Defect detection using machine learning algorithms

Abstract

Systems and methods are provided that include obtaining a first inspection image acquired by an inspection tool that provides information of a first region of a sample; at least the first inspection image is fed to a machine learning algorithm configured for each given pixel of the plurality of pixels of the first inspection image or for each given pixel group of the plurality of pixel groups of the first inspection image, determining one or more given parameters of a given model providing information of the pixel intensity distribution; for each given pixel or group of given pixels, at least some of the one or more given parameters, or a given model associated with the one or more given parameters, and the measured pixel intensity of the given pixel or group of pixels are used to determine whether there is a defect in the given pixel or group of given pixels.

Description

Technical Field

[0001] The presently disclosed subject matter relates generally to the field of inspecting semiconductor specimens, and more particularly to automating the inspection of the specimens. Background Art

[0002] The current demands for high density and performance associated with ultra-large scale integration of manufactured devices require sub-micron features, increased transistor and circuit speeds, and improved reliability. These demands require that device features be formed with high precision and uniformity, which in turn requires careful monitoring of the manufacturing process, including automated inspection of the device while it is still in semiconductor wafer form.

[0003] Inspection processes are used at various steps during semiconductor manufacturing to detect and classify defects on a sample (eg, automatic defect classification (ADC), automatic defect review (ADR), etc.). Summary of the invention

[0004] According to certain aspects of the presently disclosed subject matter, a system is provided, the system comprising one or more processing circuit systems configured to: obtain a first inspection image acquired by an inspection tool and providing information of a first area of ​​a semiconductor sample, feed at least the first inspection image to a machine learning algorithm, the machine learning algorithm configured to determine, for each given pixel among a plurality of pixels of the first inspection image or for each given pixel group among a plurality of pixel groups of the first inspection image, one or more given parameters of a given model providing information of pixel intensity distribution, and for each given pixel or each given pixel group, determine, using at least one of the one or more given parameters, or at least a portion of the given model associated with the one or more given parameters, and the measured pixel intensity of the given pixel or the given pixel group, whether there is a defect in the given pixel or in the given pixel group.

[0005] According to some embodiments, the system is configured to determine, for each given pixel or each given group of pixels, the probability of a defect existing in the given pixel or in the given group of pixels using at least some of the one or more given parameters, or at least a portion of the given model associated with the one or more given parameters.

[0006] According to some embodiments, for each given pixel or for each given group of pixels, at least a portion of the given model associated with the one or more given parameters provides information of a pixel intensity probability distribution, which can be used to determine the probability that the measured pixel intensity of the given pixel or the given group of pixels corresponds to a defect.

[0007] According to some embodiments, for each given pixel or each given group of pixels, at least some of the one or more given parameters, or at least a portion of the given model associated with the one or more given parameters, can be used to determine the expected pixel intensity in the absence of defects in the given pixel or in the given group of pixels.

[0008] According to some embodiments, for each given pixel or each given group of pixels, at least some of the one or more given parameters, or at least a portion of the given model associated with the one or more given parameters, can be used to determine an expected pixel intensity in the absence of a defect in the given pixel or in the given group of pixels, and the probability that a deviation from the expected pixel intensity corresponds to a defect.

[0009] According to some embodiments, for each given pixel or each given pixel group, at least some of the one or more given parameters, or at least a portion of the given model associated with the one or more given parameters, provides information on the pixel intensity in the absence of defects in the given pixel or in the given pixel group, and the noise in the given pixel or in the given pixel group.

[0010] According to some embodiments, the system is configured for using the given model to detect, on average in different partitions having different noise levels, a maximum number of defects below a same threshold for the different partitions.

[0011] According to some embodiments, the system is configured to, for each given pixel or each given group of pixels, distinguish between the presence of defects and the presence of noise in the given pixel or in the given group of pixels using at least some of the one or more given parameters, or at least a portion of the given model associated with the one or more given parameters.

[0012] According to some embodiments, the one or more given parameters are determined by the machine learning algorithm specifically for each given pixel or each given pixel group, and include: data providing information on the expected pixel intensity in the given pixel or in the given pixel group when no defects are present in the given pixel or in the given pixel group, and at least one of the following: data providing information on the noise present in the given pixel or in the given pixel group, data achieving normalization of the noise present in the given pixel or in the given pixel group, or data achieving differentiation between defects and noise in the given pixel or in the given pixel group.

[0013] According to some embodiments, the system is configured for feeding, in addition to the first inspection image, one or more reference images providing information of one or more other regions of the semiconductor sample or another semiconductor sample to the machine learning algorithm.

[0014] According to some embodiments, for each given pixel or each given group of pixels, the system is configured to use the output of the machine learning algorithm to determine at least one of the following: data providing information about the noise present in the given pixel or in the given group of pixels, data enabling normalization of the noise present in the given pixel or in the given group of pixels, or data enabling differentiation between defects and noise in the given pixel or in the given group of pixels.

[0015] According to some embodiments, the machine learning algorithm has been trained using at least one training image and a loss function, wherein for each given pixel among multiple pixels of the training image or for each pixel group among multiple pixel groups of the training image, the loss function includes one or more parameters of a model that models the pixel intensity associated with the given pixel or with the given pixel group.

[0016] According to some embodiments, the machine learning algorithm is configured to simultaneously determine, for each given pixel among multiple pixels of the first inspection image or for each given pixel group among multiple pixel groups of the first inspection image: data providing information on the expected pixel intensity in the given pixel or in the given pixel group if no defects are present in the given pixel or in the given pixel group, and data providing information on the noise present in the given pixel or in the given pixel group.

[0017] According to some embodiments, the machine learning algorithm is configured to determine, for each given pixel among multiple pixels of the first inspection image or for each given pixel group among multiple pixel groups of the first inspection image: data providing information on the expected pixel intensity in the given pixel or in the given pixel group if no defects are present in the given pixel or in the given pixel group, and data providing information on the noise present in the given pixel or in the given pixel group.

[0018] According to some embodiments, the machine learning algorithm is configured to generate, for each given pixel or each given pixel group, data providing information on the expected pixel intensity in the given pixel or in the given pixel group if no defect is present in the given pixel or in the given pixel group; and data providing confidence information associated with the data providing information on the expected pixel intensity in the given pixel or in the given pixel group; wherein the system is configured to determine whether there is a defect in the given pixel or in the given pixel group using the measured pixel intensity of the given pixel or the given pixel group and the data providing the confidence information.

[0019] According to some embodiments, the system is configured to: obtain a single inspection image acquired by the inspection tool and providing information of an area of ​​a semiconductor sample, determine one or more defects in the area based on the single inspection image, and the determination includes: feeding the single inspection image to the machine learning algorithm, the machine learning algorithm being configured to determine, for each given pixel among multiple pixels of the single inspection image or for each given pixel group among multiple pixel groups of the single inspection image, one or more given parameters of a given model providing information of pixel intensity distribution, and for each given pixel or each given pixel group, using at least one of the one or more given parameters, or at least a portion of the given model associated with the one or more given parameters, and the measured pixel intensity of the given pixel or the given pixel group to determine whether there is a defect in the given pixel or in the given pixel group.

[0020] According to some embodiments, one or more training images used to train the machine learning algorithm have at least one of a smaller height or a smaller width than the first verification image.

[0021] According to some embodiments, the system is configured to generate a new image using at least some of the one or more given parameters, or at least a portion of the given model associated with the one or more given parameters, wherein at least one of (i) or (ii) is satisfied: (i) the new image is noise-free or contains less noise than the first inspection image; (ii) the new image is defect-free or contains fewer defects than the first inspection image.

[0022] According to certain aspects of the presently disclosed subject matter, a method is provided that includes one or more processing circuitry performing one or more of the features described with respect to the system (thus not repeating such features).

[0023] According to other aspects of the presently disclosed subject matter, there is provided a non-transitory computer-readable medium comprising instructions that, when executed by one or more processing circuitry, cause the one or more processing circuitry to perform the operations described with reference to the above method.

[0024] According to other aspects of the presently disclosed subject matter, a system is provided, comprising one or more processing circuit systems, the one or more processing circuit systems being configured to: obtain a first training image acquired by an inspection tool and providing information of a first region of a semiconductor sample, feed at least the first training image to the machine learning algorithm to train the machine learning algorithm to: determine, for each given pixel among a plurality of pixels of the first training image or for each given pixel group among a plurality of pixel groups of the first training image, one or more given parameters of a given model providing information of pixel intensity distribution, wherein the one or more given parameters, or the given model associated with the one or more given parameters, can be used to detect the presence of defects in the given pixel or in the given pixel group.

[0025] According to some embodiments, for each given pixel or each given group of pixels, the one or more given parameters, or the given model associated with the one or more given parameters, can be used to determine the pixel intensity in the absence of defects in the given pixel or in the given group of pixels.

[0026] According to some embodiments, the one or more training images used to train the machine learning algorithm correspond to one or more inspection images of a semiconductor sample to which one or more artificial defects have been added.

[0027] According to certain aspects of the presently disclosed subject matter, a method is provided that includes one or more processing circuitry performing one or more of the features described with respect to the system (thus not repeating such features).

[0028] According to other aspects of the presently disclosed subject matter, there is provided a non-transitory computer-readable medium comprising instructions that, when executed by one or more processing circuitry, cause the one or more processing circuitry to perform the operations described with reference to the above method.

