Method for the automated classification of defects on photolithography masks

Abstract

The invention relates to a method (100) for user-side training of an artificial intelligence (1) for automated classification of defects on photolithography masks based on an aerial image of a photolithography mask, and a method (200) for using the artificial intelligence thus trained. The artificial intelligence (1) is trained based on defect candidates identified in an aerial image and on simulated defect candidates generated according to existing rules. By grouping at least the identified defect candidates into defect groups, the user effort required during training can be significantly reduced. The application of the artificial intelligence (1) is based on a captured aerial image on which defect candidates can be identified, which are then classified by the artificial intelligence (1) trained according to the invention.

Description

[0001] The invention relates to a method for user-side training of an artificial intelligence for the automated classification of defects on photolithography masks based on an aerial image of a photolithography mask, and a method for using the artificial intelligence thus trained.

[0002] Photolithography is used in the fabrication of microstructured components, such as integrated circuits. The photolithography process is carried out in a projection exposure system, which includes an illumination unit and a projection unit. The image of a mask (also called a "reticule") illuminated by the illumination unit is projected by the projection unit onto a substrate, such as a silicon wafer, coated with a photosensitive layer (called a "photoresist") and positioned in the image plane of the projection unit. This transfers the mask structure onto the photosensitive coating of the substrate. In subsequent manufacturing steps, the transferred structure is incorporated into the substrate, for example, by etching.

[0003] Even if the projection devices of projection exposure systems have a reduction factor of, for example, 8:1, the mask structures themselves must exhibit high accuracy due to the ongoing miniaturization in the semiconductor industry and the transition of the wavelength during exposure from DUV (e.g., 193 nm) to EUV (e.g., 13.5 nm). To ensure that a mask meets these accuracy requirements and that a microstructured component produced with it also exhibits the desired properties and functionality, the mask is checked for deviations from the intended structure using suitable methods before being used in the projection exposure system.

[0004] To perform this inspection, mask inspection devices are known. With such devices, masks can be inspected before operation in a microlithographic projection exposure system or during an operational downtime to detect potential defects or contaminants that could lead to the rejection of semiconductors manufactured based on the mask. For this purpose, one or more so-called aerial images of a section of the photomask are generated, which can then be examined for defects and contaminants. The term "aerial image" generally refers to the intensity distribution in the image plane of an optical system.To generate aerial images, the mask is illuminated by a lighting system with radiation of a suitable wavelength, and the radiation transformed by the mask is focused by an optical system with one or more optical elements onto an image sensor suitable for the selected wavelength in an image plane. In optical systems that have a multitude of image planes, a corresponding number of aerial images can be defined as needed, or a single aerial image can have a corresponding number of dimensions or planes.

[0005] The aerial images obtained in this way are then examined for possible defects. Potential defects can be identified simply by analyzing the aerial images, for example by recognizing unusual structures or missing areas, or by comparing them with a reference, such as a "target aerial image" showing the desired defect-free mask.

[0006] Subsequently, any potential defect must be classified. This classification involves determining the type of defect and assessing its impact on a microstructured component manufactured using the mask. While the type of defect can provide clues to its source in the mask's manufacturing and handling process and / or suggestions for its elimination, the assessment of the defect's impact reveals the need to either eliminate the defect or—if repair is not possible—dispose of the mask as scrap.

[0007] This classification is performed by trained personnel according to the state of the art, and this classification can be supported by individual computer-implemented tools. However, it has been shown that further automation of the classification of potential defects on photolithography masks fails fundamentally because manufacturers of microstructured components apply highly individual criteria for evaluating potential defects, which cannot be consolidated into a common scheme, for example, by the manufacturer of mask inspection equipment. Consequently, the classification for each potential defect is performed individually by the aforementioned trained personnel, which is very time-consuming and resource-intensive.

[0008] Furthermore, the assessment of the impact of a mask defect on a microstructured component manufactured with that mask is particularly dependent on the experience of the personnel and can vary considerably. The established procedure is therefore almost always subject to the risk of subjective misjudgment.

[0009] The object of the present invention is to create a method in which the disadvantages of the prior art no longer occur or only occur to a reduced extent.

[0010] This problem is solved by a method according to claim 1. Advantageous further developments are the subject of the dependent claims.

