METHOD FOR CHARACTERIZING DEFECTS ON SEMICONDUCTOR WAFERS
CROSS-REFERENCES TO RELATED APPLICATIONS
The present application is related to the following commonly owned, co-pending U.S. Patent Applications: 1. "Laser Imaging System For Inspection and
Analysis of Sub-Micron Particles," by Bruce W. Worster, Dale E. Crane, Hans J. Hansen, Christopher R. Fairley, and Ken K. Lee, application serial number 08/080,014, filed on June 17, 1993;
2. "A Method and Apparatus for Performing an Automatic Focus Operation," by Timothy V. Thompson, Christopher R. Fairley, and Ken K. Lee, application serial number 08/183,536, filed on January 18, 1994;
3. "A Method and Apparatus for Automatic Focusing of a Confocal Laser Microscope," by Christopher R. Fairley, Timothy V. Thompson, and Ken K. Lee, application serial number 08/373,145, filed on January 17, 1995;
4. "Surface Extraction from a Three-Dimensional Data Set," by Ken K. Lee, application serial number 08/079,193, filed on June 17, 1993;
5. "Surface Data Processor," by Abigail A. Moorhouse, Christopher R. Fairley, Phillip R.
Rigg, and Alan Helgesson, application serial number 08/198,751, filed on February 18, 1994; and 6. "Automated Surface Acquisition For a Confocal Microscope," by Ken Kinsun Lee, application serial number 08/483,234, filed on June 7, 1995. The foregoing applications are incorporated herein by reference.
BACKGROUND
Defects in the form of structural flaws, process residues, and external contamination occur during the production of semiconductor wafers. Defects are typi- cally detected by a class of instruments called defect scanners. Such instruments automatically scan wafer surfaces and detect optical anomalies using a variety of techniques. The location of these anomalies with respect to the pattern of semiconductor devices on the wafer surface is recorded. This information, or "de¬ fect map," is stored in a computer file and sent to a defect review station.
Using the defect map to locate each defect, a human operator observes each defect under a microscope and characterizes each defect according to type (e.g., particle, pit, scratch, or contaminant) . Information gained from this process is used to correct the source
of defects, and thereby improve the efficiency and yield of the semiconductor production process. Unfortunately, people are relatively slow and are quickly fatigued by the highly repetitive task of observing and characterizing defects.
Methods of automatically characterizing defects, collectively known as Automatic Defect Characterization, or "ADC," have been developed to overcome the disadvantages of manual defect characterization. Conventional white-light-microscope- based review stations are automated to load a wafer that has been mapped for defect location by a defect scanner. Once the mapped wafer is loaded, the review station: l. positions the wafer to image the site of a defect, as indicated by the defect map;
2. focuses on the site of the defect;
3. captures a digital image of the site using a digital TV camera; 4. processes and analyzes the captured image of the site to locate the defect; and 5. further analyzes the data to characterize the defect. The above process is repeated for each defect (or a predetermined subset of defects) on the wafer. The wafer is then unloaded and the process is repeated for another wafer. By eliminating a fatiguing and highly
repetitive task, such automated review stations reduce labor costs and provide improved consistency and accuracy over human operators.
Conventional ADC systems capture a conventional white-light microscope image as an array A representing a two-dimensional image. The image is an x-y array of n by m pixels, where typical values might be n=640, m=480, or n=512, m=512. This array may be represented as:
A(x, y, lr, Ig, lb) ,
where x and y are pixel coordinates, and Ir, Ig, and lb represent the intensities of the red, green, and blue image components, respectively. Of course, grey scale images may also be used, as may other color schemes, such as those of the YUV and YIQ commercial standard formats. In the case of a gray scale image, a single intensity parameter Ig is used. In addition to imaging the defect site, at least one reference image A^ is also stored. The reference image may be a previously stored data-base image of a known-good area of the same or a similar die on the same or on a similar wafer, or it may be a specific image taken from, e.g., an adjacent die. The reference image is compared with the image containing the defect. Any differences measured between the two images will
indicate the location and extent of the defect.
Multiple reference images are usually required because slight differences in focus position between the reference and test images may cause false discrepancies to appear. In some cases, a separate reference image is not taken, and instead the reference image is a portion of the same image containing the defect, but from a region of the image where no defect occurs. In general, this latter method is faster but less reliable than methods that use a separate reference image, and works only for images containing repetitive structures or patterns.
