Method for defect inspection measurement of a substrate, apparatus for imaging a substrate and method for operating the same

Abstract

A method for defect classification is described. The method comprises the following steps: storing a plurality of defect categories in a multi-dimensional feature space according to a plurality of classification rules, wherein the plurality of classification rules define a boundary of a region associated with the defect category in the multi-dimensional feature space for each defect category in the plurality of defect categories; receiving one or more electron beam image data associated with a plurality of defects detected in the one or more display devices on the inspected large area substrate; applying, by a processor, an automatic classifier to the electron beam image data, the automatic classifier based on the plurality of classification rules; and identifying the plurality of defects, each defect of the plurality of defects being classified with at least a first confidence level based on at least one confidence threshold.

Description

Technical Field

[0001] The present disclosure relates to methods for defect review (DR) measurements on substrates, particularly for display manufacturing, i.e., methods for defect review measurements on large area substrates. Further, embodiments relate to a method for defect review classification for display manufacturing. Embodiments of the present disclosure relate generally to automated inspection, and particularly to methods and systems for analyzing manufacturing defects. Background Art

[0002] In many applications, it is beneficial to inspect substrates to monitor the quality of the substrates. For example, glass substrates with layers of coating materials deposited thereon are manufactured for the display market. Since defects may occur during processing of the substrates, such as during coating of the substrates, it is beneficial to inspect the substrates to review for defects and monitor the quality of the displays.

[0003] Displays are typically manufactured on large area substrates, and the size of substrates is constantly increasing. Furthermore, displays, such as TFT displays, are also constantly being improved. Inspection of substrates can be achieved by optical systems. However, defect review (DR) measurements, such as defect review of TFT arrays, are advantageously performed with a higher resolution, which can no longer be provided by optical inspection. For example, DR measurements can provide information related to previously detected defects. Therefore, DR measurements are valuable for process control because countermeasures can be taken to prevent or reduce the probability of defects.

[0004] Automatic defect classification (ADC) can be performed based on optical defect images. However, smaller design rules in modern display manufacturing make yield more sensitive to defects with a size of 10 microns or less. Optical images do not provide enough information due to the reduced resolution.

[0005] Defect detection or re-detection in a "defect review system" can be provided by comparing a reference image with a defect image (ie, the image to be inspected). A defect is considered as a deviation between the reference image and the defect image exceeding a given threshold.

[0006] Automatic Defect Classification (ADC) technology is used to inspect and measure defects on patterned wafers in the semiconductor industry. ADC technology detects the presence of defects and automatically classifies them by type to provide more detailed feedback on the production process and reduce the load on human inspectors. For example, ADC technology can be used to distinguish between defect types caused by particle contaminants on the wafer surface and defect types associated with irregularities in microcircuit patterns, and can also identify specific types of particles and irregularities.

[0007] Substrates used for display manufacturing are typically glass substrates, with an area of, for example, 1 square meter or more. Taking high-resolution images on such large substrates is inherently very challenging, and most of the research results from the wafer industry are not applicable. Furthermore, the DR measurement options available on semiconductor wafers may not be suitable for large-area substrates. In general, manufacturing tolerances that are of the same or similar order of magnitude as the defect size to be detected may result in false defect detection or low threshold settings. Therefore, the research results from the wafer industry are not applicable.

[0008] Therefore, in view of, for example, increasing requirements on defect review quality, there is a need to improve DR and defect classification, for example, without having to break down a display substrate into smaller samples and allowing the manufacturing process of the display substrate to continue after DR measurement. Summary of the invention

[0009] In summary, a method for defect classification, a method for generating multiple classification rules, and an automatic defect classification system are provided. Other aspects, advantages, and features of the present disclosure are apparent from the description and the accompanying drawings.

[0010] According to one embodiment, a method for defect classification is provided. The method includes: storing a plurality of defect categories in a multidimensional feature space according to a plurality of classification rules, wherein the plurality of classification rules define, for each defect category in the plurality of defect categories, boundaries of a region associated with the defect category in the multidimensional feature space; receiving one or more electron beam image data associated with a plurality of defects detected in one or more display devices on an inspected large area substrate; applying, by a processor, an automatic classifier to the electron beam image data, the automatic classifier being based on the plurality of classification rules; and identifying the plurality of defects, each of the plurality of defects being classified with at least a first confidence level based on at least one confidence threshold.

[0011] According to one embodiment, a method for generating a plurality of classification rules is provided. The method includes: receiving a plurality of electron beam image data associated with a plurality of defects detected in one or more display devices on a large area substrate; receiving defect categories, each defect category being associated with one or more of the plurality of electron beam image data; and generating the plurality of classification rules in a multi-dimensional feature space, wherein the plurality of classification rules define, for each defect category in the plurality of defect categories, a boundary of a region associated with the defect category in the multi-dimensional feature space.

[0012] According to one embodiment, an automatic defect classification system is provided, wherein the system comprises: a memory comprising instructions, and a processor, wherein the instructions, when executed by the processor, cause the automatic defect classification system to perform the method as described in any embodiment of the present disclosure. BRIEF DESCRIPTION OF THE DRAWINGS

[0013] In order to be able to understand the above-mentioned features of the present disclosure in detail, a more detailed description of the present disclosure briefly summarized above can be obtained by reference to the embodiments, some of which are illustrated in the accompanying drawings. However, it should be noted that the drawings only illustrate exemplary embodiments and therefore should not be considered as limiting the scope, and other equally effective embodiments are acceptable.

[0014] Figure 1 A schematic diagram of a display manufacturing process with inline automatic defect classification according to an embodiment of the present disclosure is shown.

[0015] Figure 2 Shown is a side view of an apparatus for imaging portions of a substrate according to embodiments described herein.

[0016] Figure 3 Shown is a side view of another apparatus for imaging portions of a substrate according to embodiments described herein.

[0017] Figure 4 A flow chart is shown illustrating a method for defect classification on a large area substrate (eg, for display manufacturing) according to an embodiment of the present disclosure.

[0018] Figure 5 is a schematic representation of a feature space for defect classification according to an embodiment of the present disclosure.

[0019] Figure 6 is a block diagram of an automatic defect classification system according to an embodiment of the present disclosure.

[0020] Figure 7 is a schematic diagram illustrating the relationship between purity and rejection level according to an embodiment of the present disclosure.

[0021] Figure 8 is a block diagram of an exemplary computer system that can perform one or more of the operations described herein.

[0022] Figures 9A to 9D Images are shown to illustrate exemplary defect review measurements used in some embodiments in accordance with the present disclosure.

[0023] Fig.10 is an exemplary category image.

