Method for retraining a machine learning model for defect detection in an image of a photolithography mask

Abstract

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

Description

Field of invention

[0001] The invention relates to methods and systems for quality control and quality assurance in photolithography masks, in particular a method for retraining a machine learning model and a corresponding method and inspection system for defect detection in an image of a photolithography mask. The methods and systems can be used for quantitative metrology, process monitoring, defect detection, and defect verification in photolithography masks. Background of the invention

[0002] Semiconductor manufacturing involves the precise manipulation, such as etching, of materials like silicon or oxide at extremely fine scales in the nanometer range. Therefore, a quality management process encompassing quality assurance and quality control is essential to ensure high quality standards for the manufactured wafers. Quality assurance refers to a set of activities aimed at ensuring high-quality products by preventing any defects that may occur during the development process. Quality control refers to a system for inspecting the final quality of the product. Quality control is a component of the quality assurance process.

[0003] A wafer, made from a thin silicon disc, serves as the substrate for microelectronic devices containing semiconductor structures embedded in and on the wafer. These semiconductor structures are built up layer by layer through repeated processing steps involving chemical, mechanical, thermal, and optical processes. The dimensions, shapes, and placement of the semiconductor structures and patterns are subject to several influences. One of the most important steps is the photolithography process.

[0004] Photolithography is a process used to produce patterns on a substrate. The patterns to be printed onto the substrate surface are created using computer-aided design (CAD). From the design, a photolithography mask is created for each layer, containing a magnified image of the computer-generated pattern to be etched into the substrate. The photolithography mask can be further adapted, for example, using optical near-field correction techniques. During the printing process, an illuminated image projected from the photolithography mask is focused onto a thin photoresist film formed on the substrate. A semiconductor chip that powers mobile phones or tablets, for example, comprises approximately 80 to 120 structured layers.

[0005] Due to the increasing integration density in the semiconductor industry, photolithography masks must map ever smaller structures onto wafers. The aspect ratio and the 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 decreasing. 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 decreasing. To create the small feature dimensions mapped onto the wafer, photolithographic masks or templates for nanoimprint photolithography with ever smaller structures or pattern elements are required.The production process for photolithographic masks and templates for nanoimprint photolithography is therefore becoming increasingly complex, time-consuming, and ultimately more expensive. With the advent of EUV photolithography scanners, the type of mask changed from transmission-based to reflection-based structuring.

[0006] Due to the minute structural sizes of the pattern elements in photolithography masks or templates, it is impossible to eliminate errors during mask or template production. The resulting defects can arise, for example, from the degradation of photolithography masks or particle contamination. Of the various defects that occur during semiconductor fabrication, photolithography-related defects account for almost half of the total. Therefore, the inspection, verification, and metrology of photolithography masks play a crucial role in semiconductor process control for monitoring 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 adjust the manufacturing process parameters, such as...To improve exposure time, focus variation, etc.

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

[0008] To analyze large amounts of data requiring extensive measurements, machine learning techniques can be employed. Machine learning is a branch of artificial intelligence. Generally, machine learning techniques create a parametric machine learning model based on training data consisting of a large number of samples. After training, the model is able to generalize the knowledge gained from the training data to new, previously unexplored samples, thereby making predictions for new data. Many machine learning techniques exist, such as approaches using linear regression, k-means, support vector machines, decision trees, random forests, neural networks, and deep learning.

[0009] Deep learning is a class of machine learning that uses artificial neural networks with numerous hidden layers between the input and output layers. Because of this complex internal structure, the networks are also able to continuously extract higher-level features from the raw input data. Each layer learns to transform its input data into a more abstract and compositional presentation, thus deriving low- and high-level insights from the training data. The hidden layers can have different sizes and tasks, such as convolutional or pooling layers.

[0010] Nevertheless, defect detection methods suffer from the problem of changing conditions. For example, the optical properties of photolithography masks change during their use in photolithography processes, e.g., due to EUV exposure tools. Furthermore, different photolithography masks can vary not only in their design but also in their material. Therefore, it is important to adapt existing defect detection methods to changing conditions, such as a different inspection system and / or a different photolithography mask and / or modified conditions in the inspection system and / or modifications to the optical properties of the photolithography mask, etc.For example, reference images used in defect detection methods must be adapted to changing conditions to prevent high numbers of false positive defect detections, and images from different machines or even different types of machines can vary considerably (e.g., in contrast or appearance), for example, due to different materials or changing optical properties of the photolithography mask. All these scenarios require rapid adaptation of the defect detection method to changing conditions. Currently, machine learning models require large amounts of training data, which are not typically available in these cases.

[0011] To mitigate this problem, domain adaptation methods have been proposed in the literature. Domain adaptation methods attempt to adapt the model directly to the changing conditions or to adapt the model's training data to the changing conditions. However, due to the lack of knowledge about the image formation process, many training images of the new domain are still required for training; otherwise, the adaptation cannot be robustly estimated.

[0012] Therefore, an objective of the invention is to adapt a trained machine learning model for defect detection in a photolithography mask image to modified conditions. In particular, an objective of the invention is to adapt the machine learning model in such a way that only a small amount of training data, reflecting the modified conditions, is required with minimal user effort and minimal image acquisition time. Furthermore, an objective is to increase the accuracy of the predictions of the pre-trained machine learning model.

[0013] The objectives 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

[0014] Embodiments of the invention relate to methods and systems for retraining pre-trained machine learning models for defect detection in images of photolithography masks.

[0015] One embodiment of the invention includes a method for retraining a pre-trained machine learning model for defect detection in an image of a photolithography mask, wherein the method comprises: a. capturing fitting images of one or more photolithography masks using an inspection system, the inspection system being configured to capture images of photolithography masks; b. adjusting at least one parameter of a simulation of the inspection system using the fitting images; c. generating training images of one or more photolithography 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 of the retrained machine learning model for defect detection in an image of a photolithography mask captured by the inspection system.

[0016] One embodiment of the invention includes a method for defect detection in an image of a photolithography mask using a pre-trained machine learning model, wherein the method comprises: a. capturing fitting images of one or more photolithography masks using an inspection system, the inspection system being configured to capture images of photolithography masks; b. adjusting at least one parameter of a simulation of the inspection system using the fitting images; c. generating training images of one or more photolithography 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 photolithography mask captured by the inspection system.

[0017] By using a simulation of the inspection system, the number of required training images for the new domain is low compared to domain adaptation methods, as only a few simulation parameters require adaptation. This is because the prior knowledge about the image formation process is already contained within the simulation of the inspection system. The adapted simulation of the inspection system can then be used to automatically generate vast quantities of training images for training the defect detection algorithm, without the need to manually acquire these training images.

