Method for adjusting the focal plane of an image of an object comprising integrated circuit patterns

Abstract

The invention relates to a method for adjusting the focal plane of an image of an object comprising integrated circuit patterns, the method comprising: acquiring an input image (20) of an object comprising integrated circuit patterns using an inspection system; and applying a machine learning model (26) to the input image (20), wherein the machine learning model (26) is trained to adjust a focal plane of an input image (20) of an object comprising integrated circuit patterns. The invention also relates to methods for detecting defects (22) using such methods, methods for training a corresponding machine learning model, and a corresponding system for detecting defects (22).

Description

Field of the invention

[0001] The invention relates to methods and systems for adjusting the focal plane of images of objects comprising integrated circuit patterns obtained by an inspection system, in particular to methods using a machine learning model. The methods and systems can be used for image enhancement, defect detection, quality control, and process monitoring for objects comprising integrated circuit patterns, e.g., photolithography masks, reticles, or wafers. Background of the invention

[0002] Semiconductor manufacturing involves precise manipulation, such as etching, of materials such as silicon or oxide at very fine scales in the nanometer range. Therefore, a quality management process that includes quality assurance and quality control is important to ensure high quality standards of the manufactured wafers. Quality assurance refers to a set of activities designed to ensure 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 part of the quality assurance process.

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

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

[0005] Due to the growing integration density in the semiconductor industry, photolithography masks must image ever smaller structures on wafers. The aspect ratio and number of layers of integrated circuits are constantly increasing, and the structures are growing into the third (vertical) dimension. The current height of memory stacks is more than a dozen micrometers. In contrast, the feature size is becoming smaller. The minimum feature size or critical dimension is below 10 nm, for example, 7 nm or 5 nm, and will approach feature sizes below 3 nm in the near future. As the complexity and dimensions of semiconductor structures grow into the third dimension, the lateral dimensions of integrated semiconductor structures are becoming smaller. To create the small feature dimensions imaged on the wafer, photolithographic masks or templates for nanoimprint photolithography with increasingly smaller structures or pattern elements are required.The production process of photolithographic masks and templates for nanoimprint photolithography is therefore becoming increasingly complex and, consequently, more time-consuming and ultimately more expensive. With the advent of EUV photolithography scanners, the type of masks changed from transmission-based to reflection-based patterning.

[0006] Due to the tiny feature sizes of the pattern elements of photolithography masks or templates, it is not possible to exclude defects during mask or template 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 structure fabrication, photolithography-related defects represent almost half of the number of defects. Therefore, the inspection, verification, and metrology of photolithography masks play an important role in semiconductor process control to monitor systematic defects. Defects detected during quality assurance processes can be used for root cause analysis, for example, to modify or repair the photolithography mask. The defects can also serve as feedback to optimize the process parameters of the manufacturing process, e.g.exposure time, focus variation, etc.

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

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

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

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

[0011] Deep learning is a class of machine learning techniques that uses artificial neural networks with numerous hidden layers between the input and output layers. Due to this complex internal structure, the networks are also capable of continuously extracting higher-level features from the raw input data. Each layer learns to transform its input data into a somewhat more abstract and compositional representation, and low- and high-level insights are derived from the training data accordingly. The hidden layers can have different sizes and tasks, such as convolutional layers or pooling layers.

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

[0013] The die-to-die principle compares an acquired image of an object with another acquired image of an identical or similar object, e.g., the same section or a different section, or an object containing similar structures. Detected deviations are treated as defects.

[0014] The die-to-database principle compares an image of an object with a reference image from a database, e.g., a simulated image, a golden reference, or a model of the object, such as a CAD model, thereby detecting deviations from the ideal data. Unexpected patterns in the image are detected due to large differences and are treated as defects.

[0015] Both die-to-die and die-to-database approaches rely on comparing images, such as acquired images and / or simulated images. However, detecting defects with high accuracy and sensitivity requires high-quality images, which can be difficult to generate, especially for EUV systems. Image quality is strongly influenced by the focal plane of the images. For example, images could be defocused due to inaccurate placement of the inspection system's stage during image acquisition. Furthermore, some patterns may require pattern-dependent focus, which is difficult to implement using autofocus and can easily result in defocused portions of the image. Local height profile deviations of the photolithography mask as well as inaccurate stage movements can also lead to defocus of the acquired images.Apart from the focal plane, image quality can be further affected, e.g., by noise, contrast or brightness variations, or shadowing effects due to different EUV beam configurations.

[0016] Therefore, one object of the invention is to improve the quality of images of objects comprising integrated circuit patterns. In particular, one object of the invention is to automatically adjust the focal plane of an image of an object comprising integrated circuit patterns with high accuracy. It is also one object of the invention to adjust the focal plane of an image of an object comprising integrated circuit patterns while requiring low computation times. Another object of the invention is to improve the accuracy and sensitivity of defect detection methods for objects comprising integrated circuit patterns.

[0017] The objects are achieved by the invention specified in the independent claims. Advantageous embodiments and further developments of the invention are specified in the dependent claims. Brief description of the invention

[0018] Embodiments of the invention relate to methods and systems for improving the image quality, in particular the focal plane, of an image of an object comprising integrated circuit patterns.

[0019] A first embodiment includes a computer-implemented method for adjusting the focal plane of an image of an object comprising integrated circuit patterns, the method comprising: acquiring an input image of an object comprising integrated circuit patterns using an inspection system; and applying a machine learning model to the input image, the machine learning model trained to adjust a focal plane of an input image of an object comprising integrated circuit patterns.

[0020] By adjusting the focal plane of the captured input image, image quality is improved without having to acquire another image of the same object with a different focal plane. This reduces processing time. Furthermore, it is often not possible to acquire a second image of an object during the production cycle because the object moves on to a different stage. By using a machine learning model to adjust the focal plane of the input image, the focal plane is adjusted automatically. Furthermore, the focal plane is adjusted quickly because the inference step of a machine learning model typically requires only a single forward pass of the input data.Finally, using a machine learning model to adjust the focal plane results in highly accurate focus-adjusted images, as machine learning models derive all important information directly from training data instead of using rules defined by humans, which typically do not cover special cases or do not use all the underlying information available for the task.

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

[0022] The term "inspection system" refers to a system configured to inspect objects comprising integrated circuit patterns, e.g., photolithography masks or wafers, by capturing an image of the object and detecting defects in the image. In the case of a photolithography mask, an inspection system may be, for example, a mask verification system or a mask repair system. In particular, inspection systems include actinic photolithography mask inspection systems. In the case of a wafer, an inspection system may be, for example, a bright-field or dark-field microscopy tool, an electron beam wafer inspection tool, or a focused ion beam scanning electron microscopy (FIB-SEM) tool.

[0023] If the object is a photolithography mask, the photolithography mask may have an aspect ratio between 1:1 and 1:4, preferably between 1:1 and 1:2, more preferably 1:1 or 1:2. The photolithography mask may have a nearly rectangular shape. The photolithography mask may preferably be 5 to 7 inches long and wide, more preferably 6 inches long and wide. Alternatively, the photolithography mask may be 5 to 7 inches long and 10 to 14 inches wide, preferably 6 inches long and 12 inches wide.

[0024] An image, or input image, or reference image of an object comprising integrated circuit patterns can refer to various types of images of the object, e.g., a two-dimensional image, a three-dimensional image, a stack of images, or a volumetric three-dimensional image, which can be processed, for example, slice by slice. An image can be acquired using an inspection system, or it can be simulated, e.g., from a design of the photolithography mask. Images can be of various modalities, e.g., structured electron microscopy (SEM) images, aerial images, optical images, X-ray images, computed tomography (CT) images, focused ion beam (FIB) images, atomic force microscopy (AFM) images, ultrasound images, or multimodal images, e.g., acquired using a combination of X-ray imaging and SEM.An image of an object comprising integrated circuit patterns may image the entire object or one or more subsections thereof.

[0025] The input image may include an image acquired using the inspection system and / or the reference image.

[0026] An aerial image of a photolithography mask indicates the radiation intensity distribution of the photolithography system in a wafer plane. Accordingly, the aerial image simulates the structures on the surface of a wafer when the wafer is printed using the photolithography mask in the inspection system. A wafer plane refers to a plane within the photoresist on the wafer in a photolithography system. An aerial image can be acquired using an inspection system and a photolithography mask. An aerial image can be simulated using a photolithography mask design and an aerial image simulation method, such as the rigorous coupled wave analysis (RCWA) method or the thin element approximation (TEA) method.

