Training machine learning process for evaluating substrate
By increasing the number of labeled examples and adopting supervised training methods, the problem of insufficient performance of machine learning processes in semiconductor wafer evaluation is solved, achieving higher detection accuracy and adaptability.
Patent Information
- Application Number
- CN202510009767.9
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Priority Date
- 2024-01-11
- Filing Date
- 2025-01-03
- Publication Date
- 2025-07-11
AI Technical Summary
When existing machine learning processes are used to evaluate semiconductor wafers, due to the limited number of training examples, the performance is insufficient and it is difficult to accurately detect defects of interest.
By increasing the number of labeled examples, using supervised training methods, generate and use more labeled examples to train machine learning processes, including obtaining patterned feature markers from different substrates, layers, and process formulations, filling data sets, and training machine learning processes.
It significantly improves the accuracy of the machine learning process, can more accurately detect defects of interest on semiconductor chips, and adapt to different substrates and process changes.
Smart Images
Figure CN120297433A_ABST
Abstract
Description
Background Art
[0001] Machine learning processes are used to solve many problems in many technical fields.
[0002] The performance (e.g., accuracy) of a machine learning process is at least partially based on the quantity and quality of examples used during training of the machine learning process. The examples can be mislabeled information.
[0003] It has been found that machine learning processes for evaluating semiconductor wafers are affected by a very limited training process based on a very small number of examples.
[0004] There is an increasing need to improve the performance of machine learning processes for evaluating semiconductor wafers. Summary of the Invention
[0005] An evaluation system utilizing a machine learning process is provided, the evaluation system including: (a) an imager configured to obtain an image of a region of a substrate under evaluation, the substrate under evaluation including an evaluated pattern; and (b) a processor including one or more processing circuits, the processor being configured to apply a machine learning process to the image of the region to find an interested defect associated with a given evaluated pattern in the evaluated pattern. The machine learning process is trained to find the interested defect, wherein training the machine learning process includes feeding an interested defect data set and another data set to the machine learning process. The interested defect data set is generated by: (i) obtaining a feature marker of a substrate pattern of a training-related substrate; (ii) finding a first feature marker of a reference pattern that is similar to the feature marker of the substrate pattern; wherein the reference pattern is associated with a defect previously defined as an interested defect; wherein the finding is performed regardless of one or more parameters affecting the generation of the first reference feature marker; and (iii) filling the interested defect data set with a second feature marker of the reference pattern, the second feature marker of the reference pattern conveying more information about the defect previously defined as an interested defect than the first feature marker of the reference pattern. The another data set is filled with additional feature markers of additional patterns of the substrate.
[0006] Provided is a non-transitory computer-readable medium for training a machine learning process, the non-transitory computer-readable medium storing instructions which, when executed by a processor comprising one or more processing circuits, cause the processor to: (a) obtain a feature signature of a substrate pattern of a training-related substrate; (b) find a first feature signature of a reference pattern that is similar to the feature signature of the substrate pattern; wherein the reference pattern is associated with a defect previously defined as a defect of interest; wherein the finding is performed regardless of one or more parameters affecting the generation of the first reference feature signature; (c) populate a defect-of-interest dataset with a second feature signature of the reference pattern, the second feature signature of the reference pattern conveying more information about the defect previously defined as a defect of interest than the first feature signature of the reference pattern; (d) populate another dataset with additional feature signatures of additional patterns of the substrate; and (e) train the machine learning process in a supervised manner to find the defect of interest, wherein the training includes feeding the defect-of-interest dataset and the other dataset into the machine learning process.
