Feature size parameter extraction method and system, terminal equipment and storage medium

Abstract

The invention discloses a feature size parameter extraction method and system, terminal equipment and a storage medium. The method comprises the following steps: providing a plurality of feature sizes of a to-be-detected feature and simulation spectrums corresponding to the feature sizes; the feature size and the simulation spectrum serve as a data set of a neural network model, the neural network model is trained until the training error of the neural network model is smaller than a preset threshold value, and it is determined that training of the neural network model is completed; measuring a measurement spectrum of the to-be-measured feature by using measurement equipment; predicting the feature size of the to-be-measured feature according to the measurement spectrum and the trained neural network model to obtain a predicted size range of the feature size; and fitting the measured spectrum with the corresponding simulation spectrum in the predicted size range to obtain a simulation spectrum which is optimally fitted with the measured spectrum, and extracting a theoretical feature size corresponding to the simulation spectrum which is optimally fitted as a final feature size. According to the invention, the measurement of the to-be-measured feature size can be rapidly realized, and the requirement of online measurement is met.

Description

Technical Field

[0001] The present invention relates to the technical field of feature size detection, and in particular to a feature size parameter extraction method, system, terminal equipment and storage medium based on a neural network. Background Art

[0002] In the existing semiconductor manufacturing field, it is necessary to measure the characteristic dimensions of the nanostructures in the semiconductor. For example, in the field of semiconductor nanostructure characteristic dimension measurement based on optical scattering measurement, ellipsometer is a common characteristic dimension measurement device. The process of inverting and extracting the parameters of the nanostructure to be measured from the spectrum measured by the ellipsometer is a typical inverse problem solving process. The solution of the inverse problem in optical scattering measurement is one of the key technologies. The traditional methods for solving inverse problems are mainly nonlinear fitting and library matching methods.

[0003] Generally, it is necessary to use numerical solution algorithms (such as complex simulation algorithms such as rigorous coupled wave analysis (RCWA)) to construct an optical property model for simulation. After obtaining the simulated spectrum, it is fitted with the measured spectrum. Each fitting requires the optical property model to be simulated once until the best fitting result is obtained. The characteristic size is then obtained based on the fitting result, which takes a long time and requires a lot of calculation, and cannot achieve online measurement of the characteristic size. Summary of the invention

[0004] The main technical problem solved by the present invention is that the existing measurement method cannot realize online measurement of characteristic dimensions.

[0005] According to the first aspect, an embodiment provides a method for extracting characteristic size parameters based on a neural network, comprising:

[0006] Providing multiple characteristic dimensions of the characteristic to be measured and the simulation spectrum corresponding to the characteristic dimensions;

[0007] The characteristic size and the simulated spectrum are used as data sets of the neural network model to train the neural network model until the training error of the neural network model is less than a preset threshold, and it is determined that the training of the neural network model is completed;

[0008] Measuring a measurement spectrum of the characteristic to be measured using a measurement device;

[0009] Predicting the characteristic size of the feature to be measured based on the measured spectrum and the trained neural network model to obtain a predicted size range of the characteristic size;

[0010] The measured spectrum is fitted with the corresponding simulation spectrum within the predicted size range to obtain the simulation spectrum that best fits the measured spectrum, and the theoretical characteristic size corresponding to the best fitting simulation spectrum is extracted as the final characteristic size.

[0011] According to the second aspect, an embodiment provides a neural network-based feature size parameter extraction system, comprising:

[0012] A spectrum library module, used to provide multiple characteristic sizes of the characteristics to be measured and simulated spectra corresponding to the characteristic sizes;

[0013] A measuring device, used for measuring the spectrum of the characteristic to be measured to obtain a measured spectrum;

[0014] A neural network module is used to predict the characteristic size of the feature to be measured based on the measured spectrum and the trained neural network model to obtain a predicted size range of the characteristic size;

[0015] The fitting module fits the measured spectrum with the corresponding simulation spectrum within the predicted size range to obtain the simulation spectrum that best fits the measured spectrum, and extracts the theoretical characteristic size corresponding to the best fitting simulation spectrum as the final characteristic size.

[0016] According to the third aspect, an embodiment provides a terminal device, including:

[0017] Memory, used to store programs;

[0018] A processor is used to implement the method described in the first aspect by executing a program stored in a memory.

[0019] According to a fourth aspect, an embodiment provides a computer-readable storage medium, on which a program is stored, and the program can be executed by a processor to implement the method described in the first aspect.

