Apparatus and methods for generating denoising model

Abstract

To provide a method for training a denoising model.SOLUTION: This method includes obtaining a first set of simulated images based on design patterns. The simulated images may be clean, and noise can be added to the simulated images to generate noisy simulated images. The clean and noisy simulated images are used as training data to generate a denoising model.SELECTED DRAWING: Figure 2

Description

[Technical Field]

[0001] CROSS-REFERENCE TO RELATED APPLICATIONS

[0001] This application claims priority to U.S. Patent Application No. 63 / 051,500, filed July 14, 2020, the entire contents of which are incorporated herein by reference.

[0002]

[0002] The description herein relates generally to processing images acquired by inspection or metrology tools, and more particularly to denoising images by using machine learning. [Background technology]

[0003] Lithographic projection apparatuses may be used, for example, in the manufacture of integrated circuits (ICs). In such cases, a patterning device (e.g., a mask) may contain or be provided with a pattern (a "design layout") that corresponds to an individual layer of the IC, and this pattern may be transferred onto a target portion (e.g., comprising one or more dies) on a substrate (e.g., a silicon wafer) that is coated with a layer of radiation-sensitive material ("resist"), such as by irradiating the target portion through the patterning device. Typically, a single substrate will contain several adjacent target portions (one target portion at a time) onto which the pattern is successively transferred by the lithographic projection apparatus. In one type of lithographic projection apparatus, the pattern on the entire patterning device is transferred onto one target portion at a time; such an apparatus is commonly referred to as a stepper. In an alternative apparatus, commonly referred to as a step-and-scan apparatus, the projection beam is moved parallel to or anti-parallel to a given reference direction (the "scan" direction) in synchronization with scanning the patterning device in this reference direction. Different portions of a pattern on the patterning device are progressively transferred onto one target portion. In general, the lithographic projection apparatus will have a de-magnification ratio M (e.g., 4), so that the speed F at which the substrate is moved will be 1 / M times the speed at which the projection beam scans the patterning device. More information on lithographic devices as described herein can be found, for example, in U.S. Pat. No. 6,046,792, incorporated herein by reference.

[0004] Before transferring the pattern from the patterning device to the substrate, the substrate may undergo various procedures, such as priming, resist coating, and a soft bake. After exposure, the substrate may undergo other procedures ("post-exposure procedures"), such as a post-exposure bake (PEB), development, a hard bake, and measurement / inspection of the transferred pattern. This multitude of procedures is used as a basis for creating an individual layer of a device, e.g., an IC. The substrate may then undergo various processes, such as etching, ion implantation (doping), metallization, oxidation, chemical-mechanical polishing, etc., all intended to finish off an individual layer of the device. If several layers are required for a device, the entire procedure, or a variant thereof, is repeated for each layer. Eventually, a device is present on each target portion of the substrate. The devices are then separated from each other by techniques such as dicing or sawing, so that the individual devices can be mounted on a carrier, connected to pins, etc.

[0005]

[0005] Thus, manufacturing devices such as semiconductor devices typically involves processing a substrate (e.g., a semiconductor wafer) using multiple fabrication processes to form various features and multiple layers of the device. Such layers and features are typically produced and processed using, for example, deposition, lithography, etching, chemical-mechanical polishing, and ion implantation. Multiple devices may be fabricated on multiple dies on a substrate and then separated into individual devices. This device manufacturing process can be considered a patterning process. The patterning process includes a patterning step, such as optical and / or nanoimprint lithography, that uses a patterning device in a lithography apparatus to transfer a pattern on the patterning device to the substrate, and typically (but optionally) includes one or more associated pattern processing steps, such as developing the resist in a developer, baking the substrate using a bake tool, or etching using the pattern in an etcher. Summary of the Invention

[0006] According to one embodiment, a method for training an image denoising model for processing an image is provided. The method includes converting a design pattern into a first set of simulated images and training a denoising model based on the first set of simulated images and image noise. After training, the denoising model is operable to remove noise from an input image to generate a denoised image.

[0007] In one embodiment, a system is provided that includes electron beam optics configured to capture an image of a patterned substrate and one or more processors configured to generate a de-noised image of the input image. The one or more processors are configured to execute a trained model configured to generate a simulated image from a design pattern on the substrate. In one embodiment, the one or more processors are configured to execute the de-noising model using the captured image as input to generate a de-noised image of the patterned substrate.

[0008] In one embodiment, one or more non-transitory computer-readable media are provided for storing a denoising model. In one embodiment, the one or more non-transitory computer-readable media are configured to generate a denoising image according to the stored denoising model. In particular, the one or more non-transitory computer-readable media store instructions that, when executed by one or more processors, provide the denoising model. In one embodiment, the denoising model is generated by executing instructions for: obtaining a first set of simulated images based on a design pattern (e.g., by converting a GDS pattern into simulated images using a trained GAN); providing the first set of simulated images as inputs to a base denoising model to obtain a second set of simulated images, where the second set of simulated images are denoising images associated with the design pattern; and updating one or more configurations of the reference denoising model using a reference denoising image as feedback, where the one or more configurations are updated based on a comparison of the reference denoising image and the second set of simulated images.

[0009] In some embodiments, a GAN is trained to convert GDS pattern images into clean simulated SEM images. First, noise features are extracted from scanned SEM images, and then noise is added to these clean images to generate noisy simulated images. The clean simulated images and noisy images are used in combination with the scanned SEM images to train a denoising model. The denoising model can be further fine-tuned using captured SEM images. Once trained, the denoising model is operable to remove noise from input SEM images to generate denoised images.

[0010] According to an embodiment of the present disclosure, a denoising model is trained by using simulated images converted from a design pattern through a generator model as described above. The training data, including such simulated images, can collectively cover significantly more patterns than images captured by an SEM. As a result of the improved pattern coverage, training can advantageously significantly improve the effectiveness and accuracy of the denoising model. The need for retraining can be significantly reduced or even eliminated.

[0011]

[0011] The above aspects and other aspects and features will become apparent to those skilled in the art upon review of the following description of specific embodiments in conjunction with the accompanying drawings. [Brief explanation of the drawings]

[0012] [Figure 1] FIG. 1 shows a block diagram of various subsystems of a lithography system, according to one embodiment. [Figure 2]

[0013] 1 is a method for training a denoising model. In one embodiment, a denoising model is trained to transform design patterns. [Figure 3]

[0014] 10 is a flowchart of a variation of a method for training a denoising model, according to one embodiment. [Figure 4]

[0015] 1 illustrates an example of training a model, according to one embodiment. [Figure 5]

[0016] 5 illustrates an example of using the trained model of FIG. 4 to obtain a first set of simulated SEM images, according to one embodiment. [Figure 6]

[0017] 6 illustrates an example of adding noise to the simulated SEM image of FIG. 5, according to one embodiment. [Figure 7]

[0018] 1 illustrates an example of training a denoising model, according to one embodiment. [Figure 8]

[0019] 1 illustrates an example of a trained denoising model used to generate a denoised input SEM image, according to one embodiment. [Figure 9]

[0020] 1 schematically depicts an embodiment of a scanning electron microscope (SEM), according to one embodiment. [Figure 10]

[0021] 1 illustrates a schematic representation of an embodiment of an electron beam inspection apparatus, according to one embodiment. [Figure 11]

[0022] FIG. 1 is a block diagram of an exemplary computer system, according to one embodiment. [Figure 12]

[0023] 1 depicts a schematic diagram of a lithographic projection apparatus, according to one embodiment; [Figure 13]

[0024] 1 is a schematic diagram of another lithographic projection apparatus, according to an embodiment; [Figure 14]

[0025] FIG. 13 is a more detailed diagram of the device of FIG. 12, according to one embodiment. [Figure 15]

[0026] 15 is a more detailed diagram of the source collector module SO of the apparatus of FIGS. 13 and 14, according to one embodiment. FIG. DETAILED DESCRIPTION OF THE INVENTION

[0013]

[0027] Before describing the embodiments in detail, it is helpful to present an exemplary environment in which the embodiments may be implemented.

[0014]

[0028] While specific reference may be made herein to the manufacture of ICs, it should be expressly understood that the description herein has many other possible applications. For example, the description herein may be used in the manufacture of integrated optical systems, guidance and detection patterns for magnetic domain memories, liquid crystal display panels, thin film magnetic heads, etc. Those skilled in the art will understand that, in the context of such alternative applications, any use of the terms "reticle," "wafer," or "die" herein should be considered interchangeable with the more general terms "mask," "substrate," and "target portion," respectively.

[0015]

[0029] In this document, the terms "radiation" and "beam" may be used to encompass 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 about 5-100 nm).

[0016]

[0030] A patterning device may include or form one or more design layouts. Design layouts can be generated using computer-aided design (CAD) programs, a process often referred to as electronic design automation (EDA). Most CAD programs follow a set of predetermined design rules to create functional design layouts / patterning devices. These rules are set by process and design constraints. For example, design rules define the space tolerance between devices (gates, capacitors, etc.) or interconnect lines to ensure that the devices or lines do not interact with each other in an undesirable way. One or more of the design rule constraints may be referred to as a "critical dimension" (CD). A critical dimension of a device may be defined as the smallest width of a line or hole, or the smallest space between two lines or two holes. Thus, the CD determines the overall size and density of the designed devices. Of course, one of the goals of device fabrication is to faithfully reproduce the original design intent on a substrate (via a patterning device).

