Random polarization scattering and multimodal feature extraction for pr gel surface particle detection method

Abstract

The present application belongs to the technical field of semiconductor manufacturing process detection, and aims at the technical problems of fixed polarization information collection and random particle spatial distribution in the prior art, and provides a PR glue surface particle detection method based on random polarization scattering and multi-modal feature extraction, n non-repeated angle polarizations are randomly generated in the range of 0-180 degrees, the wafer to be measured is pre-scanned, the surface particle density distribution is predicted through the pre-scanned image, and is partitioned to generate a density adaptive scanning path; the wafer is moved according to the path, multi-polarization state scattering images of different resolutions corresponding to different density areas are synchronously collected, and multi-modal feature fusion processing is performed, so that the detection adaptability and detection efficiency of the glue layer with different thicknesses are overall improved.

Description

Technical Field

[0001] This invention belongs to the field of semiconductor manufacturing process inspection technology, and particularly relates to a method for detecting tiny particles on the surface of photoresist (PR). Background Technology

[0002] Particle detection on the surface of semiconductor photoresist is crucial throughout the entire process, from the raw photoresist solution to patterning on the wafer. In photolithography, micron- and submicron-sized particles adhering to the surface of the photoresist can cause distortion of the exposure pattern, severely impacting chip yield.

[0003] Existing detection technologies identify defects by utilizing the difference in polarization scattering characteristics between particles and the PR adhesive substrate. The light source emits linearly polarized light, employing an industry-standard fixed 4-angle sequence (0° / 45° / 90° / 135°), incident obliquely onto the PR adhesive surface at an angle of 15°~30°. The detector simultaneously acquires the intensity of scattered light at four fixed polarization angles and calculates the single characteristic parameter, degree of polarization (DoP). It also uses a globally fixed threshold to determine whether a pixel's polarization exceeds a preset threshold and marks it as a particle defect.

[0004] The aforementioned existing technologies are completely inadequate for adapting to the dynamic variations in PR adhesive thickness from 300 to 5000 nm. Different thicknesses of PR adhesive produce different phase delays, causing a drift in the polarization characteristics of the background substrate. Calculations of polarization degree at a fixed angle will result in systematic errors, leading to missed detections in thin adhesive areas and severe misjudgments in thick adhesive areas. Furthermore, the scanning path planning relies on a priori defect models and does not consider the randomness of particle spatial distribution, resulting in an excessively high proportion of invalid scanning areas (over 30%). Feature extraction is limited to a single polarization parameter (such as DoP), which is insufficient for distinguishing the reflection at the interface between transparent particles and PR adhesive, resulting in a detection accuracy of less than 85% for particles below 50 nm. Summary of the Invention

[0005] To address the technical problems of fixed polarization information and random spatial distribution of particles in existing technologies, this invention provides a method for detecting particles on the surface of PR adhesive based on random polarization scattering and multimodal feature extraction, specifically including the following steps: S1. Pre-processing and parameter initialization of the wafer to be tested, and measuring the thickness of the PR adhesive; S2. Randomly generate n non-repeating polarization angles within the range of 0~180°. The angle sequence is randomly updated for each detection. Adjust the light source wavelength according to the PR adhesive thickness measured in S1. S3. Pre-scan the wafer under test. Predict the surface particle density distribution and divide it into regions based on the pre-scan image, and generate a density-adaptive scanning path. Move the wafer according to the path and simultaneously acquire multi-polarization scattering images of different density regions at corresponding resolutions. S4. Multimodal feature fusion processing: The polarization channel is designed to extract the polarization features of the scattering image obtained in S3 at random polarization angles, and the morphology channel is designed to extract the morphological features of particles. The features are then fused through an adversarial training mechanism. S5. Adaptive threshold decision and output detection results.

