Wafer polishing method, polishing equipment and storage medium

Abstract

The embodiments of the present application provide a wafer polishing method, polishing equipment, and storage medium. The wafer polishing method includes: obtaining a first value range of a polishing time for polishing a wafer, a second value range of a polishing pressure of a polishing head, and a target removal amount of the wafer; updating a preset constraint condition based on the first value range and the second value range to obtain a target constraint condition; wherein the constraint condition is used to constrain the range of the polishing time and polishing pressure; using the polishing time and polishing pressure as variables to be optimized in an optimization model, at least minimizing the difference between the predicted removal amount and the target removal amount as the optimization target of the optimization model, and constructing an optimization model based on a preset removal rate prediction model; determining the polishing time and polishing pressure based on the target constraint condition and the optimization model; and controlling the polishing head to polish the wafer based on the determined polishing time and polishing pressure.

Description

Technical Field

[0001] The embodiments of the present application relate to the field of semiconductor manufacturing technology, and in particular to a wafer polishing method, polishing equipment, and storage medium. Background Art

[0002] Chemical Mechanical Polishing (CMP) is an ultra-precision surface processing technology for global flattening, which can complete the chemical mechanical polishing of wafers under the combined action of chemical and mechanical forces.

[0003] The chemical mechanical polishing process is susceptible to factors such as equipment consumables, exhibits nonlinear and time-varying characteristics, and interference is difficult to measure. Existing techniques typically construct linear regression models to simulate wafer variations during the chemical mechanical polishing process, thereby controlling polishing time. However, the chemical mechanical polishing process is a complex nonlinear system, and linear regression models cannot accurately fit the nonlinear variations, adversely affecting the control accuracy of polishing time using existing techniques. Summary of the Invention

[0004] In view of this, embodiments of the present application provide a wafer polishing method, a polishing device, and a storage medium to at least partially solve the above-mentioned problems.

[0005] According to a first aspect of an embodiment of the present application, a wafer polishing method is provided, comprising: obtaining a first value range of the polishing time for polishing the wafer, a second value range of the polishing pressure of the polishing head, and a target removal amount of the wafer; updating a preset constraint condition according to the first value range and the second value range to obtain a target constraint condition; wherein the constraint condition is used to constrain the range of the polishing time and the polishing pressure; taking the polishing time and the polishing pressure as variables to be optimized in an optimization model, at least minimizing the difference between the predicted removal amount and the target removal amount as the optimization target of the optimization model, and constructing the optimization model according to a preset removal rate prediction model; wherein the removal rate prediction model is a machine learning model whose input variables include the polishing pressure and whose output variable is the removal rate of the polishing head, and the predicted removal amount is equal to the product of the removal rate of the polishing head and the polishing time; determining the polishing time and the polishing pressure based on the target constraint condition and the optimization model; and controlling the polishing head to polish the wafer according to the determined polishing time and the polishing pressure.

[0006] According to a second aspect of an embodiment of the present application, a chemical mechanical polishing device is provided, comprising: a polishing disk, a polishing head, a polishing pad, a liquid supply device and a controller; the polishing head loads a wafer to be polished and abuts the polishing pad above the polishing disk, and the liquid supply device supplies polishing liquid between the polishing pad and the wafer; the controller is used to perform the following steps: obtaining a first value range of a polishing time for polishing the wafer, a second value range of a polishing pressure of the polishing head and a target removal amount of the wafer; updating a preset constraint condition according to the first value range and the second value range to obtain a target constraint condition; wherein the constraint condition is used to constrain the polishing time and the polishing pressure; pressure range; taking polishing time and polishing pressure as variables to be optimized in the optimization model, at least minimizing the difference between the predicted removal amount and the target removal amount as the optimization goal of the optimization model, and constructing the optimization model according to a preset removal rate prediction model; wherein the removal rate prediction model is a machine learning model whose input variables include the polishing pressure and the output variable is the removal rate of the polishing head, and the predicted removal amount is equal to the product of the removal rate of the polishing head and the polishing time; based on the target constraint conditions and the optimization model, determining the polishing time and the polishing pressure; according to the determined polishing time and polishing pressure, controlling the polishing head to polish the wafer.

[0007] According to a third aspect of the present application, a computer storage medium is provided, on which a computer program is stored. When the program is executed by a processor, the method described in any of the above embodiments is implemented.

[0008] In an embodiment of the present application, a first value range of the polishing time for polishing a wafer, a second value range of the polishing pressure of the polishing head, and a target removal amount of the wafer can be obtained; the preset constraints are updated according to the first value range and the second value range to obtain the target constraints, thereby constraining the polishing time and polishing pressure. The polishing time and polishing pressure are used as the variables to be optimized in the optimization model, and at least the minimum difference between the predicted removal amount and the target removal amount is used as the optimization target of the optimization model. The optimization model is constructed according to a preset removal rate prediction model, wherein the predicted removal amount is equal to the product of the removal rate and the polishing time. Based on the target constraints and the optimization model, the polishing time and polishing pressure are determined; according to the determined polishing time and polishing pressure, the polishing head is controlled to polish the wafer. The removal rate prediction model is a machine learning model. Compared with a conventional linear model, the removal rate prediction model can better fit the nonlinear changes in the polishing process to more accurately characterize the relationship between the polishing pressure and the removal rate, so that the optimization model constructed according to the removal rate prediction model can determine a more accurate polishing time and polishing pressure, thereby improving the control accuracy of the wafer polishing. BRIEF DESCRIPTION OF THE DRAWINGS

[0009] In order to more clearly illustrate the embodiments of the present application or the technical solutions in the prior art, the following briefly introduces the drawings required for use in the embodiments or the description of the prior art. Obviously, the drawings described below are only some embodiments recorded in the embodiments of the present application. For ordinary technicians in this field, other drawings can also be obtained based on these drawings.

[0010] Figure 1 It is a schematic diagram of the structure of a chemical mechanical polishing device;

[0011] Figure 2 This is a flowchart of the steps of a wafer polishing method provided in an optional embodiment of the present application;

[0012] Figure 3 This is a flowchart of the steps of training a sub-prediction model provided in an optional embodiment of the present application;

[0013] Figure 4 This is a flowchart of the steps of training another sub-prediction model provided in an optional embodiment of the present application;

[0014] Figure 5 This is a flowchart of the steps of training another seed prediction model provided in an optional embodiment of the present application;

[0015] Figure 6 This is a structural block diagram of a polishing device for polishing a wafer provided in an optional embodiment of the present application. DETAILED DESCRIPTION

[0016] In order to enable those skilled in the art to better understand the technical solutions in the embodiments of the present application, the technical solutions in the embodiments of the present application will be clearly and completely described below in conjunction with the drawings in the embodiments of the present application. Obviously, the described embodiments are only part of the embodiments of the present application, not all of the embodiments. Based on the embodiments in the embodiments of the present application, all other embodiments obtained by ordinary technicians in this field should fall within the scope of protection of the embodiments of the present application.

[0017] According to a first aspect of an embodiment of the present application, a wafer polishing method applicable to chemical mechanical polishing is provided to solve the above-mentioned problem.