[0024] To facilitate understanding, identical reference numerals have been used, where possible, to designate identical elements that are common to the figures. It is contemplated that elements and features of one embodiment may be beneficially incorporated in other embodiments without further recitation. DETAILED DESCRIPTION

[0025] Reference will now be made in detail to various exemplary embodiments, one or more of which are illustrated in each of the figures. Each example is provided by way of explanation and is not meant to be limiting. For example, features illustrated or described as part of one embodiment may be used in other implementations or used with other embodiments to produce further embodiments. The present disclosure is intended to include such modifications and variations.

[0026] In the following description of the figures, the same reference numerals refer to the same components. Only the differences related to the various embodiments are described. The structures shown in the figures are not necessarily depicted according to the true scale, but are used to better understand the embodiments.

[0027] Electron beam review (EBR), particularly for large area substrates, is a relatively new technology in which measurements are taken of the entire substrate or of areas distributed over the entire substrate, such that, for example, the display to be manufactured is not destroyed during or for the review process. Achieving resolutions of, for example, 20 nanometers or less, such as 10 nanometers or less, is very challenging, and previous findings from wafer imaging may not apply in view of the large differences in substrate sizes. For example, positioning a platform (i.e., a substrate table) in an arbitrary area of ​​the entire substrate under the electron beam may be advantageously suitable, and the positioning must be very accurate over a large area. For large area substrates, the area to be measured is larger, and the various areas may be further apart from each other, such as compared to a wafer imaging device. Therefore, simple upscaling cannot be successful due to, among other reasons, different throughput requirements. Still further, manual or semi-automatic processes may not be suitable in view of the required throughput and repeatability of the measurement positions distributed over an area of ​​a large area substrate.

[0028] Further, manufacturing tolerances are greater for display manufacturing on large area substrates than for semiconductor manufacturing on wafers. Thus, the acceptable deviation of an image at a first location relative to an image at a second location having the same pattern is greater than for semiconductor manufacturing. Thus, the size of the defect may be within the same order of magnitude as the acceptable deviation, or the size of the defect may be only one or two orders of magnitude greater than the acceptable deviation. Generally speaking, for display manufacturing and in the semiconductor industry, defect review may be based on image comparison and image deviation thresholds. Such comparisons may be limited as long as the defect size is close to the size of the acceptable deviation (e.g., deviation based on manufacturing tolerances). Therefore, embodiments of the present disclosure relate particularly to defect review and defect classification for display manufacturing, such as defect review on large area substrates.

[0029] Embodiments of the present disclosure provide automatic defect classification (ADC) based on scanning electron microscope images, which was not possible in the past for the display industry. For the display industry, ADC is provided based on optical defect images. However, the smaller design rules in modern display manufacturing make the yield more sensitive to defects with a size of 10 microns or less. Due to the reduced resolution, the optical image does not provide enough information. Embodiments of the present disclosure allow the use of scanning electron microscope (SEM) images from EBR and enable ADC to be used in batch production lines in the display industry, particularly for in-line SEM electron beam inspection of large area substrates and display devices provided on large area substrates.

[0030] Figure 1 A process flow for exemplary display manufacturing is shown, and includes automatic defect classification according to an embodiment of the present disclosure. A yield management system may typically include one or more inspection operations after a processing operation. For example, thin film deposition may be provided on a large area substrate, particularly a large area substrate with one or more structures provided thereon. Further substrate processing steps may include deposition of photoresist, etching, structuring and / or deposition of a material layer (e.g., polysilicon). After one or more substrate processing operations, automated optical inspection (AOI) and / or EBT testing may be provided. Automated optical inspection uses light to detect portions of a large area substrate. For example, an entire large area substrate may be inspected. Inspection may result in irregular structures or irregular material properties at one or more substrate locations. Additionally or alternatively, an electron beam test (EBT) tool may test pixel defects, line defects, driver defects, or other defects of a display on a large area substrate. Given that an electron beam test tool has a larger field of view (FOV) than an electron beam inspection tool (i.e., SEM), the entire large area substrate may be tested. Testing may result in possible defects at one or more substrate locations.

[0031] A method for defect review images can receive a list of defects or defect candidates. As described above, a large area substrate can be tested with an AOI tool to obtain a list of defects or defect candidates. Additionally or alternatively, pixels of a display can be tested with a display test method. Pixel defects, line defects, driver defects, or other defects can be tested with an electron beam test system and an optical test system or other measurements (such as electrical measurements). Therefore, the location of the defective pixel can be provided for defect review measurement and / or can be provided to a device for defect review measurement.

[0032] An area with a defective pixel is imaged to provide a defect image. Another area (e.g., a corresponding area of ​​an adjacent pixel) is measured to provide a reference image. Using DR measurements, defect reviews of defects from previous metrology tools can be evaluated. Due to the size of substrates used for display manufacturing and the challenges this poses to the manufacturing process, the locations for defect review measurements on large area substrates described in the embodiments of the present disclosure may be distributed over the large area substrate. For example, a display may have 5 million pixels or more, such as approximately 8 million pixels. Large displays may have even higher pixel counts. For each pixel, at least one electrode for red, one electrode for green, and one electrode for blue (RGB) are provided. Therefore, defects that are considered critical to the manufacturing process may appear over a very large area. As shown in FIG. 9A to 9D As described, embodiments may include providing a DR measurement based on a first operation with a mask pattern and a subsequent second operation without the mask pattern. The DR measurement is provided in the structure of a defect image and is provided with a reference image.

[0033] According to some embodiments that can be combined with other embodiments described herein, the defect image is generated at a defective pixel on the substrate, and the reference image is generated at a pixel adjacent to the defective pixel or at a pixel adjacent to the defective pixel. Defect review measurements can be repeated on one or more areas of the substrate, which are distributed over an area of ​​at least 1 square meter.

[0034] According to some embodiments, which can be combined with other embodiments described herein, the substrate described herein relates to a large area substrate, in particular a large area substrate for the display market. According to some embodiments, the size of the large area substrate or the corresponding substrate support may be at least 1 square meter, for example at least 1.375 square meters. The size may be from about 1.375 square meters (1100 mm x 1250 mm - Gen 5) to about 9 square meters, more specifically from about 2 square meters to about 9 square meters, or even up to 12 square meters. The substrate or substrate receiving area (for which the structures, devices and methods according to the embodiments described herein are provided) may be a large area substrate as described herein. For example, the large area substrate or carrier may be GEN 5 (corresponding to approximately 1.375 square meters of substrate (1.1 meters x 1.25 meters)), GEN 7.5 (corresponding to approximately 4.39 square meters of substrate (1.95 meters x 2.25 meters)), GEN 8.5 (corresponding to approximately 5.7 square meters of substrate (2.2 meters x 2.5 meters)), or even GEN 10 (corresponding to approximately 9 square meters of substrate (2.88 meters x 3130 meters)). Larger generations, such as GEN 11 and GEN 12, and the corresponding substrate areas may be similarly implemented. It must be considered that even though the size of the GEN 5 substrate may vary slightly between different display manufacturers, these substrate size generations still provide a fixed industry standard. For example, an embodiment of an apparatus for testing may have a GEN 5 substrate support or a GEN 5 substrate receiving area such that GEN 5 substrates of many display manufacturers may be supported by the support. The same applies to other substrate size generations.