Several conventional techniques are available to process images for automatic defect characterization. One such technique is described by Youling Lin, M.S., in Techniques for Syntactic Analysis of Images with Application for Automatic Visual Inspection, a dissertation in business administration submitted in December of 1990 to the graduate faculty of Texas Tech University in partial fulfillment of the requirements of the degree of doctor of philosophy, which is incorporated herein by this reference.
Lin describes ADC techniques for processing a two- dimensional microscope image. According to Lin, low- level image processing enhances surface features and reduces noise. This process is performed on intensity (gray scale) variations of the image. Lin describes an
extreme-median digital filter to accomplish this task.
Next, Lin describes techniques for identifying feature boundaries and converting the boundaries into a list of symbolic geometric "primitives." Suppose, for example, that a surface feature has the shape of a half-circle. Such a feature will have a boundary shaped approximately like the letter "D." This boundary could be converted into two geometric primitives; a line segment (specified by length and direction) representing the vertical portion of the "D," and an arc (specified by position and radius) representing the curved portion of the letter "D." More complex shapes may be similarly represented using a large number of connected line segments, angles, and arcs.
Symbolic geometric primitive extraction is performed, for example, by statistical comparison of the edge data with a representation of geometric primitives, such as line segments, arcs, or angles. The surface-feature boundary data is replaced with a set of primitives that best describes the boundary. The preceding steps are performed both for at least one reference image and for a test image. Then, using techniques derived from compiler theory, the set of reference primitives is compared, primitive by primitive, with the set of test primitives. When a discrepancy is encountered between the sets of
reference and test primitives, a rule-based expert system notes the discrepancy and continues the comparison. The discrepancies (i.e., the differences between the sets of reference and test primitives) define the location of a defect.
Alternatively, the defect area may be located by overlaying the test and reference images, aligning them by correlation techniques, and subtracting the images one from the other. Defects will show up as areas where the test and reference images have large difference values.
Having identified the location of a defect, the boundaries of the defect are identified and represented by a set of primitives in the manner described above for the test and reference images. In one embodiment, where more than one defect is located in a single image, only the defect with the largest area is selected for further processing.
Next, the set of primitives representing the image portion containing the defect is used to develop a set of defect parameters, each defect parameter representing a single feature of the defect. For example, one defect parameter may represent the area of the defect and another the shape of the defect. Moreover, characteristics of the area defined by the defect boundaries may be used to derive additional defect parameters. For example, the defect area may be
analyzed for average intensity, variations in intensity from one pixel to the next or within a small region ("texture") , color, or color coordinates. The defect parameters are conventionally expressed in a normalized form so that they run from, e.g., 0 to 1 or -1 to 1. A defect-parameter vector is then defined by these parameters.
The defect-parameter vector is compared, using conventional fuzzy logic techniques, with typical vectors for each known type of defect. Based on this comparison, the ADC system characterizes the defect and estimates the probability that the selected characterization is accurate. For a more detailed description of one method of developing a defect- parameter vector, see "Techniques for Syntactic
Analysis of Images with Application for Automatic Visual Inspection, " which is incorporated herein by reference.
For further discussion of conventional ADC techniques, see the IBM technical disclosure entitled "Automated Classification of Defects in Integrated Circuit Manufacturing, " by Frederick Y. Wu, et al., which is incorporated herein by this reference. Conventional ADC images have a number of shortcomings. For example, small pits versus particles cannot be distinguished, shallow structures are not discernible, and subsurface defects cannot be
characterized. And, if a defect or structure on a surface is "tall," focusing on one level leaves other levels out of focus. Accuracy of the automatic focus between the test and reference image then becomes critical because small variations in focus cause the boundary between two structures of different heights to change in appearance. A conventional ADC system may then interpret this variation as a potential defect when it is not. Human operators can compensate for this to some degree by, e.g. , moving the focus up and down and interpreting three-dimensional aspects of the images, but this wastes valuable time. Moreover, if there are low optical contrasts between the defect and the surrounding material (e.g., the defect is of approximately the same color or reflective intensity as the surrounding surface of the semiconductor) , an ADC scheme can fail to detect the true shape — or even the existence — of the defect. Therefore, what is needed is a more accurate method of automatically characterizing defects.
