Systems and methods for inspecting photomasks using machine learning

Abstract

Embodiments of the present application relate to systems and methods of inspecting reticles using machine learning. Near-field reticle images are generated via a deep learning process based on reticle database images generated from a design database, and far-field reticle images at an image plane of an inspection system are simulated via a physics-based process based on the near-field reticle images. The deep learning process includes training a deep learning model based on minimizing differences between the far-field reticle images and a plurality of corresponding training reticle images acquired by imaging training reticles fabricated from the design database. Test regions of test reticles fabricated from the design database are inspected for defects via a die-to-database process that includes comparing a plurality of reference images from reference far-field reticle images to a plurality of test images acquired by the inspection system from the test reticles. The reference far-field reticle images are simulated based on reference near-field reticle images generated by the trained deep learning model.

Description

[0001] Information related to divisional application

[0002] This application is a divisional application of Chinese invention patent application filed on December 14, 2018, with application number "201880081260.2" and invention title "System and Method for Verifying Photomasks Using Machine Learning".

[0003] Cross-reference of related applications

[0004] This application claims priority to existing U.S. Provisional Application No. 62 / 611,321, filed December 28, 2017, entitled "Inspection of Reticles Using Machine Learning," by Hawren Fang et al., pursuant to 35 U.S.SC §119, the entire contents of which are incorporated herein by reference for all purposes. Technical Field

[0005] This invention generally relates to the field of semiconductor inspection (e.g., photomask inspection). More specifically, this invention relates to die-to-database inspection and the like. Background Technology

[0006] Generally, semiconductor manufacturing involves highly complex techniques for fabricating integrated circuits using semiconductor materials (such as silicon) that are layered and patterned onto a substrate. Integrated circuits are typically fabricated using multiple photomasks. First, circuit designers provide circuit pattern data or a design database (which describes a specific integrated circuit (IC) design) to a photomask production system or photomask writer. The circuit pattern data is typically in the form of a representation layout of the physical layers of the manufactured IC device. The representation layout contains representation layers for each physical layer of the IC device (e.g., gate oxide, polysilicon, metallized, etc.), where each representation layer consists of a patterned plurality of polygons that define the layers of the specific IC device. The photomask writer uses the circuit pattern data to write (e.g., typically using an electron beam writer or laser scanner to expose the photomask pattern) multiple photomasks that will later be used to fabricate the specific IC design.

[0007] Each photomask is generally an optical element containing at least transparent and opaque areas and sometimes semi-transparent and phase-shifted areas (which together define a pattern of coplanar features in an electronic device, such as an integrated circuit). During photolithography, photomasks are used to define designated areas of a semiconductor wafer for etching, ion implantation, or other manufacturing processes.

[0008] Photomask inspection systems can inspect for defects that can occur during photomask production or after photolithography. Due to large-scale circuit integration and the miniaturization of semiconductor devices, manufactured devices have become increasingly sensitive to defects. That is, defects that cause device malfunctions are becoming smaller and smaller. Therefore, there is a need to continuously improve inspection techniques for monitoring photomask characteristics. Summary of the Invention

[0009] The following is a simplified summary of the invention to provide a basic understanding of specific embodiments thereof. This summary is not a detailed overview of the invention and does not identify key / essential elements or limit the scope of the invention. Its sole purpose is to present some of the concepts disclosed herein in a simplified form as a prelude to the more detailed description that will follow.

[0010] In one embodiment, a method and apparatus for inspecting a photomask are disclosed. A near-field mask image is generated via a deep learning process based on a mask database image generated from a design database, and a far-field mask image at an image plane of the inspection system is simulated via a physics-based process based on the near-field mask image. The deep learning process includes training a deep learning model based on minimizing the difference between the far-field mask image and multiple corresponding training mask images acquired by imaging a training mask manufactured from the design database, wherein such training mask images are selected based on pattern diversity and are defect-free. Defects in a test area of ​​a test mask manufactured from the design database are inspected via a die-to-database process, which includes comparing multiple reference images from a reference far-field mask image with multiple test images acquired by the inspection system from the test mask. The reference far-field mask image is simulated based on the reference near-field mask image generated by the trained deep learning model. In a particular embodiment, the test mask and the training mask are the same mask, and the test image is acquired from a region of this same mask that is different from the region where the training image was acquired.

[0011] In another embodiment, the physics-based process is based on the Hopkins method for generating the far-field mask image on the image plane of the testing tool based on the near-field mask image, and the deep learning process includes mapping the mask database image to a near-field mask image generated by light interacting with a mask manufactured from the design database. In another embodiment, the deep learning model is a convolutional neural network (CNN) that does not incorporate mask imaging into the image plane.

[0012] In another aspect, the deep learning model does not include perturbations of the far-field mask image caused by field dependence variations in the testing tool and is independent of the testing tool. In another aspect, the deep learning model is trained to minimize the difference between the far-field mask image and the corresponding training mask image by adjusting specific parameters (including weights and / or deviations) of multiple layers. In one embodiment, the layers being adjusted include convolutional layers with non-linear activation. In another aspect, the deep learning model is trained without adjusting parameters in one or more low-pass filter layers used for downsampling operations.

[0013] In an alternative embodiment, the CNN includes one or more convolutional layers for compensating for deviations between the mask database image and the physical mask generated from the mask database image, one or more layers for generating multiple downsampled images, and one or more layers for implementing a sparse representation with near-field resolution. In another embodiment, the test image is aligned with the reference image, and a dynamic compensation process is applied relative to the test image to the reference image to compensate for variations in the testing tool (including focus fluctuations and / or field-dependent variations).

[0014] In an alternative embodiment, the present invention relates to an inspection system for inspecting photolithography masks. The system includes at least one memory and at least one processor configured to perform one or more of the above-described operations. In another aspect, the present invention relates to a computer-readable medium having instructions stored thereon for performing one or more of the above-described operations.

[0015] These and other aspects of the invention will be further described below with reference to the drawings. Attached Figure Description

[0016] Figure 1 This is a graphical representation of a process for obtaining a reference photomask image from a design database according to an embodiment of the present invention.

[0017] Figure 2 A graphical representation of a neuron in a convolutional layer for receiving an image from a photomask database, according to an exemplary embodiment of the present invention.

[0018] Figure 3 This invention describes a convolutional neural network (CNN) process for generating near-field images of a photomask from a photomask database image, according to an embodiment of the present invention.

[0019] Figure 4 This paper describes a detailed CNN process for generating near-field images of a photomask from a photomask database image, according to a specific embodiment of the present invention.

[0020] Figure 5This describes a defect detection process according to an embodiment of the present invention.

[0021] Figure 6 This is a graphical representation of an example verification system for implementing the technology of this invention.