[0017]

[0031] The pattern layout design may include, for example, the application of resolution enhancement techniques such as optical proximity correction (OPC). OPC addresses the fact that the final size and location of the image of the design layout projected onto the substrate will not be identical to, or solely depend on, the size and location of the design layout on the patterning device. Note that the terms “mask,” “reticle,” and “patterning device” are used interchangeably herein. Also, in the context of RET, a physical patterning device is not necessarily used to represent a physical patterning device, but rather a design layout can be used, and those skilled in the art will recognize that the terms “mask,” “patterning device,” and “design layout” can be used interchangeably. Given the small size and high density of features present in some design layouts, the location of a particular edge of a given feature will be affected to some extent by the presence or absence of other adjacent features. These proximity effects arise from the small amount of radiation coupled from one feature to another or from non-geometric optical effects such as diffraction and interference. Similarly, proximity effects can arise from diffusion and other chemical effects during post-exposure bake (PEB), resist development, and etching, which typically follow lithography.

[0018]

[0032] To increase the likelihood that the projected image of a design layout will conform to the requirements of a given target circuit design, proximity effects can be predicted and compensated for using advanced numerical models, design layout correction, or predistortion. The paper "Full-Chip Lithography Simulation and Design Analysis—How OPC Is Changing IC Design," C. Spence, Proc. SPIE, Vol. 5751, pp. 1-14 (2005), provides an overview of current "model-based" optical proximity correction processes. In a typical high-end design, nearly every feature in the design layout undergoes some kind of modification to achieve high fidelity of the projected image to the target design. These modifications may include shifting or biasing edge positions or line widths, as well as the application of "assist" features intended to aid in the projection of other features.

[0019]

[0033] Assist features can be thought of as the difference between a feature on a patterning device and a feature in a design layout. The terms "primary feature" and "assist feature" do not imply that a particular feature on a patterning device must be labeled as one or the other.

[0020]

[0034] The terms "mask" or "patterning device" as used herein may be broadly interpreted as referring to any general patterning device that can be used to impart an incoming radiation beam with a patterned cross section that corresponds to the pattern to be produced in a target portion of a substrate, and the term "light valve" may also be used in this context. In addition to traditional masks (transmissive or reflective; binary, phase-shifting, hybrid, etc.), examples of other such patterning devices include: - a programmable mirror array. An example of such a device is a matrix-addressable surface having a viscoelastic control layer and a reflective surface. The basic principle behind such an apparatus is that (for example) addressed areas of the reflective surface will reflect incident radiation as diffracted radiation, while unaddressed areas will reflect the incident radiation as undiffracted radiation. Using an appropriate filter, the said undiffracted radiation can be filtered out of the reflected beam, leaving only the diffracted radiation behind; in this way the beam is patterned according to the addressing pattern of the matrix-addressable surface. The required matrix addressing can be performed using suitable electronic means. - A programmable LCD array. An example of such a construction is given by US Patent No. 5,229,872, which is incorporated herein.

[0021]

[0035] 1 shows an exemplary lithographic projection apparatus 10A, the main components of which are a radiation source 12A (as discussed herein, the lithographic projection apparatus itself need not have a radiation source), which may be a deep ultraviolet excimer laser source or other type of source, including an extreme ultraviolet (EUV) source, illumination optics which may include optics 14A, 16Aa, and 16Ab for shaping the radiation from source 12A, e.g., defining the partial coherence (denoted as sigma), a patterning device 18A, and transmission optics 16Ac which project an image of the patterning device pattern onto a substrate surface 22A. An adjustable filter or aperture 20A at the pupil plane of the projection optical system may limit the range of beam angles that can impinge on the substrate surface 22A, where the maximum possible angle defines the numerical aperture of the projection optical system, NA=n sin(Θmax), where n is the refractive index of the medium between the substrate and the last element of the projection optical system, and Θmax is the maximum angle of the beam exiting the projection optical system that can still impinge on the substrate surface 22A.

[0022]

[0036] In a lithographic projection apparatus, a source provides illumination (i.e., radiation) to a patterning device, and a projection optical system guides and shapes the illumination through the patterning device and onto a substrate. The projection optical system may include at least some of components 14A, 16Aa, 16Ab, and 16Ac. The aerial image (AI) is the radiation intensity distribution at substrate level. A resist layer on the substrate is exposed, and the aerial image is transferred to the resist layer as a latent "resist image" (RI). The resist image (RI) can be defined as the spatial distribution of resist solubility in the resist layer. A resist model can be used to calculate the resist image from the aerial image, examples of which can be found in U.S. Patent Application Publication No. 2009 / 0157360, the disclosure of which is incorporated herein by reference in its entirety. The resist model is only concerned with the properties of the resist layer (e.g., the effects of chemical processes occurring during exposure, PEB, and development). The optical properties of a lithographic projection apparatus (e.g. properties of the source, patterning device, and projection optics) determine the aerial image. Because the patterning devices used in a lithographic projection apparatus can be varied, it may be desirable to decouple the optical properties of the patterning device from the optical properties of the rest of the lithographic projection apparatus, which includes at least the source and projection optics.

[0023]

[0037] Although reference may be made herein to the use of a lithographic apparatus in the manufacture of ICs, it should be understood that the lithographic apparatus described herein may have other applications, such as, for example, the manufacture of integrated optical systems, guidance and detection patterns for magnetic domain memories, liquid crystal display panels (LCDs), thin-film magnetic heads, etc. Those skilled in the art will understand that, in the context of such alternative applications, the use of the terms “wafer” or “die” herein may be considered synonymous with the more general terms “substrate” or “target portion,” respectively. Substrates referred to herein may be processed, before or after exposure, in, for example, a track (e.g., a tool that typically applies a layer of resist to a substrate and develops the exposed resist) or a metrology or inspection tool. Where applicable, the disclosure herein may be applied to such and other substrate processing tools. Furthermore, a substrate may be processed multiple times, for example to create multi-layer ICs, and the term substrate as used herein may also refer to a substrate that already includes multiple processed layers.

[0024]

[0038] As used herein, the terms "radiation" and "beam" encompass all types of electromagnetic radiation, including ultraviolet (UV) radiation (e.g., having a wavelength of 365, 248, 193, 157, or 126 nm) and extreme ultraviolet (EUV) radiation (e.g., having a wavelength in the range of 5 to 20 nm), as well as particle beams such as ion beams or electron beams.

[0025]

[0039] Existing training methods for denoising models require a large number of images (e.g., SEM images) of a patterned substrate as training data. In such training methods, the pattern coverage of a design layout is not limited to the patterns in the SEM images. In one embodiment, pattern coverage refers to the number of unique patterns in a design layout. Typically, a design layout may have hundreds of millions to billions of patterns and millions of unique patterns. Measuring millions of patterns on a patterned substrate for training purposes is impractical because of the significant measurement time and computing resources required for training. Thus, for example, training data including SEM images is typically far less than sufficient data to train a machine learning model. Therefore, it may be necessary to retrain a trained model in real time using new patterns.

[0026]

[0040] The disclosed method has several advantages. For example, pattern coverage of a design layout can be significantly increased during offline training. Only limited SEM images (e.g., 10-20 actual SEM images) can be used for training and validation purposes. After training, the trained model can be used to denoise captured metrology images (e.g., SEM images) at run time. Because a relatively large number of patterns are covered during training, the amount of retraining of an existing model is significantly less than with existing models. Model fine-tuning can be achieved quickly, for example, by acquiring 10-20 actual SEM images. Therefore, significant machine scan time and online model training time can be reduced compared to existing models. For example, with the present method, scan time can be limited to 20 SEM images as opposed to thousands of SEM images, and online model training time can be limited to approximately 0.5 hours as opposed to 4-8 hours.

[0027]

[0041] 2 is an exemplary method 200 for training a denoising model according to an embodiment of the present disclosure. In one embodiment, for training purposes, another model converts design patterns (e.g., design layouts in GDS file data) into clean simulated SEM images. Further, noise may be added to these clean images to generate noisy simulated images. The clean simulated images and noisy images are used in combination with scanned SEM images to train the denoising model. In one embodiment, the method includes processes P201 and P203, which are discussed in detail below.

[0028]

[0042] Process P201 includes converting a design pattern into a first set of simulated images 201, such as simulated SEM images. In one embodiment, the design pattern DP is in a graphics data signal (GDS) file format. For example, a design layout containing millions of design patterns is represented as a GDS data file.

[0029]

[0043] In one embodiment, obtaining the first set of simulated images 201 includes running a trained model MD1 using the design pattern DP as input to generate the simulated images 201. In one embodiment, the trained model MD1 is trained based on the design pattern DP and captured images of the patterned substrate, where each captured image is associated with the design pattern. In one embodiment, the captured images are SEM images obtained via a scanning electron microscope (SEM) (e.g., FIGS. 10-11).

[0030]

[0044] In one embodiment, the trained model MD1 may be any model, such as a machine learning model, that can be trained using an existing training method that uses training data as discussed herein. For example, the trained model MD1 may be a convolutional neural network (CNN) or a deep convolutional neural network (DCNN). The present disclosure is not limited to a particular training method or a particular neural network. By way of example, the model MD1 may be a first deep learning model (e.g., a DCNN) trained using a training method such as a generative adversarial network (GAN), in which the design pattern DP and the SEM image are used as training data. In this example, the trained model MD1 (e.g., a DCNN) is referred to as a generative model configured to generate simulated SEM images from a given design pattern, e.g., a GDS pattern.

