A semiconductor adhesive tape surface cleanliness detection method and system based on Raman spectrum

Abstract

The application belongs to the technical field of cleanliness detection, and particularly relates to a semiconductor adhesive tape surface cleanliness detection method and system based on Raman spectrum, which comprises the following steps: S1, obtaining a multi-spectrum image of the semiconductor adhesive tape surface; for each pixel point in the multi-spectrum image, calculating the weighted sum of the Mahalanobis distance of the pixel point spectral dimension and the Gabor texture response of the spatial dimension to obtain a spatial spectrum-texture anomaly index; determining the excitation points of the Raman spectrum based on the spatial spectrum-texture anomaly index, and collecting the Raman spectrum data of each excitation point; S2, extracting the multi-spectrum feature vector and the Raman spectrum feature vector of the excitation points. The application generates a full field of view cleanliness distribution map through interpolation based on high-quality sparse measurement results with confidence weight, realizes comprehensive evaluation of the adhesive tape surface pollution condition, and solves the contradiction between the detection efficiency, recognition accuracy and representation comprehensiveness of the traditional method.

Description

Technical Field

[0001] This invention belongs to the field of cleanliness testing technology, specifically relating to a method and system for testing the surface cleanliness of semiconductor tape based on Raman spectroscopy. Background Technology

[0002] In semiconductor manufacturing processes, semiconductor tapes serve as temporary support and protective materials for critical steps such as wafer dicing, thinning, and packaging. Microparticles, organic residues, and metal ions on the tape surface can transfer to the wafer surface during processing, causing fatal defects such as short circuits, leakage, and wire bonding failures. Traditional detection methods, such as manual visual inspection or ordinary optical microscopy, suffer from high subjectivity, low efficiency, and difficulty in detecting nanoscale or transparent thin-film contaminants. While machine vision-based optical inspection technologies primarily rely on optical contrast information such as the morphology, size, or color of contaminants, they lack the ability to analyze the chemical composition of contaminants, struggle to distinguish different types of contaminants, and cannot identify chemical residues without obvious morphological characteristics, easily leading to missed detections and misjudgments.

[0003] Molecular spectroscopy techniques such as Raman spectroscopy and Fourier transform infrared spectroscopy can provide the "fingerprint" spectrum of substances, accurately identifying the molecular structure of pollutants. However, these spectroscopic techniques are essentially single-point detection techniques. To achieve comprehensive detection of the entire tape surface, time-consuming point-by-point scanning or imaging is required. For large-sized tapes, the detection efficiency is low, failing to meet the high-throughput, fast-paced production demands of the semiconductor industry. Currently, there are attempts to combine imaging and spectroscopy techniques, but how to quickly locate suspicious areas requiring precise spectral measurement to avoid the inefficiency of global scanning, and how to perform weighted fusion based on the quality and confidence level of different modal data to accurately quantify the type and concentration of pollutants and generate an intuitive full-field cleanliness distribution map, are pressing challenges that need to be addressed. Summary of the Invention

[0004] This invention provides a method and system for detecting the surface cleanliness of semiconductor tape based on Raman spectroscopy, in order to solve the technical problems of low detection efficiency and inability to quantify contaminants and their concentration in existing technologies.

[0005] In a first aspect, the present invention provides a method for detecting the surface cleanliness of semiconductor tape based on Raman spectroscopy, comprising the following steps:

[0006] S1. Acquire a multi-spectral image of the semiconductor tape surface; for each pixel in the multi-spectral image, calculate the weighted sum of the Mahalanobis distance in the spectral dimension and the Gabor texture response in the spatial dimension to obtain the spatial spectrum-texture anomaly index; determine the excitation point of the Raman spectrum based on the spatial spectrum-texture anomaly index, and collect the Raman spectrum data of each excitation point.

[0007] S2, extract the multi-spectral feature vector and Raman spectral feature vector of the excitation point; calculate the ratio of the main characteristic peak intensity to the baseline noise standard deviation of the Raman spectral data to obtain the Raman signal quality factor; calculate the fusion confidence coefficient based on the spatial spectral-texture anomaly index of the excitation point and the Raman signal quality factor; use the fusion confidence coefficient to weight the Raman spectral feature vector and concatenate it with the multi-spectral feature vector to generate the fusion feature vector;

[0008] S3 calculates the spectral distance between the fused feature vector and each standard feature vector in the preset pollutant standard feature library, and uses the pollutant type corresponding to the smallest distance as the identification result; based on the identification result, it calls the corresponding concentration regression model to calculate the pollutant concentration at the excitation point;

[0009] S4. For other pixels that are not identified as excitation points, radial basis function interpolation is performed based on the contaminant concentration and spatial location of all excitation points, and the fusion confidence coefficient of each excitation point is used as the influence weight to calculate the contaminant concentration of other pixels, thereby generating a surface cleanliness distribution map of the entire field of view.

