A wafer defect detection method based on generative adversarial networks

Abstract

This invention discloses a wafer defect detection method based on generative adversarial networks (GANs), belonging to the field of wafer defect detection technology. The method includes the following steps: dividing the surface of the wafer to be inspected into multiple independent regions; constructing a data mapping set; performing pixel-level alignment on a first-resolution image and extracting standard background features; using these features as conditional inputs to the generator of the GAN to generate a third-resolution image of the reference region; performing adversarial training through a discriminator until the generator converges and is optimized; inputting the first-resolution image of the wafer to be inspected at the target detection region and its process parameters into the optimized generator to generate a defect-free third-resolution reference image of the target region; acquiring the actual high-resolution detection image of the target region and performing a differential comparison with the generated reference image. Using this scheme, the true defects of the wafer can be accurately distinguished in complex environments with dynamically changing process characteristics, improving the accuracy of the detection results.

Description

Technical Field

[0001] This invention relates to the field of image generation technology, and in particular to a wafer defect detection method based on generative adversarial networks. Background Technology

[0002] Currently, optical scanning and image processing technologies are widely used in the field of semiconductor wafer defect detection. Existing defect detection systems typically include high-resolution optical imaging modules, precision motion control platforms, and image comparison algorithms. The basic steps involve capturing the original image of the wafer surface under test using an optical sensor, and identifying abnormal pixels using preset threshold segmentation or reference image comparison methods. This enables automated detection of macroscopic defects such as particle contamination and scratches.

[0003] To further improve detection sensitivity and reduce false alarm rate, existing technologies have proposed an improved background suppression method based on reference histogram matching or statistical modeling. For example, by averaging the image features of multiple "Golden Dies" in the same batch, a standard reference background is constructed, and the differential subtraction algorithm is used to eliminate texture noise on the wafer surface, thereby enhancing the contrast between the real defect signal and the background circuit pattern to a certain extent.

[0004] However, the aforementioned improvements still suffer from difficulties adapting to process variations in practical applications. This leads to global changes in the film thickness, reflectivity, or critical dimensions of the wafer surface when there are slight fluctuations in the photolithography, etching, or thin film deposition processes. These fluctuations, in turn, cause nonlinear changes in background features. In the absence of absolute standard samples, these background fluctuations are easily misjudged as defects, or real defects may be masked by fluctuation artifacts.

[0005] Therefore, how to accurately distinguish between real defects and process fluctuation interference in a complex environment with dynamically changing process characteristics, and improve the accuracy of detection results, has become a technical problem that urgently needs to be solved in this field. Summary of the Invention

[0006] The purpose of this invention is to provide a wafer defect detection method based on generative adversarial networks, which solves the technical problem in the prior art that it is difficult to accurately distinguish between real defects and process fluctuation interference in complex environments with dynamically changing process characteristics, thus making it difficult to improve the accuracy of detection results.

[0007] To achieve the above objectives, the technical solution of the present invention is as follows:

[0008] A wafer defect detection method based on generative adversarial networks includes the following steps: The surface of the wafer to be inspected is divided into multiple independent regions; Select one region as a reference region, obtain first resolution images and corresponding first process parameters of multiple reference wafers of the same specifications in the reference region, and construct a data mapping set; The first resolution image is aligned at the pixel level, the first process parameters are normalized to eliminate process drift error, and the standard background features of the reference area are extracted. The standard background features and the first process parameters are used as conditions to input into the generator of the generative adversarial network to generate a third resolution image of the reference region. Using the second-resolution image of the reference region as a real baseline sample, the generator is trained adversarially against the third-resolution image through a discriminator until the generator converges and is optimized. The first resolution image of the wafer to be inspected at the target inspection area and its process parameters are input into the optimized generator to generate a defect-free third resolution reference image of the target area. A high-resolution detection image of the target region is acquired and compared with the generated reference image. The difference result is used to determine whether there is a defect in the target region.

