Method for re-training a machine learning model for defect detection in an image of a photolithography mask and method for defect detection in an image of a photolithography mask using a pre-trained machine learning model

Abstract

The present invention relates to a method for retraining a pre-trained machine learning model (28) for defect detection in lithography images, the method comprising: acquiring adjusted images (52) of one or more lithography masks using an inspection system (50); adjusting at least one parameter of a simulation (44) of the inspection system (50) using the adjusted images (52); generating training images (33) of one or more lithography masks using the adjusted simulation (44) of the inspection system (50); retraining the machine learning model (28) using the training images (33) for defect detection; and applying the retrained machine learning model (28) for defect detection to an image (26) of a lithography mask (14) acquired by the inspection system (50).

Description

Technical Field

[0001] [Interactive reference to related applications]

[0002] This application claims priority to German Patent Application No. 102024120809.0, filed on July 22, 2024, which is hereby incorporated herein by reference in its entirety.

[0003] This invention relates to methods and systems for quality control and quality assurance in lithography masks, and more particularly to a machine learning model retraining method for defect detection in lithography mask images, along with a corresponding method and inspection system. These methods and systems can be used for quantitative measurement, process monitoring, defect detection, and defect re-inspection (review) in lithography masks. Prior Technology

[0004] Semiconductor manufacturing involves precise manipulation (such as etching) of materials like silicon or oxides at very fine scales in the nanometer (nm) range. Therefore, quality management procedures that include quality assurance and quality control are crucial to ensuring the high quality standards of the manufactured wafers. Quality assurance refers to a series of activities used to ensure high-quality products by preventing any defects that may occur during the development process. Quality control refers to the system for inspecting the final quality of the product. Quality control is part of the quality assurance process.

[0005] Wafers made of thin silicon sheets serve as substrates for microelectronic devices containing semiconductor structures built into and on the wafer. These semiconductor structures are constructed layer by layer using repetitive processing steps involving repeated chemical, mechanical, thermal, and optical procedures. The size, shape, and placement of these semiconductor structures and patterns are influenced by several factors. One of the most critical steps is the lithography process.

[0006] Lithography is a process used to generate patterns on a substrate. The patterns to be printed on the surface of the substrate are generated by computer-aided design (CAD). From this design, a lithography mask is generated for each layer, containing a magnified image of the computer-generated pattern to be etched into the substrate. The lithography mask can be further adjusted, for example, using optical proximity correction techniques. During the printing process, the illuminated image projected from the lithography mask is focused onto a photoresist film formed on the substrate. Semiconductor chips that power mobile phones or tablet computers, for example, contain approximately 80 to 120 patterned layers.

[0007] As the integrated density in the semiconductor industry continues to grow, lithography masks must image increasingly smaller structures onto wafers. The aspect ratios and number of layers in integrated circuits are constantly increasing, and these structures are increasingly becoming a third (vertical) dimension. Currently, the height of these memory stacks exceeds tens of micrometers. In contrast, feature sizes are becoming increasingly smaller. The smallest feature size, or critical dimension, is below 10 nm (e.g., 7 nm or 5 nm), and is increasingly approaching below 3 nm in the near future. While the complexity and size of these semiconductor structures are increasingly becoming this third dimension, the lateral dimensions of integrated semiconductor structures are becoming increasingly smaller. Generating these small structural sizes imaged onto the wafer requires lithography masks or templates for nanoimprint lithography with increasingly smaller structural or patterned elements. Therefore, the processes for generating lithography masks and templates for nanoimprint lithography are becoming increasingly complex, time-consuming, and ultimately more expensive. With the advent of EUV lithography scanners, the nature of photomasks changed from transmissive to reflective, defining patterns.

[0008] Due to the minute structural dimensions of the pattern elements in a lithography mask or template, errors cannot be eliminated during mask or template generation. The resulting defects may be due to, for example, degeneration of the lithography mask or particulate contamination. Of the various defects occurring during semiconductor structure manufacturing, lithography-related defects account for nearly half of the total. Therefore, in semiconductor process control, lithography mask inspection, re-inspection, and metrology play a crucial role in monitoring systemic defects. Defects detected during quality assurance procedures can be used, for example, for root cause analysis to modify or repair the lithography mask. These defects can also be used as feedback to improve process parameters in the manufacturing process, such as exposure time and focus variation.

[0009] Each defect in the lithography mask can lead to adverse behavior of the resulting wafer, or significant damage to the wafer system. Therefore, each defect must be identified and repaired if possible and necessary. Thus, reliable and rapid defect detection methods are crucial for lithography masks.

[0010] Machine learning methods can be used to analyze large amounts of data that require extensive processing. Machine learning is a field of artificial intelligence. Generally, machine learning methods build parameterized machine learning models based on training data containing a large number of samples. After training, the method can generalize the knowledge gained from the training data to new, unencountered samples, thereby making predictions on new data. Many machine learning methods exist, such as linear regression, k-means algorithm, support vector machines, decision trees, random forests, neural networks, and 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, these 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 complex representation, thus deriving low- and high-level knowledge from the training data. These hidden layers can have different sizes and tasks, such as convolutional or pooling layers.

[0012] Furthermore, defect detection methods are plagued by the problem of changing conditions. For example, the optical properties of a lithography mask change during its use in the lithography process, such as with EUV exposure tools. In addition, different lithography masks can vary not only in their design but also in their materials. Therefore, it is important to adapt existing defect detection methods to changes in conditions, such as for different inspection systems and / or different lithography masks and / or modified conditions in the inspection system and / or modifications to the optical properties of the lithography mask. For example, the reference image system used in defect detection methods must be adapted to changes in conditions to prevent a large number of false positive defect detections. Images from different machines or even different types of machines can vary significantly (e.g., in contrast or appearance), due to different materials or changes in the optical properties of the lithography mask. All these situations require the defect detection method to adapt quickly to changes in conditions. Moreover, machine learning models require large amounts of training data that are typically unavailable in these situations.

[0013] To mitigate this problem, domain adaptation methods have been proposed in the literature. These methods attempt to directly adapt the model to changes in conditions, or to adapt the model's training data to these changes. However, due to a lack of knowledge about the image formation process, many training images in the new domain still require training; otherwise, the adaptation cannot be robustly estimated.

[0014] Therefore, one objective of this invention is to adapt a trained machine learning model to modify conditions for defect detection in lithography images. In particular, one objective is to adapt this machine learning model to require only a small amount of acquired training data (reflecting the modified conditions), with minimal user effort and time required for image acquisition. Furthermore, the objective is to improve the accuracy of the predictions made by the retrained machine learning model.

[0015] These objectives are achieved by the invention as expressly described in the independent claims. Advantageous embodiments and further developments of the invention are clearly described in the appended claims. Summary of the Invention

[0016] Specific embodiments of the present invention relate to a method and system for retraining a pre-trained machine learning model for defect detection in lithography images.

[0017] Specific embodiments of the present invention relate to a method for retraining a pre-trained machine learning model for defect detection in lithography images. The method includes: a. acquiring adjusted images of one or more lithography masks using an inspection system configured to acquire images of the lithography masks; b. adjusting at least one parameter of a simulation of the inspection system using the adjusted images; c. generating training images of one or more lithography masks using the adjusted simulation of the inspection system; and d. retraining the machine learning model for defect detection using the generated training images, for use with the retrained machine learning model for defect detection in an image of a lithography mask acquired by the inspection system.

