Systems and methods for generating predicted images for wafer inspection using machine learning

Abstract

A system and method for generating predicted images for wafer inspection using machine learning is provided. Some embodiments of the system and method include acquiring a wafer after photoresist applied to the wafer has been developed; imaging a segment of the developed wafer; acquiring the wafer after the wafer has been etched; imaging a segment of the etched wafer; training a machine learning model using the imaged portion of the developed wafer and the imaged segment of the etched wafer; and applying the trained machine learning model using the imaged segment of the etched wafer to generate a predicted image of the developed wafer. Some embodiments include imaging a segment of the developed wafer; imaging a portion of a segment of the etched wafer; training a machine learning model; and applying the trained machine learning model to generate a predicted post-etch image of the developed wafer.

Description

[0001] Cross Reference to Related Applications

[0002] This application claims priority to U.S. Application 62 / 903,568, filed September 20, 2019, and U.S. Application 63 / 001,055, filed March 27, 2020, which are incorporated by reference herein in their entirety. TECHNICAL FIELD

[0003] The present disclosure relates generally to the field of inspecting wafers using a charged particle beam system that uses machine learning to generate predicted images of the wafers. BACKGROUND

[0004] In the manufacturing process of integrated circuits (ICs), unfinished or finished circuit components are inspected to ensure that they are manufactured according to the design and are free of defects. Inspection systems that utilize optical microscopes typically have a resolution of down to a few hundred nanometers; and the resolution is limited by the wavelength of light. As the physical size of IC components continues to decrease to below 100 nanometers or even 10 nanometers, there is a need for inspection systems that have a higher resolution than inspection systems that utilize optical microscopes.

[0005] Charged particle (e.g., electron) beam microscopes, such as scanning electron microscopes (SEMs) or transmission electron microscopes (TEMs), have a resolution down to less than 1 nanometer, making them a viable tool for inspecting IC components having feature sizes below 100 nanometers. With an SEM, the electrons of a single primary electron beam or multiple primary electron beams can be focused at a location of interest of a wafer to be inspected. The primary electrons interact with the wafer and can be backscattered or can cause the wafer to emit secondary electrons. The intensity of the electron beam, including the backscattered electrons and the secondary electrons, can vary based on the characteristics of the internal and external structures of the wafer, which can indicate whether the wafer has a defect. SUMMARY

[0006] Embodiments in accordance with the present disclosure include systems and methods for inspecting wafers in a charged particle system using machine learning. The system includes a controller including circuitry to: image a portion of a segment of a wafer after photoresist applied to the wafer has been developed; image the segment of the wafer after the wafer has been etched; and train a machine learning model using the imaged portion of the developed wafer and the imaged portion of the etched wafer.

[0007] A method for inspection includes obtaining a wafer after photoresist applied to the wafer has been developed; imaging a portion of a segment of the developed wafer; obtaining the wafer after the wafer has been etched; imaging the segment of the etched wafer; and training a machine learning model using the imaged portion of the developed wafer and the imaged segment of the etched wafer.

[0008] A method for training a machine learning model includes inputting a first set of images of a portion of a segment of photoresist applied to a wafer that has been developed into the machine learning model; inputting a second set of images of the segment of the wafer after the wafer has been etched into the machine learning model; and adjusting weights of the machine learning model based on the first set of images and the second set of images, wherein the trained machine learning model is configured to generate predicted images of other portions of the segment of the developed wafer.

[0009] A non-transitory computer readable medium stores a set of instructions executable by at least one processor of a computing device to cause the computing device to perform a method for inspecting a wafer, the method including imaging a portion of a segment of a wafer after photoresist applied to the wafer has been developed; imaging the segment of the wafer after the wafer has been etched; and training a machine learning model using the imaged portion of the developed wafer and the imaged segment of the etched wafer.

[0010] Embodiments in accordance with the present disclosure include systems and methods for inspecting a wafer in a charged particle system using machine learning. The system includes a controller including circuitry to: image a segment of a wafer after photoresist applied to the wafer has been developed; image a portion of a segment of the wafer after the wafer has been etched; and train a machine learning model using the imaged segment of the developed wafer and the imaged portion of the etched wafer.

[0011] A method for inspection includes obtaining a wafer after photoresist applied to the wafer has been developed; imaging a segment of the developed wafer; obtaining the wafer after the wafer has been etched; imaging a portion of a segment of the etched wafer; and training a machine learning model using the imaged segment of the developed wafer and the imaged portion of the etched wafer.

[0012] A method for training a machine learning model includes inputting a first set of images of a portion of a segment of photoresist applied to a wafer that has been etched into the machine learning model; inputting a second set of images of the segment of the wafer after the wafer has been developed into the machine learning model; and adjusting weights of the machine learning model based on the first set of images and the second set of images, wherein the trained machine learning model is configured to generate predicted post-etch images of other portions of the segment of the developed wafer.

[0013] The non-transient computer-readable medium storage can be executed by at least one processor of a computing device to cause the computing device to execute a set of instructions for inspecting a wafer, the method comprising: imaging a segment of the wafer after a photoresist applied to the wafer has been developed; imaging a portion of a segment of the wafer after the wafer has been etched; and training a machine learning model using the imaged segment of the developed wafer and the imaged portion of the etched wafer.

[0014] Further objects and advantages of the disclosed embodiments will be set forth in part in the description which follows, and in part will be apparent from the description, or may be learned by practice of these embodiments. The objects and advantages of the disclosed embodiments may be realized and obtained by means of the elements and combinations set forth in the claims.

[0015] It should be understood that the foregoing general description and the following detailed description are exemplary and illustrative only, and not intended to limit the claimed disclosed embodiments. Attached Figure Description

[0016] Figure 1 This is a schematic diagram illustrating an exemplary electron beam tool according to an embodiment of the present disclosure.

[0017] Figure 2A This is a flowchart illustrating an exemplary method for wafer inspection according to an embodiment of the present disclosure.

[0018] Figure 2B This is a flowchart illustrating an exemplary method for wafer inspection according to an embodiment of the present disclosure.

[0019] Figure 3 This is an exemplary distribution map of critical dimensions obtained from SEM images according to embodiments of the present disclosure.

[0020] Figure 4 This is an exemplary top view of a chip having a region for machine learning, according to an embodiment of the present disclosure.

[0021] Figure 5A Exemplary SEM images of a wafer according to embodiments of the present disclosure are depicted.

[0022] Figure 5B Exemplary SEM images of a wafer according to embodiments of the present disclosure are depicted.

[0023] Figure 6 This is an exemplary system for generating predictive SEM images according to embodiments of the present disclosure.

[0024] Figure 7AThis is a schematic diagram illustrating an exemplary machine learning model training architecture according to an embodiment of the present disclosure.

[0025] Figure 7B This is a schematic diagram illustrating an exemplary machine learning model training architecture according to an embodiment of the present disclosure.

[0026] Figure 8A This is a schematic diagram illustrating an exemplary machine learning model application architecture according to an embodiment of the present disclosure.

[0027] Figure 8B This is a schematic diagram illustrating an exemplary machine learning model application architecture according to an embodiment of the present disclosure.

[0028] Figure 9A This is a flowchart illustrating an exemplary method for wafer inspection according to an embodiment of the present disclosure.

[0029] Figure 9B This is a flowchart illustrating an exemplary method for wafer inspection according to an embodiment of the present disclosure.

[0030] Figure 10 This is a flowchart illustrating an exemplary method for wafer inspection according to an embodiment of the present disclosure. Detailed Implementation

[0031] Reference will now be made in detail to exemplary embodiments, examples of which are illustrated in the accompanying drawings. The following description refers to the accompanying drawings, wherein the same reference numerals in different drawings denote the same or similar elements unless otherwise stated. The implementations set forth in the following description of the exemplary embodiments do not represent all implementations consistent with the invention. Rather, they are merely examples of devices and methods consistent with the aspects of the invention relevance described in the appended claims.

[0032] Electronic devices consist of circuits formed on silicon wafers called substrates. Many circuits can be formed together on the same silicon wafer and are called integrated circuits or ICs. The size of these circuits has been reduced so significantly that many of them can be mounted on the substrate. For example, the IC chip in a smartphone can be as small as a thumbnail and can include more than 2 billion transistors, each smaller than 1 / 1000th the size of a human hair.

[0033] Manufacturing these extremely small ICs is a complex, time-consuming, and expensive process, typically involving hundreds of individual steps. Even a mistake in one step can result in a defect that renders the completed IC useless. Therefore, one goal of the manufacturing process is to avoid such defects in order to maximize the number of functional ICs manufactured in the process, i.e., to increase the overall yield of the process.

[0034] One component of improving yield is monitoring the chip manufacturing process to ensure that a sufficient number of functional integrated circuits are produced. One way to monitor this process is to inspect the chip circuit structure at various stages of its formation. This can be done using a scanning electron microscope (SEM). SEMs can be used to image these extremely small structures, essentially taking "photographs" of them. This image can be used to determine if the structure is formed correctly and in the correct location. If the structure is defective, the process can be adjusted to make the defect less likely to recur.

[0035] As mentioned above, inspection can be performed at various stages. For example, images of the wafer can be acquired after the photoresist applied to the wafer has been developed (e.g., after photolithography), after etching, and at other stages. Inspection of the wafer after development is often desirable because it provides a direct link between the SEM images acquired during inspection and the various exposure process conditions (e.g., focal length, dose, etc.) that define patterning performance. Inspection of the wafer after development can allow for optimization of exposure process conditions. The SEM images obtained after development can be used for defect detection (e.g., necking, bridging, etc.), roughness / randomness characterization (e.g., line edge / width roughness, local critical size uniformity, edge placement error), process window optimization, calibration of computational lithography models, and process performance monitoring and control. The SEM images obtained after development can also be used to generate predictive images of the wafer after etching. Inspection of the wafer after development can allow for reprocessing of the wafer based on inspection measurements. That is, after etching, the wafer can be reprocessed to improve inspection results instead of being returned to its previous processing state. Therefore, it is important to obtain SEM images of the printed pattern on the wafer after development and before etching.

[0036] However, post-development SEM inspection is constrained by the inherent limitations of SEM itself. SEM inspection involves bombarding the sample with electrons. Because photoresist can be electron-sensitive, it may shrink or compress by a non-negligible amount as a result of SEM inspection. In other words, SEM inspection can damage the photoresist on the wafer, altering the wafer pattern. This alteration can result in modified critical performance parameters (e.g., defects, line edge roughness, linewidth roughness, local critical dimension uniformity, etc.) that do not reflect the wafer's true critical performance parameters when not inspected by SEM after development.

[0037] Some models use parameters known based on the physics of etching and the materials to generate predicted images of the wafer after etching. However, these models may not generate accurate predicted images because the actual physical parameters of the etching process may be unknown. Some models adjust or calibrate parameters to minimize the difference between the physics-based model and the predicted image. However, because these models lack the ability to represent the complex relationships present in the data, they may still fail to generate accurate predicted images.

[0038] The disclosed embodiments provide systems and methods that include machine learning models that address some or all of these drawbacks. The disclosed embodiments provide systems and methods that can use machine learning models to generate predicted SEM images of wafers after development and SEM images of wafers after etching, thereby allowing for less or no wafer damage, more accurate wafer measurements, and higher throughput.

[0039] Figure 1 The illustration depicts a charged particle beam device in which the electron beam system may include a single primary beam that can be configured to generate a secondary beam. A detector may be positioned along optical axis 105, as shown below. Figure 1 As shown. In some embodiments, the detector can be arranged off-axis.

[0040] like Figure 1 As shown, the electron beam tool 104 may include a wafer holder 136 supported by a motorized stage 134 to hold a wafer 150 to be inspected. The electron beam tool 104 may be a single-beam system or a multi-beam system. The electron beam tool 104 includes an electron beam source, which may include a cathode 103, an anode 120, and a bore 122. The electron beam tool 104 also includes a beam confinement aperture 125, a condenser lens 126, a column aperture 135, an objective lens assembly 132, and an electron detector 144. In some embodiments, the objective lens assembly 132 may be a modified oscillating objective decelerating immersion lens (SORIL), which includes a pole piece 132a, a control electrode 132b, a deflector 132c, and an excitation coil 132d. During imaging, the electron beam 161 emitted from the tip of the cathode 103 can be accelerated by the voltage of the anode 120, pass through the gun hole 122, the beam-limiting aperture 125, the condenser lens 126, and be focused into a probe point by a modified SORIL lens before incident on the surface of the mask 170. The probe spot can be scanned across the surface of the wafer 150 by deflectors (such as deflector 132c in the SORIL lens or other deflectors). Secondary electrons emitted from the wafer can be collected by the detector 144 to form an image of the region of interest on the wafer 150.