[0009] Furthermore, the process of optimizing the generator until it converges also includes the following steps: Select at least one second reference region that is different from the aforementioned reference region; Obtain first resolution images of multiple reference wafers of the same specifications in the second reference region and the corresponding first process parameters; The first resolution image of the second reference region is pixel-level aligned, the corresponding first process parameters are normalized, and the standard background features of the second reference region are extracted. The standard background features of the second reference region and the first process parameters are used as input conditions, and the network weights of the optimized generator are loaded as initial weights and input into the generator of the generative adversarial network to generate a third resolution image of the second reference region. Using the second resolution image of the second reference region as a real benchmark sample, adversarial training continues between the generated third resolution image of the second reference region and the discriminator until the generator converges again for the second reference region and is further optimized.

[0010] Furthermore, the first resolution accuracy is lower than the second resolution, and the third resolution accuracy is greater than or equal to the second resolution.

[0011] Furthermore, the first resolution image is acquired by high-speed coarse scanning performed by a large field-of-view optical measurement device, while the second resolution image and the actual high-resolution detection image are acquired by fixed-point fine scanning performed by a scanning electron microscope or a high-resolution optical device with a confocal displacement sensing module.

[0012] Further, the step of performing a differential comparison with the generated reference image and determining whether the target region has defects based on the difference results includes the following steps: Calculate the pixel-to-pixel absolute residual between the actual high-resolution detected image and the generated defect-free third-resolution reference image, and generate a residual feature map; The residual feature map is binarized using an adaptive threshold segmentation algorithm, and connected regions in the binarized image are extracted. When the calculated area threshold of a single connected region or the average pixel intensity within that region exceeds the set defect tolerance threshold, it is determined that there is a defect in the target detection region, and the location coordinates and geometric features of the defect are recorded.

[0013] Furthermore, the normalization process for the first process parameters to eliminate process drift error includes: Extract process parameter sequences of the same dimension from x reference wafers of the same specification, and map them to a unified feature space using Z-score normalization; the normalized process parameters... Satisfying the formula:

[0014] in, The original process parameter values ​​are input for the i-th wafer. Let x be the statistical mean of the process parameters corresponding to the x wafers in the reference region. This represents the corresponding standard deviation.

[0015] Furthermore, the first process parameter is real-time physical state data of the processing machine or chamber where the wafer is located, including at least one or more of the following combined data: chamber pressure, radio frequency power, etching or deposition gas flow rate, and electrostatic chuck temperature distribution characteristics.

[0016] Further, the step of performing pixel-level alignment on the first resolution image and extracting standard background features of the reference region includes: A spatial transformation matrix is ​​calculated using a registration algorithm based on feature points or phase correlation, and the transformation matrix is ​​used to perform affine transformation registration on x images of the first resolution. The registered x images are subjected to weighted averaging and principal component analysis at the pixel channel level to filter out randomly distributed imaging noise and extract low-frequency image information with stable topological structure as the standard background features.

[0017] Furthermore, the step of using standard background features and the first process parameter as conditional input includes: The standard background features are used as the main image input of the network. At the same time, the normalized first process parameters are mapped to scaling and translation parameters of affine transformation through a multilayer perceptron. The scaling and translation parameters are used to dynamically modulate the feature map in the adaptive instance normalization layer to cross-fuse one-dimensional process features with two-dimensional spatial features.

[0018] Furthermore, during the process of the generator loading the optimized network weights and continuing adversarial training on the second reference region, to prevent the generator from forgetting the learned features of the reference region, an elastic weight consolidation regularization term is introduced into the training loss function. Perform parameter update constraints:

[0019] in, For the current generator network iteration One parameter, For the first reference region, the training convergence and solidification are performed to solidify the first reference region. Optimal parameters The first element obtained by calculating the diagonal elements of the Fisher information matrix The importance weights of each parameter for the first reference region task This is the regularization intensity coefficient.