[0018] At least one parameter of the inspection system is adjusted using fitting images. Only a small number of fitting images are required to adapt the inspection system simulation. For example, the number of fitting 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 adapting the inspection system simulation to the fitting images, a vast number of training images can be simulated that closely resemble the acquired images of the inspection system. Accordingly, large numbers of training images can be generated automatically with minimal user effort, requiring only a very small number of actually acquired images in the inspection system.This procedure enables, for example, the development of a trained machine learning inspection system at a first location, which is then adapted to different conditions at a second location. Instead of retraining the inspection system at the second location solely based on acquired training images from that location—which is often impossible due to the limited number of available images—only the parameters of the inspection system's simulation need to be adjusted using a very small number of adaptation images. The simulation can then be used to generate realistic training images for retraining the machine learning model in the inspection system at the second location. In this way, retraining the machine learning model becomes feasible even under modified conditions when only a few acquired training images (in this case, adaptation images) are available.At the same time, the user effort required to adapt the machine learning model to the new conditions, e.g., at the second location, is minimized. Furthermore, the time required to adapt the inspection system's machine learning model to the new conditions is significantly reduced using the adapted simulation of the inspection system. Additionally, the time required to acquire the necessary training images is reduced. Finally, the energy consumption of the inspection system, as well as wear and tear on the inspection system and one or more photolithography masks, are reduced.

[0019] The term "inspection system" refers to a system configured to inspect photolithography masks by capturing an image of the mask and detecting defects in the image. An inspection system can be, for example, a verification system or a repair system.

[0020] The photolithography mask can have an aspect ratio between 1:1 and 1:4, preferably between 1:1 and 1:2, and particularly preferably between 1:1 or 1:2. The photolithography mask can have an almost rectangular shape. The photolithography mask can preferably be 5 to 7 inches long and wide, and particularly preferably 6 inches long and wide. Alternatively, the photolithography mask can be 5 to 7 inches long and 10 to 14 inches wide, and preferably 6 inches long and 12 inches wide.

[0021] An image of a photolithography mask can refer to various types of images of the photolithography mask, such as two-dimensional or volumetric three-dimensional images that can be processed, for example, slice by slice. An image can be acquired using an inspection system or it can be simulated, for example, from a design of the photolithography mask. Images can be of various modalities, such as structured electron microscopy (SEM) images, aerial photographs, optical images, X-ray images, computed tomography (CT) images, focused ion beam (FIB) images, atomic force microscopy (AFM) images, ultrasound images, or multimodal images acquired, for example, using a combination of X-ray imaging and SEM. An image of a photolithography mask can depict the entire photolithography mask or one or more sections thereof.Preferably, an image of a photolithography mask refers to an aerial photograph of it.

[0022] A training image, a pre-training image, and an adaptation image each include at least one image of a photolithography mask. It may also include defect annotations or a design of the photolithography mask.

[0023] The term "defect" refers to a localized deviation of an integrated circuit pattern from a priori defined standard for that pattern. For example, a defect in an integrated circuit pattern, such as a semiconductor structure, can lead to a malfunction of an associated semiconductor device. Depending on the detected defect, the photolithography process can be improved, or photolithography masks or wafers can be repaired or discarded. The standard of the structure or pattern can be defined by one or more corresponding reference photolithography masks or reference datasets, such as design datasets, simulated datasets, or captured defect-free datasets.

[0024] A machine learning model for defect detection can perform various tasks, such as defect detection (presence or absence of a defect), defect localization (locating a defect), defect segmentation (calculating the area, volume, or outline of a defect), defect classification (assigning a defect class), and so on. The machine learning model can be supervised, unsupervised, or semi-supervised. It can use reference datasets—primarily defect-free captured or simulated images—to detect defects, for example, a die-to-die model. Alternatively, it can be a reference-free machine learning model, such as a single-die model, which detects defects by identifying deviations from the norm, for example, from prior knowledge or knowledge derived from the die itself.A pre-trained machine learning model refers to a machine learning model that has been trained at least once before, i.e., a machine learning model whose parameters have been adapted at least once using training images. Retraining a machine learning model refers to training a pre-trained machine learning model, i.e., subsequently adapting the model's parameters using further training images.

[0025] An inspection system simulation refers to a parametric model of the inspection system that simulates the image acquisition process. The inspection system simulation can be used to generate simulated images of the photolithography mask. The parameters of the parametric model of the inspection system, i.e., the simulation parameters, can be adapted, for example, using adaptation images, to simulate a specific inspection system or specific conditions of the inspection system or photolithography mask. The simulation parameters can include, for example: - Image properties, such as image intensity, image contrast, noise level, image distortion, maximum illuminance difference, maximum focus drift, maximum misalignment, - Defect information, - Machine settings, such as light source intensity, light source parameters, - Photolithography mask properties, such as layer thickness of the photolithography mask, layer materials of the photolithography mask, - Design modification parameters, such as line edge roughness, CD variations, pattern thickness, optical proximity correction (OPC) structures, - Parameters of a machine learning model, etc.

[0026] An aerial image represents the radiation intensity distribution of a photolithography system in a wafer plane for a given photolithography mask. The aerial image thus simulates the structures on the surface of a wafer when the wafer is printed in the photolithography system using the photolithography mask. A wafer plane refers to a layer within the photoresist on the wafer in the photolithography system. An aerial image can be generated by applying an aerial image measurement system or a metrology system to a photolithography mask. An aerial image can be simulated using a photolithography mask design and an aerial image simulation method.

[0027] An aerial photograph can refer to an aerial image of a complete photolithography mask or to an aerial image of a section of the photolithography mask. A design can refer to the design of a complete photolithography mask or to the design of a section of the photolithography mask.

[0028] In the case of aerial imagery, the inspection system simulation can simulate the generation of an aerial image of a photolithography mask from a photolithography mask design. The simulation can use physics-based models, such as physical models of the photolithography mask and / or the propagation of electromagnetic waves through the photolithography mask. The simulation can also use non-physics-based models, such as machine learning models trained to generate aerial images from designs using training data. Hybrid methods that use both physical and machine learning models can also be used to simulate aerial imagery.

[0029] In one example, the procedure further includes, prior to step a., pre-training the machine learning model using pre-training images, which comprise at least one simulated image generated using the inspection system simulation. The inspection system simulation can, for example, be used with initial simulation parameters. In this way, the same inspection system simulation used in step c. can be used to generate the pre-training images, thereby significantly reducing the user effort required to provide the necessary pre-training images, the time required to acquire the pre-training images, the energy consumption of the inspection system, and wear and tear on the inspection system and the photolithography mask, etc.

[0030] Additionally or alternatively, the procedure can further include, prior to step a., pre-training the machine learning model using pre-training images, which include at least one image captured by an inspection system. In this way, realistic training images are used for pre-training, resulting in higher accuracy of the machine learning model's predictions.

[0031] In a preferred example, at least some training images include defect annotations. Defect annotations can be provided, for example, by a yes / no indication, by bounding boxes of any shape and size that enclose the defect, by pixel-wise or voxel-wise segmentation of the defect, by a description, by one or more items from a defect list, etc.