SUMMARY
The present invention involves Automatic Defect Characterization (ADC) with a resulting improved accuracy and efficiency over the prior art. In one embodiment, ADC is based on three-dimensional data, including white-light confocal images and laser-based
confocal images. The present invention also involves further extension of these techniques to Laser Feedback Microscopy (LFM) derived arrays, and "image" arrays developed from other techniques. In accordance with the present invention, a test surface is defined by a set of points on the test surface, the set of points being described by a Cartesian coordinate system having x, y, and z axes such that each point has a unique location described by x, y, and z coordinates.
To so define the test surface, the test surface is contained with a rectangular test volume described by the x, y, and z axes used to define the test surface. The rectangular test volume contains a superset of test points defined by incremental x, y, and z coordinates. Using confocal opticε, the test volume is scanned by a focussed beam of light so that the focal point of the beam coincides, in turn, with each point within the test volume. The intensity of reflected light returned for each point in the test volume is measured to obtain a data value representing the reflected intensity for that point.
Next, the Z value that resulted in a maximum reflected intensity value is determined for each column of z values (each represented by a unique x, y coordinate in the test volume) . In accordance with the principles of confocal optics, the measured intensity
of reflected light is greatest when the focal point of the beam is coincident with the surface. Therefore, the Z value that resulted in a maximum reflected intensity value for a given column of z values indicates the location of the surface point along the z axis (i.e., the elevation of the point).
In addition to the Z value corresponding to the maximum reflected intensity, the ADC system also determines and stores a value representing the maximum reflected intensity of each point.
The maximum reflected intensity value and the location along the z axis of each of the points on the test surface are stored as a set of test data representing a three-dimensional image of the test surface. From this three-dimensional image, the system extracts a set of geometric constructs, or "test primitives," that approximate features of the three- dimensional image of the test surface. This set of test primitives is compared to a set of reference primitives derived from a reference image to determine whether the set of test primitives is different from the set of reference primitives.
Differences between the test and reference primi¬ tives indicate the presence of a defect. When such differences exist, the ADC system generates a set of defect parameters from the differences between the set of test primitives and the set of reference primitives.
The defect parameters define a defect-parameter vector, which is matched with a knowledge base of reference defect-parameter vectors to determine the type of de¬ fect.
BRIEF DESCRIPTION OF THE FIGURES
Figure 1 is a flow chart that depicts the process of characterizing defects using a three-dimensional surface image in accordance with a first embodiment of the invention;
Figure 2 is a flow chart that depicts the process of characterizing defects using a two-dimensional top- view surface image in accordance with a second embodi¬ ment of the invention; and Figures 3A and 3B combine to provide a flow chart that depicts the process of characterizing defects using a three-dimensional volumetric data set in accordance with a third embodiment of the invention.
DETAILED DESCRIPTION
I. Three Dimensional Laser Confocal Surface Image ADC
Figure 1 is a flow chart that depicts the process of characterizing defects using a three-dimensional surface image obtained using a laser imaging system (LIS) in accordance with a first embodiment of the invention. (An LIS in accordance with the present invention is described in the above-identified patent
application entitled "Laser Imaging System For Inspec¬ tion and Analysis of Sub-Micron Particles," incorporated herein by reference.) Once the ADC pro¬ cess is initiated, the process begins at Step 11, "Image Acquisition."
Steps IIA and 11B: Image acquisition Images of a test surface and at least one reference surface are obtained using a confocal microscope, preferably a laser-based confocal microscope. The following describes the process of obtaining an image using a confocal microscope, and is applicable to both test and reference images.
Beginning at Step IIA, to ensure the correct settings of the z starting position and the z range, a preliminary scan is taken prior to the capturing of the test or reference image to determine the optimal z starting position and z range. The starting position and the z range are important. If the first surface scan begins too far above or below the surface of interest, the image slices may not cover the defect. If the range is too large, defects may be captured with insufficient resolution in the z direction. In one embodiment, this preliminary scan is accomplished using the set-Z function described in the above-identified application entitled "Automated Surface Acquisition For a Confocal Microscope," which is incorporated by
reference.
Next, at Step 11B the LIS generates an image of the test surface using a confocal microscope. To obtain an image of a surface using a confocal micro- scope, a beam of light passes through an objective lens and is scanned across the test surface from a number locations along a z axis. At each z location, the scanned laser beam generates a number of signals, each of the signals representing an intensity of light reflected through the objective lens from a given point on the surface. The group of signals provided by an x- y scan from a single z location of the objective lens is called a "slice" of intensity data. Slices taken from a number of locations along the z axis overlap to form a three-dimensional set of reflected intensity data, hereafter referred to as a "volumetric data set."
