A wafer hybrid defect identification method based on a dual-branch dynamic tag sensing network
By using a dual-branch dynamic label perception network, combined with convolutional neural networks and lightweight Transformer networks, high-precision, lightweight, and real-time identification of mixed defects in wafers is achieved. This addresses the shortcomings of existing technologies in dynamic adaptive attention weight adjustment and defect coexistence correlation modeling, improving identification accuracy and meeting the real-time inspection requirements of semiconductor production lines.
Patent Information
- Application Number
- CN202611244208.7
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2026-08-17
- Publication Date
- 2026-11-17
AI Technical Summary
Existing wafer hybrid defect identification methods have shortcomings in dynamic adaptive attention weight adjustment, defect coexistence correlation modeling, and local and global feature extraction, and cannot meet the requirements of high-precision, lightweight, and real-time identification under advanced processes.
A method based on a dual-branch dynamic label perception network is adopted, which extracts local and global features in parallel through convolutional neural networks and lightweight Transformer networks. Combined with a dynamic adaptive attention mechanism and a label association inference module, adaptive feature enhancement and association modeling for multiple defect types are achieved.
It significantly improves the accuracy of mixed defect identification, reduces the false negative and false positive rates, meets the real-time deployment requirements of low-power embedded edge detection in semiconductor production lines, and has good process interpretability.
Smart Images

Figure CN123118484A_ABST
Abstract
Description
Technical Field
[0001] This application belongs to the field of semiconductor manufacturing technology, specifically relating to a wafer hybrid defect identification method based on a dual-branch dynamic tag sensing network. Background Technology
[0002] As the core substrate for chip manufacturing, wafer surface defect detection is a crucial step in ensuring semiconductor product quality and improving chip manufacturing yield. Wafer maps, composed of characteristic patterns formed by the spatial distribution of electrical defects in bare wafers, can intuitively reflect process anomalies during wafer manufacturing and are a core technical means for troubleshooting equipment malfunctions and optimizing manufacturing processes. In recent years, with the continuous evolution of integrated circuit process nodes towards 5nm and below, the complexity of chip manufacturing processes has increased significantly. The proportion of mixed samples with two or more defects on a single wafer has continued to rise, making high-precision automatic identification of mixed defects a rigid requirement for semiconductor production lines. Traditional single-defect detection methods are insufficient to handle complex and varied defect morphologies, urgently requiring intelligent identification technology that can simultaneously and accurately identify multiple defect types and possesses strong robustness.
[0003] Among them, the wafer hybrid defect identification method aims to achieve simultaneous detection and classification of multiple defect types on the same wafer image, representing a deep cross-disciplinary integration of computer vision and semiconductor manufacturing. This type of method uses deep learning models to extract features and perform semantic analysis on the wafer image, outputting the presence and spatial location results of various defects, providing process engineers with decision-making basis for defect warnings and root cause analysis.
[0004] However, systematic retrieval and analysis revealed that existing wafer hybrid defect identification methods still have inherent flaws and are difficult to meet the requirements of industrial applications under advanced processes. A hybrid defect identification method using selective interactive modeling of wafer feature blocks (publication number CN121937459A) enhances the global correlation modeling capability of high-response defect regions through response-driven pixel-level priority sorting and cross-block co-position pixel hierarchical interaction mechanisms. However, this technical solution uses a fixed-order multi-head attention mechanism for selective interactive calculation, which cannot dynamically adjust attention weights according to the input wafer defect morphology, resulting in a static and fixed attention operation process. Furthermore, although this solution includes a multi-label decomposition and identification module that outputs the existence determination results of each basic defect category, the defect categories are still predicted independently, failing to model the natural coexistence patterns of defects brought about by semiconductor processes and severely ignoring label correlation. In addition, after feature block division, local modeling is performed separately, lacking a complete feature extraction architecture that takes into account both large-scale global defects and small local defects. Another publication, CN121810704A, describes a method for decoupling and identifying mixed defects in high-noise wafer images based on defect prototype inversion. This method achieves spatial decoupling and reliable identification of mixed defects in high-noise environments through grain diagram structure modeling, multi-scale structural consistency scoring functions, and defect prototype inversion mechanisms. However, this approach, based on structural graph representation and node embedding, primarily focuses on local neighborhood statistical information, making it difficult to fully capture large-scale global defect features such as edge rings and full coverage. The construction of defect prototype representations relies on static structural consistency scoring, failing to dynamically adjust feature enhancement strategies according to different defect morphologies. Although decoupling and identification of mixed defects are achieved, independent prototype matching mechanisms are still used between different defect types, without introducing prior knowledge of label coexistence, making it impossible to distinguish visually similar mixed defects using process priors. These technical problems indicate that existing mixed defect identification methods still have inherent shortcomings in dynamic adaptive attention weight adjustment, defect coexistence association reasoning, and the ability to consider both local and global feature extraction. They cannot simultaneously meet the practical requirements of high-precision, lightweight, and real-time identification of mixed wafer defects in industrial mass production scenarios. Summary of the Invention
[0005] The purpose of this invention is to provide a wafer hybrid defect identification method based on a dual-branch dynamic label perception network, which can effectively solve the technical problems in the prior art such as insufficient dynamic adaptive attention weight adjustment capability, lack of defect coexistence correlation modeling, and difficulty in simultaneously extracting local and global features.
