Wafer film thickness calculation method and device based on GAN and adaptive feedback mechanism

Abstract

The invention discloses a wafer film thickness calculation method and device based on a GAN and an adaptive feedback mechanism, and belongs to the field of semiconductor manufacturing. According to the wafer film thickness calculation method based on the GAN and the self-adaptive feedback mechanism, the data set is expanded through the GAN, and the model can adapt to more diversified spectral data, especially the situation that the film thickness range is wide. And a self-adaptive feedback mechanism further improves the learning ability of the model for special spectral data. Besides, virtual data and a feedback mechanism enable the model to respond to abnormal data, and the model is dynamically optimized, so that the prediction stability under different film thickness ranges and spectrum conditions is improved. Through combination with a deep learning algorithm, the final film thickness value is further calibrated, and the accuracy and physical rationality of a prediction result are ensured. According to the wafer film thickness calculation method based on the GAN and the adaptive feedback mechanism, high-precision and high-robustness film thickness prediction can be realized, and the method is suitable for various different practical application scenes.

Description

Technical Field

[0001] The embodiments of the present application relate to the field of semiconductor manufacturing, and in particular to a method and device for calculating wafer film thickness based on GAN and an adaptive feedback mechanism. Background Art

[0002] In semiconductor manufacturing, accurate measurement of wafer epitaxial thickness is crucial. Slight deviations in epitaxial thickness may lead to inconsistencies in device performance. Existing epitaxial thickness calculation methods are mostly based on physical models or empirical formulas, which are insufficient when faced with complex and diverse spectral data, especially when dealing with noise interference and non-standard data.

[0003] The reasons include insufficient data diversity, difficulty in handling non-standard spectra, and poor model generalization. For example, the number of experimental samples is limited and cannot cover all potential situations. Non-standard spectra may be caused by process deviations and equipment errors, which are difficult to handle with existing methods. The static models used in conventional technologies cannot adapt to complex working conditions and require frequent manual intervention. Summary of the invention

[0004] The embodiments of the present application provide a method and device for calculating wafer film thickness based on GAN and an adaptive feedback mechanism, which utilizes a generative adversarial network (GAN) to generate virtual spectral data to improve data diversity, and combines an adaptive feedback mechanism to optimize model performance in real time, thereby breaking through the bottleneck of traditional methods and achieving accurate and efficient wafer film thickness calculation.

[0005] On the one hand, a method for calculating wafer film thickness based on GAN and adaptive feedback mechanism is provided, the method comprising:

[0006] Generative adversarial network (GAN) is used to expand the training data of the spectral thickness prediction model to obtain an expanded training data set.

[0007] Inputting the training data set into a spectral thickness prediction model to perform film thickness prediction training;

[0008] Introducing an adaptive feedback mechanism to dynamically update the spectral thickness prediction model and the training data set, wherein the adaptive feedback mechanism is used to label feedback data in the prediction training;

[0009] The film thickness prediction result of the spectral thickness prediction model is calibrated and corrected.

[0010] Optionally, the use of a generative adversarial network (GAN) to expand the training data of the spectral thickness prediction model to obtain an expanded training data set includes:

[0011] Obtain wafer epitaxial spectral data, including spectral samples of different film thicknesses;

[0012] Generate virtual spectral data according to the wafer epitaxial spectral data using a generative adversarial network (GAN), wherein the virtual spectral data corresponds to a thickness label and a spectral variation pattern, and the thickness label is used to indicate the wafer epitaxial film thickness corresponding to each spectral data;

[0013] The virtual spectral data is added as new training data to the actual spectral data corresponding to the spectral thickness prediction model to obtain the expanded training data set.

[0014] Optionally, inputting the training data set into a spectral thickness prediction model to perform film thickness prediction training includes:

[0015] Inputting the virtual spectrum data into the spectrum thickness prediction model to predict the film thickness value;

[0016] Comparing the predicted film thickness value with the corresponding thickness label and calculating the error to obtain a corrected thickness label;

[0017] The corrected thickness labels are matched one by one with the corresponding virtual spectral data to obtain the trained spectral thickness prediction model.

