Automatic identification system for color of surface film layer of 12-inch regenerated wafer

Abstract

The invention discloses a 12-inch regenerated wafer surface film layer color automatic identification system and method. The system comprises: an adjustable multispectral illumination module configured with an annular LED array with a dynamic color temperature (3000K-6500K) and an incident angle (30-80 degrees); the air floating type rotary scanning mechanism realizes 360-degree dead-corner-free imaging; and the intelligent analysis module is used for carrying out color classification by adopting an Inception-ResNet network trained by transfer learning, and realizing process closed-loop control in combination with elliptical polarization film thickness verification. Through a multispectral fusion algorithm: # imgabs0 # dynamic compensation material optical property influence, roughness compensation and a dual verification mechanism are innovatively introduced, so that the color recognition accuracy reaches 99.3%, the detection efficiency is improved by 15 times, and the method is particularly suitable for high-precision detection of semiconductor regenerated wafers.

Description

Technical Field

[0001] The present invention relates to the field of semiconductor manufacturing detection technology, and in particular to a system and method for automatically identifying the color of a film layer on the surface of a 12-inch regenerated wafer based on multispectral imaging and artificial intelligence. Background Art

[0002] In the semiconductor wafer regeneration process, accurate identification of the surface film color is directly related to the monitoring of film thickness, composition, and process parameters. Traditional detection technology has the following significant drawbacks:

[0003] 1. Limitations of manual visual inspection

[0004] The current industry still generally uses the inspection method of operators holding SEMI F121 standard color cards for visual comparison. The technical bottleneck is:

[0005] The human eye has insufficient color sensitivity: According to CIE 1931 colorimetry experimental data, the average human eye's recognition threshold for ΔE*ab color difference is 2.3±0.5 (JND value), while high-end regenerated wafers require an inspection accuracy of ΔE≤1.5 (SEMI STD F121-1108).

[0006] Significant environmental interference: A 500K fluctuation in the color temperature of the lighting source can result in a visual judgment error of ΔE>1.8 (experimental data see J.Opt.Soc.Am.A, 2019);

[0007] Inefficiency: Statistics show that full-surface inspection of a 12-inch wafer takes 3-5 minutes per wafer, and the consistency of manual inspection is only 82.7% (data from SEMI's annual quality report).

[0008] 2. Interference from the special surface state of the regenerated wafer

[0009] The following characteristic defects occur in regenerated wafers due to repeated use:

[0010] Oxidation residue layer: SiO2 nanoparticles (50-200nm in diameter) left over from the CMP process can induce Mie scattering, causing a color measurement offset of ΔE of 2.5-3.8 (measured data).

[0011] Micro-damage interference: surface scratches (depth> 50nm) and pits (density> 10 / cm 2 ) produces a light diffraction effect that increases the color recognition error rate of traditional RGB imaging systems by 37%;

[0012] Surface roughness variation: The Ra value of regenerated wafers fluctuates between 0.2-0.8 μm (new wafer Ra < 0.1 μm), resulting in a 45% difference in diffuse reflected light intensity (SPIE Proc. 11245, 2020).

[0013] 3. Technical defects of existing automatic detection equipment

[0014] Although automatic detection equipment represented by CN202510300131.X has partially replaced manual labor, it has the following key technical bottlenecks:

[0015] Optical system design flaws:

[0016] The ring-shaped LED light source with a fixed color temperature (5600K±200K) cannot adapt to the reflective characteristics of different film materials (such as SiO2 refractive index n=1.46, SiNx=2.05);

[0017] The spectral response range of RGB cameras (400-650nm) cannot capture long-wavelength features (such as the key absorption peak of polysilicon film layers at 650-780nm);

[0018] Insufficient algorithm adaptability:

[0019] Without considering the influence of surface roughness, when Ra>0.5μm, diffuse reflection causes the deviation of the V component of the HSV color space to be greater than 15%;

[0020] The traditional SVM-based classification model is sensitive to oxidation interference in regenerated wafers, and its measured accuracy is only 93.2% (F1-score = 0.89);

[0021] The above technical defects make it difficult for existing solutions to meet the high-standard inspection requirements of 12-inch recycled wafers. Summary of the Invention

[0022] The purpose of the present invention is to provide an automatic recognition system for the color of the surface film layer of a 12-inch regenerated wafer to solve the problems raised in the above background technology.