[0041] An image processing system 199 may also be provided, comprising an image acquirer 200, a storage device 130, and a controller 109. The image acquirer 200 may include one or more processors. For example, the image acquirer 200 may include a computer, server, mainframe, terminal, personal computer, any type of mobile computing device, or a combination thereof. The image acquirer 200 may be connected to the detector 144 of the electron beam tool 104 via a medium such as an electrical conductor, fiber optic cable, portable storage medium, IR, Bluetooth, the Internet, wireless network, radio, or a combination thereof. The image acquirer 200 may receive signals from the detector 144 and may construct an image. Thus, the image acquirer 200 may acquire an image of the wafer 150. The image acquirer 200 may also perform various post-processing functions, such as generating contours, overlaying indicators on the acquired image, etc. The image acquirer 200 may be configured to perform adjustments to the brightness and contrast of the acquired image, etc. The storage device 130 may be a storage medium such as a hard disk, random access memory (RAM), cloud storage, or other types of computer-readable storage. Storage device 130 may be coupled to image acquirer 200 and may be used to save scanned raw image data as an initial image and to save post-processed images. Image acquirer 200 and storage device 130 may be connected to controller 109. Controller 109 may be electrically connected to electron beam tool 104. Controller 109 may be a computer configured to perform various controls on electron beam tool 104. In some embodiments, image acquirer 200, storage device 130, and controller 109 may be integrated together as a single control unit.

[0042] In some embodiments, the image acquirer 200 may acquire one or more images of the wafer based on imaging signals received from the detector 144. The imaging signals may correspond to a scanning operation for imaging charged particles. The acquired image may be a single image comprising multiple imaging regions, which may contain various features of the wafer 150. The single image may be stored in the storage device 130. Imaging may be performed based on imaging frames.

[0043] The concentrator and illumination optics of an electron beam tool may include or be supplemented by an electromagnetic quadrupole electron lens. For example, such as Figure 1 As shown, the electron beam tool 104 may include a first quadrupole lens 148 and a second quadrupole lens 158. In some embodiments, the quadrupole lenses are used to control the electron beam. For example, the first quadrupole lens 148 may be controlled to adjust the beam current, and the second quadrupole lens 158 may be controlled to adjust the beam spot size and beam shape. It should be understood that any number of poles and any number of lenses may be used appropriately.

[0044] althoughFigure 1 Electron beam tool 104 is shown as a single-beam inspection tool that can use only one primary electron beam to scan one location on wafer 150 at a time, but embodiments of the invention are not limited thereto. For example, electron beam tool 104 could also be a multi-beam inspection tool that uses multiple primary electron beams to simultaneously scan multiple locations on wafer 150.

[0045] Figure 2A A flowchart illustrating an exemplary method for wafer inspection is shown. Figure 2A As shown, the wafer inspection process 200A is typically used to inspect wafers (e.g., Figure 1 The wafer 150). The wafer manufacturing process may include, but is not limited to, development (i.e., photolithography), etching, chemical mechanical polishing, or interconnect formation processes. In step 202A, the inspection system (using, for example...) Figure 1 An electron beam tool 104 can acquire a developed wafer. In step 204A, the system can then acquire an SEM image of a portion of the developed wafer. This portion can be the entire developed wafer. Immediately after step 204A, just before step 205A, the photoresist on this portion of the wafer may have been damaged by the SEM image capture. SEM inspection may damage the photoresist on the wafer, thus altering the pattern of the developed wafer. This alteration to the wafer pattern may result in modified critical performance parameters (e.g., defects, line edge roughness, line width roughness, local critical dimension uniformity, etc.) that do not reflect the true critical performance parameters of the wafer when not inspected by SEM after development (e.g., see...). Figure 3 ).

[0046] Although post-development imaging may damage parts of the inspected wafer, it may be necessary to use the SEM image acquired in step 204A to inspect the wafer and observe the direct effects of various exposure process conditions (e.g., focal length, dose, etc.) on the wafer after development. For example, the SEM image acquired in step 204A can be used for defect detection (e.g., necking, bridging, etc.), roughness / randomness characterization (e.g., line edge / width roughness, local critical size uniformity, edge placement error), process window optimization, calibration of computational lithography models, and process performance monitoring and control. Furthermore, the SEM image acquired in step 204A may be needed to generate a predictive image of the wafer after etching, allowing for wafer reprocessing based on inspection measurements. Since the wafer cannot be returned to its previous processing state after etching, reprocessing can improve post-development inspection results. Therefore, it is important to obtain SEM images of the wafer's printed pattern after development and before etching.

[0047] In step 206A, the system can acquire the wafer after it has been etched. In step 208A, the system can then acquire an SEM image of the etched wafer. Using the SEM image acquired in step 208A to examine the wafer may be desirable, to observe the etching effect on the wafer.

[0048] Figure 2B A flowchart illustrating an exemplary method for wafer inspection is shown. Figure 2B As shown, wafer inspection process 200B is ideal. Similar to process 200A, in step 202B, the inspection system can acquire the developed wafer. Process 200B can then proceed to step 206B, where the system can acquire the etched wafer. In step 208B, the system can acquire a SEM image of the etched wafer. Potential damage to the wafer is avoided in wafer inspection process 200B because this process does not include acquiring a SEM image of the wafer after development.

[0049] To obtain a SEM image of the developed wafer in process 200B, a machine learning model can be used. Because this SEM inspection does not damage or affect the wafer after etching, the SEM image obtained in step 208B can be used to generate a predicted SEM image of the wafer after development.

[0050] Figure 3 The diagram illustrates the distribution of critical dimensions obtained from SEM images of a wafer. Distribution 302 represents the measurement results of the wafer's critical dimensions based on SEM images obtained after etching but without any SEM inspection after development. Since SEM inspection after etching does not affect the wafer, the measurement results in distribution 302 also represent the measurement results of an undamaged wafer after development. Distribution 304 represents the measurement results of the wafer's critical dimensions based on SEM images obtained after both etching and development-related SEM inspections. Since SEM inspection of the wafer after development may affect and damage the wafer, the measurement results in distribution 304 are also affected by SEM inspection of the wafer after development. Figure 3 As shown, distribution 304 is offset from distribution 302. Distribution 306 represents the measurement result of the wafer's critical dimensions based on the SEM image sampled from distribution 304 in a manner that replicates distribution 302. Figure 3 As shown, distribution 306 has fewer points than distribution 302, but the critical dimensions of distributions 302 and 306 are the same. The etched SEM image from distribution 306 and its corresponding developed pair can be used to train a machine learning model. Therefore, a machine learning model using the etched SEM image from distribution 306 can accurately generate the developed SEM image of the wafer without SEM-induced damage.

[0051] Figure 4 The illustration shows a top view of wafer 400, which is segmented to show points for wafer inspection. It should be noted that the lines and dots on wafer 400 are for illustrative purposes and not physical portions of the wafer. Wafer 400 may have clusters 402 (i.e., segments of the wafer) for wafer inspection. For example, SEM images of segment 404 can be obtained for clusters after development following the application of photoresist to the wafer and after etching of the wafer. SEM images of different segments 406 can be obtained for clusters after etching of the wafer. Since post-development SEM inspection may damage the photoresist on the wafer, resulting in altered critical performance indicators (CPIs) that do not reflect the true CPIs of the wafer after development, a single post-development image can be obtained from segment 404. An etched image can also be obtained from segment 404, while only an etched image can be obtained from segment 406. Advantageously, only one post-development image can be obtained from a single cluster 402.

[0052] As described above, the SEM image of portion 406 can be undamaged and can be used in a machine learning model to generate a predicted SEM image of the wafer after development. SEM images of portion 404 (the first image obtained after development and the second image obtained after etching) can still be obtained to train the machine learning model. That is, process 200A (see...) Figure 2A This can be illustrated as a process for acquiring SEM images used to train machine learning models, while process 200B (see...) Figure 2B The process for acquiring SEM images for applying machine learning models can be illustrated. These processes can be implemented for at least some of the clusters 402 of wafer 400, since each cluster 402 can be affected differently by wafer etching. Processes 200A and 200B can be applied only to clusters 402 in the radial direction, because the wafer etching effect on wafer 400 can differ only in the radial direction. Advantageously, only a small portion of the wafer may be damaged by post-development SEM inspection, because for each cluster, only one pair of training images, one developed image, and one etched image are acquired.

[0053] In another embodiment, the machine learning model can be tuned to include parameters characterizing the wafer etching process (e.g., temperature, pressure, bias voltage, etc.). In this case, inputting the etching parameters into the model takes into account etching variability, making clustering unnecessary. That is, only one post-development image and its corresponding post-etched image can be acquired for training the model. Advantageously, since only one pair of training images is acquired for the entire wafer, even smaller portions of the wafer may be damaged by post-development SEM inspection.

[0054] In another embodiment, as described above, a machine learning model trained on a first wafer (for each cluster of wafers or by adjusting parameters to include characterization of the wafer etching process) can be applied to a second wafer. The second wafer can be imaged after etching, and the post-etched SEM image can be used to generate a predicted image of the second wafer after development by applying the machine learning model trained on the first wafer. Advantageously, the second wafer is not damaged during inspection because post-development SEM inspection is not required in this embodiment.

[0055] In some embodiments, SEM images of portion 404 can be obtained for clusters after development following the application of photoresist to the wafer and after etching of portion 404. SEM images of different portions 406 can be obtained for clusters after development following the application of photoresist to the wafer. Since portions of the wafer cannot be returned to their previous processing state after etching, only portion 404 can be etched, and a single post-etched image can be obtained from portion 404, allowing the wafer to be reprocessed to improve post-etched inspection results. A developed image can also be obtained from portion 404, while only a developed image can be obtained from portion 406. Advantageously, only a portion of cluster 402 can be etched, and only one post-etched image can be obtained from one cluster 402.

[0056] As described above, the SEM image of portion 406 can be left unetched and used in a machine learning model to generate a predicted SEM image of the wafer after etching. SEM images of portion 404 (the first image obtained after development and the second image obtained after etching) can still be obtained to train the machine learning model. That is, process 200A (see...) Figure 2A The process for acquiring SEM images for training a machine learning model can be illustrated. This process can be implemented for at least some of the clusters 402 of wafer 400, since each cluster 402 may be affected differently by wafer etching. Process 200A can be applied only to clusters 402 in the radial direction, because the wafer etching effect on wafer 400 can differ only in the radial direction. Advantageously, since only one pair of training images, one developed image, and one etched image are acquired for each cluster, only a small portion of the wafer can be etched (e.g., it may not be reprocessed).

[0057] In some embodiments, the machine learning model can be tuned to include parameters characterizing the wafer etching process (e.g., temperature, pressure, bias voltage, etc.). In this case, inputting the etching parameters into the model takes into account etching variability, making clustering unnecessary. That is, only one post-development image and its corresponding post-etched image can be acquired for training the model. Advantageously, even smaller portions of the wafer can be etched after development, since only a pair of training images are acquired for the entire wafer.

[0058] In some embodiments, as described above, a machine learning model trained on a first wafer (for each cluster of wafers or by adjusting parameters to include characterization of the wafer etching process) can be applied to a second wafer. In some embodiments, the machine learning model can be trained on the first wafer after the entire wafer has been etched. The second wafer can be imaged after development of the photoresist applied to the wafer, and the post-development SEM image can be used to generate a predicted image of the second wafer after etching by applying the machine learning model trained on the first wafer. Advantageously, the second wafer can be reprocessed during inspection before etching, since post-etching SEM inspection is not required in this embodiment.

[0059] Figure 5A The illustration shows SEM images of a wafer. SEM image 502 (i.e., the post-development image) depicts the vias on the wafer after development but before etching. SEM image 504 depicts the vias on the wafer after etching. SEM image 504 is used in a trained machine learning model to generate SEM image 506, which is a predicted SEM image of the vias on the wafer after development. Although in Figure 5A While not immediately obvious, the predicted feature size in SEM image 506 is essentially similar to the feature size in SEM image 502, without the damage caused by post-development SEM examination.

[0060] Figure 5B The illustration shows SEM images of a wafer. SEM image 512 (i.e., the post-development image) depicts vias in the photoresist layer on the wafer after development but before etching. SEM image 514 depicts vias in the layer on the wafer below the etched photoresist layer. SEM image 512 is used in a trained machine learning model to generate a binarized SEM image 516, which is a predicted SEM image of the vias on the etched wafer. The feature sizes in the predicted SEM image 516 are substantially similar to the feature sizes in the binarized SEM image 514, indicating that the trained machine learning model can be used to reprocess wafers after development but before etching.

[0061] Figure 6The illustration shows a predicted SEM image used to generate a wafer (e.g., Figure 5A SEM image 506, Figure 5B A system 600 (for binarized SEM images 516) is provided. System 600 may include a wafer inspection system 610, a model training server 620, and an image prediction server 630. System 610, model training server 620, and image prediction server 630 may be physically (e.g., via cable) or remotely electrically coupled to each other. The wafer inspection system 610 may be a reference... Figure 1 The described system is used to acquire wafers (e.g., see...). Figure 1 Image of chip 150.

