Method and system for detecting flatness of CMP polishing pad

Abstract

The invention discloses a method and system for detecting the flatness of a CMP polishing pad, and the method comprises the following steps: S1, obtaining the information of a to-be-detected region of the polishing pad, and determining a radial multi-path scanning planning scheme; s2, synchronously collecting near-infrared absorption spectrum data and eddy current induction data on the radial multiple paths through a multi-spectrum near-infrared-eddy current cooperative detection module; and S3, constructing a signal correction prediction model based on the generative adversarial network. The invention relates to the technical field of chemical mechanical polishing. According to the method and system for detecting the flatness of the CMP polishing pad, automation and integration of the detection process are achieved, all the modules work cooperatively, the detection efficiency is improved, manual intervention is reduced, the influence of human factors on the detection result is reduced, and the detection result is more objective and accurate.

Description

Technical Field

[0001] This invention relates to the field of chemical mechanical polishing, and more specifically, to a method and system for detecting the flatness of CMP polishing pads. Background Technology

[0002] In the chemical mechanical polishing (CMP) process of semiconductor manufacturing, the flatness of the polishing pad plays a crucial role in the flatness and uniformity of the wafer surface. Poor polishing pad flatness can lead to defects such as scratches and uneven thickness on the wafer surface, which in turn affects the performance and yield of semiconductor devices.

[0003] Currently, traditional CMP polishing pad flatness testing methods have several limitations. On one hand, some methods employ contact testing, which can easily damage the polishing pad surface, affecting its subsequent performance. Furthermore, contact testing is slow, making it difficult to meet the efficiency requirements of large-scale production. On the other hand, while non-contact testing methods avoid contact damage, existing non-contact technologies often only acquire single types of data. They lack effective correction methods for signal drift caused by polishing pad wear or dielectric wetting, thus affecting the accuracy and reliability of the test results. In addition, existing methods need improvement in the accuracy and comprehensiveness of constructing three-dimensional morphology models of the polishing pad surface and extracting flatness feature parameters, failing to meet the urgent needs of the semiconductor manufacturing industry for high-precision, high-reliability testing. Therefore, developing a CMP polishing pad flatness testing method and system that overcomes these shortcomings is of significant practical importance. Summary of the Invention

[0004] The purpose of this invention is to provide a method and system for detecting the flatness of CMP polishing pads. It solves the problem that existing non-contact detection technologies can only obtain a single type of data and lack effective correction methods for signal drift caused by factors such as polishing pad wear or media wetting, thus affecting the accuracy and reliability of the detection results and failing to meet the application requirements.

[0005] This invention achieves the above objective through the following technical solution: a method for detecting the flatness of a CMP polishing pad, the method comprising the following steps: S1. Obtain the information of the area to be inspected on the polishing pad and determine the radial multi-path scanning plan; S2. Through the multi-band near-infrared-eddy current collaborative detection module, near-infrared absorption spectrum data and eddy current sensing data on the radial multi-path are collected simultaneously; S3. Construct a signal correction prediction model based on generative adversarial network. Input the near-infrared absorption spectrum data and eddy current induction data on radial multipaths synchronously collected by the multi-band near-infrared-eddy current collaborative detection module to train the model and automatically correct the signal drift caused by polishing pad wear or medium wetting. S4. Construct a three-dimensional morphological model of the polishing pad surface using the corrected fused data, extract flatness feature parameters, and realize dynamic non-contact detection of polishing pad flatness.

[0006] Furthermore, the multi-band near-infrared-eddy current collaborative detection module is an integrated design, including: The multi-band near-infrared scanning unit and the eddy current sensing unit maintain their relative positions during the data acquisition process through a fixed structure. The multi-band near-infrared scanning unit operates in spectral bands covering key near-infrared bands, and includes at least five continuously adjustable spectral band channels, with each channel having a wavelength adjustment accuracy of not less than [specified value]. ; The eddy current sensing unit adopts a high-frequency excitation design, with an adjustable excitation frequency and a sampling frequency not lower than [missing information]. .

[0007] Furthermore, the determination of the radial multipath scan planning scheme includes: With the center of the polishing pad as the origin, the polishing pads are evenly distributed radially. Scan path, The angle between adjacent scan paths is equal; The spacing between sampling points on each scanning path is no greater than [missing value]. The scanning path covers the effective working area of ​​the polishing pad. above; Set up a round-trip scanning mode to eliminate abnormal data by checking for data redundancy in overlapping areas of the path.

[0008] Furthermore, the simultaneous acquisition of near-infrared absorption spectral data and eddy current induction data along multiple radial paths includes: A high-precision timestamp synchronization mechanism is adopted to ensure the timestamp error between the two types of data. ; Baseline correction and normalization were performed on the near-infrared absorption spectroscopy data, and the data was compressed to [size missing]. interval; An adaptive filtering algorithm is used to suppress noise in the eddy current induction data while retaining the effective signal.

[0009] Furthermore, the normalization calculation formula for the near-infrared absorption spectral data is as follows: in, For a certain wavelength The original absorption value below, This is the baseline value obtained by polynomial fitting at this wavelength.

