Method for training neural network model and ellipsometry method

Abstract

The invention discloses a neural network model training method and an ellipsometry method, and relates to the technical field of ellipsometry, the method comprises the following steps: obtaining a reference spectrum and a to-be-calibrated spectrum corresponding to the reference spectrum, the reference spectrum being obtained by measuring a sample through a reference ellipsometer, and the to-be-calibrated spectrum being obtained by measuring the sample through a to-be-calibrated ellipsometer; the to-be-calibrated spectrum and the simulation spectrum serve as training input, the to-be-calibrated spectrum and the simulation spectrum serve as training labels, a neural network model is trained, a loss function of the neural network model is converged, the trained neural network model is obtained, the loss function comprises a first loss item and a second loss item, and the first loss item and the second loss item correspond to each other; the first loss item is obtained based on the difference between the training output of the spectrum to be calibrated and the corresponding reference spectrum, and the second loss item is obtained based on the difference between the training output of the simulation spectrum and the simulation spectrum. The model can adjust the measurement spectrum of the ellipsometer to be calibrated so as to eliminate the measurement difference of the ellipsometers to be calibrated.

Description

A method for training a neural network model and an ellipsometry measurement method Technical Field

[0001] This invention relates to the field of ellipticity measurement, and more particularly to a method for training a neural network model and an ellipticity measurement method. Background Technology

[0002] Ellipsometry is an instrument that uses the optical properties of materials to measure optical constants, key optical dimensions of micro and nanostructures, and thin film thickness. The measurement process of ellipsometer does not require direct contact with the sample, will not damage the material surface, and does not require a vacuum environment. It is widely used in fields such as semiconductor measurement.

[0003] Current ellipsometers suffer from the following problems: slight differences in the internal optical components, assembly processes, and measurement microenvironments of different ellipsometers lead to variations in measurement results. To improve measurement consistency among ellipsometers, existing technologies employ measures such as maximizing the standardization of internal optical component selection, standardizing assembly processes, and establishing a unified measurement microenvironment to maximize hardware consistency. However, the demanding component selection, assembly processes, and environmental requirements necessitate substantial human and material resources, increasing the manufacturing cost of the ellipsometers.

[0004] In addition, existing technologies also employ methods such as fine-tuning spectral weights and loosening some fixed parameters to make the measured spectra obtained by each ellipsometer more consistent. However, these methods require a large amount of manual intervention in parameter tuning, and generally speaking, the generalization performance of algorithms with well-tuned parameters is poor. Summary of the Invention

[0005] This invention provides a method for training a neural network model and an ellipsometric measurement method. The aim is to train a neural network model that adjusts the measurement spectrum of the ellipsometer to be calibrated, using the measurement spectrum of the reference ellipsometer as a reference, to maximize the elimination of measurement differences among the ellipsometers to be calibrated, reduce the differences in measurement results among the ellipsometers to be calibrated, and improve the consistency of sample detection results, thereby solving the shortcomings of high cost and poor generalization performance of the existing technology.

[0006] According to one aspect of the present invention, a method for training a neural network model is provided, the method comprising:

[0007] Obtain the reference spectrum and the spectrum to be calibrated corresponding to the reference spectrum. The reference spectrum is obtained by measuring the sample with a reference ellipsometer, and the spectrum to be calibrated is obtained by measuring the sample with an ellipsometer to be calibrated.

[0008] Obtain the simulated spectrum of the sample;

[0009] Using the spectrum to be calibrated and the simulated spectrum as training inputs, and the spectrum to be calibrated and the simulated spectrum themselves as training labels, the neural network model is trained so that the loss function of the neural network model converges, resulting in a trained neural network model.

[0010] The loss function includes a first loss term and a second loss term. The first loss term is derived based on the difference between the training output of the spectrum to be calibrated and the corresponding reference spectrum, while the second loss term is derived based on the difference between the training output of the simulated spectrum and the simulated spectrum itself.

