Control Method and Device for Automatically Adjusting Thinning Thickness of Wafer

Abstract

The present application provides a control method and device for automatically adjusting the thinning thickness of a wafer, which relates to the field of semiconductor manufacturing technology. The method includes: obtaining an initial TTV value measured by an on-line TTV measurement module during the wafer thinning process; obtaining a corrected TTV value based on the initial TTV value by using a TTV compensation model; if the corrected TTV value is not within a set specification, determining an optimal TTV angle parameter through a pre-trained prediction model; controlling the angle of the grinding disk in the wafer thinning processing equipment to be adjusted to the optimal TTV angle parameter so as to act on the next wafer thinning thickness processing. This method can achieve real-time precise control and continuous optimization of the wafer thinning process, improving production efficiency and product quality.

Description

Technical Field

[0001] This application relates to the field of semiconductor manufacturing technology, and particularly to a control method and device for automatically adjusting the thinning thickness of a wafer. Background Art

[0002] In the field of semiconductor manufacturing, especially in the wafer thinning process, precisely controlling the thickness variation (TTV) of the wafer is crucial for ensuring the performance and quality of the wafer. Although the wafer thinning technology in the related art has been relatively mature, there are still some deficiencies.

[0003] Currently, the common practice is to conduct a sample running test before each shift, and adjust parameters such as the inclination angle of the carrier plate through manual judgment to control TTV. However, this method has obvious defects: on the one hand, the manual judgment is greatly affected by subjective factors, resulting in the inability to guarantee the accuracy of the results; on the other hand, the single adjustment method cannot adapt to the dynamic changes in the subsequent production process and is difficult to meet the requirements of higher-precision TTV control. In addition, the determination of TTV inclination adjustment in the related art often relies on empirical formulas, which further increases the uncertainty of the results. These problems not only affect the accuracy of wafer thinning, but also may lead to a reduction in production efficiency and an increase in production costs. Summary of the Invention

[0004] Based on this, in view of the above technical problems, it is necessary to provide a control method for automatically adjusting the thinning thickness of a wafer, which can achieve real-time and precise control of the wafer thinning process.

[0005] In a first aspect, this application provides a control method for automatically adjusting the thinning thickness of a wafer. This method is applied to a control module in a wafer thinning system. The wafer thinning system further includes an on-line TTV measurement module and a wafer thinning processing device. The method includes:

[0006] Obtain the initial TTV value measured by the on-line TTV measurement module during the wafer thinning process; wherein, the number of measurement points collected by the on-line TTV measurement module is less than that collected by the FRT measurement device, and the sampling resolution is less than the sampling resolution of the FRT measurement device. The FRT measurement device measures the TTV value of the same wafer after thinning is completed;

[0007] Based on the initial TTV value, use the TTV compensation model to obtain the corrected TTV value; wherein, the TTV compensation model is at least trained using the TTV values measured by the on-line TTV measurement module and the FRT measurement device for the same wafer;

[0008] If the corrected TTV value is not within the set specifications, determine the optimal TTV angle parameter through a pre-trained prediction model; wherein, the environmental feature data during the wafer thinning process is used as a constant of the prediction model, the TTV angle parameter is used as the optimization target of the prediction model, and the expected TTV value is used as the output target of the prediction model to traverse the feasible set of TTV angles to obtain the optimal angle parameter; wherein, the environmental feature data includes process parameter features;

[0009] Control the angle adjustment of the grinding disk in the wafer thinning processing equipment to the optimal TTV angle parameter to act on the next wafer thinning thickness processing.

[0010] In one embodiment, obtaining the corrected TTV value by using the TTV compensation model based on the initial TTV value includes:

[0011] Input the initial TTV value into the high-low resolution mapping model to obtain the virtual TTV value measured by the virtual FRT measurement device; the high-low resolution mapping model is obtained by mapping the TTV values measured by at least the online TTV measurement module and the FRT measurement device for the same wafer;

[0012] Concatenate the initial TTV value and the virtual TTV value and input them into the TTV compensation model to obtain the corrected TTV value.

[0013] In one embodiment, training the TTV compensation model includes:

[0014] Establish a TTV compensation model based on a neural network;

[0015] Concatenate the first resolution data and the second resolution data of the same wafer to obtain concatenated data; wherein, for the same wafer, the first resolution data at least includes the TTV value measured by the online TTV measurement module, and the second resolution data at least includes the TTV value measured by the FRT measurement device;

[0016] Use the concatenated data as the input of the initial TTV compensation model, and use the TTV value in the second resolution data used for concatenation as the output to train the TTV compensation model.

[0017] In one embodiment, determining the high-low resolution mapping model includes:

[0018] Establish an encoder f from the second resolution data to the first resolution data, and establish a decoder g from the first resolution data to the second resolution data;

[0019] Train the encoder f and the decoder g based on the first resolution data and the second resolution data of the same wafer, and obtain the high-low resolution mapping model based on the trained encoder f and decoder g.

[0020] In one embodiment, training the prediction model includes:

[0021] Obtaining sample features of a sample wafer, where, for any sample wafer, its sample features include environmental feature data, TTV angle parameters, and TTV values;

[0022] Constructing a prediction model based on the sample features;

[0023] Introducing an L1 regularization coefficient to shrink the coefficients of non-associated feature data in the prediction model;

[0024] Introducing an L2 regularization term to limit the weights of feature data;

[0025] Training the prediction model using the gradient descent method to gradually adjust the parameters of the prediction model.

[0026] In one embodiment, the optimization of the TTV angle parameters includes:

[0027] Determining the value range of the TTV angle parameters;

[0028] Traversing a parameter grid within the value range, and for each set of parameter combinations in the parameter grid, calculating the TTV value using the prediction model; the TTV angle parameters include the first angle parameter of the first support foot and the second angle parameter of the second support angle, and the parameter combination is composed of the first angle parameter and the second angle parameter;

[0029] In traversing the parameter grid, selecting the parameter combination that minimizes the TTV value as the optimal TTV angle parameter.

[0030] In one embodiment, the environmental feature data includes time domain features and frequency domain features in addition to process parameter features;

[0031] Among them, obtaining the environmental feature data includes:

[0032] Setting up a data acquisition device, which can collect process parameters, process parameters, and TTV values during the wafer thinning process in real time. The data acquisition device includes a PLC data real-time acquisition module and an external acquisition sensing module built into the wafer thinning processing equipment;

[0033] Performing feature extraction on the collected process parameters and process parameters to obtain time domain features, frequency domain features, and process parameter features.

[0034] In one embodiment, the PLC data real-time acquisition module built into the wafer thinning processing equipment is used to collect the process parameters and process parameters inside the wafer thinning processing equipment through the TCP / IP protocol;

[0035] An external acquisition and sensing module is used to implement A / D conversion of analog signals through a data acquisition card and encapsulate digital signals through the Modbus protocol.