[0010] Furthermore, the signal correction prediction model constructed based on generative adversarial networks includes: The model consists of four steps: model structure design, training data preparation, training process execution, and model optimization. The generator in the signal correction prediction model built on generative adversarial networks adopts an improved encoder-decoder network structure, which includes skip connections and attention mechanism modules; The discriminator in the signal correction prediction model based on generative adversarial networks adopts a lightweight convolutional neural network structure. The training data covers a variety of working conditions, including polishing pad materials, wear levels, and media wetting states. After preprocessing, the data is divided into training, validation, and test sets. The training process adopts an alternating training mode, and the model parameters are optimized by a hybrid loss function. After optimization, the model meets the preset performance indicators.

[0011] Furthermore, the formula for calculating the hybrid loss function is as follows: in, For mean square error loss, To help the generator combat loss, The weighting coefficient has a range of values. The optimal value is determined through cross-validation.

[0012] Furthermore, the automatic correction of signal drift caused by polishing pad wear or dielectric wetting includes: Feature extraction is performed on the preprocessed near-infrared absorption spectral data to construct a surface interference feature vector; The feature vector and the original eddy current induction data are input into the trained model, and the gain correction coefficient is output. With offset correction factor ; Based on the aforementioned correction coefficient, offset compensation is performed on the original eddy current data. The compensation formula is as follows: in, This is the original eddy current data after filtering. The data is the corrected eddy current data; when the correction deviation does not meet the threshold requirement, the online fine-tuning mechanism of the model is triggered.

[0013] Furthermore, the construction of the three-dimensional morphology model of the polishing pad surface and the extraction of flatness feature parameters include: Based on the corrected fused data, a 3D point cloud model is constructed using a spatial interpolation algorithm. An interpolation error threshold is set, and when the theoretical error of the interpolation result exceeds the threshold, the sampling points are automatically encrypted. Feature parameters are extracted from the 3D point cloud model, including: maximum peak-to-valley difference, flatness error, local undulation standard deviation, and radial thickness variation rate. The qualified thresholds for each parameter are determined based on the application scenario, industry standards, and statistical process control methods.

[0014] A system for detecting the flatness of CMP polishing pads, applied to the above-described method for detecting the flatness of CMP polishing pads, the system comprising: The module includes a scanning planning module, a data acquisition module, a signal correction module, a 3D modeling and feature extraction module, and a result output module. The scanning planning module is used to acquire information about the area to be inspected on the polishing pad and generate a radial multipath scanning planning scheme; The data acquisition module is a multi-band near-infrared-eddy current collaborative detection module, used to simultaneously acquire near-infrared absorption spectrum data and eddy current sensing data; The signal correction module has a built-in signal correction prediction model based on generative adversarial networks, which is used to correct signal drift. The 3D modeling and feature extraction module is used to construct a 3D morphology model of the polishing pad surface and extract flatness feature parameters; The results output module is used to output the flatness test results and characteristic parameter analysis report.

[0015] The beneficial effects of this invention are as follows: 1. By using a multi-band near-infrared-eddy current collaborative detection module, near-infrared absorption spectrum data and eddy current induction data on multiple radial paths are collected simultaneously. The signal correction prediction model constructed by generative adversarial network is used to automatically correct signal drift, effectively eliminating the interference of factors such as polishing pad wear or medium wetting on the detection signal, improving the accuracy and reliability of data acquisition, and thus realizing high-precision detection of CMP polishing pad flatness.

[0016] 2. This invention adopts a non-contact detection method, which avoids the damage to the polishing pad surface caused by traditional contact detection, ensuring the integrity of the polishing pad and its subsequent performance. At the same time, it can dynamically acquire the flatness information of the polishing pad in real time, meeting the requirements of detection efficiency and real-time performance in large-scale production. This helps to detect changes in the flatness of the polishing pad in a timely manner, adjust process parameters promptly, and improve production efficiency and product quality.

[0017] 3. By constructing a three-dimensional morphological model of the polishing pad surface and extracting various flatness characteristic parameters such as the maximum peak-to-valley difference, flatness error, local undulation standard deviation, and radial thickness variation rate, the flatness of the polishing pad can be comprehensively and accurately evaluated. These characteristic parameters provide rich data support for in-depth analysis of polishing pad performance and optimization of polishing process, which helps to further improve the quality and stability of semiconductor manufacturing.

[0018] 4. Determine the radial multi-path scanning plan, evenly distribute multiple scanning paths with the center of the polishing pad as the origin, and set reasonable sampling point spacing and reciprocating scanning mode to ensure that the scanning path covers most of the effective working area of ​​the polishing pad. At the same time, abnormal data is eliminated by data redundancy verification in the overlapping areas of the paths, which improves the comprehensiveness and accuracy of the detection and reduces the possibility of missed detection and false detection.

[0019] 5. During the data acquisition process, a high-precision timestamp synchronization mechanism is adopted to ensure the time synchronization of the two types of data. The near-infrared absorption spectrum data and eddy current induction data are preprocessed separately to improve the data quality. When constructing the signal correction prediction model, an improved encoder-decoder network structure generator and a lightweight convolutional neural network structure discriminator are adopted, as well as a hybrid loss function to optimize the model parameters, so that the model has better performance and generalization ability and can complete the signal correction task quickly and accurately.