[0010] Furthermore, for each pixel in the multispectral image, a weighted sum of the Mahalanobis distance in the spectral dimension and the Gabor texture response in the spatial dimension is calculated to obtain the spatial spectral-texture anomaly index, including:

[0011] The spectral mean vector of all pixels in the entire multispectral image Covariance Matrix For background statistics, calculate pixel points. Spectral Mahalanobis distance The multispectral image is filtered using a Gabor filter bank with multiple scales and orientations, and the maximum value of each pixel in all Gabor filter responses is taken as the Gabor texture response. ;

[0012] Setting spectral weights and texture weight ,and Calculate the spatial spectral-texture anomaly index .

[0013] Furthermore, the ratio of the main characteristic peak intensity to the baseline noise standard deviation of the Raman spectral data is calculated to obtain the Raman signal quality factor, including:

[0014] Baseline correction was performed on the original Raman spectral data using an iterative reweighted penalized least squares method. In the corrected Raman spectra, the peak with the highest intensity was identified as the principal characteristic peak, and its peak intensity was recorded. In the corrected Raman spectrum, regions without characteristic peaks are selected, and the standard deviation of the signal intensity within these regions is calculated as the baseline noise standard deviation. ;

[0015] Calculate the Raman signal quality factor .

[0016] Furthermore, based on the spatial spectral-texture anomaly index of the excitation point and the Raman signal quality factor, the fusion confidence coefficient is calculated, including:

[0017] Spatial spectral-texture anomaly index AI and Raman signal quality factor were respectively evaluated. Normalization is performed using the logistic function. Mapping to the [0, 1] interval yields the normalization exponent. and normalization factor The fusion confidence coefficient is obtained by multiplying the normalization exponent and the normalization factor. ;

[0018] Where S(v) represents the normalized value obtained after mapping by the logistic function, v represents the original input variable, k represents the shape parameter of the logistic function, and v0 represents the center point parameter of the logistic function.

[0019] Furthermore, the Raman spectral feature vector is weighted using the fusion confidence coefficient and concatenated with the multi-spectral feature vector to generate a fusion feature vector, including:

[0020] The multispectral data of the excitation point is used as the multispectral feature vector. The baseline-corrected Raman spectral data is downsampled or interpolated to a preset dimension and used as the Raman spectral feature vector. ; Raman spectral eigenvectors Multiplying each dimension by the fusion confidence coefficient C yields the weighted Raman feature vector. ; to multi-spectral feature vector with weighted Raman eigenvectors The features are concatenated and spliced ​​in sequence to generate a fused feature vector. .

[0021] Furthermore, by calculating the spectral distance between the fused feature vector and each standard feature vector in the preset pollutant standard feature library, the pollutant type corresponding to the minimum distance is used as the identification result, including:

[0022] Cosine distance is used as the metric for spectral distance; for the fused feature vector to be measured... and each standard feature vector in the pollutant standard feature library Calculate the cosine distance ,in, Represents the cosine distance. express The length of the mold, express The modulus is calculated; all calculated cosine distances are compared, and the pollutant type corresponding to the standard feature vector with the smallest cosine distance is determined as the identification result of the excitation point.

[0023] Furthermore, based on the identification results, the corresponding concentration regression model is invoked to calculate the pollutant concentration at the trigger point, including:

[0024] For each pollutant type in the pollutant standard feature library, a partial least squares regression model is established using the fusion feature vectors of a series of pollutant samples with known concentrations and their corresponding concentration values. Based on the identified pollutant type, the corresponding partial least squares regression model is selected and loaded. The fusion feature vector of the excitation point to be tested is input into the partial least squares regression model, and the output predicted value is the pollutant concentration at the excitation point.

[0025] Furthermore, radial basis function interpolation is performed to calculate the contaminant concentration of other pixels, including:

[0026] Pollutant concentrations at all N excitation points Spatial location and fusion confidence coefficient As an interpolation condition; a Gaussian function is used. As a radial basis function; for any pixel to be calculated within the field of view. Contaminant concentration of pixels The pollutant concentrations at all trigger points are calculated using a weighted average. The calculation formula is as follows: .

[0027] Furthermore, in S1, pixels with a spatial spectral-texture anomaly index greater than a first preset threshold are identified as excitation points for Raman spectroscopy.

[0028] Secondly, the present invention provides a semiconductor tape surface cleanliness detection system based on Raman spectroscopy, including a memory and a processor. The memory stores computer program instructions, and when the computer program instructions are executed by the processor, the above-mentioned semiconductor tape surface cleanliness detection method based on Raman spectroscopy is implemented.