[0011] According to another aspect of the present invention, an ellipticity measurement method is provided, the method comprising:

[0012] The sample was measured using an ellipsometer to be calibrated, and the measured spectrum was obtained.

[0013] By inputting the measured spectrum into any trained neural network model in this invention, a predicted spectrum is obtained;

[0014] The theoretical spectrum of the sample is obtained, and the measured values ​​of the sample's parameters are obtained by fitting the predicted spectrum with the theoretical spectrum.

[0015] According to another aspect of the present invention, an ellipticity measuring device is provided for implementing the ellipticity measuring method of the present invention. The device includes:

[0016] The measurement module is used to acquire the measurement spectrum;

[0017] The processing module is used to input the measured spectrum into any trained neural network model in this invention to obtain the predicted spectrum;

[0018] The parameter acquisition module is used to fit the predicted spectrum and the theoretical spectrum of the sample to obtain the measured values ​​of the parameters to be measured in the sample.

[0019] The significant advantages of this invention are: by adjusting the measurement spectra obtained from each ellipsometer based on a trained neural network model, the measurement spectra of each ellipsometer are made more consistent, thus ensuring the measurement consistency of each ellipsometer. Compared with existing technologies, this invention has the advantages of low cost and good generalization performance. Furthermore, by introducing simulated spectra as training data, the amount of data in the training set is expanded, avoiding underfitting due to insufficient data during model training; simultaneously, it overcomes the problem of high cost in obtaining real samples.

[0020] It should be understood that the description in this section is not intended to identify key or essential features of the embodiments of the present invention, nor is it intended to limit the scope of the invention. Other features of the invention will become readily apparent from the following description. Attached Figure Description

[0021] To more clearly illustrate the technical solutions in the embodiments of the present invention, the accompanying drawings used in the description of the embodiments will be briefly introduced below. Obviously, the accompanying drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0022] Figure 1 is a flowchart illustrating a method for training a neural network model provided by the present invention;

[0023] Figure 2 is a schematic diagram of the relationship between a value point and a sampling space provided by the present invention;

[0024] Figure 3 is a schematic diagram of the first type of spectral relationship provided by the present invention;

[0025] Figure 4 is a schematic diagram of the architecture of the first neural network model provided by the present invention;

[0026] Figure 5 is a schematic diagram of the architecture of the second neural network model provided by the present invention;

[0027] Figure 6 is a schematic diagram of the variation of a loss function provided by the present invention;

[0028] Figure 7 is a schematic diagram of model evaluation information provided by the present invention;

[0029] Figure 8 is a schematic diagram of the second type of spectral relationship provided by the present invention;

[0030] Figure 9 is a flowchart illustrating an ellipticity measurement method provided by the present invention;

[0031] Figure 10 is a schematic diagram of an elliptic measurement process provided by the present invention;

[0032] Figure 11 is a schematic diagram of the structure of an elliptic measurement device provided by the present invention. Detailed Implementation

[0033] To enable those skilled in the art to better understand the present invention, the technical solutions of the present invention will be clearly and completely described below with reference to the accompanying drawings of the embodiments. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort should fall within the scope of protection of the present invention.

[0034] It should be noted that the terms "first," "second," etc., in the specification, claims, and accompanying drawings of this invention are used to distinguish similar objects and are not necessarily used to describe a specific order or sequence. It should be understood that such data can be interchanged where appropriate so that the embodiments of the invention described herein can be implemented in orders other than those illustrated or described herein. Furthermore, the terms "comprising" and "having," and any variations thereof, are intended to cover a non-exclusive inclusion; for example, a process, method, system, product, or apparatus that comprises a series of steps or units is not necessarily limited to those steps or units explicitly listed, but may include other steps or units not explicitly listed or inherent to such processes, methods, products, or apparatus.