[0023] To achieve the above object, the present invention provides the following technical solution: an automatic recognition system for the color of a film layer on the surface of a 12-inch regenerated wafer, comprising:

[0024] Image acquisition module: This module includes a ring-shaped LED light source array and a high-resolution industrial camera. The light source array can dynamically adjust the color temperature (3000K-6500K) and the angle of incidence (30°-80°). The camera is equipped with a narrow-band filter to isolate ambient light interference.

[0025] Color analysis module: Converts the collected RGB image into HSV through multispectral reflectance analysis algorithm

[0026] Color space, and establish a color-film thickness correlation model based on the prior database of film materials;

[0027] Classification output module: A convolutional neural network (CNN) trained by transfer learning is used to classify the film color of the preprocessed image and output the recognition result with a color difference value ΔE ≤ 1.5 compared with the SEMI standard color card.

[0028] Preferably, the image acquisition module includes an air-floating wafer rotation stage, which cooperates with a linear array CMOS camera to achieve 360° full-circle scanning, the rotation stage positioning accuracy is ≤±0.1°, and the surface is configured with a vacuum adsorption hole array with a hydrophilic coating.

[0029] Preferably, the color analysis module adopts an adaptive noise reduction algorithm, specifically including:

[0030] Separate the high-frequency components of the image through wavelet transform to eliminate micro-scratch noise on the wafer surface;

[0031] Morphological opening operation was applied to eliminate the interference of particles with a size less than 5 μm;

[0032] The region growing method is used to segment the effective detection area and exclude the invalid area of 5mm at the edge.

[0033] Preferably, the CNN network structure comprises:

[0034] Input layer: receives a normalized image of 512×512 pixels;

[0035] Feature extraction layer: contains 4 groups of Inception-ResNet modules, each containing 32-128 convolution kernels;

[0036] Classification layer: The output layer uses the Softmax function to generate probability distribution corresponding to the 21 standard color classifications in the SEMI F121 standard.

[0037] Preferably, the system also includes a film thickness verification unit, which measures the film thickness in real time through an ellipsometer. When the combination of color and thickness data deviates from a preset process window, an alarm signal is triggered and the abnormal coordinates are recorded.

[0038] Preferably, a systematic film color identification method comprises the following steps:

[0039] S1. Wafer surface pretreatment: Remove the surface oxide layer through a plasma cleaning device, and the processing parameters are

[0040] Ar gas flow rate 50 sccm, RF power 300 W, processing time 120 s ± 5 s;

[0041] S2. Multispectral imaging: Collect reflectance images at three characteristic wavelengths: 450nm, 550nm, and 650nm;

[0042] S3. Color space fusion: Perform weighted fusion on the three-channel images, with the weight coefficient dynamically adjusted according to the refractive index of the film material;

[0043] S4. Credibility verification: When the CNN classification confidence is less than 90%, the secondary recognition process is started and the SVM classifier is used for cross-validation.

[0044] Preferably, the weight coefficient in step S3 is calculated using the following formula:

[0045]

[0046] where n λ is the refractive index of the film material at wavelength λ, α λ is the absorption coefficient, and d is the nominal film thickness.

[0047] Preferably, the secondary identification process includes:

[0048] Extract image LBP texture features and color histogram features;

[0049] The final classification is obtained by combining texture similarity and color difference values through fuzzy logic algorithm;

[0050] When the two recognition results are inconsistent, the manual review interface is activated and the disputed area is marked.

