Training method of prediction model of wafer grinding removal rate

Abstract

The invention provides a training method of a prediction model of a wafer grinding removal rate, and relates to the field of wafer grinding. The method comprises the following steps: acquiring grinding parameter characteristics of a plurality of batches of wafers; the service life of auxiliary materials for grinding the multiple batches of wafers is obtained, wherein the auxiliary materials at least comprise grinding pads; the grinding pad is divided into a plurality of life cycle stages based on the service life, and a performance degradation model is established in each life cycle stage; for each batch of wafer grinding processing, performance degradation characteristics of the batch of wafer grinding processing are determined based on the service life of a grinding pad in the batch of wafer grinding processing and a performance degradation model, and each batch of wafer grinding processing at least comprises one performance degradation characteristic; combining the performance degradation features with the parameter features, and constructing a reinforced data set; and training a machine learning model through the enhanced data set to obtain a removal rate prediction model. The prediction model obtained by the training method can more accurately predict the removal rate in wafer grinding processing.

Description

Technical Field

[0001] The present application relates to the field of wafer grinding processing, and in particular to a training method for a prediction model of wafer grinding processing removal rate. Background Art

[0002] In the wafer production process of the semiconductor and photovoltaic industries, the grinding process is a key link. The main purpose is to remove the saw marks and processing deterioration layers caused by cutting, so that the wafer surface can reach the required flatness. With the increase in wafer size and the improvement of precision requirements, it is crucial to accurately control the removal rate during the grinding process.

[0003] At present, the removal rate prediction methods in related technologies have certain limitations. On the one hand, the mechanism-based modeling method is difficult to fully consider the complex interactions of various factors in the grinding process, resulting in low model prediction accuracy. On the other hand, although the data-driven statistical modeling method can use historical data for learning and prediction, most of the existing methods are aimed at the polishing process, which has large differences in parameters and auxiliary material characteristics from the grinding process. When directly applied to the grinding process, the prediction accuracy is not ideal. Summary of the invention

[0004] Based on this, it is necessary to provide a training method for a prediction model of wafer grinding removal rate to address the above technical problems. The prediction model obtained by the training method can accurately predict the removal rate in wafer grinding.

[0005] In a first aspect, the present application provides a method for training a prediction model for wafer grinding removal rate, the method comprising: Obtain parameter characteristics of multiple batches of wafer grinding processes; wherein the parameter characteristics of each batch are extracted from process parameters during the wafer grinding process of the batch, and each batch has multiple parameter characteristics and at least includes a removal rate; Obtaining the service life of auxiliary materials for multiple batches of wafer grinding processes, wherein the auxiliary materials at least include a grinding pad; wherein the grinding pad includes multiple life cycle stages divided based on its own service life, and a performance degradation model is established for each life cycle stage; wherein the performance degradation model is a functional relationship between the removal rate and the service life; For each batch of wafer grinding processes, determining a performance degradation feature of the batch of wafer grinding processes based on the service life and performance degradation model of the grinding pad in the batch of wafer grinding processes, wherein each batch of wafer grinding processes includes at least one performance degradation feature; Merge the performance degradation features of multiple batches with the parameter features of multiple batches to construct an enhanced dataset; By strengthening the data set to train the machine learning model, a removal rate prediction model was obtained.

[0006] In one embodiment, the polishing pad is divided into multiple life cycle stages, including: The polishing pad is divided into three life cycle stages based on its service life, namely, the first cycle stage, the second cycle stage and the third cycle stage. The polishing pad in the first cycle stage has not been worn out, while the polishing pad in the second cycle stage and the third cycle stage has been worn out. Among them, the first cycle stage is associated with the first service life range, the second cycle stage is associated with the second cumulative service life range, and the third cycle stage is associated with the third cumulative service life range; the maximum value of the first cumulative service life range is less than or equal to the minimum value of the second cumulative service life range, and the maximum value of the second cumulative service life range is less than or equal to the minimum value of the third cumulative service life range.

[0007] In one embodiment, a performance degradation model is established for each life cycle stage, including: A first degradation model is established for the first cycle stage, in which the removal rate and the first cumulative service life of the polishing pad are in a nonlinear relationship; A second degradation model is established for the second cycle stage, in which an exponential decay function is used to express the relationship between the removal rate and the second cumulative service life of the polishing pad; A third degradation model is established for the third cycle stage. A linear function relationship is used in the third degradation model to express the relationship in which the removal rate decreases with the increase of the third cumulative service life of the polishing pad.

[0008] In one embodiment, for each batch of wafer grinding processes, determining the performance degradation characteristics of the batch of wafer grinding processes based on the service life and performance degradation model of the grinding pad in the batch of wafer grinding processes includes: For each batch of wafer grinding processes, the life cycle stages of the grinding pad are determined based on the service life of the grinding pad in the batch of wafer grinding processes, and the life cycle stages are defined as a life cycle stage set, and the life cycle stage set includes at least one life cycle stage; For any life cycle stage in the life cycle stage set, the integral of the removal rate of the life cycle stage over the service life is calculated; The integrals of each life cycle stage in the life cycle stage set are added together and divided by the cumulative service life to obtain the performance degradation characteristics of the wafer grinding process in this batch.

[0009] In one embodiment, the method further comprises: Obtain equipment markings and processing technology for multiple batches of wafer grinding processes; For each batch of wafer grinding processes, a first category feature is determined based on the equipment tag of the batch, and a second category feature is determined based on the processing technology of the batch; Multiple batches of first category features and second category features are merged into the enhanced dataset.

[0010] In one embodiment, for each batch of wafer grinding processes, determining a first category of features based on a device tag of the batch, and determining a second category of features based on a processing technique of the batch include: Label encoding is used to map the device identifier to a unique integer to obtain the first category feature, and one-hot encoding is used to convert the processing technology into a binary vector to obtain the first category feature. The binary vector can cover different wafer types and sizes.

