Training method and device of semiconductor structure prediction model, and optical critical dimension measurement method and device

Abstract

The one or more embodiments of the specification provide a training method and device of a semiconductor structure prediction model and an optical critical dimension measurement method and device. The training method comprises: obtaining original training samples, the original training samples comprising original spectrum values and a plurality of groups of target parameters corresponding to semiconductor structures; performing the following steps in a loop until a condition is met: training a structure prediction model using the original training samples as training samples used in a first round of training; after the training is completed, determining whether the accuracy of the structure prediction model meets a standard; in the case where the accuracy does not meet the standard, determining a current round of prediction effect of each group of target parameters; for each group of target parameters, in the case where the current round of prediction effect is better than a previous round of prediction effect, updating a model weight corresponding to the target parameters; determining incremental training samples used in a next round of training; and determining the original training samples and the incremental training samples as training samples used in the next round of training to continue training the structure prediction model.

Description

Technical Field

[0001] This specification relates to one or more embodiments in the field of semiconductor manufacturing, and more particularly to a method and apparatus for training a semiconductor structure prediction model and a method and apparatus for measuring optical critical dimensions. Background Technology

[0002] Optical critical dimension (OCD) measurement is a technique used to measure and control the dimensions of microstructures, and it is widely used in fields such as semiconductor manufacturing. OCD measurement technology can predict the target parameters of a sample structure by measuring the spectral data of the sample (such as a semiconductor structure under test).

[0003] The target parameters of the structure of the sample under test (e.g., the semiconductor under test) typically include multiple dimensions, and the training methods for structure prediction models provided by related technologies struggle to simultaneously achieve training accuracy for various target parameters across different dimensions. To ensure appropriate accuracy for each target parameter dimension, it is usually necessary to train the target parameters for each dimension separately, i.e., constructing multiple structure prediction models, each corresponding to one target parameter dimension, and then training each model separately. However, this approach often consumes several times the training time and computational resources. Summary of the Invention

[0004] In view of the above, one or more embodiments of this specification provide the following technical solutions:

[0005] According to a first aspect of one or more embodiments of this specification, a method for training a semiconductor structure prediction model is proposed. The structure prediction model is used to predict target parameters of multiple sets of not entirely identical semiconductor structures, each set of target parameters corresponding to a set of model weights. The method includes:

[0006] Obtain original training samples, which include original spectral values ​​and multiple sets of target parameters corresponding to semiconductor structures, wherein the original spectral values ​​are sample features and the multiple sets of target parameters are corresponding multiple sets of sample labels that are not completely identical;

[0007] Repeat the following steps until the condition is met:

[0008] The original training samples are used as the training samples for the first round of training to train the structure prediction model;

[0009] After training is completed, determine whether the accuracy of the structure prediction model meets the standard.

[0010] If the accuracy of the structural prediction model is not up to standard, determine the prediction effect of each set of target parameters in this round.

[0011] For each set of target parameters, if the prediction effect in this round is better than the prediction effect in the previous round, the model weights corresponding to the target parameters are updated.

[0012] Determine the incremental training samples to be used in the next round of training. The incremental training samples include incremental spectral values ​​and corresponding sets of target parameters.

[0013] The original training samples and the incremental training samples are selected as the training samples for the next round of training to continue training the structure prediction model.

[0014] Optional, also includes:

[0015] If the prediction result in this round is not better than the prediction result in the previous round, the model weights corresponding to the target parameters are retained.

[0016] Optionally, the process of determining whether the accuracy of the structural prediction model meets the standard includes:

[0017] Calculate the training score of the structure prediction model;

[0018] When the training score reaches the training score threshold, the structure prediction model is tested using test samples to obtain a test score.

[0019] If the test score reaches the test score threshold, the accuracy of the structural prediction model is determined to be up to standard.

[0020] If the test score does not reach the test score threshold, it is determined that the accuracy of the structural prediction model is substandard.

[0021] Optionally, the training score or the test score is obtained by weighting the prediction performance of each group of target parameters.

[0022] Optionally, the conditions include: the accuracy of the structural prediction model meets the standard or reaches a preset number of training rounds.

[0023] According to a second aspect of one or more embodiments of this specification, a method for measuring optical critical dimensions is provided, the method comprising:

[0024] Obtain the measurement spectral values ​​of the semiconductor under test;

[0025] The measured spectral values ​​are input into a structure prediction model with sufficient accuracy to obtain the target parameter prediction results of the semiconductor structure under test output by the structure prediction model.

[0026] The structural prediction model is trained using the aforementioned method.

[0027] According to a third aspect of one or more embodiments of this specification, a training apparatus for a semiconductor structure prediction model is provided. The structure prediction model is used to predict target parameters of multiple sets of not entirely identical semiconductor structures, each set of target parameters corresponding to a set of model weights. The apparatus includes:

[0028] The sample acquisition unit acquires original training samples, which include original spectral values ​​and multiple sets of target parameters corresponding to semiconductor structures. The original spectral values ​​are sample features, and the multiple sets of target parameters are corresponding multiple sets of not completely identical sample labels.

[0029] The model training unit repeatedly executes the following steps until the conditions are met:

[0030] The original training samples are used as the training samples for the first round of training to train the structure prediction model;

[0031] After training is completed, determine whether the accuracy of the structure prediction model meets the standard.

[0032] If the accuracy of the structural prediction model is not up to standard, determine the prediction effect of each set of target parameters in this round.

[0033] For each set of target parameters, if the prediction effect in this round is better than the prediction effect in the previous round, the model weights corresponding to the target parameters are updated.

[0034] Determine the incremental training samples to be used in the next round of training. The incremental training samples include incremental spectral values ​​and corresponding sets of target parameters.

[0035] The original training samples and the incremental training samples are selected as the training samples for the next round of training to continue training the structure prediction model.

[0036] According to a fourth aspect of one or more embodiments of this specification, an optical critical dimension measuring device is provided, the device comprising:

[0037] The spectral value acquisition unit acquires the measured spectral values ​​of the semiconductor under test.

[0038] The structure prediction unit inputs the measured spectral values ​​into a structure prediction model with sufficient accuracy to obtain the target parameter prediction results of the semiconductor structure under test output by the structure prediction model.

[0039] The structural prediction model is trained using the aforementioned method.

[0040] According to a fifth aspect of one or more embodiments of this specification, an electronic device is provided, comprising: a processor; a memory for storing processor-executable instructions; wherein the processor performs the steps of the method as described above by executing the executable instructions.