[0020] According to the characteristic size parameter extraction method, system, terminal device and storage medium of the above-mentioned embodiment, a spectrum library is first established using the preset characteristic size and the corresponding simulation spectrum, and then the neural network model is trained using the data of the spectrum library. Finally, the characteristic size is predicted by the trained neural network model, and the actual measured spectrum is fitted to obtain the simulation spectrum that best fits the measured spectrum, and the theoretical characteristic size corresponding to the best fitting simulation spectrum is extracted as the final characteristic size. Therefore, the use of the neural network model for prediction and fitting can quickly realize the measurement of the characteristic size to be measured, meeting the needs of online measurement. BRIEF DESCRIPTION OF THE DRAWINGS

[0021] Figure 1 A flowchart of a method for extracting characteristic size parameters provided in one embodiment of the present application (I);

[0022] Figure 2 Flow chart (II) of a characteristic size parameter extraction method provided for one embodiment of the present application;

[0023] Figure 3 A schematic diagram of the structure of a characteristic size parameter extraction system provided in one embodiment of the present application.

[0024] Reference numerals: 10 - spectral library module; 20 - measurement device; 30 - neural network module; 40 - fitting module. DETAILED DESCRIPTION

[0025] The present invention is further described in detail below by specific embodiments in conjunction with the accompanying drawings. Wherein similar elements in different embodiments adopt associated similar element numbers. In the following embodiments, many detailed descriptions are for making the present application better understood. However, those skilled in the art can easily recognize that some features can be omitted in different situations, or can be replaced by other elements, materials, methods. In some cases, some operations related to the present application are not shown or described in the specification, this is to avoid the core part of the present application being overwhelmed by too much description, and for those skilled in the art, it is not necessary to describe these related operations in detail, and they can fully understand the related operations according to the description in the specification and the general technical knowledge in the art.

[0026] In addition, the features, operations or characteristics described in the specification can be combined in any appropriate manner to form various implementations. At the same time, the steps or actions in the method description can also be interchanged or adjusted in a manner that is obvious to those skilled in the art. Therefore, the various sequences in the specification and the drawings are only for the purpose of clearly describing a certain embodiment and are not meant to be a required sequence, unless otherwise specified that a certain sequence must be followed.

[0027] The serial numbers of the components in this document, such as "first", "second", etc., are only used to distinguish the objects described and do not have any order or technical meaning. The "connection" and "coupling" mentioned in this application, unless otherwise specified, include direct and indirect connections (couplings).

[0028] In the field of detection, the workpiece to be tested needs to be inspected in multiple processing steps. Among them, the detection of the characteristic size of the nanostructure (i.e., the feature to be tested) in the workpiece to be tested is a necessary detection process to ensure the processing accuracy, and the detection can be performed after each processing step.

[0029] In semiconductor manufacturing, the part to be tested generally refers to the wafer. In this case, the feature to be tested can refer to any structure in the processed wafer pattern that needs to be detected. The feature to be tested is not limited to a specific size. Since the current semiconductor processing accuracy is already at 3nm (such as TSMC technology), and the conventional processing accuracy is at the nanometer or micron level, the feature to be tested in the wafer is generally called a nanostructure, or micro-nanostructure.

[0030] At the same time, the DUT may also refer to a display panel, such as a display panel based on micro-LED technology, in which case the nanostructure is a display unit inside the display panel or a structure inside the display unit. Of course, the DUT may also be other products that need to be tested.

[0031] Therefore, the present application does not limit the specific forms of the test piece and the nanostructure, and may also be products such as a glass panel and a mask.

[0032] In the existing detection methods, in order to calculate the spectral response and performance of nanostructures, complex simulation algorithms such as the finite element method (FEM), finite difference time domain method (FDTD) and rigorous coupled wave analysis (RCWA) can be used to simulate the spectrum. These optical characteristic model algorithms require a long calculation time.

[0033] Among them, RCWA is an algorithm based on electromagnetic field theory, which can accurately calculate the diffraction efficiency of periodic grating structures. When extracting the parameters to be measured of the nanostructure, the optical characteristic model established by RCWA can be used to calculate the simulated spectrum of the nanostructure first; then the measured spectrum and the simulated spectrum are fitted in real time by methods such as the Levenberg-Marquarelt (LM) nonlinear regression algorithm, so as to obtain the measured parameter value corresponding to the best fitting simulated spectrum.

[0034] However, since fitting algorithms such as the LM algorithm need to call optical characteristic models such as RCWA to calculate the simulated spectrum of the nanostructure in each iteration process, calculations based on the optical characteristic models require a lot of time, making its parameter extraction process very time-consuming and difficult to meet the needs of online measurement.

[0035] In order to solve the problem of slow real-time fitting speed, the industry has proposed a library matching method to extract the parameters to be measured. This method first uses the established optical property model to pre-establish a spectral library of the characteristic dimensions of the features to be measured, and then matches the measured spectrum with the simulated spectrum in the spectral library to find the simulated spectrum that best matches the measured spectrum. The parameter value corresponding to the simulated spectrum is the parameter value to be measured for the nanostructure. However, when there are many parameters to be measured for the nanostructure, the amount of library data will increase dramatically, occupying a large amount of storage space, and the time spent on library matching search will also increase, which is also difficult to meet the needs of online measurement.