[0011] In one embodiment, the machine learning model is an XGBoost regression model, and the hyperparameters of the XGBoost regression model are optimized by combining a grid search algorithm with a genetic algorithm; wherein the hyperparameters of the XGBoost regression model are optimized by combining a grid search algorithm with a genetic algorithm, including: Search the predefined hyperparameter space based on the grid search algorithm to obtain the benchmark hyperparameter combination; Taking the benchmark hyperparameter combination as the center point, determining the preselected hyperparameter range, and determining the initial parameter set based on the preselected hyperparameter range; The optimal hyperparameter combination is selected from the initial parameter set through the selection, crossover and mutation operations of the genetic algorithm.

[0012] In the second aspect, an embodiment of the present application provides a method for predicting the removal rate of a wafer grinding process, the prediction method comprising: obtaining the real-time process parameters of a current batch of wafer grinding processes and the real-time service life of the grinding pad, inputting a removal rate prediction model, and obtaining a predicted removal rate of the wafer grinding process; wherein the removal rate prediction model is obtained by a training method of a prediction model of the removal rate of a wafer grinding process.

[0013] In a third aspect, the present application provides an electronic device comprising a memory and a processor, wherein the memory stores a computer program, and when the processor executes the computer program, the training method of the prediction model of the wafer grinding process removal rate of the first aspect or the method of predicting the wafer grinding process removal rate of the second aspect is implemented.

[0014] In a fourth aspect, the present application provides a training device for a prediction model of wafer grinding removal rate, the device comprising: A first acquisition module is used to acquire parameter characteristics of multiple batches of wafer grinding processes; wherein the parameter characteristics of each batch are extracted from process parameters in the wafer grinding process of the batch, and each batch has multiple parameter characteristics and at least includes a removal rate; The second acquisition module is used to obtain the service life of auxiliary materials for multiple batches of wafer grinding processing, wherein the auxiliary materials at least include a grinding pad; wherein the grinding pad includes multiple life cycle stages divided based on its own service life, and a performance degradation model is established for each life cycle stage; wherein the performance degradation model is a functional relationship between the removal rate and the service life; A feature extraction module is used to determine, for each batch of wafer grinding processes, a performance degradation feature of the batch of wafer grinding processes based on the service life and performance degradation model of the grinding pad in the batch of wafer grinding processes, wherein each batch of wafer grinding processes includes at least one performance degradation feature; The training module is used to merge the performance degradation characteristics of multiple batches with the parameter characteristics of multiple batches to construct an enhanced data set; the training module is also used to train the machine learning model through the enhanced data set to obtain a removal rate prediction model.

[0015] The training method of the above-mentioned wafer grinding removal rate prediction model focuses on multiple life cycle stages of the grinding pad and its performance degradation model by obtaining parameter characteristics and auxiliary material service life of multiple batches of wafer grinding processes. For each batch, the performance degradation characteristics are determined in combination with the service life and performance degradation model of the grinding pad. These characteristics are combined with the parameter characteristics to construct an enhanced data set, and the machine learning model is trained with the data set to obtain a removal rate prediction model. The prediction model obtained by this training method can more accurately predict the removal rate in the wafer grinding process. BRIEF DESCRIPTION OF THE DRAWINGS

[0016] Figure 1 A flowchart of a method for training a prediction model for wafer grinding removal rate in one embodiment; Figure 2 A flowchart for establishing a performance degradation model for each life cycle stage in one embodiment; Figure 3 A flow chart for determining performance degradation characteristics of a batch of wafer grinding processes in one embodiment; Figure 4 A flowchart of determining first category features and second category features and merging them into an enhanced data set in one embodiment; Figure 5 A flowchart of optimizing the hyperparameters of an XGBoost regression model by combining a grid search algorithm with a genetic algorithm in one embodiment; Figure 6 is a structural diagram of an electronic device in one embodiment; Figure 7 FIG. 1 is a diagram of a training device for predicting a wafer grinding removal rate in another embodiment. DETAILED DESCRIPTION

[0017] In order to make the purpose, technical solution and advantages of the present application more clearly understood, the present application is further described in detail below in conjunction with 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.

[0018] In one embodiment, Figure 1 As shown, a method for training a prediction model for wafer grinding removal rate is provided, the method comprising the following steps: Step 101: Obtain parameter characteristics of multiple batches of wafer grinding processes; wherein the parameter characteristics of each batch are extracted from process parameters during the wafer grinding process of the batch, and each batch has multiple parameter characteristics and at least includes a removal rate.

[0019] The parameter characteristics of each batch include at least a removal rate, wherein the removal rate of each batch can be obtained by measuring the thickness difference of the wafers before and after grinding, obtaining an average removal amount, and dividing it by the grinding time to obtain the average removal rate of each batch of wafers.

[0020] For example, assuming that the number of wafers in a batch is m, the thickness of the wafer before grinding is: .

[0021] The thickness of the wafer after grinding is: .

[0022] The average removal amount for this batch is The average removal rate of this batch is further obtained as . Where t is the grinding time.

[0023] Where m is the number of wafers in a batch. thk_pre is the thickness of the wafer before grinding, which is an array containing the thickness of each wafer in the form of thk_after is the thickness of the wafer after grinding, which is also an array, corresponding to the thickness of each wafer after grinding, in the form of MR is the average removal amount of the batch. MRR is the average removal rate of the batch.

[0024] There may be multiple process parameters in the wafer grinding process of each batch, such as pressure, grinding liquid flow, disk temperature, upper disk current and lower disk current, and the statistical characteristics such as mean and standard deviation of each process parameter are calculated.

[0025] Step 102: Obtain the service life of auxiliary materials for multiple batches of wafer grinding processing, where the auxiliary materials at least include grinding pads; wherein the grinding pads include multiple life cycle stages divided based on their own service life, and a performance degradation model is established for each life cycle stage; wherein the performance degradation model is a functional relationship between removal rate and service life.

