Polishing endpoint detection method, device, electronic equipment and storage medium

Abstract

The embodiments of the present application disclose a polishing endpoint detection method, device, electronic device and storage medium, which are applied to chemical mechanical polishing equipment, wherein the chemical mechanical polishing equipment includes a drive motor and a polishing disk, wherein the drive motor is used to drive the polishing disk to rotate, and the method includes: in the process of chemical mechanical polishing of a wafer, collecting a torque signal corresponding to a preset number of rotations of the polishing disk driven by the drive motor; extracting torque characteristics according to the torque signal to obtain a torque characteristic set; the torque characteristic set includes at least one of the frequency domain characteristics, time domain characteristics and time-frequency domain characteristics corresponding to the torque signal; analyzing the torque characteristic set through a polishing endpoint detection model to determine a polishing endpoint detection result, wherein the polishing endpoint detection result is used to characterize whether the polishing endpoint has been reached, so that it is possible to accurately identify whether the polishing endpoint has been reached during the process of chemical mechanical polishing of the wafer, thereby improving the accuracy of polishing endpoint detection.

Description

Technical Field

[0001] The present application relates to the field of polishing technology, and in particular to a polishing endpoint detection method, device, electronic equipment and storage medium. Background Art

[0002] Chemical Mechanical Polish (CMP) is one of the key processes in the semiconductor device manufacturing process. It is used to remove multiple layers of materials and flatten the wafer surface to ensure device performance and reliability. However, in actual production, over-polishing or under-polishing will lead to serious quality problems in the product. Therefore, how to monitor the polishing status in real time and accurately determine whether the polishing endpoint has been reached has become a technical problem that needs to be solved urgently. Summary of the invention

[0003] The embodiments of the present application disclose a polishing endpoint detection method, device, electronic device and storage medium, which can accurately identify whether the polishing endpoint is reached during the chemical mechanical polishing of a wafer, thereby improving the accuracy of polishing endpoint detection.

[0004] In a first aspect, an embodiment of the present application discloses a polishing endpoint detection method, which is applied to a chemical mechanical polishing device, wherein the chemical mechanical polishing device comprises a drive motor and a polishing disk, wherein the drive motor is used to drive the polishing disk to rotate, and the method comprises:

[0005] During chemical mechanical polishing of the wafer, a torque signal corresponding to a preset number of revolutions of the polishing disk driven by the drive motor is collected;

[0006] Extracting torque features according to the torque signal to obtain a torque feature set; the torque feature set includes at least one of a frequency domain feature, a time domain feature, and a time-frequency domain feature corresponding to the torque signal;

[0007] The torque feature set is analyzed by a polishing endpoint detection model to determine a polishing endpoint detection result; the polishing endpoint detection result is used to indicate whether the polishing endpoint has been reached.

[0008] As an optional implementation, in the first aspect of the embodiment of the present application, the torque feature set includes a combined feature vector of the time domain features and the frequency domain features corresponding to the torque signal; the step of extracting the torque feature according to the torque signal to obtain the torque feature set includes:

[0009] Extracting the time domain feature from the torque signal to obtain a time domain feature vector;

[0010] Extracting the frequency domain feature from the torque signal to obtain a frequency domain feature vector;

[0011] The time domain feature vector is combined with the frequency domain feature vector to obtain a combined feature vector.

[0012] As an optional implementation, in the first aspect of the embodiment of the present application, the torque feature set includes a time-frequency domain diagram corresponding to the torque signal, and the time-frequency domain diagram is used to characterize the time-frequency domain features corresponding to the torque signal; the step of extracting the torque feature according to the torque signal to obtain the torque feature set includes:

[0013] Based on the wavelet function, the torque signal is converted into a time-frequency domain diagram.

[0014] As an optional implementation, in the first aspect of the embodiment of the present application, the polishing endpoint detection model includes a first channel and a second channel; the step of analyzing the torque feature set through the polishing endpoint detection model to determine the polishing endpoint detection result includes:

[0015] Inputting the combined feature vector into the first channel, performing convolution processing and pooling processing on the combined feature vector through the first channel to obtain a first feature vector;

[0016] Inputting the time-frequency domain image into the second channel, performing convolution processing and pooling processing on the time-frequency domain image through the second channel to obtain a second feature vector;

[0017] Fusing the first feature vector and the second feature vector to obtain a fused feature vector;

[0018] A polishing endpoint detection result is determined according to the fused feature vector.

[0019] As an optional implementation, in the first aspect of the embodiment of the present application, the polishing endpoint detection model is obtained by training a training set, the training set includes multiple groups of training torque data, each group of training torque data includes a training torque signal and an endpoint label corresponding to the training torque signal, and the endpoint label is used to indicate whether the polishing endpoint has been reached; in the process of chemical mechanical polishing of the wafer, before collecting the torque signal corresponding to the drive motor driving the polishing disk to rotate a preset number of times, the method further includes:

[0020] Extracting torque features according to the training torque signals contained in each group of the training torque data to obtain a training torque feature set corresponding to each group of the training torque data; wherein the training torque feature set includes at least one of a frequency domain feature, a time domain feature, and a time-frequency domain feature corresponding to the training torque signal;

[0021] The polishing endpoint detection model to be trained is trained according to the training torque feature sets and the corresponding endpoint labels respectively corresponding to the multiple groups of training torque data.

[0022] As an optional implementation, in the first aspect of the embodiment of the present application, before extracting the torque feature according to the training torque signal contained in each group of the training torque data to obtain the training torque feature set corresponding to each group of the training torque data, the method further includes:

[0023] Normalizing the training torque signal contained in each group of the training torque data to map the training torque signal to a target range;

[0024] The step of extracting torque features according to the training torque signals contained in each group of the training torque data to obtain a set of training torque features corresponding to each group of the training torque data includes:

[0025] According to the normalized training torque signal contained in each group of the training torque data, the torque feature is extracted to obtain a training torque feature set corresponding to each group of the training torque data.

[0026] As an optional implementation, in the first aspect of the embodiment of the present application, the polishing endpoint detection model to be trained is trained according to the training torque feature sets and the corresponding endpoint labels corresponding to the multiple sets of training torque data, including:

[0027] Analyzing a set of training torque features corresponding to one or more sets of training torque data by using a polishing endpoint detection model to be trained, and determining a polishing endpoint prediction probability corresponding to each set of training torque data;

[0028] Calculating the classification loss of the polishing endpoint detection model to be trained according to the polishing endpoint prediction probability and the corresponding endpoint label corresponding to each group of the training torque data;

[0029] According to the classification loss, the model parameters of the polishing endpoint detection model to be trained are updated until the polishing endpoint detection model converges.

[0030] As an optional implementation, in the first aspect of the embodiment of the present application, after the polishing endpoint detection model to be trained is trained according to the training torque feature sets and the corresponding endpoint labels respectively corresponding to the multiple sets of training torque data, the method further includes:

[0031] The trained polishing endpoint detection model is tested by a test set to obtain model performance information corresponding to the trained polishing endpoint detection model; the test set includes multiple groups of test torque data, each group of test torque data includes a test torque signal and an endpoint label corresponding to the test torque signal;

[0032] If the model performance information does not meet the performance condition, expanding the training set, and continuing to train the polishing endpoint detection model according to the expanded training set;

[0033] If the model performance information meets the performance conditions, the polishing endpoint detection model training is completed.

[0034] A second aspect of an embodiment of the present application discloses a polishing endpoint detection device, which is applied to a chemical mechanical polishing device, wherein the chemical mechanical polishing device comprises a driving motor and a polishing disc, wherein the driving motor is used to drive the polishing disc to rotate, and the device comprises:

[0035] An acquisition module, used for acquiring a torque signal corresponding to a preset number of revolutions of the polishing disk driven by the drive motor during chemical mechanical polishing of the wafer;

[0036] an extraction module, configured to extract torque features according to the torque signal to obtain a torque feature set; the torque feature set includes at least one of a frequency domain feature, a time domain feature, and a time-frequency domain feature corresponding to the torque signal;

[0037] The determination module is used to analyze the torque feature set through a polishing endpoint detection model to determine a polishing endpoint detection result, wherein the polishing endpoint detection result is used to indicate whether the polishing endpoint is reached.