[0036] The embodiment of the present application proposes a characteristic size parameter extraction method based on a neural network, which can quickly obtain the characteristic size and its corresponding ideal simulation spectrum through the neural network model, and can meet the high-precision and high-yield requirements of online measurement. The method can quickly and effectively process a large amount of data and has good generalization performance.

[0037] like Figure 1 As shown, the embodiment of the present application provides a method for extracting characteristic size parameters based on a neural network, which may include:

[0038] Step 1: Provide multiple feature sizes of the features to be tested and the simulated spectra corresponding to the feature sizes. You can use the data of a matched spectral library or use an optical property model for one-to-one matching to build a spectral library that matches the current device to be tested and the measuring equipment. It is not difficult to understand that if the wafer to be tested is a batch product, one-to-one matching is required when the spectral library is first built, but subsequent tests can continue to use the matched spectral library.

[0039] In some embodiments, the above step 1 may include:

[0040] Step 110: Determine the measurement parameters of the measuring device according to the theoretical range of the characteristic size of the feature to be measured; the characteristic size may include at least two parameters to be measured.

[0041] For example, the characteristic size may include at least two parameters to be measured among the diameter (or width), length and thickness of the characteristic to be measured, that is, there are multiple characteristic sizes; the spectral feature may include at least one of the reflectivity, ellipsometric parameters, Mueller matrix, central wavelength, peak amplitude, half-height width, Bayesian peak area and wavelength shift. According to the processing accuracy of the product and the designed process, the theoretical range of any characteristic size of the characteristic to be measured can be known, and the appropriate measuring equipment and measuring parameters can be selected according to this theoretical range.

[0042] For another example, the measuring device may be an ellipsometer or some measuring device based on optical scattering measurement, such as an ultraviolet-visible absorption spectrometer, a laser Raman spectrometer, a Fourier transform infrared spectrometer, or a radial diffractometer. The measuring parameters of the measuring device may include a measuring wavelength, an incident angle, and an azimuth angle, which are used to provide parameters for simulation use of the optical characteristic model.

[0043] Step 120, construct an optical characteristic model according to the measurement parameters, and use the optical characteristic model to simulate and obtain corresponding simulated spectra according to the characteristic sizes within the theoretical range; wherein each set of characteristic sizes corresponds to a simulated spectrum, and all simulated spectra constitute a spectrum library.

[0044] In some embodiments, in the above step 120, constructing the optical characteristic model according to the measurement parameters may include:

[0045] The optical characteristic model is constructed according to the measured parameters and by using a numerical solution algorithm. In this step, how to construct the optical characteristic model can refer to existing implementation methods, and this application does not limit it.

[0046] For example, the numerical solution algorithm may be a rigorous coupled wave analysis algorithm, a finite element method, a finite difference time domain method, or a finite difference frequency domain method.

[0047] In the simulation calculation of nanostructure spectra, there are many numerical simulation methods to choose from, such as RCWA, FEM, FDTD, FDFD and so on.

[0048] Among them, RCWA (Rigorous Coupled Wave Analysis) is a rigorous coupled wave analysis method based on Fourier. It is a very commonly used method for calculating the reflection and transmission spectra of periodic optical structures. This method can simulate nanostructures with complex morphology and periodicity, and is widely used in semiconductor devices, solar cells, nano-biosensors and other fields.

[0049] FEM (Finite Element Method) and FDTD (Finite Difference Time Domain Method) are numerical solutions to electromagnetic wave equations and are both used in solving spectral problems. Generally speaking, the FEM method is more suitable for structures with complex shapes, while the FDTD method is more suitable for time domain problems.

[0050] Compared with the RCWA method, the FEM and FDTD methods usually have higher requirements for computing resources, so when dealing with large or complex structures, the running time and amount of calculation may be more. In addition, the selection of a specific numerical simulation method also needs to take into account aspects such as the accuracy and stability of the algorithm. Therefore, choosing an appropriate numerical simulation method requires considering many factors, including model accuracy, solution time, computing resources, and so on.

[0051] In the above step 120, an optical characteristic model is constructed by using a suitable numerical solution algorithm, and the characteristic to be measured that needs to be measured is simulated, so that an accurate simulation spectrum can be obtained.

[0052] For example, the theoretical range of a parameter to be measured in the characteristic size is 9nm-15nm, and the data in this range is simulated to obtain the corresponding spectrum library.

[0053] Step 2: Use the characteristic size and the simulated spectrum as the data set of the neural network model, and train the neural network model until the training error of the neural network model is less than a preset threshold, and determine that the neural network model has completed training.