[0006] Technical effects: This invention achieves high-speed and high-precision detection of particles on the surface of PR adhesive by employing a random polarization information acquisition mode, constructing a particle density prediction model, and designing a feature fusion algorithm. By switching non-periodic polarization angles, the scattered light intensity matrix of the PR adhesive surface is acquired, avoiding the drift phenomenon of existing technologies with fixed polarization angle sequences in response to dynamic changes in the film layer. By predicting the surface particle density distribution, a density-adaptive scanning path is generated, thereby enabling targeted scanning of the wafer surface under test and improving efficiency. A polarization-morphology dual-channel feature extraction network is designed, and features are fused through an adversarial training mechanism to solve the problem of difficulty in distinguishing transparent particles from the background. Simultaneously, an adaptive threshold decision step is added, dynamically adjusting the defect judgment threshold based on the PR adhesive thickness measured in real time by a spectrometer. Specifically, when the thickness h < 1000 nm, a polarization entropy threshold (>0.6) is used; when the thickness h ≥ 1000 nm, a light intensity fluctuation threshold (>0.3) is used, replacing the existing fixed threshold judgment method and improving the detection adaptability of adhesive layers of different thicknesses. Attached Figure Description

[0007] Figure 1 This is a flowchart illustrating the overall process of this invention.

[0008] Figure 2 This is a flowchart of the adaptive path execution and scattering image acquisition process in S3 of this embodiment. Detailed Implementation

[0009] To make the technical means, other features and advantages of the present invention easier to understand, a more detailed description is given below in conjunction with the embodiments.

[0010] like Figure 1 As shown, this embodiment provides a method for detecting particles on the surface of PR adhesive based on random polarization scattering and multimodal feature extraction, specifically including the following steps: S1. Pre-processing and parameter initialization of the wafer under test: The wafer under test is fixed on a six-degree-of-freedom fine-tuning platform by electrostatic adsorption, with a positioning accuracy of ±0.5μm. The thickness h of the PR adhesive is measured by a spectroscopic ellipsometry, and the sampling interval is set to 1mm×1mm. S2. The random polarization angle generation unit randomly generates multiple non-repeating angles within the range of 0~180°. For example, the seven non-repeating angles are 17°, 43°, 68°, 92°, 125°, 151°, and 173°. The angle sequence is randomly updated with each detection. The wavelength of the light source is adjusted according to the h value: 380nm is selected when h<1000nm, and 650nm is selected when h≥1000nm. The wavelength is adaptively adjusted by the film thickness.

[0011] S3. Density Prediction and Scan Path Generation: S31, ultra-fast pre-scan, completes low-resolution scanning of the entire wafer under test at a speed of 500mm / s, outputs 256×256 low-resolution pre-scan image, single frame acquisition time is 10ms; adopts a simplified edge detection algorithm (Canny operator, threshold dynamic adjustment) to obtain preliminary defect candidate points.

[0012] S32. Particle density prediction: Input a 32×32 pixel pre-scanned image into the Transformer model. The Transformer model training data includes 1000 wafer images with different PR adhesive thicknesses, labeled with the actual particle density (0~50 particles / mm²). Output the predicted particle density ρ for each image patch and the high / medium / low density region segmentation results, where: High-density areas (ρ>10 areas / mm²): 512×512 resolution. Medium density zone (3≤ρ≤10 pixels / mm²): 256×256 resolution. Low-density areas (ρ<3 areas / mm²): 128×128 resolution, boundary values ​​can be dynamically adjusted according to process requirements.

[0013] S33. Generate an optimized scanning path based on the above particle density partitioning results: The scanning step size for high-density areas is 1 μm, and the dwell time is 20 μs. The scanning step size for the medium-dense region is 2 μm, and the dwell time is 10 μs. The scanning step size for the low-density region is 5 μm, and the dwell time is 5 μs. With a jump distance of ≤10mm between adjacent areas, the total scanning time of this optimized scanning path is reduced by ≥50% compared to the uniform scanning time, achieving ≤25 seconds for a 300mm wafer.

[0014] S34. Adaptive Path Execution and Image Acquisition: The top-level control unit sends the final scan path data to the six-degree-of-freedom fine-tuning platform, controlling the platform to move the wafer along this path. The light source is obliquely incident on the PR adhesive surface at a 15° incident angle, with a beam diameter of 2mm and a power density of 5mW / cm². A high-frame-rate CMOS sensor simultaneously acquires scattering images at multiple polarization angles, with a resolution of 2048×2048 and an exposure time of 5~20μs, adaptively adjusted according to the reflected light intensity. Dark field correction is used to eliminate the influence of dark current, wavelet transform is used to denoise the db4 wavelet, and three-layer decomposition is performed to complete the scattering image preprocessing for direct input to the polarization + morphology feature fusion module.