[0038] A third aspect of the present application discloses an electronic device, including:

[0039] A memory storing executable program code;

[0040] a processor coupled to the memory;

[0041] The processor calls the executable program code stored in the memory to execute the method described in any one of the above embodiments.

[0042] A fourth aspect of an embodiment of the present application discloses a computer-readable storage medium, which stores a computer program. When the computer program is executed by a processor, the processor executes the method described in any one of the above embodiments.

[0043] Compared with the related art, the embodiments of the present application include at least the following beneficial effects:

[0044] The polishing endpoint detection method, device, electronic device and storage medium disclosed in the embodiment of the present application, in the process of chemical mechanical polishing of the wafer, collects the torque signal corresponding to the preset number of revolutions driven by the driving motor to drive the polishing disk; extracts the torque feature according to the torque signal to obtain a torque feature set; the torque feature set includes at least one of the frequency domain feature, time domain feature and time-frequency domain feature corresponding to the torque signal; analyzes the torque feature set through the polishing endpoint detection model to determine the polishing endpoint detection result, and the polishing endpoint detection result is used to characterize whether the polishing endpoint has been reached. In the embodiment of the present application, in the process of chemical mechanical polishing of the wafer, the torque signal corresponding to the preset number of revolutions driven by the driving motor to drive the polishing disk can be collected, and the torque feature set corresponding to the torque signal can be extracted. By inputting the torque feature set corresponding to the torque signal into the polishing endpoint detection model for analysis, it is possible to accurately determine whether the polishing endpoint has been reached, thereby improving the accuracy and efficiency of polishing endpoint detection. BRIEF DESCRIPTION OF THE DRAWINGS

[0045] In order to more clearly illustrate the technical solutions in the embodiments of the present application, the drawings required for use in the embodiments will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present application. For ordinary technicians in this field, other drawings can be obtained based on these drawings without paying creative work.

[0046] Figure 1 It is a schematic diagram of an application scenario of a polishing endpoint detection method disclosed in an embodiment of the present application;

[0047] Figure 2 It is a schematic flow chart of a polishing endpoint detection method disclosed in an embodiment of the present application;

[0048] Figure 3 It is a schematic diagram of a process of extracting torque features according to a torque signal to obtain a torque feature set disclosed in an embodiment of the present application;

[0049] Figure 4 It is a schematic diagram of a process for analyzing a torque feature set through a polishing endpoint detection model to determine a polishing endpoint detection result disclosed in an embodiment of the present application;

[0050] Figure 5 It is a schematic diagram of the process of training the polishing endpoint detection model disclosed in the embodiment of the present application;

[0051] Figure 6 It is a flow chart of training a polishing endpoint detection model to be trained according to a plurality of training torque feature sets and corresponding endpoint labels respectively corresponding to a plurality of sets of training torque data disclosed in an embodiment of the present application;

[0052] Figure 7It is a schematic diagram of a process for testing a trained polishing endpoint detection model disclosed in an embodiment of the present application;

[0053] Figure 8 It is a structural schematic diagram of a polishing endpoint detection device disclosed in an embodiment of the present application;

[0054] Fig. 9 It is a structural schematic diagram of an electronic device disclosed in an embodiment of the present application. DETAILED DESCRIPTION

[0055] The following will be combined with the drawings in the embodiments of the present application to clearly and completely describe the technical solutions in the embodiments of the present application. Obviously, the described embodiments are only part of the embodiments of the present application, not all of the embodiments. Based on the embodiments in the present application, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of this application.

[0056] It should be noted that the terms "including" and "having" and any variations thereof in the embodiments of the present application and the accompanying drawings are intended to cover non-exclusive inclusions. For example, a process, method, system, product or device including a series of steps or units is not limited to the listed steps or units, but may optionally include steps or units not listed, or may optionally include other steps or units inherent to these processes, methods, products or devices.

[0057] In related technologies, the polishing endpoint is determined by a friction-based endpoint detection method. When the polishing material changes, the friction coefficient between the wafer surface and the polishing pad will change, and the rotational torque of the drive motor will change accordingly. Therefore, it is usually possible to determine whether the wafer polishing has reached the endpoint based on the time domain change of the torque. Due to the complex contact mechanism of the polishing process and the changeable and complex operating environment of the mechanical equipment, the torque signal usually has non-stationary and nonlinear characteristics. How to extract information closely related to the polishing state from the torque signal and accurately identify the polishing endpoint is a major challenge in current CMP technology.

[0058] The embodiments of the present application disclose a polishing endpoint detection method, device, electronic device and storage medium, which can accurately identify whether the polishing endpoint is reached during the chemical mechanical polishing of a wafer, thereby improving the accuracy of polishing endpoint detection.

[0059] See also Figure 1 , Figure 1 Schematic diagram of an application scenario of a polishing endpoint detection method disclosed in an embodiment of the present application. Figure 1As shown, the chemical mechanical polishing equipment 100 may include a driving motor 101, a polishing disc 102, a polishing pad 103, and a polishing head 104, etc., which is not specifically limited in the embodiment of the present application.

[0060] The driving motor 101 can be used to drive the polishing plate 102 to rotate. Optionally, the driving motor can include but is not limited to a servo motor and a direct drive motor.

[0061] In some embodiments, the driving motor 101 can drive the polishing disc 102 to rotate via a transmission device, wherein the transmission device may include but is not limited to gears and belts.

[0062] The polishing disc 102 refers to a rotating disc that can be rotated under the action of the driving motor 101. The rotation of the polishing disc 102 provides mechanical friction for the polishing process to help remove the material on the surface of the wafer 105. Optionally, the rotation of the polishing disc 102 can be unidirectional or bidirectional to achieve a more uniform polishing effect, which is not specifically limited in the embodiment of the present application.

[0063] The polishing pad 103 is located on the polishing plate 102, and its surface has a specific texture and porosity, which can store and distribute the polishing liquid and is in direct contact with the wafer 105 during the polishing process.

[0064] The polishing head 104 can be used to clamp the wafer 105 and press the wafer 105 tightly against the polishing pad 103 .

[0065] When the chemical mechanical polishing equipment 100 performs chemical mechanical polishing on the wafer 105, the driving motor 101 drives the polishing disk 102 to rotate through the transmission device. The polishing pad 103 is located on the polishing disk 102 and rotates driven by the polishing disk 102. During the polishing process, the wafer 105 is pressed against the polishing pad 103 by the polishing head 104 and is in direct contact with the polishing pad 103. Therefore, when the polishing pad 103 rotates with the polishing disk 102, the surface of the wafer 105 can generate friction with the polishing pad 103, thereby removing the surface material and achieving polishing of the wafer 105.

[0066] In some embodiments, the chemical mechanical polishing device 100 may further include a controller (not shown), which may be communicatively connected to the driving motor 101 and perform data exchange with the driving motor 101 .

[0067] In the embodiment of the present application, during the process of chemical mechanical polishing of the wafer 105, the chemical mechanical polishing device 100 collects the torque signal corresponding to the preset number of revolutions driven by the driving motor 101 to drive the polishing disc 102; extracts the torque feature according to the torque signal to obtain a torque feature set; the torque feature set includes at least one of the frequency domain feature, time domain feature and time-frequency domain feature corresponding to the torque signal. The torque feature set can be analyzed by the polishing endpoint detection model to determine the polishing endpoint detection result, which is used to indicate whether the polishing endpoint has been reached.

[0068] Please read further Figure 2 , Figure 2 1 is a flow chart of a polishing endpoint detection method disclosed in an embodiment of the present application, and the polishing endpoint detection method can be applied to the chemical mechanical polishing device 100. Figure 2 As shown, the polishing endpoint detection method may include the following steps:

[0069] Step 210 , during the process of chemical mechanical polishing of the wafer, a torque signal corresponding to a preset number of revolutions of the polishing disk driven by the driving motor is collected.

[0070] During the chemical mechanical polishing process of the wafer, the driving motor in the chemical mechanical polishing equipment can drive the polishing disk to rotate and grind the wafer surface. When the material on the wafer surface changes, the friction coefficient between the wafer surface and the polishing pad will change, and accordingly, the rotational torque of the driving motor will also change. Therefore, the chemical mechanical polishing equipment can collect the torque signal corresponding to the preset number of revolutions of the polishing disk driven by the driving motor, and determine whether the polishing end point has been reached by analyzing the torque signal.

