Defect location identification based on active learning

Abstract

The present invention provides a method and apparatus for optimally identifying the location to be inspected on a substrate in an inspection apparatus that uses multiple charged particle beams. [Solution] The defect location prediction model is trained using a training dataset associated with other substrates and process-related data associated with the substrate, with confidence scores associated with each prediction for each location, to generate predictions of defects or non-defects. Locations that the defect location prediction model determines have confidence scores that satisfy a confidence threshold are added to a set of locations to be inspected by the inspection system. After the inspection of a set of locations, inspection result data is acquired, and the defect location prediction model is trained stepwise using the inspection result data and process-related data for the set of locations as training data.

Description

Technical Field

[0001] Cross - reference to Related Applications

[0001] This application claims the priority of U.S. Patent Application No. 63 / 113347, filed on November 13, 2020, the entire disclosure of which is incorporated herein by reference.

[0002]

[0002] The embodiments provided herein relate to semiconductor manufacturing, and more particularly to the inspection of semiconductor substrates.

Background Art

[0003]

[0003] In the manufacturing process of integrated circuits (ICs), incomplete or complete circuit components are inspected to ensure that they are manufactured according to design and are free of defects. Inspection systems that utilize optical microscopes or charged - particle (e.g., electron) beam microscopes, such as scanning electron microscopes (SEM), can be employed. As the physical size of IC components continues to shrink, accuracy and yield in defect detection become more important.

[0004]

[0004] However, the imaging resolution and throughput of inspection tools are struggling to keep up with the miniaturization of IC component features. The accuracy, resolution, and throughput of such inspection tools can be limited by a lack of accuracy in detecting wafer displacement.

Summary of the Invention

[0005]

[0005] The embodiments provided herein disclose a particle - beam inspection apparatus, and more particularly, an inspection apparatus that uses a plurality of charged - particle beams.

[0006]

[0006] In some embodiments, a non-temporary computer-readable medium is provided which, when executed by a computer, causes the computer to execute a method for identifying locations on a substrate to be inspected. This method includes: selecting a plurality of locations on the substrate to be inspected based on a first submodel of a defect location prediction model trained with an initial training dataset associated with other substrates to generate a prediction of whether each location is defective or not; generating a confidence score for each location based on process-related data associated with the substrate, using a second submodel of the defect location prediction model trained with the initial training dataset, wherein the confidence score indicates the confidence of the prediction for the corresponding location; adding each of the locations whose confidence score satisfies one of a plurality of confidence thresholds to a set of locations to be inspected by the inspection system; acquiring inspection result data; and progressively training the defect location prediction model by providing the inspection result data and process-related data for a set of locations as training data to the defect location prediction model.

[0007]

[0007] In some embodiments, a non-temporary computer-readable medium is provided which, when executed by a computer, causes the computer to execute a method for identifying locations on a first substrate to be inspected using a machine learning model, and for training the machine learning model to identify locations on a second substrate to be inspected based on the inspection results of the locations on the first substrate. The method includes inputting process-related data associated with the substrate into a defect location prediction model; generating a defect or non-defect prediction for each of a plurality of locations on the substrate using the defect location prediction model, wherein each prediction is associated with a confidence score indicating the confidence of the prediction for the corresponding location; adding each of the locations whose confidence score satisfies one of a plurality of confidence thresholds to a set of locations to be inspected by the inspection system; obtaining inspection result data for the set of locations from the inspection system; and inputting the inspection result data and process-related data for the set of locations into the defect location prediction model to train the defect location prediction model.

[0008]

[0008] In some embodiments, methods are provided for using a machine learning model to identify locations on a first substrate to be inspected, and for training the machine learning model to identify locations on a second substrate to be inspected based on the inspection results of locations on the first substrate. The method includes inputting process-related data associated with the substrate into a defect location prediction model; generating a defect or non-defect prediction for each of a plurality of locations on the substrate using the defect location prediction model, wherein each prediction is associated with a confidence score indicating the confidence of the prediction for the corresponding location; adding each location whose confidence score satisfies a confidence threshold to a set of locations to be inspected by an inspection system; obtaining inspection result data for the set of locations from the inspection system; and inputting the inspection result data and process-related data for the set of locations into a defect location prediction model to train the defect location prediction model.

[0009]

[0009] In some embodiments, a device is provided for using a machine learning model to identify locations on a first substrate to be inspected, and for training the machine learning model to identify locations on a second substrate to be inspected based on the inspection results of locations on the first substrate. The device includes a memory for storing a set of instructions and at least one processor configured to execute a set of instructions for causing the device to perform a method, the method including inputting process-related data associated with a substrate into a defect location prediction model; using the defect location prediction model to generate a defect or non-defect prediction for each of a plurality of locations on the substrate, wherein each prediction is associated with a confidence score indicating the confidence of the prediction for the corresponding location; adding each of the locations whose confidence score satisfies a confidence threshold to a set of locations to be inspected by the inspection system; obtaining inspection result data for the set of locations from the inspection system; and inputting the inspection result data and process-related data for the set of locations into the defect location prediction model to train the defect location prediction model.

[0010]

[0010] In some embodiments, the non-temporary computer-readable medium stores a set of instructions that can be executed by at least one processor of the computing device to cause the computing device to carry out the method described above.

[0011]

[0011] Other advantages of embodiments of the present disclosure will become apparent from the following description, which will be taken in conjunction with the accompanying drawings. The following description will describe specific embodiments of the present invention, as examples and illustrative. [Brief explanation of the drawing]

[0012] [Figure 1]

[0012] This is a schematic diagram showing an exemplary electron beam inspection (EBI) system consistent with embodiments of the present disclosure. [Figure 2]

[0013] This is a schematic diagram showing an exemplary electron beam tool that may be part of the electron beam inspection system of Figure 1, consistent with embodiments of the present disclosure. [Figure 3]

[0014] This is a schematic diagram showing a semiconductor processing system consistent with embodiments of the present disclosure. [Figure 4]

[0015] This is a block diagram of a system for predicting defect locations on a substrate, consistent with embodiments of the present disclosure. [Figure 5A]

[0016] This is a block diagram for determining a confidence score using a random forest model, consistent with the embodiments of the present disclosure. [Figure 5B]

[0017] This is a block diagram for determining a confidence score using the QBC (query by committee) method, consistent with the embodiments of this disclosure. [Figure 6]

[0018] This block diagram shows the training of a defect location prediction tool using an initial training dataset, consistent with the embodiments of this disclosure. [Figure 7]

[0019] This is a flow diagram of a process for predicting the location of defects on a substrate, consistent with the embodiments of this disclosure. [Figure 8]

[0020] This is a block diagram showing a computer system that can assist in the implementation of a method, flow, module, component, or apparatus disclosed herein. [Modes for carrying out the invention]

[0013]

[0021] Electronic devices are constructed from circuits formed on a piece of silicon called a substrate. Many circuits can be formed together on the same piece of silicon and are called integrated circuits or ICs. The size of these circuits has been dramatically reduced, allowing even more circuits to fit onto a substrate. For example, a smartphone IC chip is about the size of a thumbnail, yet it can contain over 2 billion transistors, each transistor being less than 1 / 1000th the size of a human hair. Creating these extremely small ICs is a complex and expensive process that takes a great deal of time and often involves hundreds of individual steps. Even an error in a single step can result in a defect in the finished IC, rendering it useless. Therefore, one of the goals of the manufacturing process is to avoid such defects in order to maximize the number of functional ICs produced by the process, i.e., to improve the overall yield of the process.

[0014]

[0022] One factor in improving yield is monitoring the chip fabrication process to ensure the production of a sufficient number of functional integrated circuits. One way to monitor the process is to inspect the chip circuit structure at various stages of its formation. This inspection can be done using a scanning electron microscope (SEM). SEMs can be used to image these extremely small structures, essentially taking a "picture" of them. The images can be used to determine whether the structure was formed correctly and whether it was formed in the correct location. If there are defects in the structure, the process can be adjusted to reduce the likelihood of the defects recurring.

[0015]

[0023] Inspecting circuit boards is a resource-intensive process, and inspecting every location on a board can consume a significant amount of both computational resources and time. For example, it may take several days to inspect an entire board. One way to make the inspection process more efficient (e.g., minimizing resource consumption) is to identify locations on the board that are likely to be defective and inspect only those identified locations, rather than all locations. For example, conventional methods used machine learning (ML) models to predict locations that were likely to be defective. Conventional methods determine whether or not a particular location on the board is defective. However, conventional methods have drawbacks. For example, some of these methods are inaccurate; for instance, they may miss defective locations or identify locations that are not defective as defective. Because the predictions are inaccurate, the inspection system may fail to inspect such defective locations, potentially resulting in a defective finished IC. In another example, such conventional methods are not self-correcting. In other words, if a method predicts that a defect exists at a specific location on a particular substrate, it will continue to predict that similar locations on subsequent substrates being inspected will also be defective, regardless of whether defects actually exist at those locations, rendering the inspection process useless or ineffective.