[0027] According to an example of the first embodiment, the machine learning model is trained to adjust the focal plane of the input image to a specified target focal plane. Instead of adjusting the focal plane to optimize some image quality metric, a single target focal plane can be predefined for all input images, or a target focal plane can be specified separately for each input image. In this way, the machine learning model can be flexibly adapted to different application requirements. For example, a user can specify the target focal plane, or images with a desired defocus can be generated, e.g., for training purposes.

[0028] According to one aspect of the example, the focal plane of a reference image is determined, and the target focal plane is derived from the focal plane of the reference image. In this way, the target focal plane of the input image can be adjusted with respect to the focal plane of a reference image, thereby improving, for example, the comparability of the input image and the reference image or the image quality of the input image.

[0029] According to one aspect of the example, the target focal plane is provided as an additional input to the machine learning model. This improves the accuracy and applicability of the machine learning model.

[0030] In one example of the first embodiment, the focal plane of the input image is adjusted using an image quality metric. In this way, the focal plane can be adjusted automatically with reference to any predefined image quality metric to optimize image quality.

[0031] In an example of the first embodiment, the focal plane of the input image is provided as an additional input to the machine learning model. Measuring a focal plane of an input image is much simpler than modifying a focal plane of an input image. Accordingly, valuable additional information about the input image is exploited and provided to the machine learning model, simplifying training and increasing the accuracy of the predicted focus-adjusted images.

[0032] According to one example, a photolithography mask design is provided as an additional input to the machine learning model. The design helps eliminate ambiguities and accordingly improves the prediction accuracy of the machine learning model.

[0033] In one example, the machine learning model is trained to remove additional image quality degradations from the input image, such as focus-specific misalignment. This further improves the quality of the predicted focus-adjusted images.

[0034] According to an example of the first embodiment of the invention, the machine learning model comprises an encoder-decoder architecture, in particular a U-Net architecture. An encoder-decoder architecture, such as a U-Net, transforms an input image into an output image using an encoder and a decoder, thereby extracting the most relevant information for the task. Accordingly, the accuracy of the predicted focus-adjusted images is improved.

[0035] According to one aspect of the example, the focal plane of the input image is provided as an additional input to the encoder-decoder architecture in one of its hidden layers, in particular in one of the encoder layers, in particular in the bottleneck. This enables particularly simple and efficient processing of the additional information compared to adding additional information to the input of the machine learning model.

[0036] According to one aspect of the example, a target focal plane of the input image is provided as an additional input to the encoder-decoder architecture in one of its hidden layers, in particular in one of the decoder layers, in particular in the bottleneck. This enables particularly simple and efficient processing of the additional information compared to adding additional information to the input of the machine learning model.

[0037] In an example of the first embodiment of the invention, the machine learning model comprises a conditional diffusion model that sequentially reverses a stochastic process and is trained to increase a focal plane of the input image in each stochastic process step or to decrease a focal plane of the input image in each stochastic process step. Due to their iterative nature, diffusion models can perform complex tasks in a highly accurate manner and are more robust than one-step approaches that perform image-to-image transformations in a single step. Due to the use of conditional information, conditional diffusion models yield more accurate and task-specific results than standard diffusion models.

[0038] According to a preferred example of the first embodiment of the invention, a reference image of the object comprising integrated circuit patterns is provided as an additional input to the machine learning model, and the machine learning model is trained to adjust the focal plane of the input image and / or the focal plane of the reference image so that the focal planes approximately match, or so that the focal planes match. Accordingly, only the focal plane of the input image is adjusted to (approximately) match the focal plane of the reference image, or only the focal plane of the reference image is adjusted to (approximately) match the focal plane of the input image, or both the focal plane of the input image and the focal plane of the reference image are adjusted to (approximately) the same focal plane, which is different from the focal plane of the input image and the focal plane of the reference image.In this way, the focal planes of the input image and the reference image can be simultaneously aligned and optimized, for example, using an image quality metric. This improves the comparability of both images, e.g., in a die-to-die or die-to-database defect detection process. At the same time, image quality can be optimized.

[0039] An alternative method for adjusting the focal plane of an image of an object comprising integrated circuit patterns according to the first embodiment of the invention comprises: acquiring an input image of an object comprising integrated circuit patterns using an inspection system; determining the focal plane of the acquired input image; selecting a machine learning model from a set of machine learning models using the determined focal plane, wherein each machine learning model of the set of machine learning models for adjusting the focal plane of an input image of an object comprising integrated circuit patterns is trained for an input image having a focal plane within a specific focal plane interval; applying the selected machine learning model to the input image.By training a set of machine learning models that are specialized with respect to specific focal plane intervals, the machine learning models are easier to train and produce predictions with higher accuracy because the learning task is less complex due to the specialization.

[0040] An alternative method for adjusting the focal plane of an image of an object comprising integrated circuit patterns according to the first embodiment of the invention comprises: acquiring an input image of an object comprising integrated circuit patterns using an inspection system; applying two or more machine learning models from a set of machine learning models to the acquired input image, resulting in a set of focus-adjusted images, each machine learning model of the set of machine learning models for adjusting the focal plane of an input image of an object comprising integrated circuit patterns being trained for an input image having a focal plane within a specific focal plane interval; selecting a focus-adjusted image from the set of focus-adjusted images using an image quality measure.By training a set of machine learning models specialized with respect to specific focal plane intervals, the machine learning models are easier to train and yield predictions with higher accuracy because the learning task is less complex due to the specialization. The image quality of the focus-adjusted image is also improved due to the selection with respect to the image quality metric.

[0041] A method for detecting defects in an image of an object comprising integrated circuit patterns according to a second embodiment of the invention comprises: acquiring an input image of an object comprising integrated circuit patterns using an inspection system; adjusting the focal plane of the input image using a method according to any one of the preceding claims; and applying a defect detection method to the focus-adjusted input image to detect defects. Preferably, the defect detection method comprises a machine learning model for defect detection. Due to the focus adjustment of the input image, the quality of the input image is improved, and defects can be detected with higher accuracy and sensitivity.

[0042] The term "defect" refers to a localized deviation of an integrated circuit pattern from an a priori-defined norm of the integrated circuit pattern. For example, a defect in an integrated circuit pattern, e.g., a semiconductor structure, can lead to a malfunction of an associated semiconductor device. Depending on the detected defect, for example, the photolithography process can be improved, or photolithography masks or wafers can be repaired or discarded. The norm of the structure or pattern can be defined by one or more corresponding reference photolithography masks or reference data sets, e.g., design data sets, simulated data sets, or acquired defect-free data sets.

[0043] A method for detecting defects in an image of an object comprising integrated circuit patterns according to a second embodiment of the invention comprises: acquiring an input image of an object comprising integrated circuit patterns using an inspection system; obtaining a reference image of the object comprising integrated circuit patterns; adjusting the focal plane of the input image and / or the focal plane of the reference image so that the focal planes approximately match or so that the focal planes match, using a method according to claim 13 or 14; and applying a defect detection method to the focus-adjusted input image and the focus-adjusted reference image.Due to the focus adjustment of the input image and the reference image, the image quality is comparable and improved, and defects can be detected with greater accuracy and sensitivity in a die-to-die or die-to-database approach. The reference image can, for example, refer to a captured image of the same or a different object containing integrated circuit patterns (die-to-die), a simulated image of the object containing integrated circuit patterns, e.g., from a CAD model, or a golden reference (die-to-database).

[0044] According to a third embodiment of the invention, a computer-implemented method for training a machine learning model for adjusting the focal plane of an image of an object comprising integrated circuit patterns, preferably a machine learning model according to the second embodiment of the invention, comprises: obtaining training data comprising pairs of defocused images of objects comprising integrated circuit patterns and corresponding focused images obtained by adjusting the focal plane of the defocused images; and training the machine learning model for adjusting the focal plane of an input image of an object comprising integrated circuit patterns using the training data. In this way, the training of the machine learning model can be achieved efficiently and yields highly accurate predictions. The term "focused image" refers to an image without defocus, i.e.An image in focus, or an image whose focal plane corresponds to a target focal plane. The term "defocused image" refers to an image with a defocus, whose focal plane is not optimal with respect to any image quality measure or does not correspond to a target focal plane.

