Computer-implemented method for defect detection in imaging datasets of a section of an object comprising integrated circuit patterns, and corresponding computer-readable medium, computer program, and system.

Abstract

The invention relates to a computer-implemented method (26) for defect detection, comprising: obtaining a first imaging data set (22) of a section of an object (72) comprising integrated circuit patterns; obtaining at least a second imaging data set (28) and a third imaging data set (30) comprising substantially the same integrated circuit patterns as the section of the object (72); and jointly processing at least the first imaging data set (22), the second imaging data set (28), and the third imaging data set (30) to detect defects (24). The invention also relates to a corresponding computer program, computer-readable medium, and defect detection system.

Description

Field of the invention

[0001] The invention relates to systems and methods for quality assurance of objects comprising integrated circuit patterns, in particular to a computer-implemented method, a computer-readable medium, a computer program product, and a corresponding system for defect detection in an imaging data set of such an object. Defects can be detected by comparing the first imaging data set with a second imaging data set and a third imaging data set. The method, the computer-readable medium, the computer program product, and the system can be used for quantitative metrology, process monitoring, defect detection, and defect assessment in objects comprising integrated circuit patterns, e.g., in photolithography masks, reticles, or wafers. Background of the invention

[0002] A wafer made of a thin silicon disc serves as the substrate for microelectronic devices containing semiconductor structures built into and on the wafer. The semiconductor structures are constructed layer by layer through repeated processing steps that include repeated chemical, mechanical, thermal, and optical processes. The dimensions, shapes, and placement of the semiconductor structures and patterns are subject to multiple influences. One of the most important steps is the photolithography process.

[0003] Photolithography is a process for creating patterns on the substrate. The patterns to be printed on the substrate surface are created using computer-aided design (CAD). From the design, a photolithography mask is created for each layer. This mask contains a magnified image of the computer-generated pattern to be etched into the substrate. The photolithography mask can be further adjusted, for example, using optical proximity correction techniques. During the printing process, an illuminated image projected by the photolithography mask is focused onto a photoresist thin film formed on the substrate. A semiconductor chip that powers mobile phones or tablets, for example, comprises approximately 80 to 120 patterned layers.

[0004] Due to the growing integration density in the semiconductor industry, photolithography masks must image ever smaller structures on wafers. The aspect ratio and number of layers of integrated circuits are constantly increasing, and the structures are growing into the third (vertical) dimension. The current height of memory stacks is more than a dozen micrometers. In contrast, the feature size is becoming smaller. The minimum feature size or critical dimension is below 10 nm, for example, 7 nm or 5 nm, and will approach feature sizes below 3 nm in the near future. As the complexity and dimensions of semiconductor structures grow into the third dimension, the lateral dimensions of integrated semiconductor structures are becoming smaller. To create the small feature dimensions imaged on the wafer, photolithography masks or templates for nanoimprint photolithography with increasingly smaller structures or pattern elements are required.The manufacturing process of photolithography masks and templates for nanoimprint photolithography is therefore becoming increasingly complex, time-consuming, and ultimately expensive. With the advent of EUV photolithography scanners, the type of masks changed from transmission-based to reflection-based patterning.

[0005] Due to the tiny feature sizes of the pattern elements of photolithography masks or templates, it is not possible to exclude defects during mask or template fabrication. The resulting defects can arise, for example, from the degradation of photolithography masks or particle contamination. Of the various defects that occur during the fabrication of a semiconductor structure, photolithography-related defects account for almost half of the number of defects. Therefore, the inspection, evaluation, and metrology of photolithography masks plays an important role in semiconductor process control to monitor systematic defects. Defects detected during quality assurance processes can be used for root cause analysis, for example, to modify or repair the photolithography mask. The defects can also serve as feedback to optimize the process parameters of the manufacturing process, e.g.exposure time, focus variation, etc.

[0006] Photolithography mask inspection must be performed at multiple points in time to improve the quality of the photolithography masks and maximize their service life. Once the photolithography mask is manufactured according to requirements, an initial quality assessment of the photolithography mask is performed at the mask house before it is sent to the wafer fab. The semiconductor device design and the quality of the photolithography mask fabrication are verified through various procedures before the photolithography mask is sent to a semiconductor manufacturing facility to begin integrated circuit production. The semiconductor device design is tested through software simulation to verify that all features are printed correctly after photolithography during manufacturing.The photolithography mask is inspected and measured for defects to ensure that the features are within specification. The data collected during this process becomes the golden baseline or reference for further inspections to be performed at the mask house or wafer fab. Any defects found on the photolithography mask are validated using an evaluation tool, followed by a decision to send the photolithography mask for repair or to decommission the mask and order a new one. At the wafer fab, the photolithography mask is scanned to find additional defects, called "adders," compared to the last scan performed at the mask house. Each of these adders is analyzed using an evaluation tool. In the case of a particle defect, the particle is removed.In the event of a pattern-based defect, the photolithography mask is either repaired, if possible, or replaced with a new one. The inspection process is repeated every few photolithography cycles.

[0007] Any defect in the photolithography mask can lead to undesirable behavior of the produced wafer, or a wafer can be significantly damaged. Therefore, every defect must be found and, if possible and necessary, repaired. Reliable and fast defect detection methods are therefore important for photolithography masks.

[0008] In addition to defect detection in photolithography masks, defect detection in wafers is also important for quality management. During wafer production, many defects other than photolithography mask defects can occur, for example, during etching or deposition. For example, bridge defects can indicate insufficient etching, line breaks can indicate excessive etching, persistent defects can indicate a defective mask, and missing features indicate non-ideal material deposition, etc. Therefore, a quality assurance process and a quality control process are important to ensure high quality standards of the manufactured wafers.

[0009] Aside from quality assurance and quality control, defect detection in wafers is also important during process window qualification (PWQ). This process is used to define windows for a number of process parameters, primarily related to different focus and exposure conditions, to prevent systematic defects. In each iteration, a test wafer is manufactured based on a number of selected process parameters, e.g., exposure time, focus variation, etc., with different dies of the wafer being subjected to different manufacturing conditions. By detecting and analyzing the defects in the different dies based on the quality assurance process, the best manufacturing process parameters can be selected, and a window or range can be established for each process parameter from which the respective process parameter can be selected.Additionally, a high-precision quality control process and equipment for metrology of semiconductor structures in wafers are required. The detected defects can thus be used to monitor wafer quality during production or to establish a process window. Reliable and fast defect detection methods are therefore important for objects containing integrated circuit patterns.

[0010] Machine learning methods can be used to analyze large amounts of data that require large amounts of measurements. Machine learning is a field of artificial intelligence. Machine learning methods generally create a parametric machine learning model based on training data consisting of a large number of samples. After training, the method is able to generalize the knowledge gained from the training data to new, previously unencountered samples, thereby making predictions for new data. There are many machine learning methods, such as linear regression, K-means, support vector machines, neural networks, or deep learning approaches.

[0011] Deep learning is a class of machine learning that uses artificial neural networks with numerous hidden layers between the input and output layers. Due to this complex internal structure, the networks are able to progressively extract higher-level features from the raw input data. Each layer learns to transform its input data into a slightly more abstract and composite representation, thereby deriving low- and high-level knowledge from the training data. The hidden layers can have different sizes and perform different tasks, such as convolutional or pooling layers.

[0012] Methods for the automatic detection of defects in objects comprising integrated circuit patterns include defect detection algorithms that are often based on a die-to-die or die-to-database principle.

[0013] The die-to-die principle compares an imaging dataset of sections of one object with a reference dataset of the same sections of another identical object or the same object. Detected deviations are treated as defects.

[0014] The die-to-database principle compares an imaging dataset of an object with a simulated reference dataset or a design, such as a CAD file, thereby detecting deviations from the ideal data.

[0015] For example, US 2019 / 0130551 A1 discloses a die-to-die method for defect detection. In a first step, a reference data set is generated from a number of scan images of an identical wafer, e.g., using a median filter. Imaging data sets are obtained from a target wafer, and defects are detected based on pixel value differences between an imaging data set and the reference data set.

[0016] Both the die-to-die and die-to-database approaches rely on comparing an imaging dataset with a reference dataset, such as an acquired imaging dataset or a simulated or design dataset. Depending on the quality of the datasets being compared, such as the noise level or the similarity of the appearance of the design datasets and imaging datasets, defects may be missed or false-positive defect detections may occur.

[0017] To improve the accuracy of defect detection, US 2006 / 204109 A1 discloses a method for detecting defects in a plurality of images that have substantially the same image content. In a first step, differences between pairs of images are calculated. By comparing the pairwise differences, defects are detected in a second step. However, calculating pairwise differences and comparing them is time-consuming, especially for large data sets such as those obtained from objects comprising integrated circuit patterns. In addition, the results of subsequent comparisons may be ambiguous if one comparison results in a defect but another does not.

[0018] It is therefore an object of the invention to detect defects in objects comprising integrated circuit patterns with a high recall and precision. Another object of the invention is to improve the accuracy of the detected defects. Another object of the invention is to achieve short runtimes and high throughput of the defect detection method. Another object of the invention is to provide methods for defect detection that do not require expert knowledge from the users.