[0026] During the wafer grinding process, in order to accurately predict the removal rate and optimize the processing technology, it is necessary to obtain the service life of auxiliary materials for multiple batches of wafer grinding processes, where the auxiliary materials at least include the grinding pad.

[0027] It should be noted that a corresponding performance degradation model can be established for each life cycle stage. Different performance degradation models can describe the functional relationship between removal rate and service life. For example, at the beginning of the use of the grinding pad, the performance of the grinding pad is relatively new and the removal rate is higher. As the use time increases, the grinding pad enters the use stage, the performance gradually stabilizes, and the removal rate changes relatively slowly. When the grinding pad enters the excessive wear stage, the performance drops significantly, and the removal rate can gradually decrease over time.

[0028] Step 103: for each batch of wafer grinding processes, determine the performance degradation characteristics of the batch of wafer grinding processes based on the service life and performance degradation model of the grinding pad in the batch of wafer grinding processes, and each batch of wafer grinding processes includes at least one performance degradation characteristic.

[0029] The service life of the polishing pad is divided into multiple life cycle stages, each stage corresponds to a performance degradation model, and different performance degradation models can describe the functional relationship between removal rate and service life. By combining the actual service life of the polishing pad with the corresponding performance degradation model, the performance degradation characteristics of the wafer grinding process in this batch can be determined.

[0030] Each batch of wafer grinding processing includes at least one performance degradation feature, which is used to reflect the performance change of the grinding pad during the batch processing, thereby providing an important basis for predicting the removal rate and optimizing the grinding process.

[0031] Step 104: Merge the performance degradation features of multiple batches with the parameter features of multiple batches to construct an enhanced data set.

[0032] The performance degradation characteristics are determined based on the service life of the polishing pad in each batch and its corresponding performance degradation model. The performance degradation characteristics can reflect the performance changes of the polishing pad at different stages of its life cycle. The parameter characteristics include the statistical characteristics of key parameters in the polishing process of each batch, such as pressure, polishing liquid flow, disk temperature, upper disk current and lower disk current, as well as the classification characteristics of equipment and process.

[0033] By merging the performance degradation features with the parameter features in batches, a comprehensive dataset can be formed, which can be named as an enhanced dataset. The enhanced dataset can contain detailed information for each batch, providing more comprehensive and richer data support for subsequent model training and prediction, which helps to improve the accuracy of prediction and the generalization ability of the model.

[0034] For example, assume there are three batches of data. Batch 1 has parameter features P 11 , P 12 ,..., P 1n , the performance degradation characteristics are F 1. Batch 2 has the following parameter characteristics: P 21 , P 22 ,..., P 2n , the performance degradation characteristics are F 2. Batch 3 has the following parameter characteristics: P 31 , P 32 ,..., P 3n , the performance degradation characteristics are F 3.

[0035] Furthermore, the merged enhanced dataset contains all the features of each batch, that is, the enhanced dataset can be [ P 11 , P 12 ,..., P 1n , F 1],[ P 21 , P 22 ,..., P 2 n , F 2],[ P 31 , P 32 ,..., P 3n , F 3] Step 105: Train the machine learning model by strengthening the data set to obtain a removal rate prediction model.

[0036] The previously constructed enhanced dataset containing multiple batches of performance degradation features and parameter features is used as input, and machine learning algorithms (such as XGBoost, random forest, etc.) are applied for model training.

[0037] During the model training process, the model learns the patterns and relationships in the data, especially the complex associations between removal rate and various features. The model parameters are optimized by minimizing the prediction error (such as mean square error) so that it can predict the removal rate based on the input feature data as accurately as possible. After sufficient training and verification, the final removal rate prediction model will be able to effectively predict the removal rate in new grinding batches based on real-time or historical performance degradation characteristics and parameter characteristics, thereby providing strong support for optimizing the grinding process.

[0038] In this embodiment, the performance degradation characteristics of each batch are determined by obtaining parameter characteristics of multiple batches of wafer grinding processes, including key indicators such as removal rate, and the service life of auxiliary materials, and combining the life cycle stage and performance degradation model of the grinding pad. These performance degradation characteristics are combined with parameter characteristics to construct an enhanced data set for training a machine learning model to obtain a removal rate prediction model. The prediction model obtained by this training method can more accurately predict the removal rate in wafer grinding.

[0039] In one embodiment, the polishing pad is divided into a plurality of life cycle stages, including: dividing the polishing pad into three life cycle stages based on the service life of the polishing pad, which are divided into a first cycle stage, a second cycle stage and a third cycle stage, wherein the polishing pad in the first cycle stage has not been worn out, and the polishing pad in the second cycle stage and the third cycle stage is worn out; wherein the first cycle stage is associated with a first service life range, the second cycle stage is associated with a second cumulative service life range, and the third cycle stage is associated with a third cumulative service life range; the maximum value of the first cumulative service life range is less than or equal to the minimum value of the second cumulative service life range, and the maximum value of the second cumulative service life range is less than or equal to the minimum value of the third cumulative service life range.

[0040] The service life of the grinding pad is divided into three life cycle stages, namely the first cycle stage, the second cycle stage and the third cycle stage. In the first cycle stage, the grinding pad has not experienced wear and tear and is in a brand new state. The service life range of the first cycle stage is defined as the first service life range. As the grinding pad is used, it gradually enters the second cycle stage. In the second cycle stage, the grinding pad begins to wear and its service life range is the second cumulative service life range. Finally, the grinding pad will enter the third cycle stage. The degree of wear in the third cycle stage is more serious, and the corresponding service life range is the third cumulative service life range.

