Virtual measurement method, product, equipment and medium

Abstract

The invention discloses a virtual measurement method, a product, equipment and a medium, and relates to the technical field of measurement. The scheme is specifically obtained through feature mapping, feature output based on an explainable dot product attention mechanism and numerical prediction construction training based on a regression sampling convolution interaction mechanism according to a plurality of polishing process variables and corresponding material removal rates. In the process of constructing the virtual measurement model, the features output by the explainable dot product attention mechanism are used as the input of a regression sampling convolution interaction mechanism, so that feature selection and regression prediction are combined, and the two steps of feature screening and numerical prediction can be completed only by using the virtual measurement model. The method solves the potential mismatching problem of a current feature selection algorithm and a regression numerical value prediction algorithm, can give play to the interpretable advantage of the model while carrying out regression prediction on a virtual measurement target value, quantitatively measures the relative importance degree of input data, and greatly improves the credibility of the model.

Description

Technical Field

[0001] The present invention relates to the field of measurement technology, and in particular to a virtual measurement method, product, equipment and medium. Background Art

[0002] In recent years, the development and competition of Chemical Mechanical Polishing (CMP) technology has become the focus of scientific research and industry. As a key step in many precision manufacturing processes, CMP technology is widely used in production and life. In the CMP process, the Material Removal Rate (MRR) is an important indicator to measure the process quality. MRR represents the amount of material removed from the material surface per unit time and is a key parameter for evaluating the polishing effect. However, it is challenging to accurately measure MRR in real time because direct measurement may damage the material surface. Therefore, it is feasible to use virtual measurement (VM) technology to indirectly estimate MRR.

[0003] In the application field of CMP technology, there is a VM method that uses machine learning and statistical analysis techniques to establish a prediction model for MRR. It extracts features related to MRR by analyzing a large amount of historical data, and uses a regression algorithm to predict MRR. However, in the VM method based on feature extraction and feature selection, the feature selection and regression prediction steps use different algorithms, which may cause the selected features to not fully match the regression model, introducing uncertainty.

[0004] In view of the above problems, how to solve the problem that in the current VM method based on feature extraction and feature selection, the feature selection and regression prediction steps use different algorithms, resulting in the selected features not fully matching the regression model and introducing uncertainty is an urgent problem to be solved by technicians in this field. Summary of the invention

[0005] The purpose of the present invention is to provide a virtual measurement method, product, equipment and medium to solve the problem that in the current VM method based on feature extraction and feature selection, different algorithms are used in the two steps of feature selection and regression prediction, resulting in incomplete matching of the selected features with the regression model, thus introducing uncertainty.

[0006] In order to solve the above technical problems, the present invention provides a virtual measurement method, comprising:

[0007] Obtaining various target polishing process variables that affect the polishing quality of the target material;

[0008] Inputting each of the target polishing process variables into a virtual measurement model to output a target material removal rate corresponding to the target material;

[0009] Among them, the construction process of the virtual measurement model includes: mapping each polishing process variable into a mapping feature; the mapping feature retains the mathematical meaning of the corresponding polishing process variable; determining the output result of the dot product attention structure according to the mapping feature and the interpretable dot product attention mechanism; and constructing a regression sampling convolution interaction model according to the output result of the dot product attention structure.

[0010] On the one hand, mapping each of the polishing process variables into the mapping feature comprises:

[0011] Mapping each of the polishing process variables in the training data set to a feature space;

[0012] The position corresponding to each of the polishing process variables in the feature space is determined to generate the mapping feature.

[0013] On the other hand, determining the output result of the dot product attention structure according to the mapping features and the interpretable dot product attention mechanism includes:

[0014] Obtaining a query, key, and value of an interpretable dot product attention mechanism according to the mapping feature; wherein the query, key, and value of the interpretable dot product attention mechanism retain the mathematical meaning of the polishing process variable;

[0015] According to the query, key and value of the interpretable dot product attention mechanism, an interpretable attention weight distribution matrix and the output result of the dot product attention structure are obtained.

[0016] On the other hand, the regression sampling convolution interaction model is constructed according to the output result of the dot product attention structure, including:

[0017] Constructing the regression sampling convolution interaction model based on the binary tree structure according to the output result of the dot product attention structure; wherein each node of the binary tree structure contains complete input features;

[0018] A single value output by the regression sampling convolution interaction model is determined to complete the construction of the virtual measurement model.

[0019] On the other hand, determining the position corresponding to each polishing process variable in the feature space to generate the mapping feature includes:

[0020] Extracting primary features and their corresponding feature dimensions based on each of the polishing process variables;

[0021] According to each of the primary features and the corresponding feature dimensions, obtaining a position of each of the primary features on the corresponding feature dimensions;

[0022] Each of the primary features and its position on the corresponding feature dimension are added together to obtain the mapping feature of each of the polishing process variables.

[0023] On the other hand, obtaining a query, key, and value of an interpretable dot product attention mechanism according to the mapping features includes:

[0024] Three one-dimensional convolution filters are set; the kernel width and step size of the one-dimensional convolution filters are both 1;

[0025] The mapped features are respectively input into three one-dimensional convolutional filters to obtain the query, key and value of the interpretable dot product attention mechanism.

[0026] On the other hand, according to the query, key and value of the interpretable dot product attention mechanism, an interpretable attention weight distribution matrix and the output result of the dot product attention structure are obtained, including:

[0027] Determine the product of the query and the key of the interpretable dot product attention mechanism to obtain the interpretable attention weight distribution matrix;

[0028] Determine the product of the interpretable attention weight distribution matrix and the value of the interpretable dot product attention mechanism to obtain the dot product attention structure output result.

[0029] On the other hand, the regression sampling convolution interaction model based on the binary tree structure is constructed according to the output result of the dot product attention structure, including:

[0030] The output result of the dot product attention structure is copied to obtain a first copy and a second copy as inputs of the first-layer binary tree nodes;

[0031] Determine the weighted feature output result of the binary tree node at the current layer according to the first copy and the second copy;

[0032] The weighted feature output result is copied to obtain a first copy and a second copy as inputs to a binary tree node of a next layer;

[0033] Return to the step of determining the weighted feature output results of the binary tree nodes of the current layer according to the first copy and the second copy, until the binary tree structure reaches a preset number of layers to obtain the regression sampling convolution interaction model.

[0034] On the other hand, the weighted feature output result of the binary tree node at the current layer is determined according to the first copy and the second copy, including:

[0035] Performing convolution operations on the first copy and the second copy respectively to generate a convolution result corresponding to the first copy and a convolution result corresponding to the second copy;

[0036] Performing exponential operations on the convolution result corresponding to the first copy and the convolution result corresponding to the second copy respectively to generate an exponential result corresponding to the first copy and an exponential result corresponding to the second copy;

[0037] The weighted feature output result of the binary tree node at the current layer is generated according to the first copy and its corresponding index result, and the second copy and its corresponding index result.

[0038] On the other hand, performing convolution operations on the first copy and the second copy respectively to generate a convolution result corresponding to the first copy and a convolution result corresponding to the second copy, including:

[0039] Performing a one-dimensional convolution operation on the first copy through a first one-dimensional convolution module to obtain a first convolution result;

[0040] Performing a one-dimensional convolution operation on the second copy through a second one-dimensional convolution module to obtain a second convolution result;

[0041] Correspondingly, performing exponential operations on the convolution result corresponding to the first copy and the convolution result corresponding to the second copy respectively to generate an exponential result corresponding to the first copy and an exponential result corresponding to the second copy, including:

[0042] Applying a natural exponential function to the first convolution result to obtain a first exponential result;

[0043] Applying a natural exponential function to the second convolution result obtains a second exponential result.