[0032] The defect annotations can be used for supervised training of the machine learning model, 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.

[0033] According to one aspect of the invention, the simulation of the inspection system includes simulating the image acquisition process for a photolithography mask, for example by using a design of the photolithography mask. In this way, very accurate simulations of images can be generated, e.g., from the design of the photolithography mask.

[0034] In one example, the training images generated for training the machine learning model for defect detection include simulated images of defective designs—that is, designs containing one or more defects—and corresponding defect annotations. The simulated images of the defective designs are obtained by applying the inspection system's simulation to these designs. Defective designs can be created, for example, by modifying defect-free designs. The modifications correspond to atypical design structures, i.e., simulated defects. In this way, the simulated defects can be controlled with respect to their location, size, severity, type, frequency, and so on.Using simulated images of the defective designs and corresponding defect annotations, the machine learning model for defect detection can be trained in a supervised manner, resulting in highly accurate defect detections. Optionally, the defect-free design, which corresponds to the defective design, or a simulated image of the defect-free design can be used as an additional input to the machine learning model for defect detection.

[0035] According to one example, adjusting at least one parameter of the inspection system simulation in step b involves solving an optimization problem. This allows the accuracy of the adjusted parameters to be improved.

[0036] An optimization problem involves an objective function that is to be maximized or minimized. The optimization problem may also include constraints. Solving the optimization problem means applying some kind of mathematical procedure to calculate a point with an objective function value that is better than the objective function values ​​for several other points. Solving the optimization problem might, for example, mean calculating the global optimum or a local optimum of the objective function. The mathematical procedure might involve calculating an analytical solution or applying an iterative method, such as gradient descent, a simplex method, a variational approach, a combinatorial optimization approach, and so on.

[0037] According to one aspect of the example, a design of the corresponding photolithography mask is available for at least one fitting image. The simulation of the inspection system simulates the image acquisition process for a photolithography mask using a photolithography mask design and includes solving the optimization problem of minimizing the deviation of one or more fitting images from the simulated images of the corresponding designs. Accordingly, the accuracy of one or more fitted parameters can be improved.

[0038] According to another aspect of the example, solving the optimization problem involves maximizing the similarity between the distribution of fitted images and a distribution of simulated images obtained using the simulation of the inspection system. Accordingly, the accuracy of one or more fitted parameters can be improved.

[0039] According to another aspect of the example, gradients of the inspection system simulation are derived with respect to at least one parameter, and the optimization problem is solved using a gradient descent approach with the derived gradients. Accordingly, the accuracy of one or more fitted parameters can be improved.

[0040] In one example, prior knowledge is used to adjust at least one parameter of the inspection system simulation in step b. This prior knowledge can include, for example, known parameters such as mask thickness or lighting settings, parameter ranges, or parameter distributions. Such prior knowledge can be used to simplify the optimization problem, allowing solutions to be obtained with shorter computation times or enabling improvements to the optimization problem solution to increase the accuracy of the adjusted at least one parameter.

[0041] As an example, most of the adaptation images used for pretraining the machine learning model differ in at least one aspect from the group that includes the inspection system, the image acquisition time period, and the photolithography mask. At least some of the pretraining images used for pretraining the machine learning model can be acquired using an inspection system or simulated using an inspection system. Differences can lie, for example, in the type of inspection system, the specific instance of the inspection system, the parameters of the inspection system, the acquisition time period, the design, the material, or the specific instance of the photolithography mask, etc. In this way, the pretrained machine learning model can be adapted to modified conditions.

[0042] In one example, the pretraining of the machine learning model is performed on a first computer system, and at least step d. is performed on a second computer system. This allows the machine learning model to be adapted to modified conditions between the first and second computer systems. For example, the computer systems might be located in different places or have different characteristics, such as different hardware configurations, different confidentiality standards, different computation time requirements, etc. The first and second computer systems might also belong to different inspection systems. In this case, the machine learning model can be adapted to modified conditions between the inspection systems.

[0043] According to one aspect of the invention, steps a. to d. are repeated. In this way, a newly trained machine learning model can be retrained to adapt it to further modified conditions, for example, when a different photolithography mask is loaded into an inspection system.

[0044] For example, the simulation of the inspection system includes a physical simulation of the propagation of electromagnetic waves within a photolithography mask, e.g., using rigorous simulations, Kirchhoff's laws, or other simulation techniques. In this way, the image acquisition process can be simulated with high accuracy or with low computation time, thereby improving the accuracy of the inspection system simulation or reducing computation time.

[0045] Additionally or alternatively, the simulation of the inspection system includes the application of a trained machine learning model. This allows for improvements to the accuracy and / or computation time of the inspection system simulation.

[0046] An inspection system for detecting defects in an image of a photolithography mask according to a third embodiment of the invention comprises the following: an image acquisition unit configured to acquire images of photolithography masks; and a data analysis device comprising 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 photolithography mask according to an embodiment of the invention.

[0047] A system for retraining a machine learning model for defect detection in an image of a photolithography mask according to a fourth embodiment of the invention comprises the following: an image acquisition unit configured to acquire images of photolithography masks; and a data analysis device comprising at least one memory and at least one processor configured to perform the steps of a method for retraining a machine learning model for defect detection in an image of a photolithography mask according to an embodiment of the invention.

[0048] The invention described by embodiments, examples and aspects is not limited to the embodiments, examples and aspects, but can be implemented by a person skilled in the art through various combinations or modifications thereof. Brief description of the drawings Fig. Figure 1 illustrates an exemplary transmission-based photolithography system, e.g., a deep ultraviolet (DUV) photolithography system; Fig. Figure 2 illustrates an exemplary reflection-based photolithography system, e.g., an extreme ultraviolet (EUV: Extreme Ultraviolet) photolithography system; Fig. Figure 3 shows an imaging dataset of an object comprising an integrated circuit pattern in the form of a photolithography mask encompassing a defect; Fig. 4 shows a flowchart illustrating the steps of a method according to an embodiment of the invention; Fig. Figure 5 illustrates the application of a trained machine learning model for defect detection; Fig. Figure 6 illustrates the training of a machine learning model that simulates the image acquisition process in an inspection system; Fig. Figures 7a-c illustrate the method for retraining a machine learning model for defect detection according to one embodiment of the invention; and Fig. Figure 8 illustrates an inspection system for detecting defects in a photolithography mask according to a third embodiment of the invention. Detailed description

[0049] Advantageous exemplary embodiments of the invention are described below and shown schematically in the figures. The same reference numerals are used throughout the figures and the description to describe identical features or components. Dashed lines indicate optional features.

[0050] The present methods and systems can be used with a variety of photolithography systems, e.g. transmission-based photolithography systems 10 or reflection-based photolithography systems 10'.