[0071] The torque signal can be used to characterize the change in the rotational torque corresponding to the drive motor when the drive motor drives the polishing disk to rotate a preset number of times, and further reflect the change in the surface flatness of the wafer during the process, so that the polishing end point can be detected according to the torque signal in the subsequent steps. In the process of the drive motor driving the polishing disk to rotate, the chemical mechanical polishing equipment can collect the torque signal corresponding to the drive motor according to a preset period, and the preset period can correspond to the preset number of turns. For example, the preset period can be the product of the preset number of turns and the rotation speed of the polishing disk.

[0072] Optionally, the collected torque signal may be discrete signal data, and multiple torque signals may be collected when the driving motor drives the polishing disc to rotate a preset number of times.

[0073] Optionally, the collected torque signal may also be a continuous waveform signal, and the continuous waveform signal may be sampled to obtain a plurality of discrete torque signals when the driving motor drives the polishing disc to rotate a preset number of times.

[0074] The preset number of revolutions can be set according to actual needs, for example, the preset number of revolutions can be half a revolution, one revolution or two revolutions, etc., but is not limited thereto. For example, a chemical mechanical polishing device can collect a torque signal corresponding to a polishing disc rotating one revolution driven by a driving motor.

[0075] Step 220, extracting torque features according to the torque signal to obtain a torque feature set; the torque feature set includes at least one of frequency domain features, time domain features, and time-frequency domain features corresponding to the torque signal.

[0076] The chemical mechanical polishing device can extract torque features according to the currently collected torque signal to obtain a torque feature set, wherein the torque feature may refer to a feature that can reflect the signal change of the torque signal.

[0077] The time domain features corresponding to the torque signal can be used to characterize the change of the torque signal over time. Specifically, the time domain feature values ​​corresponding to the torque signal may include but are not limited to one or more of the mean value, variance, peak-to-peak value, and effective value corresponding to the torque signal. Among them, the peak-to-peak value refers to the difference between the maximum value and the minimum value of the torque signal. It can be understood that the more types of time domain feature values ​​extracted, the more comprehensive the feature description of the torque signal, and the more accurate the polishing endpoint detection result of the torque signal.

[0078] The frequency domain features corresponding to the torque signal can be used to characterize the corresponding relationship between the amplitude value and frequency of the torque signal. Specifically, the frequency domain feature values ​​corresponding to the torque signal may include but are not limited to one or more of the average frequency, the center frequency, the root mean square frequency, and the root variance frequency. It can be understood that the more types of frequency domain feature values ​​extracted, the more comprehensive the feature description of the torque signal, and the more accurate the polishing endpoint detection result of the torque signal.

[0079] The time-frequency domain features corresponding to the torque signal can be used to characterize the energy distribution of the torque signal at different times and frequencies. Energy refers to the intensity or activity of the torque signal at a specific time and frequency.

[0080] In some embodiments, since the torque signal has strong noise and randomness, which has a great impact on the quality of the torque signal, the chemical mechanical polishing equipment can reduce the noise of the torque signal before extracting the torque feature of the torque signal.

[0081] In an optional embodiment, the chemical mechanical polishing device can use a low-pass filter to filter out high-frequency signals in the torque signal and retain low-frequency signals. Among them, high-frequency signals refer to signals that change very quickly, and low-frequency signals refer to signals that change relatively slowly. The low-pass filter can allow torque signals below a preset frequency threshold to pass, while weakening or blocking torque signals above the preset frequency threshold. Optionally, the low-pass filter can be an analog low-pass filter or a digital low-pass filter.

[0082] In an optional implementation, a multi-scale Gaussian filtering method of a one-dimensional time series signal can be used to reduce the noise of the torque signal. Among them, Gaussian filtering is a linear smoothing filter. The embodiment of the present application can obtain the filtered torque signal by weighted averaging. Specifically, the expression of the one-dimensional Gaussian filter is shown in formula (1):

[0083] Formula (1);

[0084] Among them, t can represent the time position corresponding to the torque signal, is the standard deviation of the Gaussian function, also known as the fuzziness coefficient, It can represent the torque signal after Gaussian filtering at time position t. To calculate the Gaussian kernel function, the standard deviation can be taken separately , and find the Gaussian kernel at different scales , and then according to the preset weights , weighted average the Gaussian kernels at different scales. Finally, the filtered torque signal can be expressed as: ,in, .

[0085] In the embodiment of the present application, by using a Gaussian filter to reduce the noise of the torque signal, the torque signal can be smoothed, high-frequency noise can be effectively removed, and important features of the torque signal can be retained, which helps to improve the accuracy of feature extraction.

[0086] In some embodiments, the chemical mechanical polishing device may normalize the torque signal before extracting the torque feature of the torque signal, so as to map the torque signal to the same scale range.

[0087] In an optional embodiment, the chemical mechanical polishing device may use a minimum-maximum scaling method to scale the torque signal to a fixed range, such as [0, 1]. The specific calculation formula of the minimum-maximum scaling method is shown in formula (2):

[0088] Formula (2);

[0089] Where X represents the torque signal, Indicates the minimum value in the torque signal, Indicates the maximum value in the torque signal, Represents the normalized torque signal.

[0090] In an optional embodiment, the chemical mechanical polishing equipment may use a Z-score standardization method to convert the torque signal into a standard normal distribution with a mean of 0 and a standard deviation of 1. The specific formula of Z-score standardization is shown in formula (3):

[0091] Formula (3);

[0092] Where X represents the torque signal, represents the mean value of the torque signal, represents the standard deviation of the torque signal, and Z represents the value after Z-score standardization.

[0093] By performing reasonable normalization on the torque signal, the scale differences between the torque signals can be eliminated and overfitting can be prevented, thereby improving the prediction accuracy and robustness of the polishing endpoint detection model.

[0094] Step 230, analyzing the torque feature set through the polishing endpoint detection model to determine a polishing endpoint detection result; the polishing endpoint detection result is used to indicate whether the polishing endpoint has been reached.

[0095] Among them, the polishing endpoint detection model can be used to judge whether the current wafer has reached the polishing endpoint. Optionally, the polishing endpoint detection model can be a neural network model. For example, convolutional neural networks (CNN), recurrent neural networks (RNN), long-short term memory networks (LSTM), deep neural networks (DNN), etc., can be set according to actual conditions, and the embodiments of the present application do not specifically limit this.

[0096] In some embodiments, the polishing endpoint detection result can be expressed by the polishing endpoint probability, that is, the predicted probability that the current wafer reaches the polishing endpoint. In the process of the drive motor driving the polishing disk to rotate, the chemical mechanical polishing device can analyze the current torque feature set through the polishing endpoint detection model to determine the current polishing endpoint probability. The chemical mechanical polishing device can pre-set the polishing endpoint probability threshold, and determine whether the wafer reaches the polishing endpoint by judging the relationship between the current polishing endpoint probability and the polishing endpoint probability threshold. Specifically, if the polishing endpoint probability is greater than or equal to the polishing endpoint probability threshold, the wafer reaches the polishing endpoint, and the chemical mechanical polishing device can control the drive motor to stop driving the polishing disk; if the polishing endpoint probability is less than the polishing endpoint probability threshold, the wafer has not reached the polishing endpoint, and the chemical mechanical polishing device can control the drive motor to continue to drive the polishing disk to rotate until it is detected that the current polishing endpoint probability is greater than or equal to the polishing endpoint probability threshold.

[0097] In some embodiments, the polishing endpoint detection result may be represented by a value of 1 and a value of 0. The value of 1 may indicate that the wafer has reached the polishing endpoint, and the value of 0 may indicate that the wafer has not reached the polishing endpoint.

[0098] In some embodiments, the chemical mechanical polishing equipment can be communicatively connected with a terminal device. After analyzing the current torque feature set through a polishing endpoint detection model and determining a current polishing endpoint detection result, the chemical mechanical polishing equipment can send the polishing endpoint detection result to the terminal device. The terminal device can receive the polishing endpoint result and display it on a display interface. The staff can view the polishing endpoint detection result through the endpoint device to achieve real-time monitoring of the wafer polishing status.

