Wafer polishing defect prediction method and device, electronic equipment and storage medium

Abstract

The invention provides a wafer polishing defect prediction method and device, electronic equipment and a storage medium. The method comprises the following steps: acquiring parameter information of each component in the chemical mechanical planarization equipment within a preset time period; determining an overall state matrix of the chemical mechanical planarization equipment according to the use sequence and the parameter information of each component in the wafer polishing process; wherein the overall state matrix is a matrix formed by a two-dimensional time sequence with continuous time dimensions; the overall state matrix is input into a preset polishing defect prediction model, and a defect type prediction result, output by the polishing defect prediction model, of wafer polishing conducted by the chemical mechanical planarization equipment within a preset time period is obtained; wherein the polishing defect prediction model is obtained by training according to historical state matrixes of the chemical mechanical planarization equipment under different defect types and the corresponding defect types. According to the method, the wafer polishing defects can be quickly and accurately obtained, and the polishing cost is reduced.

Description

Technical Field

[0001] The present invention relates to the field of semiconductor technology, and in particular to a method, device, electronic equipment and storage medium for predicting wafer polishing defects. Background Art

[0002] Chemical-Mechanical Planarization (CMP) is a global surface flattening technology, which is mainly used to grind the thin film on the surface of the wafer. This technology usually makes the wafer and the polishing pad move relative to each other under a certain pressure and in the presence of polishing liquid, and uses the organic combination of the mechanical grinding effect of the abrasive and the chemical corrosion effect of the polishing liquid to peel off the material on the surface of the wafer layer by layer, thereby achieving a highly flattened effect.

[0003] During the grinding process, due to factors such as wear of consumables, loss of components, fluctuation of parameters and uneven distribution of grinding fluid, various defects will appear on the ground wafers, resulting in a decrease in the yield of the wafers.

[0004] In the related art, the polishing process generally involves multiple rounds of polishing. After multiple rounds of polishing, the quality of the wafer, the defect status and yield rate of the wafer can only be obtained during the subsequent wafer inspection. The defects of wafer polishing cannot be obtained in time, which increases the polishing cost of CMP. Summary of the invention

[0005] The embodiments of the present invention provide a method, device, electronic device and storage medium for predicting wafer polishing defects, so as to predict wafer polishing defects, quickly and accurately obtain wafer polishing defects, reduce the generation of defective wafers, and reduce polishing costs.

[0006] In a first aspect, an embodiment of the present invention provides a method for predicting wafer polishing defects, comprising:

[0007] Obtaining parameter information of each component in the chemical mechanical planarization device within a preset time period;

[0008] Determine the overall state matrix of the chemical mechanical planarization device according to the usage order of each component in the wafer polishing process and the parameter information; wherein the overall state matrix is ​​a matrix formed by a two-dimensional time series with a continuous time dimension;

[0009] Inputting the overall state matrix into a preset polishing defect prediction model to obtain a defect type prediction result of the chemical mechanical planarization device performing wafer polishing within the preset time period output by the polishing defect prediction model;

[0010] The polishing defect prediction model is obtained by training based on the historical state matrix of the chemical mechanical planarization equipment under different defect types and the corresponding defect types.

[0011] In a possible implementation, determining the overall state matrix of the chemical mechanical planarization device according to the usage order of each component in the wafer polishing process and the parameter information includes:

[0012] Connecting the time dimensions in the parameter information of each component according to the order in which each component in the chemical mechanical planarization device is used during the wafer polishing process;

[0013] Determine the parameter dimensions in the parameter information according to the common parameters and unique parameters of each component;

[0014] Based on the time dimension and the parameter dimension, a state matrix of the chemical mechanical planarization device is determined.

[0015] In a possible implementation, the overall state matrix is ​​input into a preset polishing defect prediction model to obtain a defect type prediction result of the chemical mechanical planarization device performing wafer polishing within the preset time period output by the polishing defect prediction model, including:

[0016] Inputting the overall state matrix into a feature extraction module in a preset polishing defect prediction model to obtain an overall feature matrix of the chemical mechanical planarization device output by the feature extraction module;

[0017] The first vector of the overall feature matrix is ​​input into the classification module in the polishing defect model to obtain the defect type prediction result of the chemical mechanical planarization equipment performing wafer polishing within the preset time period output by the classification module.

[0018] In a possible implementation, the overall state matrix is ​​input into a feature extraction module in a preset polishing defect prediction model to obtain an overall feature matrix of the chemical mechanical planarization device output by the feature extraction module, including:

[0019] Inputting the overall state matrix into the first conversion coding layer of the feature extraction module in the preset polishing defect prediction model to obtain the state features output by the first conversion coding layer;

[0020] For each conversion coding layer of the second layer and thereafter, the state feature output by the previous layer is input to the conversion coding layer to obtain the state feature output by the conversion coding layer;

[0021] Based on the state features output by the last conversion coding layer, the overall feature matrix is ​​obtained;

[0022] In which, in the attention sublayer of each conversion coding layer, multiple local attention mechanisms are set, and the output result of the attention sublayer is determined by the output result of the local attention mechanism.

[0023] In a possible implementation, the processing process of the attention sublayer is:

[0024] For each time dimension of the input, determine a preset number of attention time dimensions corresponding to the time dimension, and determine the local attention of the data corresponding to the attention time dimension according to the local attention mechanism;

[0025] Based on the local attention of all time dimensions, the output result of the attention sublayer is determined.

[0026] In a possible implementation, after inputting the overall state matrix to the first conversion coding layer of the feature extraction module in the preset polishing defect prediction model to obtain the state features output by the first conversion coding layer, the method further includes:

[0027] Performing a time dimension reduction process on the state features output by the first conversion coding layer;

[0028] Correspondingly, for each conversion coding layer of the second layer and thereafter, after inputting the state feature output by the previous layer into the conversion coding layer and obtaining the state feature output by the conversion coding layer, the method further includes:

[0029] The state features output by the conversion coding layer are subjected to dimensionality reduction processing in the time dimension.

