Wafer defect multi-angle equivalent detection method based on rotation invariance of light
By leveraging the rotational invariance of light, combined with physical modeling and deep learning algorithms, multi-angle equivalent detection of domestically produced wafer inspection equipment has been achieved. This solves the problems of missed detection of tiny particles and false detection of deep defects, reducing costs and improving detection accuracy and generalization ability.
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- NORTHEASTERN UNIV AT QINHUANGDAO
- Filing Date
- 2026-01-30
- Publication Date
- 2026-05-15
AI Technical Summary
Existing domestic wafer inspection equipment suffers from a single oblique incidence angle, resulting in a high rate of missed detection for tiny particles and a high rate of false detection for deep defects. Foreign equipment is too expensive, and existing algorithms lack physical constraints, leading to poor generalization ability.
Based on the rotation invariance of light, a mapping relationship between the oblique incident angle and the intensity of scattered light is established through physical modeling. A multi-angle equivalent detection method is designed, and MobileNetV2 and E(2)-CNN network models are adopted. Combined with data augmentation and multi-task learning, multi-angle detection without hardware modification is achieved.
It achieves high-precision, low-cost multi-angle detection, with a recall rate of 98% for microparticles, 90% for deep defects, and a false detection rate of 2%, meeting industrial throughput requirements and adapting to industrial detector noise and polarization errors.
Smart Images
Figure FT_1 
Figure FT_2 
Figure FT_3
Abstract
Description
Technical Field
[0001] This invention relates to the field of wafer defect detection technology, and in particular to a multi-angle equivalent detection method for wafer defects based on the rotational invariance of light. Background Technology
[0002] In semiconductor wafer defect detection, the angle of the oblique incident light source is a core parameter: small angles (15°-30°) are suitable for detecting tiny particles of 20-200nm, large angles (60°-75°) are suitable for detecting deep defects, and medium angles (45°) can only be used in certain scenarios. Existing domestic inspection equipment (such as the FSD series from Zhongke Feice) is limited by hardware and only supports a single 45° angle; foreign equipment (such as KLA) adopts a multi-angle design of 15°-75°, but its cost is 3-5 times that of domestic equipment. Among the existing optimization solutions, hardware modifications require adjustments to the mechanical structure, which is costly; pure data-driven algorithms lack physical constraints, have poor generalization ability, and cannot adapt to industrial noise.
[0003] Existing technological defects and their causes:
[0004] (1) Domestic equipment: The single 45° angle leads to a false detection rate of more than 30% for small particles and a false detection rate of more than 10% for deep defects. The root cause is that the hardware design does not cover multiple incident angles.
[0005] (2) Foreign equipment: High cost, difficult for domestic enterprises to use on a large scale;
[0006] (3) Existing algorithms: Pure data-driven models have no physical constraints and are not good at generalizing the scattering characteristics at different angles. Summary of the Invention
[0007] To address the shortcomings of existing technologies, a multi-angle equivalent detection method for wafer defects based on the rotation invariance of light is proposed. This method solves the problems of missed detection of tiny particles and false detection of deep defects caused by the single oblique incident angle (only 45°) of domestic wafer inspection equipment. At the same time, it avoids the high cost of multi-angle equipment and realizes multi-angle equivalent detection without hardware modification.
[0008] On one hand, the present invention provides a method for multi-angle equivalent detection of wafer defects based on the rotational invariance of light, comprising the following steps:
[0009] Step 1: Perform physical modeling based on the variation law of scattered light intensity generated by wafer defects under different oblique incident angles after incident light irradiates the wafer;
[0010] Specifically, based on Rayleigh scattering Mie scattering Establish a mapping relationship between the oblique incident angle and the intensity of the scattered light, where I R For Rayleigh scattering intensity, I Mλ is the intensity of the Mie scattering light, λ is the wavelength of the incident light, n is the refractive index of the wafer material, d is the defect size, θ is the oblique angle of incidence between the incident light and the wafer surface, and S1(θ) and S2(θ) are the Mie scattering amplitude functions.
[0011] Step 2: Construct a wafer defect image dataset;
[0012] Specifically, the dataset is established by collecting real wafer defect images at an oblique incidence angle of 45° between the incident light and the wafer surface, and expanding the dataset by generating multi-angle wafer defect images through simulation to obtain a wafer defect image dataset.