[0029] The beneficial effects are as follows: This invention rapidly screens through multi-spectral methods and calculates anomaly indices by combining spectral and texture information, thus identifying suspicious contamination points requiring precise measurement. When identifying contaminants, this invention constructs a fusion confidence coefficient, which simultaneously evaluates the reliability of the initial Raman spectral signal used for anomaly detection. By weighted fusion of multi-source features, the weight of key information is enhanced, improving the accuracy and stability of contaminant type identification and concentration quantification. Based on high-quality sparse measurement results with confidence weights, a full-field cleanliness distribution map is generated through interpolation, achieving a comprehensive assessment of the contamination status of the tape surface and resolving the contradiction between detection efficiency, identification accuracy, and comprehensive representation in traditional methods. Attached Figure Description

[0030] Figure 1 This is a flowchart of a method for detecting the surface cleanliness of semiconductor tape based on Raman spectroscopy. Detailed Implementation

[0031] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some, not all, of the embodiments of the present invention. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0032] An embodiment of the method for detecting the surface cleanliness of semiconductor tape based on Raman spectroscopy provided by this invention:

[0033] like Figure 1 As shown, the method for detecting the surface cleanliness of semiconductor tape based on Raman spectroscopy includes the following steps:

[0034] S1. Acquire a multi-spectral image of the semiconductor tape surface; for each pixel in the multi-spectral image, calculate the weighted sum of the Mahalanobis distance in the spectral dimension and the Gabor texture response in the spatial dimension to obtain the spatial spectrum-texture anomaly index; determine the pixels with the spatial spectrum-texture anomaly index greater than a first preset threshold as excitation points for Raman spectroscopy, and collect Raman spectral data for each excitation point.

[0035] Specifically, an industrial camera equipped with a tunable filter or filter wheel is used to scan and image the surface of a semiconductor tape in the visible to near-infrared band, such as 400nm to 1000nm, acquiring a series of two-dimensional images in different bands, and stacking the two-dimensional images into a three-dimensional data cube, where two dimensions are spatial coordinates x and y, and the third dimension is the spectral wavelength.

[0036] A clean region in the image is selected as the background reference. The mean vector and covariance matrix of the spectral vectors of all pixels within this region are calculated. For each pixel to be tested, the Mahalanobis distance is calculated using the pixel's spectral vector, the mean vector of the background, and the inverse of the covariance matrix. The Mahalanobis distance represents the degree of spectral anomaly. Simultaneously, the multispectral image is converted to a grayscale image, and a set of Gabor filters with different directions and frequencies are used to filter the grayscale image. The maximum or average value of each filter response is taken as the Gabor texture response for that pixel, representing spatial texture anomalies. The normalized Mahalanobis distance and the Gabor texture response value are then weighted and summed according to preset weights, such as 0.6 and 0.4, to obtain the spatial spectral-texture anomaly index.

[0037] The first preset threshold is obtained through statistical analysis of the abnormal indices of a large number of known clean and contaminated samples. For example, it can be set as the mean of the abnormal indices in the clean area plus three times the standard deviation. The coordinates of all pixels in the image whose abnormal indices exceed this threshold are recorded. A high-precision displacement platform is controlled to move the semiconductor tape sequentially to the laser focus of the confocal Raman spectrometer. For example, a 532nm laser is used to excite each coordinate point, and the Raman spectral data of each point is collected and recorded.

[0038] In an optional embodiment, for each pixel in the multispectral image, a weighted sum of the Mahalanobis distance in the spectral dimension and the Gabor texture response in the spatial dimension is calculated to obtain the spatial spectral-texture anomaly index, including:

[0039] The spectral mean vector of all pixels in the entire multispectral image Covariance Matrix For background statistics, calculate pixel points. Spectral Mahalanobis distance The multispectral image is filtered using a Gabor filter bank with multiple scales and orientations, and the maximum value of each pixel in all Gabor filter responses is taken as the Gabor texture response. ;

[0040] Setting spectral weights and texture weight ,and Calculate the spatial spectral-texture anomaly index .

[0041] Specifically, for example, a 100×100 pixel multispectral image containing 16 spectral bands is used to calculate a 16-dimensional mean vector μ and a 16×16 covariance matrix Σ using the spectral data of all 10,000 pixels. For a pixel x to be analyzed in the image, the pixel's spectral vector is [0.8, 0.7, ..., 0.9]. The difference between the vector and the background is calculated using the Mahalanobis distance formula, resulting in a value such as 15.3. This value represents the degree to which this pixel deviates from the overall background spectrally.

[0042] A filter bank consisting of 40 Gabor filters across 5 scales and 8 directions is used to filter the first principal component image of the multispectral image. For the same pixel x, 40 response values ​​are obtained after processing by the 40 filters. The largest value among these response values, for example, 0.95, is selected as the Gabor texture response for that point. Set spectral weights The texture weight is 0.6. If the value is 0.4, then the spatial spectral-texture anomaly index AI(x) for that pixel is 9.56. Repeating this process for each pixel in the image generates a complete anomaly index map.