[0019] The objects are achieved by the invention specified in the independent claims. Advantageous embodiments and further developments of the invention are specified in the dependent claims. Brief description of the invention

[0020] Embodiments of the invention relate to computer-implemented methods, computer-readable media, and systems that implement defect detection methods for objects comprising integrated circuit patterns.

[0021] An integrated circuit pattern, for example, may include semiconductor structures. An object including integrated circuit patterns may refer, for example, to a photolithography mask, a reticle, or a wafer. In a photolithography mask or reticle, the integrated circuit patterns may refer to mask structures used to create semiconductor patterns in a wafer during the photolithography process. In a wafer, the integrated circuit patterns may refer to semiconductor structures printed onto the wafer during the photolithography process.

[0022] The object comprising integrated circuit patterns may be a photolithography mask. The photolithography mask may have an aspect ratio between 1:1 and 1:4, preferably between 1:1 and 1:2, most preferably 1:1 or 1:2. The photolithography mask may have a nearly rectangular shape. The photolithography mask may preferably be 5 to 7 inches long and wide, most preferably 6 inches long and wide. Alternatively, the photolithography mask may be 5 to 7 inches long and 10 to 14 inches wide, preferably 6 inches long and 12 inches wide.

[0023] One embodiment of the invention includes a computer-implemented method for defect detection, comprising: obtaining a first imaging data set of a portion of an object comprising integrated circuit patterns; obtaining at least a second imaging data set and a third imaging data set comprising primarily the same integrated circuit patterns as the portion of the object; and jointly processing at least the first imaging data set, the second imaging data set, and the third imaging data set to detect defects. Additional imaging data sets, for example, a fourth imaging data set, or a fourth imaging data set and a fifth imaging data set, may be used to increase the precision and hit rate of the defect detection method and the accuracy of the detected defects.Additional imaging datasets may be used, for example, depending on the difficulty of the defect detection task.

[0024] By using at least three imaging datasets instead of just one imaging dataset and a reference dataset, the precision and accuracy of the defect detection method are improved due to the additional information available. For example, robustness to noise can be improved in this way. Furthermore, the ambiguity created by two or more separate comparisons of different input datasets is reduced by considering all input datasets together. Thus, the accuracy and accuracy of the defect detection method and the accuracy of the detected defects are improved. In addition, the use of three or more imaging datasets enables the detection of defects in all imaging datasets, for example, if an acquired imaging dataset presumably free of defects (the so-called golden reference) is used that nevertheless contains defects.For defect detection methods that perform pairwise comparisons between two imaging data sets, runtime is reduced because each imaging data set is considered only once by the defect detection method. This increases throughput.

[0025] The term "defect" refers to a localized deviation of an integrated circuit pattern from an a priori-defined norm of the integrated circuit pattern. For example, a defect in an integrated circuit pattern, e.g., a semiconductor structure, can lead to a malfunction of an associated semiconductor device. Depending on the detected defect, for example, the photolithography process can be improved, or photolithography masks or wafers can be repaired or discarded. The norm of the integrated circuit pattern can be defined by one or more corresponding reference objects or reference data sets, e.g., design data sets, simulated data sets, or acquired defect-free data sets.

[0026] A section of an object can refer to a part of the object, such as one or more areas of the object that may or may not be connected, or to the entire object.

[0027] According to the techniques described herein, various imaging modalities can be used to acquire one or more of the imaging datasets (the first imaging dataset, the second imaging dataset, the third imaging dataset, and potentially additional imaging datasets) for defect detection. An imaging dataset can comprise one or more single-channel images or multi-channel images, e.g., focus stacks. For example, the imaging dataset can include one or more 2D images. It is possible to employ a multi-beam scanning electron microscope (mSEM). mSEM employs multiple beams to simultaneously acquire images in multiple fields of view. For example, as many as 50 beams or even as many as 90 beams could be used. Each beam covers a separate portion of a surface of the object comprising integrated circuit patterns.This allows a large imaging dataset to be acquired within a short period of time. Typically, 4.5 gigapixels are acquired per second using contemporary machines. To illustrate, one square centimeter of a wafer can be imaged with a pixel size of 2 nm, resulting in 25 terapixels of data. Other examples of imaging datasets that include 2D images would relate to imaging modalities such as optical imaging, phase-contrast imaging, X-ray imaging, etc. It would also be possible for one or more of the imaging datasets to be volumetric 3D datasets that can be processed slice-by-slice or as a three-dimensional volume. A crossbeam imaging device including a focused ion beam (FIB) source, an atomic force microscope (AFM), or a scanning electron microscope (SEM) could be used here.Multimodal imaging datasets can be used, e.g., a combination of X-ray imaging and SEM. An imaging dataset can include one or more aerial images acquired by an aerial imaging system. An aerial image is the radiation intensity distribution at the substrate level. It can be used to simulate the radiation intensity distribution generated by the photolithography mask during the photolithography process. The aerial imaging measurement system can, for example, be equipped with a rigid array sensor, a line-scan sensor, or a time-delayed integration (TDI) sensor.

[0028] An imaging dataset may also be obtained by acquiring images of a reference object comprising integrated circuit patterns. The reference object comprising integrated circuit patterns may, for example, be another instance of the same type of object, or it may be a different type but comprise at least a portion of the same integrated circuit patterns as the object. The imaging dataset may also be obtained from one or more portions of the (same) object comprising integrated circuit patterns, e.g., from a different die of the object, for example, in the case of repeating structures, or the same die at a different time, e.g., previously acquired imaging datasets. Alternatively, the imaging dataset may be artificially generated.

[0029] One or more of the imaging datasets (the first imaging dataset, the second imaging dataset, the third imaging dataset, and potentially additional imaging datasets) may be artificially generated datasets, e.g., a simulated dataset or a design dataset. In one example, a simulated dataset of the object comprising integrated circuit patterns is obtained using rigorous physical simulations of the electromagnetic wave propagation process in the photolithography mask, such as FDTD (Finite Difference Time Domain) and RWCA (Rigorous Coupled Wave Analysis). A simulated dataset may also be obtained using approximations of the wave propagation process, such as TEA (Thin Element Approximation) or Kirchhoff's approach assuming a thin photolithography mask.A simulated dataset can also be obtained using machine learning models trained to simulate the wave propagation process, e.g., using pairs of photolithography mask descriptions and corresponding aerial photographs. An imaging dataset can also be obtained from a design file, such as a CAD file, e.g., by extracting polygons from the design file. The appearance of an artificially generated imaging dataset, e.g., a design dataset or a simulated imaging dataset, can be modified to resemble the appearance of an acquired imaging dataset, e.g., using machine learning.For example, a generative adversarial neural network comprising a generator and a discriminator can be trained so that the generator learns to mimic the appearance of an acquired imaging dataset, while the discriminator learns to distinguish between "real" acquired imaging datasets and artificially generated imaging datasets with modified appearance. After training, the generator can be used to modify the appearance of artificially generated imaging datasets.

[0030] One or more of the imaging datasets (the first imaging dataset, the second imaging dataset, the third imaging dataset, and potentially additional imaging datasets) may be acquired using an imaging system as described above. One or more of the imaging datasets may be loaded from a database, memory, or cloud storage.

[0031] One or more of the imaging data sets are preferably substantially defect-free and comprise no or only a few defects (i.e., less than 10%, preferably less than 5% of an imaging data set comprises a defect).

[0032] At least the second imaging data set and the third imaging data set comprise mainly the same integrated circuit patterns as the portion of the object. Mainly means that at least 50%, preferably at least 75%, more preferably at least 90%, most preferably at least 95% of the integrated circuit patterns of the second imaging data set and the third imaging data set are the same as the integrated circuit patterns of the portion of the object. In a preferred embodiment, the second imaging data set and the third imaging data set comprise the same integrated circuit patterns as the portion of the object.

[0033] At least the first imaging data set, the second imaging data set, and the third imaging data set are processed together to detect defects. A method that processes at least the first imaging data set, the second imaging data set, and the third imaging data set together to detect defects derives defect information by comparing data obtained directly from at least the first imaging data set, the second imaging data set, and the third imaging data set. Thus, the first imaging data set, the second imaging data set, and the third imaging data set are processed directly in a single step of the defect detection method. They are processed simultaneously.In contrast, for example, a method that does not process the input data sets jointly processes only a subset of the input data sets sequentially to obtain intermediate defect detection results, e.g., by applying the same subroutine to different subsets of the input data sets, and merges the intermediate results at the end, e.g., by combining, selecting, or filtering the detected defects.For example, a defect detection method that does not process the input data sets jointly detects intermediate defects by comparing the first imaging data set with the second imaging data set, followed by detecting intermediate defects by comparing the first imaging data set with the third imaging data set and / or followed by detecting intermediate defects by comparing the second imaging data set with the third imaging data set, and calculating the final defect detection by combining the intermediate defect detections.

[0034] The first imaging dataset, the second imaging dataset, and the third imaging dataset can be preprocessed before the defect detection method is applied. Preprocessing refers to any type of processing of an input dataset that improves, simplifies, or accelerates a subsequent defect detection method. Preprocessing is not a mandatory step of a method; it is always optional, as it only serves to improve the input dataset. Thus, preprocessing steps can be skipped without affecting the functionality of the defect detection method. For example, preprocessing includes registering the input datasets to align them before defect detection, which simplifies the comparison of the input datasets and improves the defect detection results.For example, preprocessing includes filters that can be applied to the first imaging dataset and / or the second imaging dataset and / or the third imaging dataset, e.g., contrast- or brightness-modifying filters, edge detectors, feature extractors, etc. In one example, preprocessing refers to any type of processing that enhances, simplifies, or accelerates a subsequent defect detection method and that does not utilize the other input datasets.