[0045] According to an example of the third embodiment of the invention, the machine learning model is trained by minimizing a loss function comprising an adaptation loss function that measures the deviation of focus-adjusted images predicted by the machine learning model when presented with defocused images of the training data from the corresponding focused images of the training data. In this way, the prediction accuracy of the trained machine learning model is improved.

[0046] According to one example or aspect, the machine learning model is trained by minimizing a loss function that includes an adversarial loss function that measures the distance between a distribution of focus-matched images predicted by the machine learning model when presented with defocused images of the training data and a distribution of the corresponding focused images. In this way, the machine learning model learns to imitate focus-matched images with high accuracy, thereby improving the prediction accuracy of the trained machine learning model and the training speed.

[0047] In one example, the machine learning model is trained by minimizing a loss function that includes an adaptation loss function and an adversarial loss function. Thus, the focus-adapted images exhibit an appearance of captured focus-adapted images according to the learned distribution of focus-adapted images.

[0048] A joint loss function can be used to train a machine learning model for focus adjustment and a machine learning model for defect detection according to the second embodiment of the invention.

[0049] A computer-implemented method for training a machine learning model for adjusting the focal plane of an image of an object comprising integrated circuit patterns and a machine learning model for detecting defects in an image of an object comprising integrated circuit patterns according to the second embodiment of the invention uses a joint loss function for jointly training the machine learning model for adjusting the focal plane of an image and the machine learning model for detecting defects.

[0050] For example, the machine learning model for focus adaptation and the machine learning model for defect detection can be trained by minimizing a joint loss function that includes an adaptation loss function and / or an adversarial loss function and a defect detection loss function. In this way, the focus is adapted as if by the focused images and / or the appearance of the focus-adapted images follows the learned distribution. At the same time, the focus-adapted images are well suited for defect detection, and the machine learning model for defect detection is adapted to the focus-adapted images.

[0051] A fourth embodiment of the invention relates to a computer-readable medium on which a computer program executable by a computing device is stored, wherein the computer program comprises code for carrying out a method according to the third embodiment of the invention.

[0052] A fifth embodiment of the invention relates to a computer program product comprising instructions which, when the program is executed by a computer, cause the computer to carry out a method according to the third embodiment of the invention.

[0053] An inspection system for detecting defects in an object comprising integrated circuit patterns according to a sixth embodiment of the invention comprises: an imaging device configured to provide an image of the object comprising integrated circuit patterns; one or more processing devices; one or more machine-readable hardware storage devices comprising instructions executable by one or more processing devices to perform operations comprising one of the methods for detecting defects in an image of an object comprising integrated circuit patterns according to the second embodiment of the invention.

[0054] 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. Short description of the drawings Fig. 1 illustrates an exemplary transmission-based photolithography system, e.g., a deep ultraviolet (DUV) photolithography system; Fig. 2 illustrates an exemplary reflection-based photolithography system, e.g., an extreme ultraviolet (EUV) photolithography system; Fig. 3 shows an image of an object comprising integrated circuit patterns in the form of a photolithography mask comprising a defect; Fig. 4 is a flowchart illustrating the steps of a method for adjusting the focal plane of an image of an object including integrated circuit patterns according to the first embodiment of the invention; Fig. 5 illustrates the method for adjusting the focal plane of an image of an object including integrated circuit patterns according to the first embodiment of the invention; Fig. Figure 6 illustrates a machine learning model in the form of an encoder-decoder structure, here a U-Net, used to adjust the focal plane of an image of an object comprising integrated circuit patterns and which can process additional information provided; Fig. 7 illustrates a variant of the machine learning model for focus adjustment that receives an input image and a reference image and adjusts the focus plane of the input image to that of the reference image, or that adjusts the focus plane of the reference image to that of the input image, or that adjusts both the focus plane of the input image and the focus plane of the reference image to the same focus plane; Fig. 8 illustrates a machine learning model in the form of a conditional diffusion model for adjusting the focal plane of an image of an object comprising integrated circuit patterns; Fig. 9 illustrates a flowchart of a method for adjusting the focal plane of an image of an object including integrated circuit patterns using a set of specialized machine learning models; Fig. Figure 10 shows a normal density function approximating statistics on focal plane frequencies used to define focal plane intervals; Fig. 11 illustrates a flowchart of a method for adjusting the focal plane of an image of an object comprising integrated circuit patterns using a set of specialized machine learning models without determining the focal plane of the input image; Fig. 12 shows a flowchart of a method for detecting defects in an image of an object comprising integrated circuit patterns according to a second embodiment of the invention; Fig. 13 shows a flowchart of a method for detecting defects in an image of an object comprising integrated circuit patterns using a reference image according to a second embodiment of the invention; and Fig. Figure 14 illustrates the construction of a suitable loss function for training a machine learning model for focus adaptation, in particular an encoder-decoder CNN; Fig. Figure 15 illustrates the construction of a suitable loss function for jointly training a machine learning model for focus adjustment and a machine learning model for defect detection; Fig. 16 illustrates the application of a trained machine learning model for focus adjustment and a trained machine learning model for defect detection to a defocused image; Fig. 17 illustrates a system for detecting defects in an object comprising integrated circuit patterns according to a sixth embodiment of the invention. Detailed description

[0055] Advantageous exemplary embodiments of the invention are described below and schematically shown in the figures. In the figures and in the description, the same reference numerals are used to describe identical features or components. Dashed lines indicate optional features.

[0056] The methods and systems described herein can be used to inspect images from a variety of photolithography systems, such as transmission-based photolithography systems 10 or reflection-based photolithography systems 10'.

[0057] Fig. 1 illustrates an exemplary transmission-based photolithography system 10, e.g., a DUV photolithography system. Main components are a radiation source 12, which may be a deep ultraviolet (DUV) excimer laser source, imaging optics that define, for example, partial coherence and that may include optics that shape radiation from the radiation source 12, a photolithography mask 14, illumination optics 16 that illuminate the photolithography mask 14, and projection optics 17 that project 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 impinge on the wafer plane 18, with the largest possible angle defining 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 impinge on the wafer plane 18.

[0058] In this document, the terms “radiation” or “ray” are used to include all types of electromagnetic radiation, including ultraviolet radiation (e.g., having a wavelength of 365, 248, 193, 157, or 126 nm) and EUV (extreme ultraviolet radiation, e.g., having a wavelength in the range of approximately 3-100 nm).

[0059] 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 photolithography mask 14. The projection optics 17 may include optical components for shaping, adjusting, and / or projecting the radiation after the radiation has passed through the photolithography mask 14. The illumination optics 16 excludes the light source 12, and the projection optics excludes the photolithography mask 14.

[0060] The illumination optics 16 and the projection optics 17 may comprise various types of optical systems, including, for example, refractive optics, reflective optics, apertures, and catadioptric optics. The illumination optics 16 and the projection optics 17 may also include components that operate according to any of these design types to collectively or individually direct, shape, or control the projection beam of radiation.

[0061] Fig. Figure 2 illustrates an exemplary reflection-based photolithography system 10', e.g., a lithography system for extreme ultraviolet (EUV) light. Main components are a radiation source 12, which may be a laser plasma light source, illumination optics 16, which may, for example, define partial coherence and may 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 impinge on the wafer plane 18, with the largest possible angle defining 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 impinge on the wafer plane 18.

[0062] Fig. Figure 3 illustrates an image 20 of a photolithography mask 14 containing a defect 22. According to the techniques described herein, various imaging modalities can be used to acquire the image 20. The images 20 can include single-channel images or multi-channel images, e.g., focus stacks. For example, the image may include 2D images. A multi-beam scanning electron microscope (mSEM) may be employed. mSEM employs multiple beams to simultaneously acquire images in multiple fields of view. For example, as many as 50 beams or even as many as 90 beams could be used. Each beam covers a separate portion of a surface of the photolithography mask. This allows a large image to be acquired within a short period of time. Typically, current machines acquire 4.5 gigapixels per second.Other examples of images that include 2D images concern imaging modalities such as optical imaging, phase-contrast imaging, X-ray imaging, etc. It is also possible for the image to be a volumetric 3D dataset that can be processed slice by slice or as a three-dimensional volume. 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 can be an aerial photograph acquired by an aerial measurement system. An aerial photograph is the radiation intensity distribution at the substrate level.It can be used to simulate the radiation intensity distribution generated by the photolithography mask during the photolithography process. The aerial imaging measurement system can be equipped with, for example, a staring array sensor, a line scan sensor, or a time-delayed integration (TDI) sensor.