[0030] In a possible implementation, before inputting the overall state matrix into a preset polishing defect prediction model to obtain a defect type prediction result of the chemical mechanical planarization device performing wafer polishing within the preset time period output by the polishing defect prediction model, the method further includes:

[0031] Acquiring historical parameter information of the chemical mechanical planarization under different defect types and corresponding defect types;

[0032] Determining a historical state matrix of the chemical mechanical planarization device according to the historical parameter information;

[0033] According to the historical state matrix and the corresponding defect types, the polishing defect prediction model is trained, and the model parameters are updated to obtain a trained polishing defect prediction model; wherein the polishing defect prediction model includes a feature extraction module and a classification module.

[0034] In a second aspect, an embodiment of the present invention provides a device for predicting wafer polishing defects, comprising:

[0035] An acquisition module, used to acquire parameter information of each component in the chemical mechanical planarization device within a preset time period;

[0036] A determination module, used to determine the overall state matrix of the chemical mechanical planarization device according to the use order of each component in the wafer polishing process and the parameter information; wherein the overall state matrix is ​​a matrix formed by a two-dimensional time series with a continuous time dimension;

[0037] A prediction module is used to input the overall state matrix into a preset polishing defect prediction model to obtain a defect type prediction result of the chemical mechanical planarization equipment performing wafer polishing within the preset time period output by the polishing defect prediction model; wherein the polishing defect prediction model is trained based on the historical state matrix of the chemical mechanical planarization equipment under different defect types and the corresponding defect types.

[0038] In a third aspect, an embodiment of the present invention provides an electronic device, comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein when the processor executes the computer program, the steps of the method described in the first aspect or any possible implementation method of the first aspect are implemented.

[0039] In a fourth aspect, an embodiment of the present invention provides a computer-readable storage medium, wherein the computer-readable storage medium stores a computer program, and when the computer program is executed by a processor, the steps of the method described in the first aspect or any possible implementation method of the first aspect are implemented.

[0040] Compared with the prior art, the embodiments of the present invention have the following beneficial effects:

[0041] The embodiment of the present invention can accurately obtain the processing status of the chemical mechanical planarization equipment during the wafer polishing process by acquiring the parameter information of each component in the chemical mechanical planarization equipment within a preset time period; determine the overall state matrix of the chemical mechanical planarization equipment according to the use order and parameter information of each component during the wafer polishing process, and connect the parameter information according to the time sequence of the components and collect them into the overall state matrix, so as to facilitate the subsequent prediction of the polishing defects of the wafer. By inputting the state matrix into the polishing defect prediction model, the defect type prediction result of the wafer polishing of the chemical mechanical planarization equipment within the preset time period is obtained, and the defects of wafer polishing can be predicted, and the defects of wafer polishing can be obtained quickly and accurately, and the process of chemical mechanical planarization can be adjusted in time to reduce the generation of defective wafers and reduce the polishing cost. BRIEF DESCRIPTION OF THE DRAWINGS

[0042] In order to more clearly illustrate the technical solutions in the embodiments of the present invention, the drawings required for use in the embodiments or the description of the prior art will be briefly introduced below. 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.

[0043] Figure 1 is a flow chart for implementing a method for predicting wafer polishing defects provided by an embodiment of the present invention;

[0044] Figure 2 is a schematic diagram of the structure of a polishing defect prediction model provided by an embodiment of the present invention;

[0045] Figure 3 is a schematic diagram of an implementation of the attention mechanism provided by an embodiment of the present invention;

[0046] Figure 4 is a schematic structural diagram of a device for predicting wafer polishing defects provided by an embodiment of the present invention;

[0047] Figure 5 is a schematic diagram of an electronic device provided by an embodiment of the present invention. DETAILED DESCRIPTION

[0048] In the following description, specific details such as specific system structures, technologies, etc. are provided for the purpose of illustration rather than limitation, so as to provide a thorough understanding of the embodiments of the present invention. However, it should be clear to those skilled in the art that the present invention may be implemented in other embodiments without these specific details. In other cases, detailed descriptions of well-known systems, devices, circuits, and methods are omitted to prevent unnecessary details from obstructing the description of the present invention.

[0049] Chemical mechanical planarization is a type of global surface planarization technology and a chemically enhanced mechanical polishing technology. This technology uses abrasive and corrosive grinding fluids, and cooperates with polishing pads and polishing heads to grind the wafer. During the grinding process, the material and irregular structures on the surface of the wafer are removed to achieve the purpose of planarization. The flattened wafer surface makes it easier to form patterns in dry etching and makes smaller metal patterns possible, thereby improving circuit integration.

[0050] The quality of wafers after polishing by chemical mechanical planarization equipment is affected by many factors, among which the polishing process will have a significant impact on the quality of wafers. Different characteristics of wafers require different polishing parameters, which are manually set through the grinding menu.

[0051] The inventors have found that the values ​​of the above parameters are usually set by manual experience, and the quality of the wafers is not known until the wafers are inspected after multiple groups of wafers are polished, which results in low efficiency in problem solving. In the actual polishing process, the actual values ​​of the polish component parameters will fluctuate within the set value range, and after a period of polishing, due to factors such as the wear of consumables, various losses of components, fluctuations in parameters, and uneven distribution of the polishing liquid, the current set values ​​will no longer be suitable for the processing of the current wafer, resulting in a decrease in the yield rate of the subsequent polished wafers and an increase in the probability of defects, resulting in a decrease in the yield rate.

[0052] In order to accurately obtain the defects after wafer polishing, in an embodiment of the present invention, after a round of polishing, the polishing defects of the wafer are predicted. If defects appear on the surface of the wafer, an alarm can be issued to perform machine inspection or update the grinding recipe. Among them, the corresponding overall state matrix is ​​constructed through the parameter information of each component in the chemical mechanical planarization equipment, and the complete processing situation of the wafer during polishing can be accurately obtained, and then input into the preset polishing defect prediction model for prediction, and the defect type prediction result of wafer polishing can be accurately obtained, so as to timely adjust the chemical mechanical planarization process, reduce the generation of defective wafers, save polishing costs, and improve the yield rate of wafer polishing.