[0035] According to one example, the processing includes estimating one or more defect indicators, and defects are detected using the one or more defect indicators. A defect indicator includes information about properties of a potential defect, e.g., the location, appearance, shape, measurements of the presence of the defect, or the imaging dataset containing the defect. In this way, potential defects can be easily characterized, filtered, compared, and evaluated to detect defects.

[0036] In one example, at least one defect indicator may comprise a pixel-wise map or a voxel-wise map, e.g., a segmentation map, a probability map, a confidence map, etc. By using pixel-wise maps or voxel-wise maps, highly localized information for the defects can be derived. Additionally, pixel-wise maps or voxel-wise maps can be easily compared and processed with other pixel-wise maps or voxel-wise maps. Voxel-wise maps are particularly interesting in the case of 3D input datasets.

[0037] In one example, at least one defect indicator may include a list of defect descriptions, such as bounding boxes, contours, coordinates, measurements, etc. Lists of defect descriptions may be used if pixel-by-pixel information is not required or unavailable, thus saving memory and computation time.

[0038] In one example, at least one defect indicator is a probabilistic defect indicator. A probabilistic defect indicator comprises probabilistic information about defects. Probabilistic information can, for example, include probability values ​​or probability distributions, e.g., about properties of the defect such as intensity, color, or location, or moments of probability distributions, e.g., a mean, a variance, or a higher-order moment of a probability distribution. By using a probabilistic defect indicator, a measure of uncertainty is incorporated into the defect indicator. Uncertainty measures can be used to make more qualified decisions about the presence or absence of a defect, e.g., using additional information such as location, size, shape, etc. They are also useful for further processing routines, e.g.,for routines that combine different defect indicators to determine the presence of a defect. This improves the precision and accuracy of defect detection.

[0039] In one example, at least one defect indicator comprises a defect and a unique identifier of an imaging data set containing the defect. The defect indicator thus contains information about defects and about the imaging data sets from which the defects originate. The unique identifier can, for example, refer to a number, e.g., a "1" for the first imaging data set, a "2" for the second imaging data set, a "3" for the third imaging data set, etc. Thus, using the one or more defect indicators, the imaging data sets from which the defects originate can be identified. In this way, the defects in the imaging data sets can be localized. Additionally, defects can be detected with increased accuracy because information about the source data set can be used to define the probability that a defect is present or not.

[0040] The defect detection method may yield a single defect indicator, or it may yield multiple defect indicators. According to one aspect of the invention, a defect indicator is obtained for at least each of the first imaging data set, the second imaging data set, and the third imaging data set. The defect indicators may be of the same type or different types, i.e., pixel-wise maps or lists of defect descriptions, or combinations thereof. Defects may then be detected by comparing at least the defect indicators for the first imaging data set, the second imaging data set, and the third imaging data set. By using a defect indicator for at least each of the first imaging data set, the second imaging data set, and the third imaging data set, more qualified decisions about the presence of defects can be made.For example, rare defects can be found in imaging datasets that are usually defect-free, such as golden reference datasets. These locations can be ignored when defects are detected in the first imaging dataset, thereby reducing false positives and improving the precision of the defect detection method and the accuracy of the detected defects. In another example, the regularity of the integrated circuit patterns of an imaging dataset can be investigated, and irregular integrated circuit patterns can be ignored when defects are detected in the imaging dataset, thereby reducing false positives and improving the precision of the defect detection method and the accuracy of the detected defects.In another example, if the defect indicators show no defect in two of the imaging datasets, but a defect is likely present in the other imaging dataset, then the probability of a defect is increased. This can prevent false negatives and improve the hit rate of the defect detection method and the accuracy of the detected defects.

[0041] According to one aspect of the invention, at least the first imaging data set, the second imaging data set, and the third imaging data set are registered. Registration is the process of transforming different data sets into a coordinate system. For example, the first imaging data set and the third imaging data set can be transformed into the coordinate system of the second imaging data set, or the first imaging data set and the second imaging data set can be transformed into the coordinate system of the third imaging data set, or the second imaging data set and the third imaging data set are transformed into the coordinate system of the first imaging data set. Alternatively, the first imaging data set, the second imaging data set, and the third imaging data set are transformed into a different coordinate system.The registration of at least the first imaging data set, the second imaging data set, and the third imaging data set can be performed as a preprocessing step before applying the defect detection method. By registering the first imaging data set, the second imaging data set, and the third imaging data set, corresponding integrated circuit patterns can be found at the same coordinates and can be compared more easily and accurately. Thus, the hit rate and precision of the defect detection method and the accuracy of the detected defects are improved.

[0042] In one example, the processing includes evaluating the probability of defects from properties of the layout of the integrated circuit patterns in the section of the object, e.g., by evaluating the regularity of the integrated circuit patterns. In the case of an irregular pattern in an imaging dataset, the probability of a defect is increased, whereas in the case of a regular pattern, the probability of a defect is decreased. The resulting defect probability map can be used as a separate defect indicator, or it can be used to modify another defect indicator.

[0043] In a preferred example of the invention, at least the first imaging dataset, the second imaging dataset, and the third imaging dataset are input datasets of a machine learning model. Machine learning models have several advantages. First, the machine learning model is trained using training data comprising samples. Thus, the user does not have to elaborate suboptimal decision rules or mappings themselves. Instead, the machine learning model learns the complex interrelationships between input data and desired output data in a data-driven manner. Thus, the application is simplified for the user, and even non-expert users can train the machine learning model. Additionally, the mapping between the input data and the output data is optimal with respect to predefined criteria expressed in the loss function.Therefore, the precision and accuracy of the defect detection method, as well as the accuracy of the detected defects, are improved. Furthermore, the machine learning model can be easily adapted if the input data sets change over time. Additionally, although training may take some time, applying a machine learning model to input data—i.e., a forward pass—can be parallelized using GPUs and is very fast, thus reducing runtime and increasing the throughput of the defect detection method.

[0044] Using a machine learning model, the inputs may be processed jointly. In one example, the machine learning model comprises a neural network with at least three input paths, where the first imaging dataset is processed along a first input path, the second imaging dataset is processed along a second input path, and the third imaging dataset is processed along a third input path. An input path refers to a sequence of layers of a neural network that only processes information from one of the input datasets. Each input dataset may be processed directly in the input path, or it may be preprocessed before the machine learning model is applied. The first imaging dataset, the second imaging dataset, and the third imaging dataset are processed independently in separate input paths.By using separate input paths for the input datasets, each input dataset can be processed individually using different processing steps in the input path to derive input dataset-specific information relevant for defect detection. This allows very different types of input datasets to be used in the same machine learning model, e.g., acquired imaging datasets, simulated datasets, design datasets, and / or imaging datasets from different modalities, e.g., SEM, AFM, simulated SEM, etc. Alternatively, each input dataset can be processed using the same processing steps in the input path, by having each input path have the same layers and the same weights. This ensures consistent processing regardless of the input path into which an imaging dataset is fed.

[0045] In a preferred example, the at least three input paths are merged into a single merging layer. By merging the three input paths into a single merging layer, the information of the input paths is combined or fused in the merging layer, e.g., by stacking the intermediate processing results. From the merging layer onwards, the information obtained from the three input data sets is processed together simultaneously. The machine learning model includes at least one step that derives defect information by directly comparing data obtained from at least the first imaging data set, the second imaging data set, and the third imaging data set. Thus, the first imaging data set, the second imaging data set, and the third imaging data set are processed together.Since information from all input data sets for defect detection is available at the same time (instead of sequentially), the hit rate and precision of the defect detection method and the accuracy of the detected defects are improved.

[0046] In one example, the neural network generates a defect indicator for at least each of the first imaging data set, the second imaging data set, and the third imaging data set. By generating two or more defect indicators, defect information specific to the input data sets can be generated. In this way, defects or potential defect locations or potential false-positive defect locations can be identified and used to improve defect detection in the imaging data sets. The three or more defect indicators can be combined to generate a more accurate defect detection result to improve the hit rate and precision of the defect detection method and the accuracy of the detected defects.

[0047] According to one aspect of the invention, the neural network comprises at least three output paths, wherein each defect indicator is generated by a different output path. In this way, a defect indicator is generated for at least each of the first imaging data set, the second imaging data set, and the third imaging data set. Each input path of the neural network corresponds to an output path of the neural network, wherein the corresponding output path generates the defect indicator for the input data set processed by the input path. By generating different defect indicators, input data set-specific defect information can be generated, which can be used to improve the precision and hit rate of the defect detection method and the accuracy of the detected defects.

[0048] In one example, the neural network comprises skip connections, each configured to connect a layer of an input path to a layer of an output path. In particular, each skip connection connects a layer of an input path to a layer of a corresponding output path. In this way, the skip connections can be used to directly access information from the layers of the input paths, from layers of the output paths, in particular from layers of the corresponding output paths. Thus, input dataset-specific information in the output paths can be used in addition to the jointly processed information to generate the defect indicators. In this way, the precision and hit rate of the defect detection method and the accuracy of the detected defects are improved. The neural network can also comprise further skip connections between other layers.

