Wafer back-end-of-line online wat parameter prediction model training method and system

Abstract

The application discloses a wafer back-end process online WAT parameter prediction model training method and system, and the method comprises the following steps: obtaining historical measurement parameters and corresponding historical WAT parameters of multiple stations in a wafer back-end process, and inputting the parameters into a pre-constructed WAT parameter prediction model; the WAT parameter prediction model evaluates the importance of each measurement parameter through an attention mechanism and assigns corresponding attention weights; the correlation between each measurement parameter and the corresponding WAT parameter is calculated according to the attention weights and through a recursive modeling algorithm; and the key measurement parameters are gradually included in the WAT parameter prediction model in order of correlation strength for iterative training; until the model meets a preset performance index, a trained WAT parameter prediction model and a final key measurement parameter combination are obtained. The application can improve the iterative efficiency of the model and increase the prediction accuracy of the model.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of semiconductor processing manufacturing, and in particular to a wafer back-end-of-line (BEOL) online WAT parameter prediction model training method and system. BACKGROUND

[0002] With the rapid development of AI technology, the cooperation between semiconductors and AI technology has become increasingly close. Semiconductor mass production data provides rich and large training and testing materials for AI modeling, greatly facilitating the iterative optimization of AI models. However, in the research and development field, due to the characteristics of high value and low yield of data, it is difficult to rely on massive data to build a powerful model. Therefore, exploring and breaking through existing modeling methods has become the only way to promote the development of semiconductor and AI integration.

[0003] WAT (Wafer Acceptance Test) is an important station in wafer manufacturing, which is used to detect (also known as test) the process of each wafer product on the wafer that has been manufactured, to evaluate the quality and stability of the semiconductor manufacturing process, and to determine whether the wafer product meets the electrical specification requirements of the process technology platform. WAT data can be used as a quality certificate for wafer product delivery. In addition, WAT data can also reflect the actual production situation of the production line. By collecting and analyzing WAT data, the production line situation can be monitored, and the trend of production line changes can be judged to provide early warning for possible situations.

[0004] BEOL (Back End Of Line) Inline parameters are mainly real-time monitoring of metal interconnection layers and dielectric layers in the BEOL of semiconductor manufacturing, including thickness parameters (such as dielectric thickness, metal thickness, etc.), size parameters (such as lithography CD, etched CD, etc.), material parameters (contact metal liner material, metal type, etc.), and product parameters (such as Product ID / Suffix NO / Site No / Key Recipe, etc.). There is a complex nonlinear coupling relationship between BEOL Inline parameters and WAT parameters. Traditional deep learning modeling methods often lead to overfitting in small sample scenarios, especially in the new process development stage, which cannot capture the complex nonlinear relationship between parameters. For example, by directly modeling using a deep learning model (such as LSTM, CNN), the redundancy between parameters will interfere with the model accuracy. And when traditional attention mechanisms are directly used for industrial time series data, there is a lack of feature enhancement design specific to semiconductor process characteristics. SUMMARY

[0005] The application mainly aims to provide a wafer back-end process online WAT parameter prediction model training method and system, which screens site measurement parameters more relevant to the prediction results, thereby improving the prediction accuracy of the prediction model.

[0006] The technical solution adopted by the application is:

[0007] The application provides a wafer back-end process online WAT parameter prediction model training method, which comprises the following steps:

[0008] Obtain historical measurement parameters and corresponding historical WAT parameters of multiple sites in a wafer back-end process.

[0009] Input the historical measurement parameters and corresponding historical WAT parameters into a pre-constructed WAT parameter prediction model, which is a deep neural network based on an attention mechanism. The attention mechanism is used to evaluate the importance of each measurement parameter and assign corresponding attention weights. According to the attention weights and through a recursive modeling algorithm, the correlation between each measurement parameter and the corresponding WAT parameter is calculated. The key measurement parameters are gradually included in the WAT parameter prediction model in order of correlation strength for iterative training.