[0041] According to a sixth aspect of one or more embodiments of this specification, a computer-readable storage medium is provided that stores computer instructions thereon, which, when executed by a processor, implement the steps of the method as described above.

[0042] According to a seventh aspect of one or more embodiments of this specification, a computer program product is provided, comprising a computer program / instructions that, when executed by a processor, implement the steps of the method as described above.

[0043] As can be seen from the above embodiments, by adopting the above implementation method, after training the structural prediction model once, the model weights can be updated based on the prediction performance of different dimensional target parameters. If the prediction performance in the current round is better than the previous round, the corresponding model weights can be retained, i.e., the corresponding model weights are not updated, even if the prediction performance is not better than the previous round. This achieves simultaneous optimization of the prediction accuracy of multi-dimensional target parameters. By training a single structural prediction model, the goal of training multiple dimensional parameters separately can be achieved, with almost no change in training time. Compared to building multiple models and training them separately, this method significantly saves training time and computational resources. Attached Figure Description

[0044] Figure 1 This is a schematic diagram of a grating structure provided in an exemplary embodiment.

[0045] Figure 2 This is a schematic diagram of another grating structure provided in an exemplary embodiment.

[0046] Figure 3 This is a schematic diagram of another semiconductor structure under test provided in an exemplary embodiment.

[0047] Figure 4 This is a flowchart of a training method for a semiconductor structure prediction model provided in an exemplary embodiment.

[0048] Figure 5 This is a flowchart of another method for training a semiconductor structure prediction model, provided in an exemplary embodiment.

[0049] Figure 6 This is a schematic diagram of the structure of a device provided in an exemplary embodiment.

[0050] Figure 7 This is a block diagram of a training apparatus for a semiconductor structure prediction model, provided in an exemplary embodiment.

[0051] Figure 8 This is a block diagram of an optical critical dimension measuring device provided in an exemplary embodiment. Detailed Implementation

[0052] Optical critical dimension (OCD) is a technique used to measure and control the dimensions of microstructures, and it is widely used in fields such as semiconductor manufacturing. The general principle of OCD measurement technology can be described as follows: First, a theoretical spectral database corresponding to the morphological model of the sample is established. Then, the specific morphological parameters of the sample are estimated by matching the scattering signal (measured spectrum) of the periodic structure of a specific measured region of the sample with the parameters of the theoretical spectral database.

[0053] In establishing a theoretical spectral database, the algorithm used to calculate the theoretical spectral data is generally the rigorous coupled wave analysis (RCWA) algorithm. The RCWA algorithm works by substituting the sample's dielectric function and the Fourier series expansion of the electromagnetic field in the sample region into Maxwell's equations. Using the continuity condition of the electromagnetic field, the electric field in the incident region is solved, thus obtaining the sample's reflection coefficient. This process is repeated for each wavelength point, using the RCWA algorithm to calculate the reflection coefficient of the sample at that corresponding wavelength, thereby obtaining the sample's theoretical spectral data.

[0054] Modern OCD systems typically combine machine learning and / or deep learning methods, training models with large amounts of real-world data to improve the efficiency and accuracy of predicting target parameters of the structure of a sample based on measured spectral data. However, this approach is suitable for situations with a large number of training samples. When the number of training samples is small, overfitting can easily occur, reducing the accuracy of target parameter predictions.

[0055] Based on this, embodiments of the present invention provide a technique for training a model to predict target parameters of semiconductor structures based on measured spectral data, which can solve the problem in the traditional method where the prediction of structural parameters is inaccurate when training a model using machine learning techniques when there are few training samples.

[0056] Specifically, the OCD measurement technology involved in the embodiments of the present invention can be roughly divided into the following aspects:

[0057] 1) Establish a theoretical spectral database:

[0058] A morphological model of the semiconductor structure under test is constructed based on process parameters (such as materials and exposure conditions). This model includes all possible morphologies within the corresponding structural parameter range. Theoretical spectral data corresponding to all possible morphologies are fitted to the model, and a theoretical spectral database is established through the correspondence between "possible morphology" and "theoretical spectrum." The construction process of the theoretical spectral database is as follows: A grating model is pre-constructed based on the process parameters, and the values ​​of several sets of structural parameters are determined. Several structural points are constructed through permutations and combinations, each structural point consisting of one or more structural parameters. Then, the corresponding spectral values ​​are calculated based on each structural point, and the theoretical spectral database is constructed. The quality of the theoretical spectral database construction is a key factor affecting the accuracy of OCD measurements. Therefore, when constructing the theoretical spectral database, it is necessary to accurately construct the structural model of the grating structure and accurately calculate the theoretical spectral values ​​corresponding to each structural point to ensure the reliability and accuracy of the measurement results.

[0059] 2) Obtain the measured spectral values:

[0060] The light signal scattered or diffracted from the sample surface is obtained by an OCD measurement device and converted into spectral information, which is then used to obtain the measurement spectral value of the semiconductor structure under test.

[0061] Specifically, OCD measurement systems typically employ broadband light sources, such as halogen lamps or supercontinuum lasers, covering the ultraviolet to near-infrared range (e.g., 200-1000 nm) to provide rich spectral information. During OCD measurements, the light emitted from the OCD system's light source is polarized into two polarization modes: TE and TM. In the TE mode, the electric field direction is perpendicular to the incident plane, while in the TM mode, the electric field direction is parallel to the incident plane. When polarized light illuminates the surface of the sample, the sample's three-dimensional structure (e.g., lines, holes) acts as a periodic grating, causing the incident light to produce diffraction orders (e.g., 0th order, ±1st order, etc.) at specific angles. The propagation path of light at the sample edges and surface varies depending on structural parameters (e.g., height, linewidth, etc.), thus affecting the phase and intensity distribution of the light. During the measurement process, a detector collects the light signals scattered or diffracted from the sample surface and converts them into spectral information, which is the measured spectral value. In some embodiments, polarized light is also provided to illuminate the sample surface from multiple rotation angles for testing to obtain more comprehensive spectral information.

[0062] 3) Perform spectral matching to obtain the target parameters of the actual structure:

[0063] The measured spectra are compared and matched with the theoretical spectra in the theoretical spectral database. The theoretical spectrum with the highest similarity to the measured spectrum is determined by the matching algorithm (such as genetic algorithm or neural network), and the corresponding target parameters are used as the structural parameters of the grating to be tested.