[0041] In order to ensure that the service life ranges of different cycle stages are orderly and non-overlapping, the maximum value of the first cumulative service life range is less than or equal to the minimum value of the second cumulative service life range, indicating that the service life of the first cycle stage is completely included in the service life range of the second cycle stage. Similarly, the maximum value of the second cumulative service life range is less than or equal to the minimum value of the third cumulative service life range, ensuring that the service life of the second cycle stage is completely included in the service life range of the third cycle stage.

[0042] In this embodiment, by dividing the polishing pad into different cycle stages, the use status of the polishing pad can be understood more accurately, and the performance degradation of the polishing pad can be tracked more precisely.

[0043] In one embodiment, Figure 2 As shown in the figure, a performance degradation model is established for each life cycle stage, including the following steps: Step 201: establishing a first degradation model for a first cycle stage, wherein the removal rate and the first accumulated service life of the polishing pad in the first degradation model are in a nonlinear relationship; A first degradation model is established for the first cycle stage. For example, it is assumed that the first degradation model adopts a quadratic function form .in R ( T ) represents the removal rate, T Indicates the first cumulative service life of the polishing pad, a , b , c are the model parameters obtained by fitting historical data.

[0044] The quadratic regression model contains Item, so that the removal rate R ( T ) and usage time T The relationship between the removal rate and the polishing pad is no longer a simple linear relationship, but can show a trend of first decreasing and then increasing, or first increasing and then decreasing. The nonlinear relationship can more truly reflect the change in the removal rate of the polishing pad when it is first used.

[0045] Step 202: establishing a second degradation model for the second cycle stage, wherein an exponential decay function is used in the second degradation model to express the relationship between the removal rate and the second accumulated service life of the polishing pad; For the second cycle stage, a second degradation model is established, and an exponential decay function is used to describe the relationship between the removal rate and the second cumulative service life of the polishing pad. The specific formula of the second degradation model is: .

[0046] in, R ( T) indicates the usage time T The removal rate of r is the removal rate of the polishing pad at the beginning of the second cycle stage, k is the decay constant, which indicates the rate at which the removal rate decreases. e is the base of natural logarithms.

[0047] The second degradation model selects an exponential decay function, which can describe the decreasing trend of the removal rate gradually slowing down over time.

[0048] Step 203: establishing a third degradation model for the third cycle stage, in which a linear function relationship is used to express the relationship between the decrease in removal rate and the increase in the third cumulative service life of the polishing pad.

[0049] In the third cycle, the polishing pad has been used for a while and its performance has significantly decreased. The removal rate gradually decreases with the increase of cumulative service life. A linear function can be used to concisely express the linear decline trend of the removal rate with the use time.

[0050] The third degradation model can be expressed as: R ( T )= a - B .in, R ( T ) indicates the time T The removal rate of a is the initial removal rate at the third cycle stage, b is the removal rate over time T The negative sign indicates that as T increase, R ( T ) gradually decreases.

[0051] In this embodiment, the performance degradation model of the polishing pad in different life cycles can more accurately describe the performance changes of the polishing pad in different life cycle stages, thereby providing an important basis for predicting the removal rate and optimizing the polishing process.

[0052] In one embodiment, Figure 3 As shown, for each batch of wafer grinding processes, the performance degradation characteristics of the batch of wafer grinding processes are determined based on the service life and performance degradation model of the grinding pad in the batch of wafer grinding processes, including the following steps: Step 301: For each batch of wafer grinding processing, determine the life cycle stage that the grinding pad has gone through based on the service life of the grinding pad in the batch of wafer grinding processing, which is defined as a life cycle stage set, and the life cycle stage set includes at least one life cycle stage.

[0053] For each batch of wafer grinding, the life cycle stages it has passed are determined based on the service life of the grinding pads in the batch, and these stages are defined as a life cycle stage set. For example, assuming that the total service life of the grinding pad is 100 hours, it is divided into three life cycle stages: the first cycle stage (0-20 hours), the second cycle stage (20-60 hours), and the third cycle stage (60-100 hours). If the grinding pad has been used for 45 hours during a batch of processing, the life cycle stage set of the grinding pad is the second cycle stage. If a batch of grinding pads has been used for 75 hours, its life cycle stage set is the third cycle stage.

[0054] Step 302: For any life cycle stage in the life cycle stage set, calculate the integral of the removal rate of the life cycle stage over the service life.

[0055] The first cycle stage in the life cycle stage set, the first cycle stage removal rate , calculated from T =0 to T = T Points for 1: .

[0056] In the second cycle stage of the life cycle stage set, the removal rate function of the second cycle stage is: , calculated from T = T 1 to T = T 2 points: .

[0057] The third cycle stage in the life cycle stage set, the removal rate function of the third cycle stage is R ( T )= a - B , calculated from T = T 2 to T = T 3 points: .

[0058] Step 303: Add up the integrals of each life cycle stage in the life cycle stage set and divide by the accumulated service life to obtain the performance degradation characteristics of the wafer grinding process of the batch.

[0059] Calculate the first cycle phase in the life cycle phase set (from T =0 to T = T 1) Points ; Calculate the second cycle phase in the life cycle phase set (from T = T 1 to T = T 2) Points ; Calculate the third cycle stage in the life cycle stage set (from T = T 2 to T = T 3) Points .

[0060] Add the three points to get the total point = + + , divided by the cumulative service life , the performance degradation characteristics of the wafer grinding process of this batch are obtained, that is, Among them, the performance degradation characteristics of wafer grinding processing D It can reflect the average impact of the removal rate of the polishing pad at different target life cycle stages on the total service life.

[0061] In this embodiment, the life cycle stage of the polishing pad in the batch processing, that is, the life cycle stage set, is determined; the integral of the life cycle stage removal rate over the service life in each life cycle stage set is calculated respectively, the integrals of the life cycle stages in all life cycle stage sets are added together, and divided by the accumulated service life to obtain the performance degradation characteristics of the batch wafer polishing processing. The performance degradation characteristics comprehensively reflect the performance changes of the polishing pad at different life cycle stages, providing an important basis for evaluating the polishing effect and optimizing the process.