[0099] In an embodiment of the present application, during the process of chemical mechanical polishing of a wafer, the chemical mechanical polishing equipment collects a torque signal corresponding to the preset number of revolutions driven by the driving motor to drive the polishing disk; extracts torque features based on the torque signal to obtain a torque feature set; the torque feature set includes at least one of the frequency domain features, time domain features, and time-frequency domain features corresponding to the torque signal; the torque feature set is analyzed by a polishing endpoint detection model to determine a polishing endpoint detection result, which is used to characterize whether the polishing endpoint has been reached. In an embodiment of the present application, during the process of chemical mechanical polishing of a wafer, a torque signal corresponding to the preset number of revolutions driven by the driving motor to drive the polishing disk can be collected, and a torque feature set corresponding to the torque signal can be extracted. By inputting the torque feature set corresponding to the torque signal into the polishing endpoint detection model for analysis, it is possible to accurately determine whether the polishing endpoint has been reached, thereby improving the accuracy and efficiency of polishing endpoint detection.

[0100] In some embodiments, the torque feature set includes a combined feature vector of the time domain features and the frequency domain features corresponding to the torque signal, such as Figure 3 As shown, the step of extracting the torque feature according to the torque signal to obtain a torque feature set may include the following steps:

[0101] Step 310: extracting time domain features from the torque signal to obtain a time domain feature vector.

[0102] The chemical mechanical polishing device can extract multiple time domain features from the torque signal, and the time domain feature values ​​corresponding to the multiple time domain features respectively constitute the time domain feature vector corresponding to the torque signal. Exemplarily, the chemical mechanical polishing device can use the average value, variance, peak-to-peak value and effective value corresponding to the torque signal as the time domain feature vector corresponding to the torque signal.

[0103] For example, the time domain feature vector can be expressed as Represented by, where n can represent the total number of time domain eigenvalues ​​in the time domain eigenvector, It can represent the i-th time domain eigenvalue.

[0104] Step 320: extract frequency domain features from the torque signal to obtain a frequency domain feature vector.

[0105] The chemical mechanical polishing equipment can convert the torque signal from the time domain to the frequency domain to obtain the frequency spectrum corresponding to the torque signal, extract multiple frequency domain features corresponding to the torque signal from the frequency spectrum, and form the frequency domain feature vector corresponding to the torque signal from the frequency domain feature values ​​corresponding to the multiple frequency domain features. Exemplarily, the chemical mechanical polishing equipment can use the average frequency, center frequency, root mean square frequency, and root variance frequency obtained from the analysis of the frequency spectrum corresponding to the torque signal as the frequency domain feature vector corresponding to the torque signal.

[0106] For example, the frequency domain feature vector can be expressed as Represented by, where m can represent the total number of frequency domain eigenvalues ​​in the frequency domain eigenvector, It can represent the jth frequency domain eigenvalue.

[0107] In some embodiments, the chemical mechanical polishing device can use Fourier transform to convert the torque signal from the time domain to the frequency domain to obtain the frequency spectrum corresponding to the torque signal. Specifically, the mathematical expression of Fourier transform is shown in formula (4):

[0108] Formula (4);

[0109] Wherein, x(t) represents the torque signal in the time domain, X(f) represents the torque signal in the frequency domain, f represents the frequency, and j represents the imaginary unit.

[0110] Step 330: Combine the time domain feature vector and the frequency domain feature vector to obtain a combined feature vector.

[0111] The chemical mechanical polishing equipment can combine the time domain feature vector with the frequency domain feature vector to obtain a combined feature vector.

[0112] In some embodiments, the chemical mechanical polishing device may directly cascade the time domain feature vector and the frequency domain feature vector to obtain a combined feature vector.

[0113] For example, it is assumed that the time domain feature vector corresponding to the torque signal is , where n represents the total number of time domain eigenvalues ​​in the time domain eigenvector. It can represent the i-th time domain eigenvalue, and the frequency domain eigenvector corresponding to the torque signal is , where m represents the total number of frequency domain eigenvalues ​​in the frequency domain eigenvector. can represent the jth frequency domain eigenvalue, then the combined eigenvector obtained by combining the time domain eigenvector and the frequency domain eigenvector can be .

[0114] In the embodiment of the present application, the chemical mechanical polishing device can extract time domain features from the torque signal to obtain a time domain feature vector; extract frequency domain features from the torque signal to obtain a frequency domain feature vector; and combine the time domain feature vector with the frequency domain feature vector to obtain a combined feature vector. By combining the time domain feature vector with the frequency domain feature vector to obtain a combined feature vector, all information of the original features can be retained, thereby improving the accuracy of the polishing endpoint detection result obtained by analyzing the torque feature set through the polishing endpoint detection model in the subsequent step.

[0115] In some embodiments, the torque feature set includes a time-frequency domain diagram corresponding to the torque signal, and the time-frequency domain diagram is used to characterize the time-frequency domain characteristics corresponding to the torque signal. The step of extracting the torque feature according to the torque signal to obtain the torque feature set may include the step of: converting the torque signal into a time-frequency domain diagram based on a wavelet function.

[0116] Specifically, the chemical mechanical polishing equipment can use continuous wavelet transform to convert the torque signal into a time-frequency domain diagram. The chemical mechanical polishing equipment can obtain the wavelet coefficient by convolving the torque signal with the scaled and translated wavelet function, and then use the wavelet coefficient as the grayscale value of the time-frequency domain diagram to generate a time-frequency domain diagram corresponding to the torque signal. In the time-frequency domain diagram, the horizontal axis represents time, the vertical axis represents frequency, and the grayscale value represents the energy of the torque signal at a specific time and frequency. The specific mathematical expression is shown in the following formula (5):

[0117] Formula (5);

[0118] Where t represents time, Represents the torque signal corresponding to time t; represents the mother wavelet function, which can define the basic shape of the wavelet; a represents the scaling factor, which can be used to control the width of the wavelet. A larger a value will make the wavelet wider, and a smaller a value will make the wavelet narrower; b represents the translation factor, which can be used to control the position of the wavelet. By changing the value of b, the wavelet can be moved on the time axis; Represents the wavelet coefficients of the torque signal corresponding to time t.

[0119] In some embodiments, after obtaining the time-frequency domain image, the chemical mechanical polishing equipment may compress the time-frequency domain image.

[0120] In an optional embodiment, the chemical mechanical polishing equipment can perform principal component analysis on the time-frequency domain diagram to obtain a compressed time-frequency domain diagram, and input the compressed time-frequency domain diagram into the polishing endpoint detection model. Among them, principal component analysis is a data dimensionality reduction technology that can convert original high-dimensional data into a new low-dimensional space while retaining the information of the original data as much as possible. Specifically, the chemical mechanical polishing equipment can calculate the covariance matrix or correlation coefficient matrix corresponding to the time-frequency domain diagram, and then perform eigenvalue decomposition on the covariance matrix or correlation coefficient matrix to obtain eigenvalues ​​and corresponding eigenvectors, and select k eigenvectors from large to small according to the eigenvalues ​​to form a matrix to form a compressed time-frequency domain diagram, thereby achieving compression of the time-frequency domain diagram.

[0121] In an optional embodiment, the chemical mechanical polishing equipment can use a neural network model to compress the time-frequency domain image. The neural network model may include autoencoders (AE) and generative adversarial networks (GANs). Among them, the autoencoder is composed of an encoder and a decoder, and the time-frequency domain image can be compressed into a low-dimensional representation by the encoder, and then the time-frequency domain image can be reconstructed from this low-dimensional representation by the decoder. The generative adversarial network is composed of a generator and a discriminator. The generator maps the low-dimensional latent vector to the high-dimensional image space, and the discriminator attempts to distinguish the generated time-frequency domain image from the original time-frequency domain image. Through this adversarial training process, the time-frequency domain image can be compressed.

[0122] By compressing the time-frequency domain graph, the chemical mechanical polishing equipment can obtain a time-frequency domain graph of appropriate size. At the same time, it helps to remove noise in the time-frequency domain graph and improve the accuracy of the polishing endpoint probability obtained by the polishing endpoint detection model.

[0123] In the embodiment of the present application, the chemical mechanical polishing device converts the torque signal into a time-frequency domain diagram based on the wavelet function. The torque signal usually has non-stationary and nonlinear characteristics. Through the wavelet transform, the time-varying characteristics in the torque signal can be more accurately analyzed, and the accuracy of the torque signal analysis can be improved; in addition, the wavelet transform can capture the instantaneous changes and complex results in the signal. By converting the torque signal into a time-frequency domain diagram, the details and anomalies in the signal can be seen more clearly.