[0006] To achieve the above objectives, the technical solution adopted by the present invention is as follows:
[0007] A wafer hybrid defect identification method based on a dual-branch dynamic label sensing network includes the following specific steps: Step 1: Wafer image preprocessing: The original grayscale wafer image output by the wafer electrical testing equipment is preprocessed, where pixel value 0 represents the background area, pixel value 1 represents a normal wafer, and pixel value 2 represents a defective wafer. The original single-channel grayscale image is scaled to a preset size and copied and expanded into a three-channel image to adapt to the network input channel requirements; Step 2: Dual-branch local-global feature extraction: The local branch of the convolutional neural network and the global branch of the lightweight Transformer are set in parallel. The preprocessed three-channel wafer image is input, and the local branch of the convolutional neural network undergoes initial processing. The three-level progressive feature extraction stage extracts local fine texture features. The global branch of the Transformer converts the image into several sets of independent image patches through a convolutional word generator, embeds them, and models the global long-distance spatial dependency through several layers of lightweight Transformer encoders. The outputs of the two branches are local feature maps and global feature maps with a preset number of channels, respectively. After being concatenated along the channel dimension, cross-branch feature interaction fusion is completed through convolution, normalization layers, and activation functions to output a fused feature map. Step 3: Dynamic adaptive attention feature enhancement: A dynamic adaptive attention module is embedded in the back end of each level of feature extraction in the local branch of the convolutional neural network, and... The algorithm computes channel attention and spatial attention, dynamically fusing the two output features by inputting relevant learnable gating coefficients. The channel attention branch uses global average pooling to compress the spatial dimension and generates feature weights for each channel through a multilayer perceptron. The spatial attention branch performs global average pooling and global max pooling along the channel dimension, concatenates the results, and generates pixel-level spatial weights through convolution. Adaptive gating fusion uses the global average pooling features, learnable weights, and biases to calculate gating coefficients, with values within a preset range. The final enhanced feature equals the gating coefficient multiplied by the channel weighted feature plus 1 minus the gating coefficient multiplied by the spatial weighted feature. Step 4: Label association inference: Set several basic missing values. The independent learnable label embedding matrix of the trap is used to model the defect-image feature association and the intrinsic association between defects through a two-level attention mechanism. The label-feature cross attention uses the defect label embedding as the query and the wafer image fusion feature as the key to perform multi-head cross attention. The label self attention uses the label embedding enhanced by cross attention to perform multi-head self attention to learn the coexistence association strength between various defects. Each set of optimized label embeddings is output as the original classification score through an independent fully connected layer. Step 5 Defect classification output: The classification score is converted into a preset interval probability value by the Sigmoid function, a preset judgment threshold is set, and the label vector is output to mark the existence of various defects.
[0008] Preferably, in step 1, the wafer electrical testing equipment acquires the original grayscale wafer image, where pixel value 0 represents the background area, pixel value 1 represents a normal die, and pixel value 2 represents a defective die. The original image has a size of 52×52 pixels.
[0009] Preferably, in step 1, the original single-channel grayscale image is scaled to a size of 224×224 pixels. The scaling is performed using bilinear interpolation. After copying and expanding to a three-channel image, the pixel values of the three channels are exactly the same.
[0010] Preferably, in step 2, the local branch of the convolutional neural network includes an initialization module and a three-level progressive feature stage. The initialization module includes a 7×7 convolutional layer and a 3×3 pooling layer with 32 output channels, a convolution stride of 2, and an output feature map spatial resolution of 1 / 4 of the input image (i.e., 56×56 pixels) with 32 output channels.
[0011] Preferably, each level of the three-level progressive feature stage in step 2 includes a pointwise convolutional channel expansion operation, a depthwise separable convolutional downsampling operation with a stride of 2, and a dynamic adaptive attention module, with the number of channels increasing to 256 channels at each level.
[0012] Preferably, in step 2, the local branches of the convolutional neural network ultimately output a local fine feature map with a preset number of channels, which fully preserves the texture details of scratch defects and point-like aggregation defects.
[0013] Preferably, in step 2, the convolutional terminator of the global branch of the Transformer contains three layers of stride convolutions, each with a stride of 2. Multiple downsampling operations convert the input image into 49 independent image blocks for embedding, each image block being 16×16 pixels in size.
[0014] Preferably, in step 2, the Transformer global branch superimposed learnable location embedding matrix has dimensions that match the image block embedding dimensions, and is used to encode wafer pixel spatial location information.
[0015] Preferably, in step 2, both lightweight Transformer encoders have built-in multi-head self-attention mechanism and GELU activation function. The multi-head self-attention mechanism models global long-distance spatial dependencies in four different subspaces.
[0016] Preferably, in step 2, the channel compression ratio of the channel attention branch in the dynamic adaptive attention module is set to 4, and the hidden layer dimension of the multilayer perceptron is a preset dimension.
[0017] Preferably, in step 2, the convolutional layer of the spatial attention branch in the dynamic adaptive attention module is a 7×7 convolution, which outputs a single-channel spatial weight map to accurately focus on the pixel region where the defect is located.
[0018] Preferably, in step 2, the adaptive gating fusion of the dynamic adaptive attention module utilizes global average pooling features to calculate gating coefficients using learnable weights and biases. The learnable weights are initialized to 0.5 and the biases are initialized to -0.5 to ensure a balanced contribution ratio between channel attention features and spatial attention features in the initial stage.
[0019] Preferably, in step 2, the adaptive gating fusion of the dynamic adaptive attention module automatically adjusts the gating coefficient according to the defect morphology of the input wafer. When the input image is dominated by local compact point defects and scratch defects, the gating coefficient value is automatically reduced to focus on spatial attention enhancement. When the input image is dominated by full-coverage defects and random diffuse global defects, the gating coefficient value is automatically increased to focus on channel feature calibration.
[0020] Preferably, in step 4, the label association reasoning module sets up independent learnable label embedding matrices for each of the 8 basic defect categories, and the label embedding vector dimension of each defect category is 128 dimensions.
[0021] Preferably, in step 4, the label-feature cross attention adopts a multi-head attention mechanism, using the defect label embedding as the query vector and the vector of the fused features after global pooling and linear projection as the key vector and value vector to perform multi-head cross attention calculation.
[0022] Preferably, in step 4, the self-attention between labels adopts a multi-head attention mechanism, using the label embedding enhanced by cross-attention as input, and learning the coexistence correlation strength between pairs of various defects through the self-attention mechanism, and outputting a weight matrix to store the defect coupling process rules.
[0023] Preferably, in step 4, the classification output uses an independent fully connected layer to embed and map the optimized label of each defect category to a one-dimensional classification score. The defect classification judgment threshold is 0.5. When the output probability of the Sigmoid function is greater than or equal to 0.5, the defect of that type is determined to exist and the corresponding label is output. When the output probability of the Sigmoid function is less than 0.5, the defect of that type does not exist and the corresponding label is output.
[0024] Preferably, the model training step also includes a unified binary classification cross-entropy loss function for various defects, treating each defect as an independent binary classification task for unified optimization, and conducting end-to-end joint training of all modules. Data augmentation strategies are introduced during the model training phase, including random horizontal flipping, random vertical flipping, and random rotation within a ±15° angle range, to expand sample diversity and improve model generalization ability. No augmentation operations are performed during the testing and inference phases.
[0025] Preferably, the model training process also includes using the AdamW optimizer, setting the initial learning rate to 1e-3, the weight decay coefficient to 1e-4, the learning rate to decay gradually using a cosine annealing strategy, the total number of training rounds to 30, the batch size to 64, and monitoring the model weights on the validation set using preset evaluation metrics and saving the optimal weight file.
[0026] Preferably, the process also includes inference steps, such as loading the trained model weights from the optimal weight file, inputting the wafer image to be detected, performing preprocessing, bi-branch feature extraction, dynamic attention feature enhancement, and label association inference in sequence, and finally outputting various defect recognition results.