[0064] OCD measurements typically involve measuring the dimensions of the grating structure to be measured on the wafer. For example... Figure 1 As shown, Figure 1 This is a schematic diagram of a grating structure provided in this specification. The grating structure is composed of multiple periodically arranged semiconductor structures. In one or more embodiments of this specification, the grating structure refers to the cross-sectional shape of the periodically arranged semiconductor structures constituting the grating, i.e. Figure 1 Each semiconductor structure in the diagram has a trapezoidal cross-section. Of course, besides this, the cross-section of a semiconductor structure can also be composed of multiple trapezoids, such as... Figure 2 As shown, Figure 2 This is a schematic diagram of another grating structure provided in this specification. Figure 1 The diagram lists parameters used to describe the cross-sectional characteristics of the semiconductor structure in the grating, such as top width, bottom width, sidewall angle, height, and depth. These are structural parameters used to describe the periodically arranged semiconductor structures that make up the grating. Generally, each semiconductor structure requires at least two sets of structural parameters and / or a set of combinations of operations (e.g., addition, subtraction, multiplication, and division) to describe or determine the target parameters. In rare cases, all possible sets of a single structural parameter may be used as the target parameters to describe or determine a semiconductor structure. These structural parameters are the optical critical dimensions to be measured during OCD measurement. Of course, the specific types and number of structural parameters to be measured depend on the types of structural parameters contained in the structural points stored in the theoretical spectral database used, and can be selected according to actual needs.

[0065] In one embodiment, a grating structure includes three structural parameters: top width, height, and sidewall angle. If a specific grating structure has a top width of 3 nm, a height of 4 nm, and a sidewall angle of 53°, then the structural parameters of a specific structural point are expressed as [3 nm, 4 nm, 53°]. Then, based on the determined structural points, the corresponding spectral values ​​are calculated, and the corresponding spectra are determined based on these spectral values. The spectral values ​​are multidimensional feature values ​​used to characterize the corresponding spectral features; that is, the number of spectral dimensions is often very high, for example, [0.1533 0.1236 0.1766 …… 0.4650 0.5579 0.4802]. This multidimensional value represents a spectral value containing 100 spectral dimensions, with each value representing one spectral dimension. The spectrum corresponding to the structural point can be determined based on the multidimensional values. Finally, a theoretical spectral database is constructed based on each structural point and its corresponding spectrum.

[0066] For further details, please refer to... Figure 3 , Figure 3The illustration shows an example of a minimum-period semiconductor structure under test in a grating structure, comprising two trapezoidal prisms. The upper trapezoidal prism includes structural parameters TCD1, BCD1, HT1, etc., and the lower trapezoidal prism includes structural parameters BCD1, BCD2, HT2, etc. The target parameter Reference A can be... Figure 1 The target parameter Reference B is a set of permutations and combinations of specific structural parameters (such as BCD1, BCD2, HT2, etc.) included in the trapezoidal prism below, similar to {BCD1, BCD2, HT2}, or / and permutations and combinations of operations (e.g., addition, subtraction, multiplication, division, etc.) involving BCD1, BCD2, HT2. Figure 1 The set of permutations and combinations of specific structural parameters (such as TCD1, BCD1, HT1, etc.) included in the trapezoidal prism above, similar to {TCD1, BCD1, HT1}, or / and the set of combined operations (e.g., addition, subtraction, multiplication, division, etc.) of the specific structural parameters TCD1, BCD1, HT1 included in the trapezoidal prism above. The specific settings can be based on business needs, and this manual does not impose any special restrictions on this.

[0067] The target parameters of the semiconductor under test typically include multiple dimensions, and the structural prediction model training methods provided by related technologies struggle to simultaneously achieve the training accuracy for various target parameters across different dimensions. To ensure appropriate accuracy for each dimension, it is usually necessary to train the target parameters for each dimension separately, i.e., constructing multiple structural prediction models, each corresponding to one target parameter dimension, and then training each model separately. However, this approach often consumes several times the training time and computational resources.

[0068] The dimension of the target parameter typically refers to the combination / arrangement of different structural parameters, such as different combinations of at least two different structural parameters in a structural point. For example, if a specific grating structure has a top width of 3nm, a height of 4nm, and a sidewall angle of 53°, then the target parameter can include dimensions such as [3nm, 4nm, 53°], [3nm, 53°], [4nm, 53°], and [3nm, 4nm].

[0069] This specification provides a training method for a structural prediction model, which can achieve the purpose of training multi-dimensional target parameters separately by training a single structural prediction model, with almost no change in training time. Compared with the method of building multiple models and training them separately, this method greatly saves training time and computing resources.

[0070] In this specification, the training method of the structure prediction model can be applied to servers, as well as to terminal devices such as PCs, mobile phones, tablets, and laptops. This specification does not impose any special restrictions on the application of the training method of the semiconductor structure prediction model.

[0071] In one exemplary embodiment, the structure prediction model can be used to predict multiple sets of not entirely identical target parameters of the semiconductor structure under test. That is, the structure prediction model can be used to predict target parameters in multiple dimensions of the semiconductor structure. Each set of target parameters can correspond to a target parameter in one dimension and a set of model weights. Please refer to [reference needed]. Figure 4 The training method for this structural prediction model may include the following steps:

[0072] Step 402: Obtain the original training samples, which include original spectral values ​​and multiple sets of target parameters corresponding to the semiconductor structure. The original spectral values ​​are sample features, and the multiple sets of target parameters are multiple sets of sample labels that are not completely identical.

[0073] In this embodiment, the original training samples may include original spectral values ​​and multiple sets of target parameters corresponding to the semiconductor structure. That is, the original training samples include labels with multiple dimensions.

[0074] In a simpler case, taking the original training samples as examples that include target parameters in two dimensions, these two target parameters can be denoted as Ref1 and Ref2 respectively.

[0075] Step 404: Repeat the following steps until the conditions are met: Use the original training samples as the training samples for the first round of training to train the structure prediction model; After training, determine whether the accuracy of the structure prediction model meets the standard; If the accuracy of the structure prediction model does not meet the standard, determine the prediction effect of each set of target parameters in this round; For each set of target parameters, if the prediction effect in this round is better than the prediction effect in the previous round, update the model weights corresponding to the target parameters; Determine the incremental training samples for the next round of training, the incremental training samples including incremental spectral values ​​and corresponding multiple sets of target parameters; Use the original training samples and the incremental training samples as the training samples for the next round of training to continue training the structure prediction model.