[0062] The model training server 620 may include a processor 622 and a storage device 624. The model training server 620 may also include a communication interface 626 to receive wafer images from the wafer inspection system 610. The processor 622 may be configured to analyze the received wafer images and transform the images into one or more feature maps containing visual features of the wafer by applying a convolutional neural network (CNN). The processor 622 may also be configured to map visual features associated with a first wafer image to visual features associated with a second wafer image. The processor 622 may be configured to train a CNN encoder to construct a second wafer image based on the first wafer image by using the mapping. Training features of the wafer images can be extracted from the structure and input into the CNN decoder. The processor 622 can adjust the weights of the CNN decoder by training the CNN decoder on the training features of the wafer images. Training data may be stored in the storage device 624.

[0063] Image prediction server 630 may include processor 632 and storage device 634. Image prediction server 630 may also include communication interface 636 to receive wafer images from wafer inspection system 610. Processor 632 may be configured to analyze the received wafer images and transform the images into one or more feature maps containing visual features of the wafer by applying a CNN. Processor 632 may also be configured to extract visual features associated with the wafer images using a CNN encoder, wherein extraction includes learning hierarchical features from the wafer images using one or more convolutional pooling layers. Processor 632 may identify the training features used from the extracted visual features and input the training features into CNN decoder. Processor 632 may then generate a predicted image corresponding to the wafer images received from wafer inspection system 610. Prediction data may be stored in storage device 634. As described above, a predicted image for a second wafer can be generated by applying a CNN trained on a first wafer.

[0064] Figure 7AThe illustration shows a schematic diagram of an exemplary machine learning model training architecture. This architecture may include a CNN 700A. The CNN 700A may include a chip (e.g., Figure 1 The image includes a developed image 702A of a portion of the wafer 150 and one or more etched images 704A of a segment of the wafer. The portion may be the entire wafer, and the segment may be the entire wafer. While the developed image 702A may include multiple images 706A, and the etched image 704A may include multiple images 708A, one or more etched images cover more of the wafer segment than a developed image that only covers a portion of the segment. As described above, the portion of the developed wafer to be imaged can be minimized in an attempt to avoid potential damage to the wafer.

[0065] CNN 700A can be trained using a developed image 706A and an etched image 708A. The etched image 708A can be a first portion of the image fragment of the etched wafer, where the first portion corresponds to the image portion of the wafer in the developed image 706A. Training can include transforming the developed image 706A into one or more feature maps containing visual features of the developed image 706A. Training can also include transforming the etched image 708A into one or more feature maps containing visual features of the image portion of the etched wafer. Visual features associated with the etched image 708A can be mapped to visual features associated with the developed image 706A.

[0066] The values ​​of the model parameters in each layer of the CNN 700A deep learning architecture can be learned based on the mapping. Each label in each layer can represent an attribute of the SEM input image. The mapping can be used to train the CNN encoder to construct the developed image 706A. This construction can output a reconstructed developed image 710A. Training features can be extracted from the construction and fed into the CNN decoder, where the weights of the CNN decoder are adjusted by training the CNN decoder on the training features. That is, the weights of the CNN decoder are adjusted such that the difference between the developed image 706A and the reconstructed developed image 710A is minimized. In some embodiments, the CNN 700A can use a loss function learned during training (e.g., using a generative adversarial network). In some embodiments, other loss functions can be used (e.g., mean squared error, quadratic loss, L2 loss, mean absolute error, L1 loss, mean bias error, etc.). In some embodiments, the training features can include the etching effect between the development and etching processes of the wafer. SEM noise can be averaged out during training because the noise is random, resulting in virtually or substantially noise-free predictions during the application of the machine learning model.

[0067] Figure 7BThe illustration shows a schematic diagram of an exemplary machine learning model training architecture. This architecture may include a CNN 700B. The CNN 700B may include a chip (e.g., Figure 1 The image includes an etched image 702B of a portion of the wafer 150 and one or more developed images 704B of a segment of the wafer. The portion may be the entire wafer, and the segment may be the entire wafer. While the etched image 702B may include multiple images 706B and the developed image 704B may include multiple images 708B, one or more developed images may cover more of the wafer segment than an etched image that only covers a portion of the segment. As described above, the portion of the wafer to be etched and imaged after etching can be minimized in an attempt to reprocess the wafer before it is fully etched.

[0068] CNN 700B can be trained using an etched image 706B and a developed image 708B. The developed image 708B can be a first portion of the image fragment of the developed wafer, where the first portion corresponds to the image portion of the wafer in the etched image 706B. Training can include transforming the etched image 706B into one or more feature maps containing visual features of the etched image 706B. Training can also include transforming the developed image 708B into one or more feature maps, which contain visual features of the image portion of the developed wafer. Visual features associated with the developed image 708B can be mapped to visual features associated with the etched image 706B.

[0069] The values ​​of the model parameters in each layer of CNN 700B can be learned based on the mapping. Each label in each layer can represent an attribute of the SEM input image. The mapping can be used to train the CNN encoder to construct the etched image 706B. This construction can output a reconstructed etched image 710B. Training features can be extracted from the construction and fed into the CNN decoder, where the weights of the CNN decoder are adjusted by training the CNN decoder on the training features. That is, the weights of the CNN decoder are adjusted such that the difference between the etched image 706B and the reconstructed etched image 710B is minimized. In some embodiments, CNN 700B can use a loss function learned during training (e.g., using a generative adversarial network). In some embodiments, other loss functions can be used (e.g., mean squared error, quadratic loss, L2 loss, mean absolute error, L1 loss, mean bias error, etc.). In some embodiments, the training features can include the etching effect between the wafer development process and the etching process. SEM noise is averaged out during training because the noise is random, resulting in virtually noise-free predictions during the application of the machine learning model.

[0070] In some embodiments, a linear activation function can be used in the output layer to train the CNN 700B to output a SEM image that provides grayscale values ​​(e.g., Figure 5A Image 506). Grayscale SEM images may be needed to generate predicted images that simulate actual SEM images. In some embodiments, a sigmoid activation function can be used in the output layer to train the CNN 700B to output a binarized SEM image (e.g., Figure 5B Image 516). A binarized SEM image may be needed to generate a prediction image that allows for easy extraction of key performance indicators (e.g., key dimensions) from the generated prediction image. For example, the black portion of the generated prediction image may indicate a 100% probability that the model predicts the contact hole exists in the corresponding portion of the wafer. The white portion of the generated prediction image may indicate a 100% probability that the model predicts the contact hole does not exist in the corresponding portion of the wafer. The gray portion of the generated prediction image (e.g., the edge of the contact hole) may indicate a probability between 0% and 100% that the model predicts the contact hole exists in the corresponding portion of the wafer (e.g., the model is somewhat uncertain).

[0071] Figure 8A The illustration shows a schematic diagram of an exemplary machine learning model application architecture. This architecture may include a CNN 800A. The CNN 800A may include a trained model (e.g., see...). Figure 7A (CNN 700A), and may include one or more etched images 804A of a segment of the wafer. Etched image 808A may be a second portion of an imaging segment of the wafer, wherein the second portion differs from the first portion (e.g., see...). Figure 7A The first part of the etched image 708A).

[0072] The trained model can include a deep neural network topology with multiple layers. For example, such a network architecture could include an etched image 808A as input, a first convolution, a first pooling, a second convolution, a second pooling, one or more hidden layers, activations, and an output. Depending on the nature of the defect pattern's characteristics, each layer of the architecture can have a different number of generated subsamples. For example, after the first convolution operation, fewer than ten subsamples can be generated in the first pooling. After the second convolution operation, the second layer can have more than ten subsamples generated in the second pooling.

[0073] After the model parameters have been determined from training using the imaging portions of the developed wafer and the corresponding imaging portions of the etched wafer, a trained network of the deep learning architecture can process new images of other portions of the etched wafer. The etched image 808A can be processed through multiple layers of the trained network with the determined model parameters to generate feature vectors at the ends of the architecture. At each layer of convolutional computation, the etched image 808A can be analyzed for the architecture to determine if training features exist, and if so, their geometric location. Each received image with training features can be processed by computation layer by layer across multiple layers. Therefore, a vector representation of the input image can be generated by the trained network.

[0074] Applying a trained CNN 800A may include transforming a second portion of an etched image 808A into one or more feature maps by applying the CNN 800A, wherein the one or more feature maps contain visual features of the second portion of the etched image 808A. The etched image 808A, as input to the network 800A, may include a pattern image that includes identifying features of the etched wafer. The application may also include extracting visual features associated with the second portion of the etched image 808A using a CNN encoder, wherein extraction includes learning hierarchical features based on the etched image 808A using one or more convolutional pooling layers. The application may also include using the extracted visual features to label training features. Training features of the etched image 808A can be automatically labeled.

[0075] The application of the trained model can also include inputting training features into a CNN decoder and generating a predicted wafer image 810A, wherein the predicted image 810A includes other portions of the wafer fragments developed before etching. As described above, this can be achieved by applying it to the first wafer (e.g., see...). Figure 7A CNNs trained on (e.g.) Figure 8A This is used to generate a predicted image for the second wafer.

[0076] In some embodiments, statistical analysis can be performed based on the etched image, the developed image, and the predicted image (e.g., see...). Figure 3 ), to generate statistical characteristics of SEM damage to the wafer.

[0077] In some embodiments, the CNN 800A can use a linear activation function in the output layer to output a SEM image that provides grayscale values ​​(e.g., Figure 5A Image 506). Grayscale SEM images may be needed to generate predicted images that simulate the actual SEM images.

[0078] Figure 8BThe illustration shows a schematic diagram of an exemplary machine learning model application architecture. This architecture may include a CNN 800B. The CNN 800B may include a trained model (e.g., see...). Figure 7B (CNN 700B), and may include one or more post-development images 804B of a segment of the wafer. Post-development image 808B may be a second portion of an imaging segment of the wafer, wherein the second portion differs from the first portion (e.g., see...). Figure 7B The first part of the developed image 708B).

[0079] The trained model can include a deep neural network topology with multiple layers. For example, such a network architecture could include a post-development image 808B (as input), a first convolution, a first pooling, a second convolution, a second pooling, one or more hidden layers, activations, and an output. Depending on the nature of the defect pattern's characteristics, each layer of the architecture can have a different number of generated subsamples. For example, after the first convolution operation, fewer than ten subsamples can be generated in the first pooling. After the second convolution operation, the second layer can have more than ten subsamples generated in the second pooling.

[0080] After the model parameters have been determined from training using the imaging portions of the developed wafer and the corresponding imaging portions of the etched wafer, a trained network of a deep learning architecture can process new images of the developed wafer or other portions of another wafer. The developed image 808B can be processed through multiple layers of the trained network with the determined model parameters to generate feature vectors at the ends of the architecture. At each layer of convolutional computation, the developed image 808B can be analyzed by architecture analysis to determine whether training features exist, and if so, their geometric location. Each received image with training features can be processed by computation layer-by-layer across multiple layers. Therefore, a vector representation of the input image can be generated by the trained network.

[0081] Applying a trained CNN 800B may include transforming a second portion of a developed image 808B into one or more feature maps by applying the CNN 800B, wherein the one or more feature maps contain visual features of the developed image of the second portion 808B. The developed image 808B, as input to the network 800B, may include a pattern image that includes recognizable features of the developed wafer. The application may also include using a CNN encoder to extract visual features associated with the developed image of the second portion 808B, wherein extraction includes using one or more convolutional pooling layers to learn hierarchical features based on the developed image 808B. The application may also include using the extracted visual features to identify training features. The training features of the developed image 808B may be automatically labeled.

[0082] The application of the trained model can also include inputting training features into a CNN decoder and generating a predicted wafer image 810B, wherein the predicted image 810B includes other portions of the etched wafer fragment. As described above, this can be achieved by applying the model to a first wafer (see, for example...). Figure 7B CNNs trained on (e.g.) Figure 8B This is used to generate a predicted image for the second wafer.

[0083] In some embodiments, statistical analysis may be performed based on the etched image, the developed image, and the predicted image (see, for example...). Figure 3 ), to generate statistical characteristics of SEM damage to the wafer.

[0084] In some embodiments, the CNN 800B can use a sigmoid activation function in the output layer to output a binarized SEM image (e.g., Figure 5B Image 516). A binarized SEM image may be needed to generate a prediction image that allows for easy extraction of key performance indicators (e.g., key dimensions) from the generated prediction image. For example, the black portion of the generated prediction image may indicate a 100% probability that the model predicts the contact hole exists in the corresponding portion of the wafer. The white portion of the generated prediction image may indicate a 100% probability that the model predicts the contact hole does not exist in the corresponding portion of the wafer. The gray portion of the generated prediction image (e.g., the edge of the contact hole) may indicate a probability between 0% and 100% that the model predicts the contact hole exists in the corresponding portion of the wafer (e.g., the model is somewhat uncertain).