[0011] Accordingly, the invention relates to a method for user-side training of an artificial intelligence for the automated classification of defects on photolithography masks based on an aerial image of a photolithography mask, comprising the following steps: a) Identifying potential defects in the aerial photograph; b) Grouping the identified defect candidates into defect groups; c) Querying the classification of defect candidates and / or defect groups from a user, for which the defect candidates and / or at least one defect candidate representative of a defect group are displayed to the user; d) Generating simulated defect candidates by modifying the classified defect candidates according to rules existing for the respective classification; e) Querying the classification of simulated defect candidates from a user; f) Training the artificial intelligence using the classified defect candidates and classified simulated defect candidates to automatically classify a defect candidate not classified by a user.

[0012] The invention further relates to a method for the automatic classification of defects on photolithography masks based on an aerial image of a photolithography mask, comprising the steps a) Identifying potential defects in the aerial photograph; and b) Use of the artificial intelligence trained using the method according to the invention for classifying the identified defect candidates.

[0013] The invention utilizes artificial intelligence to (partially) automate the classification of defects on photolithography masks, thereby achieving a faster and less error-prone classification of potential defects. The invention recognizes that, due to application- and user-specific requirements, as well as the unavailability of relevant training data due to high confidentiality, it is not possible to train an artificial intelligence generally to classify defects on photolithography masks. Instead, an application- and / or user-specific artificial intelligence must be trained based on a manageable number of potential training datasets to enable the rapid, at least partial, automation of the classification of potential defects.

[0014] The inventive method is based on an aerial photograph of a photolithography mask. The aerial photograph can be created using any known mask inspection device. To create an aerial photograph, the photolithography mask is exposed – usually section by section – with the same exposure settings and wavelength as those used later in the actual photolithography process. In contrast to photolithography, where the mask structure is reduced in size and projected onto a substrate, to generate an aerial photograph the mask structure is enlarged, projected onto an image acquisition sensor, and digitally stored. The aerial photograph, either section by section or assembled from the captured images, thus corresponds to the image or its corresponding section that would be produced on the substrate in a projection exposure system under production conditions.

[0015] Aerial images can be generated in various ways, for example, through imaging techniques. Suitable imaging techniques include: the use of a particle beam system such as a scanning electron microscope (SEM), a fine ion beam microscope (FIB), or an atomic force microscope (AFM); the use of an aerial imaging measurement system, for example, equipped with a staring array sensor, a line scanning sensor, or a time-delay integration (TDI) sensor; or the use of a camera adapted to capture images at predetermined wavelengths. The camera can be an EUV camera or a camera incorporating a TDI sensor. Accordingly, the camera's image sensor can be an EUV image sensor, meaning a sensor sensitive to EUV light.EUV light is light in the extreme ultraviolet spectrum with wavelengths between 5 nm and 100 nm, particularly between 5 nm and 30 nm. Most preferably, the EUV light can have a wavelength of 13.5 nm. The EUV camera can be adapted for use in a photolithography mask inspection system, wherein the photolithography mask is projected onto an EUV image sensor of the EUV camera. In preferred embodiments, the imaging method comprises illuminating a photolithography mask with actinic radiation whose wavelength lies within the EUV wavelength range. EUV radiation reflected from the photolithography mask is then projected onto an image sensor of the EUV camera by means of appropriately adapted projection optics. In a first step (a), defect candidates are identified on the aerial image. "Defect candidates" are areas or...Areas on the aerial photograph that may represent a defect and therefore require inspection. For the sake of clarity, both the actual defect candidate and a section of the image in which the defect candidate is visible are collectively referred to as the "defect candidate," with the specific meaning of the term becoming immediately clear to the expert from the context. The image section in question typically also includes the surrounding structure of the photolithography mask.

[0016] Since the sheer volume of data makes it highly theoretical for a user to identify potential defects in an aerial photograph, various methods for automating defect identification are already known in the art. One such method involves identifying potential defects through image comparison. This is achieved by comparing a captured aerial photograph, or parts thereof, with at least one reference image—for example, a previously captured aerial photograph or a section thereof of fundamentally or at least partially identical photolithography masks. Any deviation indicates a potential defect in one of the aerial photographs. The term "deviation" is to be interpreted broadly and refers to all differences arising from the image data being compared.

