset corrections have been made. Each site is reported as either no defects found within the specified search area; defects found but no pattern match; a match at a location that is significantly further from the actual site position, compared to the average distance for all of the pattern positions, and has been rejected as an outlier; and the distance in microns for the accepted patterns from the actual site positions, with the mean and standard deviation for these positions. If the results of the first scan are satisfactory, subsequent scans can be read. Since one purpose of this process is to obtain average positions of the patterns that deviate from the actual site positions by essentially just the systematic errors made by the scanner, ten scans are a reasonable number to use. The dx, dy and dθ values are locked at the values set during the initial alignment, but the scan area can be reduced, resulting in a somewhat faster pattern search.
Table 1: Report of alignment results for the first scan. Once a second scan has been studied, the Site Map button becomes sensitive, and clicking changes it to Scatter Plot. Clicking on a site on the wafer map generates a plot of the accepted pattern positions for each scan, at that site. This plot is centered at the average of the pattern positions, with a circle drawn at a radius equal to the two-sigma value for the scatter (so that ninety-five percent of the accepted positions lie within the circle, if the errors are normally distributed). After six scans, the scatter at one site is shown in Figure 3, with a two-sigma radius of 49.6 microns. Figure 3: A scatter plot at one site, showing the relative positions of the pattern at that site for six scans. Note that this plot is independent of which scan was studied first, even though the alignment based on the first scan is used for all subsequent scans. If Composite is clicked, the plots of the accepted patterns for all sites are superimposed, with a two-sigma radius equal to the average value for all sites where a pattern was detected, as in Figure 4. This two-sigma value, 41.2 microns, represents a figure of merit for the scanner, a measure of the random errors that the scanner makes in reporting defect positions.
Figure 4: The composite plot of all sites and scans, showing the figure of merit value of 41.2 microns. The wafer map shows the deviation of the average predicted position of the pattern at each site, compared to the actual position of the pattern. The average deviation is 21.2 microns (not shown). This plot is dependent on the choice of the initial scan, and its alignment. Figure 5: The errors of the first scan, compared to the average of the scans.
If Composite is clicked again, the scatter plot is redrawn, with the positions of all but the first scan plotted in blue (black), as shown in Figure 5. The first scan errors, compared to the average of all of the scans, are shown. This scan has a relatively large rotational error, as indicated by the area encompassed by the white dots. Figure 6 shows the results for the fourth scan. Here the rotational error is much less. Figure 6: The results for the fourth scan. These results can be studied to determine which scanner gives the most reproducible results from scan to scan, what the random errors are that the scanner is making, and what might be done to minimize them. For example, many scanners show good x and y positioning but large errors in rotational alignment, indicating a need for a better notchfinding procedure. Five inspectors were recently evaluated to quantify their random errors. (See Figure 7.) These inspectors consisted of both DF (dark field) and BF (bright field) equipment. The first inspector evaluated was a legacy DF inspector, capable of scanning only unpatterned wafers. Notch alignment is compensated for through software. Its laser is at 45, and is rastered across the wafer. The second inspector is a patterned/unpatterned DF, normally used post metal deposition, though it is useful in unpatterned wafer inspection. This inspector performs notch alignment off of the stage, and moves the wafer under a fixed laser, also at 45. The third inspector is a current generation BF tool, mainly used for patterned wafers, but capable of unpatterned wafer inspection. This inspector performs notch alignment on the stage, and the wafer is moved across the stage under the stationary source. The fourth inspector is a dedicated unpatterned wafer DF inspector, effectively replacing the legacy tool from Company A. Notch alignment occurs off the stage. The recipe selected a 45 incident beam. The stage spins the wafer to locate defects. Lastly, a fifth inspector is studied which is a DF patterned/unpatterned wafer inspector from Company B. Notch alignment is compensated for through software. The laser is at 60. The wafer spins on the stage during defect detection. The results are shown in the table, below.
140 120 100 80 um 60 40 20 0 Average x, y Displacement Inspector Degrees 0.07 0.06 0.05 0.04 0.03 0.02 0.01 0 Rotation Repeatability from Average Inspector 2-Sigma Precision / Figure of Merit 140 120 100 80 um 60 40 20 0 Inspector Legacy DF; Dedicated Unpatterned Modern DF; Patterned/Unpatterned Company "A" Modern BF; Patterned/(Unpatterned) Modern DF; Dedicated Unpatterned Modern DF; Patterned/Unpatterned Company "B" Figure 7: Comparisons of commercial defect scanners. Write File outputs two files: a summary of the results that can be used to redisplay any of the scatter plots; and a defect file, in the same format as the input data, with eighty positions reported. Each position is either the average position of the detected patterns at that site, or, if the pattern was never detected, the site position itself. The classification code for each site position is the number of times the pattern was detected at that site. This file, in conjunction with the wafer, can be used with a JEOL SEM to relocate the patterns and record their positions as found in the microscope. As the patterns are found and centered, a points table is generated that records both the average position as reported by the scanner, and the position in the SEM. JEOL defect review SEMs may include a nonlinear least-squares procedure, based on the Levenberg-Marquardt method 4, that adjusts a set of parameters that, when applied to the positions of defects as reported by a particular defect scanner, give a best fit to the positions of the defects in the SEM. The parameters include a x and y origin shift, a θ rotation, independent scaling factors in the x and y directions, and a non-orthogonality correction, adjusting for any difference in the direction of the y axes for the two stages when the x axes are superimposed. Figure 8 shows the refinement program, with a points table loaded and an initial refinement set for just x, y and θ set for refinement. The average difference between the predicted and observed positions is almost 250 microns.