[0085] Figure 9A The illustration shows a flowchart illustrating an exemplary method of wafer inspection 900A according to an embodiment of the present disclosure. Method 900A can be performed by a system (e.g., Figure 6 The system 600) is executed. This system may include an inspection system (e.g., that performs one or more steps of method 900A). Figure 6 Inspection system 610). For example, the inspection system can use an electron beam tool (e.g., Figure 1 Electron beam tool 104) for wafers (e.g., Figure 1 Imaging of different regions of the chip 150.

[0086] In step 901A, the system (e.g., Figure 6 The system 600 can acquire a wafer after the photoresist applied to the wafer has been developed.

[0087] In step 903A, the system may use an image acquirer (e.g., Figure 1 The image acquisition device 200) captures a portion of a segment of the developed wafer (e.g., see image acquisition device 200).Figure 4 Imagery can be performed on part 404. This part can be the entire segment. The image acquirer can obtain the image from the detector of the electron beam tool (e.g., Figure 1 The detector 144 receives the signal and constructs an image, thereby acquiring the developed wafer (e.g., Figure 1 An image of a portion of a segment of a wafer (150). The portion of the wafer to be imaged after development can be minimized in an attempt to avoid potential damage to the wafer. Figure 5A An example of the developed image is provided at position 502 in the image.

[0088] In step 905A, the system (e.g., Figure 6 The system (600) can acquire the wafer after it has been etched.

[0089] In step 907A, the system may use an image acquirer (e.g., Figure 1 The image acquisition unit 200) is used to image the portion of the developed wafer (e.g., see image acquisition unit 200). Figure 4 Imagers can image fragments of the etched wafer corresponding to (402). The imager can acquire images from the detector of an electron beam tool (e.g., Figure 1 The detector 144 receives the signal and constructs an image, thereby acquiring the etched wafer (e.g., Figure 1 An image of a fragment of chip 150. Figure 5A An example of the etched image is provided at image 504. The etched image covers more of the crystal fragment than the developed image, which only covers a portion of the fragment, to minimize potential damage to the wafer. As mentioned above, the developed image is used only to train the machine learning model, while the etched image is used to train the machine learning model and generate a predicted image of the developed wafer.

[0090] In step 909A, the system can use the imaged portion of the developed wafer and the imaged fragment of the etched wafer to train a machine learning model. The system including the inspection system may also include a model training server (e.g., Figure 6 The model training server 620 may include a processor (e.g., Figure 6 The processor 622), storage device (e.g., Figure 6 Storage device 624) and communication interface (e.g., Figure 6 The communication interface 626 is used to communicate with the inspection system (e.g., Figure 6 The inspection system 610 receives wafer images. The processor can be configured to analyze the received wafer images and apply a CNN (e.g., ...) to them. Figure 7AThe processor (using a CNN 700A) transforms an image into one or more feature maps containing visual features of the wafer. The processor can also be configured to map visual features associated with the etched image to visual features associated with the developed image. The processor can be configured to train a CNN encoder using the mapping to construct a developed image from the etched image. Training features of the wafer image can be extracted from the construction and fed into the CNN decoder. The processor can adjust the weights of the CNN decoder by training the CNN decoder on the training features of the wafer image. That is, adjusting the weights of the CNN decoder to minimize the difference between the developed image and the constructed developed image. Training data can be stored in a storage device.

[0091] In step 911A, the system can use the image fragment of the etched wafer to apply a trained machine learning model to generate other portions of the developed wafer fragment (e.g., see...). Figure 4 The predicted image (part 406) or an image of another chip. A system including an inspection system and a model training server may also include an image prediction server (e.g., Figure 6 The image prediction server 630. The inspection system, model training server, and image prediction server can be electrically coupled to each other, either physically (e.g., via cable) or remotely. The image prediction server may include a processor (e.g., Figure 6 The processor 632), storage device (e.g., Figure 6 Storage device 634) and communication interface (e.g., Figure 6 The communication interface 636 is used to communicate with the inspection system (e.g., Figure 6 The inspection system 610 receives wafer images. The processor can be configured to analyze the received wafer images and apply a CNN (e.g., ...) to them. Figure 8A The processor uses a CNN 800A to transform the image into one or more feature maps containing visual features of the wafer. The processor can also be configured to use a CNN encoder to extract visual features associated with an etched image, which is different from the etched image used for training. Extraction involves using one or more convolutional pooling layers to learn hierarchical features from the etched image. The processor can identify the trained features used from the extracted visual features and feed the trained features into the CNN decoder. The processor can then generate a predicted image of the developed wafer corresponding to the etched image received from the inspection system. The predicted data can be stored in a storage device. Figure 5A An example of the predicted image is provided at position 506 in the image.

[0092] Figure 9BThe illustration shows a flowchart illustrating an exemplary method of wafer inspection 900B according to an embodiment of the present disclosure. Method 900B can be performed by a system (e.g., Figure 6 The system 600) is executed. This system may include an inspection system (e.g., that performs one or more steps of method 900B). Figure 6 Inspection system 610). For example, the inspection system can use an electron beam tool (e.g., Figure 1 Electron beam tool 104) for wafers (e.g., Figure 1 Imaging of different regions of the chip 150.

[0093] In step 901B, the system (e.g., Figure 6 The system 600 can acquire a wafer after the photoresist applied to the wafer has been developed.

[0094] In step 903B, the system may use an image acquirer (e.g., Figure 1 Image acquisition device 200) for segments of the developed wafer (e.g., see image acquisition device 200). Figure 4 Imaging is performed using the 402 (electron beam tool). This segment can be a portion of or the entire developed wafer. The image acquirer can obtain the image from the detector of the electron beam tool (e.g., Figure 1 The detector 144 receives the signal and constructs an image, thereby acquiring the developed wafer (e.g., Figure 1 An image of a segment or the entire chip (150). Figure 5B An example of the developed image is provided at position 512 in the image.

[0095] In step 905B, the system (e.g., Figure 6 The system (600) can acquire the wafer after it has been etched.

[0096] In step 907B, the system may use an image acquirer (e.g., Figure 1 The image acquisition unit 200) is used to image segments of the developed wafer (e.g., see image acquisition unit 200). Figure 4 Imager can image a portion of the etched wafer segment corresponding to part 404. This portion can be the entire segment. The imager can acquire the image from the detector of an electron beam tool (e.g., Figure 1 The detector 144 receives the signal and constructs an image, thereby acquiring the etched wafer (e.g., Figure 1 An image of a portion of a fragment of chip 150. Figure 5BAn example of a post-etched image is provided at image 514. The etched portion of the wafer and the portion of the post-etched wafer to be imaged can be minimized in an attempt to allow for further processing of the wafer prior to etching. As mentioned above, the post-etched image is used to train a machine learning model and generate a predicted image of the post-etched wafer, while the post-etched image is only used to train the machine learning model.

[0097] In step 909B, the system can use the etched imaging portion of the wafer and the developed imaging fragment of the wafer to train a machine learning model. The system including the inspection system may also include a model training server (e.g., Figure 6 The model training server 620 may include a processor (e.g., Figure 6 The processor 622), storage device (e.g., Figure 6 Storage device 624) and communication interface (e.g., Figure 6 The communication interface 626 is used to communicate with the inspection system (e.g., Figure 6 The inspection system 610 receives wafer images. The processor can be configured to analyze the received wafer images and apply a CNN (e.g., ...) to them. Figure 7B The processor (using a CNN 700B) transforms an image into one or more feature maps containing visual features of the wafer. The processor can also be configured to map visual features associated with the developed image to visual features associated with the etched image. The processor can be configured to train a CNN encoder using the mapping to construct the etched image from the developed image. Training features of the wafer image can be extracted from the construction and fed into the CNN decoder. The processor can adjust the weights of the CNN decoder by training the CNN decoder on the training features of the wafer image. That is, adjusting the weights of the CNN decoder to minimize the difference between the etched image and the constructed etched image. Training data can be stored in a storage device.

[0098] In step 911B, the system can use the image fragment of the developed wafer to apply a trained machine learning model to generate other parts of the developed wafer fragment (e.g., see...). Figure 4 The system includes a predicted etched image (part 406) or an image of another wafer. The system may also include an image prediction server (e.g., ...). Figure 6 The image prediction server 630. The inspection system, model training server, and image prediction server can be electrically coupled to each other, either physically (e.g., via cable) or wirelessly. The image prediction server may include a processor (e.g., Figure 6 The processor 632), storage device (e.g., Figure 6 Storage device 634) and communication interface (e.g., Figure 6The communication interface 636 is used to communicate with the inspection system (e.g., Figure 6 The inspection system 610 receives wafer images. The processor can be configured to analyze the received wafer images and apply a CNN (e.g., ...) to them. Figure 8B The processor uses a CNN 800B to transform the image into one or more feature maps containing visual features of the wafer. The processor can also be configured to use a CNN encoder to extract visual features associated with a post-development image, which is different from the post-development image used for training. Extraction involves using one or more convolutional pooling layers to learn hierarchical features from the post-development image. The processor can identify the training features used from the extracted visual features and feed the training features into the CNN decoder. The processor can then generate a predicted post-etched image of the wafer corresponding to the post-development image received from the inspection system. The predicted data can be stored in a storage device. Figure 5B An example of the predicted image is provided at position 516 in the image.

[0099] Figure 10 The illustration shows a flowchart illustrating an exemplary method of wafer inspection 1000 according to an embodiment of the present disclosure. Method 1000 can be provided by a system (e.g., Figure 6 The system 600 is executed. This system may include an inspection system (e.g., [system name missing]) that performs one or more steps of method 1000. Figure 6 Inspection system 610). For example, the inspection system can use an electron beam tool (e.g., Figure 1 Electron beam tool 104) for wafers (e.g., Figure 1 Imaging of different regions of the chip 150.

[0100] In step 1001, the system (e.g., Figure 6 The system 600 can obtain a first set of images of the features of the wafer after the photoresist layer has been developed. The system can use an image acquisition device (e.g., Figure 1 The image acquisition device 200) captures a portion of a segment of the developed wafer (e.g., see image acquisition device 200). Figure 4 Imagery can be performed from the detector of the electron beam tool (e.g., 404). Figure 1 The detector 144 receives the signal and constructs an image, thereby acquiring the developed wafer (e.g., Figure 1 An image of a portion of a segment of a wafer (150). The portion of the wafer to be imaged after development is minimized in an attempt to avoid potential damage to the wafer. An example of a developed image is provided at image 502 in Figure 5.

[0101] In step 1003, the system (e.g.) Figure 6The system 600 can obtain a second set of images of the features of the wafer after it has been etched. This system can use an image acquisition device (e.g., Figure 1 The image acquisition unit 200) is used to image the portion of the developed wafer (e.g., see image acquisition unit 200). Figure 4 Imagers can image fragments of the etched wafer corresponding to (402). The imager can acquire images from the detector of an electron beam tool (e.g., Figure 1 The detector 144 receives the signal and constructs an image, thereby acquiring the etched wafer (e.g., Figure 1 An image of a fragment of wafer 150 is shown. An example of the etched image is provided at image 504 in Figure 5. The etched image covers more of the wafer fragment than the developed image, which only covers a portion of the fragment, to minimize potential damage to the wafer. As mentioned above, the developed image is used only to train the machine learning model, while the etched image is used to train the machine learning model and generate a predicted image of the developed wafer.

[0102] In step 1005, the system can use the first set of images and the second set of images to train a machine learning model. The system including the inspection system may also include a model training server (e.g., Figure 6 The model training server 620 may include a processor (e.g., Figure 6 The processor 622), storage device (e.g., Figure 6 Storage device 624) and communication interface (e.g., Figure 6 The communication interface 626 is used to communicate with the inspection system (e.g., Figure 6 The inspection system 610 receives wafer images. The processor can be configured to analyze the received wafer images and apply a CNN (e.g., ...) to them. Figure 7A The processor (using a CNN 700A) transforms an image into one or more feature maps containing visual features of the wafer. The processor can also be configured to map visual features associated with the etched image to visual features associated with the developed image. The processor can be configured to train a CNN encoder using the mapping to construct a developed image from the etched image. Training features of the wafer image can be extracted from the construction and fed into the CNN decoder. The processor can adjust the weights of the CNN decoder by training the CNN decoder on the training features of the wafer image. That is, adjusting the weights of the CNN decoder to minimize the difference between the developed image and the constructed developed image. Training data can be stored in a storage device.