[0062] In one embodiment, Figure 4 As shown, the method also includes: Step 401: Obtain equipment markings and processing techniques for multiple batches of wafer grinding processes.

[0063] Equipment tag can refer to a unique identifier used to distinguish different grinding equipment, such as equipment number or name, which can be obtained from equipment management system or production records. Processing technology can refer to specific process parameters used in the grinding process, including type of grinding liquid, grinding pressure, grinding speed, grinding time, etc., and process parameter information is usually recorded in process documents or production instructions.

[0064] Step 402: For each batch of wafer grinding processes, determine a first category feature based on the equipment mark of the batch, and determine a second category feature based on the processing technology of the batch.

[0065] For each batch of wafer grinding processes, the first category feature is determined based on the equipment tag of the batch. By encoding the equipment tag, such as using integer encoding or label encoding, the equipment information can be converted into a feature that can be recognized by the model. For example, if the factory has three different grinding equipment, the different grinding equipment can be labeled as equipment 1, equipment 2, and equipment 3, and these tags can be converted into integer category features.

[0066] The second category of features is determined based on the processing technology of the batch. Processing technology refers to the specific process parameters and processes used in the grinding process. By classifying and encoding the process parameters, the processing technology information can be converted into features that can be recognized by the model. For example, if the factory has four different processing technologies, each process can be represented as a binary vector using one-hot encoding.

[0067] Step 403: Merge multiple batches of first category features and second category features into the enhanced dataset.

[0068] The first category features (equipment mark) and the second category features (processing technology) of each batch are aligned with other related features (such as removal rate, grinding time, etc.) by batch number and integrated into one dataset. Among them, this dataset is the enhanced dataset R n×z ,in n is the number of processing batches, z The number of data features for each batch.

[0069] In this embodiment, by obtaining the equipment marks and processing techniques of multiple batches of wafer grinding processes, converting them into first category features and second category features respectively, and merging them into the enhanced data set, the diversity and representativeness of the data set can be significantly enhanced. This process helps the prediction model of wafer grinding removal rate to more comprehensively learn the impact of different equipment and processes on the grinding effect, thereby improving the generalization ability and prediction accuracy of the prediction model, so that it can better adapt to various actual production scenarios.

[0070] In one embodiment, for each batch of wafer grinding processes, determining a first category feature based on the equipment mark of the batch, and determining a second category feature based on the processing technology of the batch include: Label encoding is used to map the device identifier to a unique integer to obtain the first category feature, and one-hot encoding is used to convert the processing technology into a binary vector to obtain the first category feature. The binary vector can cover different wafer types and sizes.

[0071] Specifically, for each batch of wafer grinding processing, the first category feature is determined by the equipment tag of the batch. Specifically, label coding is used to map the equipment identifier to a unique integer, for example, equipment A is mapped to 1, equipment B is mapped to 2, and so on, so as to obtain the first category feature.

[0072] The second category features are determined by the processing technology of the batch, and the processing technology is converted into a binary vector using one-hot encoding. For example, 6-inch standard wafer grinding can be represented as [1,0,0,0], 8-inch standard wafer grinding can be represented as [0,1,0,0], 6-inch seed crystal grinding can be represented as [0,0,1,0], and 8-inch seed crystal grinding can be represented as [0,0,0,1], thereby obtaining the second category features. The binary vector can cover different wafer types and sizes.

[0073] In this embodiment, in this way, the information of equipment and process can be effectively encoded into numerical features that can be processed by the model, thereby better reflecting the impact of equipment and process on the grinding effect.

[0074] In one embodiment, Figure 5 As shown, the machine learning model is an XGBoost regression model, and the hyperparameters of the XGBoost regression model are optimized by combining a grid search algorithm with a genetic algorithm; wherein, the hyperparameters of the XGBoost regression model are optimized by combining a grid search algorithm with a genetic algorithm, including the following steps: Step 501: Searching a predefined hyperparameter space based on a grid search algorithm to obtain a benchmark hyperparameter combination; Perform a more extensive grid search algorithm to traverse the pre-defined hyperparameter space, obtain preliminary results, and calculate the performance of each hyperparameter combination. The following grid search algorithm formula can be used: .

[0075] in, represents the benchmark hyperparameter combination obtained by the grid search algorithm; Represents a predefined hyperparameter space, including all possible hyperparameter combinations; Represents the loss function calculated by cross validation, usually the mean square error (MSE) or the root mean square error (RMSE). It should be noted that the benchmark hyperparameter combination can refer to the optimal hyperparameter combination.

[0076] Step 502: Taking the benchmark hyperparameter combination as the center point, determining a preselected hyperparameter range, and determining an initial parameter set based on the preselected hyperparameter range; By comparing the average validation losses of all possible hyperparameter combinations, the hyperparameter combination with the smallest average validation loss is selected as the center point of the preselected hyperparameter range. A smaller range is set around the center point of the preselected hyperparameter range as the preselected hyperparameter range.

[0077] Assume that the benchmark hyperparameter combination obtained by the grid search algorithm is: {learning_rate=0.1, max_depth=6, n_estimators=100, gamma=0.1}.

[0078] Based on the optimal hyperparameter combination, narrow down the range of each hyperparameter: Reduce the range of learning_rate to [0.05, 0.1, 0.15]; Reduce the range of max_depth to [5,6,7]; Reduce the range of n_estimators to [80, 100, 120]; Reduce the range of gamma to [0.09, 0.1].

[0079] Using each narrowed hyperparameter range, a preselected hyperparameter range can be obtained. Further, within the preselected hyperparameter range, the grid search algorithm is used again to generate all possible hyperparameter combinations within the preselected hyperparameter range, and each hyperparameter combination is cross-validated, the average validation loss is calculated, and the hyperparameter combination with the smallest average validation loss is selected as the initial parameter set.

