Wafer multi-bin map synthesis and profile map regression analysis method based on data augmentation

Abstract

The application provides a wafer multi-Bin graph synthesis and profile graph regression analysis method based on data enhancement, and belongs to the field of semiconductor manufacturing. First, wafer single-Bin graphs, wafer multi-Bin graphs and wafer thickness profile graphs are collected and pretreated; second, a data enhancement model based on Pix2Pix is constructed to realize the generation of synthetic multi-Bin graphs from single-Bin graphs; finally, the synthetic multi-Bin graphs and wafer thickness profile graphs are input into a conditional regression model with a category condition multi-head structure to optimize model parameters until convergence, so as to obtain a trained conditional regression model, realize the generation of predicted thickness profile graphs from multi-Bin graphs, and quickly deduce the problem steps on the production line process. The application can effectively improve the efficiency of joint analysis of defect graphs and thickness graphs, locate the thickness abnormal area by using the prediction result, guide the production line process optimization, realize defect diagnosis and yield improvement.

Description

Technical Field

[0001] This invention relates to the field of semiconductor manufacturing, and in particular to a method for wafer multi-bin map synthesis and contour map regression analysis based on data enhancement. Technical Background

[0002] A wafer is a thin sheet made of semiconductor material (usually silicon) and is the foundation of integrated circuit manufacturing. Multiple dies can be cut from each wafer, and each die contains an integrated circuit. During wafer testing, automated inspection equipment generates wafer maps (also called mapping maps) by performing functional tests on each chip on the wafer. These wafer maps reflect the distribution patterns of qualified and unqualified chips, thus helping to identify problems in the manufacturing process. In actual production, wafer maps often use Bin mapping to represent different test results of chips: CP testing categorizes chips by test item (Bin) and forms a chip state map (BinMAP), and chips that fail the test are marked as failure Bins in the corresponding BinMap.

[0003] However, the following problems still exist:

[0004] (1) It is difficult to obtain wafer defect data. There is very little labeled defect image data on actual production lines. Although there are some publicly available wafer image datasets, such as WM-811K, which contains 811,457 wafer images, the proportion of "defect-free" images is as high as 85.2%, while the proportion of various defect patterns (such as center defects, donut defects, etc.) is very small, resulting in a serious imbalance in the sample distribution. When applied to deep learning models, it is difficult to obtain enough training samples.

[0005] (2) Wafer contour map visualizes the thickness distribution of wafer thin film, which is usually obtained by scanning the wafer surface point by point. This method based on manual statistics or morphological features is simple and fast, but due to the limited feature expression ability, it is difficult to capture the complex spatial distribution pattern in the wafer defect image.

[0006] (3) The frequency of multiple Bin diagrams on wafers is low in actual production lines. The number of samples is limited and the distribution is uneven, making it difficult for the model to fully learn diverse defect patterns during training. At the same time, the large differences between samples under different process conditions further exacerbate the modeling difficulties caused by data scarcity, thus limiting the model's generalization ability.

[0007] In summary, while existing technologies provide classification and visualization methods for wafer defect analysis, they face problems such as insufficient data volume, limited feature expression capabilities, and unbalanced categories. Summary of the Invention

[0008] To overcome the shortcomings of existing technologies, this invention proposes a wafer single-bin image to multi-bin image mapping and thickness profile regression analysis method based on generative adversarial networks (GANs). This method utilizes image data augmentation techniques based on deep learning and GANs to construct a wafer defect multi-bin image by learning the mapping relationship from single-bin wafer images to multi-bin wafer images based on paired images. This image is used to enhance the training sample library of the defect diagnosis model. Furthermore, it performs regression modeling on the wafer thin film thickness profile. By constructing the mapping relationship between multi-bin wafer images and wafer thin film thickness profile images, it enables reverse engineering of the production line process flow to assist in yield analysis and process optimization.

[0009] In a first aspect, the present invention provides a method for wafer multi-bin map synthesis and contour map regression analysis based on data enhancement, comprising the following steps:

[0010] S1, collect real wafer single-bin images and real wafer multi-bin images from the industrial production line;

[0011] S2, expand the number of wafer multi-bin maps based on the actual wafer multi-bin maps to obtain a first set containing wafer multi-bin maps; process the wafer multi-bin maps in the first set to obtain a second set containing wafer single-bin maps. The wafer multi-bin maps in the first set and the wafer single-bin maps in the second set are then combined. Figure 1 One-to-one correspondence; using images from the first and second sets, train a data augmentation model to convert wafer single-bin maps into wafer multi-bin maps;

[0012] S3, input the real wafer single-Bin map from S1 into the data augmentation model and the contour map generation model respectively to obtain the synthetic wafer multi-Bin map and wafer thickness contour map corresponding to the real wafer single-Bin map.

