WO2020112922A1

WO2020112922A1 - Defect classification by fitting optical signals to a point-spread function

Info

Publication number: WO2020112922A1
Application number: PCT/US2019/063478
Authority: WO
Inventors: Soren KONECKY; Bjorn BRAUER
Original assignee: KLA Corp
Current assignee: KLA Corp
Priority date: 2018-11-30
Filing date: 2019-11-27
Publication date: 2020-06-04
Anticipated expiration: 2021-05-30
Also published as: TWI832931B; KR20210087100A; IL282902B1; IL282902B2; US20200175664A1; CN113039631A; TW202038192A; CN113039631B; IL282902A; US10957035B2; KR102828290B1

Abstract

A semiconductor die is inspected using an optical microscope to generate a test image of the semiconductor die. A difference image between the test image of the semiconductor die and a reference image is derived. For each defect of a plurality of defects for the semiconductor die, a point-spread function is fit to the defect as indicated in the difference image and one or more dimensions of the fitted point-spread function are determined. Potential defects of interest in the plurality of defects are distinguished from nuisance defects, based at least in part on the one or more dimensions of the fitted point-spread function for respective defects of the plurality of defects.

Description

Defect Classification by Fitting Optical Signals to a Point-Spread Function

RELATED APPLICATION

[0001] This application claims priority to US Provisional Patent Application No.

62/773,834, filed November 30, 2018, titled“Method to Distinguish Point-Like Defects of Interest From Extended Process Variation by Fitting Defect Signals to a Point Spread Function,” which is hereby incorporated by reference in its entirety for all purposes.

TECHNICAL FIELD

[0002] This disclosure relates to semiconductor inspection, and more specifically to classifying defects detected by semiconductor inspection.

BACKGROUND

[0003] Modem optical semiconductor-inspection tools use wavelengths that are significantly longer than the dimensions of a typical defect, often by an order of magnitude or more. As such, inspection tools cannot resolve the defects arid tints cannot provide images showing the defects; instead, the inspection tools merely provide an indication that a defect has been detected. Furthermore, many of the detected defects are so-called nuisance defects that do not impact device functionality and are not of interest to process-integration and yield- improvement engineers. In addition, nuisance defects may outnumber defects of interest, for example by a factor of 1000 or more. The high volume of nuisance defects makes it impractical to perform subsequent failure analysis (e.g., visualization using a scanning electron microscope) on all identified defects. The high volume of nuisance defects also makes it impossible to determine whether a wafer should be scrapped or reworked due to a high number of defects of interest. SUMMARY

[0004] Accordingly , there is a need for improved methods and systems of distinguishing defects of interest from nuisance defects. Such methods and systems may involve fitting optical- inspection results to a point-spread function.

[0005] In some embodiments, a method of identifying semiconductor defects of interest includes inspecting a semiconductor die using an optical microscope to generate s test image of the semiconductor die. The method also includes deriving a difference image between the test image of the semiconductor die and a reference image and, for each defect of a plurality of defects for the semiconductor die, fitting a point-spread function to the defect as indicated in the difference image and determining one or more dimensions of the fitted point-spread function.

The method further includes distinguishing potential defects of interest in the plurality of defects from nuisance defects, based at least in part on the one or more dimensions of the fitted point- spread function for respective defects of the plurality of defects.

[0006] In some embodiments, 8 method of identifying semiconductor defects of interest includes inspecting a semiconductor die using an optical microscope to generate a test image of the semiconductor die. The method also includes deriving a difference image between the test image of the semiconductor die and a reference image, and fitting a summation of a plurality of point-spread functions to the difference image. Each point-spread function of the plurality of point-spread functions is centered on a distinct predefined location in the semiconductor die and has a fixed width associated with the optical microscope. Performing the fitting includes determining parameters (e.g., coefficients) of respective point-spread functions of the plurality of point-spread functions. The method further includes distinguishing potential defects of interest in a plurality of defects for the semiconductor die from nuisance defects, based at least in part on the parameters (e.g., coefficients) of the respective point-spread functions.