[0035] According to embodiments of the present disclosure, electron beam inspection can be provided on large area substrates, particularly without cutting the large area substrate into smaller pieces. Electron beam inspection can provide images of the one or more substrate locations, particularly SEM images, such as provided by an automated optical inspection tool or an electron beam test tool. SEM images can be used for defect inspection, particularly defect classification.

[0036] According to some embodiments, which can be combined with other embodiments described herein, defect classification can be used for root cause analysis, wherein substrate processing can be adapted based on the defect categories detected during defect classification. Defect classification can include different defect categories, different defect sizes, and / or interconnections of defects with other patterns or structures on the display. Properties associated with defects can be used to distinguish defects that are called killer defects or not called killer defects. Killer defects can be understood as defects that cause the display to be non-functional or unusable.

[0037] According to some embodiments that can be combined with other embodiments described herein, it can be determined whether a defect is a killer defect, for example, after the display manufacturing process is fully completed and the display can be turned on, after some processing steps, and / or after the defect is manufactured to a state where the pixel can be driven. For example, an EBT test can be provided on a substrate having a display in a state where the display can be driven. According to some embodiments that can be combined with other embodiments described herein, directly visible defective pixels can be associated with recorded defect review data from an EBR and corresponding defect classifications. According to some embodiments, it can be determined that a defect is a killer defect by correlating EBR data with EBT data. EBT detects electrical defects that are killer defects in most cases.

[0038] According to some embodiments, which may be combined with other embodiments described herein, an ADC based on electron beam image data on a display device may be used for process control. For example, the results of the ADC may be fed back to the process and the process recipe may be adapted. Statistical evaluation of defect type distribution and general process and equipment optimization by SPC (statistical process control) may be provided.

[0039] Figure 1 An EBR measurement operation is shown after different manufacturing operations of a display device on a large area substrate. The EBR information can be fed back to the processing operation for process control. Further, an EBT measurement is shown, where, for example, a substrate can be transferred to a repair station after one or more killer defects are found.

[0040] According to some embodiments, which can be combined with other embodiments described herein, the defect type can be selected from the group consisting of scratches, particles, bridges, and opens. Defect sizes can be provided in size categories such as small, medium, and large, where the size categories can correspond to a different absolute size for each defect type. As another characteristic of the defect, connection to another pattern may or may not occur. Other defect properties can be provided.

[0041] In particular, for killer defects, repair tools can be provided in display factory equipment to correct and / or repair defects between processing operations of large area substrates. For example, shorts or bridges can be cut with an ion beam or laser. After defect classification, the large area substrate can be provided to another processing tool. According to an embodiment of the present disclosure, defect classification based on electron beam review images on large area substrates can be provided within a production line of a display manufacturing process. In particular, defect classification can be provided after processing a large area substrate in a first production chamber and before further processing the large area substrate in a second production chamber. Therefore, the overall yield can be improved. The yield can be improved by correcting or repairing defects at the end of a production operation or between production operations. The yield can be improved by associating the classified defects with the root cause of the defect and adapting the manufacturing conditions of the large area substrate to be processed subsequently.

[0042] According to embodiments of the present disclosure, an electron beam inspection tool (eg, a SEM tool configured to generate images on a large area substrate) for defect classification may be provided for large area substrates. Figure 2 An electron beam inspection tool is shown. The electron beam inspection tool includes a substrate support 110 extending along an x-direction 150. Figure 2 In the drawing plane of , the x-direction 150 is the left-right direction. The substrate 160 is disposed on the substrate support 110. The substrate support 110 is movable along the x-direction 150 so as to displace the substrate 160 in the vacuum chamber 120 relative to the first imaging charged particle beam microscope 130 and the second imaging charged particle beam microscope 140. Each of the first imaging charged particle beam microscope 130 and the second imaging charged particle beam microscope can produce an SEM image of a small portion of a large area substrate. According to some embodiments that can be combined with other embodiments described herein, the field of view size of the imaging charged particle beam microscope used for the method and apparatus according to the present disclosure can be 500 microns or less and / or 5 microns or more. The resolution of the image can be about 100 nanometers or less, such as 20 nanometers or less, such as 10 nanometers or less.

[0043] An area of ​​the substrate 160 can be positioned below the first imaging charged particle beam microscope 130 or below the second imaging charged particle beam microscope 140 for DR measurement. The area may include a structure for DR measurement, which is contained in or on a coating on the substrate. The substrate support 110 can also move in the y direction (not shown) so that the substrate 160 can move in the y direction. By appropriately moving the substrate support 110 holding the substrate 160 in the vacuum chamber 120, portions of the entire range of the substrate 160 can be measured in the vacuum chamber 120.

[0044] The first imaging charged particle beam microscope 130 may be spaced apart from the second imaging charged particle beam microscope 140 by a distance 135 along the x-direction 150. Figure 2 In the illustrated embodiment, the distance 135 is the distance between the center of the first imaging charged particle beam microscope 130 and the center of the second imaging charged particle beam microscope 140. In particular, the distance 135 is the distance along the x-direction 150 between the first optical axis 131 defined by the first imaging charged particle beam microscope and the second optical axis 141 defined by the second imaging charged particle beam microscope 140. The first optical axis 131 and the second optical axis 141 extend along the z-direction 151. For example, the first optical axis 131 can be defined by the objective of the first imaging charged particle beam microscope 130. Likewise, for example, the second optical axis 141 can be defined by the objective of the second imaging charged particle beam microscope 140.

[0045] like Figure 2 As shown, the vacuum chamber 120 has an inner width 121 along the x-direction 150. The inner width 121 may be a distance obtained when passing through the vacuum chamber 120 along the x-direction from a left side wall 123 of the vacuum chamber 120 to a right side wall 122 of the vacuum chamber 120. One aspect of the present disclosure relates to a dimension of the apparatus 100 relative to, for example, the x-direction 150. According to an embodiment, a distance 135 between the first imaging charged particle beam microscope 130 and the second imaging charged particle beam microscope 140 along the x-direction 150 may be at least 30 centimeters, for example at least 40 centimeters. According to further embodiments, which may be combined with other embodiments described herein, the inner width 121 of the vacuum chamber 120 may be in a range of 250% to 450% of the distance 135 between the first imaging charged particle beam microscope 130 and the second imaging charged particle beam microscope 140.