[0053] In addition, the inventor has previously directly predicted wafer polishing defects through the parameter information of each component in the chemical mechanical planarization equipment. However, the inventor found that in this defect prediction method, due to the different number of parameters in each component in the chemical mechanical planarization equipment and the different corresponding usage times, it is impossible to directly construct the entire chemical mechanical planarization equipment, and only the separate parameter information of each component can be obtained. Moreover, when making predictions, only each component can be processed separately, and the various components and steps in the chemical mechanical planarization process are interrelated for the wafer. If each component is separated and considered, it is difficult to accurately predict wafer defects, and there is still room for further prediction accuracy. The specific content will be described in detail in the following embodiments.

[0054] In order to make the purpose, technical solutions and advantages of the present invention more clear, specific embodiments will be described below in conjunction with the accompanying drawings.

[0055] Figure 1 The implementation flow chart of the method for predicting wafer polishing defects provided by the embodiment of the present invention is described in detail as follows:

[0056] Step S101 , obtaining parameter information of each component in a chemical mechanical planarization device within a preset time period.

[0057] In this embodiment, the wafer quality is affected by the polish process. The polish process is composed of several components. Each component has several sensors and recipe parameters. The information corresponding to each recipe parameter can be obtained through the sensor, thereby obtaining the parameter information.

[0058] Here, the preset time period may include multiple time frames, and each parameter is sampled once at intervals of the preset time period as a time frame.

[0059] For a component, suppose the component has n i parameters, corresponding to t i time frames, all parameter information of the component can be obtained by an n i ×t i The matrix X i To express, it can be specifically:

[0060]

[0061] Among them, X i Indicates the parameter information of the i-th component, n i represents the set of all parameters in the i-th component, t i represents the set of time frames corresponding to the i-th component, x i represents the i-th component.

[0062] Accordingly, the parameter information of each component in the chemical mechanical planarization device can form a tuple, for example: X = (X1, X2, X3, ..., X k ), wherein X represents a tuple consisting of parameter information of each component in the chemical mechanical planarization device, X i represents the parameter information of the i-th component, and k represents the total number of all components in the chemical mechanical planarization equipment.

[0063] Optionally, after obtaining the parameter information of each component in the chemical mechanical planarization device within a preset time period, the parameter information may be subjected to denoising and normalization processing.

[0064] In this embodiment, the parameter information of each component can constitute multivariate time series data. When performing denoising, the time series of each dimension of the multivariate time series data can be Fourier transformed to convert it into frequency domain data, and then the noise frequency is filtered through a fixed Hz high-pass filter, and then converted into a filtered time series through an inverse Fourier transform, thereby achieving denoising.

[0065] Since the dimensions and distributions of each dimension of multivariate time series data are different, the data of each dimension are normalized to zero mean, and the mean of the data of all dimensions is set to 0 and the variance to 1.

[0066] The specific formula can be: Among them, x j represents the j-th dimension of multivariate time series data, x ju represents the u-th data in the j-th dimension of multivariate time series data, represents the normalized x ju ,mean(x j ) represents x j The mean, std(x j ) represents x j The variance of x is assumed to be normal when the number is large enough.

[0067] Step S102, determining the overall state matrix of the chemical mechanical planarization device according to the usage sequence and parameter information of each component in the wafer polishing process; wherein the overall state matrix is ​​a matrix formed by a two-dimensional time series with a continuous time dimension.

[0068] In this embodiment, the various components and steps in the chemical mechanical planarization process are interrelated and mutually influential for the wafer, and the acquired parameter information is continuous in the time dimension and should not be separated according to the components, but should be processed continuously as a whole. Therefore, in this embodiment, the parameter information is processed according to the order in which the components are used in the wafer polishing process, and an overall state matrix is ​​constructed to form a continuous sequence of the parameter information of the various components in the chemical mechanical planarization process, so that the adjacent or related parameter information can be fully considered in the subsequent prediction.

[0069] Here, by processing the parameter information, a continuous time series can be formed in the time dimension, thereby forming a corresponding matrix.

[0070] In the prediction method previously discovered by the inventors, defect prediction is performed directly after obtaining the parameter information of each component. At this time, the parameter information of each component is separated and exists independently. In the present embodiment, the acquired parameter information is further processed to obtain an overall state matrix of a chemical mechanical planarization device, so that the parameter information of each component is continuous and related in the time dimension, thereby improving the accuracy of defect prediction.

[0071] Step S103, input the overall state matrix into a preset polishing defect prediction model, and obtain the defect type prediction result of the chemical mechanical planarization equipment performing wafer polishing within a preset time period output by the polishing defect prediction model; wherein the polishing defect prediction model is trained based on the historical state matrix and corresponding defect types of the chemical mechanical planarization equipment under different defect types.

[0072] In this embodiment, the overall state matrix is ​​processed through the polishing defect prediction model, and the parameter information of each component in the chemical mechanical planarization process can be processed at the same time, fully considering the correlation between each component, and accurately obtaining the defect type prediction result of wafer polishing.

[0073] Here, assuming that there are p types of defect types, the polishing defect prediction model can obtain the probability results of each defect type to form a p-dimensional vector, that is, Y = [y1, y2, y3, ..., y p ]; where Y represents the vector formed by the probability results of each defect type, y p Represents the probability corresponding to the pth defect type.

[0074] The calculation process of the polishing defect prediction model can be regarded as a nonlinear multivariate function f θ , where is the parameter to be learned, then the mapping relationship between Y and the overall state matrix X′ can be expressed as:

[0075] Y=f θ (X′) = f θ ([X1′,X2′,X3′,…,X k ′]);

[0076] Among them, X′ represents the overall state matrix, X k ′ represents the matrix composed of the data corresponding to the kth component in the overall state matrix.