[0007] In some embodiments, a non-transitory computer-readable storage medium stores one or more programs for execution by one or more processors of a semiconductor-inspection system that includes an optical microscope (i.e., a semiconductor-inspection tool). The one or more programs include instructions for causing the optical microscope to inspect a semiconductor die, to generate a test image of the semiconductor die. The one or more programs also include instructions for performing the other steps of either or both of the above methods.

BRIEF DESCRIPTION OF THE DRAWINGS

[0008] For a better understanding of die various described embodiments, reference should be made to the Detailed Description below, in conjunction with the following drawings.

[0009] Figures 1-3 show flowcharts of methods of identifying semiconductor defects of interest in accordance with some embodiments.

[0010] Figure 4 is a scatterplot showing a distribution of standard deviations m two directions for defects on a die, in accordance with some embodiments.

[0011] Figure 5 is a block diagram of a semiconductor-inspection system in accordance with some embodiments.

[0012] Like reference numerals refer to corresponding parts throughout the drawings and specification.

DETAILED DESCRIPTION

[0013[ Reference will now be made in detail to various embodiments, examples of which are illustrated in the accompanying drawings. In the following detailed description, numerous specific details are set forth in order to provide a thorough understanding of the various described embodiments. However, it will be apparent to one of ordinary skill in the art that tire various described embedments may be practiced without these specific details. In other instances, well-known methods, procedures, components, circuits, and networks have not been described in detail so as not to unnecessarily obscure aspects of the embodiments.

[0014] An optical semiconductor-inspection tool (i.e., an optical microscope) identifies defects on semiconductor die by illuminating portions of the die with an optical beam, which is scattered by structures on the surface of the die. The scattered optical beam is collected and imaged. The resulting image typically does not resolve defects, because die defects are much smaller than the diffraction limit for die wavelength(s) of light used in the optical beam. However, the presence of a defect causes variation in the optical signal of the scattered optical beam. This variation can be detected and the location of potential defects can thus be identified.

[OOlSj Defects identified in this manner include both defects of interest and nuisance defects. Defects of interest are typically isolated and localized (e.g., point-like), whereas nuisance defects may be spatially extended (e.g., as the result of process variation). This distinction may be exploited to distinguish defects of interest from many nuisance defects, by fitting the optical signal as imaged by the: optica! microscope to a point-spread function. (A point-spread function is an indication of the impulse response of the optical inicroscope, with a shape in the image that represents an unresolved defect.) The resulting fit gives an indication of the dimensions, and thus the size, of the defect. Defects can then be classified as either defects of interest or nuisance defects, based at least in part on their sizes as inferred through the fit

[0016] Figure I shows a flowchart of a method 100 of identifying semiconductor defects of interest in accordance with some embodiments. The method 100 may be performed using the semiconductor-inspection system 500 (Figure 5). Steps in the method 100 may be combined or broken out.

[0017] In the method 100, a semiconductor die is inspected (102) using an optical microscope (e.g., die inspection tool 504, Figure 5) to generate a test image of the semiconductor die. (The test image may be an image of the entire die or a portion of the die.) A difference image between the test image and a reference image of the semiconductor die is derived (104). The difference image, which may be derived by subtracting the reference image from the test image or vice versa, indicates degrees of difference between the test image and the reference image on a pixel-by-pixel basis. These degrees of difference may be referred to as gray levels. The larger the magnitude of a degree of difference (i.e., the larger the gray-level magnitude), the more likely a given pixel corresponds to a defect on the semiconductor die.

[0018] In some embodiments, the reference image is an image of a neighboring die on the same wafer, an image of any other die on the same wafer, or an image of a die on another wafer. In some embodiments, the reference image is a combination (e.g., a median) of images of multiple die (e.g., multiple adjacent die, die on the same wafer, and/or die on other wafers). In some embodiments, the reference image is derived from the design of the die, for example by running the die’s layout (e.g., as specified in a gds file) through a simulator that simulates imaging by fee optical microscope in some embodiments, for an array area on the die (i.e., an area on fee die in which the same pattern is repeated), fee reference image is a neighboring instance of fee pattern or a combination of neighboring instances of the pattern. The repeating pattern may be a memory cell or group of memory cells.