[0018] Specific embodiments of the present invention relate to a method for defect detection in lithography images using a pre-trained machine learning model. The method includes: a. acquiring adjusted images of one or more lithography masks using an inspection system configured to acquire images of the lithography masks; b. adjusting at least one parameter of a simulation of the inspection system using the adjusted images; c. generating training images of one or more lithography masks using the adjusted simulation of the inspection system; d. retraining the machine learning model for defect detection using the generated training images; and e. applying the retrained machine learning model for defect detection to an image of a lithography mask acquired by the inspection system.

[0019] By using the simulation of this inspection system, the number of training images required for the new domain is less compared to the domain adaptation method, because only a few parameters of the simulation need to be adapted. This is because the prior knowledge of the image formation process is already included in the simulation of the inspection system. The adapted simulation of the inspection system can then be used to automatically generate a large number of training images to train the defect detection algorithm without acquiring such a large number of training images.

[0020] At least one parameter of the inspection system is adjusted using adjusted images. For this purpose, adjusting the simulation of the inspection system requires only a small number of adjusted images. For example, the number of adjusted images is less than 1% of the number of pre-training images used to pre-train the machine learning model, preferably less than 0.1%, more preferably less than 0.01%, and most preferably less than 0.001%. After adjusting the simulation of the inspection system with these adjusted images, a large number of training images very similar to the images acquired by the inspection system can be simulated. Therefore, a large number of training images can be automatically generated on the inspection system with minimal user effort requiring only a small number of actually acquired images. For example, this process enables the development of a trained machine learning inspection system at a first site adapted to different conditions at a second site. The parameters of the simulated inspection system must be adjusted using only a small number of adjusted images, rather than retraining the inspection system at the second site solely based on the training images acquired from the second site (which is often impossible due to the limited amount of training images available there). The simulation system can then be used to generate realistic training images for retraining the machine learning model in the inspection system at the second site. In this way, retraining the machine learning model becomes feasible under modified conditions if only a few acquired training images (i.e., adjusted images) are available. Simultaneously, the user's workload for adjusting the machine learning model to these new conditions (such as at the second site) is minimized. Furthermore, the time required to adjust the machine learning model of the inspection system to these new conditions is significantly reduced using the adjusted simulation of the inspection system. In addition, the time for acquiring the required training images is reduced. Finally, the energy consumption of the inspection system, and the wear and tear on the inspection system and the one or more lithography masks are reduced.

[0021] The term "inspection system" refers to a system configured to inspect a photomask by acquiring an image of the photomask and detecting defects in that image. An inspection system may be, for example, a review system or a repair system.

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

[0023] The image of a lithography mask can refer to various types of images of the lithography mask, such as, for example, two-dimensional or volumetric three-dimensional images that can be processed piece by piece. The image can be acquired using an inspection system or can be simulated from the design of the lithography mask. The image can be various modalities acquired using a combination of X-ray imaging and SEM, such as structured electron microscopy (SEM) images, aerial images, optical images, X-ray images, computer tomography (CT) images, focused-ion beam (FIB) images, atomic force microscopy (AFM) images, ultrasound images, or multimodal images. The image of a lithography mask can image the entire lithography mask or one or more subsections thereof. Preferably, the image of a lithography mask refers to its aerial image.

[0024] The training image, pre-training image, and adjustment image each contain at least one image of a lithography mask. It may further contain defect annotations or a design of the lithography mask.

[0025] The term "defect" refers to the deviation of an integrated circuit pattern from its a priori defined norm. For example, a defect in an integrated circuit pattern (such as a semiconductor structure) can lead to the failure of associated semiconductor devices. Depending on, for example, the detected defect, the lithography process can be improved, or the lithography mask or wafer can be repaired or discarded. The norm for the structure or pattern can be defined by one or more corresponding reference lithography masks or reference datasets, such as design datasets, simulation datasets, or defect-free datasets.

[0026] Machine learning models used for defect detection can perform various tasks, such as defect detection (the presence or absence of a defect), defect localization (locating a defect), defect segmentation (calculating the area, volume, or contour of a defect), and defect classification (assigning a defect category to a defect). These machine learning models can be supervised, unsupervised, or semi-supervised. They can use a reference dataset (i.e., images primarily obtained without defects or simulated images) to detect defects, such as in die-to-die models; or they can be referenceless machine learning models that detect defects by detecting deviations from a benchmark (such as from prior knowledge or knowledge derived from the grain itself), such as in single-grain models. A pre-trained machine learning model refers to a machine learning model that has been trained at least once, meaning its parameters have been tuned at least once using training images. A retrained machine learning model refers to the training of a pre-trained machine learning model, i.e., subsequent tuning of the model's parameters using more training images.

[0027] The simulation of the inspection system refers to a parametric model of the inspection system that simulates the image acquisition procedure of the inspection system. The simulation of the inspection system can be used to generate a simulated image of the lithography mask. The parameters of the parametric model of the inspection system (i.e., the simulation parameters) can be adjusted using such adjustment images to simulate a specified inspection system or specified conditions of the inspection system or the lithography mask. These simulation parameters may include, for example: - Image properties, such as image intensity, image contrast, noise level, image distortion, maximum illumination difference, maximum focus drift, and maximum misalignment; - Defect information; - Machine settings, such as light source intensity and light source parameters; - Properties of the photomask, such as the thickness of one layer of the photomask and the material of the photomask layers; - Design modification parameters, such as edge roughness, critical dimension (CD) variation, pattern thickness, and optical proximity correction (OPC) structure; - Parameters of a machine learning model, etc.

[0028] Aerial imagery indicates the radiation intensity distribution of a lithography system within a wafer plane used for a given lithography mask. Therefore, when the lithography mask is printed on the wafer using this lithography system, the aerial imagery simulates the structures on that surface of the wafer. The wafer plane refers to the plane within the photoresist on the wafer surface in the lithography system. Aerial imagery can be generated by applying an aerial imagery measurement or metrology system to the lithography mask. Aerial imagery can be simulated using lithography mask design and aerial imagery simulation methods.

[0029] Aerial imagery can refer to the aerial imagery of the entire photomask, or it can refer to the aerial imagery of a section of the photomask. Design can refer to the design of the entire photomask, or it can refer to the design of a section of the photomask.

[0030] In the case of aerial imagery, the simulation of the inspection system can simulate the generation of aerial imagery from the design of the lithography mask. The simulation can use physical models, such as the lithography mask and / or a physical model of the electromagnetic waves propagating through it. The simulation can also use non-physical models, such as machine learning models trained to generate aerial imagery from the design using training data. A hybrid approach using physical models and machine learning models can also be used to simulate aerial imagery.

[0031] In this example, the method further includes, prior to step a., pre-training the machine learning model using pre-training images comprising at least one simulated image generated by the simulation of the inspection system. The simulation of the inspection system may, for example, use initial simulation parameters. In this way, the same simulation of the inspection system can be used to generate such pre-training images for pre-training as used in step c., thereby significantly reducing the user workload for providing such required pre-training images, reducing the time for acquiring such pre-training images, reducing the energy consumption of the inspection system, and reducing wear and tear on the inspection system and the lithography mask.

[0032] Alternatively, the method may further include, prior to step a, pre-training the machine learning model using pre-training images comprising at least one image acquired by an inspection system. In this way, more realistic training images are used for pre-training, thereby resulting in higher accuracy of the predictions made by the machine learning model.