[0076] Based on step 402 above, after obtaining the original training samples, the initial structure prediction model can be trained using the original training samples. After training, it can be determined whether the accuracy of the structure prediction model meets the standard, and if the accuracy of the structure prediction model does not meet the standard, the incremental training samples for the next round of training are determined. Similar to the original training samples, the incremental training samples include incremental spectral values ​​and corresponding multiple sets of target parameters. Then, the original training samples and the incremental training samples are determined as the training samples for the next round of training to continue training the structure prediction model. And after training is completed, it is possible to continue to determine whether the model accuracy meets the standard.

[0077] In determining whether the model accuracy meets the standard, it is possible to first determine whether the model's training score has reached the training score threshold. If the training score has reached the training score threshold, the original test samples can be used to test the structure prediction model to obtain a test score. It is also possible to further determine whether the evaluation score has reached the test score threshold. The processing and implementation of this part will be described in detail in subsequent embodiments.

[0078] In this embodiment, if the accuracy of the structural prediction model does not meet the standard, the prediction effect of each set of target parameters in this round can be determined separately. For example, the prediction effects of target parameters Ref1 and Ref2 can be determined separately. The evaluation method of the prediction effect will be described in detail in subsequent embodiments.

[0079] In this embodiment, for a set of target parameters, if the prediction performance in the current round is better than the prediction performance in the previous round, the model weights corresponding to the target parameters can be updated. For example, assuming that the prediction performance of target parameter Ref1 in the current round is better than the prediction performance in the previous round, the model weights corresponding to target parameter Ref1 can be updated (for example, the weights corresponding to Ref1 can be changed from...). Updated to ).

[0080] The comparison between the prediction effect of the current round and the prediction effect of the previous round can be determined by the test score. If the test score of the structural parameter Ref1 in the current round is higher than the test score of the structural parameter Ref1 in the previous round, it can be determined that the prediction effect of the current round is better than the prediction effect of the previous round.

[0081] In this embodiment, for a set of target parameters, if the prediction performance in the current round is not better than the prediction performance in the previous round, the model weights corresponding to the target parameters are retained, that is, the model weights corresponding to the target parameters are not updated. For example, assuming that the prediction performance of target parameter Ref2 in the current round is not better than the prediction performance in the previous round, the model weights corresponding to the target parameters are retained (for example, the weights corresponding to Ref2 are retained). ).

[0082] The prediction performance in this round can also be compared with that in the previous round by comparing the test scores. If the test score of structural parameter Ref2 in this round is lower than or equal to the test score of structural parameter Ref2 in the previous round, it can be determined that the prediction performance in this round is not better than that in the previous round.

[0083] In this embodiment, the training score or test score of the structural prediction model can be obtained by weighting the prediction performance of each dimension's target parameters. The weights used for weighting can be preset; for example, the same weight can be set for the prediction performance of each dimension's target parameter. Alternatively, different weights can be set for the prediction performance of different dimension's target parameters; specifically, relatively larger weights can be set for important target parameters, and relatively smaller weights for other target parameters, etc. This specification does not impose any special limitations on this.

[0084] Therefore, by adopting the above implementation method, after one round of training of the structural prediction model, the model weights can be updated based on the prediction performance of different dimensional target parameters. If the prediction performance in the current round is better than the previous round, the corresponding model weights can be updated; conversely, if the prediction performance is not better than the previous round, the corresponding model weights can be retained without updating them. This achieves simultaneous optimization of the prediction accuracy of multi-dimensional target parameters. By training a single structural prediction model, the goal of separately training multi-dimensional target parameters can be achieved, with almost no change in training time. Compared to building multiple models and training them separately, this method significantly saves training time and computational resources.

[0085] The training method of semiconductor structure prediction model will be fully introduced below from two aspects: the preprocessing of original samples and the overall training process of the structure prediction model.

[0086] I. Preprocessing of raw samples

[0087] In one exemplary embodiment, before training the semiconductor structure prediction model, raw samples can be obtained, which include spectral values ​​and corresponding target parameters of the semiconductor structure. The spectral values ​​can be sample features, and the target parameters can be corresponding sample labels.

[0088] After obtaining the original samples, the original samples can be divided into training samples (referred to as original training samples) and test samples (referred to as original test samples). Both the original training samples and the original test samples can include original spectral values ​​and corresponding target parameters.

[0089] In one example, the original samples can be divided by clustering the spectral values ​​of the original samples.

[0090] Specifically, the number of first-type cluster centers can be determined first, and the original samples can be clustered using the spectral values ​​of the original samples as clustering objects to obtain the first-type cluster centers.

[0091] The number of class centers of the first type can be preset based on the number of original samples.

[0092] Optionally, due to the high dimensionality of spectral values, the spectral values ​​in the original samples can be reduced in dimensionality before clustering. Then, the reduced spectral values ​​can be clustered to obtain the first type of class centers. In this example, dimensionality reduction of the spectral values ​​effectively reduces their dimensionality, thereby lowering the cost of subsequent calculations and improving computational efficiency.

[0093] Next, the original sample with the shortest distance to the class center of the first type is determined as the original training sample.

[0094] Specifically, for each first type class center, the original sample that is closest to the first type class center can be determined as the original training sample.

[0095] For example, if clustering yields 9 first-type cluster centers, for each first-type cluster center, an original sample with the shortest distance to the first-type cluster center can be found in the original samples. This original sample with the shortest distance is then determined as the original training sample, resulting in 9 original training samples.

[0096] Then, the remaining original samples were determined as the original test samples.

[0097] As described above, this embodiment uses a clustering algorithm to cluster the spectral values ​​in the original samples and selects the original sample closest to the center of the first type as the original training sample. This ensures that the resulting original training samples are representative and relatively dispersed, guaranteeing a balanced distribution of the original training samples. This eliminates the need for manual selection of representative training data, improving the efficiency of sample data partitioning. Furthermore, using the remaining original samples as original test samples improves the testing performance of the test set, effectively ensuring the accuracy of the subsequent structure prediction model.

[0098] In another example, the original samples can be divided by clustering the target parameters of the original samples.

[0099] Specifically, the number of second-type cluster centers can be determined first, and the original samples can be clustered using the target parameters of the original samples as the clustering objects to obtain the second-type cluster centers.

[0100] The number of class centers of the second type can also be preset based on the number of original samples.

[0101] Next, the original sample with the shortest distance to the center of the second type of class in the original samples is determined as the original training sample.