[0027] Preferably, the dual-branch local-global feature extraction module captures both the local fine texture features and the global spatial distribution features of wafer defects through a parallel architecture. The convolutional neural network branch extracts the texture details of tiny point defects and scratch defects, while the lightweight Transformer branch models the global distribution patterns of edge ring defects and full-coverage defects. After the two branches are spliced along the channel dimension, cross-branch feature interaction fusion is completed through convolution, normalization layers, and activation functions, and the fused feature map is output while retaining both local detail information and global distribution information.
[0028] Preferably, the lightweight dynamic tag perception network model has a total of 2.09M parameters and a single wafer image inference computation of 164.2M FLOPs, which can be deployed in low-power embedded edge detection equipment in semiconductor production lines to meet the low-latency inference deployment requirements of industrial online real-time detection.
[0029] Preferably, the dynamic adaptive attention module abandons the fixed-sequence attention calculation process in the prior art, and dynamically balances the contribution ratio of channel attention and spatial attention by inputting relevant learnable gating coefficients, thereby achieving adaptive feature enhancement for all types of defects.
[0030] Preferably, the label association reasoning module explicitly models the process coexistence relationship between different defects, uses a cross-attention mechanism to allow each type of defect to autonomously extract image features of the corresponding region, and uses a self-attention mechanism to learn the coexistence relationship strength between each pair of defects, effectively distinguishing visually similar mixed defects and reducing the false detection rate and false negative rate of mixed defects.
[0031] Compared with the prior art, the present invention has the following beneficial effects:
[0032] This invention constructs a lightweight dynamic label perception network. Through a dual-branch structure, it simultaneously extracts local fine texture features of wafer defects and models global spatial distribution features. The convolutional neural network branch extracts the texture details of micro-scratches and dot-like cluster defects, while the lightweight Transformer branch models the global distribution patterns of edge ring defects and full-coverage defects. Ablation experiments verify that the addition of the global Transformer branch can significantly improve the model's preset evaluation indicators, thus solving the inherent defect of single-branch networks in losing key features.
[0033] This invention employs a dynamic adaptive attention mechanism, which dynamically balances the contribution ratio of channel attention and spatial attention by inputting relevant learnable gating coefficients. It abandons the attention mechanism with a fixed computation process in the prior art. Replacing the dynamic adaptive attention module alone can significantly improve the preset evaluation index, and the enhancement effect of various defect-specific features is significantly improved, realizing adaptive feature enhancement for all types of defects.
[0034] This invention explicitly models the process coexistence relationship between different defects through a label association reasoning module. It enables each type of defect to autonomously extract image features of the corresponding region through a cross-attention mechanism and learns the coexistence relationship strength between each pair of defects through a self-attention mechanism. This effectively distinguishes visually similar mixed defects, significantly reduces the gap in recognition accuracy between single and mixed defects, and greatly reduces the false negative rate of mixed samples.
[0035] This invention adopts a lightweight network architecture, with the total number of model parameters and the computational cost of inference per wafer map both controlled within a range far lower than that of heavy-duty networks. It can be deployed in low-power embedded edge detection devices on production lines, meeting the low-latency real-time inference deployment requirements of embedded edge devices in semiconductor production lines.
[0036] On the MixedWM38 public dataset, the model's macro-average preset evaluation index reaches a high level, comprehensively outperforming lightweight baseline models, heavy convolutional neural networks, pure Transformer networks, and wafer-specific detection networks, with industry-leading overall recognition performance.
[0037] This invention generates a unique thermal activation map for each type of defect using a multi-label gradient weighted activation mapping method, accurately locating the regions corresponding to various defects on the wafer. It has good process interpretability, and process engineers can quickly trace the root causes of various process failures such as lithography misalignment, abnormal thin film uniformity, and mechanical handling scratches based on the thermal map and defect combination. Attached Figure Description
[0038] Figure 1 This is a schematic diagram of the overall technical architecture of the wafer hybrid defect identification method based on a dual-branch dynamic tag sensing network proposed in this invention.
[0039] Figure 2 This is a schematic diagram of the core principle framework of the dual-branch local-global feature extraction module in this invention;
[0040] Figure 3 This is a schematic diagram of the core principle framework of the dynamic adaptive attention feature enhancement module in this invention;
[0041] Figure 4 This is a schematic diagram of the logical flow framework of the tag association reasoning module in this invention;
[0042] Figure 5 This is a schematic diagram of the overall data flow and processing flow of the wafer hybrid defect identification method provided by the present invention. Detailed Implementation
[0043] Example 1
[0044] To make the objectives, technical solutions, and advantages of this invention clearer, the invention will be further described in detail below with reference to specific embodiments.
[0045] In the aforementioned wafer hybrid defect identification method based on a dual-branch dynamic label perception network, step 1 is the wafer image preprocessing stage. This stage standardizes the original grayscale wafer image output by the wafer electrical testing equipment to provide a feature representation conforming to the input specifications for the subsequent deep neural network. The original grayscale wafer image acquired by the wafer electrical testing equipment uses a 52×52 pixel size, where pixel value 0 represents the background area, pixel value 1 represents a normal die, and pixel value 2 represents a defective die. These three types of pixel values form a feature pattern with a specific spatial distribution on the wafer image, which directly reflects the process status and defect distribution during wafer manufacturing. The original single-channel grayscale image is first scaled to 224×224 pixels using bilinear interpolation to adapt to the standard input specifications of the network in this invention. During the scaling process, the bilinear interpolation method effectively maintains the clarity and texture continuity of the wafer defect edges by performing a weighted average operation on four adjacent pixels. The scaled single-channel image is then copied and expanded into a three-channel image along the channel dimension. The copying operation makes the pixel values of the three channels exactly the same, thereby generating an RGB format wafer image that meets the requirements of the mainstream three-channel input format of convolutional neural networks. This operation does not introduce additional information loss and only meets the channel compatibility requirements of the network architecture.
[0046] In the aforementioned wafer hybrid defect identification method based on a dual-branch dynamic label perception network, step 2 is a dual-branch local-global feature extraction stage. This stage uses a parallel convolutional neural network local branch and a lightweight Transformer global branch to simultaneously capture the local fine texture features and global spatial distribution features of wafer defects. The input is a three-channel 224×224 wafer image preprocessed in step 1. The two branches of the dual-branch structure independently extract features and then fuse them along the channel dimension to output a fused feature map containing multi-scale information.