[0019] A plurality of defects for fee semiconductor die is identified (106) in the difference image of step 104, or in a separate difference image. If a separate difference image is used, the separate difference image may have been generated using a different reference image

(e.g., of any of the types of reference images described for step 104) and/or results of a different inspection (e.g., using a different optical mode than fee optical mode used in the inspection of step 102). Defects may be identified by determining whether fee corresponding gray levels in fee difference image satisfy one or more criteria (e.g., have magnitudes that exceed, or equal or exceed, specified thresholds). While shown as being performed after step 104, step 106 may alternatively be performed before step 102 or between steps 102 and 104, ifa separate difference image is used.

[0020] For each defect of the plurality of defects, a point-spread function is fit (108) to the defect as indicated in the difference image of step 104. One or more dimensions of fee fitted point-spread function are determined. The one or more dimensions provide a measure of fee size of fee defect (e.g., of widths of the defect in respective directions).

[0021] In some embodiments, the point-spread function ½ (110) a two-dimensional Gaussian function:

where A is a coefficient feat indicates the maximum intensity, x₀ and y₀ are coordinates of fee center of fee point-spread function and of fee defect (and correspond to coordinates of fee pixel on which fee defect is centered in fee difference image), and a_x and o_y are standard deviations of fee point-spread function in two orthogonal directions. The one or more dimensions include first and second dimensions (e.g., standard deviations a_x and <r_y, or full widths at half-maximum in fee x- and y-directions) indicative of widths of the fitted Gaussian function in respective first and second directions. [0022] In some embodiments, the point-spread function is (112) a sine function (i.e., cardinal sine function), polynomial function, or other analytical function. In some embodiments, the point-spread function is (112) a numeric simulation, such that step 108 includes fitting the numeric simulation to the defect as indicated by the corresponding signal in the difference image of step 104.

[0023] In some embodiments, the one or more dimensions include (114) first and second distances between maximal gradients of die fitted point-spread function in respective first and second directions (e.g., in the x- and y-directidns). Hie distance between maximal gradients in a particular direction is another indicator of defect width.

[0024] In some embodiments, the one or more dimensions include (116) first and second areas under cross-sections of the fitted point-spread function in respective first and second directions (e.g., in the x- and y-directions) through a maximum of the fitted point-spread function, normalized by the height of die maximum.

[0025] Based at least in part on the one or more dimensions of the fitted point-spread function for respective defects of the plurality of defects, potential defects of interest in the plurality of defects are distinguished (118) from nuisance defects.

[0026] In some embodiments, distinguishing (118) potential defects of interest from nuisance defects includes determining, for a particular defect, whether a first dimension of the one or more dimensions of the fitted point-spread function exceeds, or equals or exceeds, a first threshold. If the first dimension exceeds, or equals or exceeds, the first threshold, then the defect is classified as a nuisance defect. Distinguishing (118) potential defects of interest from nuisance defects may also or alternatively include determining, for a particular defect, whether the first dimension is less than, or less titan or equal to, a second threshold, wherein the second threshold is less than the first threshold. If the first dimension is less than, or less titan or equal to, the second threshold, then the defect is classified as a nuisance defect. Similar determinations may be made for a second dimension of the one or more dimensions (e.g., with first and second thresholds equal to or distinct from the first and second thresholds used for the first dimension), with defects being classified accordingly.

[0027] Figure 4 is a scatterplot 400 showing a distribution of dimensions <t_x and a_y for defects on a die, as determined in step 108, in accordance with some embodiments. Defects of interest ate shown as black asterisks, while nuisance defects arc shown as gray dots. The defects of interest are situated in a rectangle defined by upper a_y threshold 402, lower a_y threshold 404, lower s_c threshold 406, and upper s_c threshold 408. The thresholds 402, 404, 406, and 408, which may also be referred to as cut lines, may be applied as described above for the first and second thresholds. Accordingly, in some embodiments, the method 100 classifies as a nuisance defect any defect that falls outside of the rectangle. Any defect that fells within the rectangle may be classified as a potential defect of interest, or may be further analyzed based on other factors. While some of the defects within die rectangle are actually nuisance defects, application of the thresholds 402, 404, 406, and/or 408 still dramatically reduces the number of nuisance defects from the candidate set of potential defects of interest.