[0017] If a photolithography mask contains a large-scale repeating structure, an area of ​​this structure in an aerial photograph can be compared with another area of ​​the same aerial photograph as a reference image. Any deviation then indicates a potential defect in one of the compared areas. It is also possible to compare a captured aerial photograph with an image generated from the design data of the photolithography mask. Deviations then directly indicate potential defects in the aerial photograph. Various methods for generating such images from design data are known in the prior art, for example, using a virtual model of the device used to capture the aerial photograph. The desired image then corresponds to the mapping of a virtual photolithography mask generated from the design data by the virtual model onto its virtual image plane.

[0018] The term "deviation" should be understood in a broad sense.

[0019] If the artificial intelligence to be trained using the inventive method, which is explained in more detail below, is already sufficiently trained, the identification of defect candidates can preferably also be supported or even carried out by the artificial intelligence: If a sufficient number of defect candidates were already taken into account during the training of the artificial intelligence according to the invention, or if the training intentionally or unintentionally included a sufficiently large number of training sets of defect-free aerial images or areas thereof, the artificial intelligence can identify defect candidates or defect-free areas on the aerial image with sufficient reliability – possibly even without a comparison image being available.

[0020] Following the identification of defect candidates, step (b) involves grouping the identified defect candidates into defect groups. Defect candidates are grouped together based on their similarity to each other in the aerial image. Of course, it is also possible that one, several, or even all defect groups may each contain only a single defect candidate, for example, because there is insufficient similarity between one defect candidate and another.

[0021] The comparison of identified defect candidates for similarity, necessary for grouping them into defect groups, can be performed using known image analysis or image comparison methods. For example, the size, shape, or detail of defect candidates in aerial images can be captured and automatically compared, so that if there is sufficient similarity, defect candidates can be grouped together.

[0022] If the artificial intelligence to be trained using the inventive method, which is explained in more detail below, is already sufficiently trained, the grouping of identified defect candidates into defect groups can preferably also be supported or even carried out by the artificial intelligence: If the artificial intelligence can already determine a first proposal for a possible classification of the defect candidates, the defect candidates can be grouped on the basis of the classification proposals.

[0023] The user is then prompted to classify the defect candidates grouped into defect groups. For this purpose, the defect candidates of a defect group, or at least one defect candidate representative of the defect group, are displayed to the user. The user can also be shown a selection of possible classifications from which to choose the appropriate one. Naturally, the user can also add one or more classifications to the list. The classification can be user- and / or application-specific; that is, the classification offered is not a fixed, standardized set of classifications. Rather, a user or, for example, a company can define its own individual classification, possibly even for specific types of photolithography masks.

[0024] To make the classification of defect candidates as efficient as possible for the user, the classification is performed at the defect group level. In other words, the user is prompted for the classification for a defect group, which is then applied to all defect candidates within that group. The user is given the option to remove defect candidates that do not belong to the defect group beforehand, so that the classification performed for the remaining defect group is not applied to those previously removed. This prevents the misclassification of a defect candidate if it is incorrectly assigned to a defect group, without requiring the user to revert to individually classifying each defect candidate.

[0025] The classifications offered to the user preferably include classifications that identify a defect candidate or group as not containing an actual defect. This also includes defect candidates where apparent defects are merely imaging effects in aerial image acquisition that appear as defects. Such a classification can improve artificial intelligence's ability to identify defect candidates in aerial images—provided it is used for this purpose, as described above. Furthermore, subsequent steps can be omitted for defect candidates classified within this framework, thus reducing both computing power and the user's workload.

[0026] Once the artificial intelligence (AI) to be trained using the inventive method, which will be explained in more detail below, has been sufficiently trained, the AI ​​can preferably also assist the user in querying the classification of the identified defect candidates. In this way, the AI ​​can suggest to the user the classification it considers most likely. It is essential that, even though the AI ​​can assist with the classification, the final confirmation of which classification a defect group should be assigned to is made by the user. This is the only way to ensure that the AI ​​is trained quickly and effectively.

[0027] Based on the classification requested by the user, simulated defect candidates are then generated from the classified defect candidates by modifying the existing defect candidates according to rules specific to the respective classification. The rules governing which modifications can and should be made to the defect candidates of the corresponding classification to generate a large number of simulated defect candidates may overlap or even be at least partially identical across the different classifications. However, because there are fundamentally classification-specific rules, modifications that lead to meaningful simulated defect candidates for a defect candidate in a particular classification, but are meaningless for defect candidates in other classifications, can be limited to the relevant defect candidates.