[0124] like Figure 4 As shown, in some embodiments, the polishing endpoint detection model includes a first channel and a second channel. The step of analyzing the torque feature set through the polishing endpoint detection model to determine the polishing endpoint detection result may include the following steps:

[0125] Step 410: input the combined feature vector into the first channel, perform convolution processing and pooling processing on the combined feature vector through the first channel, and obtain a first feature vector.

[0126] It should be noted that the polishing endpoint detection model includes a first channel and a second channel, the first channel and the second channel respectively include a convolution layer, a pooling layer and a fully connected layer, the first channel and the second channel intersect at the feature fusion layer, and then pass through the fully connected layer for classification output. It is understandable that the number of convolution layers and pooling layers can be adjusted according to actual needs.

[0127] The chemical mechanical polishing device can input the combined feature vector into the first channel, and perform convolution processing and pooling processing on the combined feature vector through the first channel to obtain the first feature vector.

[0128] Among them, the first channel can be used to perform convolution and pooling processing on the combined feature vector.

[0129] Convolution processing of the combined feature vector refers to an operation of multiplying the convolution kernel corresponding to the first channel and the combined feature vector element by element and then adding them. The convolution kernel corresponding to the first channel refers to a weight matrix corresponding to the first channel, and the convolution kernel corresponding to the first channel can be pre-set and stored in the chemical mechanical polishing device. By convolution processing of the combined feature vector, a feature map corresponding to the combined feature vector can be obtained.

[0130] Pooling the combined feature vector refers to using a pooling window to slide on the feature map corresponding to the combined feature vector according to a preset step size, and processing the data in the area covered by the pooling window according to the pooling type.

[0131] The pooling window refers to the size of the area where the data is viewed each time; the preset step size refers to the number of pixels that the pooling window crosses when it moves horizontally or vertically; optionally, the pooling type may include but is not limited to maximum pooling, average pooling, etc., which are not specifically limited in the embodiments of the present application. Maximum pooling refers to finding the maximum value of the data in the area covered by the pooling window; average pooling refers to calculating the average value of the data in the area covered by the pooling window.

[0132] In some embodiments, the chemical mechanical polishing device may input the pooled combined feature vector into the fully connected layer in the first channel for linear transformation. Specifically, the chemical mechanical polishing device may multiply the pooled combined feature vector with the weight matrix of the fully connected layer in the first channel and add a bias term to obtain the first feature vector after dimensionality reduction.

[0133] Step 420: input the time-frequency domain image into the second channel, perform convolution processing and pooling processing on the time-frequency domain image through the second channel, and obtain a second feature vector.

[0134] The chemical mechanical polishing equipment can input the time-frequency domain image into the second channel, perform convolution processing and pooling processing on the time-frequency domain image through the second channel, and obtain a second eigenvector.

[0135] Among them, the second channel can be used to perform convolution and pooling processing on the time-frequency domain image.

[0136] Convolution processing of the time-frequency domain image refers to the operation of multiplying the convolution kernel corresponding to the second channel and the time-frequency domain image element by element and then adding them. The convolution kernel corresponding to the second channel refers to the weight matrix corresponding to the second channel, and the convolution kernel corresponding to the second channel can be pre-set and stored in the chemical mechanical polishing device. By convolution processing of the time-frequency domain image, a feature map corresponding to the time-frequency domain image can be obtained.

[0137] Pooling the time-frequency domain image refers to using a pooling window to slide on the feature map corresponding to the time-frequency domain image according to a preset step size, and processing the data in the area covered by the pooling window according to the pooling type.

[0138] In some embodiments, the chemical mechanical polishing device may input the pooled time-frequency domain image into the fully connected layer in the second channel for linear transformation. Specifically, the chemical mechanical polishing device may multiply the pooled time-frequency domain image with the weight matrix of the fully connected layer in the second channel and add a bias term to obtain a second eigenvector after dimensionality reduction.

[0139] Step 430: fuse the first feature vector and the second feature vector to obtain a fused feature vector.

[0140] The chemical mechanical polishing equipment can perform a flattening operation on the first eigenvector and the second eigenvector, map the first eigenvector and the second eigenvector from a multi-dimensional space to a one-dimensional space, and then connect the flattened first eigenvector and the second eigenvector in sequence to obtain a fused feature vector, which can contain all information of the first eigenvector and the second eigenvector.

[0141] Step 440: Determine the polishing endpoint detection result according to the fused feature vector.

[0142] The chemical mechanical polishing equipment can determine the polishing endpoint detection result according to the fused feature vector.

[0143] In some embodiments, the chemical mechanical polishing device may input the fused feature vector into a fully connected layer for linear transformation. Specifically, the chemical mechanical polishing device may multiply the fused feature vector with the weight matrix of the fully connected layer and add a bias term to obtain a fused feature vector after dimension reduction.

[0144] In some embodiments, the chemical mechanical polishing device may use an activation function to map the fused feature vector after dimension reduction to obtain a polishing endpoint detection result. Exemplarily, the chemical mechanical polishing device may use a sigmoid activation function to map the fused feature vector after dimension reduction to a value between 0 and 1, which may represent the probability of the current wafer reaching the polishing endpoint, and use this value as the polishing endpoint detection result. The mathematical expression of the sigmoid function is shown in the following formula (6):

[0145] Formula (6);

[0146] Among them, x represents the fused feature vector after dimensionality reduction, Indicates the probability that the current wafer reaches the polishing endpoint.

[0147] In the embodiment of the present application, the chemical mechanical polishing equipment can input the combined feature vector into the first channel, perform convolution and pooling processing on the combined feature vector through the first channel to obtain the first feature vector; input the time-frequency domain image into the second channel, perform convolution and pooling processing on the time-frequency domain image through the second channel to obtain the second feature vector; fuse the first feature vector and the second feature vector to obtain a fused feature vector; and determine the polishing endpoint detection result based on the fused feature vector. Convolution processing can capture local features, and pooling processing can reduce the spatial dimension of features while retaining important information. Feature fusion can further integrate information so that the fused feature vector can more comprehensively reflect the data features in the torque signal, thereby obtaining a more accurate polishing endpoint detection result based on the fused feature vector.

[0148] like Figure 5As shown, in some embodiments, the chemical mechanical polishing equipment may first train a polishing endpoint detection model. The process of training the polishing endpoint detection model may include the following steps:

[0149] Step 510, extracting torque features according to the training torque signals contained in each group of training torque data, and obtaining a training torque feature set corresponding to each group of training torque data; wherein the training torque feature set includes at least one of frequency domain features, time domain features, and time-frequency domain features corresponding to the training torque signals.

[0150] Polishing endpoint detection can be obtained through training with a training set, which includes multiple groups of training torque data. Each group of training torque data may include a training torque signal and an endpoint label corresponding to the training torque signal. The endpoint label is used to indicate whether the polishing endpoint has been reached.

[0151] In some embodiments, the process of acquiring a training set may include the following steps: during chemical mechanical polishing of a wafer, collecting a training torque signal corresponding to a preset number of revolutions of the polishing disk driven by a driving motor, and marking the corresponding endpoint label for the training torque signal, using the training torque signal and the endpoint label corresponding to the training torque signal as a set of training torque data, and repeating the above process to obtain multiple sets of training torque data.

[0152] In some embodiments, before extracting torque features based on the training torque signals contained in each group of training torque data to obtain a set of training torque features corresponding to each group of training torque data, the chemical mechanical equipment may also normalize the training torque signals contained in each group of training torque data to map the training torque signals to a target range.

[0153] For example, the chemical mechanical polishing equipment may use Min-Max normalization processing to map the data to a specified range interval [0,1], as shown in the following formula (7):

[0154] Formula (7);

[0155] in, represents the normalized training torque signal, which is between 0 and 1; X is the training torque signal, represents the minimum value in the training torque signal, Indicates the maximum value in the training torque signal.

[0156] After obtaining the normalized training torque signal, the chemical mechanical polishing equipment can extract the torque feature according to the normalized training torque signal contained in each group of training torque data to obtain a training torque feature set corresponding to each group of training torque data.

