Structural parameter inversion method, measuring device and computer readable storage medium

Abstract

This application discloses a structural parameter inversion method, measurement device, and computer-readable storage medium. By inputting the initial structural parameters of the sample to be tested into a pre-trained diffraction intensity prediction model, the spatial distribution of diffraction intensity corresponding to the initial structural parameters is obtained. Based on the spatial distribution of diffraction intensity corresponding to the initial structural parameters, a simulated diffraction image of the sample to be tested is generated and compared with the measured diffraction image. With the goal of minimizing the difference between the simulated diffraction image and the measured diffraction image, the initial structural parameters are iteratively optimized to obtain the final structural parameters of the sample to be tested. Thus, there is no need to separate diffraction spots or integrate diffraction intensity, avoiding the information loss and error amplification problems faced in scenarios with highly overlapping or diffuse diffraction spots when separating diffraction spots and integrating diffraction intensity. This enables high-precision structural parameter measurement to be achieved even in such scenarios.

Description

Technical Field

[0001] This invention relates to the field of semiconductor testing technology, specifically to a method for structural parameter inversion, a measuring device, and a computer-readable storage medium. Background Technology

[0002] In advanced semiconductor manufacturing, scattering-based optical metrology is a core tool for monitoring the critical dimensions and three-dimensional morphology of semiconductor devices. Traditional scattering measurement methods rely on manually or semi-automatically identifying and separating diffraction spots from diffraction images acquired by a detector, and then integrating the intensity of the discrete diffraction spots to extract the diffraction efficiency of each diffraction order. This is then used to fit and invert structural parameters through a pre-built physical model. For example, in the technique of characterizing the morphology of two-dimensional periodic nanostructures in integrated circuits using small-angle X-ray scattering (SAXS), when measuring nanostructures with small period lengths, their scattering spectra exhibit diffraction peaks with a distinct two-dimensional periodic distribution. The peak position information of these diffraction peaks directly reflects the two-dimensional periodic arrangement and period size of the nanostructure, and structural parameters can be inverted based on this information.

[0003] However, as the critical dimensions of semiconductor devices increase (corresponding to large-periodic structures) or the structure becomes more aperiodic (such as contact holes, multilayer stacking, etc.), the spacing between diffraction spots often shrinks to the sub-pixel level, resulting in high aliasing or even complete dispersion into a continuous scattering spectrum. At this point, traditional integration methods cannot accurately distinguish the diffraction intensity of each diffraction order, not only losing key physical information but also introducing significant errors due to signal aliasing. This is especially problematic in dealing with the complex three-dimensional geometry and material coupling effects in advanced manufacturing processes. Summary of the Invention

[0004] This application provides a structural parameter inversion method, measurement equipment, and computer-readable storage medium to improve the accuracy of structural parameter measurement in scenarios with highly overlapping or diffused diffraction spots.

[0005] In a first aspect, this application provides a structural parameter inversion method, including:

[0006] Scattering measurements are performed on the sample to be tested to obtain the measured diffraction image of the sample to be tested;

[0007] The initial structural parameters of the sample to be tested are obtained, and the initial structural parameters are input into a pre-trained diffraction intensity prediction model to obtain the spatial distribution of diffraction intensity corresponding to the initial structural parameters.

[0008] A simulated diffraction image of the sample under test is generated based on the spatial distribution of the diffraction intensity.

[0009] Based on the difference between the simulated diffraction image and the measured diffraction image, with the goal of minimizing the difference, the initial structural parameters are iteratively optimized until a preset stopping condition is reached, thereby obtaining the final structural parameters of the sample to be tested.

[0010] In some embodiments, the training method for the diffraction intensity prediction model includes:

[0011] Obtain the structural parameters of the reference sample;

[0012] The actual spatial distribution of diffraction intensity is calculated based on the structural parameters of the reference sample.

[0013] Based on the structural parameters of the reference sample and its corresponding actual diffraction intensity spatial distribution, a neural network model is trained to obtain the diffraction intensity prediction model.

[0014] In some embodiments, calculating the corresponding spatial distribution of diffraction intensity based on the structural parameters of the reference sample includes:

[0015] Based on the structural parameters of the reference sample, the diffraction intensity information of different diffraction orders of the reference sample at different rotation angles is solved to obtain the actual diffraction intensity information of each diffraction order of the reference sample at each rotation angle.

[0016] In some embodiments, training a neural network model based on the structural parameters of the reference sample and its corresponding actual diffraction intensity spatial distribution to obtain the diffraction intensity prediction model includes:

[0017] For each rotation angle, the structural parameters of the reference sample and the rotation angle are input into the neural network model to obtain the predicted diffraction intensity information of all diffraction orders at that rotation angle. Based on the difference between the predicted diffraction intensity information and the corresponding actual diffraction intensity information, the neural network model is trained to obtain the diffraction intensity prediction model.