[0010] Perform accuracy evaluation on the WAT parameter prediction model after each iteration until the preset performance indicators are met, stop iteration, and obtain the trained WAT parameter prediction model and the final key measurement parameter combination.

[0011] According to the above technical solution, the deep neural network comprises an input layer, a feature enhancement layer, a hidden layer and an output layer. The feature enhancement layer introduces an attention mechanism, and the hidden layer introduces a recursive modeling algorithm.

[0012] According to the above technical solution, the key measurement parameter combination screening process is as follows:

[0013] Calculate the recursive value of the measurement parameter with the highest attention weight through the recursive modeling algorithm. Then, gradually increase the measurement parameters according to the attention weight size and calculate the corresponding recursive weight value. If the recursive weight value is greater than the preset threshold, the measurement parameter is retained, otherwise it is deleted. The key measurement parameters are screened out for subsequent iterative training of the WAT parameter prediction model.

[0014] According to the above technical solution, the feature enhancement layer adaptively evaluates the importance of different input parameters and the weight distribution ratio according to the actual correlation degree between the actual input measurement parameters.

[0015] According to the above technical solution, the method further comprises the following steps: preprocessing the historical measurement parameters and corresponding historical WAT parameters of multiple sites in a wafer back-end process to construct a training sample set, which is used for training the WAT parameter prediction model.

[0016] According to the above technical solution, the preprocessing specifically comprises: data cleaning and labeling are performed on historical measurement parameters and corresponding historical WAT parameters of multiple sites in the wafer back-end process.

[0017] According to the above technical solution, the historical measurement parameters of the multiple sites in the back-end process include thickness parameters, size parameters, material parameters and product parameters.

[0018] The application further provides a wafer back-end process online WAT parameter prediction method, comprising the following steps:

[0019] Obtaining online measurement parameters of multiple sites in a wafer back-end process;

[0020] The online measurement parameters are input into a WAT parameter prediction model, and the predicted WAT parameters are output; wherein the WAT parameter prediction model is obtained by training the wafer back-end process online WAT parameter prediction model training method according to the above technical solution.

[0021] The application further provides a wafer back-end process online WAT parameter prediction model training system, comprising:

[0022] A training data acquisition module is configured to acquire historical measurement parameters and corresponding historical WAT parameters of multiple sites in a wafer back-end process.

[0023] An iterative training module is configured to input the historical measurement parameters and corresponding historical WAT parameters into a pre-constructed WAT parameter prediction model, wherein the WAT parameter prediction model is a deep neural network based on an attention mechanism, which evaluates the importance of each measurement parameter and assigns corresponding attention weights through the attention mechanism, calculates the correlation between each measurement parameter and the corresponding WAT parameter according to the attention weights and through a recursive modeling algorithm, and gradually includes key measurement parameters in the WAT parameter prediction model for iterative training according to the order of correlation strength.

[0024] A model evaluation module is configured to evaluate the accuracy of the WAT parameter prediction model after each iteration until the preset performance indicators are met, stop iteration, obtain the trained WAT parameter prediction model and the final key measurement parameter combination.

[0025] The application further provides a computer storage medium having a computer program stored therein, which can be executed by a processor, and the computer program executes the wafer back-end process online WAT parameter prediction model training method according to the above technical solution.

[0026] The beneficial effects generated by the present application are: the WAT parameter prediction model constructed by the present application evaluates the importance of each measurement parameter through the attention mechanism and assigns the corresponding attention weight, then calculates the correlation between each measurement parameter and the corresponding WAT parameter according to the attention weight and through the recursive modeling algorithm, and then gradually includes the key measurement parameters in the WAT parameter prediction model for iterative training according to the order of correlation strength. The unexpected effect is that through the effective combination of attention mechanism and recursive modeling algorithm, after finding the weight relationship between the input measurement parameters, the factor (i.e. the input measurement parameter) is gradually increased according to the attention weight, thereby gradually widening the input dimension, screening out the measurement parameter combination most related to the WAT parameter prediction result, improving the iteration efficiency of the model and increasing the prediction accuracy of the model.