[0018] like Figure 1As shown, the chemical mechanical polishing equipment 10 may include: a polishing disk 11, a polishing head 12, a polishing pad 13, a liquid supply device 14 and a controller 15. Among them, the polishing head 12 loads the wafer to be polished and abuts it against the polishing pad 13 above the polishing disk 11, and the liquid supply device 14 supplies polishing liquid between the polishing pad and the wafer. The controller 15 is used to control the polishing process of the wafer. The wafer polishing method provided in the embodiment of the present application can be used in the controller 15, so that it executes the wafer polishing method during the polishing process to control the polishing process of the wafer. It should be noted that the model training process in the subsequent embodiments may not be executed by the controller 15, but may also be executed by other servers or other devices that can train machine learning models.

[0019] The method provided in the embodiments of the present application is described in detail below with reference to the accompanying drawings.

[0020] like Figure 2 As shown, the wafer polishing method provided in the embodiment of the present application includes:

[0021] S210: Obtain a first value range of a polishing time for polishing a wafer, a second value range of a polishing pressure of a polishing head, and a target removal amount of the wafer.

[0022] The polishing pressure of the polishing head is the pressure applied to the wafer during the polishing process. The polishing pressure of the polishing head can be kept constant during the polishing process. The target removal volume of the wafer is the volume of wafer material that is planned to be removed during the polishing process.

[0023] S220: updating the preset constraint conditions according to the first value range and the second value range to obtain target constraint conditions; wherein the constraint conditions are used to constrain the range of polishing time and polishing pressure.

[0024] The constraint conditions may include value ranges of the polishing time and the polishing pressure. The value ranges of the polishing time and the polishing pressure may be updated through the first value range and the second value range.

[0025] S230: Polishing time and polishing pressure are used as variables to be optimized in the optimization model. Minimizing the difference between the predicted removal amount and the target removal amount is used as the optimization objective of the optimization model. The optimization model is constructed based on a preset removal rate prediction model. The removal rate prediction model is a machine learning model whose input variables include polishing pressure and whose output variable is the removal rate of the polishing head. Specifically, the removal rate prediction model is a machine learning model that can predict the removal rate of the polishing head based on the polishing pressure. The training process of the removal rate prediction model is not further described here. The predicted removal amount is equal to the product of the removal rate and the polishing time.

[0026] The removal rate is also known as the material removal rate (MRR). The removal rate of the polishing head is the amount of material removed from the wafer by the polishing head per unit time.

[0027] The optimization model can be:

[0028] Min S(y, A, T)......Formula 1

[0029] y represents the removal rate prediction model, A represents the target removal amount, and T represents the polishing time. S(y, A, T) is a function that calculates the difference between the predicted removal amount (the output variable of the removal rate prediction model y, i.e., the polishing head's removal rate) and the target removal amount A, obtained by multiplying the polishing time T. Min S(y, A, T) specifies the minimum value of S(y, A, T).

[0030] It should be noted that the input variables of the removal rate prediction model may also include parameters such as the life of the retaining ring and the life of the polishing pad (also known as the grinding pad). The retaining ring can be used to fix the wafer, and can flatten the polishing pad at the edge of the wafer and the polishing pad below the wafer to the same height. The retaining ring is usually set on a polishing head, and the polishing head can adsorb the wafer into the retaining ring. The polishing pad is used to grind the wafer. When polishing the wafer, the polishing head will load the wafer to be polished and abut it against the polishing pad. The life of the retaining ring can be represented by the time the retaining ring has been used or the number of times the retaining ring has participated in polishing, and the life of the polishing pad can be represented by the time the polishing pad has been used or the number of times the polishing pad has participated in polishing. The life of the retaining ring and the life of the polishing pad can be recorded in a memory. Therefore, when using the removal rate prediction model, the life of the retaining ring and the life of the polishing pad can be obtained from the memory.

[0031] Optionally, the removal rate prediction model can be a machine learning model such as a random forest model, a GBDT (Gradient Boosting Decision Tree) model, an XGBoost (eXtreme Gradient Boosting) model, a LightGBM (Light Gradient Boosting Machine) model or an artificial neural network model, or a combination model of the above machine learning models. The specific training method can refer to the relevant technology and will not be repeated here.

[0032] S240: Determine polishing time and polishing pressure based on target constraints and the optimization model;

[0033] Based on the target constraint conditions and the optimization model, the polishing time and polishing pressure values ​​that simultaneously meet the target constraint conditions and the optimization model are calculated, thereby determining the polishing time and polishing pressure.

[0034] It should be noted that in this embodiment, when determining the polishing time and polishing pressure, the target constraint conditions can be incorporated into the optimization model as part of the optimization model, and the optimization model is then solved to obtain the polishing time and polishing pressure. Of course, the polishing time and polishing pressure can also be calculated using other methods, all of which are within the scope of protection of this application.

[0035] S250: Control the polishing head to polish the wafer according to the determined polishing time and polishing pressure.

[0036] In an embodiment of the present application, a first value range of the polishing time for polishing a wafer, a second value range of the polishing pressure of the polishing head, and a target removal amount of the wafer can be obtained; the preset constraints are updated according to the first value range and the second value range to obtain the target constraints, thereby constraining the polishing time and polishing pressure. The polishing time and polishing pressure are used as the variables to be optimized in the optimization model, and at least the minimum difference between the predicted removal amount and the target removal amount is used as the optimization target of the optimization model. The optimization model is constructed according to a preset removal rate prediction model, wherein the predicted removal amount is equal to the product of the removal rate and the polishing time. Based on the target constraints and the optimization model, the polishing time and polishing pressure are determined; according to the determined polishing time and polishing pressure, the polishing head is controlled to polish the wafer. The removal rate prediction model is a machine learning model. Compared with a conventional linear model, the removal rate prediction model can better fit the nonlinear changes in the polishing process to more accurately characterize the relationship between the polishing pressure and the removal rate, so that the optimization model constructed according to the removal rate prediction model can determine a more accurate polishing time and polishing pressure, thereby improving the control accuracy of the wafer polishing.

[0037] In some optional embodiments, the polishing head for polishing the wafer may have a plurality of partitions, and each partition may be sequentially arranged along the radial direction of the polishing head from the center to the edge.

[0038] Polishing pressure includes the zone pressure of each zone, which is the pressure applied to the wafer during the polishing process. Each zone pressure can be controlled separately.

[0039] The second value range includes the value range of the partition pressure of each partition, and the value range of the partition pressure of each partition may be different.

[0040] The removal rate prediction model includes multiple sub-prediction models. The number of sub-prediction models can be the same as the number of partitions. The sub-prediction models correspond one-to-one to the partitions, and each sub-prediction model can be used to predict the removal rate of its corresponding partition. The input variables of the sub-prediction model include the partition pressure of the corresponding partition, and the output variable is the removal rate of the corresponding partition. Each sub-prediction model can be used to predict the removal rate of the partition based on the partition pressure of the corresponding partition. It should be understood that the removal rate of a partition is the amount of wafer material removed from the wafer by that partition per unit time. The removal rate of the polishing head is the average removal rate of each partition of the polishing head.