[0018] Optionally, the feedback data includes the virtual spectral data and the corresponding real spectral data. The adaptive feedback mechanism is introduced to dynamically update the spectral thickness prediction model, and the training data set is dynamically updated, including:

[0019] Marking the virtual spectrum data whose error exceeds a preset threshold as the feedback data;

[0020] Re-inputting the feedback data into the spectral thickness prediction model for training;

[0021] The spectral thickness prediction model is dynamically updated according to new prediction training, and the training data set is dynamically updated.

[0022] Optionally, calibrating and correcting the film thickness prediction result of the spectral thickness prediction model includes:

[0023] After the spectral thickness prediction model outputs a prediction result about the film thickness value, the output film thickness value is used as a preliminary estimation value and enters a calibration step;

[0024] In the calibration step, the preliminary estimated value is calibrated using a peak counting method, FFT or a fitting algorithm.

[0025] Optionally, in the calibration step, the preliminary estimated value is calibrated using a peak counting method, FFT or a fitting algorithm, including:

[0026] In the calibration step, the film thickness is corrected based on mathematical properties in the virtual spectral data.

[0027] Optionally, the method further includes:

[0028] The spectral thickness prediction model is designed based on Transformer.

[0029] On the other hand, a wafer film thickness calculation device based on GAN and adaptive feedback mechanism is also provided, the device comprising:

[0030] The first processing module is used to expand the training data of the spectral thickness prediction model by using a generative adversarial network (GAN) to obtain an expanded training data set;

[0031] A second processing module is used to input the training data set into a spectral thickness prediction model to perform film thickness prediction training;

[0032] A third processing module is used to introduce an adaptive feedback mechanism to dynamically update the spectral thickness prediction model and the training data set, wherein the adaptive feedback mechanism is used to mark feedback data in the prediction training;

[0033] The fourth processing module is used to calibrate and correct the film thickness prediction result of the spectral thickness prediction model.

[0034] On the other hand, a computer-readable storage medium is provided, wherein the storage medium stores at least one instruction, and the at least one instruction is used to be executed by a processor to implement the wafer film thickness calculation method based on GAN and adaptive feedback mechanism as described in the above aspects.

[0035] On the other hand, a computer program product is also provided, which stores at least one instruction, and the at least one instruction is loaded and executed by the processor to implement the wafer film thickness calculation method based on GAN and adaptive feedback mechanism described in the above aspect.

[0036] The technical effects brought about by this application are at least as follows.

[0037] The embodiment of the present application provides a wafer film thickness calculation method based on GAN and adaptive feedback mechanism. By expanding the data set through GAN, the model can adapt to more diverse spectral data, especially when the film thickness range is wide. The adaptive feedback mechanism further enhances the model's learning ability for special spectral data. In addition, virtual data and feedback mechanisms enable the model to respond to abnormal data and dynamically optimize the model, thereby improving the prediction stability under different film thickness ranges and spectral conditions. By combining a deep learning algorithm, the final film thickness value is further calibrated to ensure the accuracy and physical rationality of the prediction results. The wafer film thickness calculation method based on GAN and adaptive feedback mechanism can achieve high-precision and high-robustness film thickness prediction, and adapt to a variety of different practical application scenarios. BRIEF DESCRIPTION OF THE DRAWINGS

[0038] Figure 1 A flowchart of a method for calculating wafer film thickness based on GAN and an adaptive feedback mechanism provided by an exemplary embodiment of the present application is shown. DETAILED DESCRIPTION

[0039] In order to make the objectives, technical solutions and advantages of the present application clearer, the implementation methods of the present application will be further described in detail below with reference to the accompanying drawings.

[0040] The term "multiple" as used herein refers to two or more than two. "And / or" describes the relationship between related objects, indicating that three relationships can exist. For example, A and / or B can mean: A exists alone, A and B exist at the same time, and B exists alone. The character " / " generally indicates that the related objects are in an "or" relationship.

[0041] Example 1

[0042] Please refer to Figure 1 , which shows a flow chart of a wafer film thickness calculation method based on GAN and adaptive feedback mechanism provided by an exemplary embodiment of the present application, the method comprising:

[0043] Step 1: Use the generative adversarial network (GAN) to expand the training data of the spectral thickness prediction model to obtain an expanded training data set.