[0020] 6. The system for detecting the flatness of CMP polishing pads provided by this invention integrates multiple modules such as scan planning, data acquisition, signal correction, three-dimensional modeling and feature extraction, and result output. It realizes the automation and integration of the detection process. The modules work together to improve detection efficiency, reduce manual intervention, reduce the impact of human factors on the detection results, and make the detection results more objective and accurate. Attached Figure Description

[0021] The accompanying drawings, which are included to provide a further understanding of this application and form part of this application, illustrate exemplary embodiments and are used to explain this application, but do not constitute an undue limitation of this application. In the drawings: Figure 1 This is a flowchart illustrating the overall method of the present invention; Figure 2 Flowchart for constructing the signal correction prediction model of this invention; Figure 3 This is a flowchart of the three-dimensional shape model construction and feature extraction process of the present invention; Figure 4 This is a system block diagram of the present invention. Detailed Implementation

[0022] The present application will now be described in further detail with reference to the accompanying drawings. It should be noted that the following specific embodiments are only used to further illustrate the present application and should not be construed as limiting the scope of protection of the present application. Those skilled in the art can make some non-essential improvements and adjustments to the present application based on the above application content.

[0023] Example 1: Please see Figures 1-3 This invention provides a technical solution: a method for detecting the flatness of a CMP polishing pad, the method comprising: S1. Obtain the information of the area to be inspected on the CMP polishing pad and determine the radial multi-path scanning plan; Among them, CMP polishing pads are polishing pads used in chemical mechanical polishing processes. In precision machining fields such as semiconductor manufacturing, they are used to planarize the surfaces of workpieces such as wafers. The information of the area to be inspected refers to the relevant data of the specific area on the CMP polishing pad that needs to be inspected for flatness, which may include the location and size of the area. The radial multi-path scanning planning scheme is a scheme for designing multiple scanning paths along the radial direction of the CMP polishing pad. The purpose is to comprehensively and systematically obtain information from different positions on the polishing pad in order to accurately detect flatness. S2. Through the multi-band near-infrared-eddy current collaborative correction module, near-infrared absorption spectrum data and eddy current sensing data on multiple radial paths are collected simultaneously. The multi-band near-infrared-eddy current co-correction module is a device or system that integrates multi-band near-infrared detection technology and eddy current detection technology. Multi-band near-infrared detection utilizes the absorption characteristics generated by the interaction of near-infrared light of different wavelengths with matter to obtain information; eddy current detection is based on the principle of electromagnetic induction. When an alternating magnetic field approaches a conductive object, eddy currents are generated inside the object. By detecting changes in eddy currents, relevant information about the object is obtained. This module can simultaneously leverage the advantages of both detection technologies, working synergistically to improve the accuracy and reliability of detection. Near-infrared absorption spectral data shows that matter absorbs near-infrared light of different wavelengths to varying degrees. The data obtained by measuring this absorption degree reflects the molecular structure and composition information of the matter, and can be used to analyze the surface characteristics of polishing pads. Eddy current induction data utilizes eddy current detection technology. When the detection module approaches a CMP polishing pad, eddy currents are generated due to changes in the conductivity or surface characteristics of the polishing pad. By detecting relevant parameters of the eddy currents, such as amplitude and phase, the data obtained can be used to reflect the surface physical state of the polishing pad. S3. Construct a near-infrared absorption difference-eddy current induction deviation prediction model based on generative adversarial network, input synchronously collected data for model training, and automatically correct signal drift caused by polishing pad wear or slurry wetting. Generative Adversarial Networks (GANs) are deep learning model architectures consisting of two neural networks: a generator and a discriminator. The generator is responsible for generating data that is as realistic as possible, while the discriminator is responsible for determining whether the input data is real data or data generated by the generator. Both networks are continuously optimized during adversarial training, ultimately enabling the generator to generate high-quality data, which is used in this method to build the prediction model. The near-infrared absorption difference-eddy current induction deviation prediction model is a model built based on GANs. Its input is synchronously acquired near-infrared absorption spectrum data and eddy current induction data, and its output is the prediction result of signal drift caused by factors such as polishing pad wear or slurry wetting. This model can automatically correct the acquired data and improve data accuracy. Signal drift occurs during the detection process when the acquired near-infrared absorption spectrum data and eddy current induction data deviate from the true values ​​due to factors such as polishing pad wear, changes in surface physical structure, or slurry wetting, changes in surface chemical composition and physical state. S4. Construct a three-dimensional morphology model of the polishing pad surface using the corrected fused data, extract flatness feature parameters, and realize dynamic non-contact detection of polishing pad flatness. Among them, the fused data consists of near-infrared absorption spectral data and eddy current induction data after correction by the above prediction model. These data combine the advantages of the two detection technologies and more accurately reflect the surface characteristics of the polishing pad. The three-dimensional morphology model of the polishing pad surface is a three-dimensional model constructed using the corrected fused data, which can intuitively display the undulations and shape characteristics of the CMP polishing pad surface, and helps to fully understand the surface condition of the polishing pad. The flatness characteristic parameters are indicators extracted from the three-dimensional morphology model of the polishing pad surface that can quantitatively describe the flatness of the polishing pad, such as surface roughness and flatness. These parameters can accurately evaluate the flatness of the polishing pad. Dynamic non-contact detection does not require direct contact with the CMP polishing pad during the entire detection process and can detect the flatness of the polishing pad in real time and continuously, and obtain its flatness changes in a timely manner.