[0028] The modifications required to generate simulated defect candidates can include image-technical changes to the defect candidate or its image, such as altering brightness and / or contrast, scaling and / or rotating the entire image, adding or reducing noise, etc. Alternatively or additionally, the modifications can also include subject manipulation, i.e., manipulating the actual defect candidate while the surrounding structure remains fundamentally unchanged. In this way, the actual defect candidate can be modified in isolation from the surrounding structure, for example, by scaling, translation, and / or rotation.Motif manipulation also includes isolating the potential defect from a defect candidate and then copying it onto another, actually defect-free area of ​​the aerial image, thus altering the structure of the photolithography surrounding the potential defect.

[0029] The rules governing the modification of classified defect candidates to generate simulated defect candidates must be adjusted in light of the subsequent query of the classification of the simulated defect candidates. For example, if an existing rule specifies that the contrast of a classified defect candidate can be reduced by up to 50%, and simulated defect candidates with a 50% reduction in contrast are regularly not analyzed identically to the underlying classified defect candidate (whereas this is the case for simulated defect candidates with only a 40% reduction in contrast), the rule for the affected classification can be adjusted so that the contrast is reduced by a maximum of 40% in the future. Confidence requirements can, of course, be specified for this rule adjustment.It is also possible to occasionally slightly exceed the ranges specified by the rules, such as for contrast reduction, in order to test whether the existing rules remain plausible. The described adjustment of the existing rules for modifying the classified defect candidates to generate simulated defect candidates can be carried out using artificial intelligence or at least supported by it. The artificial intelligence used for this purpose does not necessarily have to be the same AI that is being trained, or at least partially trained, to classify defect candidates.

[0030] It is also possible that simulated defect candidates are generated by generative artificial intelligence based on classified defect candidates.

[0031] For the simulated defect candidates generated based on the existing rules, the user is prompted to select the appropriate classification, for which the defect candidates are displayed. The user can then choose the correct classification for the displayed defect candidate.

[0032] The classification can be queried individually for each simulated defect candidate. However, it is also possible, analogous to querying the classification of identified defect candidates grouped into defect groups, to group the simulated defect candidates into defect groups. The simulated defect candidates are then displayed to the user in groups, with the classification of the simulated defect candidates being queried at the defect group level. Ideally, the user can remove simulated defect candidates that do not fit into the defect group. For further explanation, please refer to the preceding sections.

[0033] For querying the classification of simulated defect candidates and / or defect candidate groups, the user can also be presented with a suggested classification. This suggested classification can be derived from the classification of the identified defect candidate on which the simulated defect candidate to be classified is based. If the artificial intelligence, which is trained using the inventive method and will be explained in more detail below, is already sufficiently trained, the user's query for the classification of the simulated defect candidates can also be supported by the artificial intelligence. During the query, the user is presented with the classification that the artificial intelligence considers most probable.It is also essential here that, regardless of how a classification proposal is generated, the final confirmation of which classification a simulated defect candidate or group of defect candidates should be assigned to is made by the user. Only in this way can rapid, targeted training of the artificial intelligence be ensured.

[0034] To further reduce the user's workload, it is preferable to refrain from querying the classification of simulated defect candidates (including those identified in aerial imagery) that offer little or no learning potential for the artificial intelligence. If the classification of a specific type of defect candidate is already sufficiently refined, a user classification of that same type will typically have little or no impact on the classification accuracy of the AI ​​being trained. In cases where a large number of identified or simulated defect candidates are available, it can be advantageous to only query the user for the classification of those candidates where a high "learning effect" for the AI ​​can be expected.

[0035] Upon completion of the classification of the simulated defect candidates, in addition to the identified defect candidates, there are also simulated defect candidates derived from them, each of which has been classified by the user. This results in a number of user-verified training data points that exceeds the number of defect candidates actually identified in the aerial imagery – namely, increased by the number of simulated defect candidates. The inventive method ensures that even with a potentially manageable number of identified defect candidates, a larger amount of training data is available, while the effort required by the user for classification generally remains manageable.