[0041] The optimization model can be:

[0042] Min S(y1,...,y n , A, T)......Formula 2

[0043] n is the number of partitions included in the polishing head, and n is an integer greater than 1. y1, ..., y n are the sub-prediction models corresponding to the n partitions included in the polishing head, A is the target removal amount, and T is the polishing time. n , A, T) is used to indicate the calculation of sub-prediction models y1, ..., y n The function of the difference between the predicted removal amount and the target removal amount A obtained by multiplying the output variable (removal rate of the partition) by the polishing time T. Min S(y1, ..., y n , A, T) is used to indicate the pair S(y1, ..., y n , A, T) takes the minimum value.

[0044] As a feasible implementation method, S(y1, ..., y n , A, T) is expressed as follows:

[0045] S(y1,...,y n , A, T) = W1|y p1 *T-A1|+...+W n |y pn *TA n |......Formula 3

[0046] As another feasible implementation, S(y1, ..., y n , A, T) is expressed as follows:

[0047] S(y1,...,y n , A, T) = W1(y p1 *T-A1) 2 +...+W n (y pn*TA n ) 2 ...Formula 4

[0048] In Equations 3 and 4, W1, ..., W n are weight coefficients for n partitions, which can be set according to the experience or needs of those skilled in the art. p1 ,...,y pn are the sub-prediction models y1,...,y n Output variables. A1...A n The target removal amount of each of the n partitions is respectively corresponding to each other. According to the ratio of the surface area of ​​the n partitions, A can be divided into A1...A n , A1+...+A n =A.

[0049] Optionally, corresponding to the removal rate prediction model, the sub-prediction model included in the removal rate prediction model may also be a machine learning model such as a random forest model, a GBDT model, an XGBoost model, a LightGBM model or an artificial neural network.

[0050] In an embodiment of the present application, the polishing head has multiple partitions; the polishing pressure includes a partition pressure of each partition, where the partition pressure of a partition is the pressure applied to the wafer by the partition during the polishing process; the second value range includes a value range of the partition pressure of each partition; the removal rate prediction model includes multiple sub-prediction models, and the sub-prediction models correspond one-to-one to the partitions; wherein the input variables of the sub-prediction models include the partition pressure of the corresponding partition, and the output variable is the removal rate of the corresponding partition. An optimization model can be constructed based on the multiple sub-prediction models included in the removal rate prediction model, so that the partition pressure of each partition of the polishing head can be determined based on the target constraint conditions and the optimization model, enabling more flexible control of the polishing head and, thereby, more accurately controlling the removal rate of the polishing head.

[0051] In some optional embodiments, the constraint condition includes a first constraint condition and a second constraint condition. The first constraint condition includes a value range of the polishing time and a value range of the polishing pressure. The constraint condition can be updated using the first value range as the value range of the polishing time and the second value range as the value range of the polishing pressure, thereby updating the constraint condition to obtain the target constraint condition.

[0052] The second constraint condition includes a non-uniformity constraint function, which is used to minimize intra-wafer non-uniformity of the wafer.

[0053] Intra-wafer non-uniformity, also known as intra-wafer uniformity, indicates the difference in performance between different locations on a wafer. The smaller the absolute value of intra-wafer non-uniformity, the smaller the difference in performance between different locations on the wafer.

[0054] It should be understood that the target constraint condition is the constraint condition that the value range of the polishing time in the first constraint condition is the first value range, and the value range of the polishing pressure is the second value range.

[0055] As a feasible implementation method, the inconsistency constraint function is:

[0056] minT WIWNU Formula 5

[0057] in,

[0058]

[0059] T WIWNU is the value of intra-chip non-uniformity, y max is the maximum removal rate of each partition of the polishing head, y min is the minimum value of the removal rate of each partition, y avg It is the average removal rate of each partition of the polishing head.

[0060] In an embodiment of the present application, the constraint conditions include a first constraint condition and a second constraint condition. The first constraint condition includes a value range of the polishing time and a value range of the partition pressure. The second constraint condition includes a non-consistency constraint function, which is used to minimize the intra-wafer non-consistency. The first constraint condition in the constraint condition is updated according to the first value range and the second value range, and a target constraint condition can be obtained, so that the target constraint condition can not only constrain the value range of the polishing time and the polishing pressure, but also minimize the intra-wafer non-consistency of the wafer, which can effectively improve the uniformity of the wafer.

[0061] In some optional embodiments, the polishing time and polishing pressure may be determined using a multi-objective optimization algorithm based on target constraints and an optimization model.

[0062] A multi-objective optimization algorithm is an algorithm used to solve optimization problems with multiple objective functions. In an embodiment of the present application, a set of objective functions can be constructed based on the objective constraints and the removal rate prediction model, and a multi-objective optimization algorithm can be used to calculate the polishing time and polishing pressure that simultaneously meet all sub-objective functions in the objective function set.

[0063] As a feasible implementation method, the multi-objective optimization algorithm adopted in the embodiment of the present application is the NSGA-II algorithm (Non-dominated Sorting Genetic Algorithm II, a non-dominated sorting genetic algorithm with an elite strategy), so that the polishing time and polishing pressure that simultaneously meet all sub-objective functions in the objective function set are calculated by the NSGA-II algorithm. For example, when using the NSGA-II algorithm, the initial population size can be set to 200, the number of evaluations is 500, the polynomial mutation (PM) probability is (0.25, 5.0), the simulated binary crossover operator (SBX) is set to 0.1, etc. Based on the above settings, the polishing time and polishing pressure that simultaneously meet all sub-objective functions in the objective function set are calculated.

[0064] The objective function set constructed based on the target constraints and the optimization model can include a first sub-objective function, a second sub-objective function, and a third sub-objective function. The optimization model can serve as the first sub-objective function, the functional expression of the first constraint in the target constraints can serve as the second sub-objective function, and the second constraint in the target constraints, including the inconsistency constraint function, can serve as the third sub-objective function. A target function set including the first, second, and third sub-objective functions is constructed, thereby utilizing a multi-objective optimization algorithm to determine the polishing time and polishing pressure that simultaneously satisfies the first, second, and third sub-objective functions. This allows the polishing head to be controlled to polish the wafer based on the determined polishing time and polishing pressure.

[0065] In some embodiments, the polishing time can be recorded as T, and the first value range can be 15s≤T≤20s. Along the radial direction of the polishing head, 7 partitions can be set on the polishing head. When the side of the polishing head in contact with the wafer is circular, the partition located in the center of the 7 partitions can be circular, and the other 6 partitions can be rings arranged in sequence along the radial direction of the polishing head. From the center of the polishing head to the outside, the partial pressures of each partition can be recorded as P1, P2, P3, P4, P5, P6 and P7 respectively. The second value range can be the following formula 7:

[0066]

[0067] Thus, the second sub-objective function can be obtained as:

[0068]

[0069] In the embodiment of the present application, the target constraint conditions and the optimization model can be used as sub-objective functions to construct an objective function set. Using a multi-objective optimization algorithm, the polishing time and polishing pressure that simultaneously meet the sub-objective functions in the objective function set can be determined, thereby achieving the optimal selection of polishing time and polishing pressure.

[0070] like Figure 3 As shown, in some optional embodiments, the polishing method provided in the embodiment of the present application further includes a sub-prediction model training step, including:

[0071] S310: Establish a sample set according to the partition pressure and removal rate of the partition corresponding to the sub-prediction model during the historical polishing process.