[0044] In this step, virtual data is mainly generated through GAN to expand the existing training data set to improve the generalization ability of the model.

[0045] In a possible implementation, step 1 includes steps S11 to S13.

[0046] Step S11, acquiring wafer epitaxial spectral data, the data including spectral samples of different film thicknesses.

[0047] Use spectrum analysis equipment to collect real wafer epitaxial film thickness spectrum data. The data includes spectrum samples at different film thicknesses (e.g., 5μm, 10μm, 20μm, etc.) to ensure that the data is representative.

[0048] In one example, a spectrometer or other spectral analysis technique is used to measure spectra of wafers of different thicknesses (e.g., 5 μm, 10 μm, 15 μm, etc.) Each spectra is accompanied by a film thickness label.

[0049] Step S12, using a generative adversarial network (GAN) to generate virtual spectral data based on the wafer epitaxial spectral data, the virtual spectral data corresponding to a thickness label and a spectral change pattern, the thickness label is used to indicate the wafer epitaxial film thickness corresponding to each spectral data.

[0050] Based on the existing actual spectral data, the GAN model is used to generate new virtual spectral data. GAN generates new spectral samples by simulating the spectral changes of different film thicknesses and assigns the correct film thickness label to each sample.

[0051] In one example, if the original dataset has only limited film thicknesses (e.g., 10 μm), GAN can generate spectral data for other film thicknesses (e.g., 5 μm, 15 μm, etc.), making the dataset more diverse.

[0052] Step S13, adding the virtual spectral data as new training data to the actual spectral data corresponding to the spectral thickness prediction model to obtain an expanded training data set.

[0053] The virtual spectral data generated in step S12 are combined with the original data to form a new training data set. These virtual data will be used as training input to help the prediction model better adapt to the spectral characteristics of different film thicknesses.

[0054] In one example, suppose the real dataset contains 1,000 samples and GAN generates 500 new samples. The final training dataset contains 1,500 samples, increasing the amount and diversity of training data for the model.

[0055] Step 2: Input the training data set into the spectral thickness prediction model to perform film thickness prediction training.

[0056] In this step, the training dataset will be used to train the spectral thickness prediction model so that it can predict the film thickness of the wafer based on the spectral data.

[0057] In a possible implementation, step 2 includes steps S21 to S23.

[0058] Step S21, inputting virtual spectrum data into a spectrum thickness prediction model to predict the film thickness value.

[0059] The virtual spectral data in the training dataset is input into a prediction model (e.g., a Transformer-based model) for preliminary prediction of film thickness. The model will generate a predicted value of film thickness by learning the relationship between spectral features and film thickness.

[0060] In one example, each spectral data is fed into the Transformer model to predict the film thickness. For example, one spectral sample might predict a film thickness of 12.3 μm.

[0061] Step S22, comparing the predicted film thickness value with the corresponding thickness label and calculating the error to obtain a corrected thickness label.

[0062] The predicted value is compared with the corresponding true film thickness label, and the prediction error (e.g., mean square error MSE) is calculated. If the error is large, it is marked as "feedback data".

[0063] In one example, if the model predicts a film thickness of 12.3 μm and the actual film thickness is 10 μm, the error is 2.3 μm. When this error exceeds a preset threshold, the sample will be processed as feedback data.

[0064] Step S23, matching the corrected thickness labels with the corresponding virtual spectrum data one by one to obtain a trained spectrum thickness prediction model.

[0065] Based on the error calculation results, the film thickness labels are corrected and updated to the training dataset together with the virtual spectrum data. After correction, the model is updated to make it more accurate in predicting film thickness.

[0066] In one example, the film thickness labels of samples whose errors exceed a threshold are adjusted to actual values ​​and re-input into the model for retraining.

[0067] Step 3: Introduce an adaptive feedback mechanism to dynamically update the spectral thickness prediction model and the training data set. The adaptive feedback mechanism is used to label the feedback data in the prediction training.

[0068] By introducing an adaptive feedback mechanism, this application can continuously optimize the model and dynamically update the training data set to ensure that the model has higher accuracy in different scenarios.

[0069] In a possible implementation, step 3 includes steps S31 to S33.

[0070] Step S31: marking virtual spectrum data whose error exceeds a preset threshold as feedback data.