[0027] Further, in the key measurement parameter combination screening process, the recursive value of the measurement parameter with the highest attention weight is calculated through the recursive modeling algorithm, and then the measurement parameters are gradually increased according to the attention weight and the corresponding recursive weight value is calculated. If the recursive weight value is greater than the preset threshold, the measurement parameter is retained, otherwise it is deleted. It can be seen that through threshold screening, the input parameters unrelated to the output prediction value can be more effectively screened out, further improving the model prediction accuracy.

[0028] Further, by adding a feature enhancement layer based on the attention mechanism between the input layer and the hidden layer of the model, the relationship between each input parameter can be found in advance, and the more important input parameters are assigned higher attention weights. The feature enhancement layer can adaptively evaluate the importance of different input parameters and the weight distribution ratio according to the actual correlation degree between the actual input measurement parameters, providing a calculation basis for the subsequent recursive modeling method.

[0029] Further, the historical data is preprocessed through the pre-constructed AI filtering model, so as to effectively filter the abnormal values caused by measurement errors, equipment failures or other abnormal operations, and at the same time reduce the random noise in the data, so that the signal is clearer, which helps to improve the accuracy and efficiency of model training. Through the fine processing of input data, the model can better understand the data connotation.

[0030] Of course, implementing any product of the present application does not necessarily need to achieve all the advantages described above. BRIEF DESCRIPTION OF DRAWINGS

[0031] In order to more clearly illustrate the technical solutions in the embodiments of the present application or the prior art, the following will briefly introduce the drawings needed to be used in the embodiments or prior art description. Obviously, the drawings in the following description are some embodiments of the present application, and those skilled in the art can also obtain other drawings according to these drawings without creative labor.

[0032] Figure 1 This is a flowchart of a method for training an online WAT parameter prediction model in a wafer back-end process according to an embodiment of the present invention;

[0033] Figure 2 This is a model R of an embodiment of the present invention. 2 Training effect trend chart;

[0034] Figure 3 This is a schematic diagram of the structure of an online WAT parameter prediction model training system for wafer back-end processes according to an embodiment of the present invention;

[0035] Figure 4 This is a flowchart of an online WAT parameter prediction method for wafer back-end processes according to an embodiment of the present invention. Detailed Implementation

[0036] To make the objectives, technical solutions, and advantages of this invention clearer, the invention 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 merely illustrative and not intended to limit the invention.

[0037] It should be noted that the illustrations provided in the embodiments of the present invention are only schematic representations of the basic concept of the present invention. Therefore, the drawings only show the components related to the present invention and are not drawn according to the number, shape and size of the components in actual implementation. In actual implementation, the form, quantity and proportion of each component can be arbitrarily changed, and the layout of the components may also be more complex.

[0038] In this invention, it should also be noted that the terms "center," "upper," "lower," "left," "right," "vertical," "horizontal," "inner," and "outer," etc., indicate the orientation or positional relationship based on the orientation or positional relationship shown in the accompanying drawings. They are used only for the convenience of describing this application and simplifying the description, and do not indicate or imply that the device or element referred to must have a specific orientation, or be constructed and operated in a specific orientation. Therefore, they should not be construed as limitations on this application. Furthermore, the terms "first" and "second" are used only for descriptive and distinguishing purposes and should not be construed as indicating or implying relative importance.

[0039] Furthermore, it should be noted that the features of the various embodiments of the present invention can be combined or integrated in whole or in part, and as those skilled in the art will understand, they can interact and operate in different ways. Each embodiment can be implemented independently of each other or in association with one another.