[0029] According to certain aspects of the presently disclosed subject matter, a non-transitory computer-readable medium is provided that includes instructions that, when executed by one or more processing circuit systems, cause the one or more processing circuit systems to perform: obtaining a first inspection image acquired by an inspection tool that provides information about a first area of ​​a semiconductor sample, feeding at least the first inspection image to a machine learning algorithm, the machine learning algorithm being configured to determine, for each given pixel among a plurality of pixels of the first inspection image or for each given pixel group among a plurality of pixel groups of the first inspection image, one or more given parameters of a given model that provides information about pixel intensity distribution, for each given pixel or each given pixel group, determining, using at least one of the one or more given parameters or at least a portion of the given model associated with the one or more given parameters, a new given pixel intensity value, thereby obtaining a set of new given pixel intensity values, and generating a new image using the set of new given pixel intensity values.

[0030] According to some embodiments, each new given pixel intensity value corresponds to a given expected pixel intensity in the absence of defects in the given pixel or in the given pixel group, and the set of new given pixel intensity values ​​corresponds to a set of given expected pixel intensity values, wherein at least one of (i) or (ii) is satisfied: (i) the new image is noise-free or contains less noise than the first inspection image; (ii) the new image is defect-free or contains fewer defects than the first inspection image.

[0031] According to some embodiments, the above features may be equivalently implemented by methods and / or systems (thus, these features are not repeated).

[0032] The proposed solution provides various technical advantages. At least some of these technical advantages are listed below.

[0033] According to some examples, the proposed solution enables efficient and accurate detection of defects in inspection images of semiconductor samples.

[0034] According to some examples, the proposed solution is operable to distinguish defects from noise.

[0035] According to some examples, the proposed solution is a computationally efficient method for detecting defects.

[0036] According to some examples, the proposed solution enables pin-pointed defect detection. In particular, the proposed solution enables prediction of whether a defect is present in each pixel of an inspection image of a semiconductor sample.

[0037] According to some examples, the proposed solution directly estimates the pixel intensity of each pixel if each pixel has no defects.

[0038] According to some examples, the proposed solution is operable to estimate the noise level present in each pixel and use this estimate to distinguish defects from noise, thereby increasing robustness.

[0039] According to some examples, the proposed solution is automatic and requires no user interaction.

[0040] According to some examples, the proposed solution is operable to determine pixel intensities in an image of a sample without defects and / or without noise, without requiring a priori knowledge about the sample or operator input.

[0041] According to some examples, the proposed solution does not require prior knowledge about the sample to detect defects and may rely only on an image of the sample.

[0042] According to some examples, the proposed solution enables detection of defects in an image of a region of a sample without requiring additional reference images of other regions of the sample (or of another sample). This enables increased throughput and enables operation in scenarios where no reference images are available.

[0043] According to some examples, the proposed solution uses a machine learning algorithm to detect defects in inspection images, where the dimensions of the training images used to train the machine learning algorithm may be smaller than the dimensions of the inspection images.

[0044] According to some examples, the proposed solution is operable to generate a defect-free image based on an inspection image of a semiconductor sample.

[0045] According to some examples, the proposed solution is operable to generate a denoised image based on an inspection image of a semiconductor sample. BRIEF DESCRIPTION OF THE DRAWINGS

[0046] In order to understand the present disclosure and to see how it may be implemented in practice, various embodiments will now be described, by way of non-limiting examples only, with reference to the accompanying drawings, in which:

[0047] Figure 1 A generalized block diagram of an inspection system in accordance with certain embodiments of the presently disclosed subject matter is shown.

[0048] Figure 2 A generalized flow chart of a method of detecting defects in each pixel of an inspection image is shown.

[0049] Figure 3Non-limiting examples of inspection images providing information of an area of ​​a sample, such as a die, are shown.

[0050] Figure 4 A non-limiting example of a probabilistic model for modeling pixel intensity in pixels of an inspection image is shown, where one or more parameters of the model provide information about the distribution of pixel intensities in the absence of a defect in the pixel.

[0051] FIG5 shows a generalized flow chart of another method of detecting defects in each pixel of an inspection image.

[0052] Figure 6 A non-limiting example of a probabilistic model for modeling pixel intensity in pixels of an inspection image is shown, where parameters of the model provide information of expected pixel intensity in the absence of defects in the pixel and in the absence of noise in the pixel.

[0053] Figure 7 Another non-limiting example of a probabilistic model for modeling pixel intensity in pixels of an inspection image is shown, where parameters of the model provide information of expected pixel intensity in the absence of defects in the pixel and in the absence of noise in the pixel.

[0054] Figure 8 A generalized flow chart of another method of detecting defects in each pixel of an inspection image is shown.

[0055] Fig. 9 Non-limiting examples of multiple inspection images providing information of different areas of a sample, such as different dies, are shown.

[0056] Fig.10 A generalized flow chart of a method of training a machine learning algorithm to detect defects in every pixel of an inspection image is shown.

[0057] Fig.11 A non-limiting example of multiple training images providing information of different regions of a sample, such as different dies, is shown.

[0058] Fig.12 Shows Fig.10 Non-limiting example of a method.

[0059] Fig.13 A generalized flow chart of a method of generating training images is shown.

[0060] Fig.14 Shows Fig.13 Non-limiting example of a method.

[0061] Fig.15 A generalized flow chart of a method of generating training images is shown.

[0062] Fig.16 A comparison is shown between the training images used in the training of the machine learning algorithm and the verification images fed to the machine learning algorithm during runtime checking.

[0063] Fig.17 A generalized flow chart of a method of generating a new image with less noise (denoised image) and / or with fewer defects is shown.

[0064] Fig.18 Shows Fig.17 An example of a method.

[0065] Fig.19 A generalized flow chart of a method of generating a new image is shown.

[0066] Fig. 20 Shows Fig.19 An example of a method. DETAILED DESCRIPTION

[0067] Runtime inspection can employ a two-stage process, for example, inspecting a sample and then reviewing sample locations for potential defects. In the first stage, a defect map is generated to show suspected locations on the sample that have a high probability of being defects. During the second stage, at least some of the suspected locations are more thoroughly analyzed at a relatively high resolution.

[0068] In at least some conventional prior art methods, during a first stage, images of a plurality of dies are acquired. To detect whether a defect exists in a current image of a given die of a sample, the current image is compared with images of other dies of the sample. The differences between images of different dies are used to detect potential defects. However, this approach is not always accurate. New methods and systems for detecting defects are presented below. Additional applications of these new methods and systems are also described below.

[0069] Attention shifts Figure 1 , Figure 1 A functional block diagram of an inspection system 100 according to some examples of the presently disclosed subject matter is shown. Figure 1The inspection system 100 shown can be used for inspection of samples (e.g., wafers and / or portions thereof) as part of a sample manufacturing process. The inspection system 100 shown includes a computer-based system 103 that is capable of automatically determining defect-related information using images obtained during sample manufacturing. The system 103 can be operably connected to one or more low-resolution inspection tools 101 and / or one or more high-resolution inspection tools 102 and / or other inspection tools. The inspection tool is configured to capture images and / or review the captured (multiple) images and / or enable or provide measurements related to the captured (multiple) images. The system 103 can be further operably connected to a CAD server 110 and a data repository 109.

[0070] System 103 includes processing circuitry 104, which includes one or more processors and one or more memories. Processing circuitry 104 is configured to provide all processing required to operate system 103, as described in further detail below (see Figure 2 Figure 5 Figure 8 , Fig.10 , Fig.13 , Fig.15 , Fig.17 and Fig.19 The method described in , which can be at least partially performed by system 103 and / or system 100).

[0071] Processing circuit system 104 is configured to execute one or more functional modules according to computer-readable instructions implemented on a non-transitory computer-readable memory of processing circuit system 104 (or operatively coupled to processing circuit system 104). The one or more functional modules include at least one machine learning algorithm 112 (also referred to as a machine learning model), such as a deep neural network (DNN).

[0072] As a non-limiting example, the layers of the machine learning algorithm 112 (eg, DNN) may be organized according to a convolutional neural network (CNN) architecture, such as a fully convolutional neural network (CNN). This is not limiting.

[0073] In other examples, the layers of the machine learning algorithm 112 (e.g., DNN) can be organized according to a recurrent neural network architecture, a recursive neural network architecture, a generative adversarial network (GAN) architecture, etc. Optionally, at least some of the layers can be organized into multiple DNN sub-networks. Each layer of the DNN can include multiple basic computing elements (CEs), which are generally referred to in the art as dimensions, neurons, or nodes.

[0074] Typically, the computational elements of a given layer can be connected to the CE of the previous layer and / or the next layer. Each connection between the CE of the previous layer and the CE of the next layer is associated with a weight value. A given CE can receive input from the CE of the previous layer via a corresponding connection, and each given connection is associated with a weight value that can be applied to the input of the given connection. The weight value can determine the relative strength of the connection, thereby determining the relative influence of the corresponding input on the output of the given CE. A given CE can be configured to calculate an activation value (e.g., a weighted sum of the inputs), and further derive an output by applying an activation function to the calculated activation. The activation function can be, for example, an identity function, a deterministic function (e.g., a linear function, a sigmoid function, a threshold function, etc.), a random function, or other suitable function. The output from a given CE can be transmitted to the CE of the next layer via a corresponding connection. Similarly, as described above, each connection at the CE output can be associated with a weighted value, which can be applied to the output of the CE before it is received as an input to the CE of the next layer. In addition to the weight value, there can also be a threshold (including a limit function) associated with the connection and the CE.