[0016]

[0024] Embodiments of this disclosure describe an inspection method that assigns confidence scores indicating the reliability of defect predictions to each location on a substrate and selects all locations having confidence scores that satisfy a confidence threshold for inspection. For example, a first prediction model may predict that a particular location is defect-free, and a second prediction model may determine a confidence score for the particular location that indicates low confidence in the prediction (e.g., a confidence score below a certain confidence threshold). By selecting locations with low confidence scores, the embodiments can avoid overlooking defect locations for inspection (or overlook fewer than conventional methods). The inspection methods of the disclosed embodiments are also self-healing. After locations with low confidence scores are inspected by an inspection system (e.g., SEM), inspection result data obtained from the inspection system (e.g., SEM images of the inspection locations, information on whether there are defects at a certain location or whether the inspection is not based on actual inspection) is fed back to the prediction models to adjust the predictions for those locations. By inputting actual inspection results from locations with low confidence scores into the predictive model, the predictive model is further trained to predict with greater accuracy the likelihood of defects at such locations on substrates to be inspected later. By progressively training the predictive model with inspection results from all substrates to be inspected later, the predictive model begins to generate predictions for such locations with higher confidence scores, minimizing the number of locations to be inspected and thereby improving yield.

[0017]

[0025] Here, exemplary embodiments will be described in detail, and examples of the exemplary embodiments are shown in the accompanying drawings. In the following description, reference is made to the accompanying drawings, in which the same numbers in different drawings represent the same or similar elements unless otherwise indicated. The implementations described in the following description of the exemplary embodiments do not represent all implementations. Rather, they are merely examples of apparatuses and methods that coincide with aspects related to the disclosed embodiments recited in the appended claims. For example, although some embodiments are described in the context of using an electron beam, the present disclosure is not limited thereto. Other types of charged particle beams may be similarly applied. Furthermore, other imaging systems such as optical imaging, light detection, X-ray detection, etc. may also be used.

[0018]

[0026] Although the present document may specifically refer to the manufacture of ICs, it should be explicitly understood that the description herein has many other possible applications. For example, the description herein may be used in the manufacture of integrated optical systems, guidance patterns and detection patterns for magnetic domain memories, liquid crystal display panels, thin film magnetic heads, etc. Those skilled in the art will understand that in the context of such alternative applications, the use of the terms "reticle", "wafer", or "die" in this document should be considered interchangeable with the more general terms "mask", "substrate", and "target portion", respectively.

[0019]

[0027] In this document, the terms "radiation" and "beam" are used to encompass all types of electromagnetic radiation, including ultraviolet radiation (e.g., having wavelengths of 365, 248, 193, 157, or 126 nm) and EUV (extreme ultraviolet radiation, e.g., having wavelengths in the range of 5 - 20 nm).

[0020]

[0028] Referring now to FIG. 1, FIG. 1 shows an exemplary electron beam inspection (EBI) system 100 that is consistent with an embodiment of the present disclosure. As shown in FIG. 1, the charged particle beam inspection system 100 includes a main chamber 10, a load lock chamber 20, an electron beam tool 40, and an equipment front end module (EFEM) 30. The electron beam tool 40 is located within the main chamber 10. Although the description and drawings are directed to electron beams, it is understood that the embodiments are not used to limit the present disclosure to specific charged particles.

[0021]

[0029] The EFEM 30 includes a first load port 30a and a second load port 30b. The EFEM 30 may include additional load ports. The first load port 30a and the second load port 30b receive a wafer front opening unified pod (FOUP) containing a wafer to be inspected (e.g., a semiconductor wafer or a wafer made of other materials) or a sample (hereinafter, wafers and samples are collectively referred to as "wafers"). One or more robot arms (not shown) of the EFEM 30 transfer the wafer to the load lock chamber 20.

[0022]

[0030] The load lock chamber 20 is connected to a load / lock vacuum pump system (not shown), and the load / lock vacuum pump system removes gas molecules in the load lock chamber 20 to reach a first pressure below atmospheric pressure. After reaching the first pressure, one or more robot arms (not shown) transfer the wafer from the load lock chamber 20 to the main chamber 10. The main chamber 10 is connected to a main chamber vacuum pump system (not shown), and the main chamber vacuum pump system removes gas molecules in the main chamber 10 to reach a second pressure below the first pressure. After reaching the second pressure, the wafer is inspected by the electron beam tool 40. In some embodiments, the electron beam tool 40 may include a single beam inspection tool. In other embodiments, the electron beam tool 40 may include a multi-beam inspection tool.

[0023]

[0031] The controller 50 can be electronically connected to the electron beam tool 40 and also to other components. The controller 50 may be a computer configured to perform various controls of the charged particle beam inspection system 100. The controller 50 may also include processing circuits configured to perform various signal and image processing functions. In Figure 1, the controller 50 is shown as an external component of the structure including the main chamber 10, the loading and locking chamber 20, and the EFEM 30, but it is understood that the controller 50 may also be part of the structure.

[0024]

[0032] This disclosure provides an example of a main chamber 10 for housing an electron beam tool, but it should be noted that aspects of this disclosure are not limited in a broad sense to chambers housing electron beam tools. Rather, it is understood that the principles described above can be applied to other chambers as well.

[0025]

[0033] Referring here to Figure 2, Figure 2 shows a schematic diagram of an exemplary electron beam tool 40, which may be part of the exemplary charged particle beam inspection system 100 of Figure 1, consistent with embodiments of the present disclosure. The electron beam tool 40 (also referred to herein as apparatus 40) includes an electron source 101, a gun aperture plate 171 having a gun aperture 103, a pre-beamlet forming mechanism 172, a condenser lens 110, a source conversion unit 120, a primary projection optical system 130, a sample stage (not shown in Figure 2), a secondary imaging system 150, and an electron detection device 140. The primary projection optical system 130 may include an objective lens 131. The electron detection device 140 may include a plurality of detection elements 140_1, 140_2, and 140_3. A beam separator 160 and a deflection scanning unit 132 may be located within the primary projection optical system 130. It can be understood that other commonly known components of apparatus 40 may be added / omit as needed.

[0026]

[0034] The electron source 101, gun aperture plate 171, condenser lens 110, supply source conversion unit 120, beam separator 160, deflection scanning unit 132, and primary projection optical system 130 can be aligned with the principal optical axis 100_1 of the device 100. The secondary imaging system 150 and electron detection device 140 can be aligned with the secondary optical axis 150_1 of the device 40.

[0027]

[0035] The electron source 101 may include a cathode, an extractor, or an anode, and primary electrons can be emitted from the cathode and then extracted or accelerated to form a primary electron beam 102 that forms a primary beam crossover (virtual or real image) 101s. The primary electron beam 102 can be visualized once it is emitted from the crossover 101s.

[0028]

[0036] The source conversion unit 120 may include an image forming element array (not shown in Figure 2), an aberration compensator array (not shown), a beam limiting aperture array (not shown), and a pre-bending micro-deflector array (not shown). The image forming element array may include multiple micro-deflectors or microlenses to form multiple parallel images (virtual or real images) of the crossover 101s using multiple beamlets of the primary electron beam 102. Figure 2 shows three beamlets 102_1, 102_2, and 102_3 as an example, and it is understood that the source conversion unit 120 can handle any number of beamlets.

[0029]

[0037] In some embodiments, the source conversion unit 120 may comprise a beam limiting aperture array and an image forming element array (neither shown). The beam limiting aperture array may include beam limiting apertures. It is understood that any number of apertures may be used as needed. The beam limiting apertures may be configured to limit the size of the beamlets 102_1, 102_2, and 102_3 of the primary electron beam 102. The image forming element array may include image forming deflectors (not shown) configured to deflect the beamlets 102_1, 102_2, and 102_3 by changing their angle toward the principal optical axis 100_1. In some embodiments, the deflectors can deflect the beamlets more significantly the further they are from the principal optical axis 100_1. Furthermore, the image forming element array may comprise multiple layers (not shown), and the deflectors may be located in separate layers. The deflectors may be configured to be controlled independently of each other. In some embodiments, the deflector may be controlled to adjust the pitch of probe spots (e.g., 102_1S, 102_2S, and 102_3S) formed on the surface of sample 1. Where referenced herein, the probe spot pitch may be defined as the distance between two directly adjacent probe spots on the surface of sample 1.

[0030]

[0038] A central deflector located in the image-forming element array can be aligned with the principal optical axis 100_1 of the electron beam tool 40. Therefore, in some embodiments, the central deflector can be configured to maintain a straight trajectory for the beamlet 102_1. In some embodiments, the central deflector can be omitted. However, in some embodiments, the primary electron source 101 does not necessarily have to be aligned with the center of the source conversion unit 120. Furthermore, although Figure 2 shows a side view of the apparatus 40 with the beamlet 102_1 on the principal optical axis 100_1, it is understood that the beamlet 102_1 can be off-axis when viewed from a different side. That is, in some embodiments, all of the beamlets 102_1, 102_2, and 102_3 may be off-axis. Off-axis components can be offset relative to the principal optical axis 100_1.