[0040] like Figure 1 As shown, the online WAT parameter prediction model training method for wafer back-end processes according to an embodiment of the present invention includes the following steps:

[0041] S101, acquire historical measurement parameters and corresponding historical WAT parameters of multiple sites in a wafer back-end process;

[0042] S102, input the historical measurement parameters and corresponding historical WAT parameters into a pre-constructed WAT parameter prediction model for training, the WAT parameter prediction model being a deep neural network based on an attention mechanism, which evaluates the importance of each measurement parameter and assigns corresponding attention weights through the attention mechanism, calculates the correlation between each measurement parameter and the corresponding WAT parameter according to the attention weights and through a recursive modeling algorithm, and gradually includes key measurement parameters in the WAT parameter prediction model for iterative training according to the order of correlation strength;

[0043] S103, perform precision evaluation on the WAT parameter prediction model after each iteration training;

[0044] S104, determine whether the preset performance indicators are met;

[0045] S105, if yes, stop iteration, obtain the trained WAT parameter prediction model and the final key measurement parameter combination; if no, continue training.

[0046] The deep neural network of the embodiment can include an input layer, a feature enhancement layer, a hidden layer and an output layer, wherein the feature enhancement layer introduces an attention mechanism (Attention), and the hidden layer introduces a recursive modeling algorithm. The deep neural network can be selected from a multi-layer perceptron, a convolutional neural network, a recurrent neural network, a transformer architecture, etc.

[0047] It can be seen that the present application introduces an Attention mechanism, and adds a feature enhancement layer between the input layer and the hidden layer for processing complex correlation relationships. The feature enhancement layer can effectively identify and emphasize important input factors, while reducing the influence of redundant information, and ultimately effectively cope with complex correlation relationships between input elements, thereby improving the explainability of model prediction. The feature enhancement layer can also adaptively evaluate the importance of different input parameters and the weight distribution ratio according to the actual correlation degree between the actual input measurement parameters, to more accurately capture the nonlinear relationship between parameters. The attention mechanism dynamically adjusts the importance weight of different input factors, so that the model can more accurately understand and respond to the complex correlation between input data. This is particularly important for processing high-dimensional and mutually entangled data in the semiconductor manufacturing process.

[0048] Further, the recursive modeling strategy is adopted in the hidden layer, the correlation between each parameter is calculated first, and then the strongly correlated factors of the WAT parameter are gradually mined through recursive method. Most existing models use one-time feature selection (such as PCA, Lasso) or fixed structure modeling (such as CNN, Transformer), which lack dynamic adjustment process. The present application gradually introduces the factors strongly correlated with the target output (WAT) through recursive method, so that the model can dynamically focus on key features during the training process and gradually mine potential relationships. The recursive modeling strategy mainly includes: 1) correlation analysis, calculate the correlation between each input parameter, find out the factors strongly correlated with WAT output. 2) Step-by-step recursive modeling, according to the order of correlation strength, gradually introduce key factors into the model for training, so as to broaden the context vector dimension and improve the model's understanding ability of complex patterns. 3) Pooling operation, the matrix is converted into a form suitable for the output of the hidden layer through the pooling technology, to ensure the effective transmission of information.

[0049] Further, the key measurement parameter combination screening process is as follows: the recursive value of the measurement parameter with the highest attention weight is calculated by the recursive modeling algorithm, then the measurement parameters are gradually increased according to the size of the attention weight, and the corresponding recursive weight value is calculated, if the recursive weight value is greater than the preset threshold, the measurement parameter is retained, otherwise it is deleted, and the key measurement parameter is screened out for the iterative training of the subsequent WAT parameter prediction model.

[0050] Specifically, the preprocessed input historical measurement parameters can be converted into a data matrix :

[0051]

[0052] Wherein, X 11 , …, X 1n represent the online measurement data of the first group of back-end process parameters corresponding to n sites; X m1 , …, X mn represent the online measurement data of the mth group of back-end process parameters corresponding to n sites.