[0047] The local branch of the convolutional neural network comprises two functional components: an initialization module and a three-level progressive feature stage. The initialization module uses a 7×7 convolutional kernel for spatial convolution, with 32 output channels and a stride of 2 to achieve spatial downsampling. A 3×3 max-pooling layer is then applied to further compress the spatial dimension, ultimately outputting a 56×56 basic feature map with a spatial resolution of 1 / 4 of the input image and 32 channels. The large 7×7 kernel in the initialization module effectively expands the receptive field of feature extraction by covering a larger pixel neighborhood in a single convolutional layer, fully capturing the initial contour features and edge information of wafer defects in the shallow layers of the network. The downsampling operation with a stride of 2 ensures a balance between computational efficiency and feature preservation. Each level of the three-level progressive feature stage sequentially includes three sub-modules: pointwise convolutional channel expansion, depthwise separable convolutional downsampling with a stride of 2, and the dynamic adaptive attention module of this invention.
[0048] In the first-level feature extraction stage, the pointwise convolutional channel expansion operation increases the number of input channels from 32 to 64 to increase the feature representation dimension and non-linear expressive power. Subsequently, a 3×3 depthwise separable convolution is used for spatial feature extraction, and a stride of 2 is used to downsample the spatial dimension, reducing the feature map size from 56×56 to 28×28 to reduce subsequent computational complexity. Depthwise separable convolution significantly reduces the number of parameters and computational overhead by decomposing the standard convolution into two independent steps: channel-wise convolution and pointwise convolution. Finally, a dynamic adaptive attention module adaptively enhances the channel and spatial dimensions of the features extracted at the current level to strengthen defect-related features and suppress background interference. The second-level feature extraction stage continues the same structural design as the first stage, increasing the number of channels from 64 to 128 and downsampling the feature map size from 28×28 to 14×14. The third-level feature extraction stage also uses the same structural design, increasing the number of channels from 128 to 256 and downsampling the feature map size from 14×14 to 7×7. After three levels of progressive feature extraction, the local branch of the convolutional neural network finally outputs a 256-channel, 7×7-size local fine feature map, which fully preserves the texture details of minor defects such as scratch defects and point cluster defects, providing a rich local feature foundation for subsequent defect recognition.
[0049] The lightweight Transformer global branch's convolutional word segmenter contains three convolutional layers with a stride of 2. Three downsampling operations transform the input 224×224×3 image into 49 independent 16×16 pixel image patch embeddings. Each image patch embedding has a 768-dimensional dimension. Mathematically, the three downsampling operations are represented as 224 divided by 2 cubed, which equals 28. 28×28 equals 784 pixel locations. After processing by the word segmenter, these 784 pixel locations form 49 image patch embedding vector sequences. The lightweight Transformer global branch stacks a learnable positional embedding matrix to encode the spatial location information of wafer pixels. The dimension of the learnable positional embedding matrix matches the dimension of the image patch embeddings, both being 49×768. The positional embeddings are added element-wise to the image patch embeddings before the Transformer encoder processes the image patch sequence, enabling the model to distinguish image patches at different spatial locations.
[0050] Both lightweight Transformer encoders incorporate a multi-head self-attention mechanism and a GELU activation function. The multi-head self-attention mechanism uses a 4-head design to model global long-range spatial dependencies in parallel across four different subspaces, enabling the model to capture semantic associations between distant pixels on the wafer image. The first layer of the two-layer lightweight Transformer encoder performs self-attention computation on the sequence of image patches after position embedding to model the relationships between local neighborhood image patches. The second layer further models global dependencies between more distant image patches based on the output of the first layer. Both encoder layers include a feedforward neural network layer consisting of two linear transformations with a GELU nonlinear activation function inserted in between to increase the model's nonlinear expressive power. After the output of the lightweight Transformer global branch is normalized by the layer, it is first projected to 256 dimensions by a linear projection layer from 768 embedded dimensions. Then, the projected features with a sequence length of 49 are reshaped into a global feature map with a spatial size of 7×7 and 256 channels to capture the feature information of globally distributed defects such as edge ring defects and full coverage defects. The spatial size of the global feature map is consistent with the local feature map output by the local branch of the convolutional neural network so as to facilitate subsequent concatenation and fusion of channel dimensions.
[0051] In the cross-branch feature fusion stage, the feature maps output from the two branches are concatenated along the channel dimension to form a 512-channel fused feature map. The concatenation operation connects the 256-channel local features from the CNN local branch output with the 256-channel global features from the Transformer global branch output end-to-end along the channel dimension to form a fusion tensor with a total of 512 channels. The concatenated fused features are then passed through a 1×1 convolutional layer, a batch normalization layer, and a ReLU activation function to complete the cross-branch feature interaction fusion. The 1×1 convolutional layer performs linear projection and integration on the channel dimension of the concatenated 512-channel features, compressing the number of channels to the original number to reduce the computational complexity of subsequent modules. The batch normalization layer normalizes the convolutional output features along the channel dimension, making the feature distribution mean 0 and variance 1 to stabilize the network training process and accelerate model convergence. The ReLU activation function introduces a non-linear transformation, enabling the fusion module to learn complex interaction patterns between channels. The final output fused feature map retains both the local texture details extracted by the convolutional neural network and the global distribution pattern information modeled by the Transformer, taking into account the feature requirements of both minor defects and global defects in a single fused feature representation.
[0052] In the aforementioned wafer hybrid defect identification method based on a dual-branch dynamic label sensing network, step 3 is the dynamic adaptive attention feature enhancement stage. The dynamic adaptive attention module described in this stage is physically embedded in the back end of each level of the three-level progressive feature stage of the local branch of the convolutional neural network described in step 2; that is, after feature extraction at each level, it undergoes feature enhancement through this module. For ease of explanation of the module's internal principles, it is listed as a separate step for detailed description. The dynamic adaptive attention module abandons the fixed-sequence attention calculation process of existing technologies by dynamically balancing the contribution ratio of channel attention and spatial attention through inputting relevant learnable gating coefficients. Internally, the module sets up two independent attention calculation paths in parallel: a channel attention branch and a spatial attention branch. The outputs of the two branches are dynamically weighted and combined through an adaptive gating fusion mechanism to generate the final enhanced features.