[0075] The weights and / or thresholds of the machine learning algorithm 112 (e.g., DNN) can be initially selected before training, and can be further iteratively adjusted or modified during training to achieve an optimal set of weights and / or thresholds in the trained DNN. After each iteration, the difference (also referred to as a loss function) between the actual output produced by the machine learning algorithm 112 (e.g., DNN) and the target output associated with the corresponding training data set can be determined. The difference can be referred to as an error value. Training can be determined to be complete when the cost or loss function indicating the error value is less than a predetermined value, or when a limited change in performance between iterations is achieved. Optionally, at least some of the DNN subnetworks (if any) can be trained separately before training the entire DNN.

[0076] The system 103 is configured to receive input data. The input data may include data generated by an inspection tool (and / or derivatives thereof and / or metadata associated therewith) and / or data generated and / or stored in one or more data repositories 109 and / or CAD servers 110 and / or another related data repository. Note that the input data may include images (e.g., captured images, images derived from captured images, simulated images, synthesized images, etc.) and associated digital data (e.g., metadata, handcrafted attributes, etc.). It should be further noted that the image data may include data related to a layer of interest and / or one or more other layers of the sample.

[0077] As a non-limiting example, the sample may be inspected by one or more low-resolution inspection machines 101 (e.g., an optical inspection system, a low-resolution SEM, etc.). The resulting data (low-resolution image data 121) providing information of a low-resolution image of the sample may be transmitted to the system 103 - directly or via one or more intermediate systems. Alternatively or additionally, the sample may be inspected by a high-resolution machine 102, such as a scanning electron microscope (SEM), an atomic force microscope (AFM), or an optical inspection tool (such as, but not limited to, Applicant's Enlight optical inspection system). The resulting data (high-resolution image data 122) providing information of a high-resolution image of the sample may be transmitted to the system 103 - directly or via one or more intermediate systems.

[0078] Note that image data may be received and processed along with metadata associated therewith (eg, pixel size, text description of defect type, parameters of the image capture process, etc.).

[0079] While processing input data (e.g., low-resolution image data and / or high-resolution image data, along with other data, such as design data, composite data, etc.), system 103 can send instructions 123 and / or 124 to any of (multiple) inspection tools, store results (such as data providing information about defect locations) in storage system 107, present results via computer-based graphical user interface GUI 108, and / or send results to an external system.

[0080] Those skilled in the art will readily appreciate that the teachings of the presently disclosed subject matter are not limited to Figure 1 Constraints of the systems shown; equivalent and / or modified functionality may be combined or divided in another manner and implemented in any suitable combination of software and firmware and / or hardware.

[0081] It should also be noted that, without limiting the scope of the present disclosure in any way, the inspection tool can be implemented as various types of inspection machines, such as optical imaging machines, electron beam inspection machines, etc. In some cases, the same inspection tool can provide low-resolution image data and high-resolution image data. In some cases, at least one inspection tool can have metrology capabilities.

[0082] It should be noted that Figure 1 The inspection system shown can be implemented in a distributed computing environment, where Figure 1The aforementioned functional modules shown may be distributed across several local and / or remote devices and may be linked via a communication network. Note further that in other embodiments, at least some of the inspection tools 101 and / or 102, the data repository 109, the storage system 107 and / or the GUI 108 and / or the CAD server 110 may be external to the inspection system 100 and operate in data communication with the system 103. The system 103 may be implemented as a (multiple) stand-alone computer used in conjunction with the inspection tool. Alternatively, the corresponding functions of the system may be at least partially integrated with one or more inspection tools.

[0083] Attention now turns to Figure 2 , which describes a method capable of detecting defects in images of semiconductor samples.

[0084] Figure 2 The method includes (operation 200) obtaining a first inspection image 300 providing information of a first region 310 (or regions) of a semiconductor sample 320. The first inspection image 300 is acquired by an inspection tool, such as inspection tool 101 and / or inspection tool 102. Note that the first inspection image 300 may have been acquired by an optical inspection tool, or by an electron beam inspection tool, or by another adapted inspection tool.

[0085] The first inspection image 300 includes a plurality of pixels. Each pixel is associated with a measured pixel intensity. The measured pixel intensity can be represented as, for example, a grayscale intensity, or any other suitable convention is used. The measured pixel intensity can be provided by the inspection tool 101 and / or 102.

[0086] Figure 2 The method further includes feeding (operation 210) at least the first inspection image 300 to the trained machine learning algorithm 112. The method for training the machine learning algorithm 112 is described below.

[0087] The machine learning algorithm 112 determines (operation 220) data that can be used to determine an expected pixel intensity in the given pixel or in the given pixel group if no defects are present in the given pixel or in the given pixel group for each given pixel in the plurality of pixels of the first inspection image 300 (such as pixels providing information of the first region 310) or for each pixel group in the plurality of pixel groups of the first inspection image 300. The pixel group may correspond to several adjacent pixels in the first region 310.

[0088] The machine learning algorithm 112 determines (at operation 220) one or more given parameters of a given model (also referred to as a probability distribution, a statistical model, a probability model, or a probability function) that provides information about a pixel intensity distribution for each given pixel in a plurality of pixels of the first inspection image 300 (such as a pixel providing information about the first region 310) or for each pixel group in a plurality of pixel groups of the first inspection image 300. The pixel intensity distribution may correspond to a function that models the distribution of pixel intensities. For example, the function may link (multiple) pixel intensity values ​​to the number of non-defective pixels associated with these (multiple) pixel intensity values. Equivalently, the function may link (multiple) pixel intensity values ​​to the probability of a non-defective pixel associated with the pixel intensity. This is not limiting. The probability distribution may be used to determine the probability that a measured pixel intensity for a given pixel or a given pixel group corresponds to a defect.

[0089] In particular, at least a portion of a given model (that is, at least a portion of a probability distribution of a given model) can model a pixel intensity distribution (also referred to as a pixel probability distribution or a pixel statistical distribution) in the absence of a defect in a given pixel or in a given group of pixels. The pixel intensity distribution can model the expected pixel intensity of each pixel in the absence of a defect, and the probability that a deviation from the expected pixel intensity corresponds to a defect.

[0090] Since a given model (obtained for a given pixel or a group of pixels) can provide information on the expected pixel intensity in a given pixel or in a given group of pixels in the absence of defects in the given pixel in the group of pixels, the given model can be used to determine the expected pixel intensity in the absence of defects in the given pixel (or in a given group of pixels).

[0091] In some examples, the method enables prediction of the (estimated) pixel intensity that would be present in the pixel if the pixel was not defective for each pixel of the first inspection image 300. In some examples, one or more of the parameters output by the machine learning model 112 for each pixel directly corresponds to the (estimated) pixel intensity that would be present in the pixel if the pixel was not defective. In other examples, one or more of the parameters output by the machine learning model 112 for each pixel can be processed for estimating the pixel intensity that would be present in the pixel if the pixel was not defective.

[0092] The type of parameters estimated by the machine learning algorithm 112 may be the same for multiple pixels (e.g., model mean, model variance, etc.). However, the values ​​of the parameters estimated by the machine learning algorithm 112 may be different for each pixel. Note that the machine learning algorithm 112 is capable of explicitly estimating and outputting the values ​​of the parameters of the model that models the pixel intensity in each given pixel. The type and number of parameters are defined during the training phase of the machine learning algorithm 112. Similarly, a model may be defined during the training phase for which the parameters are output by the machine learning algorithm 112. In particular, these parameters may be present in the loss function used to train the machine learning algorithm 112. Thus, during the prediction phase, the machine learning algorithm 112 is capable of predicting the values ​​of these parameters for each pixel (or group of pixels).

[0093] In some examples, the expected pixel intensity (in the absence of defects) can correspond to the mean (or average) of the probability distribution of the model.

[0094] In some examples, the model can correspond to a Gaussian function (a normally distributed random variable). The Gaussian function is only an example of a probability model that can be used, or various other probability models (which can be used, for example, to model the distribution of values) can be used. Another non-limiting example of a probability model that can be used is a gamma distribution.

[0095] For each pixel, the machine learning algorithm 112 estimates the values ​​of the parameter(s) of the model. In some examples where the model is a Gaussian function, the parameters may include a mean value (also called an expected value) labeled "μ" (which corresponds to the expected pixel intensity in the absence of defects) and a standard deviation labeled "σ" (which is an estimate of the noise present in the pixel). As described above, different values ​​(μ and / or σ) may be obtained for each pixel.

[0096] In some examples where the model is a gamma distribution, the parameters may include parameters α (called a shape parameter) and β (called an inverse scale parameter), or parameters k (called a shape parameter) and θ (also called a scale parameter). The expected pixel intensity (in the absence of defects) may correspond to the mean kθ or α / β.

[0097] Note that the probability model can be a symmetric model or an asymmetric model.

[0098] Note that even if two pixels have the same measured grayscale intensity, this method can determine different values ​​for estimating the pixel intensity if each pixel were not defective.

[0099] Figure 2The method further includes: (operation 230) for each given pixel or each given pixel group among a plurality of pixels, using at least some of the one or more given parameters, or at least a portion of a given model (defined by the one or more given parameters), determining whether there is a defect in the given pixel or in the given pixel group.

[0100] Operation 230 enables a determination, for each pixel (or group of pixels), whether a defect is present in the pixel (or group of pixels). Operation 230 may generate a prospect (also referred to as a probability or score) of the presence of a defect for each pixel of the first inspection image 300. If the prospect (probability or score) is above or equal to a threshold, a defect is detected. If the prospect (probability or score) is below a threshold, it is concluded that the pixel does not include a defect. The detected defects (along with their locations) may be output on a display. Note that this determination corresponds to the determination of candidate defects. As explained above, each candidate defect may be further reviewed by a high-resolution inspection tool (e.g., a high-resolution SEM) to confirm whether the candidate defect is an actual defect.