[0031]

[0039] The deflection angle of the deflected beamlet can be set based on one or more criteria. In some embodiments, the deflector can deflect the off-axis beamlet radially outward or away from the principal optical axis 100_1 (not illustrated). In some embodiments, the deflector can be configured to deflect the off-axis beamlet radially inward or toward the principal optical axis 100_1. The deflection angle of the beamlet can be set so that the beamlets 102_1, 102_2, and 102_3 land perpendicularly on the sample 1. Off-axis aberrations of the image caused by lenses such as the objective lens 131 can be reduced by adjusting the path of the beamlet through the lens. Thus, the deflection angles of the off-axis beamlets 102_2 and 102_3 can be set so that the probe spots 102_2S and 102_3S have small aberrations. To reduce aberrations in the off-axis probe spots 102_2S and 102_3S, the beamlets may be deflected to pass through or near the front focal point of the objective lens 131. In some embodiments, the deflector may be configured so that the beamlets 102_1, 102_2, and 102_3 land perpendicularly on the sample 1, while the probe spots 102_1S, 102_2S, and 102_3S have small aberrations.

[0032]

[0040] The condenser lens 110 is configured to focus the primary electron beam 102. The currents in the beamlets 102_1, 102_2, and 102_3 downstream of the source conversion unit 120 can be varied by adjusting the focusing force of the condenser lens 110 or by changing the radial size of the corresponding beam limiting apertures in the beam limiting aperture array. The currents can be varied by both changing the radial size of the beam limiting apertures and the focusing force of the condenser lens 110. The condenser lens 110 may be an adjustable condenser lens, which can be configured such that its first principal plane is movable. The adjustable condenser lens may be configured to be magnetic, so that the off-axis beamlets 102_2 and 102_3 can irradiate the source conversion unit 120 with a rotation angle. The rotation angle may vary depending on the focusing force or the position of the first principal plane of the adjustable condenser lens. Accordingly, the condenser lens 110 may be an anti-rotation condenser lens configured to maintain a constant rotation angle while the focusing force of the condenser lens 110 is changed. In some embodiments, the condenser lens 110 may be an adjustable anti-rotation condenser lens in which the rotation angle remains unchanged when the focusing force of the condenser lens 110 and the position of the first principal plane change.

[0033]

[0041] The electron beam tool 40 may include a pre-beamlet forming mechanism 172. In some embodiments, the electron source 101 may be configured to emit primary electrons and form a primary electron beam 102. In some embodiments, the gun aperture plate 171 may be configured to block electrons around the primary electron beam 102 to reduce the Coulomb effect. In some embodiments, the pre-beamlet forming mechanism 172 may further cut off electrons around the primary electron beam 102 to further reduce the Coulomb effect. After passing through the pre-beamlet forming mechanism 172, the primary electron beam 102 may be trimmed into three primary electron beamlets 102_1, 102_2, and 102_3 (or any other number of beamlets). The electron source 101, the gun aperture plate 171, the pre-beamlet forming mechanism 172, and the condenser lens 110 may be aligned with the principal optical axis 100_1 of the electron beam tool 40.

[0034]

[0042] The pre-beamlet formation mechanism 172 may include a Coulomb aperture array. The central aperture of the pre-beamlet formation mechanism 172, also referred to herein as the on-axis aperture, and the central deflector of the source conversion unit 120 can be aligned with the principal optical axis 100_1 of the electron beam tool 40. The pre-beamlet formation mechanism 172 may comprise a plurality of pre-trimming apertures (e.g., a Coulomb aperture array). In Figure 2, the three beamlets 102_1, 102_2, and 102_3 are generated when the primary electron beam 102 passes through the three pre-trimming apertures, blocking most of the remaining primary electron beam 102. That is, the pre-beamlet formation mechanism 172 can trim many or most of the electrons from the primary electron beam 102, which does not form the three beamlets 102_1, 102_2, and 102_3. The pre-beamlet formation mechanism 172 can block electrons that are not ultimately used to form probe spots 102_1S, 102_2S, and 102_3S before the primary electron beam 102 enters the source conversion unit 120. In some embodiments, a gun aperture plate 171 is provided near the electron source 101 to block electrons early, while the pre-beamlet formation mechanism 172 is also provided to further block electrons in the vicinity of multiple beamlets. Figure 2 shows three apertures of the pre-beamlet formation mechanism 172, but it is understood that there can be any number of apertures as needed.

[0035]

[0043] In some embodiments, the pre-beamlet forming mechanism 172 may be located below the condenser lens 110. By placing the pre-beamlet forming mechanism 172 closer to the electron source 101, the Coulomb effect can be reduced more effectively. In some embodiments, the gun aperture plate 171 can be omitted if the pre-beamlet forming mechanism 172 can be placed close enough to the electron source 101 and is still manufacturable.

[0036]

[0044] The objective lens 131 may be configured to focus beamlets 102_1, 102_2, and 102_3 onto sample 1 for inspection, thereby forming three probe spots 102_1s, 102_2s, and 102_3s on the surface of sample 1. To reduce the Coulomb interaction effect, the gun aperture plate 171 can block peripheral electrons of the unused primary electron beam 102. The Coulomb interaction effect can enlarge the size of each of the probe spots 102_1s, 102_2s, and 102_3s, and thus degrade the inspection resolution.

[0037]

[0045] The beam separator 160 may be a Wien filter type beam separator that includes an electrostatic deflector that generates an electrostatic dipole field E1 and a magnetic dipole field B1 (neither of which are shown in Figure 2). When these fields are applied, the force exerted by the electrostatic dipole field E1 on the electrons of beamlets 102_1, 102_2, and 102_3 is equal in magnitude and opposite in direction to the force exerted by the magnetic dipole field B1 on the electrons. Therefore, beamlets 102_1, 102_2, and 102_3 can pass through the beam separator 160 in a straight line with zero deflection angle.

[0038]

[0046] The deflection scanning unit 132 can deflect the beamlets 102_1, 102_2, and 102_3 to scan probe spots 102_1s, 102_2s, and 102_3s across three small scanning areas in a section of the surface of sample 1. In response to the incidence of beamlets 102_1, 102_2, and 102_3 at probe spots 102_1s, 102_2s, and 102_3s, three secondary electron beams 102_1se, 102_2se, and 102_3se may be emitted from sample 1. Each of the secondary electron beams 102_1se, 102_2se, and 102_3se may contain electrons with an energy distribution including secondary electrons (energy ≤ 50 eV) and backscattered electrons (energy between 50 eV and the landing energies of beamlets 102_1, 102_2, and 102_3). The beam separator 160 can guide the secondary electron beams 102_1se, 102_2se, and 102_3se toward the secondary imaging system 150. The secondary imaging system 150 can focus the secondary electron beams 102_1se, 102_2se, and 102_3se onto the detection elements 140_1, 140_2, and 140_3 of the electron detection device 140. The detection elements 140_1, 140_2, and 140_3 can detect the corresponding secondary electron beams 102_1se, 102_2se, and 102_3se and generate corresponding signals to construct, for example, an image of the corresponding scanning area of ​​sample 1.

[0039]

[0047] In Figure 2, the three secondary electron beams 102_1se, 102_2se, and 102_3se, generated by the three probe spots 102_1S, 102_2S, and 102_3S respectively, move upward along the principal optical axis 100_1 toward the electron source 101 and successively pass through the objective lens 131 and the deflection scanning unit 132. The three secondary electron beams 102_1se, 102_2se, and 102_3se are redirected by a beam separator 160 (such as a Wien filter) so that they enter the secondary imaging system 150 along the secondary optical axis 150_1. The secondary imaging system 150 focuses the three secondary electron beams 102_1se to 102_3se onto an electron detection device 140 which includes three detection elements 140_1, 140_2, and 140_3. Therefore, the electron detection device 140 can simultaneously generate images of three scanning regions scanned by three probe spots 102_1S, 102_2S, and 102_3S, respectively. In some embodiments, the electron detection device 140 and the secondary imaging system 150 form a detection unit (not shown). In some embodiments, electron-optical elements along the path of the secondary electron beam, such as (but not limited to) the objective lens 131, the deflection scanning unit 132, the beam separator 160, the secondary imaging system 150, and the electron detection device 140, can form a detection system.

[0040]

[0048] In some embodiments, the controller 50 may include an image processing system, which includes an image acquirer (not shown) and storage (not shown). The image acquirer may include one or more processors. For example, the image acquirer may include a computer, server, mainframe host, terminal, personal computer, any kind of mobile computing device and the like, or a combination thereof. The image acquirer may be communicably coupled to the electronic detection device 140 of the apparatus 40 through a medium such as a conductor, fiber optic cable, portable storage medium, IR, Bluetooth, the Internet, wireless network, wireless radio, or a combination thereof. In some embodiments, the image acquirer may receive signals from the electronic detection device 140 and construct an image. Thus, the image acquirer may acquire an image of sample 1. The image acquirer may also perform various post-processing functions such as contour generation, superimposition of indicators onto the acquired image and the like. The image acquirer may be configured to perform adjustments such as brightness and contrast of the acquired image. In some embodiments, the storage may be a storage medium such as a hard disk, flash drive, cloud storage, random access memory (RAM), other types of computer-readable memory, and the like. The storage can be used in conjunction with an image acquirer to store scanned raw image data as the original image and to store the post-processed image.