[0013] S4. A conditional regression model is trained using the synthetic wafer multi-bin map and wafer thickness contour map from S3. The conditional regression model includes an image encoder, a category embedding module, and a category conditional multi-head module composed of multiple regression heads. The conditional regression model takes the synthetic wafer multi-bin map and its defect labels as input, processes them using the image encoder and the category embedding module respectively, and then maps them to the target value of the wafer thickness contour map by the category conditional multi-head module.

[0014] S5 collects real wafer multi-bin maps to be predicted on the industrial production line, and uses a conditional regression model to generate a predicted wafer thickness profile map.

[0015] Furthermore, the data augmentation model is a Pix2pix network.

[0016] Furthermore, the methods for obtaining the first and second sets used to train the data augmentation model include:

[0017] S201, normalizes the real wafer multi-bin map collected from the industrial production line;

[0018] S202, Data augmentation is applied to the normalized real wafer multi-Bin map to expand one real wafer multi-Bin map into several wafer multi-Bin maps while keeping the defect mode distribution unchanged. The wafer multi-Bin maps before and after expansion constitute the first set.

[0019] S203, binarize the multi-bin images of each wafer in the first set to obtain the multi-bin images of the wafers. Figure 1 A corresponding wafer single-bin diagram constitutes the second set.

[0020] Furthermore, the contour map generation model is a convolutional network combined with a Gaussian mixture model, trained using real wafer single-bin maps and corresponding real wafer thickness contour maps.

[0021] Furthermore, the category embedding module maps the wafer defect category label into a continuous vector, which is then fused and spliced ​​with the wafer multi-bin map encoded by the image encoder and input into the category condition multi-head module. The category condition multi-head module is a regression head structure composed of batch regression sub-networks, with different regression sub-networks corresponding to different defect labels in the wafer multi-bin map.

[0022] Furthermore, the wafer defect category label is further refined based on the labels in the public dataset, using global statistics, radial density distribution, angular distribution, shape invariant moments, symmetry indices, and spatial clustering; the public dataset is the WM-811K wafer image dataset.

[0023] Furthermore, the training loss of the conditional regression model includes mean squared error loss and structural similarity loss.

[0024] Secondly, the present invention provides a data-enhanced wafer multi-bin map synthesis and contour map regression analysis system, used to implement the aforementioned data-enhanced wafer multi-bin map synthesis and contour map regression analysis method.

[0025] The beneficial effects of this invention are:

[0026] (1) Significantly improve regression accuracy: By generating multi-Bin wafer maps through data augmentation models and combining them with end-to-end conditional regression models, more refined feature mapping relationships can be learned. Compared with simple baseline methods, this invention can significantly reduce the mean square error of the regression model and significantly improve the structural similarity, thereby effectively reconstructing the real thickness profile map.

[0027] (2) Improve data utilization efficiency: By revealing the correlation between the distribution of multi-bin defects and the thickness of wafer thin films, a new approach is provided for wafer inspection and yield analysis. The predicted thickness profile can be used for online process monitoring or virtual inspection, saving the cost and time of actual measurement.

[0028] (3) Strong adaptability: By introducing category conditions such as defect type and distribution difference, the model can better adapt to the characteristics of different types of wafers, and improve its performance and robustness on complex and diverse datasets.

[0029] (4) Alleviating the problem of data scarcity: Using data augmentation models to synthesize rich multi-bin defect images, thereby expanding the training dataset, improving the generalization ability of the regression model, and reducing the dependence on a large amount of real measurement data. Attached Figure Description

[0030] Figure 1 This is an overall framework diagram of a specific embodiment of the present invention.

[0031] Figure 2 This is a schematic diagram of a data augmentation model according to a specific embodiment of the present invention.

[0032] Figure 3 The flowcharts for the data augmentation model and conditional regression model are shown in specific embodiments of the present invention.

[0033] Figure 4 The wafer multi-bin diagram generated for the pix2pix model in a specific embodiment of the present invention.

[0034] Figure 5 The wafer profile diagram generated by the regression model in a specific embodiment of the present invention.