[0071] Mark the data whose prediction error exceeds the threshold. These data are considered as "feedback data" that the model needs to relearn.

[0072] In one example, if the predicted film thickness error exceeds ±1 μm, the sample is marked as feedback data.

[0073] Step S32, re-inputting the feedback data into the spectral thickness prediction model for training.

[0074] The data marked as feedback is re-input into the training process, and the model is optimized again on this feedback data through incremental learning.

[0075] In one example, for prediction results with large errors, the model will focus on learning these samples with large errors, making the model more accurate in these special scenarios.

[0076] Step S33: dynamically updating the spectral thickness prediction model according to the new prediction training, and dynamically updating the training data set.

[0077] After the feedback data is retrained by the model, the model weights and parameters will be dynamically adjusted. At this time, the training data set will also be updated with the addition of feedback data.

[0078] In one example, if the feedback data indicates that the spectral signature of a certain film thickness is prone to being misjudged by the model, the model will pay special attention to this situation and optimize the parameters to reduce future errors.

[0079] Step 4: calibrate and correct the film thickness prediction result of the spectral thickness prediction model.

[0080] Finally, after obtaining the preliminary film thickness prediction results, the accuracy of the results is further improved through traditional calibration methods.

[0081] In a possible implementation, step 4 includes steps S41 to S42.

[0082] Step S41, after the spectral thickness prediction model outputs the prediction result about the film thickness value, the output film thickness value is used as a preliminary estimation value and enters the calibration step.

[0083] The prediction model outputs a preliminary film thickness value as a preliminary estimated film thickness value.

[0084] In one example, the model predicted a film thickness of 12.3 μm, which was used as a preliminary estimate and entered the calibration phase.

[0085] Step S42, in the calibration step, the preliminary estimated value is calibrated using a peak counting method, FFT or a fitting algorithm.

[0086] In the calibration step, the film thickness is corrected based on the mathematical properties in the virtual spectral data. The film thickness is calibrated using the peak counting method, FFT or fitting algorithms based on the frequency characteristics in the spectral data. The prediction results of the deep learning model are corrected by physical properties.

[0087] In one example, for certain spectral data, the film thickness value may be corrected according to the frequency component by using an FFT algorithm, for example, a film thickness value predicted to be 12.3 μm may be corrected to 12.0 μm.

[0088] In summary, the embodiment of the present application provides a method for calculating wafer film thickness based on GAN and adaptive feedback mechanism. By expanding the data set through GAN, the model can adapt to more diverse spectral data, especially when the film thickness range is wide. The adaptive feedback mechanism further enhances the model's learning ability for special spectral data. In addition, virtual data and feedback mechanisms enable the model to respond to abnormal data and dynamically optimize the model, thereby improving the prediction stability under different film thickness ranges and spectral conditions. By combining a deep learning algorithm, the final film thickness value is further calibrated to ensure the accuracy and physical rationality of the prediction results. The wafer film thickness calculation method based on GAN and adaptive feedback mechanism can achieve high-precision and high-robustness film thickness prediction, and adapt to a variety of different practical application scenarios.

[0089] Example 2

[0090] In addition, the invention also provides information on how to design a spectral thickness prediction model. In one possible implementation, the spectral thickness prediction model is designed based on Transformer.

[0091] In Example 2, the process of designing a spectral thickness prediction model based on Transformer can be described in detail through the following specific steps. We will combine spectral data and film thickness labels to build a deep learning model that can capture the complex nonlinear relationship between spectrum and film thickness.

[0092] Data preparation and preprocessing.

[0093] Before model training, the collected spectral data must first be preprocessed to ensure that it is compatible with the input requirements of the Transformer model. The spectral data is normalized or standardized to ensure that the input spectral feature values ​​are in the same numerical range. A common practice is to standardize each channel of the spectral data (i.e., the intensity of each wavelength) with zero mean and unit variance.

[0094] In one example, assuming that the spectral data contains intensity values ​​of 500 wavelengths, for each wavelength intensity value, calculate its mean and standard deviation, and then standardize the data. Organize each spectral sample into a vector of fixed length. Each sample contains spectral information of 500 wavelengths and the corresponding film thickness label. Each spectral data (500 wavelength intensity values) is converted into a vector and paired with the corresponding film thickness label (such as 10μm, 15μm, etc.) to form a training sample.