[0043] To further improve detection accuracy and reduce pollutant omissions, a sliding window is used to dynamically update local background statistics. Combined with the Welford algorithm, the covariance matrix is ​​incrementally updated. Outliers exceeding three standard deviations within the window are removed before calculating the Mahalanobis distance, preventing global background averaging from masking weak anomalous signals from low-concentration pollutants. Surface-enhanced Raman scattering (SERS) technology is introduced, using a ReS2 / ZnO / Au ternary composite substrate to construct multiple charge transfer channels. The Raman signal is enhanced to a higher level through the synergistic effect of electromagnetic and chemical enhancement. This doubles the detection limit for trace organic pollutants, reducing it to [a higher level]. M, combined with ultraviolet resonance Raman technology, effectively separates the Raman signal from the fluorescence background using filters, improving the identification of characteristic peaks of weak-signal pollutants; optimizes the fusion mechanism of spectral and texture features, adopts an adaptive weight allocation strategy, and increases the texture weight to 0.6 when the Mahalanobis distance is in the warning range, enhancing the texture feature contribution of transparent film pollutants; at the same time, a low-frequency filter channel is added to the Gabor filter bank to capture the weak texture differences of low-concentration pollutants; implements a dual-threshold dynamic calibration mechanism, based on the statistical distribution of calibration samples collected before detection, using the mean index of clean samples + 2 standard deviations as the warning threshold and the mean index of low-concentration pollutants - 1 standard deviations as the excitation threshold, and links low-power Raman spectra to pre-calibrate candidate points in the warning range, combined with connected region area screening to eliminate isolated noise points, comprehensively reducing the risk of missed detection, and ensuring that the detection system can efficiently identify pollutants of different concentrations and types.

[0044] S2. For each excitation point of the Raman spectrum, extract the multi-spectral feature vector and Raman spectral feature vector of the excitation point respectively; calculate the ratio of the main characteristic peak intensity to the baseline noise standard deviation of the Raman spectral data to obtain the Raman signal quality factor; calculate the fusion confidence coefficient based on the spatial spectral-texture anomaly index of the excitation point and the Raman signal quality factor; use the fusion confidence coefficient to weight the Raman spectral feature vector and concatenate it with the multi-spectral feature vector to generate the fusion feature vector.

[0045] Specifically, for a given excitation point in a Raman spectrum, the multispectral feature vector of that excitation point is the spectral intensity sequence of the corresponding pixel in the original multispectral data cube. The Raman spectral feature vector of the spectral intensity sequence is obtained by preprocessing the acquired raw Raman spectral data. This preprocessing includes baseline correction using asymmetric least squares and smoothing / denoising using a Savitzky-Gore filter. The resulting spectral intensity sequence is the Raman spectral feature vector.

[0046] In the baseline-corrected Raman spectrum, the highest intensity characteristic peak is identified and its intensity is recorded. A flat region without obvious characteristic peaks is selected in the Raman spectrum, and the standard deviation of all data points within this region is calculated as the baseline noise standard deviation. The ratio of the main characteristic peak intensity to the baseline noise standard deviation is the Raman signal quality factor.

[0047] The spatial spectral-texture anomaly indices of all excitation points are normalized to a value between 0 and 1. Similarly, the Raman signal quality factor of all excitation points is also normalized to a value between 0 and 1. The fusion confidence coefficient of the excitation point is obtained by multiplying the normalized spatial spectral-texture anomaly index and the Raman signal quality factor for the same point.

[0048] Each element of the Raman spectral feature vector at a given excitation point is multiplied by the fusion confidence coefficient for that point, resulting in a weighted Raman feature vector. The multi-spectral feature vector of that point is then concatenated with the weighted Raman feature vector in dimensionality to form a longer fusion feature vector containing information from both modalities.

[0049] In an optional embodiment, the ratio of the intensity of the main characteristic peak of the Raman spectral data to the standard deviation of the baseline noise is calculated to obtain the Raman signal quality factor, including:

[0050] Baseline correction was performed on the original Raman spectral data using an iterative reweighted penalized least squares method. In the corrected Raman spectra, the peak with the highest intensity was identified as the principal characteristic peak, and its peak intensity was recorded. In the corrected Raman spectrum, regions without characteristic peaks are selected, and the standard deviation of the signal intensity within these regions is calculated as the baseline noise standard deviation. ;

[0051] Calculate the Raman signal quality factor .

[0052] Specifically, suppose that the acquired raw Raman spectrum has an arched background baseline due to sample fluorescence effects, which interferes with the interpretation of the true signal. By applying an iterative reweighted penalized least squares method, the algorithm can fit and subtract the changing background baseline, thereby obtaining a Raman spectrum with a flat baseline and clear peak characteristics.

[0053] On the baseline-corrected Raman spectrum, for example, in the wavenumber range of 400 to 1800, search for the maximum signal intensity. Assuming the strongest characteristic peak is found at 1003 wavenumbers, with an intensity of 1500 count units, this value is the intensity of the main characteristic peak. In a Raman spectral region lacking characteristic peaks, such as between 1700 and 1750 wavenumbers, calculate the standard deviation of the intensity of all data points within this region. Assuming the obtained value is 5.2 count units, this value is the baseline noise standard deviation. Substituting the values ​​into the formula, we obtain the Raman signal quality factor. It is approximately 288.5, and the higher the value, the better the signal quality.