[0036] All defect candidates classified in the process, whether simulated or identified in aerial imagery, are then used to train an artificial intelligence (AI) to automatically classify a defect candidate not classified by a user. By providing an AI with training data containing defect candidates and their corresponding classifications, it can be trained to independently classify a defect candidate not classified by a user at a later time (after receiving a sufficient amount of training data).

[0037] In particular, if the artificial intelligence to be trained using the inventive method has already been used in one of the preceding steps to identify defect candidates, the identification of defect candidates on the aerial image can be carried out again after training based on the previously acquired training data. The aerial image is the one that was analyzed previously; however, due to the further training of the artificial intelligence, it is possible that previously unrecognized defect candidates will be identified. If new defect candidates are indeed identified, the steps described above, up to and including the training of the artificial intelligence for these newly identified defect candidates, can be carried out.

[0038] Even if the method according to the invention serves the purpose of user- and / or application-specific training of the artificial intelligence, the artificial intelligence can – particularly if the at least partially trained artificial intelligence is to be used in one of the steps of the method – be initially trained with a user- and / or application-independent basic training dataset. By training the artificial intelligence with a corresponding basic training dataset, it can be used for various tasks even during the first run of the method. This eliminates the need for potentially complex provision of alternative algorithms for tasks that are to be performed at a later time by an at least partially trained artificial intelligence. The basic training dataset is preferably designed in such a way that it influences the artificial intelligence as little as possible.Any potential initial bias is suppressed as quickly as possible by the inventive training process. The artificial intelligence can also be trained to discard the basic training data as soon as sufficient training data is available through the inventive method.

[0039] It is preferred if at least one or at least part of an artificial intelligence is based on a machine learning algorithm, preferably random decision forests or k-nearest neighbor, or a deep learning algorithm, preferably convolutional neural networks or vision transformer networks. It is also possible to use diffusion models. The corresponding algorithms are known from the prior art for other applications than the one presented here, but can, in principle, be adapted to the present application by a qualified professional.

[0040] For an explanation of the inventive method for the automatic classification of defects on photolithography masks based on an aerial image of a photolithography mask, reference is made to the preceding statements.

[0041] If a defect candidate from an aerial image is presented to an artificial intelligence trained (fully) according to the invention, the artificial intelligence can perform the correct classification of the defect candidate with a reliability at least comparable to that of a human user.

[0042] The potential defect can, in principle, be identified in any way on an aerial photograph. However, it is preferred if the artificial intelligence trained according to the invention is used to identify potential defects on an aerial photograph. Reference is made to the corresponding descriptions of how to identify potential defects on an aerial photograph using the artificial intelligence – which is only partially trained there – in the course of the user-side training process for the artificial intelligence.

[0043] The photolithography mask can have an aspect ratio between 1:1 and 1:3, preferably between 1:1 and 1:2, and particularly preferably 1:1 or 1:2. The photolithography mask can be substantially rectangular. The photolithography mask can preferably be 12.7 cm (5 inches) to 17.8 cm (7 inches) long and wide, and particularly preferably 15.2 cm (6 inches) long and wide. Alternatively, the photolithography mask can be 12.7 cm (5 inches) to 17.8 cm (7 inches) long and 25.4 cm (10 inches) to 35.6 cm (14 inches) wide, and preferably 15.2 cm (6 inches) long and 30.5 cm (12 inches) wide.

[0044] The invention will now be described by way of example using an advantageous embodiment with reference to the accompanying drawings. These show: Fig. 1: A schematic representation of a method according to the invention for user-side training of an artificial intelligence for the automated classification of defects on photolithography masks based on an aerial image of a photolithography mask; and Fig. 2: a schematic representation of a method according to the invention for the automatic classification of defects on photolithography masks based on an aerial image of a photolithography mask.

[0045] In Fig. Figure 1 is an embodiment of a method 100 according to the invention for user-side training of an artificial intelligence 1 for automated classification of defects on photolithography masks based on an aerial image of a photolithography mask, schematically represented.

[0046] Method 100 is based on one or more aerial photographs of one or more photolithography masks, which are acquired in a preparatory step 90 (not part of the actual method 100) using suitable mask inspection devices. Since such devices and acquisition methods are well known in the prior art, further explanation is unnecessary here.