[0102] Specifically, for each second-type class center, the original sample that is closest to the second-type class center in the original samples can be determined as the original training sample.

[0103] For example, if clustering yields 9 second-type cluster centers, for each second-type cluster center, an original sample with the shortest distance to the second-type cluster center can be found in the original samples. This original sample with the shortest distance is then determined as the original training sample, resulting in 9 original training samples.

[0104] Then, the remaining original samples were determined as the original test samples.

[0105] When the original sample corresponds to multiple sets of semiconductor structural parameters, one set of structural parameters can be used as the clustering object for clustering. This specification does not impose any special restrictions on this.

[0106] As described above, this embodiment uses a clustering algorithm to cluster the structural parameters in the original samples and selects the original sample closest to the center of the second type as the original training sample. This ensures that the resulting original training samples are representative and relatively dispersed, guaranteeing a balanced distribution of the original training samples. This eliminates the need for manual selection of representative training data, improving the efficiency of sample data partitioning. Furthermore, using the remaining original samples as original test samples improves the testing performance of the test set, effectively ensuring the accuracy of the subsequent structural prediction model.

[0107] In one exemplary embodiment, the original samples may be preprocessed for optimization before training the structure prediction model using the original samples, in order to optimize the internal training time of the subsequent model.

[0108] Taking the original training samples as an example, the process of preprocessing the original training samples to optimize the subsequent model training time can begin by determining the number of samples and the feature dimension of the original spectral values. Then, the relationship between the number of samples and the feature dimension is compared. If the number of samples is greater than the feature dimension, the eigenvalue method can be used to solve for the original spectral values ​​to obtain the model input corresponding to the original spectral values.

[0109] When the number of samples exceeds the feature dimension, it indicates a large number of original training samples. The eigenvalue method can be used to solve for the original spectral value matrix composed of the original sample data, yielding the corresponding eigenvalues ​​and eigenvectors. In this matrix, each row represents the original spectral value corresponding to one original training sample, and each column represents the original spectral value of one dimension.

[0110] In this embodiment, the score matrix obtained by the eigenvalue calculation method can be used as the model input corresponding to the original training sample. In the subsequent model training process, it can replace the original spectral values ​​in the original training sample with the input of the structure prediction model to train the structure prediction model, realize feature dimensionality reduction, thereby optimizing the training time of the structure prediction model and improving training efficiency.

[0111] When the number of samples is less than or equal to the feature dimension, the original spectral values ​​can be decomposed into low-dimensional features using matrix factorization as the model input corresponding to the original spectral values.

[0112] If the number of samples is less than or equal to the feature dimension, it indicates that the number of original training samples is small. Matrix factorization can be used to reduce the dimension of the original training samples, thereby optimizing the training time of the structure prediction model and improving training efficiency.

[0113] For example, the original spectral value matrix can be decomposed into the product of two low-rank matrices using Singular Value Decomposition (SVD) to obtain a score matrix. This score matrix can then be used to replace the original spectral values ​​from the original training samples and input into the structure prediction model for training. When using SVD, the left orthogonal matrix can be omitted, further saving computation time and memory. In this example, the computational complexity of the low-dimensional matrix is ​​much smaller than that of the original high-dimensional matrix (i.e., the original spectral value matrix), thus reducing computational complexity.

[0114] II. Overall Training Process of Structural Prediction Model

[0115] In one exemplary embodiment, please refer to Figure 5 The training process for a semiconductor structure prediction model may include the following steps:

[0116] Step 502: Use the original training samples as the training samples for the first round of training to train the structure prediction model.

[0117] In this embodiment, the structure prediction model can be a neural network model, etc. This specification does not impose any special restrictions on the structure prediction model.

[0118] Step 504: After training is completed, calculate the training score of the structure prediction model.

[0119] If the original training samples are used to train the structure prediction model during the first round of training, and the structure prediction model has converged, then the training can be considered complete. Next, a training score for the structure prediction model can be calculated to evaluate the effectiveness of the first round of training.

[0120] If, during subsequent training iterations, the structure prediction model is trained using the original training samples and corresponding incremental training samples, and the structure prediction model has converged, then training can be considered complete. Next, a training score for the structure prediction model can be calculated to evaluate the model's performance in this training iteration.

[0121] In one example, training samples (including original training samples and incremental training samples) can be input into a pre-trained structure prediction model to obtain the prediction results (i.e., the predicted target parameters) output by the structure prediction model. Then, the deviation between the predicted target parameters and the sample labels (i.e., the target parameters corresponding to the training samples) can be calculated. The number of first samples whose deviation does not exceed a preset deviation threshold can then be counted, and the model's training score can be determined based on this number of first samples. The training score is positively correlated with the number of first samples; that is, the more first samples whose deviation does not exceed the deviation threshold, the higher the model's training score; conversely, the fewer first samples whose deviation does not exceed the deviation threshold, the lower the model's training score.

[0122] If only this method of calculating bias is used to calculate the training score, it can be directly determined that the training score has reached the training score threshold when the number of the first sample reaches more than 80% of the number of training samples, and it can be directly determined that the training score has not reached the training score threshold when the number of the first sample is less than 80% of the number of training samples.

[0123] In another example, Design of Experiments (DoE) can be used to analyze the differences between different wafers, thereby determining the training score of the model.

[0124] Specifically, the target parameter mean values ​​of each wafer can first be sorted based on sample labels to obtain a label sequence. The sample labels are the structural parameters of the training samples. Then, the label sequence and the predicted sequence can be compared item by item. If the two sequences are completely identical, it indicates that the model's predicted values ​​are completely consistent with the actual values, the model's prediction performance is good, and the training score can be determined as the value that has reached the training score threshold. If the two sequences are not completely identical, it indicates that there is a difference between the model's predicted values ​​and the actual values, the model's prediction performance is poor, and the training score can be determined as the value that has not reached the training score threshold.

[0125] In another example, the training score of the model can be calculated by comparing the difference between the maximum and minimum values ​​of the target parameters for each wafer, as well as the range of fluctuation between the model's predicted values ​​and the sample labels.

[0126] Specifically, the label range of the target parameters can be calculated first based on the sample labels, and a preset floating parameter can be obtained. The corresponding prediction range range is then determined based on the label range range and the floating parameter. The floating parameter can be a percentage or a fixed value. Next, for the training samples used for training, it can be determined whether the predicted parameters of the structure prediction model are within the aforementioned prediction range range. If the predicted structure parameters are within the prediction range range, the prediction for that sample is considered satisfactory.