[0031]

[0045] In one embodiment, a generative adversarial network (GAN) includes two deep learning models, namely, a generative model (e.g., a CNN or a DCNN) and a classifier model (e.g., another CNN or another DCNN), that are trained together, particularly adversarially. The generator model can receive a design pattern DP and a captured image (e.g., an SEM image) as input and output a simulated image (e.g., a simulated SEM image). In one embodiment, the output simulated image can be labeled as a fake image or a real image. In one example, a fake image is an image of a specific class that has not previously existed in reality (e.g., a denoised image of an SEM image). Meanwhile, a real image used as a reference (or ground truth) is a previously existing image (e.g., an SEM of a printed circuit board) that can be used during training of the generator model and the classifier model. The training goal is to train the generator model to generate a fake image that closely resembles the real image. For example, the features of the fake image are at least 95% consistent with the features of the real image (e.g., the denoised SEM image). As a result, the trained generator model is capable of producing realistic simulation images with a high level of accuracy.

[0032]

[0046] In one embodiment, the generator model (G) may be a convolutional neural network. The generator model (G) receives the design pattern (z) as input and generates an image. In one embodiment, the image may be referred to as a false image or a simulated image. The false image is a representation of the X fake =G(z). The generator model (G) may be associated with a first cost function. The first cost function allows for tuning of the parameters of the generator model such that the cost function is improved (e.g., maximized or minimized). In one embodiment, the first cost function includes a first log-likelihood term that determines the probability that the simulated image is a false image given the input vector.

[0033]

[0047] An example of the first cost function can be expressed as Equation 1 below. L S =E[logP(S=fake|X fake )]…(1)

[0034]

[0048] In the above equation, the log likelihood of the conditional probability is calculated. In this equation, S refers to the source assignment as fake by the discriminator model, and X fake is the output of the generator model, i.e., the false image. Thus, in one embodiment, the training method minimizes a first cost function (L). As a result, the generator model generates false images (i.e., simulated images) such that the conditional probability that the classifier model will recognize the false image as a fake is low. In other words, the generator model gradually generates more realistic images or patterns.

[0035]

[0049] 4 illustrates an exemplary process for training a model (e.g., MD1) as discussed in process P201. In one embodiment, the trained generator model MD1 is trained using a generative adversarial network. Generative adversarial network-based training involves two deep learning models GM1 and DM1 that are trained together so that the generator model GM1 gradually produces more accurate and robust results.

[0036]

[0050] In one embodiment, generator model GM1 can receive, for example, design patterns DP1 and DP2 and SEM images SEM1 and SEM2 corresponding to design patterns DP1 and DP2 as inputs, and output simulated images such as images DN1 and DN2.

[0037]

[0051] The simulated image of GM1 is received by another CNN, a classifier model DM1. The classifier model DM1 also receives real images SEM1 and SEM2 (or a set of real patterns) in the form of pixelated images. Based on the real images, the classifier model DM1 determines whether the simulated image is fake (e.g., label L1) or real (e.g., label L2) and assigns a label accordingly. If the classifier model DM1 classifies the simulated image as fake, the parameters (e.g., biases and weights) of GM1 and DM1 are modified based on a cost function (e.g., the first cost function described above). The models GM1 and DM1 are iteratively modified until the classifier DM1 consistently classifies the simulated images generated by GM1 as real. In other words, the generator model GM1 is configured to generate realistic SEM images for any input design pattern.

[0038]

[0052] 5 illustrates an example of obtaining a first set of simulated SEM images using the trained generator model MD1 of FIG. 4, according to one embodiment. In one embodiment, arbitrary design patterns DP11, DP12, ... DPn can be input to the trained generator model MD1 to generate simulated SEM images S1, S2, ... Sn, respectively. Although the present disclosure is not limited thereto, these simulated SEM images S1, S2, ... Sn can be clean images that are typically free of noise or contain very low noise.

[0039]

[0053] In some embodiments, image noise may be added to generate the first set of simulated images 201 in Figure 2. Figure 6 shows how noise can be added to simulated SEM images S1, S2, ... Sn to generate the first set of simulated images S1, S2, ... Sn for training a denoising model. * , S2 * , …Sn * where the denoising model is another machine learning model configured to denoise the input image. Training of the denoising model is discussed below with respect to process P203.

[0040]

[0054] Process P203 includes training a denoising model, using the first set of simulated images 201 and the image noise as training data. In one embodiment, additionally, a captured image 205 of the patterned substrate may be included in the training data. For example, the captured image 205 may be an SEM image of the patterned substrate.

[0041]

[0055] In one embodiment, the denoising model MD2 is a second machine learning model. For example, the second machine learning model may be a second CNN or a second DCNN. This disclosure is not limited to a particular deep learning or machine learning training method. In one embodiment, training is an iterative process performed until the second set of simulated images falls within a specified threshold of a ground truth, such as the first set of simulated images 201 (e.g., simulated SEM images S1, S2, ... Sn of FIG. 5 ) before adding noise or a reference image. In one embodiment, training the denoising model includes using the first set of simulated images, image noise, and captured images as training data.

[0042]

[0056] In one embodiment, image noise is extracted from a captured image of the patterned substrate, for example, a captured SEM image. For example, a noise filter may be applied to extract noise from an SEM image captured by an SEM tool. The extracted noise may be represented as image noise. In one embodiment, the image noise is Gaussian noise, white noise, or salt and pepper noise characterized by user-specified parameters. In one embodiment, the image noise comprises pixels whose intensity values ​​are statistically independent of one another. In one embodiment, the Gaussian noise may be generated by varying parameters of a Gaussian distribution function.

[0043]

[0057] In one embodiment, referring to FIG. 6, image noise, such as Gaussian noise, is added to the simulated images S1, S2, . . . Sn, for example, S1 * , S2 * , …, Sn* The noisy images may be generated as follows: The noisy images are the first set of simulated images 201 used to train the denoising model.

[0044]

[0058] In conventional methods, denoising models are trained by using captured SEM images as training images. Limited by the imaging throughput of the SEM system, and therefore the amount of images captured by the SEM, the training images can collectively cover only a relatively small number of patterns, rendering the trained denoising model useless for denoising input images, which may have a variety of patterns. Unfortunately, the trained denoising model must be retrained to process images with new patterns. According to embodiments of the present disclosure, the denoising model is trained by using simulated images converted from design patterns through a generator model, as described above. The training data, including the simulated images, can collectively cover significantly more patterns than images captured by the SEM. As a result of the improved pattern coverage, training can advantageously significantly improve the effectiveness and accuracy of the denoising model. The need for retraining can be significantly reduced or even eliminated.

[0045]

[0059] FIG. 7 illustrates an exemplary process for training a denoising model (eg, MD2) according to an embodiment of the present disclosure.

[0046]

[0060] The first simulated image 201, the image noise, and / or the reference image REF as training data for training the denoising model MD2. In one embodiment, additionally, captured images 710 (e.g., SEM images) of the patterned substrate may be included in the training data. In one embodiment, the number of captured images 710 may be relatively smaller than the number of simulated images 201. In one embodiment, the captured images 710 may be used to update the trained denoising model MD2.

[0047]

[0061] In one embodiment, the model MD2 is generated from the noisy image S1 discussed with respect to, for example, FIGS. * , S2 * , …, Sn * Model MD2 may receive as input a denoised image (e.g., generated using model MD1) and image noise. Model MD2 outputs denoised images such as images DN11, D12, ...DNn. During the training process, one or more model parameters of model MD2 (e.g., weights and biases of different layers of a DCNN) may be modified until convergence is achieved or the denoised image falls within a specified threshold of the reference image REF. In one embodiment, convergence is achieved when changes in model parameter values ​​do not result in a significant improvement in the model output compared to the previous model output.

[0048]

[0062] In one embodiment, the method 200 further includes acquiring an SEM image of the patterned substrate via a metrology tool, and running the trained denoising model MD2 using the SEM image as an input image to generate a denoised SEM image.

[0049]

[0063] In one embodiment, the second machine learning model MD2 may also be trained using the GAN training method as discussed above, using the inputs discussed in process P203. For example, the first simulated image 201, image noise, and a reference image REF are used as training data. In one embodiment, the reference image REF may be the first simulated image 201.

[0050]

[0064] FIG. 8 illustrates an exemplary process for generating denoised SEM images using trained denoising model MD2. Exemplary SEM images 801 and 802 of a patterned substrate are captured via an SEM tool. Note that SEM images 801 and 802 have complex patterns that are very different from the patterns used in the images in the training data. Because trained model MD2 is trained based on simulated images related to the design patterns, it can advantageously cover a large number of patterns. Thus, trained model MD2 can generate highly accurate denoised images 811 and 812 of SEM images 801 and 802, respectively. The results in FIG. 8 demonstrate that trained model MD2 can handle new patterns without additional training. In one embodiment, further fine-tuning of denoising model MD2 can be performed using newly captured SEM images to further improve the quality of the denoised images.

[0051]

[0065] 3 is a flowchart of another exemplary method 300 for generating a denoising model according to an embodiment of the present disclosure. Method 300 includes processes P301, P303, and P305, which are discussed below.