[0047] The one or more aerial photographs thus acquired are then supplied to the inventive method 100. In a first step 110, potential defects are identified on the at least one aerial photograph by searching the aerial photographs for areas on which possible defects are depicted. For this purpose, two different methods 111, 112 are applied in parallel in the illustrated embodiment.

[0048] In the first method 111, the aerial image to be examined is compared with an image created from the design data for the photolithography mask depicted on the aerial image, and any differences that exceed a predetermined limit for a permissible deviation are identified as defect candidates.

[0049] The second method 112 utilizes the artificial intelligence 1 trained using method 100. If the artificial intelligence 1 has already been initially trained, either by having run the inventive method 100 several times or by having been trained in its basic principles using a provided basic training dataset, it can search an aerial image for areas where, at the current stage of training, it would suggest a classification for a defect with a probability above a predetermined value. The potential defect thus identified is then considered a defect candidate. If the artificial intelligence 1 is not yet sufficiently trained when running the method 100, the application of the method 112 based on it can be omitted.

[0050] The defect candidates are then passed to the next step 120, regardless of the method used to identify them, avoiding duplicate defect candidates found in both methods.

[0051] In step 120, the identified defect candidates are grouped into defect groups. Two different methods, 121 and 122, are provided for this purpose; however, they are not applied in parallel, but rather exclusively.

[0052] Method 121 uses automated image comparison to form groups. For this purpose, the defect candidates, or more precisely, the sections of the aerial photograph containing the individual defect candidates, are examined for similarities using known image analysis methods. For example, if foreign particles are present on a photolithography mask whose aerial photograph was processed using Method 100, the individual foreign particles can be identified as defect candidates, with the corresponding images being characterized by a larger contiguous area with a specific shading. Through automated image comparison, defect candidates with corresponding characteristics can be grouped together.

[0053] Alternatively, in method 122, the already sufficiently trained artificial intelligence 1 can determine a proposal for the classification of each of the defect candidates, whereby those defect candidates for which the artificial intelligence 1 has determined the same proposal for a classification are then grouped together in a defect group.

[0054] To avoid an assignment conflict, procedure 100 provides for the grouping of the identified defect candidates into defect groups using method 121 until the artificial intelligence 1 has been sufficiently trained to enable an improved assignment to defect groups compared to method 121. In this case, only method 122 is then applied.

[0055] Subsequently, in step 130, the identified defect candidates are displayed to the user, grouped into defect groups, for classification.

[0056] The user is shown all identified defect candidates within a defect group and presented with a list of possible classifications, which the user can expand as needed. The user can quickly check whether any of the identified defect candidates displayed are incorrectly assigned to the defect group defined by the other defect candidates and can then remove them from that group. Subsequently, the user can specify the appropriate classification for the remaining identified defect candidates by selecting from the suggested (or, if applicable, user-supplemented) classifications, including options for "not identifiable" or "no defect." The selected classification is then assigned to each defect candidate remaining in the defect group, so that the classification ultimately takes place at the defect group level.Once the defect candidates of a defect group have been classified, the user is shown the next defect group for classification, until all defect groups or the defect candidates contained therein have been classified.

[0057] Provided that the artificial intelligence (AI) 1 has already been sufficiently trained, it can also suggest a suitable classification to the user. If the identified defect candidates were grouped together in step 120 according to method 122, the classification underlying the formation of the defect group can also be suggested to the user. In any case, however, it is important to note that the actual classification is still carried out by the user, even when suggestions are made by the AI ​​1.

[0058] In the next step, 140, simulated defect candidates are generated based on the previously classified identified defect candidates—at least those not classified as "unidentifiable" or "not a defect." The classified identified defect candidates are modified according to the rules applicable to each classification, with the fundamental goal that the simulated defect candidates can be assigned to the same classifications as the underlying identified defect candidate. Alternatively or additionally, individual modifications can be made in such a way that a simulated defect candidate might "just barely" no longer be assigned to the classification of the underlying identified defect candidate, thus enabling the artificial intelligence to later make a sharper distinction between defect and non-defect.

[0059] The modifications to the identified defect candidates to obtain simulated defect candidates include image-technical changes and subject manipulations, such as changing the brightness and / or contrast, scaling and / or rotating the entire image, adding or reducing noise, and modifying the actual defect candidate in isolation from the surrounding structure by scaling, translation and / or rotation.