[0041]

[0049] In some embodiments, the image acquirer can acquire one or more images of a sample based on one or more imaging signals received from the electronic detection device 140. The imaging signals may correspond to a scanning operation for performing charged particle imaging. The acquired image may be a single image containing multiple imaging areas, or it may consist of multiple images. A single image can be stored in storage. A single image may be an original image that can be divided into multiple regions. Each region may contain one imaging area containing features of sample 1. The acquired image may consist of multiple images of a single imaging area of ​​sample 1 sampled multiple times over time, or it may consist of multiple images of different imaging areas of sample 1. Multiple images can be stored in storage. In some embodiments, the controller 50 may be configured to perform image processing steps using multiple images of the same location of sample 1.

[0042]

[0050] In some embodiments, the controller 50 may include a measurement circuit (e.g., an analog-to-digital converter) to obtain the distribution of detected secondary electrons. The electron distribution data collected during the detection time window can be used in combination with the corresponding scanning path data of the primary beamlets 102_1, 102_2, and 102_3 incident on the wafer surface to reconstruct an image of the wafer structure under inspection. The reconstructed image can be used to reveal various features of the internal or external structure of sample 1, and thus can be used to reveal any defects that may be present in the wafer.

[0043]

[0051] In some embodiments, the controller 50 can control an electric stage (not shown) to move the sample 1 during inspection. In some embodiments, the controller 50 can cause the electric stage to continuously move the sample 1 in a certain direction at a constant speed. In other embodiments, the controller 50 can cause the electric stage to change the speed at which the sample 1 moves over time according to the steps of the scanning process. In some embodiments, the controller 50 can adjust the configuration of the primary projection optical system 130 or the secondary imaging system 150 based on images of the secondary electron beams 102_1se, 102_2se, and 102_3se.

[0044]

[0052] Figure 2 shows that the electron beam tool 40 uses three primary electron beams, but it is understood that the electron beam tool 40 may use two or more primary electron beams. This disclosure does not limit the number of primary electron beams used in the device 40.

[0045]

[0053] Here, we refer to Figure 3, a schematic diagram showing a semiconductor processing system. Figure 3 shows a conventional semiconductor processing system 300 having a scanner 305, a developing tool 320, an etching tool 325, an ash tool 330, a monitoring tool 335, a point determination tool 345, and a verification unit 350. The scanner 305 may include a control unit 310. The semiconductor processing system 300 can assist in computer-guided inspection of a substrate, as described below.

[0046]

[0054] The scanner 305 can expose a photoresist-coated substrate to a circuit pattern to be transferred to the substrate. The control unit 310 can control the exposure recipe used to expose the substrate. The control unit 310 can adjust various exposure recipe parameters, such as exposure time, radiation source intensity, and exposure dose. A high-density focus map (HDFM) 315 can be recorded in response to the exposure.

[0047]

[0055] The developing tool 320 can develop a pattern on an exposed substrate by removing photoresist from unwanted areas. In the case of positive-type photoresist, the portion of the photoresist exposed to light by the scanner 305 becomes soluble in the photoresist developer, while the unexposed portion of the photoresist remains insoluble in the photoresist developer. In the case of negative-type photoresist, the portion of the photoresist exposed to light by the scanner 305 becomes insoluble in the photoresist developer, while the unexposed portion of the photoresist remains soluble in the photoresist developer.

[0048]

[0056] The etching tool 325 can transfer a pattern to one or more films beneath the photoresist by etching the film from the portion of the substrate from which the photoresist has been removed. The etching tool 325 may be a dry etching tool or a wet etching tool.

[0049]

[0057] The Ash Tool 330 can remove any remaining photoresist from the etched substrate, allowing the pattern transfer process to the film on the substrate to be completed.

[0050]

[0058] The monitoring tool 335 can inspect the processed substrate at one or more locations on the substrate to generate monitoring results. The monitoring results may be based on spatial pattern determination, size measurement of different pattern features, or positional shifts at different pattern features. The inspection locations can be determined by the point determination tool 345. In some embodiments, the monitoring tool may be part of the EBI system 100 in Figure 1, or it may be the electron beam tool 40.

[0051]

[0059] The point determination tool 345 may include one or more predictive models for determining inspection locations on the substrate based on the HDFM 315 and the weak point information 340. In some embodiments, the point determination tool 345 can generate predictions for each location on the substrate that predict the likelihood that the location is defective (or not defective). For example, the point determination tool 345 may assign a probability value to each location that indicates the probability that the location is defective (or not defective).

[0052]

[0060] The weakness information 340 may include information about locations where problems related to the patterning process are highly likely. The weakness information 340 may be based on various process parameters and characteristics of the transfer pattern, wafer, scanner 305, or etching tool 325.

[0053]

[0061] The verification unit 350 may compare monitoring results from the monitoring tool 335 with corresponding design parameters to generate verification results. The verification unit 350 may provide the verification results to the control unit 310 of the scanner 305. The control unit 310 may adjust the exposure recipe for subsequent substrates based on the verification results. For example, the control unit 310 may reduce the exposure dose of the scanner 305 for several locations on the subsequent substrate based on the verification results.

[0054]

[0062] While the above description states that the semiconductor processing system 300 includes a scanner 305, a developing tool 320, an etching tool 325, and an ash tool 330, the semiconductor processing system 300 is not limited to the aforementioned tools and may include additional tools to assist in printing patterns onto a substrate. In some embodiments, two or more tools may be combined to form a composite tool that provides the functionality of multiple tools. Further details relating to the semiconductor processing system 300 are available in U.S. Patent Application Publication No. 2019 / 0187670, which is incorporated herein by reference in whole.

[0055]

[0063] The following paragraphs describe an improved defect location prediction model 405 that predicts defect locations on a substrate with higher accuracy than conventional tools (e.g., point determination tool 345). In some embodiments, the defect location prediction model 405 is trained using active learning techniques to generate predictions with higher accuracy. In the active learning techniques, the trained defect location prediction model 405 (e.g., trained with an initial dataset) is not only used to generate predictions about defect locations on the substrate to be inspected, but is also further trained using actual inspection results of the predicted locations (e.g., obtained from an inspection system) to update the defect location prediction model 405 based on actual inspection results of the predicted locations. Such a training process may be carried out stepwise, for example, using actual inspection results for all substrates subsequently analyzed by the defect location prediction model 405, which can result in improved prediction accuracy of the defect location prediction model 405. An active learning-based defect location identification method is described with reference to at least Figures 4 and 7 below.

[0056]

[0064] Figure 4 is a block diagram of a system 400 for predicting defect locations on a substrate 410, consistent with various embodiments of the present disclosure. The system 400 includes a defect location prediction model 405, an inspection tool 465, and a feedback tool 470. The defect location prediction model 405 includes a location prediction model 450, a confidence model 455, and a location selection component 460. In some embodiments, before generating predictions about a substrate (e.g., substrate 410), the defect location prediction model 405 is trained using an initial training dataset, at least as described with reference to Figure 6.

[0057]

[0065] In some embodiments, the location prediction model 450 is a machine learning (ML) model, similar to the point determination tool 345 in Figure 3. The location prediction model 450 generates predictions 415a~n for the number of locations n on the substrate 410, indicating whether a given location is likely to be a defective location or a non-defective location. A prediction 415a associated with "location a" on the substrate 410 may include the likelihood that "location a" is a defective location or a non-defective location. For example, a prediction may include a probability of "0.8", indicating that there is an "80%" likelihood that "location a" is defective and a "20%" likelihood that "location a" is not defective. Thus, the location prediction model 450 may classify "location a" as a defective location. Locations can be classified as defective or non-defective using other types of classification methods that do not use probability values. In some embodiments, the location prediction model 450 generates predictions 415a based on process-related data 435 associated with the substrate 410. In some embodiments, process-related data 435 may be analogous to weakness information 340. Process-related data 435 may include data associated with various tools and processes of the semiconductor processing system 300, such as developing tool 320, etching tool 325, ash tool 330, or other processes. For example, process-related data 435 may include metronome data such as critical dimension (CD) measurements, aberrations, edge placement error (EPE), film thickness on substrate 410, or other such data that may contribute to defects.

[0058]

[0066] In some embodiments, the confidence model 455 is an ML model. The confidence model 455 analyzes the process-related data 435 and generates confidence scores 420a to n, which indicate the confidence levels of the predictions 415a to n generated for each location by the location prediction model 450. For example, confidence score 420a indicates the confidence level of the prediction 415a that "location a" is defective. The confidence model 455 may use any of several measures when generating the confidence scores. For example, confidence score 420a can be a value in the range of "0" to "1", within this range, a higher value indicates higher confidence in the prediction. In some embodiments, the confidence model 455 may assign a higher confidence score if the process-related data 435 is similar to any of the previously analyzed process-related data, or a lower confidence score if the process-related data 435 is not similar to any of the previously analyzed process-related data. The confidence scores may be determined using any of several active learning methods. For example, the confidence score may be determined using a random forest model, as described below with reference to Figure 5A, or using the QBC (querying by committee) active learning method, as described below with reference to Figure 5B.

[0059]

[0067] Figure 5A is a block diagram for determining confidence scores using a random forest model, consistent with embodiments of the present disclosure. In the random forest model, a location prediction model 450 generates several predictions for each location, e.g., predictions 501 to 509, and a confidence model 455 determines a confidence score for that location as a function of predictions 501 to 509, for example, based on the variance 511 of all predictions. Further details regarding the random forest model can be found in the paper GASusto, "A dynamic sampling strategy based on confidence level of virtual metrology predictions," Proc. 28th Annu. SEMI Adv. Semiconductor Manuf. Conf. (ASMC), May 2017, which is incorporated herein by reference in whole.