[0015] S4. Multimodal Feature Fusion Processing: A polarization channel is designed to extract light intensity fluctuation features under random polarization angles, such as the rate of change of light intensity between adjacent angles and polarization entropy. A morphology channel extracts geometric features such as the fractal dimension and edge gradient of particles, and features are fused through an adversarial feature fusion network.

[0016] Furthermore, the polarization fluctuation feature extraction specifically involves calculating σ, E, and ΔI using the light intensity data from all seven polarization angles at that location: Calculate the standard deviation of light intensity σ under multiple polarization angles to reflect the stability of polarization state. In the granular region, σ>0.15 (normalized value). Extracting polarization entropy ,in The light intensity percentage of a pixel at multiple angles, where the particle region E>0.6; Constructing the fluctuation characteristic matrix: ,in This represents the difference in light intensity between adjacent angles.

[0017] In this embodiment, adjacent angles refer to 7 randomly generated, non-repeating polarization angles (0~180°), numbered 1, 2, 3, 4, 5, 6, and 7 in the order of acquisition. The fluctuation feature matrix contains only 3 sets of ΔI, which strictly correspond to: ΔI 1-2 The intensity difference between the first polarization angle and the second polarization angle. ΔI 3-4 The intensity difference between the 3rd and 4th polarization angles. ΔI 5-6 The intensity difference between the 5th and 6th polarization angles. The 7th polarization angle is not involved in the ΔI calculation and serves as a redundancy check bit, primarily used to improve the calculation stability of σ and E and reduce the impact of random noise. The intervals of the 7 random angles are inherently non-uniform; pairing them in pairs is sufficient to cover a wide enough polarization interval range to capture the core wave characteristics of particle scattering. If all 6 adjacent differences (1-2, 2-3, 3-4, 4-5, 5-6, 6-7) are used, the feature dimension will double, resulting in a more than 30% increase in the inference time of the feature fusion network, which cannot meet the 28-second online detection requirement for a 300mm wafer.

[0018] In one possible implementation, if the number of random polarization angles n (5~8) is adjusted due to process requirements, the pairing rule of ΔI remains consistent: pair them in pairs according to the acquisition order, take all odd-even adjacent pairs in the first n-1 angles, and use the last angle only for the calculation of σ and E.

[0019] Example: When n=5 angles, the wave characteristic matrix is ​​[σ,E,ΔI]. 1-2 ,ΔI 3-4 ], Example: When n=8 angles, the wave characteristic matrix is ​​[σ,E,ΔI]. 1-2 ,ΔI 3-4 ,ΔI 5-6 ,ΔI 7-8 ].

[0020] Furthermore, the morphological feature extraction specifically involves using the multi-polarization scattering image at any polarization angle at that location for watershed segmentation, ensuring that the particle morphology is consistent across all polarization angles. Watershed segmentation is performed on the multi-polarization scattering image described in S3 to obtain the morphological feature parameters of the candidate regions, which are then normalized to obtain normalized morphological features. The fractal dimension D represents the surface roughness of the particles, where D > 1.2; The edge gradient change rate G represents the steepness of the transition between particles and the background, G>0.8; The area-to-perimeter ratio S / L represents the roundness of the particles; S / L > 0.1.

[0021] Furthermore, the adversarial feature fusion network fuses the fluctuation feature matrix with normalized morphological feature parameters. This network includes a generator network and a discriminator network. The generator network maps polarization and morphological features to a joint feature space with 128 dimensions. The discriminator network receives preset standard particle feature samples from a real particle feature sample library to distinguish real particle features from background interference features. Simultaneously, the discriminator network guides the generator network to optimize the fusion weights, and the final fused features are input into an adaptive threshold decision module.

[0022] When the polarization entropy E>0.7, the polarization feature weight is increased to 0.7, which is suitable for transparent particles; when the edge gradient change rate G>0.9, the morphological feature weight is increased to 0.6, which is suitable for opaque particles.