[0054] In some embodiments, the neural network model may be a feedforward neural network model, a recursive neural network model, or a convolutional neural network model. The loss function of the neural network model may be a mean square error loss function, a mean absolute error loss function, or a divergence loss function.

[0055] Among them, Feedforward Neural Networks (FNN): FNN is one of the most basic neural network models, which includes an input layer, several hidden layers and an output layer. FNN can automatically identify the nonlinear mapping relationship between data and make predictions by learning a given set of sample data. The advantage of FNN is that it can be trained and optimized end-to-end, which is convenient for batch processing and prediction.

[0056] Recurrent Neural Networks (RNN): RNN is mainly used to process data with sequence properties, such as time series data or text data. The neurons in RNN not only receive the input of the current moment, but also receive the output of the previous moment, so that the previous state information can be retained. In the task of semiconductor nanostructure spectrum prediction, RNN variant models such as LSTM (Long Short-Term Memory) can be used.

[0057] Convolutional Neural Networks (CNN): CNN is mainly used to process data with spatial structure, such as images, sounds, etc. CNN can automatically extract features of different scales through convolution and pooling operations, and use the fully connected layer to make the final prediction. In the task of semiconductor nanostructure spectrum prediction, the spectrum data can be regarded as a two-dimensional image, and convolutional neural networks can be used for feature extraction and prediction.

[0058] When training a neural network, it is necessary to select a suitable loss function to measure the difference between the predicted results and the actual results, so as to perform the back propagation algorithm to optimize the weight parameters and deviations of the neural network. For the nanostructure characterization problem, the following loss function can be used to train the neural network.

[0059] Among them, Mean Squared Error (MSE): This loss function is often used in regression problems to calculate the square error between the predicted value and the actual value. For this problem, the output value of the neural network is compared with the actual spectral reflectance calculated by the RCWA model to evaluate the error between the prediction and the actual.

[0060] Mean Absolute Error (MAE): This loss function is also commonly used in regression problems to calculate the absolute error between the predicted value and the actual value. Compared with MSE, MAE pays more attention to the prediction accuracy of small errors and has less weight than large errors.

[0061] KL divergence loss function (Kullback-Leibler Divergence, KL): This loss function is often used in classification problems and probability distribution fitting problems to measure the distance and difference between two probability distributions. For this problem, KL divergence can measure the difference between the predicted spectral distribution and the actual distribution, and can be used to optimize the prediction accuracy of the neural network.

[0062] Therefore, the embodiments of the present application do not limit the specific neural network model and the loss function used for training it, and can be adjusted according to the actual training results.

[0063] In some embodiments, the simulated spectrum may include spectral features in multiple dimensions; in the above step 2, training the neural network model may include:

[0064] Step 210: Perform dimensionality reduction processing on multiple spectral features to screen out spectral features whose eigenvalues ​​are greater than a preset condition.

[0065] In some embodiments, in the above step 210, performing dimensionality reduction processing on the multiple spectral features may include:

[0066] Step 211: Use principal component analysis or linear discriminant analysis to perform dimensionality reduction processing on multiple spectral features.

[0067] Dimensionality reduction can effectively improve the calculation speed, while reducing the dimension and complexity of the data, making it easier for subsequent data analysis and processing.

[0068] In the spectral characterization of semiconductor nanostructures, dimensionality reduction methods can include principal component analysis (PCA) and linear discriminant analysis (LDA). These methods can linearly combine multiple highly correlated measurement parameters to obtain a set of uncorrelated new variables, called principal components. By selecting an appropriate number of principal components, the dimensionality of the original data can be reduced to a lower dimension to achieve the effect of reducing the amount of computation and optimizing training.

[0069] When using neural networks for spectral prediction, the principal components after dimensionality reduction can be used as input features to achieve more efficient calculations. At the same time, the use of dimensionality reduction methods can also reduce the interference of noise and redundant information and improve the accuracy and reliability of measurements.

[0070] In some embodiments, in the above step 210, screening out spectral features whose characteristic values ​​are greater than a preset condition may include:

[0071] Step 212: When the principal component analysis method is used, the spectral features whose variance resolution is greater than the first preset threshold are selected; when the linear discriminant analysis method is used, the spectral features whose data distribution discrimination is greater than the second preset threshold are selected. The preset threshold is adjusted according to the actual calculation situation and the training result of the neural network model.

[0072] Determining the feature with the largest eigenvalue can be achieved through dimensionality reduction methods such as principal component analysis (PCA) or linear discriminant analysis (LDA). In these methods, the eigenvalue represents the variance of the principal component and is an indicator for evaluating the contribution of each principal component to the explanation of the original data. The principal component with a larger eigenvalue contributes more to the explanation of the data.