[0052]

[0066] Process P301 includes obtaining a first set of simulated images 301 based on the design pattern. In one embodiment, each image in the first set of simulated images 301 is a combination of a simulated SEM image and image noise (e.g., S1 in FIG. 6). * , S2 * , …Sn * (See

[0053]

[0067] In one embodiment, obtaining the simulated SEM image includes running a trained model using the design pattern as input to generate the simulated SEM image. For example, running trained model MD1 as discussed with respect to FIG. 5 . In one embodiment, the trained model (e.g., MD1) is trained based on the design pattern and captured images of the patterned substrate, where each captured image is associated with the design pattern. In one embodiment, the captured images are SEM images obtained via a scanning electron microscope (SEM). In one embodiment, the image noise is noise extracted from the captured images of the patterned substrate. In one embodiment, the image noise is Gaussian noise, white noise, or salt and pepper noise characterized by user-specified parameters.

[0054]

[0068] In one embodiment, the trained model (e.g., MD1) is a first machine learning model. In one embodiment, the first machine learning model is a CNN or DCNN trained using a generative adversarial network. In one embodiment, the trained model MD1 is a generative model configured to generate simulated SEM images for a given design pattern. For example, trained model MD1 as discussed with respect to FIGS. 5 and 6. In one embodiment, the reference denoised images are simulated SEM images associated with the design pattern. For example, the reference images can be S1, S2, ... Sn generated by MD1 of FIG. 5.

[0055]

[0069] Process P303 includes providing the first set of simulated images 301 as input to a base denoising model BM1 to obtain an initial second set of simulated images, where the initial second set of simulated images are denoised images associated with a design pattern. In one embodiment, the base model can be an untrained model or a trained model that needs to be fine-tuned. In one embodiment, captured images of the patterned substrate can be used to train or fine-tune the denoising model. Process P305 includes updating one or more configurations of the base denoising model BM1 using the reference denoising image as feedback. The one or more configurations are updated based on a comparison between the reference denoising image and the second set of simulated images. For example, updating the one or more configurations includes modifying model parameters of the base model. At the end of the training process, the base model with the updated configuration becomes the denoising model. Such a denoising model can, for example, use an SEM image as input to generate the second set of simulated images.

[0056]

[0070] In an embodiment, the denoising model is a second machine learning model. In an embodiment, the second deep learning model is trained using a deep learning method or a machine learning method. In an embodiment, the denoising model may also be trained using a generative adversarial network training method. In an embodiment, the denoising model is a convolutional neural network or other machine learning model. In an embodiment, the denoising model is MD2, as discussed with respect to FIG. 7.

[0057]

[0071] As discussed herein, an example of a denoising model is a machine learning model. Both unsupervised and supervised machine learning models can be used to generate denoised images from input noisy images, such as SEM images of patterned substrates. Without limiting the scope of the present invention, the application of supervised machine learning algorithms is described below.

[0058]

[0072] Supervised learning is the machine learning task of inferring a function from labeled training data. The training data includes a set of training examples. In supervised learning, each example is a pair of an input object (typically a vector) and a desired output value (also called a supervisory signal). A supervised learning algorithm analyzes the training data and generates an inferred function that can be used to map new examples. The optimal scenario would allow the algorithm to correctly determine the class label of unseen instances. This requires the learning algorithm to generalize from the training data to unseen situations in a "reasonable way."

[0059]

[0073] x i is the feature vector of the i-th example, and y i is its label (i.e., class) {(x1,y1), (x2,y2), ..., (x N ,y N Given a set of N training examples of the form {\displaystyle \mathbb {X}}, a learning algorithm finds a function g:X→Y, where X is the input space and Y is the output space. A feature vector is an n-dimensional vector of numerical features that represent some object. Many algorithms in machine learning require a numerical representation of an object because it facilitates processing and statistical analysis. When representing an image, the feature values ​​might correspond to the pixels of the image; when representing text, the feature values ​​might correspond perhaps to term occurrence frequencies. The vector space associated with these vectors is often called the feature space. The function g is an element of some space of possible functions G, usually called the hypothesis space. g is the function that finds the highest score:

number

number

[0060]

[0074] G and F can be any space of functions, but many learning algorithms are probabilistic models where g takes the form of a conditional probability model g(x) = P(y|x) or f takes the form of a joint probability model f(x,y) = P(x,y). For example, Naive Bayes and Linear Discriminant Analysis are joint probability models, and logistic regression is a conditional probability model.

[0061]

[0075] There are two basic approaches to selecting f or g: empirical risk minimization and structural risk minimization. Empirical risk minimization finds the function that best fits the training data. Structural risk minimization includes a penalty function that controls the bias / variance tradeoff.

[0062]

[0076] In both cases, the training set consists of independent and identical distribution pairs (x i ,y i ) samples. To measure how well the function fits the training data, we use a loss function L:

number

number

number

[0063]

[0077] The risk R(g) of a function g is defined as the expected loss of g.

number

[0064]

[0078] Exemplary models of supervised learning include decision trees, ensemble methods (bagging, boosting, random forests), k-NN, linear regression, naive Bayes, neural networks, logistic regression, perception, support vector machines (SVM), relevance vector machines (RVM), and deep learning.

[0065]

[0079] SVM is an example of a supervised learning model that can analyze data, recognize patterns, and be used for classification and regression analysis. Given a set of training examples, each marked as belonging to one of two categories, an SVM training algorithm builds a model that assigns new examples to one or the other category, making it a non-probabilistic binary linear classifier. An SVM model is a representation of examples as points in space, mapped so that examples from different categories are separated by as wide an apparent gap as possible. New examples are then mapped into the same space and predicted to belong to a category based on which side of the gap they fall on.

[0066]

[0080] In addition to performing linear classification, SVMs can efficiently perform non-linear classification using so-called kernel methods to implicitly map their inputs into a high-dimensional feature space.

[0067]

[0081] Kernel methods require a user-specified kernel, i.e., a similarity function for pairs of data points in the raw representation. Kernel methods get their name from the use of kernel functions, which allow them to operate in a high-dimensional implicit feature space without ever computing the coordinates of the data in that space, but simply by computing the dot product between the images of all pairs of data in the feature space. This operation is often computationally cheaper than explicitly computing the coordinates. This technique is called the "kernel trick."

[0068]

[0082] The effectiveness of an SVM depends on the choice of kernel, the kernel parameters, and the soft margin parameter C. A common choice is a Gaussian kernel with a single parameter γ. The best combination of C and γ is often an exponentially increasing sequence of C and γ (e.g., C∈{2 -5 , 2 -4 , …, 2 15 , 2 16};γ∈{2 -15 , 2 -14 , …, 2 4 , 2 5}) by grid search (also known as "parameter sweep").

[0069]

[0083] Grid search is an exhaustive search through a manually specified subset of a learning algorithm's hyperparameter space. Grid search algorithms are guided by some performance metric, typically measured by cross-validation on a training set or evaluation on a provided validation set.

[0070]

[0084] Each combination of parameter selection may be checked using cross-validation, and the parameters with the best cross-validation accuracy are chosen.

[0071]

[0085] Cross-validation, sometimes called rotational estimation, is a model validation technique that evaluates how the results of a statistical analysis generalize to an independent dataset. It is primarily used in situations where the goal is prediction and one wants to estimate how accurately a predictive model will perform in practice. In prediction problems, a model is typically given a dataset of known data (a training dataset) on which training is performed and a dataset of unknown (or unseen) data (a test dataset) on which the model is tested. The goal of cross-validation is to define a dataset (i.e., a validation dataset) to "test" the model during the training phase, to limit problems like overfitting and provide insight into how the model generalizes to an independent dataset (i.e., an unknown dataset from a real problem, for example). One round of cross-validation involves partitioning data samples into complementary subsets, conducting an analysis on one subset (called the training set), and validating the analysis on the other subset (called the validation or test set). To reduce variability, multiple rounds of cross-validation are performed using different partitions, and the validation results across the rounds are averaged.

[0072]

[0086] A final model that can be used for testing and to classify new data is then trained on the entire training set using the selected parameters.

[0073]

[0087] Another example of supervised learning is regression. Regression infers the relationship between a dependent variable and one or more independent variables from a set of values ​​of the dependent variable and corresponding values ​​of the independent variables. Regression may estimate the conditional expectation of the dependent variable given the independent variables. The inferred relationship is sometimes called a regression function. The inferred relationship may be probabilistic.

[0074]

[0088] In one embodiment, a system is provided that can generate a de-noised image using model MD2 after the system captures an image of a patterned substrate. In one embodiment, the system can be, for example, the SEM tool of FIG. 9 or the inspection tool of FIG. 10 configured to include models MD1 and / or MD2 discussed herein. For example, the metrology tool includes an electron beam generator for capturing an image of the patterned substrate and one or more processors that include the MD1 and MD2 models. The one or more processors are configured to execute a trained model configured to generate a simulated image based on a design pattern used to pattern the substrate, and to execute a de-noising model using the captured image and the simulated image as inputs to generate a de-noised image of the patterned substrate. As previously mentioned, the de-noising model (e.g., MD2) is a convolutional neural network.

[0075]

[0089] Additionally, in one embodiment, the one or more processors are further configured to update the denoising model based on the captured image of the patterned substrate, hi one embodiment, updating the denoising model includes running the denoising model using the captured image to generate a denoised image, and updating one or more parameters of the denoising model based on a comparison of the denoised image and a reference denoised image.