[0020] Furthermore, the discriminator of the generative adversarial network includes: The image discrimination module is used to divide the input image into multiple local receptive field patches and determine the authenticity of each patch; A consistency discrimination module is used to evaluate the conditional consistency between the patch features received by the image discrimination module and the input process parameters; The final output of the discriminator is a weighted sum of the outputs of the image discrimination module and the conditional consistency discrimination module.

[0021] The present invention provides a wafer defect detection method based on generative adversarial networks, which has the following core advantages compared with the prior art: This scheme constructs a data mapping set by acquiring a first-resolution image of a reference wafer of the same specifications and the corresponding first-process parameters, and then normalizes the process parameters. This enables the establishment of a quantitative correlation between "process fluctuations and image features," eliminating process drift errors at the underlying data level. By extracting standard background features from the reference region, the previously difficult-to-quantify nonlinear background changes are transformed into trainable feature vectors, thus solving the interference of background feature changes caused by fluctuations in film thickness and reflectivity in traditional methods, facilitating accurate differentiation between real defects and process artifacts. Furthermore, by utilizing a generative adversarial network (GAN) to input standard background features and real-time process parameters as conditional inputs to the generator, a defect-free third-resolution reference image of the corresponding target region is directly generated, freeing detection from the physical limitation of a "golden die." Even in cases of subtle changes in the production environment or the lack of a perfect reference wafer, the generator can infer and predict the ideal defect-free background under specific process conditions based on the current process parameters. Moreover, through adversarial training optimization and high-resolution comparison, the detection accuracy of minute defects is significantly improved. Adversarial training is then conducted using a second-resolution image and the generated third-resolution image until the generator converges. The aforementioned closed-loop optimization process ensures that the generated reference image closely approximates the actual physical imaging in terms of texture detail and noise distribution. Finally, by performing a differential comparison between the actual high-resolution detected image and the generated defect-free image, background noise can be significantly suppressed and subtle defect signals highlighted. This detection mode based on deep learning feature reconstruction effectively solves the problems of false positives and false negatives caused by process fluctuations, achieving high sensitivity and high accuracy in detecting wafer defects in complex environments. Attached Figure Description

[0022] Figure 1 This is a flowchart of a wafer defect detection method based on generative adversarial networks according to the present invention; Figure 2 This is a schematic diagram of the architecture of the adversarial network generated in an embodiment of the present invention.

[0023] Figure reference numerals: 1. Generator; 11. Multilayer perceptron; 12. Elastic weight consolidation regularization term; 2. Discriminator; 21. Consistent discrimination module; 22. Image discrimination module; 3. Standard background features; 4. Normalized process parameters; 5. Defect-free reference image; 6. Discrimination result. Detailed Implementation

[0024] To make the objectives, technical solutions, and advantages of the embodiments of the present invention clearer, the technical solutions in the embodiments of the present invention will be clearly and completely described below. Obviously, the described embodiments are only some embodiments of the present invention, not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention. Unless otherwise defined, the technical or scientific terms used herein should have the ordinary meaning understood by those skilled in the art. The terms "comprising" and similar expressions used herein mean that the element or object preceding the word covers the element or object listed after the word and its equivalents, but does not exclude other elements or objects.

[0025] Combined with appendix Figure 1 - Appendix Figure 2 The specific embodiments of the present invention will be further described in detail below.

[0026] Reference Figure 1 and Figure 2 In some embodiments of this invention, a wafer defect detection method based on generative adversarial networks is proposed. This method aims to address the problem in existing technologies where subtle fluctuations in processes such as photolithography, etching, or thin film deposition cause nonlinear changes in wafer background features, easily leading to misjudgments of background undulations as defects or masking actual defects. The execution flow of the above detection method mainly includes the following steps: The surface of the wafer to be inspected is reasonably divided into multiple independent regions, and one region is selected as a reference region. Then, the first resolution images and corresponding first process parameters of multiple reference wafers of the same specifications at the reference region are obtained to construct a basic data mapping set.