[0080] Step 503: Filter out the optimal hyperparameter combination from the initial parameter set through selection, crossover and mutation operations of the genetic algorithm.

[0081] In genetic algorithms, individuals are usually represented by chromosomes, which are candidate solutions in a solution space. Each chromosome can be represented as a vector or a string, which is a hyperparameter combination θ here. The quality of each hyperparameter combination θ is evaluated based on the fitness function, and the combination with high fitness is selected to enter the next generation. The fitness function is defined as the negative root mean square error. The fitness function is: .

[0082] Select the value that maximizes the fitness function. The fitness function is the negative root mean square error, and the parameter combination with high fitness is selected through the selection operation.

[0083] Calculate the selection probability P(x i ) is the hyperparameter combination x i The probability of being selected is calculated as: .

[0084] in, is the hyperparameter combination x i The probability of being selected, is the fitness value of the jth hyperparameter combination, and N is the number of hyperparameter combinations.

[0085] Perform a crossover operation on the selected hyperparameter combinations. For example, select two hyperparameter combinations, exchange some parameter values, and generate new hyperparameter combinations. For example, assuming that the current hyperparameter combination is θ t , generate new combinations θ through crossover operation t+1 The calculation formula is:

[0086] Perform mutation operations to randomly change some parameter values ​​in a certain hyperparameter combination to increase the diversity of the population. For example, make a slight adjustment to a parameter in a certain hyperparameter combination to generate a new hyperparameter combination. Mutation operations help avoid falling into local optimal solutions and improve global search capabilities. Through multiple generations of iterative optimization, the hyperparameter combinations in the population will continue to improve and eventually approach the optimal solution. In each generation, repeat the selection, crossover, and mutation operations to gradually select hyperparameter combinations with better performance. After multiple generations of evolution, the optimal hyperparameter combination is obtained. Use the optimal hyperparameter combination to train the XGBoost regression model. Through the selection, crossover, and mutation operations of the genetic algorithm, the optimal hyperparameter combination is selected from the initial parameter set to improve the prediction performance of the model, and finally obtain an optimal model under the goal of minimizing the root mean square error.

[0087] In this embodiment, the hyperparameters of the XGBoost regression model are optimized by combining grid search with genetic algorithm, and the hyperparameter range is narrowed by grid search to obtain a pre-selected range, and the initial parameter set is obtained within the range, and the selection, crossover and mutation operations of the genetic algorithm are used to screen out the optimal hyperparameter combination. This method effectively improves the efficiency and accuracy of hyperparameter optimization, enables the model to achieve better performance under the goal of minimizing the root mean square error, and enhances its generalization ability.

[0088] In one embodiment, the real-time process parameters of the current batch of wafer grinding process and the real-time service life of the grinding pad are obtained, and a removal rate prediction model is input to obtain the predicted removal rate of the wafer grinding process; wherein the removal rate prediction model is obtained by a training method of a prediction model of the wafer grinding process removal rate.

[0089] Obtain the real-time process parameters of the batch processing, including grinding pressure, grinding liquid flow, disk temperature, upper disk current, lower disk current, etc. Obtain the real-time service life of the grinding pad, that is, the time the grinding pad has been used. The real-time process parameters and the real-time service life of the grinding pad are used as inputs to the pre-trained removal rate prediction model. The removal rate prediction model can predict the removal rate of the current batch of wafer grinding processing based on the input data. Predicting the removal rate can help optimize the grinding process and improve processing efficiency and quality control.

[0090] In this embodiment, by obtaining the real-time process parameters and the real-time service life of the grinding pad during the current batch of wafer grinding processing and inputting them into the removal rate prediction model, the predicted removal rate of the batch of wafer grinding processing can be obtained, thereby realizing real-time evaluation and prediction of the current grinding processing effect, and providing timely reference basis for process adjustment and optimization.

[0091] Based on the same concept, the present application also provides an electronic device, including a memory and a processor, wherein the memory stores a computer program, and when the processor executes the computer program, it can implement a training method for a prediction model of wafer grinding removal rate or a method for predicting wafer grinding removal rate.

[0092] In one embodiment, an electronic device is provided, including a memory and a processor. The electronic device may be a terminal, and its internal structure diagram may be as shown in FIG. Figure 6 As shown. The electronic device includes a processor, a memory, a communication interface, a display screen and an input device connected through a system bus. Among them, the processor of the electronic device is used to provide computing and control capabilities. The memory of the electronic device includes a non-volatile storage medium and an internal memory. The non-volatile storage medium stores an operating system and a computer program. The internal memory provides an environment for the operation of the operating system and the computer program in the non-volatile storage medium. The communication interface of the electronic device is used to communicate with an external terminal in a wired or wireless manner, and the wireless manner can be implemented through WIFI, a mobile cellular network, NFC (near field communication) or other technologies. When the computer program is executed by the processor, a training method for a prediction model of a wafer grinding removal rate or a method for predicting a wafer grinding removal rate is implemented. The display screen of the electronic device can be a liquid crystal display screen or an electronic ink display screen, and the input device of the electronic device can be a touch layer covered on the display screen, or a key, trackball or touchpad set on the housing of the electronic device, or an external keyboard, touchpad or mouse.