[0051] Fig. Figure 1 illustrates an exemplary transmission-based photolithography system 10, e.g., a DUV photolithography system. The main components are a radiation source 12, which can be a deep ultraviolet (DUV) excimer laser source; imaging optics, which, for example, define the partial coherence and may include optics that shape radiation from the radiation source 12; a photolithography mask 14; illumination optics 16, which illuminate the photolithography mask 14; and projection optics 17, which projects an image of the photolithography mask model 92, e.g., the design pattern, onto a wafer plane 18.An adjustable filter or aperture at the pupil plane of the projection optics 17 can limit the range of beam angles that strike the wafer plane 18, where the largest possible angle defines the numerical aperture of the projection optics NA = n sin(Gmax), where n is the refractive index of the media between the substrate and the last element of the projection optics 17 and Gmax is the largest angle of the beam exiting the projection optics 17 that can still strike the wafer plane 18.

[0052] In this document, the terms “radiation” or “ray” are used to include all types of electromagnetic radiation, including ultraviolet radiation (e.g., with a wavelength of 365, 248, 193, 157 or 126 nm) and EUV (extreme ultraviolet radiation, e.g., with a wavelength in the range of about 3-100 nm).

[0053] The illumination optics 16 can include optical components for shaping, adjusting, and / or projecting radiation from the radiation source 12 before the radiation passes through the photolithography mask 14. The projection optics 17 can 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 exclude the light source 12, and the projection optics exclude the photolithography mask 14.

[0054] The illumination optics 16 and the projection optics 17 can 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 17 can also include components that operate according to one of these design types to direct, shape, or control the projection beam of radiation collectively or individually.

[0055] Fig. Figure 2 illustrates an exemplary reflection-based photolithography system 10', e.g., a lithography system for extreme ultraviolet (EUV) light. The main components are a radiation source 12, which can be a laser plasma light source; illumination optics 16, which, for example, define the partial coherence and can include optics that shape radiation from the radiation source 12; a photolithography mask 14; and projection optics 17, which projects an image of the photolithography mask model 92, the design pattern, onto a wafer plane 18.An adjustable filter or aperture at the pupil plane of the projection optics 17 can limit the range of beam angles that strike the wafer plane 18, where the largest possible angle defines the numerical aperture of the projection optics NA = n sin(Gmax), where n is the refractive index of the media between the substrate and the last element of the projection optics 17 and Gmax is the largest angle of the beam exiting the projection optics 17 that can still strike the wafer plane 18.

[0056] Fig. Figure 3 illustrates an image 20 of a photolithography mask 14 containing a defect 22. According to the techniques described here, various imaging modalities can be used to acquire the image. The images can include single-channel images or multi-channel images, such as focus stacks. For example, the image may contain 2D images. It is possible to employ a multi-beam scanning electron microscope (mSEM). mSEM uses multiple beams to simultaneously acquire images in several fields of view. For example, a number of no fewer than 50 or even no fewer than 90 beams could be used. Each beam covers a separate portion of a surface of the photolithography mask. This allows a large image to be acquired within a short time. Typically, current machines acquire 4.5 gigapixels per second.Other examples of images that include 2D images involve imaging modalities such as optical imaging, phase-contrast imaging, X-ray imaging, etc. It is also possible that the image is a volumetric 3D 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) could be used. Furthermore, magnetic resonance (MR) images, ultrasound images, or computed tomography (CT) images could be used. Multimodal images can be used, e.g., a combination of X-ray imaging and SEM. The image could be an aerial image acquired by an aerial survey 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 image measurement system can be equipped with, for example, a staring array sensor, a line scan sensor, or a time-delayed integration (TDI) sensor.

[0057] For defect detection, machine learning models are a popular choice for obtaining high-quality results with short computation times. Machine learning models are trained using training data, i.e., examples, and thus derive their knowledge independently from this data, rather than requiring a user to define defect detection rules. In this way, optimal defect detection results can be obtained automatically in a data-driven manner. Consequently, the use of machine learning models increases the recall and precision of defect detection procedures and reduces user effort.

[0058] However, machine learning models, especially deep learning models, require large amounts of training data that are not always available. Furthermore, training a machine learning model can take up to several days or weeks, depending on its complexity. Additionally, the machine learning model can only apply the knowledge derived from the training data. If conditions change, the machine learning model must be retrained.

[0059] Conditions can change, for example, if the optical properties of photolithography masks change during their use in photolithography processes, such as through EUV exposure tools. Conditions can also change if the photolithography mask is replaced, as different photolithography masks can vary not only in their design but also in their material. Conditions also change if the defect detection procedure is performed on a different inspection system, such as the same type of inspection system or a different type. In all these cases, rapid retraining of the machine learning model for defect detection is required. However, training data is usually unavailable or very scarce when conditions change.

[0060] To enable retraining of the machine learning model, it shows Fig. 4 A flowchart of a method 24 for retraining a pre-trained machine learning model for defect detection in an image of a photolithography mask according to an embodiment of the invention. The method for retraining a pre-trained machine learning model for defect detection in an image of a photolithography mask comprises the following: a. Acquiring, in a step M1, fitting images of one or more photolithography masks using an inspection system, wherein the inspection system is configured to acquire images of photolithography masks; b. Fitting, in a step M2, at least one parameter of a simulation of the inspection system using the fitting images; c. Generating, in a step M3, training images of one or more photolithography masks using the fitted simulation of the inspection system; and d.In step M4, the machine learning model is retrained for defect detection using the generated training images. In step M6, the retrained machine learning model can then be used for defect detection in an image of a photolithography mask captured by the inspection system.

[0061] As in Fig. As illustrated in Figure 5, the trained machine learning model 28 for defect detection receives an image 26, in this case an aerial image, as input and assigns the image 26 to no defects, one or more defects 22 or representations thereof. A supervised machine learning model for defect segmentation can use well-known machine learning segmentation architectures, such as a U-Net or a Segformer, as described in the scientific publication "SegFormer: Simple and Efficient Design for Semantic Segmentation with Transformers," Enze Xie, Wenhai Wang, Zhiding Yu, Anima Anandkumar, Jose M. Alvarez, Ping Luo, arXiv 2105.15203, 2021. A supervised machine learning model for defect localization can use well-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.A bounding frame suggestion is retained if an additional keypoint of the same class is found in the central region of the bounding frame. YOLO is a very fast, machine learning-based object detector that uses a fully convolutional neural network for bounding frame prediction. The image is divided into grid cells, and for each grid cell, a specified number of bounding frames are predicted, which may be larger than the grid cell. Bounding frames are retained based on their class probabilities and bounding frame confidences. By using bounding frame-based object detectors, areas containing a defect can be distinguished from defect-free areas. For defect detection, reference-based registration and detection machine learning architectures, such as those described in German patent application 10 2023 104 378, can be used.1 reveals, or reference-based die-to-database machine learning architectures can be used via known design-to-image simulation algorithms. Furthermore, self-monitoring or reference-free anomaly detection machine learning models that do not require annotated training images can be used for defect detection, e.g., autoencoders.