[0022] Figure 7 A schematic diagram of a photomask inspection apparatus according to a specific embodiment is provided. Detailed Implementation

[0023] In the following description, numerous specific details are set forth to provide a thorough understanding of the invention. The invention may be practiced without some or all of these specific details. In other instances, well-known processes have not been described in detail to avoid unnecessarily obscuring the invention. Although the invention will be described in conjunction with specific embodiments, it should be understood that it is not intended to limit the invention to these embodiments.

[0024] The inspection techniques described herein can be applied relative to any suitable type of photomask or photomask. In one example, extreme ultraviolet (EUV) lithography uses an EUV-type photomask designed to facilitate patterning on a wafer at EUV wavelengths (e.g., 13.5 nm). EUV photomasks typically comprise a substrate, such as a low thermal expansion (LTE) or ultra-low expansion (ULE) glass plate (e.g., fused silica). The substrate is covered by multiple layers of material to provide moderate reflectivity (e.g., 60% to 70% or greater) at EUV wavelengths for performing lithographic exposure at EUV wavelengths. The multilayer (ML) stack acts as a Bragg reflector that maximizes reflection of EUV radiation while simultaneously being a poor absorber of EUV radiation. Reflection generally occurs at interfaces between materials with different refractive indices that have a high difference, resulting in greater reflectivity. Although the refractive index of the material exposed to extremely low wavelengths is approximately 1, significant reflection can be achieved by using multiple layers with alternating layers having different refractive indices. ML stacks include low absorption properties, such that shock radiation is reflected with almost no loss. In a particular embodiment, the multiple layers comprise approximately 30 to approximately 40 (or 40 to 50) alternating pairs of molybdenum (Mo) layers and silicon (Si) layers arranged at a pitch of approximately 7 nanometers. Other suitable layers may comprise alternating layers of Mo₂C and Si, Mo and beryllium (Be), or molybdenum-ruthenium (MoRu) and Be.

[0025] Multiple layers may include a capping layer (e.g., Ru) to prevent oxidation. In other embodiments, the EUV mask may include quartz, an anti-reflective coating (ARC), and other features. Patterning is formed in an absorber layer disposed on multiple layers. For example, a tantalum boron nitride (TaBN) film topped with a thin anti-reflective oxide (e.g., tantalum boron oxide (TaBO)) acts as an EUV absorber. The materials(s) used for the mask pattern may be selected to have near-zero etch deviation to achieve ultra-fine resolution features.

[0026] Besides EUV photomasks, the terms "photomask" and "photomask" can also include a transparent substrate on which an opaque material, such as glass, borosilicate glass, quartz, or fused silica, is formed. The opaque (or substantially opaque) material can include any suitable material that completely or partially blocks photolithography light (e.g., deep UV). Examples of such materials include chromium, molybdenum silicide (MoSi), tantalum silicide, tungsten silicide, opaque MoSi on glass (OMOG), and so on. A polycrystalline silicon film can also be added between the opaque layer and the transparent substrate to improve adhesion. Low-reflectance films (e.g., molybdenum oxide (MoO2), tungsten oxide (WO2), titanium oxide (TiO2), or chromium oxide (CrO2)) can be formed on the opaque material.

[0027] The term "mask" can refer to different types of masks, including (but not limited to) brightfield masks, darkfield masks, binary masks, phase-shifting masks (PSMs), alternating PSMs, attenuating or halftone PSMs, ternary attenuating PSMs, and chromium-free phase lithography PSMs. Brightfield masks have transparent field or background regions, while darkfield masks have opaque field or background regions. Binary masks are masks with patterned regions that are either transparent or opaque. For example, a photomask made from a transparent fused silica blank with a pattern defined by a chromium metal absorbing film can be used. Binary masks differ from phase-shifting masks (PSMs) (one type of which may contain a film that only partially transmits light), and these masks can be collectively referred to as halftone or embedded phase-shifting masks (EPSMs). If a phase-shifting material is placed in the alternating bright space of the mask, then the mask is called an alternating PSM, ALT PSM, or Levenson PSM. Phase-shifting materials applied to arbitrary layout patterns are called attenuating or halftone PSMs, which can be manufactured by replacing the opaque material with a partially transmissive or "halftone" film. Ternary attenuating PSMs are attenuating PSMs that also include fully opaque features.

[0028] Returning to the EUV lithography process, the light source can generate any suitable radiation suitable for use with an EUV photomask. For example, EUV wavelengths between about 11 and 14 nm or lower soft X-ray wavelengths can be utilized. In a particular embodiment, a wavelength of about 13.5 nm is generated. During lithography, radiation reflected from multiple layers of the EUV photomask is absorbed into a photoresist formed on the wafer substrate. The absorbed radiation generates photoacids (H+) and amplified photoacids, which, when the resist is developed, form an exposure pattern in the photoresist of the wafer substrate corresponding to the absorber pattern layer of the EUV photomask.

[0029] From a process yield management perspective, controlling the defect rate of EUV photomasks that define patterns printed on silicon wafers plays a crucial role. However, due to the lack of photoluminescent EUV photomask inspectors that optically inspect photomasks at wavelengths identical to those used in EUV scanners (e.g., 13.5 nm), defect detection has been considered a high-risk area in EUV lithography development. Electron beam inspection tools (which may offer good sensitivity) typically have inspection throughput several orders of magnitude lower than the required throughput and are therefore not a feasible solution for full mask inspection. Currently and for the foreseeable future, the inspection of patterned EUV masks must rely on more readily available, higher-throughput inspection tools operating in the deep UV (DUV) wavelength range (190 to 260 nm).

[0030] One type of inspection technique utilizes a die-to-database approach, which typically involves calculating a reference image based on a database. The database contains a series of polygonal shapes to be written onto a photomask. Calculating the reference image typically involves: (i) modeling the polygons through which they are written onto the photomask using electron beam lithography; (ii) characterizing the pupil illumination pattern of the photomask inspection microscope; (iii) calculating how the illumination interacts with the patterned photomask to form a diffractive near-field; and (iv) modeling how the diffractive near-field is imaged and recorded at the image plane by an array sensor.

[0031] When using physics-based methods (such as the Hopkins method), calculating how illumination interacts with a patterned mask to form a diffraction near field is computationally difficult. Generally, the Hopkins method is based on an exchange of point source contributions and integral order over diffraction amplitudes, allowing a given optical system with fixed illumination, numerical aperture, defocus, and other aberrations to be described with a transmission cross-correlation coefficient (TCC). The TCC can be calculated only once and subsequently reused for repeated image simulations of different mask patterns imaged by the same optical system. One problem is that applying the Hopkins method to extreme ultraviolet (EUV) masks is computationally extremely expensive. For example, known methods cannot rigorously solve the entire mask within practical inspection time. Another difficulty in calculating how illumination interacts with a patterned mask is the uncertainty in the size of the mask pattern, the contours of the sidewalls of the etched pattern, and the optical properties of the mask material.