[0023] S5. Adaptive threshold decision and output detection results.

[0024] Based on the real-time measurement of PR adhesive thickness using a spectrometer, the defect judgment threshold is dynamically adjusted. When the thickness h < 1000 nm, the polarization entropy threshold (> 0.6) is used; when h ≥ 1000 nm, the light intensity fluctuation threshold (> 0.3) is used, replacing the fixed threshold judgment method and improving the detection adaptability of adhesive layers of different thicknesses.

[0025] Specifically: For thick adhesive regions (h≥1000nm), the light intensity fluctuation threshold is used, and particles are identified when σ>0.2 and S / L>0.12. The thin film region (h<1000nm) was identified as particles using the polarization entropy threshold. When E>0.65 and D>1.25, the particles were considered. At the same time, a time series consistency check is introduced. When the same position in three consecutive frames of images meets the judgment condition, it is confirmed as a real particle.

[0026] Result Quantification and Output: Particle parameter calculation: Based on sub-pixel edge detection, size measurement accuracy ±8nm, position accuracy ±1μm; Output format: Conforms to SEMIE152 standard, including: Wafer ID, PR adhesive thickness distribution, and inspection time; Particle list (coordinates, size, type (transparent / opaque), confidence level); Density heatmap (spatial resolution 1mm×1mm).

[0027] Furthermore, a system self-calibration mechanism is added, which uses standard calibration sheets (containing 30nm, 50nm, and 100nm standard particles) for daily calibration; the calibration parameters are polarization angle measurement deviation (≤1°) and light intensity response nonlinearity (≤2%); environmental compensation: the light source power is automatically corrected for every 1°C change in temperature (compensation coefficient 0.015 / °C).

[0028] The following example, using a 300mm wafer to be tested with a PR adhesive thickness of 800nm, further illustrates the process.

[0029] Test object: 300mm silicon-based wafer, with an 800nm ​​thick positive photoresist coating on the surface, containing 30~100nm transparent / opaque particles, with a higher density in the central 50mm area; System parameters: 380nm light source (suitable for thin film detection), 7 randomly generated polarization angles (23°, 51°, 77°, 103°, 136°, 162°, 178°), CMOS sensor resolution 2048×2048.

[0030] Testing process: 1. Fast pre-scan: The whole wafer low-resolution (256×256) scan is completed within 2 seconds, and the density prediction model identifies the central 50mm area as a high-density area (ρ≈12 / mm²).

[0031] 2. Adaptive scanning: Central high-density area: 512×512 resolution, 1μm step size, time taken 8.5 seconds; Low-density edge area: 128×128 resolution, 5μm step size, time taken 4.3 seconds; Total scan time: 22 seconds (51% more efficient than uniform scan).

[0032] 3. Feature processing: For the characteristics of 800nm ​​thin film, polarization entropy features are the main feature (weight 0.6), and morphological features such as fractal dimension are integrated.

[0033] Test results: 30nm transparent particle detection rate 98.3%, 50nm and above particle detection rate ≥99.5%; size measurement error ≤±7nm, position error ≤±0.8μm; accuracy 99.1%, false positive rate 0.4%, total detection time 28 seconds, meeting the online detection requirements of photolithography process.

[0034] All content not described in detail in this specification belongs to the prior art known to those skilled in the art. Furthermore, for those skilled in the art, there will be changes in specific implementation methods and application scope based on the ideas of this invention. Therefore, the content of this specification should not be construed as a limitation of this invention.

Claims

1. A method for detecting particles on the surface of PR adhesive based on random polarization scattering and multimodal feature extraction, characterized in that, Specifically, the steps include the following: S1. Pre-processing and parameter initialization of the wafer to be tested, and measuring the thickness of the PR adhesive; S2. Randomly generate n non-repeating polarization angles within the range of 0~180°. The angle sequence is randomly updated for each detection. Adjust the light source wavelength according to the PR adhesive thickness measured in S1. S3. Pre-scan the wafer under test. Predict the surface particle density distribution and divide it into regions based on the pre-scan image, and generate a density-adaptive scanning path. Move the wafer according to the path and simultaneously acquire multi-polarization scattering images of different density regions at corresponding resolutions. S4. Multimodal feature fusion processing: The polarization channel is designed to extract the polarization features of the scattering image obtained in S3 at random polarization angles, and the morphology channel is designed to extract the morphological features of particles. The features are then fused through an adversarial training mechanism. S5. Adaptive threshold decision and output detection results.