[0036] In a second aspect, the present application further provides a control device for automatically adjusting the thinning thickness of a wafer. The device includes:

[0037] An acquisition module is used to acquire the initial TTV value measured by an on-line TTV measurement module during the wafer thinning process. Among them, the number of measurement points collected by the on-line TTV measurement module is less than that collected by the FRT measurement device, and the sampling resolution is less than the sampling resolution of the FRT measurement device. The FRT measurement device measures the TTV value of the same wafer after thinning is completed.

[0038] A compensation module is used to obtain a corrected TTV value based on the initial TTV value using a TTV compensation model. Among them, the TTV compensation model is at least trained using the TTV values measured by the on-line TTV measurement module and the FRT measurement device for the same wafer.

[0039] An optimization module is used to determine the optimal TTV angle parameter through a pre-trained prediction model if the corrected TTV value is not within the set specifications. The environmental characteristic data during the wafer thinning process is used as a constant of the prediction model, the TTV angle parameter is used as the optimization target of the prediction model, and the expected TTV value is used as the output target of the prediction model to traverse the feasible set of the TTV angle to obtain the optimal angle parameter. Among them, the environmental characteristic data includes process parameter characteristics.

[0040] An adjustment module is used to control the angle of the grinding disc in the wafer thinning processing equipment to be adjusted to the optimal TTV angle parameter to act on the next wafer thinning thickness processing.

[0041] In a third aspect, the present application further provides a wafer thinning system. The wafer thinning system includes:

[0042] A wafer thinning processing equipment is used to thin the wafer.

[0043] An on-line TTV measurement module is used to collect the thickness data of the wafer in real time during the wafer thinning process by the wafer thinning processing equipment.

[0044] A control module includes a memory and a processor. The memory stores a computer program. When the processor executes the computer program, it implements the control method for automatically adjusting the wafer thinning thickness in the first aspect.

[0045] The above control method for automatically adjusting the wafer thinning thickness obtains the initial TTV value in real time using an on-line TTV measurement module during the wafer thinning process; corrects the initial TTV value using a TTV compensation model trained based on the measurement data of the on-line TTV measurement module and the FRT measurement device to obtain a corrected TTV value closer to the FRT measurement result; if the corrected TTV value exceeds the set specification, determines the optimal TTV angle parameter through a pre-trained prediction model. The prediction model takes environmental characteristic data (including process parameter characteristics) as constants and the TTV angle parameter as the optimization target, and obtains the optimal angle parameter by traversing the feasible set; adjusts the angle of the grinding disk in the wafer thinning processing equipment to the optimal TTV angle parameter to optimize the next wafer thinning process. This method can achieve real-time precise control and continuous optimization of the wafer thinning process, improving production efficiency and product quality. Description of the Drawings

[0046] Figure 1 It is a flowchart of the control method for automatically adjusting the wafer thinning thickness in an embodiment;

[0047] Figure 2 It is a flowchart of obtaining the corrected TTV value using the TTV compensation model based on the initial TTV value in an embodiment;

[0048] Figure 3 It is a structural diagram of the TTV correction value obtained using the TTV compensation model in an embodiment;

[0049] Figure 4 It is a flowchart of training the TTV compensation model in an embodiment;

[0050] Figure 5 It is a flowchart of determining the high-low resolution mapping model in an embodiment;

[0051] Figure 6 It is a structural diagram of the high-low resolution mapping model in an embodiment;

[0052] Figure 7 It is a flowchart of training the prediction model in an embodiment;

[0053] Figure 8 It is a flowchart of the optimization of the TTV angle parameter in an embodiment;

[0054] Figure 9 It is a flowchart of the wafer thinning run-to-run control in an embodiment;

[0055] Figure 10 It is a device diagram of the control device for automatically adjusting the wafer thinning thickness in an embodiment;

[0056] Figure 11It is a structural diagram of a wafer thinning system in an embodiment. Detailed implementation manners

[0057] In order to make the objectives, technical solutions and advantages of the present application more clear and understandable, the present application will be further described in detail below with reference to the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are only used to explain the present application and are not used to limit the present application.

[0058] In one embodiment, as Figure 1 shown, a control method for automatically adjusting the wafer thinning thickness is provided. This method is applied to the control module in the wafer thinning system. The wafer thinning system further includes an on-line TTV measurement module and a wafer thinning processing device. The method includes the following steps:

[0059] Step 101: Obtain the initial TTV value measured by the on-line TTV measurement module during the wafer thinning process; wherein, the number of measurement points collected by the on-line TTV measurement module is less than that collected by the FRT measurement device, and the sampling resolution is less than the sampling resolution of the FRT measurement device. The FRT measurement device measures the TTV value of the same wafer after thinning is completed;

[0060] A wafer is a thin slice made of high-purity semiconductor material used in semiconductor manufacturing and is the basic material for manufacturing semiconductor devices such as integrated circuits and chips. Common wafer materials include silicon wafers, silicon carbide wafers, and sapphire wafer slices, etc.

[0061] The on-line TTV measurement module can obtain the initial TTV value of the wafer in real time during the thinning process, and both the number of measurement points and the sampling resolution are relatively low. When measuring in real time on the production line, it is necessary to quickly obtain data to detect obvious problems in a timely manner. If the number of measurement points is too large or the resolution is too high, it will greatly increase the measurement time and calculation amount, affecting production efficiency. Therefore, the on-line TTV measurement module improves the measurement speed by reducing the number of measurement points and the sampling resolution.

[0062] The FRT measurement device performs a more detailed and accurate TTV measurement on the same wafer after the wafer thinning is completed. Therefore, higher requirements are placed on the accuracy and comprehensiveness of the wafer measurement. The FRT measurement device can collect more measurement points and has a higher sampling resolution, and thus can provide more accurate TTV value information. Among them, the TTV value refers to the wafer thickness change value and is used to evaluate the quality of the wafer after thinning processing.

[0063] Step 102: Based on the initial TTV value, use the TTV compensation model to obtain the corrected TTV value; wherein, the TTV compensation model is at least trained using the TTV values measured by the on-line TTV measurement module and the FRT measurement device for the same wafer;

[0064] During the process of training the TTV compensation model, an online TTV measurement module and an FRT measurement device can be used to measure multiple wafers and collect TTV data of multiple wafers. Further, the TTV data of these wafers includes the TTV values of low resolution and few measurement points measured online, and the TTV values of high resolution and a large number of measurement points measured by FRT. By comparing these two measurement results, the TTV compensation model can learn the deviation between the online measurement value and the actual high-precision FRT measurement value.

[0065] Further, for the training of the TTV compensation model, paired online and FRT measurement values can be used as training data and input into a neural network. By adjusting the parameters of the TTV compensation model, the output of the TTV compensation model can be made as close as possible to the high-precision TTV value measured by FRT, thereby minimizing the error between the predicted value and the true value. During the training process, the TTV compensation model can learn how to map the low-resolution TTV value measured online to the high-resolution TTV value measured by FRT. After the TTV compensation model is trained, the initial TTV value obtained by online measurement can be input into the TTV compensation model, and the TTV compensation model can output the corrected TTV value to make it closer to the true high-precision FRT measurement result.