[0027] The first resolution image is precisely aligned at the pixel level, and the first process parameters are normalized to eliminate the underlying process drift error. Based on this, the standard background features 3 of the reference area are extracted.

[0028] The aforementioned standard background feature 3 and the first process parameter are combined as conditions and input into generator 1 in the generative adversarial network to generate a third resolution image of the corresponding reference region.

[0029] During the training phase, a second-resolution image pre-acquired from the reference region is used as a real benchmark sample and input together with the generated third-resolution image into the discriminator 2 to carry out adversarial training until the generator 1 model converges and is fully optimized.

[0030] During the actual inspection phase, the first resolution image of the wafer to be inspected at the target inspection area and its corresponding real-time process parameters are input into the optimized generator 1 to directly generate a third resolution reference image of the target area without defects.

[0031] Finally, the actual high-resolution detection image of the target area is obtained, and it is compared with the generated reference image with high precision. Based on the difference calculation results, it is accurately determined whether there are micro-defects in the target area.

[0032] In some specific embodiments of the present invention, the resolution accuracy configuration for the above-mentioned image follows a specific hierarchical constraint, that is, the accuracy of the first resolution is lower than that of the second resolution, while the accuracy of the third resolution is greater than or equal to that of the second resolution.

[0033] In some specific embodiments of the present invention, at least one second reference region different from the first reference region is further selected during training, and first resolution images of multiple reference wafers of the same specification at the second reference region and corresponding first process parameters are obtained. After performing pixel-level alignment, parameter normalization, and extraction of standard background features 3 on the image data of the second reference region, the background features and process parameters of the new region are used as conditional inputs, and the network weights of the optimized generator 1 are directly loaded as the initial weights for this training. The generator 1 generates a third resolution image of the second reference region. Subsequently, the high-precision second resolution image of the second reference region is used as a real benchmark sample, and adversarial training is continued with the generated image through the discriminator 2 until the generator 1 converges again for the second reference region and is further optimized.

[0034] The above method involves dividing and selecting multiple non-overlapping reference regions on a small number of reference wafers for feature extraction and model training. This multi-region selection fully utilizes the repeatability and symmetry of the semiconductor wafer surface structure, enabling the extraction of sufficiently rich process variation features from a limited number of physical wafer samples. Compared to traditional methods that obtain training samples by physically stacking a large number of wafers from different batches, this method uses background undulations at different spatial coordinates within the wafer as a data source, significantly reducing the dependence on the total number of physical reference wafer samples, substantially shortening the data collection cycle, and lowering measurement costs.

[0035] In some specific embodiments of the present invention, at least one second reference region different from the first reference region is further selected during the training phase, and first resolution images and corresponding first process parameters of multiple reference wafers of the same specification are obtained at the second reference region. Since semiconductor processes often exhibit regular feature distributions at different radial positions or azimuth angles on the same wafer, parallel or sequential sampling of different regions is equivalent to introducing multi-dimensional boundary condition samples into the model training. Under the above measurements, the generative adversarial network can selectively learn the fine background texture of local regions without processing the massive global data of the entire wafer. Employing detection based on local region partitioning effectively reduces the data throughput and complexity of a single training task, thereby reducing the computation and convergence time of the model during training.

[0036] Furthermore, for the feature processing of the second reference region, the first resolution image of the acquired second reference region is first pixel-level aligned, and the corresponding first process parameters are normalized to extract the standard background feature 3 of the second reference region. Subsequently, the standard background feature 3 and the first process parameters of the second reference region are used as conditional inputs to the current network layer of generator 1. At this point, generator 1 does not learn from scratch, but rather makes minor adjustments based on the general physical features learned from existing regions (e.g., the first region). A third resolution image of the second reference region is generated by generator 1, and the second resolution image corresponding to the second region is used as a real benchmark sample for adversarial training. This allows for faster local feature differentiation during the optimization process for subsequent regions, further validating the effectiveness of multi-region sampling in significantly reducing the overall training workload.