2. The method for detecting particles on the surface of PR adhesive based on random polarization scattering and multimodal feature extraction according to claim 1, characterized in that, Adjust the light source wavelength according to the thickness of the PR adhesive: use 380nm when the thickness is <1000nm, and use 650nm when the thickness is ≥1000nm.

3. The method for detecting particles on the surface of PR adhesive based on random polarization scattering and multimodal feature extraction according to claim 1, characterized in that, The specific definition of particle density partitioning in S3 is as follows: High-density areas (ρ>10 areas / mm²): 512×512 resolution, 1μm scan step, 20μs dwell time; Medium-dense area (3≤ρ≤10 cells / mm²): 256×256 resolution, 2μm scan step, 10μs dwell time; Low-density areas (ρ<3 areas / mm²): 128×128 resolution, 5μm scan step, 5μs dwell time; The jump distance between adjacent areas is ≤10mm.

4. The method for detecting particles on the surface of PR adhesive based on random polarization scattering and multimodal feature extraction according to claim 1, characterized in that, The synchronous acquisition of multi-polarization scattering images specifically involves: the top-level control unit sending the final scanning path data to the six-degree-of-freedom fine-tuning platform, controlling the six-degree-of-freedom fine-tuning platform to drive the wafer movement along this path; the light source obliquely incident on the PR adhesive surface at a 15° incident angle, with a beam diameter of 2mm and a power density of 5mW / cm²; the CMOS sensor synchronously acquires scattering images at multiple polarization angles, with a resolution of 2048×2048 and an exposure time of 5~20μs, adaptively adjusted according to the reflected light intensity; and employing dark field correction to eliminate the influence of dark current, wavelet transform to denoise the db4 wavelet, and three-layer decomposition to complete the scattering image preprocessing, so that it can be directly input into the polarization + morphological feature fusion module.

5. The method for detecting particles on the surface of PR adhesive based on random polarization scattering and multimodal feature extraction according to claim 1, characterized in that, The polarization characteristics include calculating the standard deviation of light intensity σ and polarization entropy at multiple polarization angles. ,in Given the light intensity percentage of a pixel at multiple angles, with the particle region E > 0.6, a fluctuation feature matrix is ​​constructed: ,in This represents the difference in light intensity between adjacent angles.

6. The method for detecting particles on the surface of PR adhesive based on random polarization scattering and multimodal feature extraction according to claim 1, characterized in that, The morphological feature extraction specifically involves performing watershed segmentation and normalization on a multi-polarization state scattering image at any polarization angle at that location to obtain normalized morphological features: Fractal dimension D, particle surface roughness, D>1.2; Edge gradient change rate G, the steepness of the transition between particles and background, G>0.8; Area-to-perimeter ratio S / L, particle roundness, S / L>0.

1.

7. The method for detecting particles on the surface of PR adhesive based on random polarization scattering and multimodal feature extraction according to claim 5 or 6, characterized in that, The adversarial training mechanism utilizes an adversarial feature fusion network, which includes a generator network and a discriminator network. The generator network maps polarization features and morphological features to a joint feature space. The discriminator network is input with preset standard particle feature samples from a real particle feature sample library to distinguish real particle features from background interference features. Simultaneously, the discriminator network guides the generator network to optimize the fusion weights, and the final fused features are input into an adaptive threshold decision module.

8. The method for detecting particles on the surface of PR adhesive based on random polarization scattering and multimodal feature extraction according to claim 7, characterized in that, When the polarization entropy E>0.7, the polarization feature weight is increased to 0.7, which is suitable for transparent particles; when the edge gradient change rate G>0.9, the morphological feature weight is increased to 0.6, which is suitable for opaque particles.

Images