A wafer inspection data conversion method and system based on physical constraint loss

Abstract

This invention provides a method and system for wafer inspection data conversion based on physical constraint loss, relating to the field of semiconductor manufacturing technology. The method specifically includes: acquiring wafer inspection data at a source wavelength and determining the target wavelength to be converted; based on the wafer inspection data at the source wavelength and the target wavelength, matching a mathematical model from a pre-set physical law model library to describe the physical laws of the wafer inspection scenario; and based on the mathematical model, using a CycleGAN network based on physical constraint loss to map the wafer inspection data at the source wavelength to the target wavelength, obtaining the wafer inspection data at the target wavelength. This invention significantly improves the physical accuracy of cross-wavelength conversion through multi-task loss function design and physical feature embedding layer design in the generator, thereby adapting to the needs of wafer defect detection.

Description

Technical Field

[0001] This invention relates to the field of semiconductor manufacturing technology, and in particular to a wafer inspection data conversion method and system based on physical constraint loss. Background Technology

[0002] Currently, cross-wavelength wafer image conversion often employs image conversion methods based on Cycle-Consistent Generative Adversarial Networks (CycleGANs). This method learns the mapping relationship from the source domain (low wavelength) to the target domain (high wavelength) through adversarial training, enabling the generation of visually realistic images. However, CycleGAN-based image conversion methods suffer from drawbacks in wafer inspection, including indirect modeling of physical laws and accuracy loss. Their implementation relies on adversarial training between the generator G and the discriminator D, along with cycle consistency constraints, to achieve cross-wavelength mapping through visual similarity constraints, but it does not explicitly encode the physical laws of optical scattering. Specifically, it has the following limitations:

[0003] 1. Design flaws of the generator and discriminator: The generator learns the pixel-level mapping from 266nm to 355nm through a convolutional network, and the optimization goal is only to distinguish between real and generated images, rather than constraining physical parameters; the discriminator only judges the visual realism of the image, while ignoring whether the scattered light intensity conforms to physical laws such as scattering intensity.

[0004] 2. Limitations of loss functions: adversarial loss The generated image distribution is consistent with the real image, but physical consistency of wavelength conversion cannot be guaranteed (e.g., the contrast of a 355nm scratch should be lower than that of a 266nm scratch); cyclic consistency loss. It ensures that the bidirectional mapping is invertible; however, the defect characteristics of the intermediate domain (355nm) (such as the attenuation ratio of Mie scattering intensity) are not constrained in the total loss function.

[0005] 3. Unencoded physical parameters: The key physical properties of wafer inspection (such as scattering coefficient and refractive index) are not embedded in the model, causing the scattering intensity, contrast and other parameters of the generated image to deviate from the true optical properties. Summary of the Invention

[0006] To address the shortcomings of the existing technologies, this invention proposes a wafer inspection data conversion method and system based on physical constraint loss through the design of a multi-task loss function and the design of a physical feature embedding layer in the generator. The aim is to explicitly model the physical laws of optical scattering and wavelength dependence, and then improve the physical accuracy of cross-wavelength conversion through the design of the physical constraint loss function, thereby adapting to the needs of defect detection models.

[0007] On the one hand, this invention proposes a wafer inspection data conversion method based on physical constraint loss, which includes the following process:

[0008] Acquire wafer inspection data at the source wavelength and determine the target wavelength to be converted from the wafer inspection data;

[0009] Based on the wafer inspection data at the source wavelength and the target wavelength, a mathematical model for describing the physical laws of the wafer inspection scenario is matched from a pre-set physical law model library.

[0010] Based on the mathematical model, the wafer inspection data at the source wavelength is mapped to the target wavelength using the CycleGAN network based on physical constraint loss, thus obtaining the wafer inspection data at the target wavelength.

[0011] Furthermore, the specific method for matching a mathematical model describing the physical laws of the wafer inspection scenario from a pre-set physical law model library based on the wafer inspection data at the source wavelength and the target wavelength is as follows:

[0012] A mathematical model is constructed to describe the physical laws of wafer inspection scenarios; wherein the mathematical model includes: a scattering intensity wavelength dependence model and a contrast attenuation model based on defect type and scattering mechanism;

[0013] Define a physical parameter library for storing predefined parameters; wherein the predefined parameters include: the source wavelength of wafer inspection data conversion, the target wavelength of wafer inspection data conversion, the scattering intensity scaling factor, and the contrast attenuation coefficient for different defect types;

[0014] Based on the wafer inspection data at the source wavelength, determine the defect type corresponding to the wafer inspection data;

[0015] Based on the wafer inspection data at the source wavelength, the target wavelength, and the defect type, and in conjunction with the physical parameter library, the scattering intensity wavelength dependence model and contrast attenuation model corresponding to the wafer inspection data are selected.