[0044] On the other hand, generating the weighted feature output result of the current layer binary tree node according to the first copy and its corresponding index result, and the second copy and its corresponding index result, includes:

[0045] Performing a Hadamard product operation on the first copy and the second exponential result to obtain a first process quantity;

[0046] Performing a Hadamard product operation on the second copy and the first exponential result to obtain a second process quantity;

[0047] Performing a one-dimensional convolution operation on the second process quantity through a fourth one-dimensional convolution module to obtain a fourth convolution result, and adding the first process quantity and the fourth convolution result to obtain a first weighted feature output result;

[0048] A third one-dimensional convolution module is used to perform a one-dimensional convolution operation on the first process quantity to obtain a third convolution result, and the second process quantity is subtracted from the third convolution result to obtain a second weighted feature output result.

[0049] On the other hand, the structures of the first one-dimensional convolution module, the second one-dimensional convolution module, the third one-dimensional convolution module and the fourth one-dimensional convolution module are the same, and are composed of a hyperbolic tangent activation function, a one-dimensional convolution layer, a random loss layer, a leaky linear rectification activation function and a one-dimensional convolution layer in sequence.

[0050] On the other hand, determining a single value output by the regression sampling convolution interaction model to complete the construction of the virtual measurement model includes:

[0051] Splicing multiple feature matrices corresponding to the regression sampling convolution interaction model to obtain a spliced ​​matrix; wherein the size of each feature matrix corresponding to the regression sampling convolution interaction model is equal to the size of the mapping feature;

[0052] The concatenated matrix is ​​stacked with fully connected layers and the matrix average value is calculated to obtain the single value, thereby completing the construction of the virtual measurement model.

[0053] On the other hand, the training process of the virtual measurement model includes:

[0054] Obtaining a training data set consisting of a plurality of polishing process variables and their corresponding material removal rates;

[0055] The virtual measurement model is trained according to the data in the training data set until the virtual measurement model converges to obtain the virtual measurement model.

[0056] On the other hand, after the virtual measurement model is trained according to the data in the training data set until the virtual measurement model converges, the method further includes:

[0057] Acquire a test data set consisting of a plurality of polishing process variables and their corresponding material removal rates;

[0058] Obtaining an interpretable attention weight distribution matrix corresponding to the current virtual measurement model;

[0059] Testing the current virtual measurement model according to the data in the test data set to obtain the virtual measurement accuracy;

[0060] Determining whether the virtual measurement accuracy reaches a preset accuracy value;

[0061] If not, filtering data in the training data set according to the interpretable attention weight distribution matrix to obtain a new training data set, and returning to the step of training the virtual measurement model according to the data in the training data set;

[0062] If so, the current virtual measurement model is used as the final virtual measurement model.

[0063] On the other hand, after inputting each of the target polishing process variables into the virtual measurement model to output the target material removal rate corresponding to the target material, the method further includes:

[0064] Storing the target polishing process variables and the corresponding target material removal rates in a training data set to update the training data set;

[0065] The virtual measurement model is retrained based on the updated training data set.

[0066] In order to solve the above technical problem, the present invention also provides a computer program product, including a computer program or instructions, which implement the steps of the above virtual measurement method when executed by a processor.

[0067] In order to solve the above technical problems, the present invention further provides a virtual measurement device, comprising:

[0068] Memory for storing computer programs;

[0069] A processor is used to implement the steps of the above-mentioned virtual measurement method when executing the computer program.

[0070] In order to solve the above technical problem, the present invention further provides a computer-readable storage medium, on which a computer program is stored, and when the computer program is executed by a processor, the steps of the above virtual measurement method are implemented.

[0071] The virtual measurement method provided by the present invention obtains each target polishing process variable that affects the polishing quality of the target material; each target polishing process variable is input into the virtual measurement model to output the target material removal rate corresponding to the target material; wherein the construction process of the virtual measurement model includes: mapping each polishing process variable into a mapping feature; the mapping feature retains the mathematical meaning of the corresponding polishing process variable; determining the output result of the dot product attention structure according to the mapping feature and the interpretable dot product attention mechanism; and constructing a regression sampling convolution interaction model according to the output result of the dot product attention structure.

[0072] The beneficial effect of the present invention is that a virtual measurement model for material removal rate prediction is provided, which is specifically obtained by feature mapping, feature output based on an interpretable dot product attention mechanism, and numerical prediction based on a regression sampling convolution interaction mechanism, in sequence, according to multiple polishing process variables and their corresponding material removal rates; in the process of constructing the virtual measurement model, the mathematical meaning of the polishing process variables is retained in the mapping features, and at the same time, the features output by the interpretable dot product attention mechanism are used as the input of the regression sampling convolution interaction mechanism, thereby retaining the mathematical meaning of all input features and combining feature selection with regression prediction, and only using the virtual measurement model as one model to complete the two steps of feature screening and numerical prediction, thereby solving the potential mismatch problem between the current feature selection algorithm and the regression numerical prediction algorithm, and while regressing and predicting the virtual measurement target value, the advantage of the model's interpretability can be brought into play, and the relative importance of the input data can be quantitatively measured, which greatly improves the credibility of the model and facilitates real-time and effective supervision of the model in industrial production.

[0073] On the other hand, the present invention specifically maps each polishing process variable in the training data set to a mapping feature; obtains the query, key and value of the interpretable dot product attention mechanism according to the mapping feature; obtains the interpretable attention weight distribution matrix and the output result of the dot product attention structure according to the query, key and value of the interpretable dot product attention mechanism; constructs a regression sampling convolution interaction model based on a binary tree structure according to the output result of the dot product attention structure; and determines a single value output by the regression sampling convolution interaction model, thereby realizing the construction of a virtual measurement model. A training data set consisting of multiple polishing process variables and their corresponding material removal rates is obtained; the virtual measurement model is trained according to the data in the training data set until the virtual measurement model converges, thereby realizing the training of the virtual measurement model. If the virtual measurement accuracy does not reach the preset accuracy value, the data in the training data set is screened according to the interpretable attention weight distribution matrix to obtain a new training data set, and the step of training the virtual measurement model according to the data in the training data set is returned to, so as to find more relevant new training data by data screening, and the model is trained again based on the new training data to improve the model accuracy.

[0074] In addition, the present invention also provides a computer program product, a virtual measurement device and a medium, which have the same effects as above. BRIEF DESCRIPTION OF THE DRAWINGS

[0075] In order to more clearly illustrate the embodiments of the present invention, the following briefly introduces the drawings required for use in the embodiments. Obviously, the drawings described below are only some embodiments of the present invention. For ordinary technicians in this field, other drawings can be obtained based on these drawings without paying creative work.

[0076] Figure 1 A flowchart of a virtual measurement method provided by an embodiment of the present invention;

[0077] Figure 2 A schematic diagram of a framework of a virtual measurement model provided by an embodiment of the present invention;

[0078] Figure 3 A schematic diagram of a framework of an interpretable dot product attention mechanism provided by an embodiment of the present invention;

[0079] Figure 4 A schematic diagram of the structure of a sampling convolution interaction block provided in an embodiment of the present invention;

[0080] Figure 5 A schematic diagram of a virtual measurement device provided by an embodiment of the present invention;

[0081] Figure 6 A schematic diagram of a virtual measurement device provided by an embodiment of the present invention. DETAILED DESCRIPTION

[0082] The following will be combined with the drawings in the embodiments of the present invention to clearly and completely describe the technical solutions in the embodiments of the present invention. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of the present invention.

[0083] The core of the present invention is to provide a virtual measurement method, product, equipment and medium to solve the problem that in the current VM method based on feature extraction and feature selection, the two steps of feature selection and regression prediction use different algorithms, resulting in incomplete matching of the selected features with the regression model, thus introducing uncertainty.