[0101] In some examples, operation 230 may include determining (for each given pixel or pixel group) an expected pixel intensity in the given pixel (or pixel group) using given parameters and / or a given model (defined by the given parameters) if the given pixel (or pixel group) is not defective. It may further include comparing, for each given pixel (or given pixel group), the measured pixel intensity of the given pixel (or given pixel group) with the expected pixel intensity (in the absence of a defect) generated by the machine learning algorithm 112. If the comparison indicates that the difference between the measured pixel intensity and the expected pixel intensity is above or equal to a threshold, this may be considered a defect. If the comparison indicates that the difference between the measured pixel intensity and the expected pixel intensity is below a threshold, this may be considered to be non-defective. This is not limiting, and other methods may be used to determine whether a defect is present, as explained below.

[0102] Figure 2 The method may be performed during a runtime scan of a sample by an inspection tool. Each time a new image is provided by the inspection tool (e.g., for each new die), the image may be obtained based on Figure 2 The new image is processed by this method.

[0103] In some examples, in addition to the first inspection image providing information of the first region, the machine learning algorithm 112 may be fed with one or more additional inspection images providing information of additional region(s) of the same sample or another sample. The additional region(s) may match similarity criteria with the first region. For example, they may correspond to similar dies. This will be referred to below Figure 8 Further discussion.

[0104] Figure 2 The method can also be performed with a single image. In other words, feeding a single inspection image of a region to the machine learning algorithm 112 is sufficient to detect defects in that region without the need to feed additional reference images of other regions to the machine learning algorithm 112. This constitutes a technical advantage because, in at least some cases, no additional reference images are available. For example, some samples are made of a single die (such as, but not limited to, a graphics processing unit).

[0105] Figure 4 A non-limiting example of a model 400 that models the pixel intensity distribution for a given pixel is shown. At least a portion of the model models the pixel intensity distribution in the absence of defects. As described above, the values ​​of the parameters of the model may vary from pixel to pixel. The abscissa 440 represents probability and the ordinate 450 represents pixel intensity.

[0106] exist Figure 4 In the example of , model 400 is a Gaussian model. The machine learning algorithm 112 can be trained to determine the mean 410 ("μ") of the model and the standard deviation 420 (labeled "σ") of the model. The mean 410 ("μ") corresponds to the expected pixel intensity of the given pixel when there is no defect in the pixel. In other words, the machine learning algorithm 112 generates a prediction of the pixel intensity of the given pixel if there is no defect.

[0107] exist Figure 4 In the example of FIG. 4 , the measured pixel intensity of a given pixel (for which the values ​​of the parameters of the model 400 have been estimated by the machine learning algorithm 112) is denoted by reference numeral 430. Figure 4 As can be seen in FIG. 4 , the measured pixel intensity differs from the expected pixel intensity 410 (when no defect is present). The deviation between the measured pixel intensity and the model's mean 410 can be used to determine the probability that the measured pixel intensity corresponds to a defect. As will be explained below, in some examples, a given model can be used to distinguish deviations due to defects from deviations due to noise.

[0108] Attention now turns to Figure 5A .

[0109] Figure 5A The methods include those already referenced Figure 2 Operations 200, 210, and 220 are described.

[0110] Figure 5AThe method further includes: (operation 505) using at least some of the one or more given parameters, or at least a portion of a given model (associated with the one or more given parameters), and the measured pixel intensity of a given pixel or a given pixel group to distinguish the presence of a defect in a given pixel or a given pixel group from the presence of noise.

[0111] In practice, it may happen that, although there is a difference between the measured pixel intensity and the expected pixel intensity (in the absence of a defect), this does not correspond to a defect, but rather to the noise present in that pixel. The noise may be caused by various factors, such as (but not limited to) noise caused by the inspection tool, noise caused by the manufacturing process (but not constituting a defect to be identified), etc.

[0112] Figure 5A The method enables the same (maximum) number of defects to be detected between different regions, even if these different regions are associated with different noise levels. In particular, the same (maximum) number of defects can be detected between regions with a high noise level and regions with a low noise level.

[0113] On average, for a given surface, a certain number of defects is expected (hereinafter referred to as the expected number of defects). Note that it may happen that the sample contains partitions with an abnormal number of defects, which is higher than the expected number. However, the method enables the detection of the same (maximum) number of expected defects on average between various different areas (some areas are associated with high noise levels and some areas are associated with low noise levels). This enables a constant false alarm rate to be obtained.

[0114] In some examples, Figure 5A The method includes using a given model (obtained for each given pixel or given group of pixels and defined by one or more given parameters output by the machine learning model) and the measured pixel intensity of a given pixel (or given group of pixels) to determine a score (probability) that the given pixel (or given group of pixels) corresponds to a defect.

[0115] In some examples, the following (non-limiting) scores may be used: - In the case where the measured pixel intensity c is less than the median or mean of the distribution pdf(x), the score - In the case where the measured pixel intensity c is less than the median or mean of the distribution pdf(x), the score

[0116] In this equation, pdf(x) is a probability distribution (e.g., Gaussian, gamma, or other probability distribution, etc.) that models the pixel intensity. The integral of the distribution is performed over an interval that depends on the measured pixel intensity for a given pixel. The probability distribution models the expected pixel intensity (in the absence of defects and noise), and the probability that a deviation from the expected pixel intensity corresponds to a defect (rather than noise). By calculating a score using the integral of the probability distribution, the probability that the pixel intensity corresponds to a defect rather than noise can be determined. In other words, a distinction between defects and noise is performed.

[0117] If the score is higher than or equal to the threshold, the defect is detected. If the score is lower than the threshold, the defect is not detected.

[0118] In some examples, data indicating the noise level for each pixel may be explicitly calculated. This will refer to Figure 5B Describe, Figure 5B Describes Figure 5A Another non-limiting example of an implementation of the method. Figure 5B The method may include determining a confidence level for a prediction of an expected pixel intensity (in the absence of a defect), and using the confidence level to determine whether a defect is present.

[0119] The data that provides information about the confidence of the prediction of the expected pixel intensity (in the absence of defects) is an estimate of the noise present in the pixel. If there is a small confidence (corresponding to a large standard deviation σ in the case of a Gaussian model), this indicates that there is a high noise level in the pixel. In other words, the pixel intensity is expected to likely experience large variations around the expected pixel intensity (in the absence of defects), but these variations do not indicate a defect. If there is a high confidence (corresponding to a small standard deviation σ in the case of a Gaussian model), this indicates that there is a low noise level in the pixel. In other words, the pixel intensity is not expected to experience significant variations around the expected pixel intensity (in the absence of defects), and the presence of significant variations is likely to indicate a defect. In view of the foregoing, Figure 5B The method can perform normalization of the noise so that real defects can be distinguished from noise in each pixel.

[0120] Figure 5B The method includes obtaining (operation 500 ) a first inspection image acquired by an inspection tool, the first inspection image providing information of a first region of a semiconductor sample. Operation 500 is similar to operation 200 .

[0121] Figure 5B The method further includes feeding (operation 510 ) at least the first inspection image 300 to the trained machine learning algorithm 112 . Operation 510 is similar to operation 210 .

[0122] The machine learning algorithm 112 generates (operation 520) one or more given parameters of a given model for each given pixel of the plurality of pixels of the first inspection image 300 or for each given pixel group of the plurality of pixel groups of the first inspection image 300, wherein at least a portion of the given model provides information of the distribution of pixel intensities in the absence of defects. At least some of the given parameters (multiple) and / or the given model (defined by the given parameters (multiple)) can be used to determine, for each given pixel, an expected non-defective pixel intensity of the given pixel. As described above, the non-defective pixel intensity can correspond to a mean (or average) of the given model. In addition, the given parameters (multiple) can correspond to data that provides information of confidence associated with the expected non-defective pixel intensity generated based on the output of the machine learning algorithm. The data also provides information of noise present in each given pixel or each given pixel group (which does not constitute a defect to be detected).

[0123] In some examples, the method enables the generation of data providing information about the expected pixel intensity for each pixel (in the absence of defects), and data providing information about the confidence of the prediction for each pixel (data providing information about the noise in each pixel) based on the output of the same (single) machine learning algorithm 112.

[0124] A graph may be output which includes, for each pixel of the first inspection image 300, the (estimated) pixel intensity that would exist in the pixel if the pixel were free of defects, and data providing confidence information associated with data providing information about the expected pixel intensity (the data also providing information about the noise present in the pixel).

[0125] The machine learning algorithm 112 can specifically estimate the noise in each pixel, or provide parameters specific to each pixel, so that information about the noise present in each pixel can be obtained. The estimate of the noise can vary from pixel to pixel.

[0126] As explained above, the data providing information about the confidence level indicates to what extent the expected pixel intensity (corresponding to a configuration without defects) is reliable. If the confidence level is high, this indicates that small deviations of the measured pixel intensity from the expected pixel intensity have a high probability of corresponding to a defect. This is in the Figure 6 It is shown in Figure 6 , the standard deviation 620 (which indicates the confidence level in the expected pixel intensity and, therefore, the noise level) is smaller. Figure 6 As shown, small deviations of measured pixel intensity 630 from expected pixel intensity 610 (relative to standard deviation 620) are likely to indicate a defect.