[0063] Due to inaccuracies in the image acquisition process, such as inaccurate stage movements, variable height profiles of the integrated circuit pattern, or pattern-dependent focuses, the focal plane of the acquired image may be suboptimal. Accordingly, image quality can be improved by adjusting the image's focal plane.

[0064] For this purpose, Fig. 4 illustrates a method for adjusting the focal plane of an image of an object comprising integrated circuit patterns, according to a first embodiment of the invention. The method comprises: acquiring an input image of an object comprising integrated circuit patterns using an inspection system in a step M1; and applying a machine learning model to the input image, the machine learning model trained to adjust a focal plane of an input image of an object comprising integrated circuit patterns, in a step M2.

[0065] Fig. 5 illustrates the computer-implemented method 24 for adjusting the focal plane of an image 20 of an object comprising integrated circuit patterns, according to the first embodiment of the invention. An input image 20 is acquired using an inspection system. The input image 20 is used as input to a machine learning model 26. The machine learning model 26 is trained to adjust the focal plane of the input image 20, resulting in a focus-adjusted image 28. The focus adjustment results in improved contrast of the structures in the image and accordingly improves the detectability of defects 22.

[0066] The method can be used to adjust the focal plane of an image to an optimized focal plane. Optimality can be measured, for example, using any image quality measure. An image quality measure can, for example, include an average contrast of an image, an average edge contrast, a quantity of high frequencies (e.g., frequencies above some defined threshold) in Fourier or wavelet space, a structural similarity index (SSIM), a signal-to-noise ratio (SNR), a response of a machine learning model trained to measure the quality of an image, etc. An image quality measure can also be specifically designed to measure the detectability of one or more specific defect types, in particular compared to defect-free areas or compared to other defect types.Accordingly, the machine learning model can be trained to adjust the focal plane of an input image to improve the detectability of specific defect types, e.g., contrast or brightness.

[0067] The method can also be used to adjust the focal plane of an image to a target focal plane. According to one example, the machine learning model is trained to adjust the focal plane of the input image to a specified target focal plane. The machine learning model can be trained to adjust any input image to the same specified target focal plane. In this case, the majority of training images contain focused images with the target focal plane. The target focal plane can, for example, be specified by a user, it can be obtained from a database, or it can be measured in a reference image. The target focal plane can be an optimized focal plane with respect to some image quality metric. The target focal plane can also be a suboptimal focal plane used to generate defocus images.

[0068] In one example, a reference image is provided, the focal plane of the reference image is determined, and the target focal plane is derived from the focal plane of the reference image. The focal plane of the reference image can be measured, for example. The target focal plane can be set to the focal plane of the reference image or it can be calculated from the focal plane of the reference image using some calculation rule, e.g., twice the focal plane of the reference image. The focal plane of the input image can be approximated or equalized to the focal plane of the reference image, for example. In this way, the comparability of the input image and the reference image is improved. Furthermore, the image quality of the reference image, which strongly depends on the focal plane, can also be retained in the input image.

[0069] The target focal plane can also vary for each input image. It can be used as additional input to the machine learning model, so that the target focal plane is image-specific, i.e., varies with the input image. The target focal plane can, for example, be specified by a user for the input image, it can be obtained from a database, or it can be measured in a reference image. The target focal plane can be an optimized focal plane with respect to some image quality metric. The target focal plane can also be a suboptimal focal plane used to generate defocus images.

[0070] Measuring the focal plane of an image is actually much simpler than adjusting the focal plane of the image. The focal plane of an image can be measured using, for example, an interferometer that measures the distance between the photolithography mask and the sensor. Alternatively, time-of-flight sensors or tactile sensors could be used to measure the distance between the photolithography mask and the sensor. Alternatively, the focal plane of an image can be derived from the contrast of the image using gradients or convolutions with Sobel filters, as described, for example, in US 2022 / 0405903.

[0071] According to one example, a design of a photolithography mask is provided as an additional input 30 to the machine learning model 26. If the object comprising integrated circuit patterns is a photolithography mask, the design of the photolithography mask may be provided. If the object comprising integrated circuit patterns is a wafer, the design of the photolithography mask used to fabricate the wafer may be provided. The design may be used to eliminate ambiguities in a defocused input image. The defocused input image is typically blurred and / or has a lower contrast, resulting in a loss of information. Therefore, there is no clear mapping from a defocused input image to a corresponding focused image. There may be multiple focused images that could result in the same defocused input image.To obtain an unambiguous solution—i.e., a unique focused image from the defocused input image—the photolithography mask design can be used as an additional source of information. For example, if the defocused input image contains a blob—i.e., a round, out-of-focus object that does not exhibit sharp contrast with the background—and the photolithography mask design contains a triangle at the same location, then the focused image should also contain a triangle at that location. Accordingly, the machine learning model can be trained using the photolithography mask design to eliminate ambiguities in the defocused input image.

[0072] The design can be provided, for example, as a rasterized image.

[0073] In a preferred example, a focal plane of the input image 20 is provided as additional input 30 to the machine learning model 26. The focal plane of the input image 20 may, for example, be specified by a user, it may be obtained from a database, or it may be measured in the input image 20.

[0074] Different machine learning models can be used for focus adjustment, such as deep learning models, encoder-decoder architectures, CNNs, U-Nets, transformers, diffusion models, etc. Deep learning models are preferably used as the machine learning model for focus adjustment to achieve high accuracy of the machine learning model's predictions.

[0075] In Fig. 6, the input image 20 is used as input to the machine learning model 26. The machine learning model is a deep learning machine learning model, specifically an encoder-decoder architecture, in this case a U-Net. The U-Net comprises an encoder 34, a decoder 36, and a bottleneck 32. The bottleneck 32 is an interface between the encoder 34 and the decoder 36. Accordingly, it belongs to the encoder 34 and the decoder 36. The encoder 34 maps the input to a code, and the decoder 36 maps the code to an output, here the predicted focus-adjusted image 28. The encoder 34 and the decoder 36 are trained to minimize a difference between the predicted focus-adjusted images 28 and corresponding focused training images.The encoder 34 gradually reduces the dimensionality of the input image 20 down to the bottleneck 32, thereby compressing the information contained in the input image 20 to the most relevant information for the focus adjustment task. The code generated in the bottleneck 32 is a lower-dimensional representation of the input image 20 and can accordingly be viewed as a compressed version of the input image 20. The decoder 36 gradually transforms the code in the bottleneck 32 into the output, i.e., the focus-adjusted image 28. Skip connections 38 allow the decoder 36 to directly access different abstraction levels of the input image, thereby enabling small details of the input image 20 to be preserved in the output.

[0076] In a preferred example, additional information is provided as additional input 30 to the machine learning model 26. Additional information may include a target focal plane, a focal plane of the input image (possibly an inaccurate measurement of the focal plane), a photolithography mask design, image acquisition information such as an image type, a machine type, an acquisition time, a photon count, a stage speed (from which an expected image blur can be derived), or photolithography mask information such as one or more photolithography mask materials, refractive indices, a maximum or minimum feature size, etc. The additional input 30 may be provided at different locations in the machine learning model 26, for example, in the input layer or a hidden layer of a neural network, e.g., an encoder-decoder architecture, e.g., in the bottleneck.

[0077] In a preferred example, the encoder-decoder architecture receives the target focal plane as additional input 30 in one of its hidden layers, in particular in one of the layers of the decoder 36, in particular in the bottleneck 32. The target focal plane can, for example, be incorporated as an additional channel in the bottleneck 32.

[0078] As in Fig. 6, the encoder-decoder architecture in the form of a U-Net receives the target focal plane as additional input 30 in one of its hidden layers, in particular in one of the layers of the encoder 34, in particular in the bottleneck 32. The focal plane can be incorporated, for example, as an additional channel in the bottleneck 32.