[0032] Specific embodiments of the present invention use machine learning to map a pattern on a mask to a diffraction field. In a particular embodiment, a convolutional neural network (CNN) is used to learn this mapping from the mask pattern to the diffraction field (referred to as the "mask near field"). Although the following example embodiments are described primarily in the context of CNNs, other machine learning or neural network processes can be used to learn mask near-field images.

[0033] Beyond the interaction between illumination and the photomask, physically based models can be used to accurately characterize and rapidly simulate image formation. Therefore, the method for generating the reference image for the photomask maintains a physically based model process. That is, the learning process (e.g., via a CNN) does not incorporate photomask image formation onto the image plane (e.g., an inspection tool detector).

[0034] In one embodiment, the convolutional neural network is trained using an actual photomask and its actual image acquired by the inspection tool. In this mode of operation, training is performed by minimizing the difference between the final simulated image on the detector and the actual image acquired by the tool.

[0035] The properties of an optical system can vary over time and at different locations within the imaging field. Advantageously, once the field diffracted by the photomask is known from the CNN process, a physics-based model can be used to calculate the perturbation of the image relative to the optical system. Therefore, the CNN does not bear this task, making the training process much simpler than when the detector image is learned by the CNN. Compared to physics-based methods, the neural network process provides significantly more accurate computation of the near-field photomask image. Furthermore, the neural network process is highly efficient because the input and output images of the neural network are not significantly different from each other, and the learning task is simplified. This efficiency is achieved by not having the neural network process learn far-field images. In summary, combining an accurate neural network method for computation of the near-field photomask image with a physics-based method for computation of the photomask image from the near-field photomask image results in a more efficient and sensitive inspection process.

[0036] Figure 1 This is a schematic representation of a process 100 for obtaining a reference photomask image from a design database according to an embodiment of the present invention. First, a database image 102 is provided to a deep learning model 104. The database image can be provided based on a design database used to manufacture the photomask to be tested. For example, any suitable technique can be used to reproduce the binary photomask image from the polygonal description of the design database. That is, dark and bright intensity values ​​can be assigned to the absorber and multilayer regions respectively, or vice versa.

[0037] A deep learning model 104 is configured to generate near-field photomask images via a learning model that inputs a photomask image 102 from a database. During training, the near-field photomask image is output by the deep learning model 104 and input to a physics-based simulation process 106 based on physics-based modeling parameters 108. For example, the physics-based method involves the Hopkins method for simulating light from the near-field photomask image through the collection path of the inspection tool to the far-field image on the detector of this tool. The physics-based modeling parameters may include a pre-calculated transmission cross coefficient (TCC) of the light behavior from the near field through the inspection system to the detector. The physics-based simulation process 106 outputs a simulated optical photomask image, which is input to a training optimizer 110 that optimizes the difference between the optical photomask image and a defect-free training optical image 112 obtained by imaging a photomask manufactured from a design database.

[0038] In one embodiment, a small area of ​​the photomask (e.g., less than 0.01%) is examined and verified or assumed to be defect-free. An image is obtained from this defect-free portion of the photomask and used as a training image. After training the model, the output from the trained model can be used to examine the remaining portion of the photomask (e.g., the other 99.99%).

[0039] The deep learning model works in conjunction with a training optimizer 110 to train a thick mask diffraction model to output a near-field mask image from a database of mask images based on a defect-free training optical image 112. That is, the optimizer trains or tunes the deep learning model parameters while keeping the physically based modeling parameters 108 constant.

[0040] Any suitable deep learning model can be used to determine the near field of a photomask. In a particular implementation, a convolutional neural network (CNN) can be used. CNNs are a class of deep feedforward neural networks most commonly used for analyzing visual imagery, making them well-suited for analyzing photomask database images. Generally, a CNN contains multiple interconnected layers of "neurons," and each neuron is designed to mimic the visual cortex because each neuron only receives and transforms a small input domain (e.g., a small portion of the photomask database image). The input region of a neuron is called its receptive field. In a convolutional layer, the receptive field is smaller than the entire previous layer. Figure 2 A schematic representation of a convolutional layer 204 having neurons 204a for receiving a receptive field 202a of a photomask database image 202, according to an exemplary embodiment of the present invention. Layer 204 will also contain other neurons (not shown) for other receptive fields of the photomask database image 202.

[0041] In a neural network, each neuron applies a function to input values ​​from its receptive domain in a previous layer and computes a set of output values. The function applied to the input values ​​is specified by a weight vector and a bias (usually real numbers). Learning in a neural network progresses by incrementally adjusting the bias and weights. In convolutional layers, the weight vector and bias can be called filters, and a key feature of convolutional layers is that many neurons can share the same filters. This sharing of filters reduces memory footprint because a single filter can be used across all receptive domains, rather than each receptive domain having its own bias and weight vector in a fully connected layer.

[0042] In a neural network, each neuron receives input from a number of locations in previous layers. In a fully connected layer, each neuron receives input from every element in the previous layer. In a convolutional layer, neurons receive input only from a restricted subregion of the previous layer. Even a fully connected layer for a small image (e.g., 100×100 pixels) would require 10,000 weights for each neuron in the receiving layer. Conversely, using convolutional layers reduces the number of free parameters, allowing the network to go deeper with fewer parameters. For example, a 5×5 stitched region can each use the same shared weights (e.g., only 25 learnable parameters).

[0043] CNNs typically consist of stacked layers of different types, some locally connected and some fully connected. These stacked layers can be used to generate accurate near-field images of a mask from a database of images. For example, several techniques for learning accurate images using CNNs are further described in the paper "Learning a Deep Convolutional Network for Image Super-Resolution" by Chao Dong et al. (presented at the 2014 European Conference on Computer Vision (ECCV), which is incorporated herein by reference). Generally, images from the mask database are fed into CNN layers to generate near-field images, which are then used as input to a physics-based model that outputs a far-field image of the mask (e.g., at a test detector).

[0044] Figure 3 A simplified CNN process 300 for generating a near-field image 312 of a photomask from a photomask database image 302, according to an embodiment of the present invention, is described. In this example, the CNN will typically include additional layers and additional filters and mappings at each layer, which are not shown in the illustrative figures for simplicity.