[0013] Step 3: Design two parallel and independent detection algorithm schemes, respectively adapted to the two types of industrial scenarios of "real-time high throughput" and "high-precision detection", and construct the corresponding network models; specifically including:
[0014] Solution 1: Real-time detection of 45° real wafer defect images: Taking the projection point of the incident light on the wafer surface as the origin, the Cartesian coordinates (x, y) of the image are transformed into polar coordinates (ρ, φ), where ρ is the distance from the pixel to the origin, and φ is the angle between the line connecting the pixel and the origin and the reference axis; then the angle φ is normalized and its value range is limited to [0°, 180°] to achieve a unified representation of scattering characteristics at different angles;
[0015] The polar coordinate transformed image is input into the MobileNetV2 network; the MobileNetV2 network contains 13 inverted residual blocks, and the rotation-invariant features of wafer defects are extracted through the network's depthwise convolution operation; the network ends with a global average pooling layer to compress and integrate the features, outputting a feature vector with uniform dimensionality; physical constraint loss is introduced. L p The physical constraint loss is used to quantify the difference between the scattered light intensity predicted by the model and the theoretical scattered light intensity based on physical modeling; h is the loss weight coefficient, y is the loss calculation term, and I(θ) is the theoretical scattered light intensity corresponding to the oblique incident angle θ. model (θ) represents the intensity of scattered light corresponding to the same oblique incident angle θ predicted by the network; To indicate the summation sign, the total loss function is cross-entropy loss + 0.3 × physical constraint loss L. p The total loss is minimized through backpropagation, and the defect type, size, and confidence level are finally output.
[0016] Scheme 2: Construct an E(2)-CNN model, specifically including: the input layer receives a 45° real wafer defect image and outputs the original image data to the equivariant convolutional layer; after the equivariant convolutional layer inputs the original image, it first undergoes a 0° / 90° / 180° / 270° transformation through the rotation group E(2), and then extracts features through convolution operations, maintaining the equivariance of "input rotation θ → output synchronous rotation θ"; at the same time, the features are decomposed into radial features r and angular features p, and the decomposed feature map is output to the feature coupling layer; the feature coupling layer inputs radial features r and angular features p, dynamically allocates weights through the attention mechanism, strengthens key features, suppresses noise, and outputs the fused high-dimensional features to the voting fusion classification head; after the voting fusion classification head inputs the high-dimensional features, it classifies the defects corresponding to each angle from 15° to 75° through multiple parallel branches, and then integrates the results of each branch through the voting fusion strategy, selecting the classification result with the highest support as the final output; the output layer outputs the defect type, size and confidence.
[0017] Step 4: Perform data augmentation and multi-task learning: For the two parallel and independent detection algorithm schemes in Step 3, train the corresponding network models respectively to achieve equivalent detection of wafer defects from multiple angles.
[0018] Step 4.1: Pre-training stage: Use multi-angle wafer defect images to pre-train the two network models respectively; enable the network models to learn the mapping relationship between angle and scattering features;
[0019] Step 4.2: Fine-tuning stage: After pre-training, fine-tune the two network models using real wafer defect images to adapt them to the hardware noise of the device.
[0020] Step 4.3: Training parameters: The Adam optimizer is used uniformly, with β1=0.5 and β2=0.999, where β1 is the exponential decay rate of the first-order momentum estimate and β2 is the exponential decay rate of the second-order momentum estimate. The learning rate is 0.0002, the batch size is 8, and the total number of training epochs is 200.
[0021] Step 4.4: Synchronization optimization operation, specifically including:
[0022] (1) Data augmentation: Affine transformations are performed in parallel during training, including random rotation, scaling, translation, horizontal and vertical flipping, to expand the range of data distribution;
[0023] (2) Multi-task learning: Defect classification task and angle regression task are executed in parallel. The defect classification task allows the network model to identify the defect type, and the angle regression task allows the network model to learn the optimal oblique incidence angle corresponding to the defect.
[0024] Step 4.5: After pre-training, fine-tuning and synchronous optimization, two trained network models are output respectively, namely the real-time detection model and the high-precision detection model, which can be called on demand in industrial scenarios to realize the equivalent detection of wafer defects from multiple angles.