[0079] The target focal plane and / or the focal plane of the input image 20 may also be incorporated into the input layer, e.g., by concatenating the input image 20 and the target focal plane and / or the focal plane, or by adding an additional channel to the input image 20 that contains the target focal plane everywhere, and / or by adding an additional channel to the input image 20 that contains the focal plane of the input image 20 everywhere.

[0080] An additional input in the form of a scalar value, e.g., a target focal plane, a plane of the input image, an image type, a machine type, an acquisition time, etc., can be provided to the machine learning model using an encoding of the scalar value based on sine or cosine functions with different frequencies, similar to position encodings for x / y coordinates in transformer architectures. A sequence of sine and cosine functions can be generated, for example, for a scalar value τ as follows: S(τ,2i)=sinτn2i / d S(τ,2i+1)=cosτn2i+1 / d where i = 0, ..., N is used to map the sine and cosine functions to positions within a sequence, d is the dimension of the output embedding the sequence S, and n is a specified scalar, e.g., 10,000. The resulting sequence S for a scalar τ can optionally be transformed into a feature vector, e.g., by a multilayer perceptron. The multilayer perceptron can, for example, have a linear input layer, a hidden Gaussian error linear unit layer, and a linear output layer. The resulting feature vector is an encoding of the scalar value τ. The encoded scalar can, for example, be concatenated with the input of the input layer or the input of a hidden layer, e.g., the bottleneck, or with the input of a layer of a conditional diffusion model.

[0081] Additional inputs 30 can also be provided to the machine learning model using, for example, cross-attention layers. Cross-attention layers transform their input into a new representation, called an attention-based representation, by processing or paying attention to another data source, here the additional input 30. Compared to CNNs, cross-attention layers are not limited to convolutions within local neighborhoods, but consider large portions of the entire second source of information. Furthermore, the weights of the cross-attention layers are not fixed after training, but depend on the second source of information, i.e., the additional input 30 of the machine learning model.Accordingly, cross-attention layers are particularly flexible in considering a second source of information, resulting in highly accurate prediction results from the machine learning model.

[0082] A design image can be provided as additional input 30 in the input layer of a neural network, e.g., a coder-decoder architecture such as a U-Net. The rasterized design image can be combined with the input image 20, for example, via cross-attention layers.

[0083] Multiple additional pieces of information may be provided simultaneously as additional inputs 30 to the machine learning model 26. For example, the target focal plane and the focal plane of the input image may both be provided as additional inputs 30 to the machine learning model 26.

[0084] According to an example of the invention described in Fig. 7, a reference image 40 of the object comprising integrated circuit patterns is provided as an additional input 30 to the machine learning model 26, and the machine learning model 26 is trained to adjust the focal plane of the input image 20 and / or the focal plane of the reference image 40 such that the focal planes approximately match. The focal plane of the input image 20 may be adjusted to the focal plane of the reference image 40, or the focal plane of the reference image 40 may be adjusted to the focal plane of the input image 20, or the focal plane of the input image 20 and the focal plane of the reference image 40 may both be adjusted to a different focal plane, e.g., to a target focal plane or to an optimized focal plane with respect to some image quality metric, e.g., improved contrast or improved defect detectability.In the latter case, the focal plane of the input image 20 and the focal plane of the reference image 40 are adjusted using the image quality measure. The focal plane of the input image 20 and / or the focal plane of the reference image 40 and / or a target focal plane may be provided as an additional input 30 to the machine learning model, as described with reference to FIG. Fig. 6. The focus-adjusted input image 28 and the focus-adjusted reference image 42 have matching foci. Accordingly, they can be compared with greater accuracy, enabling defect detection with higher accuracy and increased sensitivity. For example, a difference image of the focus-adjusted input image 28 and the focus-adjusted reference image 42 would contain defects and a low noise level, whereas a difference image of the input image 20 and the reference image 40 would reveal large differences across the entire image due to the contrast difference and due to different sharpness levels at image edges.

[0085] In another example, the machine learning model comprises a conditional diffusion model that sequentially reverses a stochastic process and that is trained to increase a focal plane of the input image in each stochastic process step or that is trained to decrease a focal plane of the input image in each stochastic process step.

[0086] One in Fig. The diffusion model 44 illustrated in Figure 8 comprises a generative machine learning model configured to sequentially reverse a stochastic process 50, preferably a diffusion process, using a reverse stochastic process 52. In this case, the stochastic process is a focus plane increasing or focus plane decreasing process. The diffusion model 44 is configured to learn a distribution of images, here focused images. During training, the diffusion model 44 applies one or more stochastic process steps 46 to a focused image 58. This is known as the stochastic process 50, which is used only during training of the diffusion model 44. The stochastic process 50 gradually results in samples 48 that are farther removed from the learned distribution of focused images 58, e.g., "more defocused."The stochastic process 50 is then reversed in a reverse stochastic process 52 to recover the original focused image 58 by sequentially applying a reverse stochastic process step 54 to the sample 48, resulting in a generated focused image 60. In this way, the diffusion model 44 learns to gradually remove the effect of the stochastic process 50, the defocus, from the samples 48. During inference, only the reverse stochastic process 56 is applied to randomly generated initial samples 48 to generate focused images 58. Diffusion models are described, for example, in "Denoising Diffusion Probabilistic Models," J. Ho, A. Jain, P. Abbeel, 2020, arXiv 2006.11239.Since the invention does not aim at generating arbitrary focused images 58, but rather focused images for a specific input image 20, a conditional diffusion model conditioned on the input image 20 is preferably used. This is achieved by using the input image 20 as input to the reverse stochastic process 52. Apart from the input image 20, additional information can optionally also be provided as a condition for the conditional diffusion model 44, i.e., as additional inputs 30 to the reverse stochastic process 52.

[0087] In one example, the machine learning model is trained to remove additional image quality degradations from the input image, such as misalignment, noise, or shadowing effects. Some of these image quality degradations may be correlated with defocus of the input image. One specific image quality degradation caused by defocus is a focus-specific misalignment of the input image. This focus-specific misalignment can be corrected along with focus adjustment and, optionally, other image quality degradations.To obtain a machine learning model for focus adjustment that simultaneously removes additional image quality degradations, the machine learning model's training data can be configured as follows: the training data includes defocused images and focused images, where the defocused images contain additional image quality degradations, whereas the focused images do not contain these additional image quality degradations. When trained with this type of training data, the machine learning model learns to simultaneously remove image quality degradations from the input image with the focus adjustment without significant computational overhead.

[0088] According to a Fig. In the flowchart shown in Figure 9, instead of training a single machine learning model to adjust the focal plane of an input image, a set of machine learning models is trained. Each machine learning model in the set of machine learning models is trained for a specific focal plane interval of the input image.The method 62 for adjusting the focal plane of an image of an object comprising integrated circuit patterns comprises: acquiring, in a step S1, an input image of an object comprising integrated circuit patterns using an inspection system; determining, in a step S2, the focal plane of the acquired input image; selecting, in a step S3, a machine learning model from a set of machine learning models using the determined focal plane, wherein each machine learning model of the set of machine learning models for adjusting the focal plane of an input image of an object comprising integrated circuit patterns is trained for an input image having a focal plane within a specific focal plane interval; and applying, in a step S4, the selected machine learning model to the input image.

[0089] The set of machine learning models covers several focal plane intervals. These focal plane intervals can be of the same size or they can differ in size. They can be uniformly distributed over a predefined interval of expected focal planes of the input image. Alternatively, they can be distributed over the interval of expected focal planes according to the probability that the focal planes occur in an input image. For example, as in Fig. 10 illustrates, a statistic about the frequency 64 of the focal planes 62 in input images can be obtained. A probability density function 66, e.g., a normal density function, could optionally be fitted to the statistic and the focal plane intervals 68, e.g., I1, I2, I3, I4 in Fig. 10, could be defined to occur with approximately the same probability, e.g., by covering approximately the same area under the probability density function. The focal plane intervals 68 could alternatively be randomly distributed within the interval of expected focal planes, etc. In step S2, the focal plane of the acquired input image may be determined, for example, using a measurement method, or it may be specified by a user or obtained from a database, etc. Selecting a machine learning model from the set of machine learning models in step S3 may, for example, comprise selecting the machine learning model whose focal plane interval covers the particular focal plane of the input image. Alternatively, two, three, or more machine learning models could be selected for focal plane intervals closest to the particular focal plane of the input image.After applying these machine learning models to the input image, the best result with respect to any image quality measure could be selected as the focus-adjusted image.