[0095] Build the Transformer model architecture.

[0096] The core of the Transformer architecture is the "Self-Attention Mechanism", which can take into account the dependencies between different parts when processing the input sequence. In the case of spectral data, the self-attention mechanism can help the model recognize the complex relationship between spectral wavelengths. Input the spectral data into the input layer of the Transformer model. Each spectral data is represented as a 500-dimensional vector, with each dimension corresponding to the intensity value of a wavelength.

[0097] In one example, a spectral data sample [s1, s2, ..., s500], where each si represents the intensity value of the corresponding wavelength. Because the Transformer model itself does not have sequential information, position encoding needs to be added to preserve the positional relationship of each wavelength. The position encoding is added to the input vector to identify the position of each wavelength in the sequence. Assuming that the input spectral data is a 500-dimensional vector, the position encoding will be added to each element to preserve the position information of the wavelength. Through the self-attention mechanism, the model can learn the correlation between each wavelength and other wavelengths, thereby capturing the complex relationship between spectral features. For a wavelength position in a spectral sample, the model dynamically adjusts the representation of each position by calculating the correlation (attention weight) with other wavelengths. For example, a wavelength (such as 500μm) may have a strong correlation with another wavelength (such as 600μm), and the model will dynamically adjust the weight according to this correlation.

[0098] In addition, through the multi-head self-attention mechanism, the model can calculate multiple different attention weights in parallel, capturing different levels of correlations and dependencies. Through the multi-head mechanism, the model can learn multiple different correlations in parallel (for example, some wavelengths have a strong effect on film thickness, while some wavelengths have a weaker effect on film thickness).

[0099] The data processed by the self-attention layer will be passed to the feedforward neural network for further processing to extract the high-order features of the spectral data. The representation of each spectral data sample after the self-attention mechanism will be passed to a small fully connected neural network for further feature abstraction.

[0100] In order to better learn the complex spectrum-film thickness relationship, multiple self-attention layers and feedforward neural network layers can be stacked to form a deep model, such as stacking 3-6 layers of Transformer layers, so that the model can capture the relationship between deeper spectral features and film thickness.

[0101] Model output with film thickness prediction.

[0102] The final output of the Transformer model is the predicted value of the film thickness corresponding to each spectral data sample. The output layer of the model is usually a fully connected layer that outputs a single value (i.e., film thickness).

[0103] The output layer of the model converts the data processed by the Transformer network into a predicted film thickness value. Usually, it is mapped through a linear layer and finally outputs the film thickness (for example: 10.2μm). For example, if the model outputs a value of 10.2 after processing by the Transformer layer, it means that the predicted film thickness is 10.2μm.

[0104] Loss function and model training.

[0105] In order to train the model and optimize its prediction performance, a loss function needs to be defined. Usually, the mean square error (MSE) is used as the loss function to calculate the difference between the film thickness predicted by the model and the actual film thickness.

[0106] The mean square error (MSE) is used to calculate the difference between the predicted value and the actual film thickness label. The model is optimized through the backpropagation algorithm, the parameters of the Transformer model are updated, and the loss function is minimized. For example, the Adam optimizer is used for training, the weight of the model is adjusted, and the error between the predicted film thickness and the actual film thickness is gradually reduced.

[0107] Model evaluation and calibration.

[0108] After training is complete, the model can be evaluated on the test set to check its performance on unseen data. If the model performance is poor, you can consider using algorithms (such as FFT, peak counting, etc.) to post-process and calibrate the prediction results.

[0109] In one example, by calculating the MSE or other evaluation indicators (such as R 2) to evaluate the prediction performance of the model. For those predictions with large errors, the film thickness value can be calibrated using FFT or the number peak method. For example, assuming that the film thickness predicted by the model is 12.5μm, the calibration value obtained through FFT analysis is 12.0μm, and the final output is 12.0μm as the final film thickness.