The overlapping slices of data create a column of data values for each point on the surface, each data value representing a reflected intensity of light from that point from a particular Z location. For each such column, data values are compared to determine the loca¬ tion along the z axis that resulted in a maximum reflected intensity. Because the intensity of the reflected light from a particular point is greatest when that point on the surface is coincident with the focal plane of the objective lens, the location of the objective lens on the z axis that corresponds to the
maximum reflected intensity gives an indication of the z coordinate of that point on the surface. In this way, the x, y, and z Cartesian coordinates are determined for each point on the surface. An image of the surface may then be generated from this information.
The volumetric data set from a laser imaging system (LIS) or a white-light confocal microscope may be represented as an array A(x, y, z, Iλ) . This array contains data representing reflected intensity (i.e.,
Iλ) for every x-y-z coordinate scanned. Different plane images P may be derived from the volumetric data set by making cuts through the array, such as a single horizontal slice PH(*X", y, 2, Iλ) , where Z represents a single value of z, or vertical planes such as PV(X, y, z, Iλ) or Pv(x, Y , z, Iλ) , where X and Y represent single values of x and y, respectively. Other planes may also be specified at any arbitrary orientation to the original x, y, and z axes. In addition to the Cartesian coordinate system described above, other systems, such as those using spherical or cylindrical coordinates, can be used to describe the volumetric data set.
A pair of surface arrays, Sj and Sz, may be derived from a volumetric data set extracted by a confocal microscope by determining, for each x-y coordinate pair, the maximum intensity value, I,^, and the Z
coordinate corresponding to the maximum intensity value. (For simple reflective surfaces, the confocal response of the LIS is a maximum at the surface.) The surface intensity array S, may be represented as:
Sι( , y, I ) ,
and the array of z coordinates corresponding to the maximum intensity values may be represented as:
Sz(- r Y , Zlmax) ,
where Z^^ represents the Z coordinate corresponding to the point of maximum reflected intensity for a given x- y coordinate.
In one embodiment, surface arrays Sj and Sz are updated as the ADC system scans the test surface from each Z position. Each intensity value of each slice of intensity data is compared to a maximum intensity value corresponding to the same x-y coordinate in the array of maximum intensity values. If the intensity value of the slice is greater than the corresponding maximum intensity value, then the intensity value of the array of maximum intensity values, Sj(x, y, I-^^) , is updated with a new maximum intensity value for that x-y coordinate and the array of Z values, Sz(x, y, Z^^) , is updated with the Z location of the new maximum
intensity value. Because the point of maximum reflected intensity gives an indication of the location of the surface, the array of Z values provides an indication of the surface contour. This second method is faster and requires less memory than is required for generating a complete volumetric data set.
For a more detailed description of a laser imaging system that employs a confocal microscope, see the co¬ pending application entitled "Laser Imaging System For Inspection and Analysis of Sub-Micron Particles," the content of which is incorporated herein by reference.
The aforementioned embodiments use three- dimensional images, such as surface array Sz(x, y, ■^imax) i derived from either a LIS or a white-light confocal microscope, to perform improved ADC. The use of three-dimensional images overcomes the x-y resolution, vertical resolution, and low-contrast problems that limit the performance of conventional ADC systems. It will become apparent to those skilled in the art that the present invention may be used to perform defect detection only, or, alternatively, to perform both defect detection and identification.
Step 12 : Low-level image processing After the test image and reference image are taken using the image extraction method described in connection with Step 11, the images are processed by
digital filtering to reduce noise and enhance boundaries. Here "texture" occurs both in intensity contrast and in surface height variations (also expressed as "roughness" or, alternately, "smoothness") . Also, structures are characterized by height differences as well as by contrast differences across structure boundaries. Hence, the filtering must handle both the intensity and the z-dimension values stored in the surface image array. Examples of appropriate filters include conventional low-pass filters, conventional median filters (as described in "The Image Processing Handbook, Second Edition, " by John C. Russ, CRC Press, 1995, which is incorporated herein by this reference) , and extreme-median filters, as described in "Techniques for Syntactic Analysis of Images with Application for Automatic Viεual Inspection . "
Step 13: Extraction of Geometric Primitives The intensity variables of the test image and reference image are processed to locate and reduce the boundaries of surface features. To accomplish this, the test and reference image intensity arrays, which are expressed as Sj(x, y, 1^^) as discussed above, are subjected to "thresholding and skeletonization," a conventional edge-detection technique. Other edge- detection techniques are commonly available, such as
"automatic tracing." These and other edge-detection techniques are detailed in "The Image Processing Handbook . " Such detectors are used to define the boundaries of the various surface features. Next, the defined feature boundaries are converted into a list of symbolic primitives that are combined to represent the boundaries. Conventional approaches to extraction of geometric primitives and analysis and characterization of shape boundaries are described in "Techniques for Syntactic Analysis of Images with Application for Automatic Visual Inspection . "
Next, the z variables of the test image and reference image are processed to locate and reduce the boundaries of surface features. The test and reference image z arrays, which are expressed as surface arrays in the form Sz(x, y, £faιax) as discussed above, are subjected to a conventional edge-detection technique, such as "automatic tracing" or "thresholding and skeletonization," to define the boundaries of the various surface features. Next, using conventional methods, the defined feature boundaries are converted into a list of symbolic primitives that are combined to represent the boundaries.