[0127] Then, the number of samples within the predicted range for the predicted parameters is used as a second sample size, which reflects the predictive performance of the structural prediction model. Furthermore, the training score or the test score is determined based on the second sample size, and the training score or the test score is positively correlated with the second sample size. That is, the larger the second sample size, the higher the training score.

[0128] In another example, the training score can be calculated based on the degree of linear correlation between the sample labels and the predicted parameters. Generally, the closer this linear correlation is to 1, the higher the training score.

[0129] In another example, the training score can be calculated based on the slope of the linear regression line between the sample labels and the prediction parameters. Generally speaking, the closer the slope of the linear regression line is to 1, the higher the training score.

[0130] In another example, the training score can be calculated using the Gauge Repeatability and Reproducibility (GRR) method. If the 3xSigma (standard deviation) does not exceed a preset threshold, it indicates that the model has good repeatability and stability, and the training score is acceptable. If the 3xSigma (standard deviation) exceeds the preset threshold, it indicates that the model has high variability, poor repeatability and stability, and the training score is unacceptable.

[0131] It should be noted that in practical applications, one or more of the above calculation methods can be used to calculate the training score of the structural prediction model. If multiple calculation methods are used, the training scores obtained by each method can be weighted to determine the final training score. Of course, other methods can also be used to determine the training score, and this specification does not impose any special restrictions on this.

[0132] Step 506: If the training score reaches the training score threshold, the original test sample is used to test the structure prediction model to obtain a test score.

[0133] Based on the calculation result of the training score in step 504 above, it can be determined whether the training score has reached a preset training score threshold. If the training score has not reached the training score threshold, step 512 can be executed. If the training score has reached the training score threshold, it indicates that the training effect of the structure prediction model is good, and the original test samples can be used to test the structure prediction model and calculate the test score.

[0134] The calculation process for the test score can refer to the aforementioned calculation process for the training score. For example, the test score can be calculated by calculating the bias, or by designing an experiment with DoE, or by combining multiple calculation methods. These will not be elaborated on in this specification.

[0135] Step 508: If the test score does not reach the test score threshold, determine the incremental training samples.

[0136] Based on the calculation result of the test score in step 506 above, it can be determined whether the test score has reached the preset test score threshold. If the test score has not reached the test score threshold, it indicates that the structure prediction model is performing poorly in the test set, and incremental training samples can be determined to continue training the structure prediction model.

[0137] Optionally, a threshold for the number of incremental iterations can be set for incremental training. If the test score still does not reach the threshold when the threshold is reached, training can be terminated, and the model with the best training performance among the multiple training iterations can be selected as the output. For example, the model with the highest test score among the multiple training iterations can be selected as the output of the trained structure prediction model.

[0138] For example, if the test score does not reach the test score threshold, it can be first determined whether the number of incremental training iterations has reached the incremental iteration threshold. If not, the step of determining incremental training samples can be executed. If the incremental iteration threshold has been reached, training can be terminated, and the model with the highest test score obtained during training can be selected as the trained result prediction model and output to the user.

[0139] Step 510: The original training samples and the incremental training samples are determined as the training samples to be used in the next round of training to continue training the structure prediction model. After training is completed, return to step 504 to calculate the training score.

[0140] In this embodiment, if the test score of the structure prediction model does not reach the test score threshold, it indicates that the accuracy of the structure prediction model is not up to standard. The original training samples and the determined incremental training samples can be used to continue training the structure prediction model. After training is completed, the process can return to step 504 to calculate the model training score in order to evaluate the training effect of the model.

[0141] The process of determining incremental training samples will be described in detail later.

[0142] Step 512: If the training score does not reach the training score threshold, adjust the regularization factor and continue training the structure prediction model. After training is completed, return to step 504 to calculate the training score.

[0143] In this embodiment, if the training score of the structure prediction model does not reach the training score threshold, the regularization factor can be adjusted to continue training the structure prediction model. For example, the regularization factor can be increased when the structure prediction model is overfitting, and decreased when the structure prediction model is underfitting.

[0144] In this embodiment, after the structure prediction model has been retrained, step 504 can be executed to calculate the training score. If the number of adjustments to the regularization factor reaches the adjustment threshold, and the training score still does not reach the training score threshold, then model training can be terminated.

[0145] Step 514: If the test score reaches the test score threshold, determine that the model accuracy meets the standard and end the model training.

[0146] In this embodiment, when the test score of the structure prediction model reaches the test score threshold, it can be determined that the accuracy of the structure prediction model meets the standard, and the model training can be terminated. Subsequently, the structure prediction model with the qualified accuracy can be used to predict semiconductor target parameters.

[0147] Specifically, the spectral values ​​of the semiconductor under test can be obtained first, referred to as the measured spectral values. These spectral values ​​are then input into a structure prediction model with sufficient accuracy to obtain the target parameter prediction results of the semiconductor structure under test, output by the structure prediction model. After obtaining the target parameter prediction results, a determined correction function can be used to correct these prediction results, thereby obtaining the target parameters of the semiconductor under test.

[0148] As can be seen from the above description, the structural prediction model training method provided in this manual allows for the calculation of the model's training score after training is complete. Furthermore, once the training score meets the required standard, the model can be tested using the original test samples, and a test score can be calculated. If the test score also meets the required standard, then the accuracy of the structural prediction model can be determined. This automated method of determining whether the model's accuracy meets the standard improves the efficiency of model accuracy assessment.

[0149] At the same time, by combining training scores and test scores, it can be ensured that the structure prediction model not only performs well on training data, but also maintains good performance on unseen test data. This dual evaluation mechanism can improve the generalization ability of the structure prediction model and make it more reliable in practical applications.

[0150] Furthermore, when calculating model training and testing scores, various calculation methods such as bias calculation and experimental design are used to comprehensively evaluate the performance of the structural prediction model and improve the accuracy of model performance evaluation results.

[0151] In one exemplary embodiment, a regression method can be used to find the standard parameters corresponding to the original spectral values ​​in the original training samples in the optical critical size model library. Each set of original spectral values ​​can correspond to a set of standard parameters. The optical critical size model library can be pre-constructed, storing spectral values ​​and their corresponding structural parameters. For ease of distinction, the spectral values ​​stored in the optical critical size model library can be referred to as standard spectral values ​​and standard structural parameters.

[0152] Next, the parameter boundaries can be determined based on the standard parameters.