[0060]

[0068] Figure 5B is a block diagram for determining a confidence score using the QBC method, consistent with an embodiment of the present disclosure. The QBC method can generate predictions for each location on the substrate 410, e.g., predictions 521 to 529, using several location prediction models 450a to n (e.g., a diverse committee of location prediction models 450a to n). The confidence model 455 can determine a confidence score as a function of predictions 521 to 529, for example, based on the variance 531 of predictions 521 to 529. For example, the confidence model 455 obtains predictions for "location a" from each location prediction model 450a to n of the committee, and then calculates a confidence score 531 as the variance of predictions 521 to 529 obtained from the committee. Further details regarding the QBC active learning method and other active learning methods can be found in the following papers: “Committee-based sampling for training probabilistic classifiers,” Dagan, I., & Engelson, SP (1995), Proc. of 12th Intl.Conf. on Machine Learning (ICML-95); “Employing EM and pool-based active learning for text classification,” McCallum, A., & Nigam, K. (1998), Proc. of 15th Intl.Conf. on Machine Learning (ICML-98); “Query learning strategies using boosting and bagging,” Abe, N., & Mamitsuka, H. (1998), Proc. of 15th Intl.Conf. on Machine Learning (ICML-98); and the ebook “An introduction to active learning,” Jennifer Prendki, (2018) (all of which are incorporated herein by reference in their entirety).

[0061]

[0069] Referring again to Figure 4, the location selection component 460 selects all locations on the substrate 410 that are associated with predictions having confidence scores that satisfy the location selection criteria. For example, the location selection component 460 may select all locations that are predicted to have defects and that are associated with confidence scores exceeding a first confidence threshold. In another example, the location selection component 460 may select all locations associated with confidence scores below a second confidence threshold, regardless of whether those locations are predicted to have defects or not. The location selection component 460 may add the selected locations to a sampling plan 425, which can be input to an inspection tool 465 for inspecting the selected locations. The sampling plan 425 may include information (e.g., (x,y) coordinates) about the locations on the substrate 410 to be inspected by the inspection tool 465. The inspection tool 465 may inspect the locations on the substrate 410 based on the sampling plan 425 and may output actual inspection results 430 (e.g., unpredicted) for the inspected locations. In some embodiments, the inspection result 430 may include an image of the inspected location (e.g., an SEM image), location information of the inspected location (e.g., (x,y) coordinates), and whether a defect was found or not at that location. In some embodiments, the inspection tool 465 may include the monitoring tool 335 of Figure 3, or the electron beam tool 40 of Figure 1 for performing the inspection, or a verification unit 350 that generates the inspection result 430 by comparing the inspection result 430 with the design parameters of a pattern printed on the substrate 410.

[0062]

[0070] The feedback tool 470 can further train the defect location prediction model 405 using the actual inspection results 430 for selected locations by inputting the inspection results 430 along with process-related data for those locations back into the defect location prediction model 405. By training the defect location prediction model 405 with the actual inspection results from the inspection tool 465, the cost function of the defect location prediction model 405 can be reduced, and the prediction accuracy of the defect location prediction model 405 can be improved (e.g., increased). In some embodiments, the cost function may represent the deviation between the prediction and the actual inspection results 430, and the prediction accuracy may represent the number of correct predictions compared to the total number of predictions. By training the defect location prediction model 405 incrementally (e.g., training the defect location prediction model 405 with the actual inspection results from the inspection tool 465 each time a prediction is made for a new or different substrate), the cost function is minimized, and therefore the prediction accuracy is maximized. As the prediction accuracy improves, the defect location prediction model 405 can predict locations that are likely to be defective with greater confidence.

[0063]

[0071] In some embodiments, the location selection component 460 may be configured to control the selection of locations for inspection (e.g., by adjusting one or more confidence thresholds). For example, if the prediction accuracy of the defect location prediction model 405 falls below an accuracy threshold, the location selection component 460 may select, for inspection, locations predicted to have a defect with a high confidence score (e.g., s > x, where s is the score and x is a first confidence threshold), and may have a larger first confidence threshold such that locations predicted to have a defect with a lower confidence score (e.g., s < x) are ignored. As the prediction accuracy improves, the location selection component 460 may decrease the first confidence threshold such that locations predicted to have a defect with an even lower confidence score (e.g., s > y and y < x, where y is the adjusted first confidence threshold) are selected for inspection. In another example, if the prediction accuracy of the defect location prediction model 405 falls below an accuracy threshold, the location selection component 460 may have a larger second confidence threshold such that locations associated with a lower confidence score (e.g., s < a and a < x, where a is a second confidence threshold) are selected for inspection regardless of whether they are predicted to have a defect or not. As the prediction accuracy improves, the location selection component 460 may decrease the second confidence threshold such that locations predicted to have a defect with a very low confidence score (e.g., s < b and b < a, where b is the second confidence threshold) are selected for inspection. In some embodiments, the location selection component 460 may also be configured to control the selection of locations for inspection based on resources available for inspection (e.g., the time and computational resources of the inspection tool 465). The location selection component 460 may adjust the confidence threshold according to the available resources. For example, the fewer the available resources, the fewer the number of locations selected for inspection. In some embodiments, the confidence threshold, accuracy threshold, available resources, or the number of locations to be inspected may be configurable by the user.

[0064]

[0072] Figure 6 is a block diagram showing the training of a defect location prediction model 405 using an initial training dataset, consistent with various embodiments of the present disclosure. The defect location prediction model 405 may need to be trained using an initial training dataset 605 before it can be used to generate predictions for substrates such as the substrate 410 in Figure 4. The initial training dataset 605 may be a labeled dataset, which includes process-related data 610a-n and inspection results 615a-n for n substrates. For example, for substrate "A", the initial training dataset 605 may include process-related data 610a and inspection results 615a associated with substrate "A". In some embodiments, the process-related data 610a may be similar to process-related data 435 and may include metronome data such as CD measurements, aberrations, EPE, film thickness on substrate "A", or other such data that may contribute to defects. In some embodiments, the inspection result 615a may be similar to the inspection result 430 and may include an image of the inspected location (e.g., an SEM image), location information of the inspected location (e.g., (x,y) coordinates), and whether a defect was found at that location or not. Labeled datasets may be obtained from various sources, including the tools of the semiconductor processing system 300 shown in Figure 3.

[0065]

[0073] The location prediction model 450 and the confidence model 455 may be ML models, as described above with reference to at least Figure 4. Training the defect location prediction model 405 may be an iterative process in which each iteration may include analyzing process-related data 610 associated with the substrate, determining a cost function, and updating the configuration of the defect location prediction model 405 based on the cost function, all of which are described in more detail below. In some embodiments, the defect location prediction model 405 may be trained in a "batch" manner rather than as an iterative process. For example, a training dataset 605 having process-related data 610a~n and inspection results 615a~n for "n" substrates may be input all at once. Upon inputting the process-related data 610a and inspection results 615a, the location prediction model 450 generates predictions 625a1~625ax for "x" locations on substrate "A", and the confidence model assigns confidence scores 630a1~630ax to the predictions 625a1~625ax, respectively. Next, the defect location prediction model 405 determines a cost function 650 that can show the deviation between the predicted results 625a1-625ax and the actual inspection results 615a by comparing the predicted results with the inspection results 615a. The defect location prediction model 405 may minimize the cost function 650 by updating its configuration (e.g., weights, biases, or other parameters of the location prediction model 450 or confidence model 455) based on the cost function 650 or other reference feedback information (e.g., user indications of accuracy, reference labels, or other information). The above process is repeated iteratively using process-related data and inspection results associated with different substrates in each iteration until a termination condition is met. The termination condition may include a predefined number of iterations, the cost function meeting a specific threshold, or other such conditions. After the termination condition is met, the defect location prediction model 405 may be considered "trained" and can be used to identify or predict defect locations on new substrates (e.g., substrates not yet analyzed using the defect location prediction model 405).

[0066]

[0074] In some embodiments, the trained defect location prediction model 405 can be used to predict defect locations on a new substrate, such as substrate 410. However, the trained defect location prediction model 405 can be further trained using an active learning ML method to further improve prediction accuracy. In the active learning ML method, the trained defect location prediction model 405 is trained using selectively labeled data, for example, actual inspection results of the locations from which predictions were generated using the trained defect location prediction model 405. This further improves prediction accuracy, for example, when the defect location prediction model 405 is analyzing process-related data that is not similar to any of the process-related data previously analyzed (either during training the defect location prediction model 405 or during actual prediction of defect locations). Such an active learning method can overcome the "conceptual drift" problem, a situation in which an ML model can become outdated and its accuracy can degrade if it is not regularly updated with new training data. In the field of semiconductor processing, manufacturing processes can change continuously, and therefore, process-related data associated with the substrate can also change. In some embodiments, even if the process-related data has not drifted, the relationship between process-related data and defective / non-defective labels may drift as a function of time (e.g., due to some hidden process variable that may not be available to the ML model). If the trained defect location prediction model 405 is input with process-related data that is not similar to, or significantly different from, the previously analyzed process-related data, the predictions generated by the defect location prediction model 405 may be inaccurate. The "conceptual drift" problem can be overcome and prediction accuracy can be improved by training the trained defect location prediction model 405 stepwise using actual inspection results of the locations from which predictions were generated using the trained defect location prediction model 405 (e.g., as described with reference to at least Figure 4).