Step 14: Merging Intensity and Z Primitives
The two lists of primitives created in the last step are merged into a combined primitive list, so that
the combined list of primitives contains information indicative of the size, intensity, orientation, and location in x-y-z space of image features.
Step 15: X-Y Alignment
Using conventional auto-alignment techniques, the merged list of primitives from the test surface is shifted, compared, and matched with a similar list of primitives from the reference image. The x-y offset between test image and reference image is calculated using the x-y locations of the matched primitives. The x-y offset is then used to calculate the overlapping area between test image and the reference image.
Step 16: Defect Detection and Z Alignment
Once aligned, the lists of test and reference image intensity primitives are compared to determine whether differences exist between them. Any difference (e.g., any missing, extra, or misplaced geometric primitives) between the test list and the reference list indicates the presence of a defect.
The z axes of the test and reference surfaces are then aligned by subtracting the lowest z value in the overlapping area (i.e., the area shared by the test and reference arrays where no difference is detected) of the test array from all other z values in the overlapping area of the test array, and subtracting the
lowest z value in the overlapping area of the reference array from all other z values in the overlapping area of the reference array. Thus, the lowest z value in the overlapping area of each array is 0, so that absolute height is converted into relative height.
Step 17: Extraction of Defect Parameters Having identified a defect, defect boundaries are located, using conventional thresholding and skeletonization techniques. The boundaries are then conventionally represented by geometric primitives as described above. Neighboring defect primitives are grouped together and combined with their interior pattern to represent the area defined by the defect in x-y space as a single "high-level" defect.
The surface image provides up to three different lists of parameters for enhanced boundary detection and characterization: the intensity list, the z list, and the merged list of intensity and z lists. In one embodiment, the z list (and consequently the merged list) is not generated unless the intensity contrast of the surface under test is low, resulting in poor intensity image quality. Such selective use save valuable time. Any two or all three may be used, depending on the desired level of analysis.
Step 18: Defect Characterization Next, defect parameters are developed by analyzing various characteristics of the defect area as defined by a defect parameter dictionary. Such characteristics include image intensity, texture (i.e., local variations in brightness from one pixel to the next or within a small region) , boundary curvature, surface curvature, height (in z) , area, and shape. For example, if height is used as a defect parameter, then the height of the defect is assigned a magnitude conventionally expressed in a normalized form so that values run from, for example, 0 to 1 or -1 to 1. Other parameters are similarly assigned magnitudes, and these magnitudes are used to define a defect-parameter vector corresponding to the defect.
Similar defects have similar defect-parameter vectors. In other words, two defects with similar size, shape, texture, surface curvature, etc. , will have similar defect-parameter vectors, and are likely to be similar defects. Based on this premise, the present invention characterizes a defect by comparing the defect's defect-parameter vector, using conventional fuzzy-logic comparison techniques, with previously stored defect vectors for different types of defects. The closest vector match is used to characterize the defect. The present invention further estimates the degree of confidence in the
characterization based on the precision of the defect- parameter vector match.
For a more exhaustive analysis, additional defect parameters may be obtained by analyzing a vertical slice of data taken through the defect so that the vertical slice represents a vertical, two-dimensional cross-section of the defect. Such a cross-section may be used to define defect parameters representing, for example, cross-section area and boundary features. In another embodiment, additional defect parameters are defined by providing one or more silhouettes of the defect by looking "sideways" at the defect to determine the defect profile. To ensure that only the defect is considered, only those columns of data points (i.e., data points sharing a common x-y location) within the x-y defect boundary are considered.