[0066] Step 103: If the corrected TTV value is not within the set specifications, determine the optimal TTV angle parameter through a pre-trained prediction model; wherein, the environmental feature data during the wafer thinning process is used as a constant of the prediction model, the TTV angle parameter is used as the optimization target of the prediction model, and the expected TTV value is used as the output target of the prediction model to traverse the feasible set of TTV angles to obtain the optimal angle parameter; wherein, the environmental feature data includes process parameter features.

[0067] The prediction model can be trained based on a large amount of historical data and actual measurement results. The prediction model can predict the corresponding TTV value according to the input environmental feature data and TTV angle parameter. Among them, the environmental feature data can be used as a constant of the prediction model. The environmental feature data can include process parameter features, such as temperature, pressure, rotation speed, etc. during the thinning process. The environmental feature data is relatively stable during the production process and has an important impact on the TTV value.

[0068] The TTV angle parameter can be the optimization target of the prediction model. It is necessary to adjust the TTV angle parameter so that the TTV value output by the prediction model reaches the expected target. The specific operation is to traverse the feasible set of TTV angles, that is, to conduct a comprehensive search and attempt within the possible value range of the TTV angle parameter, calculate the predicted TTV value corresponding to each possible angle parameter combination, and then screen out the optimal angle parameter combination that makes the predicted TTV value closest to the expected value among each possible angle parameter combination.

[0069] Step 104: Adjust the angle of the polishing pad in the wafer thinning processing equipment to the optimal angle parameter of TTV, so as to act on the next wafer thinning thickness processing.

[0070] During the wafer thinning process, the angle of the polishing pad can affect the TTV of the wafer. After determining the optimal angle parameter of TTV through the prediction model, it is necessary to adjust the angle of the polishing pad in the wafer thinning processing equipment accordingly. Further, adjust the angle of the polishing pad to the optimal angle parameter to ensure that the expected TTV value can be achieved during the next wafer thinning thickness processing, meeting the quality requirements of wafer thinning.

[0071] In this embodiment, during the wafer thinning process, the method uses a TTV compensation model trained based on the measurement data of the on-line TTV measurement module and the FRT measurement equipment to correct the initial TTV value, and obtains a corrected TTV value closer to the FRT measurement result; if the corrected TTV value exceeds the set specification, the optimal TTV angle parameter is determined through a pre-trained prediction model. The prediction model takes the environmental characteristic data (including process parameter characteristics) as constants and the TTV angle parameter as the optimization target, and obtains the optimal angle parameter by traversing the feasible set; adjust the angle of the polishing pad in the wafer thinning processing equipment to the optimal TTV angle parameter to optimize the next wafer thinning processing. This method can realize real-time precise control and continuous optimization of the wafer thinning process, improving production efficiency and product quality.

[0072] In one embodiment, as Figure 2 shown, obtaining the corrected TTV value by using the TTV compensation model based on the initial TTV value includes the following steps:

[0073] Step 201: Input the initial TTV value into the high-low resolution mapping model to obtain the virtual TTV value measured by the virtual FRT measurement equipment; the high-low resolution mapping model is obtained by mapping at least the TTV values measured by the on-line TTV measurement module and the FRT measurement equipment for the same wafer;

[0074] The high-low resolution mapping model can be obtained by mapping and training the TTV values measured by the on-line TTV measurement module and the FRT measurement equipment for the same wafer. It should be noted that the on-line TTV measurement module can obtain the initial TTV value of the wafer in real time during the thinning process, but due to the small number of measurement points and the low sampling resolution, the accuracy of the initial TTV value is limited. The FRT measurement equipment can perform more detailed and accurate measurements on the same wafer after the wafer thinning is completed, with a large number of measurement points collected and a high sampling resolution, and can provide a more accurate TTV value.

[0075] Furthermore, the high-low resolution mapping model can be trained using the relationship between the two measurement results to learn how to map the low-resolution TTV value measured online to the high-resolution FRT measurement value. Exemplarily, the initial TTV value can be input into the high-low resolution mapping model, and the high-low resolution mapping model can output a virtual high-precision and high-resolution TTV value according to the previously learned mapping relationship. This virtual TTV value can be closer to the true high-precision measurement result, which helps to more accurately evaluate and control the flatness of the wafer during the production process.

[0076] Step 202: Concatenate the initial TTV value and the virtual TTV value and input them into the TTV compensation model to obtain the corrected TTV value.

[0077] The initial TTV value is the low-resolution data obtained by the online TTV measurement module, and the virtual TTV value is the high-resolution data obtained by mapping the initial TTV value through the high-low resolution mapping model. Concatenating the two TTV values can form a more comprehensive input vector.

[0078] It should be noted that the TTV compensation model can adopt a neural network architecture. Through forward propagation and backpropagation algorithms, loss functions such as mean square error are used to optimize the model parameters. During the training process, the TTV compensation model can continuously adjust the weights and biases to minimize the difference between the predicted value and the true FRT measurement value.

[0079] Input the concatenated initial TTV value and virtual TTV value into the trained TTV compensation model, and the TTV compensation model can output the corrected TTV value. This corrected TTV value can be closer to the true high-precision FRT measurement result, which helps to more accurately evaluate and control the flatness of the wafer during the production process.

[0080] In this embodiment, this method can make the online measurement result closer to the actual high-precision FRT measurement result.

[0081] In one embodiment, as Figure 3 shown, the TTV value measured online can be the high-resolution virtual TTV value obtained by mapping through the high-low resolution mapping model. The high-resolution virtual TTV value and the low-resolution TTV value are concatenated, and after concatenating the initial TTV value and the virtual TTV value, they are input into the TTV compensation model, and the TTV compensation model combines these data and outputs the corrected TTV value.

[0082] In one embodiment, as Figure 4 shown, training the TTV compensation model includes the following steps:

[0083] Step 401: Establish a TTV compensation model based on a neural network;

[0084] Collect the on-line TTV measurement values and FRT measurement values of the same wafer. The data of the measured TTV values can be used to train the TTV compensation model. Further, to construct the TTV compensation model, the data needs to be cleaned, normalized, and divided into a training set and a test set. Select a suitable neural network architecture, such as a multi-layer perceptron or a convolutional neural network, define the number of neurons in the input layer, hidden layer, and output layer, select the activation function and loss function, and use an optimization algorithm to adjust the model parameters so that the corrected TTV value output by the TTV compensation model is as close as possible to the true FRT measurement value. After training, the performance of the TTV compensation model can be evaluated using the test set, and the TTV compensation model can be adjusted and optimized if necessary.