[0090] Alternatively, the set of machine learning models can be used to adjust the focus of an input image without determining the focal plane of the input image, as shown in the flowchart in Fig. 11. The method 70 for adjusting the focal plane of an image of an object comprising integrated circuit patterns comprises: acquiring, in a step T1, an input image of an object comprising integrated circuit patterns using an inspection system; applying, in a step T2, two or more machine learning models from a set of machine learning models to the acquired input image, resulting in a set of focus-adjusted images, each machine learning model of the set of machine learning models for adjusting the focal plane of an input image of an object comprising integrated circuit patterns being trained for an input image having a focal plane within a specific focal plane interval; and selecting, in a step T3, a focus-adjusted image from the set of focus-adjusted images using an image quality measure.

[0091] The two or more applied machine learning models may include all machine learning models of the set of machine learning models. Alternatively, the two or more applied machine learning models may be randomly selected from the set of machine learning models. Alternatively, the two or more applied machine learning models may be selected from the set of machine learning models according to some pattern, e.g., a machine learning model may be selected for every second, third, fourth, etc. focal plane interval. Alternatively, a fixed number of machine learning models may be selected from the set of machine learning models according to a decreasing probability of the corresponding focal plane interval, i.e., a decreasing probability that a focal plane of an input image falls within the focal plane interval, etc.

[0092] The machine learning model for adjusting a focal plane of an image of an object including integrated circuit patterns can be used for defect detection in the image.

[0093] Fig. 12 shows a flowchart of a method 71 for detecting defects in an image of an object comprising integrated circuit patterns, according to a second embodiment of the invention. The method comprises: acquiring, in a step D1, an input image of an object comprising integrated circuit patterns using an inspection system; adjusting, in a step D2, the focal plane of the input image using a method according to one of the examples or aspects of the first embodiment of the invention; and applying, in a step D3, a defect detection method to the focus-adjusted input image to detect defects.

[0094] Alternatively, the focal plane of a reference image used for defect detection can be adjusted together with the focal plane of the input image. Fig. 13 shows a flowchart of a method 71' for detecting defects in an image of an object comprising integrated circuit patterns, according to a second embodiment of the invention. The method comprises: acquiring, in a step F1, an input image of an object comprising integrated circuit patterns using an inspection system; obtaining, in a step F2, a reference image of the object comprising integrated circuit patterns; adjusting, in a step F3, the focal plane of the input image and / or the focal plane of the reference image such that the focal planes approximately match, using a method according to claim 13 or 14; and applying, in a step F4, a defect detection method to the focus-adjusted input image and the focus-adjusted reference image.

[0095] Various defect detection methods for images of objects comprising integrated circuit patterns are known to those skilled in the art, in particular die-to-die and die-to-database methods. For example, a machine learning model could be trained to detect defects from pairs of the focus-adjusted input image and the focus-adjusted reference image and / or from their difference image. Optionally, the original input image and reference image could be used as additional input to the machine learning model, as they may contain additional information. Alternatively, defects could be detected by applying simple thresholds or adaptive thresholds to the difference image through template matching approaches or filtering approaches.

[0096] The training of the machine learning model for focus adaptation in this application can be performed in several ways.

[0097] According to one example, a computer-implemented method for training a machine learning model to adjust the focal plane of an image of an object comprising integrated circuit patterns according to a third embodiment of the invention comprises: obtaining training data comprising pairs of defocused images of objects comprising integrated circuit patterns and corresponding focused images obtained by adjusting the focal plane of the defocused images; and training the machine learning model to adjust the focal plane of an input image of an object comprising integrated circuit patterns using the training data. The focal plane of the defocused images may be adjusted, for example, by adjusting the position of the stage. Alternatively, the focal plane of the defocused image may be adjusted by applying a known image manipulation.

[0098] To generate the training data, pairs of defocused images and corresponding in-focus images can be obtained. These pairs can be acquired using different focus settings across one or more objects comprising integrated circuit patterns in the inspection system, e.g., by acquiring a defocused image with any defocus and a focused image of the same object at the same location without defocus. The focal planes of the defocused images and / or the in-focus images can be stored together in pairs to be used as additional input to the machine learning model. The different focus settings can be realized, for example, by sequentially recording images with different focus settings, i.e., by scanning a strip multiple times with different focus offsets.From each recorded pair of defocused image and focused image, overlapping sub-images, e.g., with a size of 2048 x 2048, can be extracted to be used as training data.

[0099] The training data preferably includes different types of integrated circuit patterns, e.g., different types of semiconductor structures or photolithography mask structures, such as lines and spaces, contact holes, logic patterns, etc., to achieve process reliability across different structure types. The training data may include defocused images and corresponding in-focus images of different objects, e.g., of different photolithography masks or wafers, or of the same object. Preferably, the defocused images contain integrated circuit patterns at different locations on the objects, e.g., the photolithography masks or wafers, to learn spatially dependent defocusing options. The defocused images should be acquired or simulated for different focal planes, preferably covering the range of expected focal planes in acquired input images.For example, the distribution of focal planes in the defocused images can reflect the distribution of focal planes in the acquired input images. This ensures the reliability of the method for different focal planes of the input image.

[0100] The training data can include captured images of objects containing integrated circuit patterns and / or simulated images of objects containing integrated circuit patterns. Captured images are more realistic, e.g., they include noise and image quality degradations, but their acquisition is often time-consuming, requires significant user effort, and carries the risk of not covering all relevant patterns, defects, focal planes, or image acquisition conditions to achieve sufficient generalization capability of the machine learning model. Simulated input images are less realistic but are easily generated automatically and quickly with little user effort. They also enable systematic and dense image generation for different focal plane ranges, types of integrated circuit patterns, defect types, image acquisition conditions, image quality degradations, etc.The simulated images can be obtained, for example, using simulations based on physical models, such as RCWA or TEA, or using simulations based on machine learning models, e.g., diffusion models, or from design data, e.g., from CAD models, for example, by using a generative machine learning model conditioned on the design of a photolithography mask. Preferably, the training data includes both acquired input images and simulated input images to achieve high prediction accuracy.

[0101] Preferably, the training data is generated in a systematic manner. For example, the same object comprising integrated circuit patterns, or a portion thereof, should be used to acquire images at different focal planes to generate defocused images and corresponding focused images. Images can be acquired under different focus conditions, for example, by scanning the same object comprising integrated circuit patterns with different focus offsets or by varying the image sensor acquisition settings, e.g., by mounting some of the image sensors, e.g., TDI sensors, in and out of focus. Images can also be acquired under different focus settings by arranging multiple sensors above and below the focal plane with different fixed specifications in the inspection system.

[0102] The training data is split into training data, test data, and validation data, e.g., using a split ratio of 70% / 15% / 15%. The training data is split layer by layer to ensure that all structure types and focal levels are present in the training data, test data, and validation data. The validation data is used to measure the performance of the machine learning model on unknown data during training. It is used to guide the training process and prevent overfitting. The test data is used to measure the performance of the trained machine learning model.

[0103] Using the training data, the machine learning model can be trained to adjust the focal plane of an image of an object containing integrated circuit patterns. A machine learning model suitable for this task is, for example, an encoder-decoder architecture, such as a U-Net, as described in Fig. 6, or a conditional diffusion model as in Fig. 8 shown.

[0104] During the training of the machine learning model for focus adaptation, a loss function is minimized. The construction of a suitable loss function for training the machine learning model, e.g., an encoder-decoder CNN model, is described in Fig. 14. The loss function 82 may include an adaptation loss function 78 and / or an adversarial loss function 80. The adaptation loss function 78 measures the deviation of a predicted focus-adjusted image 28, obtained by applying the machine learning model 26 to a defocused image 72, from the corresponding focused image 76. The deviation may be measured, for example, using some norm of the difference between the predicted focus-adjusted image 28 and the focused image 76, etc. The adversarial loss function 80 refers to a discriminator loss function of a discriminator trained in parallel to distinguish between predicted focus-adjusted images 28 and focused images 76. The adversarial loss function 80 measures the distance between a distribution of predicted focus-adjusted images 28 and a distribution of corresponding focused images 76.Accordingly, the discriminator learns to statistically distinguish “fake images” predicted by the machine learning model from “real images” that are captured.