[0024] It should be noted that when using this technology, a radial multi-path scanning plan should be determined to fully cover the area to be inspected, avoid blind spots, and improve the integrity of the inspection. The multi-band near-infrared-eddy current collaborative correction module collects data simultaneously, combining the advantages of the two detection technologies. It utilizes near-infrared absorption spectroscopy to reflect the material composition information and eddy current induction to obtain the surface physical state, providing richer and more accurate information. Based on a generative adversarial network, a prediction model is built, which can automatically correct signal drift and eliminate the influence of interference factors such as polishing pad wear and slurry wetting, thereby improving data reliability. The corrected fused data is used to construct a three-dimensional morphology model and extract flatness feature parameters to achieve dynamic non-contact inspection. The inspection results are not only intuitive and visual, but also allow for real-time continuous monitoring to promptly grasp changes in flatness, providing strong support for the quality control and process optimization of polishing pads.

[0025] In one embodiment, the multi-band near-infrared-eddy current co-correction module is an integrated design module. The core components include a multi-band near-infrared scanner and an embedded eddy current sensor. The two are fixed into an integrated structure by a mechanical bracket to ensure that their relative positions remain unchanged during the acquisition process and to ensure data synchronization. Among them, the working spectrum of the multi-band near-infrared scanner covers the key near-infrared band of 1000-2500nm, and includes at least 5 continuously adjustable spectral channels. The wavelength adjustment accuracy of each channel can reach ±1nm, which can accurately distinguish the differences between surface interference such as slurry residue and moisture adhesion on the surface of the polishing pad and internal characteristics such as fiber structure and pore distribution inside the polishing pad. The embedded eddy current sensor adopts a high-frequency excitation design with an adjustable excitation frequency range of 100kHz-1MHz. It supports dynamic adaptation of the optimal frequency according to the polishing pad material, which includes polyurethane-based and diamond particle-doped materials. Its sampling frequency is not less than 10kHz, which can quickly capture the minute thickness changes and conductivity shifts of the polishing pad during the polishing process, providing high-precision raw data support for flatness detection.

[0026] This integrated design ensures that the core components are fixed in one unit, maintaining the relative positions of the multi-band near-infrared scanner and the embedded eddy current sensor during data acquisition, thus guaranteeing data synchronization. The multi-band near-infrared scanner covers key wavelengths and has multiple continuously adjustable channels, enabling precise differentiation between surface interference and internal characteristic differences. The embedded eddy current sensor features a high-frequency excitation design with adjustable frequency, providing high sampling frequency and rapid capture of minute changes in the polishing pad. The two sensors work together, combining the advantages of both technologies to provide rich, accurate, and high-precision raw data for flatness detection, effectively improving the reliability and accuracy of the detection and meeting the needs for comprehensive detection of polishing pad surface characteristics.

[0027] In one embodiment, determining a radial multipath scan planning scheme includes: With the center of the polishing pad as the origin, N scanning paths are evenly arranged radially, N≥8, and the included angle between two adjacent scanning paths is equal to ensure the uniformity of scanning coverage and avoid missing detection in local areas. The spacing between sampling points on each scanning path is no greater than 0.5 mm. This spacing is set based on the industry accuracy requirements for polishing pad flatness detection, specifically the detection accuracy of semiconductor-grade polishing pads. It can effectively capture minute undulation defects, while the scanning path needs to completely cover more than 95% of the effective working area of ​​the polishing pad. The effective working area of ​​the polishing pad is the area that actually participates in wafer polishing, taking into account both the comprehensiveness and efficiency of the detection. The scanning path is configured with a round-trip scanning mode, where each path first scans from the inside out, then scans back from the outside in. Data redundancy is checked in overlapping areas, and the differences in source data between the two scans are compared. A data difference threshold is set. Sampling points exceeding this threshold are identified as abnormal data and removed, further improving the reliability and accuracy of the collected data.

[0028] This design, with the center of the polishing pad as the origin, evenly distributes scanning paths, sets the sampling point spacing and reciprocating scanning mode, and evenly distributes multiple paths radially with equal adjacent angles. This ensures comprehensive and uniform coverage of the polishing pad, avoiding localized missed detections. The reasonable sampling point spacing effectively captures minute undulations and defects, balancing comprehensiveness and efficiency. The reciprocating scanning mode uses data redundancy verification to compare differences in source data, eliminating abnormal data and further improving the reliability and accuracy of the collected data, providing a solid data foundation for subsequent precise flatness testing.

[0029] In one embodiment, the simultaneous acquisition of near-infrared absorption spectral data and eddy current induction data along radial multipath paths includes: A high-precision timestamp synchronization mechanism is employed to synchronize the data acquisition triggering of the two sensors. The start time of acquisition is controlled by a hardware trigger signal to ensure the timestamp error between the near-infrared absorption spectroscopy data and the eddy current induction data. This ensures that the time consistency requirements during data fusion are met. The acquired near-infrared absorption spectral data were preprocessed. First, baseline correction was performed using a polynomial fitting algorithm to eliminate baseline shifts caused by instrument noise and ambient light interference. Then, normalization was performed to compress the data. The interval is used to facilitate subsequent feature extraction and model calculation. The normalization calculation formula is: in, For a certain wavelength The original absorption value below, This is the baseline value obtained by polynomial fitting at this wavelength; The eddy current induction data is subjected to targeted filtering. Considering the high-frequency noise generated by the slurry flow during polishing, an adaptive Kalman filter algorithm is adopted, and the filter convergence threshold is set to [value missing]. When the fluctuation amplitude of the filtered data is less than the threshold, the size of the filtering window is fixed; when the fluctuation amplitude is greater than the threshold, the window size is dynamically adjusted by 1.2 times the sampling frequency. While retaining the effective signal, high-frequency interference is suppressed to the maximum extent. The effective signal is the change in the thickness and conductivity of the polishing pad, which ensures the stability of the eddy current data.