[0085] Step 402: Stitch the first-resolution data and the second-resolution data of the same wafer to obtain stitched data; wherein, for the same wafer, the first-resolution data includes at least the TTV values measured by the on-line TTV measurement module, and the second-resolution data includes at least the TTV values measured by the FRT measurement device;

[0086] The first-resolution data includes at least the TTV values measured by the on-line TTV measurement module. These data are obtained in real time during the thinning process, with a small number of measurement points and a low sampling resolution, mainly used for quickly detecting the flatness of the wafer. The second-resolution data includes at least the TTV values measured by the FRT measurement device. These data are obtained after the wafer thinning is completed by conducting more detailed and precise measurements on the same wafer, with a large number of measurement points and a high sampling resolution, capable of providing more accurate and detailed wafer flatness information.

[0087] Stitching these two types of resolution data can combine their advantages, making use of both the real-time nature of on-line measurement and the high precision of FRT measurement, providing a richer and more comprehensive data basis for subsequent analysis and modeling. Exemplarily, the low-resolution TTV values measured on-line and the high-resolution TTV values measured by FRT can be matched and stitched according to the unique identifier (wafer ID) of the wafer to form a complete data set for training the TTV compensation model, thereby improving the accuracy and reliability of the TTV compensation model.

[0088] Step 403: Use the stitched data as the input of the initial TTV compensation model, and use the TTV values in the second-resolution data used for stitching as the output to train the TTV compensation model.

[0089] The spliced data is input into the TTV compensation model, and the second-resolution data with high precision, that is, the TTV value measured by FRT, is used as the target output. Further, by adjusting the parameters of the TTV compensation model, the TTV compensation model can learn how to relate the low-resolution online measurement value and the high-resolution FRT measurement value, and then can output a more accurate corrected TTV value based on the initial TTV value and the virtual TTV value measured online, making it closer to the true high-precision FRT measurement result.

[0090] In this embodiment, the method can generate a high-precision FRT measurement result close to the actual value by establishing a TTV compensation model for the initial TTV value measured online, which helps to more accurately evaluate and control the flatness of the wafer.

[0091] In one embodiment, as Figure 5 shown, determining the high-low resolution mapping model includes the following steps:

[0092] Step 501: Establish an encoder f from the second-resolution data to the first-resolution data, and establish a decoder g from the first-resolution data to the second-resolution data;

[0093] The second-resolution data may refer to the high-resolution, multi-measurement-point TTV values collected by the FRT measurement device. The first-resolution data refers to the low-resolution, small-number-of-measurement-point TTV values collected by the online TTV measurement module.

[0094] The function of the encoder f can be to convert the high-resolution second-resolution data into low-resolution first-resolution data. This can be achieved by operations such as downsampling, feature extraction, or dimensionality reduction on the high-resolution data. Exemplarily, convolutional layers or pooling layers in a neural network can be used to extract the main features of the high-resolution data and map them into a low-resolution space. Further, the encoder f can learn how to extract key information from the high-resolution data and represent it in the form of low-resolution data.

[0095] The decoder g converts the low-resolution first-resolution data back into high-resolution second-resolution data. This process is relatively complex and requires upsampling, transposed convolution, or other generative methods to restore the high-resolution details of the data. For example, transposed convolution layers or pixel shuffle operations in a neural network can be used to gradually magnify the low-resolution data and fill in the details to make it close to the original high-resolution data. Among them, the goal of the decoder g is to reconstruct the high-resolution data as accurately as possible while retaining the main features in the low-resolution data.

[0096] Step 502: Train the encoder f and the decoder g based on the first-resolution data and the second-resolution data of the same wafer. Based on the trained encoder f and decoder g, obtain the high-low resolution mapping model.

[0097] After sufficient training, the encoder f can learn how to extract key information from high-resolution data and map it to the low-resolution space, while the decoder g can recover the details of high-resolution data from low-resolution data. Combining the trained encoder f and decoder g can form the high-low resolution mapping model. That is, when the online measured low-resolution TTV value is obtained, the high-low resolution mapping model can be used to convert the low-resolution TTV value into a virtual high-resolution TTV value.

[0098] Exemplarily, for the first-resolution data and the second-resolution data of the same wafer, a data mapping model between high and low resolutions can be constructed. Assume that there are pieces of first-resolution data, and the data is .

[0099] There are pieces of second-resolution data, , and there is . For this wafer, the first-resolution data has fewer points and belongs to a low-dimensional sample; the second-resolution data has more points and belongs to a high-dimensional sample. For the mapping management between high- and low-dimensional samples, establish to encoder f , and reverse train to decoder g , and obtain the high-low resolution mapping model.

[0100] Furthermore, extract the encoder g and the TTV compensation model in the trained high-low resolution mapping model to form an online low-resolution TTV correction model. Use the online measured as the encoder input to obtain the virtual high-dimensional input , and input and into the compensation model to obtain the corrected TTV value.

[0101] In this embodiment, the trained encoder and decoder can accurately convert the online measured low-resolution data into high-resolution virtual FRT measurement data, thereby improving the accuracy and reliability of online measurement and making it closer to the actual high-precision FRT measurement result.

[0102] In one embodiment, as Figure 6As shown in the figure, it mainly includes an encoder f, a decoder g, and a TTV compensation model. The high-resolution data is converted into low-resolution data by the encoder f, and the decoder g restores the low-resolution data to high-resolution data. The encoder f and the decoder g are trained based on the high-resolution data and the low-resolution data of the same wafer. Based on the trained encoder f and decoder g, a high-low resolution mapping model is obtained.

[0103] In one embodiment, as Figure 7 shown, the steps of training the prediction model are as follows:

[0104] Step 701: Obtain the sample features of the sample wafer. For any sample wafer, its sample features include environmental feature data, TTV angle parameters, and TTV values.

[0105] For any sample wafer, its sample features include environmental feature data, TTV angle parameters, and TTV values. Among them, the environmental feature data may include various process parameters during wafer thinning, such as temperature, pressure, rotation speed, etc. The process parameters have an important impact on the thinning effect of the wafer. The TTV angle parameter can affect the thinning thickness and surface flatness of the wafer. The TTV value may refer to the initial TTV value and the final high-precision TTV value measured by the online TTV measurement module and the FRT measurement device respectively, and the more accurate wafer flatness data obtained after being corrected by the TTV compensation model. These data together constitute the sample features of the sample wafer. By analyzing and modeling the sample features of the sample wafer, the TTV of the wafer can be predicted and controlled more accurately, the thinning process can be optimized, and the product quality can be improved.

[0106] Step 702: Based on the sample features, construct a prediction model.