[0053] The channel attention branch utilizes global average pooling to compress the spatial dimension, reducing the spatial dimension of the input feature map from H×W to 1×1 to generate a global statistical feature representation for each channel. The global average pooling operation iterates through all spatial locations of each channel in the feature map, calculating the arithmetic mean of all pixel values within that channel as the global descriptive statistic for that channel. This statistic encodes the average activation intensity of the channel feature across the entire image space. The globally average pooled features are then fed into a two-layer perceptron network for non-linear mapping of channel weights. The perceptron network includes a compression layer and an expansion layer, with a channel compression ratio set to 4 to balance computational efficiency and feature representation capability. Specifically, assuming the number of channels in the input feature map is C, the output dimension of the compression layer is C divided by 4, and the expansion layer restores the dimension to C to match the number of input channels. The compression layer of the perceptron network first compresses the channel dimension to force the model to learn a more discriminative channel statistical representation, and the expansion layer projects the compressed representation back to the original dimension to generate the channel attention weight vector. A ReLU activation function is used to introduce a nonlinear transformation between the compression layer and the expansion layer. The final output layer uses a Sigmoid activation function to normalize the channel weight vectors to the range of 0 to 1, which serves as the feature modulation coefficients for each channel. The weight vectors generated by the channel attention branch are multiplied channel by channel with the input feature map to achieve weighted enhancement of the features of each channel, automatically enhancing the defect-specific feature channels while suppressing irrelevant background channels.
[0054] The spatial attention branch performs two spatial statistical operations along the channel dimension: global average pooling and global max pooling, to encode the importance of spatial locations from different perspectives. Global average pooling calculates the average value of each spatial location along the channel dimension, generating an average feature map with a channel dimension of 1. Global max pooling extracts the maximum value of each spatial location along the channel dimension, generating a maximum feature map with a channel dimension of 1. The average feature map encodes the average activation intensity of each spatial location, while the maximum feature map encodes the most salient activation response of each spatial location. The two spatial statistical feature maps are concatenated along the channel dimension to form a dual-channel spatial feature representation, which is then fed into a 7×7 convolutional layer for non-linear mapping of spatial weights. The large spatial receptive field of the 7×7 convolutional kernel can capture the contextual relationships between pixels in the spatial neighborhood. The convolutional layer outputs a single-channel spatial weight map, which is then normalized to the 0-1 range using the Sigmoid activation function to serve as the feature modulation coefficients for each spatial location. The weight map generated by the spatial attention branch is multiplied with the input feature map location-by-location to achieve precise focusing on the pixel region where the defect is located.
[0055] The mathematical expression for adaptive gating fusion is:
[0056] ,
[0057] Where Fenhanced F represents the final enhanced feature. channel F represents the weighted feature of the channel attention branch output. spatial The weighted feature of the spatial attention branch output is represented by α, which is a gating coefficient. The gating coefficient α is a scalar value, and the same fusion ratio is used uniformly for all channels and all spatial locations of the current feature map.
[0058] The adaptive gating fusion performs global average pooling on the input feature map F_input of the current dynamic adaptive attention module, and calculates the gating coefficients by passing the pooled one-dimensional feature vector through learnable weights w and bias b. The gating coefficient α is constrained to the range of 0 to 1 by the Sigmoid function. The learnable weight w is initialized to 0.5 and the bias b is initialized to −0.5 to ensure a balanced contribution ratio between channel attention features and spatial attention features in the initial stage. When the input image is dominated by local compact point defects and scratch defects, the gating coefficient α is automatically reduced to below 0.5, causing the network to automatically focus on spatial attention enhancement to better capture the spatial distribution pattern and texture edge information of defects. When the input image is dominated by full-coverage defects and random diffuse global defects, the gating coefficient α is automatically increased to above 0.5, causing the network to automatically focus on channel attention calibration to better model the global channel activation pattern of defects.
[0059] In the aforementioned wafer hybrid defect identification method based on a dual-branch dynamic label perception network, step 4 is the label association inference stage. This stage models defect-image feature association and defect-defect intrinsic association through a two-level attention mechanism to explicitly mine the process coexistence associations between different defects. The label association inference module sets up independent learnable label embedding matrices for each of the eight basic defect categories. The label embedding vector dimension for each defect category is set to 128 dimensions. During network training, the learnable label embedding matrix is automatically optimized through a backpropagation algorithm to learn the semantic representation that best distinguishes each type of defect.
[0060] The first level employs a label-feature cross-attention mechanism, which enables cross-modal interaction modeling between defect label semantics and wafer image features. The fused features are first compressed to a 1×1×512 one-dimensional vector using global average pooling, and then further reduced to 128 dimensions through a linear projection layer to match the label embedding dimension. Using the label embeddings of eight defect classes as query vectors and the projected image features as key and value vectors, a four-head cross-attention computation is performed. This mechanism allows each defect class to proactively extract image features of the corresponding defect region from the entire wafer image based on its label semantics, rather than passively relying on network learning. During the cross-attention computation, the relevance score between the query vector of each defect class and the position key vectors of all image features is calculated, normalized using Softmax, and used as the attention weight. Finally, it is weighted and summed with the value vector to generate the enhanced feature representation of that defect class. The four-head cross-attention design allows each defect class to interact with image features in four different semantic subspaces to capture the multi-dimensional relationship between defects and images.
[0061] The second level is the inter-label self-attention mechanism, which models the pairwise coexistence correlation strength among the eight defect label embeddings. The inter-label self-attention mechanism uses the eight defect label embeddings (enhanced by cross-attention) as input sequences to perform a four-head self-attention calculation. This mechanism allows each defect category's label embedding to interact with all other defect category embeddings to learn coexistence correlation patterns. During the self-attention calculation, each defect label embedding acts as a query and all other defect label embeddings act as keys to calculate correlation strength matrices between defects. These matrices encode the natural coexistence patterns of defects introduced by semiconductor manufacturing processes. For example, scratch defects and point-like aggregate defects may coexist during mechanical handling, thus exhibiting a strong coexistence correlation; edge-loop defects and full-coverage defects may coexist during thin-film deposition process anomalies, also showing a strong coexistence correlation. The correlation weight matrix output by the self-attention mechanism can be extracted for process analysis to help process engineers understand the root causes of defects.
[0062] In the aforementioned wafer hybrid defect identification method based on a dual-branch dynamic label sensing network, step 5 is the defect classification output stage. This stage generates the final defect existence determination result based on the optimized label embedding output by the label association inference module in step 4. The optimized label embedding for each type of defect is mapped to a one-dimensional classification score through an independent fully connected layer. The output dimension of the fully connected layer is 1 to match the requirements of binary classification tasks. Each of the eight defect types has an independent fully connected layer to ensure that the classification decisions for each type of defect do not interfere with each other. The classification score is converted into a probability value within the range of 0 to 1 using the Sigmoid function. The mathematical form of the Sigmoid function compresses and maps any real value to the 0 to 1 range to represent the prediction confidence of the existence of that type of defect. The defect classification determination threshold is set to 0.5. When the output probability of the Sigmoid function is greater than or equal to 0.5, the defect type is determined to exist and a binary label 1 is output. When the output probability of the Sigmoid function is less than 0.5, the defect type is determined to not exist and a binary label 0 is output. Step 5 finally outputs an 8-dimensional 0 / 1 label vector indicating whether each of the eight types of defects exists. This label vector is directly used as the final output result of the wafer hybrid defect identification method.