[0035] Figure 1 is a flowchart illustrating a method for training a neural network model according to the present invention. This method is applicable to training a neural network model for adjusting the measurement spectra of ellipsometers to be calibrated, so that the measurement spectra of each ellipsometer approach a reference spectrum. This method can be executed by the apparatus for training neural network models provided by the present invention. This apparatus can be implemented in hardware and / or software. In a specific embodiment, the apparatus can be integrated into an electronic device. The following embodiments will illustrate this method using the integration of the apparatus into an electronic device as an example. Referring to Figure 1, the method specifically includes the following steps:

[0036] S101. Obtain the reference spectrum and the spectrum to be calibrated corresponding to the reference spectrum. The reference spectrum is obtained by measuring the sample with a reference ellipsometer, and the spectrum to be calibrated is obtained by measuring the sample with a ellipsometer to be calibrated.

[0037] The reference spectrum and the corresponding spectrum to be calibrated are obtained by measuring the same micro / nano structure of the sample using a reference ellipsometer and a ellipsometer to be calibrated, respectively. Based on the above method, this invention can use a reference ellipsometer and a ellipsometer to be calibrated to measure multiple micro / nano structures of the sample, obtaining multiple reference spectra and corresponding spectra to be calibrated.

[0038] S102. Obtain the simulated spectrum of the sample.

[0039] The simulated spectrum of the sample can be understood as the spectrum simulated based on the structural parameters of the sample. Since the process of obtaining real samples is complex and costly, and training the model requires a large number of samples, the method in this step can quickly obtain a large number of training samples, thereby solving the problem of overfitting of the autoencoder neural network model caused by insufficient real samples.

[0040] The number of simulated spectra is greater than the number of reference spectra and the spectra to be calibrated. In this embodiment, the number of reference spectra and corresponding spectra to be calibrated is 100-1000; in this embodiment, 700 reference spectra and corresponding spectra to be calibrated were actually measured. The number of simulated spectra is 10,000-200,000; in this embodiment, 100,000 simulated spectra were actually acquired.

[0041] In one embodiment, S102 includes: establishing an optical property model of the sample, adjusting the structural parameters in the optical property model, and then performing optical simulation on the optical property model based on simulation algorithms such as RCWA (Rigorous Coupled Wave Analysis) or FDTD (Finite-Difference Time-Domain) to obtain the simulation spectrum corresponding to each set of structural parameters.

[0042] The specific steps for adjusting the structural parameters to obtain the corresponding simulated spectrum are as follows: Based on the number and value range of the structural parameters in the optical property model, a sampling space is constructed. The dimension of the sampling space is the same as the number of structural parameters, and the side length of the sampling space corresponds to the value range of the structural parameters. A preset number of value points are arranged in the sampling space, and the value points are randomly and uniformly distributed in the sampling space. The simulated spectrum corresponding to each value point is obtained.

[0043] Since each value point corresponds to a specific value in each dimension of the sampling space, each value point can correspond to a specific set of structural parameters. These structural parameters are input into the optical property model of the sample, and then optical simulation is performed on the optical property model using simulation algorithms such as RCWA and FDTD to obtain the simulated spectrum corresponding to each set of structural parameters.

[0044] By randomly and uniformly distributing the value points within the sampling space, it is possible to obtain parameter combinations that cover the entire sampling space and avoid the "curse of dimensionality" problem caused by uniform sampling.

[0045] The "curse of dimensionality" refers to the phenomenon where, under uniform sampling, the total number of sampled combinations increases exponentially with the number of structural parameters. For example, assuming there are 10 structural parameters, even if each parameter is uniformly selected with only 3 values, there will still be a total of 3... 10 A combination of sampling methods.

[0046] Figure 2 is a schematic diagram of the relationship between a value point and a sampling space provided by the present invention. The sampling space shown in Figure 2 includes two structural parameter variables. The sampling space is presented as a two-dimensional rectangle. Specifically, S in Figure 2 represents the sampling space. Each circle P in the sampling space S represents a value point. Each value point P located in the two-dimensional rectangle S can correspond to two specific structural parameters. Based on the specific structural parameters, the simulated spectrum corresponding to the value point can be obtained.