[0084] In order to enable those skilled in the art to better understand the solution of the present invention, the present invention is further described in detail below in conjunction with the accompanying drawings and specific implementation methods.

[0085] At present, in the application field of CMP technology, there is a VM method that uses machine learning and statistical analysis techniques to establish a prediction model for MRR. Features related to MRR are extracted by analyzing a large amount of historical data, and a regression algorithm is used to predict MRR. However, in the VM method based on feature extraction and feature selection, different algorithms are used in the two steps of feature selection and regression prediction, which may cause the selected features to not fully match the regression model, introducing uncertainty. Therefore, in order to solve the above problems, the present invention provides a virtual measurement method.

[0086] Figure 1 A flowchart of a virtual measurement method provided by an embodiment of the present invention. Figure 1 As shown, the method includes:

[0087] S10: Obtaining target polishing process variables that affect the polishing quality of the target material.

[0088] Specifically, in order to predict the MRR of the target material, it is first necessary to obtain various target polishing process variables that affect the polishing quality of the target material.

[0089] It should be noted that the target material in this embodiment is the material for chemical mechanical polishing, and the specific type of the target material is not limited in this embodiment. In addition, the polishing process variables refer to various parameters and conditions that affect the polishing effect and quality during the polishing process, and the target polishing process variables vary with different target materials; the polishing process variables generally include pressure, speed, flow, swing parameters and other parameters. In this embodiment, there is no limitation on the target polishing process variables involved in the target material, which depends on the specific implementation situation.

[0090] S11: Input each target polishing process variable into the virtual measurement model to output a target material removal rate corresponding to the target material.

[0091] Among them, the construction process of the virtual measurement model includes: mapping each polishing process variable into mapping features; the mapping features retain the mathematical meaning of the corresponding polishing process variables; determining the output results of the dot product attention structure based on the mapping features and the interpretable dot product attention mechanism; and constructing a regression sampling convolution interaction model based on the output results of the dot product attention structure.

[0092] Furthermore, each target polishing process variable is input into the virtual measurement model, so that the target material removal rate corresponding to the target material is predicted and output through the virtual measurement model, thereby realizing the prediction of the MRR of the target material.

[0093] It should be noted that the virtual measurement model in this embodiment is an interpretable sample convolution and interaction former (SCI-Former) model, which can numerically measure the importance of all input features while outputting with high precision. Figure 2 Schematic diagram of the framework of the virtual measurement model provided by the embodiment of the present invention. Figure 2 As shown, the virtual measurement model consists of two parts, one is the interpretable dot product attention module as an encoder, and the other is the regression sample convolution and interaction network module as a decoder. The interpretable dot product attention measures the importance of the embedded input features and outputs weighted features according to the attention weights. Compared with the traditional dot product attention structure, the information flow of the interpretable dot product attention provided by the present invention is more straightforward to explain. The regression sample convolution and interaction network module is improved based on the sample convolution and interaction network model, and is an efficient regressor for MRR prediction based on weighted features.

[0094] In addition, in this embodiment, each polishing process variable is mapped to a mapping feature; the mapping feature retains the mathematical meaning of the corresponding polishing process variable; the dot product attention structure output result is determined according to the mapping feature and the interpretable dot product attention mechanism; and a regression sampling convolution interaction model is constructed according to the output result of the dot product attention structure to realize the construction of a virtual measurement model. Therefore, the virtual measurement model can complete the two steps of feature screening and numerical prediction, solving the potential mismatch problem between the feature screening algorithm and the numerical prediction algorithm. At the same time, the virtual measurement model is also interpretable, and can numerically measure the importance of all input features while outputting with high precision, greatly improving the credibility of the model. In this embodiment, there is no restriction on the specific construction method and training process of the virtual measurement model, which depends on the specific implementation situation.

[0095] In this embodiment, according to multiple polishing process variables and their corresponding material removal rates, they are constructed and trained in sequence through feature mapping, feature output based on an interpretable dot product attention mechanism, and numerical prediction based on a regression sampling convolution interaction mechanism; in the process of constructing the virtual measurement model, the mathematical meaning of the polishing process variables is retained in the mapping features, and the features output by the interpretable dot product attention mechanism are used as the input of the regression sampling convolution interaction mechanism, thereby retaining the mathematical meaning of all input features and combining feature selection with regression prediction. Only the virtual measurement model is used to complete the two steps of feature screening and numerical prediction, which solves the potential mismatch problem between the current feature selection algorithm and the regression numerical prediction algorithm. While regressing and predicting the virtual measurement target value, the advantage of the model's interpretability can be brought into play, and the relative importance of the input data can be quantitatively measured, which greatly improves the credibility of the model and facilitates real-time and effective supervision of the model in industrial production.

[0096] The construction of the virtual measurement model in the present invention is divided into three stages:

[0097] First, the first stage is a mapping layer, which maps the data to the feature space while retaining the independence of the input data. Therefore, based on the above embodiments, in some embodiments, each polishing process variable is mapped to a mapping feature, specifically including:

[0098] S12: Mapping each polishing process variable in the training data set to a feature space;

[0099] S13: Determine the position corresponding to each polishing process variable in the feature space to generate a mapping feature.

[0100] In this embodiment, each polishing process variable in the training data set is specifically mapped to the feature space, and the mathematical meaning of each polishing process variable can be retained. It can be understood that the training set data is composed of multiple polishing process variables and their corresponding material removal rates. The position corresponding to each polishing process variable in the feature space is further determined, so that each polishing process variable is marked, and finally a mapping feature is generated. It should be noted that the specific process of determining the position corresponding to each polishing process variable in the feature space is not limited in this embodiment.

[0101] Furthermore, the second stage is an interpretable dot product attention mechanism, which extracts necessary feature information, shields invalid feature information, outputs high-order features, and accurately measures the importance of each feature. Therefore, based on the above embodiments, in some embodiments, the output result of the dot product attention structure is determined according to the mapping features and the interpretable dot product attention mechanism, including:

[0102] S14: Get the query, key and value of the interpretable dot-product attention mechanism based on the mapped features;

[0103] Among them, the query, key and value of the interpretable dot product attention mechanism retain the mathematical meaning of the polishing process variables;

[0104] S15: According to the query, key and value of the interpretable dot product attention mechanism, obtain the interpretable attention weight distribution matrix and the dot product attention structure output result.

[0105] In this embodiment, the query, key and value of the interpretable dot product attention mechanism are obtained specifically according to the mapping features, and the interpretable attention weight distribution matrix and the dot product attention structure output result are obtained according to the query, key and value of the interpretable dot product attention mechanism.

[0106] It should be noted that the query, key and value of the interpretable dot product attention mechanism in this embodiment retain the mathematical meaning of the polishing process variables. At the same time, the interpretable dot product attention mechanism no longer adds the output of the attention mechanism to the original input features at the element level. Discarding the residual connection can prevent the original features that have not been adjusted by the attention weight from entering the deep network, ensuring the importance of the variables measured by the attention weight. In this embodiment, there is no restriction on the specific acquisition process of the query, key and value, the interpretable attention weight distribution matrix and the output results of the dot product attention structure, which depends on the specific implementation situation.

[0107] Finally, the third stage is regression sampling convolution interaction, which consists of convolution layers and makes predictions based on the high-order features of the previous stage. Therefore, based on the above embodiments, in some embodiments, a regression sampling convolution interaction model is constructed based on the output results of the dot product attention structure, including:

[0108] S16: Construct a regression sampling convolution interaction model based on a binary tree structure according to the output results of the dot product attention structure;

[0109] Among them, each node of the binary tree structure contains complete input features;

[0110] S17: Determine a single value output by the regression sampling convolution interaction model to complete the construction of the virtual measurement model.

