Defect offset correction for semiconductor sample inspection

Abstract

A system and method for inspecting a semiconductor sample is provided. The method includes obtaining a set of defect candidates associated with respective inspection positions represented in an inspection coordinate system; providing, for each defect candidate, a probability that the defect candidate is a defect of interest (DOI) using a trained machine learning (ML) model, and sorting the group of defect candidates into an ordered list according to the respective probabilities of the defect candidates; in response to a portion of the ordered list of defect candidates being reviewed by the review tool according to an order of the ordered list of defect candidates, receiving a predefined number of DOIs associated with respective review locations represented in a review coordinate system; and calculating an offset between the review coordinate system and the inspection coordinate system based on the respective inspection and review positions associated with the predefined number of DOIs.

Description

Technical Field

[0001] The presently disclosed subject matter generally relates to the field of inspecting semiconductor samples, and more particularly to position correction of defects on a sample. Background Art

[0002] Current demands for high density and performance associated with the very large scale integration of fabricated devices require sub-micron features, increased transistor and circuit speeds, and improved reliability. As semiconductor processes advance, pattern dimensions such as line widths and other types of critical dimensions continue to shrink. These demands require the formation of device features with high precision and uniformity, which in turn requires careful monitoring of the manufacturing process, including automated inspection of devices while they are still in the form of semiconductor wafers.

[0003] Non-destructive inspection tools can be used to provide inspection during or after the manufacture of a sample to be inspected. As non-limiting examples, various non-destructive inspection tools include scanning electron microscopes, atomic force microscopes, optical inspection tools, and the like.

[0004] The inspection process can include multiple inspection steps. The manufacturing process of semiconductor devices can include various processes such as etching, deposition, planarization, growth (such as epitaxial growth), implantation, and the like. The inspection steps can be performed multiple times, for example, after certain process steps and / or after manufacturing certain layers, etc. Additionally or alternatively, each inspection step can be repeated multiple times, for example, for different wafer positions or for the same wafer position with different inspection settings.

[0005] Inspection processes are used at various steps during semiconductor manufacturing to detect and classify defects on a sample and to perform metrology-related operations. The effectiveness of inspection can be improved through the automation of processes such as, for example, defect detection, automatic defect classification (ADC), automatic defect review (ADR), image segmentation, automated metrology-related operations, and the like.

[0006] Automated inspection systems ensure that the fabricated parts meet the expected quality standards and provide useful information regarding possible adjustments to manufacturing tools, equipment, and / or ingredients based on the identified defect types. In some cases, machine learning techniques can be used to assist the automated inspection process in order to promote higher yields. Summary of the Invention

[0007] According to certain aspects of the presently disclosed subject matter, there is provided a computerized system for inspecting a semiconductor sample, the system including a processing circuitry configured to: obtain a group of defect candidates from a defect map generated by inspecting the sample by an inspection tool, the group of defect candidates being associated with respective inspection positions represented in an inspection coordinate system; use a trained machine learning (ML) model to provide, for each defect candidate in the group, a probability that the defect candidate is a defect of interest (DOI), and rank the group of defect candidates into an ordered list of defect candidates according to the respective probabilities of the defect candidates; in response to a portion of the ordered list of defect candidates being reviewed by a review tool according to the order of the ordered list of defect candidates, receive from the review tool a predefined number of DOIs identified from the portion of the ordered list and associated with respective review positions represented in a review coordinate system; and calculate an offset between the inspection coordinate system and the review coordinate system based on the respective inspection positions and review positions associated with the predefined number of DOIs, wherein the offset can be used by the review tool to review at least a portion of the group of defect candidates that have not yet been reviewed.

[0008] In addition to the above features, the system according to this aspect of the presently disclosed subject matter may also include one or more of the following features (i) to (xii) in any desired combination or permutation that is technically possible: (i). The ML model may be previously trained using a training set of defect candidates, each defect candidate being characterized by one or more inspection attributes and associated with a respective ground truth (GT) label indicating whether the defect candidate is a DOI or an interference. (ii). One or more inspection attributes characterizing each defect candidate may include at least one of the following: rank, volume, polarity, intensity, size, and the probability that the defect candidate is a DOI. (iii). Training of the ML model may include: for each given defect candidate in the training set, processing the given defect candidate through the ML model to obtain a predicted class of the given defect candidate, and optimizing the ML model using a loss function based on the predicted class and the GT label associated with the given defect candidate. (iv). The loss function may be configured based on selection purity, which represents the percentage of the number of actual DOIs among the number of defect candidates identified as DOIs. (v). Only the defect candidates in the training set that are DOIs and have a relatively high correlation level for offset correction may be labeled as DOIs. (vi). The offset may be represented by a transformation matrix that includes a plurality of transformation coefficients corresponding to a plurality of degrees of freedom of the transformation between the inspection coordinate system and the review coordinate system. (vii). The plurality of transformation coefficients can include two or more of the following: X offset, Y offset, rotation, perpendicularity, X scaling and Y scaling, X parabolic and Y parabolic. (viii). The predefined number of selected DOIs can be associated with the number of transformation coefficients in the transformation matrix. (ix). The offset can be calculated by performing a linear regression based on the corresponding inspection positions and review positions associated with the predefined number of DOIs. (x). The processing circuitry can further be configured to obtain the review positions of at least a portion of the group of defect candidates not yet reviewed using the offset, such that a review tool can review at least a portion of the group of defect candidates at the review positions. (xi). Reviewing at least a portion of the group of defect candidates not yet reviewed using the offset enables an increase in the throughput (TpT) of the review tool. (xii). The ML model is implemented as one of a random forest, a logistic regression model, or a neural network.

[0009] According to other aspects of the presently disclosed subject matter, a computerized method for inspecting a semiconductor sample is provided, the method comprising: obtaining a group of defect candidates from a defect map generated by inspecting a semiconductor sample by an inspection tool, the group of defect candidates being associated with corresponding inspection positions represented in an inspection coordinate system; using a trained machine learning (ML) model to provide, for each defect candidate in the group, a probability that the defect candidate is a defect of interest (DOI), and sorting the group of defect candidates into an ordered list of defect candidates according to the corresponding probabilities of the defect candidates; in response to a portion of the ordered list of defect candidates being reviewed by a review tool in the order of the ordered list of defect candidates, receiving from the review tool a predefined number of DOIs identified from the portion of the ordered list and associated with corresponding review positions represented in a review coordinate system; and calculating an offset between the inspection coordinate system and the review coordinate system based on the corresponding inspection positions and review positions associated with the predefined number of DOIs, wherein the offset can be used by the review tool to review at least a portion of the group of defect candidates not yet reviewed.

[0010] These aspects of the presently disclosed subject matter can be modified as necessary, in any desired combination or arrangement technically possible, to include one or more of the features (i) to (xii) listed above with respect to the system.

[0011] In other aspects of the presently disclosed subject matter, there is provided a non-transitory computer-readable storage medium including instructions that, when executed by a computer, cause the computer to perform a method of inspecting a semiconductor sample, the method including: obtaining a group of defect candidates from a defect map generated by inspecting a semiconductor sample by an inspection tool, the group of defect candidates being associated with respective inspection locations represented in an inspection coordinate system; using a trained machine learning (ML) model to provide, for each defect candidate in the group, a probability that the defect candidate is a defect of interest (DOI), and sorting the group of defect candidates into an ordered list of defect candidates according to the respective probabilities of the defect candidates; in response to a portion of the ordered list of defect candidates being reviewed by a review tool in the order of the ordered list of defect candidates, receiving from the review tool a predefined number of DOIs identified from the portion of the ordered list and associated with respective review locations represented in a review coordinate system; and calculating an offset between the inspection coordinate system and the review coordinate system based on the respective inspection locations and review locations associated with the predefined number of DOIs, wherein the offset can be used by the review tool to review at least a portion of the group of defect candidates that have not yet been reviewed.

[0012] These aspects of the subject matter of the present disclosure may be modified as necessary to include one or more of the features (i) through (xii) listed above with respect to the system in any desired combination or arrangement that is technically possible. BRIEF DESCRIPTION OF THE DRAWINGS

[0013] To understand the present disclosure and to see how it may be implemented in practice, embodiments will now be described by way of non-limiting example only with reference to the accompanying drawings, in which:

[0014] Figure 1 A generalized block diagram of an inspection system in accordance with certain embodiments of the presently disclosed subject matter is shown.

[0015] Figure 2 A generalized flowchart of automatic defect offset correction between an inspection tool and a review tool in accordance with certain embodiments of the presently disclosed subject matter is shown.

[0016] Figure 3 A generalized flowchart of training an ML model in accordance with certain embodiments of the presently disclosed subject matter is shown.

[0017] Figure 4 A schematic diagram of two different coordinate systems and an example of a defect offset in accordance with certain embodiments of the presently disclosed subject matter is shown.

[0018] Figure 5 A schematic diagram of an example of a training set in accordance with certain embodiments of the presently disclosed subject matter is shown.

[0019] Figure 6Shows a schematic diagram of the runtime deployment of an ML model according to certain embodiments of the present disclosure.