[0007] Provided is a method for training a machine learning process, the method comprising: (a) obtaining a feature signature of a substrate pattern of a training-related substrate; (b) finding a first feature signature of a reference pattern that is similar to the feature signature of the substrate pattern; wherein the reference pattern is associated with a defect previously defined as a defect of interest; wherein the finding is performed regardless of one or more parameters affecting the generation of the first reference feature signature; (c) populating a defect-of-interest dataset with a second feature signature of the reference pattern, the second feature signature of the reference pattern conveying more information about the defect previously defined as a defect of interest than the first feature signature of the reference pattern; (d) populating another dataset with additional feature signatures of additional patterns of the training-related substrate; and (e) training the machine learning process in a supervised manner to find the defect of interest, wherein the training includes feeding the defect-of-interest dataset and the other dataset into the machine learning process. BRIEF DESCRIPTION OF THE DRAWINGS
[0008] The subject matter that is specifically pointed out and distinctly claimed as an embodiment is set forth at the end of the specification. However, the embodiments as to the organization and method of operation, as well as the examples, features, and advantages of the embodiments, may be most thoroughly understood by reference to the following detailed description when read in conjunction with the accompanying drawings, in which:
[0009] Figure 1 Examples of the method are illustrated;
[0010] Figure 2 Examples of the method are illustrated;
[0011] Figure 3 Examples of the method are illustrated;
[0012] Figure 4Illustrates an example of an evaluation system and the environment of the evaluation system; and
[0013] Figure 5 Illustrates an example of one or more substrates.
[0014] It should be understood that, for simplicity and clarity of illustration, the elements shown in the drawings are not necessarily drawn to scale. For example, for clarity, the dimensions of some of the elements may be enlarged relative to other elements. In addition, where considered appropriate, reference numerals may be repeated in the drawings to indicate corresponding or similar elements. Detailed Description
[0015] Provides solutions related to the evaluation of evaluated substrates (such as, but not limited to, wafers, especially wafers having at least one semiconductor portion).
[0016] Evaluating an evaluated substrate may include: inspecting the evaluated substrate to find defects, or inspecting the evaluated substrate to find suspected defects.
[0017] An evaluated substrate may exhibit multiple types of defects. Only some of the multiple types are of interest. Defects belonging to the types of interest are referred to as defects of interest. The definition of which defects are defects of interest can be determined in a variety of ways, for example, by the consumer of the evaluated substrate, the manufacturer of the evaluated substrate, etc.
[0018] Provides solutions for significantly increasing the number of labeled examples provided to a machine learning process during supervised training, thereby enhancing the performance (e.g., accuracy) of the machine learning process.
[0019] The increase in the number of labeled examples allows the machine learning process to be trained in a supervised manner and provides a more accurate machine learning process than the corresponding machine learning process trained using an unsupervised training process or a semi-supervised training process.
[0020] Examples of unsupervised or semi-supervised trained machine learning processes include:
[0021] a. Autoencoders for anomaly detection.
[0022] b. Statistically modeling "typical" pixel values and detecting strong deviations from the model.
[0023] c. Self-supervised learning for distinguishing valid and anomalous transitions.
[0024] d. Semi-supervised training with artificially implanted defects.
[0025] Solutions are provided for evaluating patterns of substrates such as semiconductor wafers using machine learning processes with high accuracy. Other substrates may be provided. The accuracy of the machine learning process is improved at least in part by using the machine learning process to train the machine learning process in a supervised manner to detect defects of interest using an increasing number of labeled examples.
[0026] The reasons for the lack of samples are:
[0027] a. The size of the pattern of the substrate continues to shrink and reaches the sub-micron level and even the nano level.
[0028] b. Evaluating the substrate involves irradiating the pattern with radiation having a wavelength of the same order of magnitude as the size of the pattern, which results in detection signals that are difficult to understand.
[0029] c. The detection signals change significantly from one inspection recipe to another, change significantly between one semiconductor and another semiconductor, and vary significantly between one layer and another layer of the semiconductor.
[0030] d. The training of the machine learning process for evaluating a specific layer of a specific substrate and using a specific recipe is limited to labeled examples pre-obtained from a specific layer of one or more reference substrates (expected to be the same as the substrate being evaluated) and using a specific recipe.
[0031] It has been surprisingly found that, contrary to the above limitation (d), when searching for labeled examples obtained from a much larger number of reference substrates, the number of labeled examples can be significantly increased by ignoring at least one of the following:
[0032] a. The differences between the training-related substrates (used with one or more reference substrates during the construction of at least one dataset for training) and the one or more reference substrates. Thus, labeled examples can be obtained from reference substrates that are different from the training-related substrates in terms of manufacturer, design, and / or layout, etc.
[0033] b. At least one difference between different layers of the training-related substrates and different layers of the one or more reference substrates. Thus, labeled examples can be obtained from another layer of the reference substrate.
[0034] c. Ignoring at least one difference between the recipe used for inspection and the recipe used. Thus, labeled examples can be obtained using a recipe different from the recipe used to obtain information from the training-related substrate.