[0016] Furthermore, the method for constructing the mathematical model used to describe the physical laws of the wafer inspection scenario is as follows:

[0017] Based on the universal law of Rayleigh scattering, a wavelength-dependent model of scattering intensity is constructed;

[0018] A contrast attenuation model based on defect type and scattering mechanism is constructed, including: a contrast attenuation model based on Mie scattering and a contrast attenuation model based on Rayleigh scattering; wherein the contrast attenuation model based on Mie scattering is applicable to defects with an aspect ratio greater than a preset threshold; and the contrast attenuation model based on Rayleigh scattering is applicable to defects whose size characteristics satisfy the Rayleigh scattering condition.

[0019] Furthermore, based on the mathematical model, the CycleGAN network based on physical constraint loss is used to map the wafer inspection data at the source wavelength to the target wavelength, resulting in the following specific content of the wafer inspection data at the target wavelength:

[0020] Based on the mathematical model, a CycleGAN network based on physical constraint loss is constructed; wherein the CycleGAN network based on physical constraint loss is: replacing the generator in the traditional CycleGAN network with an improved generator, and constructing a total loss function based on physical constraint according to the mathematical model used to describe the physical laws of the wafer inspection scenario.

[0021] We obtained wafer defect detection datasets at the source wavelength and at the target wavelength, respectively, and constructed a training set.

[0022] The training set is used to train a CycleGAN network based on physical constraint loss, resulting in a trained CycleGAN network based on physical constraint loss.

[0023] The wafer inspection data at the source wavelength is input into the trained CycleGAN network based on physical constraint loss, and the output is the wafer inspection data at the target wavelength.

[0024] Furthermore, the improved generator includes: a physical feature embedding layer, a splicing layer, and an encoder-decoder network;

[0025] The physical feature embedding layer is used to encode the source wavelength and the target wavelength respectively, and to use convolution to map the obtained wavelength encoding into a physical modulation vector;

[0026] The splicing layer is used to splice the physical modulation vector with the wafer detection data at the source wavelength in the channel dimension to obtain a fused feature map;

[0027] The encoder-decoder network is used to extract and reconstruct features from the fused feature map based on the physical modulation vector to obtain wafer detection data at the target wavelength.

[0028] Furthermore, the encoder-decoder network includes, in sequence, a first encoder, a second encoder, a third encoder, a first decoder, a second decoder, and a third decoder; wherein the second encoder is skipped by the first decoder; and the first encoder and the second decoder are skipped by each other.

[0029] The first encoder, second encoder, third encoder, first decoder, second decoder, and third decoder are each implemented using a different number of filters.

[0030] Furthermore, the total loss function based on physical law constraints includes:

[0031] Based on the wafer inspection data at the source wavelength, the ideal target wafer inspection data is calculated. Then, based on the preset constraint of consistent scattered light intensity distribution, the scattering intensity loss between the wafer inspection data at the target wavelength and the ideal target wafer inspection data is calculated.

[0032] Based on the selected contrast attenuation model, the defect region contrast at the source wavelength is calculated according to the wafer inspection data at the source wavelength, and then the defect region contrast at the target wavelength is calculated according to the wafer inspection data at the target wavelength. Finally, the defect type contrast loss between the defect region contrast at the source wavelength and the defect region contrast at the target wavelength is calculated.

[0033] Based on the improved generator, the scattering intensity scaling factor and defect type contrast attenuation coefficient currently learned by the improved generator are obtained, and the physical feature embedding loss is calculated by combining the physical parameter library.

[0034] Based on the CycleGAN network with physical constraint loss, calculate the cycle consistency loss and generative adversarial loss respectively;

[0035] The total loss function based on physical law constraints is obtained by weighted summation of scattering intensity loss, defect type contrast loss, physical feature embedding loss, cycle consistency loss, and generative adversarial loss.

[0036] Furthermore, the specific content of the trained CycleGAN network based on physical constraint loss, obtained by training the training set, is as follows:

[0037] Set the initial weights for scattering intensity loss, defect type contrast loss, physical feature embedding loss, and generative adversarial loss in the total loss function, and set the target thresholds for peak signal-to-noise ratio and structural similarity.