[0020] Figure 7 Shows a schematic diagram of the training process of an ML model according to certain embodiments of the present disclosure.

[0021] Figure 8 Shows a schematic diagram of offset calculation according to certain embodiments of the present disclosure.

[0022] Figure 9 Shows a graph demonstrating an improvement in the efficiency of anchor defect selection according to certain embodiments of the present disclosure.

[0023] Figure 10 Shows an example of an improvement in throughput (TpT) for inspecting samples according to certain embodiments of the present disclosure. Detailed Description

[0024] Semiconductor manufacturing processes typically require multiple sequential processing steps and / or layers, some of which may result in errors, which may in turn lead to yield loss. Examples of various processing steps can include lithography, etching, deposition, planarization, growth (such as, for example, epitaxial growth), and implantation, among others. Various inspection operations, such as defect-related inspections (e.g., defect detection, defect review, and defect classification, etc.) and / or metrology-related inspections (e.g., critical dimension (CD) measurement, etc.), can be performed at different processing steps / layers during the manufacturing process to monitor and control the process. The inspection operations can be performed multiple times, for example, after certain processing steps and / or after manufacturing certain layers, etc.

[0025] Defect-related inspections typically can employ a two-stage process, e.g., inspecting a sample and then reviewing the sampled locations of potential defects. During the first stage, the surface of the sample is inspected by an inspection tool at a relatively high speed and relatively low resolution. Defect detection is typically performed by applying a defect detection algorithm to the inspection output. Various detection algorithms can be used to detect defects on the sample, such as die-to-die (D2D), die-to-history (D2H), die-to-database (D2DB), cell-to-cell (C2C), etc. As an example, in D2D, an inspection image of a target die is captured, and one or more reference images are captured from one or more reference dies (e.g., one or more adjacent dies) of the target die. The inspection image and the reference images are aligned and compared to each other. One or more difference images (and / or their derivatives, such as a grading image) can be generated based on the differences between the pixel values of the inspection image and the pixel values derived from one or more reference images. Then a detection threshold is applied to the difference map, and a defect map is produced to show the suspicious locations on the target die that have a high probability of being true defects (also referred to as defects of interest (DOI)).

[0026] During the second stage, at least some of the suspect locations on the defect map are more thoroughly analyzed by a review tool with a relatively high resolution to determine whether the defect candidate is indeed a DOI, and / or to determine different parameters of the DOI, such as category, thickness, roughness, size, etc.

[0027] When a defect candidate identified by an inspection tool is sent to a review tool for review, the defect candidate often cannot be observed by the review tool at the assumed location indicated in the defect map. This is mainly because the inspection tool and the review tool are different inspection machines and may be produced by different companies in some cases. Naturally, these two tools have different coordinate systems, which results in the inspection coordinates representing the location of the defect candidate in the defect map not being automatically usable by the review tool without correction / adjustment. Generally, when the review tool acquires a review image at the location of the defect candidate specified in the defect map, due to this mismatch between the two different coordinate systems, it is common for the review image not to capture the defect candidate at the exact location.

[0028] Figure 4 A schematic diagram showing two different coordinate systems and an example of defect offset according to certain embodiments of the presently disclosed subject matter is shown.

[0029] The coordinate system 402 is illustrated as the inspection coordinate system of an inspection tool (such as, for example, an optical tool), while the coordinate system 404 is illustrated as the review coordinate system of a review tool (such as, for example, an SEM). As shown, there is a mismatch between the two coordinate systems, in one or more degrees of freedom of the transformation between the two systems, such as, for example, X offset, Y offset, rotation, perpendicularity, X scaling, Y scaling, X parabola, and Y parabola, etc. Due to this mismatch, the inspection location of the defect in the inspection coordinate system 402 is not aligned with the review location of the same defect in the review coordinate system 404 ( Figure 4 Three defect examples are respectively marked with circles, stars, and triangles in), and the review location is usually shifted by an offset from the inspection location.

[0030] Therefore, to ensure capturing the defect candidate, the review tool needs to be configured with an enlarged field of view (FOV) to ensure that the defect candidate is within the FOV. As shown at 408, when the review tool navigates based on the inspection location, that is, when determining the center of the FOV through the inspection location, the actual location of the defect may deviate from the center by a certain offset. In this example, the defect appears with an offset to the upper right corner of the FOV. Therefore, the FOV of the review tool should be enlarged to a certain extent to ensure that the shifted defect will be captured. Since the review tool usually scans the sample at a relatively low speed and high resolution, the enlargement of the FOV will necessarily result in a long image capture time, thus having a negative impact on the throughput (TpT) of the system.

[0031] In some cases, in addition to increasing the size of the FOV, the coordinates between the inspection tool and the review tool can also be manually aligned, which is called manual defect offset (MDO) correction. The traditional MDO process involves a repetitive process of accessing defect candidates in the defect map one by one, and verifying each candidate with the review tool to locate a qualified defect for alignment (the defect is also referred to as an "anchoring" defect). Since the defect map usually contains a large number of defect candidates, most of which are actually false alarms, this trial-and-error of repeatedly manually selecting defects to locate the "anchoring" defect is cumbersome and time-consuming.

[0032] Some efforts have been devoted to trying to establish automatic offset correction between different tools, such as, for example, reference-based offset correction. In such cases, reference features on the wafer are used to identify the offset between the tools, and then the offset between the tools can be applied to the defect candidates on the defect map. For the purpose of having unique patterns as registration reference features, in some cases, virtual patterns need to be designed and fabricated on the physical wafer, such as in the scribe region. Such physical changes to the wafer inevitably affect the flexibility of the integrated circuit (IC) design, increase process complexity, and may even affect the yield. In addition, the reference-based correction method does not provide a direct measurement of the defect coordinate difference (but via the reference coordinate offset), so it cannot always ensure the measurement accuracy level.

[0033] With the continuous progress of semiconductor manufacturing processes, semiconductor devices have evolved into increasingly complex structures with shrinking feature sizes, which increases the sensitivity of semiconductor processing to defect offsets, and thus makes it more critical to correct the misalignment between different inspection tools and match the coordinates between different coordinate systems (such as aligning the inspection coordinate system 402 and the review coordinate system 404, as shown in 406) to provide satisfactory inspection performance.

[0034] Accordingly, certain embodiments of the presently disclosed subject matter propose an automatic defect offset correction system that does not have one or more of the disadvantages described above. The present disclosure proposes sampling only a small number of defect candidates in order to identify "anchoring" defects that can be used to establish the more accurate positions of all defect candidates. The present disclosure provides an automated system that is configured to correct defect offsets in an automated manner without the need for manual intervention or suffering from the trial-and-error of defect selection for locating qualified defects for alignment, thereby significantly improving the review efficiency, as will be detailed below.

[0035] With this in mind, attention turns to Figure 1 , Figure 1 FIG. shows a functional block diagram of an inspection system according to certain embodiments of the presently disclosed subject matter.

[0036] Figure 1 The inspection system 100 shown can be used for inspecting semiconductor samples (e.g., wafers, dies, or portions thereof) as part of a sample manufacturing process. As described above, the inspection referred to herein can be interpreted to cover any kind of operation related to defect inspection / detection, defect review, defect classification, interference filtering, segmentation, and / or metrology operations, etc. on the sample. The system 100 includes one or more inspection tools, which are configured to scan the sample and capture an image of the sample for further processing for various inspection applications.

[0037] The term "(s) inspection tool" used herein should be interpreted broadly to cover any tool that can be used for inspecting the relevant process, as non-limiting examples, including scanning (in single or multiple scans), imaging, sampling, reviewing, measuring, classifying, and / or other processes provided on the sample or a portion thereof. Without limiting the scope of the present disclosure in any way, it should also be noted that the inspection tool can be implemented as various types of inspection machines, such as optical inspection machines, electron beam inspection machines (e.g., scanning electron microscope (SEM), atomic force microscope (AFM), or transmission electron microscope (TEM), etc.), and so on.

[0038] One or more inspection tools may include one or more inspection tools 120 and one or more review tools 121. In some cases, the inspection tool 120 can be configured to scan the sample (e.g., the entire wafer, the entire die, or a portion thereof) to capture inspection images (usually at a relatively high speed and / or low resolution) for detecting potential defects (i.e., defect candidates). During inspection, the wafer can be moved relative to the detector of the inspection tool at a certain step size during exposure (or the wafer and the tool can move relative to each other in opposite directions), and the wafer can be scanned step by step along a strip of the wafer by the inspection tool, where the inspection tool images the components / portions (within the strip) of the sample at a time. As an example, the inspection tool can be an optical inspection tool. In each step, light can be detected from a rectangular portion of the wafer, and this detected light is converted into multiple intensity values at multiple points in the portion, thereby forming an image corresponding to the components / portions of the wafer. For example, in optical inspection, an array of parallel laser beams can scan the wafer surface along the strip. The strips are placed in adjacent parallel rows / columns to build an image of the wafer surface strip by strip. For example, the tool can scan the wafer from top to bottom along the strip, then switch to the next strip and scan the wafer from bottom to top, and so on, until the entire wafer is scanned and the inspection image of the wafer is collected.