[0033] In a preferred example, at least some training images contain defect annotations. Defect annotations may be given by, for example, yes / no indications, bounding boxes of any shape and size covering the defect, pixel-wise or voxel-wise segmentation of the defect, descriptive content, or one or more items from a defect list.

[0034] These defect annotations can be used for supervised training of machine learning models, thereby improving the accuracy of the detected defects. Even if only a few defect annotations are available, they can still be used to improve the results of unsupervised training of the machine learning model.

[0035] According to the present invention, the simulation of the inspection system includes, for example, simulating the image acquisition procedure for a lithography mask using a design employing such a lithography mask. In this way, a highly accurate simulation of the image can be generated from the design of the lithography mask.

[0036] In this example, the training images generated for training the machine learning model for defect detection include simulated images of defective designs (i.e., designs containing one or more defects) and corresponding defect annotations. These simulated images of defective designs are obtained by applying the simulation of the inspection system to the defective designs. The defective designs can be obtained, for example, by modifying defect-free designs. These modifications correspond to atypical design structures, i.e., simulated defects. In this way, the simulated defects can be controlled in terms of their location, size, intensity, type, and frequency. Using these simulated images of defective designs and corresponding defect annotations, the machine learning model for defect detection can be trained in a supervised manner to achieve high-accuracy defect detection. Depending on the needs, the defect-free design or its simulated image, corresponding to the defective design, can be used as additional input to the machine learning model for defect detection.

[0037] According to the example, adjusting at least one parameter of the simulation of the inspection system in step b. involves solving an optimization problem. In this way, the accuracy of the adjusted parameters can be improved.

[0038] An optimization problem involves maximizing or minimizing an objective function. This optimization problem may also include constraints. Solving the optimization problem means applying a mathematical procedure to compute a point with an objective function value that is superior to that of multiple other points. Solving the optimization problem may, for example, mean compute the global optimum or a local optimum of the objective function. The mathematical procedure may involve compute an analytical solution or apply an iterative method, such as gradient descent, a simplex method, a variational approach, or a combinatorial optimization approach.

[0039] Based on this example, for at least one adjusted image, a corresponding lithography mask design can be provided. The simulation of the inspection system uses this lithography mask design to simulate the image acquisition procedure for the lithography mask, and solving the optimization problem involves minimizing the deviation between one or more adjusted images and one of the simulated images of the corresponding design. Therefore, the accuracy of the adjusted one or more parameters can be improved.

[0040] According to another aspect of this paradigm, solving the optimization problem involves maximizing the similarity between the distribution of the adjusted image and the distribution of one of the simulated images obtained using the simulation of the inspection system. Therefore, the accuracy of these adjusted one or more parameters can be improved.

[0041] Based on another aspect of this paradigm, the gradient of the simulation of the inspection system with respect to at least one parameter is derived, and solving the optimization problem involves using a gradient descent method (using the derived gradients). Therefore, the accuracy of the adjusted one or more parameters can be improved.

[0042] In the example, prior knowledge is used in step b. to adjust at least one parameter of the simulation of the inspection system. Prior knowledge may include, for example, known parameters such as a photomask thickness or an illumination setting parameter, parameter range, or parameter distribution. Such prior knowledge can be used to simplify the optimization problem, thereby enabling a solution to be obtained with fewer computations, or to improve the solution to the optimization problem to increase the accuracy of the adjusted at least one parameter.

[0043] According to the example, most of the adjusted images differ from the pre-trained images used to pre-train the machine learning model in at least one sample from the group including the inspection system, the image acquisition time, and the lithography mask. At least some of these pre-trained images used to pre-train the machine learning model can be acquired using the inspection system, or can be simulated using a simulation of the inspection system. Differences may lie, for example, in the type of inspection system, the instance of the inspection system, the parameters of the inspection system, the acquisition time, the design, material, or instance of the lithography mask, etc. In this way, the pre-trained machine learning model can be adapted to the modified conditions.

[0044] In this example, the pre-training of the machine learning model is performed on a first computer system, and at least step d. is performed on a second computer system. In this way, the machine learning model can be adapted to the modified conditions between the first and second computer systems. These computer systems may, for example, be located in different locations or have different characteristics, such as different hardware configurations, different security standards, different computation time requirements, etc. The first and second computer systems may also belong to different inspection systems. In this case, the machine learning model can be adapted to the modified conditions between these inspection systems.

[0045] According to the present invention, steps a. to d. are iterative. In this way, the retrained machine learning model can be retrained again to adapt to more modified conditions, such as when different lithography masks are loaded into the inspection system.

[0046] According to the example, the simulation of the inspection system includes a physical simulation of the transmission of electromagnetic waves within the photomask, such as using rigorous simulation or Kirchhoff's method or other simulation techniques. In this way, the image acquisition procedure can be simulated with high accuracy or low computation time, thereby improving the accuracy or computation time of the inspection system's simulation.

[0047] Alternatively, the simulation of the inspection system may include the application using a trained machine learning model. In this way, the accuracy and / or computation time of the simulation of the inspection system can be improved.

[0048] An inspection system for detecting defects in a lithography mask image according to a third embodiment of the present invention includes: an image acquisition unit configured to acquire an image of a lithography mask; a data analysis device including at least one memory; and at least one processor configured to perform the steps of a method for detecting defects in an image of a lithography mask according to a third embodiment of the present invention.

[0049] A machine learning model retraining system for defect detection in a lithography mask image according to a fourth embodiment of the present invention includes: an image acquisition unit configured to acquire an image of a lithography mask; a data analysis device including at least one memory; and at least one processor configured to perform the steps of a machine learning model retraining method for defect detection in an image of a lithography mask according to a fourth embodiment of the present invention.

[0050] The invention illustrated by the specific embodiments, examples, and embodiments is not limited to these specific embodiments, examples, and embodiments, but can be practiced by those skilled in the art through various combinations or modifications thereof. Simple Explanation of the Diagram

[0051] Figure 1 illustrates an exemplary transmission lithography system, such as a deep ultraviolet (DUV) lithography system; Figure 2 illustrates an exemplary reflective lithography system, such as an extreme ultraviolet (EUV) lithography system; Figure 3 shows an imaging dataset of an object containing an integrated circuit pattern in the form of a photomask containing a defect; Figure 4 shows a flowchart illustrating the steps of a method according to a specific embodiment of the present invention; Figure 5 illustrates this application of a trained machine learning model for defect detection; Figure 6 illustrates the training of a machine learning model that simulates the image acquisition procedure in an inspection system; Figures 7a and 7c illustrate the retraining method for a machine learning model used for defect detection according to a specific embodiment of the present invention; and Figure 8 illustrates an inspection system for detecting defects in a photomask according to a third specific embodiment of the present invention. Implementation

[0052] In the following description, advantageous exemplary embodiments of the invention are illustrated and schematically shown in the figures. Throughout these figures and descriptions, the same reference numerals are used to describe the same features or parts. Dashed lines indicate features as needed.

[0053] The methods and systems described herein can employ various lithography systems, such as a transmissive lithography system 10 or a reflective lithography system 10'.

[0054] Figure 1 illustrates an exemplary transmissive lithography system 10, such as a DUV lithography system. The main components are a radiation source 12 (which may be a deep ultraviolet (DUV) excimer laser source), imaging optics (which, for example, define the coherence of the portion and may include optics that shape the radiation from the radiation source 12), a lithography mask 14, an illumination optics 16 (which illuminates the lithography mask 14), and a projection optics 17 (which projects an image of the lithography mask model 92 (such as the design pattern) onto a wafer plane 18). At the pupil plane of the projection optics 17, an adjustable filter or aperture can limit the range of beam angles illuminating the wafer plane 18. The maximum possible angle is defined as the numerical aperture NA of the projection optics NA = n sin(Gmax), where n is the refractive index of the medium between the substrate and the last element of the projection optics 17, and Gmax is the maximum angle at which the beam emitted from the projection optics 17 can still illuminate the wafer plane 18.