[0054] In an optional embodiment, the fusion confidence coefficient is calculated based on the spatial spectral-texture anomaly index of the excitation point and the Raman signal quality factor, including:

[0055] Spatial spectral-texture anomaly index AI and Raman signal quality factor were respectively evaluated. Normalization is performed using the logistic function. Mapping to the [0, 1] interval yields the normalization exponent. and normalization factor ;

[0056] The fusion confidence coefficient is obtained by multiplying the normalization exponent and the normalization factor. ;

[0057] Where S(v) represents the normalized value obtained after mapping by the logistic function, v represents the original input variable, k represents the shape parameter of the logistic function, and v0 represents the center point parameter of the logistic function.

[0058] Specifically, assuming the spatial spectral-texture anomaly index AI at a certain excitation point is 9.56, and the Raman signal quality factor... The value is 288.5. The normalization parameter k of the logistic function for AI is set to 0.5, and the center point... The normalization exponent is calculated to be 10. It is approximately 0.77. This value represents the probability that the point is considered an outlier in both spatial and spectral terms.

[0059] Considering The numerical range is large, so different logistic function parameters are set for normalization, for example, k is 0.01, and the center point... The value is 100. The normalization factor is calculated. The value is approximately 0.87, which indicates the reliability of the Raman spectral data. Multiplying the two normalized values ​​yields the fused confidence coefficient C, which is 0.6699. This coefficient combines spatial anomalies and spectral signal quality and is used for subsequent feature weighting.

[0060] Specifically, the core of logistic function calibration is determining the shape parameter k and the center point parameter v0. This needs to be done in conjunction with the sample data and accuracy requirements of the detection scenario. The key process is as follows:

[0061] First, clearly define the calibration target and accurately map the original variables such as the spatial spectrum-texture anomaly index (AI) and Raman signal quality factor to the [0,1] interval to ensure that the normalized values ​​truly reflect the degree of anomaly or signal reliability.

[0062] Next, the parameters are determined using the core parameter calibration method:

[0063] First, a large number of semiconductor tape samples with known clean / contaminated states are collected to obtain the original variables and divide them into normal and abnormal intervals. Then, the center point parameter v0 is calibrated and the midpoint between the upper limit of the normal interval and the lower limit of the abnormal interval is taken. For example, when the value of A is in the normal interval of 0-8 and the abnormal interval of 12-20, v0=10. The signal quality factor is taken as the acceptable signal quality threshold, such as the empirical value of 100.

[0064] Subsequently, the shape parameter k is calibrated, and a k value that is approximately 0.1 or less after normalization of normal samples and approximately 0.9 or more after abnormal samples is selected through trial and error or cross-validation, such as A1 mapping k=0.5 and signal quality factor mapping k=0.01.

[0065] Finally, the test set samples are used for verification. If there is confusion in the normalized values, such as normalized samples being greater than 0.3, v0 and k are adjusted until the classification accuracy requirements are met.

[0066] In an optional embodiment, the Raman spectral feature vector is weighted using a fusion confidence coefficient and concatenated with the multi-spectral feature vector to generate a fusion feature vector, including:

[0067] The multispectral data of the excitation point is used as the multispectral feature vector. The baseline-corrected Raman spectral data is downsampled or interpolated to a preset dimension and used as the Raman spectral feature vector. ; Raman spectral eigenvectors Multiplying each dimension by the fusion confidence coefficient C yields the weighted Raman feature vector. ; to multi-spectral feature vector with weighted Raman eigenvectors The features are concatenated and spliced ​​in sequence to generate a fused feature vector. .

[0068] Specifically, assuming the multispectral data from the excitation point has 16 bands, then the multispectral feature vector of the data... It's a 16-dimensional vector, for example, [0.8, 0.7, ..., 0.9]. Meanwhile, the baseline-corrected Raman spectrum contains 1024 data points. To reduce the data dimensionality, the data is downsampled to 256 dimensions by averaging every four points, resulting in the Raman spectral feature vector. For example, [50, 55, ..., 1500, ..., 80].

[0069] Using the calculated fusion confidence coefficient C, for example 0.6699, the Raman spectral eigenvectors are analyzed. Weighting will be applied. (Soon) Each element in the vector is multiplied by 0.6699 to obtain the weighted Raman eigenvector. For example, a point with an original intensity of 1500 becomes 1004.85 after weighting. This can reduce the weight of low-confidence Raman data in the features. The 16-dimensional multispectral feature vector... With 256-dimensional weighted Raman eigenvectors The features are concatenated to form a 272-dimensional fused feature vector. This vector integrates information from both modalities.