[0062] In an optional step M0 prior to step M1, the procedure can further include pretraining the machine learning model using pretraining images. The pretraining images can include at least one image captured by an inspection system. This at least one image can be captured using the same inspection system as in step M1 for capturing the adaptation images, or it can be captured using a different inspection system, or even a different type of inspection system. For example, the machine learning model can be pretrained using images captured by one inspection system at a first location, and it can be retrained using images captured by a different inspection system at a second location.Even if the inspection systems are of the same type, the parameters or conditions at the two locations may differ, requiring retraining of the machine learning model.

[0063] Alternatively or additionally, the pre-training images 31 can include at least one simulated image generated using the simulation of the inspection system. The inspection system simulation can be used with initial simulation parameters to generate the pre-training images. These initial simulation parameters can later be adapted to new conditions when using the adaptation images in step M2.

[0064] In an optional step M5, a final retraining of the newly trained machine learning model can be performed after step d of the procedure using fewer acquired training images of the inspection system, preferably containing defect annotations. This procedure can be particularly advantageous if the simulation of the inspection system is not entirely realistic or if the adjusted simulation parameters are not a perfect match and the simulated images are therefore not entirely realistic. In these cases, a final retraining with acquired training images of the inspection system would reduce potential deficiencies of the newly trained machine learning model due to suboptimal simulated training images. Consequently, the accuracy of the predictions of the newly trained machine learning model could be improved.

[0065] According to one aspect of the invention, the simulation 44 of the inspection system 50 comprises simulating the image acquisition process for a photolithography mask 14, in particular by using a design 32 of the photolithography mask 14. The design 32 is used to generate an image 26, e.g., an aerial photograph, of the corresponding photolithography mask. Simulating the image acquisition process for a photolithography mask can, for example, include using an optically inspired simulation based on physics-based models, e.g., rigorous simulations or Kirchhoff simulations. For these simulations, the simulation parameters could include optical parameters, such as a light source intensity or a mask thickness. Alternatively, simulating the image acquisition process could include using a data-driven model, e.g., a machine learning model trained to predict an image from a design of a photolithography mask.In this case, the simulation parameters could include the trainable parameters of the machine learning model. Alternatively, hybrid models could be used that combine an optically inspired simulation based on a physical model with a machine learning model. In this case, the simulation parameters could include a combination of optical and machine learning parameters.

[0066] Aerial images can be simulated, for example, using a physical model to generate an aerial image from a design. This leads to accurate results but is often time-consuming. Among these methods are rigorous simulation techniques, such as Finite Difference Time Domain (FDTD) or Rigorous Coupled Wave Analysis (RWCA), which are familiar to those skilled in the art. Because these require long computation times, fast approximations, such as Thin Element Approximation (TEA), can be used. Thin Element Approximation (TEA) assumes that the thickness of the structures on the photolithography mask is very small compared to the wavelength and that the widths of the structures on the photolithography mask are very large compared to the wavelength.However, since photolithographic processes use radiation of increasingly shorter wavelengths, and the structures on the pattern-forming device become smaller and grow in the vertical dimension, these assumptions no longer hold true, and mask 3D effects must be taken into account. Therefore, the results of the TEA method are less accurate but much faster to obtain than rigorous simulation results.

[0067] The simulation 44 of the inspection system 50 can also include a machine learning model to simulate the image acquisition process. Fig. Figure 6 shows a machine learning model 38 trained to simulate an image acquisition process in order to generate simulated pre-training images 31 in step M0 or training images 33 in step M3. The training images 30 for the machine learning model simulating the image acquisition process comprise designs 32 and corresponding images 26, e.g., aerial photographs. The corresponding images 26 are preferably acquired using any image acquisition system. During training, an objective function is minimized. The objective function can, for example, minimize the deviation of the images generated by the machine learning model from the corresponding images 26.

[0068] In another example, physics-based models can be used to simulate the image acquisition process, such as the propagation of incident electromagnetic waves through photolithography masks in the case of aerial images. Machine learning models can then optionally be used to refine these results. To obtain realistic simulated images, for example, noise, focus variations, or deviations from ideal structures, such as line edge roughness, variations in structure thickness, or corner rounding, can be applied to the simulated images.

[0069] Preferably, at least some of the training images (30 simulated images of designs) contain unusual structures. Unusual structures are structures that are not usually part of a design, e.g., defects, design deviations that are not necessarily classified as defects, such as minor deviations from design structures (e.g., variations in structure thickness, corner rounding, etc.), or types of structures not included in the training images. For example, if only designs containing lines and spaces are included in the training images, unusual structures could include holes, intersections, auxiliary features, complex polygons, etc. For example, the rightmost design (32) contains Fig. 6 a defect 22 in the form of a bulge or intrusion. In this way, the simulation 44 of the inspection system 50 is able to simulate the image acquisition process for all types of designs, including designs containing unusual structures.

[0070] Using the simulation 44 of the inspection system 50, training images for the machine learning model for defect detection can then be quickly and easily simulated using the simulated image acquisition process (e.g., using physical models and / or machine learning models). Artificial defects 22 can be added to the designs 32 to obtain defective designs 34. From these artificial defects 22, defect annotations 36 can be easily generated along with the 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, other unusual structures can also be added to the designs 32 to generate training images. In this way, the machine learning model 28 for defect detection learns to distinguish between defects 22 and unusual structures that are not considered defects 22.

[0071] Fig. Figures 7a to c illustrate the method according to the invention for effectively retraining a machine learning model 28 for defect detection in an image of a photolithography mask.

[0072] Fig. Figure 7a illustrates a pre-training stage 40 according to the optional step M0 for pre-training the machine learning model 28 using pre-training images 31. The simulation 44 of the inspection system 50 includes a simulation of the image acquisition process when a design 32 of a photolithography mask is given, e.g., a physics-based model or a machine learning model 38, as in Fig. 6 illustrates.

[0073] To generate the pre-training images 31, the simulation 44 of the inspection system 50 with initial simulation parameters 42 is preferably used, e.g., the trained machine learning model 38 with initial parameters in Fig. 6 or a physics-based model with initial parameters. The initial simulation parameters can be selected in various ways. For example, parameters of inspection system 50 with a high probability can be used, or average parameters of inspection system 50 can be used, or expected parameters of inspection system 50 can be used, or random parameters drawn from pre-specified ranges of the parameters of the inspection system can be used. If possible or probable modifications of the conditions are already known, e.g., the type of inspection system or commonly used types of inspection systems, lighting conditions at a second location or typical lighting conditions, the types of photolithography masks used at a second location or typical types of photolithography masks, etc.These can be taken into account when the initial simulation parameters 42 are selected to generate pre-training 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 diverse pre-training images 31 for pre-training the machine learning model 28. In this way, the machine learning model can already be closely adapted to the modified conditions.