[0072] The historical polishing process is the polishing process that the polishing head has experienced. The partition pressure and removal rate of the partition of the polishing head in each historical polishing process can be used as a set of sample data. According to the partition pressure and removal rate of the partition of the polishing head in multiple historical polishing processes, multiple sets of sample data can be obtained to establish a sample set.

[0073] As a feasible implementation, the input variables of each sub-prediction model can also include retaining ring life and polishing pad life. A sample set can then be established based on the pressure, retaining ring life, polishing pad life, and removal rate of the partition corresponding to the sub-prediction model during historical polishing processes. The pressure, retaining ring life, polishing pad life, and removal rate of each partition of the polishing head during each historical polishing process can be used as a set of sample data. Multiple sets of sample data can be obtained based on the pressure, retaining ring life, polishing pad life, and removal rate of each partition of the polishing head during multiple historical polishing processes, thereby establishing a sample set.

[0074] S320: Constructing an initial model of a sub-prediction model, wherein the sub-prediction model is a machine learning model.

[0075] It should be understood that the input variables and output variables of the initial model of the sub-prediction model are the same as those of the sub-prediction model, and the initial model of the sub-prediction model is the initial model of the machine learning model.

[0076] S330: Train the initial model based on the sample set to obtain a sub-prediction model.

[0077] The initial model's input variables can include partition pressure, and its output variable can be the partition removal rate. Accordingly, the initial model can be trained by using the partition pressure and removal rate of the sample data in the sample set as input variables and output variables, thereby obtaining a sub-prediction model for the partition corresponding to the sample set.

[0078] In some optional embodiments, the training step of the sub-prediction model further includes: preprocessing the sample set, where the preprocessing includes at least one of the following: data cleaning, normalization, feature construction, and feature extraction.

[0079] By performing data cleaning on the sample set, the data of the sample set can be checked and verified, specifically, the consistency of the sample data of the sample set can be checked, and invalid values ​​and missing values ​​of the sample data in the sample set can be processed. By performing normalization on the sample set, the sample data in the sample set can be mapped to the same dimension, and each sample data can be mapped to a fixed interval (for example, 0 to 1) to facilitate subsequent corresponding operations based on the sample set. By performing feature construction on the sample set, new features can be constructed based on the sample data of the sample set to expand the sample set. By performing feature extraction on the sample set, features can be extracted from the sample data of the sample set to organize the sample data and reduce the amount of data in the sample set. In addition, in the embodiment of the present application, the sample set is preprocessed, which can be after the sample set is divided into a training set and a validation set, the training set and the validation set divided from the sample set are preprocessed, or it can be directly preprocessed before the sample set is divided into a training set and a validation set.

[0080] In an embodiment of the present application, a sample set can be established based on the pressure and removal rate of the partition corresponding to the sub-prediction model during historical polishing processes. After constructing an initial model for the sub-prediction model, the initial model can be trained based on the sample set to obtain a sub-prediction model. The sub-prediction model is a machine learning model. Through the above process, the nonlinear behavior of the partitions of the polishing head during the polishing process can be fitted, allowing the sub-prediction model to more accurately fit the relationship between the pressure and removal rate of the partition.

[0081] In some optional embodiments, each partition may have multiple initial models for the sub-prediction model, and each of the initial models may be initial models of different types of machine learning models. For example, the sub-prediction model may include five initial models, and each initial model may be an initial model of a random forest model, a GBDT model, an XGBoost model, a LightGBM model, and an artificial neural network model.

[0082] like Figure 4 As shown, in the embodiment of the present application, training the initial model based on the sample set to obtain the sub-prediction model may include:

[0083] S331. Divide the sample set into a training set and a validation set.

[0084] S332: Train the initial model based on the training set to obtain a candidate model corresponding to each initial model.

[0085] S333. Based on the validation set, determine the value of the first evaluation index of each candidate model, and determine a candidate model from each candidate model as a sub-prediction model according to the value of the first evaluation index.

[0086] In some optional embodiments, the first evaluation index includes at least one of the following: determination coefficient (R-Square, R 2 ), Mean Squared Error (MSE), and Mean Absolute Error (MAE). The specific calculation methods of each indicator can be referred to related technologies and will not be described here.

[0087] It should be noted that the determination coefficient indicates the fitting effect of the candidate model, the mean square error can indicate the accuracy of the prediction result of the candidate model, and the mean absolute error can indicate the prediction error of the candidate model. According to the value of the first evaluation index of each candidate model, the performance of the candidate model in terms of fitting effect, accuracy and prediction error can be obtained. When determining the sub-prediction model, the candidate model with the best performance in one of the parameters of fitting effect, accuracy and prediction error can be selected from the obtained candidate models as the sub-prediction model. For example, the candidate model with the largest determination coefficient can be selected as the sub-prediction model; or, the candidate model with the best comprehensive performance in fitting effect, accuracy and prediction error can be selected from the obtained candidate models as the sub-prediction model. For example, weights can be set for parameters such as fitting effect, accuracy and prediction error respectively, so that the sum of the weighted fitting effect, accuracy and prediction error indicates the comprehensive performance of the candidate model, and the sub-prediction model is determined from the candidate model according to the size of the sum.

[0088] In an embodiment of the present application, there can be multiple initial models of the sub-prediction model corresponding to each partition, and the multiple initial models can be initial models of different types of machine learning models. Each initial model can be trained based on the divided training set to obtain a candidate model corresponding to each initial model. Based on the validation set, the value of the first evaluation index of each candidate model can be determined, so that the performance of the candidate model can be determined according to the value of the first evaluation index of each candidate model, so that the candidate model with better performance can be selected from multiple different types of candidate models as the sub-prediction model to ensure that the sub-prediction model can accurately simulate the relationship between the partition pressure and the removal rate of the partition.

[0089] like Figure 5 As shown, in some optional embodiments, the initial model is trained based on the training set to obtain a candidate model corresponding to each initial model, including:

[0090] S3321. Train the initial models using the training set according to preset hyperparameters to obtain at least one intermediate model corresponding to each initial model. Determine the value of a second evaluation metric for the intermediate model based on the validation set; and determine whether the value of the second evaluation metric meets a preset metric condition.

[0091] Hyperparameters are external model parameters that need to be set before model learning or training. They control the behavior of model training and greatly affect the model's prediction performance. For example, the sub-prediction model can be an XGBoost model. In this case, hyperparameters may include the initial learning rate, number of trees, maximum tree depth, and sampling ratio of the XGBoost model.

[0092] During the candidate model training process, the initial model can be trained based on the training set, and the resulting model can be used as an intermediate model. It should be noted that, according to different partitioning methods, multiple training sets and corresponding validation sets can be partitioned from the sample set. When there are multiple training sets, multiple training sets can be used to train an initial model separately, with each training set producing an intermediate model, thereby obtaining multiple intermediate models corresponding to the initial model.

[0093] In some optional embodiments, the second evaluation index includes at least one of the following index parameters: coefficient of determination (R-Square, R2), mean squared error (MSE), and mean absolute error (MAE). It should be understood that the first evaluation index is an evaluation index for the candidate model, and the second evaluation index is an evaluation index for the intermediate model. The first evaluation index and the second evaluation index evaluate different objects, but the first evaluation index and the second evaluation index may include the same index parameters. Of course, the first evaluation index and the second evaluation index may also include different index parameters.