[0070] S3 compares the fused feature vector with a preset pollutant standard feature library, which contains standard feature vectors for various pollutant types; by calculating the spectral distance between the fused feature vector and each standard feature vector, the pollutant type corresponding to the smallest distance is used as the identification result; based on the identification result, the corresponding concentration regression model is called to calculate the pollutant concentration at the excitation point.

[0071] Specifically, the standard feature library of contaminants is pre-established. Standard feature vectors are obtained and stored from known pure contaminant samples, such as silicone oil, cutting fluid, and photoresist residue, using the same process as for the samples to be tested. During comparison, for the fused feature vector generated at the test point, the Euclidean distance or cosine similarity between the fused feature vector and the standard feature vector of each contaminant in the library is calculated. The contaminant type corresponding to the smallest calculated distance value or the highest similarity value is determined as the contaminant identification result for that excitation point.

[0072] For each known contaminant type, a regression model, such as a partial least squares regression model, is pre-established between concentration and fusion feature vector. The training data for the model is obtained by preparing a series of standard samples containing the contaminant at known concentrations. When an excitation point is identified as a contaminant, such as photoresist residue, the pre-trained photoresist residue concentration regression model is invoked, using the fusion feature vector of that point as input. The predicted value output by the model is the contaminant concentration at that point.

[0073] In an optional embodiment, the pollutant type corresponding to the minimum distance is used as the identification result by calculating the spectral distance between the fused feature vector and each standard feature vector, including:

[0074] Cosine distance is used as the metric for spectral distance; for the fused feature vector to be measured... and each standard feature vector in the pollutant standard feature library Calculate the cosine distance ,in, Represents the cosine distance. express The length of the mold, express The modulus is calculated; all calculated cosine distances are compared, and the pollutant type corresponding to the standard feature vector with the smallest cosine distance is determined as the identification result of the excitation point.

[0075] Specifically, assuming a 272-dimensional fused feature vector for a point to be tested has already been generated. Simultaneously, there is a pre-established standard feature library containing standard feature vectors for various known pollutants, such as Rhodamine B, Malachite Green, and Sudan Red. .

[0076] Calculate the fusion feature vector to be tested one by one. With each standard feature vector in the library The cosine distance between them. For example, calculating... The cosine distance to the standard eigenvector of Rhodamine B is 0.08, to the standard eigenvector of malachite green is 0.65, and to the standard eigenvector of Sudan Red is 0.81. The smaller the cosine distance value, the closer the two vectors are in direction, meaning the more similar their Raman spectral characteristics. Comparing the three distance values, 0.08 is the smallest; therefore, the pollutant type at this excitation point is determined to be Rhodamine B.

[0077] In an optional embodiment, based on the identification results, the corresponding concentration regression model is invoked to calculate the pollutant concentration at the trigger point, including:

[0078] For each pollutant type in the pollutant standard feature library, a partial least squares regression model is established using the fusion feature vectors of a series of pollutant samples with known concentrations and their corresponding concentration values. Based on the identified pollutant type, the corresponding partial least squares regression model is selected and loaded. The fusion feature vector of the excitation point to be tested is input into the partial least squares regression model, and the output predicted value is the pollutant concentration at the excitation point.

[0079] Specifically, before detection, a series of standard samples with known concentrations are prepared for each pollutant in the feature library, such as Rhodamine B, for example, from 1 ppm to 100 ppm. Data is collected for each standard sample, and a fused feature vector is generated. Using the feature vector and their corresponding true concentration values, a partial least squares regression model specifically for predicting Rhodamine B concentration is trained and established. Similarly, independent concentration regression models are established for other pollutants such as malachite green and Sudan Red. When a certain excitation point is identified as Rhodamine B, it calls the previously trained Rhodamine B concentration regression model from the model library. The 272-dimensional fused feature vector of the excitation point is used as input and fed into the specific model for calculation. After calculation, the model outputs a predicted concentration value, such as 12.5. This value is determined as the pollutant concentration of Rhodamine B at that excitation point, in ppm.

[0080] S4. For other pixels that are not identified as excitation points, radial basis function interpolation is performed based on the contaminant concentration and spatial location of all excitation points, and the fusion confidence coefficient of each excitation point is used as the influence weight to calculate the contaminant concentration of other pixels, thereby generating a surface cleanliness distribution map of the entire field of view.

[0081] Specifically, all excitation points detected by Raman spectroscopy are used as control points for interpolation calculations. The spatial coordinates and concentration values ​​of these control points are known, and each point also has a fusion confidence coefficient as a reliability weight for the measurement results. For any undetected pixel within the field of view, the contaminant concentration value is calculated using a radial basis function interpolation algorithm. This algorithm comprehensively considers the influence of all control points during calculation; control points closer to the pixel have a greater influence, and control points with higher fusion confidence coefficients also have a greater weight in the interpolation calculation. After performing this interpolation calculation for every pixel within the field of view, the concentration values ​​of all pixels are obtained. Different colors are assigned according to the concentration values ​​to generate a visualized full-field surface cleanliness distribution map.