[0055] In this invention document, the terms “radiation” or “beam” are used to encompass all types of electromagnetic radiation, including ultraviolet radiation (such as those with wavelengths of 365, 248, 193, 157, or 126 nm) and EUV (extreme ultraviolet radiation, such as those with wavelengths in the range of about 3-100 nm).

[0056] The illumination optics 16 may include optical components for shaping, adjusting, and / or projecting radiation from the radiation source 12 before the radiation passes through the lithography mask 14. The projection optics 17 may include optical components for shaping, adjusting, and / or projecting the radiation after it has passed through the lithography mask 14. The illumination optics 16 do not include the light source 12, and the projection optics do not include the lithography mask 14.

[0057] The illumination optics 16 and projection optics 17 may comprise various types of optical systems, including, for example, refractive optics, reflective optics, aperture optics, and catadioptric optics. The illumination optics 16 and projection optics 17 may also include components that operate according to any of these design types for guiding, shaping, or controlling the radiation of the projection beam collectively or individually.

[0058] Figure 2 illustrates an exemplary reflective lithography system 10', such as an extreme ultraviolet (EUV) lithography system. The main components are a radiation source 12 (which may be a laser plasma source), an illumination optics 16 (which, for example, defines the partial coherence and may include optics that shape the radiation from the radiation source 12), a lithography mask 14, and a projection optics 17 (which projects an image of the lithography mask model 92 (the design pattern) onto a wafer plane 18). At the pupil plane of the projection optics 17, an adjustable filter or aperture can limit the range of beam angles illuminating the wafer plane 18, where the maximum possible angle defines the numerical aperture NA of the projection optics as NA = n sin(Gmax), where n is the refractive index of the medium between the substrate and the last element of the projection optics 17, and Gmax is the maximum angle of the beam emitted from the projection optics 17 that can still illuminate the wafer plane 18.

[0059] Figure 3 illustrates an image 20 of a lithography mask 14 containing defect 22. Various imaging modalities can be used to acquire this image according to the techniques described herein. The image may contain single-channel or multi-channel images, such as focus stacking. For example, the image may include a two-dimensional image. A multi-beam scanning electron microscope (mSEM) may be used. mSEM uses multiple beams to simultaneously acquire images in multiple fields of view. For example, at least 50 beams or even at least 90 beams may be used. Each beam covers a separate portion of the surface of the lithography mask. Thus, large images are acquired in a short time. Typically, modern machines acquire 4.5 billion pixels per second. Other examples of images including two-dimensional images relate to imaging modalities such as optical imaging, phase-contrast imaging, X-ray imaging, etc. The image may also be a volumetric three-dimensional dataset that can be processed slice by slice or as a three-dimensional volume. Here, a crossbeam imaging system including a focused ion beam (FIB) source, an atomic force microscope (AFM), or a scanning electron microscope (SEM) may be used. Furthermore, magnetic resonance (MR) imaging, ultrasound imaging, or computed tomography (CT) imaging may be used. Multimodal imaging, such as a combination of X-ray imaging and SEM, can be used. The image can be an aerial image acquired by an aerial image measurement system. The aerial image shows the radiation intensity distribution at the substrate level. It can be used to simulate the radiation intensity distribution generated by the lithography mask during the lithography process. The aerial image measurement system may be equipped with, for example, a staring array sensor, a line scan sensor, or a time-delayed integration (TDI) sensor.

[0060] For defect detection, machine learning models are a popular choice for achieving high-quality results with short computation times. Machine learning models are trained using training data (i.e., examples) and thus independently derive their knowledge from that data, without requiring users to define defect detection rules. In this way, optimal defect detection results can be automatically obtained in a data-driven manner. Therefore, using machine learning models improves the recall and precision of defect detection methods while reducing the user's workload.

[0061] However, machine learning models (especially deep learning models) require a large amount of training data that may not be readily available. Furthermore, training a machine learning model can take days or weeks, depending on its complexity. Moreover, the model can only apply the knowledge derived from the training data. If conditions change, the model must be retrained.

[0062] For example, if the optical properties of the lithography masks change during the lithography process (e.g., by EUV exposure tools), the conditions may change. If the lithography mask is replaced, the conditions may also change, as different lithography masks can vary not only in their design but also in their materials. If the defect detection method is performed on different inspection systems (e.g., on the same type of inspection system or on different types of inspection systems), the conditions also change. In all these cases, rapid retraining of the machine learning model used for defect detection is required. However, training data is often unavailable or scarce when conditions change.

[0063] To enable retraining of the machine learning model, Figure 4 shows a flowchart of a pre-trained machine learning model retraining method 24 for defect detection in lithography images according to a specific embodiment of the present invention. The retraining method for the pre-trained machine learning model for defect detection in lithography images includes: a. acquiring adjusted images of one or more lithography masks using an inspection system, wherein the inspection system is configured to acquire images of the lithography masks in step M1; b. adjusting at least one parameter of a simulation of the inspection system using the adjusted images in step M2; c. generating training images of one or more lithography masks using the adjusted simulation of the inspection system in step M3; and d. retraining the machine learning model for defect detection using the generated training images in step M4. The retrained machine learning model can be used for defect detection in the lithography images acquired by the inspection system in step M6.

[0064] As illustrated in Figure 5, the trained machine learning model 28 for defect detection takes image 26 (in this case, aerial image) as input and maps image 26 to zero, one, or more defects 22 or their representations. The supervised machine learning model for defect segmentation can use known machine learning segmentation architectures such as U-Net or Segformer, as described in the paper "SegFormer: Simple and Efficient Design for Semantic Segmentation with Transformers," Enze Xie, Wenhai Wang, Zhiding Yu, Anima Anandkumar, Jose M. Alvarez, Ping Luo, 2021 arXiv 2105.15203. The supervised machine learning model for defect localization can use known machine learning object detection architectures such as CenterNet or YOLO. CenterNet is a machine learning-based object detector based on keypoint triplets, where two keypoints represent opposite corners of a bounding box. If additional keypoints of the same category are found in the central region of the bounding box, the bounding box proposal is retained. YOLO is a very fast machine learning-based object detector that uses fully convolutional neural networks for bounding box prediction. The image is subdivided into grid cells, and for each grid cell, a specified number of bounding boxes larger than that grid cell can be predicted. Bounding boxes are retained based on their category probabilities and bounding box confidences. By using a bounding box-type object detector, the area containing a defect can be distinguished from the area without defects. For defect detection, reference-based registration and detection machine learning architectures, such as those disclosed in German patent application 10 2023 104 378.1, can be used, or reference-based grain-to-database machine learning architectures known to be designed into image simulation algorithms can be used. Furthermore, self-supervised or referenceless anomaly detection machine learning models that do not require annotated training of images can also be used for defect detection, such as autoencoders.