[0153] After identifying the standard parameters, the parameter boundaries of the structure can be determined based on these parameters. Generally, the semiconductor structure corresponding to the parameters within these boundaries will be similar to the semiconductor structure under test, and can be used to construct incremental training samples.

[0154] In practice, the range of the structural parameter boundaries is related to the fluctuation range of the structural parameter. This fluctuation range is generally closely related to specific process and material parameters. Specifically, the critical value of a structural parameter can be determined based on expert experience, historical data thresholds, semiconductor structure design values, and specific semiconductor process parameters, thus determining the fluctuation range of that structural parameter. In most cases, the value is less than or equal to the fluctuation range of the structural parameter; in a few cases, actual process and measurement results in a small number of samples exceeding the fluctuation range. In this embodiment, the range constraining the structural parameter boundaries is less than or equal to the fluctuation range of the structural parameter.

[0155] Then, several parameters can be determined as incremental parameters within the range of the parameter boundaries.

[0156] For example, the target parameters in the original sample data are obtained from real measurement data and are usually accurate structural parameters. However, the structural parameters and spectral values ​​in the optical critical dimension model library are obtained through theoretical calculations or model inference using methods such as the RCWA algorithm, grid point construction, or pre-trained model construction (e.g., neural network models). Therefore, the corresponding standard parameters found in the optical critical dimension model library using the original spectral values ​​may contain errors. Consequently, a correction function can be determined based on the target parameters and the corresponding standard parameters. Then, the randomly determined parameters are corrected based on the correction function to obtain more accurate incremental parameters.

[0157] Specifically, for ease of distinction, several parameters determined within the parameter boundary range can be called candidate parameters. Then, a correction function is used to correct each candidate parameter, and the resulting correction result is called the incremental parameter, which is used to construct incremental training samples.

[0158] This embodiment can determine the correction function based on the target parameters and standard parameters. Subsequently, the correction function is used to correct the candidate parameters to obtain incremental parameters, which can improve the accuracy and reliability of the incremental parameters, thereby improving the accuracy of the subsequent structural prediction model.

[0159] Furthermore, after determining the incremental parameters, the spectral values ​​corresponding to each incremental structural parameter can be determined as the corresponding incremental spectral values, thereby constructing incremental training samples.

[0160] Specifically, the incremental spectral value corresponding to the incremental parameter can be determined using the following method:

[0161] In this embodiment, the theoretical spectral value corresponding to the incremental parameter can be determined first.

[0162] In one example, the spectral values ​​of several adjacent parameters (referred to as library spectral values) of the incremental parameter can be searched in the optical critical dimension model library. By searching for adjacent parameters, library spectral values ​​that are close to the incremental parameter can be obtained. Then, interpolation calculations can be performed on the incremental parameter based on the library spectral values ​​to obtain its corresponding theoretical spectral value. This example constructs a theoretical spectral value for the new parameter by searching for library spectral values ​​of adjacent parameters and performing interpolation calculations, thereby effectively utilizing existing spectral data, reducing reliance on experimental measurements, and improving the calculation efficiency and accuracy of spectral values.

[0163] In another example, the incremental parameter can also be used as model input, and a trained spectral value prediction model can be used to predict the theoretical spectral value corresponding to the incremental parameter. The spectral value prediction model can be a neural network model or other model structure, which can be trained based on the parameters and their corresponding spectral values, and used to predict spectral values ​​based on the parameters.

[0164] In another example, the Rigorous Coupled-Wave Analysis (RCWA) algorithm can be used to calculate the theoretical spectral values ​​corresponding to the incremental parameters. Specifically, a geometric model of the periodic structure can be established based on the incremental parameters, and the theoretical spectral values ​​corresponding to the incremental parameters can be obtained by solving the partial differential equations.

[0165] Then, the incremental spectral values ​​can be obtained by interpolating between the wavelength points corresponding to the original spectral values ​​based on the theoretical spectral values.

[0166] Specifically, interpolation calculations can be performed between wavelength points corresponding to the original spectral values ​​to obtain incremental spectral values. This interpolation is performed in the wavelength dimension and is used to convert the theoretical spectral values ​​into incremental spectral values ​​with the same wavelength points as the original spectral values.

[0167] When the original sample corresponds to multiple sets of target parameters of the semiconductor structure, incremental training samples can be constructed separately for each set of target parameters when determining the incremental training samples. This specification does not impose any special restrictions on this.

[0168] It can be seen that this embodiment can construct incremental training samples using the original training samples without experimental measurement, which can improve the efficiency and accuracy of incremental training sample construction.

[0169] Figure 6 This is a schematic structural diagram of a device provided in an exemplary embodiment. Please refer to... Figure 6 At the hardware level, the device includes a processor 602, an internal bus 604, a network interface 606, memory 608, and non-volatile memory 610, and may also include other hardware required for its functions. One or more embodiments of this specification can be implemented in software, such as the processor 602 reading the corresponding computer program from the non-volatile memory 610 into memory 608 and then running it. Of course, in addition to software implementation, one or more embodiments of this specification do not exclude other implementation methods, such as logic devices or a combination of hardware and software, etc. That is to say, the execution subject of the following processing flow is not limited to each logic unit, but can also be hardware or logic devices.

[0170] Please refer to Figure 7 The training device 700 for semiconductor structure prediction models can be applied to, for example... Figure 6The device shown is used to implement the technical solution of this specification. The training device 700 for the semiconductor structure prediction model may include:

[0171] The sample acquisition unit 701 acquires original training samples, which include original spectral values ​​and multiple sets of target parameters corresponding to semiconductor structures. The original spectral values ​​are sample features, and the multiple sets of target parameters are corresponding multiple sets of not completely identical sample labels.

[0172] Model training unit 702 performs the following steps repeatedly until the conditions are met:

[0173] The original training samples are used as the training samples for the first round of training to train the structure prediction model;

[0174] After training is completed, determine whether the accuracy of the structure prediction model meets the standard.

[0175] If the accuracy of the structural prediction model is not up to standard, determine the prediction effect of each set of target parameters in this round.

[0176] For each set of target parameters, if the prediction effect in this round is better than the prediction effect in the previous round, the model weights corresponding to the target parameters are updated.

[0177] Determine the incremental training samples to be used in the next round of training. The incremental training samples include incremental spectral values ​​and corresponding multiple sets of incremental structural parameters.

[0178] The original training samples and the incremental training samples are selected as the training samples for the next round of training to continue training the structure prediction model.