[0111] In this embodiment, a regression sampling convolution interaction model based on a binary tree structure is constructed specifically according to the output result of the dot product attention structure; and a single value output by the regression sampling convolution interaction model is determined to complete the construction of the virtual measurement model.

[0112] It should be noted that in this embodiment, there is no restriction on the specific construction process of constructing a regression sampling convolution interaction model based on a binary tree structure according to the output results of the dot product attention structure, and there is no restriction on the method of determining the single value output by the regression sampling convolution interaction model, which depends on the specific implementation situation.

[0113] The construction process of the virtual measurement model is described in detail below with reference to specific embodiments:

[0114] (1) Feature mapping layer;

[0115] Based on the above embodiments, in some embodiments, determining the position corresponding to each polishing process variable in the feature space to generate a mapping feature includes:

[0116] S131: extracting primary features and their corresponding feature dimensions based on each polishing process variable;

[0117] S132: According to each primary feature and its corresponding feature dimension, obtain the position of each primary feature on the corresponding feature dimension;

[0118] S133: Add each primary feature and its position in the corresponding feature dimension to obtain a mapping feature of each polishing process variable.

[0119] Specifically, the feature map layer It consists of multiple one-dimensional convolutional layers with a kernel width of 1 and a stride of 1. Based on the polishing process variables, the feature mapping layer Extract primary features and their corresponding feature dimensions A one-dimensional convolutional layer that only computes a single input variable can effectively preserve the mathematical meaning of the input variable.

[0120] Furthermore, according to each primary feature and its corresponding feature dimension, the position of each primary feature on the corresponding feature dimension is obtained. Specifically, assuming Primary Features ; The positions of each primary feature in the corresponding feature dimension are as follows:

[0121] ;

[0122] ;

[0123] in, and Respectively represent the even and odd positions of the primary features on the feature dimension. .

[0124] Finally, each primary feature is summed with its position in the corresponding feature dimension to obtain the mapping features of each polishing process variable, as follows:

[0125] ;

[0126] in, is the mapping feature, For the primary feature, are the primary features and their positions in the corresponding feature dimensions.

[0127] (2) Interpretable dot-product attention mechanism;

[0128] Based on the above embodiments, in some embodiments, obtaining a query, a key, and a value of an interpretable dot product attention mechanism according to a mapping feature includes:

[0129] S141: setting three one-dimensional convolution filters; the kernel width and step size of the one-dimensional convolution filters are both 1;

[0130] S142: Input the mapped features into three one-dimensional convolutional filters respectively to obtain the query, key and value of the interpretable dot product attention mechanism respectively.

[0131] It is understandable that the attention value of the transformer structure is positively correlated with the importance of the variable. The attention mechanism of the neural network multiplies the non-zero weight with the input, and the sum is 1. The attention value of the variable numerically represents the relative importance of the variable. However, for efficiency and accuracy, the transformer structure breaks the mathematical meaning of the input features, making it impossible to measure the importance of the original input features. The classic dot product attention mechanism cannot be used as a feature filter either. There are two specific reasons: First, the query, key, and value in the traditional transformer structure are all obtained through the fully connected layer, which destroys the mathematical meaning of the feature. The query, key, and value obtained through the fully connected layer are a mixture of all input variables, and it is difficult to trace back to the original input variables they represent, which is very cumbersome. Second, the traditional transformer structure adds its original input to the output of the attention mechanism at the element level. This addition retains the original data features, resulting in the original features continuing to participate in the subsequent calculation process without attention weight adjustment.

[0132] Figure 3 A schematic diagram of the framework of the interpretable dot product attention mechanism provided by the embodiment of the present invention. Figure 3 As shown, this embodiment provides an interpretable multi-head dot product attention mechanism. Specifically, three one-dimensional convolution filters are first set; the kernel width and step size of the one-dimensional convolution filter are both 1. The mapping features are respectively input into the three one-dimensional convolution filters to obtain the query, key and value of the interpretable dot product attention mechanism respectively.

[0133] Further, after obtaining the query, key, and value of the interpretable dot product attention mechanism, the interpretable attention weight distribution matrix and the dot product attention structure output result are obtained according to the query, key, and value of the interpretable dot product attention mechanism, specifically including:

[0134] S151: Determine the product of the query and key of the interpretable dot product attention mechanism to obtain an interpretable attention weight distribution matrix. The formula is as follows:

[0135] ;

[0136] in, is the interpretable attention weight distribution matrix, For query, is the key.

[0137] S152: Determine the product of the explainable attention weight distribution matrix and the value of the explainable dot product attention mechanism to obtain the dot product attention structure output result. The formula is as follows:

[0138] ;

[0139] in, Output result for the dot product attention structure, is the interpretable attention weight distribution matrix, for , For the value, for The row vector of .

[0140] Therefore, compared with the classic attention mechanism, the main improvement of the interpretable multi-head dot product attention mechanism provided by this scheme is that the query, key and value are obtained through a one-dimensional convolution filter. Such a filter ensures that the mathematical meaning of the obtained query, key and value is clear, and the mathematical meaning is directly inherited from the input variables. At the same time, the residual connection structure is discarded. The interpretable dot product attention mechanism no longer adds the output of the attention mechanism to the original input features at the element level. Discarding the residual connection can prevent the original features that have not been adjusted by the attention weight from entering the deep network, ensuring the importance of the variables measured by the attention weight.

[0141] (3) Regression sampling convolution interaction model;

[0142] On the basis of the above embodiments, in some embodiments, a regression sampling convolution interaction model based on a binary tree structure is constructed according to the output result of the dot product attention structure, including:

[0143] S161: Copy the output result of the dot product attention structure to obtain a first copy and a second copy as the input of the first-layer binary tree node;

[0144] S162: Determine the weighted feature output result of the binary tree node at the current layer according to the first copy and the second copy;

[0145] S163: Copy the weighted feature output result to obtain the first copy and the second copy as the input of the next layer of binary tree nodes; return to step S162 until the binary tree structure reaches a preset number of layers to obtain the regression sampling convolution interaction model.

[0146] Figure 4 A schematic diagram of the structure of the sampling convolution interaction block provided in an embodiment of the present invention. Since the original sampling convolution interaction block (SCINet Block) downsamples the input sequence according to the characteristics of the time series to retain basic information, this operation is not recommended when processing non-time series input. Therefore, in this embodiment, Figure 4As shown, the output result of the dot product attention structure is specifically copied to obtain the first copy and the second copy as the input of the first-layer binary tree node. The weighted feature output result of the current-layer binary tree node is further determined based on the first copy and the second copy, and the weighted feature output result is copied to obtain the first copy and the second copy as the input of the next-layer binary tree node; return to the step of determining the weighted feature output result of the current-layer binary tree node based on the first copy and the second copy, until the binary tree structure reaches the preset number of layers, and the regression sampling convolution interaction model is obtained.

[0147] It should be noted that each binary tree node is a sampling convolution interaction block. The regression sampling convolution interaction model constructed with sampling convolution interaction blocks as nodes has a total of L layers of binary tree structure, where L is a positive integer. In this embodiment, there is no restriction on the specific value of L. Compared with the general convolution-based neural network model, the most significant advantage of the regression sampling convolution interaction model is that it has a rich set of independent convolution layers, which improves the model's ability to extract features. The binary tree with sampling convolution interaction blocks as nodes is the core part of the regression sampling convolution interaction model. Thanks to the copy operation of the sampling convolution interaction block, each sampling convolution interaction block node contains a complete set of input features.