[0060] For the various classifications, there are fundamentally different rules for modifying the identified defect candidates in order to approximate the basic objective of creating simulated defect candidates of the same classification as the respective underlying identified defect candidate. Since certain modifications suffice for defect candidates of a particular classification to achieve this objective, but the same modifications do not result in a suitable simulated defect candidate for other classifications, it is advantageous to provide the rules for modifications individually for each classification, even if there may be significant overlaps between the respective rules, sometimes even to the point of identity.

[0061] Furthermore, the existing rules can be optimized using the additional artificial intelligence 2. For example, the results from the subsequent step 150, concerning the query of the classification of the simulated defect candidates, can be used to adjust the existing rules as needed, specifically to avoid simulated defect candidates that are not assigned to the same classification as the underlying identified defect candidate. The additional artificial intelligence 2 can also be integrated into the artificial intelligence 1, so that the functionality of the additional artificial intelligence 2 is implemented by the artificial intelligence 1.

[0062] Once the simulated defect candidates have been generated, the user is then prompted to select their respective classification (step 150). This prompt is essentially analogous to the classification prompt for the identified defect candidates in step 130, so please refer to the explanations there – particularly regarding the prompt at the defect group level, where the simulated defect candidates are grouped. Before the defect candidates are displayed to the user, those simulated defect candidates that offer no learning potential for the artificial intelligence are filtered out – the number of which is significantly higher than the number of identified defect candidates. Therefore, no classification prompt appears for these simulated defect candidates.In this way, the number of simulated defect candidates to be classified by the user can be kept low, while at the same time ensuring rapid training of the artificial intelligence 1.

[0063] Based on the classified identified defect candidates and the classified simulated defect candidates—which can be considered training sets—artificial intelligence 1 can then be trained so that it gradually matures into an AI capable of automatically classifying a defect candidate classified by a user. The methods for this are well-known for the various algorithms suitable for artificial intelligence 1 in this case and require no further explanation here. The same applies to the other artificial intelligence 2, which can also be trained appropriately.

[0064] If, as in the present case, the identification of defect candidates on the aerial image (step 110) is also performed using artificial intelligence 1 (see method 112), it is advantageous to repeat the identification of defect candidates on the aerial image (step 110) after completing the training according to step 160. Should additional defect candidates not previously considered be identified, steps 120 to 160 can be repeated for these newly identified defect candidates. This ensures that as many defect candidates as possible can be identified on an aerial image.

[0065] In the illustrated embodiment, the artificial intelligences 1 and 2 are initially trained using a user- and / or application-independent basic training dataset so that the use of the artificial intelligences 1 and 2 in the various steps 110, 120, and 140 of the procedure 100 is possible even on the first run. The basic training dataset is designed in terms of scope and characteristics such that the training data it contains is sufficient after only a few runs of the procedure. Fig. The methods described in 1 no longer shape artificial intelligences 1, 2.

[0066] After a sufficient number of runs of procedure 100 according to Fig. 1. The artificial intelligence 1 is sufficiently trained to actually be used. The corresponding procedure 200 is in Fig. 2 shown.

[0067] In method 200 for the automatic classification of defects on photolithography masks based on an aerial image of a photolithography mask, an aerial image of the photolithography mask to be analyzed is first acquired using a suitable mask inspection device (step 90) before the actual method 200 is carried out. Since corresponding devices and acquisition methods are well known in the prior art, further explanation is unnecessary here.

[0068] The captured aerial image is fed into procedure 200, in which, using the previously trained artificial intelligence, one potential defect is identified in the aerial image. Step 210 is identical to step 110 of procedure 100 according to Fig. 1, which is why reference is made to the explanations given there - also with regard to alternative design variants.

[0069] For the identified defect candidates, the artificial intelligence 1 then performs the classification (step 220). Since the artificial intelligence 1 was trained specifically for this use case, the classification can be performed with a high degree of reliability.