[0038] Set the training cycle and use the training set to train the CycleGAN network based on physical constraint loss;

[0039] During the training process, at the end of each training cycle, the current scattering intensity loss, defect type contrast loss, structural similarity index, and peak signal-to-noise ratio are calculated respectively.

[0040] For each of the following losses: scattering intensity loss, defect type contrast loss, physical feature embedding loss, and generative adversarial loss, the current weight of the loss is adjusted based on the calculated scattering intensity loss, defect type contrast loss, structural similarity index, and peak signal-to-noise ratio, combined with the preset dynamic adjustment rules and the target thresholds for peak signal-to-noise ratio and structural similarity.

[0041] Based on the adjusted weights of each loss, the total loss function is adjusted until the adjusted total loss function meets the preset training completion conditions. Then the training ends, and the trained CycleGAN network based on physical constraint loss is obtained.

[0042] On the other hand, this invention proposes a wafer inspection data conversion method and system based on physical constraint loss, the system comprising:

[0043] The detection data acquisition module is used to acquire wafer detection data at the source wavelength and determine the target wavelength to be converted from the wafer detection data;

[0044] The mathematical model selection module is used to match a mathematical model from a pre-set physical law model library to describe the physical laws of the wafer inspection scenario based on the wafer inspection data at the source wavelength and the target wavelength.

[0045] The detection data conversion module is used to map the wafer detection data at the source wavelength to the target wavelength based on the mathematical model and using a CycleGAN network based on physical constraint loss, so as to obtain the wafer detection data at the target wavelength.

[0046] Thirdly, the present invention proposes a computer program product, including a computer program or instructions, which, when executed by a processor, implements the aforementioned wafer inspection data conversion method based on physical constraint loss.

[0047] The beneficial effects of adopting the above technical solution are as follows:

[0048] This invention injects physical laws into the generator model, transforming them into computable features through numerical translation. This allows the generator to directly learn the quantitative relationships between physical parameters, significantly reducing distortion caused by the lack of physical laws in the generated results. By fusing physical features into the feature generation chain during the generation process, physical laws are synchronously followed during generation, avoiding the lag in traditional methods where physical constraints are only verified at the output end, and significantly improving the physical rationality of the generated results.

[0049] Meanwhile, the method of this invention performs multi-constraint joint optimization of the loss function. By adding physical constraint terms such as scattering intensity loss and defect type contrast loss to the traditional CycleGAN loss function, it avoids missed defects caused by inconsistent scattering patterns and forces the contrast between the generated defect region and the background to conform to the real situation, thereby improving the "distinguishability" of defects. By dynamically adjusting the weights of each loss, it coordinates the contradiction between "physical constraints" and "detection performance", achieving joint optimization of complex tasks and making the generated data more consistent with physical laws. Attached Figure Description

[0050] Figure 1 This is a flowchart of a wafer inspection data conversion method based on physical constraint loss in this embodiment;

[0051] Figure 2 This is a schematic diagram of the CycleGAN network based on physical constraint loss in this embodiment;

[0052] Figure 3 This is a schematic diagram of the improved generator in this embodiment;

[0053] Figure 4 This is a structural diagram of a wafer inspection data conversion system based on physical constraint loss in this embodiment. Detailed Implementation

[0054] To facilitate understanding of this application, specific embodiments of the present invention will be described in further detail below with reference to the accompanying drawings and embodiments. The following embodiments are illustrative of the invention but are not intended to limit its scope. Rather, these embodiments are provided to provide a more thorough and complete understanding of the disclosure of this application.

[0055] Example 1:

[0056] This embodiment presents a wafer inspection data conversion method based on physical constraint loss, such as... Figure 1 As shown, the method includes the following steps:

[0057] Acquire wafer inspection data at the source wavelength and determine the target wavelength to be converted from the wafer inspection data.

[0058] In this embodiment, wafer inspection data conversion is performed for the wafer inspection scenario involving wavelength conversion from 266nm to 355nm, i.e., the source wavelength is 266nm and the target wavelength is 355nm. Furthermore, this embodiment uses wafer inspection images as wafer inspection data.

[0059] Based on the wafer inspection data at the source wavelength and the target wavelength, a mathematical model describing the physical laws of the wafer inspection scenario is matched from a pre-set physical law model library.