[0148] In addition, in order to enhance the nonlinear fitting ability of the model and enhance the interactivity between data, in some embodiments, the sampling convolution interaction block includes four groups of one-dimensional convolution modules. and two exponential nonlinear modules, where is the first one-dimensional convolution module, is the second one-dimensional convolution module, is the third one-dimensional convolution module, It is the fourth one-dimensional convolution module. It should be noted that the four groups of one-dimensional convolution modules The structures of are the same, which are composed of a hyperbolic tangent (Tanh) activation function, a one-dimensional convolution layer, a random loss layer, a leaky linear rectification (LeakyRelu) activation function, and a one-dimensional convolution layer. Based on this, the process of determining the weighted feature output result of the current layer binary tree node according to the first copy and the second copy specifically includes:

[0149] S164: performing convolution operations on the first copy and the second copy respectively to generate a convolution result corresponding to the first copy and a convolution result corresponding to the second copy;

[0150] S165: performing exponential operation on the convolution result corresponding to the first copy and the convolution result corresponding to the second copy respectively, so as to generate an exponential result corresponding to the first copy and an exponential result corresponding to the second copy;

[0151] S166: Generate a weighted feature output result of the binary tree node at the current layer according to the first copy and its corresponding index result, and the second copy and its corresponding index result.

[0152] In order to determine the weighted feature output results, specifically for the first copy and the second copy Convolution operations are performed respectively to generate the convolution results corresponding to the first copy and the convolution results corresponding to the second copy; further exponential operations are performed on the convolution results corresponding to the first copy and the convolution results corresponding to the second copy to generate the exponential results corresponding to the first copy and the exponential results corresponding to the second copy. Finally, based on the first copy and its corresponding exponential results, and the second copy and its corresponding exponential results, the weighted feature output results of the binary tree nodes at the current layer are generated. The calculation process is described in detail below with reference to the specific formula:

[0153] The first copy and the second copy are respectively convolved to generate a convolution result corresponding to the first copy and a convolution result corresponding to the second copy. Specifically, the first copy is convolved by the first one-dimensional convolution module. Perform a one-dimensional convolution operation to obtain the first convolution result ; Through the second one-dimensional convolution module, the second copy Perform a one-dimensional convolution operation to obtain the second convolution result ;

[0154] Correspondingly, the convolution result corresponding to the first copy and the convolution result corresponding to the second copy are respectively subjected to exponential operation to generate the exponential result corresponding to the first copy and the exponential result corresponding to the second copy. Specifically, the first convolution result Apply the natural exponential function to get the first exponential result ; For the second convolution result Apply the natural exponential function to get the second exponential result .

[0155] Furthermore, according to the first copy and its corresponding index result, and the second copy and its corresponding index result, a weighted feature output result of the binary tree node at the current layer is generated, including:

[0156] S167: Perform a Hadamard product operation on the first replica and the second exponential result to obtain a first process quantity; the formula is as follows:

[0157] ;

[0158] in, is the first process quantity, For the first copy, For the second copy, is the second convolution result, is the second index result, is the Hadamard product.

[0159] S168: performing a Hadamard product operation on the second copy and the first exponential result to obtain a second process quantity;

[0160] ;

[0161] in, is the second process quantity, For the first copy, For the second copy, is the first convolution result, is the first index result, is the Hadamard product.

[0162] S169: Perform a one-dimensional convolution operation on the second process quantity through a fourth one-dimensional convolution module to obtain a fourth convolution result, and add the first process quantity and the fourth convolution result to obtain a first weighted feature output result; the formula is as follows:

[0163] ;

[0164] in, Output result for the first weighted feature, is the result of the fourth convolution.

[0165] S170: Perform a one-dimensional convolution operation on the first process quantity through a third one-dimensional convolution module to obtain a third convolution result, and subtract the second process quantity from the third convolution result to obtain a second weighted feature output result. The formula is as follows:

[0166] ;

[0167] in, Output result for the second weighted feature, is the result of the third convolution.

[0168] In summary, the calculation of weighted feature output results is realized.

[0169] Based on the above embodiments, in some embodiments, determining a single value output by the regression sampling convolution interaction model to complete the construction of the virtual measurement model includes:

[0170] S171: concatenating multiple feature matrices corresponding to the regression sampling convolution interaction model to obtain a concatenated matrix; wherein the size of each feature matrix corresponding to the regression sampling convolution interaction model is equal to the size of the mapping feature;

[0171] S172: After stacking the concatenated matrix with fully connected layers, an average value of the matrix is ​​calculated to obtain a single value, thereby completing the construction of the virtual measurement model.

[0172] Finally, a regression sampling convolution interaction model based on an L-layer binary tree was generated, and the mapping feature 2 of the same size L Feature matrices. The multiple feature matrices corresponding to the regression sampling convolution interaction model are further concatenated to obtain a concatenated matrix. After the concatenated matrix is ​​stacked through two fully connected layers, a matrix is ​​obtained, and the matrix is ​​averaged, and the mean is taken as the single value of the final output of the virtual measurement model. The average calculation enhances the robustness and generalization ability of the model. In this way, the construction of the virtual measurement model is completed.

[0173] Based on the above embodiments, in some embodiments, the training process of the virtual measurement model includes:

[0174] S18: obtaining a training data set consisting of a plurality of polishing process variables and their corresponding material removal rates;

[0175] S19: training the virtual measurement model according to the data in the training data set until the virtual measurement model converges to obtain the virtual measurement model.

[0176] In order to realize the training of the virtual measurement model, a training data set consisting of multiple polishing process variables and their corresponding material removal rates is obtained, and the virtual measurement model is trained according to the data in the training data set until the virtual measurement model converges to obtain the virtual measurement model.

[0177] In order to determine whether the virtual measurement model has converged, you can choose to observe the change of the loss function in the specific implementation. Specifically, during the training process, by observing the change of the loss function (loss) on the training set and the validation set, you can judge whether the model has converged. For example, if both the training loss (train loss) and the validation loss (val loss) decrease, it means that the network is still learning; if the training loss decreases, but the validation loss is stable, it may indicate that the network has begun to overfit; if the training loss is stable, but the validation loss decreases, it may indicate that there is a problem with the data set; if both the training loss and the validation loss are stable, it may indicate that the model has converged, or that the learning has encountered a bottleneck, you can try to reduce the learning rate; if both the training loss and the validation loss increase, it may indicate that there is a problem with the network structure design, or that the training parameters are improperly set, and you need to stop learning and adjust the code in time. You can also judge based on the training curve. By drawing the training curve, you can observe the convergence of the model more intuitively. For example, underfitting is usually manifested as high training loss and validation loss, and there is no obvious downward trend, indicating that the model has not fully learned the data features; overfitting is usually manifested as low training loss, but the validation loss starts to rise after a certain point, indicating that the model fits the training data too well and the generalization ability deteriorates; model convergence is usually manifested as both training loss and validation loss drop to a stable point, and the difference between the two is not large. In addition, the change of the gradient can be observed. Specifically, by calculating the norm of the gradient, it can be determined whether the gradient is close to zero, and thus whether the model has converged; if the norm of the gradient is very small and does not change much, it means that the model has converged. Finally, the above methods can be used to comprehensively determine whether the virtual measurement model has converged, and make corresponding adjustments and optimizations as needed to ensure the prediction accuracy of the model.

[0178] In order to improve the prediction accuracy of the virtual measurement model, based on the above embodiments, in some embodiments, after the virtual measurement model is trained according to the data in the training data set until the virtual measurement model converges, the method further includes:

[0179] S20: obtaining a test data set consisting of a plurality of polishing process variables and their corresponding material removal rates;

[0180] S21: Obtain the interpretable attention weight distribution matrix corresponding to the current virtual measurement model;

[0181] S22: testing the current virtual measurement model according to the data in the test data set to obtain the virtual measurement accuracy;

[0182] S23: Determine whether the virtual measurement accuracy reaches the preset accuracy value; if not, proceed to step S24; if yes, proceed to step S25;

[0183] S24: Filter data in the training data set according to the interpretable attention weight distribution matrix to obtain a new training data set, and return to step S19;

[0184] S25: Using the current virtual measurement model as the final virtual measurement model.