[0065] In step M0, which precedes step M1 as needed, the method may further include pre-training the machine learning model using pre-trained images. These pre-trained images may include at least one image acquired by an inspection system. This at least one image may be acquired using the same inspection system used to acquire the adjusted images in step M1, or it may be acquired using a different inspection system or even a different type of inspection system. For example, the machine learning model may be pre-trained using images acquired by an inspection system at a first site, and retrained using images acquired by a different inspection system at a second site. Even if the inspection systems are of the same type, the parameters or conditions at the two sites may still be different, thus requiring retraining of the machine learning model.

[0066] Alternatively or additionally, the pre-trained images 31 may include at least one simulated image generated using the simulation of the inspection system. The simulation of the inspection system may use initial simulation parameters to generate the pre-trained images. These initial simulation parameters may later be used in step M2 to adapt to new conditions using the adjustment images.

[0067] In step M5, which may be necessary, after step d, the final retraining of the retrained machine learning model can be performed using a small number of training images acquired by the inspection system (preferably including defect annotations). This process is particularly advantageous if the simulation of the inspection system is not perfectly realistic, or if the adjusted simulation parameters are not perfectly fitted and the simulated images are therefore not perfectly realistic. In these cases, final retraining using the training images acquired by the inspection system will reduce the potential drawbacks of the retrained machine learning model due to suboptimal simulated training images. Therefore, the accuracy of the predictions made by the retrained machine learning model can be improved.

[0068] According to the present invention, the simulation 44 of the inspection system 50 includes simulating an image acquisition process for a lithography mask 14, particularly by using a design 32 of the lithography mask 14. Design 32 is used to generate an image 26, such as an aerial image, of the corresponding lithography mask. The simulation of the image acquisition process for the lithography mask may, for example, include an optically inspired simulation based on a physical model, such as a rigorous simulation or a Kirchhoff simulation. For these simulations, the simulation parameters may include optical parameters, such as a light source intensity or a mask thickness. Alternatively, the simulation of the image acquisition process may include using a data-driven model, such as a machine learning model trained to predict an image from a design of a lithography mask. In this case, the simulation parameters may include the trainable parameters of the machine learning model. Alternatively, a hybrid model combining a physical model and a machine learning model with optically inspired simulation may be used. In this case, the simulation parameters may include a combination of optical parameters and machine learning parameters.

[0069] Aerial imagery can be simulated, for example, using a solid model to generate aerial imagery from a design. This leads to accurate results, but is often time-consuming. Among these methods are rigorous simulation methods known to those skilled in the art, such as the Finite-Difference Time Domain (FDTD) method or Rigorous Coupled Wave Analysis (RWCA). Since these methods require long computation times, fast approximation methods such as the Thin Element Approximation (TEA) can be used. The TEA assumes that the thickness of the structures on the lithography mask is small compared to the wavelength, while the width of the structures on the lithography mask is large compared to the wavelength. However, as lithography processes use increasingly shorter wavelengths of radiation, and the structures on the patterning device become smaller and extend to the vertical dimension, these assumptions no longer hold, and the three-dimensional effects of the lithography mask must be taken into account. Therefore, the results obtained by the TEA method are less accurate, but much faster than rigorous simulation results.

[0070] The simulation 44 of the inspection system 50 may also include a machine learning model for simulating the image acquisition procedure. Figure 6 shows a machine learning model 38 trained to simulate the image acquisition procedure, so as to generate a simulated pre-trained image 31 in step M0 or a training image 33 in step M3. The training images 30 for the machine learning model used to simulate the image acquisition procedure include a design 32 and corresponding images 26, such as aerial images. These corresponding images 26 are preferably acquired using some image acquisition system. During training, an objective function is minimized. This objective function may, for example, minimize the deviation between the images generated by the machine learning model and the corresponding images 26.

[0071] In another example, a physical model can be used to simulate the image acquisition process, such as the transmission of incident electromagnetic waves through lithography masks in the case of aerial images. Machine learning models can then be used to refine these results, depending on the needs. For instance, to obtain realistic simulated images, noise, focal length variations, or deviations from ideal structures (such as line edge smoothness, structural thickness variations, or corner rounding) can be applied to these simulated images.

[0072] Preferably, at least some of the training images 30 are simulated images of designs containing uncommon structures. Uncommon structures are typically structures that are not part of the design, such as defects, design deviations that do not need to be classified as minor variations in the design structure (e.g., variations in structural thickness, corner rounding, etc.), or types of structures not included in the training images. For example, if designs containing only lines and spaces are included in the training images, uncommon structures may include holes, intersections, auxiliary features, complex polygons, etc. For instance, the design 32 on the far right of Figure 6 contains a defect 22 that protrudes or extends inward. In this way, the simulation 44 of the inspection system 50 can simulate the image acquisition process for all types of designs (including designs containing uncommon structures).

[0073] Using the simulation 44 of the inspection system 50, the training images for the machine learning model used for defect detection can then be quickly and easily simulated using the simulated image acquisition procedure (such as using an entity model and / or a machine learning model). Artificial defects 22 can be added to design 32 to obtain a defective design 34. From these artificial defects 22, defect annotations 36 can be easily generated together with these training images. These can be used to effectively retrain the machine learning model 28 for defect detection in a supervised manner. In addition to defects 22, more uncommon structures can also be added to these designs 32 to generate training images. In this way, the machine learning model 28 for defect detection learns to distinguish between defects 22 and uncommon structures that do not meet the conditions of defects 22.

[0074] Figures 7a to 7c illustrate the effective retraining method of machine learning model 28 for defect detection in lithography images according to the present invention.

[0075] Figure 7a illustrates the pre-training phase 40 of the optional step M0 for pre-training the machine learning model 28 using the pre-trained image 31. The simulation 44 of the inspection system 50 includes a simulation of the image acquisition procedure given a design 32 of a lithography mask, such as a physical model or a machine learning model 38 illustrated in Figure 6.

[0076] To generate these pre-trained images 31, it is preferable that the simulation 44 of the inspection system 50 uses initial simulation parameters 42, such as the trained machine learning model 38 using initial parameters or the physical model using initial parameters as shown in FIG6. These initial simulation parameters can be selected in various ways. For example, parameters of the inspection system 50 with high probability can be used, or the average parameters of the inspection system 50 can be used, or the expected parameters of the inspection system 50 can be used, or random parameters drawn from a pre-specified range of these parameters of the inspection system can be used. If the possible or probable modifications of these conditions are known, such as the type of the inspection system or a common type of the inspection system, the lighting conditions or typical lighting conditions at the second site, the type of lithography mask used at the second site or a typical type of lithography mask, etc., these can be taken into consideration when selecting the initial simulation parameters 42 to generate the pre-trained images 31 for training the simulation 44 of the inspection system 50. Different sets of initial simulation parameters 42 can also be used to generate a variety of pre-trained images 31 for pre-training the machine learning model 28. In this way, the machine learning model can be tightly adapted to modified conditions.

[0077] Alternatively or additionally, the images acquired by inspection system 50 can be used as pre-training images 31 for pre-training machine learning model 28. Several options are conceivable here. For example, at least some of the pre-training images 31 and at least some of the adjusted images 52 can be acquired using different inspection systems 50 of the same type, but with slightly different performance due to variations within predefined tolerances of the inspection systems 50. In another example, at least some of the pre-training images 31 and at least some of the adjusted images 52 can be acquired using different types of inspection systems 50, resulting in images of different modalities, such as aerial images and SEM images. Even though these images are of different modalities, they can still contain information about the same lithography mask. In yet another example, at least some of the pre-training images 31 and at least some of the adjusted images 52 can be acquired by the same inspection system 50, but used for different lithography masks, such as those using different materials or designs. In another example, at least some of the pre-trained images 31 and at least some of the adjusted images 52 are acquired by the same inspection system 50 for the same lithography mask, but at different times, such as at least some of the pre-trained images 31 being acquired days, weeks, months, or even years before at least some of the adjusted images 52, or vice versa. In this way, the acquired images can be used in addition to or in place of the simulated pre-trained images 31 used to pre-train the machine learning model.