[0039] In some cases, the review tool 121 can be configured to capture review images of at least some of the defect candidates detected by the inspection tool for determining whether the defect candidates are indeed defects of interest (DOI). Such review tools are typically configured to inspect segments of the sample one at a time (usually at a relatively low speed and / or high resolution). As an example, the review tool can be an electron beam tool such as, for example, a scanning electron microscope (SEM). An SEM is an electron microscope that produces an image of a sample by scanning the sample with a focused electron beam. The electrons interact with the atoms in the sample, generating various signals that contain information about the surface topography and / or composition of the sample. The SEM is capable of precisely inspecting and measuring features during semiconductor wafer manufacturing.

[0040] The inspection tool 120 and the review tool 121 can be different tools located at the same or different locations, or a single tool operating in two different modes. In some cases, the same inspection tool can provide low-resolution image data and high-resolution image data. The resulting image data (low-resolution image data and / or high-resolution image data) can be transmitted - directly or via one or more intermediate systems - to the system 101. The present disclosure is not limited to any specific type of inspection tool and / or the resolution of the image data generated by the inspection tool. In some cases, at least one of the inspection tools has metrology capabilities and can be configured to capture images and perform metrology operations on the captured images. Such an inspection tool is also referred to as a metrology tool.

[0041] According to certain embodiments of the presently disclosed subject matter, the inspection system 100 includes a computer-based system 101 that is operably connected to the inspection tool 120 and the review tool 121 and is capable of performing automatic defect offset correction between the inspection tool and the review tool. The system 101 is also referred to as a defect offset correction system.

[0042] The system 101 includes processing circuitry 102 that is operably connected to a hardware-based I / O interface 126 and is configured to provide the processing necessary for an operating system, as further detailed in Figures 2 to 3 reference. The processing circuitry 102 can include one or more processors (not shown separately) and one or more memories (not shown separately). One or more processors of the processing circuitry 102 can be configured to perform a number of functional modules individually or in any suitable combination according to computer-readable instructions implemented on a non-transitory computer-readable memory included in the processing circuitry. Such functional modules are hereinafter referred to as being included in the processing circuitry.

[0043] One or more processors as referred to herein may represent one or more general-purpose processing devices, such as a microprocessor, a central processing unit, 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.

[0044] The memory as referred to herein may include one or more of the following: internal memory (such as, for example, processor registers and caches, 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.).

[0045] According to certain embodiments, one or more functional modules included in the processing circuitry 102 of system 101 may include a machine learning (ML) module 106 (which was previously trained during the training / setup phase) and an offset calculation module 108 that are operatively connected to each other.

[0046] Specifically, the processing circuitry 102 may be configured to obtain a group of defect candidates from a defect map generated by inspecting a sample by an inspection tool via the I / O interface 126. The group of defect candidates is associated with corresponding inspection positions represented in an inspection coordinate system. The trained ML model 106 may be used to provide a probability that each defect candidate in the group is a defect of interest (DOI), and rank the group of defect candidates into an ordered list of defect candidates according to the corresponding probabilities of the defect candidates.

[0047] The ordered list of defect candidates may be sent to the review tool 121. The review tool 121 may review at least a portion of the ordered list of defect candidates according to the order of the ordered list of defect candidates in order to identify a predefined number of DOIs associated with corresponding review positions represented in a review coordinate system. After receiving the predefined number of DOIs identified from the portion of the ordered list from the review tool and associating them with the corresponding review positions represented in the review coordinate system, the offset calculation module 108 may be configured to calculate an offset between the review coordinate system and the inspection coordinate system based on the corresponding inspection positions and review positions associated with the predefined number of DOIs. The review tool may use the offset to review at least a portion of the group of defect candidates that have not been reviewed yet.

[0048] In some cases, the ML model 106 and the offset calculation module 108 can be regarded as part of a defect inspection recipe, which can be used to perform runtime defect inspection operations, particularly defect review operations, on the acquired runtime images of defect candidates.

[0049] In some cases, the system 101 can be configured as a training system capable of training an ML model using a specific training set during a training / setup phase. In such cases, one or more functional modules included in the processing circuitry 102 of the system 101 can include a training module 104 and an ML model 106 to be trained. Specifically, the training module 104 can be configured to obtain a training set and use the training set to train the ML model, which will be described in detail below with reference to Figure 3 As described above, after being trained, the ML model 106 can be used to rank a group of defect candidates into an ordered list of defect candidates according to the corresponding probabilities of the defect candidates at runtime.

[0050] According to certain embodiments, the ML model 106 can be implemented as various types of machine learning models. As an example, the ML model can be implemented as one of the following: various decision trees, regression models, neural networks, transformers, and / or an integration / combination of the above. The learning algorithm used by the ML model can be any one of the following: supervised learning, unsupervised learning, self-supervised, semi-supervised learning, or a combination of the above, etc. The presently disclosed subject matter is not limited to a specific type of ML model or a specific type of learning algorithm used by the ML model.

[0051] As an example, in some cases, the ML model can be implemented as a deep neural network (DNN). The DNN can include multiple layers organized according to a corresponding DNN architecture. As a non-limiting example, the layers of the DNN can be organized according to the architecture of a convolutional neural network (CNN), a recurrent neural network, a recursive neural network, a generative adversarial network (GAN), or otherwise. Optionally, at least some of the layers can be organized into multiple DNN sub-networks. Each layer of the DNN can include multiple basic computational elements (CEs), which are commonly referred to as dimensions, neurons, or nodes in the art.

[0052] The weights and / or thresholds associated with the CEs of the DNN and their connections 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 between the actual output generated by the DNN module and the target output associated with the corresponding training data set can be determined. The difference can be referred to as an error value. When the loss / cost function indicating the error value is less than a predetermined value, or when a limited change in performance between iterations is achieved, it can be determined that the training is complete. The input data set used to adjust the weights / thresholds of the DNN is referred to as the training set.

[0053] It should be noted that the teachings of the presently disclosed subject matter are not bound by the specific architecture of the ML model as described above.

[0054] It should be noted that although some embodiments of the present disclosure relate to a processing circuitry 102 configured to perform the operations described above, the functions / operations of the above functional modules can be performed in various ways by one or more processors in the processing circuitry 102. As an example, the operations of each functional module can be performed by a specific processor, or by a combination of processors. Thus, the operations of the various functional modules (such as the ML model and offset calculation, etc.) can be performed by the corresponding processor (or combination of processors) in the processing circuitry 102, and optionally, these operations can be performed by the same processor. The present disclosure should not be construed as being limited to a single processor that always performs all operations.

[0055] In some cases, in addition to the system 101, the inspection system 100 may further include one or more inspection modules, such as, for example, a defect detection module, an interference filtering module, an automatic defect review module (ADR), an automatic defect classification module (ADC), a metrology operation module, and / or other inspection modules that can be used to inspect semiconductor samples. One or more inspection modules can be implemented as stand-alone computers, or their functions (or at least some of them) can be integrated with the inspection tools 120 and 121. In some cases, the output of the system 101 (such as, for example, the calculated offset, an ordered list of defect candidates, and / or further defect review results) can be provided to one or more inspection modules (such as ADR, ADC, etc.) for further processing.

[0056] According to certain embodiments, the system 100 may include a storage unit 122. The storage unit 122 can be configured to store any data required by the operating system 101, such as, for example, data related to the input and output of the system 101, and intermediate processing results generated by the system 101. As an example, the storage unit 122 can be configured to store the sample images generated by the inspection tool 120 and / or their derivatives, such as, for example, the defect maps as described above, the inspection images and review images of defect candidates, and the training set. Thus, different types of required input data can be retrieved from the storage unit 122 and provided to the processing circuitry 102 for further processing. The output of the system 101 can be sent to the storage unit 122 for storage, such output as, for example, the calculated offset, an ordered list of defect candidates, and / or further defect review results.

[0057] In some embodiments, system 100 may optionally include a computer-based graphical user interface (GUI) 124 configured to implement user-specified inputs related to system 101. For example, a visual representation of a sample (e.g., via a display forming part of GUI 124) may be presented to the user, including an image of the sample, a defect map, etc. Options for defining certain operation parameters may be provided to the user via the GUI, such as, for example, the target number of DOIs required as anchor defects, the number of degrees of freedom for the transformation between the inspection coordinate system and the review coordinate system, etc. The user may also view operation results or intermediate processing results on the GUI, such as, for example, calculated offsets, an ordered list of defect candidates, and / or further defect review results, etc.

[0058] In some cases, system 101 may further be configured to send operation results to inspection tools 120 and 121 via I / O interface 126 for further processing. In some cases, system 101 may further be configured to send results to storage unit 122 and / or an external system (e.g., a yield management system (YMS) of a fabrication facility (fab)). A yield management system (YMS) in the context of semiconductor manufacturing is a data management, analysis, and tool system that collects data from a fabrication facility, especially during manufacturing acceleration, and helps engineers find ways to improve yield. YMS helps semiconductor manufacturers and fabrication facilities manage high-volume production analysis with fewer engineers. These systems analyze yield data and generate reports. YMS can be used by integrated device manufacturers (IDMs), fabrication facilities, fabless semiconductor companies, and outsourced semiconductor assembly and test (OSAT).