[0067]

[0075] Figure 7 is a flow diagram of a process 700 for predicting defect locations on a substrate, consistent with embodiments of the present disclosure. In some embodiments, process 700 may be implemented in the system 400 of Figure 4. In operation P701, process-related data associated with the substrate is input to the defect location prediction model 405. Process-related data 435 associated with the substrate 410, including metronome data such as CD measurements, aberrations, EPE, film thickness on the substrate 410, or other such data that may contribute to defects, may be input to the defect location prediction model 405.

[0068]

[0076] In operation P703, a location 705 on the substrate 410 to be inspected may be selected based on predictions generated by the location prediction model 450. For example, the location prediction model 450 generates predictions 415a~n for the number of locations n on the substrate 410, indicating whether a given location is likely to be a defect location or a non-defect location. In some embodiments, the location prediction model 450 is initially trained using an initial training dataset to predict defect locations, as illustrated with reference to at least Figure 6.

[0069]

[0077] In operation P705, confidence scores 420a~n are generated for each prediction associated with location 705. The confidence scores may indicate the confidence level of the corresponding prediction. For example, confidence score 420a indicates the confidence level of prediction 415a that "location a" is defective. In some embodiments, a higher confidence score indicates higher confidence in the associated prediction. In some embodiments, the confidence model 455 may assign a higher confidence score if the process-related data 435 is similar to any of the previously analyzed process-related data, and a lower confidence score otherwise. The confidence scores may be determined using any of several active learning methods. For example, the confidence scores may be determined using a random forest model, as described with reference to at least Figure 5A, or using the QBC active learning method, as described with reference to at least Figure 5B.

[0070]

[0078] In operation P707, locations 705 associated with predictions that have confidence scores that satisfy the location selection criteria are added to a set of locations 707 to be inspected by the inspection tool 465. For example, the location selection component 460 may add all locations 705 that are predicted to be defective and are associated with confidence scores exceeding a first confidence threshold to a set of locations 707. In another example, the location selection component 460 may add all locations associated with confidence scores below a second confidence threshold to a set of locations 707, regardless of whether the predictions for those locations are defective or not.

[0071]

[0079] In operation P709, an inspection result 430 for a set of locations 707 is obtained from the inspection tool 465. The location selection component 460 may add information about the set of locations 707 (e.g., (x,y) coordinates) to the sampling plan 425, and input the sampling plan 425 to the inspection tool 465. The inspection tool 465 may inspect the set of locations 707 on the substrate 410 and output the actual inspection result 430. In some embodiments, the inspection result 430 may include an image of the inspected location (e.g., an SEM image), location information of the inspected location (e.g., (x,y) coordinates), and whether a defect was found or not at that location.

[0072]

[0080] In operation P711, the inspection results 430 of a set of locations 707 and process-related data for those locations are fed back to the defect location prediction model 405, thereby further training the defect location prediction model 405 using the actual inspection results 430 of a set of locations. In some embodiments, the defect location prediction tool is trained incrementally by performing operations P701 to P711 each time a prediction is made for a new or different substrate. That is, the defect location prediction model 405 is trained using the actual inspection results from the inspection tool 465 each time a prediction is made for a new or different substrate. By incrementally training the defect location prediction model 405, the cost function associated with the defect location prediction model 405 is minimized, and therefore the prediction accuracy of the defect location prediction model 405 is maximized. As prediction accuracy improves, the defect location prediction model 405 can predict locations that are likely to be defective with greater confidence.

[0073]

[0081] Figure 8 is a block diagram showing a computer system 800 that can assist in the implementation of a method, flow, module, component, or apparatus disclosed herein. The computer system 800 includes a bus 802 or other communication mechanism for communicating information and a processor 804 (or a plurality of processors 804 and 805) coupled to the bus 802 for processing information. The computer system 800 also includes main memory 806, such as random access memory (RAM) or other dynamic storage device, coupled to the bus 802 for storing information and instructions executed by the processor 804. The main memory 806 may also be used to store temporary variables or other intermediate information during the execution of instructions executed by the processor 804. The computer system 800 further includes read-only memory (ROM) 808 or other static storage device coupled to the bus 802 for storing static information and instructions for the processor 804. A storage device 810, such as a magnetic disk or optical disk, is provided and coupled to the bus 802 for storing information and instructions.

[0074]

[0082] The computer system 800 may be coupled via bus 802 to a display 812, such as a cathode ray tube (CRT), flat panel, or touch panel display, for displaying information to the computer user. An input device 814, including alphanumeric and other keys, is coupled to bus 802 to communicate information and command selections to the processor 804. Another type of user input device is a cursor control unit 816, such as a mouse, trackball, or cursor directional keys, for communicating directional information and command selections to the processor 804 and for controlling cursor movement on the display 812. This input device typically has two degrees of freedom, allowing the device to pinpoint its position in a plane along two axes (a first axis (e.g., x) and a second axis (e.g., y)). A touch panel (screen) display may also be used as an input device.

[0075]

[0083] According to one embodiment, some of the methods described herein may be implemented by a computer system 800 in response to a processor 804 executing one or more sequences of one or more instructions contained in main memory 806. Such instructions may be read into main memory 806 from another computer-readable medium, such as a storage device 810. By executing the sequence of instructions contained in main memory 806, the processor 804 implements the process steps described herein. Alternatively, one or more processors in a multiprocessing configuration may be used to execute the sequence of instructions contained in main memory 806. In alternative embodiments, hardwired circuits may be used instead of, or in combination with, software instructions. Thus, the description herein is not limited to any particular combination of hardware circuits and software.

[0076]

[0084] As used herein, the term “computer-readable medium” refers to any medium involved in providing instructions to processor 804 for execution. Such mediums can take many forms, but are not limited to, non-volatile media, volatile media, and transmission media. Non-volatile media include, for example, optical or magnetic disks such as storage device 810. Volatile media include dynamic memory such as main memory 806. Transmission media include coaxial cables, copper wires, and optical fibers, including wires including bus 802. Transmission media can also take the form of sound waves or light waves, such as those generated during radio frequency (RF) and infrared (IR) data communications. Common forms of computer-readable media include, for example, floppy disks, flexible disks, hard disks, magnetic tapes, any other magnetic media, CD-ROMs, DVDs, any other optical media, punch cards, paper tapes, any other physical media having a pattern of holes, RAM, PROMs, and EPROMs, FLASH-EPROMs, any other memory chips or cartridges, carrier waves as described below, or any other medium that a computer can read.

[0077]

[0085] Various forms of computer-readable media may be involved in transporting one or more sequences of one or more instructions to processor 804 for execution. For example, instructions may first be transported to a magnetic disk of a remote computer. The remote computer may load the instructions into dynamic memory and transmit them over a telephone line using a modem. A modem local to computer system 800 may receive data over the telephone line and convert the data into an infrared signal using an infrared transmitter. An infrared detector coupled to bus 802 may receive the data transported by the infrared signal and load that data onto bus 802. Bus 802 transports the data to main memory 806, from which processor 804 retrieves and executes the instructions. Instructions received by main memory 806 may optionally be stored in storage device 810 before or after execution by processor 804.

[0078]

[0086] The computer system 800 may also include a communication interface 818 coupled to bus 802. The communication interface 818 provides bidirectional data communication coupled to a network link 820 connected to a local network 822. For example, the communication interface 818 may be an ISDN (Integrated Services Digital Network) card or modem that provides data communication connectivity to a corresponding type of telephone line. As another example, the communication interface 818 may be a local area network (LAN) card that provides data communication connectivity to a compatible LAN. Wireless links may also be implemented. In such implementations, the communication interface 818 transmits and receives electrical, electromagnetic, or optical signals that carry digital data streams representing various types of information.

[0079]

[0087] Network link 820 typically provides data communication to other data devices through one or more networks. For example, network link 820 may provide connection to a host computer 824 or to data equipment operated by an Internet service provider (ISP) 826 through a local network 822. The ISP 826 then provides data communication services through the Worldwide Packet Data Communications Network (now commonly referred to as the “Internet” 828). Both the local network 822 and the Internet 828 use electrical, electromagnetic, or optical signals to carry digital data streams. Signals carrying digital data to and from computer system 800, as well as signals on network link 820 and through communication interface 818, are exemplary forms of information carriers.

[0080]

[0088] The computer system 800 can send messages and receive data, including program code, through one or more networks, network links 820, and communication interfaces 818. In the Internet example, server 830 may send request code for an application program through the Internet 828, ISP 826, local network 822, and communication interfaces 818. Such a downloaded application may provide all or part of the methods described herein, for example. The received code may be executed by processor 804 upon receipt and / or stored in storage device 810 or other non-volatile storage for later execution. In this way, the computer system 800 may obtain application code in carrier form.