[0035] Figure 6 This is a schematic diagram of the electronic device terminal structure of the present invention. Detailed Implementation

[0036] To make the above-mentioned objects, features, and advantages of the present invention more apparent and understandable, the specific embodiments of the present invention will be described in detail below with reference to the accompanying drawings. Many specific details are set forth in the following description to provide a thorough understanding of the present invention. However, the present invention can be practiced in many other ways different from those described herein, and those skilled in the art can make similar modifications without departing from the spirit of the present invention. Therefore, the present invention is not limited to the specific embodiments disclosed below. Technical features in the various embodiments of the present invention can be combined accordingly without mutual conflict.

[0037] like Figure 1As shown, this invention mainly consists of wafer data acquisition and processing, wafer data augmentation through a pix2pix-based data augmentation model, embedding the augmented wafer dataset into the model for training and optimization to obtain the mapping relationship between the wafer multi-bin map and the contour map, thereby enabling the prediction of the corresponding wafer contour map based on the multi-bin map and realizing the reverse engineering of the production line process.

[0038] Step 1: Data Construction and Augmentation Phase

[0039] Specifically, the steps include the following:

[0040] S1. Data Collection: First, collect multi-bin images of wafers and corresponding wafer thickness profile images (representing the thin film thickness of the wafer in the form of a heat map) from the actual production line. The format is as follows: The PNG images are used. Next, single-bin images of the wafers are extracted from the publicly available WM-811K dataset and used as input to the trained data augmentation model in step three. The WM-811K wafer image dataset contains 811,457 wafer images, of which 172,950 are labeled with artificial defects, representing 9 defect patterns (numbered 0–8, with label 8 indicating no defects).

[0041] S2. Data Standardization: Normalize all images using the following formula:

[0042]

[0043] in and Here, denoted as the pixel mean and standard deviation, respectively, and 'a' represents the image pixel value. These are the normalized image pixel values.

[0044] S3. Data Expansion: In order to increase the training data of the data augmentation model, this invention keeps the defect pattern distribution of the wafer multi-Bin map unchanged, and uses augmentation methods such as rotation, flipping, and translation to expand a single wafer multi-Bin map into multiple multi-Bin maps to obtain an expanded multi-Bin map image dataset, thereby expanding the initial multi-Bin image dataset.

[0045] S4. Construction of training data pairs: The expanded multi-Bin image dataset is binarized to obtain a one-to-one correspondence of single-Bin datasets, which are then used to construct training sample pairs for supervised learning of the data augmentation model. Each training sample pair consists of a single-Bin input image and a multi-Bin target image containing the corresponding defect pattern.

[0046] Step 2: Building a data augmentation model based on Pix2Pix

[0047] like Figure 2 As shown, the specific steps are as follows:

[0048] S1. Construct a Pix2Pix-based data augmentation model: The data augmentation model consists of pix2pix, with the input being single-bin images from the training samples. The output is a multi-bin image synthesized by the generator. The training sample pairs correspond to the real wafer multi-bin diagrams. As a validation object for the model discriminator. This is achieved through training the condition generator. This makes it possible to input a single Bin graph. Synthetic multi-bin graphs that approximate real multi-bin graphs can be generated. ,in It is a random noise vector, discriminator judge The authenticity of the statement.

[0049] The generator G employs a U-Net architecture, consisting of a symmetrical encoder and decoder. The encoder comprises several convolutional blocks, each including 2D convolution (Conv2d) + batch normalization (BatchNorm) + LeakyReLU activation, used for progressive feature extraction and downsampling. The decoder uses transposed convolution (ConvTranspose2d) or upsampling layers to restore the image size. Each block contains convolution + BatchNorm + ReLU activation, and skip connections are used between the corresponding encoder and decoder layers to concatenate feature maps, allowing high-resolution low-level features to be directly passed to the decoder, helping to preserve the details of the output image. Dropout layers (e.g., with a dropout rate of 0.5) can be added to the bottom of the network to increase the randomness of the generation. The last layer of the generator typically uses Tanh or Sigmoid activation functions to map the output back to the range of image pixel values.

[0050] Discriminator D: Employs a Markovian PatchGAN architecture to distinguish between real and fake local blocks in the image. Specifically, it directly concatenates the single-bin input image with the corresponding (real or generated) multi-bin image along the channel dimension using left and right concatenation. The image is used as input to the discriminator. The discriminator consists of a series of convolutional layers, BatchNorm, and LeakyReLU. With an input image size of 64×64, the discriminator outputs a 6×6 two-dimensional feature map. Each output unit corresponds to a local region of approximately 46×46 pixels in the original image, used to determine the probability of authenticity for that region. The global output of PatchGAN... The authenticity of the entire image can be determined by averaging the results of all local discriminations.