[0179] Optionally, if the model training unit 702 does not achieve a better prediction result in the current round than in the previous round, it retains the model weights corresponding to the target parameters.

[0180] Optionally, the process of determining whether the accuracy of the structural prediction model meets the standard includes:

[0181] Calculate the training score of the structure prediction model;

[0182] When the training score reaches the training score threshold, the structure prediction model is tested using test samples to obtain a test score.

[0183] If the test score reaches the test score threshold, the accuracy of the structural prediction model is determined to be up to standard.

[0184] If the test score does not reach the test score threshold, it is determined that the accuracy of the structural prediction model is substandard.

[0185] Optionally, the training score or the test score is obtained by weighting the prediction performance of each group of target parameters.

[0186] Optionally, the conditions include: the accuracy of the structural prediction model meets the standard or reaches a preset number of training rounds.

[0187] Please refer to Figure 8 The optical critical dimension measuring device 800 can also be applied to, for example... Figure 5 The device shown is used to implement the technical solution of this specification. The optical critical dimension measuring device 800 may include:

[0188] The spectral value acquisition unit 801 acquires the measured spectral values ​​of the semiconductor under test;

[0189] The structure prediction unit 802 inputs the measured spectral values ​​into a structure prediction model with sufficient accuracy to obtain the target parameter prediction results of the semiconductor structure under test output by the structure prediction model.

[0190] The structural prediction model is trained using the method described above in this specification.

[0191] Based on the same concept as the methods described above, this specification also provides an electronic device, including: a processor; a memory for storing processor-executable instructions; wherein the processor performs the steps of the method as described in any of the above embodiments by executing the executable instructions.

[0192] Based on the same concept as the methods described above, this specification also provides a computer-readable storage medium having computer instructions stored thereon that, when executed by a processor, implement the steps of the methods as described in any of the above embodiments.

[0193] Based on the same concept as the methods described above, this specification also provides a computer program product, including a computer program / instructions that, when executed by a processor, implement the steps of the methods as described in any of the above embodiments.

Claims

1. A training method for a semiconductor structure prediction model, characterized in that, The structure prediction model is used to predict target parameters of multiple sets of not entirely identical semiconductor structures. Each set of target parameters corresponds to a dimension, and each dimension represents the combination or permutation of different structural parameters. Each set of target parameters corresponds to a set of model weights. The method includes: Obtain original training samples, which include original spectral values ​​and multiple sets of target parameters corresponding to semiconductor structures, wherein the original spectral values ​​are sample features and the multiple sets of target parameters are corresponding multiple sets of sample labels that are not completely identical; Repeat the following steps until the condition is met: The original training samples are used as the training samples for the first round of training to train the structure prediction model; After training is completed, determine whether the accuracy of the structure prediction model meets the standard. If the accuracy of the structural prediction model is not up to standard, determine the prediction effect of each set of target parameters in this round. For each set of target parameters, if the prediction effect in this round is better than the prediction effect in the previous round, the model weights corresponding to the target parameters are updated. Determine the incremental training samples to be used in the next round of training. The incremental training samples include incremental spectral values ​​and corresponding sets of target parameters. The original training samples and the incremental training samples are selected as the training samples for the next round of training to continue training the structure prediction model.

2. The method according to claim 1, characterized in that, Also includes: If the prediction result in this round is not better than the prediction result in the previous round, the model weights corresponding to the target parameters are retained.

3. The method according to claim 1, characterized in that, The process of determining whether the accuracy of the structural prediction model meets the standard includes: Calculate the training score of the structure prediction model; When the training score reaches the training score threshold, the structure prediction model is tested using test samples to obtain a test score. If the test score reaches the test score threshold, the accuracy of the structural prediction model is determined to be up to standard. If the test score does not reach the test score threshold, it is determined that the accuracy of the structural prediction model is substandard.

4. The method according to claim 3, characterized in that, The training score or the test score is obtained by weighting the prediction performance of each group of target parameters.

5. The method according to claim 1, characterized in that, The conditions include: the accuracy of the structural prediction model meets the standard or reaches the preset number of training rounds.

6. A method for measuring critical optical dimensions, characterized in that, The method includes: Obtain the measurement spectral values ​​of the semiconductor under test; The measured spectral values ​​are input into a structure prediction model with sufficient accuracy to obtain the target parameter prediction results of the semiconductor structure under test output by the structure prediction model. The structural prediction model is trained using the method described in any one of claims 1-5.

7. A training device for a semiconductor structure prediction model, characterized in that, The structure prediction model is used to predict target parameters of multiple sets of not entirely identical semiconductor structures. Each set of target parameters corresponds to a dimension, and each dimension represents the combination or permutation of different structural parameters. Each set of target parameters corresponds to a set of model weights. The device includes: The sample acquisition unit acquires original training samples, which include original spectral values ​​and multiple sets of target parameters corresponding to semiconductor structures. The original spectral values ​​are sample features, and the multiple sets of target parameters are corresponding multiple sets of not completely identical sample labels. The model training unit repeatedly executes the following steps until the conditions are met: The original training samples are used as the training samples for the first round of training to train the structure prediction model; After training is completed, determine whether the accuracy of the structure prediction model meets the standard. If the accuracy of the structural prediction model is not up to standard, determine the prediction effect of each set of target parameters in this round. For each set of target parameters, if the prediction effect in this round is better than the prediction effect in the previous round, the model weights corresponding to the target parameters are updated. Determine the incremental training samples to be used in the next round of training. The incremental training samples include incremental spectral values ​​and corresponding sets of target parameters. The original training samples and the incremental training samples are selected as the training samples for the next round of training to continue training the structure prediction model.

8. An optical critical dimension measuring device, characterized in that, The device includes: The spectral value acquisition unit acquires the measured spectral values ​​of the semiconductor under test. The structure prediction unit inputs the measured spectral values ​​into a structure prediction model with sufficient accuracy to obtain the target parameter prediction results of the semiconductor structure under test output by the structure prediction model. The structural prediction model is trained using the method described in any one of claims 1-5.

9. An electronic device, characterized in that, include: processor; A memory for storing processor-executable instructions; wherein the processor implements the steps of the method as described in any one of claims 1-6 by executing the executable instructions.

10. A computer-readable storage medium, characterized in that, It stores computer instructions that, when executed by a processor, implement the steps of the method as described in any one of claims 1-6.

11. A computer program product, characterized in that, Includes a computer program / instructions that, when executed by a processor, implement the steps of the method as described in any one of claims 1-6.

Images