[0077] In the prediction method previously discovered by the inventor, since the parameter information of each component in the process of chemical mechanical planarization is separated, the number of parameters of each component is different, and the corresponding time dimension is different, therefore, when using the model for prediction, only the parameter information of each component can be processed separately to achieve prediction, and it is difficult to consider the association between the parameter information of the components and the corresponding implicit information. However, this embodiment inputs the overall state matrix into the polishing defect prediction model. Since the overall state matrix is ​​a matrix composed of a two-dimensional time series, the matrix is ​​a whole and continuous in time, the overall state matrix can be processed as a whole, without processing each component separately, and the influence between the parameters of each component can be fully considered to accurately predict the defects of wafer polishing.

[0078] Optional, see Figure 2The structural schematic diagram of the polishing defect prediction model shown in the figure, in this embodiment, the overall state matrix is ​​input into the preset polishing defect prediction model to obtain the defect type prediction result of the chemical mechanical planarization equipment performing wafer polishing within a preset time period output by the polishing defect prediction model, and the overall state matrix X′ can be first input into the feature extraction module in the preset polishing defect prediction model to obtain the overall feature matrix A of the chemical mechanical planarization equipment output by the feature extraction module; and then the first vector B of the overall feature matrix A is input into the classification module in the polishing defect model to obtain the defect type prediction result Y of the chemical mechanical planarization equipment performing wafer polishing within the preset time period output by the classification module.

[0079] In this embodiment, the polishing defect prediction model may include two parts, one part is a feature extraction module, which may be composed of an encoder, and the other part is a classification module, which may be composed of two fully connected layers.

[0080] Among them, the feature extraction module can encode the overall state matrix, realize feature extraction, extract the main features in the state matrix, and obtain the corresponding overall feature matrix.

[0081] Since the overall feature extraction is performed on the overall state matrix, the first row of the overall feature matrix contains the main features of each component in the chemical mechanical planarization process in the corresponding time dimension. Therefore, only the vector corresponding to the first row can be used for classification, without inputting the entire feature matrix into the classification module, so as to reduce the amount of data input to the classification module, reduce the complexity of classification, and improve the calculation efficiency of classification, thereby improving the prediction efficiency of wafer polishing defects.

[0082] Here, the classification module can calculate the probability of each defect type according to the input vector, obtain the probability value corresponding to each defect type, determine the defect type with the highest probability value, use it as the final defect type prediction result, and output it.

[0083] The embodiment of the present invention can accurately obtain the processing status of the chemical mechanical planarization equipment during the wafer polishing process by acquiring the parameter information of each component in the chemical mechanical planarization equipment within a preset time period; determine the overall state matrix of the chemical mechanical planarization equipment according to the use order and parameter information of each component during the wafer polishing process, and connect the parameter information according to the time sequence of the components and collect them into the overall state matrix, so as to facilitate the subsequent prediction of the polishing defects of the wafer. By inputting the state matrix into the polishing defect prediction model, the defect type prediction result of the wafer polishing of the chemical mechanical planarization equipment within the preset time period is obtained, and the defects of wafer polishing can be predicted, and the defects of wafer polishing can be obtained quickly and accurately, and the process of chemical mechanical planarization can be adjusted in time to reduce the generation of defective wafers and reduce the polishing cost.

[0084] In some embodiments, the overall state matrix of the chemical mechanical planarization equipment is determined based on the usage order and parameter information of each component in the wafer polishing process. The time dimension in the parameter information of each component can be connected according to the usage order of each component in the chemical mechanical planarization equipment during the wafer polishing process; the parameter dimension in the parameter information is then determined based on the common parameters and unique parameters of each component; finally, the state matrix of the chemical mechanical planarization equipment is determined based on the time dimension and the parameter dimension.

[0085] In this embodiment, in order to construct the input of the model into a matrix formed by a two-dimensional time series, the time dimensions of each component are connected according to the order of each component in the polishing process to form the time dimension of the overall state matrix.

[0086] Here, we can use the formula: Calculate the time dimension, where t represents the time dimension of the calculated overall state matrix, t i It represents the time dimension corresponding to the i-th component, that is, the sum of the time frames corresponding to the i-th component. When the component is used, flag = true, and the corresponding time dimension is added. Otherwise, flag = false, and the corresponding time dimension is not calculated.

[0087] In addition, since each component has parameters with the same meaning and parameters unique to each component,

[0088] Here, we can use the formula: Calculate the parameter dimension, where n represents the parameter dimension of the calculated overall state matrix, n share Indicates the parameter dimension of the same meaning parameters contained in each component, n unique,i Indicates the unique parameter dimension corresponding to the i-th component. Similarly, when the component is used, flag = true, and the corresponding parameter dimension is added, otherwise flag = false, and the corresponding parameter dimension is not calculated.

[0089] Based on the above method, the parameter dimension and time dimension of the overall state matrix can be obtained, and then filled or connected according to the parameter information of the corresponding components to obtain the final overall state matrix. Here, the overall state matrix can be a matrix composed of an n×t two-dimensional time series.

[0090] In some embodiments, the overall state matrix is ​​input into a feature extraction module in a preset polishing defect prediction model to obtain an overall feature matrix of the chemical mechanical planarization device output by the feature extraction module. The overall state matrix can be first input into the first conversion coding layer of the feature extraction module in the preset polishing defect prediction model to obtain the state features output by the first conversion coding layer; for the second and subsequent conversion coding layers, the state features output by the previous layer are input into the conversion coding layer to obtain the state features output by the conversion coding layer; finally, based on the state features output by the last conversion coding layer, the overall feature matrix is ​​obtained; wherein, in the attention sublayer of each conversion coding layer, multiple local attention mechanisms are set, and the output results of the attention sublayer are determined by the output results of the local attention mechanisms.

[0091] In this embodiment, the feature extraction module includes multiple conversion coding layers, each of which encodes and extracts features from the input (the overall state matrix or the state features) in turn, thereby obtaining a final overall feature matrix. The conversion coding layer can be composed of a transformer-encoder layer.