[0028] In some embodiments in which the one or more dimensions include two dimensions (e.g., corresponding to the x- and y-axes), distinguishing (118) potential defects of interest from nuisance defects includes determining, for a particular defect, whether a metric that is a function of the two dimensions satisfies a threshold. If the metric does not satisfy the threshold, then the defect is classified as a nuisance defect. The metric may correspond to a particular shape (e.g., circle, ellipse, etc.) in a plane (e.g., the a_x-c_y plane of Figure 4), where defects that do not satisfy foe threshold are those defects feat fell outside of fee shape.

[0029] In some embodiments, distinguishing (118) potential defects of interest from nuisance defects includes providing the one or more dimensions of the fitted point-spread function for the respective defects of the plurality of defects as input to a machine-learning algorithm trained to distinguish potential defects of interest from nuisance defects. The machine- learning algorithm may have been trained using a training set of dimensions determined from fitted point-spread functions for known defects of interest and known nuisance defects (e.g., as identified through scanning-electron microscopy and/or other failure-analysis techniques).

[0030] In the method 100, the defects are identified (i.e., detected) before die fitting is performed. Alternatively, fitting may be performed before the defects are detected. Figure 2 shows a flowchart of such a method 200 of identifying semiconductor defects of interest in accordance with some embodiments. The method 200, like the method 100, may be performed using the semiconductor-inspection system 500 (Figure 5). Steps in the method 200 may be combined or broken out. [0031] In die method 200, a semiconductor die is inspected (102) and a difference image between the test image and a reference image is derived (104), as described for the method 100 (Figure 1).

)0032) For each pixel of a plurality of pixels in the difference image, a point-spread function is fit (208) to a location in the difference image corresponding to the pixel. The location includes the pixel and and other nearby pixels (e.g., surrounding pixels) within the width of the point-spread function of the optical microscope. One or more dimensions of tire fitted point- spread function ate determined. The fitting of step 208 may be performed in the same manner as for the step 108 of the method 100 (Figure 1). For example, step 208 may include step 110, 112, 114, or 116. The plurality of pixels for which the point-spread function is fit may be every pixel in the reference image or a subset of pixels that are identified as candidate pixels (e.g., ail pixels in die reference image with local maxima and/or minima gray-level values). Because the plurality of pixels in step 208 includes pixels that correspond to defects, step 208 is an example of step 108 of the method 100 (Figure 1).

[0033] Based at least in part on the one dr more dimensions, pixels in the plurality of pixels with signals due to potential defects of interest in a plurality of defects for the

semiconductor die are distinguished (218) from pixels with signals due to nuisance defects. Step 218 provides a way to distinguish potential defects of interest from nuisance defects, as in step 118 of the method 100 (Figure 1).

[0034] In some embodiments, pixels that do not satisfy a criterion are removed (220) from a set of candidate pixels in the reference image. The criterion is based on at least one of the one or more dimensions. For example, the criterion may be whether pixels satisfy one or more of the thresholds 402, 404, 406, and 408 (Figure 4), or whether a metric that is a function of multiple dimensions satisfies a criterion (e.g., whether pixels fell within a corresponding shape in a plane, such as the s_c-s_g plane of Figure 4). In some embodiments, degrees of difference for respective pixels of the difference image in the set of candidate pixels are adjusted (222) using a criterion that is based on at least one of the one or more dimensions of the fitted point-spread function for the respective pixels. The magnitude of a degree of difference may be decreased if the one or more dimensions of the fitted point-spread function correspond to a likelihood that a defect is not a defect of interest, and/or may be increased if the one or more dimensions of the fitted point-spread function correspond to a likelihood that a defect is a defect of interest. After performing steps 220 and/or 222, the potential defects of interest are identified (224) from the set of candidate pixels. Step 220 removes pixels that probably correspond to nuisance defects from the set of candidate pixels. Step 222 makes it less likely that pixels that probably correspond to nuisance defects are identified in step 224, and/or makes it more likely that pixels that probably correspond to defects of interest are identified in step 224. in either or both of these manners, potential defects of interest are distinguished from nuisance defects, since the ratio of defects of interest to nuisance defects identified in step 224 is improved.