[0059] Those skilled in the art will readily understand that the teachings of the presently disclosed subject matter are not limited by Figure 1 the systems shown in Figure 1 Each system component and module in Figure 1 may consist of any combination of relevant software, hardware, and / or firmware executed on a suitable one or more devices that perform the functions defined and explained herein. Equivalent and / or modified functions as described for each system component and module may be combined or divided in another way. Thus, in some embodiments of the presently disclosed subject matter, the system may include fewer, more, modified, and / or different components, modules, and functions than

[0060] Figure 1Each component in can represent multiple specific components that are adapted to operate independently and / or cooperatively to process various data and electrical inputs and are used to implement operations related to a computerized inspection system. In some cases, multiple instances of a component may be utilized for reasons of performance, redundancy, and / or availability. Similarly, in some cases, multiple instances of a component may be utilized for reasons of functionality or application. For example, different parts of a specific function may be placed in different instances of the component.

[0061] It should be noted that Figure 1 the inspection system shown can be implemented in a distributed computing environment where Figure 1 one or more of the aforementioned components and functional modules shown may be distributed across several local and / or remote devices. As an example, inspection tool 120 and inspection tool 121 and system 101 may be located at the same entity (hosted by the same device in some cases) or distributed across different entities. As another example, as described above, in some cases, system 101 may be configured as a training system for training an ML model, while in some other cases, system 101 may be configured as a runtime defect offset correction system that uses the trained ML model. Depending on the specific system configuration and implementation requirements, the training system and the runtime offset correction system may be located at the same entity (hosted by the same device in some cases) or distributed across different entities.

[0062] In some examples, certain components are implemented using the cloud, for example, implemented in a private cloud or a public cloud. In cases where the various components of the inspection system are not all located in one location or one physical entity, communication between the various components can be achieved through any signaling system or communication components, modules, protocols, software languages, and drive signals, and can be wired and / or wireless, as the case may be.

[0063] It should also be noted that in some embodiments, at least some of inspection tool 120 and inspection tool 121, storage unit 122, and / or GUI 124 may be external to inspection system 100 and communicate data with systems 100 and 101 via I / O interface 126. System 101 may be implemented as a (multiple) standalone computer used in conjunction with inspection tools and / or additional inspection modules as described above. Alternatively, the various functions of system 101 may be at least partially integrated with one or more of inspection tool 120 and inspection tool 121, thereby facilitating and enhancing the functions of the inspection tools in inspection-related processes.

[0064] Although not necessarily so, the operating processes of system 101 and system 100 may correspond to some or all of the stages of the method described with respect to Figures 2 to 3 Similarly, with respect to Figures 2 to 3The described methods and their possible implementations can be implemented by system 101 and system 100. Therefore, it should be noted that the embodiments discussed in relation to Figures 2 to 3 the described methods can also be implemented, with necessary modifications, as various embodiments of system 101 and system 100, and vice versa.

[0065] Referring to Figure 2 , a generalized flowchart of automatic defect offset correction between an inspection tool and a review tool is shown in accordance with certain embodiments of the presently disclosed subject matter.

[0066] As described above, semiconductor samples are typically made up of multiple layers. The inspection process of the sample can be performed multiple times during the manufacturing process of the sample, such as after a processing step of a particular layer. In some cases, based on the known impact of the processing step on device characteristics or yield, a set of sampling processing steps can be selected for inline inspection. Images of the sample or portions thereof can be acquired at the set of sampling processing steps to be inspected.

[0067] For illustrative purposes only, the images of a given processing step / layer with respect to the set of sampling processing steps describe certain embodiments described below. Those skilled in the art will readily understand that the teachings of the presently disclosed subject matter, such as the process of automatic defect offset correction described below, can be performed after any layer and / or processing step of the sample. The present disclosure should not be limited to the number of layers included in the sample and / or the specific layer(s) to be inspected.

[0068] A defect map can be generated (202) that results from inspecting a semiconductor sample by an inspection tool (e.g., by inspection tool 120). The defect map indicates the distribution of defect candidates on the sample (e.g., suspected defects that are defects of interest (DOI) and their corresponding inspection locations on the sample). Specifically, the defect map typically can contain a large number of defect candidates, and in some cases, the number of defect candidates can reach on the order of tens of thousands to millions. In addition to the inspection locations, the defect map can also provide information on one or more inspection attributes associated with each given defect candidate and represent the defect characteristics of the given defect candidate, such as, for example, intensity, size, grade, volume, and polarity, etc.

[0069] A group of defect candidates can be obtained (204) from the defect map (e.g., by processing circuitry 102 of system 101). In some embodiments, the group of defect candidates can be selected from the large number of candidates in the defect map as the group of defect candidates that have a higher probability of being a DOI.

[0070] The selection can be performed in various ways. As an example, a defect classifier or an interference filter based on inspection attributes of defect candidates can be used to select a group of defect candidates. For example, the defect classifier can provide a probability score indicating the probability that the candidate is a DOI for each defect candidate, and classify the defect candidates as DOIs or interferences based on the probability scores of the defect candidates. The group of defect candidates can be selected based on the probability scores and / or classifications of the defect candidates. In some cases, the probability score of a given defect candidate provided in such a process can be added as one of the inspection attributes characterizing the given defect candidate.

[0071] In some embodiments, the selection of the group of defect candidates can be regarded as a preprocessing step performed before the currently disclosed defect offset correction process. Such preprocessing can reduce the number of potential defect candidates from the defect map to a lower magnitude, thus enabling the proposed defect offset correction process to work on a smaller group of candidates that are more likely to be DOIs. In some embodiments, the selection function of the group of defect candidates can be integrated as part of the inspection tool 120, or alternatively, it can be regarded as part of the system 101.

[0072] As described above, the selected group of defect candidates is associated with the corresponding inspection positions represented in the inspection coordinate system of the inspection tool.

[0073] To automatically identify the anchor defects (i.e., the defects that meet the defect offset correction / alignment purposes) from the group of defect candidates (instead of going through the trial and error of repeated defect selection as described above), the present disclosure proposes to use a (206) trained machine learning (ML) model (e.g., the ML module 106) to provide the probability that a defect candidate is a defect of interest (DOI) for each defect candidate in the group, and sort the group of defect candidates into an ordered list of defect candidates according to the corresponding probabilities of the defect candidates.

[0074] For the purpose of efficiently identifying a sufficient number of anchor defects, the ordered list of defect candidates can be sent from the system 101 to a review tool (e.g., the review tool 121) for review. The defects that meet the anchor defect criteria should be real defects (i.e., DOIs) in order to capture real defects during both inspection and review. Additionally, in some cases, such as in terms of the geometric characteristics of the anchor defects, the anchor defects can have a high level of correlation / applicability for offset correction, as will be described below.

[0075] A sufficient number of anchor defects need to be identified for aligning the multiple degrees of freedom of the transformation between the inspection coordinate system and the review coordinate system. In some cases, the number of required anchor defects (also referred to as the predefined number of DOIs or the target number of DOIs) can be predefined based on the number of degrees of freedom of the transformation between the two coordinate systems.

[0076] Thus, at least a portion of the defect candidate ordered list can be reviewed (208) by the review tool 121 according to the order of the candidates in the list to identify a predefined number of DOIs associated with corresponding review positions represented in the review coordinate system. Reviewing the defect candidates according to an order (the order being sorted based on the probability that a defect candidate is a DOI) enables the review tool to sample a relatively small number of defect candidates and efficiently locate the required predefined number of anchored defects as compared to traversing and validating each defect candidate in an unordered manner (e.g., randomly).

[0077] As described above, the ML model described with reference to the reference frame 206 is used to sort a group of defect candidates into a defect candidate ordered list according to the probability that a defect candidate is a DOI to assist in the selection of anchored defects. The ML model has been previously trained during a training phase. Figure 3 A generalized flowchart of training an ML model in accordance with certain embodiments of the present disclosure is shown.

[0078] A training set of defect candidates can be obtained (302) (e.g., by the training module 104 in the processing circuitry 102).

[0079] According to certain embodiments, the training set of defect candidates can be derived from inspecting one or more semiconductor samples that share the same design as the sample to be inspected at runtime. In some cases, the set of defect candidates can refer to the entire population of defect candidates revealed during inspection, while in some other cases, the training set can refer to a selected set of defect candidates selected by processing the entire population of defect candidates using a defect classifier or interference filter.