[0157] By normalizing the training torque signals contained in each group of training torque data, each training torque signal is converted to the same scale, which is beneficial to the training of the polishing endpoint detection model to be trained in the subsequent steps, thereby improving the accuracy of the polishing endpoint detection result.

[0158] Step 520 , training the polishing endpoint detection model to be trained according to the training torque feature sets and the corresponding endpoint labels corresponding to the multiple sets of training torque data.

[0159] like Figure 6 As shown, in some embodiments, the step of training the polishing endpoint detection model to be trained according to the training torque feature sets and the corresponding endpoint labels corresponding to the multiple sets of training torque data may include the following steps:

[0160] Step 610: Analyze the training torque features corresponding to one or more sets of training torque data using the polishing endpoint detection model to be trained to determine the polishing endpoint prediction probability corresponding to each set of training torque data.

[0161] The polishing endpoint prediction probability corresponding to each set of training torque data may represent the prediction probability that the set of training torque data is the polishing endpoint. For example, when the polishing endpoint prediction probability corresponding to a set of training torque data is 50%, the prediction probability that the set of training torque data is the polishing endpoint is 50%.

[0162] Step 620 , calculating the classification loss of the polishing endpoint detection model to be trained according to the polishing endpoint prediction probability and the corresponding endpoint label corresponding to each group of training torque data.

[0163] The end point label can be represented by a value of 1 or a value of 0. The value of 1 represents the polishing end point, and the value of 0 represents not the polishing end point.

[0164] The chemical mechanical polishing equipment can determine the prediction probability of the polishing endpoint detection model to be trained for the correct category based on the polishing endpoint prediction probability and the corresponding endpoint label corresponding to each group of training torque data. Specifically, the chemical mechanical polishing equipment can determine the prediction probability of the polishing endpoint detection model to be trained under the endpoint label based on the polishing endpoint prediction probability and the corresponding endpoint label corresponding to each group of training torque data. Exemplarily, when the polishing endpoint prediction probability corresponding to a certain group of training torque data is 75% and the corresponding endpoint label is 1, it can be determined that the prediction probability of the polishing endpoint detection model to be trained for the correct category is 75%; when the polishing endpoint prediction probability corresponding to a certain group of training torque data is 80% and the corresponding endpoint label is 0, it can be determined that the prediction probability of the polishing endpoint detection model to be trained for the correct category is 20%.

[0165] After determining the prediction probability of the correct category of the polishing endpoint detection model to be trained, the chemical mechanical polishing equipment can use the focal loss function to calculate the classification loss. The focal loss function includes a balance factor and a modulation factor. Among them, the balance factor can be used to solve the problem of class imbalance. By adjusting the balance factor according to the ratio of positive and negative samples, a greater weight is given to the minority class; the modulation factor can be used to distinguish between samples that are easy to classify and samples that are difficult to classify, amplify or reduce the loss, weaken the focus on samples that are easy to classify, and pay more attention to samples that are difficult to classify. The expression of the focal loss function is shown in the following formula (8):

[0166] Formula (8);

[0167] in, It represents the predicted probability of the correct category by the polishing endpoint detection model to be trained, reflecting the closeness to the correct category. The larger it is, the closer it is to the correct category and the more accurate the classification.

[0168] is the balancing factor, is the modulation factor.

[0169] represents the classification loss, The larger the value is, the less accurate the prediction of the training torque data by the polishing endpoint detection model to be trained is; conversely, The smaller it is, the more accurate the prediction of the training torque data by the polishing endpoint detection model to be trained is.

[0170] Positive Category Negative ,in, It can be used to control the weights of positive and negative samples. The value range is [0,1]. The positive class indicates that the polishing endpoint has been reached, and the negative class indicates that the polishing endpoint has not been reached.

[0171] Through the modulation factor, the loss contribution of easy-to-classify samples is reduced, making the model more concerned with difficult-to-classify samples, thereby improving the generalization ability and robustness of the model; through the balancing factor, the imbalance problem of positive and negative samples is effectively alleviated, so that the model treats each type of sample more fairly during the training process, so that the trained model can more accurately identify the polishing state corresponding to the new torque signal, thereby improving the prediction accuracy of the polishing end point.

[0172] Step 630: Update the model parameters of the polishing endpoint detection model to be trained according to the classification loss until the polishing endpoint detection model converges.

[0173] The chemical mechanical polishing equipment updates the model parameters of the polishing endpoint detection model to be trained according to the classification loss until the polishing endpoint detection model converges.

[0174] In some embodiments, the chemical mechanical polishing device may update the model parameters of the polishing endpoint detection model to be trained at intervals of a preset number of cycles. For example, the chemical mechanical polishing device may update the model parameters of the polishing endpoint detection model to be trained at intervals of 10 iteration cycles.

[0175] In the embodiment of the present application, the chemical mechanical polishing equipment analyzes one or more sets of training torque features through the polishing endpoint detection model to be trained, determines the polishing endpoint prediction probability corresponding to each set of training torque data; calculates the classification loss of the polishing endpoint detection model to be trained according to the polishing endpoint prediction probability; updates the model parameters of the polishing endpoint detection model to be trained according to the classification loss until the polishing endpoint detection model converges. By learning the relationship between the training torque feature and the actual polishing endpoint, the model can more accurately identify the polishing state corresponding to the new torque signal, thereby improving the prediction accuracy of the polishing endpoint.

[0176] In the embodiment of the present application, the chemical mechanical polishing equipment extracts torque features according to the training torque signals contained in each set of training torque data, and obtains a set of training torque features corresponding to each set of training torque data; wherein the set of training torque features includes at least one of frequency domain features, time domain features, and time-frequency domain features corresponding to the training torque signals; and according to the training torque feature sets and corresponding endpoint labels corresponding to the multiple sets of training torque data, the polishing endpoint detection model to be trained is trained. By training the polishing endpoint detection model to be trained, the accuracy of the polishing endpoint detection model in analyzing the polishing endpoint detection results can be improved.

[0177] like Figure 7 As shown, in some embodiments, after the step of training the polishing endpoint detection model to be trained according to the training torque feature sets and the corresponding endpoint labels corresponding to the multiple sets of training torque data, the following steps may also be included:

[0178] Step 710, testing the trained polishing endpoint detection model through a test set to obtain model performance information corresponding to the trained polishing endpoint detection model; the test set includes multiple groups of test torque data, each group of test torque data includes a test torque signal and an endpoint label corresponding to the test torque signal.

[0179] The chemical mechanical polishing equipment can test the trained polishing endpoint detection model through a test set to obtain model performance information corresponding to the trained polishing endpoint detection model; the test set includes multiple groups of training torque data, each group of training torque data includes a test torque signal and an endpoint label corresponding to the test torque signal.

[0180] Among them, the model performance information can be used to characterize whether the trained polishing endpoint detection model can obtain accurate polishing endpoint detection results.

[0181] In some embodiments, the model performance information may include at least one of classification accuracy, classification recall, and classification precision. Exemplarily, the classification accuracy may be calculated based on the ratio of the number of correctly classified samples to the sum of the number of correctly classified samples and the number of incorrectly classified samples; the classification recall may be calculated based on the ratio of the number of positive samples correctly classified as positive to the sum of the number of positive samples correctly classified as positive and the number of positive samples incorrectly classified as negative; the classification precision may be calculated based on the ratio of the number of positive samples correctly classified as positive to the sum of the number of positive samples correctly classified as positive and the number of negative samples incorrectly classified as positive.

[0182] For example, the calculation formula of classification accuracy can refer to formula (9):

[0183] Formula (9);

[0184] For example, the calculation formula of classification recall rate (Recall) can refer to formula (10):

[0185] Formula (10);

[0186] For example, the calculation formula of classification accuracy (Precision) can refer to formula (11):

[0187] Formula (11);

[0188] Among them, TP represents the number of positive samples correctly classified as positive, TN represents the number of negative samples correctly classified as negative, FP represents the number of negative samples incorrectly classified as positive, FN represents the number of positive samples incorrectly classified as negative, positive samples represent torque signals that have reached the polishing end point, negative samples represent torque signals that have not reached the polishing end point, positive class represents that the polishing end point has been reached, and negative class represents that the polishing end point has been reached.