[0018] The step of inputting the initial structural parameters into a pre-trained diffraction intensity prediction model to obtain the spatial distribution of diffraction intensity corresponding to the initial structural parameters includes:

[0019] By iterating through all rotation angles, for each rotation angle, the initial structural parameters and the rotation angle are input into the diffraction intensity prediction model to obtain the diffraction intensity information of all diffraction orders under that rotation angle, so as to obtain the diffraction intensity information of all diffraction orders under all rotation angles after iterating through all rotation angles.

[0020] In some embodiments, training a neural network model based on the structural parameters of the reference sample and its corresponding actual diffraction intensity spatial distribution to obtain the diffraction intensity prediction model includes:

[0021] For each diffraction order, the structural parameters of the reference sample and the diffraction order are input into the neural network model to obtain the predicted diffraction intensity information of the diffraction order at all rotation angles. Based on the difference between the predicted diffraction intensity information and the corresponding actual diffraction intensity information, the neural network model is trained to obtain the diffraction intensity prediction model.

[0022] The step of inputting the initial structural parameters into a pre-trained diffraction intensity prediction model to obtain the spatial distribution of diffraction intensity corresponding to the initial structural parameters includes:

[0023] By iterating through all diffraction orders, for each diffraction order, the initial structural parameters and the diffraction order are input into the diffraction intensity prediction model to obtain the diffraction intensity information of the diffraction order at all rotation angles, so as to obtain the diffraction intensity information of all diffraction orders at all rotation angles after iterating through all diffraction orders.

[0024] In some embodiments, generating a simulated diffraction image of the sample under test based on the spatial distribution of diffraction intensity includes:

[0025] Based on the diffraction intensity information of each diffraction order at each rotation angle, each diffraction order at each rotation angle is mapped to the reciprocal space to obtain the position of each diffraction order at each rotation angle in the reciprocal space.

[0026] After adding simulated interference information to the positions of each diffraction order at each rotation angle in the reciprocal lattice space, the information is projected onto the detector plane to obtain a simulated diffraction image of the sample under test.

[0027] In some embodiments, the iterative optimization of the initial structural parameters based on the difference between the simulated diffraction image and the measured diffraction image, with the goal of minimizing the difference, includes:

[0028] Based on a loss function aimed at minimizing the difference, a gradient optimization algorithm is used to iteratively optimize the initial structural parameters.

[0029] In some embodiments, the structural parameters include at least one of linewidth, aperture, depth, sidewall angle, ellipticity, and film thickness.

[0030] Secondly, this application provides a measuring device, comprising:

[0031] An X-ray source is used to emit X-rays onto a sample to be tested.

[0032] A detector is used to receive X-rays scattered by the sample under test and obtain the corresponding measured diffraction image;

[0033] A processor is configured to execute the structural parameter inversion method described in the first aspect above to obtain the structural parameters of the sample under test.

[0034] Thirdly, this application provides a computer-readable storage medium storing a computer program that can be executed by a processor to implement the structural parameter inversion method described in the first aspect above.

[0035] The structural parameter inversion method, measurement device, and computer-readable storage medium provided in this application utilize a diffraction intensity prediction model for structural parameter inversion. By inputting the initial structural parameters of the sample under test into a pre-trained diffraction intensity prediction model, the spatial distribution of diffraction intensity corresponding to the initial structural parameters is obtained. Based on the spatial distribution of diffraction intensity corresponding to the initial structural parameters, a simulated diffraction image of the sample under test is generated and compared with the measured diffraction image. With the goal of minimizing the difference between the simulated and measured diffraction images, the initial structural parameters are iteratively optimized to obtain the final structural parameters of the sample under test. This eliminates the need to separate diffraction spots or integrate diffraction intensity, avoiding the problems of information loss and error amplification faced by methods that separate diffraction spots and integrate diffraction intensity in scenarios with highly overlapping or diffuse diffraction spots. Thus, high-precision structural parameter measurement can still be achieved in scenarios with highly overlapping or diffuse diffraction spots. Furthermore, the diffraction intensity prediction model achieves a forward mapping from structural parameters to diffraction intensity, avoiding the ill-conditioned inverse mapping from image to structural parameters. Attached Figure Description

[0036] Figure 1 These are flowcharts of the structural parameter inversion methods in some embodiments;

[0037] Figure 2 This is a flowchart of generating simulated diffraction images of the sample under test based on the spatial distribution of diffraction intensity in some embodiments;

[0038] Figure 3 This is a flowchart of the training method for the diffraction intensity prediction model in some embodiments;

[0039] Figure 4 These are schematic diagrams of the measuring device in some embodiments. Detailed Implementation

[0040] The present invention will now be described in further detail with reference to specific embodiments and accompanying drawings. Similar elements in different embodiments are referred to by associated similar element reference numerals. In the following embodiments, many details are described to facilitate a better understanding of this application. However, those skilled in the art will readily recognize that some features may be omitted in different situations, or may be replaced by other elements, materials, or methods. In some cases, certain operations related to this application are not shown or described in the specification. This is to avoid obscuring the core parts of this application with excessive description. For those skilled in the art, detailed description of these related operations is not necessary; they can fully understand the related operations based on the description in the specification and general technical knowledge in the art.