[0078] Preferably, at least some of these designs are defective designs 34 including one or more defects 22. Simulated images of these defective designs 34 can be used to pre-train a machine learning model 28 for defect detection. To obtain a defective design 34, a defect-free design 32 can be modified by adding one or more artificial defects 22. Alternatively, a design including a defect can be used. Using the simulation 44 of the inspection system 50, for example with initial simulation parameters 42, images of the defective design 34 can be simulated as pre-trained images 31 for pre-training a machine learning model 28 for defect detection. Preferably, at least some of these simulated pre-trained images 31 include defect annotations 36. When the defects 22 are artificially generated, the corresponding defect annotations 36 can be easily included in the pre-trained images 31. For example, the modification of the defect-free design 32 can be used to obtain defect annotations 36 for the simulated image of the defective design 34. Alternatively, the difference between the simulated image of the defect-free design and the simulated image of the defective design 34 can be used to obtain defect annotations 36. Defect annotations 36 indicate the presence, location, degree, and / or type of a defect, such as in the form of bounding boxes, pixel annotations, defect categories, etc. The pre-trained images 31 may also include acquired images with or without defect annotations 36. The pre-trained images 31 may also include acquired or simulated images without defects.

[0079] The pre-trained images 31 are used to pre-train the machine learning model 28 for defect detection in the pre-training step 46, thereby deriving the pre-trained machine learning model 47. The defect-free design 32 corresponding to the defective design 34 can be used as additional input to the machine learning model 28 for defect detection, as needed.

[0080] Figure 7b illustrates an adjustment stage 48 for adjusting the parameters of the simulation 44 of the inspection system 50 under modified conditions (e.g., at different sites) when using different inspection systems 50 or different types of inspection systems 50, different machine settings (e.g., lighting settings), different lithography masks, different materials, different designs 32, and different image properties (e.g., noise, distortion, intensity, brightness, contrast). For this purpose, the adjusted image 52 is acquired using the inspection system 50. When no retraining of the pre-trained machine learning model 47 is required at this stage, and only the adjustment of the simulation parameters of the inspection system 50 is performed, a small amount of adjusted image 52 is acquired, such as less than 1% of the number of pre-trained images used to pre-train the machine learning model 28 or the number of training images used to retrain the machine learning model 28, preferably less than 0.1%, more preferably less than 0.01%, and most preferably less than 0.001%. Preferably, at least some of the adjusted images 52 are acquired using a lithography mask having the same design 32 as the underlying design 32 of the pre-trained images 31 used for pre-training the machine learning model 28. In this way, adjusting the parameters is simplified and yields more accurate results. The simulation parameters of the simulation 44 of the checking system 50 are then adjusted in the parameter adjustment step 54 on the acquired adjusted images 52 to obtain the adjusted simulation parameters 56.

[0081] The parameter adjustment step 54 preferably includes solving an optimization problem. If such adjustments are made to the image... Including the design of such corresponding photomasks Then, the parameters of the simulation T of the inspection system are used to adjust these parameters. For example, parameter adjustment step 54 can be performed by solving the following optimization problem that minimizes a loss function L: The loss function measures the adjusted image. For the corresponding design The simulated image The difference between them is considered, and one or more simulation parameters θ are given. Therefore, these simulation parameters 42 are adjusted so that the simulated images from these designs match the adjusted images as well as possible. The loss function may contain more terms, such as a regularization term. Or, especially if for the adjusted images... Corresponding design If this cannot be provided, then the loss function can maximize the adjusted image. The distribution of these simulated images can be compared to the distribution of other simulated images for design purposes. This can be achieved by minimizing the Kullback-Leibler divergence or other stochastic measures, or by using comparative image statistics such as signal-to-noise ratio, contrast ratio, blurriness, and edge strengths. Alternatively, classification methods such as discriminators used in Generative Adversarial Networks (GANs) can be used to learn and distinguish between different simulated images. With simulated images The parameters θ of the analog 44 of the inspection system 50 can be modified until the discriminator can no longer distinguish the adjusted images. With these simulated images That is, regarding the two sets of images and The confusion of the classifier is maximized. The discriminator can be trained on patches of these images.

[0082] If the simulation 44 of the inspection system 50 enables the calculation of gradients for the simulation parameters, then the simulation parameters can be adjusted by solving the optimization problem using the gradient descent method.

[0083] To simplify the optimization problem, prior knowledge can be used. For example, not all simulation parameters need to be estimated, as some may be known beforehand, such as lighting parameters and photomask materials. These parameters may be loaded from a database or specified by the user. Furthermore, for many simulation parameters, a range of parameter values ​​or a probability distribution of those values ​​can be specified. The range of parameter values ​​can be used in the optimization problem to constrain those values ​​of the simulation parameters using constraints. The probability distribution of the parameter values ​​may be included in the objective function of the optimization problem. This probability distribution can be maximized, or its negative logarithmic probability can be minimized. Here, Indicates the parameterized loss function, Indicates a parameterized objective function, while Indicates the parameter maximized by the probability of the negative logarithm. The probability distribution.

[0084] Figure 7c illustrates the retraining of the pre-trained machine learning model 47 in the retraining phase 58. Using the adjusted simulation parameters 56, the simulation 44 of the inspection system 50 is applied to generate a large number of training images 33 from the design 32 of the lithography mask. These training images 33 are used to retrain the pre-trained machine learning model 47 for defect detection. The pre-trained images 31 and the training images 33 preferably rely at least partially on the same design. However, they may also rely on different designs. When simulating the training images 33, they can be automatically generated with a low number of computations and require less user effort. The generated training images 33 can be targeted at a specific defect 22 or design type. In this way, it is possible to retrain the machine learning model for a specific defect type or design type. Alternatively, the generated training images 33 can systematically cover a series of defect types or design types, thereby enabling efficient and rapid retraining of the machine learning model 28.

[0085] In step c., at least some of the generated training images 33 are simulated images containing defective designs 34 and corresponding defect annotations 36, which are obtained by applying the simulation 44 of the inspection system 50 to the defective designs 34. To obtain the defective design 34, the defect-free design 32 can be modified by adding one or more artificial defects 22. Alternatively, a design including a defect can be used. After adjusting the parameters, the image of the defective design 34 can be simulated using the simulation 44 of the inspection system 50. Preferably, at least some of the simulated training images 33 contain defect annotations 36. When the defects 22 are artificially generated, the corresponding defect annotations 36 can be easily included in the training images 33. For example, the modification example of the defect-free design 32 can be used to obtain the defect annotation 36 for the simulated image of the defective design 34. Alternatively, the difference between the simulated image of the defect-free design and the simulated image of the defective design 34 can be used to obtain the defect annotation 36. Defect annotation 36 indicates the presence, location, degree, and / or type of a defect, such as in the form of a bounding box, pixel annotation, defect category, etc. The training images 33 may also include acquired images with or without defect annotation 36. The training images 33 may also include acquired or simulated images without defects.