[0094] The preset indicator conditions are used to limit the second evaluation indicator to define the conditions that need to be met to obtain the sub-prediction model. The preset indicator conditions can be set based on the experience of people in this field or as needed, and will not be repeated here.

[0095] S3322. If the values ​​of the second evaluation indicators all meet the indicator conditions, then according to at least one intermediate model corresponding to each initial model, obtain a candidate model corresponding to each initial model.

[0096] When the values ​​of the second evaluation indicators of the intermediate models corresponding to all initial models meet the indicator conditions, for each initial model, a candidate model corresponding to the initial model can be obtained based on at least one intermediate model corresponding to the initial model. Exemplarily, if the initial model corresponds to only one intermediate model, the intermediate model is used as the candidate model corresponding to the initial model; if the initial model corresponds to only multiple intermediate models, a fusion model obtained by fusing the multiple intermediate models can be used as the candidate model corresponding to the initial model. For example, the average values ​​of the parameters of the multiple intermediate models can be taken respectively, and the average values ​​of the parameters can be used as the parameters of the fusion model to obtain the fusion model. Alternatively, the training set and the test set can be merged into a new training set, and the hyperparameters of the intermediate models can be used as the hyperparameters of the initial model. The initial model can be retrained using the new training set, and the retrained model can be used as the candidate model.

[0097] S3323. If there is a second evaluation indicator that does not meet the indicator conditions, the hyperparameters are adjusted, and a new initial model is constructed based on the adjusted hyperparameters, and the step of training the initial model based on the training set is repeated until a candidate model corresponding to each initial model is obtained.

[0098] It should be understood that there may be multiple new initial models corresponding to the original initial models.

[0099] After constructing a new initial model, the initial model can be trained based on the training set to obtain an intermediate model. Based on the validation set, the value of the second evaluation index of the intermediate model is determined to determine whether the value of the second evaluation index meets the preset index condition. Similar to the above embodiment, if the values ​​of the second evaluation index obtained this time all meet the index condition, the candidate model corresponding to the new initial model can be obtained based on the intermediate model; if there is a second evaluation index that does not meet the index condition this time, the hyperparameters are adjusted again, and based on the adjusted hyperparameters, a new initial model is constructed again, and the above steps of training the initial model based on the training set are repeated until the values ​​of the second evaluation index of the intermediate models corresponding to all the initial models meet the index condition, thereby obtaining the candidate model corresponding to each initial model.

[0100] As a feasible implementation method, multiple groups of preset hyperparameters can be set. Before starting to train the initial model of the sub-prediction model, a group of hyperparameters is selected from multiple groups of preset hyperparameters as the hyperparameters for training the initial model. When adjusting the hyperparameters, the next group of hyperparameters can be selected in sequence from multiple groups of preset hyperparameters as the adjusted hyperparameters to achieve the adjustment of the hyperparameters. Alternatively, a group of hyperparameters that have not been used can be randomly selected from multiple groups of preset hyperparameters as the adjusted hyperparameters to achieve the adjustment of the hyperparameters. Of course, other methods can also be used to adjust the hyperparameters, and the embodiments of the present application do not limit the specific implementation method of adjusting the hyperparameters.

[0101] In an embodiment of the present application, the initial model can be trained based on the training set to obtain at least one intermediate model corresponding to each initial model; based on the validation set, the value of the second evaluation index of the intermediate model is determined; it is judged whether the value of the second evaluation index meets the preset index condition, and when the value of the second evaluation index meets the index condition, the candidate model corresponding to the initial model is obtained based on the at least one intermediate model corresponding to the initial model. When the value of the second evaluation index does not meet the index condition, the hyperparameters can be adjusted, and based on the adjusted hyperparameters, a new initial model can be constructed, and the step of training the initial model based on the training set is repeated until the candidate model corresponding to each initial model is obtained. When the value of the second evaluation index does not meet the index condition, the hyperparameters can be continuously adjusted with the goal of ensuring that the value of the second evaluation index of the intermediate model meets the index condition to obtain an intermediate model that meets the index condition, thereby ensuring the performance of the candidate model obtained based on the intermediate model.

[0102] In some optional embodiments, dividing the sample set into a training set and a validation set includes:

[0103] S3311. Divide the sample data in the sample set into K sub-sample sets, and use the set of K-1 sub-sample sets in the K sub-sample sets as the training set, and use the other sub-sample set as the validation set, to obtain K training sets and a validation set corresponding to each training set; K is a positive integer greater than 1.

[0104] The initial model is trained using the training set to obtain at least one intermediate model corresponding to each initial model; based on the validation set, a value of a second evaluation indicator of the intermediate model is determined; and whether the value of the second evaluation indicator meets a preset indicator condition is determined, including:

[0105] K training sets are used to train the initial model respectively to obtain K intermediate models corresponding to the initial model; based on the validation sets corresponding to the K training sets, the values ​​of the second evaluation indicators of the K intermediate models are determined, and it is judged whether the values ​​of the second evaluation indicators meet the preset indicator conditions; wherein the indicator conditions are: the average value of the second evaluation indicators of the K intermediate models is within a predetermined range.

[0106] For example, K can be 5, meaning the sample data in the sample set can be divided into five subsample sets, and four of the five subsample sets are used as training sets, with the other subsample set used as a validation set, resulting in five training sets and a validation set corresponding to each training set. Based on the five training sets, the initial model can be trained independently using each training set. Training the initial model with each training set yields one intermediate model, resulting in five intermediate models.

[0107] It should be understood that the second evaluation indicator in the embodiment of the present application can be multiple. For example, the second evaluation indicator can include multiple indicators such as the coefficient of determination, mean square error, and mean absolute error. When the second evaluation indicator is multiple, a predetermined range of the second evaluation indicator can be preset for each second evaluation indicator. Only when the average value of each second evaluation indicator of the K intermediate models is within the predetermined range corresponding to the second evaluation indicator, the value of the second evaluation indicator of the K intermediate models meets the indicator condition.

[0108] As a feasible implementation, determining whether the value of the second evaluation index satisfies the preset index condition may include: determining an average value of the second evaluation index of the K intermediate models based on the value of the second evaluation index of each intermediate model, and determining whether the average value is within a predetermined range. If the average value is within the predetermined range, the value of the second evaluation index satisfies the index condition; if the average value is not within the predetermined range, the value of the second evaluation index does not satisfy the index condition.

[0109] It should be noted that in the embodiment of the present application, if the values ​​of the second evaluation indicators all meet the indicator conditions, then the candidate model corresponding to each initial model can be obtained based on the K intermediate models corresponding to each initial model. The specific implementation method can refer to the above embodiment and will not be repeated here. If there is a second evaluation indicator that does not meet the indicator conditions, the hyperparameters are adjusted, and based on the adjusted hyperparameters, a new initial model is constructed, and the step of training the initial model based on the training set is repeated to obtain K intermediate models for each initial model until a candidate model corresponding to each initial model is obtained.