[0063] During model training, a unified binary cross-entropy loss function is used for all eight defect classes, treating each as an independent binary classification task for unified optimization. All modules are jointly trained end-to-end to ensure optimal overall performance through parameter co-optimization. The mathematical expression of the loss function is the sum and average of the binary cross-entropy losses for each defect class, yielding the final training loss value used for backpropagation to update the network. Data augmentation strategies are introduced during model training, including random horizontal flipping, random vertical flipping, and random rotations within ±15° angles, to expand sample diversity and improve model generalization ability. No augmentation operations are performed during the testing and inference phases. The AdamW optimizer, combined with an initial learning rate of 1e-3 and a weight decay coefficient of 1e-4, controls the regularization strength of the parameters to prevent overfitting. The AdamW optimizer combines the advantages of Adam adaptive learning rate optimization and weight decay regularization, adaptively adjusting the learning rate of each parameter while effectively preventing excessive parameter amplitude. The learning rate is gradually decayed using a cosine annealing strategy. A larger learning rate is maintained in the early stages of training for rapid convergence, while the learning rate is reduced in the later stages to fine-tune the parameters to the optimal solution. The cosine annealing strategy smoothly decays the learning rate from an initial 1e-3 to near zero over 30 training epochs. The total number of training epochs is set to 30, and the batch size is 64. The model weights are monitored on the validation set for the macro-average F1 score; when this score improves, the optimal weight file is automatically saved to preserve the model parameters that yielded the best validation performance during training.
[0064] The inference process loads the trained model weights from the optimal weight file, inputs them into the wafer image to be detected, and sequentially executes steps 1 (preprocessing), 2 (bi-branch feature extraction), 3 (dynamic attention feature enhancement), and 4 (label association inference), ultimately outputting 8 types of defect identification results. No data augmentation is performed during the inference process to ensure the determinism of the inference results. The inference results include an 8-dimensional 0 / 1 label vector and a Sigmoid probability value for each type of defect, used for subsequent process analysis and root cause analysis.
[0065] This embodiment uses the MixedWM38 hybrid wafer public dataset as a verification basis to comprehensively illustrate the complete technical implementation process and performance verification results of the wafer hybrid defect identification method based on a dual-branch dynamic label perception network. The MixedWM38 hybrid wafer public dataset contains a total of 38,015 wafer grayscale images, including 1 class of normal defect-free samples, 8 classes of single defect samples, and 29 classes of binary, ternary, and quaternary mixed defect samples, totaling 38 defect type combinations. The pixel values of the original images are labeled according to the encoding rule that 0 equals background, 1 equals normal die, and 2 equals defect die. The experimental hardware equipment uses a single NVIDIA RTX5070 discrete graphics card, the deep learning development framework uses PyTorch 2.8, the input images are uniformly scaled to 224×224, the training batch size is set to 64, and the total number of training rounds is 30. The dataset is randomly divided into training set, validation set, and test set in a fixed ratio of 8:1:1, and a fixed random seed is used to ensure the reproducibility of the experimental results.
[0066] The core parameters of the model architecture are configured as follows: The CNN local branch employs a three-level inverted residual block structure, with the number of channels progressively increasing to 256 dimensions. Depth-separable convolutions ensure lightweight parameter and computational complexity. The DAA dynamic adaptive attention module's channel compression ratio is set to 4 to balance attention modeling capability and computational efficiency. The Transformer global branch converts the input image into 49 16×16 pixel image blocks, embedded in a two-layer lightweight Transformer encoder. A 4-head self-attention mechanism is configured internally to model global spatial dependencies in four different subspaces. The LCR label association inference module's defect label embedding dimension is set to 128 dimensions. Both label-feature cross-attention and label-to-label self-attention employ a 4-head design to achieve multi-dimensional semantic interaction. The training optimizer uses the AdamW algorithm, with an initial learning rate of 1e-3 and a weight decay coefficient of 1e-4. The learning rate uses a cosine annealing strategy, gradually decaying to near zero over 30 training epochs. The defect classification threshold is set to 0.5, ultimately outputting an 8-dimensional 0 / 1 label vector representing the presence or absence of eight basic defect types.
[0067] To verify the performance contribution of each innovative module of this invention, this embodiment designed a system module ablation control experiment. Using only a convolutional neural network with the standard CBAM attention mechanism and an independent Sigmoid classifier as the base blank model, the three innovative modules of this invention were progressively superimposed to verify their respective performance gains. The macro-average F1 score of the base blank model was 96.42%, and the F1 score difference between single-defect and mixed-defect recognition was 2.22%, reflecting the performance bottleneck of existing methods in mixed-defect recognition tasks. After superimposing the dynamic adaptive attention module of this invention to replace the standard CBAM attention mechanism, the F1 score improved to 97.18%, an improvement of 0.76%, verifying the improved effect of the dynamic gating attention mechanism on feature enhancement. Further superimposing the Transformer global branch on the basis of only replacing the DAA module improved the F1 score to 98.06%, an improvement of 0.88% compared to the previous step, fully demonstrating the important contribution of global feature modeling to mixed-defect recognition tasks. After loading the complete tag association reasoning module, the macro average F1 score reached 99.55%, and the F1 score difference between single defect and mixed defect identification was reduced to 0.21%. The three innovative modules worked together to achieve the best identification effect, verifying the effectiveness of the technical solution of this invention.
[0068] To comprehensively evaluate the recognition performance of the technical solution of this invention, this embodiment conducts a systematic comparative verification experiment between DLA-Net and eight mainstream baseline models across four categories under unified datasets and training conditions. Lightweight convolutional neural network baseline models include three representative architectures: MobileNetV2, MobileNetV3-Small, and ShuffleNetV2. These models significantly reduce the number of parameters and computational complexity using techniques such as depthwise separable convolution and grouped convolution, but their optimal F1 score is only 97.58%, lower than the solution of this invention. Heavyweight convolutional neural network baseline models include two architectures: ResNet50 and EfficientNetV2-S. These models improve feature extraction capabilities through residual connections and compound scaling strategies, but their parameter count exceeds 20M, resulting in huge computational overhead. Their optimal F1 score is only 98.38%, with both accuracy and lightweightness inferior to the solution of this invention. Pure Transformer architecture baseline models include two architectures: Swin-T and ViT-Tiny. These models possess powerful global modeling capabilities through self-attention mechanisms, but their ability to capture texture details of minor defects is insufficient, with an optimal F1 score of only 98.27%. While the DC-Net, a dedicated inspection network for the wafer field, is specifically designed for wafer defect detection tasks with 4.87M parameters and a sample accuracy of 98.46%, slightly higher than this invention, its multi-label core F1 score is still 1.04 percentage points lower than this invention. In contrast, the DLA-Net model of this invention has a total parameter count of only 2.09M and a single wafer image inference computation of only 164.2M FLOPs, achieving significant weight reduction while achieving a macro-average F1 score of 99.55%, comprehensively outperforming all baseline models and verifying the overall performance advantages of the technical solution of this invention.