[0127] If the confidence level is high, this indicates that small deviations of the measured pixel intensity from the expected pixel intensity have a low probability of corresponding to a defect. Only certain deviations of the measured pixel intensity from the expected pixel intensity (above the threshold) have a high probability of corresponding to a defect. Figure 7 It is shown in Figure 7 , standard deviation 720 (which indicates the confidence level of the expected pixel intensity) is small. Therefore, only large deviations of measured pixel intensity 730 from expected pixel intensity 710 (relative to standard deviation 720) are likely to indicate a defect.

[0128] Figure 5B The method further includes: (operation 530) for each given pixel (or given pixel group) among multiple pixels, determining whether there is a defect using data providing information about expected pixel intensity in the given pixel (or given pixel group), data providing confidence information associated with data providing the expected pixel intensity of the given pixel (or given pixel group) and data providing information about measured pixel intensity (data providing information about noise).

[0129] In particular, for each given pixel, the measured pixel intensity can be compared to the expected pixel intensity for the given pixel. The comparison can be normalized using data that provides information about confidence (data that provides information about noise). In other words, the machine learning algorithm 112 outputs built-in data that implements noise normalization.

[0130] According to some examples, the following non-limiting equation may be used to calculate a score for each given pixel (this score is particularly useful for symmetric distributions):

[0131] In this equation, pdf(x) is a probability distribution (e.g., a Gaussian distribution, a gamma distribution, or other probability distribution, etc.) associated with one or more given parameters output by the machine learning algorithm 112, c is the measured pixel intensity of a given pixel, and μ is the mean of the probability distribution.

[0132] If the score (as mentioned in Equation 2) is above or equal to the threshold, a defect may be detected in the given pixel (or in the given pixel group). If the score is below the threshold, this indicates that there is no defect in the given pixel (or in the given pixel group).

[0133] When the probability distribution is Gaussian, the score of each pixel (as shown in Equation 2) is:

[0134] In this equation, c is the measured pixel intensity for a given pixel, μ is the mean of the Gaussian distribution, which corresponds to the expected pixel intensity (in the absence of a defect, as output by the machine learning algorithm 112 for each pixel), and σ is the standard deviation of the Gaussian model, which corresponds to the data providing information about the confidence (as output by the machine learning algorithm 112 for each pixel, as described above, σ provides information about the noise). If the score (as mentioned above) is above or equal to a threshold, then a defect may be detected in the given pixel (or in a given pixel group). If the score is below the threshold, then this indicates that there is no defect in the given pixel (or in a given pixel group).

[0135] The use of data providing information on the noise in each given pixel (denoted for example σ) enables normalization of the noise and therefore enables differentiation between noise and defects in each given pixel.

[0136] If σ is high, this indicates that the level of noise present in the given pixel is high. Therefore, even if the measured pixel intensity is different from the expected pixel intensity, this does not necessarily mean that there is a defect, but may be due to noise. This can be evaluated individually for each given pixel using data providing information about the noise (e.g. σ in the Gaussian model) generated by the machine learning algorithm 112 for each given pixel. Therefore, it can be used to distinguish between defects and noise in a given pixel or in a given group of pixels. If the data providing information about the noise (e.g. σ in the Gaussian model) is low, this indicates that the level of noise present in the given pixel is small. Therefore, a small difference between the measured pixel intensity and the expected pixel intensity can already indicate the presence of a defect.

[0137] Figure 5A and / or Figure 5B The method may be performed during a runtime scan of a sample by an inspection tool. Each time a new image is provided by the inspection tool (e.g., for each new die), the image may be obtained based on Figure 5A and / or Figure 5B The new image is processed by this method.

[0138] Attention now turns to Figure 8 and Fig. 9 , which describes a method for detecting defects in images using multiple images fed to a machine learning algorithm 112.

[0139] Figure 8 The method includes obtaining (operation 800) a first inspection image 910 acquired by an inspection tool providing information of a first region 910 of a semiconductor sample 1 , and obtaining (operation 805) at least a second inspection image (905) providing information of a second region (905, 920) of the semiconductor sample that is different from the first region 9101 ,920 1 In some examples, multiple additional inspection images may be obtained (each providing information for multiple different regions of the sample, all different than the first region). In some examples, each inspection image provides information for a region corresponding to a die of the sample.

[0140] In some examples, the first region and the second region meet a similarity criterion. For example, they can provide information of similar structural elements (such as similar dies). In some examples, the first region is located at a given position in a first sample, and the second region is located at the same given position in a second sample different from the first sample (the second sample is manufactured using the same manufacturing process as the first sample).

[0141] The second inspection image and / or the plurality of additional inspection images may be used as reference images by the machine learning algorithm 112 in order to detect whether the first inspection image includes (a plurality of) defects. The first inspection image and the plurality of additional inspection images may belong to the same image that has been segmented, or they may have been acquired separately by the inspection tool. The first inspection image and the plurality of additional inspection images may provide information on regions of the same sample or different samples.

[0142] According to some examples, the inspection tool scans various regions of the sample according to a recipe that specifies the path of the electron beam on the sample. The recipe dictates the order in which various dies / regions of the sample are acquired.

[0143] The first area 910 may correspond to an area that must (currently) be inspected for defects, and the second area 920 may correspond to an area acquired by the inspection tool after (e.g., immediately after) the first area 910. In some examples, the second area may correspond to an area 905 acquired by the inspection tool before the first area 910. The first area 910 may correspond to an area acquired immediately after the second area 905. For example, the first area 910 corresponds to a given die, the second area 905 corresponds to a die located above the given die, and the second area 920 corresponds to a die located below the given die.

[0144] In some examples, the method includes obtaining a first inspection image (such as first inspection image 910) providing information of the first region. 1 ), a first additional inspection image (such as inspection image 905) that provides information of an area acquired before the first area 1 ), and a second additional inspection image (such as inspection image 920) that provides information about an area acquired after the first area. 1 ).

[0145] Figure 8The method includes feeding (operation 810) at least a first inspection image and a second inspection image into a trained machine learning algorithm 112. If multiple additional inspection images have been obtained, the method can include feeding the first inspection image and the multiple additional inspection images into the trained machine learning algorithm 112.

[0146] Based on the first inspection image and one or more additional inspection images, the machine learning algorithm 112 generates (operation 820) one or more given parameters of a given model for each given pixel among the multiple pixels of the first inspection image or for each given pixel group among the multiple pixel groups of the first inspection image (such as the pixels providing information of the first region 910). The given model provides information on the pixel intensity distribution. In particular, at least a part of the given model provides information on the pixel intensity distribution in the absence of defects. The given model also provides information on the expected pixel intensity in the given pixel or given pixel group in the absence of defects, and the probability that the deviation from the expected pixel intensity corresponds to a defect. As described above, one or more given parameters or at least a part of the given model (as defined by the given parameters) can be used to distinguish the presence of a defect from the presence of noise in the given pixel or given pixel group.

[0147] Figure 8 The method further includes: (operation 830) for each given pixel or for each given pixel group, using at least some of the one or more given parameters, or at least a part of the given model, and the measured pixel intensity of the given pixel or given pixel group to determine whether there is a defect in the given pixel or given pixel group.

[0148] A score can be calculated for each pixel based on the given model (see equations 1, 2, or 3).

[0149] If the score is higher than or equal to the threshold, a defect can be detected in the given pixel (or in the given pixel group). If the score is lower than the threshold, this indicates that there is no defect in the given pixel (or in the given pixel group).

[0150] In some examples, for each given pixel, the measured pixel intensity can be compared with the expected pixel intensity of the given pixel. This comparison can be normalized by using data providing information on confidence (data providing information on noise). The result of the comparison can be compared with a threshold to detect whether there is a defect.

[0151] Figure 8 The method can be performed during the runtime scan of a sample by an inspection tool. Each time a new image is provided by the inspection tool (e.g., for each new die), the new image can be processed according to Figure 8 the method.

[0152] Attention shifts Fig.10 , which describes a method for training a machine learning algorithm 112.

[0153] Fig.10 The method includes obtaining (operation 1000) a first training image providing information of a first area of ​​a sample. According to some examples, the first training image has been acquired by an inspection tool (see reference numeral 101 and / or reference numeral 102). Note that the first inspection image may have been acquired by an optical inspection tool, or by an electron beam inspection tool, or by another adapted inspection tool. The same inspection tool may be used, or different inspection tools may be used.

[0154] In some examples, operation 1000 may include obtaining one or more (additional) training images in addition to the first training image, each (additional) training image providing information of a different region of the sample (or, in some examples, one or more other samples). In some examples, each region corresponds to a different die of the sample (or, in some examples, one or more other samples). In some examples, the first training image provides information of the first region, and the additional training images provide information of regions that meet a similarity criterion with the first region. For example, the similarity criterion may require that the region corresponds to a similar die of the same sample (or another sample).

[0155] In some examples, the first training image is an image in which a defect may exist, and the one or more additional training images correspond to a reference image that is assumed to be free of defects. However, this is not limiting.

[0156] Fig.11 1 shows a non-limiting example of a plurality of training images that may be fed to the machine learning algorithm 112. A first training image 1100 provides a first region 1100 1 The second training image 1110 provides the second area 1110 1 According to some examples, the first area 1100 1 and the second area 1110 1 The similarity criterion is satisfied. For example, the similarity criterion may require that both regions correspond to similar dies of the same sample (or another sample). Similarly, if the third region 1120 is provided 1 According to some examples, the first region 1100 1 , second area 1110 1 and the third area 1120 1 The similarity criteria can be met.

[0157] Fig.10The method includes feeding (operation 1010) at least a first training image to the machine learning algorithm 112 for training of the machine learning algorithm 112. If one or more additional training images are used, one or more additional training images may be fed to the machine learning algorithm 112 in addition to the first training image for training of the machine learning algorithm 112.