[0051] Compared with the prior art, the present invention has the following beneficial effects:

[0052] 1. Detection accuracy is significantly improved

[0053] Color Difference Quantification Control: Multi-spectral dynamic illumination (450 / 550 / 650nm) and narrowband filters synergistically suppress ambient light interference. Combined with HSV color space conversion and transfer learning CNN classifiers, the ΔE color difference detection accuracy is ≤1.5 (traditional visual ΔE>2.0), meeting the highest level requirements of SEMI standards. Enhanced Anti-Interference Capability: An adaptive noise reduction algorithm using wavelet transform and morphological opening operation effectively eliminates interference from surface micro-scratches (depth <50nm) and particles (particle size <5μm), maintaining an identification accuracy of ≥99.3% even on rough surfaces with Ra = 0.8μm (traditional methods drop to 85% when Ra > 0.5μm).

[0054] 2. Breakthrough in detection efficiency and coverage

[0055] High-speed full-circle scanning: An air-bearing rotary stage (positioning accuracy ±0.1°) coupled with a linear array CMOS camera achieves 360° imaging with no blind spots. Single-wafer inspection time is ≤ 15 seconds, 15 times more efficient than manual inspection. Edge area optimization: A region growing algorithm precisely segments the effective inspection area, achieving a 100% rejection rate for invalid areas within 5mm of the edge, reducing missed inspections by 82% compared to existing equipment.

[0056] 3. Multi-parameter collaborative analysis and process closed loop

[0057] Real-time correlation between color and film thickness: An integrated spectroscopic ellipsometer measures film thickness online and establishes a correlation model based on the material's optical constants (n, k). The film thickness inversion error is ±3nm (traditional offline measurement delay is >30 minutes).

[0058] Fast response to abnormalities: When the color difference or film thickness deviates from the process window, the system can trigger an alarm and mark the coordinates within 0.5 seconds, increasing the detection rate of process abnormalities to 99.8%, reducing the risk of batch scrapping of wafers.

[0059] 4. Complex surface adaptive capability

[0060] Dynamic roughness compensation: Targeting the surface characteristics of recycled wafers with Ra=0.2-0.8μm, anisotropic diffusion filtering and V component enhancement algorithms are used to increase color recognition confidence to over 96% (only 85% without compensation).

[0061] Material universality optimization: Built-in (n, k) optical database of six common film layers such as SiO2 / SiNx / Poly, supports automatic matching of refractive index range of 1.4-3.5, and applicability covers 98% of recycled wafer processes. BRIEF DESCRIPTION OF THE DRAWINGS

[0062] Figure 1 Schematic diagram of the system structure of the present invention;

[0063] Figure 2 : Color analysis algorithm flow chart;

[0064] Figure 3 This is the schematic diagram of the convolutional neural network (CNN);

[0065] Figure 4 This is the SVM principle and calculation formula diagram.

[0066] In the picture: 1. Industrial camera, 2. Dynamic LED light source, 3. Air-bearing rotation stage, 4. Linear array CMOS. DETAILED DESCRIPTION

[0067] The following will clearly and completely describe the technical solutions in the embodiments of the present invention in conjunction with the accompanying drawings. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without making creative efforts are within the scope of protection of the present invention.

[0068] See also Figure 1-2 , the present invention provides a technical solution:

[0069] An automatic recognition system for the color of a film layer on a 12-inch regenerated wafer surface, comprising:

[0070] Image acquisition module: This module includes a ring-shaped LED light source array and a high-resolution industrial camera. The light source array can dynamically adjust the color temperature (3000K-6500K) and the angle of incidence (30°-80°). The camera is equipped with a narrow-band filter to isolate ambient light interference.

[0071] Color analysis module: This module converts the collected RGB images into the HSV color space through a multispectral reflectance analysis algorithm, and establishes a color-film thickness correlation model based on a priori database of film materials.

[0072] Classification output module: A convolutional neural network (CNN) trained by transfer learning is used to classify the film color of the preprocessed image and output the recognition result with a color difference value ΔE ≤ 1.5 compared with the SEMI standard color card.

[0073] The image acquisition module includes an air-floating wafer rotation stage, which is combined with a linear array CMOS camera to achieve 360° full-circle scanning. The rotation stage positioning accuracy is ≤±0.1°, and the surface is equipped with a vacuum adsorption hole array with a hydrophilic coating.