[0069] This invention achieves a macro-average F1 score of 99.55% on the MixedWM38 public dataset, comprehensively outperforming lightweight baseline models, heavyweight convolutional neural networks, pure Transformer networks, and wafer-specific detection networks, leading the industry in overall recognition performance. The model has only 2.09M parameters, and the inference computation per wafer image is only 164.2MFLOPs, far lower than heavyweight network architectures such as ResNet, EfficientNet, and Swin-T. It can be deployed in low-power embedded edge detection equipment on production lines, meeting the low-latency inference deployment requirements for industrial online real-time detection. This invention generates a unique thermal activation map for each type of defect using a multi-label gradient weighted activation mapping method, accurately locating the corresponding regions of various defects on the wafer. Process engineers can quickly trace the root causes of various process faults such as lithography misalignment, thin film uniformity anomalies, and mechanical handling scratches based on the thermal map and defect combination, demonstrating excellent process interpretability.
[0070] Example 2
[0071] To further verify the technical solution of this invention, this embodiment uses a specific industrial production line deployment scenario as an example to illustrate the application process and technical details of this invention in an actual production environment. In the industrial production line deployment scenario, the wafer electrical testing equipment collects wafer electrical test data in real time and outputs a 52×52 pixel grayscale wafer image. The wafer image is transmitted to the edge embedded detection terminal of the production line via an industrial fieldbus. The edge terminal adopts an ARM architecture processor and an NPU acceleration chip, which has the real-time inference capability to process multiple wafer images per second.
[0072] After receiving the original grayscale wafer image, the edge detection terminal performs step 1, preprocessing. First, it uses bilinear interpolation to scale the 52×52 image to 224×224. Bilinear interpolation is computationally efficient and simple to implement on industrial-grade embedded devices, meeting real-time processing requirements. Then, the single-channel grayscale image is copied and expanded to a three-channel RGB format to adapt to the input format requirements of deep neural networks. The preprocessed wafer image is then input to the DLA-Net model deployed on the edge terminal for forward inference. Step 2, the dual-branch feature extraction module, simultaneously extracts local texture features and global distribution features. The CNN local branch uses a three-level progressive feature extraction method, combined with a dynamic adaptive attention module to adaptively enhance each level of feature. The Transformer global branch uses a two-layer lightweight encoder to model the semantic relationships between pixels that are far apart on the wafer image.
[0073] The DLA-Net model weight file loaded on the edge terminal is approximately 8.4MB in size. After INT8 quantization compression, the weight file size can be further reduced to approximately 2.1MB to adapt to the storage limitations of embedded devices. The model inference latency, accelerated by the NPU, can be controlled within 10 milliseconds, meeting the timeliness requirements of real-time inspection on the production line. The inference output includes an 8-dimensional 0 / 1 label vector indicating the existence of 8 types of defects and the corresponding Sigmoid probability confidence value for each type of defect. The edge terminal simultaneously calls the gradient-weighted class activation mapping algorithm to generate a unique heatmap for each type of defect. The spatial resolution of the heatmap is consistent with the input image, identifying the spatial distribution location of each type of defect on the wafer with pixel-level precision.
[0074] The edge terminal uploads the defect identification results and heatmaps to the central server in real time via the factory's MES (Manufacturing Execution System) interface. The uploaded data includes complete information such as wafer batch number, inspection timestamp, presence markers for eight types of defects, defect confidence scores, and heatmap pixel data. The factory's MES system aggregates and analyzes the uploaded data to extract defect co-occurrence patterns. For example, the simultaneous presence of edge ring defects and full-coverage defects in a batch of wafers may indicate a temperature uniformity issue in the thin-film deposition process; the simultaneous presence of scratch defects and point-like cluster defects in a batch of wafers may indicate contact damage during mechanical handling.
[0075] Process engineers can use the MES system's visual interface to view defect identification results and heatmaps, combined with the defect association weight matrix learned by the LCR module, to perform comprehensive analysis and pinpoint the root causes of process anomalies. The defect association weight matrix output by the LCR module stores the coexistence association strength between each pair of defects in an 8×8 symmetric matrix format. The numerical value ranges from 0 to 1. The larger the value, the stronger the coexistence association between the two types of defects. For example, the association strength between scratch defects and point cluster defects may reach 0.87, while the association strength between scratch defects and full coverage defects may only be 0.12. Process engineers can quickly identify possible sources of process failures based on the association strength matrix and perform targeted process parameter adjustments and equipment maintenance.
[0076] This embodiment verifies the complete application process of the technical solution of the present invention in a real-world industrial production line deployment scenario, demonstrating the technical feasibility and engineering practical value of the solution. The lightweight, real-time, high-precision, and interpretable characteristics of the present invention fully meet the actual needs of semiconductor manufacturing companies for wafer mixed defect identification.
[0077] The foregoing has shown and described the basic principles, main features, and advantages of the present invention. The scope of the present invention is not limited to the above embodiments; any modifications, equivalent substitutions, or improvements made to the above embodiments based on the technical essence of the present invention should be included within the protection scope of the present invention.
Claims
1. A wafer mixed defect identification method based on a dual-branch dynamic label-aware network, characterized in that, Includes the following steps: Step 1, Wafer Image Preprocessing: Standardize the original grayscale wafer image, scale it to a preset size, and copy it as a three-channel image; Step 2, Dual-branch local-global feature extraction: The local branch of the convolutional neural network and the global branch of the lightweight Transformer are set in parallel to extract local texture features and global spatial dependencies respectively. The feature maps output by the two paths are concatenated and fused along the channel dimension to output a fused feature map. Dynamic Adaptive Attention Feature Enhancement: An embedded dynamic adaptive attention module is used to compute channel attention and spatial attention in parallel. The two output features are dynamically fused using a learnable gating coefficient α as input, enhancing the features as follows: The gating coefficients are calculated by performing global average pooling on the current input feature map, then calculating the learnable weights and biases, and finally constraining them to the 0~1 range by the Sigmoid function. Step 4, Label Association Inference: An independent, learnable label embedding is set for each type of defect. A two-level attention mechanism is used to model defect-image feature association and defect-defect coexistence association: The first level is label-feature cross-attention, where multi-head cross-attention is performed using the label embedding as the query and the global pooled projection of the fused features as the key and value; the second level is label self-attention, where multi-head self-attention is performed on the enhanced label embeddings to learn the coexistence association strength between each pair of defects, outputting a weight matrix; each optimized set of label embeddings is then processed through an independent fully connected layer to output a classification score. Step 5, Defect Classification Output: The classification score is converted into a probability value by the Sigmoid function, and a label vector indicating the presence or absence of each type of defect is output according to a preset threshold. 2.The wafer mixed defect identification method based on the dual-branch dynamic tag-aware network according to claim 1, wherein, In step 3, the adaptive gating fusion uses global average pooling features to calculate the gating coefficient through learnable weights and biases. The range of the gating coefficient α is constrained to the interval between 0 and 1 by the Sigmoid function. The learnable weights are initialized to 0.5 and the biases are initialized to -0.