[0046] Embodiments described herein may provide a device for imaging portions of a substrate. The substrate is processed as a whole in a vacuum chamber. In particular, embodiments described herein do not require destruction of the substrate or etching of the substrate surface. Thus, high resolution images for defect review measurements may be provided. One advantage of having a vacuum chamber with reduced size provided by some embodiments described herein is that one or more vibrations of the vacuum chamber may be reduced, as vibration levels increase as a function of vacuum chamber size. Thus, the amplitude of vibrations of the substrate may also be advantageously reduced.

[0047] The distance along the first direction between the exemplary first imaging charged particle beam microscope and the second imaging charged particle beam microscope is in the range of 30% to 70% of the first receiving area size of the substrate receiving area. More specifically, the distance along the first direction can be in the range of 40% to 60% of the first receiving area size, for example, about 50% of the first receiving area size. For example, referring to Figure 2In the illustrated embodiment, the distance along the first direction may refer to the distance 135 between the first imaging charged particle beam microscope 130 and the second imaging charged particle beam microscope 140. Figure 2 In the exemplary embodiment shown, the distance 135 is approximately 50% of the width of the substrate receiving area.

[0048] The substrate support can be movable in the vacuum chamber relative to the first imaging charged particle beam microscope and / or relative to the second imaging charged particle beam microscope. According to an embodiment that can be combined with other embodiments described herein, the distance between the second imaging charged particle beam microscope and the first imaging charged particle beam microscope is at least 30 centimeters, more particularly at least 40 centimeters, for example, about 50% of the size of the first receiving area. Minimizing the distance between the first imaging charged particle beam microscope and the second imaging charged particle beam microscope, that is, greater than the distance of redundancy achieved by simply copying two adjacent imaging charged particle beam microscopes (e.g., two adjacent SEMs), has the advantage of reducing the distance traveled by the substrate inspected by the device. This allows the size of the vacuum chamber to be reduced, so that the vibration of the vacuum chamber can also be advantageously reduced.

[0049] According to embodiments, an apparatus for imaging a portion of a substrate for defect review and defect classification may be provided. The apparatus includes a vacuum chamber and a substrate support disposed in the vacuum chamber. According to some embodiments, which may be combined with other embodiments described herein, the substrate support may optionally provide a substrate receiving area of ​​at least 1 square meter.

[0050] Figure 3 1 shows a side view of another apparatus for imaging portions of a substrate according to embodiments described herein. Apparatus 100 includes a vacuum chamber 120. Apparatus 100 further includes a substrate support 110 on which a substrate 160 may be supported. Apparatus 100 includes a first imaging charged particle beam microscope 130. Figure 2 on the contrary, Figure 3 A single imaging charged particle beam microscope is shown provided above the substrate support 110. Although this may result in reduced imaging capabilities, such as reduced resolution, the resulting resolution may be sufficient for some DR measurements. Further, for semiconductor wafer applications where the defect size of the defects to be detected is small compared to acceptable manufacturing tolerances, an apparatus for imaging portions of a substrate may be provided, the apparatus having a single imaging charged particle beam microscope. Figure 2 similar, Figure 3The device shown may include a controller and a deflection assembly. The controller may be connected to a substrate support, in particular a displacement unit of the substrate support. Further, the controller may be connected to a deflection assembly of an imaging charged particle beam microscope.

[0051] Defect review measurements are typically provided on various areas of a substrate, such as a wafer in semiconductor manufacturing or a large area glass substrate for display manufacturing for example. Thus, a statistical analysis of structural defect review can be performed over the entire substrate area and over multiple processed substrates. For small substrates such as wafers, this can be done with sufficient throughput using methods known from the semiconductor industry. In the semiconductor industry, matching of measurement capabilities is provided from tool to tool. For electron beam review (EBR) of display substrates, two imaging charged particle beam microscopes (see Figure 2 ) can be matched to each other. This is related to relative position and measurement capability. Single column device (see Figure 3 ) can avoid matching multiple columns in a system while accepting lower resolution. Multi-column devices can advantageously include column matching and have higher resolution.

[0052] According to some embodiments that can be combined with other embodiments described herein, a method of operating an apparatus for imaging of the present disclosure may include: matching a first coordinate system on a large area substrate of a first imaging charged particle beam microscope with a second coordinate system on a large area substrate of a second imaging charged particle beam microscope.

[0053] Both options, the single-column method and the multi-column method, enable the improved DR measurement process described herein, wherein sufficient detection sensitivity and sufficient throughput are provided, especially on large-area substrates. According to embodiments of the present disclosure, the DR measurement described herein can be provided in various regions of a large-area substrate. For example, two or more regions, such as 5 regions to 100 regions, can be distributed on the substrate.

[0054] The imaging charged particle beam microscope used in this article can be adapted to produce a low-energy charged particle beam with a landing energy of 2keV or less (particularly 1keV or less). Compared with a high-energy beam, a low-energy beam will not affect or deteriorate the display backplane structure during defect review measurement. According to a further embodiment that can be combined with other embodiments described herein, between the particle beam source and the substrate, the charged particle energy (e.g., electron energy) can be increased to 5keV or more, such as 10keV or more. Accelerating charged particles in a column reduces the interaction between charged particles, reduces the aberration of optoelectronic components, and thus improves the resolution of the imaging scanning charged particle beam microscope.

[0055] According to another embodiment that can be combined with other embodiments described herein, the term "substrate" used herein includes both non-flexible substrates (such as glass substrates or glass plates) and flexible substrates (such as coils or foils). The substrate can be a coated substrate, in which one or more thin layers of material are coated or deposited on the substrate, for example by a physical vapor deposition process (PVD) or a chemical vapor deposition process (CVD). Substrates used for display manufacturing typically include insulating materials, such as glass. Therefore, in contrast to semiconductor wafer SEMs, devices for imaging portions of large area substrates do not allow biasing of the substrate. According to embodiments described herein that can be combined with other embodiments described herein, the substrate is grounded. The substrate cannot be biased to a potential that affects the landing energy or other optoelectronic aspects of the scanning electron beam microscope. This is another example of the difference between EBR systems for large area substrates and semiconductor wafer SEM inspections. This may further lead to electrostatic discharge (ESD) problems when handling the substrate on a substrate support. Therefore, it can be seen that the wafer inspection scheme may not be easily applied to DR measurements of substrates used for display manufacturing.