[0073] In PCA, methods such as singular value decomposition (SVD) or eigenvector decomposition (EVD) can be used to calculate the principal components and corresponding eigenvalues ​​of the data. During the calculation process, the obtained eigenvalues ​​can be sorted by size, and the principal component with the largest eigenvalue can be selected as the representation of the main information of the data. It can also be used to calculate the eigenvectors of the covariance matrix. The covariance matrix can describe the relationship between different variables and reflect the variance and correlation of the data. The eigenvectors and eigenvalues ​​of the covariance matrix are the key to determining the principal components. The eigenvalue represents the variance of each principal component, and the eigenvector represents the direction of the corresponding principal component. By calculating the eigenvalues ​​and eigenvectors of the covariance matrix, the principal components of the data and the corresponding variance contribution rates can be obtained. In the process of principal component analysis, the principal components with higher variance contribution rates are usually selected to describe the main information of the data.

[0074] In LDA, the between-class scattering matrix and the within-class scattering matrix can be used to calculate the linear discriminant direction and the corresponding eigenvalue. The within-class scattering matrix measures the difference between classes within the data, and the between-class scattering matrix measures the difference between classes between the data. By calculating the ratio of the within-class scattering matrix to the within-class scattering matrix, the eigenvalue of the linear discriminant analysis can be obtained. The linear discriminant direction with the largest eigenvalue corresponds to the highest data distribution discrimination.

[0075] Step 3: Use a measuring device to measure the measurement spectrum of the feature to be measured.

[0076] For example, in an embodiment of the present application, an ellipsometer may be used for measurement to obtain an actual measured spectrum for subsequent fitting.

[0077] Step 4: predict the characteristic size of the feature to be measured based on the measured spectrum and the trained neural network model to obtain a predicted size range of the characteristic size.

[0078] For example, based on the trained neural network model, the output can be predicted to obtain the predicted range of the output. For example, the characteristic size of the feature to be measured is its thickness, and the spectral feature is the reflectivity. Assuming that the input (characteristic size) of the neural network model is 10nm-12nm, the reflectivity range corresponding to the spectral feature of the theoretical simulation spectrum is 1.1-1.3, then when the reflectivity of the measured spectrum is 1.2, it can be predicted that the predicted size range of the characteristic size falls within 10nm-12nm. In other words, the predicted size range of the characteristic size corresponding to the measured spectrum can be obtained through the neural network model.

[0079] Step 5: Fit the measured spectrum to the corresponding simulation spectrum within the predicted size range to obtain the simulation spectrum that best fits the measured spectrum, and extract the theoretical characteristic size corresponding to the best fitting simulation spectrum as the final characteristic size.

[0080] In the above step 5, based on the predicted size range of the characteristic size, a characteristic size can be input into the neural network model starting from the middle value or the maximum value of this range to obtain a corresponding simulation spectrum, and the measured spectrum is fit with the simulation spectrum, and the input amount is adjusted according to the measurement results until the best fitting simulation spectrum is obtained.

[0081] For example, when the predicted size range of the characteristic size is 10nm-12nm, 11nm is input into the neural network model to obtain the simulation spectrum corresponding to 11nm, and the input amount is adjusted according to the fitting result. It can be adjusted in the form of 11±0.1nm until the input is adjusted to the best fit.

[0082] In some embodiments, in step 5 above, fitting the measured spectrum with the corresponding simulated spectrum within the predicted size range may include:

[0083] A curve fitting algorithm is used to fit the measured spectrum to the corresponding simulated spectrum within the predicted size range.

[0084] For example, the curve fitting algorithm may be a nonlinear regression algorithm, a least squares algorithm, a polynomial fitting algorithm, or a kernel function regression algorithm.

[0085] For example, in the nonlinear regression algorithm (LM algorithm), during the fitting process, because the LM algorithm is an iterative optimization algorithm, each iteration needs to calculate the prediction results of the model according to the current parameters and evaluate the loss function, and then update the parameters. The calculation of the RCWA optical model is relatively time-consuming, so the LM algorithm needs to call RCWA to calculate the simulated spectrum in each iteration, which may result in a high calculation cost in practical applications. In the embodiment of the present application, in each iteration of the LM algorithm, the neural network model is called for calculation, and the calculation speed is fast, which can meet the requirements of online measurement.

[0086] In summary, the method provided by the embodiment of the present application can solve the problems of slow speed, complex calculation, and large storage space occupied in the inverse problem solving process of characteristic size measurement of nanostructures. The present application proposes a characteristic size parameter extraction method based on a neural network, which can quickly and accurately extract nanostructure parameters to meet the high precision and high yield requirements of online measurement. Compared with traditional nonlinear fitting and library matching methods, the method proposed in the present application has significant advantages in real-time fitting speed and computational efficiency. At the same time, the method can process a large amount of data and has good generalization performance. Therefore, the present application has significant beneficial effects in improving the characteristic size measurement accuracy and efficiency of nanostructures, while also reducing the large amount of storage space required for the library matching method.