[0103] In step 1007, the system can generate a third set of images based on approximate features of the photoresist layer applied to the wafer, wherein the approximate features exclude damage caused by SEM imaging after development of the photoresist layer. The system can use an image fragment of the etched wafer to apply a trained machine learning model to generate other portions of that fragment of the developed wafer (e.g., see...). Figure 4 The predicted image (part 406). A system including an inspection system and a model training server may also include an image prediction server (e.g., Figure 6 The image prediction server 630. The inspection system, model training server, and image prediction server can be electrically coupled to each other, either physically (e.g., via cable) or remotely. The image prediction server may include a processor (e.g., Figure 6 The processor 632), storage device (e.g., Figure 6 Storage device 634) and communication interface (e.g., Figure 6 The communication interface 636 is used to communicate with the inspection system (e.g., Figure 6 The inspection system 610 receives wafer images. The processor can be configured to analyze the received wafer images and apply a CNN (e.g., ...) to them. Figure 8A The processor uses a CNN 800A to transform the image into one or more feature maps containing visual features of the wafer. The processor can also be configured to use a CNN encoder to extract visual features associated with an etched image, which is different from the etched image used for training. Extraction involves using one or more convolutional pooling layers to learn hierarchical features from the etched image. The processor can identify the used training features from the extracted visual features and feed the training features into the CNN decoder. The processor can then generate a predicted image of the developed wafer corresponding to the etched image received from the inspection system. The predicted data can be stored in a storage device. An example of the predicted image is provided at image 506 in Figure 5.

[0104] In step 1009, the system can perform statistical analysis based on the first set of images, the second set of images, and the third set of images to generate statistical characteristics of SEM damage to the wafer. The system (e.g., Figure 6 The system 600 can perform statistical analysis by measuring the features of the wafer based on a first set of images, a second set of images, and a third set of images. An example of statistical analysis is provided by... Figure 3 The distribution map is provided. For example, statistical analysis may include a distribution map of critical dimensions obtained from SEM images of the wafer. The first distribution (e.g., Figure 3Distribution 302) can represent the measurement results of the critical dimensions of the wafer based on SEM images obtained after etching but without any SEM inspection after development. Since SEM inspection of the etched wafer does not affect the wafer, the measurement results of the first distribution also represent the measurement results of the wafer that is not damaged after development. The second distribution (e.g., Figure 3 Distribution 304 can represent the measurement results of the wafer's critical dimensions based on SEM images obtained after etching and after development SEM inspection. Since SEM inspection of the developed wafer can affect and damage the wafer, it also affects the measurement results of the second distribution. The second distribution can be shifted from the first distribution. A third distribution (e.g., Figure 3 The distribution 306 can represent the measurement result of the critical size of the wafer based on the SEM image sampled from the second distribution in a manner that replicates the first distribution. The third distribution may have fewer points than the first distribution, but the critical sizes of the first and third distributions can be the same. The etched SEM image from the third distribution and its corresponding developed pair can be used to train a machine learning model. Therefore, a machine learning model using the etched SEM image obtained from the third distribution can accurately generate the developed SEM image of the wafer without SEM-induced damage.

[0105] These embodiments may be further described using the following terms:

[0106] 1. A method for inspecting a wafer, comprising:

[0107] Obtain the wafer after the photoresist applied to it has been developed;

[0108] To image a portion of a developed wafer fragment;

[0109] Obtain the wafer after it has been etched;

[0110] Imaging fragments of an etched wafer; and

[0111] Machine learning models are trained using imaging portions of developed wafers and imaging fragments of etched wafers.

[0112] 2. The method according to Clause 1, wherein the imaging fragment of the etched wafer comprises multiple images.

[0113] 3. The method according to any one of Clauses 1 and 2, wherein the imaging portion of the developed wafer comprises multiple images.

[0114] 4. The method according to any one of clauses 1-3, wherein training the machine learning model includes:

[0115] An image fragment of an etched wafer is transformed into one or more feature maps by applying a convolutional neural network (CNN), wherein the one or more feature maps contain visual features of the image fragment of the etched wafer.

[0116] The image portion of the developed wafer is transformed into one or more feature maps by applying a CNN, wherein the one or more feature maps contain visual features of the image portion of the developed wafer.

[0117] Visual features associated with an image segment of an etched wafer are mapped to visual features associated with an image portion of a developed wafer.

[0118] Mapping is used to train a CNN encoder to construct an image portion of a developed wafer based on an image fragment of an etched wafer;

[0119] Extract training features from the construction.

[0120] The training features are input into the CNN decoder; and

[0121] The weights of the CNN decoder are adjusted by training the CNN decoder on the training features.

[0122] 5. The method according to Clause 4, wherein mapping visual features includes: mapping visual features associated with an imaging portion of the developed wafer to visual features associated with a first portion of an imaging segment of the etched wafer.

[0123] 6. The method according to Clause 5, wherein the imaged portion of the developed wafer corresponds to a first portion of the imaged segment of the etched wafer.

[0124] 7. The method according to any one of clauses 1-6, wherein the application of the trained machine learning model includes: using a second portion of the imaging fragment of the etched wafer.

[0125] 8. The method according to Clause 7, wherein a first portion of the image segment of the etched wafer is different from a second portion of the image segment of the etched wafer.

[0126] 9. The method according to any one of clauses 1-8, wherein the application of a trained machine learning model includes:

[0127] An image fragment of an etched wafer is transformed into one or more feature maps by applying a convolutional neural network (CNN), wherein the one or more feature maps contain visual features of the image fragment of the etched wafer.

[0128] A CNN encoder is used to extract visual features associated with an image fragment of an etched wafer, wherein the extraction includes using one or more convolutional pooling layers to learn hierarchical features based on the image fragment of the etched wafer.

[0129] Use the extracted visual features to identify the training features;

[0130] The training features are input into the CNN decoder; and

[0131] Generate predicted images of other parts of the developed wafer fragment.

[0132] 10. The method according to any one of clauses 1-9, wherein the machine learning model includes multiple machine learning models.

[0133] 11. The method according to Clause 10, wherein each machine learning model corresponds to a different segment of the wafer.

[0134] 12. The method according to any one of clauses 1-9, wherein the application of the trained machine learning model includes a machine learning model and input etch parameters that characterize the etching of the wafer.

[0135] 13. The method according to any one of clauses 1-9, wherein the application of a trained machine learning model includes: using an imaging fragment of the etched wafer to generate a predicted image of other portions of a fragment of the developed wafer.

[0136] 14. The method according to any one of clauses 1-6, wherein the wafer is a first wafer.

[0137] 15. The method pursuant to Clause 14 includes:

[0138] Obtain the second wafer after the second wafer has been etched;

[0139] Imaging a fragment of the etched second wafer; and

[0140] An image fragment of the etched second wafer is used to apply a trained machine learning model to generate a predicted image of the second wafer fragment after the photoresist applied to the second wafer has been developed.

[0141] 16. The method according to any one of clauses 14 and 15, wherein the application of a trained machine learning model includes:

[0142] An image fragment of the etched second wafer is transformed into one or more feature maps by applying a convolutional neural network (CNN), wherein the one or more feature maps contain visual features of the image fragment of the etched second wafer.

[0143] A CNN encoder is used to extract visual features associated with the image fragments of the etched second wafer, wherein the extraction includes using one or more convolutional pooling layers to learn hierarchical features based on the image fragments of the etched second wafer.

[0144] Use the extracted visual features to identify the training features;

[0145] The training features are input into the CNN decoder; and

[0146] Generate a predicted image of a fragment of the developed second wafer.

[0147] 17. A method for training a machine learning model, the method comprising:

[0148] The first set of images of a segment of the wafer after the photoresist applied to the wafer has been developed are input into the machine learning model;

[0149] A second set of images of fragments of the wafer after it has been etched will be input into the machine learning model; and

[0150] Based on the first set of images and the second set of images, the weights of the machine learning model are adjusted, wherein the trained machine learning model is configured to generate predicted images of other parts of the developed wafer fragment.

[0151] 18. The method pursuant to Clause 17 also includes:

[0152] A second set of images of an etched wafer fragment is transformed into one or more feature maps by applying a convolutional neural network (CNN), wherein the one or more feature maps contain visual features of the second set of images of the etched wafer fragment.

[0153] A first set of images of a portion of the developed wafer is transformed into one or more feature maps by applying a CNN, wherein the one or more feature maps contain visual features of the first set of images of the portion of the developed wafer.

[0154] Visual features associated with a second set of images of a fragment of the etched wafer are mapped to visual features associated with a first set of images of a portion of the developed wafer.

[0155] The CNN encoder is trained using mapping to construct a first set of images of a portion of the developed wafer from a second set of images of fragments of the etched wafer.

[0156] Extract training features from the construction.

[0157] The training features are input into the CNN decoder; and

[0158] The weights of the CNN decoder are adjusted by training the CNN decoder on the training features.

[0159] 19. The method according to Clause 18, wherein mapping visual features includes: mapping visual features associated with the set of images of a portion of the developed wafer to visual features associated with a set of images of a first portion of a fragment of the etched wafer.

[0160] 20. The method according to Clause 19, wherein a first set of images of a portion of a developed wafer corresponds to a second set of images of a first portion of a segment of an etched wafer.

[0161] 21. A charged particle multi-beam system for inspecting wafers, the system comprising:

[0162] A controller, comprising circuitry for the following:

[0163] Imaging a segment of a wafer after the photoresist applied to the wafer has been developed;

[0164] Imaging fragments of a wafer after it has been etched; and

[0165] Machine learning models are trained using imaging portions of developed wafers and imaging fragments of etched wafers.

[0166] 22. The system according to Clause 21, wherein the imaging fragment of the etched wafer comprises multiple images.

[0167] 23. A system according to any one of clauses 21 and 22, wherein the imaging portion of the developed wafer comprises multiple images.

[0168] 24. A system according to any one of clauses 21-23, wherein training a machine learning model comprises:

[0169] An image fragment of an etched wafer is transformed into one or more feature maps by applying a convolutional neural network (CNN), wherein the one or more feature maps contain visual features of the image fragment of the etched wafer.

[0170] The image portion of the developed wafer is transformed into one or more feature maps by applying a CNN, wherein the one or more feature maps contain visual features of the image portion of the developed wafer.

[0171] Visual features associated with an image segment of an etched wafer are mapped to visual features associated with an image portion of a developed wafer.

[0172] Mapping is used to train a CNN encoder to construct an image portion of a developed wafer based on an image fragment of an etched wafer;

[0173] Extract training features from the construction.

[0174] The training features are input into the CNN decoder; and

[0175] The weights of the CNN decoder are adjusted by training the CNN decoder on the training features.

[0176] 25. The system according to Clause 24, wherein mapping visual features includes: mapping visual features associated with an imaged portion of a developed wafer to visual features associated with a first portion of an imaged segment of an etched wafer.

[0177] 26. The system according to Clause 25, wherein the imaged portion of the developed wafer corresponds to a first portion of the imaged segment of the etched wafer.

[0178] 27. A system pursuant to any one of clauses 21-25, wherein the application of a trained machine learning model includes: using a second portion of an image fragment of an etched wafer.

[0179] 28. A system according to Clause 27, wherein a first portion of an image segment of an etched wafer is different from a second portion of an image segment of an etched wafer.

[0180] 29. A system according to any one of clauses 21-28, wherein the application of a trained machine learning model includes:

[0181] An image fragment of an etched wafer is transformed into one or more feature maps by applying a convolutional neural network (CNN), wherein the one or more feature maps contain visual features of the image fragment of the etched wafer.

[0182] A CNN encoder is used to extract visual features associated with an image fragment of an etched wafer, wherein the extraction includes using one or more convolutional pooling layers to learn hierarchical features based on the image fragment of the etched wafer.

[0183] Use the extracted visual features to identify the training features;

[0184] The training features are input into the CNN decoder; and

[0185] Generate predicted images of other parts of the developed wafer fragment.

[0186] 30. A system pursuant to any one of clauses 21-29, wherein the machine learning model comprises multiple machine learning models.

[0187] 31. A system according to Clause 30, wherein each machine learning model corresponds to a different segment of a chip.

[0188] 32. A system according to any one of clauses 21-29, wherein the application of a trained machine learning model includes a machine learning model and input etch parameters that characterize the etching of the wafer.

[0189] 33. A system according to any one of clauses 21-29, wherein the application of a trained machine learning model includes: using an image fragment of an etched wafer to generate a predicted image of other portions of a fragment of a developed wafer.

[0190] 34. A system pursuant to any one of clauses 21-26, wherein the chip is the first chip.

[0191] 35. The system pursuant to Clause 34 includes:

[0192] Obtain the second wafer after the second wafer has been etched;

[0193] Imaging a fragment of the etched second wafer; and

[0194] An image fragment of the etched second wafer is used to apply a trained machine learning model to generate a predicted image of the second wafer fragment after the photoresist applied to the second wafer has been developed.

[0195] 36. A system pursuant to any one of clauses 34 and 35, wherein the application of a trained machine learning model includes:

[0196] An image fragment of the etched second wafer is transformed into one or more feature maps by applying a convolutional neural network (CNN), wherein the one or more feature maps contain visual features of the image fragment of the etched second wafer.

[0197] A CNN encoder is used to extract visual features associated with the image fragments of the etched second wafer, wherein the extraction includes using one or more convolutional pooling layers to learn hierarchical features based on the image fragments of the etched second wafer.