[0035] In some embodiments, the methods 100 and/or 200 further include determining a goodness of fit of the fitted point-spread function for respective defects of the plurality of defects. For example, the goodness of fit may be determined by the sum of the square of residuals (the“R-squared” metric) for the fit or by the largest single residual for the fit.

Distinguishing (118) the potential defects of interest from nuisance defects may be further based at least in part on the goodness of fit for the respective defects: increased goodness of fit is directly correlated with an increased probability that a defect is a defect of interest and not a nuisance defect, and defects are classified accordingly. For example, if the goodness of fit does not satisfy a threshold, the corresponding defect may be classified as a nuisance defect or may be excluded from a set of candidate pixels. In another example, a degree of difference may be adjusted based on the goodness of fit, by analogy to step 222 (e.g., such that a low goodness of fit causes one or more corresponding pixels to have their gray levels adjusted so feat they are less likely to be identified as defects).

[0036] Figure 3 shows a flowchart of yet another method 300 of identifying

semiconductor defects of interest in accordance with some embodiments. The method 300, like the methods 100 (Figure 1) and 200 (Figure 2), may be performed using the semiconductor- inspection system 500 (Figure 5). Steps in the method 300 may be combined or broken out.

[0037] In the method 300, a semiconductor die is inspected ( 102) and a difference image between fee test image of the semi conductor die and a reference image is derived (104), as described for the method 100 (Figure 1 ).

[0038] A summation of a plurality of point-spread functions is fit (308) to fee difference image. Each point-spread function of fee plurality of point-spread functions is centered on a distinct predefined location in the semiconductor die and has a fixed width associated with die optical microscope. (The method may allow for a degree movement of die predefined locations in accordance with process variation.) Determining the fit includes determining coefficients of respective point-spread functions of the plurality of point-spread functions in the summation.

[0039] For example, a summation of a plurality of Gaussian point-spread functions may be used:

where A* is a coefficient to be determined, Xot and $¾t «re coordinates for a respective predefined location, s_c and <r_y are standard deviations of the point-spread function in the x- and y- directions, and N is the number of distinct predefined locations and thus the number of point- spread functions. The values of s_c and o_y are fixed (e.g., known values for the inspection tool 504, Figure 5, are used). For example:

where A is the optical-inspection wavelength,

is die numerical aperture of the illumination system in the inspection tool, and N is the numerical aperture of the

collection system in the inspection tool.

[0040] Gaussian point^spread functions are merely one example of point-spread functions that may be summed in the method 300. Other examples (e.g., as described for foe step 108 of he method 100, Figure 1) are possible.

[0041] Based at least in part on die coefficients of the respective point-spread functions, potential defects of interest in a plurality of defects for the semiconductor die are distinguished (318) from nuisance defects. A coefficient (e.g., an At value) with a large magnitude as compared to other coefficients (e.g., other A, values) suggests that a defect of interest is present at the location for the corresponding point-spread function. A coefficient that is in the same range as other coefficients, however, is suggestive of a nuisance defect (e.g., resulting from process variation). A defect therefore may be classified as a defect of interest if its

corresponding coefficient (or the magnitude thereof) satisfies (e.g„ exceeds, or equals or exceeds) a threshold, and may be classified as a nuisance defect if its corresponding coefficient (or foe magnitude thereof) does not satisfy foe threshold. The threshold may be predefined or may be a function (e.g., mean or median) of other coefficient values (e.g., other Ai values).

[0042] The method 300 may be performed in conjunction with die method 100 or 200, such that defects may be classified as defects of interest based, at least in part, on dimensions of fitted point-spread functions as determined in the method 100 or 200 and also on point-spread- function coefficients as determined in the method 300.