[0087] like Figure 2 As shown, a specific embodiment is provided below for further explanation.

[0088] The first step is to determine the value range of each parameter to be measured in the characteristic size of the feature to be measured, determine the value range of the measurement parameters such as the measurement wavelength, incident angle, azimuth angle, etc. of the measuring equipment, use RCWA to establish an optical characteristic model to calculate the simulated spectrum of each combination of the characteristic size of the feature to be measured, and establish a spectral library corresponding to the characteristic size of the feature to be measured. The simulated spectrum can be specifically represented by reflectivity, ellipsometric parameters, and Mueller matrix.

[0089] In the second step, the established spectral library is used as the data set of the neural network. According to the characteristic size parameters of the features to be measured and the characteristic distribution of the simulated spectrum, a fully connected neural network model that can map nonlinear functions is established, and the mean square error (MSE) is used as the loss function for training.

[0090] Before training, the data set needs to be preprocessed, and data cleaning and normalization are performed to remove any outliers and ensure that the data is within a consistent range. Due to the large dimensionality of spectral features, principal component analysis (PCA) can be used to transform the data set to compress the spectral features to a lower feature space without losing too much information. By calculating the covariance matrix of the spectral matrix, and then obtaining the eigenvalue eigenvector of the covariance matrix, the matrix consisting of the eigenvectors corresponding to the k features with the largest eigenvalue (i.e., the largest variance) is selected. The spectral matrix is ​​converted to a new space to achieve dimensionality reduction of the spectral features. The number of spectral data features is determined by the variance interpretation. Generally, k features with a variance interpretation of (1-10-7)*100% are selected to prevent large data errors after dimensionality reduction.

[0091] By adopting the exhaustive grid search method, the hyperparameters such as the number of neurons, activation function and optimizer are finely adjusted, and the MSE of different neural network models are compared to optimize the model performance. When the data set is small or the prediction error is large, the feature size parameter extraction method based on the neural network proposed in this application can also be trained and predicted using deep learning network models such as CNN (convolutional neural network) or RNN (recurrent neural network) to obtain the preset size range of the feature size corresponding to the current measured spectrum. These deep learning network models have the ability to process complex data and can better capture the relationship between the measured parameters and the spectral data, thereby improving the accuracy and reliability of the model. For example, using the CNN model, local features in the spectral data can be gradually extracted through the convolution layer and the pooling layer, and these features are combined in the fully connected layer for tasks such as classification or regression. Using the RNN model, time series information can be taken into account, and the spectral data can be modeled using a cyclic structure, retaining the correlation between the data, thereby better realizing the prediction of the spectral data.

[0092] The third step is to first use the trained neural network model to predict the spectrum of the characteristic size of the feature to be measured, and after obtaining the prediction result, use it as the spectral data of the LM algorithm iteration for nonlinear fitting to extract the parameters to be measured. The LM algorithm is a commonly used nonlinear least squares optimization algorithm with high accuracy and robustness when estimating parameters. The LM algorithm is used as the core algorithm for inverse problem solving to quickly and accurately extract the parameters to be measured. When the parameters to be measured of the nanostructure are more, the traditional library matching method needs to establish a large amount of spectral data, which takes up a lot of storage space, and the search time will also increase, which is difficult to meet the needs of online measurement. In contrast, the neural network model occupies less memory, and the speed of using the neural network to predict a set of spectra is faster, which can usually be completed in about 1ms. Therefore, in the characteristic size parameter extraction method proposed in this application, the method of using the neural network prediction can significantly improve the calculation efficiency and accuracy, thereby meeting the requirements of online measurement.

[0093] In summary, through the above method, when using scattering measurement equipment such as ellipsometers to detect semiconductor nanostructures, a more accurate simulation spectrum and fitting structure can be obtained by selecting appropriate optical feature models and fitting algorithms. For example, the RCWA algorithm is used to establish a spectral library of the nanostructure to be measured as a data set for the neural network; a neural network model is constructed with the mapping input as the characteristic size of the feature to be measured and the output as the theoretical simulation spectrum, and a nonlinear relationship between the nanostructure parameters and the spectral data is established. The spectral features are compressed using a dimensionality reduction method based on principal component analysis, and the parameters are tuned and the neural network is quickly trained. The trained neural network model is used to predict the simulation spectrum of the parameter to be measured, and the spectrum as the iterative spectrum of the LM algorithm is nonlinearly fitted, and the best fitting simulation spectrum is used to extract the parameter to be measured. This method can achieve fast and accurate measurement and meet the high precision and high yield requirements of online measurement.