[0110] In an embodiment of the present application, K training sets and validation sets can be obtained by dividing the sample set into K sub-sample sets and using each sub-sample set as a validation set. K intermediate models can be obtained by training the initial model based on each training set. Based on the validation set corresponding to each training set, the value of the second evaluation index of the corresponding intermediate model can be determined. Since the validation set corresponding to each training set is different, the intermediate models trained by each training set can calculate the value of the second evaluation index according to different validation sets, which can fully utilize the data of the sample set, and the performance of the intermediate model can be comprehensively verified by the value of the second evaluation index obtained according to different validation sets.

[0111] In some optional embodiments, the hyperparameters include a first hyperparameter, which is a hyperparameter common to all initial models. The first hyperparameter may include a learning rate, a number of iterations, etc. It should be understood that the hyperparameters may also include the first hyperparameter, and the second hyperparameter is a hyperparameter other than the first hyperparameter among the hyperparameters of the initial models.

[0112] Adjusting the hyperparameter includes adjusting the first hyperparameter.

[0113] When adjusting hyperparameters, when there are multiple initial models, the hyperparameters used in training all initial models can be adjusted uniformly, without having to adjust the hyperparameters used in training each initial model separately, which can simplify the hyperparameter adjustment procedure.

[0114] like Figure 6 As shown, the embodiment of the present application provides a polishing device 400 for polishing a wafer, comprising:

[0115] An acquisition module 410 is configured to acquire a first value range of a polishing time for polishing a wafer, a second value range of a polishing pressure of a polishing head, and a target removal amount of the wafer;

[0116] The constraint module 420 is used to update the preset constraint conditions according to the first value range and the second value range to obtain target constraint conditions; wherein the constraint conditions are used to constrain the range of polishing time and polishing pressure;

[0117] Model acquisition module 430 is configured to use polishing time and polishing pressure as variables to be optimized in an optimization model, minimize the difference between the predicted removal amount and the target removal amount as the optimization objective of the optimization model, and construct an optimization model based on a preset removal rate prediction model. The removal rate prediction model is a machine learning model whose input variables include polishing pressure and whose output variable is the removal rate of the polishing head. The predicted removal amount is equal to the product of the removal rate of the polishing head and the polishing time.

[0118] A calculation module 440 is used to determine the polishing time and polishing pressure based on the target constraint conditions and the optimization model;

[0119] The control module 450 is used to control the polishing head to polish the wafer according to the determined polishing time and polishing pressure.

[0120] The polishing apparatus for polishing wafers in this embodiment is based on the same inventive concept as the aforementioned wafer polishing method embodiments, is used to implement the corresponding polishing methods in the aforementioned multiple method embodiments, and has the beneficial effects of the corresponding method embodiments, and will not be described in detail here. Furthermore, the functional implementation of each unit in the polishing apparatus of this embodiment can be referenced to the corresponding descriptions in the aforementioned method embodiments, and will not be described in detail here.

[0121] The present application also provides a chemical mechanical polishing device, such as Figure 1 As shown, the chemical mechanical polishing equipment 10 may include: a polishing plate 11, a polishing head 12, a polishing pad 13, a liquid supply device 14, and a controller 15. The polishing head 12 loads the wafer to be polished and places it against the polishing pad 13 above the polishing plate 11. The liquid supply device 14 supplies polishing liquid between the polishing pad and the wafer. The controller 15 is configured to perform the following steps:

[0122] A first value range of the polishing time for polishing the wafer, a second value range of the polishing pressure of the polishing head 12, and a target removal amount of the wafer are obtained.

[0123] The preset constraint conditions are updated according to the first value range and the second value range to obtain target constraint conditions; wherein the constraint conditions are used to constrain the range of polishing time and polishing pressure.

[0124] Polishing time and polishing pressure are used as variables to be optimized in the optimization model, and at least minimizing the difference between the predicted removal amount and the target removal amount is used as the optimization goal of the optimization model. The optimization model is constructed according to a preset removal rate prediction model; wherein the removal rate prediction model is a machine learning model whose input variables include polishing pressure and whose output variable is the removal rate of the polishing head 12, and the predicted removal amount is equal to the product of the removal rate of the polishing head 12 and the polishing time.

[0125] Based on the target constraints and the optimization model, the polishing time and polishing pressure are determined.

[0126] According to the determined polishing time and polishing pressure, the polishing head 12 is controlled to polish the wafer.

[0127] In some optional embodiments, the polishing head 12 has multiple partitions; the polishing pressure includes the partition pressure of each partition, and the partition pressure of the partition is the pressure applied by the partition to the wafer during the polishing process; the second value range includes the value range of the partition pressure of each partition; the removal rate prediction model includes multiple sub-prediction models, and the sub-prediction models and partitions correspond one to one; wherein, the input variables of the sub-prediction model include the partition pressure of the corresponding partition, and the output variable is the removal rate of the corresponding partition.

[0128] In some optional embodiments, the constraints include a first constraint and a second constraint; the first constraint includes a value range of the polishing time and a value range of the partition pressure; the second constraint includes a non-consistency constraint function, which is used to minimize the intra-wafer non-consistency of the wafer.

[0129] In some optional embodiments, the steps executed by the controller also include a sub-prediction model training step, including: establishing a sample set based on the partition pressure and removal rate of the partition corresponding to the sub-prediction model during the historical polishing process; constructing an initial model of the sub-prediction model, wherein the sub-prediction model is a machine learning model; training the initial model based on the sample set to obtain the sub-prediction model.

[0130] In some optional embodiments, there are multiple initial models of the sub-prediction model, and the multiple initial models are initial models of different types of machine learning models; the initial model is trained based on the sample set to obtain the sub-prediction model, including: dividing the sample set into a training set and a validation set; training the initial model based on the training set to obtain a candidate model corresponding to each initial model; based on the validation set, determining the value of the first evaluation index of each candidate model, and according to the value of the first evaluation index, determining a candidate model from each candidate model as the sub-prediction model.

[0131] In some optional embodiments, the initial model is trained based on the training set to obtain a candidate model corresponding to each initial model, including: training the initial model with the training set according to preset hyperparameters to obtain at least one intermediate model corresponding to each initial model; determining the value of the second evaluation indicator of the intermediate model based on the validation set; judging whether the value of the second evaluation indicator meets the preset indicator condition; if the value of the second evaluation indicator meets the indicator condition, obtaining the candidate model corresponding to each initial model based on at least one intermediate model corresponding to each initial model; if there is a second evaluation indicator that does not meet the indicator condition, adjusting the hyperparameters, and constructing a new initial model based on the adjusted hyperparameters, and repeating the step of training the initial model based on the training set until the candidate model corresponding to each initial model is obtained.

[0132] In some optional embodiments, the sample set is divided into a training set and a validation set, including: dividing the sample data in the sample set into K sub-sample sets, and using a set of K-1 sub-sample sets in the K sub-sample sets as a training set, and using another sub-sample set as a validation set, to obtain K training sets and a validation set corresponding to each training set; K is a positive integer greater than 1.

[0133] The initial model is trained using a training set to obtain at least one intermediate model corresponding to each initial model; based on the validation set, the value of the second evaluation index of the intermediate model is determined; and whether the value of the second evaluation index meets the preset index condition is judged, including: using K training sets to train the initial model respectively to obtain K intermediate models corresponding to the initial model; based on the validation sets corresponding to the K training sets, the value of the second evaluation index of the K intermediate models is determined, and whether the value of the second evaluation index meets the preset index condition; wherein the index condition is: the average value of the second evaluation index of the K intermediate models is within a predetermined range.