[0056] According to further embodiments that can be combined with other embodiments described herein, defect review measurements on large-area displays for display manufacturing can be further distinguished from semiconductor wafer DR based on scanning technology. In general, analog scanning technology and digital scanning technology can be distinguished. Analog scanning technology can include providing an analog sawtooth signal with a predetermined frequency to a scanning deflection assembly. The sawtooth signal can be combined with a continuous or quasi-continuous substrate movement to reach the scanning area of ​​the substrate. Digital scanning technology provides discrete values ​​for the x-positioning and y-positioning of the charged particle beam on the substrate, and each pixel of the scanned image is addressed pixel by pixel (i.e., digitally) by the coordinate value. Analog scanning technology ("flying platform") may be considered to be preferred for semiconductor wafer SEM inspection due to scanning speed and lower complexity, but it is not conducive to DR measurement on large-area substrates. Due to the size of the substrate, the area to be scanned is scanned digitally, i.e., by providing a desired beam position and a target list. That is, the image is scanned using digital scanning technology (i.e., a digital scanner). Due to the size of the substrate, this scanning process provides better throughput and accuracy.

[0057] According to one embodiment of the present disclosure, a method for defect classification is provided. Figure 4As shown in operation 410 in , a plurality of defect categories are stored, for example, in a memory, or provided according to a plurality of classification rules in a multidimensional feature space. The plurality of classification rules define, for each defect category in the plurality of defect categories, boundaries of an area associated with the defect category in the multidimensional feature space. One or more electron beam image data are received and associated with a plurality of defects detected in one or more display devices on a large area substrate under inspection (e.g., see operation 420). In operation 430, a processor applies an automatic classifier to the electron beam image data. The automatic classifier is based on the plurality of classification rules. In operation 440, a plurality of defects are identified, each defect being classified with at least a first confidence level based on at least one confidence threshold. According to some embodiments that may be combined with other embodiments described herein, the electron beam image data may be converted to a second data format by an interface algorithm in order to apply the automatic classifier (see operation 422).

[0058] For example, the first data format received during operation 420 may be converted to a second data format in operation 422. The first data format may include a reference image and a defect image. According to yet further optional modifications, the first data format may further include a category image. According to some embodiments that may be combined with other embodiments described herein, and as shown in operation 442, the method may further include: adding, by the interface algorithm, a defect classification to the first data format for each defect in the plurality of defects identified with at least the first confidence level.

[0059] Embodiments of the present disclosure provide electron beam image data, such as a result file, which may include a high-resolution SEM image of each predefined substrate position. The electron beam image data or result file may be provided by an EBR tool. For example, an interface algorithm provided by the interface software may convert data from a first data format associated with data from display manufacturing to a second data format. The second data format may be provided to an automatic defect classification (ADC) server. The ADC server provides automatic defect classification based on a machine learning algorithm (e.g., using artificial intelligence).

[0060] Depending on the embodiment, the machine learning algorithm may combine a multi-class classifier and a single-class classifier. Further, the purity of the classification may be used for automatic defect classification. The classification purity may be the percentage of remaining defects (e.g., defects found by the ADC system to be classifiable and not rejected) that are correctly classified. The system operator may specify a classification performance metric, such as a desired purity and / or a certain maximum rejection rate. The classification performance metric may be the percentage of defects that the ADC system cannot confidently classify and thus returns to a human expert (e.g., a system operator) for classification.

[0061] Figure 5is an exemplary schematic representation of a feature space 40 into which a set of defects 42, 44, 50, 51, 56 are mapped according to one embodiment of the present disclosure. Although the feature space 40 is represented as being two-dimensional, the classification process is typically implemented in a higher dimensional space. Figure 5 The defects in are assumed to belong to two different classes, one associated with defect 42 (hereinafter referred to as "Class I") and the other associated with defect 44 (hereinafter referred to as "Class II"). Defect 42 is bounded by boundary 52 in feature space 40, and defect 44 is bounded by boundary 54 in feature space 40. The boundaries may overlap.

[0062] The ADC machine can classify defects using two types of classifiers: a multi-class classifier, and at least one single-class classifier. The multi-class classifier distinguishes between Class I and Class II. For example, the multi-class classifier can be a binary classifier that defines a boundary 46 between regions associated with two classes. In some embodiments, the ADC machine performs multi-class classification by superimposing multiple binary classifiers (each classifier corresponds to a different pair of classes) and assigning each defect to the class that receives the most votes from the multiple binary classifiers. Once the defects are classified by the multi-class classifier, the single-class classifier represented by the boundaries 52 and 54 identifies the defects that can be reliably assigned to the corresponding class, while rejecting the defects outside the boundaries as "unknown" defects.

[0063] In some embodiments, a system operator of the ADC machine provides confidence thresholds to determine the loci of the boundaries of regions in feature space 40 associated with defect classes. Setting the confidence thresholds for multi-class classification can be equivalent to placing boundaries 48 on either side of boundary 46. In some embodiments, the higher the confidence threshold, the farther away boundary 48 will be. The ADC machine can reject defects 51 that are between boundaries 48 but within boundaries 52 or 54 as "undecidable" defects because the ADC machine may not be able to automatically assign these defects to one class or the other with the desired confidence level. Undecidable defects can be determined by a human operator or by a different classification module with different input values.

[0064] In some embodiments, the confidence threshold controls the shape of the boundary of the single-class classifier. The shape can refer to the geometric form of the boundary or the extent of the boundary. The shape can be associated with the parameters of the kernel function used to implement the single-class classifier. For each value of the confidence threshold, the ADC machine selects an optimal parameter value. In some embodiments, the extent of the boundary decreases as the confidence threshold increases, and the geometry of the boundary can also change as different kernel parameter values ​​are selected.

[0065] Back to Figure 5 , defect 56 falls outside boundaries 52 and 54 and can therefore be classified as an "unknown" defect, even though defect 56 may have been determined by the multi-class classifier. Defect 50, which is both outside boundaries 52 and 54 and between boundaries 48, is also considered "unknown" because defect 50 falls outside boundaries 52 and 54. In some embodiments, setting a lower confidence threshold can fully expand boundaries 52 and / or 54 to include defect 50 and / or defect 56, resulting in the ADC machine rejecting fewer defects. However, by setting a lower confidence, the ADC machine may make more classification errors, thereby reducing the purity of the classification. In some embodiments, increasing the confidence threshold can increase the purity of the classification, but may result in a higher rejection rate (more defects will be rejected by the ADC machine as unknown defects).