[0060] The specific content of the mathematical model used to describe the physical laws of the wafer inspection scenario, matched from a pre-set physical law model library based on the wafer inspection data at the source wavelength and the target wavelength, is as follows:

[0061] A mathematical model is constructed to describe the physical laws of wafer inspection scenarios; wherein the mathematical model includes: a scattering intensity wavelength dependence model and a contrast attenuation model based on defect type and scattering mechanism.

[0062] In this embodiment, for the wafer inspection scenario of wavelength conversion from 266nm to 355nm, two types of physical laws are identified and mathematical formulas are established respectively.

[0063] The method for constructing the mathematical model used to describe the physical laws of the wafer inspection scenario is as follows:

[0064] Based on the universal law of Rayleigh scattering, a wavelength-dependent model of scattering intensity is constructed.

[0065] The wavelength-dependent model for scattering intensity is expressed as:

[0066] ;

[0067] in, As the source wavelength, in this embodiment This corresponds to a deep ultraviolet light source; For the target wavelength, in this embodiment This corresponds to a near-ultraviolet light source; The intensity of the scattered light at the target wavelength; The intensity of the scattered light at the source wavelength.

[0068] A contrast attenuation model based on defect type and scattering mechanism is constructed, including: a contrast attenuation model based on Mie scattering and a contrast attenuation model based on Rayleigh scattering; wherein the contrast attenuation model based on Mie scattering is applicable to defects with an aspect ratio greater than a preset threshold; and the contrast attenuation model based on Rayleigh scattering is applicable to defects whose size characteristics satisfy the Rayleigh scattering condition.

[0069] In this embodiment, for any defect in the wafer image to be inspected, the aspect ratio of the defect is obtained, and when the aspect ratio... At this point, the defect is considered a scratch defect; therefore, a contrast attenuation model based on Mie scattering is chosen, expressed as:

[0070] ;

[0071] in, The contrast of the scratch defect in the target image (355nm); The contrast of the scratch defect in the source image (266nm); is the contrast attenuation coefficient for scratch defects, and This is the experimental calibration value, meaning the scratch contrast at 355nm is approximately 266nm. .

[0072] In this embodiment, the diameter of the defect is obtained. As a size characteristic, when diameter At that time, the defect was considered to be a particle defect. Let be the wavelength of the incident ultraviolet light; in this case, a contrast attenuation model based on Rayleigh scattering is chosen, expressed as:

[0073] ;

[0074] in, The contrast of particle defects in the target image (355nm); The contrast of particle defects in the source image (266nm); is the contrast attenuation coefficient for particle defects, and This is the experimental calibration value, meaning the particle contrast at 355nm is approximately 266nm. .

[0075] In this embodiment, the contrast decay of scratch defects is faster than that of particle defects.

[0076] Define a physical parameter library for storing predefined parameters; wherein the predefined parameters include: the source wavelength of wafer inspection data conversion, the target wavelength of wafer inspection data conversion, the scattering intensity scaling factor, and the contrast attenuation coefficient for different defect types.

[0077] In this embodiment, a physical parameter library is defined to store predefined parameters for use by the loss function, including: the source wavelength for wafer inspection data conversion. The target wavelength for wafer inspection data conversion. Scattering intensity scaling factor Contrast attenuation coefficient of scratch defects and the contrast attenuation coefficient of particle defects. .

[0078] Based on the wafer inspection data at the source wavelength, determine the defect type corresponding to the wafer inspection data.

[0079] In this embodiment, the defect type corresponding to the wafer detection data is determined by obtaining the aspect ratio and diameter of any defect in the wafer image to be inspected.

[0080] Based on the wafer inspection data at the source wavelength, the target wavelength, and the defect type, and in conjunction with the physical parameter library, the wavelength dependence model of scattering intensity and the contrast attenuation model corresponding to the wafer inspection data are obtained.

[0081] Based on the mathematical model, the wafer inspection data at the source wavelength is mapped to the target wavelength using the CycleGAN network based on physical constraint loss, thus obtaining the wafer inspection data at the target wavelength.

[0082] like Figure 2 As shown, based on the mathematical model, the CycleGAN network based on physical constraint loss maps the wafer inspection data at the source wavelength to the target wavelength, resulting in the following specific content of the wafer inspection data at the target wavelength:

[0083] Based on the mathematical model, a CycleGAN network based on physical constraint loss is constructed; wherein the CycleGAN network based on physical constraint loss is: replacing the generator in the traditional CycleGAN network with an improved generator, and constructing a total loss function based on physical constraint according to the mathematical model used to describe the physical laws of the wafer inspection scenario.