[0025] On the other hand, this application proposes a computer-readable storage medium storing executable instructions that, when executed, cause a processor to perform the aforementioned method for multi-angle equivalent detection of wafer defects based on the rotational invariance of light.
[0026] Thirdly, this application proposes a computer program product, including a computer program or instructions, which, when executed by a processor, implements the aforementioned method for multi-angle equivalent detection of wafer defects based on the rotational invariance of light.
[0027] The beneficial effects of adopting the above technical solution are as follows:
[0028] This invention provides a multi-angle equivalent detection method for wafer defects based on the rotational invariance of light, which has the following advantages compared with the prior art:
[0029] (1) No hardware modification is required; it is compatible with existing domestic equipment, and the cost is less than 1 / 10 of the cost of existing equipment.
[0030] (2) Detection performance: accuracy > 95%, micro-particle recall rate > 98%, deep defect recall rate > 90%, false detection rate < 2%, reaching the level of foreign equipment;
[0031] (3) Generalization ability: physical constraints + rotation invariant feature extraction, adapting to industrial detector noise and polarization error;
[0032] (4) Real-time performance: The inference time for a single image is <5ms, which meets the industrial throughput of 300 images / hour. Attached Figure Description
[0033] Figure 1 Overall flowchart of the multi-angle equivalent detection method for wafer defects according to an embodiment of the present invention;
[0034] Figure 2 Schematic diagram of oblique incidence angle and scattered light intensity distribution in an embodiment of the present invention;
[0035] Figure 3 Polar coordinate transformation + CNN algorithm architecture diagram of this invention embodiment;
[0036] Figure 4 The architecture diagram of the E(2)-CNN algorithm in this embodiment of the invention;
[0037] Figure 5 Flowchart of model training in an embodiment of the present invention. Detailed Implementation
[0038] The specific implementation methods of this application will be further described in detail below with reference to the accompanying drawings and embodiments.
[0039] Example 1:
[0040] On the one hand, this invention provides a method for multi-angle equivalent detection of wafer defects based on the rotational invariance of light, such as... Figure 1 As shown, it includes the following steps:
[0041] Step 1: Perform physical modeling based on the variation law of scattered light intensity generated by wafer defects under different oblique incident angles after incident light irradiates the wafer;
[0042] Specifically, based on Rayleigh scattering Mie scattering A mapping relationship between the oblique incident angle and the scattered light intensity is established, with fixed parameters λ = 266nm / 355nm, n = 1.5, and d = 50nm / 5μm, where I R Rayleigh scattering intensity is a term specifically used to characterize the intensity of scattered light from tiny particle defects in the 20-200 nm range. M λ is the Mie scattering intensity, specifically characterizing the scattered light intensity of deep defects (such as 5μm scratches and 50nm voids). λ is the incident light wavelength, with 266nm specifically adapted for microparticle detection and 355nm specifically adapted for deep defect detection. n is the refractive index of the wafer material, fixed at 1.5, a common refractive index for semiconductor wafers. d is the defect size, with 50nm corresponding to the magnitude of microparticle defects and 5μm corresponding to the magnitude of larger defects (scratches and voids). θ is the oblique incident angle between the incident light and the wafer surface, with a value range covering 15°-75°, including the fixed 45° of existing domestic equipment. S1(θ) and S2(θ) are the Mie scattering amplitude functions, which are directly related to the incident angle θ and are the core parameters for calculating the scattered light intensity distribution of deep defects.
[0043] Step 2: Construct a wafer defect image dataset;
[0044] Specifically, a dataset is established by acquiring real wafer defect images when the incident light is incident at an oblique angle of 45° to the wafer surface. The oblique angle of incident light to the wafer surface is the fixed incident angle of existing domestic inspection equipment, and it is also the incident angle benchmark for acquiring real wafer defect images in this invention. The dataset is expanded by generating multi-angle wafer defect images through simulation to obtain a wafer defect image dataset.
[0045] In this embodiment, 50,000 real wafer defect images (including 20nm particles, 5μm scratches, etc.) were collected, and 250,000 multi-angle images at 15°, 30°, 60°, and 75° were generated and expanded.