[0030] This design employs a high-precision timestamp synchronization mechanism and preprocesses the acquired data separately. Hardware trigger signals ensure synchronized acquisition of data from both sensors, resulting in minimal timestamp errors and meeting the consistency requirements for data fusion. Preprocessing of near-infrared absorption spectroscopy data eliminates baseline offset and normalizes it for easier subsequent processing. Targeted filtering of eddy current sensing data adaptively adjusts the window size, preserving effective signals while suppressing high-frequency interference, ensuring data stability and guaranteeing the quality of the acquired data. This provides accurate and reliable data for subsequent signal correction and model calculations.

[0031] In one embodiment, a near-infrared absorption difference-eddy current induction deviation prediction model is constructed based on a generative adversarial network, including four core steps: model structure design, training data preparation, training process execution, and model optimization, as detailed below: (1) Model structure design: The generator uses an improved U-Net network structure; The input channel has 130 channels, consisting of a 128-dimensional near-infrared absorption difference feature vector and 2-dimensional eddy current raw data. The 2-dimensional eddy current raw data contains amplitude and phase information. The output channel has 2 channels, corresponding to the corrected eddy current amplitude and phase data. The encoding path contains four convolutional blocks. Each convolutional block consists of two 3×3 convolutional layers, a batch normalization layer, and a ReLU activation function. The convolutional stride is 2, which enables feature downsampling. The decoding path corresponds to 4 deconvolution blocks. Each deconvolution block consists of a 2×2 deconvolution layer, a batch normalization layer, and a ReLU activation function. The deconvolution stride is 2, which realizes feature upsampling. Skip connections are set between corresponding levels of the encoding and decoding paths to preserve low-level detailed features; attention mechanism modules are inserted in the last two layers of the decoding path to enhance the capture of signal drift-sensitive features through channel attention and spatial attention weighting. The attention mechanism module is CBAM. The discriminator employs a lightweight convolutional neural network structure. The input is a 64×64 two-dimensional data matrix, which consists of local feature maps of the corrected eddy current data or standard data. The network structure is as follows: 3×3 convolutional layer, BN layer, LeakyReLU activation function, 3×3 convolutional layer, BN layer, LeakyReLU activation function, 3×3 convolutional layer, BN layer, LeakyReLU activation function, global average pooling layer, 2 fully connected layers, and Sigmoid activation function. The 3×3 convolutional layer outputs 64 channels with a stride of 2, the second 3×3 convolutional layer outputs 128 channels with a stride of 2, and the third 3×3 convolutional layer outputs 256 channels with a stride of 2. The LeakyReLU activation function has a negative slope of 0.2. The two fully connected layers output dimensions of 128 and 1, respectively. The Sigmoid activation function outputs the probability of data authenticity, which is in the range of 0-1. (2) Training data preparation: Training sample sets were collected under different working conditions, covering 3 mainstream polishing pad materials, 5 wear levels, and 4 slurry wetting states. The 3 mainstream polishing pad materials include polyurethane-based, diamond-doped, and ceramic-reinforced. The 5 wear levels include brand new, 10% wear, 30% wear, 50% wear, and 70% wear. The 4 slurry wetting states include no wetting, light wetting, moderate wetting, and heavy wetting, for a total of 3×5×4=60 working condition combinations. 100 valid samples were collected under each working condition, for a total of 6000 samples. Each sample set includes: raw multi-band near-infrared absorption data, raw eddy current induction data, and standard calibration data. The raw multi-band near-infrared absorption data contains 5 spectral bands with 1024 data points per band. The raw eddy current induction data contains amplitude and phase data, with 1024 data points each. The standard calibration data is obtained through laser interferometry measurements and has high accuracy. ; The sample data were preprocessed as follows: near-infrared data underwent baseline correction and normalization using the methods described above, and the absorption difference between different spectral bands was calculated and reduced to a 128-dimensional feature vector using PCA; eddy current data were filtered by Kalman to extract the time series features of amplitude and phase; standard calibration data were used as labels for model training supervision. The sample set was divided into training set, validation set and test set in a ratio of 7:2:1. The training set contained 4200 samples, the validation set contained 1200 samples and the test set contained 600 samples. Stratified sampling was used in the partitioning process to ensure that the distribution ratio of each working condition in the three sets of data was consistent. (3) Training process execution: Initialization parameters: The weights of both the generator and discriminator are initialized using a He normal distribution, and the bias term is initialized to 0; the optimizer is set to the AdamW optimizer, and the learning rates of the generator and discriminator are respectively... The weight decay coefficient is ; Set the training hyperparameters as follows: batch size is 32, training epochs are 200, and the patience of the early stopping strategy is 20, that is, if the validation set loss does not decrease for 20 consecutive epochs, training will stop. The training process adopts an alternating training mode, and the specific steps are as follows: ① Fix the generator and train the discriminator: Randomly select a batch of samples from the training set, input the near-infrared absorption difference feature vector and the original eddy current data into the generator to obtain the corrected generated data; input the generated data and the corresponding standard data into the discriminator respectively, and calculate the adversarial loss of the discriminator. The discriminator parameters are updated through backpropagation; ② Fix the discriminator and train the generator: Keep the discriminator parameters unchanged, input the same batch of samples into the generator to obtain generated data, input the generated data into the discriminator to obtain the discrimination result, and calculate the generator's mixture loss. The hybrid loss includes mean squared error loss and adversarial loss, and the generator parameters are updated through backpropagation; ③ After every 5 training rounds, calculate the mean absolute error and root mean square error of the model using the validation set, and record the optimal model parameters; The formula for calculating the hybrid loss function is: in, , For standard data, To generate the data, N is the number of sample points. , For the number of samples in the batch, The probability of authenticity output by the discriminator. The weighting coefficient has a range of values. The determination rule is as follows: based on the 5-fold cross-validation results of the training set, with the goal of minimizing the mean absolute error of the corrected data, a grid search method is used to traverse five candidate values ​​(0.6, 0.65, 0.7, 0.75, and 0.8) and select the value that minimizes the mean absolute error. The value is used as the optimal weight coefficient; (4) Model optimization: After training, the model's performance is evaluated using a test set. Evaluation metrics include mean absolute error, root mean square error, and calibration accuracy, where calibration accuracy is the deviation between the calibrated data and the standard data. The sample proportion, requiring the mean absolute error of the test set. Root mean square error Correction accuracy ; If performance requirements are not met, the following optimization strategies will be adopted: ① If the mean absolute error is too large, adjust Value increases The weights; ②If the calibration accuracy is insufficient, increase the proportion of samples with low wear and high wetting conditions in the training samples and retrain; ③ If the model is overfitting, and the loss on the training set is much smaller than the loss on the validation set, it indicates that the model is overfitting. Add a Dropout layer to the encoding layer of the generator with a dropout rate of 0.2, and retrain. The final optimized model is saved as a lightweight inference model. The model size is compressed to less than 50MB through model quantization. The model quantization uses INT8 quantization to meet the deployment requirements of embedded systems.