[0051] S2. Model Training Optimization: A training strategy of alternating updates to the generator and discriminator is adopted. In each iteration, the generator parameters are fixed first, and the discriminator is updated. The discriminator receives the true pairs in each batch. and generating pairs ( Using binary cross-entropy loss, let... , This maximizes the adversarial loss that distinguishes real samples from generated samples.

[0052] Combat losses The calculation formula is as follows:

[0053]

[0054] in, For real sample pairs Expectations For input samples The expectation.

[0055] Then, the discriminator parameters are fixed and the generator is updated. The generator is then optimized through joint optimization. Get as close to 1 as possible, while reducing the difference between L1 loss.

[0056] The L1 loss calculation formula is as follows: The difference between the generated image and the real image in pixel space is constrained:

[0057]

[0058] The model is optimized with the goal of minimizing the combined loss function until the model converges.

[0059] Combination loss function The calculation formula is:

[0060]

[0061] Among them, hyperparameters This is a weighting factor used to balance L1 loss and adversarial loss. In this specific embodiment, weights are set. This ensures that L1 loss accounts for a major proportion of the combined loss, which helps to preserve image structural information.

[0062] The training and optimization of the data-augmented model are implemented in the PyTorch framework, and the model parameters are updated through backpropagation and gradient descent. In common setups, the Adam optimizer (learning rate) is used. , , The training process involves approximately 100–200 epochs, with a batch size of 16 and weight decay set to 0. The learning rate is linearly decayed during the later stages of training to stabilize model convergence. These configurations can be fine-tuned based on specific circumstances to achieve optimal generation results, ultimately yielding a fully trained data-augmented model.

[0063] Step 3: Single Bin Batch Generation Stage

[0064] First, the WM-811K single-bin images acquired in step one are standardized to the same size. The data is then input into the data augmentation model trained in step two, and the corresponding synthetic multi-bin image is output.

[0065] Secondly, a contour map generation model composed of a convolutional network and a Gaussian mixture model is constructed. This model takes a real single-bin image as input and is trained and optimized using the maximum likelihood estimation criterion and the real contour map corresponding to the real single-bin image as the supervision signal. The trained contour map generation model is then obtained by inputting the WM-811K single-bin image into the trained contour map generation model.

[0066] Finally, the resulting synthetic multi-bin map and the corresponding synthetic wafer profile map are used for subsequent conditional regression model training.

[0067] Step 4: Establishing the Conditional Regression Model

[0068] To explore the quantitative relationship between multi-bin wafer maps and corresponding wafer profile maps (reflecting thin film thickness distribution), this invention employs a conditional depth end-to-end network to construct a conditional regression model. Based on the existing depth encoder structure, a type-conditional modeling mechanism is introduced. The overall process is as follows: Figure 3 As shown, by enhancing feature representation capabilities and modeling flexibility, the correlation strength between multi-bin maps and contour maps on wafers is comprehensively evaluated, thereby improving the model's ability to learn complex and diverse wafer map distribution patterns.

[0069] The conditional regression model uses ResNet as its backbone network, consisting of an image encoder, a type embedding module, and a regression head. The regression head employs a multi-head conditional structure, meaning that an independent regression sub-network is designed for each wafer with different defect categories. This multi-head structure effectively separates interference between categories, improving the model's ability to model in multi-category mixed scenarios. The regression sub-network is a small neural network module that concatenates the feature vectors obtained from the encoder and the continuous vectors obtained from the type embedding module, mapping them to the target value (e.g., a film thickness indicator, regression score, or contour similarity). In the feature encoding stage, Conditional BatchNorm is used to normalize the intermediate feature maps extracted by the encoder, integrating category conditional information into the feature distribution modulation process.

[0070] The input data for the conditional regression model consists of three-channel multi-bin images of a wafer and defect category labels that include different categories of conditions such as defect type and distribution differences. The encoder outputs the corresponding contour feature map prediction result. After inputting a multi-bin image of the wafer, the encoder outputs a compressed high-dimensional feature vector. Meanwhile, defect category labels via the embedding matrix in the category embedding module Mapped to continuous vectors To express the conditional prior information of wafer types with different defect categories, the continuous vector is then combined with the eigenvector. The feature vectors are fused and spliced ​​to form a feature vector that incorporates category conditional information. It is used as input to the regression head to predict the wafer profile through the regression network.