[0107] Collect the historical sample features of multiple sample wafers, and sort and preprocess these sample feature data, including operations such as data cleaning, normalization, and feature selection, to ensure the quality and consistency of the data. Further, an appropriate algorithm can be selected to construct a prediction model, using the environmental feature data and TTV angle parameters in the sample features as input variables, and the corrected TTV value as the target output variable. By adjusting the parameters of the prediction model, the prediction model can learn the mapping relationship between the input variables and the output variables. Further, after the prediction model is trained, a part of the sample data that did not participate in the training can be used to verify and test the prediction model, and evaluate the accuracy and generalization ability of the prediction model. For example, the prediction performance of the prediction model can be measured by calculating metrics such as root mean square error and mean absolute error.

[0108] Exemplarily, based on the sample features of the sample wafer and the corrected TTV value, a wafer thinning TTV prediction model is established. The constructed prediction model can be model Among them, is the eigenvector, is the target variable, and

[0109] are the model parameters.

[0109] Step 703: Introduce the L1 regularization coefficient to shrink the coefficients of the non-associated feature data in the prediction model;

[0110] The training process of the prediction model is based on a large number of sample feature data. Among the sample feature data, some sample feature data has a relatively close relationship with the TTV value and is the key factor affecting the TTV value; while some sample feature data has little relationship with the TTV value or even no direct association, that is, non-associated feature data.

[0111] The L1 regularization coefficient can be referred to as adding a penalty term to the loss function, making the prediction model tend to shrink the coefficients of these non-associated feature data to zero during the training process. Features that do not contribute significantly to the TTV value prediction will be automatically excluded from the prediction model, thus achieving the purpose of feature selection. For example, during the training process of the prediction model, if the influence of a certain environmental feature data (such as temperature) on the TTV value is negligible, after introducing the L1 regularization, the coefficient corresponding to the temperature feature will be shrunk to zero (to the extent that it does not affect the prediction model effect), and the temperature feature will no longer participate in the prediction calculation of the prediction model.

[0112] Step 704: Introduce the L2 regularization term to limit the weights of the feature data;

[0113] The L2 regularization term adds a penalty term proportional to the sum of the squares of the weights of the feature data to the loss function, making the prediction model tend to reduce the absolute value of the weights during the training process. The weights of the features with less influence on the TTV value prediction will be further reduced, thereby reducing the interference to the output of the prediction model. For example, during the training process of the prediction model, if the influence of a certain environmental feature data (such as pressure) on the TTV value is relatively small, after introducing the L2 regularization term, the weight corresponding to the pressure feature will be restricted within a small range to avoid its excessive influence on the output of the prediction model.

[0114] Step 505: Use the gradient descent method to train the prediction model to gradually adjust the parameters of the prediction model.

[0115] Gradient descent is an iterative optimization algorithm. Its core idea is to calculate the gradient of the loss function with respect to the parameters of the prediction model, update the parameters in the opposite direction of the gradient, thereby gradually reducing the value of the loss function and making the prediction results of the prediction model closer to the actual values. During training, a suitable loss function is defined, such as the mean squared error (MSE), to measure the difference between the predicted TTV value of the prediction model and the actual TTV value. Further, the parameters of the prediction model are initialized, usually using random initialization or zero initialization. The sample features in the training dataset are input into the prediction model, the predicted output of the prediction model is calculated, and the loss value under the current parameters is calculated according to the loss function. The gradient of the loss function with respect to each parameter is calculated through the backpropagation algorithm, and the parameters of the prediction model are updated according to the learning rate and the gradient value. This process is repeated multiple times until the value of the loss function converges to a small range or reaches the preset maximum number of iterations.

[0116] Training the prediction model using gradient descent can effectively optimize the parameters of the model and improve the prediction accuracy and stability of the prediction model. For example, in the prediction model of wafer thinning, by adjusting the parameters through gradient descent, the prediction model can more accurately capture the complex relationship between the environmental feature data, the TTV angle parameter, and the TTV value.

[0117] Exemplarily, a training dataset is collected, which contains the input feature vector x and the target variable y. A loss function J(β) is defined, usually composed of a data fitting error term and a regularization term, such as L1 and L2 regularization terms, to prevent overfitting and improve the stability of the prediction model. The form of the loss function is:

[0118] .

[0119] where m is the number of samples, y i is the true value of the i-th sample, x i is the feature vector of the i-th sample, β is the model parameter, λ 1 and λ 2 are the regularization coefficients.

[0120] The training process is as follows: Initialize the model coefficient β, usually using a zero vector for initialization. Traverse all coefficients β, and successively take a single coefficient β j as a variable, lock the other coefficients as the results of the previous calculation and treat them as constants, and find the optimal solution of β j under the current conditions. At the k-th iteration, that is, when updating the weight coefficient it needs to satisfy:

[0121] .

[0122] In each iteration, calculating the gradient of the loss function with respect to each parameter β j involves calculating the derivative of the data fitting error and the derivative of the regularization term:

[0123] .

[0124] Update each parameter β j according to the gradient value, making it gradually approach the optimal value. Repeat the above process iteratively until the value of the loss function converges to a stable state or reaches the preset maximum number of iterations.

[0125] In this embodiment, the method can effectively remove the interference of non - related features, simplify the structure of the prediction model, limit the feature weights, avoid overfitting, and improve the stability and generalization ability of the prediction model through the combination of L1 regularization and L2 regularization. The trained prediction model can more accurately predict the TTV value of the wafer according to the input environmental feature data and TTV angle parameters.

[0126] In one embodiment, as Figure 8 shown, the optimization of the TTV angle parameter includes the following steps:

[0127] Step 801: Determine the value range of the TTV angle parameter;

[0128] According to the physical limitations and process requirements of the wafer thinning processing equipment, the maximum and minimum values of the TTV angle parameter in actual operation can be determined. The physical limitations can be provided by the manufacturer to ensure the safe operation of the physical limitations and the stability of the process. For example, the angle adjustment mechanism of the grinding disk may only be physically adjustable within a certain range, and exceeding this range may cause equipment damage or abnormal operation.

[0129] Different wafer materials and thinning targets may require different angle settings. The process requirements come from actual production experience and a large amount of experimental data. Historical data and experience in actual production can also be referred to for further refinement and adjustment of the theoretically determined value range to ensure that in actual applications, the value of the TTV angle parameter can meet both the physical limitations of the equipment and the optimization requirements of the process, thus providing a reliable basis for subsequent model optimization and process parameter adjustment.

[0130] Step 802: Traverse the parameter grid within the value range. For each set of parameter combinations in the parameter grid, use the prediction model to calculate the TTV value; the TTV angle parameter includes the first angle parameter of the first support foot and the second angle parameter of the second support angle, and the parameter combination is composed of the first angle parameter and the second angle parameter;

[0131] The TTV angle parameters include the first angle parameter of the first support foot (the first angle parameter can be ), and the second angle parameter of the second support foot (the second angle parameter can be ). The parameter combination can be composed of the first angle parameter and the second angle parameter.