[0037] In some specific embodiments of the present invention, the reduction in sample size and training load is manifested in the following ways: by defining sampling point arrays of multiple regions on a limited number of reference wafers, a virtual sample library of different regions can be quickly constructed. When the wafer to be inspected enters the inspection process, the optimized generator 1 already possesses a knowledge base of background noise distribution in different regions of the wafer. At this time, only the first resolution image of the target inspection area and its process parameters need to be input into the generator 1 to accurately generate a defect-free third resolution reference image of that specific region under the current process state. Finally, by obtaining the actual high-resolution inspection image of the region and performing differential comparison with the generated reference image, high-sensitivity capture of micro-defects can be achieved with low sample size and low training cost.

[0038] In some other embodiments of the present invention, from the perspective of actual hardware acquisition methods, the first resolution image can be quickly acquired by a high-speed coarse scan performed by a large field-of-view optical measurement device to meet the throughput requirements of large-area wafer inspection; while the second resolution image used as the training benchmark and the actual high-resolution detection image used for final comparison are acquired by a scanning electron microscope or a high-resolution optical device with a confocal displacement sensing module through a fixed-point fine scan to ensure the absolute fidelity of the microscopic morphology features.

[0039] In some other embodiments of the present invention, in order to completely eliminate process drift errors caused by the evolution of the manufacturing process at the underlying data level, a strict standardization algorithm is used for the normalization processing of the first process parameter. The aforementioned first process parameter essentially reflects the real-time physical state data of the processing machine or chamber where the wafer is located, and it includes at least one or more combined data of chamber pressure, radio frequency power, etching or deposition gas flow rate, and electrostatic chuck temperature distribution characteristics.

[0040] In practice, process parameter sequences of the same dimension are extracted from x reference wafers of the same specifications, and then normalized using the Z-score normalization algorithm to map them uniformly into a dimensionless feature space. The normalized process parameters are then... Strictly satisfy the following mathematical statistics formula: ,in, The original process parameter values ​​are input for the i-th wafer. Let x be the statistical mean of the process parameters corresponding to the x wafers in the reference region. This represents the standard deviation of the parameter set, reflecting the degree of dispersion in the parameter distribution. Specifically, it involves extracting the process parameter sequence of the same dimension from x reference wafers of the same specifications. This represents the real-time physical state of the processing equipment or chamber where the wafer is located, specifically including chamber pressure, RF power, etching or deposition gas flow rate, and electrostatic chuck temperature distribution characteristics. (Normalized result) That is, the original parameters are converted into dimensionless standardized values ​​with a mean of 0 and a standard deviation of 1. The normalized process parameter 4, that is... Instead of being used directly for detection, it is used as a "condition" input to generator 1 along with the standard background feature 3 of the wafer. This allows for the cross-fusion of one-dimensional process features and two-dimensional image space features within the multilayer perceptron 11. By employing this scheme, standardization smooths out the differences in absolute values, enabling the model to focus on process variations and eliminating the nonlinear interference caused by process drift on imaging features at the underlying data level. Furthermore, normalization transforms the originally discrete, dimensionlessly variable physical data into trainable feature vectors.

[0041] In some specific embodiments of the present invention, for image pixel-level alignment and standard background feature 3, this method employs a registration algorithm based on feature points or phase correlation to accurately calculate the spatial transformation matrix of the wafer surface, and uses this transformation matrix to perform high-precision affine transformation registration on x acquired first-resolution images. After registration, these x images are subjected to weighted averaging and principal component analysis at the pixel channel level to filter out randomly distributed imaging noise caused by grain surface roughness or optical sensors, thereby extracting low-frequency image information with a stable topological structure, which is used as the standard background feature 3 for subsequent generation tasks. Through this process, the nonlinear changes of background features that were originally difficult to quantify directly are successfully transformed into feature vectors that can be stably trained by the network.