Figure 8: The Levenberg-Marquardt Least-Squares program. Clicking Refine to Convergence reduces the average error to about 12 microns, as in Figure 9. Figure 9: After refinement of x, y and θ. If the x and y scaling factors and the non-orthogonality parameters are selected in addition, the refinement proceeds as shown in Figure 10, with a slight decrease in the average error.
Figure 10: Final results of refinement. Click on Last Cycle to show the standard deviations of the parameters, as shown in Figure 11. Figure 11: The final results of the refinement, with the alignment parameters and their standard deviations.
These parameters can then be saved to a file that the SEM will access when another defect file is loaded from the same scanner, applying the correction parameters to the predicted positions. The new defect file will still contain a set of random errors, but the systematic errors will have been essentially eliminated. As an example, consider the following. If the patterned wafer is loaded and scanned several times in a scanner, and the outputs compared in the DSA program, the results are shown in Figure 12. The figure of merit is 52.1 microns. If the average positions of the patterns are loaded in a review SEM, the pattern positions relocated and the two sets of positions saved to a file, the LMLS (SSBWA) program can compare them. Without any corrections to the predicted positions, the average discrepancy is 646 microns (see Figure 13). This is a measure of the systematic errors between. Figure 12: Figure of merit for the scanner Figure 13: Discrepancies between data sets before corrections. Figure 14: Agreement between predicted and actual defect coordinates after corrections.
Figure 15: Agreement between corrected scanner positions and positions relocated in the SEM. SUMMARY the coordinate systems of the scanner and the review SEM. If the six alignment parameters are now refined, adjusting the predicted coordinates to match the SEM coordinates more closely, the discrepancies average less than two microns (see Figure 14). If these corrections are saved and then applied to a wafer scanned in the same scanner, the average error between the corrected predicted positions and the positions of the defects as relocated in the SEM are, in this example, only 22 microns (see Figure 15). This error is almost entirely due to the random alignment errors made when the wafer was scanned, and is consistent with the two-sigma figure of merit for the scanner of 52 microns. Decreasing minimum defect sizes, absence of large defects and the need to eliminate operator intervention have all made defect relocation on an SEM more and more difficult. Two major problems are systematic differences between the defect review scanner and the SEM coordinate systems, and the random alignment errors made each time a wafer is processed by a defect scanner. A fast, accurate method is provided for evaluating and possibly aiding in reducing the random errors made by a defect scanner, and a method is also available that will correct systematic position differences between a particular defect scanner and a particular review SEM. These methods involve the use of a custom wafer with a standard pattern of defects at eighty sites across the wafer. If the defects are detected by the scanner, a computer program can locate the patterns and analyze the predicted positions for multiple scans. The same wafer can be examined in an SEM, and the pattern positions recorded. A second computer program can then compare the predicted and measured positions of the patterns, and calculate a set of six correction parameters that, when applied to the predicted positions, minimize the differences between the corrected and measured positions. Data is presented showing the range of random errors made by typical defect scanners, and how these errors have been reduced in newer scanners. Data is also given that shows the improvement in accuracy of the predicted positions of defects when corrections are made in advance for the systematic differences between the defect scanner and the SEM coordinate systems. REFERENCES 1. A. S. Parkes, W. M. LeMay, Method for Detection and Relocation of Wafer Defects, Pat. Pending. 2. N. Nagai, M. Tomohiro, Method for Inspecting Surface Foreign Matter of Semiconductor Wafers, Japan Patent No. JP10012686, January 16, 1998. 3. P. D. Kinney, Y. S. Uritsky, H. Q. Lee, Multiple-Scan Method for Wafer Particle Analysis, U. S. Patent No. 5,422,742, June 6, 1995. 4. W. H. Press, S. A. Teukolsky, W. T. Vetterling, B. P. Flannery, Numerical recipes in C, The Art of Scientific Computing, Second Edition, Cambridge University Press, 1988, 1992. * parkes@jeol.com, phone 978-536-2367, fax 978-536-2205. ** marek@jeol.com; phone 978-536-2494, fax 978-536-2205.