[0049] According to one aspect of the invention, the machine learning model is designed to process the input paths asynchronously. Asynchronous means that processing of an input path can start as soon as the corresponding input data set is available, without waiting for the availability of the other input data sets. Since an input path does not depend on the other input data sets, computing time and resources can be saved by processing the input paths when the input data sets are available. The common part of the neural network can then be processed after the results of all input paths are available.

[0050] In one example, the first imaging dataset, the second imaging dataset, and the third imaging dataset are registered using an additional machine learning model. The additional machine learning model can be trained to register the input datasets. It can be integrated into the neural network architecture as a preprocessing step. By registering the input datasets, corresponding integrated circuit patterns are located at the same location within each input dataset. Thus, comparing the imaged integrated circuit patterns is simplified and therefore more accurate. Therefore, the precision and hit rate of the defect detection method and the accuracy of the detected defects are improved.

[0051] In one example, at least one imaging dataset includes additional information about the probability that a defect is present in the imaging dataset. The probability that a defect is present in the imaging dataset can be derived, for example, from the method used to generate the imaging dataset. This type of information can be used in defect detection to determine whether a defect is present or not. In this way, the precision and hit rate of the defect detection method and the accuracy of the detected defects can be improved.

[0052] According to one embodiment of the invention, a computer-implemented method for training a machine learning model for defect detection according to any of the above-described embodiments, examples, or aspects comprises: providing training images of objects comprising integrated circuit patterns, wherein the training images comprise at least triples of first imaging datasets, second imaging datasets, and third imaging datasets including annotated defects; and training the machine learning model using the provided training images by minimizing a loss function configured for defect detection. The loss function may comprise a cross-entropy loss function and / or a focus loss function, etc., to compare the defect indicators or the detected defects with the annotated defects.

[0053] A computer program according to an embodiment of the invention comprises instructions which, when the program is executed by a computer, cause the computer to perform a method according to one of the embodiments, examples or aspects described above.

[0054] A computer-readable medium according to an embodiment of the invention stores a computer program that can be executed by a computing device, wherein the computer program comprises code for carrying out a method according to one of the embodiments, examples, or aspects described above.

[0055] A defect detection system according to an embodiment of the invention comprises: an imaging device configured to provide one or more imaging data sets of a portion of an object comprising integrated circuit patterns; one or more processing devices; and one or more machine-readable hardware storage devices comprising instructions executable by one or more processing devices to apply a defect detection method according to one of the above-described embodiments, examples, or aspects.

[0056] The invention described by examples and embodiments is not limited to the embodiments and examples, but can be implemented by a person skilled in the art through various combinations or modifications thereof. Short description of the drawings Fig. 1 illustrates an exemplary transmission-based photolithography system, e.g., a deep ultraviolet (DUV) photolithography system; Fig. 2 illustrates an exemplary reflection-based photolithography system, e.g., an extreme ultraviolet (EUV) photolithography system; Fig. 3 shows a first imaging data set of an object comprising integrated circuit patterns in the form of a photolithography mask comprising a defect; Fig. 4 illustrates a flowchart of a computer-implemented method for defect detection according to an embodiment of the invention; Fig. 5 and Fig. 6 illustrate examples of the computer-implemented method for defect detection according to an embodiment of the invention; Fig. 7 illustrates a machine learning model in the form of a neural network that jointly processes the first imaging data set, the second imaging data set, and the third imaging data set; Fig. 8 illustrates a machine learning model in the form of a neural network that jointly processes the first imaging dataset, the second imaging dataset, and the third imaging dataset and includes multiple output paths and skip connections; Fig. 9 illustrates a flowchart of a computer-implemented method for training a machine learning model for defect detection according to an embodiment of the invention; and Fig. 10 illustrates a defect detection system according to an embodiment of the invention. Detailed description

[0057] Advantageous exemplary embodiments of the invention are described below and schematically shown in the figures. In the figures and in the description, the same reference numerals are used to describe identical features or components. Dashed lines indicate optional features.

[0058] The methods described herein may be used, for example, with transmission-based photolithography systems 10 or reflection-based photolithography systems 10', as in Fig. 1 and Fig. 2 shown.

[0059] Fig. 1 illustrates an exemplary transmission-based photolithography system 10, e.g., a DUV photolithography system. Key components include a light source 12, which may be a deep ultraviolet (DUV) excimer laser source, imaging optics, which may include, for example, optics that shape the radiation from the light source 12, a photolithography mask 14, illumination optics 16 that illuminates the photolithography mask 14, and projection optics 18 that projects an image of the photolithography mask pattern onto a photoresist layer of a wafer 20. An adjustable filter or pupil plane aperture of the projection optics 18 can limit the range of beam angles incident on the wafer 20.

[0060] In this document, the terms ‘radiation’ or ‘ray’ refer to all types of electromagnetic radiation, including ultraviolet radiation (e.g. with wavelengths of 365, 248, 193, 157 or 126 nm) and EUV (extreme ultraviolet radiation, e.g. with a wavelength in the range of approximately 3-100 nm).

[0061] The illumination optics 16 may include optical components for shaping, adjusting, and / or projecting radiation from the light source 12 before the radiation passes through the photolithography mask 14. The projection optics 18 may include optical components for shaping, adjusting, and / or projecting the radiation after the radiation has passed through the photolithography mask 14. The illumination optics 16 excludes the light source 12, and the projection optics excludes the photolithography mask 14.

[0062] The illumination optics 16 and the projection optics 18 may comprise various types of optical systems, including, for example, refractive optics, reflective optics, apertures, and catadioptric optics. The illumination optics 16 and the projection optics 18 may also include components that operate according to any of these design types to collectively or individually direct, shape, or control the projection beam of radiation.

[0063] Fig. Figure 2 illustrates an exemplary reflection-based photolithography system 10', e.g., an extreme ultraviolet (EUV) photolithography system 10'. Main components are a light source 12, which may be a laser plasma light source, illumination optics 16, which may include, for example, optics that shape the radiation from the light source 12, a photolithography mask 14, and projection optics 18 that project an image of the photolithography mask pattern onto a photoresist layer of a wafer 20. An adjustable filter or pupil plane aperture of the projection optics 18 can limit the range of beam angles incident on the wafer 20.

[0064] The production of objects containing integrated circuit patterns, such as photolithography masks and wafers, requires great care due to the small feature sizes of the integrated circuit patterns. Defects cannot be prevented but can lead to the malfunction of semiconductor devices. Therefore, an accurate and rapid method for defect detection in objects containing integrated circuit patterns is required for quality control.

[0065] Fig. 3 shows a first imaging data set 22 of an object comprising integrated circuit patterns, in the form of a photolithography mask 14 comprising a defect 24. Methods known in the art often use die-to-die or die-to-database methods to detect such defects 24. Die-to-die methods compare a portion of the first imaging data set 22 with another portion of the same or a different first imaging data set 22 to detect the defects 24. Die-to-database methods provide an artificial reference data set, e.g., a design data set or a simulated data set, which is compared with the first imaging data set 22 of the object comprising integrated circuit patterns. The hit rate and precision of the defect detection method and the accuracy of the detected defects depend on the quality of the reference data set, e.g.,the noise level, the accuracy of the structures, the alignment of the reference dataset and the first imaging dataset 22, etc. It is therefore important to prevent false-negative and false-positive defect detections.

[0066] By using three or more imaging datasets, the additional information can be used to make a qualified decision about the presence of defects. However, known methods that use three or more imaging datasets usually rely on pairwise comparisons of the first imaging dataset 22 and each other imaging dataset. This process is suboptimal because the available information is only considered sequentially, and because it can lead to ambiguities if contradictory results are obtained from the two or more comparisons. In addition, the computing time increases with the number of imaging datasets. It is therefore an object of the invention to improve the hit rate and precision of the defect detection method and the accuracy of the detected defects, while still achieving low runtimes and high throughput of the defect detection method.

[0067] Fig. 4 illustrates a flowchart of a computer-implemented method 26 for defect detection according to an embodiment of the invention. The method comprises the following steps: obtaining a first imaging data set of a portion of an object comprising integrated circuit patterns in a step M1; obtaining at least a second imaging data set and a third imaging data set comprising mainly the same integrated circuit patterns as the portion of the object in a step M2; jointly processing at least the first imaging data set, the second imaging data set, and the third imaging data set to detect defects in a step M3.

[0068] By using at least the first imaging data set 22, the second imaging data set, and the third imaging data set, additional information is available for defect detection, thereby improving the hit rate and precision of the defect detection method and the accuracy of the detected defects. Instead of performing sequential pairwise comparisons, e.g., by comparing the first imaging data set 22 with the second imaging data set, followed by a comparison of the first imaging data set 22 with the third imaging data set, at least the first imaging data set 22, the second imaging data set, and the third imaging data set are processed together. This saves computing time and increases throughput.Furthermore, jointly considering information from the first imaging data set 22, the second imaging data set, and the third imaging data set allows for a more qualified decision regarding the presence of a defect. Additionally, defects can be detected in each of the imaging data sets, particularly in golden reference data sets or artificially generated imaging data sets that are usually defect-free. Thus, false-negative and false-positive defect detections due to the simultaneous use of at least three imaging data sets can be prevented.