5. When the gating coefficient α is lower than 0.5, the dynamic adaptive attention module emphasizes spatial attention more than channel attention to enhance the capture of spatial distribution patterns and texture edge information of defects. When the gating coefficient α is higher than 0.5, the dynamic adaptive attention module emphasizes channel attention more than spatial attention in order to strengthen the global channel activation mode that addresses modeling defects. 3.The wafer mixed defect identification method based on the dual-branch dynamic tag-aware network according to claim 1, wherein, In step 1, the original grayscale wafer image is 52×52 pixels in size. The original single-channel grayscale image is scaled to 224×224 pixels in size. The scaling is done using bilinear interpolation. After copying and expanding to a three-channel image, the pixel values of the three channels are exactly the same. Data augmentation strategies are introduced during the model training phase, including random horizontal flipping, random vertical flipping, and random rotation within a range of ±15°. 4.The wafer mixed defect identification method based on the dual-branch dynamic tag-aware network of claim 1, wherein, In step 2, the local branch of the convolutional neural network includes an initialization module and a three-level progressive feature stage. The initialization module includes a 7×7 convolutional layer and a 3×3 pooling layer with a convolution stride of 2, 32 output channels, and an output feature map spatial resolution of 1 / 4 of the input image. Each level of the three-level progressive feature stage includes a pointwise convolutional channel expansion operation, a depthwise separable convolutional downsampling operation with a stride of 2, and a dynamic adaptive attention module. The number of channels increases progressively to 256 channels. Specifically, the number of channels in the first level increases from 32 to 64, and the feature map size is downsampled from 56×56 to 28×28; the number of channels in the second level increases from 64 to 128, and the feature map size is downsampled from 28×28 to 14×14; the number of channels in the third level increases from 128 to 256, and the feature map size is downsampled from 14×14 to 7×7. 5.The wafer mixed defect identification method based on the dual-branch dynamic tag-aware network of claim 1, wherein, In step 2, the global branch of the Transformer's convolutional word generator contains three layers of stride convolutions, each with a stride of 2, which convert the input image into 49 independent image block embeddings, each image block being 16×16 pixels in size; a learnable position embedding matrix is superimposed to encode spatial position information; and a two-layer lightweight Transformer encoder has a built-in multi-head self-attention mechanism, which uses a 4-head design to model global long-distance spatial dependencies.
6. The wafer mixed defect identification method based on the dual-branch dynamic label-aware network of claim 1, wherein, In step 3, the channel compression ratio of the channel attention branch in the dynamic adaptive attention module is set to 4. The spatial dimension is compressed by global average pooling and the feature weights of each channel are generated by a multilayer perceptron. The spatial attention branch performs global average pooling and global max pooling along the channel dimension and then concatenates them. The pixel-level spatial weights are generated by 7×7 convolution.
7. The wafer mixed defect identification method based on the dual-branch dynamic label-aware network according to claim 1, wherein, In step 4, the label association reasoning module sets up independent learnable label embedding matrices for each of the 8 basic defect categories. The label embedding vector of each defect category has a dimension of 128. The label-feature cross attention uses the defect label embedding as the query and the vector of the fused features after global pooling and linear projection as the key value to perform multi-head cross attention. The multi-head cross attention adopts a 4-head design. The relevance score between the query vector of each defect category and the position key value vector of all image features is calculated. After being normalized by Softmax, it is used as the attention weight. Finally, it is weighted and summed with the value vector to generate the enhanced feature representation of the defect category. The self-attention between labels is performed by multi-head self-attention using the label embeddings enhanced by cross-attention. The self-attention adopts a 4-head design. Each defect label embedding is used as a query and all other defect label embeddings are used as keys to calculate the correlation between defects, generating an 8×8 correlation strength matrix between defects. The output weight matrix stores the defect coupling process rules. The classification output uses an independent fully connected layer to embed and map the optimized label of each defect category to a one-dimensional classification score, with a defect classification threshold of 0.
5. 8.The wafer mixed defect identification method based on the dual-branch dynamic tag-aware network of claim 1, wherein, It also includes model training steps, using a unified binary classification cross-entropy loss function for various defects, treating each type of defect as an independent binary classification task for unified optimization, and conducting end-to-end joint training of all modules; using the AdamW optimizer, with an initial learning rate set to 1e-3 and a weight decay coefficient set to 1e-4, the learning rate gradually decaying using a cosine annealing strategy, a total of 30 training rounds, a batch size of 64, monitoring preset evaluation metrics on the validation set and saving the optimal weight file. 9.The wafer mixed defect identification method based on the dual-branch dynamic tag-aware network of claim 1, wherein, The lightweight dynamic tag sensing network model has a total of 2.09M parameters and a single wafer image inference computation of 164.2M FLOPs. It can be deployed in low-power embedded edge detection equipment in semiconductor production lines, meeting the low-latency inference deployment requirements for industrial online real-time detection.
10. The wafer mixed defect identification method based on the dual-branch dynamic label-aware network of claim 1, wherein, The dual-branch local-global feature extraction module captures both the local fine texture features and the global spatial distribution features of wafer defects simultaneously through a parallel architecture; the dynamic adaptive attention module dynamically balances the contribution ratio of channel attention and spatial attention by inputting relevant learnable gating coefficients, thereby achieving adaptive feature enhancement for all types of defects. The label association reasoning module explicitly models the process coexistence relationship between different defects, effectively distinguishes visually similar mixed defects, and reduces the false detection rate and false negative rate of mixed defects.
Citation Information
Patent Citations
Defect prototype inversion-based high-noise wafer graph mixed defect decoupling identification method
CN121810704A
Mixed defect identification method for selective interactive modeling of feature blocks of wafer graph
CN121937459A