[0076]

[0090] The present disclosure is not limited to applications in which denoised images are used. In the semiconductor industry, denoised images can be used, for example, for inspection and metrology. In one embodiment, the denoised image can be used to determine hot spots on a patterned substrate. The hot spots are determined based on absolute CD values ​​measured from the denoised image. Alternatively, the hot spots can be determined based on a set of predetermined rules, such as those used in a design rule checking system, including, but not limited to, line end pullback, corner rounding, proximity of adjacent features, necking or pinching of the pattern, and other metrics of pattern deformation relative to the desired pattern.

[0077]

[0091] In one embodiment, the denoised image can be used to improve the patterning process. For example, the denoised image can be used in a simulation of the patterning process to predict, for example, contours, CDs, edge placement (e.g., edge placement error), etc. in the resist and / or etched image. The goal of the simulation is to accurately predict, for example, the edge placement and / or aerial image intensity gradient and / or CD of the printed pattern. These values ​​can be compared to the intended design to, for example, correct the patterning process, identify where defects are expected to occur, etc. The intended design is typically defined as a pre-OPC design layout and can be provided in a standard digital file format such as GDSII or OASIS, or other file format.

[0078]

[0092] In some embodiments, the inspection or metrology tool may be a scanning electron microscope (SEM) that obtains an image of the exposed or transferred structure on the substrate (e.g., some or all of the structures of a device). Figure 9 depicts an embodiment of an SEM tool. A primary electron beam EBP emitted from an electron source ESO is focused by a condenser lens CL and then passes through a beam deflector EBD1, an ExB deflector EBD2, and an objective lens OL to illuminate at a focal point a substrate PSub on a substrate table ST.

[0079]

[0093] When the substrate PSub is irradiated with the electron beam EBP, secondary electrons are generated from the substrate PSub. The secondary electrons are deflected by the E×B deflector EBD2 and detected by the secondary electron detector SED. For example, a two-dimensional electron beam image can be obtained by detecting the electrons generated from the sample in synchronization with two-dimensional scanning of the electron beam by the beam deflector EBD1, or with repeated scanning of the electron beam EBP by the beam deflector EBD1 in the X or Y direction together with continuous movement of the substrate PSub by the substrate table ST in the other of the X or Y directions.

[0080]

[0094] The signal detected by the secondary electron detector SED is converted into a digital signal by an analog-to-digital (A / D) converter ADC, and the digital signal is sent to an image processing system IPU. In one embodiment, the image processing system IPU may have a memory MEM for storing all or part of the digital image for processing by the processing unit PU. The processing unit PU (e.g., specially designed hardware or a combination of hardware and software) is configured to convert or process the digital image into a data set representative of the digital image. Furthermore, the image processing system IPU may have a storage medium STOR configured to store the digital image and the corresponding data set in a reference database. A display device DIS may be connected to the image processing system IPU, so that an operator can perform the required operations of the equipment with the aid of a graphical user interface.

[0081]

[0095] As described above, SEM images can be processed to extract contours that depict the edges of objects in the image, representing device structures. These contours are quantified using metrics such as CD. Therefore, images of device structures are typically compared and quantified using simple metrics, such as edge-to-edge distance (CD) or simple pixel differences between images. Typical contour models for detecting object edges in images to measure CD use image gradients. Indeed, these models rely on strong image gradients. However, in reality, images typically have noisy and discontinuous boundaries. Techniques such as smoothing, adaptive thresholding, edge detection, erosion, and dilation can be used to process the results of image gradient contour models to address noisy and discontinuous images, but these techniques ultimately result in low-resolution quantification of high-resolution images. Therefore, mathematical manipulation of device structure images to reduce noise and automate edge detection often leads to a loss of image resolution, thereby resulting in a loss of information. The result, therefore, is low-resolution quantification, resulting in a simplified representation of complex, high-resolution structures.

[0082]

[0096] Thus, it is desirable to have a mathematical representation of a structure (e.g., a circuit feature, alignment mark, or metrology target portion (e.g., grating feature), etc.) that is produced or expected to be produced using a patterning process, whether the structure is in a latent resist image, in a developed resist image, or transferred, e.g., by etching, to a layer on a substrate that maintains resolution and yet can represent the general shape of the structure. In the context of lithography or other patterning processes, the structure may be a device or portion thereof under fabrication, and the image may be an SEM image of the structure. In some cases, the structure may be a feature of a semiconductor device, e.g., an integrated circuit. In this case, the structure may be referred to as a pattern including multiple features of the semiconductor device, or as a desired pattern. In some cases, the structure may be an alignment mark, or a portion thereof (e.g., a grating of an alignment mark), used in an alignment measurement process to determine the alignment of an object (e.g., a substrate) with another object (e.g., a patterning device), or may be a metrology target, or a portion thereof (e.g., a grating of a metrology target), used to measure a parameter of the patterning process (e.g., overlay, focus, dose, etc.) In one embodiment, the metrology target is a diffraction grating, for example used to measure overlay.

[0083]

[0097] 10 schematically illustrates a further embodiment of an inspection apparatus. The system is used to inspect a sample 90 (such as a substrate) on a sample stage 88 and includes a charged particle beam generator 81, a condenser lens module 82, a probe-forming objective lens module 83, a charged particle beam deflection module 84, a secondary charged particle detector module 85, and an image formation module 86.

[0084]

[0098] The charged particle beam generator 81 generates a primary charged particle beam 91. The condenser lens module 82 collects the generated primary charged particle beam 91. The probe forming objective lens module 83 focuses the collected primary charged particle beam into a charged particle beam probe 92. The charged particle beam deflection module 84 scans the formed charged particle beam probe 92 across the surface of an area of ​​interest on a sample 90 fixed on a sample stage 88. In one embodiment, the charged particle beam generator 81, the condenser lens module 82, and the probe forming objective lens module 83, or equivalent designs, alternatives, or any combination thereof, together form a charged particle beam probe generator that generates the scanning charged particle beam probe 92.

[0085]

[0099] The secondary charged particle detector module 85 detects secondary charged particles 93 emitted from the sample surface (possibly along with other charged particles reflected or scattered from the sample surface) and generates a secondary charged particle detection signal 94 upon being struck by the charged particle beam probe 92. The imaging module 86 (e.g., a computing device) is coupled to the secondary charged particle detector module 85 to receive the secondary charged particle detection signal 94 from the secondary charged particle detector module 85 and, in response, form at least one scanned image. In one embodiment, the secondary charged particle detector module 85 and the imaging module 86, or equivalent designs, alternatives, or any combination thereof, together form an imaging device that forms a scanned image from the detected secondary charged particles emitted from the sample 90 struck by the charged particle beam probe 92.

[0086]

[0100] In one embodiment, the monitoring module 87 is coupled to the imaging module 86 of the imaging device to use scanned images of the sample 90 received from the imaging module 86 to monitor and control the patterning process and / or to derive parameters for designing, controlling, monitoring, etc., the patterning process. As such, in one embodiment, the monitoring module 87 is configured or programmed to perform the methods described herein. In one embodiment, the monitoring module 87 comprises a computing device. In one embodiment, the monitoring module 87 comprises a computer program for providing the functionality described herein, the computer program encoded on a computer-readable medium forming or disposed within the monitoring module 87.

[0087]

[0101] In one embodiment, similar to the electron beam inspection tool of Figure 9 that uses a probe to inspect a substrate, the electron current in the system of Figure 10 is significantly larger than, for example, a CD-SEM as depicted in Figure 9, so the probe spot may be sufficiently large and, as a result, the inspection speed may be fast. However, the resolution may not be as high as a CD-SEM due to the large probe spot. In one embodiment, the inspection apparatus discussed above may be a single beam or a multi-beam apparatus without limiting the scope of the present disclosure.

[0088]

[0102] For example, SEM images from the systems of Figures 9 and / or 10 can be processed to extract contours that delineate the edges of objects in the images, representing device structures. These contours are then typically quantified using metrics such as CD at user-defined cut lines. Images of device structures are therefore typically compared and quantified using metrics such as edge-to-edge distance (CD) measured on the extracted contours or simple pixel differences between images.

[0089]

[0103] In one embodiment, one or more procedures of processes 200 and / or 300 may be implemented as instructions (e.g., program code) in a processor of a computer system (e.g., process 104 of computer system 100). In one embodiment, the procedures may be distributed across multiple processors (e.g., parallel computing) to increase computational efficiency. In one embodiment, a computer program product including a non-transitory computer-readable medium has instructions stored on the non-transitory computer-readable medium, which, when executed by a computer hardware system, perform the methods described herein.

[0090]

[0104] According to the present disclosure, combinations and subcombinations of the disclosed elements constitute separate embodiments. For example, a first combination may include determining a denoising model based on a simulated image relating to the design pattern and the noisy image. A subcombination may include determining the denoising image using the denoising model. In another combination, the denoising image may be used in an inspection process to determine OPC or SMO based on variance data generated with the model. In another example, the combination may include determining process adjustments to a lithography process, a resist process, or an etch process based on inspection data based on the denoising image to improve yield of the patterning process.