[0066] Figure 6 is an exemplary block diagram of a classifier. The classifier may include a multi-class classifier 62 and one or more single-class classifiers 64. The classifier may include a single-class classifier 64 for each defect class in the ADC system. In another embodiment, the single-class classifier 64 may be used for more than one defect class. The multi-class classifier 62 may process the feature value vector of each defect to select a defect class for the defect, or to eliminate the defect as an undeterminable or unknown defect. In one embodiment, the multi-class classifier 62 is a support vector machine. In an alternative embodiment, the multi-class classifier 62 is a classifier having properties similar to a support vector machine. The single-class classifier 64 may examine the characteristics of the defect according to one or more elimination rules of one or more classes represented by the single-class classifier 64.

[0067] Figure 7 is an exemplary schematic graph of the functional relationship between classification purity and rejection rate according to an embodiment of the present disclosure. The ADC machine can generate a graph based on the actual classification results of the training data. To this end, the ADC machine can compare the automatic classification results of a set of defects in the training data with the "gold standard" (validation set) visual classification performed by a human inspector. The comparison can be performed for different confidence thresholds (corresponding to different rejection rates). When all defects are automatically classified by the ADC machine and the rejection rate is zero, the purity of the classification will be very low because the machine needs to classify many problematic defects. However, selecting a high rejection rate can improve the purity of the classification, but may cause human inspectors to spend more time visually classifying defects classified as "unknown" by the ADC machine.

[0068] An operator of an ADC machine (such as a human inspector) can use Figure 5The graph shown can be used to select a reject rate that will achieve a desired purity level, or to estimate the assortment purity resulting from setting a certain reject rate.

[0069] An embodiment of the present disclosure provides an ADC system using a multi-class classifier and a single-class classifier. The multi-class classifier divides a multidimensional feature space into multiple defect categories and assigns each defect to one of the categories depending on the position of the defect in the feature space. The multi-class classifier identifies defects in the overlapping areas between the categories as undeterminable defects. The multi-class classifier can identify defects in the overlapping areas by using a confidence threshold. For each defect category, the single-class classifier applies category-specific rules to identify defects that belong to the defect category and defects that do not belong to the category. Defects that do not belong to the category can be identified using the confidence threshold of the category and can be identified as unknown defects. The single-class classifier and the multi-class classifier are used together to classify defects with high purity. The outer boundaries of different categories and the range of overlapping areas are adjusted by variable confidence thresholds to maximize purity while keeping the rejection rate within a predefined threshold.

[0070] The confidence thresholds for single-class classifiers (for distinguishing known defects from unknown defects) and multi-class classifiers (for distinguishing decidable defects from undecidable defects) can be adjusted during the training process using a set of defects that have been manually classified in advance by a human operator. The result of the training process can be a set of classification rules (also called rejection rules) that define the boundaries of each defect class in the feature space. This set of classification rules can define the corresponding ranges of inspection feature values ​​that characterize the class. The classification rules also provide a confidence metric that gives the confidence level associated with each single-class classification or multi-class classification of the defect based on the location of the defect in the feature space.

[0071] Embodiments of the present disclosure may include the following steps to apply an automatic classifier: applying a multi-class classifier to the electron beam image data to classify the plurality of defects, and applying a single-class classifier to identify the plurality of defects with the at least one confidence threshold. For example, the purity level and / or confidence threshold may be set to adapt the rejection rate. When classifying actual production defects whose classification is unknown, the confidence threshold for each classifier may be selected to achieve a desired level of performance.

[0072] In an embodiment of the present disclosure, defects rejected by the automatic classifier (e.g., classified as undeterminable or unknown) are passed to one or more other inspection modalities different from the inspection modality used to produce the first classification result in order to be classified into a defect category, thereby obtaining a second classification result. In one embodiment, the inspection modality is a human inspector who assigns the rejected defects to the appropriate defect category. In an alternative embodiment, the rejected defects are classified based on additional inspection data (such as X-ray inspection data, etc.), which provide additional information about the material located at and / or near the defect location. The updated defect assignment (second classification result) of the rejected defects can be transmitted back to the ADC system. In one embodiment, the ADC system integrates the updated defect assignment (second classification result) with the automatically classified defects (first classification result) into a combined data set. Therefore, the ADC system can present a complete and unified report on the distribution of defects in a set of samples. Due to the high purity of the automatic classification results, the unified report can provide the system operator with the most comprehensive and accurate view of the defect distribution.

[0073] In some embodiments, the second classification result together with the corresponding defect image can be used to improve the automatic classifier. For example, a multi-class classifier for common defects in the pre-classified training set usually has higher accuracy and purity, while a classifier for less common defect categories has lower accuracy and higher rejection rate (because the number of defects in these categories is small, so they are not well represented in the training data). The second classification result may be particularly helpful in improving the classifier for less common defect categories. Once a sufficient number of defects have been accumulated for a less common defect category, the less common defect category can be added to the training set to improve the accuracy and purity of the less common defect category. As less common defect categories are added, the confidence of each defect category will also increase, ultimately reducing the number of defects that are rejected.

[0074] According to embodiments of the present disclosure, automatic defect classification (ADC) may be provided for the display industry. According to some embodiments, which may be combined with other embodiments described herein, a large area substrate having one or more display devices provided thereon may be measured or inspected with at least one of an automated optical inspection tool or an electron beam test tool to obtain a plurality of locations associated with potential defects. Thus, the defect coordinates may be provided to an EBR tool. As shown below, Figures 9A to 9DAccording to the exemplary description, the EBR tool can determine a high-resolution defect image. According to some embodiments that can be combined with other embodiments described herein, additional defect information can be provided, such as material, size and / or connection to adjacent patterns. According to some embodiments that can be combined with other embodiments described herein, the electron beam image data can provide information about material differences. According to some modifications, the electron beam image data can include the material of each area in the image. For example, the material of the particle (i.e., contaminant) can be included in the electron beam, image date. According to some embodiments, EDX (energy dispersive X-ray spectroscopy) measurements can be performed using an electron beam microscope.

[0075] The EBR tool can provide a result file in a first data format, i.e., electron beam image data with optional additional information. The result file includes a high-resolution SEM image. A high-resolution SEM image can be provided for each substrate position determined by the EBR tool. An interface algorithm (e.g., interface software) converts the first data format into a second data format for automatic defect classification by the ADC server. As described above, the ADC server provides automatic defect classification using a machine learning algorithm. For example, the classification results can be provided to the interface algorithm so that the defect classification data is added to the first data format. The updated information can be reported to a host, such as a display manufacturing plant network. Defect classification is used for processing tools and yield optimization.