[0032] This design constructs the model from multiple aspects, including model structure design and training data preparation. A reasonable model structure can better learn data features. The generator adopts an improved U-Net structure, and the discriminator adopts a lightweight convolutional neural network structure. Each part is designed with specific goals in mind. The rich training data covers a variety of working conditions. After preprocessing and partitioning, it can improve the model's generalization ability. Alternating training modes and reasonable hyperparameter settings make model training more efficient. The model optimization strategy is adjusted for different problems, and finally a lightweight inference model is obtained, which meets the deployment requirements of embedded systems and can accurately predict and correct signal drift.

[0033] In one embodiment, automatic correction of signal drift caused by polishing pad wear or slurry wetting includes: Feature extraction was performed on the preprocessed near-infrared absorption spectrum data. Principal component analysis was used to screen out the characteristic wavelength ranges that are sensitive to surface interference, including slurry wetting and wear debris adhesion. The principal component contribution rate threshold was set to 85%. The wavelength ranges corresponding to the top 3 principal components whose cumulative contribution rate reached the threshold were screened out. The absorption difference between different spectral bands in the range was calculated, and a 128-dimensional surface interference feature vector was constructed to comprehensively characterize the surface state of the polishing pad. The surface disturbance feature vector and the original eddy current induction data are input into the trained generator model. The model, through the learned nonlinear mapping relationship, outputs the gain correction coefficient for the current operating condition in real time. With offset correction factor The output frequency is consistent with the sensor sampling frequency. ; Based on the two correction coefficients mentioned above, offset compensation is performed on the original eddy current data. The compensation formula is as follows: in, This is the original eddy current data after filtering. For the corrected eddy current data, a threshold for determining the effectiveness of the correction is set as follows: When the deviation between the corrected data and the standard data is less than the threshold, the correction is considered effective; if the deviation is greater than the threshold, the online fine-tuning mechanism of the model is triggered: 50 similar samples are extracted from the historical data of the current working condition, and the parameters of the last two fully connected layers of the generator are fine-tuned using the Mini-Batch Gradient Descent method. The Mini-Batch Gradient Descent method is Mini-BatchSGD with a learning rate of 1e-6 and 10 rounds of fine-tuning iterations, and then the model is re-optimized. and The coefficient is adjusted until the correction deviation meets the threshold requirement; This compensation process effectively eliminates signal drift caused by changes in sensor distance due to polishing pad wear and changes in dielectric constant due to slurry wetting.

[0034] This design extracts features from near-infrared data to construct vectors, inputs them into the model to obtain correction coefficients and perform compensation, and uses principal component analysis to screen characteristic wavelength ranges. The constructed surface interference feature vectors can comprehensively characterize the surface state of the polishing pad. The model outputs correction coefficients in real time, which can quickly compensate for the offset of the original eddy current data. A threshold for judging the effectiveness of correction is set. When the deviation is large, the online fine-tuning mechanism of the model is triggered to re-optimize the coefficients and ensure the correction effect. This process effectively eliminates signal drift caused by wear and slurry wetting, improves the accuracy of detection data, and provides a guarantee for the subsequent construction of an accurate three-dimensional morphology model.