[0035] Ignoring one or more differences is very beneficial when inspecting new semiconductor wafers of types or manufacturers that have not been inspected before.
[0036] It should be noted that the training or retraining of the machine learning process can occur in one or more situations, such as:
[0037] a. When the training-related substrate is new, i.e., the training-related substrate has not been evaluated by the machine learning process in the past.
[0038] b. When the inference result is considered inaccurate, and the amount of inaccuracy can be defined in any way (e.g., by the consumer of the evaluated substrate, the manufacturer of the evaluated substrate, etc.).
[0039] c. When at least a predefined time period (or at least multiple evaluations) has elapsed since the last training, and the amount of the predefined time period and / or the sufficient evaluations eligible for training or retraining can be defined in any way (e.g., by the consumer of the evaluated substrate, the manufacturer of the evaluated substrate, etc.).
[0040] d. After receiving an indication of a significant process change. The indication can be provided by a third party unrelated to the evaluation process. The amount of the significant process change can be defined in any way (e.g., by the consumer of the evaluated substrate, the manufacturer of the evaluated substrate, etc.).
[0041] e. When receiving a new batch of substrates.
[0042] It should be noted that the training or retraining of the machine learning process can be triggered for a specific one or more layers of the evaluated substrate. For example, the evaluation of one layer can be considered accurate enough and does not require retraining, while the evaluation of another layer can be considered inaccurate, and the machine learning process should be adapted to better evaluate the other layer.
[0043] Figure 1 is an example of method 100 for constructing a data set including labeled examples for training a machine learning process. The construction of the data set provides multiple examples.
[0044] According to an embodiment, method 100 begins at step 110: obtaining a signature of the substrate pattern of the training-related substrate. The substrate is related to the training used during the construction of the data set. According to an embodiment, the use includes generating a signature of the substrate pattern. The generating may include: irradiating the training-related substrate during step 110 to provide a detection signal; and generating a signature of the substrate pattern based on the detection signal.
[0045] The signature of the substrate pattern is information representing the substrate pattern.
[0046] Examples of the signature of the reference pattern include:
[0047] a. An image of the substrate pattern. The image may include any number of pixels, such as an image of 64×64 pixels or an image of any other size.
[0048] b. Embedding of the substrate pattern.
[0049] c. Compressed representation of the substrate pattern.
[0050] d. Any other representation of the substrate pattern.
[0051] According to an embodiment, after step 110, step 120 is performed: finding a first feature marker of the reference pattern that is similar to the feature marker of the substrate pattern.
[0052] According to an embodiment, the finding is performed regardless of one or more parameters that affect the generation of the first feature marker of the reference pattern.
[0053] According to an embodiment, the one or more parameters include the image acquisition process used during the generation of the first reference feature marker. For example, any recipe such as the polarization of the illumination path, the polarization of the reception path, the wavelength of the illumination, the wavelength that can pass through the reception path, the intensity of the illumination, the attenuation of the reception path, etc.
[0054] According to an embodiment, the one or more parameters include the source of the substrate pattern. The source includes the substrate, the layers of the substrate, and combinations thereof.
[0055] According to an embodiment, the reference pattern is associated with a defect previously defined as a defect of interest. The reference pattern can belong to any substrate (including a substrate different from the training-related substrate), and even belong to a substrate manufactured or designed by another manufacturer or designer different from the manufacturer or designer of the training-related substrate.
[0056] According to an embodiment, the reference pattern belongs to a layer different from the layer of the substrate pattern of the training-related substrate.
[0057] According to an embodiment, a different recipe from the recipe used to generate the feature marker of the substrate pattern of the training-related substrate is used to obtain the first feature marker of the reference pattern. Alternatively, the recipe used to generate the feature marker of the substrate pattern of the training-related substrate is used to obtain the first feature marker of the reference pattern.
[0058] According to an embodiment, after step 120, step 130 is performed: filling the defect dataset of interest with the second feature marker of the reference pattern.
[0059] The second feature marker of the reference pattern conveys more information about the defect previously defined as a defect of interest than the first feature marker of the reference pattern.