[0185] Specifically, after the model converges, a test data set consisting of multiple polishing process variables and their corresponding material removal rates is obtained, and an interpretable attention weight distribution matrix corresponding to the current virtual measurement model is obtained. It can be understood that the interpretable attention weight distribution matrix here is the interpretable attention weight distribution matrix obtained in the above embodiment.

[0186] The current virtual measurement model is further tested according to the data in the test data set to obtain the virtual measurement accuracy. It is determined whether the virtual measurement accuracy reaches the preset accuracy value. It should be noted that the preset accuracy value is not limited in this embodiment and depends on the specific implementation situation.

[0187] If the virtual measurement accuracy does not reach the preset accuracy value, the current virtual measurement model is directly used as the final virtual measurement model. If the virtual measurement accuracy does not reach the preset accuracy value, the data in the training data set is filtered according to the interpretable attention weight distribution matrix to obtain a new training data set, and the step of training the virtual measurement model according to the data in the training data set is returned to, so as to find more relevant new training data by data screening, and train the model again based on the new training data, thereby improving the model accuracy.

[0188] Based on the above embodiments, in some embodiments, after inputting each target polishing process variable into the virtual measurement model to output the target material removal rate corresponding to the target material, the method further includes:

[0189] S26: storing the target polishing process variables and the corresponding target material removal rates in the training data set to update the training data set;

[0190] S27: retraining the virtual measurement model based on the updated training data set.

[0191] In a specific implementation, in order to maintain the prediction accuracy of the virtual measurement model, after the virtual measurement model completes the prediction of the target material removal rate each time, the target polishing process variable and the corresponding target material removal rate can be stored in the training data set to update the training data set; and the virtual measurement model can be retrained based on the updated training data set. In this way, the virtual measurement model after iterative training can more accurately predict the target material removal rate corresponding to the target material.

[0192] In the above embodiments, the virtual measurement method is described in detail. The present invention also provides corresponding embodiments of the virtual measurement device.

[0193] Figure 5 Schematic diagram of a virtual measurement device provided by an embodiment of the present invention. Figure 5 As shown, the device comprises:

[0194] The acquisition module 10 is used to acquire various target polishing process variables that affect the polishing quality of the target material.

[0195] The prediction module 11 is used to input each target polishing process variable into the virtual measurement model to output a target material removal rate corresponding to the target material.

[0196] Among them, the construction process of the virtual measurement model includes: mapping each polishing process variable into mapping features; the mapping features retain the mathematical meaning of the corresponding polishing process variables; determining the output results of the dot product attention structure based on the mapping features and the interpretable dot product attention mechanism; and constructing a regression sampling convolution interaction model based on the output results of the dot product attention structure.

[0197] In some embodiments, in order to realize the construction of a virtual measurement model, each polishing process variable is mapped to a mapping feature, specifically mapping each polishing process variable in a training data set to a feature space; determining the position corresponding to each polishing process variable in the feature space to generate a mapping feature. The output result of the dot product attention structure is determined according to the mapping feature and the interpretable dot product attention mechanism, and the query, key and value of the interpretable dot product attention mechanism are obtained according to the mapping feature; wherein the query, key and value of the interpretable dot product attention mechanism retain the mathematical meaning of the polishing process variable; according to the query, key and value of the interpretable dot product attention mechanism, the interpretable attention weight distribution matrix and the output result of the dot product attention structure are obtained. A regression sampling convolution interaction model is constructed according to the output result of the dot product attention structure, and a regression sampling convolution interaction model based on a binary tree structure is constructed according to the output result of the dot product attention structure; wherein each node of the binary tree structure contains complete input features; determining a single value output by the regression sampling convolution interaction model to complete the construction of the virtual measurement model.

[0198] In some embodiments, the position corresponding to each polishing process variable in the feature space is determined to generate a mapping feature, specifically, primary features and their corresponding feature dimensions are extracted based on each polishing process variable; the position of each primary feature on the corresponding feature dimension is obtained according to each primary feature and its corresponding feature dimension; each primary feature and its position on the corresponding feature dimension are added to obtain the mapping feature of each polishing process variable.

[0199] In some embodiments, the query, key, and value of the interpretable dot product attention mechanism are obtained according to the mapping features, and three one-dimensional convolution filters are specifically set; the kernel width and step size of the one-dimensional convolution filter are both 1; the mapping features are respectively input into the three one-dimensional convolution filters to obtain the query, key, and value of the interpretable dot product attention mechanism respectively.

[0200] In some embodiments, based on the query, key and value of the interpretable dot product attention mechanism, an interpretable attention weight distribution matrix and a dot product attention structure output result are obtained, and specifically the product of the query and the key of the interpretable dot product attention mechanism is determined to obtain the interpretable attention weight distribution matrix; the product of the interpretable attention weight distribution matrix and the value of the interpretable dot product attention mechanism is determined to obtain the dot product attention structure output result.

[0201] In some embodiments, a regression sampling convolution interaction model based on a binary tree structure is constructed according to the output results of the dot product attention structure. Specifically, the output results of the dot product attention structure are copied to obtain a first copy and a second copy as the input of the first-layer binary tree node; the weighted feature output results of the current layer binary tree nodes are determined according to the first copy and the second copy; the weighted feature output results are copied to obtain a first copy and a second copy as the input of the next layer binary tree node; and the step of determining the weighted feature output results of the current layer binary tree nodes according to the first copy and the second copy is returned to, until the binary tree structure reaches a preset number of layers to obtain the regression sampling convolution interaction model.

[0202] In some embodiments, the weighted feature output results of the binary tree nodes of the current layer are determined based on the first copy and the second copy, and specifically, convolution operations are performed on the first copy and the second copy, respectively, to generate a convolution result corresponding to the first copy and a convolution result corresponding to the second copy; exponential operations are performed on the convolution results corresponding to the first copy and the convolution results corresponding to the second copy, respectively, to generate an exponential result corresponding to the first copy and an exponential result corresponding to the second copy; the weighted feature output results of the binary tree nodes of the current layer are generated based on the first copy and its corresponding exponential result, and the second copy and its corresponding exponential result.

[0203] In some embodiments, convolution operations are performed on the first copy and the second copy, respectively, to generate a convolution result corresponding to the first copy and a convolution result corresponding to the second copy. Specifically, a first one-dimensional convolution operation is performed on the first copy by a first one-dimensional convolution module to obtain a first convolution result; a second one-dimensional convolution operation is performed on the second copy by a second one-dimensional convolution module to obtain a second convolution result; correspondingly, exponential operations are performed on the convolution result corresponding to the first copy and the convolution result corresponding to the second copy, respectively, to generate an exponential result corresponding to the first copy and an exponential result corresponding to the second copy, including: applying a natural exponential function to the first convolution result to obtain a first exponential result; applying a natural exponential function to the second convolution result to obtain a second exponential result.

[0204] In some embodiments, based on the first copy and its corresponding exponential result, and the second copy and its corresponding exponential result, a weighted feature output result of the current layer binary tree node is generated, specifically, a Hadamard product operation is performed on the first copy and the second exponential result to obtain a first process quantity; a Hadamard product operation is performed on the second copy and the first exponential result to obtain a second process quantity; a one-dimensional convolution operation is performed on the second process quantity through a fourth one-dimensional convolution module to obtain a fourth convolution result, and the first process quantity is added to the fourth convolution result to obtain a first weighted feature output result; a one-dimensional convolution operation is performed on the first process quantity through a third one-dimensional convolution module to obtain a third convolution result, and the second process quantity is subtracted from the third convolution result to obtain a second weighted feature output result.