[0069] In one example, at least the first imaging data set 22, the second imaging data set, and the third imaging data set comprise acquired imaging data sets obtained by applying an imaging device, e.g., an SEM, a FIB-SEM, an AFM, an aerial image acquisition system, etc., equipped with a rigid array sensor, a line scan sensor, or a time-delayed integration sensor, an X-ray image, etc. Alternatively, the first imaging data set 22, the second imaging data set, and / or the third imaging data set may be acquired by another imaging device. Alternatively, the first imaging data set 22, the second imaging data set, and / or the third imaging data set may be artificially generated, e.g., by means of simulation or from designs of the integrated circuit patterns.

[0070] In one example, one or more of the imaging datasets are acquired imaging datasets that primarily comprise the same integrated circuit patterns as the portion of the object and are substantially defect-free, so-called "golden reference datasets." In another example, one or more of the imaging datasets are obtained from designs that primarily comprise the same integrated circuit patterns as the portion of the object, e.g., from the design of the object itself or from other, but similar, designs. The appearance of the design can be modified to match the appearance of an acquired imaging dataset, e.g., by applying machine learning techniques such as generative adversarial neural networks.In a preferred example, the first imaging data set is an acquired imaging data set of the portion of the object, the second imaging data set is an acquired imaging data set of another die of the same object, mainly comprising the same integrated circuit patterns, and the third imaging data set is a golden reference data set obtained from a different object, or a simulated data set, or a design data set. In another preferred example, the first imaging data set is an acquired imaging data set, the second imaging data set is obtained from a simulation or a design, and the third imaging data set is a golden reference data set. In this case, rarely occurring defects in the golden reference data set can also be detected.In another preferred example, the first imaging data set, the second imaging data set, and the third imaging data set are acquired imaging data sets obtained from different sections of the same object, which primarily comprises the same integrated circuit patterns. By using three imaging data sets, additional information is available that can be used to increase the precision and hit rate of the defect detection method and the accuracy of the detected defects. This approach can be referred to as a die-to-die defect detection method.

[0071] Fig. 5 illustrates an example of the steps of the computer-implemented method 26 for defect detection according to an embodiment of the invention. A first imaging data set 22 of a portion of an object comprising integrated circuit patterns in the form of a photolithography mask is obtained, e.g., using an SEM. The first imaging data set 22 includes a defect 24. Additionally, a second imaging data set 28 and a third imaging data set 30 are obtained. The second imaging data set 28 includes an acquired first imaging data set of another die of the same object, primarily comprising the same integrated circuit patterns. The third imaging data set 30 is obtained from a design of the photolithography mask. It may include polygons modeling the integrated circuit patterns. Additional imaging data sets, e.g.,A fourth imaging data set, or a fourth and a fifth imaging data set, may also be obtained. At least the first imaging data set 22, the second imaging data set 28, and the third imaging data set 30 are processed together. The processing 32 may be performed by some type of defect detection method that derives defect information by directly comparing data obtained from at least the first imaging data set 22, the second imaging data set 28, and the third imaging data set 30.In contrast, a defect detection method that compares the first imaging data set 22 with the second imaging data set 28 in a first step and then compares the first imaging data set 28 with the third imaging data set 30 in a second step and detects defects based on the two results does not process the first imaging data set 22, the second imaging data set 28 and the third imaging data set 30 together, but sequentially.

[0072] In an example shown in Fig. 5, the processing includes estimating one or more defect indicators 34, and defects 24 are detected using the one or more defect indicators 34. A defect indicator 34 indicates (potential) defects 24 in one or more imaging data sets, e.g., in the first imaging data set 22, the second imaging data set 28, or the third imaging data set 30, or in further imaging data sets, or in two or more imaging data sets together.

[0073] A defect indicator 34 may comprise a pixel-wise map 36 or a voxel-wise map that indicates the presence or absence of a defect 24 for each pixel. In the case of a 2D imaging data set (e.g., an image or a stack of images), such a pixel-wise map 36 may be obtained using a defect detection method that decides for each pixel in the data set whether a defect 24 is present or not, or that includes the probability of whether a defect 24 is present or not. In the case of a 3D data set (a voxel grid), a voxel-wise map may be obtained using a defect detection method that decides for each voxel in the data set whether a defect 24 is present or not, or that includes the probability of whether a defect 24 is present or not.For example, a pixel-wise map 36 or a voxel-wise map may include a confidence map indicating the probability of the absence of a defect, as in . Fig. 5. Alternatively, the pixel-wise map 36 or voxel-wise map may comprise a defect probability map or a segmentation. A pixel-wise map 36 or voxel-wise map may be obtained, for example, using a segmentation algorithm, e.g., a machine learning method for defect segmentation.

[0074] Alternatively, a defect indicator 34 may comprise a list 38 of defect descriptions 40. Using lists 38 of defect descriptions 40 may be useful if the accuracy of a pixel-wise map 36 is not required for further processing, or if pixel-wise maps 36 are not available or cannot be generated by the defect detection method. A defect description 40 contains information about a (potential) defect 24. A defect description 40 may, for example, indicate the location of the defect and / or the shape of the defect and / or measurements of the defect and / or the appearance of the defect and / or the presence of the defect ("yes" / "no" or a probability of its presence or absence), etc. For example, a defect description 40 may comprise one or more coordinates of the defect 24, e.g.Center point coordinates, boundary coordinates, centroid coordinates, bounding box coordinates, or any other coordinates related to the defect 24. A defect description 40 may include a bounding box of any shape and size, e.g., a rectangular bounding box, a circular bounding box, an elliptical bounding box, a perimeter of the defect, a free-form bounding box, a 3D bounding box, etc. A defect description 40 may include a contour or a surface of the defect 24. A defect description 40 may include one or more measurements of the defect 24, such as a length, a diameter, a radius, an area, a surface size, a contour length, etc. A defect description 40 may include one or more appearance properties of the defect 24, e.g., an intensity or color distribution, the variance of the intensity or color distribution, edge locations, etc.A defect description 40 may include one or more shape attributes, e.g., shape descriptors, eccentricity, angles, lengths, ratios, curvature, Fourier descriptors, etc. A defect description 40 may also include any other properties of the defect.

[0075] In one example, at least one defect indicator 34 is a probabilistic defect indicator. A probabilistic defect indicator includes probabilistic information about defects 24. For example, a pixel-wise map 36 or a voxel-wise map may include probabilities for the presence or absence of a defect 24. For example, a defect description 40 may include a probability for the presence or absence of 24, or it may include a probability distribution over properties of the defect, e.g., over intensities, color values, locations, etc., or it may include moments of probability distributions over properties, e.g., a mean, a variance, a kurtosis, or higher-order moments, e.g., a mean intensity value or a color variance. For example, the two pixel-wise maps 36 in Fig. 5 Probability values ​​for the absence of a defect, ie confidence maps.

[0076] The one or more defect indicators 34 may all be of the same type, e.g., pixel-wise maps 36 or lists 38 of defect descriptions 40, or they may be of different types, e.g., one pixel-wise map 36 and two lists 38 of defect descriptions 40, etc.

[0077] In one example, at least one defect indicator 34 comprises a defect and a unique identifier of an imaging data set containing the defect. For example, the defect indicator 34 contains a list 38 of defect descriptions 40, and for at least one of the defect descriptions 40, the imaging data set containing the defect to which the at least one defect description 40 refers is indicated by a unique identifier. For example, if a defect description 40 refers to a defect in the second imaging data set, the defect description 40 may include the unique identifier "2." Alternatively, the entire list of defect descriptions 40 may be associated with one or more unique identifiers of the imaging data sets from which the defects in the list originate, e.g., a "1" and a "2" indicating the first imaging data set 22 and the second imaging data set 28.If a defect indicator 34 comprises a pixel-wise map 36 or a voxel-wise map, individual pixels or sets of pixels, or individual voxels or sets of voxels in the map, may be associated with one or more unique identifiers of the one or more imaging data sets from which the pixel or voxel information is obtained. In this way, defects are associated with the specific imaging data sets from which the defect information is obtained.

[0078] Defects 24 can be detected using the one or more defect indicators 34. For example, if a single defect indicator 34 is obtained after processing the first imaging data set 22, the second imaging data set 28, and the third imaging data set 30, the defects 24 can be extracted from the defect indicator 34, e.g., from a pixel-by-pixel map 36, for example, by thresholding a confidence or probability map, or using the entries in a list 38 of defect descriptions 40. If two or more defect indicators 34 are available for different input data sets, e.g.,for the first imaging data set 22 and the second imaging data set 28 and / or the third imaging data set 30, the defect indicators 34 can be processed, combined or compared together to make a qualified decision about the presence of a defect 24.

[0079] In one example, processing 32 includes estimating a defect indicator 34 for at least each of the first imaging data set 22, the second imaging data set 28, and the third imaging data set 30. By comparing at least the defect indicators 34 of the first imaging data set 22, the second imaging data set 28, and the third imaging data set 30, the defects 24 can be detected with improved precision and accuracy. For example, potential defects 24 that appear in the defect indicator 34 of the first imaging data set 22, but not in the defect indicator of the second imaging data set 28 and the third imaging data set 30 can be marked as a defect 24 with a high probability.On the other hand, potential defects 24 that appear in the defect indicator 34 of the second imaging data set 28 or in the defect indicator 34 of the third imaging data set 30 may indicate a false-positive defect detection and may not be marked as a defect 24 in the other imaging data sets 22. Potential defects in the defect indicators 34 of imaging data sets that are usually defect-free, e.g., golden reference data sets, artificially generated imaging data sets, etc., may indicate rare defects 24. Such defects, which usually lead to false-positive defect detections in other imaging data sets, can be handled by processing three or more imaging data sets.