[0041] Furthermore, the features, operations, or characteristics described in the specification can be combined in any suitable manner to form various embodiments. At the same time, the steps or actions in the method description can be rearranged or adjusted in a manner obvious to those skilled in the art. Therefore, the various orders in the specification and drawings are only for the clear description of a particular embodiment and do not imply a necessary order, unless otherwise stated that a particular order must be followed.

[0042] The serial numbers assigned to components or physical quantities in this document, such as "first," "second," etc., are used only to distinguish the described objects and have no sequential or technical meaning. They should not be construed as indicating or implying relative importance or implicitly specifying the number of technical features indicated. "Multiple" means two or more. Unless otherwise specified, "connection" or "linkage" in this application includes both direct and indirect connections (linkages).

[0043] As semiconductor device structures continue to shrink and become more complex (such as the emergence of large-period arrays, quasi-random contact holes, and multi-layer three-dimensional stacking), their diffraction patterns exhibit two significant characteristics: First, the diffraction spots are extremely dense in reciprocal space, with their spacing much smaller than the detector pixel size, resulting in severe signal aliasing at the pixel level; second, the diffraction spots tend to diffuse due to structural complexity, and may even merge into a quasi-continuous spectrum. In such scenarios, the traditional method based on "spot detection-intensity integration" faces fundamental bottlenecks: (1) failure of physical information extraction: it is impossible to reliably separate overlapping diffraction orders at sub-pixel accuracy, resulting in significant deviations in the extracted diffraction efficiency and loss of key structural sensitive information; (2) error propagation and amplification: the error of the initial diffraction order will be directly transmitted to the downstream parameter fitting process and amplified by the parameter fitting optimization algorithm, ultimately leading to unreliable or even completely wrong structural parameter inversion results; (3) poor generalization of the method: the above methods can still cope with simple structures with good periodicity, but for emerging quasi-periodic, non-periodic or complex structures with strong three-dimensional features, the physical assumptions on which the traditional methods rely no longer hold, making it difficult to put into practical use.

[0044] To overcome these bottlenecks, the industry has developed several technical approaches. One is full-image fitting, which directly compares the measured image with the physical simulation results to determine the optimal structural parameters. Another is fitting after extracting diffraction orders, which first estimates the diffraction efficiency of each order from the image and then performs parameter inversion. These methods still have significant limitations: full-image fitting is sensitive to initial values ​​and easily gets trapped in local minima; while diffraction order extraction is unreliable in sub-pixel scenarios.

[0045] Furthermore, in recent years, methods have emerged that use artificial intelligence (AI) to predict structural parameters directly from images end-to-end. Essentially, these methods still attempt to solve the inverse problem of "from image to structural parameters." However, this is a highly ill-conditioned inverse problem. Due to the phenomenon of "parameter coupling" (i.e., different combinations of structural parameters may produce visually highly similar diffraction images), such methods suffer from inherent prediction ambiguity, making it difficult for the model to learn stable and reliable mapping relationships, severely limiting their application in high-precision industrial metrology.

[0046] Overall, existing solutions mostly focus on patching localized issues and have not yet established an efficient and stable unified framework. Therefore, how to achieve high-precision, robust, and rapid structural parameter inversion in extreme scenarios with highly aliased or diffused diffraction spots, thereby providing reliable online metrology capabilities for next-generation semiconductor manufacturing, is an urgent problem to be solved.

[0047] This application proposes a novel paradigm for structural parameter measurement. Instead of addressing signal aliasing in the reverse engineering process, this approach utilizes a novel, physically interpretable forward computation paradigm to fundamentally bypass the flawed inverse mapping from image to structural parameters. Instead, it enables AI to learn a stable forward physical process from structural parameters to diffraction intensity and then to diffraction image. Specifically, an AI model is first trained, establishing a forward mapping from the structural parameters of the semiconductor device to the diffraction physical response. This results in a forward AI model that takes the semiconductor device's structural parameters (such as linewidth, sidewall angles, film thickness, and material optical constants) as input and outputs a complete spatial distribution of diffraction intensity. This distribution is then mapped to a simulated diffraction image consistent with the detector geometry, light source conditions, and / or the detector's point spread function. This simulated diffraction image is compared with the measured diffraction image, and the difference between the two is used as the optimization objective. The structural parameters are updated through iterative optimization.