[0189] Step 720: If the model performance information cannot meet the performance condition, the training set is expanded, and the polishing endpoint detection model is continuously trained according to the expanded training set.

[0190] Step 730: If the model performance information meets the performance conditions, the polishing endpoint detection model training is completed.

[0191] Among them, the performance condition refers to the model performance information reaching a stable state. It can be understood that the chemical mechanical polishing equipment can test the trained polishing endpoint detection model through the test set at the end of each iteration cycle, obtain the model performance information corresponding to the trained polishing endpoint detection model, and record the model performance information. When it is observed that the changes between the model performance information of multiple consecutive iteration cycles are small, it can be determined that the model performance information has reached a stable state, thereby determining that the polishing endpoint detection model training is completed.

[0192] In the embodiment of the present application, the chemical mechanical polishing equipment tests the trained polishing endpoint detection model through the test set to obtain the model performance information corresponding to the trained polishing endpoint detection model; the test machine includes multiple groups of test torque data, each group of test torque data includes a test torque signal and an endpoint label corresponding to the test torque signal; if the model performance information cannot meet the performance conditions, the training set is expanded, and the polishing endpoint detection model is continued to be trained according to the expanded training set; if the model performance information meets the performance conditions, the polishing endpoint detection model training is completed. By evaluating the performance of the trained polishing endpoint detection model through the model performance information, it can be ensured that the trained polishing endpoint detection model can obtain accurate polishing endpoint detection results.

[0193] See also Figure 8 , Figure 8 800 is a schematic diagram of a polishing endpoint detection device disclosed in an embodiment of the present application. The polishing endpoint detection device 800 can be applied to the chemical mechanical polishing equipment 100. Figure 8 As shown, the polishing endpoint detection device 800 may include: a collection module 810 , an extraction module 820 and a determination module 830 .

[0194] The acquisition module 810 is used to acquire a torque signal corresponding to a preset number of revolutions of the polishing disk driven by the driving motor during the chemical mechanical polishing of the wafer.

[0195] The extraction module 820 is used to extract torque features according to the torque signal to obtain a torque feature set; the torque feature set includes at least one of frequency domain features, time domain features and time-frequency domain features corresponding to the torque signal.

[0196] The determination module 830 is used to analyze the torque feature set through the polishing endpoint detection model to determine the polishing endpoint detection result, and the polishing endpoint detection result is used to indicate whether the polishing endpoint is reached.

[0197] In some embodiments, the torque feature set includes a combined feature vector of time domain features and frequency domain features corresponding to the torque signal; the extraction module 820 is also used to extract time domain features from the torque signal to obtain a time domain feature vector; extract frequency domain features from the torque signal to obtain a frequency domain feature vector; and combine the time domain feature vector with the frequency domain feature vector to obtain a combined feature vector.

[0198] In some embodiments, the torque feature set includes a time-frequency domain diagram corresponding to the torque signal, which is used to characterize the time-frequency domain features corresponding to the torque signal; the extraction module 820 is also used to convert the torque signal into a time-frequency domain diagram based on a wavelet function.

[0199] In some embodiments, the polishing endpoint detection model includes a first channel and a second channel; the determination module 830 is also used to input the combined feature vector into the first channel, perform convolution and pooling processing on the combined feature vector through the first channel, and obtain a first feature vector; input the time-frequency domain image into the second channel, perform convolution and pooling processing on the time-frequency domain image through the second channel, and obtain a second feature vector; fuse the first feature vector and the second feature vector to obtain a fused feature vector; and determine the polishing endpoint detection result based on the fused feature vector.

[0200] In some embodiments, the polishing endpoint detection model is obtained by training a training set, which includes multiple groups of training torque data, each group of training torque data is a training torque signal and an endpoint label corresponding to the training torque signal, and the endpoint label is used to characterize whether the polishing endpoint is reached; the polishing endpoint detection device 800 also includes a training feature extraction module and a training module.

[0201] A training feature extraction module is used to extract torque features based on the training torque signals contained in each group of training torque data to obtain a training torque feature set corresponding to each group of training torque data; wherein the training torque feature set includes at least one of the frequency domain features, time domain features, and time-frequency domain features corresponding to the training torque signals.

[0202] The training module is used to train the polishing endpoint detection model to be trained according to the training torque feature sets and the corresponding endpoint labels corresponding to the multiple sets of training torque data.

[0203] In some embodiments, the polishing endpoint detection device further includes a normalization module.

[0204] The normalization module is used to perform normalization processing on each training torque signal contained in each group of training torque data so as to map each training torque signal to a target range.

[0205] The training feature extraction module is also used to extract torque features based on multiple training torque signals contained in each group of training torque data after normalization, so as to obtain a training torque feature set corresponding to each group of training torque data.

[0206] In some embodiments, the training module further includes an analyzing unit, a determining unit, a calculating unit, and an updating unit.

[0207] The analysis unit is used to analyze the training torque feature set corresponding to one or more sets of training torque data through the polishing endpoint detection model to be trained, and determine the polishing endpoint prediction probability corresponding to each set of training torque data.

[0208] The calculation unit calculates the classification loss of the polishing endpoint detection model to be trained according to the polishing endpoint prediction probability.

[0209] The updating unit updates the model parameters of the polishing endpoint detection model to be trained according to the classification loss until the polishing endpoint detection model converges.

[0210] In some embodiments, the polishing endpoint detection device 800 further includes a testing module.

[0211] A testing module is used to test the trained polishing endpoint detection model through a test set to obtain model performance information corresponding to the trained polishing endpoint detection model; the test set includes multiple groups of test torque data, each group of test torque data includes a test torque signal and an endpoint label corresponding to the test torque signal; if the model performance information cannot meet the performance conditions, the training set is expanded, and the polishing endpoint detection model is continued to be trained according to the expanded training set; if the model performance information meets the performance conditions, the polishing endpoint detection model training is completed.

[0212] In an embodiment of the present application, during the process of chemical mechanical polishing of a wafer, the chemical mechanical polishing equipment collects a torque signal corresponding to the preset number of revolutions driven by the driving motor to drive the polishing disk; extracts torque features based on the torque signal to obtain a torque feature set; the torque feature set includes at least one of the frequency domain features, time domain features, and time-frequency domain features corresponding to the torque signal; the torque feature set is analyzed by a polishing endpoint detection model to determine a polishing endpoint detection result, which is used to characterize whether the polishing endpoint has been reached. In an embodiment of the present application, during the process of chemical mechanical polishing of a wafer, a torque signal corresponding to the preset number of revolutions driven by the driving motor to drive the polishing disk can be collected, and a torque feature set corresponding to the torque signal can be extracted. By inputting the torque feature set corresponding to the torque signal into the polishing endpoint detection model for analysis, it is possible to accurately determine whether the polishing endpoint has been reached, thereby improving the accuracy and efficiency of polishing endpoint detection.

[0213] See also Fig. 9 , Fig. 9 Schematic diagram of the structure of an electronic device disclosed in the embodiment of the present application. Fig. 9 As shown, the electronic device 900 may include:

[0214] A memory 910 storing executable program codes;

[0215] a processor 920 coupled to the memory 910;

[0216] The processor 920 calls the executable program code stored in the memory 910 to execute any one of the polishing endpoint detection methods disclosed in the embodiments of the present application.

[0217] An embodiment of the present application discloses a computer-readable storage medium storing a computer program, wherein when the computer program is executed by the processor, the processor implements any one of the polishing endpoint detection methods disclosed in the embodiment of the present application.

[0218] It should be understood that the "one embodiment" or "an embodiment" mentioned throughout the specification means that the specific features, structures or characteristics related to the embodiment are included in at least one embodiment of the present application. Therefore, "in one embodiment" or "in an embodiment" appearing throughout the specification does not necessarily refer to the same embodiment. In addition, these specific features, structures or characteristics can be combined in one or more embodiments in any suitable manner. Those skilled in the art should also be aware that the embodiments described in the specification are all optional embodiments, and the actions and modules involved are not necessarily required for the present application.

[0219] In the various embodiments of the present application, it should be understood that the size of the serial numbers of the above-mentioned processes 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 embodiments of the present application.

[0220] The units described above 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.

[0221] In addition, each functional unit in each embodiment of the present application 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.