[0082] In an optional embodiment, radial basis function interpolation is performed to calculate the contaminant concentration of other pixels, including:

[0083] Pollutant concentrations at all N excitation points Spatial location and fusion confidence coefficient As an interpolation condition; a Gaussian function is used. As a radial basis function; for any pixel to be calculated within the field of view. Contaminant concentration of pixels The pollutant concentrations at all trigger points are calculated using a weighted average. The calculation formula is as follows: .

[0084] Suppose three excitation points were selected on the image for Raman spectroscopy analysis. The first excitation point is located at coordinates (20, 30), with a measured contaminant concentration of 12.5 ppm and a fusion confidence coefficient of 0.67. The second excitation point is located at coordinates (80, 50), with a contaminant concentration of 18.2 ppm and a fusion confidence coefficient of 0.81. The third excitation point is located at coordinates (45, 75), with a contaminant concentration of 9.8 ppm and a fusion confidence coefficient of 0.55. These three sets of data, including spatial location, concentration, and fusion confidence coefficient, will serve as the basis for interpolation calculations. The contaminant concentration at the pixel at coordinates (50, 40) is estimated, and the spatial distance from this pixel to the three known points is calculated. The spatial distance is then converted into spatial weights using a Gaussian function. A weighted average is calculated according to a given formula, taking into account both the distance between each known point and the target point, as well as the reliability of the measurement results at each known point. For example, the pollutant concentration at pixel (50, 40) is calculated to be 14.3 ppm. This process is repeated for all unmeasured pixels in the image to generate a pollutant concentration distribution map.

[0085] An embodiment of the semiconductor tape surface cleanliness detection system based on Raman spectroscopy provided by this invention:

[0086] The Raman spectroscopy-based semiconductor tape surface cleanliness detection system includes a processor and a memory. The memory stores computer program instructions, which, when executed by the processor, implement the aforementioned Raman spectroscopy-based semiconductor tape surface cleanliness detection method.

[0087] The Raman spectroscopy-based semiconductor tape surface cleanliness detection system also includes other components well known to those skilled in the art, such as communication interfaces. Their setup and functions are known in the art and will not be described in detail here.

[0088] In this invention, the aforementioned memory can be any tangible medium containing or storing a program that can be used or combined with an instruction execution system, apparatus, or device. For example, a computer-readable storage medium can be any suitable magnetic or magneto-optical storage medium, such as Resistive Random Access Memory (RRAM), Dynamic Random Access Memory (DRAM), Static Random Access Memory (SRAM), Enhanced Dynamic Random Access Memory (EDRAM), High-Bandwidth Memory (HBM), Hybrid Memory Cube (HMC), etc., or any other medium that can be used to store desired information and can be accessed by an application, module, or both. Any such computer storage medium can be part of a device or accessible to or connected to a device. Any application or module described in this invention can be implemented using computer-readable / executable instructions stored or otherwise maintained by such a computer-readable medium.

[0089] The above are all preferred embodiments of the present invention and are not intended to limit the scope of protection of the present invention. Therefore, all equivalent changes made in accordance with the structure, shape and principle of the present invention should be covered within the scope of protection of the present invention.

Claims

1. A method for detecting the cleanliness of a semiconductor tape surface based on Raman spectroscopy, characterized by, The method comprises the following steps: S1, acquiring a multi-spectrum image of a surface of a semiconductor adhesive tape; for each pixel point in the multi-spectrum image, a weighted sum of Mahalanobis distance in the spectral dimension and Gabor texture response in the spatial dimension of the pixel point is calculated to obtain a spatial-spectral-texture anomaly index; Based on the spatial-spectral-texture anomaly index, an excitation point of the Raman spectrum is determined, and Raman spectrum data of each excitation point is collected; S2, extracting a multi-spectrum feature vector and a Raman spectrum feature vector of the excitation point; calculating a ratio of a main characteristic peak intensity of the Raman spectrum data to a standard deviation of baseline noise to obtain a Raman signal quality factor; According to the spatial-spectral-texture anomaly index and the Raman signal quality factor of the excitation point, a fusion confidence coefficient is calculated, comprising: respectively, are normalized, mapped to the interval [0, 1] by a logistic function , and obtained as normalized index and normalized factor ; the fusion confidence coefficient is obtained by multiplying the normalized index and the normalized factor ;​ S(v) represents a normalized value obtained after mapping by a logistic function, v represents an original variable input, k and v0 represent shape parameters and center point parameters of the logistic function, respectively; The Raman spectrum feature vector is weighted by using the fusion confidence coefficient, and is spliced with the multi-spectrum feature vector to generate a fusion feature vector; S3, by calculating a spectral distance between the fusion feature vector and each standard feature vector in a preset standard feature library of pollutants, a pollutant type corresponding to a minimum distance is taken as a recognition result; according to the recognition result, a corresponding concentration regression model is called to calculate a pollutant concentration of the excitation point; S4, for other pixel points that are not determined as excitation points, based on pollutant concentrations and spatial positions of all excitation points, and taking fusion confidence coefficients of the excitation points as influence weights, a radial basis function interpolation is performed to calculate pollutant concentrations of the other pixel points, thereby generating a surface cleanliness distribution map of a full field of view; The pollutant concentration at all N excitation points , spatial position and fusion confidence coefficient as interpolation conditions; Gaussian function as radial basis function; for any pixel point to be calculated within the field of view , the pollutant concentration of the pixel point is calculated by weighted average of the pollutant concentrations of all excitation points, satisfying: .