[0198] Use the extracted visual features to identify the training features;

[0199] The training features are input into the CNN decoder; and

[0200] Generate a predicted image of a fragment of the developed second wafer.

[0201] 37. A non-transient computer-readable medium storing a set of instructions executable by at least one processor of a computing device to cause the computing device to perform a method for inspecting a wafer, the method comprising:

[0202] Imaging a segment of a wafer after the photoresist applied to the wafer has been developed;

[0203] Imaging segments of a wafer after it has been etched;

[0204] Machine learning models are trained using imaging portions of developed wafers and imaging fragments of etched wafers; and

[0205] An image fragment of an etched wafer is used to apply a trained machine learning model to generate predicted images of other parts of a developed wafer fragment.

[0206] 38. A non-transient computer-readable medium pursuant to Clause 37, wherein an etched wafer imaging fragment comprises multiple images.

[0207] 39. A non-transient computer-readable medium pursuant to any of Clauses 37 and 38, wherein the imaging portion of a developed wafer comprises a plurality of images.

[0208] 40. A non-transient computer-readable medium according to any one of clauses 37-39, wherein training a machine learning model comprises:

[0209] An image fragment of an etched wafer is transformed into one or more feature maps by applying a convolutional neural network (CNN), wherein the one or more feature maps contain visual features of the image fragment of the etched wafer.

[0210] The image portion of the developed wafer is transformed into one or more feature maps by applying a CNN, wherein the one or more feature maps contain visual features of the image portion of the developed wafer.

[0211] Visual features associated with an image segment of an etched wafer are mapped to visual features associated with an image portion of a developed wafer.

[0212] Mapping is used to train a CNN encoder to construct an image portion of a developed wafer based on an image fragment of an etched wafer;

[0213] Extract training features from the construction.

[0214] The training features are input into the CNN decoder; and

[0215] The weights of the CNN decoder are adjusted by training the CNN decoder on the training features.

[0216] 41. A non-transient computer-readable medium according to Clause 40, wherein mapping visual features includes: mapping visual features associated with an imaged portion of a developed wafer to visual features associated with a first portion of an imaged segment of an etched wafer.

[0217] 42. A non-transient computer-readable medium pursuant to Clause 41, wherein an imaged portion of a developed wafer corresponds to a first portion of an imaged segment of an etched wafer.

[0218] 43. A non-transient computer-readable medium according to any one of clauses 37-42, wherein the application of a trained machine learning model includes: using a second portion of an image fragment of an etched wafer.

[0219] 44. A non-transient computer-readable medium pursuant to Clause 43, wherein a first portion of an etched wafer imaging segment is different from a second portion of an etched wafer imaging segment.

[0220] 45. A non-transient computer-readable medium according to any one of clauses 37-44, wherein a trained machine learning model is applied, including:

[0221] An image fragment of an etched wafer is transformed into one or more feature maps by applying a convolutional neural network (CNN), wherein the one or more feature maps contain visual features of the image fragment of the etched wafer.

[0222] A CNN encoder is used to extract visual features associated with an image fragment of an etched wafer, wherein the extraction includes using one or more convolutional pooling layers to learn hierarchical features based on the image fragment of the etched wafer.

[0223] Use the extracted visual features to identify the training features;

[0224] The training features are input into the CNN decoder; and

[0225] Generate predicted images of other parts of the developed wafer fragment.

[0226] 46. ​​A non-transient computer-readable medium pursuant to any of clauses 37-45, wherein the machine learning model comprises a plurality of machine learning models.

[0227] 47. A non-transitory computer-readable medium pursuant to Clause 46, wherein each machine learning model corresponds to a different segment of a wafer.

[0228] 48. A non-transient computer-readable medium of any of the clauses 37-45, wherein the application of a trained machine learning model includes a machine learning model and input etch parameters that characterize the etching of a wafer.

[0229] 49. A non-transient computer-readable medium according to any one of clauses 37-45, wherein the application of a trained machine learning model includes: using an imaging fragment of an etched wafer to generate a predicted image of other portions of a fragment of a developed wafer.

[0230] 50. A non-transient computer-readable medium pursuant to any of clauses 37-42, wherein the wafer is a first wafer.

[0231] 51. A non-transient computer-readable medium pursuant to Clause 50, comprising:

[0232] Obtain the second wafer after the second wafer has been etched;

[0233] Imaging a fragment of the etched second wafer; and

[0234] An image fragment of the etched second wafer is used to apply a trained machine learning model to generate a predicted image of the second wafer fragment after the photoresist applied to the second wafer has been developed.

[0235] 52. A non-transient computer-readable medium according to any one of clauses 50 and 51, wherein the application of a trained machine learning model includes:

[0236] An image fragment of the etched second wafer is transformed into one or more feature maps by applying a convolutional neural network (CNN), wherein the one or more feature maps contain visual features of the image fragment of the etched second wafer.

[0237] A CNN encoder is used to extract visual features associated with the image fragments of the etched second wafer, wherein the extraction includes using one or more convolutional pooling layers to learn hierarchical features based on the image fragments of the etched second wafer.

[0238] Use the extracted visual features to identify the training features;

[0239] The training features are input into the CNN decoder; and

[0240] Generate a predicted image of a fragment of the developed second wafer.

[0241] 53. A method for generating SEM images, comprising:

[0242] Obtain a first image of the features of the wafer after it has been etched;

[0243] The first image is analyzed using a trained machine learning model; and

[0244] A machine learning model is used to generate an image of features based on the application of photoresist to the wafer, corresponding to features in a first image.

[0245] 54. The method according to Clause 53, wherein:

[0246] Multiple images based on chip features are used to train a machine learning model; and

[0247] Multiple images include images of features after photoresist has been applied to the wafer and images of corresponding features after the wafer has been etched.

[0248] 55. The method according to Clause 54 also includes training the machine learning model by:

[0249] An image of the features of an etched wafer is transformed into one or more feature maps by applying a convolutional neural network (CNN), wherein one or more feature maps contain visual features of the image of the etched wafer;

[0250] By applying a CNN, the corresponding image of the features after the photoresist is applied to the wafer is transformed into one or more feature maps, wherein the one or more feature maps contain the visual features of the image of the features after the photoresist is applied to the wafer;

[0251] Visual features associated with an image of an etched wafer are mapped to visual features associated with an image of features after photoresist has been applied to the wafer.

[0252] The CNN encoder is trained using a mapping to construct an image of the features after photoresist has been applied to the wafer, based on an image of the etched wafer.

[0253] Extract training features from the construction.

[0254] The training features are input into the CNN decoder; and

[0255] The weights of the CNN decoder are adjusted by training the CNN decoder on the training features.

[0256] 56. The method according to any one of clauses 54 and 55, wherein the first image comprises a portion of the wafer that is different from the portion of the wafer used to train the machine learning model.

[0257] 57. The method according to any one of clauses 53-56, wherein the machine learning model comprises:

[0258] The first image is transformed into one or more feature maps by applying a convolutional neural network (CNN), wherein the one or more feature maps contain the visual features of the first image;

[0259] A CNN encoder is used to extract visual features associated with a first image, wherein the extraction includes learning hierarchical features based on the first image using one or more convolutional pooling layers;

[0260] Use the extracted visual features to identify the training features; and

[0261] The training features are input into the CNN decoder.

[0262] 58. The method according to any one of clauses 53-57, wherein the first image comprises a plurality of images.

[0263] 59. The method according to any one of clauses 53-58, wherein the machine learning model includes multiple machine learning models.

[0264] 60. The method according to Clause 59, wherein each machine learning model corresponds to a different part of the chip.

[0265] 61. The method according to any one of clauses 53-58, wherein the machine learning model is performed by a machine learning model, the method further comprising input etch parameters characterizing the etching of the wafer.

[0266] 62. The method according to any one of clauses 53-55, wherein the wafer is a first wafer.

[0267] 63. The method pursuant to Clause 62 also includes:

[0268] Obtain a second image of the features of the second wafer after the second wafer has been etched;

[0269] Analyzing the second image using a trained machine learning model; and

[0270] A machine learning model is used to generate an image of features based on the application of photoresist to the second wafer, corresponding to features in the second image.

[0271] 64. The method according to Clause 63, wherein the machine learning model for the second image comprises:

[0272] The second image is transformed into one or more feature maps by applying a convolutional neural network (CNN), wherein the one or more feature maps contain the visual features of the second image;

[0273] A CNN encoder is used to extract visual features associated with a second image, wherein the extraction includes learning hierarchical features based on the second image using one or more convolutional pooling layers;

[0274] Use the extracted visual features to identify the training features; and

[0275] The training features are input into the CNN decoder.

[0276] 65. A method for characterizing damage to a wafer caused by damage to a photoresist layer, wherein the damage to the photoresist layer of the wafer is caused by SEM imaging after development of the photoresist layer, the method comprising:

[0277] Obtain the first set of images of the wafer features after the photoresist layer has been developed;

[0278] Obtain a second set of images of the wafer's features after it has been etched;

[0279] The first set of images and the second set of images were used to train the machine learning model.

[0280] A third set of images is generated based on approximate features of the photoresist layer applied to the wafer, wherein the approximate features exclude damage caused by SEM imaging after development of the photoresist layer; and

[0281] Statistical analysis was performed based on the first set of images, the second set of images, and the third set of images to generate statistical characteristics of SEM damage to the wafer.

[0282] 66. The method according to Clause 65, wherein the first set of images includes a portion of the wafer, and a subset of the second set of images includes the same portion of the wafer.

[0283] 67. The method according to clause 66, wherein training the machine learning model includes:

[0284] A subset of the second set of images is transformed into one or more feature maps by applying a convolutional neural network (CNN), wherein the one or more feature maps contain visual features of the subset of the second set of images;

[0285] The first set of images is transformed into one or more feature maps by applying a CNN, wherein one or more feature maps contain the visual features of the first set of images;

[0286] Map the visual features associated with a subset of the second set of images to the visual features associated with the first set of images;

[0287] The CNN encoder is trained using mappings to construct the first set of images from a subset of the second set of images;

[0288] Extract training features from the construction.

[0289] The training features are input into the CNN decoder; and

[0290] The weights of the CNN decoder are adjusted by training the CNN decoder on the training features.

[0291] 68. The method according to any one of clauses 66 and 67, wherein the third set of images includes portions of the wafer that differ from portions of the wafer in subsets of the first set of images and the second set of images.

[0292] 69. The method according to any one of clauses 66-68, wherein generating the third set of images includes:

[0293] A subset of a second set of images, different from the subset used to train a machine learning model, is transformed into one or more feature maps by applying a convolutional neural network (CNN), wherein the one or more feature maps contain visual features of the subset of the second set of images.

[0294] A CNN encoder is used to extract visual features associated with a subset of the second set of images, wherein the extraction includes learning hierarchical features based on a subset of the second set of images using one or more convolutional pooling layers;

[0295] Use the extracted visual features to identify the training features;

[0296] The training features are input into the CNN decoder; and

[0297] The third set of images is generated based on a subset of the second set of images.

[0298] 70. The method according to any one of clauses 65-69, wherein the machine learning model includes multiple machine learning models.

[0299] 71. The method according to Clause 70, wherein each machine learning model corresponds to a different part of the chip.

[0300] 72. The method according to any one of clauses 65-69, wherein generating the third set of images includes a machine learning model and input etch parameters characterizing the etching of the wafer.

[0301] 73. The method according to any one of clauses 9-13, 15-20 or 35-36, wherein the machine learning model uses a linear activation function in the output layer to generate a predicted image.

[0302] 74. The method according to Clause 73, wherein the generated predicted image provides grayscale values.

[0303] 75. A system according to any one of clauses 29-33, wherein the machine learning model uses a linear activation function in the output layer to generate a predicted image.

[0304] 76. A system according to Clause 75, wherein the generated predicted image provides grayscale values.

[0305] 77. A non-transient computer-readable medium according to any one of clauses 37-52, wherein a machine learning model uses a linear activation function in the output layer to generate a predicted image.

[0306] 78. A non-transient computer-readable medium pursuant to Clause 77, wherein the generated predicted image provides grayscale values.

[0307] 79. The method according to any one of clauses 53-64, wherein the machine learning model uses a linear activation function in the output layer to generate a predicted image.

[0308] 80. The method according to Clause 79, wherein the generated SEM image provides grayscale values.

[0309] 81. The method according to any one of clauses 65-72, wherein the machine learning model uses a linear activation function in the output layer to generate a third set of images.

[0310] 82. The method according to Clause 81, wherein the generated third set of images provides grayscale values.

[0311] 83. A method for inspecting a wafer, comprising:

[0312] Obtain the wafer after the photoresist applied to it has been developed;

[0313] To image segments of a developed wafer;

[0314] Obtain the wafer after it has been etched;

[0315] Imaging a portion of an etched wafer fragment; and

[0316] Machine learning models are trained using imaging fragments from developed wafers and imaging portions from etched wafers.