[0080] Each defect candidate in the training set can be characterized by one or more inspection attributes. The inspection attributes can be obtained by the inspection tool during the inspection process, e.g., based on certain characteristics of the (multiple) inspection images and / or (multiple) defect maps. As an example, during inspection, the inspection tool can capture inspection images of a sample (e.g., a wafer, die, or portion thereof). Various defect detection algorithms can be used to process the captured images of the sample to generate a defect map indicating the distribution of defect candidates on the sample. The generated defect map can provide information about inspection attributes such as, for example, the location, intensity, size, volume, grade, and polarity of the defect candidates. Optionally, in some cases, additional attributes can also be collected, including image characteristics corresponding to the defect candidates such as, for example, gray intensity, contrast, etc., and acquisition information such as acquisition time, acquisition tool ID, region ID, wafer ID, etc. As described above, in cases where the defect candidates have been previously classified / filtered using a defect classifier or interference filter, the probability score that each defect candidate is a DOI can be added to the training set as an additional inspection attribute.

[0081] Each defect candidate is associated with a corresponding ground truth (GT) label of the defect candidate, where the GT label indicates whether the defect candidate is a DOI or an interference. The GT label can be obtained in various ways, such as, for example, via manual annotation, from a review tool, etc. As an example, a review tool can be used to capture a review image with a higher resolution at the location of the defect candidate, and the review image is reviewed to determine whether the defect candidate is a DOI or an interference. The output of the review tool includes the defect category / type associated with the defect candidate respectively. The defect category / type of the candidates provided by the review tool can be regarded as the ground truth labels of these candidates.

[0082] The training set of defect candidates can be represented in different types of data representations. As an example, the training set of defect candidates and the inspection attributes associated with the defect candidates can be represented in tabular form, as illustrated in Figure 5 the example shown.

[0083] Figure 5 A schematic diagram showing an example of a training set according to certain embodiments of the presently disclosed subject matter is shown.

[0084] The training set 500 is illustrated as a tabular data set. The training set 500 includes N defect candidates stored in a table, where each row represents a specific defect candidate in the defect candidate set, and each column represents the inspection attributes of the defect candidate.

[0085] The inspection attributes obtained during defect inspection can include the location, intensity, size, volume, grade, polarity, etc. of the defect candidate. In this example, the probability score that each defect candidate is a DOI is added as an additional inspection attribute (represented as "TLF score" in the training set 500). The probability score can be previously derived by a defect classifier or an interference filter used for classifying / filtering defect candidates.

[0086] In addition to the inspection attributes, the training set also includes a column representing the GT label of the defect candidate. In this example, the defect candidate set can be reviewed by a review tool (such as, SEM). As shown, the training set 500 includes a column named "SEM review", which indicates the ground truth labels of the candidates provided by the SEM (where "true defect" represents a DOI, and "NVD" represents an interference).

[0087] The inspection attributes and the GT labels of the defect candidates derived from one or more samples constitute the training set 500. The training set 500 can be used to train the ML model 502, as will be detailed below.

[0088] It should be noted that although it is shown in tabular form as Figure 5The training set 500 shown is for exemplary purposes only and should not be considered as limiting the present disclosure. Any other suitable representation in lieu of the tabular format of such a data set can be used, including defect candidates and their attributes. For example, in some cases, any one of the following table-like structures can be used in lieu of the tabular format when appropriate: lists, charts, matrices, or general binary relations, etc. In some other cases, the training set of defect candidates can be represented in other types of data representations, such as, an image data set, e.g., image patches extracted from inspection images.

[0089] For example, for each defect candidate detected from the inspection image of a sample, an inspection patch including the defect candidate can be extracted from the inspection image. The inspection patch can be cropped from the inspection image according to a bounding box placed around the candidate. The inspection patch can be cropped in various sizes, such as, for example, 32×32 pixels, 64×64 pixels, or any other suitable size / dimension. In such cases, the training set can include a set of inspection patches corresponding to the set of defect candidates detected from the sample. It is also possible that the training set can include the inspection images of the training samples, where each defect candidate is labeled at the corresponding location of the defect candidate.

[0090] Return reference Figure 3 , after obtaining the training set, the ML model can be trained as follows: for each given defect candidate in the training set, the given defect candidate is processed (304) by the ML model to obtain the predicted class of the given defect candidate, and the ML model is optimized (306) using a loss function based on the predicted class associated with the given defect candidate and the GT label.

[0091] In particular, the ML model is designed and used for the specific purpose of sorting a group of defect candidates into such an order that when the defect candidates are reviewed according to the order, the review tool can identify a sufficient number of DOIs as efficiently / quickly as possible. As an example, assume that the group of defect candidates consists of N candidates, where M DOIs should be selected as anchor defects, then the ML model should be able to sort the N candidates into an order from 1 to N such that when the review tool reviews the candidates sequentially according to the order (e.g., starting from the candidate sorted as No.1 in the order), the review tool can sample fewer candidates and identify M DOIs as quickly as possible (i.e., with minimal trial and error).

[0092] That is to say, the proposed ML-based defect selection system focuses on selecting purity, which is the percentage of the number of actual DOIs among the defect candidates identified as DOIs by the ML model. In other words, the defect candidates ranked by the ML model as having a relatively high order / probability of being a DOI should be actually DOIs as "purely" as possible, so that M DOIs can be identified within fewer trials / samplings (e.g., the number of trials can be slightly larger than M but significantly smaller than N). In such cases, less attention is paid to whether the defect candidates ranked with a relatively low probability of being a DOI are correctly ranked (e.g., whether there may be some candidates that are actually DOIs but are somehow misranked with a lower order), because the model is not designed to capture all DOIs and thus has a higher capture rate.

[0093] Therefore, the proposed ML model aims to achieve a relatively high selection purity, which is different from the conventional defect classifiers or interference filters that usually aim to improve the capture rate or false alarm rate. In other words, the conventional defect classifiers try to capture as many DOIs as possible to avoid missing any real defects, thus affecting the yield, while the proposed ML model is used for a different purpose (i.e., anchoring defect selection). Therefore, the ML model is driven by a different motivation, that is, to improve the purity of the high-ranked candidates in order to sample fewer candidates and identify the required number of DOIs as efficiently as possible.

[0094] This difference may necessarily affect the specific training process of the ML model. The training process of the ML model can be configured in various ways to improve the selection purity. As an example, the loss function used to optimize the ML model can be configured based on the selection purity. Specifically, for example, the ML model can be customized to prioritize purity by adjusting the loss function (which is configured for binary classification). This can be achieved by amplifying the influence of the misclassified negative training samples (i.e., the training samples whose GT label is interference) within the loss function. Therefore, when a negative sample is misclassified as positive, a greater penalty can be imposed (e.g., by adjusting the weight of such negative samples in the loss function). This approach can ensure that the model is more conservative and only asserts positive predictions with high confidence, thereby enhancing the purity.

[0095] Additionally or as an alternative to the loss function configuration, the training set can also be specifically selected with respect to selection purity. As an example, when marking defect candidates in the training set, it can be decided to mark only those defect candidates that are DOIs and have a high level of correlation (e.g., applicability / suitability) for offset correction as DOIs. As an example, a defect candidate can be considered relevant / suitable for offset correction in terms of the geometric characteristics of the defect candidate, such as size and dimensions. For example, the size of a suitable defect should not be too large. A certain dimension of the defect, such as length or width, should not be too long. This is because large defects may occupy or exceed the FOV, making it difficult to measure the offset. In some cases, defect candidates that are DOIs in the training set can be further filtered based on the geometric characteristics of the defect candidates. For example, defect candidates with a size / dimension within a predetermined range can be selected. Additionally, the defect candidates in the training set should also be diverse in type. Different types / categories of DOIs should be included in the training set so that the ML model can learn to adapt to different types of defects.

[0096] Figure 7 A schematic diagram of the training process of an ML model according to certain embodiments of the presently disclosed subject matter is shown.

[0097] Figure 7 The ML model 700 is illustrated. The ML model 700 can be implemented in various types and architectures, such as, for example, decision trees (such as random forests), regression models (such as logistic regression models), and neural networks (such as CNNs), etc. To train the ML model, a training set of defect candidates is obtained. As described above, the training set can be represented in different types of data representations, such as, for example, the tabular data set and the image data set illustrated as Figure 5 in the example.

[0098] In the case where the training set is an image data set, the training set can include one or more training test patches, each training test patch containing a defect candidate marked as a DOI, such as, for example, the training patch 702 containing the DOI 704. The training set also includes one or more training test patches, each training test patch containing a defect candidate marked as interference or false alarm, such as, for example, the training patch 706 (where no DOI is marked).

[0099] The training set (e.g., the tabular data set or the image data set) can be fed into the ML model 700 for processing. The specific type of ML model to be implemented can be related to the type of input data. As an example, in the case where the training set is a tabular data set (such as Figure 5 the training set 500 illustrated as in the example), the ML model can be implemented as a decision tree (such as the random forest 502). In the case where the training set is an image data set, the ML model 700 can be implemented as a neural network, such as a DNN (e.g., a CNN).

[0100] As an example, a DNN can inherently extract representative features / attributes of training inspection patches and classify image patches based on the extracted features / attributes. Taking a CNN as an exemplary implementation of an ML model, during the forward pass, a convolution operation is performed on each training patch to learn to capture representative features. For each specific layer in the CNN, an output feature map can be generated. For example, by convolving each filter of a specific layer over the width and height of the input feature map and generating a two-dimensional activation map that gives the response of this filter at each spatial position. Stacking the activation maps of all filters along the depth dimension forms the complete output feature map of a specific layer, representing the extracted features / attributes of a given input training patch. After feature extraction through the convolutional layer, additional layers (such as fully connected layers and output layers) can convert the output feature map into probability scores for each class, and the class with the highest probability is selected as the predicted class for a given training inspection patch.