[0080] In Fig. 5, the defect indicator 34 in the form of a pixel-wise map 36 (a confidence map for the absence of a defect) for the first imaging data set 22 indicates a low confidence at the location of the defect 24 because the integrated circuit patterns in the second imaging data set 28 and the third imaging data set 30 do not correspond to the integrated circuit patterns in the first imaging data set 22. The defect indicator 34 in the form of a list 38 of defect descriptions 40 includes two defect descriptions D1 and D2 corresponding to defects 29 contained in the other die of the same object used as the second imaging data set 28. The defect descriptions 40 may include, for example, the location of the center of the defect 24, the size of the defect, the shape of the defect (e.g., using shape descriptors), properties of the object such as intensity or color, intensity or color variance, etc.The defect indicator 34 in the form of a pixel-wise map 36 (a confidence map for the absence of a defect) for the third imaging data set 30 indicates a high confidence.

[0081] Using the defect indicators 34 of the first imaging data set 22, the second imaging data set 28, and the third imaging data set 30 together in an optional defect detection 42, defects 24 can be detected with improved accuracy and precision using the one or more defect indicators 34. The defect detection 42 can use any type of defect detection method that detects defects from one or more defect indicators 34, e.g., thresholding, difference images, segmentation, contour detection, finding connected components, machine learning methods, etc. For example, the defects 29 in the second imaging data set 28 can be detected, e.g., by comparing the list of defect descriptions D1 and D2 with the pixel-by-pixel map 36 of the third imaging data set 30.For example, the corresponding locations of the potential defects 29 in the second imaging dataset 28 may be ignored in the first imaging dataset 22. The low-confidence regions in the pixel-wise map 36 of the first imaging dataset 22 may be marked as a defect 24 with increased probability because the list 38 of defect descriptions 40 does not contain an entry for this location and the pixel-wise map 36 of the third imaging dataset 30 has a high confidence at this location. For example, a machine learning model may be trained to detect defects from the one or more defect descriptors 34. If a single defect descriptor 34 comprises a pixel-wise probability map, thresholding may be used to detect defects.

[0082] In the Fig. 6, the processing includes estimating one or more defect indicators 34. As for Fig. 5, defects 24 are detected by defect detection 42 using the one or more defect indicators 34. The defect indicator 34 in the form of a pixel-wise map 36 (a confidence map for the absence of a defect) for the first imaging data set 22 indicates a low confidence at the location of the defect 24 because the integrated circuit patterns in the second imaging data set 28 and the third imaging data set 30 do not correspond to the integrated circuit patterns in the first imaging data set 22. The pixel-wise maps 36 (confidence maps for the absence of a defect) for the second imaging data set 28 and the third imaging data set 30 indicate a high confidence.The second imaging data set 28 and the third imaging data set 30 have corresponding integrated circuit patterns, while only the integrated circuit patterns of the first imaging data set 22 differ in the location of the defect 24. The low-confidence locations in the pixel-wise map 34 of the first imaging data set 22 can thus be marked as a defect 24 with high probability, since both the second imaging data set 28 and the third imaging data set 30 have a high confidence at these locations.

[0083] In one example, the one or more defect indicators 34 are filtered prior to detecting defects 24. For example, the defect indicators 34 may be filtered using prior knowledge about the defects 24 of interest, e.g., regarding size, location, shape, material, or specific properties, etc. For example, the defect indicators 34 may be filtered by removing potential defects below a predefined minimum size, in specific locations, within materials not of interest, with confidence or defect probability values ​​or average confidence or defect probability values ​​above or below a predefined value, with intensity variances above a certain threshold, with atypical shapes, with properties that fall outside confidence intervals created using primarily defect-free imaging datasets, etc.By filtering the defect indicators 34, false-positive defect detections or irrelevant defect detections can be prevented.

[0084] Types of defect indicators 34, e.g., pixel-wise maps 36 or lists 38 of defect descriptions 40, can be transformed into other types of defect indicators. For example, a pixel-wise map 36 can be transformed into a list 38 of defect descriptions 40 by, e.g., deriving bounding boxes, center or boundary coordinates, sizes, locations, or other properties of the defects 24 from the pixel-wise map 36. A list 38 of defect descriptions 40 can be transformed into a pixel-wise map 36 by, e.g., assigning specific values ​​or probabilities to the pixels depending on the defect descriptions 40. For example, a value of 0 can be assigned to all pixels that potentially belong to a defect according to the defect descriptions 40, while a value of 1 is assigned to all pixels that do not belong to any of the defect descriptions 40, or vice versa.For example, if a defect description contains bounding boxes of defects, a value of 0 can be assigned to all pixels within the bounding boxes. Intermediate values, i.e., confidence values ​​or defect probability values, can be assigned to pixels that may or may not belong to a defect, e.g., pixels near the boundary of a bounding box or pixels near a defect's center coordinate.

[0085] In one example, defects 24 are detected by applying a function to the one or more defect indicators 34. The defect indicators 34, which include lists 38 of defect descriptions 40, can be transformed into pixel-wise maps 36 before the function is applied. The function can, for example, take the following form f(i1(x),i2(x),i3(x))=i1(x)⋅max(i2(x),i3(x)), where i1(x) indicates a value of a pixel-wise map 36 for the first imaging data set 22 at a location x, i2(x) indicates a value of a pixel-wise map 36 for the second imaging data set 28 at a location x, and i3(x) indicates a value of a pixel-wise map 36 for the third imaging data set 30. Using this function, a confidence map of the first imaging data set 22 can be modified by reducing the confidence in the case of a defect in both the second and third imaging data sets 28, 30. Numerous other functions are conceivable.

[0086] To obtain the defect indicators 34, various defect detection methods can be applied that jointly process at least the first imaging data set 22, the second imaging data set 28, and the third imaging data set 30, for example, classical computer vision methods such as graph-based methods, energy optimization methods, variational methods, PDE-based methods, clustering methods, area growth methods, histogram-based methods, etc. On the other hand, machine learning methods can be used, such as support vector machines, decision trees, Bayesian classifiers, clustering algorithms, neural networks, in particular deep neural networks, etc.

[0087] For example, to obtain a pixel-wise map 36 of an imaging dataset, a machine learning model trained in an unsupervised manner on substantially defect-free images, such as a one-class support vector machine, a multi-layer perceptron with a single hidden layer, a codebook, or an autoencoder, can be used, which takes patches of the imaging dataset as input. These methods learn a subspace of substantially defect-free image patches from training data. Based on a distance of an image patch to the learned subspace, a decision can be made, for example, using a threshold, as to whether the image patch contains a defect or not. Alternatively, supervised machine learning methods such as a deep encoder-decoder neural network, e.g., a U-Net, can be used, which take an imaging dataset as input and generate a pixel-wise map 36, e.g.,a segmentation or semantic segmentation, as output.

[0088] To obtain a list 38 of defect descriptions 40, a bounding box detection method, such as CenterNet or YOLO, can be used, which takes the first imaging dataset 22, the second imaging dataset 28, and the third imaging dataset 30 as input. CenterNet is a machine learning-based object detector that represents objects by a single point at their bounding box center. Other properties such as object size, dimensions, 3D extent, orientation, and pose are directly inferred from image features at the center location. Object detection is thus considered a keypoint estimation problem. The input image is used as input to a fully convolutional network that generates a heatmap whose vertices correspond to object centers. Image features at each vertex predict the object's bounding box.YOLO is a very fast machine learning-based object detector that uses a fully convolutional neural network for bounding box prediction. The image is divided into grid cells, and for each grid cell, a specific number of bounding boxes is predicted, which can be larger than the grid cell. Bounding boxes are preserved based on their class probabilities and bounding box confidences. By using bounding box-based object detectors, areas containing a defect 24 can be quickly distinguished from defect-free areas.

[0089] In another example, defects may be detected by comparing the integrated circuit patterns in the first imaging data set 22, the second imaging data set 28, the third imaging data set 30, and potentially additional imaging data sets. If the integrated circuit patterns in the first imaging data set 22 and at least one other imaging data set are identical, a high confidence may be assigned to the corresponding location in the first imaging data set 22. If the integrated circuit patterns in the first imaging data set 22 differ from the integrated circuit patterns in all or a substantial number of the other imaging data sets 28, 30, the probability of a defect at the corresponding location in the first imaging data set 22 may be increased.To detect potential defects in a golden reference, the integrated circuit patterns in the corresponding imaging data set may be compared with the integrated circuit patterns in the one or more other imaging data sets, in particular with imaging data sets that include a design of the integrated circuit patterns of the portion of the object.