[0074] Additionally or alternatively, images acquired by an inspection system 50 can be used as pre-training images 31 for pre-training the 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 adaptation images 52 can be acquired using different inspection systems 50 of the same type, but with slightly differing behavior 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 adaptation images 52 are acquired using inspection systems 50 of different types, resulting in images of different modalities, e.g., aerial photographs and SEM images. Even if these images originate from different modalities, they can contain information about the same photolithography mask.In another example, at least some of the pre-training images 31 and at least some of the adaptation images 52 are captured by the same inspection system 50, but for different photolithography masks, e.g., with different materials, different designs, etc. In yet another example, at least some of the pre-training images 31 and at least some of the adaptation images 52 are captured by the same inspection system 52 for the same photolithography mask, but at different time periods, e.g., at least some of the pre-training images 31 were captured days, weeks, months, or even years before at least some of the adaptation images 52, or vice versa. In this way, captured images can be used in addition to or instead of simulated pre-training images 31 for pre-training the machine learning model.

[0075] Preferably, at least some of the designs are defective designs 34, which include one or more defects 22. Simulated images of the defective designs 34 can be used to pretrain the 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 that includes a defect can be used. Using the simulation 44 of the inspection system 50, e.g., with initial simulation parameters 42, an image of a defective design 34 can be simulated, which is used as a pretraining image 31 for pretraining the machine learning model 28 for defect detection. Preferably, at least some of the simulated pretraining images 31 include defect annotations 36.Since the defects 22 are artificially generated, the corresponding defect annotations 36 can easily be included in the pre-training 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 a 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, extent, and / or type of a defect, etc., e.g., in the form of bounding boxes, pixel annotations, defect classes, etc. The pre-training images 31 can also include captured images with or without defect annotations 36. The pre-training images 31 can also include captured or simulated images without defects.

[0076] The pre-training images 31 are used to pre-train the machine learning model 28 for defect detection in a pre-training step 46, which results in a pre-trained machine learning model 47. The defect-free design 32, which corresponds to the defective design 34, can optionally be used as additional input for the machine learning model 28 for defect detection.

[0077] Fig. Figure 7b illustrates an adaptation stage 48 for adjusting the parameters of the simulation 44 of the inspection system 50 to modified conditions, e.g., at a different location, using a different inspection system 50 or a different type of inspection system 50, different machine settings, e.g., lighting settings, different photolithography masks, different materials, different designs 32, different image properties, e.g., noise, distortion, intensity, brightness, contrast, etc. For this purpose, adaptation images 52 are acquired using the inspection system 50. Since no retraining of the pre-trained machine learning model 47 is intended at this stage, but only the adaptation of the simulation parameters of the inspection system 50, acquiring a small number of adaptation images 52 is sufficient, e.g.,less than 1%, preferably less than 0.1%, more preferably less than 0.01%, and particularly preferably less than 0.001% of the number of pre-training images used to pre-train the machine learning model 28 or of the training images used to retrain the machine learning model 28. Preferably, at least some of the adaptation images 52 are captured using photolithography masks with the same design 32 as the designs 32 underlying the pre-training images 31 used to pre-train the machine learning model 28. This simplifies parameter adjustment and yields more accurate results. The simulation parameters of the simulation 44 of the inspection system 50 are then adapted to the captured adaptation images 52 in a parameter adjustment step 54, resulting in adapted simulation parameters 56.

[0078] Parameter fitting step 54 preferably involves solving an optimization problem. If the fitting images x i Designs d i The corresponding photolithography masks can be included in parameter adjustment step 54 to adjust the parameters θ. k The simulation T of the inspection system can be performed, for example, by solving the following optimization problem, which minimizes a loss function L: θT=argminθ∑iL(xi,T(di,θ)).

[0079] The loss function measures the difference between the fitted image x i and the simulated image T(d i , θ) for the corresponding design d iand one or more given simulation parameters θ. Accordingly, the simulation parameters 42 are adjusted so that the simulated images from the designs match the fitted images as closely as possible. The loss function may contain further terms, e.g., regulation terms. Alternatively, especially if corresponding designs d i for the adjustment images x i are not available, the loss function the similarity of a distribution of the fitting images x iand maximize the distribution of simulated images for some available designs (or minimize dissimilarity), e.g., by minimizing Kullback-Leibler divergence or other stochastic measures, or by comparing image statistics using, for example, signal-to-noise ratios, contrast ratios, blur, edge thickness, etc. Alternatively, classification methods, such as discriminators used in approaches with a generative adversarial network (GAN), can be used to learn to distinguish between matching images x. i and simulated images T(d i , θ) are used. The parameters θ of the simulation 44 of the inspection system 50 can be modified until the discriminator no longer distinguishes between the fitting images x. i and the simulated images T(d i, θ) can distinguish, that is, if the classifier ambiguity about the two sets of images x i and T(d i , θ) is maximized. The discriminator can be trained on spots in the images.

[0080] If the simulation 44 of the inspection system 50 allows the calculation of gradients with reference to the simulation parameters, the simulation parameters can be adjusted by solving an optimization problem using gradient descent approaches.

[0081] 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 in advance, such as a lighting parameter, a mask material, etc. Such parameters can be loaded from a database, for example, or they can be specified by a user. Furthermore, parameter value ranges or probability distributions over parameter values ​​can be specified for many simulation parameters. Parameter value ranges can be used in the optimization problem to restrict the values ​​of the simulation parameters using constraints. Probability distributions over parameter values ​​can, for example, be included in the objective function of the optimization problem. The probability distributions can be maximized, or their negative log-likelihood can be minimized. L(θ)=o(θ)−log p(θk).

[0082] Here, L(θ) gives a parameterized loss function, o(θ) any parameterized objective function, and p(θ) k ) a probability distribution over a parameter θ k , which is maximized due to the negative log-likelihood.

[0083] Fig. Figure 7c illustrates the retraining of the pre-trained machine learning model 47 in a retraining stage 58. Using the adapted simulation parameters 56, the simulation 44 of the inspection system 50 is applied to generate large quantities of training images 33 from the designs 32 of the photolithography masks. The training images 33 are used to retrain the pre-trained machine learning model 47 for defect detection. The pre-training images 31 and the training images 33 are preferably based at least partially on the same designs. However, they can also be based on different designs. Since the training images 33 are simulated, they can be generated automatically with short computation times and require little user effort. The generated training images 33 can target specific defects 22 or design types. In this way, it is possible to retrain the machine learning model with reference to specific defect types or design types.Alternatively, the generated training images 33 can systematically cover a range of defect types or design types, enabling efficient and rapid retraining of the machine learning model 28.