[0046] Step 3: Design two parallel and independent detection algorithm schemes, respectively adapted to two industrial scenarios: "real-time high throughput" and "high-precision detection," and construct corresponding network models; users can flexibly choose according to actual production needs. The two schemes have no sequential execution relationship and complete the detection task independently. Specifically, this includes:
[0047] Solution 1: Adapted to "real-time high throughput" industrial scenarios; Real-time detection of 45° real wafer defect images (512×512 pixels): Taking the projection point of the incident light on the wafer surface as the origin, the Cartesian coordinates (x,y) of the image are transformed into polar coordinates (ρ,φ), where ρ is the distance from the pixel to the origin, and φ is the angle between the line connecting the pixel and the origin and the reference axis; then the angle φ is normalized and its value range is limited to [0°,180°] to achieve a unified representation of scattering characteristics at different angles and eliminate the interference of angle offset on the detection results.
[0048] The polar coordinate transformed image is input into the MobileNetV2 network. The MobileNetV2 network contains 13 inverted residual blocks. Through depthwise convolution operations, rotation-invariant features of wafer defects are extracted, including key information such as defect shape and scattered light intensity distribution. At the network's end, a global average pooling layer compresses and integrates the features, outputting a feature vector with uniform dimensionality. A physical constraint loss is introduced. L p The physical constraint loss is used to quantify the difference between the scattered light intensity predicted by the model and the theoretical scattered light intensity based on physical modeling; h is the loss weight coefficient, set to 0.3 in this scheme, used to balance the proportion of physical constraint loss and classification loss; y is the loss calculation term, obtained by accumulating the light intensity differences at different angles; I(θ) is the theoretical scattered light intensity corresponding to the oblique incident angle θ. model (θ) represents the intensity of scattered light corresponding to the same oblique incident angle θ predicted by the network; To determine the summation sign, the light intensity differences covering the entire 15°-75° angle range are accumulated; the total loss function is "cross-entropy loss (i.e., optimization defect classification) + 0.3 × physical constraint loss L". p The total loss is minimized through backpropagation, and the defect type, size, and confidence level are finally output. The inference time for a single image is <5ms, which meets the industrial throughput requirement of 300 pieces / hour.
[0049] Option 2: Adapting to the "high-precision inspection" industrial scenario: Constructing an E(2)-CNN model, specifically including: The input layer receives a 45° real wafer defect image and outputs the original image data to the equivariant convolutional layer; After the equivariant convolutional layer inputs the original image, it first undergoes a 0° / 90° / 180° / 270° transformation through the rotation group E(2), and then extracts features through convolution operations, maintaining the equivariance of "input rotation θ → output synchronous rotation θ"; At the same time, the features are decomposed into radial features r (reflecting the distance from the defect to the origin and associated with scattering angle information) and angular features p (reflecting the incident angle information), and the output is divided into radial features r (reflecting the distance from the defect to the origin and associated with scattering angle information) and angular features p (reflecting the incident angle information), and the output is divided into radial features r (reflecting the distance from the defect to the origin and associated with scattering angle information) and angular features p (reflecting the incident angle information). The solved feature map is fed to the feature coupling layer; the feature coupling layer takes radial feature r and angular feature p as input, dynamically assigns weights through an attention mechanism to strengthen key features and suppress noise, and outputs the fused high-dimensional features to the voting fusion classification head; after the high-dimensional features are input to the voting fusion classification head, the defects corresponding to each angle from 15° to 75° are classified through multiple parallel branches, and then the results of each branch are integrated through a voting fusion strategy, and the classification result with the highest support is selected as the final output; the output layer outputs the defect type, size and confidence, ensuring that the recall rate of small particles is >98% and the recall rate of deep defects is >90%.
[0050] Step 4: Perform data augmentation and multi-task learning: For the two parallel and independent detection algorithm schemes in Step 3, train the corresponding network models respectively (Model 1: polar coordinate transformation + MobileNetV2 network; Model 2: E(2)-CNN network) to achieve multi-angle equivalent detection of wafer defects.
[0051] Step 4.1: Pre-training stage: Using multi-angle wafer defect images, in this embodiment, 250,000 multi-angle simulated wafer defect images (i.e., simulated images at 15°, 30°, 60°, and 75°) generated in step 2 are used to pre-train the two network models respectively; the core objective is to enable the network models to learn the mapping relationship between angle and scattering features, and to initially grasp the characteristic representation rules of wafer defects at different angles.