[0048] This scheme eliminates the need for explicit separation or integration of subpixel-level diffraction spots. Instead, it models the entire imaging chain from structural parameters to diffraction images as an end-to-end differentiable physical process. This fundamentally avoids the problems of information loss and error amplification in traditional methods when diffraction spots are dense, overlapping, or diffuse. Furthermore, AI achieves a forward mapping from structural parameters to diffraction intensity and then to diffraction images, bypassing the ill-conditioned inverse problem of "from image to structural parameters." This is beneficial for achieving high-precision, robust, and rapid structural parameter inversion in extreme scenarios where diffraction spots are highly overlapping or diffuse.

[0049] The structural parameter inversion method provided in this application will be explained below. Please refer to [link / reference]. Figure 1 The structural parameter inversion method in some embodiments includes steps 100 to 400, which are described in detail below.

[0050] Step 100: Perform scattering measurement on the sample to be tested to obtain the measured diffraction image of the sample.

[0051] The sample under test can be a wafer, display panel, etc., with nanostructures such as through holes and gaps on its surface. Scattering measurements can be performed using scattering measurement methods such as SAXS (e.g., transmission small-angle X-ray scattering, T-SAXS). X-rays are incident on the sample at a certain angle, and are scattered by the nanostructures on the sample as they pass through it. The scattered rays are then received by a detector to form a measured diffraction image of the sample. Due to the periodicity of the structure, the diffraction image is usually composed of discrete Bragg diffraction orders, each of which corresponds to a specific (h, k) index (e.g., (0,0), (1,0), (1,1), etc.) or (h,k,l) ​​index in the reciprocal space of the crystal lattice. Typically, the sample under test is placed on a rotatable sample stage, and during the scattering measurement, the sample needs to be rotated by the sample stage at a series of rotation angles.

[0052] Step 200: Obtain the initial structural parameters of the sample to be tested, input the initial structural parameters into the pre-trained diffraction intensity prediction model, and obtain the spatial distribution of diffraction intensity corresponding to the initial structural parameters.

[0053] The structural parameters include at least one of linewidth, aperture, depth, sidewall angle, ellipticity, and film thickness, reflecting the morphology or geometric characteristics of the nanostructure. The initial structural parameters of the sample under test can be estimated based on actual diffraction images or set empirically. The spatial distribution of diffraction intensity is used to characterize the diffraction intensity at different locations in space, including diffraction intensity information of different diffraction orders at different rotation angles of the sample under test. This diffraction intensity information can be expressed using complex amplitude, intensity curves, etc.

[0054] The diffraction intensity prediction model is built upon a neural network model. Its core objective is to establish a mapping from structural parameters to the spatial distribution of diffraction intensity, i.e., a forward mapping. A neural network model can be trained using pre-collected sample structural parameters and their corresponding spatial distributions of diffraction intensity (as true values). This allows the model to learn the ability to predict the corresponding spatial distribution of diffraction intensity based on the structural parameters, thus obtaining the diffraction intensity prediction model. The specific training method will be described below.

[0055] Step 300: Generate a simulated diffraction image of the sample under test based on the spatial distribution of diffraction intensity.

[0056] By mapping the diffraction intensity information of each diffraction order at each rotation angle to the corresponding position on the detector plane, the diffraction pattern of the sample under test can be simulated, thus obtaining a simulated diffraction image. Please refer to... Figure 2 In some embodiments of this application, step 300 includes steps 310 and 320.

[0057] Step 310: Based on the diffraction intensity information of each diffraction order at each rotation angle, map each diffraction order at each rotation angle to the reciprocal space to obtain the position of each diffraction order at each rotation angle in the reciprocal space.

[0058] Let i denote the i-th diffraction order. Let j represent the j-th rotation angle, (i, Let represent the i-th diffraction order at the j-th rotation angle. Then, each (i, The corresponding position in reciprocal space, for example, the Q-space information (Q) in three-dimensional reciprocal space. x Q y Q z ), and each (i, The diffraction intensity information of ) is mapped to its corresponding position in reciprocal space.

[0059] Step 320: After adding simulated interference information to the positions of each diffraction order in reciprocal space at each rotation angle, project it onto the detector plane to obtain a simulated diffraction image of the sample under test.

[0060] In actual measurement environments, there are often various types of interference information that are not considered under ideal conditions, such as background scattering, photon noise, and the detector's point spread function. This step simulates these interference information and adds the corresponding simulated interference information to the positions of each diffraction order in reciprocal space at each rotation angle. Finally, the reciprocal space positions (e.g., Q-space information) containing diffraction intensity information and simulated interference information are projected onto the detector's detection plane, thereby obtaining a realistic, high-fidelity simulated diffraction image of the sample under test that is completely comparable to the real measurement environment.