[0081]

[0089] The embodiments can be further described using the following clauses. 1. A non-temporary computer-readable medium having instructions that, when executed by a computer, cause the computer to execute a method for identifying a location to be inspected on a circuit board, wherein the method is Selecting multiple locations on the substrate to be inspected based on a first submodel of a defect location prediction model trained using an initial training dataset associated with other substrates to generate a prediction of whether each location is defective or not, Using a second submodel of a defect location prediction model trained with an initial training dataset, generate confidence scores for each location based on process-related data associated with the substrate, wherein the confidence scores indicate the confidence of the prediction for the corresponding location. Each location whose confidence score satisfies one of several confidence thresholds is added to a set of locations to be inspected by the inspection system, Obtaining test result data, By providing the defect location prediction model with inspection result data and process-related data for a set of locations as training data, the defect location prediction model is trained step by step. Non-temporary computer-readable media, including [specific examples of such media]. 2. Training the second sub-model step by step, with each iteration, To train a first submodel using inspection result data and process-related data from different substrates that were not inspected in any of the previous iterations. A computer-readable medium as described in Clause 1, which is an iterative process including the following. 3. Adding each of the positions A computer-readable medium as described in Clause 1, which includes adding each location to a set of locations if the confidence score of the prediction of defects for the corresponding location exceeds a first confidence threshold of the confidence threshold. 4. Adding each of the positions The computer-readable media described in Clause 1, which includes adding each location to a set of locations if the confidence score of the prediction of defect or non-defect for the corresponding location falls below a second confidence threshold of the confidence threshold. 5. Determine the prediction accuracy of the defect location prediction model based on the number of correct predictions and the total number of predictions. Computer-readable media as described in Clause 1, further including the above. 6. Training the defect location prediction model stepwise improves prediction accuracy, as described in the computer-readable media of Clause 5. 7. Adjust the confidence threshold based on changes in prediction accuracy. Computer-readable media as described in Clause 5, further including the above. 8. The computer-readable media described in Clause 7, wherein adjusting the confidence threshold includes decreasing the first confidence threshold of the confidence threshold as prediction accuracy improves, and the first confidence threshold is used to select locations from among the locations associated with a confidence score in which the prediction of a defect exceeds the first confidence threshold. 9. The computer-readable media described in Clause 7, wherein adjusting the confidence threshold includes decreasing a second confidence threshold as prediction accuracy improves, the second confidence threshold being used to select locations associated with confidence scores where the prediction of defect or non-defect falls below the second confidence threshold. 10. The computer-readable media described in Clause 7, wherein adjusting the confidence threshold includes increasing a first confidence threshold as prediction accuracy decreases, and the first confidence threshold is used to select locations from among the locations associated with a confidence score in which the prediction of a defect exceeds the first confidence threshold. 11. The computer-readable media described in Clause 7, wherein adjusting the confidence threshold includes increasing a second confidence threshold as prediction accuracy decreases, the second confidence threshold being used to select locations associated with confidence scores where the prediction of defect or non-defect falls below the second confidence threshold. 12. A computer-readable medium as described in Clause 1, wherein the first submodel is configured to generate probability values ​​for each of the predictions, the probability values ​​indicating the probability that the corresponding location is a defective location or a non-defective location. 13. Generating a confidence score is A computer-readable medium as described in Clause 1, which includes generating a confidence score for a particular location based on a comparison of process-related data associated with a particular location among the locations with process-related data in an initial training dataset, or training data used to train a defect location prediction model. 14. The defect location prediction model includes multiple first submodels and generates confidence scores. From each of the first submodels, obtain the probability value associated with the prediction for a specific location among the locations, The confidence score for a specific location is generated as a function of the probability values ​​obtained from the first submodel, Computer-readable media as described in Clause 1, including the above. 15. Computer-readable media as described in Clause 1, including obtaining test result data from a testing system. 16. A computer-readable medium as described in Clause 1, in which the inspection result data includes information regarding whether or not a defect exists at each location in a set of locations. 17. A computer-readable medium as described in Clause 16, in which inspection result data indicates the presence of defects at specific locations within a set of locations, based on the number of defects detected at specific locations that satisfy a defect threshold. 18. A computer-readable medium as described in Clause 1, in which process-related data includes data associated with multiple processes involved in forming a pattern on a substrate for each of the locations. 19. Computer-readable media as described in Clause 18, including metrology data associated with multiple processes. 20. The initial training dataset is a computer-readable medium as described in Clause 1, containing process-related data associated with multiple substrates. 21. A non-temporary computer-readable medium having instructions that, when executed by a computer, cause the computer to execute a method for identifying a location to be inspected on a first substrate using a machine learning model, and for training the machine learning model to identify a location to be inspected on a second substrate based on the inspection results of the location on the first substrate, wherein the method Inputting process-related data associated with the substrate into the defect location prediction model, The process involves generating a defect or non-defect prediction for each of several locations on a substrate using a defect location prediction model, wherein each prediction is associated with a confidence score indicating the confidence level of the prediction for the corresponding location. Each location whose confidence score satisfies one of several confidence thresholds is added to a set of locations to be inspected by the inspection system, Obtaining inspection result data for a set of locations from the inspection system, The inspection result data and process-related data for a set of locations are input into the defect location prediction model to train the defect location prediction model, Non-temporary computer-readable media, including [specific examples of such media]. 22. Training a defect location prediction model stepwise, wherein each iteration of the stepwise training is To train a defect location prediction model using inspection result data and process-related data from different substrates that were not inspected in any of the previous iterations. Computer-readable media as described in Clause 21, which further includes training, which is an iterative process including the following. 23. Adding each of the positions A computer-readable medium as described in Clause 21, which includes adding each location to a set of locations if the confidence score of the prediction of defects for the corresponding location exceeds a first confidence threshold of the confidence threshold. 24. Adding each of the positions The computer-readable media described in Clause 21, which includes adding each location to a set of locations if the confidence score of the prediction of defect or non-defect for the corresponding location falls below a second confidence threshold of the confidence threshold. 25. Determine the prediction accuracy of the defect location prediction model based on the number of correct predictions and the total number of predictions. Computer-readable media as described in Clause 21, further including the above. 26. Adjust the confidence threshold based on changes in prediction accuracy. Computer-readable media as described in Clause 25, further including the above. 27. To generate predictions, The computer-readable medium described in Clause 21, which includes training a defect location prediction model using an initial training dataset associated with other substrates to generate defect or non-defect predictions for each location with respect to a corresponding substrate, before inputting process-related data of the substrate, wherein the initial training dataset includes process-related data of other substrates. 28. To generate predictions, A computer-readable medium as described in Clause 21, which includes generating a confidence score for a particular location based on a comparison of process-related data associated with a particular location among the locations with process-related data associated with other substrates used to train a defect location prediction model. 29. To generate predictions, Obtain probability values ​​associated with the prediction of defects or non-defects for a specific location from each of multiple prediction models, The process involves generating a confidence score for a specific location as a function of probability values ​​obtained from a predictive model, and Computer-readable media as described in Clause 21, including the above. 30. A method for identifying a location to be inspected on a first substrate using a machine learning model, and for training a machine learning model to identify a location to be inspected on a second substrate based on the inspection results of the location on the first substrate, wherein the method is Inputting process-related data associated with the substrate into the defect location prediction model, The process involves generating a defect or non-defect prediction for each of several locations on a substrate using a defect location prediction model, wherein each prediction is associated with a confidence score indicating the confidence level of the prediction for the corresponding location. Each location where the confidence score satisfies the confidence threshold is added to a set of locations to be inspected by the inspection system, Obtaining inspection result data for a set of locations from the inspection system, The inspection result data and process-related data for a set of locations are input into the defect location prediction model to train the defect location prediction model, Methods that include... 31. Training a defect location prediction model stepwise, wherein each iteration of the stepwise training is To train a defect location prediction model using inspection result data and process-related data from different substrates that were not inspected in any of the previous iterations. The method according to clause 30, which further includes training, which is an iterative process. 32. Adding each of the positions The method according to clause 30, which includes adding each location to a set of locations if the confidence score of the defect prediction for the corresponding location exceeds a first confidence threshold of the confidence threshold. 33. Adding each of the positions The method according to Clause 30, which includes adding each location to a set of locations if the confidence score of the prediction of defect or non-defect for the corresponding location falls below a second confidence threshold of the confidence threshold. 34. Determine the prediction accuracy of the defect location prediction model based on the number of correct predictions and the total number of predictions. The method described in clause 30, further including the following: 35. Adjust the confidence threshold based on changes in prediction accuracy. The method described in Clause 34, further including the method described in Clause 34. 36. To generate predictions, The method according to Clause 30, comprising training a defect location prediction model using an initial training dataset associated with other substrates to generate defect or non-defect predictions for each location with respect to a corresponding substrate, before inputting process-related data for the substrate. 37. To generate predictions, The method according to clause 30, comprising generating a confidence score for a particular location based on a comparison of process-related data associated with a particular location among the locations with process-related data associated with other substrates used to train a defect location prediction model. 38. To generate predictions, Obtaining probability values ​​from each of multiple prediction models that are associated with the prediction of a particular location as either defective or non-defective, The process involves generating a confidence score for a specific location as a function of probability values ​​obtained from a predictive model, and The method described in Clause 30, including the method described in Clause 30. 39. An apparatus for identifying a position to be inspected on a first substrate using a machine learning model, and for training the machine learning model to identify a position to be inspected on a second substrate based on the inspection results of the position on the first substrate, wherein the apparatus comprises: Memory for storing a set of instructions, Inputting process-related data associated with the substrate into the defect location prediction model, The process involves generating a defect or non-defect prediction for each of several locations on a substrate using a defect location prediction model, wherein each prediction is associated with a confidence score indicating the confidence level of the prediction for the corresponding location. Each location where the confidence score satisfies the confidence threshold is added to a set of locations to be inspected by the inspection system, Obtaining inspection result data for a set of locations from the inspection system, The inspection result data and process-related data for a set of locations are input into the defect location prediction model to train the defect location prediction model, A processor configured to execute a set of instructions to cause the device to perform the method, A device including a device. 40. The method is This involves training a defect location prediction model step by step, where each iteration of the step by step is... To train a defect location prediction model using inspection result data and process-related data from different substrates that were not inspected in any of the previous iterations. The apparatus described in Clause 39, which further includes training, which is an iterative process. 41. Adding each of the positions The apparatus according to Clause 39, which includes adding each location to a set of locations if the confidence score of the prediction of defects for a corresponding location exceeds a first confidence threshold of the confidence threshold. 42. Adding each of the positions The apparatus according to Clause 39, which includes adding each location to a set of locations if the confidence score of the prediction of defect or non-defect for the corresponding location falls below a second confidence threshold of the confidence threshold. 43. Determine the prediction accuracy of the defect location prediction model based on the number of correct predictions and the total number of predictions. The apparatus described in Clause 39, further including the following. 44. Adjust the confidence threshold based on changes in prediction accuracy. The apparatus described in Clause 43, further including the following. 45. To generate predictions, The apparatus according to Clause 39, which includes training a defect location prediction model using an initial training dataset associated with other substrates to generate a prediction of defect or non-defect for each location with respect to a corresponding substrate, before inputting process-related data of the substrate. 46. ​​To generate predictions, The apparatus according to Clause 39, which includes generating a confidence score for a particular location based on a comparison of process-related data associated with a particular location among the locations with process-related data associated with other substrates used to train a defect location prediction model. 47. To generate predictions, Obtain probability values ​​associated with the prediction of defects or non-defects for a specific location from each of multiple prediction models, The process involves generating a confidence score for a specific location as a function of probability values ​​obtained from a predictive model, and The apparatus described in Clause 39, including the apparatus described in Clause 39. 48. A non-temporary computer-readable medium on which instructions are recorded, wherein, when executed by a computer, the instructions perform the method described in any of the preceding clauses.