[0092] Here, since the time dimension of the parameters of each component in the polishing process is very long, the time dimension of the overall feature matrix input after the time dimensions of each component are connected is too large. In order to save training costs, avoid overfitting, and improve prediction efficiency, the attention sublayer in the transformation coding layer can be improved. By using local attention, each time dimension and the time dimension associated with it can be processed. This can reduce the data that needs to be processed and improve processing efficiency while ensuring the continuity between each component.

[0093] Optionally, the processing process of the attention sub-layer of this embodiment may be, for each time dimension of the input, determining a preset number of attention time dimensions corresponding to the time dimension, and determining the local attention of the data corresponding to the attention time dimension based on the local attention mechanism; based on the local attention of all time dimensions, determining the output result of the attention sub-layer.

[0094] In this embodiment, for each time dimension, a preset number of time dimensions related to the time dimension can be extracted from the input data to constitute the attention time dimension of the time dimension, so that attention can be calculated for the data corresponding to each attention time dimension, and then the calculated local attentions can be pieced together in the time dimension to form a two-dimensional time series as the output of the attention sub-layer.

[0095] By determining the attention time dimension of each time dimension as described above, attention can be sought for each time dimension separately, and when seeking attention, the relevant time dimensions can be fully considered, thereby considering the causal relationship between each component.

[0096] Here, the preset number of time dimensions related to the time dimension may be a number of time dimensions before and after the time dimension.

[0097] Taking the data of the i-th time frame as an example, we can take the max(1,ir)th time frame to the min(t,i+r)th time frame as its attention time dimension, and calculate the attention attn of the data of the i-th time frame on the determined attention time dimension. i Finally, the calculated t attentions are put together to form the output of the attention sub-layer attn = [attn1, attn2,…, attn t ]. Also, see Figure 3 The implementation diagram of the attention mechanism shown in Figure 1 is shown in Figure 2, where r=3 is used as an example. Figure 3 For the first time frame, the time frame and the following three time frames are selected for attention calculation; for the second time frame, the time, the time frame before the time frame (i.e., the first time frame), and the three time frames after the time frame are selected for attention calculation; for the third time frame, the time, the two time frames before the time frame (i.e., the first time frame and the second time frame), and the three time frames after the time frame are selected for attention calculation; for the fourth time frame, the time, the three time frames before the time frame (i.e., the first time frame, the second time frame, and the third time frame), and the three time frames after the time frame are selected for attention calculation; for the fifth time frame, the time, the three time frames before the time frame (i.e., the second time frame, the third time frame, and the fourth time frame), and the three time frames after the time frame are selected for attention calculation; and so on. Figure 3 The time frame contained in the brackets corresponding to the dotted line in is the data required for attention calculation in the time frame corresponding to the dotted line.

[0098] For the value of r, we can use the formula: Make your selection.

[0099] In some embodiments, after the overall state matrix is ​​input into the first conversion coding layer of the feature extraction module in the preset polishing defect prediction model to obtain the state features output by the first conversion coding layer, the state features output by the first conversion coding layer can also be subjected to dimensionality reduction processing of the time dimension.

[0100] Correspondingly, in this embodiment, for the second layer and each subsequent conversion coding layer, the state features of the previous layer output are input into the conversion coding layer. After obtaining the state features of the output of the conversion coding layer, the state features of the output of the conversion coding layer can also be subjected to dimensionality reduction processing of the time dimension.

[0101] In this embodiment, in order to avoid overfitting, after each transformer-encoder layer, dimensionality reduction processing can be performed to compress the time dimension of the parameters. Here, a maximum pooling layer or an average pooling layer can be used to downsample the parameters.

[0102] Correspondingly, when attention is calculated in each transformer-encoder layer after dimensionality reduction, the time dimension will also change accordingly.

[0103] In some embodiments, before inputting the overall state matrix into a preset polishing defect prediction model to obtain the defect type prediction result of the chemical mechanical planarization equipment performing wafer polishing within a preset time period output by the polishing defect prediction model, the historical parameter information and corresponding defect types of the chemical mechanical planarization under different defect types can be obtained first; then, based on the historical parameter information, the historical state matrix of the chemical mechanical planarization equipment is determined; finally, based on the historical state matrix and the corresponding defect types, the polishing defect prediction model is trained, and the model parameters are updated to obtain a trained polishing defect prediction model; wherein the polishing defect prediction model includes a feature extraction module and a classification module.

[0104] In this embodiment, a polishing defect prediction model can be constructed based on a transformer-encoder model and a classifier. The transformer-encoder model can include multiple layers of transformer-encoder layers to form a feature extraction module for extracting features from the overall state matrix. The classifier can include two layers of fully connected layers to form a classification module for calculating the extracted features to obtain a defect probability distribution, thereby predicting defects that may occur after wafer polishing.

[0105] Here, the constructed polishing defect prediction model is iteratively trained through the historical state matrix and the corresponding defect types, and the polishing defect prediction model parameters are updated during the iteration process to improve the closeness between the defect type prediction result and the corresponding defect type, thereby improving the accuracy of the prediction.

[0106] Optionally, the cross entropy function can be used as a loss function in the polishing defect prediction model training process. The loss function represents the similarity between the predicted result and the actual defect type label value during the training process, and iterative training is performed for the purpose of improving the similarity.

[0107] Among them, the cross entropy function can be expressed as:

[0108]

[0109] In the formula, J θ represents the cross entropy function, m represents the sample set, that is, the set of historical state matrices, q represents the qth sample, that is, the qth historical state matrix, p represents the set of defect types, r represents the rth defect type, y qr represents the probability corresponding to the label of the qth historical state matrix and the rth defect type, Represents the probability of the qth historical state matrix output predicted by the polishing defect prediction model.

[0110] In addition, during the model training process, the learning rate can be adjusted using a linear warm-up technique. During the warm-up period, the learning rate increases linearly from 0 to the initial preset learning rate, and then the learning rate is linearly reduced to the set minimum value, thereby ensuring the training effect in model training.