[0101] Accordingly, the ML model can provide a predicted class 708 based on this (in some cases, the predicted class can be associated with the prediction probability that the candidate belongs to the predicted class). The predicted class 708 can be evaluated using a loss function 710 (e.g., a classification loss, such as cross-entropy or squared hinge loss, etc.) with respect to the ground truth label 712 of the DOI 704. As described above, in some cases, the loss function can be specifically configured to attempt to improve the selection purity. The parameters of the ML model can be optimized to reduce / minimize the difference between the predicted class 708 and the ground truth label 712.

[0102] After being trained, the ML model 700 can be deployed at runtime and used to sort a group of defect candidates into an ordered list, where each defect candidate is associated with a ranking according to the order of the probability / likelihood that the defect candidate is a defect of interest (DOI), as described in detail in Figure 2 For example, if the group has N defect candidates, after being processed by the trained decision model, these N defect candidates will be sorted from 1 to N respectively, where each candidate has a unique ranking in the order.

[0103] Figure 6 FIG. shows a schematic diagram of the runtime deployment of an ML model according to certain embodiments of the present disclosure.

[0104] During runtime inspection, a group of defect candidates is generated by the inspection tool examining one or more samples. A group of defect candidates including N defect candidates is stored in Table 600, where each row represents a specific defect candidate in the group of defect candidates and each column represents an inspection attribute of the defect candidate. The group of defect candidates can be processed by a trained ML model, which provides a subsequence list of the defect candidates in the group, as shown in Table 602.

[0105] As shown, compared to Table 600, Table 602 adds a new column "True Defect Probability", in which each of the N defect candidates is associated with the probability that the defect candidate is a DOI. Thus, the defect candidates are sorted into a total order from 1 to N. The N defect candidates are sorted in descending order according to the column "True Defect Probability", such that the defect candidate with the highest likelihood of being a DOI is listed at the top. For example, a defect candidate with a 98% probability is ranked first, indicating that this candidate is the most likely to be a DOI among all candidates.

[0106] Continue Figure 2 As described above, at least a portion of the ordered list of defect candidates can be reviewed by the review tool according to the order of the candidates in the list. Compared to traversing and validating each defect candidate in an unordered manner (e.g., randomly), reviewing the defect candidates according to an order (which is sorted based on the probability that the defect candidate is a DOI) can enable the review tool to efficiently locate a predefined number of anchor defects. Figure 9 FIG. shows a graph demonstrating an improvement in the selection efficiency of anchor defects according to certain embodiments of the presently disclosed subject matter.

[0107] Figure 9 The curve in FIG. has an X-axis representing the number of defect candidates traversed / sampled and a Y-axis representing the probability of selecting a predefined number of DOIs (assuming the target number of DOIs is six in this example). Curve 902 shows such a selection scenario: without using the presently proposed ML-based defect selection system, the review tool randomly traverses / samples the defect candidates in the group. As shown, 50 defect candidates need to be reviewed by the review tool in order to achieve a probability of less than 60% of locating six DOIs. In contrast, when using the ML model specifically configured to assist in defect selection as described above, the number of defect candidates to be sampled is significantly reduced while achieving a higher probability of locating the target number of DOIs.

[0108] As an example, graph 904 shows such a selection scenario: an ML model implemented as a random forest is used to rank a group of defect candidates. In this case, the review tool only needs to sample 20 to 30 defect candidates according to the order, and has managed to achieve a relatively high probability (between 80% - 100%) of locating the target number of DOIs. Similarly, graph 906 shows such a selection scenario: the ML model is implemented as a different type of model, such as a logistic regression model, and the selection efficiency has been significantly improved in a similar manner. Therefore, the proposed ML-based correction can sample only a smaller number of defect candidates for efficiently locating the target number of anchor defects for alignment.

[0109] Specifically, when looking at the sampling of 30 defect candidates on the X-axis, the probability of locating the target number of DOIs increases from approximately 11% without using the proposed solution to approximately 95% when using the proposed ML-based solution.

[0110] Return reference Figure 2 , once the review tool identifies a predefined number of DOIs associated with the corresponding review positions represented in the review coordinate system, the predefined number of DOIs identified from a portion of the ordered list and their review positions can be sent back to system 101. After receiving the predefined number of DOIs associated with the corresponding review positions of the predefined number of DOIs from the review tool, an offset between the review coordinate system and the inspection coordinate system can be calculated (210) based on the corresponding inspection positions and review positions associated with the predefined number of DOIs (e.g., by the offset calculation module 108). The review tool can use the calculated offset to review at least a portion of the group of defect candidates that have not yet been reviewed.

[0111] The selected predefined number of DOIs are now associated with both the inspection positions (from the defect map) of the DOIs in the inspection coordinate system of the inspection tool and the review positions of the DOIs in the review coordinate system of the review tool. These DOIs can be used as anchor defects for aligning the two coordinate systems.

[0112] Figure 8 A schematic diagram of offset calculation according to certain embodiments of the presently disclosed subject matter is shown.

[0113] Sample graph 802 (e.g., a wafer map) is illustrated, in which the identified n anchor defects are marked therein (represented by circles, where n = 6 in this example). Each anchor defect is associated with the inspection position and the review position of the anchor defect. For each defect, the difference between these two positions is marked as a vector.

[0114] The offset between two coordinate systems can be calculated based on the inspection locations and review locations of n anchor defects. As an example, the offset can be derived by performing linear regression using two sets of locations. Linear regression can be used to fit a predictive model to an observed data set of values of two variables, such as the inspection locations and review locations of anchor defects. Once such a model is developed, when additional values of one variable (such as new defects having only inspection locations) are collected, the fitted model can be used to predict the values of the other variable (such as the review locations of new defects).

[0115] Graph 804 shows an example of a linear relationship 806 between the X-axis variable representing the review location of a defect and the Y-axis variable representing the inspection location of the defect. The linear relationship can be derived based on a linear regression model. The model can estimate the slope and intercept of the best-fit line that represents the linear relationship 806 between these two variables. The slope represents the change in the Y-axis variable for each unit change in the X-axis variable, and the intercept represents the predicted value of the Y-axis variable when the X-axis variable is zero.

[0116] As described above, the mismatch between two coordinate systems can be reflected in one or more degrees of freedom of the transformation between the two systems. Thus, the offset between two coordinate systems can be represented by a transformation matrix that includes a plurality of transformation coefficients corresponding to the multiple degrees of freedom of the transformation between the inspection coordinate system and the review coordinate system. As an example, the plurality of transformation coefficients can include two or more of the following: X offset, Y offset, rotation, perpendicularity, X scaling, Y scaling, X parabola, and Y parabola (where X parabola and Y parabola represent higher-order corrections (square terms)).

[0117] As an example, the transformation matrix can be represented by the following formula: where X and Y represent the inspection coordinates of a given defect, a0, b0, a1, b1, a2, b2, a3, and b3 represent the transformation coefficients corresponding to the multiple degrees of freedom of the transformation between the two coordinate systems, and X' and Y' represent the corrected coordinates, i.e., the expected review coordinates of the defect.

[0118] It should be noted that the specific examples of a - f and the transformation coefficients (such as X offset, Y offset, rotation, etc.) are listed only for illustrative purposes. Different transformation coefficients can also be used in addition to or in place of the above examples.

[0119] Once the offset (in the form of a transformation matrix) is obtained, the offset can be used to correct the locations of defect candidates in the group that have not yet been reviewed by the review tool, such that the review tool can review the un-reviewed defect candidates at the corrected locations.

[0120] It should be noted that in some cases, a predefined number of DOIs to be selected from the defect candidate group is associated with the number of transformation coefficients in the transformation matrix, and the number of transformation coefficients corresponds to the number of degrees of freedom of the transformation between the inspection coordinate system and the review coordinate system. As an example, in the case where there are eight degrees of freedom in the transformation between two coordinate systems (e.g., including X offset, Y offset, rotation, vertical, X scaling, Y scaling, X parabola, and Y parabola, as listed above), the transformation matrix includes eight transformation coefficients, and at least eight DOIs need to be selected from the defect candidate group to derive the transformation matrix, for example, based on linear regression.

[0121] Continue Figure 4 In Example 408 of, when the review tool navigates based on the inspection position of the defect candidate, that is, when determining the center of the FOV through the inspection position, the actual position of the defect deviates from the center of the FOV by a certain offset. By applying the calculated offset (represented by the transformation matrix) to the inspection position, the corrected position can be calculated, and when the FOV is centered at the corrected position, the review tool can find the defect at the center of the FOV, as Figure 8 shown in 808 of.

[0122] Similarly, for at least a portion of the defect candidate group that has not yet been reviewed, the offset can be used to obtain the review position of such defect candidates. At least a portion of the defect candidate group can be reviewed by the review tool at the review position.