[0047] S103. Using the spectrum to be calibrated and the simulated spectrum as training inputs, and the spectrum to be calibrated and the simulated spectrum themselves as training labels, train the neural network model so that the loss function of the neural network model converges, and obtain the trained neural network model.

[0048] The loss function includes a first loss term and a second loss term. The first loss term is derived based on the difference between the training output of the spectrum to be calibrated and the corresponding reference spectrum, while the second loss term is derived based on the difference between the training output of the simulated spectrum and the simulated spectrum itself. Specifically, the difference includes mean squared error or mean absolute error.

[0049] The trained neural network model can take the input spectrum to be calibrated as input and output a predicted spectrum that is close to the simulated spectrum.

[0050] The advantage of this training method lies in using readily available simulated spectra as the basic training dataset, reducing the model training cost. Furthermore, since the neural network model used for training in this embodiment has five layers—an input layer, three hidden layers (the hidden layers are layers between the input and output layers), and an output layer—with both the input and output layers having 1500 dimensions and millions of weights, if a large amount of simulated spectra is not used in model training, and the model is trained using only a small number of real spectra, the model is prone to overfitting. Therefore, by introducing simulated spectra, the amount of training data can be expanded, avoiding the overfitting problem caused by insufficient data.

[0051] The reason why the simulated spectrum's own output and input are compared in this loss function is that when the selection of internal optical components of the ellipsometer is standardized to the maximum extent, the assembly process is standardized, and a unified measurement microenvironment is established, the difference in the measured spectra of different ellipsometers will not be significant. That is, the difference between the spectrum measured by the reference ellipsometer and the spectrum measured by the ellipsometer to be calibrated will not be significant. Therefore, using the simulated spectrum to compare with itself will not cause conflict in the comparison between the spectrum measured by the reference ellipsometer and the spectrum measured by the ellipsometer to be calibrated. On the contrary, it helps to save computational resources because there is no need to label the simulated data.

[0052] In a preferred embodiment, the present invention includes two different neural network deployment frameworks: a first neural network model and a second neural network model. The first neural network model establishes a separate neural network model between each ellipsometer to be calibrated and the reference ellipsometer, while the second neural network model establishes only one neural network model between all ellipsometers to be calibrated and the reference ellipsometer.

[0053] The first neural network model offers high accuracy but requires more training costs, while the second neural network model offers lower training costs but slightly lower accuracy. To balance the advantages of both deployment frameworks, this embodiment determines whether the number of ellipsometers to be calibrated is less than or equal to a preset threshold before training the neural network model. If the number of ellipsometers is less than or equal to the preset threshold, the neural network model is designated as the first neural network model; if the number of ellipsometers is greater than the preset threshold, the neural network model is designated as the second neural network model. This approach allows for a manageable computational load when the number of ellipsometers is small, enabling the use of the first neural network model to maximize model accuracy while maintaining cost control. However, for a large number of ellipsometers, the computational load becomes excessive, requiring significant training time. Therefore, using the second neural network model reduces training costs while ensuring a certain level of accuracy.

[0054] Figure 4 is a schematic diagram of the architecture of the first neural network model provided by the present invention. In the figure, E0 represents the reference ellipsometer, E1 represents the ellipsometer to be calibrated, and N represents the neural network model. As can be seen from Figure 4, there is a neural network model between each ellipsometer to be calibrated and the reference ellipsometer. The first loss function corresponding to the first neural network model is: Where loss1 represents the first loss function, Let w represent the weights of the first neural network model, and b represent the bias term of the first neural network model. This represents the first loss term of the first loss function. This represents the spectrum to be calibrated of the nth ellipsometer to be calibrated, where n is a positive integer. Represents the reference spectrum. X represents the second loss term of the first loss function. simu This represents the simulated spectrum.