[0071] To make the conditional regression in the model more discriminative and physically meaningful, this invention is based on global statistics (total number of defects, coverage, average radial distance, radial variance, and average edge distance) and radial density distribution (dividing the wafer into 10 concentric rings outward from the center, and calculating the defect distribution density on each ring). Based on the quantification of the original eight defect criteria in the WM-811K dataset, the defect categories were further refined and a unified interpretable defect category label was formed (16–24 categories are recommended, but 20 categories are defaulted in the specific implementation of this invention). The features include angular distribution (dividing the wafer into 8 sectors and calculating the number of defects per unit area in each sector), shape invariant moments (Hu moments), symmetry index, and spatial clustering.

[0072] For example, for the global defect categories in the dataset, the classification criteria can be further quantified based on the radial density distribution. The sum is greater than 0.5 and If all values ​​are <0.1, it is classified as "central type". The sum is greater than 0.5 and If <0.1, it is classified as "marginal type". The maximum value is >0.5 and , If all values ​​are <0.1, it is classified as "ring / donut type". If the average value is less than 0.1, it is marked as "uniform / sparse". For the strip / local defect category in the dataset, it is further subdivided into single sector, double sector symmetrical or inclined strip, based on the uneven distribution characteristics of the sector in the angular distribution (a certain single sector accounts for a significant proportion). For the defect categories such as "strip, block and grid / petal" in the dataset, it is further divided according to shape and symmetry.

[0073] The model can automatically determine the defect category label of the multi-bin map from the numerical features of the input wafer multi-bin map based on the above classification rules. At the same time, these defect category labels can also be manually corrected during the early stage of model training. The defect category labels are used as conditional inputs to the category embedding module to modulate the encoder output and drive the regression head corresponding to each class, thereby maintaining the consistency of input and output during the training and inference stages and improving the prediction accuracy and interpretability in multi-class mixed scenarios.

[0074] During model training, mean squared error (MSE) is used as the main loss function, and structural similarity index (SSIM) is used as an auxiliary loss term to improve the model's ability to preserve the spatial structure of the contour map. Overall loss function. Expressed as:

[0075]

[0076] in and As a weighting factor, For the conditional regression model to predict the profile plot, This is a true outline drawing. This is the mean square error term. It is a structural similarity index.

[0077] The mean square error term The calculation formula is as follows:

[0078]

[0079] Structural similarity index The calculation formula is as follows:

[0080]

[0081] in, , These represent the average pixel value of the true contour map and the average pixel value of the predicted contour map, respectively. , These are the variance of the true profile plot and the variance of the predicted profile plot, respectively. To predict the variance of the profile plot and the covariance of the true profile plot, , This is a preset stability constant.

[0082] This composite loss design can both constrain the numerical accuracy of the predicted values ​​and maintain the spatial structural consistency of the thickness distribution map.

[0083] The conditional regression model optimizer uses AdamW, with a Dropout layer (scale 0.2–0.4) and weight decay added to the regression head. To enhance the model's generalization ability, and to prevent the pre-trained encoder from being over-updated, the encoder and head modules use a differential learning rate scheduling mechanism, i.e., the encoder uses a lower learning rate and the regression head uses a higher learning rate, in order to ensure the stability and generalization ability of the training process. At the same time, K-fold cross-validation (K=5) is used for performance verification to prevent overfitting.

[0084]

[0085]

[0086] in, For encoder learning rate, This is the learning rate for the regression head.

[0087] To evaluate the robustness and stability of the model, an early stopping mechanism can be introduced during training. When the validation set error does not decrease significantly for several consecutive cycles, the training process is automatically terminated.

[0088] To verify the effectiveness of the method of the present invention, the following experiment was further designed:

[0089] S1. Data Acquisition: Due to the current lack of large-scale publicly available paired datasets of wafer outline maps and multi-bin maps in the industry, this specific embodiment collects a certain number of wafer single-bin maps and corresponding wafer outline maps as the original data source, based on real data from a certain advanced logic device production line. The resolution of each image is uniformly scaled down to [resolution value missing]. The data is saved as a three-channel PNG format, where the wafer single-bin map reflects the device yield distribution, and the wafer contour map expresses the thin film thickness distribution in the form of a heat map.

[0090] The experimental environment configuration for this specific embodiment is shown in Table 1.