[0111] Also, the optimizer in the training process can use the AdamW optimizer, combined with the L2 regular term to make the parameters to be learned as small as possible, which can effectively avoid overfitting.

[0112] Here, the parameter update can be performed using the following formula:

[0113]

[0114] In the formula, θ represents the parameter to be learned, θ u+1 represents the parameters to be learned in the u+1th iteration, θ u represents the parameters to be learned in the uth iteration, WD θ represents the L2 regularization term, λ represents the coefficient of the regularization term, and α represents the learning rate. At each iteration, the parameters can be updated according to the above formula.

[0115] Here, when the set number of training times is reached, the iteration is terminated, and the parameters θ of the model during the iteration corresponding to the optimal accuracy rate are selected from the training process as the final parameters of the model.

[0116] Optionally, after obtaining the historical parameter information of chemical mechanical planarization under different defect types and the corresponding defect types, the historical parameter information may be subjected to denoising and normalization processing.

[0117] For details, please refer to Figure 1 The description in the corresponding embodiments will not be repeated here.

[0118] The embodiment of the present invention can accurately obtain the processing status of the chemical mechanical planarization equipment during the wafer polishing process by acquiring the parameter information of each component in the chemical mechanical planarization equipment within a preset time period; determine the overall state matrix of the chemical mechanical planarization equipment according to the use order and parameter information of each component during the wafer polishing process, and connect the parameter information according to the time sequence of the components and collect them into the overall state matrix, so as to facilitate the subsequent prediction of the polishing defects of the wafer. By inputting the state matrix into the polishing defect prediction model, the defect type prediction result of the wafer polishing of the chemical mechanical planarization equipment within the preset time period is obtained, and the defects of wafer polishing can be predicted, and the defects of wafer polishing can be obtained quickly and accurately, and the process of chemical mechanical planarization can be adjusted in time to reduce the generation of defective wafers and reduce the polishing cost. Among them, by inputting the state matrix into the feature extraction model in the polishing defect prediction model, the feature extraction module can be used to extract features from the state matrix, that is, to extract the main features in the state matrix and obtain the corresponding feature matrix; then the first vector of the feature matrix is ​​input into the classification module in the polishing defect model, and the defect type prediction result of wafer polishing can be obtained quickly and accurately. There is no need to input the entire feature matrix into the classification module. While ensuring accuracy, the amount of data input to the classification module can be reduced, further improving the prediction efficiency of wafer polishing defects. Among them, by setting a local attention mechanism in the feature extraction module, each time dimension and the time dimension associated with it can be processed, which can reduce the data to be processed and improve processing efficiency on the basis of ensuring the continuity between each component.

[0119] It should be understood that the order of execution of the steps in the above embodiment does not necessarily mean the order of execution. The execution order of each process should be determined by its function and internal logic, and should not constitute any limitation on the implementation process of the embodiment of the present invention.

[0120] The following is an embodiment of the device of the present invention. For details not described in detail therein, reference may be made to the corresponding method embodiment described above.

[0121] Figure 4 The schematic diagram of the structure of the device for predicting wafer polishing defects provided by an embodiment of the present invention is shown. For the convenience of explanation, only the part related to the embodiment of the present invention is shown, which is described in detail as follows:

[0122] like Figure 4 As shown, the wafer polishing defect prediction device 40 includes:

[0123] An acquisition module 41 is used to acquire parameter information of each component in the chemical mechanical planarization device within a preset period of time;

[0124] A determination module 42 is used to determine the overall state matrix of the chemical mechanical planarization device according to the use order and parameter information of each component in the wafer polishing process; wherein the overall state matrix is ​​a matrix formed by a two-dimensional time series with a continuous time dimension;

[0125] The prediction module 43 is used to input the overall state matrix into a preset polishing defect prediction model to obtain a defect type prediction result of the chemical mechanical planarization equipment performing wafer polishing within a preset time period output by the polishing defect prediction model; wherein the polishing defect prediction model is obtained by training based on the historical state matrix and corresponding defect types of the chemical mechanical planarization equipment under different defect types.

[0126] In a possible implementation, the determination module 42 is specifically configured to:

[0127] According to the order in which the components in the chemical mechanical planarization equipment are used during the wafer polishing process, the time dimension in the parameter information of each component is connected;

[0128] Determine the parameter dimensions in the parameter information according to the common parameters and unique parameters of each component;

[0129] Based on the time dimension and the parameter dimension, the state matrix of the chemical mechanical planarization device is determined.

[0130] In a possible implementation, the prediction module 43 is specifically used for:

[0131] Inputting the overall state matrix into a feature extraction module in a preset polishing defect prediction model to obtain an overall feature matrix of the chemical mechanical planarization device output by the feature extraction module;

[0132] The first vector of the overall feature matrix is ​​input into the classification module in the polishing defect model, and the defect type prediction result of the chemical mechanical planarization equipment performing wafer polishing within a preset time period is obtained as output by the classification module.

[0133] In a possible implementation, the prediction module 43 is specifically used for:

[0134] Inputting the overall state matrix into the first conversion coding layer of the feature extraction module in the preset polishing defect prediction model to obtain the state features output by the first conversion coding layer;

[0135] For each conversion coding layer of the second layer and thereafter, the state feature output by the previous layer is input to the conversion coding layer to obtain the state feature output by the conversion coding layer;

[0136] Based on the state features output by the last conversion coding layer, the overall feature matrix is ​​obtained;

[0137] In the attention sublayer of each conversion coding layer, multiple local attention mechanisms are set, and the output results of the attention sublayer are determined by the output results of the local attention mechanism.

[0138] In one possible implementation, the processing of the attention sublayer is:

[0139] For each time dimension of the input, determine a preset number of attention time dimensions corresponding to the time dimension, and determine the local attention of the data corresponding to the attention time dimension according to the local attention mechanism;

[0140] Based on the local attention in all time dimensions, the output of the attention sub-layer is determined.