[0055] Figure 5 is a schematic diagram of the architecture of the second neural network model provided by this invention. Similarly, in the figure, E0 represents the reference ellipsometer, E1 represents the ellipsometer to be calibrated, and N represents the neural network model. As can be seen from Figure 5, all ellipsometers to be calibrated and the reference ellipsometer have only one neural network model. The second loss function corresponding to this second neural network model is: Where loss2 represents the second loss function, Let w represent the weights of the second neural network model, and b represent the bias term of the second neural network model. This represents the first loss term of the second loss function. Let represent the spectrum to be calibrated of the nth ellipsometer to be calibrated, where n is a positive integer less than or equal to N, and N represents the total number of ellipsometers to be calibrated. Represents the reference spectrum. X represents the second loss term of the second loss function. simu This represents the simulated spectrum.

[0056] Optionally, the neural network model is trained, including: calculating the value of the loss function based on the training output of the spectrum to be calibrated and the training output of the simulated spectrum; calculating the gradient of each neuron in the neural network model based on the value of the loss function using the backpropagation algorithm; and adjusting the weights and / or biases of each neuron based on the gradients to make the loss function converge until the training termination condition is met, which includes the loss function value reaching a threshold, the number of iterations reaching a preset number, and the slope of the loss function being less than a threshold.

[0057] For example, this invention acquires 700 spectra to be calibrated and reference spectra, as well as 100,000 simulated spectra. These spectra to be calibrated and reference spectra are then divided. For instance, 500 spectra to be calibrated and their corresponding reference spectra can be grouped together with the 100,000 simulated spectra into a training set. The remaining 200 spectra to be calibrated and their corresponding reference spectra are grouped into a test set. The training set is used to train the neural network model, and the test set is used to test the trained neural network model.

[0058] The training process included: inputting 100,000 simulated spectra, 500 spectra to be calibrated, and the corresponding reference spectra from the training set into the neural network model. The output values ​​of the simulated spectra were compared with the input values ​​to calculate the mean squared error, which served as the first part of the loss term (simulation data loss) for the neural network model. The output values ​​of the spectra to be calibrated were compared with the reference spectra to calculate the mean squared error, which served as the second part of the loss term (real data loss). The sum of the first and second loss terms constituted the loss function of the neural network model. The training process iterated 500 times. The changes of the first and second loss terms with the number of iterations are shown in Figure 6. As can be seen from Figure 6, both loss terms converged with the increase of the number of iterations, indicating that the total value of the loss function also converged, and the model was well trained.

[0059] The iterative process specifically includes: 1) Inputting 100,000 simulated spectra, 500 spectra to be calibrated, and 500 corresponding reference spectra from the training set into the neural network model, and calculating the activation value of each neuron in each layer in turn until the predicted value of the output layer is obtained; 2) Calculating the value of the loss function; 3) Based on the value of the loss function, calculating the gradient of each neuron in reverse, that is, calculating the contribution of each neuron to the value of the loss function; 4) Based on the gradient of each neuron, updating the weights and / or biases of each neuron through an optimization algorithm, including but not limited to gradient descent and its variants.

[0060] In step 1), the neurons in the neural network model refer to each node in the hidden and output layers. Each neuron includes weights and an activation function. Typically, the weights of each neuron in a layer can be represented by an N*M matrix, where N represents the dimension of the input layer and M represents the dimension of the output layer. Common activation functions include, but are not limited to, the Sigmoid function and the ReLU function. To calculate the activation value of each neuron in a layer, the activation value of the feedforward layer neuron is multiplied by its weight to obtain the adjusted input value. This adjusted input value is then substituted into the activation function as the independent variable to calculate the activation value of that neuron.

[0061] In step 3), the backpropagation algorithm essentially refers to solving the gradient of the feedforward neuron in reverse by using the chain rule based on the value of the loss function of the output layer.