[0053] The weight matrix W Q of the query vector, the weight matrix W K of the key vector, and the weight matrix W V of the value vector are as follows:

[0054]

[0055]

[0056]

[0057] Then the query vector of the attention head is Q=W Q ×X M×N The bond vector K=W K ×X M×N Value vector V=W V ×X M×N .

[0058] The attention scores are transformed into a probability distribution using the Softmax function, ensuring that all scores sum to 1 (i.e., the scores in each row of the matrix sum to 1). The attention mechanism primarily performs the following calculations:

[0059]

[0060]

[0061]

[0062]

[0063]

[0064] Where, Attenton Weights: Attention weights; Q: Query vector; K: Key vector. K: Vector dimension; V: Value vector; Context Vector: Context vector (enters the fully connected layer for calculation; in the prediction model of this invention, the layers between the input and output layers are collectively called hidden layers; hidden layers may contain fully connected layers; the context vector enters the fully connected layer for calculation mainly to calculate the predicted WAT value. The resulting matrix Z is obtained by changing the dimension of the matrix in the fully connected layer. The matrix is ​​transformed into a value through calculation, for example, multiplying a 1×N matrix with an N×1 matrix will yield a value); W: Weight; C: Context Vector; b: bias, which can be understood as a buffer (bias term, increasing the flexibility of the model) weight vector; Awmax is the maximum value in Attention Weights (each column in the attention matrix represents all the measurement values ​​of a measurement station. When the measurement values ​​are converted into weights, if the sum of the weights in a certain column is the largest, then the attention weight is the largest, indicating that the measurement station is the most important); Awsecond is the Attention... The second largest value in Weights; the value of Z is judged by the ReLU activation function or LeakyRELU activation function; if it is greater than a certain preset value, it is retained; if it is less than the preset value, it is discarded (RELU 0 or Leaky RELU custom value).

[0065] Recursive modeling can be understood as a stepwise recursive selection process for measurement stations. The input data matrix is ​​calculated in the feature enhancement layer. The attention weights of each matrix element are summed, and the data in each column are sorted from highest to lowest. The column with the largest sum of attention weights is selected, and its recursive value Z is calculated. Then, a measurement parameter is added to each column in descending order of the sum of its attention weights, and the corresponding recursive weight value is calculated (e.g., first calculate the recursive value of the first-ranked column, then the sum of the first-ranked and second-ranked columns, then the sum of the first-ranked, second-ranked, and third-ranked columns, and so on). If the recursive weight value is greater than a preset threshold, the measurement parameter in that column is retained; otherwise, it is deleted. The final key measurement parameters are selected for iterative training of the subsequent WAT parameter prediction model. During model training, the actual measured WAT parameters are used as observations in the training process and compared with the model prediction results using R-squared. 2 (Coefficient of determination) is calculated to assess prediction accuracy. If adding a specific measurement data point from a new measurement station improves the overall model's R-squared... 2 If the value is positive, the measurement value of that station is retained; otherwise, it is discarded. Through this recursive selection mechanism, the model can dynamically optimize the combination of input features, thereby improving prediction performance and generalization ability.

[0066] After multiple iterations and continuous optimization of the model structure and parameter settings, the expected performance metrics (R) were achieved. 2 >0.90, MAPE <5%. In model R 2 After the MAPE requirement is met, the required WAT (Rs / Rc, where Rs is the resistance of the metal layer and Rc is the contact resistance of the wire layer connecting different metal layers) is measured for the online experimental lot (referring to a box of 0 to 25 semiconductor wafers). Then, the inline measurement values ​​of the lot are provided for prediction, and the bias between the predicted and actual values ​​is compared. MAPE, Mean Absolute Percentage Error, is a statistical indicator used to measure the accuracy of prediction or measurement, and its calculation formula is:

[0067] .