[0132] According to the physical limitations and process requirements of the device, determine and value ranges, for example range can be [0°, 10°], range can be [5°, 15°]. Further, divide the range of this and the range of into multiple equally spaced values respectively to generate a parameter grid. Exemplarily, the range of can be divided into 0°, 2°, 4°, 6°, 8°, 10°, and the range of can be divided into 5°, 8°, 11°, 14°. The parameter grid will contain 6×4 = 24 different parameter combinations.

[0133] Step 803: In traversing the parameter grid, select the parameter combination that minimizes the TTV value as the optimal TTV angle parameter.

[0134] Exemplarily, assume that at a certain grid point, the TTV value corresponding to the parameter combination ( =5°, =10°) is 2.0, and at another grid point, the TTV value corresponding to the parameter combination ( =6°, =12°) is 1.8, then the latter may be a better parameter combination.

[0135] In this embodiment, by comparing the TTV values of all combinations, select the parameter combination that minimizes the TTV value as the optimal solution. This method can comprehensively cover possible parameter combinations, avoid missing the optimal solution, and improve process accuracy and product quality.

[0136] In one embodiment, the environmental feature data includes time domain features and frequency domain features in addition to process parameter features; among them, obtaining the environmental feature data includes: building a data acquisition device, the data acquisition device can collect process parameters, process parameters and TTV values during the wafer thinning process in real time, and the data acquisition device includes a PLC data real-time acquisition module and an external acquisition sensing module built in the wafer thinning processing equipment; perform feature extraction on the collected process parameters and process parameters to obtain time domain features, frequency domain features and process parameter features.

[0137] During the wafer thinning process, the acquisition of environmental characteristic data is achieved by setting up a dedicated data acquisition device. The data acquisition device can collect various key information during the wafer thinning process in real time, providing a basis for subsequent data analysis and model construction.

[0138] The data acquisition device includes a built-in PLC data real-time acquisition module and an external acquisition sensing module. Among them, the PLC data real-time acquisition module is integrated inside the wafer thinning processing equipment and is responsible for obtaining the internal process parameters during equipment operation, such as temperature, pressure, rotation speed, etc. The internal process parameters are usually directly measured by sensors inside the equipment and can reflect the operating state and process conditions of the equipment. The external acquisition sensing module is installed outside the equipment and is used to collect other process parameters, such as current value, vacuum pressure, vibration frequency, temperature in the grinding area, etc. The process parameters can be measured by additional sensors, such as an acceleration sensor for detecting the vibration of the equipment and an ammeter for collecting the current signal during the processing.

[0139] The collected raw data needs to be preprocessed to ensure the quality and consistency of the data. This includes data cleaning to remove outliers and noise, data normalization to scale data in different ranges to a unified range, and data alignment to ensure the temporal consistency of data from different sources.

[0140] The preprocessed data will undergo feature extraction to obtain time-domain features, frequency-domain features, and process parameter features. Time-domain features include mean, variance, peak value, etc., which can reflect the distribution and change trend of data over time; frequency-domain features can be obtained through methods such as Fourier transform, such as power spectral density, frequency peak value, etc., revealing the distribution and periodic changes of data in the frequency domain. The extracted time-domain features, frequency-domain features, process parameter features, and surface profile result TTV form the sample features of each wafer. ,

[0141] .

[0142] Among them and are the time-domain feature and frequency-domain feature of the timing parameter respectively; is the process parameter feature; and are the angle adjustment values of TTV.

[0143] In this embodiment, by comprehensively acquiring and processing environmental characteristic data, rich and accurate information can be provided for subsequent model training and prediction. The extraction of time-domain features and frequency-domain features can capture the change laws of data in time and frequency, helping to discover potential process problems and optimization points.

[0144] In one embodiment, the PLC data real-time acquisition module built in the wafer thinning processing equipment is used to acquire the process parameters and process variables inside the wafer thinning processing equipment through the TCP / IP protocol;

[0145] The external acquisition sensing module is used to perform A / D conversion of analog signals through a data acquisition card and encapsulate digital signals through the Modbus protocol.

[0146] The PLC data real-time acquisition module built in the wafer thinning processing equipment acquires the process parameters and process variables inside the equipment through the TCP / IP protocol. As the core device of industrial automation, the PLC can monitor and control the operating status of the equipment in real time. During the wafer thinning process, the PLC module is responsible for collecting various process parameters inside the equipment, such as temperature, pressure, rotational speed, feed rate, etc. The process parameters can be measured by sensors inside the equipment. The PLC module converts these analog signals into digital signals and transmits them to the data storage server or control system through the TCP / IP protocol to achieve real-time monitoring of the equipment operating status and data recording.

[0147] The external acquisition sensing module is used to perform A / D conversion of analog signals through a data acquisition card and encapsulate digital signals through the Modbus protocol. The external acquisition sensing module usually includes various sensors, such as vibration sensors, sound sensors, current sensors, etc. These sensors are used to monitor the environmental parameters outside the equipment and the physical quantities during equipment operation. The data acquisition card performs A / D conversion on these analog signals, that is, converts continuous analog signals into discrete digital signals for easy computer processing and analysis. The converted digital signals are encapsulated through the Modbus protocol. Modbus is a protocol widely used in industrial communication that allows data exchange between devices of different manufacturers. The encapsulated digital signals can be transmitted to the upper computer or other control systems for further processing and analysis.

[0148] In this embodiment, through the collaborative work of these two modules, comprehensive monitoring and data acquisition of the wafer thinning process can be achieved, providing a basis for subsequent data analysis, model construction, and process optimization.

[0149] In one embodiment, as Figure 9As shown in the figure, it is the run-to-run control process for wafer thinning. The historical data of wafers on the same thinning equipment is collected through the data acquisition module, including time-domain characteristics, frequency-domain characteristics, process parameter characteristics, and TTV data of the surface profile results. The time-series characteristics of the multi-modal data are extracted, and the characteristics are extracted after time-series noise reduction and frequency-domain transformation to convert the multi-source data into structured data. The online measurement and FRT high-resolution measurement data in the historical data are used to establish a TTV correction model. According to the surface profile result of the wafer, the corrected TTV predicted value is used as the TTV value of the surface profile result. When the processed TTV of the wafer is within the specification, the current process is continued to produce the next wafer. When the processed TTV of the wafer exceeds the specification, the optimal control is used to obtain the TTV angle adjustment value.

[0150] Using the collected wafer thinning data , a high-dimensional and non-linear regression prediction model for TTV is constructed, with RMS as the loss function and L1 and L2 regularization terms introduced. Using gradient descent to optimize the parameters of the prediction model, and obtain the current optimal prediction model when the training conditions are met. In actual operation, as the wafer data accumulates, the wafer data is dynamically updated, and a new prediction model is established.