[0091]

[0105] 11 is a block diagram illustrating a computer system 100 that can assist in implementing the methods, flows, or apparatuses disclosed herein. The computer system 100 includes a bus 102 or other communication mechanism for communicating information and a processor 104 (or multiple processors 104 and 105) coupled with the bus 102 for processing information. The computer system 100 also includes a main memory 106, such as a random access memory (RAM) or other dynamic storage device, coupled to the bus 102 for storing information and instructions executed by the processor 104. The main memory 106 may also be used for storing temporary variables or other intermediate information during execution of instructions executed by the processor 104. The computer system 100 further includes a read-only memory (ROM) 108 or other static storage device coupled to the bus 102 for storing static information and instructions for the processor 104. A storage device 110, such as a magnetic or optical disk, is provided and coupled to the bus 102 for storing information and instructions.

[0092]

[0106] Computer system 100 may be coupled via bus 102 to a display 112, such as a cathode ray tube (CRT), flat panel, or touch panel display, for displaying information to a computer user. An input device 114, including alphanumeric and other keys, is coupled to bus 102 for communicating information and command selections to processor 104. Another type of user input device is a cursor control 116, such as a mouse, trackball, or cursor direction keys, for communicating directional information and command selections to processor 104 and for controlling cursor movement on display 112. This input device typically has two degrees of freedom, allowing the device to be positioned in two axes—a first axis (e.g., x) and a second axis (e.g., y)—within a plane. A touch panel (screen) display may also be used as an input device.

[0093]

[0107] According to some embodiments, portions of one or more methods herein may be performed by computer system 100 in response to processor 104 executing one or more sequences of one or more instructions contained in main memory 106. Such instructions may be read into main memory 106 from another computer-readable medium, such as storage device 110. Execution of the sequences of instructions contained in main memory 106 causes processor 104 to perform the process steps described herein. One or more processors in a multi-processing arrangement may be used to execute the sequences of instructions contained in main memory 106. In some alternative embodiments, hard-wired circuitry may be used in place of or in combination with software instructions. Thus, the description herein is not limited to any specific combination of hardware circuitry and software.

[0094]

[0108] The term "computer-readable medium," as used herein, refers to any medium that participates in providing instructions to processor 104 for execution. Such media may take many forms, including, but not limited to, non-volatile media, volatile media, and transmission media. Non-volatile media include, for example, optical or magnetic disks, such as storage device 110. Volatile media include dynamic memory, such as main memory 106. Transmission media include coaxial cables, copper wire, and fiber optics (including the wires that comprise bus 102). Transmission media may also take the form of acoustic or light waves, such as those generated during radio frequency (RF) and infrared (IR) data communications. Common forms of computer-readable media include, for example, floppy disks, flexible disks, hard disks, magnetic tape, other magnetic media, CD-ROMs, DVDs, other optical media, punch cards, paper tape, other physical media with patterns of holes, RAM, PROMs, and EPROMs, FLASH-EPROMs, other memory chips or cartridges, carrier waves, as described below, or other computer-readable media.

[0095]

[0109] Various forms of computer-readable media may be involved in carrying one or more sequences of one or more instructions to processor 104 for execution. For example, the instructions may initially reside on a magnetic disk of a remote computer. The remote computer may load the instructions into its dynamic memory and send the instructions over a telephone line using a modem. A modem local to computer system 100 may receive the data on the telephone line and use an infrared transmitter to convert the data to an infrared signal. An infrared detector coupled to bus 102 may receive the data carried in the infrared signal and place the data on bus 102. Bus 102 carries the data to main memory 106, from which processor 104 retrieves and executes the instructions. The instructions received by main memory 106 may optionally be stored on storage device 110 either before or after execution by processor 104.

[0096]

[0110] Computer system 100 may also include a communication interface 118 coupled to bus 102. The communication interface 118 provides a two-way data communication coupling to a network link 120 that is connected to a local network 122. For example, communication interface 118 may be an Integrated Services Digital Network (ISDN) card or a modem to provide a data communication connection to a corresponding type of telephone line. As another example, communication interface 118 may be a local area network (LAN) card to provide a data communication connection to a compatible LAN. A wireless link may also be implemented. In such an implementation, communication interface 118 sends and receives electrical, electromagnetic, or optical signals that carry digital data streams representing various types of information.

[0097]

[0111] Network link 120 typically provides data communication through one or more networks to other data devices. For example, network link 120 may provide a connection through local network 122 to a host computer 124 or to data equipment operated by an Internet Service Provider (ISP) 126. ISP 126 in turn provides data communication services through the world wide packet data communication network (now commonly referred to as the "Internet" 128). Local network 122 and Internet 128 both use electrical, electromagnetic, or optical signals that carry digital data streams. The signals through the various networks and on network link 120 and through communication interface 118, which carry the digital data to and from computer system 100, are exemplary forms of carrier waves transporting the information.

[0098]

[0112] Computer system 100 can send messages and receive data, including program code, through one or more networks, network link 120, and communication interface 118. In the Internet example, server 130 might send a requested code for an application program through Internet 128, ISP 126, local network 122, and communication interface 118. Such a downloaded application may provide all or a portion of the methods herein. The received code may be executed by processor 104 as received, and / or stored in storage device 110, or other non-volatile storage for later execution. In this manner, computer system 100 may obtain the application code in the form of a carrier wave.

[0099]

[0113] 12 schematically depicts an exemplary lithographic projection apparatus that can be used in conjunction with the techniques described herein. The apparatus includes: - an illumination system IL for conditioning a radiation beam B. In this particular case, the illumination system also comprises a radiation source SO; - a first object table (e.g., patterning device table) MT comprising a patterning device holder for holding a patterning device MA (e.g., a reticle) and connected to a first positioner for accurately positioning the patterning device relative to the item PS; a second object table (substrate table) WT comprising a substrate holder for holding a substrate W (e.g. a resist-coated silicon wafer) and connected to a second positioner for accurately positioning the substrate relative to the item PS; a projection system ("lens") PS (e.g. a refractive, reflective, or catadioptric optical system) that images an illuminated portion of the patterning device MA onto a target portion C (e.g. comprising one or more dies) of the substrate W.

[0100]

[0114] As depicted herein, the apparatus is of a transmissive type (i.e. has a transmissive patterning device). In general, however, it may be of a reflective type (e.g. have a reflective patterning device). The apparatus may use other types of patterning device than a conventional mask; examples include a programmable mirror array or an LCD matrix.

[0101]

[0115] A source SO (e.g. a mercury lamp or excimer laser, LPP (Laser Produced Plasma) EUV source) produces a radiation beam. This beam is fed to an illumination system (illuminator) IL, either directly or after having traversed conditioning means, such as a beam expander Ex. The illuminator IL may comprise conditioning means AD for setting the outer and / or inner radial extent (commonly referred to as σ-outer and σ-inner, respectively) of the intensity distribution of the beam. Furthermore, it will generally comprise various other components, such as an integrator IN and a condenser CO. In this way, the beam B impinging on the patterning device MA has a desired uniformity and intensity distribution in its cross-section.

[0102]

[0116] It should be noted, with regard to Figure 12, that the source SO may be located within the housing of the lithographic projection apparatus (as is often the case when the source SO is, for example, a mercury lamp), but it may also be remote from the lithographic projection apparatus and the radiation beam that it produces may be directed into the apparatus (for example by the use of appropriate directing mirrors). This latter scenario is often the case when the source SO is an excimer laser (for example based on KrF, ArF or F2 lasing).

[0103]

[0117] Subsequently, beam PB intersects the patterning device MA, which is held on a patterning device table MT. After traversing the patterning device MA, beam B passes through a lens PL, which focuses beam B onto a target portion C of the substrate W. Using the second positioning means (and interferometric measurement means IF), the substrate table WT can be precisely moved, e.g., to position a different target portion C in the path of beam PB. Similarly, the first positioning means can be used to precisely position the patterning device MA with respect to the path of beam B, e.g., after mechanical retrieval of the patterning device MA from a patterning device library, or during a scan. In general, movement of the object table MT, WT, is realized using a long-stroke module (coarse positioning) and a short-stroke module (fine positioning), which are not explicitly depicted in FIG. 12 . However, in the case of a stepper (as opposed to a step-and-scan tool), the patterning device table MT may be connected to a short-stroke actuator only, or may be fixed.

[0104]

[0118] The depicted tool can be used in two different modes: - in step mode, the patterning device table MT remains essentially stationary and the entire patterning device image is projected onto the target portion C in one go (i.e. in a single "flash"), and the substrate table WT is then shifted in the x and / or y directions so that a different target portion C can be irradiated by the beam PB; In scan mode, essentially the same scenario applies, except that a given target portion C is not exposed in a single "flash". Instead, the patterning device table MT is movable in a given direction (the so-called "scan direction", e.g. the y direction) with a speed v, such that the projection beam B is caused to scan over the patterning device image. In parallel, the substrate table WT is simultaneously moved in the same or opposite direction with a speed V = Mv, in which M is the magnification factor of the lens PL (typically, M = 1 / 4 or 1 / 5). In this way, a relatively large target portion C can be exposed, without having to compromise on resolution.

[0105]

[0119] FIG. 13 schematically depicts another exemplary lithographic projection apparatus LA that may be utilized in conjunction with the techniques described herein.