[0110] Therefore, the Transformer-based spectral thickness prediction model effectively captures the nonlinear relationship in the spectral data through the self-attention mechanism. By stacking multiple Transformer layers, the model can learn the complex dependencies between different wavelengths at multiple levels, thereby accurately predicting the film thickness. The entire process includes data preprocessing, model construction, training, evaluation and post-processing (such as FFT calibration), ensuring the high accuracy and robustness of the final prediction.

[0111] In summary, in the embodiments of the present application, the model design is based on the Transformer architecture, which can better capture the long-distance dependencies in the spectral data and improve the accuracy of the prediction. Among them, the self-attention mechanism of the Transformer allows the model to comprehensively consider the information of the entire spectral range when processing each spectral sample, rather than relying solely on local features. The multi-level structure of the Transformer enables the model to extract features layer by layer, from low-level frequency features to high-level film thickness patterns, layer by layer, and gradually accurately predict the film thickness.

[0112] An exemplary embodiment of the present application further provides a wafer film thickness calculation device based on GAN and an adaptive feedback mechanism, the device comprising:

[0113] The first processing module is used to expand the training data of the spectral thickness prediction model by using a generative adversarial network (GAN) to obtain an expanded training data set;

[0114] A second processing module is used to input the training data set into a spectral thickness prediction model to perform film thickness prediction training;

[0115] A third processing module is used to introduce an adaptive feedback mechanism to dynamically update the spectral thickness prediction model and the training data set, wherein the adaptive feedback mechanism is used to mark feedback data in the prediction training;

[0116] The fourth processing module is used to calibrate and correct the film thickness prediction result of the spectral thickness prediction model.

[0117] Optionally, the first processing module includes:

[0118] A first processing unit is used to obtain wafer epitaxial spectral data, the data including spectral samples of different film thicknesses;

[0119] A second processing unit is used to generate virtual spectrum data according to the wafer epitaxial spectrum data using a generative adversarial network (GAN), wherein the virtual spectrum data corresponds to a thickness label and a spectrum change pattern, and the thickness label is used to indicate the wafer epitaxial film thickness corresponding to each spectrum data;

[0120] The third processing unit is used to add the virtual spectral data as new training data to the actual spectral data corresponding to the spectral thickness prediction model to obtain the expanded training data set.

[0121] Optionally, the second processing module includes:

[0122] A fourth processing unit, used for inputting the virtual spectrum data into the spectrum thickness prediction model to predict the film thickness value;

[0123] a fifth processing unit, configured to compare the predicted film thickness value with the corresponding thickness label and calculate an error to obtain a corrected thickness label;

[0124] The sixth processing unit is used to match the corrected thickness labels with the corresponding virtual spectrum data one by one to obtain a trained spectrum thickness prediction model.

[0125] Optionally, the third processing module includes:

[0126] a seventh processing unit, configured to mark the virtual spectrum data whose error exceeds a preset threshold as the feedback data;

[0127] An eighth processing unit, configured to re-input the feedback data into the spectral thickness prediction model for training;

[0128] A ninth processing unit is used to dynamically update the spectral thickness prediction model according to new prediction training, and to dynamically update the training data set.

[0129] Optionally, the fourth processing module includes:

[0130] A tenth processing unit, configured to use the output film thickness value as a preliminary estimation value after the spectral thickness prediction model outputs a prediction result about the film thickness value, and enter a calibration step;

[0131] The eleventh processing unit is used to calibrate the preliminary estimated value by using a peak counting method, FFT or a fitting algorithm in the calibration step.

[0132] Optionally, the eleventh processing unit is further used to correct the film thickness according to the mathematical characteristics in the virtual spectral data in the calibration step.

[0133] Optionally, the device further comprises:

[0134] The fifth processing module is used to design the spectral thickness prediction model based on Transformer.

[0135] On the other hand, a computer-readable storage medium is provided, wherein the storage medium stores at least one instruction, and the at least one instruction is used to be executed by a processor to implement the wafer film thickness calculation method based on GAN and adaptive feedback mechanism as described in the above aspects.

[0136] On the other hand, a computer program product is also provided, which stores at least one instruction, and the at least one instruction is loaded and executed by the processor to implement the wafer film thickness calculation method based on GAN and adaptive feedback mechanism described in the above aspect.

[0137] The serial numbers of the above-mentioned embodiments of the present application are for description only and do not represent the advantages or disadvantages of the embodiments.