[0094] For example, the RCWA algorithm is very computationally intensive, especially for larger structures or high-resolution calculations, which results in very high computation time and memory requirements, making it difficult to implement online calculations. Therefore, by using a trained neural network model to replace the RCWA algorithm, such as the CNN+LM algorithm, the computation time and memory requirements can be greatly reduced, enabling faster online calculations.

[0095] The above is a description of the characteristic size parameter extraction method based on a neural network provided in an embodiment of the present application. The following is a description of a characteristic size parameter extraction system based on a neural network provided in an embodiment of the present application.

[0096] like Figure 3 As shown, a neural network-based feature size parameter extraction system provided in an embodiment of the present application is used to implement the method described in the above embodiment. The system may include:

[0097] The spectrum library module 10 is used to provide a plurality of characteristic sizes of the features to be measured and simulated spectra corresponding to the characteristic sizes.

[0098] The measuring device 20 is used to measure the spectrum of the feature to be measured to obtain a measured spectrum.

[0099] The neural network module 30 is used to predict the characteristic size of the feature to be measured based on the measured spectrum and the trained neural network model to obtain a predicted size range of the characteristic size.

[0100] The fitting module 40 fits the measured spectrum with the corresponding simulation spectrum within the predicted size range to obtain a simulation spectrum that best fits the measured spectrum, and extracts a theoretical characteristic size corresponding to the best fitting simulation spectrum as the final characteristic size.

[0101] The spectrum library module 10, the neural network module and the fitting module 40 can be implemented by one or more processors. The measuring device can be an ellipsometer or some measuring devices based on optical scattering measurement, such as a UV-visible absorption spectrometer, a laser Raman spectrometer, a Fourier transform infrared spectrometer, or a radial diffractometer.

[0102] In the above system, the neural network module 30 is also used to use the characteristic size and the simulated spectrum as the data set of the neural network model to train the neural network model until the training error of the neural network model is less than a preset threshold, and it is determined that the neural network model has completed training.

[0103] The above system is used to implement the above characteristic size parameter extraction method, which has corresponding technical effects and will not be described again here.

[0104] The characteristic size parameter extraction method provided in the present application is implemented by a terminal device, which may include a memory and a processor. For example, the terminal device may be a computer, a server, or other device with computing and data processing capabilities.

[0105] A memory is used to store a program. A processor is used to implement the feature size parameter extraction method described in the above embodiment by executing the program stored in the memory.

[0106] Those skilled in the art will appreciate that all or part of the functions of the various methods in the above-mentioned embodiments can be implemented by hardware or by computer programs. When all or part of the functions in the above-mentioned embodiments are implemented by computer programs, the program can be stored in a computer-readable storage medium, and the storage medium can include: read-only memory, random access memory, disk, optical disk, hard disk, etc., and the program is executed by a computer to implement the above-mentioned functions. For example, the program is stored in the memory of the device, and when the program in the memory is executed by the processor, all or part of the above-mentioned functions can be implemented. In addition, when all or part of the functions in the above-mentioned embodiments are implemented by computer programs, the program can also be stored in a storage medium such as a server, another computer, disk, optical disk, flash disk or mobile hard disk, and can be downloaded or copied and saved in the memory of the local device, or the system of the local device is updated, and when the program in the memory is executed by the processor, all or part of the functions in the above-mentioned embodiments can be implemented.

[0107] This document is described with reference to various exemplary embodiments. However, those skilled in the art will recognize that changes and modifications may be made to the exemplary embodiments without departing from the scope of this document. For example, various operating steps and components for performing the operating steps may be implemented in different ways (e.g., one or more steps may be deleted, modified, or incorporated into other steps) depending on the specific application or considering any number of cost functions associated with the operation of the system.

[0108] Although the principles of this invention have been shown in various embodiments, many modifications of structures, arrangements, proportions, elements, materials and components particularly suitable for specific environments and operational requirements can be used without departing from the principles and scope of this invention. The above modifications and other changes or amendments will be included in the scope of this invention.

[0109] The foregoing specific description has been described with reference to various embodiments. However, those skilled in the art will recognize that various modifications and changes can be made without departing from the scope of the present disclosure. Therefore, the consideration of the present disclosure will be illustrative rather than restrictive, and all these modifications will be included in its scope. Similarly, the advantages, other advantages and solutions to the problems of various embodiments have been described above. However, the benefits, advantages, solutions to the problems and any elements that can produce these, or make them more clear, should not be interpreted as critical, necessary or necessary. The term "include" and any other variants used in this article are all non-exclusive inclusions, so that the process, method, article or device including the list of elements not only includes these elements, but also includes other elements that are not explicitly listed or do not belong to the process, method, system, article or device. In addition, the term "coupled" and any other variants used in this article refer to physical connections, electrical connections, magnetic connections, optical connections, communication connections, functional connections and / or any other connections.