[0045] First, the mask database image 302 can be received by the tile extraction layer 314a to extract and represent tiles from the mask database image 302. Tiles can overlap. For example, the tile extraction layer 314a may contain filters 304, which are in the form of weighted kernels applied to each pixel. Each filter has a size f1×f1. In the illustrated example, filter 304 is a 9×9 kernel convolved relative to each pixel of the mask database image 302. That is, the 9×9 kernel steps across the entire mask database image to place each pixel one at a time in the center of the kernel. In the current example, filter 304 is shown as applied to pixel 302b of the mask database image 302 and its neighboring pixels 302a. The output of filter 304 on this pixel 302b and its neighboring pixels will be the value 306a (of array 306). There will be additional filters (not shown) applied to each pixel to result in an array of n1 values ​​for each pixel (e.g., array 306). Generally, the weights of each core filter are multiplied by the corresponding overlapping mask database image to result in a value for a specific pixel. The resulting image, having the same size as the mask database image, is generated from cores that have been stepped and applied across the entire mask database image. The output arrays of all filters in the patch extraction layer applied to all pixels of the mask database image are not shown to simplify the invention.

[0046] Next, as shown, the output from the tile extraction layer can be fed into the next layer. As shown, a nonlinear mapping layer 314b can be applied to the output of the tile extraction layer 314a. Next, as shown, the convolution result n1 (306) of the mask pixels 302b can be fed into a nonlinear mapping layer that outputs the nonlinear mapping result n2 (308). Any suitable number and type of nonlinear mapping layers can be used by the CNN. For example, sigmoid(x), tanh(x), ReLU (rectified linear unit), leaky ReLU, etc. The ReLU function applies a non-saturating activation function f(x) = {max(0,x)}, which is used to increase the nonlinearity of the decision function and the entire network without affecting the receptive field of the convolutional layer.

[0047] After all pixels in the mask database have undergone tile extraction and nonlinear mapping, the reconstruction process 314c reconstructs the mask near-field image result 312 by combining the filter and nonlinear mapping results. More precisely, the reconstruction process aggregates f3×f3 neighboring tiles 310 for each pixel. The resulting near-field image represents a real-world near-field image of the terrain. However, as will be further described herein, additional processing of the near-field image using physics-based methods can be performed to simulate the final far-field image of the mask, which is expected to resemble an easily obtainable real-world far-field mask image.

[0048] As mentioned in this article, CNNs may contain any number and type of convolutional layers and / or other layers. Figure 4 This section describes a detailed CNN process 400 for generating a near-field image of a mask from a mask database image, according to a specific embodiment of the present invention. First, the mask database image may include multiple images 402. For example, the mask database image 402 may include an image rasterized from a design database, an absolute X gradient image represented by |gradx(DB)|, and an absolute Y gradient image represented by |grady(DB)|. In this example, "DB" is the mask database image. The gradient image |gradx(DB)| in the X direction may be implemented by a finite difference approximation, which is performed by a convolutional filter [-1 0 1] applied to rows of the scaled DB. Next, an absolute value function may be applied element-wise to form |gradx(DB)|. Similarly, the absolute gradient image |grady(DB)| in the Y direction can be obtained.

[0049] A CNN may include one or more convolutional layers to compensate for the discrepancies between the design database images and the physical mask. For example, a 1×1 kernel may be convolved across all the mask database images 402 to result in a feature image 404, which may then be fed into another convolutional layer to produce a set of feature images 406 that increase the number of images by applying multiple filters to each feature image 404. That is, more than one filter may be applied to each pixel of the compensated image 404 to produce more feature images 406 than the input feature image 404.

[0050] In a typical cancellation example, summing the mask database image and the scaled |gradx(DB)| image has the effect of altering the CD (critical dimension) of the vertical line-space pattern. |gradx(DB)| and |grady(DB)| can also be used to relax the isometric mapping (e.g., the CD variations in the x and y directions can differ). This is because the electron beam mask writer is imperfect and the resulting mask can differ slightly from the design. For example, if the design has a 100nm CD, then due to the imperfections of the mask writer, the physical mask may actually have a 95nm CD. This deviation is often referred to as "mask deviation". Therefore, the first convolutional layer can be used to cancel this mask deviation. In effect, this layer is related to (but not equivalent to) learning the mask deviation. In this example, this layer is configured to provide a feature image(s) that aids in learning and thus improves the accuracy of the predicted near-field mask image.

[0051] Next, feature image 406 can be input into a low-pass filtering process to further downsample feature image 406 into a set of smaller feature images 408. For example, feature image 406 may have a 4X size, and the resulting feature images 408 may have a 2X size. Generally, the input image lacks grayscale and is almost binary (most pixels are either fully bright or fully dark), while the expected near-field and far-field images will exhibit fine grayscale (as expected in the ground reality image). Therefore, when the input can be downsampled throughout the entire CNN to achieve the final output, a pixel size of 55 / 4nm is desired. 2 The higher resolution feature image (4X) has a pixel size of 55nm. 2 The 1X. This first downsampling layer can be configured to reduce the pixel size to 55 / 4nm by applying a low-pass filter from signal processing and then downsampling. 2 Change to 55 / 2nm 2 (4X to 2X). That is, unlike average pooling (which is more popular in computer vision).

[0052] The subsequent layer 422 of CNN 400 can implement sparse representations at near-field resolution. This neural network technique has been applied to the paper by Chao Dong cited above and... Figure 3 Image super-resolution in CNNs. For example, a set of 9×9 filters is convolved in all feature images 408 to result in feature image 410. Since the number of 9×9 filters in this stage is higher than that of the input feature image 408, the number of output feature image 410 (or output channels) is also higher than that of the input feature image 408. The number of channels output from each layer can be selected to provide a significant improvement in accuracy that is balanced with saving processing resources and generalization (i.e., more channels tend to result in higher accuracy). For example, if a higher number of channels results in minimal improvement and increased processing, then it may not be selected; but if a higher number of channels results in a significant improvement in accuracy (even if it increases the processing burden), then it is selected. The selection of the number of filters can be based on empirical results. Output feature image 410 can be smaller by using only the "effective" region and excluding halo regions. Compared to a 4X size mask database image, the downsampling layers of the CNN can facilitate a reduction in pixel size to achieve a final 1X size.

[0053] Next, feature image 410 is received by a nonlinear layer that performs a dimensionality reduction process to output a set of feature images 412 with fewer channels than the input feature image 410. In this example, a 1×1 filter is convolved over all feature images 410 to output a reduced-channel / feature image 412 with the same size as the input image.

[0054] Next, the reconstruction layer described above can be applied to feature image 412 to output a set of near-field residual images 414. The output of the (non-linear) dimensionality reduction layer can be referred to as the "base image". To reconstruct each of the real and imaginary parts of the near-field residual image, a 5×5 trainable filter can be applied to each base image; all filtered base images are summed; then a non-linear activation function is applied to output the near-field residual (real or imaginary part). The filter size of 5×5 is chosen based on empirical results to achieve a predefined accuracy level, but other sizes can be chosen for other applications.