[0068] like Figure 2 As shown, in one embodiment of the present invention, after eight model iterations, the current version V1.0.2 model predicts R... 2 All values ​​are >0.90, meeting the model accuracy requirements of the POC (Technology Validation) stage. Figure 2In the embodiment, Metal1_Rs is a model mainly for predicting the resistance of the first metal layer in the wafer back-end-of-line process, Metal2_Rs is a model mainly for predicting the resistance of the second metal layer in the wafer back-end-of-line process, Metal3_Rs is a model mainly for predicting the resistance of the third metal layer in the wafer back-end-of-line process, Via1_Rc is a model mainly for predicting the contact resistance of the first wire layer connecting different metal layers in the wafer back-end-of-line process, and Via2_Rc is a model mainly for predicting the contact resistance of the second wire layer connecting different metal layers in the wafer back-end-of-line process.

[0069] Therefore, on the basis described above, the wafer back-end-of-line process online WAT parameter prediction model training method of the embodiment further includes the steps of: preprocessing the historical measurement parameters and the corresponding historical WAT parameters of the wafer back-end-of-line process multiple sites, and constructing a training sample set for training of the WAT parameter prediction model. The preprocessing specifically includes: data cleaning and labeling of the historical measurement parameters and the corresponding historical WAT parameters of the wafer back-end-of-line process multiple sites, wherein the data cleaning is specifically performed by filtering abnormal values and random noise through a pre-constructed AI filtering model.

[0070] The AI filtering model and the labeling model can be integrated into the input layer to achieve fine processing of the input data and help the model to better understand the data connotation. The AI filtering model mainly performs preprocessing on the input parameters, such as identifying and labeling abnormal values in the data set using statistical methods or machine learning algorithms. These abnormal values may be caused by measurement errors, equipment failures or other abnormal operations. Filtering technology or other noise reduction algorithms reduce random noise in the data, making the signal clearer and helping to improve the accuracy and efficiency of model training. The labeling model mainly labels the training sample set, and the labeling content mainly labels the physical meanings of parameters such as product, suffix, and key recipe.

[0071] In the embodiment of the present application, the historical measurement parameters of the plurality of stations in the back-end-of-line process include thickness type parameters, size type parameters, material type parameters and product type parameters. Specifically, the thickness type parameters include dielectric thickness, metal thickness, etc., the thickness of the dielectric layer affects the signal propagation speed and loss, and different metal materials affect the conductivity and reliability. The size type parameters include lithography critical dimension, etched critical dimension, etc., which directly affect the size of the interconnection line and affect the speed and quality of signal transmission; the material type parameters include contact metal liner material, metal type, etc., the dielectric constant of the dielectric layer affects the signal propagation speed and loss, and different metal materials also affect the conductivity and reliability; the product type parameters include Product ID (product code), Suffix NO (code of different versions of the same product), Site No (measurement point), Key Recipe (key program), etc. The actual measurement value of WAT is input as the observation value of the training process.

[0072] In addition to using the predicted value of WAT as the model output value, some process quantity outputs can also be designed, such as measurement station weight map, heat map, etc. Through the process quantity, users can conveniently view some important results in the modeling process. The attention weight ranking result given by the model before entering the full connection layer calculation, users can judge whether the model result meets the engineering experience, and on this basis, the model weight can also be fine-tuned, such as slightly increasing the weight of the station considered most important by the engineer, so as to improve the explainability of the model.

[0073] In the field of semiconductors, using AI technology to predict the relationship between Inline-WAT parameters is an extremely innovative attempt. The WAT parameter prediction model successfully developed by the present application not only achieves a prediction accuracy of more than 95%, but also demonstrates excellent ability and reliability in processing complex process data.