[0093] Based on the same concept, Figure 7 As shown, the present application also provides a training device for a prediction model of wafer grinding removal rate, the device comprising: The first acquisition module 701 is used to acquire parameter characteristics of multiple batches of wafer grinding processes; wherein the parameter characteristics of each batch are extracted from process parameters in the wafer grinding process of the batch, and each batch has multiple parameter characteristics and at least includes a removal rate; The second acquisition module 702 is used to obtain the service life of auxiliary materials for multiple batches of wafer grinding processing, wherein the auxiliary materials at least include a grinding pad; wherein the grinding pad includes multiple life cycle stages divided based on its own service life, and a performance degradation model is established for each life cycle stage; wherein the performance degradation model is a functional relationship between the removal rate and the service life; A feature extraction module 703 is used to determine, for each batch of wafer grinding processes, a performance degradation feature of the batch of wafer grinding processes based on the service life and performance degradation model of the grinding pad in the batch of wafer grinding processes, wherein each batch of wafer grinding processes includes at least one performance degradation feature; The training module 704 is used to merge the performance degradation characteristics of multiple batches with the parameter characteristics of multiple batches to construct an enhanced data set; the training module is also used to train the machine learning model through the enhanced data set to obtain a removal rate prediction model.

[0094] In one embodiment, the second acquisition module 702 divides the polishing pad into multiple life cycle stages, specifically used for: dividing the polishing pad into three life cycle stages based on the service life of the polishing pad, divided into a first cycle stage, a second cycle stage and a third cycle stage, the polishing pad in the first cycle stage has not been worn, and the polishing pad in the second cycle stage and the third cycle stage is worn; wherein, the first cycle stage is associated with a first service life range, the second cycle stage is associated with a second cumulative service life range, and the third cycle stage is associated with a third cumulative service life range; the maximum value of the first cumulative service life range is less than or equal to the minimum value of the second cumulative service life range, and the maximum value of the second cumulative service life range is less than or equal to the minimum value of the third cumulative service life range.

[0095] In one embodiment, the second acquisition module 702 establishes a performance degradation model for each life cycle stage, specifically for: establishing a first degradation model for the first cycle stage, in which the removal rate and the first cumulative service life of the polishing pad are in a nonlinear relationship; establishing a second degradation model for the second cycle stage, in which an exponential decay function is used to express the relationship between the removal rate and the second cumulative service life of the polishing pad; establishing a third degradation model for the third cycle stage, in which a linear function relationship is used to express the relationship between the removal rate and the decrease in the third cumulative service life of the polishing pad.

[0096] In one embodiment, the feature extraction module 703 determines the performance degradation characteristics of each batch of wafer grinding processing based on the service life and performance degradation model of the grinding pad in the batch of wafer grinding processing, and is specifically used for: for each batch of wafer grinding processing, based on the service life of the grinding pad in the batch of wafer grinding processing, determining the life cycle stage that the grinding pad has gone through since its use, which is defined as the target life cycle stage; calculating the integral of the removal rate and the service life in each target life cycle stage respectively; adding the integrals of each target life cycle stage and dividing by the cumulative service life to obtain the performance degradation characteristics of the batch of wafer grinding processing.

[0097] In one embodiment, the first acquisition module 701 acquires equipment markings and processing techniques of multiple batches of wafer grinding processes; for each batch of wafer grinding processes, a first category feature is determined based on the equipment marking of the batch, and a second category feature is determined based on the processing technique of the batch; and the first category features and second category features of multiple batches are merged into an enhanced data set.

[0098] In one embodiment, the first acquisition module 701 determines, for each batch of wafer grinding processing, a first category feature based on the equipment tag of the batch, and determines a second category feature based on the processing technology of the batch, specifically for: using label encoding to map the equipment identifier to a unique integer to obtain the first category feature, and using one-hot encoding to convert the processing technology into a binary vector to obtain the first category feature, and the binary vector can cover different wafer types and sizes.

[0099] In one embodiment, the machine learning model is an XGBoost regression model, and the hyperparameters of the XGBoost regression model are optimized by combining grid search and genetic algorithm; wherein the training module 704 optimizes the hyperparameters of the XGBoost regression model by combining grid search and genetic algorithm, specifically including: searching a predefined hyperparameter space based on a grid search algorithm to obtain a benchmark hyperparameter combination; taking the benchmark hyperparameter combination as the center point, determining a preselected hyperparameter range, and determining an initial parameter set based on the preselected hyperparameter range; and screening out the optimal hyperparameter combination from the initial parameter set through selection, crossover and mutation operations of the genetic algorithm.

[0100] An embodiment of the present application also provides a device for predicting the removal rate of wafer grinding processing, which includes a prediction module, which is used to obtain the real-time process parameters of the current batch of wafer grinding processing and the real-time service life of the grinding pad, input the removal rate prediction model, and obtain the predicted removal rate of the wafer grinding process.

[0101] The technical features of the above embodiments may be combined arbitrarily. To make the description concise, not all possible combinations of the technical features in the above embodiments are described. However, as long as there is no contradiction in the combination of these technical features, they should be considered to be within the scope of this specification.

[0102] The above embodiments only express several implementation methods of the present application, and the descriptions thereof are relatively specific and detailed, but they cannot be understood as limiting the scope of the present application. It should be pointed out that, for a person of ordinary skill in the art, several modifications and improvements can be made without departing from the concept of the present application, and these all belong to the protection scope of the present application. Therefore, the protection scope of the present application shall be subject to the attached claims.

Claims

1. A method for training a prediction model for wafer grinding removal rate, characterized in that: The method comprises: Obtain parameter characteristics of multiple batches of wafer grinding processes; wherein the parameter characteristics of each batch are extracted from process parameters during the wafer grinding process of the batch, and each batch has multiple parameter characteristics and at least includes a removal rate; Obtaining the service life of auxiliary materials for the multiple batches of wafer grinding processes, wherein the auxiliary materials at least include a grinding pad; wherein the grinding pad includes multiple life cycle stages divided based on its own service life, and a performance degradation model is established for each life cycle stage; wherein the performance degradation model is a functional relationship between the removal rate and the service life; For each batch of wafer grinding processes, determining a performance degradation feature of the batch of wafer grinding processes based on the service life of the grinding pad in the batch of wafer grinding processes and the performance degradation model, wherein each batch of wafer grinding processes includes at least one performance degradation feature; Merging the performance degradation features of multiple batches with parameter features of multiple batches to construct an enhanced data set; The machine learning model is trained by the enhanced data set to obtain a removal rate prediction model.