[0061] Step 400: Based on the difference between the simulated diffraction image and the measured diffraction image, with the goal of minimizing the difference, the initial structural parameters are iteratively optimized until a preset stopping condition is reached, and the final structural parameters of the sample to be tested are obtained.

[0062] Specifically, the simulated diffraction image is compared with the measured diffraction image to obtain the difference between them. The initial structural parameters are updated with the goal of minimizing this difference. The updated structural parameters are then input into the diffraction intensity prediction model to obtain the corresponding diffraction intensity spatial distribution. Based on this spatial distribution, a simulated diffraction image of the sample is generated, and this image is again compared with the measured diffraction image to obtain the difference. The structural parameters are updated based on this difference. This process is repeated, continuously adjusting the structural parameters input to the diffraction intensity prediction model to make the simulated diffraction image approximate the measured diffraction image, until a preset stopping condition is met, thus obtaining the final structural parameters of the sample. To improve the accuracy and precision of the structural parameters, pixel-level comparisons can be performed between the simulated and measured diffraction images. The preset stopping condition could be that the difference between the simulated and measured diffraction images is less than a difference threshold, or that a preset number of iterations has been reached.

[0063] In some embodiments, a loss function, such as a mean squared error loss function, can be constructed to minimize the difference between the simulated and measured diffraction images. Then, based on the difference between the simulated and measured diffraction images, the initial structural parameters are iteratively optimized with the goal of minimizing the difference. This includes: using a gradient optimization algorithm to iteratively optimize the initial structural parameters based on the loss function aimed at minimizing the difference.

[0064] As can be seen, the structural parameter inversion method of this application transforms the complex structural parameter inversion problem into an image matching optimization problem driven by a differentiable neural network model. This completely avoids the step of relying on manual extraction of diffraction spots or diffraction peak intensities in traditional methods. It can still achieve high-precision, fully automatic structural parameter measurement even in scenarios where diffraction spots are highly overlapping or diffuse.

[0065] The training method for the diffraction intensity prediction model is described below. Please refer to [link / reference]. Figure 3 In some embodiments, the training method for the diffraction intensity prediction model includes steps 410 to 430, which are described in detail below.

[0066] Step 410: Obtain the structural parameters of the reference sample.

[0067] The reference sample can be an actual sample or a simulated sample. Multiple reference samples are used to obtain multiple sets of structural parameters for training. In some embodiments, the structural parameters can be parameterized as a high-dimensional vector. As inputs to the diffraction intensity prediction model, where, Let n represent a real number, where n is the total number of structure parameters.

[0068] Step 420: Calculate the actual spatial distribution of diffraction intensity based on the structural parameters of the reference sample, and use it as the prediction target of the neural network, that is, the actual spatial distribution of diffraction intensity of the reference sample as the true value of the prediction.

[0069] In some embodiments, the spatial distribution of diffraction intensity is calculated based on the structural parameters of the reference sample, including: based on the structural parameters of the reference sample, solving for the diffraction intensity information of different diffraction orders when the reference sample is at different rotation angles, and obtaining the actual diffraction intensity information of each diffraction order when the reference sample is at each rotation angle.

[0070] The rotation angle and the number of diffraction orders can be determined based on different samples. In a specific embodiment, for each diffraction order, the diffraction intensity information of that diffraction order can be solved based on the layered model and Green's function when the reference sample is at different rotation angles, thus obtaining the actual diffraction intensity information of all diffraction orders at all rotation angles. This series of calculations is physically equivalent to a simulation of a complete forward scattering process, ensuring the physical correctness of the results. By performing the same calculations on each set of structural parameters, the actual diffraction intensity information of each diffraction order at each rotation angle is finally generated for each set of structural parameters.

[0071] Taking diffraction intensity information as the complex amplitude as an example, for structural parameter p, rotation angle The complex amplitude of the i-th diffraction order can be expressed as:

[0072] ,

[0073] in, For the real part, This is the imaginary part. For each set of structural parameters p, the complex amplitude of each diffraction order at each rotation angle is solved one by one based on the hierarchical model and Green's function. Finally, a structured tensor containing the complex amplitudes of all diffraction orders at all rotation angles is generated for each set of structural parameters p. As the target data for model training, that is, forming {p, The mapping relationship of}.

[0074] Step 430: Based on the structural parameters of the reference sample and its corresponding actual diffraction intensity spatial distribution, train a neural network model to obtain a diffraction intensity prediction model.

[0075] The structural parameters of a reference sample can be input into a neural network model to obtain the predicted spatial distribution of diffraction intensity. A loss function is constructed to characterize the difference between the predicted and actual spatial distributions of diffraction intensity. The neural network model is trained using supervised learning, and the network weights are optimized through gradient descent until the loss function converges, resulting in the diffraction intensity prediction model. After training, for any new set of structural parameters, the diffraction intensity prediction model can perform fast forward inference to generate its diffraction intensity spatial distribution.