[0082]

[0090] A non-temporary computer-readable medium may be provided that stores instructions for the processor of a controller (e.g., controller 50 in Figure 1) to perform, among other things, image inspection, image acquisition, stage positioning, beam focusing, electric field adjustment, beam bending, condenser lens adjustment, charged particle source activation, beam deflection, and at least a portion of methods 600 and 700. Common forms of non-temporary media include, for example, floppy disks, flexible disks, hard disks, solid-state drives, magnetic tapes or any other magnetic data storage media, compact disk read-only memory (CD-ROM), any other optical data storage media, any physical medium having a hole pattern, random access memory (RAM), programmable read-only memory (PROM), erasable programmable read-only memory (EPROM), flash EPROM or any other flash memory, non-volatile random access memory (NVRAM), caches, registers, any other memory chips or cartridges and their network-connected versions.

[0083]

[0091] The relative dimensions of components in the drawings may be exaggerated for clarity. In the description of the drawings, identical or similar reference numbers refer to identical or similar components or entities, and only differences relating to individual embodiments are described. When used herein, unless otherwise specified, the term “or” encompasses all possible combinations unless it is impossible to achieve. For example, if a component is specified to include A or B, then unless otherwise specified or impossible to achieve, the component may include A, or B, or A and B. As a second example, if a component is specified to include A, B, or C, then unless otherwise specified or impossible to achieve, the component may include A, or B, or C, or A and B, or A and C, or B and C, or A and B and C.

[0084]

[0092] It will be understood that the embodiments of this disclosure are not limited to the structures described above and shown in the accompanying drawings, and that various modifications and alterations may be made without departing from that scope. This disclosure is described in relation to various embodiments, and other embodiments of the invention will become apparent to those skilled in the art from considering the specifications and practices of the invention disclosed herein. Specifications and examples are to be considered merely illustrative, and the true scope and spirit of the invention are intended to be shown by the following claims.

[0085]

[0093] The above description is intended to be illustrative, not limiting. Therefore, it will be apparent to those skilled in the art that modifications can be made as described, without departing from the claims set forth below.

Claims

1. A non-temporary computer-readable medium having instructions that cause the computer to perform a method for identifying a location to be inspected on a circuit board, wherein the method is Selecting a plurality of locations on the substrate to be inspected based on a first submodel of a defect location prediction model trained using an initial training dataset associated with other substrates to generate a prediction of whether each of the locations is defective or not, Using a second submodel of the defect location prediction model trained with the initial training dataset, generate confidence scores for each of the locations based on process-related data associated with the substrate, wherein the confidence scores indicate the confidence of the prediction for the corresponding location. Each of the locations whose confidence score satisfies one of a plurality of confidence thresholds is added to a set of locations to be inspected by the inspection system. Obtaining test result data, The defect location prediction model is trained step by step by providing the inspection result data and process-related data for one set of the aforementioned locations as training data to the defect location prediction model. Non-temporary computer-readable media, including [specific examples of such media].

2. Training the second submodel stepwise means that each iteration is, The first submodel is trained using inspection result data and process-related data from different substrates that have not been inspected in any of the previous iterations. A computer-readable medium according to claim 1, comprising an iterative process including the following:

3. Adding each of the aforementioned positions The computer-readable medium according to claim 1, comprising adding each of the locations to a set of locations if the confidence score of the prediction of the defect for the corresponding location exceeds a first confidence threshold of the confidence threshold.

4. Adding each of the aforementioned positions The computer-readable medium according to claim 1, comprising adding each of the locations to a set of locations if the confidence score of the prediction of defect or non-defect with respect to the corresponding location falls below a second confidence threshold of the confidence threshold.

5. The prediction accuracy of the defect location prediction model is determined based on the number of correct predictions and the total number of predictions. The computer-readable medium according to claim 1, further comprising:

6. The computer-readable medium according to claim 5, wherein training the defect location prediction model in stages improves the prediction accuracy.

7. Adjust the confidence threshold based on the change in the prediction accuracy. The computer-readable medium according to claim 5, further comprising:

8. The computer-readable medium according to claim 7, wherein adjusting the confidence threshold includes decreasing a first confidence threshold of the confidence threshold as the prediction accuracy improves, the first confidence threshold is used to select a location from the locations associated with the confidence score at which the prediction of the defect exceeds the first confidence threshold.

9. The computer-readable medium according to claim 7, wherein adjusting the confidence threshold includes decreasing a second confidence threshold of the confidence threshold as the prediction accuracy improves, the second confidence threshold is used to select a location from the locations associated with a confidence score where the prediction of defect or non-defect is below the second confidence threshold.

10. The computer-readable medium according to claim 7, wherein adjusting the confidence threshold includes increasing a first confidence threshold of the confidence threshold as the prediction accuracy decreases, the first confidence threshold is used to select a location from the locations associated with the confidence score at which the prediction of the defect exceeds the first confidence threshold.

11. The computer-readable medium according to claim 7, wherein adjusting the confidence threshold includes increasing a second confidence threshold of the confidence threshold as the prediction accuracy decreases, the second confidence threshold is used to select a location among the locations associated with the confidence score where the prediction of defect or non-defect is below the second confidence threshold.

12. The computer-readable medium according to claim 1, wherein the first submodel generates probability values ​​for each of the predictions, the probability values ​​indicating the probability that the corresponding location is a defective location or a non-defective location.

13. To generate the aforementioned confidence score, The computer-readable medium according to claim 1, comprising generating the confidence score for the specific location based on a comparison of process-related data associated with a specific location among the locations, with process-related data in the initial training dataset, or the training data used to train the defect location prediction model.

14. The defect location prediction model includes a plurality of first submodels and generates the confidence score, From each of the first submodels, obtain the probability value associated with the prediction for a specific location among the locations, The confidence score for the specific location is generated as a function of the probability value obtained from the first submodel, A computer-readable medium according to claim 1, including the following:

15. An apparatus for identifying a position to be inspected on a first substrate using a machine learning model, and for training the machine learning model to identify a position to be inspected on a second substrate based on the inspection results of the position on the first substrate, wherein the apparatus comprises: Memory for storing a set of instructions, Inputting process-related data associated with the substrate into the defect location prediction model, Using the defect location prediction model, generate predictions of defects or non-defects for each of a plurality of locations on the substrate, wherein each prediction is associated with a confidence score indicating the confidence level of the prediction for the corresponding location. Each of the aforementioned locations whose confidence score satisfies the confidence threshold is added to a set of locations to be inspected by the inspection system. To obtain inspection result data for one set of the aforementioned positions from the inspection system, The inspection result data and process-related data relating to one set of the aforementioned locations are input into the defect location prediction model in order to train the defect location prediction model. At least one processor that executes a set of instructions for causing the device to perform the method, A device including a device.

Images