[0052] Step 4.2: Fine-tuning stage: After pre-training is completed, real wafer defect images are used. In this embodiment, 50,000 real wafer defect images at 45° collected in step 2 are used to fine-tune the two network models respectively. The core objective is to make the network model adapt to the hardware noise of existing domestic equipment (such as detector noise and polarization error) and improve the generalization ability of the model in real industrial scenarios.
[0053] Step 4.3: Training parameters: The Adam optimizer is used uniformly, with β1=0.5 and β2=0.999, where β1 is the exponential decay rate of the first-order momentum estimate and β2 is the exponential decay rate of the second-order momentum estimate. The learning rate is 0.0002, the batch size is 8, and the total number of training epochs is 200.
[0054] Step 4.4: Synchronization optimization operation, specifically including:
[0055] (1) Data augmentation: Affine transformations are performed in parallel during training, including random rotation (0°-360°), scaling (0.8-1.2 times), translation (±10 pixels), horizontal and vertical flipping, to expand the range of data distribution and reduce model overfitting;
[0056] (2) Multi-task learning: Defect classification task and angle regression task are executed in parallel. The defect classification task allows the network model to identify the defect type, and the angle regression task allows the network model to learn the best oblique incidence angle corresponding to the defect. Through the dual-task collaborative optimization, the model's ability to understand defect features is improved.
[0057] Step 4.5: After pre-training, fine-tuning and synchronous optimization, two trained network models are output respectively, namely the real-time detection model and the high-precision detection model, which can be called on demand in industrial scenarios to realize the equivalent detection of wafer defects from multiple angles.
[0058] like Figure 2 As shown, the light intensity of Rayleigh scattering (microparticles) and Mie scattering (deep defects) varies with the incident angle of 15°-75°, labeled "sensitive area for microparticles (15°-30°)" and "sensitive area for deep defects (60°-75°)".
[0059] like Figure 3 The diagram illustrates the module flow of "input layer → preprocessing module → feature extraction → physical constraints → output layer".
[0060] like Figure 4 The diagram illustrates the module flow of "input layer → isovariant convolutional layer → feature coupling layer → classification head → output layer".
[0061] like Figure 5 As shown, the training steps are "dataset preparation → pre-training → optimization → fine-tuning → output model".
[0062] Example 2: Real-time detection scheme using polar coordinate transformation + CNN:
[0063] Step A1: Physical Modeling
[0064] With λ=266nm (for detecting tiny particles), calculate the relative intensity of Rayleigh scattering light corresponding to 15°-75°.
[0065] Step A2: Data Preparation
[0066] 50,000 real images at 45° (including 20,000 images containing 20nm particles) were collected, and 250,000 images at 15°-75° were generated through simulation.
[0067] Step A3: Algorithm Implementation:
[0068] (1) Polar coordinate transformation: Using the projection point of the light source as the origin, the 512×512 image is converted into polar coordinates;
[0069] (2) Network setup: MobileNetV2 (13 inverted residual blocks) + global average pooling;
[0070] (3) Loss function: Cross-entropy loss + physical constraint loss with a weight of 0.3;
[0071] Step A4: Training:
[0072] Adam optimizer, learning rate 0.0002, batch size 8, training for 200 epochs, data augmentation using affine transformation;
[0073] Step A5: Detection:
[0074] Input a 45° image, output defect type + size, inference time 3.2ms.
[0075] Example 3: High-precision detection scheme of E(2)-CNN;
[0076] Step B1: Physical modeling;
[0077] With λ=355nm fixed (for detecting deep defects), calculate the relative intensity of Mie scattering light corresponding to 15°-75°.
[0078] Step B2: Data preparation;
[0079] Collect 50,000 real images at 45° (including 10,000 images with 50nm holes), and simulate to generate 250,000 images at 15°-75°.
[0080] Step B3: Algorithm Implementation;
[0081] (1) Isovariant convolutional layer: rotation group E(2) (0° / 90° / 180° / 270°);
[0082] (2) Feature coupling layer: decompose radial / angular features and weight them with attention mechanism;
[0083] (3) Classification Header: Multi-angle classification + voting fusion;
[0084] Step B4: Training;
[0085] Adam optimizer, learning rate 0.0002, batch size 8, training for 200 epochs, multi-task learning (defect classification + angle regression).