[0042] Reference Figure 2 In some other embodiments of the present invention, during the conditional input stage, the extracted standard background feature 3 is used as the main image input to the network. Simultaneously, the normalized first process parameters are mapped to scaling and translation parameters required for affine transformation via a multilayer perceptron 11 (MLP). These scaling and translation parameters are then used to dynamically modulate the feature map in the adaptive instance normalization layer, thereby achieving efficient cross-fusion of one-dimensional process features and two-dimensional spatial features in the deep network of generator 1. Specifically, the adaptive instance normalization layer (AdaIN) is used to fuse one-dimensional process features into a convolutional layer of two-dimensional background features.

[0043] In some specific embodiments of the present invention, the discriminator 2 of the generative adversarial network also adopts a multi-branch structure, specifically including an image discrimination module 22 and a consistency discrimination module 21. The defect-free reference image 5 output by the generator 1 is input into the discriminator 2. The image discrimination module 22 is used to divide the input complete image into multiple local receptive field patches and determine the authenticity of each patch one by one; the image discrimination module 22 can also discriminate the local texture of the image. The consistency discrimination module 21 is responsible for specifically evaluating whether the patch features received by the image discrimination module 22 match the conditional consistency of the synchronously input process parameters. The final feedback output of the discriminator 2 is a weighted sum of the outputs of the image discrimination module 22 and the conditional consistency discrimination module, thereby enabling the generator 1 to gradually improve its judgment consistency. The discriminator 2 outputs a discrimination result 6 after the discrimination is completed.

[0044] Furthermore, during training, at least one second reference region different from the first reference region is selected, and first-resolution images and corresponding first-process parameters of multiple reference wafers of the same specification are obtained at this second reference region. This also helps to address the forgetting phenomenon caused by differences in features across multiple regions. Specifically, during the process of generator 1 loading optimized weights and continuing adversarial training for the second reference region, an elastic weight consolidation regularization term 12 is introduced into the global loss function of training, namely... To implement strict parameter update constraints. .in, For the current generator network iteration One parameter, For the first reference region, the training convergence and solidification are performed to solidify the first reference region. Optimal parameters The first element obtained by calculating the diagonal elements of the Fisher information matrix The importance weights of each parameter for the first reference region task This is the regularization intensity coefficient.

[0045] Specifically, the system first trains the parameters until convergence in the first reference region to obtain the optimal parameters. Subsequently, when training second reference regions of different sizes or locations, the optimized weights are loaded as initial weights for continued adversarial training. The Fisher information matrix is ​​used to calculate the importance weight of each parameter for the first region task. If a parameter is crucial for preserving the features of the first region, its... The results are quite high. The above scheme effectively prevents generator 1 from forgetting the background texture features of the first region when learning the second reference region, which facilitates the accurate generation of defect-free backgrounds at different locations on the wafer, such as the central and edge regions.

[0046] In some other embodiments of the present invention, a differential comparison is performed with the generated defect-free reference image 5 to determine whether there are microscopic defects in the target region. Specifically, firstly, the pixel-to-pixel absolute residual between the actually acquired high-resolution detection image and the defect-free third-resolution reference image output by generator 1 is calculated to generate a highly sensitive residual feature map. Next, an adaptive threshold segmentation algorithm is used to perform precise binarization processing on the residual feature map, and then image morphology methods are used to extract the isolated connected regions exposed in the binarized image. When the area threshold of a single connected region in the image is calculated, or the average pixel intensity in the region exceeds the set process defect tolerance threshold, it is strictly determined that a defect truly exists inside the target detection region, and the absolute position coordinates of the defect in the wafer coordinate system and its specific geometric features are recorded.