[0055] In practice, a single CNN network can be used to predict two near-field images to improve computational efficiency. In this example, the near-field image is a composite image that can be represented by two real-valued feature images. The number of images can correspond to any number and type of images generated for different testing tool settings. In the current example, both X and Y polarization simulation images can be generated by a CNN for use with X and Y polarization optical images acquired by a Teron 640e tool purchased from KLA-Tencor in Milpitas, CA. Thus, a total of four final mask real-valued images exist to form two near-field images. Then, as shown, the near-field residual image 414 can be added to a Kirchoff field (a linear approximation of the near field) to produce a set of four predicted real-valued images 416 to form two near-field images for the two polarizations (X and Y polarizations).

[0056] The Kirchhoff field is generally a linear approximation of the near field (fast but inaccurate). To facilitate training, the CNN can first predict the near-field residual image (the difference between the desired near field and the Kirchhoff field). Then, the predicted near-field image can be obtained by summing the residual with the Kirchhoff field. In the illustrated example, the near-field residual image 414 is added to the Kirchhoff field to form a 2X near-field residual image 416 processed by a low-pass filter, where the result is downsampled to produce the predicted near-field image 420 at 1X.

[0057] Regardless of the specific CNN configuration, the predicted near-field image can be used to generate the far-field photomask image, which is then used to train the CNN, as relative to... Figure 1 As described. For example, the predicted near-field image (NF) can then be input into a physics-based simulation process 106 based on physics-based modeling parameters 108. In one implementation, a partial coherence model (PCM) is configured to generate a far-field image via the following equation:

[0058]

[0059] Where x and y are pixel indices, NF is the complex near field of the mask, and These are the physical parameters of the testing tool.

[0060] Optimizer 10 receives the generated model image I model and corresponding training image I training (112) (which is pre-screened to be defect-free (via independent measurement)). Any suitable optimization process can be used, such as stochastic gradient descent (SGD), RMSProp, momentum, Adam, K-FAC, etc. The optimizer 110 can generally be configured to minimize such training images I by adjusting certain trainable parameters relative to the deep learning model 104. training (112) and simulated far-field image I model The difference between them. In a particular implementation, the loss function can be minimized, for example, the sum of squares of the difference images as defined in the following equation [2]:

[0061]

[0062] In the above equation [2], “|||| F "This is the Frobenius norm. That is, the trainable parameters of the deep learning model are adjusted until the loss function of equation [2] is minimized, for example, after which the deep learning model is considered to be trained for a specific mask. In the deep learning model, the parameters of the convolutional layers are adjustable and trainable parameters, except for the low-pass filter layers with fixed parameters that are not trained. In a particular embodiment, the deep learning model can be trained in real time after the mask is fabricated and is considered to be defect-free before being used in the lithography process. The training data can be selected and screened by experienced engineers to achieve full coverage of the defect-free pattern.

[0063] Once a deep learning model is trained for a specific photomask, it can be used as part of the entire defect detection process, including image alignment to compensate for translational shifts and dynamic compensation (DC) mechanisms to compensate for optical fluctuations and field dependencies. Figure 5 The defect detection process 500 according to an embodiment of the present invention is described below. For completeness, the design database polygons may first be received by a DB raster image reconstruction process 502 that generates a 4X database (DB) image of the mask, and the 4X DB image is then received by a deep learning process 504 (e.g., a trained CNN described above) that includes a deep learning model 504 containing the output predicted near-field image.

[0064] The optical model image reconstruction process 506 receives a near-field photomask image and configures it to generate a far-field image. The image alignment process 508 configures the test image acquired from the inspection tool to align with the reconstructed far-field image. Specifically, the test image is moved relative to the far-field image until the difference between the two sets of images is minimal.

[0065] Test images are obtained from the inspection area of ​​the photomask. Furthermore, as the incident beam scans each patch of the photomask, the EUV inspection tool can be operated to detect and collect multiple polarized reflected light images. In non-EUV photomask inspectors, transmitted light (and / or reflected light) can be used, and the techniques described herein can be configured for such inspectors. The incident beam can scan a photomask scan strip, each comprising multiple patches. Light is collected from multiple points or sub-regions of each patch in response to this incident beam.

[0066] The testing tool is generally operable to convert this detection light into a detection signal corresponding to intensity values. The detection signal can be in the form of an electromagnetic waveform with amplitude values ​​corresponding to different intensity values ​​at different locations on the photomask. The detection signal can also be in the form of a simple list of intensity values ​​and associated photomask point coordinates. The detection signal can also be in the form of an image with different intensity values ​​corresponding to different locations or scan points on the photomask. A photomask image can be generated after scanning all locations on the photomask and detecting the light, or a portion of the photomask image can be generated while scanning each section of the photomask. Generally, test images from specific patches of the photomask can be generated by the testing tool.

[0067] In a particular embodiment, the dynamic compensation (DC) process 512 may also be performed on the far-field image and alignment test image generated by the optical model image reconstruction process 506. The dynamic compensation process can be configured to counteract specific tool variations (e.g., focus fluctuations and / or field-dependent variations) to produce a dynamically compensated (DC) mask image. The far-field images before and after dynamic compensation (DC) are referred to as the pre-DC and post-DC images, respectively.

[0068] In this example, the image before the DC can be defined by the following equation:

[0069]

[0070] Where NF is the complex near field of the mask, and the feature pairs These are the physical parameters of the testing tool.

[0071] Next, dynamic compensation may include minimizing another loss function. For example, linear least squares is:

[0072]

[0073] This consists of minimal elements λ1, λ2, ..., λ n Image I formed instead of the original feature values model This is called the DC-post image.

[0074] Next, defect detection (510) can be performed by comparing the DC-generated mask image with the test image, and an inspection report can then be obtained. The differences can then be re-examined by, for example, a defect classification process or a high-resolution tool. For example, SEM can be used to re-examine defect areas to determine if the critical dimension (CD) is out of specification. Re-examination may involve separating harmful defects from “real” defects that could affect the functionality of devices on the final wafer produced using this mask. For example, it can be determined that a particular harmful mask defect is unlikely to cause printing defects on the wafer, while other harmful mask defects could cause printing defects that affect device functionality.

[0075] Out-of-specification CDs (or other defects) can cause the photomask to fail inspection. If the photomask fails inspection, it can be discarded or repaired (if possible). For example, a specific defect can be removed from the photomask. After repair or removal, a new inspection can be performed on the repaired or removed photomask, and the process can be repeated. Alternatively, the photomask manufacturing process or photomask design can be adjusted, and a new photomask can be manufactured.

[0076] Generally, the absorber and multilayer materials of an EUV photomask are patterned and formed with a critical dimension (CD) width. A specific CD value typically affects how a particular photomask feature is transferred to the wafer during the photolithography process, and this CD is selected to optimize this transfer process. In other words, if the CD value of a particular photomask feature is within a specified CD range, then this CD value will result in the fabrication of the corresponding wafer feature that allows proper operation of the resulting integrated circuit, as intended by the circuit designer. Features are often formed with a minimum size that also results in operational circuitry, saving integrated chip area.