[0043] In some embodiments of foe methods 100, 200, and/or 300, a report specifying the potential defects of interest is generated. For example, the report may list all of the defects (e.g., with their coordinates) and specify whether each has been classified as a potential defect of interest or a nuisance defect. Alternatively, the report may list the potential defects of interest (e.g., with their coordinates) and omit die nuisance defects. The report may be graphical; for example, the report may show a map of the die with indications of the locations of the potential defects of interest. The report may be displayed (e.g., on display 511, Figure 5) and/or transmitted to a client device for display.

[0044] The methods 100, 200, and 300 thus allow potential defects of interest to be distinguished from nuisance defects despite die feet that the defects may not be resolved in the difference image. In some embodiments of the methods 100, 200, and/or 300, a decision whether to scrap, rework, or continue to process a wafer is made based at least in part on the identified potential defects of interest. In some embodiments of the methods 100, 200, and/or 300, failure analysis (e.g., including scanning-electron-microscope imaging) is subsequently performed for one or more of the potential defects of interest.

[0045] Figure 5 is a block diagram of a semiconductor-inspection sy stem 500 in accordance with some embodiments. The semiconductor-inspection system 500 includes a semiconductor-inspection tool 504 (i.e., optical microscope) and an associated computer system that includes one or more processors 502 (e.g., CPUs), user interfaces 510, memory 514, and one or more communication buses 503 interconnecting these components. The semiconductor- inspection system 500 may also include one or more network interfaces (wired and/or wireless, not shown) for communicating with (e.g., retrieving recipes from and/or transmitting data to) remote computer systems. [0046] The inspection tool 504 includes an illumination source 505, illumination and collection optics 506, a wafer chuck 507, and an image sensor 508. Semiconductor wafers are loaded onto the wafer chuck 507 for inspection. An optical beam scattered off of the wafer surface is imaged by the image sensor 508.

[0047] The user interfeces 510 may include a display 511 and one or more input devices

512 (e.g., a keyboard, mouse, touch-sensitive surface of the display 511, etc.). The display 511 may display results of defect classification.

[0048] Memory 514 includes volatile and/or non-volatile memory. Memory 514 (e.g., die non-volatile memory within memory 514) includes a non-transitory computer-readable storage medium Memory 514 optionally includes one or more storage devices remotely located from the processors 502 and/or a non-transitory computer-readable storage medium dim is removably inserted into die computer system. In some embodiments, memory 514 (e.g., the non-transitory computer-readable storage medium of memory 514) stores the following modules and data, or a subset or superset thereof: an operating system 516 that includes procedures for handling various basic system services and for performing hardware-dependent tasks, an inspection module 518 (e.g., for causing step 102 to be performed), a difference-image derivation module 520 (e.g., for performing step 104), a defect identification module 522 (e.g^ for performing step 106 and/or 224), a function-fitting module 524 (e.g., for performing steps 108, 208, and/or 308), and a defect-classification module 526 (e.g., for performing steps 118 and/or 318).

[0049] The memory 514 (e.g., the non-transitory computer-readable storage medium of the memory 514) thus includes instructions for performing all or a portion of the methods 100, 200. and/or 300 (Figures 1-3). Each of the modules stored in the memory- 514 corresponds to a set of instructions for performing one or more functions described herein. Separate modules need not be implemented as separate software programs. The modules and various subsets of the modules may be combined or otherwise re-arranged. In some embodiments, the memory 514 stores a subset or superset of the modules and/or data structures identified above.

[0050] Figure 5 is intended more as a functional description of the various features that may be present in a semiconductor-inspection system than as a structural schematic. For example, the arrangement of the components of the inspection tool 504 may vary (e.g. , in manners known in the art). Items shown separately could be combined and some items could be separated. Furthermore, the functionality of die semiconductor-inspection system 500 may be split between multiple devices. For example, a portion of the modules stored in the memory 514 may alternatively be stored in one or more computer systems communicatively coupled with the semiconductor-inspection system 500 through one or more networks.