[0047] In some other specific embodiments of the present invention, those skilled in the art should understand that the provided generative adversarial network is not limited to the above-mentioned two-dimensional light field image detection in actual industrial deployment. It can also be flexibly combined with wafer topography and film thickness distribution matrix obtained by thin film interferometer as auxiliary input conditions through feature stitching. Based on the above-mentioned deep learning feature reconstruction detection, wafer defect detection is no longer limited to the physical "gold standard". Even if the production and processing environment undergoes changes such as slow etching rate drift and fluctuations in the reflectivity of multilayer thin films, resulting in the lack of a perfect reference wafer, the generative adversarial network can still deduce the ideal defect-free background under that specific physical state based on the real-time process parameters of the machine at that time. By constructing a quantitative correlation between "process fluctuations and image features" and implementing a closed-loop generation adversarial optimization method, the generated reference image is ensured to be extremely close to the real physical imaging conditions in terms of grain texture details and background noise distribution. Finally, in the differential comparison stage, the sudden changes and fluctuations caused by complex background features are effectively suppressed, fundamentally solving the frequent misjudgments and missed judgments caused by process fluctuation artifacts in traditional detection technologies. This enables wafer defects such as microparticle contamination, bridging, or open circuits to be detected with high sensitivity and high accuracy in complex process environments.

[0048] In this invention, unless otherwise explicitly specified and limited, the terms "connection," "fixed," etc., should be interpreted broadly. For example, "fixed" can mean a fixed connection, a detachable connection, or an integral part; it can mean a mechanical connection or an electrical connection; it can mean a direct connection or an indirect connection through an intermediate medium; it can mean the internal communication of two components or the interaction between two components, unless otherwise explicitly limited. Those skilled in the art can understand the specific meaning of the above terms in this invention according to the specific circumstances.

[0049] Furthermore, if the embodiments of this invention involve descriptions such as "first" or "second," these descriptions are for descriptive purposes only and should not be construed as indicating or implying their relative importance or implicitly specifying the number of technical features indicated. Therefore, a feature defined with "first" or "second" may explicitly or implicitly include at least one of those features. Additionally, the technical solutions of the various embodiments can be combined with each other, but this must be based on the ability of those skilled in the art to implement them. If the combination of technical solutions is contradictory or impossible to implement, it should be considered that such a combination of technical solutions does not exist and is not within the scope of protection claimed by this invention.

[0050] While embodiments of the present invention have been described in detail above, it will be apparent to those skilled in the art that various modifications and variations can be made to these embodiments. However, it should be understood that such modifications and variations fall within the scope and spirit of the present invention. Furthermore, the present invention described herein may have other embodiments and can be implemented or carried out in various ways.

Claims

1. A wafer defect detection method based on generative adversarial networks, characterized in that, Includes the following steps: The surface of the wafer to be inspected is divided into multiple independent regions; Select one region as a reference region, obtain first resolution images of multiple reference wafers of the same specifications at the reference region and the corresponding first process parameters, and construct a data mapping set; The first resolution image is aligned at the pixel level, the first process parameters are normalized to eliminate process drift error, and the standard background features of the reference area are extracted. The standard background features and the first process parameters are used as conditions to input into the generator of the generative adversarial network to generate a third resolution image of the reference region. Using the second-resolution image of the reference region as a real baseline sample, the generator is trained adversarially against the third-resolution image through a discriminator until the generator converges and is optimized. The first resolution image of the wafer to be inspected at the target inspection area and its process parameters are input into the optimized generator to generate a defect-free third resolution reference image of the target area. Acquire the actual high-resolution detection image of the target area, perform a differential comparison with the generated reference image, and determine whether there is a defect in the target area based on the differential result; The process until the generator converges and is optimized also includes the following steps: Select at least one second reference region that is different from the aforementioned reference region; Acquire a first resolution image of multiple reference wafers of the same specifications in the second reference region and the corresponding first process parameters; The first resolution image of the second reference region is pixel-level aligned, the corresponding first process parameters are normalized, and the standard background features of the second reference region are extracted. The standard background features of the second reference region and the first process parameters are used as input conditions, and the network weights of the optimized generator are loaded as initial weights and input into the generator of the generative adversarial network to generate a third resolution image of the second reference region. Using the second resolution image of the second reference region as the real benchmark sample, the discriminator continues to perform adversarial training with the generated third resolution image of the second reference region until the generator converges again for the second reference region and is further optimized. During the process of the generator loading optimized network weights and continuing adversarial training on the second reference region, to prevent the generator from forgetting the learned features of the reference region, an elastic weight consolidation regularization term is introduced into the training loss function. Perform parameter update constraints: in, For the current generator network iteration One parameter, For the first solidified after training convergence in the reference region One optimal parameter The first element obtained by calculating the diagonal elements of the Fisher information matrix Each parameter has a weight representing the importance of the task in the reference region. This is the regularization intensity coefficient.