[0077] Newly manufactured photomasks may contain CD (or other film or pattern characteristics) defects. For example, a photomask may have defective CD areas. Photomasks can deteriorate over time in many different ways. Some types of CD degradation can be caused by chemical reactions between the photomask features (MoSi) and the exposure light, cleaning processes, contamination, and so on. These physical effects can also adversely affect the critical dimensions (CD) of the photomask over time.

[0078] As a result of this degradation, the feature CD value will change significantly, affecting the CD uniformity of the entire mask and adversely impacting wafer yield. For example, the mask feature width in a portion of the mask may be significantly larger than the original linewidth CD. For example, radial patterns of CD non-uniformity may exist, where the center of the mask has a CD different from that at the edges of the mask.

[0079] Critical Dimension Uniformity (CDU) maps of the mask can be generated to facilitate CD monitoring of the mask. These CDU maps are important for semiconductor chip manufacturers to understand the process windows that will be derived from the use of the mask. CDU maps allow chip manufacturers to determine the use of the mask, compensate for errors in the lithography process, or improve the mask manufacturing to form an improved next mask.

[0080] Various techniques can be used to generate CDU maps. In a die-to-database inspection method, the average intensity values ​​between corresponding regions of the test and reference images are compared to obtain a differential intensity value. The differential intensity values ​​of the entire mask can then be effectively used to form a differential intensity map, which can then be calibrated to a full CDU map. Although the inspection technique is described as being based on intensity-type signals, other types of signals can be used in alternative embodiments of the invention.

[0081] Specific embodiments of the present invention provide apparatus and techniques for significantly improving the EUV mask defect sensitivity of DUV inspection tools by providing reference mask images that are efficiently and accurately reproduced from a design database. By comparison, a combination of deep learning techniques for generating accurate near-field images and a physics-based modeling process for generating far-field mask images can be completed in less than 90 minutes, whereas generating near-field images using rigorous physics-based simulation methods would take years. It has been found that the results obtained using the combination of deep learning and physics-based modeling have significantly improved accuracy based on the difference between the final modeled mask image and the defect-free image compared to other techniques. To reduce the effect of shot noise, a 2×2 convolutional filter with all 1s can be applied, and the maximum difference in the measurable values ​​can be used to estimate the model error. In an example of a 2K×1K tile image with a 2D pattern, the maximum convolutional difference between the DC-prepared image and the test image using the deep learning method is 51 gray levels, while the maximum convolutional difference using conventional methods is 109 gray levels. For the same patch image, the maximum 88-level grayscale convolution difference between the DC-generated image using deep learning and the test image was found to be relative to the 40-level convolution difference using conventional methods. Ideally, the deep learning process used to generate the near-field mask image should be independent of the testing tool.

[0082] The technology of this invention can be implemented in any suitable combination of hardware and / or software. Figure 6This is a schematic representation of an example inspection system 600 that can implement the technology of the present invention. The inspection system 600 may receive input 602 from an inspection tool or scanner (not shown). The inspection system may also include a data distribution system (e.g., 604a and 604b) for distributing the received input 602, a processing system (e.g., tile processors and memories 606a and 606b) for processing intensity signals (or tiles) of specific portions / tiles of the received input 602, a deep learning system (e.g., a deep learning GPU and memory 612a) for learning near-field mask images, a far-field modeling and learning support system (e.g., a far-field modeling and learning support processor and memory 612b), a network (e.g., a switching network 608) for allowing communication between components of the inspection system, an optional capacity storage device 616, and one or more inspection control and / or re-inspection stations (e.g., 610) for re-inspecting images. Each processor of the inspection system 600 may typically include one or more microprocessor integrated circuits and may also include interface and / or memory integrated circuits and may be additionally coupled to one or more shared and / or global memory devices. In certain implementations, deep learning models (such as CNNs) are implemented on graphics processing units (GPUs) to improve the speed and processing power of the training process, while other processes (physical modeling, training optimization, dynamic compensation, etc.) used to generate post-DC or pre-DC photomask images are implemented by one or more CPUs and memory.

[0083] The scanner or data acquisition system (not shown) used to generate input data 602 can take the form of any suitable instrument (such as those described further herein) for obtaining an intensity signal or image of the photomask. For example, the scanner can construct an optical image or generate an intensity value of a portion of the photomask based on a portion of detected light reflected, transmitted, or otherwise directed to one or more light sensors. The scanner can then output the intensity value or image.

[0084] A photomask is typically divided into multiple patch sections from which multiple intensity values ​​are obtained from multiple points. These patch sections of the photomask can be scanned to obtain this intensity data. Depending on the specific system and application requirements, the patch sections can be of any size and shape. Generally, multiple intensity values ​​for each patch section can be obtained by scanning the photomask in any suitable manner. For example, multiple intensity values ​​for each patch section can be obtained by raster scanning the photomask. Alternatively, an image can be obtained by scanning the photomask using any suitable pattern (e.g., a circular or spiral pattern). Of course, the sensor must be arranged differently (e.g., in a pie pattern) and / or the photomask can be moved differently (e.g., rotated) during scanning to scan the circular or spiral shape from the photomask.

[0085] In the example described below, as the photomask moves past the sensor, light is detected by a rectangular area of ​​the photomask (referred to herein as a “scan band”), and this detected light is converted into multiple intensity values ​​at multiple points in each patch. In this embodiment, the scanner’s sensors are arranged in a rectangular pattern to receive light reflected and / or transmitted from the photomask and generate from it a set of intensity data for the scan band corresponding to patches of the photomask. In a particular example, each scan band may be about 1 to 2 million pixels wide and about 1,000 to 2,000 pixels high, while each patch may be about 2,000 pixels wide and about 1,000 pixels high.

[0086] Intensity values ​​for each patch can be obtained using optical inspection tools set up in any suitable manner. Optical tools typically use a set of operating parameters or a "recipe" that is substantially the same for different inspection operations to obtain intensity values. The recipe settings may include one or more of the following: settings for scanning the mask in a specific pattern and pixel size; settings for grouping adjacent signals from a single signal; focus settings; polarization settings; illumination or detection aperture settings; incident beam angle and wavelength settings; detector settings; settings for reflected or transmitted light amount; aerial modeling parameters, etc.

[0087] Intensity or image data 602 can be received by a data distribution system via network 608. The data distribution system may be associated with one or more memory devices (e.g., RAM buffers) for storing at least a portion of the received data 602. Preferably, the total memory is sufficient to store the entire scan band of data. For example, 1 gigabyte of memory is ideal for a scan band set to 1 million × 1000 pixels or dots per polarization and focus.