[0317] 84. The method according to Clause 83, wherein the imaging segment of the developed wafer comprises multiple images.

[0318] 85. The method according to any one of clauses 83-83, wherein the imaging portion of the etched wafer comprises multiple images.

[0319] 86. The method according to any one of clauses 83-85, wherein training the machine learning model comprises:

[0320] A convolutional neural network (CNN) is applied to transform an image fragment of a developed wafer into one or more feature maps, wherein the one or more feature maps contain visual features of the image fragment of the developed wafer.

[0321] The image portion of the etched wafer is transformed into one or more feature maps by applying a CNN, wherein the one or more feature maps contain visual features of the image portion of the etched wafer.

[0322] Visual features associated with the imaging segment of the developed wafer are mapped to visual features associated with the imaging portion of the etched wafer.

[0323] Mapping is used to train a CNN encoder to construct an image portion of an etched wafer based on an image fragment of a developed wafer;

[0324] Extract training features from the construction.

[0325] The training features are input into the CNN decoder; and

[0326] The weights of the CNN decoder are adjusted by training the CNN decoder on the training features.

[0327] 87. The method according to Clause 86, wherein mapping visual features includes: mapping visual features associated with an imaged portion of an etched wafer to visual features associated with a first portion of an imaged segment of a developed wafer.

[0328] 88. The method according to Clause 87, wherein the imaged portion of the etched wafer corresponds to a first portion of the imaged segment of the developed wafer.

[0329] 89. The method according to any one of clauses 83-88, wherein the application of a trained machine learning model includes: using a second portion of an imaging segment of a developed wafer.

[0330] 90. The method according to Clause 89, wherein a first portion of the image segment of the developed wafer is different from a second portion of the image segment of the developed wafer.

[0331] 91. The method according to any one of clauses 83-90, wherein the application of a trained machine learning model includes:

[0332] A convolutional neural network (CNN) is applied to transform an image fragment of a developed wafer into one or more feature maps, wherein the one or more feature maps contain visual features of the image fragment of the developed wafer.

[0333] A CNN encoder is used to extract visual features associated with imaging segments of the developed wafer, wherein the extraction includes learning hierarchical features from the imaging segments of the developed wafer using one or more convolutional pooling layers.

[0334] Use the extracted visual features to identify the training features;

[0335] The training features are input into the CNN decoder; and

[0336] Generate a predicted post-etched image of the remaining portions of the developed wafer fragment.

[0337] 92. The method according to any one of clauses 83-91, wherein the machine learning model includes multiple machine learning models.

[0338] 93. The method according to Clause 92, wherein each machine learning model corresponds to a different segment of the wafer.

[0339] 94. The method according to any one of clauses 83-91, wherein the application of the trained machine learning model includes a machine learning model and input etch parameters characterizing the etching of the wafer.

[0340] 95. The method according to any one of clauses 83-91, wherein the application of a trained machine learning model includes: using an image fragment of the developed wafer to generate a predicted post-etched image of other portions of the developed wafer fragment.

[0341] 96. The method according to any one of clauses 83-88, wherein the wafer is a first wafer.

[0342] 97. The method pursuant to Clause 95 includes:

[0343] Obtain the second wafer after the photoresist applied to the second wafer has been developed;

[0344] Imaging a fragment of the developed second wafer; and

[0345] An image fragment of the developed second wafer is used to apply a trained machine learning model to generate a predicted post-etched image of the fragment of the second wafer after the photoresist applied to the second wafer has been developed.

[0346] 98. The method according to any one of clauses 95-96, wherein the application of a trained machine learning model includes:

[0347] The imaging fragments of the developed second wafer are transformed into one or more feature maps by applying a convolutional neural network (CNN), wherein the one or more feature maps contain visual features of the imaging fragments of the developed second wafer.

[0348] A CNN encoder is used to extract visual features associated with an image fragment of the developed second wafer, wherein the extraction includes learning hierarchical features based on the image fragment of the developed second wafer using one or more convolutional pooling layers.

[0349] Use the extracted visual features to identify the training features;

[0350] The training features are input into the CNN decoder; and

[0351] Generate a predicted post-etched image of a fragment of the developed second wafer.

[0352] 99. The method according to any one of clauses 91-95 or 97-98, wherein the machine learning model uses a sigmoid activation function in the output layer to generate a predicted etched image.

[0353] 100. The method according to Clause 99, wherein the predicted etched image is binarized.

[0354] 101. A method for training a machine learning model, the method comprising:

[0355] The first set of images of a segment of the first wafer after the photoresist applied to the first wafer has been developed are input into the machine learning model;

[0356] A second set of images of a portion of a fragment of the first wafer after the first wafer has been etched is input into the machine learning model; and

[0357] Based on the first set of images and the second set of images, the weights of the machine learning model are adjusted, wherein the trained machine learning model is configured to generate a predicted post-etched image of the second wafer.

[0358] 102. The method pursuant to Clause 101 further includes:

[0359] A first set of images of a fragment of a first developed wafer is transformed into one or more feature maps by applying a convolutional neural network (CNN), wherein the one or more feature maps contain visual features of the first set of images of a fragment of a first developed wafer.

[0360] A second set of images of a portion of the etched first wafer is transformed into one or more feature maps by applying a CNN, wherein the one or more feature maps contain visual features of the second set of images of the portion of the etched first wafer.

[0361] Visual features associated with a first set of images of a segment of the developed first wafer are mapped to visual features associated with a second set of images of a portion of the etched first wafer.

[0362] A mapping is used to train a CNN encoder to construct a second set of images of etched portions of the first wafer based on a first set of images of fragments of the first wafer that have been developed.

[0363] Extract training features from the construction.

[0364] The training features are input into the CNN decoder; and

[0365] The weights of the CNN decoder are adjusted by training the CNN decoder on the training features.

[0366] 103. The method according to Clause 102, wherein mapping visual features includes: mapping visual features associated with the set of images of a portion of the etched first wafer to visual features associated with a set of images of a first portion of a fragment of the developed first wafer.

[0367] 104. The method according to Clause 103, wherein a second set of images of a portion of an etched wafer corresponds to a first set of images of a first portion of a fragment of a developed first wafer.

[0368] 105. The method according to any one of clauses 101-104, wherein the machine learning model uses a sigmoid activation function in the output layer to generate a predicted etched image.

[0369] 106. The method according to Clause 105, wherein the predicted etched image is binarized.

[0370] 107. A charged particle multi-beam system for inspection, the system comprising:

[0371] A controller, comprising circuitry for the following:

[0372] Imaging segments of a wafer after the photoresist applied to the wafer has been developed;

[0373] Imaging a portion of a wafer after it has been etched; and

[0374] Machine learning models are trained using imaging fragments from developed wafers and imaging portions from etched wafers.

[0375] 108. A system according to Clause 107, wherein the imaging fragment of the developed wafer comprises multiple images.

[0376] 109. A system according to any one of clauses 107-108, wherein the imaging portion of the etched wafer comprises multiple images.

[0377] 110. A system according to any one of clauses 107-109, wherein training a machine learning model comprises:

[0378] A convolutional neural network (CNN) is applied to transform an image fragment of a developed wafer into one or more feature maps, wherein the one or more feature maps contain visual features of the image fragment of the developed wafer.

[0379] The image portion of the etched wafer is transformed into one or more feature maps by applying a CNN, wherein the one or more feature maps contain visual features of the image portion of the etched wafer.

[0380] Visual features associated with the imaging segment of the developed wafer are mapped to visual features associated with the imaging portion of the etched wafer.

[0381] Mapping is used to train a CNN encoder to construct an image portion of an etched wafer based on an image fragment of a developed wafer;

[0382] Extract training features from the construction.

[0383] The training features are input into the CNN decoder; and

[0384] The weights of the CNN decoder are adjusted by training the CNN decoder on the training features.

[0385] 111. The system according to Clause 110, wherein mapping visual features includes: mapping visual features associated with an imaged portion of an etched wafer to visual features associated with a first portion of an imaged segment of a developed wafer.

[0386] 112. The system according to Clause 111, wherein the imaged portion of the etched wafer corresponds to a first portion of the imaged segment of the developed wafer.

[0387] 113. A system according to any one of clauses 107-112, wherein the application of a trained machine learning model includes: using a second portion of an imaging fragment of a developed wafer.

[0388] 114. A system according to Clause 113, wherein a first portion of an image segment of a developed wafer is different from a second portion of an image segment of a developed wafer.

[0389] 115. A system according to any one of clauses 107-114, wherein the application of a trained machine learning model includes:

[0390] A convolutional neural network (CNN) is applied to transform an image fragment of a developed wafer into one or more feature maps, wherein the one or more feature maps contain visual features of the image fragment of the developed wafer.

[0391] A CNN encoder is used to extract visual features associated with imaging segments of the developed wafer, wherein the extraction includes learning hierarchical features based on imaging segments of the developed wafer using one or more convolutional pooling layers.

[0392] Use the extracted visual features to identify the training features;

[0393] The training features are input into the CNN decoder; and

[0394] Generate a predicted post-etched image of the remaining portions of the developed wafer fragment.

[0395] 116. A system pursuant to any one of clauses 107-115, wherein the machine learning model comprises multiple machine learning models.

[0396] 117. A system according to Clause 116, wherein each machine learning model corresponds to a different segment of a wafer.

[0397] 118. A system according to any one of clauses 107-115, wherein the application of a trained machine learning model includes a machine learning model and input etch parameters that characterize the etching of the wafer.

[0398] 119. A system according to any one of clauses 107-115, wherein the application of a trained machine learning model includes: using an image fragment of a developed wafer to generate a predicted post-etched image of other portions of the developed wafer fragment.

[0399] 120. A system pursuant to any one of clauses 107-112, wherein the chip is a first chip.

[0400] 121. The system pursuant to Clause 120 includes:

[0401] Obtain the second wafer after the photoresist applied to the second wafer has been developed;

[0402] Imaging a fragment of the developed second wafer; and

[0403] An image fragment of the developed second wafer is used to apply a trained machine learning model to generate a predicted post-etched image of the fragment of the second wafer after the photoresist applied to the second wafer has been developed.

[0404] 122. A system according to any one of clauses 120-121, wherein the application of a trained machine learning model includes:

[0405] The imaging fragments of the developed second wafer are transformed into one or more feature maps by applying a convolutional neural network (CNN), wherein the one or more feature maps contain visual features of the imaging fragments of the developed second wafer.

[0406] A CNN encoder is used to extract visual features associated with an image fragment of the developed second wafer, wherein the extraction includes learning hierarchical features based on the image fragment of the developed second wafer using one or more convolutional pooling layers.

[0407] Use the extracted visual features to identify the training features;

[0408] The training features are input into the CNN decoder; and

[0409] Generate a predicted post-etched image of a fragment of the developed second wafer.

[0410] 123. A system according to any one of clauses 115-119 or 121-122, wherein the machine learning model uses a sigmoid activation function in the output layer to generate a predicted etched image.

[0411] 124. The system according to Clause 123, wherein the predicted etched image is binarized.

[0412] 125. A non-transient computer-readable medium storing a set of instructions executable by at least one processor of a computing device to cause the computing device to perform a method for inspecting a wafer, the method comprising:

[0413] Image a segment of the first wafer after the photoresist applied to the first wafer has been developed;

[0414] Imaging a portion of a segment of the first wafer after the first wafer has been etched;

[0415] A machine learning model is trained using an image fragment from a developed first wafer and an image portion from an etched first wafer; and

[0416] An image fragment of the developed second wafer is used to apply a trained machine learning model to generate a predicted post-etched image of the developed second wafer.

[0417] 126. A non-transient computer-readable medium pursuant to Clause 125, wherein an imaging segment of a developed first wafer comprises a plurality of images.

[0418] 127. A non-transient computer-readable medium pursuant to any of Clauses 125-126, wherein the imaging portion of an etched first wafer comprises a plurality of images.

[0419] 128. A non-transient computer-readable medium according to any one of clauses 125-127, wherein training a machine learning model comprises:

[0420] An image fragment of the developed first wafer is transformed into one or more feature maps by applying a convolutional neural network (CNN), wherein the one or more feature maps contain visual features of the image fragment of the developed first wafer.

[0421] The image portion of the etched first wafer is transformed into one or more feature maps by applying a CNN, wherein the one or more feature maps contain visual features of the image portion of the etched first wafer.

[0422] Visual features associated with the imaging segment of the developed first wafer are mapped to visual features associated with the imaging portion of the etched first wafer.

[0423] A mapping is used to train a CNN encoder to construct an image portion of an etched first wafer based on an image fragment of a developed first wafer;

[0424] Extract training features from the construction.

[0425] The training features are input into the CNN decoder; and

[0426] The weights of the CNN decoder are adjusted by training the CNN decoder on the training features.

[0427] 129. A non-transient computer-readable medium pursuant to Clause 128, wherein mapping visual features includes: mapping visual features associated with an imaging portion of an etched first wafer to visual features associated with a first portion of an imaging fragment of a developed first wafer.