[0076] During the model training phase, the core objective is to learn the mapping from structural parameters to diffraction intensity information. To adapt to different tasks and application scenarios, two specific mapping relationships can be constructed to achieve efficient learning of the network.

[0077] One type is full-order prediction for a specific rotation angle, i.e., fixed rotation angle. The diffraction intensity prediction model aims to predict the diffraction intensity information corresponding to all diffraction orders (i.e., all (h,k) or (h,k,l) ​​indices) at a given rotation angle. This strategy aims to allow the diffraction intensity prediction model to learn the global response at a specific rotation angle in one go, thereby obtaining the complete diffraction spectrum.

[0078] Specifically, in step 430, a neural network model is trained based on the structural parameters of the reference sample and its corresponding actual diffraction intensity spatial distribution. This includes: for each rotation angle, inputting the structural parameters of the reference sample and the rotation angle into a neural network model to obtain the predicted diffraction intensity information of all diffraction orders at that rotation angle; and training the neural network model based on the difference between the predicted diffraction intensity information and the corresponding actual diffraction intensity information to obtain a diffraction intensity prediction model.

[0079] The rotation angle can be input into the neural network model as an additional feature along with the structural parameters. Each rotation angle can be combined with various sets of structural parameters and input into the neural network model to obtain the predicted diffraction intensity information of all diffraction orders at that rotation angle. The predicted diffraction intensity information is then compared with the corresponding actual diffraction intensity information to train the neural network model. By performing the same operation on each rotation angle in sequence, the neural network model can learn to predict the diffraction intensity information of all diffraction orders at a specific rotation angle.

[0080] After training, the diffraction intensity prediction model can be iterated through all rotation angles to generate a complete spatial distribution of diffraction intensity. Specifically, step 200 involves inputting the initial structural parameters into the pre-trained diffraction intensity prediction model to obtain the spatial distribution of diffraction intensity corresponding to the initial structural parameters. This includes iterating through all rotation angles, and for each rotation angle, inputting the initial structural parameters and the corresponding rotation angle into the diffraction intensity prediction model to obtain the diffraction intensity information of all diffraction orders at that rotation angle. This process ensures that after iterating through all rotation angles, the diffraction intensity information of all diffraction orders at all rotation angles is obtained.

[0081] Another approach is full-angle prediction for a specific diffraction order, where the diffraction order i is fixed (i.e., a (h,k) or (h, k,l) index is defined). The diffraction intensity prediction model aims to predict the diffraction intensity information corresponding to that diffraction order at all rotation angles, such as complex amplitude curves (i.e., sequences containing real and imaginary parts). This strategy aims to enable the diffraction intensity prediction model to accurately learn the physical behavior of a single diffraction order evolving with rotation angle.

[0082] Specifically, in step 430, a neural network model is trained based on the structural parameters of the reference sample and its corresponding actual diffraction intensity spatial distribution. This includes: for each diffraction order, inputting the structural parameters of the reference sample and the diffraction order into the neural network model to obtain the predicted diffraction intensity information of the diffraction order at all rotation angles; and training the neural network model based on the difference between the predicted diffraction intensity information and the corresponding actual diffraction intensity information to obtain a diffraction intensity prediction model.

[0083] Each diffraction order can be assigned a unique identifier, which is then input into the neural network model along with the structural parameters. Each diffraction order, combined with its set of structural parameters, is input into the neural network model to obtain predicted diffraction intensity information for that order at all rotation angles. This predicted intensity is then compared with the corresponding actual diffraction intensity information to train the neural network model. By performing the same operation on each diffraction order sequentially, the neural network model learns to predict the diffraction intensity information of a specific diffraction order at all rotation angles.

[0084] After training, the diffraction intensity prediction model can be iterated through all diffraction orders to generate a complete spatial distribution of diffraction intensity. Specifically, step 200 involves inputting the initial structural parameters into the pre-trained diffraction intensity prediction model to obtain the spatial distribution of diffraction intensity corresponding to the initial structural parameters. This includes iterating through all diffraction orders, and for each diffraction order, inputting the initial structural parameters and the diffraction order itself into the diffraction intensity prediction model to obtain the diffraction intensity information for that diffraction order at all rotation angles. This process ensures that after iterating through all diffraction orders, the diffraction intensity information for all diffraction orders at all rotation angles is obtained.

[0085] This application also provides a measuring device, please refer to... Figure 4 In some embodiments, the measuring device includes an X-ray source 1, a detector 2, and a processor 3.

[0086] X-ray source 1 is used to emit X-rays to the sample 4 to be tested.

[0087] Detector 2 is used to receive the X-rays scattered by the sample 4 and obtain the corresponding measured diffraction image. Detector 2 can be a CCD (charge coupled device) detector, a single-photon counting detector, etc.