[0106]

[0120] The lithographic projection apparatus LA includes: - Source Collector Module SO an illumination system (illuminator) IL configured to condition a radiation beam B (e.g., EUV radiation); a support structure (e.g., patterning device table) MT constructed to support a patterning device (e.g., a mask or reticle) MA and connected to a first positioner PM configured to accurately position the patterning device; a substrate table (e.g., wafer table) WT constructed to hold a substrate (e.g., a resist-coated wafer) W and connected to a second positioner PW configured to accurately position the substrate; and a projection system (e.g. a reflective projection system) PS configured to project a pattern imparted to the radiation beam B by the patterning device MA onto a target portion C (e.g. comprising one or more dies) of the substrate W.

[0107]

[0121] As depicted here, the apparatus LA is reflective (e.g., uses a reflective patterning device). Note that because most materials are absorptive in the EUV wavelength range, the patterning device can have a multilayer reflector including, for example, a multistack of molybdenum and silicon. In one example, a multistack reflector has 40 layer pairs of molybdenum and silicon, with each layer being a quarter wavelength thick. Even smaller wavelengths can be produced using x-ray lithography. Because most materials are absorptive at EUV and x-ray wavelengths, a thin strip of patterned absorbing material (e.g., TaN absorber on a multilayer reflector) on the patterning device topography defines where features will print (positive resist) or not print (negative resist).

[0108]

[0122] Referring to FIG. 13 , the illuminator IL receives an extreme ultraviolet radiation (EUV) beam from a source collector module SO. Methods for generating EUV radiation include, but are not limited to, converting a material into a plasma state having at least one element (e.g., xenon, lithium, or tin) with one or more emission lines in the EUV range. In one such method, often referred to as laser-produced plasma (“LPP”), the plasma can be generated by irradiating a fuel, such as droplets, streams, or clusters of material having a line-emitting element, with a laser beam. The source collector module SO may be part of an EUV radiation system that includes a laser (not shown in FIG. 13 ) that provides a laser beam that excites the fuel. The resulting plasma emits output radiation (e.g., EUV radiation), which is collected using a radiation collector disposed in the source collector module. The laser and source collector module may be separate entities, for example, when a CO laser is used to provide the laser beam for fuel excitation.

[0109]

[0123] In such cases, the laser is not considered to form part of the lithographic apparatus, and the radiation beam is passed from the laser to the source collector module using a beam delivery system, for example comprising appropriate directing mirrors and / or beam expanders. In other cases, the source may be an integral part of the source collector module, for example when the source is a discharge produced plasma EUV generator, often referred to as a DPP source.

[0110]

[0124] The illuminator IL may include an adjuster for adjusting the angular intensity distribution of the radiation beam. Generally, at least the outer and / or inner radial extent (commonly referred to as σ-outer and σ-inner, respectively) of the intensity distribution in a pupil plane of the illuminator may be adjusted. In addition, the illuminator IL may include various other components, such as faceted field and pupil mirror devices. The illuminator may be used to condition the radiation beam to have a desired uniformity and intensity distribution in its cross-section.

[0111]

[0125] The radiation beam B is incident on a patterning device (e.g., mask) MA, which is held on a support structure (e.g., patterning device table) MT, and is patterned by the patterning device. After reflecting from the patterning device (e.g., mask) MA, the radiation beam B passes through a projection system PS, which focuses the beam onto a target portion C of a substrate W. The substrate table WT can be accurately moved using a second positioner PW and a position sensor PS2 (e.g., an interferometric device, a linear encoder, or a capacitive sensor), for example, to position a different target portion C in the path of the radiation beam B. Similarly, the first positioner PM and another position sensor PS1 can be used to accurately position the patterning device (e.g., mask) MA with respect to the path of the radiation beam B. The patterning device (e.g., mask) MA and substrate W may be aligned using patterning device alignment marks M1, M2 and substrate alignment marks P1, P2.

[0112]

[0126] The depicted apparatus LA could be used in at least one of the following modes: 1. In step mode, the support structure (e.g. patterning device table) MT and the substrate table WT remain essentially stationary while an entire pattern imparted to the radiation beam is projected onto a target portion C in one go (i.e. a single static exposure), where the substrate table WT is then shifted in the X and / or Y directions so that a different target portion C can be exposed. 2. In scan mode, the support structure (e.g. patterning device table) MT and the substrate table WT are scanned synchronously (i.e. single dynamic exposure) while a pattern imparted to the radiation beam is projected onto a target portion C. The velocity and direction of the substrate table WT relative to the support structure (e.g. patterning device table) MT may be determined by the de-magnification and image reversal characteristics of the projection system PS. 3. In another mode, the support structure (e.g. patterning device table) MT holds a programmable patterning device and remains essentially stationary, and the substrate table WT is moved or scanned, while a pattern imparted to the radiation beam is projected onto a target portion C. In this mode, a pulsed radiation source is generally used, and the programmable patterning device is updated as required after each movement of the substrate table WT, or between successive radiation pulses during a scan. This mode of operation is readily adaptable to maskless lithography, employing a programmable patterning device such as a programmable mirror array of the type referred to above.

[0113]

[0127] FIG. 14 shows in more detail the apparatus LA, including the source collector module SO, the illumination system IL, and the projection system PS. The source collector module SO is constructed and arranged such that a vacuum environment can be maintained within the enclosure 220 of the source collector module SO. The EUV radiation-emitting plasma 210 can be formed by a discharge-produced plasma source. The EUV radiation can be generated by a gas or vapor (e.g., Xe gas, Li vapor, or Sn vapor, in which a very hot plasma 210 is created to emit radiation in the EUV range of the electromagnetic spectrum). The very hot plasma 210 is created, for example, by a discharge that produces an at least partially ionized plasma. A partial pressure of, for example, 10 Pa of Xe, Li, Sn vapor, or any other suitable gas or vapor may be required for efficient generation of radiation. In an embodiment, a plasma of excited tin (Sn) is provided to generate the EUV radiation.

[0114]

[0128] Radiation emitted by the high-temperature plasma 210 passes from the source chamber 211 into the collector chamber 212 through an optional gas barrier or contaminant trap 230 (sometimes also referred to as a contaminant barrier or foil trap) located in or behind the opening of the source chamber 211. The contaminant trap 230 may include a channel structure. The contaminant trap 230 may also include a gas barrier or a combination of a gas barrier and a channel structure. The contaminant trap or contaminant barrier 230 further described herein includes at least a channel structure, as known in the art.

[0115]

[0129] The collector chamber 211 may include a radiation collector CO, which may be a so-called grazing incidence collector. The radiation collector CO has an upstream radiation collector side 251 and a downstream radiation collector side 252. Radiation traversing the collector CO may be reflected off a grating spectral filter 240 to be focused to a virtual source point IF along the optical axis indicated by the dash-dotted line "O". The virtual source point IF is commonly called the intermediate focus, and the source collector module is positioned such that the intermediate focus IF is located at or near the opening 221 of the enclosure structure 220. The virtual source point IF is an image of the radiation-emitting plasma 210.

[0116]

[0130] The radiation subsequently traverses an illumination system IL, which may include a faceted field mirror device 22 and a facetted pupil mirror device 24 arranged to provide a desired angular distribution of the radiation beam 21 at the patterning device MA, and a desired uniformity of the radiation intensity at the patterning device MA. Upon reflection of the radiation beam 21 off the patterning device MA, which is held by a support structure MT, a patterned beam 26 is formed, and the patterned beam 26 is imaged by the projection system PS via reflective elements 28, 30 onto a substrate W held by a substrate table WT.

[0117]

[0131] Generally, more elements than shown may be present in illumination optics unit IL and projection system PS. A grating spectral filter 240 may optionally be present, depending on the type of lithographic apparatus. Furthermore, more mirrors than shown in the figures may be present, for example, 1 to 6 additional reflective elements may be present in projection system PS than shown in Figure 14.

[0118]

[0132] 14 is depicted as a nested collector with grazing incidence reflectors 253, 254, and 255, just as an example of a collector (or collector mirror). The grazing incidence reflectors 253, 254, and 255 are arranged axisymmetrically about the optical axis O, and this type of collector system CO can be used in combination with a discharge produced plasma source, often referred to as a DPP source.

[0119]

[0133] Alternatively, the source collector module SO may be part of an LPP radiation system, as shown in Figure 15. A laser LA is arranged to deposit laser energy into a fuel such as xenon (Xe), tin (Sn), or lithium (Li) to create a highly ionized plasma 210 with an electron temperature of tens of eV. Energetic radiation produced during de-excitation and recombination of these ions is emitted from the plasma and collected by a near-normal incidence collector system CO and focused onto an opening 221 in an enclosure structure 220.

[0120]

[0134] The concepts disclosed herein can be used to simulate or mathematically model general imaging systems for imaging subwavelength features and can be particularly useful for new imaging technologies capable of producing shorter wavelengths. New technologies already in use include EUV (extreme ultraviolet), DUV lithography, which can produce wavelengths of 193 nm using ArF lasers and even 157 nm using fluorine lasers. EUV lithography can also produce wavelengths within this range by using synchrotrons or by bombarding materials (solids or plasmas) with high-energy electrons to generate photons within the 20-5 nm range.