[0141] In a possible implementation, the prediction module 43 is further configured to:

[0142] Perform dimensionality reduction processing on the time dimension of the state features output by the first conversion coding layer;

[0143] Correspondingly, for each conversion coding layer of the second layer and thereafter, after inputting the state feature output by the previous layer into the conversion coding layer and obtaining the state feature output by the conversion coding layer, the method further includes:

[0144] The state features output by the conversion coding layer are subjected to dimensionality reduction processing in the time dimension.

[0145] In a possible implementation, the prediction module 43 is further configured to:

[0146] Obtain historical parameter information and corresponding defect types of chemical mechanical planarization under different defect types;

[0147] Determine a historical state matrix of a chemical mechanical planarization device based on historical parameter information;

[0148] According to the historical state matrix and the corresponding defect types, the polishing defect prediction model is trained, and the model parameters are updated to obtain a trained polishing defect prediction model; wherein the polishing defect prediction model includes a feature extraction module and a classification module.

[0149] Figure 5 Schematic diagram of an electronic device provided by an embodiment of the present invention. Figure 5 As shown, the electronic device 50 of this embodiment includes: a processor 51, a memory 52, and a computer program 53 stored in the memory 52 and executable on the processor 51. When the processor 51 executes the computer program 53, the steps in the above-mentioned various wafer polishing defect prediction method embodiments are implemented, for example Figure 1Alternatively, when the processor 51 executes the computer program 53, the functions of each module in the above-mentioned device embodiments are realized, for example, Figure 4 The functions of modules 41 to 43 are shown.

[0150] Exemplarily, the computer program 53 may be divided into one or more modules / units, one or more modules / units are stored in the memory 52 and executed by the processor 51 to implement the present invention. The one or more modules / units may be a series of computer program instruction segments capable of implementing specific functions, and the instruction segments are used to describe the execution process of the computer program 53 in the electronic device 50. For example, the computer program 53 may be divided into Figure 4 Modules 41 to 43 are shown.

[0151] The electronic device 50 may include, but is not limited to, a processor 51 and a memory 52. ​​Those skilled in the art will appreciate that Figure 5 It is only an example of the electronic device 50 and does not constitute a limitation of the electronic device 50. It may include more or fewer components than shown in the figure, or a combination of certain components, or different components. For example, the electronic device may also include input and output devices, network access devices, buses, etc.

[0152] The processor 51 may be a central processing unit (CPU), or other general-purpose processors, digital signal processors (DSP), application-specific integrated circuits (ASIC), field-programmable gate arrays (FPGA) or other programmable logic devices, discrete gate or transistor logic devices, discrete hardware components, etc. A general-purpose processor may be a microprocessor or any conventional processor, etc.

[0153] The memory 52 may be an internal storage unit of the electronic device 50, such as a hard disk or memory of the electronic device 50. The memory 52 may also be an external storage device of the electronic device 50, such as a plug-in hard disk, a smart media card (SMC), a secure digital (SD) card, a flash card, etc. equipped on the electronic device 50. Further, the memory 52 may also include both an internal storage unit of the electronic device 50 and an external storage device. The memory 52 is used to store computer programs and other programs and data required by the electronic device. The memory 52 may also be used to temporarily store data that has been output or is to be output.

[0154] The technicians in the relevant field can clearly understand that for the convenience and simplicity of description, only the division of the above-mentioned functional units and modules is used as an example for illustration. In practical applications, the above-mentioned function allocation can be completed by different functional units and modules as needed, that is, the internal structure of the device can be divided into different functional units or modules to complete all or part of the functions described above. The functional units and modules in the embodiment can be integrated in a processing unit, or each unit can exist physically separately, or two or more units can be integrated in one unit. The above-mentioned integrated unit can be implemented in the form of hardware or in the form of software functional units. In addition, the specific names of the functional units and modules are only for the convenience of distinguishing each other, and are not used to limit the scope of protection of this application. The specific working process of the units and modules in the above-mentioned system can refer to the corresponding process in the aforementioned method embodiment, which will not be repeated here.

[0155] In the above embodiments, the description of each embodiment has its own emphasis. For parts that are not described or recorded in detail in a certain embodiment, reference can be made to the relevant descriptions of other embodiments.

[0156] Those skilled in the art will appreciate that the units and algorithm 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 present invention.

[0157] In the embodiments provided by the present invention, it should be understood that the disclosed devices / electronic devices and methods can be implemented in other ways. For example, the device / electronic device embodiments described above are only schematic. For example, the division of modules or units is only a logical function division. There may be other division methods in actual implementation. For example, multiple units or components can be combined or integrated into another system, or some features can be ignored or not executed. Another point is that the mutual coupling or direct coupling or communication connection shown or discussed can be through some interfaces, indirect coupling or communication connection of devices or units, which can be electrical, mechanical or other forms.

[0158] The units described as separate components may or may not be physically separated, and the components shown as units may or may not be physical units, that is, they may be located in one place or distributed on multiple network units. Some or all of the units may be selected according to actual needs to achieve the purpose of the solution of this embodiment.

[0159] In addition, each functional unit in each embodiment of the present invention may be integrated into one processing unit, or each unit may exist physically separately, or two or more units may be integrated into one unit. The above-mentioned integrated unit may be implemented in the form of hardware or in the form of software functional units.

[0160] If the integrated module / unit 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 present invention implements all or part of the processes in the above-mentioned embodiment method, and can also be completed by instructing the relevant hardware through a computer program. The computer program can be stored in a computer-readable storage medium, and the computer program can implement the steps of the above-mentioned various method embodiments when executed by the processor. Among them, the computer program includes computer program code, and the computer program code can be in source code form, object code form, executable file or some intermediate form. Computer-readable media can include: any entity or device capable of carrying computer program code, recording medium, U disk, mobile hard disk, disk, optical disk, computer memory, read-only memory (ROM), random access memory (RAM), electrical carrier signal, telecommunication signal and software distribution medium, etc.