[0088] The data distribution system (e.g., 604a and 604b) can also control the distribution of portions of the received input data 602 to the tile processors (e.g., 606a and 606b). For example, the data distribution system can route data from a first tile to a first tile processor 606a and can route data from a second tile to tile processor 606b. Multiple sets of data from multiple tiles can also be routed to each tile processor. The distribution system can also control the distribution of portions of the pre- or post-DC modeling mask image (“reference mask image”) to the tile processors (e.g., 606a and 606b).

[0089] The tile processor can receive intensity values ​​or images corresponding to at least a portion of a photomask or tiles. The tile processor may also be individually coupled to or integrated with one or more memory devices (not shown) (e.g., DRAM devices providing local memory functionality, such as storing portions of received data) or said one or more memory devices. Preferably, the memory is sufficient to store data corresponding to tiles of the photomask. For example, 8 megabytes of memory is well-suited for intensity values ​​or images corresponding to tiles that are 512 × 1024 pixels. The tile processor may also share memory.

[0090] Each set of input data 602 may correspond to a scan band of a photomask. One or more sets of data may be stored in the memory of a data distribution system. This memory may be controlled by one or more processors within the data distribution system, and the memory may be divided into multiple partitions. For example, the data distribution system may receive data corresponding to a portion of a scan band into a first memory partition (not shown), and the data distribution system may receive another set of data corresponding to another scan band into a second memory partition (not shown). Preferably, each memory partition of the data distribution system stores only a portion of the data that will be routed to the processor associated with this memory partition. For example, the first memory partition of the data distribution system may store first data and route the first data to tile processor 606a, and the second memory partition may store second data and route the second data to tile processor 606b.

[0091] Incident or detection light can pass through any suitable spatial aperture to produce any incident or detection light profile at any suitable incident angle. For example, programmable illumination or detection apertures can be used to generate specific beam profiles, such as dipole, quadrupole, quasar, ring, etc. In certain instances, source mask optimization (SMO) or any pixelated illumination technique can be implemented.

[0092] A data distribution system can define and distribute each set of data based on any suitable parameters of the data. For example, data can be defined and distributed based on the corresponding positions of tiles on a photomask. In one embodiment, each scan band is associated with a range of column positions corresponding to the horizontal positions of pixels within the scan band. For example, columns 0 to 256 of the scan band may correspond to a first tile, and the pixels in these columns will include a first set of image or intensity values ​​routed to one or more tile processors. Similarly, columns 257 to 512 of the scan band may correspond to a second tile, and the pixels in these columns will include a second set of image or intensity values ​​routed to different tile processors.

[0093] Figure 7A schematic diagram of an example inspection system 750 according to a specific embodiment is provided. The inspection system 750 has an illumination optics 751a including an imaging lens (which has a relatively large numerical aperture 751b at a photomask plane 752). The depicted inspection system 750 includes inspection optics 753a and 753b, which include microscope magnifying optics designed to provide, for example, 60 to 200X or greater magnification to enhance inspection. For example, the numerical aperture 751b at the photomask plane 752 of the inspection system can be significantly larger than the numerical aperture 701 at the photomask plane of the lithography system, which will result in a difference between the test inspection image and the actual printed image. Although the illustrated inspector includes both a reflected light assembly and a transmitted light assembly (for inspecting non-EUV photomasks), EUV photomask inspection will utilize only reflected light.

[0094] The testing techniques described herein can be implemented in various specially configured testing systems (e.g., Figure 7 The illustrated inspection system 750 includes an illumination source 760 that generates a light beam that is guided through an illumination optics 751a to a photomask M in a photomask plane 752. Examples of light sources include lasers or filter lamps. In one example, the source is a 193nm laser. As explained above, the inspection system 750 may have a numerical aperture 751b at the photomask plane 752 that is larger than the numerical aperture of the photomask plane in the corresponding lithography system. The photomask M to be inspected is placed on a mask stage at the photomask plane 752 and exposed to the source.

[0095] A patterned image is guided from a mask M through an array of optical elements 753a, which projects the patterned image onto a sensor 754a. In a reflective system, optical elements (e.g., a beam splitter 776 and a detection lens 778) guide and capture the reflected light onto a sensor 754b. Suitable sensors include charge-coupled devices (CCDs), CCD arrays, time-delay integration (TDI) sensors, TDI sensor arrays, photomultiplier tubes (PMTs), and other sensors.

[0096] The illumination optical column can be moved relative to the mask stage and / or the stage can be moved relative to the detector or camera via any suitable mechanism to scan the mask's pattern. For example, a motor mechanism can be used to move the stage. For instance, the motor mechanism can be formed by a screw drive and stepper motor, a linear drive or belt actuator with feedback position, and a stepper motor.

[0097] The signals captured by each sensor (e.g., 754a and / or 754b) may be processed by a computer system 773 or more generally by one or more signal processing devices (each of which may include an analog-to-digital converter configured to convert the analog signals from each sensor into digital signals for processing). The computer system 773 typically has one or more processors coupled to input / output ports and one or more memories via suitable buses or other communication mechanisms.

[0098] The computer system 773 may also include one or more input devices (e.g., keyboard, mouse, joystick) for providing user input (e.g., changing focus and other test protocol parameters). The computer system 773 may also be connected to a stage to control (e.g.,) sample position (e.g., focusing and scanning) and to other test system components to control other test parameters and configurations of such test system components.

[0099] Computer system 773 can be configured (e.g., using programming instructions) to provide a user interface (e.g., a computer screen) to display obtained intensity values, images, and other test results. Computer system 773 can be configured to analyze the intensity changes, phase, and / or other characteristics of the reflected and / or transmitted sense beam. Computer system 773 can be configured (e.g., using programming instructions) to provide a user interface (e.g., on a computer screen) to display obtained intensity values, images, and other test characteristics. In a particular embodiment, computer system 773 is configured to implement the testing techniques detailed above.

[0100] Because such information and program instructions can be implemented on a specially configured computer system, this system contains program instructions / computer code that can be stored on a computer-readable medium for performing the various operations described herein. Examples of machine-readable media include (but are not limited to): magnetic media, such as hard disks, floppy disks, and magnetic tapes; optical media, such as CD-ROM disks; magneto-optical media, such as optical discs; and hardware devices specially configured to store and execute program instructions, such as read-only memory devices (ROM) and random access memory (RAM). Examples of program instructions include both: machine code, for example generated by a compiler; and files containing higher-level code that can be executed by a computer using an interpreter.

[0101] In a particular embodiment, the system for inspecting a photomask includes at least one memory and at least one processor configured to perform the techniques described herein. Examples of inspection systems include a specially configured TeraScan available from KLA-Tec in Milpitas, California. TM DUV inspection system.