[0035] In one embodiment, a three-dimensional topographic model of the polishing pad surface is constructed, and flatness feature parameters are extracted, including: Based on the corrected near-infrared-eddy current fusion data, a three-dimensional point cloud model of the polishing pad surface is constructed using the Kriging interpolation algorithm. This algorithm analyzes the spatial correlation of known sampling points and fits the spatial correlation using a spherical variogram model. The variogram parameters are: sill value... Variable range Gold value Set the interpolation error threshold to ; When the theoretical error of the interpolation result exceeds the threshold, the sampling points in the corresponding region are automatically encrypted and re-interpolated. The spacing between the encrypted sampling points is adjusted accordingly. To ensure interpolation accuracy It can accurately reproduce the microscopic undulations on the surface of the polishing pad; Key flatness feature parameters are extracted from the 3D point cloud model, including: Maximum peak-to-valley difference, flatness error, local undulation standard deviation, and radial thickness variation rate. The maximum peak-to-valley difference is the height difference between the highest and lowest points in the model. Flatness error characterizes the degree of deviation between the polishing pad surface and the ideal plane. Local undulation standard deviation reflects the dispersion of surface micro-unevenness. Radial thickness variation rate describes the thickness uniformity along the radial direction. The flatness error is calculated using a least-squares plane fitting algorithm. First, the ideal plane equation is obtained by fitting all point cloud data. in, The fitting coefficients are used to calculate the absolute deviation between each point cloud height value and the ideal plane height value. The maximum value is taken as the flatness error, and the formula is: in, In point cloud models The actual height value at the coordinates, This is the height value of the coordinate on the least squares fitting plane; The rules for determining the acceptable thresholds for each flatness characteristic parameter are as follows: Based on the application scenarios and process requirements of polishing pads, including 12-inch wafer polishing and 8-inch wafer polishing, and referring to the SEMI international standard (SEMI F130-0301), and combining actual production data on the lifespan of polishing pads, a statistical process control (SPC) method was adopted. The statistical distribution of parameters of qualified polishing pads over the past three months was used as the basis, and the upper limit of the 95% confidence interval was taken as the qualified threshold, specifically: the maximum peak-to-valley difference. Flatness difference Local fluctuation standard deviation Radial thickness variation rate .

[0036] This design employs the Kriging interpolation algorithm to construct the model and extract key parameters. Through spatial correlation analysis of known sampling points, the Kriging interpolation algorithm accurately constructs a 3D point cloud model, sets an interpolation error threshold, and automatically densifies sampling points to ensure interpolation accuracy and accurately reproduce the microscopic undulations of the surface. The key flatness feature parameters extracted from the model comprehensively reflect the flatness of the polishing pad. The rules for determining the pass / fail threshold are scientifically sound and reasonable, combining application scenarios, international standards, and actual production data. This provides a clear standard for accurately assessing the flatness of the polishing pad, helping to promptly identify non-conforming products and ensure product quality.

[0037] Example 2: Please see Figure 4A system for detecting the flatness of CMP polishing pads, applied to the method described above for detecting the flatness of CMP polishing pads, the system comprising: The module includes a scanning planning module, a data acquisition module, a signal correction module, a 3D modeling and feature extraction module, and a result output module. The scanning planning module is used to acquire information about the area to be inspected on the polishing pad and generate a radial multipath scanning planning scheme; The data acquisition module is a multi-band near-infrared-eddy current collaborative detection module, used to simultaneously acquire near-infrared absorption spectrum data and eddy current sensing data; The signal correction module has a built-in signal correction prediction model based on generative adversarial networks, which is used to correct signal drift. The 3D modeling and feature extraction module is used to construct a 3D morphology model of the polishing pad surface and extract flatness feature parameters; The results output module is used to output the flatness test results and characteristic parameter analysis report.

[0038] This design incorporates multiple modules, each with a clearly defined function. The scanning planning module generates a reasonable scanning plan based on the information of the area to be inspected on the polishing pad; the data acquisition module collects accurate data synchronously; the signal correction module effectively corrects signal drift using a predictive model; the 3D modeling and feature extraction module constructs an accurate model and extracts key parameters; and the result output module promptly outputs the inspection results and analysis reports. These modules work collaboratively to form a complete and efficient inspection system, enabling automated and precise inspection of the flatness of CMP polishing pads, improving inspection efficiency and accuracy, and providing strong support for the production process.

[0039] Those skilled in the art will understand that all or part of the steps in the methods of the above embodiments can be implemented by a program instructing related hardware. Therefore, this application can take the form of a completely hardware embodiment, a completely software embodiment, or an embodiment combining software and hardware aspects. Moreover, this application can take the form of a computer program product implemented on one or more computer-usable storage media (including but not limited to disk storage, CD-ROM, optical storage, etc.) containing computer-usable program code.

[0040] The above embodiments provide a detailed description of the present invention. Specific examples have been used to illustrate the principles and implementation methods of the present invention. The descriptions of the above embodiments are only for the purpose of helping to understand the method and core ideas of the present invention. At the same time, for those skilled in the art, there will be changes in the specific implementation methods and application scope based on the ideas of the present invention. Therefore, the content of this specification should not be construed as a limitation of the present invention.

Claims

1. A method for detecting the flatness of a CMP polishing pad, characterized in that, The method includes the following steps: S1. Obtain the information of the area to be inspected on the polishing pad and determine the radial multi-path scanning plan; S2. Through the multi-band near-infrared-eddy current collaborative detection module, near-infrared absorption spectrum data and eddy current sensing data on the radial multi-path are collected simultaneously; S3. Construct a signal correction prediction model based on generative adversarial network. Input the near-infrared absorption spectrum data and eddy current induction data on radial multipaths synchronously collected by the multi-band near-infrared-eddy current collaborative detection module to train the model and automatically correct the signal drift caused by polishing pad wear or medium wetting. S4. Construct a three-dimensional morphological model of the polishing pad surface using the corrected fused data, extract flatness feature parameters, and realize dynamic non-contact detection of polishing pad flatness.