[0086] The generation of pre-trained images and the pre-training and retraining of the machine learning model can be performed, for example, on a computer, on a cluster, on a distributed system, or in the cloud. For this purpose, the use of powerful hardware, such as graphics processing units (GPUs) or tensor processing units (TPUs), is advantageous for accelerating these algorithms. Ample random access memory (RAM) and storage with fast input / output (I / O) also help reduce computation time. This hardware system does not require a physical connection to the inspection system, which is advantageous.

[0087] The pre-training system for the machine learning model can be executed in an environment drastically different from the environment used to retrain the machine learning model. For example, the pre-training system can be executed in a development environment at a first site using a large computer cluster, dedicated hardware, or even a cloud instance. In contrast, the retraining system can be executed in an application environment, such as at a second site with less hardware or where confidentiality standards need to be specified (e.g., at a customer's site). Using the method according to the invention, the pre-training system for the machine learning model can be meticulously executed in the development environment using a large number of pre-training images, while the retraining system can be executed in the application environment using only a small number of acquired adjustment images, simulated training images, and (as needed) a small number of acquired training images.

[0088] The above-described method for retraining a machine learning model for defect detection under modified conditions is iterative. The machine learning model needs to be retrained each time conditions are modified, for example, each time a different lithography mask is loaded by the inspection system. For this purpose, steps a through d can be repeated. Using a retrained machine learning model as a pre-trained machine learning model to incorporate more modified conditions is advantageous because it requires less tuning than starting from scratch or retraining the machine learning model from a pre-trained model. For example, if the machine learning model is pre-trained at a first site (e.g., a production site) and transported to a second site (e.g., a consumer), it may require less tuning if the transported pre-trained machine learning model from the first site is used for retraining, rather than if the already retrained machine learning model is used for further retraining.

[0089] An inspection system 50 for detecting defects 22 in an image 26 of a photomask 14 according to a third embodiment of the present invention is illustrated in FIG8. The inspection system 50 includes: an image acquisition unit 60 configured to acquire an image 26 of the photomask 14; a data analysis device 62 including at least one memory 64; and at least one processor 66 configured to perform the steps of a method according to one embodiment of the present invention.

[0090] Image acquisition unit 60 provides image 26 to data analysis device 62. Processor 66 may be implemented as a central processing unit (CPU), GPU, or TPU. Processor 66 may receive image 26 through interface 68. Processor 66 may load code from memory 64, such as code for executing a method for detecting defect 22 according to a specific embodiment of the present invention as described above. Processor 66 may execute the code.

[0091] A machine learning model retraining system for defect detection in a lithography mask image according to a fourth embodiment of the present invention includes: an image acquisition unit configured to acquire an image of a lithography mask; and a data analysis device including at least one memory; and at least one processor configured to perform the steps of a machine learning model retraining method for defect detection in an image of a lithography mask according to an embodiment of the present invention. The image acquisition unit can be used to acquire adjustment images and (as needed) training images. The data analysis device obtains a pre-trained machine learning model that has been retrained as described above.

[0092] Throughout this specification, references to "specific embodiment," "example," or "style" indicate that a particular feature, structure, or characteristic described in relation to that specific embodiment, example, or style is included in at least one specific embodiment, example, or style. Therefore, the use of the phrases "according to a specific embodiment," "according to an example," or "according to a style" throughout this specification does not necessarily refer to the same specific embodiment, example, or style, but may do so. Furthermore, in one or more embodiments, such specific features or characteristics can be combined in any suitable manner, as will be apparent to those skilled in the art from this disclosure.

[0093] Furthermore, although some specific embodiments, examples, or patterns described herein include, but are not included in, other features in combinations of features of different embodiments, examples, or patterns, they are intended to be within the scope of the claims and form different embodiments, as will be understood by those skilled in the art.

[0094] The following items contain preferred embodiments of the present invention: 1. A method 24 for retraining a pre-trained machine learning model 28 for defect detection in an image 26 of a lithography mask 14, the method comprising: a. Using an inspection system 50, an adjusted image 52 of one or more photomasks 14 is acquired, wherein the inspection system 50 is configured to acquire an image 26 of the photomasks 14; b. Using the adjusted image 52, adjust at least one parameter of one of the simulations 44 of the inspection system 50; c. Using the adjusted simulation 44 of the inspection system 50, one or more training images 33 of the photomask 14 are generated; and d. Use these training images 33 to retrain the machine learning model 28 for defect detection. 2. A method 24 for defect detection in an image 26 of a lithography mask 14 using a pre-trained machine learning model 28, the method comprising: a. Using an inspection system 50, an adjusted image 52 of one or more photomasks 14 is acquired, wherein the inspection system 50 is configured to acquire an image 26 of the photomasks 14; b. Using the adjusted image 52, adjust at least one parameter of one of the simulations 44 of the inspection system 50; c. Using the adjusted simulation 44 of the inspection system 50, one or more training images 33 of the photomask 14 are generated; d. Using these training images 33, the machine learning model 28 is retrained for defect detection; e. The retrained machine learning model 28 for defect detection is applied to an image 26 of a photomask 14 acquired by the inspection system 50. 3. The method of Project 1 or Project 2, wherein adjusting at least one parameter of the simulation 44 of the inspection system 50 in step b. involves solving an optimization problem. 4. The method of Project 3, wherein for at least one adjusted image 52, a design 32 of a corresponding lithography mask 14 can be provided, wherein the simulation 44 of the inspection system 50 uses the design 32 of the lithography mask 14 to simulate the image acquisition procedure for a lithography mask 14, and wherein solving the optimization problem includes minimizing the deviation between one or more adjusted images 52 and the simulated images of the corresponding design 32. 5. The method of Project 3 or Project 4, wherein solving the optimization problem involves maximizing the similarity between the distribution of the adjusted image 52 and one of the distributions of the simulated image obtained by simulation 44 of the inspection system 50. 6. The method of any of Items 3 to 5, wherein the gradient of the simulation 44 of the checking system 50 with respect to the at least one parameter is derived, and wherein solving the optimization problem includes using a gradient descent method (using the derived gradients). 7. The method of any of the foregoing items further includes, prior to step a., pre-training a machine learning model 28 using a pre-trained image 31 comprising at least one simulated image generated by simulation 44 using inspection system 50. 8. The method of any of the foregoing items further includes, prior to step a., pre-training the machine learning model 28 using a pre-trained image 31 containing at least one image acquired by the inspection system 50. 9. The method of any of the foregoing items, wherein at least some of the training images 33 contain defect annotations 36. 10. The method of any of the foregoing items, wherein the simulation 44 of the inspection system 50 simulates the image acquisition procedure for a photomask 14. 11. The method of Project 10, wherein the simulation 44 of the inspection system 50 includes a design 32 using a lithography mask 14 to simulate the image acquisition procedure. 12. The method of Item 11, wherein the generated training images 33 include simulated images of defective designs 34 and / or designs containing uncommon structures and corresponding defect annotations 36, the simulated images being obtained by applying simulation 44 of inspection system 50 to the defective designs 34 and / or designs containing uncommon structures. 13. The method of any of the foregoing items, wherein prior knowledge is used in step b. to adjust at least one parameter of the simulation of the inspection system. 14. The method of any of the foregoing items, wherein most of the adjusted images 52 are different from the pre-trained images 31 used to pre-train the machine learning model 28 in at least one sample from the group including the inspection system, the image acquisition period, and the lithography mask. 15. The method of any of the foregoing items, wherein the pre-training of the machine learning model 28 is performed on a first computer system, and at least step d. is performed on a second computer system. 16. The method of any of the foregoing items, wherein the number of the adjusted images 52 is less than 1% of the number of pre-trained images 31 used for pre-training the machine learning model 28, preferably less than 0.1%, more preferably less than 0.01%, and most preferably less than 0.001%. 17. The method of any of the foregoing items, wherein steps a. to d. are iterative. 18. The method of any of the foregoing items, wherein the simulation 44 of the inspection system 50 is a physical simulation of the transmission of electromagnetic waves within a photomask. 19. The method of any of the foregoing items, wherein the simulation 44 of the inspection system 50 includes the application of a trained machine learning model 38. 20. A method for detecting defects 22 in an image 26 of a lithography mask 14 in an inspection system 50, the method comprising retraining a machine learning model 28 for defect detection based on any of the foregoing items. 21. An inspection system 50 for detecting defects 22 in an image 26 of a photomask 14 comprises: a. An image acquisition unit 60 configured to acquire an image 26 of a photomask 14; and b. A data analysis device 62, comprising at least one memory 64, and at least one processor 66 configured to perform the steps of a method for detecting defects 22 in an image 26 of a photomask 14 according to item 20.