[0151] When using the prediction model to predict the processed TTV of the next wafer, the structure of the prediction model remains unchanged. For the structured process parameters , the set values of the process parameters are kept unchanged without adjustment; for the time-domain and frequency-domain characteristics of the time-series parameters, since the process data of the next wafer to be predicted cannot be obtained temporarily because it has not been processed, the average value of the process data of the current nearest n wafers is used as the and of the wafer to be predicted. The grid search is used to optimize the angle adjustment parameters, and the and of the calculation result are used as the recommended TTV angle adjustment values.

[0152] In one embodiment, as Figure 9As shown, through real-time measurement, online data acquisition, model prediction, and parameter optimization, a closed-loop control process is formed, aiming to continuously improve the precision and quality of wafer thinning. The high- and low-resolution TTV correction models map the low-resolution online TTV measurement data to the high-resolution FRT data to improve the precision of online measurement. The IPG thickness measurement and online TTV measurement modules obtain the thickness and TTV values of the wafer in real time during the thinning process, and these data, together with the process parameters, are used to dynamically adjust the thinning process. Among them, IPG thickness measurement refers to using integrated process thickness measurement technology to measure the thickness of the wafer in real time during wafer thinning. The FRT measurement is carried out after the wafer thinning is completed, providing high-precision flatness data for final quality assessment and model training. According to the measurement results and process parameters of the current wafer, the optimal angle parameter is calculated, and the angle of the grinding disk is adjusted to optimize the subsequent wafer thinning process. The data acquisition system runs through the whole process, collecting various data to support model training and process control.

[0153] Based on the same concept, as Figure 10 shown, the present application also provides a control device for automatically adjusting the thinning thickness of a wafer, and the device includes:

[0154] An acquisition module 1001, configured to acquire the initial TTV value measured by the online TTV measurement module during the wafer thinning process; wherein, the number of measurement points collected by the online TTV measurement module is less than that collected by the FRT measurement device, and the sampling resolution is less than the sampling resolution of the FRT measurement device, and the FRT measurement device measures the TTV value of the same wafer after the thinning is completed;

[0155] A compensation module 1002, configured to obtain the corrected TTV value by using the TTV compensation model based on the initial TTV value; wherein, the TTV compensation model is at least trained by using the TTV values measured by the online TTV measurement module and the FRT measurement device for the same wafer;

[0156] An optimization module 1003, configured to, if the corrected TTV value is not within the set specification, determine the optimal TTV angle parameter through a pre-trained prediction model, use the environmental feature data during the wafer thinning process as a constant of the prediction model, use the TTV angle parameter as the optimization target of the prediction model, and use the expected TTV value as the output target of the prediction model to traverse the feasible set of the TTV angle to obtain the optimal angle parameter; wherein, the environmental feature data includes process parameter features;

[0157] An adjustment module 1004, configured to control the angle of the grinding disk in the wafer thinning processing equipment to be adjusted to the optimal TTV angle parameter to act on the next wafer thinning thickness processing.

[0158] In one embodiment, the compensation module 1002 uses the TTV compensation model based on the initial TTV value to obtain the corrected TTV value, specifically for: inputting the initial TTV value into the high-low resolution mapping model to obtain the virtual TTV value measured by the virtual FRT measurement device; the high-low resolution mapping model is obtained by mapping the TTV values measured by at least the online TTV measurement module and the FRT measurement device on the same wafer; splicing the initial TTV value and the virtual TTV value and inputting them into the TTV compensation model to obtain the corrected TTV value.

[0159] In one embodiment, the compensation module 1002 training the TTV compensation model includes: establishing the TTV compensation model based on a neural network; splicing the first resolution data and the second resolution data of the same wafer to obtain the spliced data; wherein, for the same wafer, the first resolution data at least includes the TTV value measured by the online TTV measurement module, and the second resolution data at least includes the TTV value measured by the FRT measurement device; using the spliced data as the input of the initial TTV compensation model, and using the TTV value in the second resolution data for splicing as the output to train the TTV compensation model.

[0160] In one embodiment, the compensation module 1002 determines the high-low resolution mapping model, specifically for: establishing an encoder f from the second resolution data to the first resolution data, and establishing a decoder g from the first resolution data to the second resolution data; training the encoder f and the decoder g based on the first resolution data and the second resolution data of the same wafer, and obtaining the high-low resolution mapping model based on the trained encoder f and decoder g.

[0161] In one embodiment, the optimization module 1003 trains the prediction model to obtain the sample features of the sample wafer. Among them, for any sample wafer, its sample features include environmental feature data, TTV angle parameters, and TTV values; based on the sample features, a prediction model is constructed; an L1 regularization coefficient is introduced to shrink the coefficients of the non-associated feature data in the prediction model; an L2 regularization term is introduced to limit the weights of the feature data; the gradient descent method is used to train the prediction model to gradually adjust the parameters of the prediction model.

[0162] In one embodiment, for the optimization of the TTV angle parameters, the optimization module 1003 is specifically used to determine the value range of the TTV angle parameters; traverse the parameter grid within the value range, and for each group of parameter combinations in the parameter grid, calculate the TTV value using the prediction model; the TTV angle parameters include the first angle parameter of the first support foot and the second angle parameter of the second support angle, and the parameter combination is composed of the first angle parameter and the second angle parameter; in the process of traversing the parameter grid, select the parameter combination that makes the TTV value the smallest as the optimal TTV angle parameter.

[0163] In one embodiment, the environmental feature data includes time domain features and frequency domain features in addition to process parameter features; wherein, the environmental feature data is acquired. The optimization module 1003 is specifically configured to set up a data acquisition device, which can collect process parameters, process parameters and TTV values during the wafer thinning process in real time. The data acquisition device includes a PLC data real-time acquisition module built in the wafer thinning processing equipment and an external acquisition sensing module; feature extraction is performed on the collected process parameters and process parameters to obtain time domain features, frequency domain features and process parameter features.

[0164] In one embodiment, the optimization module 1002, the PLC data real-time acquisition module built in the wafer thinning processing equipment, is used to collect the process parameters and process parameters inside the wafer thinning processing equipment through the TCP / IP protocol; the external acquisition sensing module is used to implement A / D conversion of analog signals through a data acquisition card and encapsulate digital signals through the Modbus protocol.

[0165] Based on the same concept, as Figure 11 shown, the present application also provides a wafer thinning system 1100, which includes: a wafer thinning processing equipment 1110 for thinning the wafer; an on-line TTV measurement module 1120 for collecting the thickness data of the wafer in real time during the wafer thinning process by the wafer thinning processing equipment; a control module 1130, including a memory 1131 and a processor 1132, the memory 1131 stores a computer program, and when the processor 1132 executes the computer program, it realizes the control method for automatically adjusting the wafer thinning thickness.

[0166] The above embodiments only represent several implementation manners of the present application, and their descriptions are relatively specific and detailed, but they should not be construed as limiting the patent scope of the present application. It should be noted that for those of ordinary skill in the art, without departing from the concept of the present application, several modifications and improvements can still be made, and these all belong to the protection scope of the present application. Therefore, the protection scope of the present application should be subject to the appended claims.