[0060] According to an embodiment, even when the presence of a defect affects the similarity test used in step 120, the second feature marker includes the feature marker of the reference pattern (including the defect). For example, assume that the feature marker of a pattern includes an image of 64×64 pixels, and a defect having a size of at least 12×12 pixels may prevent finding a similar reference feature marker. In the said assumption, the first feature marker represents a defect-free pattern, and the second feature marker will represent a defective one. The defect-free pattern may be pre-inspected as defect-free or may be a pattern in which the defect is at least partially masked.
[0061] According to an embodiment, step 130 is followed by step 140: filling another data set with additional feature markers of additional patterns of the training-related substrate.
[0062] According to an embodiment, the additional feature markers are estimated or determined as the feature markers of OK patterns.
[0063] According to an embodiment, assuming that the manufacturing process exhibits a satisfactory yield (e.g., about 95%, 96%, 97%, 99%, 99.2%, 99.4%, 99.6%, 99.9%, 99.9%, 99.99%, 99.998% and higher yields), then the additional feature markers can be selected from the feature markers of the training-related substrate in any way (e.g., in a random manner).
[0064] According to an embodiment, the additional feature markers are marked as the feature markers of OK patterns, and the second feature markers are marked as the feature markers of defective patterns.
[0065] Figure 2 is an example of method 200 for training a machine learning process.
[0066] According to an embodiment, method 200 starts with step 210: obtaining (a) a data set of defects of interest filled with second feature markers of reference patterns, and (b) another data set filled with additional feature markers of additional patterns of the substrate.
[0067] According to an embodiment, step 210 includes one of the following:
[0068] a. Receiving the data set of defects of interest and receiving another data set.
[0069] b. Generating the data set of defects of interest and receiving another data set.
[0070] c. Receiving the data set of defects of interest and generating another data set.
[0071] d. Generating the data set of defects of interest and generating another data set.
[0072] According to an embodiment, after step 210 is step 220: training a machine learning process to find defects of interest, wherein the training includes feeding a dataset of defects of interest and another dataset into the machine learning process.
[0073] According to an embodiment, an additional feature marker is marked as a feature marker of an OK pattern, and a second feature marker is marked as a feature marker of a defective pattern.
[0074] According to an embodiment, obtaining a dataset, training a machine learning, or evaluating an evaluated substrate by an evaluation system is performed by a party different from the manufacturer of the evaluation system.
[0075] According to an embodiment, any of the methods illustrated in the specification includes keeping confidential the information obtained during any of the methods and even preventing the manufacturer from accessing such information.
[0076] Figure 3 is an example of method 300 for evaluating a substrate.
[0077] According to an embodiment, method 300 begins with step 310: obtaining an image of a region of an evaluated substrate. The evaluated substrate includes evaluated patterns. The region can be of any size or an area having any fraction of the total size of the substrate.
[0078] According to an embodiment, after step 310 is step 320: applying a machine learning process to the image of the region to find one or more defects of interest associated with one or more given evaluated patterns in the evaluated patterns.
[0079] According to an embodiment, the machine learning process is trained by performing method 200.
[0080] According to an embodiment, after step 320 is step 330: responding to finding one or more defects.
[0081] According to an embodiment, step 330 includes one of the following: storing the result of step 320; preventing a third party from accessing at least the result of step 320; determining whether the machine learning process should be adjusted or retrained; triggering or indicating or requesting an adjustment or retraining of the machine learning process; determining whether the result of step 320 indicates a process change, etc.
[0082] Figure 4 is an example of evaluation system 400 and the environment of the evaluation system.
[0083] A non - limiting example of evaluation system 400 is ENLIGHT from Applied Materials Inc. in Santa Clara, California, USA TM Optical inspection.
[0084] The evaluation system 400 includes an imager 410, a memory unit 417, a controller 418, and a processor 490 including one or more processing circuits 491. The imager 410 includes a detection unit 411 and an optical device 412. The optical device includes an irradiation path 413 for irradiating a substrate and a reception path 414 (also referred to as a collection path) for collecting the irradiation from the substrate.
[0085] Figure 4 Illustrated is an environment of an evaluation system including the following:
[0086] a. A reference storage system 430 for storing a reference database 431. The reference database includes reference information 432, such as a first feature marker 432-1 of a reference pattern and a second feature marker 432-2 of the reference pattern, and (ii) an interested defect metadata 432-3 that defines what is equivalent to an interested defect.