[0070] The artificial intelligence 1, trained using method 100 and used with method 200, is characterized by the fact that it is trained in a user- and / or application-specific manner; that is, the classification is performed according to the user's specifications, possibly based on specific types of structures on the photolithography masks. Method 100 of the training process is particularly distinguished by the fact that, on the one hand, a manageable number of original defect candidates in the captured aerial images is sufficient to train the artificial intelligence 1, thanks to the defect candidates simulated on this basis. On the other hand, the classification-specific rules for generating simulated defect candidates ensure that the number of defect candidates actually to be classified by the user remains manageable.This effort can be further reduced by optionally grouping defect candidates and implementing other possible measures described above. As a result, only with the inventive method 100 is it practical to train an artificial intelligence 1 for the automated classification of defects on photolithography masks in an application- and / or user-specific manner.

Claims

[1] Method (100) for user-trained artificial intelligence (1) for automated classification of defects on photolithography masks based on an aerial image of a photolithography mask, comprising the steps: a) Identifying defect candidates on the aerial image (step 110); b) Grouping the identified defect candidates into defect groups (step 120); c) Querying the classification of defect candidates and / or defect groups from a user, for which the defect candidates and / or at least one defect candidate representative of a defect group are displayed to the user (step 130); d) Generating simulated defect candidates by modifying the classified defect candidates according to rules existing for each classification (step 140); e) Querying the classification of the simulated defect candidates from a user, for which the defect candidates are displayed to the user (step 150); and f) Training the artificial intelligence (1) using the classified defect candidates and classified simulated defect candidates to automatically classify a defect candidate not classified by a user (step 160). [2] Method according to claim 1, characterized by , that the identification of defect candidates on the aerial image is carried out by comparing the aerial image or a portion thereof with a previously acquired aerial image, a comparison image generated from the design data of the photolithography mask or a portion thereof (Method 111). [3] Method according to any one of the preceding claims, characterized by, that the identification of defect candidates is supported or carried out by the artificial intelligence that has already been at least partially trained (1) (Method 112). [4] Method according to claim 3, characterized by , that after completion of the training of the artificial intelligence (1), the identification of defect candidates on the aerial image is supported or carried out by the previously trained artificial intelligence (1), and upon identification of new defect candidates, steps (b) to (f) are carried out for these newly identified defect candidates. [5] Method according to any one of the preceding claims, characterized by , that the grouping of identified defect candidates into defect groups is carried out by automated image comparison (Method 121) and / or is supported or carried out by the artificial intelligence (1) that has already been at least partially trained (Method 122). [6] Method according to any one of the preceding claims, characterized by that the classification of defect candidates takes place at the level of defect groups, whereby preferably defect candidates that do not fit the defect group can be extracted from the defect group by the user. [7] Method according to any one of the preceding claims, characterized by , that when querying the classification, the user is offered possible classifications, whereby the offered classifications preferably also include a classification that identifies a defect candidate as not containing an actual defect. [8] Method according to any one of the preceding claims, characterized by , that the modification of classified defect candidates to generate simulated defect candidates includes image-technical changes and / or motif manipulations. [9] Method according to any one of the preceding claims, characterized by, that the existing rules which are used to generate simulated defect candidates are adapted in view of the query for the classification of simulated defect candidates by a user (step 150). [10] Method according to any one of the preceding claims, characterized by that the simulated defect candidates are grouped into defect groups, the simulated defect candidates are displayed to the user in groups, and the classification of the simulated defect candidates is queried at the level of the defect groups, whereby preferably simulated defect candidates that do not fit the defect group can be extracted from the defect group by the user. [11] Method according to any one of the preceding claims, characterized by , that at least for simulated defect candidates that do not promise a learning effect for artificial intelligence (1), no classification query is performed. [12] Method according to any one of the preceding claims, characterized by , that the artificial intelligence (1) is initially trained using a user- and / or application-unspecific basic training dataset. [13] Method according to any one of the preceding claims, characterized by , that the artificial intelligence (1) is based on a machine learning algorithm, preferably random decision forests or k-nearest neighbor, or a deep learning algorithm, preferably convolutional neural networks or vision transformer networks. [14] Method (200) for automatic classification of defects on photolithography masks based on an aerial image of a photolithography mask, comprising the steps: a) Identifying defect candidates on the aerial image (step 210); and b) Use of the artificial intelligence (1) trained by the method according to any one of claims 1 to 13 for classifying the identified defect candidates (step 220). [15] Method according to claim 14, characterized by , that the identification of defect candidates on the aerial image is supported or carried out by the artificial intelligence (1) trained using the method according to one of claims 1 to 13.

Images