For example, the defect may be viewed along a line parallel to the x axiε from the first y location of defect boundary to the laεt y location of the defect boundary. Only those data columns that vertically intersect the x-y defect boundary are considered. When looking at the defect from the x direction, the x element of the array is not considered. Therefore, the side view is effectively an infinite-depth-of-focus view of one side of the defect (i.e., a silhouette of the defect) . Additional defect parameterε, εuch as
profile shape, may be defined by the silhouette. Additional defect parameters may be obtained by providing similar silhouetteε from other angleε. Software for performing two-dimenεional ADC analogouε to the three-dimensional process deεcribed with reference to εteps 12 through 18 above is available from ISOA, Inc. of Richardεon, Texas. However, the software must be modified to accommodate the additional information provided by the z data. Such modificationε will be readily understood by those skilled in the art.
Importantly, the availability of z data provides additional defect parameters. Theεe include (1) physical roughness (small z variations) of the defect area, (2) z contraεt variationε (i.e., large z variations) , (3) sign of the relative z value of the defect region compared with the reference, which may be used, for example, to distinguish between pits and particles, (4) actual z valueε of the defect region relative to the reference, (5) surface slope, determined by relative z values of adjacent x-y location. Moreover, using the defect crosε-section and silhouette techniques described above, defect parameters may be developed to represent surface curvature and cross-sectional area.
By adding one or more of the above-listed parameters to the defect vector characterizing a given
defect, additional information about the nature of the defect and its relationship to its environment can be obtained. Even small differences in surface height over a limited region with reεpect to the reεt of the image can indicate buried defectε of non-trivial thickneεε that have caused an overlying layer to bulge upward. And, small surface depressions can indicate subsurface voids in prior layers. Because the defect- parameter vector contains additional and unique information, the accuracy of the characterization is improved over two-dimenεional ADC, and the number of different defect types that can be characterized is ex¬ panded over the conventional microscope image analysis.
II. Top-View Laser Image ADC
Figure 2 is a flow chart that depictε the proceεs of characterizing defects using a two-dimensional, top- view surface image in accordance with a second embodi¬ ment of the invention. The three-dimensional surface image array S described above in connection with Figure 1 may be reduced to image array S'(x, y, Iλmax) by eliminating the z value. Image array S* is the equivalent of an "infinite" depth of field top-view image of the εurface (i.e., all points are in focus, regardleεε of the height) . Thiε new image differs from conventional two- dimensional images in that the new image has improved
resolution and infinite depth of field. The improved resolution provides more information for ADC analysis. In addition, because surface characteristics that would be out of focus in a conventional image are now in focus, the array S1 allows for evaluation of, e.g., texture, without obtaining additional imageε.
A similar "infinite" depth of focus image, either grey scale or color, can be obtained from a white-light confocal microεcope focus expoεure εeries, but chromatic aberration and other resolution limitationε of thiε procedure affect quality of the result. Still, it would be an improvement over a single conventional image, and could be used in ADC.
The top-view laser image (or top-view confocal image) may be used in place of the conventional video camera image to provide a two-dimensional array with improved reεolution and no focuε diεcrepancy problemε between teεt and reference images. Hence, only one defect-free reference image is required for non- repeating pattern images, rather than multiple references to obtain the best focuε compariεon. For repeating pattern imageε, the correlation of image characteristics will be more consiεtent from one εample to the next, resulting in more accurate defect characterization.
Upon initiation of the ADC process, the LIS begins at Step 21A, "Image acquiεition. " Thiε and the
following steps of Figure 2 are similar to those described above in Figure 1 for three-dimensional surface ADC, except that the elimination of the z data simplifies each step of the process and eliminates the step of merging primitives. The two-dimensional data may then be processed to characterize defects using ADC software available from ISOA, Inc. of Richardεon, Texas.