2. The method for training a prediction model for wafer grinding removal rate according to claim 1, characterized in that: The polishing pad is divided into multiple life cycle stages, including: The polishing pad is divided into three life cycle stages based on the service life of the polishing pad, namely, a first cycle stage, a second cycle stage and a third cycle stage. The polishing pad in the first cycle stage is not worn out, and the polishing pads in the second cycle stage and the third cycle stage are worn out. Among them, the first cycle stage is associated with a first service life range, the second cycle stage is associated with a second cumulative service life range, and the third cycle stage is associated with a third cumulative service life range; the maximum value of the first cumulative service life range is less than or equal to the minimum value of the second cumulative service life range, and the maximum value of the second cumulative service life range is less than or equal to the minimum value of the third cumulative service life range.

3. The method for training a prediction model for wafer grinding removal rate according to claim 2, characterized in that: Build performance degradation models for each life cycle stage, including: Establishing a first degradation model for the first cycle stage, wherein the removal rate in the first degradation model is in a nonlinear relationship with a first accumulated service life of the polishing pad; Establishing a second degradation model for the second cycle stage, wherein the second degradation model uses an exponential decay function to express the relationship between the removal rate and the second cumulative service life of the polishing pad; A third degradation model is established for the third cycle stage, in which a linear function relationship is used to express the relationship in which the removal rate decreases with the increase of the third cumulative service life of the polishing pad.

4. The method for training a prediction model for wafer grinding removal rate according to claim 3, characterized in that: For each batch of wafer grinding processes, the performance degradation characteristics of the batch of wafer grinding processes are determined based on the service life of the grinding pad in the batch of wafer grinding processes and the performance degradation model, including: For each batch of wafer grinding processing, determining the life cycle stage of the grinding pad based on the service life of the grinding pad in the batch of wafer grinding processing, which is defined as a life cycle stage set, wherein the life cycle stage set includes at least one life cycle stage; For any life cycle stage in the life cycle stage set, calculating the integral of the removal rate of the life cycle stage over the service life; The integrals of each life cycle stage in the life cycle stage set are added together and divided by the cumulative service life to obtain the performance degradation characteristics of the wafer grinding process of the batch.

5. The method for training a prediction model for wafer grinding removal rate according to any one of claims 1 to 4, characterized in that: The method further comprises: Obtain equipment markings and processing technology for multiple batches of wafer grinding processes; For each batch of wafer grinding processes, a first category feature is determined based on the equipment tag of the batch, and a second category feature is determined based on the processing technology of the batch; Multiple batches of first category features and second category features are merged into the enhanced dataset.

6. The method for training a prediction model for wafer grinding removal rate according to claim 5, characterized in that: For each batch of wafer grinding processing, the first category of features is determined based on the equipment mark of the batch, and the second category of features is determined based on the processing technology of the batch, including: The device identifier is mapped into a unique integer using label encoding to obtain the first category feature, and the processing technology is converted into a binary vector using one-hot encoding to obtain the first category feature, and the binary vector can cover different wafer types and sizes.

7. The method for training a prediction model for wafer grinding removal rate according to claim 1, characterized in that: The machine learning model is an XGBoost regression model, and the hyperparameters of the XGBoost regression model are optimized by combining grid search and genetic algorithm; wherein the hyperparameters of the XGBoost regression model are optimized by combining grid search and genetic algorithm, including: Search the predefined hyperparameter space based on the grid search algorithm to obtain the benchmark hyperparameter combination; Taking the benchmark hyperparameter combination as a center point, determining a preselected hyperparameter range, and determining an initial parameter set based on the preselected hyperparameter range; The optimal hyperparameter combination is screened out from the initial parameter set through the selection, crossover and mutation operations of the genetic algorithm.

8. A method for predicting wafer grinding removal rate, characterized in that: Obtain the real-time process parameters of the current batch of wafer grinding processing and the real-time service life of the grinding pad, input the removal rate prediction model, and obtain the predicted removal rate of the wafer grinding process; wherein the removal rate prediction model is obtained by the training method of the wafer grinding process removal rate prediction model described in any one of claims 1 to 7.

9. An electronic device comprising a memory and a processor, wherein the memory stores a computer program, wherein: When the processor executes the computer program, it implements the steps of the method for training a prediction model for wafer grinding removal rate according to any one of claims 1 to 7 or the method for predicting wafer grinding removal rate according to claim 8.

10. A training device for a prediction model of wafer grinding removal rate, characterized in that: The device comprises: A first acquisition module is used to acquire parameter characteristics of multiple batches of wafer grinding processes; wherein the parameter characteristics of each batch are extracted from process parameters in the wafer grinding process of the batch, and each batch has multiple parameter characteristics and at least includes a removal rate; A second acquisition module is used to obtain the service life of auxiliary materials for the multiple batches of wafer grinding processing, wherein the auxiliary materials at least include a grinding pad; wherein the grinding pad includes a plurality of life cycle stages divided based on its own service life, and a performance degradation model is established for each life cycle stage; wherein the performance degradation model is a functional relationship between the removal rate and the service life; A feature extraction module, for determining, for each batch of wafer grinding processes, a performance degradation feature of the batch of wafer grinding processes based on the service life of the grinding pad in the batch of wafer grinding processes and the performance degradation model, wherein each batch of wafer grinding processes includes at least one performance degradation feature; A training module is used to merge the performance degradation characteristics of multiple batches with the parameter characteristics of multiple batches to construct an enhanced data set; the training module is also used to train a machine learning model through the enhanced data set to obtain a removal rate prediction model.

Images