[0134] In some optional embodiments, the hyperparameter includes a first hyperparameter, which is a hyperparameter common to all initial models; and adjusting the hyperparameter includes: adjusting the first hyperparameter.

[0135] In some optional embodiments, the training step of the sub-prediction model further includes: preprocessing the sample set, where the preprocessing includes at least one of the following: data cleaning, normalization, feature construction, and feature extraction.

[0136] In some optional embodiments, the first evaluation indicator includes at least one of the following: determination coefficient, mean square error and mean absolute error; and / or, the first evaluation indicator includes at least one of the following: determination coefficient, mean square error and mean absolute error.

[0137] In some optional embodiments, determining the polishing time and polishing pressure based on the target constraint conditions and the optimization model includes: determining the polishing time and polishing pressure using a multi-objective optimization algorithm based on the target constraint conditions and the optimization model.

[0138] As a feasible implementation manner, the controller 15 may include a processor and a computer program. When the processor runs the computer program, the steps / methods in any of the aforementioned method embodiments may be implemented.

[0139] The chemical mechanical polishing apparatus of this embodiment is based on the same inventive concept as the aforementioned wafer polishing method embodiments, and is used to implement the corresponding wafer polishing methods of the aforementioned multiple method embodiments, and has the beneficial effects of the corresponding method embodiments. A detailed description thereof will not be repeated here. Furthermore, the functional implementation of each unit of the chemical mechanical polishing apparatus of this embodiment can be referenced to the corresponding descriptions of the aforementioned method embodiments, and will not be repeated here.

[0140] The present application also provides a computer storage medium having a computer program stored thereon, which, when executed by a processor, implements the method described in any of the aforementioned method embodiments. The computer storage medium includes, but is not limited to, a compact disc read-only memory (CD-ROM), random access memory (RAM), a floppy disk, a hard disk, or a magneto-optical disk.

[0141] It should be pointed out that, according to the needs of implementation, the various components / steps described in the embodiments of the present application can be split into more components / steps, or two or more components / steps or partial operations of components / steps can be combined into new components / steps to achieve the purpose of the embodiments of the present application.

[0142] The above-mentioned method according to the embodiment of the present application can be implemented in hardware, firmware, or implemented as software or computer code that can be stored in a recording medium (such as a CD-ROM, RAM, floppy disk, hard disk or magneto-optical disk), or implemented as computer code originally stored in a remote recording medium or a non-temporary machine-readable medium downloaded via a network and to be stored in a local recording medium, so that the method described herein can be stored in such software processing on a recording medium using a general-purpose computer, a dedicated processor or programmable or dedicated hardware (such as an application-specific integrated circuit (ASIC) or a field-programmable gate array (FPGA).

[0143] Those skilled in the art will appreciate that the units and method steps of each example described in conjunction with the embodiments disclosed herein can be implemented in electronic hardware, or a combination of computer software and electronic hardware. Whether these functions are performed in hardware or software depends on the specific application and design constraints of the technical solution. Professional and technical personnel can use different methods to implement the described functions for each specific application, but such implementation should not be considered to be beyond the scope of the embodiments of this application.

[0144] The above implementation methods are only used to illustrate the embodiments of the present application, and are not intended to limit the embodiments of the present application. Ordinary technicians in the relevant technical field can make various changes and modifications without departing from the spirit and scope of the embodiments of the present application. Therefore, all equivalent technical solutions also fall within the scope of the embodiments of the present application, and the scope of patent protection of the embodiments of the present application should be defined by the claims.

Claims

1. A wafer polishing method, characterized in that: include: Obtaining a first value range of a polishing time for polishing a wafer, a second value range of a polishing pressure of a polishing head, and a target removal amount of the wafer; The preset constraint condition is updated according to the first value range and the second value range to obtain a target constraint condition; wherein the constraint condition is used to constrain the range of polishing time and polishing pressure; Polishing time and polishing pressure are used as variables to be optimized in the optimization model, and at least minimizing the difference between the predicted removal amount and the target removal amount is used as the optimization goal of the optimization model. The optimization model is constructed based on a preset removal rate prediction model; wherein the removal rate prediction model is a machine learning model whose input variables include polishing pressure and whose output variable is the removal rate of the polishing head, and the predicted removal amount is equal to the product of the removal rate of the polishing head and the polishing time; Based on the target constraint condition and the optimization model, determining a polishing time and a polishing pressure that simultaneously satisfies the target constraint condition and the optimization model; The polishing head is controlled to polish the wafer according to the determined polishing time and polishing pressure.

2. The method according to claim 1, characterized in that The polishing head has a plurality of partitions; The polishing pressure includes the partition pressure of each partition, and the partition pressure of the partition is the pressure applied by the partition to the wafer during the polishing process; The second value range includes the value range of the partition pressure of each partition; The removal rate prediction model includes multiple sub-prediction models, and the sub-prediction models correspond to the partitions one by one; wherein the input variables of the sub-prediction models include the partition pressure corresponding to the partition, and the output variable is the removal rate corresponding to the partition.

3. The method according to claim 2, characterized in that The constraint conditions include a first constraint condition and a second constraint condition; The first constraint condition includes a value range of the polishing time and a value range of the partition pressure; The second constraint condition includes a non-uniformity constraint function, and the non-uniformity constraint function is used to minimize intra-wafer non-uniformity of the wafer.

4. The method according to claim 3, characterized in that The method further includes a step of training the sub-prediction model, comprising: Establishing a sample set according to the partition pressure and removal rate of the partition corresponding to the sub-prediction model during the historical polishing process; Constructing an initial model of the sub-prediction model, wherein the sub-prediction model is a machine learning model; The initial model is trained based on the sample set to obtain the sub-prediction model.

5. The method according to claim 4, characterized in that There are multiple initial models of the sub-prediction model, and the multiple initial models are initial models of different types of machine learning models; The training of the initial model based on the sample set to obtain the sub-prediction model includes: Dividing the sample set into a training set and a validation set; The initial model is trained based on the training set to obtain a candidate model corresponding to each initial model; based on the validation set, the value of the first evaluation index of each candidate model is determined, and according to the value of the first evaluation index, one candidate model is determined from each candidate model as the sub-prediction model.

6. The method according to claim 5, characterized in that The training of the initial models based on the training set to obtain candidate models corresponding to each of the initial models includes: The initial model is trained using the training set according to preset hyperparameters to obtain at least one intermediate model corresponding to each of the initial models; a value of a second evaluation indicator of the intermediate model is determined based on the validation set; and whether the value of the second evaluation indicator meets a preset indicator condition is determined; If the values ​​of the second evaluation index all meet the index condition, obtaining a candidate model corresponding to each of the initial models according to at least one intermediate model corresponding to each of the initial models; If there is a second evaluation indicator that does not meet the indicator condition, the hyperparameter is adjusted, and based on the adjusted hyperparameter, a new initial model is constructed, and the step of training the initial model based on the training set is repeated until a candidate model corresponding to each initial model is obtained.