[0091] Table 1 Experimental Environment Configuration

[0092]

[0093] S2. Data Augmentation: To compensate for the insufficient sample size, the trained data augmentation model was used to augment and map single-bin images, artificially generating multi-bin images. A triplet dataset of "single-bin image – multi-bin image – contour image" was constructed, resulting in approximately 20,000 pairs of samples, which were divided into training, validation, and test sets in an 8:1:1 ratio. The multi-bin images generated by the data augmentation model are shown below. Figure 4 As shown, the “GFA” column represents the defect category, the “WM-811K” column represents the actual single-bin graph, and the “Synthetic Graph” column represents the synthetic multi-bin graph output by the data augmentation model. It can be found that the defect distribution pattern of the synthetic multi-bin graph is basically consistent with that of the actual single-bin graph.

[0094] S3. Comparison of Different Regression Strategies: To verify and compare the quantitative relationship between different regression strategies and the contour maps of wafers from the same batch (reflecting the film thickness distribution), this specific embodiment further compared four regression schemes (denoted as E1–E4) within the same experimental framework to evaluate the differences in accuracy, robustness, and generalization ability of each scheme. E1–E3 are control methods, and E4 is the method of this invention.

[0095] E1: Manual feature extraction + LightGBM. Statistical and morphological features are extracted from multiple Bin images, including pixel mean, variance, energy, contrast, number of connected regions, area / density, centroid coordinates, boundary length, etc. The feature vectors are then input into the LightGBM regressor to output a predicted contour map.

[0096] E2: Frozen pre-trained Encoder + LightGBM. ResNet50 is pre-trained on ImageNet, the top-level classifier is removed to obtain deep features, most network layers are frozen, and only the first or shallow convolutional modules are allowed to be slightly fine-tuned. These are then fed into the LightGBM regressor to output predicted contour maps.

[0097] E3: End-to-end fine-tuning encoder + small MLP head. The encoder and regression head (MLP) are trained together, allowing partial layer unfreezing. The mapping relationship between multi-bin map features and predicted contour maps is obtained through a non-linear mapping function, and the final output is the predicted contour map.

[0098] All four regression strategies used AdamW as the optimizer, with weight decay set to... encoder lr= , regression head lr= Batch size = 16; maximum epochs = 100. The regression head uses Dropout (0.2–0.4) and BatchNorm / LayerNorm. MSE, MAE, and SSIM are calculated on the validation set for each epoch. Early termination is triggered if MSE does not significantly decrease for 10 consecutive epochs, and K=5 fold cross-validation is used to assess model robustness.

[0099] The final performance comparison results of the four regression model methods are shown in Table 2. It can be found that E4's MSE, MAE, and SSIM performance are all better than the other three regression strategies, which proves the superiority of the experimental method.

[0100] Table 2 Performance Comparison of Regression Model Methods

[0101]

[0102] Figure 5 The comparison results of the wafer profile map predicted by E4 and the actual wafer profile map are shown. "GFA" is the defect category, "WM-811K" is the actual single-bin map, and "synthetic map" is the wafer profile map predicted by E4. It can be found that the defect distribution of the wafer map is basically the same.

[0103] To further verify the stability of the data augmentation model of this invention, real data from industrial production lines was incorporated into the testing phase of the conditional regression model to verify its generalization performance. The results are shown in Table 3. It can be observed that the model trained based on synthetic samples can effectively adapt to the distribution of real production line data, maintaining high prediction accuracy even on unseen real data. This result demonstrates that this invention achieves feature distribution alignment and inter-domain transfer through generative data augmentation, significantly improving the model's robustness and adaptability to real production line data, thus exhibiting superior generalization ability in wafer chart multi-bin and contour map regression tasks.

[0104] Table 3. Performance Comparison of Regression Model Methods After Incorporating Real Data

[0105]

[0106] Based on the same inventive concept, this embodiment also provides a data-enhanced wafer multi-bin map synthesis and contour map regression analysis system for implementing the above-mentioned data-enhanced wafer multi-bin map synthesis and contour map regression analysis method. The system includes:

[0107] The data acquisition module is used to acquire real wafer single-bin images and real wafer multi-bin images from industrial production lines.

[0108] The data augmentation module is used to expand the number of wafer multi-bin maps based on real wafer multi-bin maps to obtain a first set containing wafer multi-bin maps; the wafer multi-bin maps in the first set are processed to obtain a second set containing wafer single-bin maps, and the wafer multi-bin maps in the first set and the wafer single-bin maps in the second set are combined. Figure 1 One-to-one correspondence; using images from the first and second sets, train a data augmentation model to convert wafer single-bin maps into wafer multi-bin maps;

[0109] The single-bin map input module is used to input the real wafer single-bin map into the data augmentation model and the contour map generation model respectively, to obtain the synthetic wafer multi-bin map and wafer thickness contour map corresponding to the real wafer single-bin map.