[0205] In some embodiments, the structures of the first one-dimensional convolution module, the second one-dimensional convolution module, the third one-dimensional convolution module and the fourth one-dimensional convolution module are the same, and are composed of a hyperbolic tangent activation function, a one-dimensional convolution layer, a random loss layer, a leaky linear rectification activation function and a one-dimensional convolution layer in sequence.

[0206] In some embodiments, a single value output by the regression sampling convolution interaction model is determined to complete the construction of the virtual measurement model. Specifically, multiple feature matrices corresponding to the regression sampling convolution interaction model are spliced ​​to obtain a spliced ​​matrix; wherein the size of each feature matrix corresponding to the regression sampling convolution interaction model is equal to the size of the mapping feature; the spliced ​​matrix is ​​stacked with a fully connected layer and the matrix average is obtained to obtain a single value to complete the construction of the virtual measurement model.

[0207] In some embodiments, the training process of the virtual measurement model specifically obtains a training data set consisting of multiple polishing process variables and their corresponding material removal rates; trains the virtual measurement model according to the data in the training data set until the virtual measurement model converges to obtain the virtual measurement model.

[0208] In some embodiments, after the virtual measurement model is trained according to the data in the training data set until the virtual measurement model converges, a test data set consisting of multiple polishing process variables and their corresponding material removal rates is further obtained; an interpretable attention weight distribution matrix corresponding to the current virtual measurement model is obtained; the current virtual measurement model is tested according to the data in the test data set to obtain the virtual measurement accuracy; it is determined whether the virtual measurement accuracy reaches a preset accuracy value; if not, the data is filtered in the training data set according to the interpretable attention weight distribution matrix to obtain a new training data set, and the step of training the virtual measurement model according to the data in the training data set is returned to; if so, the current virtual measurement model is used as the final virtual measurement model.

[0209] In some embodiments, it also includes:

[0210] An updating module, used for storing the target polishing process variables and the corresponding target material removal rates in the training data set to update the training data set;

[0211] The iterative training module is used to retrain the virtual measurement model based on the updated training data set.

[0212] Since the embodiments of the apparatus part correspond to the embodiments of the method part, please refer to the description of the embodiments of the method part for the embodiments of the apparatus part, which will not be repeated here.

[0213] In addition, the present invention also provides a computer program product, including a computer program or instructions, which implement the steps of the above virtual measurement method when executed by a processor.

[0214] Figure 6 A schematic diagram of a virtual measuring device provided by an embodiment of the present invention, such as Figure 6 As shown, the virtual measuring device includes:

[0215] A memory 20, used for storing computer programs;

[0216] The processor 21 is used to implement the steps of the virtual measurement method mentioned in the above embodiment when executing the computer program.

[0217] The virtual measurement device provided in this embodiment may include but is not limited to a smart phone, a tablet computer, a laptop computer, or a desktop computer.

[0218] Among them, the processor 21 may include one or more processing cores, such as a 4-core processor, an 8-core processor, etc. The processor 21 can be implemented in at least one hardware form of a digital signal processor (DSP), a field programmable gate array (FPGA), and a programmable logic array (PLA). The processor 21 may also include a main processor and a coprocessor. The main processor is a processor for processing data in an awake state, also known as a central processing unit (CPU); the coprocessor is a low-power processor for processing data in a standby state. In some embodiments, the processor 21 may be integrated with a graphics processing unit (GPU), and the GPU is responsible for rendering and drawing the content to be displayed on the display screen. In some embodiments, the processor 21 may also include an artificial intelligence (AI) processor, which is used to process computing operations related to machine learning.

[0219] The memory 20 may include one or more computer-readable storage media, which may be non-transitory. The memory 20 may also include a high-speed random access memory, and a non-volatile memory, such as one or more disk storage devices, flash memory storage devices. In this embodiment, the memory 20 is at least used to store the following computer program 201, wherein the computer program can implement the relevant steps of the virtual measurement method disclosed in any of the aforementioned embodiments after being loaded and executed by the processor 21. In addition, the resources stored in the memory 20 may also include an operating system 202 and data 203, etc., and the storage method may be temporary storage or permanent storage. Among them, the operating system 202 may include Windows, Unix, Linux, etc. Data 203 may include but is not limited to data related to the virtual measurement method.

[0220] In some embodiments, the virtual measuring device may further include a display screen 22 , an input / output interface 23 , a communication interface 24 , a power supply 25 , and a communication bus 26 .

[0221] Those skilled in the art will understand that Figure 6 The structure shown in the figure does not constitute a limitation of the virtual measuring device, and may include more or less components than those shown in the figure.

[0222] Finally, the present invention also provides an embodiment corresponding to a computer-readable storage medium. The computer-readable storage medium stores a computer program, and when the computer program is executed by a processor, the steps recorded in the above method embodiment are implemented.

[0223] It is understandable that if the method in the above embodiment is implemented in the form of a software functional unit and sold or used as an independent product, it can be stored in a computer-readable storage medium. Based on this understanding, the technical solution of the present invention, or the part that contributes to the prior art, or all or part of the technical solution can be embodied in the form of a software product, and the computer software product is stored in a storage medium to execute all or part of the steps of the method described in each embodiment of the present invention. The aforementioned storage medium includes: U disk, mobile hard disk, read-only memory (ROM), random access memory (RAM), disk or optical disk, etc. Various media that can store program codes.

[0224] The above is a detailed introduction to a virtual measurement method, product, device and medium provided by the present invention. The various embodiments in the specification are described in a progressive manner, and each embodiment focuses on the differences from other embodiments. The same and similar parts between the embodiments can be referenced to each other. For the device disclosed in the embodiment, since it corresponds to the method disclosed in the embodiment, the description is relatively simple, and the relevant parts can be referred to the method part description. It should be pointed out that for ordinary technicians in this technical field, without departing from the principle of the present invention, several improvements and modifications can be made to the present invention, and these improvements and modifications also fall within the scope of protection of the present invention.

[0225] It should also be noted that, in this specification, relational terms such as first and second, etc. are only used to distinguish one entity or operation from another entity or operation, and do not necessarily require or imply any such actual relationship or order between these entities or operations. Moreover, the terms "comprise", "include" or any other variants thereof are intended to cover non-exclusive inclusion, so that a process, method, article or device including a series of elements includes not only those elements, but also other elements not explicitly listed, or also includes elements inherent to such process, method, article or device. In the absence of further restrictions, an element defined by the statement "comprises a ..." does not exclude the presence of other identical elements in the process, method, article or device including the element.

Claims

1. A virtual measurement method, characterized in that: include: Obtaining various target polishing process variables that affect the polishing quality of the target material; Inputting each of the target polishing process variables into a virtual measurement model to output a target material removal rate corresponding to the target material; Among them, the construction process of the virtual measurement model includes: mapping each polishing process variable into a mapping feature; the mapping feature retains the mathematical meaning of the corresponding polishing process variable; determining the output result of the dot product attention structure according to the mapping feature and the interpretable dot product attention mechanism; and constructing a regression sampling convolution interaction model according to the output result of the dot product attention structure.

2. The virtual measurement method according to claim 1, characterized in that: Mapping each of the polishing process variables into the mapping feature comprises: Mapping each of the polishing process variables in the training data set to a feature space; The position corresponding to each of the polishing process variables in the feature space is determined to generate the mapping feature.