[0062] Optionally, after obtaining the trained neural network model, the method further includes: testing the trained neural network model, which includes: acquiring the test reference spectrum (i.e., the 200 reference spectra in the aforementioned test set) and the corresponding test calibration spectrum (i.e., the 200 calibrated spectra in the aforementioned test set); inputting the test calibration spectrum into the trained neural network model to obtain the test output spectrum; calculating the mean square error between the test output spectrum and the test reference spectrum based on the test loss function to obtain the value of the test loss function; if the value of the test loss function is lower than the loss threshold, the model is considered to have been trained and the trained neural network model is used for measurement; if the value of the test loss function is greater than or equal to the loss threshold, the model is considered to have been insufficiently trained and the neural network model is retrained. In this embodiment, the loss threshold is set to 10^-4.

[0063] Figure 7 is a schematic diagram of model evaluation information provided by the present invention. As can be seen from Figure 7, in the model of the present invention, the values ​​of the test loss function are all within 10^-5. Therefore, the model trained by the present invention meets the conditions and can be used to adjust the measurement spectrum of the ellipsometer to be calibrated.

[0064] To more clearly illustrate the differences between the model-generated predicted spectrum, the spectrum to be calibrated, and the reference spectrum, the sample corresponding to the red dot in Figure 7 is selected, and its related spectrum is displayed in Figure 8, which shows one of the Mueller elements in the Mueller spectrum. As can be seen from Figure 3, the difference between the spectrum to be calibrated and the reference spectrum is large, while the difference between the model-generated predicted spectrum and the reference spectrum is very small. This means that the final measurement results will also have very small differences, which is sufficient to solve the problem of differences in measurement results between ellipsometers.

[0065] The technical solution of this embodiment uses the spectrum to be calibrated and the simulated spectrum as training inputs, and the spectrum to be calibrated and the simulated spectrum themselves as training labels to train a neural network model. When the loss function of the neural network model converges, the converged model is determined to be the trained neural network model. This trained neural network model can take the measured spectrum of the reference ellipsometer as input and output a predicted spectrum close to the reference spectrum. Therefore, each ellipsometer to be calibrated can calculate the final measurement result based on essentially the same spectrum, reducing the difference in measurement results between the various ellipsometers. This solves the problems that high-requirement component selection, assembly processes, and environmental conditions require a large amount of human and material resources, which increases the manufacturing cost of the ellipsometer; and that operations such as fine-tuning spectral weights and releasing fixed parameters require a large amount of manual intervention in parameter tuning, resulting in poor generalization performance.

[0066] Figure 9 is a flowchart illustrating an ellipsometric measurement method provided by this invention. This method utilizes the trained neural network model from the above embodiments to process the measured spectrum, resolving the discrepancy between measurement results between the ellipsometer to be calibrated and the reference ellipsometer with minimal human intervention. The sample parameter measurement process is as follows:

[0067] S201. Measure the sample using the ellipsometer to be calibrated to obtain the measurement spectrum.

[0068] The measurement spectrum obtained by measuring a sample using an ellipsometer to be calibrated can be understood as using the ellipsometer to be calibrated (i.e., the ellipsometer used to measure the sample) to obtain a spectral information.

[0069] S202. Input the measured spectrum into the trained neural network model to obtain the predicted spectrum.

[0070] The trained neural network model is any one of the trained neural network models in the above embodiments, used to adjust the measurement spectrum of the ellipsometer to be calibrated. The predicted spectrum can be understood as the measurement spectrum after being processed by the trained neural network model. This predicted spectrum is almost identical to the reference spectrum. As can be seen from Figure 8, the reference spectrum and the predicted spectrum almost overlap, that is, the difference between the two spectra is very low. When the predicted spectrum of each ellipsometer to be calibrated overlaps with the reference spectrum, the difference in measurement results between the ellipsometers will be infinitesimally small, which is also the purpose of this invention.