[0123] By using the offset calculated using the above offset correction system, the throughput (TpT) of the review process for reviewing at least a portion of the un-reviewed defect candidates by the review tool can be significantly improved. Specifically, compared with the previous situation where defect offset correction was not performed, thus requiring the review tool to capture a relatively large FOV to ensure that the shifted defect would be captured, the currently proposed solution enables the review tool to capture a smaller FOV during the review, thereby reducing the image capture time and increasing the system TpT. Compared with the previous situation where defect offset correction was applied by randomly sampling defect candidates to identify a sufficient number of anchored defects, the currently proposed solution enables sampling fewer defect candidates while quickly and efficiently locating the required number of defects, thereby saving the system overhead of the trial-and-error process for searching for anchored defects.

[0124] Figure 10 Shows an example of TpT improvement for inspecting a sample according to certain embodiments of the currently disclosed subject matter.

[0125] The specific semiconductor sample being inspected has four layers. Table 1000 shows the TpT broken down into four factors including motion, image capture time, autofocus (AF), and automatic defect recognition (ADR). As shown, after using the proposed defect offset correction, the image capture times for all layers have been significantly improved (in particular, there is a 47% improvement in the capture time for layer 1). As described above, using the proposed offset correction enables the inspection tool to capture a smaller FOV, and the image capture time is typically proportional to the square of the FOV size. For example, when the FOV size is reduced by 30%, the image scan area of the tool can be reduced by 50%. The improvement in the image capture time results in a 10% - 20% improvement in the overall defects per hour (DPH, i.e., the number of defects that can be inspected in one hour).

[0126] It should be noted that the examples shown in this disclosure (such as exemplary images and defects, exemplary ML models, loss functions, training datasets, transformation matrices, etc.) are shown for illustrative purposes and should not be considered to limit this disclosure in any way. Other suitable examples / implementations can also be used in addition to or in place of the above examples.

[0127] One of the advantages of certain embodiments of the presently disclosed subject matter as described herein is to provide an automatic defect offset correction system that can correct the defect offset between two coordinate systems in an automatic manner and establish more accurate positions for all defect candidates without manual intervention. The proposed system can sample only a small number of defect candidates for efficiently locating qualified anchor defects for alignment, without suffering from the heavy overhead of trial and error in defect selection, thereby significantly enhancing the inspection efficiency.

[0128] A further advantage of certain embodiments of the presently disclosed subject matter as described herein is that, compared with other automatic defect offset correction methods (such as reference-based offset correction), the currently proposed solution does not require physical changes to the design or wafer. It provides a direct measurement of the defect offset between two coordinate systems (instead of an indirect measurement via a reference coordinate offset), thereby improving the accuracy level of the corrected defect positions.

[0129] One of the additional advantages of certain embodiments of the presently disclosed subject matter described herein is that by using the offsets calculated using the proposed offset correction system, the throughput (TpT) of the review process by the review tool can be significantly increased. Specifically, compared to the previous situation where defect offset correction was not performed and thus the review tool needed to capture a relatively large FOV to ensure that shifted defects would be captured, the presently proposed solution enables the review tool to capture a smaller FOV during review, thereby reducing the image capture time and increasing the system TpT. Compared to the previous situation where defect offset correction was applied by randomly sampling defect candidates to identify a sufficient number of anchored defects, the presently proposed solution enables sampling of fewer defect candidates to quickly and efficiently locate the required number of defects, thereby saving the system overhead for the trial-and-error process of searching for anchored defects.

[0130] One of the further advantages of certain embodiments of the presently disclosed subject matter described herein is that, compared to conventional defect classifiers or interference filters that typically aim to improve the capture rate or false alarm rate, the ML model used herein for ranking defect candidates and assisting in defect selection is specifically designed to achieve a relatively high selection purity. In other words, conventional defect classifiers attempt to capture as many DOIs as possible so as not to miss any real defects, thus affecting the yield, whereas the proposed ML model is used for a different purpose (i.e., anchored defect selection). Therefore, the proposed ML model is driven by a different motivation, namely, to improve the purity of the high-ranking candidates in order to sample fewer defect candidates and identify the required number of DOIs as quickly as possible.

[0131] This can be achieved through various configurations of the training process of the ML model to increase the selection purity. As an example, the loss function for optimizing the ML model can be configured based on the selection purity. Additionally or alternatively, the training set can be specifically selected with respect to the selection purity. The ML model trained in this way is capable of sampling a smaller number of defect candidates for locating the target number of anchored defects for alignment.

[0132] It should be understood that the present disclosure is not limited in its application to the details set forth in the description contained herein or shown in the drawings.

[0133] In this detailed description, numerous specific details are set forth to provide a thorough understanding of the present disclosure. However, those skilled in the art will understand that the presently disclosed subject matter may be practiced without these specific details. In other instances, well-known methods, procedures, components, and circuits have not been described in detail so as not to obscure the present disclosure.

[0134] Unless otherwise explicitly stated, as will be apparent from this discussion, it should be understood that throughout the specification discussion, the use of terms such as "obtaining", "inspecting", "examining", "processing", "using", "providing", "performing", "applying", "correcting", "sorting", "calculating", "reviewing", "training", "optimizing", "implementing", etc. refers to the (multiple) actions and / or (multiple) processes of a computer to manipulate and / or transform data into other data, where the data is represented as physical (such as electronic) quantities and / or the data represents physical objects. The term "computer" should be construed broadly to cover any kind of hardware-based electronic device having data processing capabilities, as non-limiting examples, including the inspection system disclosed in this application, the defect offset correction system, and various parts of the above systems.

[0135] The terms "non-transitory memory" and "non-transitory storage medium" as used herein should be construed broadly to cover any volatile or non-volatile computer memory suitable for the presently disclosed subject matter. 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 sets of instructions. These terms should also be understood to include any medium that is capable of storing or encoding a set of instructions for execution by a computer and that causes the computer to perform any one or more of the methods of the present disclosure. Thus, these terms should be understood to include, but not be limited to, read-only memory ("ROM"), random access memory ("RAM"), disk storage media, optical storage media, flash memory devices, etc.

[0136] The term "sample" as used in this specification should be construed broadly to cover any kind of physical object or substrate, including wafers, masks, reticles, and other structures, combinations, and / or parts thereof used in the manufacture of semiconductor integrated circuits, magnetic heads, flat panel displays, and other semiconductor manufactured articles. Samples are also referred to herein as semiconductor samples and can be produced by manufacturing equipment that performs the corresponding manufacturing processes.

[0137] The term "inspection" as used in this specification should be construed broadly to cover any kind of operation related to various types of defect detection, defect review, and / or defect classification, segmentation, and / or metrology operations during and / or after a sample manufacturing process. Non-destructive inspection tools are used during or after the manufacturing of the sample to be inspected to provide the inspection. As a non-limiting example, the inspection process can include runtime scans (in single or multiple scans), imaging, sampling, detection, review, measurement, classification, and / or other operations performed on the sample or a portion thereof using the same or different inspection tools. Similarly, inspection can be provided before manufacturing the sample to be inspected and can include, for example, generating (a) inspection recipe(s) and / or other setup operations. It should be noted that unless otherwise explicitly stated, the term "inspection" or its derivatives as used in this specification is not limited in terms of the resolution or size of the inspection area. As a non-limiting example, various non-destructive inspection tools include scanning electron microscopes (SEM), atomic force microscopes (AFM), optical inspection tools, etc.

[0138] The term "metrology operation" as used in this specification should be construed broadly to cover any metrology operation process for extracting metrology information related to one or more structural elements on a semiconductor sample. In some embodiments, the metrology operation can include measurement operations such as, for example, critical dimension (CD) measurements performed on certain structural elements on the sample, including but not limited to the following: dimensions (e.g., line width, line pitch, contact diameter, element size, edge roughness, gray-scale statistics, etc.), shape of the element, distance within or between elements, associated angles, overlay information associated with elements corresponding to different design levels, etc. For example, image processing techniques are employed to analyze measurement results such as measurement images. It should be noted that unless otherwise explicitly stated, the term "metrology" or its derivatives as used in this specification is not limited in terms of the measurement technique, measurement resolution, or size of the inspection area.

[0139] The term "defect" as used in this specification should be construed broadly to cover any kind of abnormal or undesired feature / function formed on the sample. In some cases, the defect may be a defect of interest (DOI), which is a real defect that has a certain impact on the function of the manufactured device, and thus detecting the DOI is in the customer's interest. For example, any "fatal" defect that may cause a yield loss can be indicated as a DOI. In some other cases, the defect may be an interference (also known as a "false alarm" defect), which can be ignored because the interference has no impact on the function of the completed device and does not affect the yield.

[0140] The term "defect candidate" as used in this specification should be construed broadly to cover a suspected defect location on a sample that is detected as having a relatively high probability of being a defect of interest (DOI). Thus, when being reviewed / tested, a DOI candidate may actually be a DOI, or in some other cases, a DOI candidate may be an interference, or random noise caused by different variations during inspection (e.g., process variations, color variations, mechanical and electrical variations, etc.).