[0161] The above embodiments are only used to illustrate the technical solutions of the present invention, rather than to limit the same. Although the present invention has been described in detail with reference to the aforementioned embodiments, those skilled in the art should understand that the technical solutions described in the aforementioned embodiments may still be modified, or some of the technical features may be replaced by equivalents. Such modifications or replacements do not deviate the essence of the corresponding technical solutions from the spirit and scope of the technical solutions of the embodiments of the present invention, and should be included in the protection scope of the present invention.

Claims

1. A method for predicting wafer polishing defects, characterized in that: include: Obtaining parameter information of each component in the chemical mechanical planarization device within a preset time period; Determine the overall state matrix of the chemical mechanical planarization device according to the usage order of each component in the wafer polishing process and the parameter information; wherein the overall state matrix is ​​a matrix formed by a two-dimensional time series with a continuous time dimension; Inputting the overall state matrix into a preset polishing defect prediction model to obtain a defect type prediction result of the chemical mechanical planarization device performing wafer polishing within the preset time period output by the polishing defect prediction model; The polishing defect prediction model is obtained by training based on the historical state matrix of the chemical mechanical planarization equipment under different defect types and the corresponding defect types.

2. The method for predicting wafer polishing defects according to claim 1, characterized in that: According to the usage order of each component in the wafer polishing process and the parameter information, an overall state matrix of the chemical mechanical planarization device is determined, including: According to the order in which the components in the chemical mechanical planarization device are used during the wafer polishing process, the time dimensions in the parameter information of the components are connected; Determine the parameter dimensions in the parameter information according to the common parameters and unique parameters of each component; Based on the time dimension and the parameter dimension, a state matrix of the chemical mechanical planarization device is determined.

3. The method for predicting wafer polishing defects according to claim 1, characterized in that: Inputting the overall state matrix into a preset polishing defect prediction model to obtain a defect type prediction result of the chemical mechanical planarization device performing wafer polishing within the preset time period output by the polishing defect prediction model, including: Inputting the overall state matrix into a feature extraction module in a preset polishing defect prediction model to obtain an overall feature matrix of the chemical mechanical planarization device output by the feature extraction module; The first vector of the overall feature matrix is ​​input into the classification module in the polishing defect model to obtain the defect type prediction result of the chemical mechanical planarization equipment performing wafer polishing within the preset time period output by the classification module.

4. The method for predicting wafer polishing defects according to claim 3, characterized in that: Inputting the overall state matrix into a feature extraction module in a preset polishing defect prediction model to obtain an overall feature matrix of the chemical mechanical planarization device output by the feature extraction module, including: Inputting the overall state matrix into the first conversion coding layer of the feature extraction module in the preset polishing defect prediction model to obtain the state features output by the first conversion coding layer; For each conversion coding layer of the second layer and thereafter, the state feature outputted by the previous layer is inputted into the conversion coding layer to obtain the state feature outputted by the conversion coding layer; Based on the state features output by the last conversion coding layer, the overall feature matrix is ​​obtained; In which, in the attention sublayer of each conversion coding layer, multiple local attention mechanisms are set, and the output result of the attention sublayer is determined by the output result of the local attention mechanism.

5. The method for predicting wafer polishing defects according to claim 4, characterized in that: The processing process of the attention sublayer is: For each time dimension of the input, determine a preset number of attention time dimensions corresponding to the time dimension, and determine the local attention of the data corresponding to the attention time dimension according to the local attention mechanism; Based on the local attention of all time dimensions, the output result of the attention sublayer is determined.

6. The method for predicting wafer polishing defects according to claim 4, characterized in that: After inputting the overall state matrix into the first conversion coding layer of the feature extraction module in the preset polishing defect prediction model to obtain the state features output by the first conversion coding layer, the method further includes: Performing a time dimension reduction process on the state features output by the first conversion coding layer; Correspondingly, for each conversion coding layer of the second layer and thereafter, after inputting the state feature output by the previous layer into the conversion coding layer and obtaining the state feature output by the conversion coding layer, the method further includes: The state features output by the conversion coding layer are subjected to dimensionality reduction processing in the time dimension.

7. The method for predicting wafer polishing defects according to any one of claims 1 to 6, characterized in that: Before inputting the overall state matrix into a preset polishing defect prediction model to obtain a defect type prediction result of the chemical mechanical planarization device performing wafer polishing within the preset time period output by the polishing defect prediction model, the method further includes: Acquiring historical parameter information of the chemical mechanical planarization under different defect types and corresponding defect types; Determining a historical state matrix of the chemical mechanical planarization device according to the historical parameter information; According to the historical state matrix and the corresponding defect types, the polishing defect prediction model is trained, and the model parameters are updated to obtain a trained polishing defect prediction model; wherein the polishing defect prediction model includes a feature extraction module and a classification module.

8. A device for predicting wafer polishing defects, characterized in that: include: An acquisition module, used to acquire parameter information of each component in the chemical mechanical planarization device within a preset time period; A determination module, used to determine the overall state matrix of the chemical mechanical planarization device according to the use order of each component in the wafer polishing process and the parameter information; wherein the overall state matrix is ​​a matrix formed by a two-dimensional time series with a continuous time dimension; A prediction module is used to input the overall state matrix into a preset polishing defect prediction model to obtain a defect type prediction result of the chemical mechanical planarization equipment performing wafer polishing within the preset time period output by the polishing defect prediction model; wherein the polishing defect prediction model is trained based on the historical state matrix of the chemical mechanical planarization equipment under different defect types and the corresponding defect types.

9. An electronic device comprising a memory and a processor, wherein the memory is used to store a computer program, and the processor is used to call and run the computer program stored in the memory, characterized in that: When the processor executes the computer program, the steps of the method according to any one of claims 1 to 7 are implemented.

10. A computer-readable storage medium storing a computer program, characterized in that: When the computer program is executed by a processor, the steps of the method according to any one of claims 1 to 7 are implemented.

Images