[0086] Step B5: Detection;
[0087] Input a 45° image, output defect type + size, with an accuracy of 97.1%.
[0088] Example 4: Adaptation schemes for different defect scenarios:
[0089] Scenario: A wafer containing both 20nm particles and 5μm scratches;
[0090] In this embodiment, the fusion inference of polar coordinate transformation + CNN and E(2)-CNN is selected;
[0091] Training parameters: learning rate 0.0001, batch size 4, training time 250 epochs;
[0092] Results: 98.5% recall rate for small particles, 92.5% recall rate for scratches, and 1.7% false positive rate.
[0093] Example 5:
[0094] This embodiment proposes a computer-readable storage medium that stores executable instructions. When these instructions are executed, if they are implemented as software functional units and sold or used as independent products, they can be stored in a computer-readable storage medium.
[0095] The computer software product is stored in a storage medium and includes several instructions to cause a computer device (which may be a personal computer, a server, or a network device, etc.) to execute all or part of the steps of the wafer defect multi-angle equivalent detection method based on the rotation invariance of light described in various embodiments of this application.
[0096] The aforementioned storage media include: flash memory, hard disk, multimedia card, card-type memory (e.g., SD (Secure Digital Memory Card) or DX (Memory Data Register, MDR) memory, random access memory (RAM), static random access memory (SRAM), read-only memory (ROM), electrically erasable programmable read-only memory (EEPROM), programmable read-only memory (PROM), magnetic memory, disk, optical disk, server, APP (Application) application store, and other media capable of storing program verification codes. These media store computer programs, and when executed by a processor, they can implement the various steps of the aforementioned method for multi-angle equivalent detection of wafer defects based on the rotational invariance of light.
[0097] Example 6:
[0098] This embodiment proposes a computer program product, including a computer program or instructions, which, when executed by a processor, implements the aforementioned method for multi-angle equivalent detection of wafer defects based on the rotational invariance of light.
[0099] Based on this understanding, the technical solution of this application, in essence, or the part that contributes to the prior art, or part of the technical solution, can be embodied in the form of a computer program product.
[0100] The various embodiments in this application are described in a progressive manner. The same or similar parts between the various embodiments can be referred to each other. Each embodiment focuses on describing the differences from other embodiments.
[0101] The scope of protection of this application is not limited to the embodiments described above. Obviously, those skilled in the art can make various modifications and variations to this disclosure without departing from the scope and spirit of this disclosure. If such modifications and variations fall within the scope of the methods disclosed herein and their equivalents, then the intent of this disclosure also includes such modifications and variations.
Claims
1. A method for multi-angle equivalent detection of wafer defects based on the rotational invariance of light, characterized in that, Includes the following steps: Step 1: Perform physical modeling based on the variation law of scattered light intensity generated by wafer defects under different oblique incident angles after incident light irradiates the wafer; Step 2: Construct a wafer defect image dataset; Step 3: Design two parallel and independent detection algorithm schemes to adapt to the two types of industrial scenarios of "real-time high throughput" and "high-precision detection" respectively, and build the corresponding network models; Step 4: Perform data augmentation and multi-task learning: For the two parallel and independent detection algorithm schemes in Step 3, train the corresponding network models respectively to achieve equivalent detection of wafer defects from multiple angles.
2. The method for multi-angle equivalent detection of wafer defects based on the rotational invariance of light according to claim 1, characterized in that, Step 1 specifically involves: based on Rayleigh scattering Mie scattering Establish a mapping relationship between the oblique incident angle and the intensity of the scattered light, where I R For Rayleigh scattering intensity, I M λ is the intensity of the Mie scattering light, λ is the wavelength of the incident light, n is the refractive index of the wafer material, d is the defect size, θ is the oblique angle of incidence between the incident light and the wafer surface, and S1(θ) and S2(θ) are the Mie scattering amplitude functions.
3. The method for multi-angle equivalent detection of wafer defects based on the rotational invariance of light according to claim 1, characterized in that, Step 2 specifically involves: acquiring real wafer defect images at an oblique incidence angle of 45° between the incident light and the wafer surface to establish a dataset, and expanding the dataset by generating multi-angle wafer defect images through simulation to obtain a wafer defect image dataset.