[0158] The machine learning algorithm 112 is trained to predict one or more given parameters of a given model for each given pixel in a plurality of pixels of a first training image or for each given pixel group in a plurality of pixel groups of a first training image. The given model provides information of a pixel intensity distribution. In particular, at least a portion of the given model provides information of a pixel intensity distribution in the absence of a defect in a given pixel or in a given pixel group (the given model can be used to determine an expected pixel intensity in the absence of a defect in a given pixel or in a given pixel group).

[0159] When one or more additional reference images are fed to the machine learning algorithm 112, they may be used as reference images by the machine learning algorithm 112 (assuming no defects are present in the reference images).

[0160] In some examples, the machine learning algorithm 112 is trained to predict, for each given pixel in a plurality of pixels of the first training image or for each given pixel group in a plurality of pixel groups of the first training image, data providing information about the noise present in the given pixel (or in the given pixel group). As described above, this may correspond to a confidence level in the prediction of the expected pixel intensity (in the absence of the defect) estimated by the machine learning algorithm 112.

[0161] In some examples, the machine learning algorithm 112 is trained to predict one or more values ​​of one or more parameters of a model (such as a probabilistic model) that models the intensity of the pixels in each pixel or each pixel group in a plurality of pixels of the first training image. In some examples, the parameters may include data providing information about the expected pixel intensity in a given pixel or in a given pixel group (in the absence of defects in the given pixel or in the given pixel group), and data providing information about the noise present in each given pixel.

[0162] Training of the machine learning algorithm 112 may be performed using methods such as back-propagation (which is not limiting).

[0163] In some examples, the model is a Gaussian model, and the machine learning algorithm 112 is trained to determine, for each given pixel of the first training image, a mean value ("μ") of the model and a standard deviation (labeled "σ") of the model. The mean value ("μ") corresponds to the expected pixel intensity for the given pixel when no defect is present in the pixel. The standard deviation (labeled "σ") corresponds to the level of noise present in the given pixel. Note that the Gaussian model is merely an example, and other probabilistic models with different parameters and / or a different number of parameters may be used.

[0164] Fig.10 The method (operations 1000 and 1010) may be repeated with different first training images (see operation 1020) that are fed to the machine learning algorithm 112. As described above, operations 1000 and 1010 may include obtaining one or more additional training images and feeding them to the machine learning algorithm 112. The one or more additional training images may be the same at each iteration of the method, or may be different. Note that the various training images may have been acquired by scanning the sample at random locations. The various training images may have been acquired by an optical inspection tool, or by an electron beam inspection tool, or by another adapted inspection tool. The method may be repeated. Fig.10 until the loss function (used to train the machine learning algorithm 112) satisfies the optimization criterion (e.g., its value is below a threshold).

[0165] Fig.12 The output of the machine learning algorithm 112 during its training is shown. For each given pixel (1110 of the first training image 1110 1,1 ,1110 1,2 etc.), the machine learning algorithm 112 determines the corresponding given pixel (labeled 1200) of the model that models the pixel intensity in the pixel. 1,1 , 1200 1,2 Thus, a graph 1200 of values ​​is obtained.

[0166] The loss function used to train the machine learning algorithm 112 may include the measured pixel intensity of each pixel of the first training image that must be determined, and the parameters of the model. As described above, in some examples, the parameters include data providing information about the expected pixel intensity to be determined for each pixel (in the absence of defects), and data providing information about the noise to be determined for each pixel. When one or more additional training images are fed to the machine learning model 112 in each iteration for training of the machine learning model 112, the machine learning algorithm 112 may also use the pixel intensities of these one or more additional training images to determine the parameters during training of the machine learning model 112.

[0167] According to some examples, for each pixel, the loss function comprises a difference between a measured pixel intensity for each pixel of the first training image and an expected pixel intensity (to be determined in the absence of defects), wherein the difference is normalized by a noise level in the pixel (to be determined).

[0168] In some examples where a Gaussian distribution is used to model pixel intensity, the following loss function may be used (this equation is not limiting):

[0169] In this equation, i corresponds to the number of the pixel in the first training image, "curr" corresponds to the measured pixel intensity of pixel number i, μ corresponds to the expected pixel intensity of pixel number i (if it has no defects), σ corresponds to the estimate of the noise in pixel number i, f is a function including a probability model (Gaussian distribution) for modeling pixel intensity, and G corresponds to, for example, a logarithmic function.

[0170] In this example, where the probability model f is a Gaussian model (however, this is not limiting), the machine learning algorithm 112 is trained to determine the mean ("μ") of the model and the standard deviation (labeled "σ") of the model. The mean corresponds to the expected pixel intensity of the pixel when no defect is present in the pixel.

[0171] Note that the number of parameters of the model can be greater than two. Furthermore, the model can be different from a Gaussian model and any other suitable probabilistic model can be used.

[0172] Attention now turns to Fig.13 and Fig.14 .

[0173] As described above, the first training image (in some examples, along with one or more additional training images) is fed to the machine learning algorithm 112 for training of the machine learning algorithm 112 .

[0174] The first training image is obtained based on acquiring the area by the inspection tool.In some examples, the first training image may correspond to an inspection image acquired by the inspection tool to which one or more artificial defects have been added.

[0175] In some examples, a test image is acquired (operation 1300) and used to generate (operation 1310) a plurality of training images. Each given training image in the plurality of training images can be obtained by adding one or more artificial defects to the test image. This can also be designated as implanting defects in the test image. The location and / or type and / or aspect of the one or more artificial defects can vary between different training images. This enables the training set of training images to be enhanced without the need to perform additional acquisitions.

[0176] pass Fig.13 Each training image generated by the method can be used as an input for training the machine learning algorithm 112 together with one or more reference training images, such as reference Fig.10 Explained.

[0177] Fig.14 Shows Fig.13 A non-limiting example of a method in which a test image 1400 is used to generate three training images 1421, 1422, and 1423. The first training image 1421 has a first artificial defect 1421 added thereto. 1 The second training image 1422 corresponds to the inspection image 1400. The second training image 1422 has a second artificial defect 1422 added thereto. 1 The third training image 1423 corresponds to the inspection image 1400. The third training image 1423 has the third artificial defect 1423 added. 1 1400. Note that more than one artificial defect may be added to the inspection image.

[0178] Note that it is not mandatory to generate training images with artificial defects in order to train the machine learning algorithm 112, and training images corresponding to images acquired from one or more samples by the inspection tool may be used. In some examples, one or more of the training images may correspond to simulated images.

[0179] Attention now turns to Fig.15 .

[0180] When the inspection tool acquires the same area of ​​a sample twice, the image between the two acquisitions may differ. This is due to the presence of noise (e.g. generated by the inspection tool). Additionally, the noise may be caused by the manufacturing process of the sample. In order to account for the presence of noise during the training process, Fig.15The method includes obtaining (operation 1500) an inspection image of a region of a sample (captured by an inspection tool), and adding (operation 1510) noise to the inspection image to generate one or more training images. Note that the inspection image may have been acquired by an optical inspection tool, or by an electron beam inspection tool, or by another adapted inspection tool. Different levels of noise may be introduced into the inspection image to generate different training images. Fig.10 The training images are used in the method to train the machine learning algorithm 112.

[0181] Attention now turns to Fig.16 .

[0182] Assume that one or more training images of a first dimension 1600 have been used to train the machine learning algorithm 120. The first dimension 1600 corresponds, for example, to the height and / or width of the training image(s). Note that the height of the training image(s) is not necessarily the same as their width, and the first dimension is not necessarily the same for all training images. During the prediction phase, in which the machine learning algorithm 112 is used to determine data that provides information about the expected pixel intensity (in the absence of defects), a test image of a second dimension 1610 may be used, wherein the second dimension 1610 is different from the first dimension 1600. The second dimension corresponds, for example, to the height and / or width of the test image(s). Note that the height of the test image(s) is not necessarily the same as their width, and the second dimension of the test image is not necessarily the same for all test images. According to some examples, the first dimension 1600 may be smaller than the second dimension 1610. Note that this ability to use test images of larger size than the training images may be ensured by using a convolutional neural network (such as a fully convolutional neural network) as the machine learning algorithm 112. However, this is not restrictive.

[0183] Attention now turns to Fig.17 .

[0184] The various methods described herein may be used not only to detect defects, but also for other applications, such as (but not limited to) generating denoised images and / or defect-free images, as explained below.

[0185] Fig.17 The method includes obtaining (operation 1700) an inspection image of an area of ​​a sample. Note that the inspection image may have been acquired by an optical inspection tool, or by an electron beam inspection tool, or by another adapted inspection tool.

[0186] Fig.17 The method includes feeding (operation 1710) at least a test image to a trained machine learning algorithm 112.

[0187] Fig.17 The method further includes: for each given pixel in a plurality of pixels of the inspection image or each given pixel group in a plurality of pixel groups of the inspection image, using (operation 1720) the machine learning algorithm 112 to generate data providing information of expected pixel intensities in the given pixel or in the given pixel group in the absence of defects in the given pixel or in the given pixel group. In particular, the machine learning algorithm 112 may generate, for each given pixel or given pixel group, one or more given parameters of a given model providing information of pixel intensity distribution. The given model (defined by the given parameters) or the given parameters may be used to determine a non-defective pixel intensity for the given pixel. In some examples, a median (or average) of the given model may be determined, and the median (or average) corresponds to the non-defective pixel intensity for the given pixel. A set of expected pixel intensities (in the absence of defects) is obtained for the pixels of the inspection image.