[0076] According to some embodiments, which can be combined with other embodiments described herein, the interface algorithm can add the ADC results to the first data format. For example, the defect type, defect nature, defect location (i.e., exact defect location), defect periodicity, defect shape (e.g., roundness) can be provided in the first data format. The corresponding variables can pre-exist in the first data format and the values ​​are provided by the interface algorithm. The corresponding data fields (e.g., variables) can also be added to the first data format.

[0077] According to some embodiments that can be combined with other embodiments described herein, in-line automatic defect classification is achieved. For example, large area substrate processing (i.e., display manufacturing) can include loading a large area substrate having one or more structures of the one or more display devices from a first production chamber into a vacuum chamber, the vacuum chamber having an electron beam microscope coupled to the vacuum chamber, the electron beam microscope being configured to measure an electron beam image for the one or more electron beam image data. After generating the SEM image, the large area substrate (i.e., the entire large area substrate) can be loaded directly or indirectly from the vacuum chamber of the EBR tool into another production chamber. According to some embodiments that can be combined with other embodiments described herein, the substrate can be loaded into a repair station, especially after determining a killer defect (e.g., with an EBT tool), the defect can be corrected in the repair station.

[0078] Embodiments of the present disclosure may further include compensation for defect locations (ie, defect location offsets or defect location arrows) that have been provided by a previous inspection system (eg, an AOI tool or an EBT tool).

[0079] Figure 8 A diagram of an automatic defect classification system in the form of an exemplary machine is illustrated, the machine being, for example, a computer system 800, within which a set of instructions for causing the machine to perform any one or more of the methodologies and inspection methods discussed herein may be executed. The machine may be connected (e.g., networked) to other machines in a LAN, an intranet, an extranet, or the Internet. The machine may operate as a server or client machine in a client and server network environment, or as a peer machine in a peer-to-peer (or distributed) network environment. Further, while only a single machine is shown, the term "machine" should also be construed to include any collection of machines that individually or collectively execute a set (or multiple sets) of instructions to perform any one or more embodiments of the present disclosure.

[0080] The exemplary computer system 800 includes a processing device (processor 802), a main memory 804 (e.g., read-only memory (ROM), flash memory, dynamic random access memory (DRAM) such as synchronous dynamic random access memory (SDRAM), double data rate SDRAM (DDR SDRAM) or Rambus DRAM (RDRAM)), etc.), a static memory 806 (e.g., flash memory, static random access memory (SRAM), etc.), and a data storage device 818 that communicate with each other via a bus 808.

[0081] The processor 802 represents one or more general processing devices such as a microprocessor, a central processing unit, etc. In more detail, the processor 802 may be a complex instruction set computing (CISC) microprocessor, a reduced instruction set computing (RISC) microprocessor, a very long instruction word (VLIW) microprocessor, or a processor that implements other instruction sets, or a processor that implements a combination of instruction sets. The processor 802 may also be one or more special-purpose processing devices, such as an application-specific integrated circuit (ASIC), a field programmable gate array (FPGA), a digital signal processor (DSP), a network processor, or the like. The processor 802 is configured to execute instructions 826 for performing the operations discussed herein. The computer system 800 may further include a network interface device 822. The computer system 800 may also include a video display unit 810 (e.g., a liquid crystal display (LCD) or a cathode ray tube (CRT)), an alphanumeric input device 812 (e.g., a keyboard), a cursor control device 814 (e.g., a mouse), and a signal generating device 816 (e.g., a speaker).

[0082] The data storage device 818 may include a computer-readable storage medium 824 having stored thereon one or more sets of instructions 826 (e.g., software) embodying any one or more of the methodologies or functionality described herein. The instructions 826 may also reside completely or at least partially within the main memory 804 and / or within the processor 802 during execution of the instructions by the computer system 800, the main memory 804 and the processor 802 also constituting computer-readable storage media. The instructions 826 may include instructions for integrating automatic defect classification and manual defect classification and / or a software library containing instructions for integrating automatic defect classification according to embodiments of the present disclosure.

[0083] Figures 9A to 9D An exemplary embodiment of a method for defect review measurement is shown, which may be used as electron beam image data for automatic defect classification according to embodiments described herein. Fig.9A Reference image 10 is shown. For example, the image may include a portion of a thin film transistor of a pixel of a display. The image may be a scanning electron microscope image. For example, signal electrons are measured when primary electrons strike a substrate, i.e., the signal intensity may be measured. The intensity signal of the signal electrons may be displayed to produce an image. Reference image 10 shows structure 14. Structure 14 corresponds to a structure manufactured during display manufacturing. According to some embodiments that may be combined with other embodiments described herein, reference image 10 may also include feature 12 (see also Figure 7 Element symbol C in A). Feature 12 may be an unwanted feature or a peculiar feature that may not cause a defect but is not intended for a perfectly manufactured structure 14.

[0084] Fig. 9B A defect image is displayed. The defect image includes a defect 22. A comparison between the reference image 10 and the defect image 20 is calculated. For example, the comparison image 30 is generated by calculating the brightness difference between the reference image 10 and the defect image 20. For example, a defect image and a reference image that exactly match will produce a black comparison image, i.e., an image without deviation. The difference between the reference image and the defect image appears as a bright spot, i.e., a brightness deviation. For example, the absolute value of the difference between the reference image and the defect image can be calculated and / or plotted. The greater the deviation of the intensity signal, the greater the absolute difference, and therefore the brighter the area in the comparison image. According to some embodiments that can be combined with other embodiments described herein, the comparison image can be additionally or alternatively generated by a filter and a further image processing routine, wherein the defect image and the reference image are compared.

[0085] like Fig.9DAs shown, a mask pattern 32 is overlaid on the comparison image 30. The mask pattern 32 is generated by the structure 14 of the reference image 10. According to an embodiment of the present disclosure, the structure may include one or more features selected from the group consisting of: a guide hole, a line, a groove, a connection, a material boundary, an etch layer structure, or the like. According to some embodiments that can be combined with other embodiments described herein, the structure can be a part of a thin film transistor or another transistor for operating a pixel of a display. According to some embodiments that can be combined with other embodiments described herein, the mask pattern 32 is generated by a pattern recognition method.

[0086] According to some embodiments, which can be combined with other embodiments described herein, mask pattern 32 can include feature 12, i.e., a singular feature of reference image 10. Since feature 12 is not intentional, it may cause a brightness difference between the reference image and the defect image. However, since the reference image does not include the defect, the brightness difference corresponding to feature 12 may cause an incorrect defect detection. Comparative image 30 is masked by mask pattern 32, and areas of the mask pattern are not considered. Therefore, the mask pattern including feature 12 prevents incorrect defect detection of the singular feature.