4. The method for multi-angle equivalent detection of wafer defects based on the rotational invariance of light according to claim 1, characterized in that, Step 3 specifically includes: Solution 1: Real-time detection of 45° real wafer defect images: Taking the projection point of the incident light on the wafer surface as the origin, the Cartesian coordinates (x, y) of the image are transformed into polar coordinates (ρ, φ), where ρ is the distance from the pixel to the origin, and φ is the angle between the line connecting the pixel and the origin and the reference axis; then the angle φ is normalized and its value range is limited to [0°, 180°] to achieve a unified representation of scattering characteristics at different angles; The polar coordinate transformed image is input into the MobileNetV2 network; the MobileNetV2 network contains 13 inverted residual blocks, and the rotation-invariant features of wafer defects are extracted through the network's depthwise convolution operation; the network ends with a global average pooling layer to compress and integrate the features, outputting a feature vector with uniform dimensionality; physical constraint loss is introduced. L p The physical constraint loss is used to quantify the difference between the scattered light intensity predicted by the model and the theoretical scattered light intensity based on physical modeling; h is the loss weight coefficient, y is the loss calculation term, and I(θ) is the theoretical scattered light intensity corresponding to the oblique incident angle θ. model (θ) represents the intensity of scattered light corresponding to the same oblique incident angle θ predicted by the network; To indicate the summation sign, the total loss function is cross-entropy loss + 0.3 × physical constraint loss L. p The total loss is minimized through backpropagation, and the defect type, size, and confidence level are finally output. Scheme 2: Construct an E(2)-CNN model, specifically including: the input layer receives a 45° real wafer defect image and outputs the original image data to the equivariant convolutional layer; after the equivariant convolutional layer inputs the original image, it first undergoes a 0° / 90° / 180° / 270° transformation through the rotation group E(2), and then extracts features through convolution operations, maintaining the equivariance of "input rotation θ → output synchronous rotation θ"; at the same time, the features are decomposed into radial features r and angular features p, and the decomposed feature map is output to the feature coupling layer; the feature coupling layer inputs radial features r and angular features p, dynamically allocates weights through the attention mechanism, strengthens key features, suppresses noise, and outputs the fused high-dimensional features to the voting fusion classification head; after the voting fusion classification head inputs the high-dimensional features, it classifies the defects corresponding to each angle from 15° to 75° through multiple parallel branches, and then integrates the results of each branch through the voting fusion strategy, selecting the classification result with the highest support as the final output; the output layer outputs the defect type, size and confidence.
5. The method for multi-angle equivalent detection of wafer defects based on the rotational invariance of light according to claim 1, characterized in that, Step 4 specifically includes the following steps: Step 4.1: Pre-training stage: Use multi-angle wafer defect images to pre-train the two network models respectively; enable the network models to learn the mapping relationship between angle and scattering features; Step 4.2: Fine-tuning stage: After pre-training, fine-tune the two network models using real wafer defect images to adapt them to the hardware noise of the device. Step 4.3: Training parameters: The Adam optimizer is used uniformly, with β1=0.5 and β2=0.999, where β1 is the exponential decay rate of the first-order momentum estimate and β2 is the exponential decay rate of the second-order momentum estimate. The learning rate is 0.0002, the batch size is 8, and the total number of training epochs is 200. Step 4.4: Synchronization optimization operation, specifically including: (1) Data augmentation: Affine transformations are performed in parallel during training, including random rotation, scaling, translation, horizontal and vertical flipping, to expand the range of data distribution; (2) Multi-task learning: Defect classification task and angle regression task are executed in parallel. The defect classification task allows the network model to identify the defect type, and the angle regression task allows the network model to learn the optimal oblique incidence angle corresponding to the defect. Step 4.5: After pre-training, fine-tuning and synchronous optimization, two trained network models are output respectively, namely the real-time detection model and the high-precision detection model, which can be called on demand in industrial scenarios to realize the equivalent detection of wafer defects from multiple angles.
6. A computer-readable storage medium, characterized in that, It stores executable instructions that, when executed, cause the processor to perform the wafer defect multi-angle equivalent detection method based on the rotational invariance of light as described in any one of claims 1-5.
7. A computer program product, characterized in that, Includes a computer program or instructions that, when executed by a processor, implement the multi-angle equivalent detection method for wafer defects based on the rotational invariance of light as described in any one of claims 1-5.