2. The wafer defect detection method based on generative adversarial networks according to claim 1, characterized in that, The first resolution accuracy is lower than the second resolution, and the third resolution accuracy is greater than or equal to the second resolution.

3. The wafer defect detection method based on generative adversarial networks according to claim 2, characterized in that, The first resolution image is acquired by high-speed coarse scanning performed by a large field-of-view optical measurement device, while the second resolution image and the actual high-resolution detection image are acquired by fixed-point fine scanning performed by a scanning electron microscope or a high-resolution optical device with a confocal displacement sensing module.

4. The wafer defect detection method based on generative adversarial networks according to claim 1, characterized in that, The step of performing a differential comparison with the generated reference image and determining whether the target region has defects based on the difference results includes the following steps: Calculate the pixel-to-pixel absolute residual between the actual high-resolution detected image and the generated defect-free third-resolution reference image, and generate a residual feature map; The residual feature map is binarized using an adaptive threshold segmentation algorithm, and connected regions in the binarized image are extracted. When the calculated area threshold of a single connected region or the average pixel intensity within that region exceeds the set defect tolerance threshold, it is determined that there is a defect in the target detection region, and the location coordinates and geometric features of the defect are recorded.

5. The wafer defect detection method based on generative adversarial networks according to claim 1, characterized in that, The normalization process for the first process parameters to eliminate process drift error includes: Extract the process parameter sequence of the same dimension from multiple reference wafers of the same specification, and map them to a unified feature space using Z-score normalization; the normalized process parameters Satisfying the formula: in, The original process parameter values ​​are input for the i-th wafer. This represents the statistical mean of the process parameters corresponding to multiple wafers in the reference region. This represents the corresponding standard deviation.

6. The wafer defect detection method based on generative adversarial networks according to claim 1, characterized in that, The first process parameter is the real-time physical status data of the processing machine or chamber where the wafer is located, including at least one or more of the following combined data: chamber pressure, radio frequency power, etching or deposition gas flow rate, and electrostatic chuck temperature distribution characteristics.

7. The wafer defect detection method based on generative adversarial networks according to claim 1, characterized in that, The step of performing pixel-level alignment on the first resolution image and extracting standard background features of the reference region includes: A spatial transformation matrix is ​​calculated using a registration algorithm based on feature points or phase correlation, and the transformation matrix is ​​used to perform affine transformation registration on multiple images of the first resolution. The registered x images are subjected to weighted averaging and principal component analysis at the pixel channel level to filter out randomly distributed imaging noise and extract low-frequency image information with stable topological structure as the standard background features.

8. The wafer defect detection method based on generative adversarial networks according to claim 7, characterized in that, The step of using the standard background features and the first process parameter as conditional inputs includes: The standard background features are used as the main image input of the network. At the same time, the normalized first process parameters are mapped to scaling and translation parameters of affine transformation through a multilayer perceptron. The scaling and translation parameters are used to dynamically modulate the feature map in the adaptive instance normalization layer to cross-fuse one-dimensional process features with two-dimensional spatial features.

9. The wafer defect detection method based on generative adversarial networks according to claim 1, characterized in that, The discriminator of the generative adversarial network includes: The image discrimination module is used to divide the input image into multiple local receptive field patches and determine the authenticity of each patch; A consistency discrimination module is used to evaluate the conditional consistency between the patch features received by the image discrimination module and the input process parameters; The final output of the discriminator is a weighted sum of the outputs of the image discrimination module and the conditional consistency discrimination module.

Images