[0188] Fig.17 The method further includes generating a new image using (operation 1730) data providing information about expected pixel intensities generated by the machine learning algorithm 112 for a plurality of pixels of the test image. In particular, each pixel (1860) of the new image 1860 1 , 1860 2 etc.) may be assigned a pixel for the corresponding pixel in the inspection image (having the same position - see, for example, 1800 1 , 1800 2 etc.) calculated from the expected pixel intensity (in the absence of defects).

[0189] The new image is a denoised image. In particular, the noise present in the new image is less than the noise present in the test image. In some examples, the noise is eliminated in the new image. The new image is defect-free or contains fewer defects than the test image. The new image can be used for various applications, such as a reference image. In some examples, the new image can be used together with the test image as a reference image provided to the machine learning algorithm 112 for use in the prediction phase (see, e.g., Figure 8 ) and / or during the training phase (see e.g. Fig.10 ) to detect (multiple) defects in the inspection image.

[0190] Attention now turns to Fig.19 .

[0191] The various methods described herein can be used not only for detecting defects, but also for other applications, such as (but not limited to) generating new images. For example, there may be a situation where the number of available inspection images is insufficient. Fig.19 The method enables artificial generation of new images that look similar to actual test images. This enables augmenting the size of the training set with realistic images.

[0192] Fig.19 The method includes obtaining (operation 1900) an inspection image 2000 of an area of ​​a sample. Note that the inspection image 2000 may have been acquired by an optical inspection tool, or by an electron beam inspection tool, or by another adapted inspection tool.

[0193] Fig.19 The method includes feeding (operation 1910) at least a test image 2000 to a trained machine learning algorithm 112.

[0194] Fig.19 The method further includes generating (operation 1920) one or more given parameters of a given model (also referred to as a pixel intensity probability distribution, a statistical model, a probability model, or a probability function) providing information of a pixel intensity distribution for each given pixel in the plurality of pixels of the inspection image 2000 or for each pixel group in the plurality of pixel groups of the inspection image 2000. The pixel intensity probability distribution can be used to determine the probability that a measured pixel intensity of a given pixel or a given pixel group corresponds to a defect.

[0195] Fig.19 The method further includes generating a new image (or multiple new images) using (operation 1930) the given model obtained for each pixel.

[0196] In particular, each given pixel (2060 1 , 2060 2 ) may be assigned a pixel intensity value based on the corresponding given pixel (2000) for the inspection image 2000. 1 , 2000 2 In fact, since the given model corresponds to a pixel intensity probability distribution, (multiple) values ​​of pixel intensity matching the probability distribution can be randomly generated.

[0197] exist Fig. 20 , an additional new image 2070 is generated. Each given pixel (2070 1 , 2070 2 ) may be assigned a pixel intensity value based on the corresponding given pixel (2000) for the inspection image 2000. 1 , 2000 2 etc.) are randomly generated from a given model.

[0198] One or more new images are obtained that look similar to the test image, but are still different from the test image. The one or more new images can be used, for example, for training of the machine learning algorithm 112, or for other applications.

[0199] Notice, Fig.19 Another specific application of the method may be to generate new images with fewer defects and / or less noise, such as reference Fig.17 Explained.

[0200] In the specific embodiments, numerous specific details have been set forth in order to provide a thorough understanding of the present disclosure. However, those skilled in the art will appreciate that the subject matter disclosed herein can be practiced without these specific details. In other instances, well-known methods, procedures, components, and circuits are not described in detail to avoid confusing the subject matter disclosed herein.

[0201] Unless otherwise expressly stated, as is apparent from the above discussion, it should be understood that throughout the specification discussion, terms such as "obtain," "apply," "determine," "execute," "use," "estimate," "train," "feed," etc., are used to refer to computer actions and / or processes that manipulate and / or transform data into other data, where the data is represented as physical (such as electronic) quantities and / or where the data represents physical objects.

[0202] The term "computer" or "computer-based system" should be broadly interpreted to include any type of hardware-based electronic device having data processing circuitry (e.g., digital signal processor (DSP), GPU, TPU, field programmable gate array (FPGA), application specific integrated circuit (ASIC), microcontroller, microprocessor, etc.), including, as non-limiting examples, Figure 1Computer-based system 103 and corresponding parts disclosed in the present application. Data processing circuit system (also designated as processing circuit system) may include, for example, one or more processors operably connected to a computer memory, the one or more processors being loaded with executable instructions for performing operations, as further described below. Data processing circuit system covers a single processor or multiple processors, the processors may be located in the same geographic area, or may be at least partially located in different areas, and may be able to communicate together. One or more processors may represent one or more general-purpose processing devices, such as microprocessors, central processing units, etc. More specifically, a given processor may be one of the following: a complex instruction set computing (CISC) microprocessor, a reduced instruction set computing (RISC) microprocessor, a very long instruction word (VLIW) microprocessor, a processor implementing other instruction sets, or a processor implementing a combination of instruction sets. One or more processors 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, etc. One or more processors are configured to execute instructions for performing the operations and steps discussed herein.

[0203] The memory referred to in this document may include one or more of the following: internal memory (such as, for example, processor registers and cache, etc.), main memory (such as, for example, read-only memory (ROM)), flash memory, dynamic random access memory (DRAM) (such as synchronous DRAM (SDRAM) or Rambus DRAM (RDRAM), etc.).

[0204] The terms "non-transient memory" and "non-transient storage medium" used herein should be broadly interpreted as covering any volatile or non-volatile computer memory suitable for the subject matter currently disclosed. These terms should be understood to include a single medium or multiple media (e.g., a centralized or distributed database, and / or associated caches and servers) that store one or more groups of instructions. These terms should also be understood to include any medium that can store or encode a group of instructions for execution by a computer and cause the computer to execute any one or more methods in the method disclosed herein. Therefore, these terms should be understood to include, but are not limited to, read-only memory ("ROM"), random access memory ("RAM"), disk storage media, optical storage media, flash memory devices, etc.

[0205] It should be noted that although the present disclosure relates to a processing circuit system 104 configured to perform various functions and / or operations, these functions / operations may be performed in various ways by one or more processors of the processing circuit system 104. As an example, the operations described below may be performed by a specific processor, or by a combination of processors. Therefore, the operations described below may be performed by a corresponding processor (or combination of processors) in the processing circuit system 104, and optionally, at least some of these operations may be performed by the same processor. The present disclosure should not be limited to being interpreted as a single processor that always performs all operations.

[0206] The term "sample" used in this specification should be broadly interpreted to cover any kind of wafers, masks and other structures used to manufacture semiconductor integrated circuits, magnetic heads, flat panel displays and other semiconductor manufactured articles, combinations and / or portions thereof.

[0207] The term "inspection" used in this specification should be broadly interpreted to cover any type of metrology-related operations and operations related to the detection and / or classification of defects in a sample during sample manufacturing. Inspection is provided using a non-destructive inspection tool during or after the manufacture of the sample to be inspected. As a non-limiting example, the inspection process may include runtime scanning (in a single or multiple scans), sampling, review, measurement, classification and / or other operations provided with respect to the sample or part thereof using the same or different inspection tools. Similarly, inspection may be provided before manufacturing the sample to be inspected, and may include, for example, generating (multiple) inspection recipes and / or other setup operations. It should be noted that, unless otherwise expressly stated, the term "inspection" or its derivatives used in this specification are not limited in terms of the resolution or size of the inspection area. As non-limiting examples, various non-destructive inspection tools include scanning electron microscopes, atomic force microscopes, optical inspection tools, and the like.

[0208] As a non-limiting example, the runtime check may employ a two-stage process, e.g., inspecting a sample and then reviewing sampled locations for potential defects. During the first stage, the surface of the sample is inspected at high speed and relatively low resolution. In the first stage, a defect map is generated to show suspected locations on the sample that have a high probability of being defects. During the second stage, at least some of the suspected locations are more thoroughly analyzed at relatively high resolution. In some cases, the two stages may be implemented by the same inspection tool, and in some other cases, the two stages may be implemented by different inspection tools.

[0209] The term "noise" may include variations in an image acquired by an inspection tool relative to an intended design. Noise may be caused, for example, by the inspection tool and / or by the manufacturing process (process variations, etc.).

[0210] The term "defect" includes an abnormal or undesirable feature formed on or within a sample. In some examples, a defect may include noise that varies in magnitude (relative to the intended design) above a certain threshold (which may be selected by the manufacturer). The threshold may vary depending on the type of feature, the type of manufacturing process, etc., or other considerations.

[0211] The term "design data" as used in this specification should be broadly interpreted to encompass any data indicative of a hierarchical physical design (layout) of a sample. The design data may be provided by a corresponding designer and / or may be derived from the physical design (e.g., through complex simulations, simple geometric and Boolean operations, etc.). The design data may be provided in different formats, such as, as non-limiting examples, GDSII format, OASIS format, etc. The design data may be presented in a vector format, a grayscale intensity image format, or otherwise.

[0212] It should be understood that, unless otherwise specifically stated, certain features of the presently disclosed subject matter described in the context of separate embodiments may also be provided in combination in a single embodiment. Conversely, various features of the presently disclosed subject matter described in the context of a single embodiment may also be provided individually or in any suitable subcombination. In the following detailed description, numerous specific details are set forth in order to provide a thorough understanding of the method and apparatus.