[0110] Those skilled in the art will appreciate that many changes may be made to the details of the above-described embodiments without departing from the basic principles of the invention. Therefore, the scope of the present invention should be determined solely by the claims.

Claims

1. A method for extracting characteristic size parameters based on neural network, characterized in that: include: Providing multiple characteristic dimensions of the characteristic to be measured and the simulation spectrum corresponding to the characteristic dimensions; Using the characteristic size and the simulated spectrum as data sets of a neural network model, training the neural network model until a training error of the neural network model is less than a preset threshold, and determining that the training of the neural network model is completed; Measuring the measurement spectrum of the feature to be measured using a measuring device; Predicting the characteristic size of the feature to be measured according to the measured spectrum and the trained neural network model to obtain a predicted size range of the characteristic size; The measured spectrum is fitted with the simulation spectrum corresponding to the predicted size range to obtain a simulation spectrum that best fits the measured spectrum, and a theoretical characteristic size corresponding to the simulation spectrum that best fits is extracted as the final characteristic size.

2. The method according to claim 1, characterized in that The providing of a plurality of characteristic sizes of the feature to be measured and the simulated spectra corresponding to the characteristic sizes comprises: Determine the measurement parameters of the measuring device according to the theoretical range of the characteristic size of the feature to be measured; the characteristic size includes at least two parameters to be measured; An optical characteristic model is constructed according to the measurement parameters, and the corresponding simulated spectrum is obtained by simulating using the optical characteristic model according to the characteristic size within the theoretical range; wherein each group of the characteristic sizes corresponds to a simulated spectrum, and all the simulated spectra constitute a spectrum library.

3. The method according to claim 2, characterized in that The step of constructing an optical characteristic model according to the measurement parameters comprises: An optical characteristic model is constructed according to the measurement parameters and using a numerical solution algorithm.

4. The method according to claim 3, characterized in that The numerical solution algorithm is a rigorous coupled wave analysis algorithm, a finite element method, a finite difference time domain method or a finite difference frequency domain method.

5. The method according to claim 1, characterized in that The fitting of the measured spectrum to the corresponding simulated spectrum within the predicted size range comprises: A curve fitting algorithm is used to fit the measured spectrum to the corresponding simulated spectrum within the predicted size range.

6. The method according to claim 5, characterized in that The curve fitting algorithm is a nonlinear regression algorithm, a least squares algorithm, a polynomial fitting algorithm or a kernel function regression algorithm.

7. The method according to claim 1, characterized in that The neural network model is a feedforward neural network model, a recursive neural network model or a convolutional neural network model.

8. The method according to claim 7, characterized in that The loss function of the neural network model is a mean square error loss function, a mean absolute error loss function or a divergence loss function.

9. The method according to claim 1, characterized in that The simulated spectrum includes spectral features in multiple dimensions; The training of the neural network model comprises: A dimension reduction process is performed on the plurality of spectral features to screen out the spectral features whose characteristic values ​​are greater than a preset condition.

10. The method according to claim 9, characterized in that The performing dimensionality reduction processing on the plurality of spectral features comprises: The principal component analysis method or the linear discriminant analysis method is used to perform dimension reduction processing on the plurality of spectral features.

11. The method according to claim 10, characterized in that The step of screening out the spectral features whose characteristic values ​​are greater than a preset condition comprises: When the principal component analysis method is used, the spectral features having a variance resolution greater than a first preset threshold are selected; When the linear discriminant analysis method is adopted, the spectral feature whose data distribution discrimination is greater than the second preset threshold is selected.

12. The method according to claim 9, characterized in that The characteristic size includes at least two of the diameter, length and thickness of the feature to be measured; The spectral characteristics include at least one of reflectivity, ellipsometric parameters, Mueller matrix, central wavelength, peak amplitude, half-width, Bayesian peak area and wavelength shift.

13. A characteristic size parameter extraction system based on neural network, characterized in that: include: A spectrum library module, used to provide multiple characteristic sizes of the characteristics to be measured and simulated spectra corresponding to the characteristic sizes; A measuring device, used to measure the spectrum of the feature to be measured to obtain a measured spectrum; A neural network module is used to predict the characteristic size of the feature to be measured according to the measured spectrum and the trained neural network model to obtain a predicted size range of the characteristic size; The fitting module fits the measured spectrum with the corresponding simulated spectrum within the predicted size range to obtain a simulated spectrum that best fits the measured spectrum, and extracts a theoretical characteristic size corresponding to the best-fitting simulated spectrum as a final characteristic size.

14. A terminal device, characterized in that: include: Memory, used to store programs; A processor, configured to implement the method according to any one of claims 1 to 12 by executing the program stored in the memory.

15. A computer-readable storage medium, characterized in that: The medium stores a program, which can be executed by a processor to implement the method according to any one of claims 1 to 12.

Images