[0087] b. An evaluated substrate storage system 433 for storing an evaluated substrate database 434. The evaluated substrate database includes evaluation information 434-1 about the substrate, such as one or more images 434-2 of one or more regions of the substrate or any other detection signals generated during scanning or other irradiation of the substrate, and feature markers 434-3 of an evaluated pattern.
[0088] c. A training storage system 435 for storing a training database 436. The training database stores training information 437. The training information includes a feature marker 438-1 of a substrate pattern of a training-related substrate, a first feature marker 438-2 of a reference pattern similar to the feature marker of the substrate pattern, an interested defect dataset 438-3 including a second feature marker 438-4 of the reference pattern, and another dataset 438-5 having an additional feature marker 438-6 of an additional pattern of the training-related substrate.
[0089] d. A machine learning process training unit 450 configured to train a machine learning process 460.
[0090] e. A processor 490 including one or more processing circuits 491. The processor 490 is configured to apply a machine learning process 460 to an image of a region of an evaluated substrate to find an interested defect associated with a given evaluated pattern in the evaluated pattern.
[0091] Figure 5 Illustrated is an example of a test-related substrate 510, one or more reference substrates 518, and an evaluated substrate 519.
[0092] Irradiate a test-related substrate 510 to generate a detection signal, and process the detection signal to provide characteristic markers of the substrate pattern of a training-related substrate 438-1. Some of the patterns are OK (see OK pattern 511), and some patterns are found to be associated with defects of interest (see defect-of-interest pattern 512).
[0093] Figure 5 Also illustrated are two versions of a second characteristic marker 532 (in this case a compact image) of a pattern including a defect of interest and first characteristic markers 533 and 534 of patterns not including a defect of interest. In one version, pixels covered by the defect of interest are set to a default value of zero gray level. In the second version, the pattern appears as if the defect of interest never existed.
[0094] According to an embodiment, the processing circuit is implemented as a central processing unit (CPU). According to an embodiment, the processing circuit is implemented as a graphics processing unit (GPU). According to an embodiment, the processing circuit is implemented as a hardware accelerator. According to an embodiment, the processing circuit includes one or more other integrated circuits, such as an application specific integrated circuit (ASIC) or a field programmable gate array (FPGA) or a fully custom integrated circuit.
[0095] In the foregoing detailed description, numerous specific details are set forth in order to provide a thorough understanding of embodiments of the present disclosure.
[0096] However, those skilled in the art will understand that embodiments of the present disclosure may be practiced without these specific details. In other instances, well-known methods, procedures, and components have not been described in detail so as not to obscure embodiments of the present disclosure.
[0097] The subject matter that is particularly pointed out and distinctly claimed as embodiments of the present disclosure is set forth in the concluding portion of the specification. However, when read in conjunction with the drawings, embodiments of the present disclosure as to the organization and method of operation, as well as the objects, features, and advantages of the embodiments, may be most thoroughly understood by reference to the following detailed description.
[0098] It should be understood that, for simplicity and clarity of illustration, the elements shown in the drawings are not necessarily drawn to scale. For example, the dimensions of some of the elements may be exaggerated relative to other elements for clarity. Additionally, where considered appropriate, reference numerals may be repeated in the drawings to indicate corresponding or analogous elements.
[0099] Since most of the illustrated embodiments of the present disclosure can be implemented using electronic components and circuits known to those skilled in the art, details will not be explained to a greater extent than considered necessary above in order to understand and recognize the basic concepts of the embodiments of the present disclosure and in order not to obscure or depart from the teachings of the embodiments of the present disclosure.
[0100] Any reference in this specification to a method shall be applied, with the necessary changes, to a system capable of performing the method and shall be applied, with the necessary changes, to a computer program product storing instructions which, when executed, cause the method to be performed.
[0101] Any reference in this specification to a system shall be applied, with the necessary changes, to a method executable by the system and shall be applied, with the necessary changes, to a computer program product storing instructions executable by the system.
[0102] Any reference in this specification to a computer program product shall be applied, with the necessary changes, to a method executable when executing the instructions stored in the computer program product and shall be applied, with the necessary changes, to a system configured to execute the instructions stored in the computer program product.
[0103] The term "and / or" means additionally or alternatively. For example, A and / or B means only A, or only B, or A and B.
[0104] In the foregoing description, numerous specific details are set forth in order to provide a thorough understanding of embodiments of the present disclosure.