[0051] The foregoing description, for purpose of explanation, has been described with reference to specific embodiments. However, the illustrative discussions above are not intended to be exhaustive or to limit the scope of the claims to the precise forms disclosed. Many modifications and variations are possible in view of the above teachings. The embodiments were chosen in order to best explain the principles underlying the claims and their practical applications, to thereby enable others skilled in the art to best use the embodiments with various modifications as are suited to the particular uses contemplated.

Claims

WHAT IS CLAIMED IS:

1. A non-transitory computer-readable storage medium storing one or more programs for execution by one or more processors of a semiconductor-inspection system that includes an optical microscope, the One or mote programs comprising instructions for:

deriving a difference image between a test image of a semiconductor die and a reference image, wherein die test image is generated by inspecting the semiconductor die using the optical microscope;

for each defect of a plurality of defects for the semiconductor die, fitting a point-spread function to the defect as indicated in the difference image and determining one or more dimensions of the fitted point-spread function; and

based at least in part on the one or more dimensions of the fitted point-spread function for respective defects of the plurality of defects, distinguishing potential defects of interest in the plurality of defects from nuisance defects.

2. The computer-readable storage medium of claim 1, wherein:

the point-spread function is a two-dimensional Gaussian function; and

the one or more dimensions comprise a first dimension indicative of a width of the fitted Gaussian function in a first direction and a second dimension indicative of a width of the fitted Gaussian function in a second direction.

3. The computer-readable storage medium of claim 2, wherein:

the first dimension is selected from the group consisting of a standard deviation of the fitted Gaussian function in the first direction and a foil width at half maximum of the fitted Gaussian function in the first direction; and

the second dimension is selected from the group consisting of a standard deviation of the fitted Gaussian function in the second direction and a foil width at half maximum of die fitted Gaussian function in the second direction.

4. The computer-readable storage medium of claim 1 , wherein die one or more dimensions comprise:

a first distance between maximal gradients of the fitted point-spread function in a first direction; and a second distance between maximal gradients of the fitted point-spread function in a second direction.

5. The computer-readable storage medium of claim 1 , wherein the one or mote dimensions comprise:

a first area under a cross-section of the fitted point-spread function in a first direction through a maximum of the fitted point-spread function, normalized by the height of the maximum; and

a second area under a cross-section of the fitted point-spread function in a second direction through the maximum, normalized by die height of the maximum.

6. The computer-readable storage medium of claim 1, wherein the point-spread function is a sine function.

7. The computer-readable storage medium of claim l, wherein the point-spread function is a polynomial function.

8. The computer-readable storage medium of claim 1 , wherein the instructions for fitting the point-spread function to the defect comprise instructions for fitting a numeric simulation of the point-spread function to the defect.

9. The computer-readable storage medium of claim 1 , wherein the one or more programs further comprise instructions for identifying die plurality of defects in the difference image before fitting the point-spread function to the defect for each of the plurality of defects.

10. The computer-readable storage medium of claim 1, wherein:

the one or more programs do not include instructions for identifying the plurality of defects before fitting the point-spread function to the defect for each of the plurality of defects; and

die instructions for fitting the point-spread function to the defect for each of the plurality of defects comprise, for each respective pixel of a plurality of pixels in the difference image, instructions for fitting the point-spread function to a respective location in the difference image, wherein the respective location corresponds to the respective pixel and comprises the respective pixel and other pixels of the plurality of pixels that surround the respective pixel.

11. The computer-readable storage medium of claim 10, wherein the instructions fin:

distinguishing foe potential defects of interest from nuisance defects comprise instructions fbr: removing, from a set of candidate pixels in the difference image, pixels that do not satisfy a criterion that is based on at least one of the one or more dimensions of foe fitted point-spread function; and

after performing the removing, identifying the potential defects of interest from die set of candidate pixels.

12. The computer-readable storage medium of claim 10, wherein the difference image indicates degrees of difference between foe test image and foe reference image on a pixel-bypixel basis, and the instructions for distinguishing foe potential defects of interest from nuisance defects comprise instructions for:

adjusting the degrees of difference for respective pixels of the difference image using a criterion that is based on at least one of foe one or more dimensions of the fitted point-spread function for foe respective pixels; and

after performing the adjusting, identifying the potential defects of interest from foe set of candidate pixels.