[0105] Training the machine learning model may be performed using any optimization algorithm known to one skilled in the art, in particular the Adam optimization algorithm for neural networks, for a number of epochs, e.g., 500. The hyperparameters of the machine learning model may be optimized using the validation set and any hyperparameter optimization method known to one skilled in the art.

[0106] The machine learning model trained using a method described above may be trained according to one of the examples of the first embodiment of the invention described above. For example, the machine learning model may use additional information as additional input, e.g., a design of a photolithography mask, a target focal plane, a focal plane of an input image, a reference image used as additional input, wherein the machine learning model is trained to adjust the focal plane of the input image and / or the focal plane of the reference image so that the focal planes approximately match or match, etc. The additional information may be provided together with the input of the machine learning model, or it may be provided in a later stage of the machine learning model, e.g., in a hidden layer of a neural network.The machine learning model can be trained to remove additional image quality degradations from the input image, such as noise, shadowing effects, or focus-dependent misalignment, etc. For this purpose, pairs of training data comprising defocused images including these image quality degradations and corresponding focused images without these image quality degradations can be used.

[0107] In a preferred example, the machine learning model for adjusting the focal plane of an image of an object comprising integrated circuit patterns and a subsequent defect detection machine learning model are trained jointly. Joint training means that a single model is trained to perform both tasks simultaneously or sequentially—in this case, focus adjustment of the input image followed by defect detection. To achieve this, the anomaly detection errors can be backpropagated not only into the defect detection machine learning model but also into the focal plane adjustment machine learning model. In this way, the focal plane adjustment machine learning model can be trained more robustly. Because both tasks were solved together, they can exploit information from each other, resulting in predictions with higher accuracy.

[0108] Fig. Figure 15 illustrates the construction of a suitable loss function for jointly training a machine learning model 26 for focus adaptation and a machine learning model 84 for defect detection. The loss function 82 may include an adaptation loss function 78 and / or an adversarial loss function 80 and a detection loss function 86. The adaptation loss function 78 and / or the adversarial loss function 80 may be configured as previously described.

[0109] The detection loss function 86 measures the deviation of a defect prediction 85 from a defect annotation 83. The defect prediction 85 can be obtained by applying the machine learning model 84 for defect detection to the focus-adjusted image 28. Defect annotations 83 can be obtained from user annotations or they can be generated if simulated defects are used. The detection loss function 86 can, for example, contain any difference measure of the defect prediction 85 and the defect annotation 83, e.g., a norm of a difference image or an intersection-over-union measure.

[0110] Instead of applying the defect detection machine learning model 84 after the focus adjustment machine learning model 26, a single machine learning model can be used that contains two branches—one for focus adjustment and the other for defect detection. In this way, the machine learning models 26 and 84 share some of the initial layers. After branching, they can process the information in different ways to obtain a focus-adjusted image 28 and a defect prediction 85.

[0111] Alternatively, a defect prediction 85 can be obtained by applying the machine learning model 84 for defect detection to the output of an intermediate layer of the machine learning model 26 for focus adjustment. In this way, intermediate feature maps are used for defect detection before the final focus adjustment. These can provide features that are more suitable for defect detection than the final focus-adjusted image 28.

[0112] During training and inference, one or more reference images could additionally be used by the focus adjustment machine learning model 26 and the defect detection machine learning model 84, e.g., a reference image used by both the focus adjustment machine learning model 26 and the defect detection machine learning model 84, e.g., a defect-free, focused image of an object comprising integrated circuit patterns. Alternatively, different reference images could be used by the focus adjustment machine learning model 26 and the defect detection machine learning model 84, e.g., a focused image of an object comprising integrated circuit patterns as a first reference image for the focus adjustment machine learning model 26 and a defect-free image of an object comprising integrated circuit patterns as a second reference image for the defect detection machine learning model 84.The second reference image may have a suboptimal focus, e.g., it may have the same focus as the defocused image 72.

[0113] By jointly training the focus adjustment machine learning model and the defect detection machine learning model 84, the focus adjustment machine learning model 26 can be trained to adjust the focus of an image in a manner that is well suited for defect detection and thus improves defect detection results. At the same time, the defect detection machine learning model 84 learns to adapt the defect detection to the focus-adjusted image 28. Furthermore, the optimal focus plane does not need to be defined in advance for the training data, but can be determined indirectly by the detection loss function 86. For example, images that are not perfectly focused might yield better defect detection results.

[0114] Any loss function 82 that includes two or more separate loss functions, such as the adaptation loss function 78, the adversarial loss function 80, or the detection loss function 86, may include weighting factors. By using a weighting factor for each of the separate loss functions, the impact of each loss function on the trained machine learning model can be controlled.

[0115] Instead of jointly training the machine learning model 26 for focus adjustment and the machine learning model 84 for defect detection, both models can also be trained separately and can still be used together.

[0116] Fig. 15 illustrates the application of a trained focus adjustment machine learning model 26 and then a trained defect detection machine learning model 84 to a defocused image 72. The defocused image 72 is transformed into a focus-adjusted image 28 by the trained focus adjustment machine learning model 26. A defect prediction 85 is then calculated from the focus-adjusted image 28 using the trained defect detection machine learning model 84. Optionally, one or both of the machine learning models may use a reference image as additional input. A single reference image could be used by both the focus adjustment machine learning model 26 and the defect detection machine learning model 84, e.g., a defect-free focused image of an object comprising integrated circuit patterns.Alternatively, different reference images could be used by the machine learning model 26 for focus adjustment and by the machine learning model 84 for defect detection, e.g., a focused image of an object comprising integrated circuit patterns as a first reference image for the machine learning model 26 for focus adjustment and a defect-free image of an object comprising integrated circuit patterns as a second reference image for the machine learning model 84 for defect detection. The second reference image could have a suboptimal focus; for example, it could have the same focus as the defocused image 72.

[0117] An inspection system 88 for detecting defects 22 in an object 100 comprising integrated circuit patterns according to a sixth embodiment of the invention is shown in Fig.17. The inspection system 88 includes: an imaging device 98 configured to provide an image 20 of the object 100 comprising integrated circuit patterns; one or more processing devices 92; one or more machine-readable hardware storage devices 94 comprising instructions executable by one or more processing devices 92 to perform operations comprising one of the methods for detecting defects in an image 20 of an object 100 comprising integrated circuit patterns, according to any of the examples or aspects according to the second embodiment of the invention.

[0118] The imaging device 98 provides the image 20 to a data analysis device 90. The one or more processing devices 92 may be implemented, for example, as a CPU, GPU, or TPU. The one or more processing devices 92 may receive the image 20 via an interface 96. The one or more processing devices 92 may load program code from memory, e.g., program code for executing a method for detecting defects 22 according to the second embodiment of the invention, as previously described. The one or more processing devices 92 may execute the program code.

[0119] Reference throughout this specification 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" in various places throughout this specification do not necessarily refer to the same embodiment, example, or aspect, but may do so. Further, the particular features or characteristics may be combined in any suitable manner in one or more embodiments, as would be apparent to one of ordinary skill in the art from this disclosure.

[0120] Furthermore, while some embodiments, examples, or aspects described herein include some features included in other embodiments, examples, or aspects, but not others, combinations of features of different embodiments, examples, or aspects are intended to be within the scope of the claims and to form different embodiments, as would be understood by one skilled in the art.