[0071] S203. Obtain the theoretical spectrum of the sample, and fit the predicted spectrum and the theoretical spectrum to obtain the measured value of the parameter to be measured.

[0072] Theoretical spectra are obtained by modeling and simulating the sample. The established model can be adjusted to accommodate the corresponding measured parameters. The model is then simulated using algorithms such as RCWA and FDTD. This simulation process yields a large number of theoretical spectra, with each spectrum corresponding to a set of measured parameters.

[0073] Based on the predicted spectrum and the theoretical spectrum, the measured value of the parameter to be measured is obtained by fitting. The theoretical spectrum that is closest to the predicted spectrum is found by a nonlinear regression algorithm, and the set of floating parameters corresponding to the theoretical spectrum is the measured value of the parameter to be measured.

[0074] The measured values ​​of the parameters to be measured can be understood as the measurement results of the sample, including but not limited to optical constants, key optical dimensions of micro / nano structures, and film thickness.

[0075] Figure 10 is a schematic diagram of an ellipsometric measurement process provided by the present invention, illustrating the specific implementation process of this embodiment. In the figure, ellipsometric measurement represents the measurement of a sample using an ellipsometric instrument to be calibrated, and the AE neural network model represents the neural network model trained in the above embodiment. As can be seen from the figure, the present invention uses an ellipsometric instrument to be calibrated to measure the sample and obtain a measurement spectrum. Then, it uses the AE neural network model to process the measurement spectrum to obtain a predicted spectrum. On the other hand, it models the sample and generates a theoretical spectrum. Finally, it fits the predicted spectrum and the theoretical spectrum to obtain the sample measurement result.

[0076] The technical solution of this embodiment can solve the problem of differences in measurement results between ellipsometers without requiring a large number of real samples and with minimal human intervention, thus greatly ensuring the measurement accuracy and efficiency of the samples.

[0077] Figure 11 is a schematic diagram of an ellipticity measurement device provided by the present invention. As shown in Figure 11, the device includes: a sample measurement module 301, a sample processing module 302, and a parameter measurement module 303.

[0078] Measurement module 301 is used to acquire the measurement spectrum.

[0079] The processing module 302 is used to input the measured spectrum into any trained neural network model in this invention to obtain the predicted spectrum.

[0080] The parameter acquisition module 303 is used to fit the predicted spectrum and the theoretical spectrum of the sample to obtain the measured value of the parameter to be measured of the sample.

[0081] The ellipticity measurement device provided in this embodiment is used to implement the ellipticity measurement method provided by the present invention, and the beneficial effects of the device are the same as those of the ellipticity measurement method.

[0082] The specific embodiments described above do not constitute a limitation on the scope of protection of this invention. Those skilled in the art should understand that various modifications, combinations, sub-combinations, and substitutions can be made according to design requirements and other factors. Any modifications, equivalent substitutions, and improvements made within the spirit and principles of this invention should be included within the scope of protection of this invention.

Claims

1. A method for training a neural network model, characterized in that, include: A reference spectrum and a corresponding calibrated spectrum are obtained. The reference spectrum is obtained by measuring the sample using a reference ellipsometer, and the calibrated spectrum is obtained by measuring the sample using an ellipsometer. Obtain the simulated spectrum of the sample; Using the spectrum to be calibrated and the simulated spectrum as training inputs, and the spectrum to be calibrated and the simulated spectrum themselves as training labels, a neural network model is trained to make the loss function of the neural network model converge, thereby obtaining a trained neural network model; wherein, the loss function includes a first loss term and a second loss term, the first loss term being derived based on the difference between the training output of the spectrum to be calibrated and the corresponding reference spectrum, and the second loss term being derived based on the difference between the training output of the simulated spectrum and the simulated spectrum itself.