[0074] The color analysis module uses an adaptive noise reduction algorithm, which includes:

[0075] Separate the high-frequency components of the image through wavelet transform to eliminate micro-scratch noise on the wafer surface;

[0076] Morphological opening operation was applied to eliminate the interference of particles with a size less than 5 μm;

[0077] The region growing method is used to segment the effective detection area and exclude the invalid area of 5mm at the edge.

[0078] The CNN network structure includes:

[0079] Input layer: receives a normalized image of 512×512 pixels;

[0080] Feature extraction layer: contains 4 groups of Inception-ResNet modules, each containing 32-128 convolution kernels;

[0081] Classification layer: The output layer uses the Softmax function to generate probability distribution corresponding to the 21 standard color classifications in the SEMI F121 standard.

[0082] The system also includes a film thickness verification unit, which measures the film thickness in real time using an ellipsometer. When the combination of color and thickness data deviates from the preset process window, an alarm signal is triggered and the abnormal coordinates are recorded.

[0083] Another key issue is credibility verification: when the CNN classification confidence is less than 90%, a secondary recognition process is initiated, using the SVM classifier for cross-validation. If the two recognition results are inconsistent, a manual review interface is activated and the disputed area is marked. Figure 4 shown.

[0084] As for why the svm algorithm model is used, the following references are given:

[0085]

[0086] A method for automatically identifying the color of a film layer on a 12-inch regenerated wafer surface by a system for automatically identifying the color of a film layer on a 12-inch regenerated wafer surface comprises the following steps:

[0087] S1. Wafer surface pretreatment: Remove the surface oxide layer through a plasma cleaning device, and the processing parameters are

[0088] Ar gas flow rate 50 sccm, RF power 300 W, processing time 120 s ± 5 s;

[0089] S2. Multispectral imaging: Collect reflectance images at three characteristic wavelengths: 450nm, 550nm, and 650nm;

[0090] S3. Color space fusion: Perform weighted fusion on the three-channel images, with the weight coefficient dynamically adjusted according to the refractive index of the film material;

[0091] S4. Credibility verification: When the CNN classification confidence is less than 90%, the secondary recognition process is started and the SVM classifier is used for cross-validation.

[0092] 7. The weight coefficient in step S3 is calculated using the following formula:

[0093]

[0094] where n λ is the refractive index of the film material at wavelength λ, α λ is the absorption coefficient, and d is the nominal film thickness. The secondary identification process includes:

[0095] Extract image LBP texture features and color histogram features;

[0096] The final classification is obtained by combining texture similarity and color difference values through fuzzy logic algorithm;

[0097] When the two recognition results are inconsistent, the manual review interface is activated and the disputed area is marked.

[0098] Example 1: SiO2 film color recognition

[0099] 1. Place the regenerated wafer on the air float rotating table and start plasma cleaning (Ar 50sccm, 300W, 120s)

[0100] 2. Set the LED light source: color temperature 5000K, incident angle 45°, and collect three-band reflection images

[0101] 3. Color analysis module calculates weight coefficient (n = 1.46 @ 550nm, α = 0.02μm -1 )

[0102] 4. CNN classifier outputs "light blue" (corresponding to SEMI F121 standard color number #07)

[0103] 5. The film thickness verified by ellipsometry is 152nm (process window: 150±5nm), and it is considered qualified

[0104] Example 2: SiNx film abnormality detection

[0105] 1. Detecting color difference ΔE = 1.8 (threshold 1.5) triggers secondary verification

[0106] 2. SVM classifier extracts LBP texture features (contrast improvement of 30%)

[0107] 3. The fuzzy logic algorithm determines it as "dark gray with a brownish tint" (color number #15).