2. The Raman spectroscopy-based semiconductor tape surface cleanliness detection method according to claim 1, characterized by, For each pixel point in the multi-spectrum image, a weighted sum of Mahalanobis distance in the spectral dimension and Gabor texture response in the spatial dimension of the pixel point is calculated to obtain a spatial-spectral-texture anomaly index, comprising: The spectral mean vector of all pixel points in the whole multi-spectrum image and the covariance matrix are taken as background statistics, the spectral Mahalanobis distance of the pixel point is calculated ; a Gabor filter bank with multiple scales and directions is used to filter the multi-spectrum image, and the maximum value of each pixel point in all Gabor filter responses is taken as the Gabor texture response ; Setting spectral weights and texture weights , and , computing spatial spectral-texture anomaly index .

3. The Raman spectroscopy-based semiconductor tape surface cleanliness detection method according to claim 1, characterized by, A ratio of a main characteristic peak intensity of the Raman spectrum data to a standard deviation of baseline noise is calculated to obtain a Raman signal quality factor, comprising: Baseline correction is applied to the original Raman spectrum data by using iterative reweighted penalized least squares method; in the corrected Raman spectrum, the peak with the maximum intensity is found as the main characteristic peak, and the peak value intensity of the main characteristic peak is recorded ; in the corrected Raman spectrum, a region without characteristic peaks is selected, and the standard deviation of signal intensity in the region is calculated as the baseline noise standard deviation ; Computing a raman signal quality factor .

4. The Raman spectroscopy-based semiconductor tape surface cleanliness detection method according to claim 1, characterized by, The Raman spectrum feature vector is weighted by using the fusion confidence coefficient, and is spliced with the multi-spectrum feature vector to generate a fusion feature vector, comprising: The multispectral data of the excitation point is taken as a multispectral feature vector The baseline-corrected Raman spectrum data is down-sampled or interpolated to a preset dimension as a Raman spectrum feature vector Each dimension of the Raman spectrum feature vector is multiplied by a fusion confidence coefficient C to obtain a weighted Raman feature vector The multispectral feature vector and the weighted Raman feature vector are concatenated in a front-back order to generate a fusion feature vector .

5. The Raman spectroscopy-based semiconductor tape surface cleanliness detection method according to claim 1, characterized by, By calculating a spectral distance between the fusion feature vector and each standard feature vector in a preset standard feature library of pollutants, a pollutant type corresponding to a minimum distance is taken as a recognition result, comprising: Cosine distance is used as the measurement standard of spectrum distance; for the fusion feature vector to be tested and each standard feature vector in the standard feature library of pollutants , the cosine distance is calculated , wherein denotes the cosine distance, denotes the module length of denotes the module length of; compare all the calculated cosine distances, and determine the pollutant type corresponding to the standard feature vector with the smallest cosine distance as the recognition result of the excitation point.

6. The Raman spectroscopy-based semiconductor tape surface cleanliness detection method according to claim 1, characterized by, According to the recognition result, a corresponding concentration regression model is called to calculate a pollutant concentration of the excitation point, comprising: For each pollutant type in the standard feature library of pollutants, a partial least squares regression model is established by using a series of fusion feature vectors of pollutant samples with known concentrations and corresponding concentration values; according to the recognized pollutant type, a corresponding partial least squares regression model is selected and loaded, and the fusion feature vector of the to-be-tested excitation point is input into the partial least squares regression model, and a predicted value output by the partial least squares regression model is the pollutant concentration of the excitation point.

7. The Raman spectroscopy-based semiconductor tape surface cleanliness detection method according to claim 1, characterized by, In S1, pixel points with a spatial-spectral-texture anomaly index greater than a first preset threshold are determined as excitation points of the Raman spectrum.

8. A Raman spectroscopy-based semiconductor tape surface cleanliness detection system, comprising: The semiconductor adhesive tape surface cleanliness detection method based on Raman spectrum comprises a memory and a processor, computer program instructions are stored in the memory, and the computer program instructions realize the semiconductor adhesive tape surface cleanliness detection method based on Raman spectrum in any one of claims 1-7 when the computer program instructions are executed by the processor.

Images