2. The method for training a neural network model according to claim 1, characterized in that: The neural network model includes a first neural network model established based on each of the ellipsometers to be calibrated and the reference ellipsometer. The loss function includes a first loss function corresponding to the first neural network model. The first loss function is: Where loss1 represents the first loss function, This represents the first neural network model. This represents the first loss term of the first loss function. This represents the spectrum to be calibrated of the nth ellipsometer to be calibrated, where n is a positive integer. This represents the reference spectrum. X represents the second loss term of the first loss function. simu This represents the simulated spectrum.

3. The method for training a neural network model according to claim 1, characterized in that: The neural network model includes a second neural network model based on multiple ellipsometers to be calibrated and a reference ellipsometer. The loss function includes a second loss function corresponding to the second neural network model, and the second loss function is: Where loss2 represents the second loss function, This represents the second neural network model. This represents the first loss term of the second loss function. Let represent the spectrum to be calibrated of the nth ellipsometer to be calibrated, where n is a positive integer less than or equal to N, and N represents the total number of ellipsometers to be calibrated. Represents the reference spectrum. X represents the second loss term of the second loss function. simu This represents the simulated spectrum.

4. The method for training a neural network model according to claim 2 or 3, characterized in that, Before training the neural network model, the method further includes: determining whether the number of ellipsometers to be calibrated is less than or equal to a preset threshold; if the number of ellipsometers to be calibrated is less than or equal to the preset threshold, then the neural network model is set as a first neural network model; if the number of ellipsometers to be calibrated is greater than the preset threshold, then the neural network model is set as a second neural network model.

5. The method for training a neural network model according to claim 1, characterized in that, Obtaining the simulated spectrum of the sample includes: establishing an optical property model of the sample; constructing a sampling space based on the number and value range of structural parameters in the optical property model, wherein the dimension of the sampling space is the same as the number of structural parameters, and the side length of the sampling space corresponds to the value range of the structural parameters; arranging a preset number of value points in the sampling space, and making the value points randomly and uniformly distributed in the sampling space; obtaining the structural parameters corresponding to each value point, inputting the corresponding structural parameters into the optical property model, and then performing optical simulation on the optical property model to obtain the simulated spectrum of the sample.

6. The method for training a neural network model according to claim 1, characterized in that, After obtaining the trained neural network model, the method further includes: testing the trained neural network model, the testing including: acquiring a test reference spectrum and a corresponding test calibration spectrum, wherein the test reference spectrum and the reference spectrum do not overlap, and the test calibration spectrum and the calibration spectrum do not overlap; inputting the test calibration spectrum into the trained neural network model to obtain a test output spectrum; obtaining the value of a test loss function based on a comparison between the test output spectrum and the test reference spectrum; if the value of the test loss function is lower than a loss threshold, then the trained neural network model is used for measurement; if the value of the test loss function is greater than or equal to the loss threshold, then the neural network model is retrained.

7. The method for training a neural network model according to claim 1, characterized in that: The difference includes mean square error or mean absolute error.

8. The method for training a neural network model according to claim 1, characterized in that, Training the neural network model includes: calculating the value of the loss function based on the training output of the spectrum to be calibrated and the training output of the simulated spectrum; calculating the gradient of each neuron in the neural network model based on the value of the loss function using a backpropagation algorithm; and adjusting the weights and / or biases of each neuron based on the gradients so that the loss function converges to a threshold.

9. The method for training a neural network model according to claim 1, characterized in that: The number of simulated spectra is greater than the number of reference spectra and the number of spectra to be calibrated; and / or, the number of reference spectra and corresponding spectra to be calibrated is 100-1000, and the number of simulated spectra is 10000-200000.

10. A method for measuring ellipticity, characterized in that, include: The sample was measured using an ellipsometer to be calibrated, and the measured spectrum was obtained. The measured spectrum is input into the trained neural network model as described in any one of claims 1 to 9 to obtain the predicted spectrum; The theoretical spectrum of the sample is obtained, and the measured value of the parameter to be measured of the sample is obtained by fitting the predicted spectrum and the theoretical spectrum.

Images