3. The virtual measurement method according to claim 2, characterized in that: Determining the output result of the dot product attention structure according to the mapping features and the interpretable dot product attention mechanism includes: Obtaining a query, key, and value of an interpretable dot product attention mechanism according to the mapping feature; wherein the query, key, and value of the interpretable dot product attention mechanism retain the mathematical meaning of the polishing process variable; According to the query, key and value of the interpretable dot product attention mechanism, an interpretable attention weight distribution matrix and the output result of the dot product attention structure are obtained.

4. The virtual measurement method according to claim 3, characterized in that: The regression sampling convolution interaction model is constructed according to the output result of the dot product attention structure, including: Constructing the regression sampling convolution interaction model based on the binary tree structure according to the output result of the dot product attention structure; wherein each node of the binary tree structure contains complete input features; A single value output by the regression sampling convolution interaction model is determined to complete the construction of the virtual measurement model.

5. The virtual measurement method according to any one of claims 2 to 4, characterized in that: Determining the position corresponding to each of the polishing process variables in the feature space to generate the mapping feature includes: Extracting primary features and their corresponding feature dimensions based on each of the polishing process variables; According to each of the primary features and the corresponding feature dimensions, obtaining a position of each of the primary features on the corresponding feature dimensions; Each of the primary features and its position on the corresponding feature dimension are added together to obtain the mapping feature of each of the polishing process variables.

6. The virtual measurement method according to claim 3 or 4, characterized in that: Obtaining the query, key, and value of the interpretable dot-product attention mechanism based on the mapped features, including: Three one-dimensional convolution filters are set; the kernel width and step size of the one-dimensional convolution filters are both 1; The mapped features are respectively input into three one-dimensional convolutional filters to obtain the query, key and value of the interpretable dot product attention mechanism.

7. The virtual measurement method according to claim 6, characterized in that: According to the query, key and value of the interpretable dot product attention mechanism, an interpretable attention weight distribution matrix and the output result of the dot product attention structure are obtained, including: Determine the product of the query and the key of the interpretable dot product attention mechanism to obtain the interpretable attention weight distribution matrix; Determine the product of the interpretable attention weight distribution matrix and the value of the interpretable dot product attention mechanism to obtain the dot product attention structure output result.

8. The virtual measurement method according to claim 4, characterized in that: Constructing the regression sampling convolution interaction model based on the binary tree structure according to the output result of the dot product attention structure, including: The output result of the dot product attention structure is copied to obtain a first copy and a second copy as inputs of the first-layer binary tree nodes; Determine the weighted feature output result of the binary tree node at the current layer according to the first copy and the second copy; The weighted feature output result is copied to obtain a first copy and a second copy as inputs to a binary tree node of a next layer; Return to the step of determining the weighted feature output results of the binary tree nodes of the current layer according to the first copy and the second copy, until the binary tree structure reaches a preset number of layers to obtain the regression sampling convolution interaction model.

9. The virtual measurement method according to claim 8, characterized in that: Determine the weighted feature output result of the binary tree node at the current layer according to the first copy and the second copy, including: Performing convolution operations on the first copy and the second copy respectively to generate a convolution result corresponding to the first copy and a convolution result corresponding to the second copy; Performing exponential operations on the convolution result corresponding to the first copy and the convolution result corresponding to the second copy respectively to generate an exponential result corresponding to the first copy and an exponential result corresponding to the second copy; The weighted feature output result of the binary tree node at the current layer is generated according to the first copy and its corresponding index result, and the second copy and its corresponding index result.

10. The virtual measurement method according to claim 9, characterized in that: Performing convolution operations on the first copy and the second copy respectively to generate a convolution result corresponding to the first copy and a convolution result corresponding to the second copy, including: Performing a one-dimensional convolution operation on the first copy through a first one-dimensional convolution module to obtain a first convolution result; Performing a one-dimensional convolution operation on the second copy through a second one-dimensional convolution module to obtain a second convolution result; Correspondingly, performing exponential operations on the convolution result corresponding to the first copy and the convolution result corresponding to the second copy respectively to generate an exponential result corresponding to the first copy and an exponential result corresponding to the second copy, including: Applying a natural exponential function to the first convolution result to obtain a first exponential result; Applying a natural exponential function to the second convolution result obtains a second exponential result.

11. The virtual measurement method according to claim 10, characterized in that: According to the first copy and its corresponding index result, and the second copy and its corresponding index result, the weighted feature output result of the binary tree node of the current layer is generated, including: Performing a Hadamard product operation on the first copy and the second exponential result to obtain a first process quantity; Performing a Hadamard product operation on the second copy and the first exponential result to obtain a second process quantity; Performing a one-dimensional convolution operation on the second process quantity through a fourth one-dimensional convolution module to obtain a fourth convolution result, and adding the first process quantity and the fourth convolution result to obtain a first weighted feature output result; A third one-dimensional convolution module is used to perform a one-dimensional convolution operation on the first process quantity to obtain a third convolution result, and the second process quantity is subtracted from the third convolution result to obtain a second weighted feature output result.

12. The virtual measurement method according to claim 11, characterized in that: The structures of the first one-dimensional convolution module, the second one-dimensional convolution module, the third one-dimensional convolution module and the fourth one-dimensional convolution module are the same, and are composed of a hyperbolic tangent activation function, a one-dimensional convolution layer, a random loss layer, a leaky linear rectification activation function and a one-dimensional convolution layer in sequence.

13. The virtual measurement method according to claim 4, characterized in that: Determining a single value output by the regression sampling convolution interaction model to complete the construction of the virtual measurement model includes: Splicing multiple feature matrices corresponding to the regression sampling convolution interaction model to obtain a spliced ​​matrix; wherein the size of each feature matrix corresponding to the regression sampling convolution interaction model is equal to the size of the mapping feature; The concatenated matrix is ​​stacked with fully connected layers and the matrix average value is calculated to obtain the single value, thereby completing the construction of the virtual measurement model.

14. The virtual measurement method according to claim 1, characterized in that: The training process of the virtual measurement model includes: Obtaining a training data set consisting of a plurality of polishing process variables and their corresponding material removal rates; The virtual measurement model is trained according to the data in the training data set until the virtual measurement model converges to obtain the virtual measurement model.

15. The virtual measurement method according to claim 14, characterized in that: After the virtual measurement model is trained according to the data in the training data set until the virtual measurement model converges, the method further includes: Acquire a test data set consisting of a plurality of polishing process variables and their corresponding material removal rates; Obtaining an interpretable attention weight distribution matrix corresponding to the current virtual measurement model; Testing the current virtual measurement model according to the data in the test data set to obtain the virtual measurement accuracy; Determining whether the virtual measurement accuracy reaches a preset accuracy value; If not, filtering data in the training data set according to the interpretable attention weight distribution matrix to obtain a new training data set, and returning to the step of training the virtual measurement model according to the data in the training data set; If so, the current virtual measurement model is used as the final virtual measurement model.

16. The virtual measurement method according to claim 1, characterized in that: After inputting each of the target polishing process variables into the virtual measurement model to output the target material removal rate corresponding to the target material, the method further includes: Storing the target polishing process variables and the corresponding target material removal rates in a training data set to update the training data set; The virtual measurement model is retrained based on the updated training data set.

17. A computer program product comprising a computer program or instructions, characterized in that When the computer program or instruction is executed by a processor, the steps of the virtual measurement method according to any one of claims 1 to 16 are implemented.

18. A virtual measuring device, characterized in that: include: Memory for storing computer programs; A processor, configured to implement the steps of the virtual measurement method according to any one of claims 1 to 16 when executing the computer program.

19. A computer-readable storage medium, characterized in that: The computer-readable storage medium stores a computer program, and when the computer program is executed by a processor, the steps of the virtual measurement method according to any one of claims 1 to 16 are implemented.

Images