[0074] In order to realize the above-mentioned method embodiment, as Figure 3 shown, the wafer back-end-of-line online WAT parameter prediction model training system of the embodiment of the present application mainly includes:

[0075] The training data acquisition module is used to acquire the historical measurement parameters of the plurality of stations in the wafer back-end-of-line process and the corresponding historical WAT parameters;

[0076] an iterative training module configured to input the historical measurement parameters and corresponding historical WAT parameters into a pre-constructed WAT parameter prediction model, the WAT parameter prediction model being a deep neural network based on an attention mechanism, the WAT parameter prediction model being configured to evaluate the importance of each measurement parameter through the attention mechanism and assign a corresponding attention weight, calculate the correlation between each measurement parameter and a corresponding WAT parameter according to the attention weight and through a recursive modeling algorithm, and gradually include key measurement parameters in the WAT parameter prediction model in order of correlation strength for iterative training;

[0077] a model evaluation module configured to evaluate the accuracy of the WAT parameter prediction model after each iteration of training until a preset performance indicator is met, stop iteration, and obtain a trained WAT parameter prediction model and a final key measurement parameter combination.

[0078] Each module is mainly used to implement each step of the above-mentioned prediction training method embodiment, and will not be described here.

[0079] The WAT parameter prediction model obtained through the above-mentioned training method embodiment can be used to predict WAT parameters. As shown in FIG. 8, the wafer back-end-of-line WAT parameter prediction method of the embodiment of the present application comprises the following steps: Figure 4

[0080] S401, obtaining online measurement parameters of multiple sites in a wafer back-end-of-line process;

[0081] S402, inputting the online measurement parameters into a WAT parameter prediction model and outputting predicted WAT parameters; wherein the WAT parameter prediction model is trained through the wafer back-end-of-line WAT parameter prediction model training method of the above-mentioned method embodiment.

[0082] The WAT parameter prediction model trained through iteration can be used directly online, and the online measurement parameters of multiple sites in a wafer back-end-of-line process can be input online to predict WAT parameters, so as to monitor the production line and judge the trend of changes in the production line and give early warning of possible situations.

[0083] ​The application further provides a computer readable storage medium, such as a flash memory, a hard disk, a multimedia card, a card-type memory (for example, an SD or DX memory, etc.), a random access memory (RAM), a static random access memory (SRAM), a read-only memory (ROM), an electrically erasable programmable read-only memory (EEPROM), a programmable read-only memory (PROM), a magnetic memory, a magnetic disk, an optical disk, a server, an App application store, and the like, which stores a computer program, and the program is executed by a processor to realize corresponding functions. The computer readable storage medium of the embodiment is executed by the processor to realize the wafer back-end process online WAT parameter prediction model training method and the wafer back-end process online WAT parameter prediction method of the method embodiment.

[0084] To sum up, the application constructs a WAT parameter prediction model and iteratively trains the same, and the unexpected effect is that, by effectively combining the attention mechanism and the recursive modeling algorithm, after the weight relationship between the input measurement parameters is found, the factors (i.e., the input measurement parameters) are gradually increased according to the attention weight strength, so as to gradually widen the input dimension, filter out the most relevant measurement parameter combination of the WAT parameter prediction result, improve the iteration efficiency of the model, and increase the prediction accuracy of the model.

[0085] It should be noted that, according to the needs of implementation, each step / component described in the present application can be split into more steps / components, or two or more steps / components or part of the operation of the steps / components can be combined into a new step / component, so as to realize the purpose of the present application.

[0086] The size of the serial number of each step in the above embodiment does not mean the execution order, and the execution order of each process should be determined according to its function and inherent logic, and should not constitute any limitation on the implementation process of the embodiments of the present application.

[0087] It should be understood that, for those skilled in the art, improvements or changes can be made according to the above description, and all these improvements and changes should belong to the protection scope of the appended claims of the present application.

Claims

1. A method for training an online WAT parameter prediction model for back-end wafer fabrication processes, characterized in that, The method comprises the following steps: obtaining historical measurement parameters and corresponding historical WAT parameters of multiple sites in a wafer back-end process; inputting the historical measurement parameters and corresponding historical WAT parameters into a pre-constructed WAT parameter prediction model, the WAT parameter prediction model being a deep neural network based on an attention mechanism, which evaluates the importance of each measurement parameter and assigns a corresponding attention weight through the attention mechanism, calculates the correlation between each measurement parameter and the corresponding WAT parameter according to the attention weight and through a recursive modeling algorithm, and gradually includes key measurement parameters in the WAT parameter prediction model for iterative training in order of correlation strength; performing precision evaluation on the WAT parameter prediction model after each iteration, until a preset performance indicator is met, stopping iteration, and obtaining a trained WAT parameter prediction model and a final key measurement parameter combination.