[0105] However, those skilled in the art will understand that embodiments of the present disclosure may be practiced without these specific details. In other instances, well-known methods, procedures, and components have not been described in detail so as not to obscure embodiments of the present disclosure.
[0106] The subject matter that is particularly pointed out and distinctly claimed as an embodiment of the present disclosure is set forth in the concluding portion of the specification. However, when read in conjunction with the drawings, the embodiments of the present disclosure as to the organization of operations and method, as well as the objects, features, and advantages of the embodiments, may be most thoroughly understood by reference to the following detailed description.
[0107] It should be understood that, for simplicity and clarity of illustration, the elements shown in the drawings are not necessarily drawn to scale. For example, the dimensions of some of the elements may be exaggerated relative to other elements for clarity. Further, where considered appropriate, reference numerals may be repeated in the drawings to indicate corresponding or analogous elements.
[0108] Any reference in the specification to a support unit shall be applied, with the necessary changes, to a method executable by the support unit.
[0109] The term "and / or" means additionally or alternatively. For example, A and / or B means only A, or only B, or A and B.
[0110] In the foregoing specification, embodiments of the present disclosure have been described with reference to specific examples of the embodiments. However, it is apparent that various modifications and changes can be made to the embodiments without departing from the broad spirit and scope of the appended claims.
[0111] Furthermore, the terms "front", "rear", "top", "bottom", "upper", "lower", etc. in the specification and claims, if any, are used for descriptive purposes and not necessarily to describe a permanent relative position. It should be understood that such terms can be interchanged where appropriate so that, for example, the embodiments of the present disclosure described herein can operate in orientations other than those shown or described herein.
[0112] Any reference to the term "comprising" or "having" or "including" shall be applied, with the necessary changes, to "consisting of" and / or shall be applied, with the necessary changes, to "consisting essentially of".
[0113] However, other modifications, variations and substitutions are also possible. Accordingly, the specification and drawings are to be regarded as illustrative rather than restrictive.
[0114] In the claims, any reference signs placed between parentheses shall not be construed as limiting the claim. The word "comprising" does not exclude the presence of other elements or steps than those listed in a claim. Further, as used herein, the term "a" or "an" is defined as one or more than one. Moreover, even when the same claim includes a recitation phrase "one or more" or "at least one" and an indefinite article such as "a" or "an", the use of a recitation phrase such as "at least one" and "one or more" in a claim shall not be construed as implying that the introduction of another claim element by the indefinite article "a" or "an" will limit any particular claim containing such introduced claim element to embodiments containing only one such element. This also applies to the use of definite articles. Unless otherwise stated, terms such as "first" and "second" are used arbitrarily to distinguish elements so described. Accordingly, these terms are not necessarily intended to denote a temporal or other priority of such elements. The fact that particular measures are recited in mutually different claims does not indicate that a combination of these measures cannot be used advantageously.
[0115] Although specific features of the embodiments have been shown and described herein, many modifications, substitutions, changes and equivalents will now occur to those skilled in the art. Accordingly, it is to be understood that the appended claims are intended to cover all such modifications and changes that fall within the true spirit of the embodiments.
Claims
1. A method for training a machine learning process, the method comprising: Obtaining a feature marker of a substrate pattern of a training-related substrate; Finding a first feature marker of a reference pattern that is similar to the feature marker of the substrate pattern; Wherein the reference pattern is associated with a defect previously defined as a defect of interest; Wherein the finding is performed regardless of one or more parameters affecting the generation of the first reference feature marker; Populating a defect-of-interest dataset with a second feature marker of the reference pattern, the second feature marker of the reference pattern conveying more information about the defect previously defined as the defect of interest than the first feature marker of the reference pattern; Populating another dataset with additional feature markers of additional patterns of the training-related substrate; and Training the machine learning process in a supervised manner to find defects of interest, wherein the training includes feeding the defect-of-interest dataset and the other dataset into the machine learning process.
2. The method according to claim 1, wherein after the training, the training-related substrate is evaluated by the machine learning process.
3. The method according to claim 1, wherein after the training, one or more evaluated substrates different from the training-related substrate are evaluated by the machine learning process.
4. The method according to claim 1, wherein the one or more parameters include an image acquisition process used during the generation of the first reference feature marker.