[0428] 130. A non-transient computer-readable medium pursuant to Clause 129, wherein an etched imaging portion of a first wafer corresponds to a first portion of an imaging fragment of a developed first wafer.

[0429] 131. A non-transient computer-readable medium pursuant to any of clauses 125-130, wherein the second wafer is the first wafer, and the application of a trained machine learning model comprises: using a second portion of an image fragment of the developed first wafer.

[0430] 132. A non-transient computer-readable medium pursuant to Clause 131, wherein a first portion of an image segment of a developed first wafer is different from a second portion of an image segment of a developed first wafer.

[0431] 133. A non-transient computer-readable medium pursuant to any of clauses 125-132, wherein a trained machine learning model is applied, comprising:

[0432] An image fragment of the developed first wafer is transformed into one or more feature maps by applying a convolutional neural network (CNN), wherein the one or more feature maps contain visual features of the image fragment of the developed first wafer.

[0433] A CNN encoder is used to extract visual features associated with an image fragment of the developed first wafer, wherein the extraction includes learning hierarchical features based on the image fragment of the developed first wafer using one or more convolutional pooling layers.

[0434] Use the extracted visual features to identify the training features;

[0435] The training features are input into the CNN decoder; and

[0436] Generate a predicted post-etched image of the remaining portions of the fragment of the first wafer that has been developed.

[0437] 134. A non-transient computer-readable medium pursuant to any of clauses 125-133, wherein the machine learning model comprises a plurality of machine learning models.

[0438] 135. A non-transitory computer-readable medium pursuant to Clause 134, wherein each machine learning model corresponds to a different segment of a first chip.

[0439] 136. A non-transient computer-readable medium according to any one of clauses 125-133, wherein the application of a trained machine learning model includes a machine learning model and input etch parameters characterizing the etching of a first wafer.

[0440] 137. A non-transient computer-readable medium pursuant to any of clauses 125-133, wherein the application of a trained machine learning model comprises: using an imaging fragment of a developed first wafer to generate a predicted post-etched image of other portions of the fragment of the developed first wafer.

[0441] 138. A non-transient computer-readable medium pursuant to any of clauses 125-130, wherein the second wafer is different from the first wafer.

[0442] 139. Non-transient computer-readable media pursuant to Clause 138, including:

[0443] Obtain the second wafer after the photoresist applied to the second wafer has been developed;

[0444] Imaging a fragment of the developed second wafer; and

[0445] An image fragment of the developed second wafer is used to apply a trained machine learning model to generate a predicted post-etched image of the fragment of the second wafer after the photoresist applied to the second wafer has been developed.

[0446] 140. A non-transient computer-readable medium pursuant to any of clauses 138-139, wherein a trained machine learning model is applied, comprising:

[0447] The imaging fragments of the developed second wafer are transformed into one or more feature maps by applying a convolutional neural network (CNN), wherein the one or more feature maps contain visual features of the imaging fragments of the developed second wafer.

[0448] A CNN encoder is used to extract visual features associated with an image fragment of the developed second wafer, wherein the extraction includes learning hierarchical features based on the image fragment of the developed second wafer using one or more convolutional pooling layers.

[0449] Use the extracted visual features to identify the training features;

[0450] The training features are input into the CNN decoder; and

[0451] Generate a predicted post-etched image of a fragment of the developed second wafer.

[0452] 141. A non-transient computer-readable medium according to any one of clauses 125-140, wherein a machine learning model uses a sigmoid activation function in the output layer to generate a predicted etched image.

[0453] 142. A non-transient computer-readable medium pursuant to Clause 141, wherein the predicted etched image is binarized.

[0454] 143. A method for generating SEM images, comprising:

[0455] Obtain a first image of the features of the wafer after photoresist has been applied to it;

[0456] The first image is analyzed using a trained machine learning model; and

[0457] A machine learning model is used to generate a predicted image of features based on an etching application to the wafer that corresponds to features in a first image.

[0458] 144. The method according to Clause 143, wherein:

[0459] The machine learning model is trained based on multiple images of the chip's features; and

[0460] Multiple images include images of features after photoresist has been applied to the wafer and images of corresponding features after the wafer has been etched.

[0461] 145. The method according to Clause 144 also includes training the machine learning model by:

[0462] An image of the features of an etched wafer is transformed into one or more feature maps by applying a convolutional neural network (CNN), wherein one or more feature maps contain visual features of the image of the etched wafer;

[0463] By applying a CNN, the corresponding image of the features after the photoresist is applied to the wafer is transformed into one or more feature maps, wherein the one or more feature maps contain the visual features of the image of the features after the photoresist is applied to the wafer;

[0464] Visual features associated with an image of features after photoresist has been applied to the wafer are mapped to visual features associated with an image of the etched wafer.

[0465] The CNN encoder is trained using a mapping to construct an image of the features after the wafer has been etched, based on an image of the features after the photoresist has been applied to the wafer.

[0466] Extract training features from the construction.

[0467] The training features are input into the CNN decoder; and

[0468] The weights of the CNN decoder are adjusted by training the CNN decoder on the training features.

[0469] 146. The method according to any one of clauses 144-145, wherein the first image comprises a portion of the wafer that is different from the portion of the wafer used to train the machine learning model.

[0470] 147. The method according to any one of clauses 143-146, wherein the machine learning model comprises:

[0471] The first image is transformed into one or more feature maps by applying a convolutional neural network (CNN), wherein the one or more feature maps contain the visual features of the first image;

[0472] A CNN encoder is used to extract visual features associated with a first image, wherein the extraction includes learning hierarchical features based on the first image using one or more convolutional pooling layers;

[0473] Use the extracted visual features to identify the training features; and

[0474] The training features are input into the CNN decoder.

[0475] 148. The method according to any one of clauses 143-147, wherein the first image comprises a plurality of images.

[0476] 149. The method according to any one of clauses 143-148, wherein the machine learning model includes multiple machine learning models.

[0477] 150. The method according to Clause 149, wherein each machine learning model corresponds to a different part of the chip.

[0478] 151. The method according to any one of clauses 143-148 further includes inputting etching parameters that characterize the etching of the wafer.

[0479] 152. The method according to any one of clauses 143-145, wherein the wafer is a first wafer.

[0480] 153. The method pursuant to Clause 152 also includes:

[0481] Obtain a second image of the features of a second wafer after the photoresist applied to the wafer has been developed;

[0482] Analyzing the second image using a trained machine learning model; and

[0483] Machine learning analysis is used to generate a predicted image of features based on an etching application to a second wafer that corresponds to features in a second image.

[0484] 154. The method according to Clause 153, wherein the machine learning model for the second image comprises:

[0485] The second image is transformed into one or more feature maps by applying a convolutional neural network (CNN), wherein the one or more feature maps contain the visual features of the second image;

[0486] A CNN encoder is used to extract visual features associated with a second image, wherein the extraction includes learning hierarchical features based on the second image using one or more convolutional pooling layers;

[0487] Use the extracted visual features to identify the training features; and

[0488] The training features are input into the CNN decoder.

[0489] 155. The method according to any one of clauses 143-154, wherein the machine learning model uses a sigmoid activation function in the output layer to compute the generated predicted image.

[0490] 156. The method according to Clause 155, wherein the generated predicted image is binarized.

[0491] 147. The method according to any one of clauses 101-106, wherein a portion of a fragment is the entire fragment.

[0492] A non-transient computer-readable medium may be provided that stores a controller for controlling wafer inspection according to embodiments of the present disclosure (e.g., Figure 1 The controller (109) administers instructions to the processor. For example, based on steps in the wafer inspection process (e.g., after development, after etching), the controller can train a machine learning model and execute the trained machine learning model to image the wafer. Common forms of non-transient media include, for example, floppy disks, flexible disks, hard disks, solid-state drives, magnetic tape or any other magnetic data storage media, optical disc read-only memory (CD-ROM), any other optical data storage media, any physical media with a perforated pattern, random access memory (RAM), programmable read-only memory (PROM) and 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 networked versions.

[0493] As used herein, unless otherwise specifically stated, the term "or" covers all possible combinations, unless impractical. For example, if a statement component may include A or B, then unless otherwise specifically stated or impractical, the component may include A or B, or A and B. As a second example, if a statement component may include A, B, or C, then unless otherwise specifically stated or impractical, 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.

[0494] It should be understood that the embodiments of this disclosure are not limited to the exact constructions described above and illustrated in the accompanying drawings, and various modifications and changes can be made without departing from the scope of this disclosure. This disclosure has been described in conjunction with various embodiments, and other embodiments of the invention will be apparent to those skilled in the art in light of the specification and practice of the invention disclosed herein. The specification and embodiments are to be considered exemplary only, and the true scope and spirit of the invention are indicated by the appended claims.

Claims

1. A method for generating SEM images, comprising: Obtain a first image of the features of the wafer after it has been etched; The first image is analyzed using a trained machine learning model; as well as The machine learning model is used to generate an image based on features of the photoresist application on the wafer corresponding to the features of the first image, wherein: The machine learning model is trained based on multiple images of features from the chip; and The multiple images include images of features after photoresist has been applied to the wafer and images of corresponding features after the wafer has been etched.

2. The method of claim 1, further comprising training the machine learning model by: The image of the features of the etched wafer is transformed into one or more feature maps by applying a convolutional neural network (CNN), wherein the one or more feature maps contain visual features of the image of the etched wafer. The CNN is applied to transform the corresponding image of the features after the photoresist is applied to the wafer into one or more feature maps, wherein the one or more feature maps contain the visual features of the image of the features after the photoresist is applied to the wafer; The visual features associated with the image of the etched wafer are mapped to the visual features associated with the image of the features after the photoresist is applied to the wafer; The mapping is used to train a CNN encoder to construct an image of the features after the photoresist is applied to the wafer, based on the image of the etched wafer. Extract training features from the constructed structure. The trained features are then input into the CNN decoder. as well as The weights of the CNN decoder are adjusted by training the CNN decoder on the training features.

3. The method of claim 1, wherein the first image comprises a portion of the wafer that is different from the portion of the wafer used to train the machine learning model.

4. The method according to claim 1, wherein the machine learning model comprises: The first image is transformed into one or more feature maps by applying a convolutional neural network (CNN), wherein the one or more feature maps contain visual features of the first image; A CNN encoder is used to extract the visual features associated with the first image, wherein the extraction includes using one or more convolutional pooling layers to learn hierarchical features based on the first image; The extracted visual features are used to identify the training features; and The trained features are then input into the CNN decoder.

5. The method according to claim 1, wherein the first image comprises a plurality of images.

6. The method according to claim 1, wherein the machine learning model comprises a plurality of machine learning models.

7. The method of claim 6, wherein each machine learning model corresponds to a different part of the wafer.

8. The method of claim 1, wherein the machine learning model is executed by a machine learning model, the method further comprising: Input etching parameters, which characterize the etching of the wafer.

9. The method of claim 1, wherein the wafer is a first wafer.

10. The method of claim 9, further comprising: Obtain a second image of the features of the second wafer after it has been etched; The trained machine learning model is used to analyze the second image; as well as The machine learning model is used to generate an image of features based on the application of photoresist to the second wafer corresponding to the features of the second image.

11. The method of claim 10, wherein the machine learning model for the second image comprises: The second image is transformed into one or more feature maps by applying a convolutional neural network (CNN), wherein the one or more feature maps contain visual features of the second image; A CNN encoder is used to extract the visual features associated with the second image, wherein the extraction includes using one or more convolutional pooling layers to learn hierarchical features based on the second image; The extracted visual features are used to identify the training features; and The trained features are then input into the CNN decoder.

12. A charged particle multi-beam system for inspecting wafers, the system comprising: A controller, the controller comprising a circuit system for: Image a portion of a wafer after the photoresist applied to the wafer has been developed; Imaging the segment of the wafer after it has been etched; and A machine learning model is trained using the developed imaging portion of the wafer and the etched imaging fragment of the wafer, wherein training the machine learning model includes: The image fragment of the etched wafer is transformed into one or more feature maps by applying a convolutional neural network (CNN), wherein the one or more feature maps contain visual features of the image fragment of the etched wafer. The CNN is applied to transform the imaged portion of the developed wafer into one or more feature maps, wherein the one or more feature maps contain visual features of the imaged portion of the developed wafer. The visual features associated with the imaging segment of the etched wafer are mapped to the visual features associated with the imaging portion of the developed wafer; The mapping is used to train a CNN encoder to construct the image portion of the developed wafer based on the image fragment of the etched wafer; Extract training features from the constructed structure. The trained features are input into the CNN decoder; and The weights of the CNN decoder are adjusted by training the CNN decoder on the training features.

13. The system of claim 12, wherein the etched imaging segment of the wafer comprises a plurality of images, and wherein the developed imaging portion of the wafer comprises a plurality of images.

Images