13. The computer-readable storage medium of claim 1 , wherein the instructions for distinguishing the potential defects of interest from nuisance defects comprise instructions for: determining, for a first defect of the plurality of defects, that a first dimension of the one or more dimensions of the fitted point-spread function exceeds, or equals or exceeds, a first threshold; and

in response to determining that the first dimension exceeds, or equals or exceeds, foe first threshold, classifying the first defect as a nuisance defect.

14. The computer-readable storage medium of claim 13, wherein the instructions for distinguishing the potential defects of interest from nuisance defects further comprise instructions for:

determining, for a second defect of foe plurality of defects, that foe first dimension is less than, or less titan or equal to, a second threshold, wherein the second threshold is less than foe first threshold; and in response to determining dial the first dimension is less than, or less than or equal to, the second threshold, classifying die second defect as a nuisance defect.

15. The computer-readable storage medium of claim 1 , wherein the one or more dimensions of the fitted point-spread function comprise two dimensions, and the instructions for

distinguishing the potential defects of interest from nuisance defects comprise instructions for: determining, for a first defect of the plurality of defects, whether a metric satisfies a threshold, wherein die metric is a function of the two dimensions; and

in response to a determination that die metric does not satisfy die threshold, classifying the first defect as a nuisance defect

16. The computer-readable storage medium of claim 1, wherein the instructions for distinguishing the potential defects of interest from nuisance defects comprise instructions for providing the one or more dimensions of the fitted point-spread function for the respective defects of the plurality of defects as input to a machine-learning algorithm trained to distinguish potential defects of interest from nuisance defects.

17. The computer-readable storage medium of claim l, wherein the one or more programs further comprise instructions for determining a goodness of fit of the fitted point-spread function for die respective defects of the plurality of defects,

wherein distinguishing the potential defects of interest from nuisance defects is further based at least in part on the goodness of fit for die respective defects.

18. The computer-readable storage medium of claim 1 , wherein the respective defects of die plurality of defects are not resolved in die difference image.

19. The computer-readable storage medium of claim 1, wherein the one or more programs further comprise instructions for generating a report specifying the potential defects of interest.

20. A semiconductor-inspection system, comprising:

an optical microscope;

one or more processors; and

memory storing one or more programs for execution by the one or more processors, the one or more programs including instructions for: deriving a difference image between a test image of a semiconductor die and a reference image, wherein fee test image is generated by inspecting fee semiconductor die using the optical microscope;

for each defect of a plurality of defects for the semiconductor die, fitting a point- spread function to the defect as indicated in the difference image and determining one or more dimensions of the fitted point-spread function; and

21. A method of identifying semiconductor defects of interest, comprising:

inspecting a semiconductor die using an optical microscope to generate a test image of the semiconductor die; and

in a computer system comprising one or more processors arid memory storing

instructions for execution by the one or more processors:

deriving a difference image between the test image of the semiconductor die and a reference image;

for each defect Of a plurality Of defects for the semiconductor die, fitting a point- spread function to the defect as indicated in fee difference image and determining one or more dimensions of the fitted point-spread function; and

based at least in part on fee one or more dimensions of the fitted point-spread function for respective defects of the plurality of defects, distinguishing potential defects of interest in the plurality of defects from nuisance defects.

22. A non-transitory computer-readable storage medium storing one or more programs for execution by one or mote processors of a semiconductor-inspection system that includes an optical microscope, the one or more programs comprising instructions for:

deriving a difference image between a test image of the semiconductor die and a reference image, wherein the test image is generated by inspecting the semiconductor die using the optical microscope;

fitting a summation of a plurality of point-spread functions to the difference image, wherein each point-spread function of the plurality of point-spread functions is centered on a distinct predefined location in the semiconductor die and has a fixed width associated with the optical microscope, the fitting comprising determining coefficients of respective point-spread functions of the plurality of point-spread functions; and

based at least in part on the coefficients of the respective point-spread functions, distinguishing potential defects of interest in a plurality of defects for the semiconductor die from nuisance defects.