7. The method according to claim 6, characterized in that The step of dividing the sample set into a training set and a validation set comprises: The sample data in the sample set is divided into K subsample sets, and a set of K-1 subsample sets in the K subsample sets is used as a training set, and another subsample set is used as a validation set, to obtain K training sets and a validation set corresponding to each training set; K is a positive integer greater than 1; The step of training the initial model using the training set to obtain at least one intermediate model corresponding to each initial model; determining a value of a second evaluation indicator of the intermediate model based on the validation set; and judging whether the value of the second evaluation indicator satisfies a preset indicator condition includes: The initial model is trained respectively using K training sets to obtain K intermediate models corresponding to the initial model; based on the validation sets corresponding to the K training sets, the values ​​of the second evaluation indicators of the K intermediate models are determined, and it is judged whether the values ​​of the second evaluation indicators meet the preset indicator conditions; wherein the indicator conditions are: the average values ​​of the second evaluation indicators of the K intermediate models are within a predetermined range.

8. The method according to claim 6 or 7, characterized in that The hyperparameters include a first hyperparameter, which is a hyperparameter common to all the initial models; The adjusting the hyperparameter includes: adjusting the first hyperparameter.

9. The method according to any one of claims 4 to 7, characterized in that The training step of the sub-prediction model also includes: preprocessing the sample set, and the preprocessing includes at least one of the following: data cleaning, normalization, feature construction and feature extraction.

10. The method according to claim 7, characterized in that The first evaluation index includes at least one of the following: determination coefficient, mean square error and mean absolute error; and / or the second evaluation index includes at least one of the following: determination coefficient, mean square error and mean absolute error.

11. The method according to any one of claims 1, 2, 4-7 and 10, characterized in that The determining, based on the target constraint condition and the optimization model, a polishing time and a polishing pressure that simultaneously satisfies the target constraint condition and the optimization model comprises: Based on the target constraint conditions and the optimization model, a multi-objective optimization algorithm is adopted to determine the polishing time and the polishing pressure.

12. A chemical mechanical polishing device, characterized in that: include: It includes a polishing disc, a polishing head, a polishing pad, a liquid supply device and a controller; The polishing head loads the wafer to be polished and contacts the polishing pad above the polishing plate; the liquid supply device supplies polishing liquid between the polishing pad and the wafer; and the controller is configured to perform the following steps: Obtaining a first value range of a polishing time for polishing a wafer, a second value range of a polishing pressure of a polishing head, and a target removal amount of the wafer; The preset constraint condition is updated according to the first value range and the second value range to obtain a target constraint condition; wherein the constraint condition is used to constrain the range of polishing time and polishing pressure; Polishing time and polishing pressure are used as variables to be optimized in the optimization model, and at least minimizing the difference between the predicted removal amount and the target removal amount is used as the optimization goal of the optimization model. The optimization model is constructed based on a preset removal rate prediction model; wherein the removal rate prediction model is a machine learning model whose input variables include polishing pressure and whose output variable is the removal rate of the polishing head, and the predicted removal amount is equal to the product of the removal rate of the polishing head and the polishing time; Based on the target constraint condition and the optimization model, determining a polishing time and a polishing pressure that simultaneously satisfies the target constraint condition and the optimization model; The polishing head is controlled to polish the wafer according to the determined polishing time and polishing pressure.

13. The chemical mechanical polishing equipment according to claim 12, characterized in that: The polishing head has a plurality of partitions; The polishing pressure includes the partition pressure of each partition, and the partition pressure of the partition is the pressure applied by the partition to the wafer during the polishing process; The second value range includes the value range of the partition pressure of each partition; The removal rate prediction model includes multiple sub-prediction models, and the sub-prediction models correspond to the partitions one by one; wherein the input variables of the sub-prediction models include the partition pressure corresponding to the partition, and the output variable is the removal rate corresponding to the partition.

14. The chemical mechanical polishing equipment according to claim 13, characterized in that: The constraint conditions include a first constraint condition and a second constraint condition; The first constraint condition includes a value range of the polishing time and a value range of the partition pressure; The second constraint condition includes a non-uniformity constraint function, and the non-uniformity constraint function is used to minimize intra-wafer non-uniformity of the wafer.

15. The chemical mechanical polishing equipment according to claim 14, characterized in that: The steps performed by the controller also include a training step of the sub-prediction model, including: Establishing a sample set according to the partition pressure and removal rate of the partition corresponding to the sub-prediction model during the historical polishing process; Constructing an initial model of the sub-prediction model, wherein the sub-prediction model is a machine learning model; The initial model is trained based on the sample set to obtain the sub-prediction model.

16. The chemical mechanical polishing equipment according to claim 15, characterized in that: There are multiple initial models of the sub-prediction model, and the multiple initial models are initial models of different types of machine learning models; The training of the initial model based on the sample set to obtain the sub-prediction model includes: Dividing the sample set into a training set and a validation set; The initial model is trained based on the training set to obtain a candidate model corresponding to each initial model; based on the validation set, the value of the first evaluation index of each candidate model is determined, and according to the value of the first evaluation index, one candidate model is determined from each candidate model as the sub-prediction model.

17. The chemical mechanical polishing equipment according to claim 16, characterized in that: The training of the initial models based on the training set to obtain candidate models corresponding to each of the initial models includes: The initial model is trained using the training set according to preset hyperparameters to obtain at least one intermediate model corresponding to each of the initial models; a value of a second evaluation indicator of the intermediate model is determined based on the validation set; and whether the value of the second evaluation indicator meets a preset indicator condition is determined; If the values ​​of the second evaluation index all meet the index condition, obtaining a candidate model corresponding to each of the initial models according to at least one intermediate model corresponding to each of the initial models; If there is a second evaluation indicator that does not meet the indicator condition, the hyperparameter is adjusted, and based on the adjusted hyperparameter, a new initial model is constructed, and the step of training the initial model based on the training set is repeated until a candidate model corresponding to each initial model is obtained.

18. The chemical mechanical polishing equipment according to claim 17, characterized in that: The step of dividing the sample set into a training set and a validation set comprises: The sample data in the sample set is divided into K subsample sets, and a set of K-1 subsample sets in the K subsample sets is used as a training set, and another subsample set is used as a validation set, to obtain K training sets and a validation set corresponding to each training set; K is a positive integer greater than 1; The step of training the initial model using the training set to obtain at least one intermediate model corresponding to each initial model; determining a value of a second evaluation indicator of the intermediate model based on the validation set; and judging whether the value of the second evaluation indicator satisfies a preset indicator condition includes: The initial model is trained respectively using K training sets to obtain K intermediate models corresponding to the initial model; based on the validation sets corresponding to the K training sets, the values ​​of the second evaluation indicators of the K intermediate models are determined, and it is judged whether the values ​​of the second evaluation indicators meet the preset indicator conditions; wherein the indicator conditions are: the average values ​​of the second evaluation indicators of the K intermediate models are within a predetermined range.

19. The chemical mechanical polishing equipment according to claim 17 or 18, characterized in that: The hyperparameters include a first hyperparameter, which is a hyperparameter common to all the initial models; The adjusting the hyperparameter includes: adjusting the first hyperparameter.

20. A computer storage medium, characterized in that A computer program is stored thereon, and when the program is executed by a processor, the method according to any one of claims 1 to 11 is implemented.

Images