[0087] Additionally or alternatively, the mask pattern 32 masks brightness differences 24 caused by manufacturing tolerances (such as edge roughness or other manufacturing tolerances) that may produce false defect detections in the comparison image 30. Therefore, an acceptable deviation of an image at a first location (such as the reference image 10) relative to an image at a second location (such as the defect image 20 having the same pattern) may not result in a defect alarm because the acceptable deviation is masked by the mask pattern 32.

[0088] like Fig. 9C As shown, defect 22 (see also Figure 7 C) shows the brightness difference between the reference image 10 and the defect image 20. The defect 22 is located outside the mask pattern 32. The defect 22 is detected by comparing the defect image and the reference image in the area outside the mask pattern. According to some embodiments that can be combined with other embodiments described herein, one or more best defect candidates outside the mask pattern are searched in the comparison image 30. After one or more defects (e.g. Fig. 9C After detecting the defect 22 in the image, defect detection at the defect location can be further provided without the mask pattern 32. Fig.10 A class image 70 of an exemplary defect is shown in FIG. 7 , which may be imaged with a higher resolution.

[0089] According to some embodiments, which can be combined with other embodiments, a class image view can be provided with a higher resolution, for example, by rescanning the required FOV with an imaging charged particle beam microscope. The class image view can be based in particular on the determined defect contour. Depending on the size of the defect contour, the class image or the class image view can show an area comprising the defect, the size of the area having a predetermined ratio compared to the size of the defect contour. Additionally or alternatively, a higher magnification image showing more defect details can be provided as a class image. The partial image can be a digital zoom of the defect image.

[0090] There are several advantages to a multi-step method including defect selection with a mask pattern and further defect re-detection of the selected defects without a mask pattern. These advantages can be tailored to the manufacturing conditions of the display. During the display manufacturing process, more obvious pattern edge roughness will not cause false defects. Due to the masking, defects in the remaining area (i.e., the unmasked area (region / area) in the image) can be searched with higher sensitivity. Defect candidates that are partially covered by the mask or separated by the mask are contour-corrected by a second local defect detection operation (i.e., a second defect detection without a mask pattern). Therefore, a correct defect contour (i.e., a defect contour without a mask) can be provided, which is beneficial for defect type classification. The correct defect contour (i.e., defect detection without a mask pattern) allows the determination of the real defect area and the real defect size.

[0091] In summary, some embodiments that can be combined with other embodiments described herein include: generating a defect image of a portion of a large-area substrate, the portion including defects; generating a reference image corresponding to the defect image; determining a mask pattern based on the reference image; comparing the defect image and the reference image in an area outside the mask pattern to detect the defects; and re-detecting the defects without the mask pattern to generate the electron beam image data.

[0092] While the foregoing is directed to certain embodiments, other and further embodiments may be devised without departing from the basic scope thereof, and the scope thereof is to be determined by the following claims.

Claims

1. A method for defect classification, comprising: storing a plurality of defect categories in a multidimensional feature space according to a plurality of classification rules, wherein the plurality of classification rules define, for each defect category in the plurality of defect categories, a boundary of an area associated with the defect category in the multidimensional feature space; receiving one or more electron beam image data associated with a plurality of defects detected in one or more display devices on a large area substrate under inspection; applying, by a processor, an automatic classifier to the electron beam image data, the automatic classifier being based on the plurality of classification rules; as well as The plurality of defects are identified, each defect in the plurality of defects being classified with at least a first confidence level based on at least one confidence threshold.

2. The method of claim 1, wherein the one or more electron beam image data are received in a first data format and converted by an interface algorithm into a second data format for application of the automatic classifier. The method of claim 2 , wherein the first data format includes a reference image and a defect image. The method of claim 3 , wherein the first data format further comprises a category image.

5. The method according to any one of claims 2 to 4, further comprising: A defect classification is added, by the interface algorithm, to the first data format for each defect in the plurality of defects identified with at least the first confidence level.

6. The method according to any one of claims 1 to 5, further comprising: generating a defect image of a portion of the large-area substrate, the portion including the defect; generating a reference image corresponding to the defect image; determining a mask pattern based on the reference image; comparing the defect image in an area outside the mask pattern with the reference image to detect the defect; as well as The defect is re-detected without the mask pattern to generate the electron beam image data.

7. The method according to any one of claims 1 to 6, further comprising: loading a large area substrate having one or more structures of the one or more display devices from a first production chamber into a vacuum chamber having an electron beam microscope coupled to the vacuum chamber, the electron beam microscope configured to measure an electron beam image for the one or more electron beam image data; and The large area substrate is loaded directly or indirectly from the vacuum chamber into a second production chamber.

8. The method of claim 7, further comprising: Prior to loading the large area substrate into the second production chamber, the large area substrate is loaded into a repair station.

9. The method according to any one of claims 1 to 8, further comprising: The one or more display devices are measured using at least one of an automated optical inspection tool or an electron beam test tool to obtain a plurality of locations associated with the plurality of defects.

10. The method of claim 9, further comprising: A defect position offset is compensated for each of the plurality of positions.

11. The method according to any one of claims 1 to 10, further comprising: Electron beam image data is provided to an operator for defects classified below the first confidence level, wherein the first confidence level indicates that the defect is located in an overlapping region between respective boundaries of at least two of the plurality of defect classes and / or outside a boundary of a single class classifier.

12. The method of any one of claims 1 to 11, wherein applying the automatic classifier comprises: applying a multi-class classifier to the electron beam image data to classify the plurality of defects; as well as A single class classifier is applied to identify the plurality of defects using the at least one confidence threshold.

13. The method of claim 12, further comprising: Set purity levels and / or confidence thresholds to adapt the reject rate.

14. A method for generating a plurality of classification rules, the method comprising: receiving a plurality of electron beam image data associated with a plurality of defects detected in one or more display devices on a large area substrate; receiving defect categories, each defect category being associated with one or more of the plurality of electron beam image data; as well as The plurality of classification rules are generated in a multi-dimensional feature space, wherein the plurality of classification rules define, for each defect category in a plurality of defect categories, boundaries of a region associated with the defect category in the multi-dimensional feature space.

15. The method of claim 14, further comprising: Prior to generating the plurality of classification rules, the plurality of electron beam image data are converted from a first data format to a second data format by an interface algorithm.

16. The method according to any one of claims 14-15, wherein the first data format comprises a reference image, a defect image and a category image.

17. An automatic defect classification system comprising: A memory comprising instructions, and a processor, wherein the instructions, when executed by the processor, cause the automatic defect classification system to perform the method of any one of claims 1 to 16.

Images