[0121]

[0135] Embodiments of the present disclosure can be further described by the following clauses. 1. One or more non-transitory computer-readable media storing a denoising model and generating a denoised image with denoising model instructions that, when executed by one or more processors, provide a denoising model, the denoising model comprising: Transforming the design patterns into a first set of simulated images; providing the first set of simulated images as inputs to a base denoising model to obtain a second set of simulated images, the second set of simulated images being denoised images associated with the design pattern; updating one or more configurations of the base denoised model using the reference denoised image as feedback, the one or more configurations being updated based on a comparison of the reference denoised image with the second set of simulated images; and A non-transitory computer-readable medium generated by 2. The medium of clause 1, wherein each image in the first set of simulated images is a combination of a simulated SEM image and image noise. 3. The medium of clause 2, wherein transforming the design pattern includes running the trained model using the design pattern as input to generate the simulated SEM image. 4. The medium of clause 3, wherein the trained model is trained based on design patterns and captured images of the patterned substrate, each captured image being associated with a design pattern. 5. The medium of clause 4, wherein the captured image is a scanning electron microscope (SEM) image obtained via an SEM. 6. The medium of clause 5, wherein the image noise is noise extracted from a captured image of the patterned substrate. 7. The medium of any one of clauses 3-6, wherein the trained model is a first machine learning model. 8. The medium of clause 7, wherein the trained model is a convolutional neural network or a deep convolutional neural network trained using a generative adversarial network. 9. The medium of clause 8, wherein the trained model is a generative model configured to generate simulated SEM images for a given design pattern. 10. The medium of any one of clauses 2-9, wherein the image noise is Gaussian noise, white noise, or salt and pepper noise characterized by user-specified parameters. 11. The medium of any one of clauses 2 to 10, wherein the reference denoised image is a simulated SEM image associated with the design pattern. 12. The medium of any one of clauses 1-11, wherein the denoising model is a second machine learning model. 13. The medium of any one of clauses 1 to 12, wherein the denoising model is a convolutional neural network or a deep convolutional neural network. 14. The medium of any one of clauses 1 to 13, wherein the design pattern is in a graphics data signal (GDS) file format. 15. Electron beam optics configured to capture an image of the patterned substrate; one or more processors configured to execute a denoising model using the captured image as input to generate a denoised image of the patterned substrate; Including, the system. 16. The system of clause 15, wherein the denoising model is a convolutional neural network. 17. The system of clause 15 or 16, wherein the one or more processors are further configured to execute the trained model using design patterns provided in a graphics data signal (GDS) file format to generate a simulated image. 18. The system of any one of clauses 15-17, wherein the one or more processors are further configured to update the denoising model based on the captured image of the patterned substrate. 19. The system of any one of clauses 15 to 18, wherein the one or more processors are further configured to update one or more parameters of the denoising model based on a comparison of the denoised image and a reference denoised image. 20. The noise reduction model Transforming the design patterns into a first set of simulated images; providing the first set of simulated images as inputs to a base denoising model to obtain a second set of simulated images, the second set of simulated images being denoised images associated with the design pattern; updating one or more configurations of the base denoised model using the reference denoised image as feedback, the one or more configurations being updated based on a comparison of the reference denoised image with the second set of simulated images; and 20. The system of any one of clauses 15 to 19, 21. The system of clause 20, wherein each image in the first set of simulated images is a combination of a simulated SEM image and image noise. 22. A method for training a denoising model, comprising: Transforming the design patterns into a first set of simulated images; training a denoising model based on the first set of simulated images and the image noise, the denoising model operable to generate a denoised image of the input image; A method comprising: 23. The method of clause 22, wherein converting the design patterns into a first set of simulated images includes running a trained model using the design patterns as input to generate the simulated images. 24. The method of clause 23, wherein the trained model is trained based on design patterns and captured images of the patterned substrate, each captured image being associated with a design pattern. 25. The method of clause 24, wherein the captured image is a scanning electron microscope (SEM) image obtained via an SEM. 26. The method of clause 25, further comprising adding image noise to the first set of simulated images to generate a second set of simulated images, the image noise being extracted from a captured image of the patterned substrate. 27. The method of clause 26, wherein training the denoising model includes using a first set of simulated images, image noise and captured images as training data. 28. The method of any one of clauses 23 to 27, wherein the trained model is a first machine learning model. 29. The method of clause 28, wherein the trained model is a convolutional neural network or a deep convolutional neural network trained using a generative adversarial network training method. 30. The method of clause 29, wherein the trained model is a generative model configured to generate simulated SEM images for a given design pattern. 31. The method of any one of clauses 22 to 30, wherein the image noise is Gaussian noise, white noise, or salt and pepper noise, characterized by user-specified parameters. 32. The method of any one of clauses 23 to 31, wherein the denoising model is a second machine learning model. 33. The method of any one of clauses 22 to 32, wherein the denoising model is a convolutional neural network. 34. The method of any one of clauses 22 to 33, wherein the design pattern is in a graphics data signal (GDS) file format. 35. Obtaining an SEM image of the patterned substrate via a metrology tool; running the trained denoising model using the SEM image as an input image to generate a denoised SEM image; 25. The method of any one of clauses 22 to 24, further comprising: 36. One or more non-transitory computer-readable media storing a denoising model and generating a denoised image with denoising model instructions that, when executed by one or more processors, provide a denoising model, the denoising model comprising: Transforming the design patterns into a first set of simulated images; training a denoising model based on the first set of simulated images and the image noise, the denoising model operable to generate a denoised image of the input image; A non-transitory computer-readable medium generated by 37. The medium of clause 36, wherein converting includes running the trained model using the design pattern as input to generate the simulated image. 38. The medium of clause 37, wherein the trained model is trained based on design patterns and captured images of the patterned substrate, each captured image being associated with a design pattern. 39. The medium according to clause 38, wherein the captured image is a scanning electron microscope (SEM) image obtained via an SEM. 40. The medium of clause 39, further comprising adding image noise to the first set of simulated images to generate a second set of simulated images, the image noise being extracted from a captured image of the patterned substrate. 41. The medium of any one of clauses 37-40, wherein the trained model is a first machine learning model. 42. The medium of clause 41, wherein the trained model is a convolutional neural network or a deep convolutional neural network trained using a generative adversarial network training method. 43. The medium of clause 42, wherein the trained model is a generative model configured to generate simulated SEM images for a given design pattern. 44. The medium of any one of clauses 36-43, wherein the image noise is Gaussian noise, white noise, or salt and pepper noise characterized by user-specified parameters. 45. The medium of any one of clauses 36-44, wherein the denoising model is a second machine learning model. 46. ​​The medium of any one of clauses 38-45, wherein training the denoising model includes using a first set of simulated images, image noise, and captured images as training data. 47. The medium of any one of clauses 36-46, wherein the denoising model is a convolutional neural network. 48. The medium of any one of clauses 36 to 47, wherein the design pattern is in a graphics data signal (GDS) file format. 49. Obtaining an SEM image of the patterned substrate via a metrology tool; running the trained denoising model using the SEM image as an input image to generate a denoised SEM image; 49. The medium of any one of clauses 36 to 48, further comprising:

[0122]

[0136] The concepts disclosed herein may be used for imaging on substrates such as silicon wafers, although it is understood that the disclosed concepts may be used in any type of lithographic imaging system (e.g., one used for imaging on substrates other than silicon wafers).

[0123]

[0137] The above description is intended to be illustrative, not limiting. Thus, it will be apparent to one skilled in the art that modifications may be made as described without departing from the scope of the claims set out below.

Claims

1. 1. A method for training a denoising model, comprising: Transforming the design patterns into a first set of simulated images; training the denoising model based on the first set of simulated images and image noise, the denoising model operable to generate a denoised image of an input image; A method comprising:

2. 2. The method of claim 1 , wherein the converting the design patterns into the first set of simulated images comprises running a trained model configured to generate the simulated images using the design patterns as inputs.

3. The method of claim 2 , wherein the trained model is trained based on the design patterns and captured images of a patterned substrate, each captured image being associated with a design pattern.

4. The method of claim 3 , wherein the captured image is a scanning electron microscope (SEM) image obtained via an SEM.

5. 5. The method of claim 4, further comprising adding the image noise to the first set of simulated images to generate a second set of simulated images, the image noise being extracted from the captured image of the patterned substrate.

6. The method of claim 5 , wherein the training the denoising model includes using the first set of simulated images, the image noise, and the captured image as training data.

7. The method of claim 2 , wherein the trained model comprises a first machine learning model.

8. 8. The method of claim 7, wherein the trained model comprises a convolutional neural network or a deep convolutional neural network trained using a generative adversarial network training method.

9. The method of claim 8 , wherein the trained model is a generative model configured to generate simulated SEM images for a given design pattern.

10. The method of claim 1 , wherein the image noise is Gaussian noise, white noise, or salt and pepper noise characterized by user-specified parameters.

11. The method of claim 1 , wherein the denoising model comprises a second machine learning model.

12. The method of claim 1 , wherein the design pattern is in a graphics data signal (GDS) file format.

13. obtaining a captured SEM image of the patterned substrate; running a trained denoising model using the captured SEM image as an input image to generate a denoised SEM image; The method of claim 1 further comprising:

14. The method of claim 1 , further comprising updating the denoising model based on a captured image of the patterned substrate.

15. A method which, when executed by a processor, comprises: Transforming the design patterns into a first set of simulated images; One or more non-transitory computer-readable media storing instructions that cause the processor to implement the method of training the denoising model based on the first set of simulated images and image noise, the denoising model being operable to generate a denoised image of an input image.

Images