2. The wafer back-end-of-line online WAT parameter prediction model training method of claim 1, wherein, The deep neural network comprises an input layer, a feature enhancement layer, a hidden layer, and an output layer, wherein the feature enhancement layer introduces an attention mechanism, and the hidden layer introduces a recursive modeling algorithm.

3. The wafer back-end-of-line online WAT parameter prediction model training method of claim 1, wherein, The key measurement parameter combination screening process is as follows: calculate the recursive value of the measurement parameter with the highest attention weight through the recursive modeling algorithm, then gradually increase the measurement parameters according to the attention weight size and calculate the corresponding recursive weight value, if the recursive weight value is greater than a preset threshold, the measurement parameter is retained, otherwise it is deleted, and the key measurement parameter is screened out for subsequent iterative training of the WAT parameter prediction model.

4. The wafer back-end-of-line online WAT parameter prediction model training method of claim 2, wherein, The feature enhancement layer adaptively evaluates the importance of different input parameters and the weight distribution ratio according to the actual correlation degree between the actual input measurement parameters.

5. The wafer back-end-of-line online WAT parameter prediction model training method of claim 1, wherein, The method further comprises the step of preprocessing the historical measurement parameters and corresponding historical WAT parameters of multiple sites in a wafer back-end process to construct a training sample set for training of the WAT parameter prediction model.

6. The wafer back-end-of-line online WAT parameter prediction model training method of claim 5, wherein, The preprocessing specifically comprises: data cleaning and labeling of the historical measurement parameters and corresponding historical WAT parameters of multiple sites in a wafer back-end process, wherein the data cleaning specifically filters abnormal values and random noise through a pre-constructed AI filtering model.

7. The wafer back-end-of-line online WAT parameter prediction model training method of any one of claims 1-6, wherein, The historical measurement parameters of multiple sites in a wafer back-end process include size parameters and material parameters.

8. A wafer back-end-of-line online WAT parameter prediction method, characterized in that, The method comprises the following steps: obtaining online measurement parameters of multiple sites in a wafer back-end process; inputting the online measurement parameters into the WAT parameter prediction model to output predicted WAT parameters; wherein the WAT parameter prediction model is trained by the wafer back-end process online WAT parameter prediction model training method of any one of claims 1-7.

9. A training system for an online WAT parameter prediction model in a wafer back-end process, characterized in that, The method comprises: a training data acquisition module for obtaining historical measurement parameters and corresponding historical WAT parameters of multiple sites in a wafer back-end process; The iteration training module is configured to input the historical measurement parameters and corresponding historical WAT parameters into a pre-constructed WAT parameter prediction model, which is a deep neural network based on an attention mechanism. The WAT parameter prediction model assesses the importance of each measurement parameter through the attention mechanism and assigns a corresponding attention weight. According to the attention weight and through a recursive modeling algorithm, the correlation between each measurement parameter and the corresponding WAT parameter is calculated. In order of correlation strength, the key measurement parameters are gradually included in the WAT parameter prediction model for iterative training. The model evaluation module is configured to evaluate the accuracy of the WAT parameter prediction model after each iteration of training until the pre-set performance indicators are met, at which point the iteration is stopped, and a trained WAT parameter prediction model and a final key measurement parameter combination are obtained.

10. A computer storage medium, characterized in that The computer program stored in the memory can be executed by the processor, and the computer program executes the wafer back-end-of-line WAT parameter prediction model training method according to any one of claims 1-7.

Images