[0090] In one example, processing 32 includes evaluating the probability of defects 24 from characteristics of the layout of the integrated circuit patterns in the portion of the object. For example, the irregularity of the integrated circuit patterns in an imaging dataset, e.g., in a design dataset, can be used to define a confidence map for a defect. In the case of irregular integrated circuit patterns, a defect 24 is likely to be detected by a defect detection method, often leading to false-positive defect detections. For example, a bridge, a broken wire, or rounded edges form irregular integrated circuit patterns in an object. Irregular integrated circuit patterns can be detected, for example, using machine learning. For example, an autoencoder can be trained using only regular integrated circuit patterns.Then, by comparing an input dataset with its reconstruction by the autoencoder, irregular integrated circuit patterns cannot be fully reconstructed and cause deviations. From these deviations, irregular integrated circuit patterns can be detected in one imaging dataset, e.g., a design dataset. Defects detected in these regions in another imaging dataset, e.g., an acquired imaging dataset, can then be discarded as false positive defect detections.

[0091] The joint processing of at least the first imaging data set 22, the second imaging data set 28 and the third imaging data set 30 can also be carried out by a machine learning model.

[0092] Fig. 7 and Fig. 8 illustrate machine learning models in the form of neural networks 44, 46 that jointly process the first imaging data set 22, the second imaging data set 28, and the third imaging data set 30. The neural networks 44, 46 comprise at least three input paths, wherein the first imaging data set 22 is processed along a first input path 48, the second imaging data set 28 is processed along a second input path 50, and the third imaging data set 30 is processed along a third input path 52. The layers of the input paths 48, 50, 52 can differ in their number, type, and / or size, etc. In this way, the at least three inputs, the first imaging data set 22, the second imaging data set 28, and the third imaging data set 30, can be processed in different ways.It is also possible for the at least three inputs to be processed in the same way using identical input paths 48, 50, 52 with identical weights. The neural networks 44, 46 may have additional input paths for processing additional imaging data sets.

[0093] In the Fig. 7 and Fig. In the examples illustrated in Figure 8, the at least three input paths are merged into a single merging layer 54. The neural network also includes a subsequent processing path that starts from the merging layer 54 and ends in the output layer, processing information from all three inputs. Thus, both neural networks 44, 46 provide at least three input data sets, including the first imaging data set 22, the second imaging data set 28, and the third imaging data set 30, which are processed through the three input paths 48, 50, 52. The neural networks 44, 46 also include at least one step, the merging layer 54, which uses information from at least the first imaging data set 22, the second imaging data set 28, and the third imaging data set 30 to derive defect information.Thus, the first imaging data set 22, the second imaging data set 28 and the third imaging data set 30 are processed together.

[0094] The subsequent processing path between the merging layer 54 and the output layer may, for example, include an encoder that maps the inputs into a feature space. In another example, the subsequent processing path includes a combination of an encoder, a bottleneck, and a decoder. The subsequent processing path may include only a decoder. The subsequent processing path may include further operations such as convolutional layers, pooling layers, dropout layers, etc.

[0095] In Fig. 7, the subsequent processing path between the merging layer 54 and the output layer of the neural network 44 results in a single defect indicator 34, e.g., a pixel-by-pixel map 36 such as a confidence map, a defect probability map, or a segmentation of an imaging dataset 22 or a list 38 of defect descriptions 40, e.g., a list of bounding boxes indicating defects 24 or center coordinates of defects 24, etc., in one of the imaging datasets. The defect indicator 34 may include a unique identifier indicating the imaging dataset including the defects 24. Alternatively, the subsequent processing path may generate a defect indicator 34 comprising at least three channels such that each channel indicates defects in one of the imaging datasets. During training of the neural network, each channel may be compared to the annotated defects within each of the imaging datasets.Using the defect indicator 34, defects can be detected, e.g., by thresholding a probability map or by extracting defect locations from a defect description 40, etc.

[0096] In Fig. 8, the subsequent processing path of the neural network 46, which starts at the merging layer 54, results in three defect indicators 34, e.g., a defect indicator 34 for the first imaging data set 22, a defect indicator 34 for the second imaging data set 28, and a defect indicator 34 for the third imaging data set 30. Each defect indicator 34 is generated by a different output path 58, 60, 62, in particular by the output path 58, 60, 62 corresponding to the input path 48, 50, 52. In Fig. 8, the defect indicator 34 for the first imaging data set 22 is generated by a first output path 58. The first output path 58 corresponds to the first input path 48. The defect indicator 34 generated by the first output path 58 indicates defects 24 in the first imaging data set 22. The defect indicator 34 for the second imaging data set 28 is generated by a second output path 60. The second output path 58 corresponds to the second input path 50. The defect indicator 34 generated by the second output path 60 indicates defects 24 in the second imaging data set 28. The defect indicator 34 for the third imaging data set 30 is generated by a third output path 62. The third output path 62 corresponds to the third input path 52. The defect indicator 34 generated by the third output path 62 indicates defects 24 in the third imaging data set 30.Instead of directly assigning input paths to output paths, the defect indicators 34 may contain a unique identifier of the imaging dataset from which they are generated. The defect indicators 34 may also include, for each defect, a unique identifier of an imaging dataset containing the defect.

[0097] In a Fig. In the example illustrated in Figure 8, the neural network includes skip connections 56 that connect layers of an input path 48, 50, 52 to layers of an output path 58, 60, 62. In particular, the skip connections 56 connect layers of an input path 48, 50, 52 to layers of a corresponding output path 58, 60, 62. For example, layers of the first input path 48 are connected to layers of the first output path 58, layers of the second input path 50 are connected to layers of the second output path 60 (not shown here), and layers of the third input path 52 are connected to layers of the third output path 62. By connecting input paths to output paths, information from the specific input path can be used directly by the output path, e.g.specific information about the first imaging data set 22 contained in the first input path 48 can be used directly by the corresponding first output path 58 to generate the defect indicator 34 for the first imaging data set 22, specific information about the second imaging data set 28 contained in the second input path 50 can be used directly by the corresponding second output path 60 to generate the defect indicator 34 for the second imaging data set 28, and / or specific information about the third imaging data set 30 contained in the third input path 52 can be used directly by the corresponding third output path 62 to generate the defect indicator 34 for the third imaging data set 22.

[0098] Furthermore, a skip connection 56 can connect a layer of an input path 48, 50, 52 to a layer of a non-corresponding output path 58, 60, 62, e.g., a layer of the first input path 48 to a layer of the second or third output path 60, 62. In this way, specific information about a data set, e.g., about the first imaging data set 22, can be used in one or more non-corresponding output paths, e.g., in the second and / or third output path 60, 62, to generate the defect indicator 34 for the second and / or third imaging data set 28, 30, etc. The skip connections 56 are optional. They can be used to connect any number of layers, but they can also be omitted, particularly for layers near the merge layer 54.

[0099] In one example, the machine learning model is designed to process the input paths asynchronously. Asynchronous means that the processing of an input path can be started as soon as the input data set is available. For example, if a second imaging data set 28 is already available, but a first imaging data set 22 and a third imaging data set 30 are not yet available, the processing of the second input path 50 can already be carried out using the available second imaging data set 28. As soon as the first imaging data set 22 or the third imaging data set 30 is available, the processing of the corresponding input path 48, 52 can also be carried out. In the case of a single merging layer 54, as in Fig. 7 and Fig. As shown in Figure 8, asynchronous processing of input paths can be performed up to the merge layer 54. Processing continues as soon as all results of the input paths are available.

[0100] In one example, at least one imaging data set 22, 28, 30 includes additional information about the probability that a defect 24 is present in the imaging data set 22, 28, 30. The probability that a defect 24 is present in the imaging data set 22, 28, 30 can be derived, for example, from the generation method of the imaging data set 22, 28, 30. For example, an imaging data set 22, 28, 30 acquired from the object comprising integrated circuit patterns may include defects 24 with a low probability, since defects 24 rarely occur in these objects. An imaging data set 22, 28, 30 of an object that has been inspected for defects (a golden reference) may still include defects 24, but with a lower probability.A simulated imaging data set or an imaging data set obtained from a CAD file is usually defect-free, and thus the probability that a defect 24 is present is close to 0. This type of information can be used in defect detection 42 to decide whether a defect is present or not. For example, if the first and second imaging data sets 22, 28 contain the same or a similar defect, a defect 24 will likely be detected in the third imaging data set 30 because it differs from the first imaging data set 22 and the second imaging data set 28. If the defect probability in the third imaging data set 30 is close to 0, but higher for the first imaging data set 22 and for the second imaging data set 28, the defect 24 can be correctly detected as present in the first imaging data set 22 and the second imaging data set 24.In this way, the precision and hit rate of the defect detection method and the accuracy of the detected defects can be improved.

[0101] According to an embodiment of the invention described in Fig. 9, a computer-implemented method 64 for training a machine learning model for defect detection, as described above, comprises: providing training images of objects comprising integrated circuit patterns, wherein the training images comprise at least triples of first imaging datasets 22, second imaging datasets 28, and third imaging datasets 30 including annotated defects 24, in a step T1; and training the machine learning model using the provided training images by minimizing a loss function configured for defect detection, in a step T2. The annotated defects may be obtained, for example, using human annotators, simulated datasets, or design datasets.The loss function may, for example, comprise a channel-wise binary cross-entropy loss function and / or a focal loss function that compares the defect indicator 34 with the annotated defects. The annotated defects may be transformed into a ground-truth defect indicator, for example, into a pixel-wise map 36, e.g., a confidence map, or into a list 38 of defect descriptions 40, e.g., bounding boxes. The ground-truth defect indicator may then be compared with the defect indicator 34 in the loss function.