III. Three-Dimensional Laεer Confocal Volume Image ADC In the caεe of opaque εurfaceε, a three- dimensional surface image obtained using a LIS contains almost as much information as the full-volume image represented by the volumetric data set. However, when viewing transparent or εemi-transparent structures, such as dielectric films, a simple surface extraction is often insufficient to obtain a realistic representation of the wafer surface. Thiε is because light from the imaging syεtem can paεs through transparent structures to be reflected from subsurface layers. Three-dimensional procesεing provideε an important advantage over conventional ADC syεtemε, which cannot distinguish a large class of defects that are embedded or lie below or within transparent layerε. Figureε 3A and 3B combine to provide a flow chart that depicts the process flow of an ADC system that uses three-dimensional volume imaging in accordance
with a third embodiment of the invention. Once the ADC process is initiated, the process begins at Step 31A, "Image Acquiεition. "
Stepε 31A and 3IB: Image Acquiεition
The set of test images (slices of x-y intensity images) , and the set of reference imageε where needed, are generated in a manner similar to that aε deεcribed for acquiring the top-view laser image. In this case, the raw, unprocessed sliceε of data, which make up a volumetric data εet, are εtored in memory aε array V(x, y, z, Iλ) .
Step 32: Low-level image proceεεing In any column of data pointε parallel to the z axis (i.e., for fixed values of x and y) , the volume image contains the confocal z response of the system convolved with the reflectance of each surface encounter, combined with the absorption and scattering of each layer traversed by the light. Local peakε in intensity appear corresponding to each εurface, and intensity is modulated by the reflectivity of each interface, as well as by attenuation of the light during its round-trip through the wafer surface media. The first order of procesεing is to sharpen the peaks of intensity correεponding to εurfaces. In one embodiment, conventional deconvolution techniques are
applied along each column of data pointε parallel to the any z axiε, removing the confocal z reεponse. In another embodiment, conventional deconvolution techniques are applied over a vertical plane cut through the volumetric data set. Finally, conventional deconvolution techniques may be performed in three dimensionε, compenεating for contributions from surrounding surfaces, and operated over subsets of the volumetric data set large enough to provide good correction but small enough to allow proceεsing in a reasonable time.
Alternately, the peaks of intenεity may be εharpened applying conventional two-dimensional sharpening filters to any set of vertical planeε cut through the volumetric data εet, or by applying conventional one-dimenεional εharpening filters along one or more columnε (i.e., unique x-y coordinate) of data. Thiε iε done to define (1) multiple εurfaceε within the volumetric data εet that delineate multiple layers and elevations of the wafer structure, and (2) the surfaces of any defects.
Steps 33-37 Stepε 33-37 of Figure 3 are performed in much the same way as steps 13-17 described above in connection with three-dimensional laser confocal surface image ADC. However, because the volumetric image typically
provides an indication of more than one surface, for example a surface underlying a transparent or semitransparent film, a principal surface must be selected for analysis. The principal surface selected is typically the top surface. Alεo, due to the need for additional volume information, the entire volumetric data εets for the test and reference surfaceε must be maintained.
The preceding steps may provide sufficient information to characterize a defect. However, the data obtained may be insufficient to characterize — or even detect — subεurface defectε. In one embodiment, if the data obtained through Step 37 iε sufficient to characterize a defect with an acceptable degree of accuracy, the ADC syεtem moves directly to Step 11, Defect Characterization.
Step 38: Volumetric Defect Detection The test and reference intenεity data sets are expresεed as arrays Vtest(x, y, z, Iλ) and Vref(x, y, z,
Iλ) , reεpectively. To find a defect in the volume of the teεt array Vtest, the data value εtored for each voxel (i.e., each x, y, z coordinate) iε εubtracted from the data value of the correεponding voxel (i.e., the same x-y-z coordinate) of the reference array Vrcf. The location and extent of one or a group of differences indicates the location and extent of a defect. In one
embodiment, a difference doeε not indicated a defect unleεε the value of the difference data exceedε a predetermined threεhold. The reεultε of the εubtraction are εtored aε a new volumetric array Vdef that represents the defect in three dimensions.
Step 39: Grouping of Defect The boundaries of the defect are defined by determining a minimum volume within the volumetric defect array Vdef that encompasses all of the difference data. This is accomplished by, for example, (1) defining the two-dimensional defect boundaries of each slice, using a conventional edge-detection technique, such as "automatic tracing" or "thresholding and skeletonization, " and (2) combining the overlapping two-dimenεional boundaries to form a three-dimensional boundary.
Step 40: Extraction of Defect Parameterε In addition to the defect parameterε extracted in Step 37, the additional volume data may be used to provide important additional defect parameters. For example, defect parameterε may be defined for the εize, εhape, reflectivity, tranεparency, and depth of a subsurface element.