[0095] In summary, this invention relates to a method for retraining a pre-trained machine learning model 28 for defect detection in lithography images. The method includes: acquiring adjusted images 52 of one or more lithography masks using an inspection system 50; adjusting at least one parameter of a simulation 44 of the inspection system 50 using the adjusted images 52; generating training images 33 of one or more lithography masks using the adjusted simulation 44 of the inspection system 50; and retraining the machine learning model 28 using the generated training images 33 for defect detection, thus providing a retrained machine learning model 28 for defect detection in an image of a lithography mask acquired by the inspection system 50. This invention also relates to a method and inspection system for defect detection.

[0096] 10: Transmissive lithography system; Lithography system 10': Reflective lithography system; Lithography system 12: Radiation source; light source 14: Microfilm Light Mask 16: Illumination Optical Devices 17: Projection Optical Devices 18: Wafer Plane 19: Projection Section 20,26: Images 22: Defect; Artificial defect 24: Method 28: Pre-trained machine learning model; machine learning model; trained machine learning model; retrained machine learning model 30: Training Videos 31: Pre-trained images; simulated pre-trained images; a variety of pre-trained images 32: Design; Defect-free design 33: Training images; simulated training images 34: Defective Design 36: Defect Notes 38: Machine learning model; trained machine learning model 40: Pre-training phase 42: Initial simulation parameters; Simulation parameters 44: Simulation; Adjusted simulation 46: Pre-training steps 47: Pre-trained machine learning models 48: Adjustment Phase 50: Inspection System 52: Adjust Image 54: Parameter Adjustment Steps 56: After adjusting the simulation parameters 58: Retraining Phase 60: Image Acquisition Unit 62: Data Analysis Device 64: Memory 66: Processor 68: Interface 92: Microfilm Lithomask Model M0: Steps as needed; Steps M1, M2, M3, M4, M6: Steps M5: Steps as needed

Claims

1. A retraining method (24) for a pre-trained machine learning model (28) for defect detection in an image (26) of a lithography mask (14), the method comprising: a. acquiring one or more adjusted images (52) of the lithography mask (14) using an inspection system (50), wherein the inspection system (50) is configured to acquire the image (26) of the lithography mask (14); b. adjusting at least one parameter of a simulation (44) of the inspection system (50) using the adjusted images (52); c. generating one or more training images (33) of the lithography mask (14) using the adjusted simulation (44) of the inspection system (50); and d. retraining the machine learning model (28) using the training images (33) for defect detection.

2. The method of request item 1, wherein adjusting at least one parameter of the simulation (44) of the inspection system (50) in step b. involves solving an optimization problem.

3. The method of claim 2, wherein for at least one of the adjusted images (52), there is a design (32) corresponding to the photomask (14), wherein the simulation (44) of the inspection system (50) uses the design (32) of the photomask (14) to simulate the image acquisition procedure for the photomask (14), and wherein solving the optimization problem includes minimizing the deviation between the one or more adjusted images (52) and the simulated images of the corresponding designs (32).

4. The method of claim 2, wherein solving the optimization problem involves maximizing the similarity between the distribution of the adjusted image (52) and the distribution of the simulated image obtained by the simulation (44) using the inspection system (50).

5. The method of claim 2, wherein the gradient system of the simulation (44) of the checking system (50) with respect to the at least one parameter is derived, and wherein solving the optimization problem includes using a gradient descent method with the derived gradients.

6. The method of claim 1 further includes, prior to step a, pre-training the machine learning model (28) with a pre-trained image (31) containing at least one simulated image generated by the simulation (44) using the inspection system (50).

7. The method of claim 1 further includes, prior to step a, pre-training the machine learning model (28) using a pre-trained image (31) containing at least one image acquired by the inspection system (50).

8. The method of claim 1, wherein at least some of the training images (33) contain defect annotations (36).

9. The method of claim 1, wherein the simulation (44) of the inspection system (50) simulates the image acquisition procedure for the photomask (14).

10. The method of claim 9, wherein the simulation (44) of the inspection system (50) includes simulating the image acquisition procedure using one of the designs (32) of the photomask (14).

11. The method of claim 10, wherein the generated training images (33) include simulated images of defective designs (34) and corresponding defect annotations (36), the simulated images being obtained by applying the simulation (44) of the inspection system (50) to the defective designs (34).

12. The method of claim 1, wherein prior knowledge is used in step b. to adjust at least one parameter of the simulation of the inspection system.

13. The method of claim 1, wherein most of the adjusted images (52) are different from the pre-trained images (31) used to pre-train the machine learning model (28) in at least one sample from a group including the inspection system, the image acquisition period, and the lithography mask.

14. The method of claim 1, wherein the pre-training of the machine learning model (28) is performed on a first computer system, and at least step d. is performed on a second computer system.

15. The method of claim 1, wherein the number of such adjusted images (52) is less than 1% of the number of pre-training images (31) used to pre-train the machine learning model (28), preferably less than 0.1%, more preferably less than 0.01%, and most preferably less than 0.001%.

16. As in request item 1, where steps a. through d. are iterated.

17. The method of claim 1, wherein the simulation (44) of the inspection system (50) comprises a physical simulation of electromagnetic wave transmission within a photomask.

18. The method of request 1, wherein the simulation (44) of the inspection system (50) includes the application of a trained machine learning model (38).

19. A method (24) for defect detection in an image (26) of a lithography mask (14) using a pre-trained machine learning model (28), the method comprising: a. acquiring one or more adjusted images (52) of the lithography mask (14) using an inspection system (50), wherein the inspection system (50) is configured to acquire the image (26) of the lithography mask (14); b. adjusting at least one parameter of a simulation (44) of the inspection system (50) using the adjusted images (52); c. generating one or more training images (33) of the lithography mask (14) using the adjusted simulation (44) of the inspection system (50); d. retraining the machine learning model (28) using the training images (33) for defect detection; e. applying the retrained machine learning model (28) for defect detection to an image (26) of the lithography mask (14) acquired by the inspection system (50).

20. An inspection system (50) for detecting a defect (22) in an image (26) of a photomask (14) comprises: a. an image acquisition unit (60) configured to acquire the image (26) of the photomask (14); and b. a data analysis device (62) comprising at least one memory (64) and at least one processor (66) configured to perform the steps of a method for detecting the defect (22) in the image (26) of the photomask (14) according to claim 19.

Images