[0102] A defect detection system 66 according to an embodiment of the invention shown in Fig.10, includes an imaging device 70 for obtaining one or more imaging data sets of a portion of an object 72 comprising integrated circuit patterns, and a data analysis device 68 comprising one or more processing devices 74 and one or more machine-readable hardware storage devices 76 comprising instructions executable by one or more processing devices 74 to perform operations comprising a computer-implemented method 26 for defect detection, as described above.

[0103] The system 66 may optionally include a database 80, e.g., for loading and / or storing first imaging data sets 22, second imaging data sets 28, third imaging data sets 30, training images, etc. The imaging device 70 for obtaining one or more imaging data sets of the object 72 comprising integrated circuit patterns may, for example, comprise a charged particle beam device, e.g., a helium ion microscope, a crossbeam device including FIB and SEM, an atomic force microscope, or any other charged particle imaging device or aerial image acquisition system.The imaging device 70 for obtaining one or more imaging data sets of the object 72 comprising integrated circuit patterns may provide a first imaging data set 22 and / or a second imaging data set 28 and / or a third imaging data set 30 and / or further imaging data sets to the data analysis device 68. Alternatively, the first imaging data set 22 and / or the second imaging data set 28 and / or the third imaging data set 30 and / or further imaging data sets may be loaded from a database or storage device or from a cloud.

[0104] The data analysis device 68 includes one or more processors 74, implemented, for example, as a CPU or GPU. The one or more processors 74 may receive the first imaging data set 22 and / or the second imaging data set 28 and / or the third imaging data set 30 and / or other imaging data sets via an interface 78. The one or more processors 74 may load program code from a hardware storage device 76, e.g., program code for executing a computer-implemented method 26 for detecting defects 24 according to an embodiment of the invention, as described above. The one or more processors may also load a first imaging data set 22 and / or a second imaging data set 28 and / or a third imaging data set 30 and / or other imaging data sets from the hardware storage device 76.The one or more processors 74 may execute the program code.

[0105] The system 66 may optionally include a user interface 82, e.g., for monitoring the training progress of a machine learning model, selecting training parameters, etc.

[0106] The methods disclosed herein can be used, for example, during research and development of objects comprising integrated circuit patterns, or during serial production of objects comprising integrated circuit patterns, or for process window qualification or improvement. Additionally, the methods disclosed herein can also be used for defect detection of first X-ray imaging data sets of objects comprising integrated circuit patterns, e.g., after packaging of the semiconductor device for delivery.

[0107] Reference throughout this specification to "one embodiment," "an example," or "an aspect" means that a particular feature, structure, or characteristic described in connection with the embodiment, example, or aspect is included in at least one embodiment, example, or aspect. Thus, appearances of the phrases "according to an embodiment," "according to an example," or "according to an aspect" in various places throughout this specification do not necessarily refer to the same embodiment, example, or aspect, but may do so. Further, the particular features or characteristics may be combined in any suitable manner in one or more embodiments, as would be apparent to one of ordinary skill in the art from this disclosure.

[0108] Furthermore, while some embodiments, examples, or aspects described herein include some, but not other, features included in other embodiments, examples, or aspects, combinations of features of different embodiments, examples, or aspects are intended to be within the scope of the claims and to form different embodiments, as would be understood by those skilled in the art.

[0109] In summary, a computer-implemented method 26 comprises: obtaining a first imaging data set 22 of a portion of an object 72 comprising integrated circuit patterns; obtaining at least a second imaging data set 28 and a third imaging data set 30 comprising primarily the same integrated circuit patterns as the portion of the object 72; and jointly processing at least the first imaging data set 22, the second imaging data set 28, and the third imaging data set 30 to detect defects 24. The invention also relates to a corresponding computer program, computer-readable medium, and system for defect detection. List of reference symbols 10, 10' photolithography system 12 Light source 14 Photolithography mask 16 Lighting optics 18 Projection optics 20 wafers 22 First imaging dataset 24 Defect 26 Computer-implemented procedure 28 Second imaging dataset 29 Defect 30 Third imaging dataset 32 Processing 34 Defect indicator 36 pixel-by-pixel map 38 List 40 Defect description 42 Defect detection 44, 46 Neural network 48 First input path 50 Second input path 52 Third input path 54 Merge layer 56 Skip connection 58 First output path 60 Second output path 62 Third output path 64 Computer-implemented procedure 66 systems 68 Data analysis device 70 imaging device 72 objects 74 Processing device 76 Hardware storage device 78 Interface 80 database 82 User interface QUOTES CONTAINED IN THE DESCRIPTION

[0000] This list of documents submitted by the applicant was generated automatically and is included solely for the convenience of the reader. This list is not part of the German patent or utility model application. The DPMA assumes no liability for any errors or omissions. Cited patent literature

[0000] US 2019 / 0130551 A1

[0015] US 2006 / 204109 A1

[0017]

Claims

[1] Computer-implemented method (26) for defect detection, comprising: - Obtaining a first imaging data set (22) of a portion of an object (72) comprising integrated circuit patterns; - Obtaining at least one second imaging data set (28) and one third imaging data set (30) comprising mainly the same integrated circuit patterns as the portion of the object (72); and - jointly processing (32) at least the first imaging data set (22), the second imaging data set (28) and the third imaging data set (30) to detect defects (24). [2] The method of claim 1, wherein the processing (32) comprises estimating one or more defect indicators (34), and wherein defects (24) are detected using the one or more defect indicators (34). [3] The method of claim 2, wherein at least one defect indicator (34) comprises a pixel-wise map (36) or a voxel-wise map. [4] Method according to claim 2 or 3, wherein at least one defect indicator (34) comprises a list (38) of defect descriptions (40). [5] Method according to one of claims 2 to 4, wherein at least one defect indicator (34) is a probabilistic defect indicator. [6] Method according to one of claims 2 to 5, wherein at least one defect indicator (34) comprises a defect and a unique identifier of an imaging data set containing the defect. [7] The method of any one of claims 2 to 6, wherein a defect indicator (34) is obtained for at least each of the first imaging data set (22), the second imaging data set (28) and the third imaging data set (30). [8] The method of claim 7, wherein defects (24) are detected by comparing at least the defect indicators (34) for the first imaging data set (22), the second imaging data set (28) and the third imaging data set (30). [9] Method according to one of the preceding claims, wherein at least the first imaging data set (22), the second imaging data set (28) and the third imaging data set (30) are registered. [10] A method according to any preceding claim, wherein the processing (32) comprises evaluating the probability of defects (24) from properties of the layout of the integrated circuit patterns in the portion of the object (72). [11] Method according to one of the preceding claims, wherein at least the first imaging data set (22), the second imaging data set (28) and the third imaging data set (30) are input data sets of a machine learning model. [12] The method of claim 11, wherein the machine learning model comprises a neural network (44, 46) having at least three input paths (48, 50, 52), wherein the first imaging data set (22) is processed along a first input path (48), the second imaging data set (28) is processed along a second input path (50), and the third imaging data set (30) is processed along a third input path (52). [13] The method of claim 12, wherein the at least three input paths (48, 50, 52) are merged in a single merging layer (54). [14] The method of claim 12 or 13, wherein the neural network (44, 46) generates a defect indicator (34) for at least each of the first imaging data set (22), the second imaging data set (28) and the third imaging data set (30). [15] The method of claim 14, wherein the neural network (44, 46) comprises at least three output paths (58, 60, 62), each defect indicator (34) being generated by a different output path (58, 60, 62). [16] The method of claim 15, wherein the neural network (44, 46) comprises skip connections (56) each configured to connect a layer of an input path (48, 50, 52) to a layer of an output path (58, 60, 62). [17] Method according to one of claims 11 to 16, wherein the machine learning model is designed to process the input paths (48, 50, 52) asynchronously. [18] The method of any one of claims 11 to 17, wherein the first imaging data set (22), the second imaging data set (28) and the third imaging data set (30) are registered using a machine learning model. [19] The method of any preceding claim, wherein the first imaging data set (22), the second imaging data set (28) and the third imaging data set (30) are selected from the group consisting of acquired imaging data sets, design data sets and simulated data sets. [20] Method according to one of the preceding claims, wherein at least one imaging data set (22, 28, 30) comprises additional information about the probability that a defect (24) is present in the imaging data set (22, 28, 30). [21] A computer-implemented method for training a machine learning model for defect detection according to any one of claims 11 to 20, the method comprising: - providing training images of objects (72) comprising integrated circuit patterns, wherein the training images comprise at least triples of first imaging data sets (22), second imaging data sets (28) and third imaging data sets (30), and contain annotated defects (24); - Train the machine learning model using the provided training images by minimizing a loss function designed for defect detection. [22] A computer program comprising instructions which, when executed by a computer, cause the computer to carry out a method according to any one of claims 1 to 21. [23] A computer-readable medium having stored thereon a computer program executable by a computer device, the computer program comprising code for carrying out a method according to any one of claims 1 to 21. [24] A defect detection system (66) comprising: - an imaging device (70) configured to provide one or more imaging data sets of a portion of an object (72) comprising integrated circuit patterns; - one or more processing devices (74); - one or more machine-readable hardware storage devices (76) comprising instructions executable by one or more processing devices to apply a defect detection method according to any one of claims 1 to 20.

Images