[0213] In embodiments of the presently disclosed subject matter, a comparison may be performed. Figure 2 Figure 5 Figure 8 , Fig.10 , Fig.13 , Fig.15 , Fig.17 and Fig.19 In an embodiment of the presently disclosed subject matter, Figure 2 Figure 5 Figure 8 , Fig.10 , Fig.13 , Fig.15 , Fig.17 and Fig.19 One or more of the stages shown in the methods may be performed in a different order, and / or one or more groups of stages may be performed simultaneously.

[0214] It is to be understood that the invention is not limited in its application to the details set forth in the description contained herein or illustrated in the drawings.

[0215] It should also be understood that the system according to the present invention can be implemented at least in part on a suitably programmed computer. Likewise, the present invention contemplates a computer program readable by a computer for performing the method of the present invention. The present invention further contemplates a non-transitory computer readable memory tangibly embodying an instruction program executable by a computer for performing the method of the present invention.

[0216] The present invention is capable of other embodiments and of being practiced and executed in various ways. Therefore, it should be understood that the wording and terminology employed herein are for descriptive purposes and should not be considered restrictive. Therefore, it will be appreciated by those skilled in the art that the concepts upon which the present disclosure is based can be easily used as the basis for designing other structures, methods and systems for achieving the several purposes of the presently disclosed subject matter.

[0217] It will be readily appreciated by those skilled in the art that various modifications and changes may be applied to the embodiments of the invention described above without departing from the scope of the invention as defined by the accompanying claims.

Claims

1. A system comprising one or more processing circuit systems, wherein the one or more processing circuit systems are configured to: obtaining a first inspection image acquired by an inspection tool providing information of a first region of the semiconductor sample, feeding at least the first test image to a machine learning algorithm, the machine learning algorithm being configured for determining, for each given pixel of a plurality of pixels of the first test image or for each given pixel group of a plurality of pixel groups of the first test image, one or more given parameters of a given model providing information of a pixel intensity distribution, For each given pixel or each given pixel group, use: at least one of the one or more given parameters, or at least a portion of the given model associated with the one or more given parameters, and the measured pixel intensity of the given pixel or the given group of pixels, To determine whether there is a defect in the given pixel or in the given pixel group.

2. A system as described in claim 1, wherein the system is configured to determine, for each given pixel or each given pixel group, the probability of a defect existing in the given pixel or in the given pixel group using at least some of the one or more given parameters, or at least a portion of the given model associated with the one or more given parameters.

3. The system of claim 1, wherein: For each given pixel or for each given group of pixels, at least a portion of the given model associated with the one or more given parameters provides information of a pixel intensity probability distribution, which can be used to determine the probability that the measured pixel intensity of the given pixel or the given group of pixels corresponds to a defect.

4. The system of claim 1, wherein: For each given pixel or each given group of pixels, at least some of the one or more given parameters, or at least a portion of the given model associated with the one or more given parameters, can be used to determine the expected pixel intensity in the absence of defects in the given pixel or in the given group of pixels.

5. The system of claim 1, wherein: For each given pixel or each given pixel group, at least one of the one or more given parameters, or at least a portion of the given model associated with the one or more given parameters, can be used to determine: - the expected pixel intensity in the absence of defects in said given pixel or in said given group of pixels, and -The probability that a deviation from the expected pixel intensity corresponds to a defect.

6. The system of claim 1, the system being configured to use the given model to detect, on average in different partitions having different noise levels, a maximum number of defects below a same threshold for the different partitions.

7. A system as described in claim 1, wherein the system is configured to use at least some of the one or more given parameters, or at least a portion of the given model associated with the one or more given parameters, to distinguish between the presence of defects and the presence of noise in the given pixel or in the given pixel group for each given pixel or each given pixel group.

8. The system of claim 1, wherein: The one or more given parameters are determined by the machine learning algorithm specifically for each given pixel or each given pixel group, wherein the one or more given parameters or data derived therefrom include: - data providing information of the expected pixel intensity in said given pixel or in said given group of pixels in the absence of a defect in said given pixel or in said given group of pixels; - At least one of the following: o data providing information on the noise present in said given pixel or in said given group of pixels; o data providing information about a confidence level associated with the data providing information about the expected pixel intensity in said given pixel or in said given group of pixels; o data enabling normalization of the noise present in said given pixel or in said given group of pixels; or o Data enabling differentiation of defects from noise in the given pixel or in the given group of pixels.

9. The system of claim 1, wherein the system is configured to feed the machine learning algorithm, in addition to the first inspection image, one or more reference images providing information of one or more other regions of the semiconductor sample or another semiconductor sample.

10. The system of claim 1, wherein: The machine learning algorithm has been trained using at least one training image and a loss function, wherein for each given pixel among multiple pixels of the training image or for each pixel group among multiple pixel groups of the training image, the loss function includes one or more parameters of a model that models the pixel intensity associated with the given pixel or with the given pixel group.

11. The system of claim 1, wherein: Satisfy at least one of (i) or (ii): (i) the machine learning algorithm is configured to simultaneously determine, for each given pixel in a plurality of pixels of the first inspection image or for each given pixel group in a plurality of pixel groups of the first inspection image: data providing information of the expected pixel intensity in said given pixel or in said given group of pixels in the absence of a defect in said given pixel or in said given group of pixels, and data providing information of noise present in said given pixel or in said given group of pixels; (ii) the same machine learning algorithm is configured to determine, for each given pixel in a plurality of pixels of the first inspection image or for each given group of pixels in a plurality of groups of pixels of the first inspection image: data providing information of the expected pixel intensity in said given pixel or in said given group of pixels in the absence of a defect in said given pixel or in said given group of pixels, and Data providing information of noise present in said given pixel or in said given group of pixels.

12. The system of claim 1, wherein the system is configured to: obtaining a single inspection image acquired by the inspection tool providing information of a region of the semiconductor sample, Determining one or more defects in the region based on the single inspection image, the determining comprising: feeding the single test image to the machine learning algorithm, the machine learning algorithm being configured to determine, for each given pixel of a plurality of pixels of the single test image or for each given pixel group of a plurality of pixel groups of the first test image, one or more given parameters of a given model providing information of pixel intensity distribution, and For each given pixel or each given pixel group, at least one of the one or more given parameters, or at least a portion of the given model associated with the one or more given parameters, and the measured pixel intensity of the given pixel or the given pixel group are used to determine whether there is a defect in the given pixel or in the given pixel group.

13. The system of claim 1, wherein: One or more training images used to train the machine learning algorithm have at least one of a smaller height or a smaller width than the first verification image.

14. The system of claim 1, wherein the system is configured to generate a new image using at least one of the one or more given parameters, or at least a portion of the given model associated with the one or more given parameters, wherein at least one of (i) or (ii) is satisfied: (i) the new image is noise-free or contains less noise than the first test image; (ii) the new image is defect-free or contains fewer defects than the first inspection image.

15. A system comprising one or more processing circuit systems configured to: get: a first training image acquired by an inspection tool providing information of a first region of a semiconductor sample, At least the first training image is fed to the machine learning algorithm to train the machine learning algorithm to: determine one or more given parameters of a given model that provides information on pixel intensity distribution for each given pixel among multiple pixels of the first training image or for each given pixel group among multiple pixel groups of the first training image, wherein the one or more given parameters, or the given model associated with the one or more given parameters, can be used to detect the presence of defects in the given pixel or in the given pixel group.

16. The system of claim 15, wherein: For each given pixel or each given pixel group, the one or more given parameters, or the given model associated with the one or more given parameters, can be used to determine the pixel intensity when no defects are present in the given pixel or in the given pixel group.

17. The system of claim 15, wherein: The one or more training images used to train the machine learning algorithm correspond to one or more inspection images of a semiconductor sample to which one or more artificial defects have been added.

18. A non-transitory computer readable medium comprising instructions that, when executed by at least one or more processing circuit systems, cause the at least one or more processing circuit systems to perform: obtaining a first inspection image acquired by an inspection tool providing information of a first region of the semiconductor sample, feeding at least the first test image to a machine learning algorithm, the machine learning algorithm being configured to determine, for each given pixel of a plurality of pixels of the first test image or for each given pixel group of a plurality of pixel groups of the first test image, one or more given parameters of a given model providing information of a pixel intensity distribution, For each given pixel or each given pixel group, use: at least one of the one or more given parameters, or at least a portion of the given model associated with the one or more given parameters, and the measured pixel intensity of the given pixel or the given group of pixels, To determine whether there is a defect in the given pixel or in the given pixel group.

19. A non-transitory computer-readable medium comprising instructions that, when executed by one or more processing circuit systems, cause the one or more processing circuit systems to perform: obtaining a first inspection image acquired by an inspection tool providing information of a first region of the semiconductor sample, feeding at least the first test image to a machine learning algorithm, the machine learning algorithm being configured to determine, for each given pixel of a plurality of pixels of the first test image or for each given pixel group of a plurality of pixel groups of the first test image, one or more given parameters of a given model providing information of a pixel intensity distribution, determining, for each given pixel or each given group of pixels, a new given pixel intensity value using at least one of the one or more given parameters, or at least a portion of the given model associated with the one or more given parameters, Thus, a new set of given pixel intensity values ​​is obtained, and A new image is generated using the new set of given pixel intensity values.

20. The non-transitory computer readable medium of claim 19, wherein: Each new given pixel intensity value corresponds to a given expected pixel intensity in the absence of a defect in the given pixel or in the given group of pixels, and The new set of given pixel intensity values ​​corresponds to a set of given expected pixel intensity values, At least one of (i) or (ii) is satisfied: (i) the new image is noise-free or contains less noise than the first test image; (ii) the new image is defect-free or contains fewer defects than the first inspection image.

Images