[0088] The processor 3 is used to execute the structural parameter inversion method provided in the embodiments of this application to obtain the structural parameters of the sample 4 to be tested.

[0089] The structural parameter inversion method and measurement device according to the above embodiments generate diffraction images based on the AI ​​forward model of diffraction intensity prediction to invert structural parameters. This eliminates the need to extract diffraction spots or peaks for diffraction intensity integration. Even in scenarios with highly overlapping or diffused diffraction spots, high-precision, fully automated structural parameter measurement can still be achieved. This solves the technical problem of traditional scattering measurement methods failing to accurately extract diffraction efficiencies at various levels when diffraction spots are highly dense, overlapping, or diffuse. Furthermore, the technical solution of this application considers the ill-conditioned difficulties of inverse mapping, while the physical process of forward mapping is clear and controllable. It constructs an end-to-end differentiable AI forward model "from structural parameters to diffraction intensity to diffraction image" to bypass the inverse solution problem, solving the problem of ambiguity in direct structural parameter prediction caused by parameter coupling in existing AI-based structural parameter measurement methods.

[0090] The technical solution of this application can be roughly divided into two stages: a forward generation stage and a reverse inversion stage. In the forward generation stage, the advantages of simulation are fully utilized. Structural parameters are input into a pre-trained diffraction intensity prediction model to predict the diffraction intensity information of each diffraction order at different rotation angles, unaffected by issues such as pixel aliasing, background scattering, and noise in the actual detector. Then, a simulated diffraction image of the sample under test is synthesized based on the spatial distribution of diffraction intensity. In some embodiments, this "ideal" physical response of the spatial distribution of diffraction intensity is intentionally and controllably superimposed with simulated non-ideal factors (such as detector pixelation, point spread function, background scattering, and photon noise) to render a high-fidelity two-dimensional diffraction image that is completely comparable to the real measurement environment. In the reverse inversion stage, the traditional diffraction spot or peak position extraction steps are directly avoided. The rendered "synthetic image" (simulated diffraction image) is directly compared and globally optimized with the measured "aliased image" (measured diffraction image), using the differences to drive the update of structural parameters.

[0091] The advantages of this technical solution are as follows: In the forward generation stage, the complete spatial distribution of diffraction intensity can be obtained losslessly at a speed of milliseconds using a diffraction intensity prediction model; while in the reverse inversion stage, it mimics the "observation perspective" of real equipment, using the overall measured diffraction image containing aliasing as the comparison object to iteratively optimize the accurate structural parameters. In addition, in some embodiments, the actual diffraction intensity spatial distribution corresponding to the structural parameters of the reference sample is calculated using a physical model as training data. This allows for the generation of high-quality training data by utilizing the precise intrinsic knowledge of the physical model, while completely avoiding the inherent bottleneck of manually extracting fragile physical signals from aliased images in practical applications.

[0092] Furthermore, since the diffraction intensity prediction model predicts the spatial distribution of diffraction intensity without being affected by issues such as pixel aliasing, background scattering, and noise in the actual detector, it is physically interpretable and more efficient than directly predicting diffraction images, but also more susceptible to adding interference factors. By generating simulated diffraction images of the sample under test using the physically interpretable diffraction intensity spatial distribution as an intermediate representation, and comparing them with measured diffraction images to retrieve inverted structural parameters, it retains the stability of traditional order fitting while fully leveraging the expressive power and computational efficiency of AI, significantly improving the robustness and accuracy of inversion for periodic, quasi-periodic, and aperiodic structures.

[0093] Those skilled in the art will understand that all or part of the functions of the various methods in the above embodiments can be implemented by hardware or by computer programs. When all or part of the functions in the above embodiments are implemented by computer programs, the program can be stored in a computer-readable storage medium, which may include: read-only memory, random access memory, disk, optical disk, hard disk, etc., and the program is executed by a computer to achieve the above functions. For example, the program can be stored in the memory of a device, and when the program in the memory is executed by the processor, all or part of the above functions can be achieved. In addition, when all or part of the functions in the above embodiments are implemented by computer programs, the program can also be stored in a server, another computer, disk, optical disk, flash drive, or external hard drive, etc., and can be downloaded or copied to the memory of a local device, or the system of the local device can be updated. When the program in the memory is executed by the processor, all or part of the functions in the above embodiments can be achieved.

[0094] The above examples illustrate the present invention only to aid in understanding it and are not intended to limit the scope of the invention. Those skilled in the art can make various simple deductions, modifications, or substitutions based on the principles of this invention.