[0102] Although the invention has been described in detail for clarity, it should be understood that certain changes and modifications may be practiced within the scope of the appended claims. It should be noted that numerous alternatives exist for the processes, systems, and apparatus used to implement the invention. Therefore, the embodiments of the invention should be considered illustrative rather than restrictive, and the invention is not limited to the details given herein.

Claims

1. A method for inspecting a photomask, the method comprising: Multiple reference far-field images are simulated by inputting multiple reference near-field images into a physics-based model, wherein the multiple reference near-field images are generated by a trained deep learning model from a test portion of a design database used to fabricate test areas of a test mask, wherein the deep learning model is trained as a convolutional neural network (CNN) that does not form and incorporate the mask images onto an image plane, and the deep learning model includes mapping each mask database image to a near-field image generated by light interacting with the mask fabricated by the design database; Defects in the test area of ​​a test mask manufactured from the design database are inspected via a die-to-database process, which includes comparing a plurality of reference far-field images simulated by the physical model with a plurality of test images acquired by the inspection system from the test area of ​​the test mask.

2. The method of claim 1, wherein the deep learning model is trained by: (i) inputting a plurality of mask database images generated by a training portion of a design database into the deep learning model to generate a plurality of near-field mask images; (ii) receiving the plurality of near-field mask images into the physics-based model to simulate a plurality of corresponding far-field mask images at an image plane of the testing system via a physics-based process; (iii) providing a plurality of corresponding training mask images obtained by imaging a portion of a training mask manufactured by the training portion of the design database, wherein such training mask images are selected for pattern diversity and are defect-free; (iv) adjusting a plurality of parameters of the deep learning model and repeating the operation of inputting the plurality of mask database images into the deep learning model with the adjusted parameters to generate a plurality of adjusted near-field images received by the physics-based model to simulate the next plurality of far-field images until the difference between the next far-field mask image and the plurality of corresponding training mask images is minimized, thereby training the deep learning model.

3. The method of claim 2, wherein the physics-based process is based on the Hopkins method for generating each far-field mask image on the image plane of the inspection system based on each corresponding near-field image.

4. The method of claim 3, wherein the CNN does not include perturbations of each far-field image caused by field-dependent changes in the testing system and is independent of the testing system.

5. The method of claim 3, wherein the adjusted plurality of parameters of the deep learning model includes weights and / or deviations of a plurality of layers of the deep learning model, wherein the deep learning model is trained without adjusting any parameters of the physics-based model.

6. The method of claim 5, wherein the layer undergoing adjustment comprises a convolutional layer with non-linear activation.

7. The method of claim 6, wherein the deep learning model is trained without adjusting the parameters of one or more low-pass filter layers used for the downsampling operation.

8. The method of claim 3, wherein the CNN comprises one or more convolutional layers for offsetting the deviation between each mask database image and the physical mask portion generated from the mask database image, one or more layers for generating a plurality of downsampled images, and one or more layers for implementing a sparse representation of near-field resolution.

9. The method of claim 1, further comprising: Align the test image with the reference image; and The dynamic compensation process is applied relative to the test image to the reference image to counteract changes in the testing system, including focal fluctuations and / or field-dependent variations.

10. An inspection system for inspecting photomasks, the system comprising at least one memory and at least one processor configured to perform the following operations: Multiple reference far-field images are simulated by inputting multiple reference near-field images into a physics-based model, wherein the multiple reference near-field images are generated by a trained deep learning model from a test portion of a design database used to fabricate test areas of a test mask, wherein the deep learning model is trained as a convolutional neural network (CNN) that does not form and incorporate the mask images onto an image plane, and the deep learning model includes mapping each mask database image to a near-field image generated by light interacting with the mask fabricated by the design database; Defects in the test area of ​​a test mask manufactured from the design database are inspected via a die-to-database process, which includes comparing a plurality of reference far-field images simulated by the physical model with a plurality of test images acquired by the inspection system from the test area of ​​the test mask.

11. The system of claim 10, wherein the deep learning model is trained by: (i) inputting a plurality of mask database images generated by a training portion of a design database into the deep learning model to generate a plurality of near-field mask images; (ii) receiving the plurality of near-field mask images into the physics-based model to simulate a plurality of corresponding far-field mask images at an image plane of the test system via a physics-based process; (iii) providing a plurality of corresponding training mask images obtained by imaging a portion of a training mask manufactured by the training portion of the design database, wherein such training mask images are selected for pattern diversity and are defect-free; (iv) adjusting a plurality of parameters of the deep learning model and repeating the operation of inputting the plurality of mask database images into the deep learning model with the adjusted parameters to generate a plurality of adjusted near-field images received by the physics-based model to simulate the next plurality of far-field images until the difference between the next far-field mask image and the plurality of corresponding training mask images is minimized, thereby training the deep learning model.

12. The system of claim 11, wherein the physical-based process is based on the Hopkins method for generating each far-field mask image on the image plane of the testing system based on each corresponding near-field image.

13. The system of claim 12, wherein the CNN does not include perturbations of each far-field mask image caused by field-dependent variations in the testing system and is independent of the testing system.

14. The system of claim 12, wherein the adjusted plurality of parameters of the deep learning model includes weights and / or deviations of a plurality of layers of the deep learning model, wherein the deep learning model is trained without adjusting any parameters of the physics-based model.

15. The system of claim 14, wherein the layer undergoing adjustment comprises a convolutional layer with non-linear activation.

16. The system of claim 15, wherein the deep learning model is trained without adjusting the parameters of one or more low-pass filter layers used for the downsampling operation.

17. The system of claim 12, wherein the CNN comprises one or more convolutional layers for offsetting the deviation between each mask database image and a physical mask portion generated from the mask database image, one or more layers for generating a plurality of downsampled images, and one or more layers for implementing a sparse representation with near-field resolution.

18. The system of claim 10, wherein the at least one memory and the at least one processor are further configured to: Align the test image with the reference image; and The dynamic compensation process is applied relative to the test image to the reference image to counteract changes in the testing system, including focal fluctuations and / or field-dependent variations.

19. A non-transitory computer-readable medium having instructions stored thereon for performing the following operations: Multiple reference far-field images are simulated by inputting multiple reference near-field images into a physics-based model, wherein the multiple reference near-field images are generated by a trained deep learning model from a test portion of a design database used to fabricate test areas of a test mask, wherein the deep learning model is trained as a convolutional neural network (CNN) that does not form and incorporate the mask images onto an image plane, and the deep learning model includes mapping each mask database image to a near-field image generated by light interacting with the mask fabricated by the design database; Defects in the test area of ​​a test mask manufactured from the design database are inspected via a die-to-database process, which includes comparing a plurality of reference far-field images simulated by the physical model with a plurality of test images acquired by the inspection system from the test area of ​​the test mask.

Images