[0110] The conditional regression module is used to train a conditional regression model using a synthetic wafer multi-bin map and a wafer thickness profile map. The conditional regression model includes an image encoder, a category embedding module, and a category conditional multi-head module consisting of multiple regression heads. The conditional regression model takes the synthetic wafer multi-bin map and its defect labels as input, processes them using the image encoder and the category embedding module respectively, and then maps them to the target values ​​of the wafer thickness profile map by the category conditional multi-head module.

[0111] For the system embodiments, since they basically correspond to the method embodiments, relevant details can be found in the descriptions of the method embodiments; the implementation methods of the remaining modules will not be repeated here. The system embodiments described above are merely illustrative. The units described as separate components may or may not be physically separate, and the components shown as units may or may not be physical units; that is, they may be located in one place or distributed across multiple network units. Some or all of the modules can be selected to achieve the purpose of the present invention according to actual needs. Those skilled in the art can understand and implement this without any creative effort.

[0112] The system embodiments of the present invention can be applied to any device with data processing capabilities, such as a computer or other similar device. The system embodiments can be implemented in software, hardware, or a combination of both. Taking software implementation as an example, as a logical device, it is formed by the processor of any data processing device loading the corresponding computer program instructions from non-volatile memory into memory for execution.

[0113] It should also be noted that the data-enhanced wafer multi-bin map synthesis and contour map regression analysis method in the above embodiments can essentially be executed by a computer program. Therefore, similarly, based on the same inventive concept, another preferred embodiment of the present invention also provides a computer electronic device corresponding to the method provided in the above embodiments, which includes a memory and a processor;

[0114] The memory is used to store computer programs;

[0115] The processor is configured to implement, when executing the computer program, a data-enhanced wafer multi-bin map synthesis and contour map regression analysis method as described in the above embodiments.

[0116] From a hardware perspective, such as Figure 6 The diagram shown is a hardware structure diagram provided in this embodiment. In addition to the processor, memory, network interface and non-volatile memory shown in the diagram, any device with data processing capabilities in the embodiment may also include other hardware depending on the actual function of the device with data processing capabilities, which will not be described in detail here.

[0117] When the logical instructions in the aforementioned memory can be implemented as software functional units and sold or used as independent products, they can be stored in a computer-readable storage medium. Based on this understanding, the technical solution of the present invention, in essence, or the part that contributes to the prior art, or a part of the technical solution, can be embodied in the form of a software product, which is stored in a storage medium.

[0118] Therefore, based on the same inventive concept, another preferred embodiment of the present invention also provides a computer-readable storage medium corresponding to the method provided in the above embodiments. The storage medium stores a computer program, which, when executed by a processor, can realize a data-enhanced wafer multi-bin map synthesis and contour map regression analysis method in the above embodiments.

[0119] It is understood that the computer-readable storage medium can be an internal storage unit of any data processing device described in any of the foregoing embodiments, such as a hard disk or memory. The computer-readable storage medium can also be an external storage device of any data processing device, such as a plug-in hard disk, smart media card (SMC), SD card, flash card, etc., equipped on the device. Furthermore, the computer-readable storage medium can include both internal storage units and external storage devices of any data processing device. The computer-readable storage medium is used to store the computer program and other programs and data required by the data processing device, and can also be used to temporarily store data that has been output or will be output.

[0120] The above examples are merely specific embodiments of the present invention. Obviously, the present invention is not limited to the above embodiments and many variations are possible. All variations that can be directly derived or conceived by those skilled in the art from the disclosure of the present invention should be considered within the scope of protection of the present invention.

Claims

1. A wafer multi-Bin map synthesis and contour map regression analysis method based on data augmentation, characterized in that, The method comprises the following steps: S1, collecting real wafer single Bin images and real wafer multi-Bin images on an industrial production line; S2, performing quantity expansion based on the real wafer multi-Bin images to obtain a first set of wafer multi-Bin images; processing the wafer multi-Bin images in the first set to obtain a second set of wafer single Bin images, the wafer multi-Bin images in the first set corresponding one-to-one to the wafer single Bin images in the second set; and training a data enhancement model for converting wafer single Bin images into wafer multi-Bin images by using the images in the first set and the second set; S3, inputting the real wafer single Bin images of S1 into the data enhancement model and a contour image generation model respectively to obtain synthetic wafer multi-Bin images and wafer thickness contour images corresponding to the real wafer single Bin images; S4, training a conditional regression model by using the synthetic wafer multi-Bin images and the wafer thickness contour images of S3, the conditional regression model comprising an image encoder, a category embedding module and a category conditional multi-head module composed of multiple regression heads; the conditional regression model taking the synthetic wafer multi-Bin images and their defect labels as inputs, processing the inputs by using the image encoder and the category embedding module respectively, and then mapping the inputs to target values of the wafer thickness contour images by the category conditional multi-head module; S5, collecting real wafer multi-Bin images to be predicted on an industrial production line and generating predicted wafer thickness contour images by using the conditional regression model.