Claims

1. A structural parameter inversion method, characterized in that, include: Scattering measurements are performed on the sample to be tested to obtain the measured diffraction image of the sample to be tested; The initial structural parameters of the sample to be tested are obtained, and the initial structural parameters are input into a pre-trained diffraction intensity prediction model to obtain the spatial distribution of diffraction intensity corresponding to the initial structural parameters. A simulated diffraction image of the sample under test is generated based on the spatial distribution of the diffraction intensity. Based on the difference between the simulated diffraction image and the measured diffraction image, with the goal of minimizing the difference, the initial structural parameters are iteratively optimized until a preset stopping condition is reached, thereby obtaining the final structural parameters of the sample to be tested.

2. The structural parameter inversion method as described in claim 1, characterized in that, The training method for the diffraction intensity prediction model includes: Obtain the structural parameters of the reference sample; The actual spatial distribution of diffraction intensity is calculated based on the structural parameters of the reference sample. Based on the structural parameters of the reference sample and its corresponding actual diffraction intensity spatial distribution, a neural network model is trained to obtain the diffraction intensity prediction model.

3. The structural parameter inversion method as described in claim 2, characterized in that, The calculation of the corresponding diffraction intensity spatial distribution based on the structural parameters of the reference sample includes: Based on the structural parameters of the reference sample, the diffraction intensity information of different diffraction orders of the reference sample at different rotation angles is solved to obtain the actual diffraction intensity information of each diffraction order of the reference sample at each rotation angle.

4. The structural parameter inversion method as described in claim 3, characterized in that, The process of training a neural network model based on the structural parameters of the reference sample and its corresponding actual diffraction intensity spatial distribution to obtain the diffraction intensity prediction model includes: For each rotation angle, the structural parameters of the reference sample and the rotation angle are input into the neural network model to obtain the predicted diffraction intensity information of all diffraction orders at that rotation angle. Based on the difference between the predicted diffraction intensity information and the corresponding actual diffraction intensity information, the neural network model is trained to obtain the diffraction intensity prediction model. The step of inputting the initial structural parameters into a pre-trained diffraction intensity prediction model to obtain the spatial distribution of diffraction intensity corresponding to the initial structural parameters includes: By iterating through all rotation angles, for each rotation angle, the initial structural parameters and the rotation angle are input into the diffraction intensity prediction model to obtain the diffraction intensity information of all diffraction orders under that rotation angle, so as to obtain the diffraction intensity information of all diffraction orders under all rotation angles after iterating through all rotation angles.

5. The structural parameter inversion method as described in claim 3, characterized in that, The process of training a neural network model based on the structural parameters of the reference sample and its corresponding actual diffraction intensity spatial distribution to obtain the diffraction intensity prediction model includes: For each diffraction order, the structural parameters of the reference sample and the diffraction order are input into the neural network model to obtain the predicted diffraction intensity information of the diffraction order at all rotation angles. Based on the difference between the predicted diffraction intensity information and the corresponding actual diffraction intensity information, the neural network model is trained to obtain the diffraction intensity prediction model. The step of inputting the initial structural parameters into a pre-trained diffraction intensity prediction model to obtain the spatial distribution of diffraction intensity corresponding to the initial structural parameters includes: By iterating through all diffraction orders, for each diffraction order, the initial structural parameters and the diffraction order are input into the diffraction intensity prediction model to obtain the diffraction intensity information of the diffraction order at all rotation angles, so as to obtain the diffraction intensity information of all diffraction orders at all rotation angles after iterating through all diffraction orders.

6. The structural parameter inversion method as described in claim 4 or 5, characterized in that, The process of generating a simulated diffraction image of the sample under test based on the spatial distribution of diffraction intensity includes: Based on the diffraction intensity information of each diffraction order at each rotation angle, each diffraction order at each rotation angle is mapped to the reciprocal space to obtain the position of each diffraction order at each rotation angle in the reciprocal space. After adding simulated interference information to the positions of each diffraction order at each rotation angle in the reciprocal lattice space, the information is projected onto the detector plane to obtain a simulated diffraction image of the sample under test.

7. The structural parameter inversion method according to any one of claims 1 to 6, characterized in that, The step of iteratively optimizing the initial structural parameters based on the difference between the simulated diffraction image and the measured diffraction image, with the goal of minimizing the difference, includes: Based on a loss function aimed at minimizing the difference, a gradient optimization algorithm is used to iteratively optimize the initial structural parameters.

8. The structural parameter inversion method according to any one of claims 1 to 7, characterized in that, The structural parameters include at least one of linewidth, aperture, depth, sidewall angle, ellipticity, and film thickness.

9. A measuring device, characterized in that, include: An X-ray source is used to emit X-rays onto a sample to be tested. A detector is used to receive X-rays scattered by the sample under test and obtain the corresponding measured diffraction image; A processor is configured to execute the structural parameter inversion method as described in any one of claims 1 to 8 to obtain the structural parameters of the sample under test.

10. A computer-readable storage medium, characterized in that, The medium stores a computer program that can be executed by a processor to implement the structural parameter inversion method as described in any one of claims 1 to 8.

Images