2. The method for detecting the flatness of a CMP polishing pad according to claim 1, characterized in that, The multi-band near-infrared-eddy current collaborative detection module is an integrated design, including: The multi-band near-infrared scanning unit and the eddy current sensing unit maintain their relative positions during the data acquisition process through a fixed structure. The multi-band near-infrared scanning unit operates in spectral bands covering key near-infrared bands, and includes at least five continuously adjustable spectral band channels, with each channel having a wavelength adjustment accuracy of not less than [specified value]. ; The eddy current sensing unit adopts a high-frequency excitation design, with an adjustable excitation frequency and a sampling frequency not lower than [missing information]. .

3. The method for detecting the flatness of a CMP polishing pad according to claim 1, characterized in that, The method for determining the radial multipath scan planning scheme includes: With the center of the polishing pad as the origin, the polishing pads are evenly distributed radially. Scan path, The angle between adjacent scan paths is equal; The spacing between sampling points on each scanning path is no greater than [missing value]. The scanning path covers the effective working area of ​​the polishing pad. above; Set up a round-trip scanning mode to eliminate abnormal data by checking for data redundancy in overlapping areas of the path.

4. The method for detecting the flatness of a CMP polishing pad according to claim 1, characterized in that, The synchronous acquisition of near-infrared absorption spectral data and eddy current induction data along radial multipath paths includes: A high-precision timestamp synchronization mechanism is adopted to ensure the timestamp error between the two types of data. ; Baseline correction and normalization were performed on the near-infrared absorption spectroscopy data, and the data was compressed to [size missing]. interval; An adaptive filtering algorithm is used to suppress noise in the eddy current induction data while retaining the effective signal.

5. The method for detecting the flatness of a CMP polishing pad according to claim 4, characterized in that, The normalization formula for the near-infrared absorption spectral data is as follows: in, For a certain wavelength The original absorption value below, This is the baseline value obtained by polynomial fitting at this wavelength.

6. The method for detecting the flatness of a CMP polishing pad according to claim 1, characterized in that, The signal correction prediction model based on generative adversarial networks includes: The model consists of four steps: model structure design, training data preparation, training process execution, and model optimization. The generator in the signal correction prediction model built on generative adversarial networks adopts an improved encoder-decoder network structure, which includes skip connections and attention mechanism modules; The discriminator in the signal correction prediction model based on generative adversarial networks adopts a lightweight convolutional neural network structure. The training data covers a variety of working conditions, including polishing pad materials, wear levels, and media wetting states. After preprocessing, the data is divided into training, validation, and test sets. The training process adopts an alternating training mode, and the model parameters are optimized by a hybrid loss function. After optimization, the model meets the preset performance indicators.

7. The method for detecting the flatness of a CMP polishing pad according to claim 6, characterized in that, The formula for calculating the hybrid loss function is as follows: in, For mean square error loss, To help the generator combat loss, The weighting coefficient has a range of values. The optimal value is determined through cross-validation.

8. The method for detecting the flatness of a CMP polishing pad according to claim 1, characterized in that, The automatic correction of signal drift caused by polishing pad wear or media wetting includes: Feature extraction is performed on the preprocessed near-infrared absorption spectral data to construct a surface interference feature vector; The feature vector and the original eddy current induction data are input into the trained model, and the gain correction coefficient is output. With offset correction factor ; Based on the aforementioned correction coefficient, offset compensation is performed on the original eddy current data. The compensation formula is as follows: in, This is the original eddy current data after filtering. The data is the corrected eddy current data; when the correction deviation does not meet the threshold requirement, the online fine-tuning mechanism of the model is triggered.

9. The method for detecting the flatness of a CMP polishing pad according to claim 1, characterized in that, The process of constructing a three-dimensional morphological model of the polishing pad surface and extracting flatness feature parameters includes: Based on the corrected fused data, a 3D point cloud model is constructed using a spatial interpolation algorithm. An interpolation error threshold is set, and when the theoretical error of the interpolation result exceeds the threshold, the sampling points are automatically encrypted. Feature parameters are extracted from the 3D point cloud model, including: maximum peak-to-valley difference, flatness error, local undulation standard deviation, and radial thickness variation rate. The qualified thresholds for each parameter are determined based on the application scenario, industry standards, and statistical process control methods.

10. A system for detecting the flatness of CMP polishing pads, characterized in that, The system is used to perform the method for detecting the flatness of a CMP polishing pad according to any one of claims 1-9, the system comprising: The module includes a scanning planning module, a data acquisition module, a signal correction module, a 3D modeling and feature extraction module, and a result output module. The scanning planning module is used to acquire information about the area to be inspected on the polishing pad and generate a radial multipath scanning planning scheme; The data acquisition module is a multi-band near-infrared-eddy current collaborative detection module, used to simultaneously acquire near-infrared absorption spectrum data and eddy current sensing data; The signal correction module has a built-in signal correction prediction model based on generative adversarial networks, which is used to correct signal drift. The 3D modeling and feature extraction module is used to construct a 3D morphology model of the polishing pad surface and extract flatness feature parameters; The results output module is used to output the flatness test results and characteristic parameter analysis report.

Images