Claims

1. A control method for automatically adjusting the wafer thinning thickness, characterized in that: The method is applied to a control module in a wafer thinning system, wherein the wafer thinning system further comprises an online TTV measurement module and a wafer thinning processing device, and the method comprises: Obtaining an initial TTV value measured by the online TTV measurement module during wafer thinning; wherein the measurement points collected by the online TTV measurement module are smaller than the measurement points collected by the FRT measurement device, and the sampling resolution is smaller than the sampling resolution of the FRT measurement device, and the FRT measurement device is a TTV value measurement performed on the same wafer after thinning is completed; Based on the initial TTV value, a TTV compensation model is used to obtain a corrected TTV value; wherein the TTV compensation model is obtained by training the TTV value measured on the same wafer using at least the online TTV measurement module and the FRT measurement device; If the corrected TTV value is not within the set specification, the optimal TTV angle parameter is determined by a pre-trained prediction model; wherein the environmental characteristic data during wafer thinning is used as a constant of the prediction model, the TTV angle parameter is used as an optimization target of the prediction model, and the expected TTV value is used as an output target of the prediction model to perform a traversal of the feasible set of the TTV angle to obtain the optimal angle parameter; wherein the environmental characteristic data includes process parameter characteristics; The grinding disc angle in the wafer thinning processing equipment is controlled to be adjusted to the TTV optimal angle parameter to act on the next wafer thinning thickness processing.

2. The control method for automatically adjusting wafer thinning thickness according to claim 1, characterized in that: Based on the initial TTV value, a TTV compensation model is used to obtain a corrected TTV value, including: Inputting the initial TTV value into a high-low resolution mapping model to obtain a virtual TTV value measured by a virtual FRT measurement device; the high-low resolution mapping model is obtained by mapping TTV values ​​measured on the same wafer using at least the online TTV measurement module and the FRT measurement device; The initial TTV value and the virtual TTV value are spliced ​​and input into a TTV compensation model to obtain a corrected TTV value.

3. The control method for automatically adjusting wafer thinning thickness according to claim 2, characterized in that: Training the TTV compensation model includes: Establishing the TTV compensation model based on a neural network; Splicing the first resolution data and the second resolution data of the same wafer to obtain spliced ​​data; wherein, for the same wafer, the first resolution data at least includes the TTV value measured by the online TTV measurement module, and the second resolution data at least includes the TTV value measured by the FRT measurement device; The spliced ​​data is used as input of the initial TTV compensation model, and the TTV value in the second resolution data used for splicing is used as output to train the TTV compensation model.

4. The control method for automatically adjusting wafer thinning thickness according to claim 3, characterized in that: Determining the high and low resolution mapping models includes: Establishing an encoder f from the second resolution data to the first resolution data, and establishing a decoder g from the first resolution data to the second resolution data; The encoder f and the decoder g are trained based on the first resolution data and the second resolution data of the same wafer, and a high-low resolution mapping model is obtained based on the trained encoder f and decoder g.

5. The control method for automatically adjusting wafer thinning thickness according to claim 1, characterized in that: Training the prediction model includes: Acquire sample characteristics of the sample wafer, wherein for any sample wafer, the sample characteristics include environmental characteristic data, TTV angle parameters and TTV values; Based on the sample characteristics, construct a prediction model; Introducing an L1 regularization coefficient to shrink the coefficients of non-correlated feature data in the prediction model; Introduce L2 regularization term to limit the weight of feature data; The prediction model is trained using a gradient descent method to gradually adjust the parameters of the prediction model.

6. The control method for automatically adjusting wafer thinning thickness according to claim 1, characterized in that: The optimization of the TTV angle parameters includes: Determine the value range of TTV angle parameters; The parameter grid is traversed within the value range, and for each parameter combination in the parameter grid, the TTV value is calculated using the prediction model; the TTV angle parameter includes a first angle parameter of the first support foot and a second angle parameter of the second support angle, and the parameter combination consists of the first angle parameter and the second angle parameter In the traversal parameter grid, the parameter combination that minimizes the TTV value is selected as the optimal TTV angle parameter.

7. The control method for automatically adjusting wafer thinning thickness according to any one of claims 1 to 6, characterized in that: The environmental characteristic data includes not only the process parameter characteristics but also time domain characteristics and frequency domain characteristics; Wherein, obtaining the environmental characteristic data includes: Building a data acquisition device, which can collect process parameters, process parameters and TTV values ​​in real time during wafer thinning, and the data acquisition device includes a built-in PLC data real-time acquisition module of the wafer thinning processing equipment and an external acquisition sensor module; The collected process parameters and procedure parameters are subjected to feature extraction to obtain time domain features, frequency domain features and process parameter features.

8. The control method for automatically adjusting wafer thinning thickness according to claim 7, characterized in that: The PLC data real-time acquisition module built into the wafer thinning processing equipment is used to collect the process parameters and process parameters inside the wafer thinning processing equipment through the TCP / IP protocol; The external data acquisition sensor module is used to realize A / D conversion of analog signals through a data acquisition card and encapsulate digital signals through the Modbus protocol.

9. A control device for automatically adjusting the wafer thinning thickness, characterized in that: The device comprises: An acquisition module, used for acquiring an initial TTV value measured by an online TTV measurement module during wafer thinning; wherein the measurement points collected by the online TTV measurement module are smaller than the measurement points collected by the FRT measurement device, and the sampling resolution is smaller than the sampling resolution of the FRT measurement device, and the FRT measurement device is a TTV value measurement performed on the same wafer after thinning is completed; A compensation module, configured to obtain a corrected TTV value based on the initial TTV value using a TTV compensation model; wherein the TTV compensation model is obtained by training the TTV values ​​measured on the same wafer using at least the online TTV measurement module and the FRT measurement device; An optimization module, for determining the optimal TTV angle parameter through a pre-trained prediction model if the corrected TTV value is not within the set specification, using the environmental characteristic data in the wafer thinning process as a constant of the prediction model, using the TTV angle parameter as an optimization target of the prediction model, and using the expected TTV value as an output target of the prediction model to perform a traversal of the feasible set of the TTV angle to obtain the optimal angle parameter; wherein the environmental characteristic data includes process parameter characteristics; The adjustment module is used to control the grinding disc angle in the wafer thinning processing equipment to be adjusted to the TTV optimal angle parameter so as to act on the next wafer thinning thickness processing.

10. A wafer thinning system, characterized in that: The wafer thinning system comprises: Wafer thinning processing equipment, used for performing thinning processing on the wafer; An online TTV measurement module collects thickness data of the wafer in real time during the wafer thinning process of the wafer thinning equipment; A control module comprises a memory and a processor, wherein the memory stores a computer program, and the processor implements the steps of the control method for automatically adjusting the wafer thinning thickness as described in any one of claims 1 to 8 when executing the computer program.

Images