[0138] A person skilled in the art will understand that all or part of the steps to implement the above embodiments may be accomplished by hardware or by instructing related hardware through a program, and the program may be stored in a computer-readable storage medium, and the above-mentioned storage medium may be a read-only memory, a disk or an optical disk, etc.

[0139] The above description is only an optional embodiment of the present application and is not intended to limit the present application. Any modifications, equivalent substitutions, improvements, etc. made within the spirit and principles of the present application shall be included in the protection scope of the present application.

Claims

1. A wafer film thickness calculation method based on GAN and adaptive feedback mechanism, characterized in that: The method comprises: Generative adversarial network (GAN) is used to expand the training data of the spectral thickness prediction model to obtain an expanded training data set. Inputting the training data set into a spectral thickness prediction model to perform film thickness prediction training; Introducing an adaptive feedback mechanism to dynamically update the spectral thickness prediction model and the training data set, wherein the adaptive feedback mechanism is used to label feedback data in the prediction training; The film thickness prediction result of the spectral thickness prediction model is calibrated and corrected.

2. The method according to claim 1, characterized in that The method of using the generative adversarial network (GAN) to expand the training data of the spectral thickness prediction model to obtain an expanded training data set includes: Obtain wafer epitaxial spectral data, including spectral samples of different film thicknesses; Generate virtual spectral data according to the wafer epitaxial spectral data using a generative adversarial network (GAN), wherein the virtual spectral data corresponds to a thickness label and a spectral variation pattern, and the thickness label is used to indicate the wafer epitaxial film thickness corresponding to each spectral data; The virtual spectral data is added as new training data to the actual spectral data corresponding to the spectral thickness prediction model to obtain the expanded training data set.

3. The method according to claim 2, characterized in that The step of inputting the training data set into the spectral thickness prediction model to perform film thickness prediction training comprises: Inputting the virtual spectrum data into the spectrum thickness prediction model to predict the film thickness value; Comparing the predicted film thickness value with the corresponding thickness label and calculating the error to obtain a corrected thickness label; The corrected thickness labels are matched one by one with the corresponding virtual spectral data to obtain the trained spectral thickness prediction model.

4. The method according to claim 3, characterized in that The introducing of the adaptive feedback mechanism to dynamically update the spectral thickness prediction model and the dynamically updating of the training data set includes: Marking the virtual spectrum data whose error exceeds a preset threshold as the feedback data; Re-inputting the feedback data into the spectral thickness prediction model for training; The spectral thickness prediction model is dynamically updated according to new prediction training, and the training data set is dynamically updated.

5. The method according to claim 2, characterized in that: The calibrating and correcting the film thickness prediction result of the spectral thickness prediction model includes: After the spectral thickness prediction model outputs a prediction result about the film thickness value, the output film thickness value is used as a preliminary estimation value and enters a calibration step; In the calibration step, the preliminary estimated value is calibrated using a peak counting method, FFT or a fitting algorithm.

6. The method according to claim 5, characterized in that In the calibration step, the preliminary estimated value is calibrated using a peak counting method, FFT or a fitting algorithm, including: In the calibration step, the film thickness is corrected based on mathematical properties in the virtual spectral data.

7. The method according to any one of claims 1 to 6, characterized in that: The method further comprises: The spectral thickness prediction model is designed based on Transformer.

8. A wafer film thickness calculation device based on GAN and adaptive feedback mechanism, characterized in that: The device comprises: The first processing module is used to expand the training data of the spectral thickness prediction model by using a generative adversarial network (GAN) to obtain an expanded training data set; A second processing module is used to input the training data set into a spectral thickness prediction model to perform film thickness prediction training; A third processing module is used to introduce an adaptive feedback mechanism to dynamically update the spectral thickness prediction model and the training data set, wherein the adaptive feedback mechanism is used to mark feedback data in the prediction training; The fourth processing module is used to calibrate and correct the film thickness prediction result of the spectral thickness prediction model.

9. A computer-readable storage medium, characterized in that: The computer-readable storage medium stores a computer program, which is loaded and executed by a processor to implement the wafer film thickness calculation method based on GAN and adaptive feedback mechanism as described in any one of claims 1 to 7.

Images