[0121] In summary, the invention relates to a method for adjusting the focal plane of an image of an object comprising integrated circuit patterns, the method comprising: acquiring an input image 20 of an object comprising integrated circuit patterns using an inspection system; and applying a machine learning model 26 to the input image 20, wherein the machine learning model 26 is trained to adjust a focal plane of an input image 20 of an object comprising integrated circuit patterns. The invention also relates to methods for detecting defects 22 using such methods, methods for training a corresponding machine learning model, and a corresponding system for detecting defects 22. List of reference symbols 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 26 Machine learning model 28 Focus-adjusted image 30 Additional input 32 Bottleneck 34 coders 36 decoders 38 Skip connection 40 reference image 42 Focus-adjusted reference image 44 Diffusion model 46 Stochastic process step 48 Sample 50 Stochastic process 52 Reverse stochastic process 54 Reverse stochastic process step 56 episodes 58 Focused image 60 Generated focused image 62 focal plane 64 Frequency 66 Probability density function 68 focal plane interval 70 Computer-implemented procedure 71, 71' Procedure 72 Defocused image 76 Focused image 78 Adaptation loss function 80 Adversarial loss function 82 Loss function 83 Defect annotation 84 Machine learning model for defect detection 85 Defect prediction 86 Detection loss function 88 Inspection system 90 Data analysis device 92 processing device 94 Hardware storage device 96 Interface 98 Imaging device 100 objects QUOTES CONTAINED IN THE DESCRIPTION

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

[0000] US 2022 / 0405903

[0070] Cited non-patent literature

[0000] Denoising Diffusion Probabilistic Models", J. Ho, A. Jain, P. Abbeel, 2020, arXiv 2006.11239

[0086]

Claims

[1] A computer-implemented method (24) for adjusting the focal plane of an image (20) of an object (100) comprising integrated circuit patterns, the method comprising: - providing an input image (20) of an object (100) comprising integrated circuit patterns, - applying a machine learning model (26) to the input image (20), wherein the machine learning model (26) is trained to adapt a focal plane of an input image (20) of an object (100) comprising integrated circuit patterns. [2] The method of claim 1, wherein the machine learning model (26) is trained to adjust the focal plane of the input image (20) to a specified target focal plane. [3] Method according to claim 2, wherein a reference image is additionally provided, wherein the focal plane of the reference image (40) is determined, and wherein the target focal plane is derived from the focal plane of the reference image (40). [4] The method of claim 2 or 3, wherein the target focal plane is provided as an additional input (30) to the machine learning model (26). [5] The method of claim 1, wherein the focal plane of the input image (20) is adjusted using an image quality measure. [6] Method according to one of the preceding claims, wherein the focal plane of the input image (20) is provided as an additional input (30) to the machine learning model (26). [7] A method according to any one of the preceding claims, wherein a design of a photolithography mask is provided as an additional input (30) to the machine learning model (26). [8] Method according to one of the preceding claims, wherein the machine learning model (26) is trained to remove additional image quality degradations from the input image (20). [9] Method according to one of the preceding claims, wherein the machine learning model (26) comprises an encoder-decoder architecture, in particular a U-Net architecture. [10] Method according to claim 9, wherein the focal plane of the input image (20) is provided as an additional input (30) to the encoder-decoder architecture in one of its hidden layers, in particular in one of the layers of the encoder (34), in particular in the bottleneck (32). [11] Method according to claim 9 or 10, wherein a target focal plane of the input image (20) is provided as an additional input (30) to the encoder-decoder architecture in one of its hidden layers, in particular in one of the layers of the decoder (36), in particular in the bottleneck (32). [12] The method of any one of claims 1 to 9, wherein the machine learning model (26) comprises a conditional diffusion model (44) that sequentially reverses a stochastic process (50) and that is trained to increase a focal plane of the input image (20) in each stochastic process step (46) or that is trained to decrease a focal plane of the input image (20) in each stochastic process step (46). [13] Method according to one of the preceding claims, wherein a reference image (40) of the object (100) comprising integrated circuit patterns is provided, and wherein the machine learning model (26) is trained to adjust the focal plane of the input image (20) and / or the focal plane of the reference image (40) so that the focal planes approximately match. [14] The method of claim 13, wherein the focal plane of the input image (20) and the focal plane of the reference image (40) are adjusted using an image quality measure. [15] A computer-implemented method (62) for adjusting the focal plane of an image (20) of an object (100) comprising integrated circuit patterns, the method comprising: - providing an input image (20) of an object (100) comprising integrated circuit patterns; - determining the focal plane of the input image (20); - selecting a machine learning model (26) from a set of machine learning models (26) using the determined focal plane, wherein each machine learning model (26) of the set of machine learning models is trained to adjust the focal plane of an input image (20) of an object (100) comprising integrated circuit patterns for an input image (20) having a focal plane within a specific focal plane interval (68); - Applying the selected machine learning model (26) to the input image (20). [16] A computer-implemented method (70) for adjusting the focal plane of an image (20) of an object (100) comprising integrated circuit patterns, the method comprising: - providing an input image (20) of an object (100) comprising integrated circuit patterns; - applying two or more machine learning models (26) from a set of machine learning models to the input image (20), resulting in a set of focus-adjusted images (28), each machine learning model (26) of the set of machine learning models being trained to adjust the focus plane of an input image (20) of an object (100) comprising integrated circuit patterns for an input image (20) having a focus plane within a specific focus plane interval (68); - selecting a focus-adjusted image (28) from the set of focus-adjusted images (28) using an image quality measure. [17] A method according to any one of the preceding claims, wherein the input image (20) is acquired using an inspection system (88). [18] A method (71) for detecting defects (22) in an image (20) of an object (100) comprising integrated circuit patterns, the method comprising: - providing an input image (20) of an object (100) comprising integrated circuit patterns; - adjusting the focal plane of the input image (20) using a method according to any one of the preceding claims; - applying a defect detection method to the focus-adjusted input image (28) to detect defects (22). [19] The method of claim 18, wherein the defect detection method comprises a machine learning model (84) for defect detection. [20] A method (71') for detecting defects (22) in an image (20) of an object (100) comprising integrated circuit patterns, the method comprising: - providing an input image (20) of an object (100) comprising integrated circuit patterns; - Obtaining a reference image (40) of the object (100) comprising integrated circuit patterns; - adjusting the focal plane of the input image (20) and / or the focal plane of the reference image (40) so that the focal planes approximately match, using a method according to claim 13 or 14; - applying a defect detection method to the focus-adjusted input image (28) and the focus-adjusted reference image (42). [21] The method of any one of claims 18 to 20, wherein the input image (20) is acquired using an inspection system (88). [22] A computer-implemented method for training a machine learning model (26) to adjust the focal plane of an image (20) of an object (100) comprising integrated circuit patterns, the method comprising: - Obtaining training data comprising pairs of defocused images (72) of objects (100) comprising integrated circuit patterns and corresponding focused images (76) obtained by adjusting the focal plane of the defocused images (72); and - training the machine learning model (26) to adjust the focal plane of an input image (20) of an object (100) comprising integrated circuit patterns using the training data. [23] The method of claim 22, wherein the machine learning model (26) is trained by minimizing a loss function (82) comprising the deviation of focus-adjusted images (28) predicted by the machine learning model (26) when presented with defocused images (72) of the training data from the corresponding focused images (76) of the training data. [24] The method of claim 22 or 23, wherein the machine learning model (26) is trained by minimizing a loss function (82) comprising an adversarial loss function (80) that measures the distance between a distribution of focus-adjusted images (28) predicted by the machine learning model (26) when presented with defocused images (72) of the training data and a distribution of the corresponding focused images (76). [25] The method of any one of claims 22 to 24, wherein the machine learning model (26) is configured to adjust the focal plane of an input image (20) of an object (100) comprising integrated circuit patterns according to any one of claims 1 to 17. [26] A computer-implemented method for training a machine learning model (26) for adjusting the focal plane of an image (20) of an object (100) comprising integrated circuit patterns and a machine learning model (84) for detecting defects in an image (20) of an object (100) comprising integrated circuit patterns, as claimed in claim 19, wherein a joint loss function is used for jointly training the machine learning model (26) for adjusting the focal plane of an image (20) and the machine learning model (84) for detecting defects. [27] A computer-readable medium having stored thereon a computer program executable by a computing device, the computer program comprising code for carrying out a method according to any one of claims 22 to 26. [28] A computer program product comprising instructions which, when executed by a computer, cause the computer to carry out a method according to any one of claims 22 to 26. [29] An inspection system (88) for detecting defects (22) in an object (100) comprising integrated circuit patterns, the system comprising: - an imaging device (98) configured to provide an image (20) of the object (100) comprising integrated circuit patterns; - one or more processing devices (92); - one or more machine-readable hardware storage devices (94) comprising instructions executable by one or more processing devices (92) to perform operations comprising one of the methods for detecting defects (22) in an image (20) of an object (100) comprising integrated circuit patterns according to any one of claims 18 to 21.

Images