[0084] At least some of the training images 33 generated in step c. include simulated images of defective designs 34 and corresponding defect annotations 36, wherein the simulated images are obtained by applying the simulation 44 of the inspection system 50 to the defective designs 34. 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 that already contains a defect can be used. Using the simulation 44 of the inspection system 50 after adjusting the parameters, an image of the defective design 34 can be simulated. Preferably, at least some of the simulated training images 33 include defect annotations 36. Since the defects 22 are artificially generated, the corresponding defect annotations 36 can be easily incorporated into the training images 33.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 a 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, extent, and / or type of a defect, etc., e.g., in the form of bounding boxes, pixel annotations, defect classes, etc. The training images 33 can also include captured images with or without defect annotations 36. The training images 33 can also include captured or simulated images without defects.

[0085] The generation of pre-training and training images, as well as the pre-training and retraining of the machine learning model, can be performed, for example, on a single computer, a cluster, a distributed system, or in the cloud. For this purpose, the use of powerful hardware is advantageous; for instance, GPUs or TPUs can be used to accelerate the algorithms. Ample RAM and storage space with fast I / O are also beneficial for reducing computation times. The hardware does not need to be physically connected to the inspection system, but this can be advantageous.

[0086] Pretraining the machine learning model can be performed in an environment that differs significantly from the environment used for retraining the machine learning model. For example, pretraining the machine learning model can be performed in a development environment at a primary location using large computer clusters, dedicated hardware, or even cloud instances. In contrast, retraining the machine learning model can be performed in an application environment, such as at a secondary location with reduced hardware or requiring specific confidentiality standards, for example, at a customer site.Using the method according to the invention, the pre-training of the machine learning model can be carried out thoroughly using large quantities of pre-training images in the development environment, whereas the retraining of the machine learning model can be carried out using only very few captured adaptation images, simulated training images and optionally a small number of captured training images in the application environment.

[0087] The previously described procedure for retraining a machine learning model for defect detection under modified conditions can be performed iteratively. Each time conditions are modified, the machine learning model requires retraining, for example, each time a different photolithography mask is loaded by the inspection system. For this purpose, steps a. to d. can be repeated. Instead of retraining the machine learning model from scratch or from the pre-trained machine learning model 47, it may be advantageous to use an already retrained machine learning model as a pre-trained machine learning model to incorporate further modified conditions, as this may require fewer adjustments. For example, if the machine learning model was pre-trained at a first location, e.g., a production facility, and then transferred to a second location, e.g.,If a machine learning model that has already been retrained is sent to a consumer, fewer adjustments might be required if a previously retrained model is used for further retraining than if the pre-trained model sent from the first location is used for retraining.

[0088] An inspection system 50 for detecting defects 22 in an image 26 of a photolithography mask 14 according to a third embodiment of the invention is in Fig. Figure 8 illustrates the inspection system 50, which comprises: an image acquisition unit 60 configured to acquire images 26 of photolithography masks 14; and 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 according to an embodiment of the invention.

[0089] The image acquisition unit 60 delivers the image 26 to the data analysis device 62. The processor 66 can be implemented, for example, as a CPU, GPU, or TPU. The processor 66 can receive the image 26 via an interface 68. The processor 66 can load program code from a memory 64, for example, program code for executing a method for detecting defects 58 according to an embodiment of the invention as described above. The processor 66 can execute the program code.

[0090] A system for retraining a machine learning model for defect detection in an image of a photolithography mask according to a fourth embodiment of the invention comprises the following: an image acquisition unit configured to acquire images of photolithography masks; and a data analysis device comprising at least one memory and at least one processor configured to perform the steps of a method for retraining a machine learning model for defect detection in an image of a photolithography mask according to an embodiment of the invention. The image acquisition unit can be used to acquire fitting images and optionally training images. The data analysis device receives a pre-trained machine learning model, which is retrained as described above.

[0091] Any reference throughout this document to "an 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" at various points throughout this document do not necessarily refer to the same embodiment, example, or aspect, but may. Furthermore, the particular features or characteristics may be combined in any suitable manner in one or more embodiments, as would be apparent to a person skilled in the art from this disclosure.

[0092] While some embodiments, examples or aspects described herein include some features that are included in other embodiments, examples or aspects, but not others, combinations of features of different embodiments, examples or aspects shall also be within the scope of protection of the claims and constitute different embodiments, as would be understood by a person skilled in the art.

[0093] The following clauses relate to preferred embodiments of the invention: 1. A method 24 for retraining a pre-trained machine learning model 28 for defect detection in an image 26 of a photolithography mask 14, wherein the method comprises the following: a. Capturing adaptation images 52 of one or more photolithography masks 14 using an inspection system 50, wherein the inspection system 50 is configured to capture images 26 of photolithography masks 14; b. Adjusting at least one parameter of a simulation 44 of the inspection system 50 using the adjustment images 52; c. Generating training images 33 of one or more photolithography masks 14 using the adapted simulation 44 of the inspection system 50; and d. Retraining the machine learning model 28 for defect detection using the generated training images 33. 2. The procedure according to clause 1, wherein the adjustment of at least one parameter of the simulation 44 of the inspection system 50 in step b. includes solving an optimization problem. 3. The method according to clause 2, wherein for at least one matching image 52 a design 32 of the corresponding photolithography mask 14 is available, wherein the simulation 44 of the inspection system 50 simulates the image acquisition process for a photolithography mask 14 using a design 32 of the photolithography mask 14, and wherein solving the optimization problem comprises minimizing a deviation of one or more matching images 52 from the simulated images of the corresponding designs 32. 4. The procedure according to clause 2 or 3, wherein solving the optimization problem includes maximizing the similarity of the distribution of fitting images 52 and a distribution of simulated images obtained using the simulation 44 of the inspection system 50. 5. The method according to one of clauses 2 to 4, wherein gradients of the simulation 44 of the inspection system 50 are derived with respect to the at least one parameter and wherein solving the optimization problem includes using a gradient descent approach using the derived gradients. 6. The method according to any of the preceding clauses, further comprising, prior to step a., pre-training the machine learning model 28 using pre-training images 31, which include at least one simulated image generated using the simulation 44 of the inspection system 50. 7. The method according to any of the preceding clauses, further comprising, prior to step a., pre-training the machine learning model 28 using pre-training images 31, which include at least one image that is captured by an inspection system 50. 8. The procedure according to one of the preceding clauses, wherein at least some training images include 33 defect annotations 36. 9. The method according to any of the preceding clauses, wherein the simulation 44 of the inspection system 50 simulates the image acquisition process for a photolithography mask 14. 10. The procedure according to clause 9, wherein the simulation 44 of the inspection system 50 comprises simulating the image acquisition process using a design 32 of the photolithography mask 14. 11. The method according to clause 10, wherein the generated training images 33 comprise simulated images of defective designs 34 and / or designs containing unusual structures and corresponding defect annotations 36, wherein the simulated images are obtained by applying the simulation 44 of the inspection system 50 to the defective designs 34 and / or the designs containing unusual structures. 12. The method according to one of the preceding clauses, wherein prior knowledge is used to adjust at least one parameter of the simulation of the inspection system in step b. 13. The method according to one of the preceding clauses, wherein the majority of the adaptation images 52 of pre-training images 31 used to pre-train the machine learning model 28 differ in at least one aspect from the group comprising the inspection system, the image acquisition time period, the photolithography mask. 14. The method according to any of the preceding clauses, wherein the pretraining 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 according to any of the preceding clauses, wherein the number of adaptation 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 particularly preferably less than 0.001%. 16. The procedure according to one of the preceding clauses, repeating steps a. to d. 17. The method according to any of the preceding clauses, wherein the simulation 44 of the inspection system 50 comprises a physical simulation of the propagation of electromagnetic waves within a photolithography mask. 18. The procedure according to one of the preceding clauses, wherein the simulation 44 of the inspection system 50 comprises the application of a trained machine learning model 38. 19. A method for detecting defects 22 in an image 26 of a photolithography mask 14 in an inspection system 50, wherein the method comprises the use of a machine learning model 28 for defect detection which has been retrained according to any of the preceding clauses. 20. An inspection system 50 for detecting defects 22 in an image 26 of a photolithography mask 14, comprising the following: a. an image acquisition unit 60 configured to acquire images 26 from photolithography masks 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 photolithography mask 14 according to clause 19.