[0222] If the above-mentioned integrated 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-accessible memory. Based on this understanding, the technical solution of the present application, in essence, 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, which is stored in a memory and includes several requests for a computer device (which can be a personal computer, a server or a network device, etc., specifically a processor in a computer device) to perform some or all of the steps of the above-mentioned methods of various embodiments of the present application.

[0223] Those skilled in the art will appreciate that all or part of the steps in the various methods of the above embodiments may be completed by instructing related hardware through a program, and the program may be stored in a computer-readable storage medium, the storage medium including a read-only memory (ROM), a random access memory (RAM), a programmable read-only memory (PROM), an erasable programmable read-only memory (EPROM), a one-time programmable read-only memory (OTPROM), an electronically erasable rewritable read-only memory (EEPROM), a compact disc (CD-ROM) or other optical disc storage, magnetic disk storage, magnetic tape storage, or any other computer-readable medium that can be used to carry or store data.

[0224] The above is a detailed introduction to a polishing endpoint detection method, device, electronic device and storage medium disclosed in the embodiment of the present application. The principle and implementation method of the present application are described in detail using specific examples. The description of the above embodiments is only used to help understand the method and core idea of ​​the present application. At the same time, for those of ordinary skill in the art, according to the idea of ​​the present application, there will be changes in the specific implementation method and application scope. In summary, the content of this specification should not be understood as limiting the present application.

Claims

1. A polishing endpoint detection method, characterized in that: Applied to a chemical mechanical polishing device, the chemical mechanical polishing device comprises a driving motor and a polishing disc, the driving motor is used to drive the polishing disc to rotate, and the method comprises: During chemical mechanical polishing of the wafer, a torque signal corresponding to a preset number of revolutions of the polishing disk driven by the drive motor is collected; Extracting torque features according to the torque signal to obtain a torque feature set; the torque feature set includes at least one of a frequency domain feature, a time domain feature, and a time-frequency domain feature corresponding to the torque signal; The torque feature set is analyzed by a polishing endpoint detection model to determine a polishing endpoint detection result; the polishing endpoint detection result is used to indicate whether the polishing endpoint has been reached; the polishing endpoint detection model is obtained by training a training set, the training set includes multiple sets of training torque data, each set of training torque data includes a training torque signal and an endpoint label corresponding to the training torque signal, and the endpoint label is used to indicate whether the polishing endpoint has been reached; The torque feature set includes a combined feature vector of the time domain feature and the frequency domain feature corresponding to the torque signal, and also includes a time-frequency domain diagram corresponding to the torque signal, wherein the time-frequency domain diagram is used to characterize the time-frequency domain features corresponding to the torque signal; the polishing endpoint detection model includes a first channel and a second channel; and analyzing the torque feature set through the polishing endpoint detection model to determine the polishing endpoint detection result includes: Inputting the combined feature vector into the first channel, performing convolution processing and pooling processing on the combined feature vector through the first channel to obtain a first feature vector; Inputting the time-frequency domain image into the second channel, performing convolution processing and pooling processing on the time-frequency domain image through the second channel to obtain a second feature vector; Fusing the first feature vector and the second feature vector to obtain a fused feature vector; A polishing endpoint detection result is determined according to the fused feature vector.

2. The method according to claim 1, characterized in that The step of extracting the torque feature according to the torque signal to obtain a torque feature set includes: Extracting the time domain feature from the torque signal to obtain a time domain feature vector; Extracting the frequency domain feature from the torque signal to obtain a frequency domain feature vector; The time domain feature vector is combined with the frequency domain feature vector to obtain a combined feature vector.

3. The method according to claim 2, characterized in that The step of extracting the torque feature according to the torque signal to obtain a torque feature set includes: Based on the wavelet function, the torque signal is converted into a time-frequency domain diagram.

4. The method according to claim 1, characterized in that: In the process of chemical mechanical polishing of the wafer, before collecting the torque signal corresponding to the polishing disk being driven by the driving motor to rotate a preset number of times, the method further includes: Extracting torque features according to the training torque signals contained in each group of the training torque data to obtain a training torque feature set corresponding to each group of the training torque data; wherein the training torque feature set includes at least one of a frequency domain feature, a time domain feature, and a time-frequency domain feature corresponding to the training torque signal; The polishing endpoint detection model to be trained is trained according to the training torque feature sets and the corresponding endpoint labels respectively corresponding to the multiple groups of training torque data.

5. The method according to claim 4, characterized in that Before extracting torque features according to the training torque signals contained in each group of the training torque data to obtain a training torque feature set corresponding to each group of the training torque data, the method further includes: Normalizing the training torque signal contained in each group of the training torque data to map the training torque signal to a target range; The step of extracting torque features according to the training torque signals contained in each group of the training torque data to obtain a set of training torque features corresponding to each group of the training torque data includes: According to the normalized training torque signal contained in each group of the training torque data, the torque feature is extracted to obtain a training torque feature set corresponding to each group of the training torque data.

6. The method according to claim 4, characterized in that The step of training the polishing endpoint detection model to be trained according to the training torque feature sets and the corresponding endpoint labels respectively corresponding to the multiple sets of training torque data includes: Analyzing a set of training torque features corresponding to one or more sets of training torque data by using a polishing endpoint detection model to be trained, and determining a polishing endpoint prediction probability corresponding to each set of the training torque data; Calculating the classification loss of the polishing endpoint detection model to be trained according to the polishing endpoint prediction probability and the corresponding endpoint label corresponding to each group of the training torque data; According to the classification loss, the model parameters of the polishing endpoint detection model to be trained are updated until the polishing endpoint detection model converges.

7. The method according to any one of claims 4 to 6, characterized in that: After training the polishing endpoint detection model to be trained according to the training torque feature sets and the corresponding endpoint labels respectively corresponding to the multiple sets of training torque data, the method further includes: The trained polishing endpoint detection model is tested by a test set to obtain model performance information corresponding to the trained polishing endpoint detection model; the test set includes multiple groups of test torque data, each group of test torque data includes a test torque signal and an endpoint label corresponding to the test torque signal; If the model performance information does not meet the performance condition, expanding the training set, and continuing to train the polishing endpoint detection model according to the expanded training set; If the model performance information meets the performance conditions, the polishing endpoint detection model training is completed.

8. A polishing endpoint detection device, characterized in that: Applied to chemical mechanical polishing equipment, the chemical mechanical polishing equipment includes a drive motor and a polishing disc, the drive motor is used to drive the polishing disc to rotate, and the device includes: An acquisition module, used for acquiring a torque signal corresponding to a preset number of revolutions of the polishing disk driven by the drive motor during chemical mechanical polishing of the wafer; an extraction module, configured to extract torque features according to the torque signal to obtain a torque feature set; the torque feature set includes at least one of a frequency domain feature, a time domain feature, and a time-frequency domain feature corresponding to the torque signal; A determination module is used to analyze the torque feature set through a polishing endpoint detection model to determine a polishing endpoint detection result; the polishing endpoint detection result is used to indicate whether the polishing endpoint has been reached; the polishing endpoint detection model is obtained by training a training set, the training set includes multiple groups of training torque data, each group of training torque data includes a training torque signal and an endpoint label corresponding to the training torque signal, and the endpoint label is used to indicate whether the polishing endpoint has been reached; The torque feature set includes a combined feature vector of the time domain feature and the frequency domain feature corresponding to the torque signal, and also includes a time-frequency domain graph corresponding to the torque signal, wherein the time-frequency domain graph is used to characterize the time-frequency domain features corresponding to the torque signal; the polishing endpoint detection model includes a first channel and a second channel; The determination module is further used to input the combined feature vector into the first channel, perform convolution and pooling processing on the combined feature vector through the first channel to obtain a first feature vector; input the time-frequency domain image into the second channel, perform convolution and pooling processing on the time-frequency domain image through the second channel to obtain a second feature vector; fuse the first feature vector and the second feature vector to obtain a fused feature vector; and determine the polishing endpoint detection result based on the fused feature vector.

9. An electronic device, characterized in that: include: A memory storing executable program code; a processor coupled to the memory; The processor calls the executable program code stored in the memory to execute the method according to any one of claims 1 to 7.

10. A computer-readable storage medium, characterized in that: The computer-readable storage medium stores a computer program, wherein when the computer program is executed by a processor, the processor is caused to perform the method according to any one of claims 1 to 7.

Images