[0108] 4. The film thickness measurement shows 183nm (standard value 170±10nm), and the system marks it as "abnormal film thickness"

[0109] 5. Optical microscope review confirmed that the abnormal PECVD process caused the film composition to shift

[0110] Example 3: Poly-Si film roughness compensation

[0111] 1. Surface roughness Ra = 0.6 μm measured by white light interferometer

[0112] 2. Enable roughness compensation algorithm:

[0113] Enhance the V component in HSV space (brightness increased by 20%)

[0114] Apply anisotropic diffusion filtering to reduce noise

[0115] 3. Color recognition confidence increased from 85% to 96%

[0116] 4. Output "Metallic Gray" (color number #12), which is consistent with the SEMI standard color card.

[0117] The system and method for automatically identifying the color of the film layer on the surface of a 12-inch regenerated wafer provided by the present invention have the following significant technical advantages over traditional manual detection and existing automatic equipment:

[0118] 1. Detection accuracy is significantly improved

[0119] Color difference quantitative control: Multi-spectral dynamic lighting (450 / 550 / 650nm) and narrow-band filters are used to synergistically suppress ambient light interference. Combined with HSV color space conversion and transfer learning CNN classifiers, the ΔE color difference detection accuracy is ≤1.5 (traditional visual ΔE>2.0), meeting the highest level requirements of SEMI standards.

[0120] Enhanced anti-interference ability: Adopting the adaptive noise reduction algorithm of wavelet transform and morphological opening operation, it effectively eliminates the interference of surface micro scratches (depth <50nm) and particles (particle size <5μm), and maintains the recognition accuracy rate ≥99.3% under the rough surface of Ra=0.8μm (traditional methods are

[0121] When Ra>0.5μm, the accuracy drops to 85%).

[0122] 2. Breakthrough in detection efficiency and coverage

[0123] High-speed full-circle scanning: The air-floating rotation stage (positioning accuracy ±0.1°) and the linear array CMOS camera achieve 360° imaging without blind spots. The single-wafer inspection time is ≤ 15 seconds, which is 15 times more efficient than manual inspection.

[0124] Edge area optimization: The effective detection area is accurately segmented through the region growing algorithm, and the 5mm invalid area at the edge is excluded at a rate of 100%, which is 82% lower than the missed detection rate of existing equipment.

[0125] 3. Multi-parameter collaborative analysis and process closed loop

[0126] Real-time correlation between color and film thickness: An integrated spectroscopic ellipsometer measures film thickness online and establishes a correlation model based on the material's optical constants (n, k). The film thickness inversion error is ±3nm (traditional offline measurement delay is >30 minutes).

[0127] Fast response to abnormalities: When the color difference or film thickness deviates from the process window, the system can trigger an alarm and mark the coordinates within 0.5 seconds, increasing the detection rate of process abnormalities to 99.8%, reducing the risk of batch scrapping of wafers.

[0128] 4. Complex surface adaptability

[0129] Dynamic roughness compensation: Targeting the surface characteristics of recycled wafers with Ra=0.2-0.8μm, anisotropic diffusion filtering and V component enhancement algorithms are used to increase color recognition confidence to over 96% (only 85% without compensation).

[0130] Material universality optimization: Built-in (n, k) optical database of six common film layers such as SiO2 / SiNx / Poly, supports automatic matching of refractive index range of 1.4-3.5, and applicability covers 98% of recycled wafer processes.

[0131] 5. Significant economic benefits

[0132] Actual application data shows that this system can reduce manual re-inspection hours by 82%, saving approximately 12 million yuan in quality costs for a single production line each year.

[0133] Through high-precision detection with ΔE≤1.5, the annual losses caused by abnormalities in the recycled wafer process have been reduced from US$230 million to US$40 million.

[0134] Comprehensive Advantage Comparison

[0135]

[0136]

[0137] Through multi-dimensional innovations in optics, algorithms, and mechanics, this invention solves the key technical bottleneck in color detection of recycled wafers, providing the semiconductor manufacturing industry with a high-precision, high-efficiency, and high-reliability detection solution.

[0138] The training process of CNN is usually carried out through the back-propagation algorithm, using the gradient descent method to minimize the loss function. The loss function (such as cross entropy loss) is used to evaluate the predictive performance of the model, and the weights are updated by calculating the gradient of each parameter. Figure 3 Shows the schematic diagram of the convolutional neural network (CNN).