[0094] In summary, the invention relates to a method for retraining a pre-trained machine learning model 28 for defect detection in an image of a photolithography mask, wherein the method comprises: acquiring adaptation images 52 of one or more photolithography masks using an inspection system 50; adapting at least one parameter of a simulation 44 of the inspection system 50 using the adaptation images 52; generating training images 33 of one or more photolithography masks using the adapted simulation 44 of the inspection system 50; and retraining the machine learning model 28 for defect detection using the generated training images 33 for use of the retrained machine learning model 28 for defect detection in an image of a photolithography mask acquired by the inspection system 50. The invention also relates to a method and an inspection system for defect detection. Reference symbol list 10, 10' photolithography system 12. Radiation source 14 Photolithography mask 16 Lighting optics 17 Projection optics 18 wafer level 19 Projection section 20 images 22 Defect 24 procedures Image 26 28 Machine learning model 30 training images 31 pre-training images 32 Design 33 training images 34 Defective Design 36 Defect annotation 38 Machine learning model 40 Pre-training level 42 Initial simulation parameters 44 Simulation 46 Pre-training step 48 Adjustment level 50 inspection system 52 Adjustment image 54 Parameter adjustment step 56 Adapted simulation parameters 58 New training level 60 image capture units 62 Data analysis device 64 memory 66 processor 68 interface QUOTES INCLUDED IN THE DESCRIPTION

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

[0000] DE 10 2023 104 378.1

[0061] Cited non-patent literature

[0000] SegFormer: Simple and Efficient Design for Semantic Segmentation with Transformers", Enze Xie, Wenhai Wang, Zhiding Yu, Anima Anandkumar, Jose M. Alvarez, Ping Luo, arXiv 2105.15203, 2021

[0061]

Claims

[1] Method (24) for defect detection in an image (26) of a photolithography mask (14) using a pre-trained machine learning model (28), wherein the method comprises: a. Capturing adaptation images (52) of one or more photolithography masks (14) using an inspection system (50), wherein the inspection system (50) is configured to capture images (26) of photolithography masks (14); b. Adjusting at least one parameter of a simulation (44) of the inspection system (50) using the fitting images (52); c. Generating training images (33) of one or more photolithography masks (14) using the adapted simulation (44) of the inspection system (50); d. Retraining the machine learning model (28) for defect detection using the generated training images (33); e. Applying the newly trained machine learning model (44) for defect detection to an image (26) of a photolithography mask (14) captured by the inspection system (50). [2] Method according to claim 1, wherein the adjustment of the at least one parameter of the simulation (44) of the inspection system (50) in step b. comprises solving an optimization problem. [3] Method according to claim 2, wherein for at least one matching image (52) a design (32) of the corresponding photolithography mask (14) is available, wherein the simulation (44) of the inspection system (50) simulates the image acquisition process for a photolithography mask (14) using a design (32) of the photolithography mask (14) and wherein solving the optimization problem comprises minimizing a deviation of one or more matching images (52) from the simulated images of the corresponding designs (32). [4] Method according to claim 2 or 3, wherein solving the optimization problem comprises maximizing the similarity of the distribution of fitting images (52) and a distribution of simulated images obtained using the simulation (44) of the inspection system (50). [5] Method according to one of claims 2 to 4, wherein gradients of the simulation (44) of the inspection system (50) are derived with respect to the at least one parameter and wherein solving the optimization problem comprises using a gradient descent approach using the derived gradients. [6] Method according to one of the preceding claims, further comprising, prior to step a., pre-training the machine learning model (28) using pre-training images (31) comprising at least one simulated image generated using the simulation (44) of the inspection system (50). [7] Method according to one of the preceding claims, further comprising, prior to step a., pre-training the machine learning model (28) using pre-training images (31) comprising at least one image that is captured by an inspection system (50). [8] Method according to any of the preceding claims, wherein at least some training images (33) include defect annotations (36). [9] Method according to any of the preceding claims, wherein the simulation (44) of the inspection system (50) simulates the image acquisition process for a photolithography mask (14). [10] Method according to claim 9, wherein the simulation (44) of the inspection system (50) comprises simulating the image acquisition process using a design (32) of the photolithography mask (14). [11] Method according to claim 10, wherein the generated training images (33) comprise simulated images of defective designs (34) and corresponding defect annotations (36), wherein the simulated images are obtained by applying the simulation (44) of the inspection system (50) to the defective designs (34). [12] Method according to one of the preceding claims, wherein prior knowledge is used to adapt the at least one parameter of the simulation of the inspection system in step b. [13] Method according to one of the preceding claims, wherein the majority of the adaptation images (52) of pre-training images (31) used to pre-train the machine learning model (28) differ in at least one aspect from the group comprising the inspection system, the image acquisition time period, the photolithography mask. [14] Method according to any of the preceding claims, wherein the pretraining 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] Method according to any of the preceding claims, wherein the number of adaptation 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 particularly preferably less than 0.001%. [16] Method according to any of the preceding claims, wherein steps a. to d. are repeated. [17] Method according to one of the preceding claims, wherein the simulation (44) of the inspection system (50) comprises a physical simulation of the propagation of electromagnetic waves within a photolithography mask. [18] Method according to one of the preceding claims, wherein the simulation (44) of the inspection system (50) comprises the application of a trained machine learning model (38). [19] Inspection system (50) for detecting defects (22) in an image (26) of a photolithography mask (14), comprising the following: a. an image acquisition unit (60) configured to acquire images (26) of photolithography masks (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 photolithography mask (14) according to one of the preceding claims.

Images