5. The method according to claim 1, wherein the one or more parameters include a source of the substrate pattern.
6. The method according to claim 1, wherein the additional pattern of the training-related substrate is a defect-free reference pattern.
7. The method according to claim 1, wherein the reference pattern exhibits a defect having less than a defined influence on the similarity between the first feature marker and the feature marker of the substrate pattern, and wherein the second feature marker exhibits a defect having greater than the defined influence on the similarity between the second feature marker and the feature marker of the substrate pattern.
8. The method according to claim 1, wherein the reference pattern belongs to one or more other substrates different from the training-related substrate.
9. The method according to claim 1, wherein the additional feature markers are generated based on a random sampling of the training-related substrate.
10. The method according to claim 1, wherein an evaluation system is used to generate the feature marker of the substrate pattern, and wherein the method further comprises preventing the manufacturer of the evaluation system from accessing at least (a) the machine learning process, (b) the defect-of-interest dataset, and (c) the other dataset.
11. The method according to any one of claims 1 to 10, wherein the method further comprises: After the training, applying the machine learning process by an evaluation system to find defects of interest of one or more evaluated substrates; And Preventing the manufacturer of the evaluation system from accessing the results of the application of the machine learning process.
12. A non-transitory computer-readable medium for training a machine learning process, the non-transitory computer-readable medium storing instructions that, when executed by a processor comprising one or more processing circuits, cause the processor to: Obtain a feature marker of a substrate pattern of a training-related substrate; Locate a first feature marker of a reference pattern that is similar to the feature marker of the substrate pattern; Wherein the reference pattern is associated with a defect previously defined as a defect of interest; Wherein the locating is performed regardless of one or more parameters affecting the generation of the first reference feature marker; Populate a defect-of-interest dataset with a second feature marker of the reference pattern, the second feature marker of the reference pattern conveying more information about the defect previously defined as a defect of interest than the first feature marker of the reference pattern; Populate another dataset with additional feature markers of additional patterns of the substrate; and Train the machine learning process in a supervised manner to locate defects of interest, wherein the training includes feeding the defect-of-interest dataset and the another dataset into the machine learning process.
13. An evaluation system utilizing a machine learning process, the evaluation system comprising: An imager configured to obtain an image of a region of an evaluated substrate; The evaluated substrate includes an evaluated pattern; A processor comprising one or more processing circuits, the processor being configured to apply the machine learning process to the image of the region to locate a defect of interest associated with a given evaluated pattern in the evaluated pattern; Wherein the machine learning process is trained to locate defects of interest, wherein training the machine learning process includes feeding a defect-of-interest dataset and another dataset into the machine learning process; Wherein the defect-of-interest dataset is generated by: Obtaining a feature marker of a substrate pattern of a training-related substrate; Locating a first feature marker of a reference pattern that is similar to the feature marker of the substrate pattern; Wherein the reference pattern is associated with a defect previously defined as a defect of interest; Wherein the locating is performed regardless of one or more parameters affecting the generation of the first reference feature marker; And Populating the defect-of-interest dataset with a second feature marker of the reference pattern, the second feature marker of the reference pattern conveying more information about the defect previously defined as a defect of interest than the first feature marker of the reference pattern; Wherein the another dataset is populated with additional feature markers of additional patterns of the substrate.
14. The evaluation system according to claim 13, wherein the one or more parameters include an image acquisition process used during the generation of the first reference feature marker.
15. The evaluation system according to claim 13, wherein the one or more parameters include a source of the substrate pattern.
16. The evaluation system according to claim 13, wherein the additional pattern of the substrate is a defect-free reference pattern.
17. The evaluation system according to claim 13, wherein the reference pattern exhibits a defect having an influence less than a defined influence on the similarity between the first feature marker and the feature marker of the substrate pattern, and wherein the second feature marker exhibits a defect having an influence greater than the defined influence on the similarity between the second feature marker and the feature marker of the substrate pattern.
18. The evaluation system according to claim 13, wherein the reference pattern belongs to one or more other substrates different from the training-related substrate.
19. The evaluation system according to claim 13, wherein the additional feature marker is generated based on a random sampling of the training-related substrate.
20. The evaluation system according to any one of claims 13 to 19, wherein the processor is configured to prevent the manufacturer of the evaluation system from accessing the result of the application of the machine learning process.