Step 41: Defect Characterization One or more εubsurface defect parameters are added to the list of defect parameterε uεed to define the defect-parameter vector. The defect-parameter vector iε then compared, uεing conventional fuzzy logic techniques, with typical vectors for each known kind of defect. Based on this comparison, the ADC εyεtem char- acterizeε the defect and estimates the probability that the selected characterization is accurate. Because the extended defect-parameter vector contains additional and unique information, an ADC syεtem in accordance with the present invention has the ability to identify and characterize εubsurface defects. Moreover, because structureε are created on a wafer over time, the depth parameter for a defect can provide an indication of the time in the process at which the defect occurred. Such timing information may prove invaluable in determining the source of a defect. The preεent invention thus provides functionality unavailable in conventional microεcope image analysis.
IV. Three Dimensional Laεer Confocal Phaεe Image ADC
In addition to laser confocal imaging, additional information and reεolution, eεpecially in microεcopic surface and volume εcanning, may be obtained by measur¬ ing the phase of reflected light. One unique method is Laser Feedback Interferometry (LFI) , also called Laser
Feedback Microscopy (LFM) or Laεer Amplified Motion Detection and Analysis (LAMDA) . In LFI, light reflect¬ ed from a surface to be imaged reenters the laser cavi¬ ty whence it originated. The reflected light is then subjected to the amplification provided by the laser gain mechanism, providing a unique method of detecting reflected light over a large dynamic range and with great precision.
Next, the phase difference Δø and amplitude dif- ference Δlλ iε measured between the original and reflected light. The phase and amplitude difference data may be stored as arrays (e.g., Aψ(x, y, z, Iλ, φ) or Aφ(x, y, Iλ, φ) ) . (Here Iλ refers to the intensity ratio at a εpecific laser wavelength, λ. Lasers having multiple wavelengths may also be utilized, with an array for each wavelength taken either simultaneouεly or consecutively.)
All of the techniques deεcribed above for LIS volume, εlice, and surface imageε may be extended to the LFM arrays and their derivatives, such as A (x, y, ø) . LFI/LFM/LAMDA technology haε been deεcribed by Bearden, et. al. in a paper entitled "Imaging and Vibrational Analysis with Laser-Feedback Interferometry, " (Univerεity of California at Berkeley, Graduate Group in Biophysics and Division of
Neurobiology, 1992.) The Bearden paper is incorporated herein by this reference.
Additional analytical information may be obtained with microεcopic images, either conventional or confo¬ cal (or in combination) , using techniques for obtaining white light or laser fluorescence images of the test and reference surfaceε, techniqueε for obtaining laεer Raman εpectra of the defect/contaminant versus its background (both confocal and conventional) , and a variety of special images in polarized light, darkfield illumination, phase contrast images. Differential In- terference Contrast (DIC, also called Nomarεki images) , etc. , each of which can provide additional parameters that can be asεociated with the confocal or confocal image arrays described above.
In the case of the three-dimensional volume image array aε obtained with the LFM (or related combined microscope/interferometer combinations) , the phaεe data provideε an additional dimenεion for the array.
V. Three-Dimenεional Laεer Confocal Volume Image ADC Supplemented with Phyεical Analvεiε
The ADC methods described above can identify many kinds of εtructural defectε and the preεence of contaminant particleε or residue by analyzing physical attributes of the defect. Another embodiment of the invention provides additional information about the chemical composition of a defect. Such information often provides important clues as to the origin of a defect.
To analyze the chemical compoεition of a defect area, scattered laser light from the affected region may be automatically analyzed with a Raman spectrometer to identify both the Raman and/or fluorescence properties of the defect. (Fluoreεcence alone can be analyzed with optical filters and detectors, and does not require an optical εpectrometer. Alεo, x-ray fluorescence, EDX, or related techniques may be performed in a scanning electron microscope (SEM) or similar instrument, providing additional compoεition information. )
Additional defect parameters are defined uεing one or more of the foregoing techniques to provide composition information. Using these parameterε to calculate the defect-parameter vector further enhances defect discrimination.
VI. Conclusion
While the present invention haε been deεcribed in connection with εpecific embodiments, variations of these embodiments will be obvious to thoεe having ordinary εkill in the art. For example, while the invention iε described in connection with microscope that reflectε a maximum intenεity to the photodetector during a focuεed condition, it iε clear that the invention may be modified to operate with a microscope that reflects a minimum intensity to the photodetector
during a focuεed condition. Further, microscopes that do not utilize visible light, εuch aε infrared, ion beam, or electron beam microεcopes, could also benefit from the foregoing methods. Therefore, the spirit and scope of the appended claims should not be limited to the description of the preferred versionε.