[0139] While embodiments of the present invention have been shown and described, it will be appreciated by those skilled in the art that various changes, modifications, substitutions, and variations may be made to these embodiments without departing from the principles and spirit of the invention, and that the scope of the invention is defined by the appended claims and their equivalents.

Claims

1. An automatic recognition system for the color of the film layer on the surface of a 12-inch regenerated wafer, characterized in that include: Image acquisition module: This module includes a ring-shaped LED light source array and a high-resolution industrial camera. The light source array can dynamically adjust the color temperature (3000K-6500K) and the angle of incidence (30°-80°). The camera is equipped with a narrow-band filter to isolate ambient light interference. Color analysis module: Converts the collected RGB image into HSV through multispectral reflectance analysis algorithm Color space, and establish a color-film thickness correlation model based on the prior database of film materials; Classification output module: A convolutional neural network (CNN) trained by transfer learning is used to classify the film color of the preprocessed image and output the recognition result with a color difference value ΔE ≤ 1.5 compared with the SEMI standard color card.

2. The automatic recognition system for the color of a film layer on the surface of a 12-inch regenerated wafer according to claim 1, characterized in that: The image acquisition module includes an air-floating wafer rotating stage and a linear array CMOS The camera achieves 360° full-circle scanning, the rotation stage positioning accuracy is ≤±0.1°, and the surface is equipped with a vacuum adsorption hole array with a hydrophilic coating.

3. The automatic recognition system for the color of a film layer on the surface of a 12-inch regenerated wafer according to claim 1, characterized in that: The color analysis module adopts an adaptive noise reduction algorithm, which specifically includes: Separate the high-frequency components of the image through wavelet transform to eliminate micro-scratch noise on the wafer surface; Morphological opening operation was applied to eliminate the interference of particles with a size less than 5 μm; The region growing method is used to segment the effective detection area and exclude the invalid area of 5mm at the edge.

4. The automatic recognition system for the color of a film layer on the surface of a 12-inch regenerated wafer according to claim 1, characterized in that: The CNN network structure includes: Input layer: receives a normalized image of 512×512 pixels; Feature extraction layer: contains 4 groups of Inception-ResNet modules, each group contains 32-128 convolution kernels; classification layer: the output layer uses the Softmax function to generate probability distribution, corresponding to the 21 standard color classifications in the SEMI F121 standard.

5. The automatic recognition system for the color of the surface film of a 12-inch regenerated wafer according to claim 1, characterized in that It also includes a film thickness verification unit, which measures the film thickness in real time through an ellipsometer. When the combination of color and thickness data deviates from the preset process window, an alarm signal is triggered and the abnormal coordinates are recorded.

6. A method for identifying the film color of a 12-inch regenerated wafer surface based on an automatic identification system for the film color of a 12-inch regenerated wafer according to any one of claims 1 to 5, characterized in that The following steps are involved: S1. Wafer surface pretreatment: Remove the surface oxide layer using a plasma cleaning device. Processing parameters are: Ar flow rate 50 sccm, RF power 300 W, and treatment time 120 s ± 5 s. S2. Multispectral imaging: Collect reflectance images at three characteristic wavelengths: 450nm, 550nm, and 650nm; S3. Color space fusion: Perform weighted fusion on the three-channel images, with the weight coefficient dynamically adjusted according to the refractive index of the film material; S4. Credibility verification: When the CNN classification confidence is less than 90%, the secondary recognition process is started and the SVM classifier is used for cross-validation.

7. The film color recognition method according to claim 6, wherein: The weight coefficient in step S3 is calculated using the following formula: where n λ is the refractive index of the film material at wavelength λ, α λ is the absorption coefficient, and d is the nominal film thickness.

8. The film color recognition method according to claim 6, wherein: The secondary identification process includes: Extract image LBP texture features and color histogram features; The final classification is obtained by combining texture similarity and color difference values through fuzzy logic algorithm; When the two recognition results are inconsistent, the manual review interface is activated and the disputed area is marked.

Images