2. The data augmentation based wafer multi-Bin graph synthesis and profile graph regression analysis method of claim 1, wherein, The data enhancement model is a Pix2pix network.

3. The data augmentation based wafer multi-Bin graph synthesis and profile graph regression analysis method of claim 1, wherein, The method for obtaining the first set and the second set for training the data enhancement model comprises: S201, normalizing real wafer multi-Bin images collected from an industrial production line; S202, expanding one real wafer multi-Bin image into several wafer multi-Bin images by using a data enhancement method while keeping the defect mode distribution unchanged, the wafer multi-Bin images before and after the expansion constituting the first set; S203, performing binarization processing on each wafer multi-Bin image in the first set to obtain wafer single Bin images corresponding one-to-one to the wafer multi-Bin images, the wafer single Bin images constituting the second set.

4. The data augmentation based wafer multi-Bin graph synthesis and profile graph regression analysis method of claim 1, wherein, The contour image generation model is a convolutional network combined with a Gaussian mixture model and is trained by using real wafer single Bin images and corresponding real wafer thickness contour images.

5. The data augmentation based wafer multi-Bin map synthesis and profile map regression analysis method of claim 1, wherein, The category embedding module maps wafer defect category labels into continuous vectors, which are fused and spliced with wafer multi-Bin images encoded by the image encoder and then input into the category conditional multi-head module; The category conditional multi-head module is a regression head structure composed of batch regression sub-networks, different regression sub-networks corresponding to different defect labels of wafer multi-Bin images.

6. The data augmentation based wafer multi-Bin graph synthesis and profile graph regression analysis method of claim 5, wherein, The wafer defect category labels are further refined based on global statistics, radial density distribution, angular distribution, shape invariant moments, symmetry index and spatial aggregation degree on the basis of labels of a public data set, the public data set being a WM-811K wafer image data set.

7. The data augmentation based wafer multi-Bin graph synthesis and profile graph regression analysis method of claim 1, wherein, The training loss of the conditional regression model comprises a mean square error loss and a structural similarity loss.

8. A wafer multi-Bin figure synthesis and profile graph regression analysis system based on data enhancement, used to implement the wafer multi-Bin figure synthesis and profile graph regression analysis method based on data enhancement in claim 1, characterized in that, The system comprises: a data collection module configured to collect real wafer single-Bin images and real wafer multi-Bin images on an industrial production line; a data augmentation module configured to perform quantity expansion based on the real wafer multi-Bin images to obtain a first set of wafer multi-Bin images; processing the wafer multi-Bin images in the first set to obtain a second set of wafer single-Bin images, the wafer multi-Bin images in the first set corresponding one-to-one to the wafer single-Bin images in the second set; and training a data augmentation model for converting wafer single-Bin images into wafer multi-Bin images using the images in the first set and the second set; a single-Bin image input module configured to input real wafer single-Bin images into the data augmentation model and the contour map generation model respectively to obtain synthetic wafer multi-Bin images and wafer thickness contour maps corresponding to the real wafer single-Bin images; a conditional regression module configured to train a conditional regression model using the synthetic wafer multi-Bin images and the wafer thickness contour maps, the conditional regression model including an image encoder, a category embedding module, and a category-conditioned multi-head module composed of multiple regression heads; the conditional regression model takes the synthetic wafer multi-Bin images and their defect labels as input, processes them using the image encoder and the category embedding module respectively, and then maps them to target values of the wafer thickness contour maps by the category-conditioned multi-head module.

9. A computer-readable storage medium, characterized in that, The storage medium has a computer program stored thereon, and when the computer program is executed by a processor, a wafer multi-Bin image synthesis and contour map regression analysis method based on data augmentation is implemented.

10. A computer electronic device, comprising: comprising a memory and a processor; the memory is configured to store a computer program; the processor is configured to implement the wafer multi-Bin image synthesis and contour map regression analysis method based on data augmentation when the computer program is executed.

Images