[0141] The term "design data" as used in this specification should be construed broadly to cover any data that indicates the hierarchical physical design (layout) of a sample. The design data can be provided by the corresponding designer and / or can be derived from the physical design (e.g., through complex simulations, simple geometric and Boolean operations, etc.). The design data can be provided in different formats, and as non-limiting examples, for instance, GDSII format, OASIS format, etc. The design data can be presented in vector format, grayscale intensity image format, or other ways.

[0142] The term "(s)image" or "image data" as used in the specification should be construed broadly to cover any original image / frame of a sample captured by an inspection tool during a manufacturing process, derivatives of the captured image / frame obtained through various preprocessing stages, and / or computer-generated synthetic images (in some cases based on design data). Depending on the specific scanning mode (e.g., one-dimensional scanning such as line scanning, two-dimensional scanning in the x and y directions, or point scanning at specific points, etc.), the image data can be represented in different formats, such as, for example, as grayscale profiles, two-dimensional images, or discrete pixels, etc. It should be noted that in some cases, in addition to the image (e.g., the captured image, the processed image, etc.), the image data referred to herein can also include digital data associated with the image (e.g., metadata, manually crafted attributes, etc.). It should be further noted that the image or image data can include data related to a processing step / layer of interest or multiple processing steps / layers of a sample.

[0143] It should be understood that unless otherwise specifically stated, certain features of the presently disclosed subject matter described in the context of separate embodiments can also be provided in combination in a single embodiment. Conversely, the various features of the presently disclosed subject matter described in the context of a single embodiment can also be provided separately or in any suitable sub-combination. In this detailed description, numerous specific details are set forth in order to provide a thorough understanding of the methods and apparatuses.

[0144] It should also be understood that the systems according to the present disclosure can be implemented, at least in part, on a suitably programmed computer. Similarly, the present disclosure contemplates a computer program readable by a computer for performing the methods of the present disclosure. The present disclosure further contemplates a non-transitory computer-readable memory tangibly embodying a program of instructions executable by a computer for performing the methods of the present disclosure.

[0145] The present disclosure is capable of having other embodiments and of being practiced or carried out in various ways. Accordingly, it is to be understood that the phraseology and terminology employed herein are for the purpose of description and should not be regarded as limiting. Thus, those skilled in the art will appreciate that the concept upon which this disclosure is based can readily be utilized as a basis for designing other structures, methods, and systems for carrying out the several purposes of the presently disclosed subject matter.

[0146] Those skilled in the art will readily appreciate that various modifications and changes can be applied to the embodiments of the present disclosure described above without departing from the scope of the present disclosure as defined by the appended claims.

Claims

1. A computerized system for inspecting a semiconductor sample, the system comprising processing circuitry configured to: Obtain a group of defect candidates from a defect map generated by inspecting the sample by an inspection tool, the group of defect candidates being associated with respective inspection positions represented in an inspection coordinate system; Use a trained machine learning (ML) model to provide, for each defect candidate in the group, a probability that the defect candidate is a defect of interest (DOI), and rank the group of defect candidates into an ordered list of defect candidates according to the respective probabilities of the defect candidates; In response to a portion of the ordered list of defect candidates being reviewed by a review tool in the order of the ordered list of defect candidates, receive from the review tool a predefined number of DOIs identified from the portion of the ordered list and associated with respective review positions represented in a review coordinate system; And Calculate an offset between the review coordinate system and the inspection coordinate system based on the respective inspection positions and review positions associated with the predefined number of DOIs, wherein the offset can be used by the review tool to review at least a portion of the group of defect candidates that have not yet been reviewed.

2. The computerized system according to claim 1, wherein, The ML model was previously trained using a training set of defect candidates, each defect candidate being characterized by one or more inspection attributes and associated with a respective ground truth (GT) label indicating whether the defect candidate is a DOI or an interference.

3. The computerized system according to claim 2, wherein, The one or more inspection attributes characterizing each defect candidate include at least one of the following: rank, volume, polarity, intensity, size, and the probability that the defect candidate is a DOI.

4. The computerized system according to claim 2, wherein, The training of the ML model includes: for each given defect candidate in the training set, processing the given defect candidate through the ML model to obtain a predicted class of the given defect candidate, and optimizing the ML model using a loss function based on the predicted class and the GT label associated with the given defect candidate.

5. The computerized system according to claim 4, wherein, The loss function is configured based on selection purity, which represents the percentage of the number of actual DOIs among the number of defect candidates identified as DOIs.

6. The computerized system according to claim 2, wherein, Only defect candidates in the training set that are DOIs and have a relatively high correlation level for offset correction are labeled as DOIs.

7. The computerized system according to claim 1, wherein, The offset is represented by a transformation matrix, the transformation matrix including a plurality of transformation coefficients corresponding to a plurality of degrees of freedom of the transformation between the inspection coordinate system and the review coordinate system.

8. The computerized system according to claim 7, wherein, The plurality of transformation coefficients include two or more of the following: X offset, Y offset, rotation, perpendicularity, X scale and Y scale, X parabola and Y parabola.

9. The computerized system according to claim 7, wherein, The predefined number of selected DOIs is associated with the number of transformation coefficients in the transformation matrix.

10. The computerized system according to claim 1, wherein, The offset is calculated by performing linear regression based on the respective inspection positions and review positions associated with the predefined number of DOIs.

11. The computerized system according to claim 1, wherein, The processing circuitry is further configured to use the offset to obtain an inspection position of at least a portion of the defect candidates that have not been inspected, so that the inspection tool can inspect at least a portion of the defect candidates at the inspection position.

12. The computerized system according to claim 1, wherein, The inspection of at least a portion of the defect candidates that have not been inspected using the offset enables an increase in the throughput (TpT) of the inspection tool.

13. The computerized system according to claim 1, wherein, The ML model is one of a random forest, a logistic regression model, or a neural network.

14. A computerized method for inspecting a semiconductor sample, the method comprising: obtaining a group of defect candidates from a defect map generated by inspecting the semiconductor sample by an inspection tool, the group of defect candidates being associated with respective inspection positions represented in an inspection coordinate system; using a trained machine learning (ML) model to provide, for each defect candidate in the group, a probability that the defect candidate is a defect of interest (DOI), and sorting the group of defect candidates into an ordered list of defect candidates according to the respective probabilities of the defect candidates; in response to a portion of the ordered list of defect candidates being inspected by an inspection tool in the order of the ordered list of defect candidates, receiving from the inspection tool a predefined number of DOIs identified from the portion of the ordered list and associated with respective inspection positions represented in an inspection coordinate system; and calculating an offset between the inspection coordinate system and the inspection coordinate system based on the respective inspection positions and inspection positions associated with the predefined number of DOIs, wherein the offset can be used by the inspection tool to inspect at least a portion of the defect candidates that have not been inspected.

15. The computerized method according to claim 14, wherein, The ML model was previously trained using a training set of defect candidates, each defect candidate being characterized by one or more inspection attributes and associated with a respective ground truth (GT) label indicating whether the defect candidate is a DOI or an interference.

16. The computerized method according to claim 15, wherein, The training of the ML model includes: for each given defect candidate in the training set, processing the given defect candidate through the ML model to obtain a predicted class of the given defect candidate, and optimizing the ML model using a loss function based on the predicted class and the GT label associated with the given defect candidate.

17. The computerized method according to claim 16, wherein the loss function is configured based on selection purity, the selection purity representing the percentage of the number of actual DOIs in the number of defect candidates identified as DOIs.

18. The computerized method according to claim 14, wherein, The offset is represented by a transformation matrix, the transformation matrix including a plurality of transformation coefficients corresponding to a plurality of degrees of freedom of the transformation between the inspection coordinate system and the inspection coordinate system.

19. The computerized method according to claim 14, wherein, The inspection of at least a portion of the defect candidates that have not been inspected using the offset enables an increase in the throughput (TpT) of the inspection tool.

20. A non-transitory computer-readable storage medium, the non-transitory computer-readable storage medium tangibly embodying an instruction program that, when executed by a computer, causes the computer to execute a method for inspecting a semiconductor sample, the method comprising: Obtain a group of defect candidates from a defect map generated by inspecting the semiconductor sample with an inspection tool, the group of defect candidates being associated with respective inspection positions represented in an inspection coordinate system; Use a trained machine learning (ML) model to provide, for each defect candidate in the group, a probability that the defect candidate is a defect of interest (DOI), and rank the group of defect candidates into an ordered list of defect candidates according to the respective probabilities of the defect candidates; In response to a portion of the ordered list of defect candidates being reviewed by a review tool in the order of the ordered list of defect candidates, receive from the review tool a predefined number of DOIs identified from the portion of the ordered list and associated with respective review positions represented in a review coordinate system; And Calculate an offset between the review coordinate system and the inspection coordinate system based on the respective inspection positions and review positions associated with the predefined number of DOIs, wherein the offset can be used by the review tool to review at least a portion of the group of defect candidates that have not yet been reviewed.

Images