[0084] The improved generator includes: a physical feature embedding layer, a splicing layer, and an encoder-decoder network.

[0085] In this embodiment, as Figure 3 As shown, the improved generator is implemented using a U-Net network model (encoder-decoder-jump connection) + physical feature embedding layer + splicing layer.

[0086] The physical feature embedding layer is used to encode the source wavelength and the target wavelength respectively, and to use convolution to map the obtained wavelength encoding into a physical modulation vector.

[0087] In this embodiment, the wavelength is one-hot encoded using a physical feature embedding layer, such as converting the source wavelength of 266nm into an encoded wavelength. Convert the target wavelength of 355nm into encoding Subsequently through Convolution maps wavelength encoding to a 64-channel physical modulation vector, thus matching the number of encoder input channels.

[0088] The splicing layer is used to splice the physical modulation vector with the wafer detection data at the source wavelength in the channel dimension to obtain a fused feature map.

[0089] In this embodiment, a splicing layer is used to splice the physical modulation vector and the wafer detection data at the source wavelength in the channel dimension. That is, the 64-channel physical modulation vector and the wafer detection image at the source wavelength, which is represented as a 1-channel grayscale image, are spliced ​​in the channel dimension to form a 65-channel fused feature map.

[0090] The encoder-decoder network is used to extract and reconstruct features from the fused feature map based on the physical modulation vector to obtain wafer detection data at the target wavelength.

[0091] The encoder-decoder network includes, in sequence, a first encoder, a second encoder, a third encoder, a first decoder, a second decoder, and a third decoder; wherein the second encoder is skipped by the first decoder; and the first encoder and the second decoder are skipped by each other.

[0092] In this embodiment, the image size of the fused feature map is progressively compressed using three encoders, and deep features are extracted to form a 256-channel fused feature map. Then, the image resolution of the fused feature map is restored step-by-step using three decoders, reducing the number of channels to 1. The skip connection connects and fuses the deep semantic features of the encoder and physical feature embedding layer with the corresponding shallow detail features of the decoder, supplementing high-frequency details, avoiding information loss during decoding, and ultimately outputting wafer detection data at the target wavelength.

[0093] The first encoder, second encoder, third encoder, first decoder, second decoder, and third decoder are each implemented using a different number of filters.

[0094] In this embodiment, the encoder portion of the encoder-decoder network uses a convolutional neural network to extract features from the input fused feature map. Specifically, the first encoder, as the initial stage of the encoder portion, is responsible for receiving the 65-channel fused feature map. The first encoder contains 128 filters for initial sampling and basic feature extraction (such as extracting local edge structures in a wafer image). The number of channels in the output feature map of the first encoder is increased to 128, providing a rich foundation of detail for subsequent stages and ensuring that physical parameters (such as wavelength dependence) are fused early. The second encoder, receiving the output of the first encoder, is used for intermediate feature abstraction and further sampling. The number of filters in the second encoder is increased to 192, and the number of channels in the output feature map of the second encoder is increased to 192, preparing for the extraction of high-level semantics and ensuring that the differences in defect types (such as scratches and particles) are encoded in the features. The third encoder, as the final stage of the encoder portion, is used to complete size compression and deep feature extraction. It contains a maximum of 256 filters to output a 256-channel fused feature map.

[0095] In this embodiment, the decoder part of the encoder-decoder network is implemented using deconvolution layers to restore image resolution and reduce the number of channels to 1, thereby outputting wafer detection data at the target wavelength. Specifically, the first decoder is used to initially sample the 256-channel fused feature map. The number of filters in the first decoder is reduced from 256 to 128, enabling the recovery of intermediate resolution from high-level semantic features and expanding the feature map size through deconvolution operations. Therefore, the number of channels in the output feature map of the first decoder is reduced to 128, initially reconstructing the scattering intensity distribution and laying the foundation for subsequent refinement. The second decoder takes over the output of the first decoder and further samples and refines it. The number of filters in the second decoder is reduced from 128 to 64 to enhance efficiency. The second decoder focuses on restoring local details, such as defect contrast and scattered light intensity gradient. Resolution is expanded through deconvolution. The number of channels in the output feature map of the second decoder is reduced to 64, preparing for the final output. The third decoder, as the final stage of the decoder part, reduces the number of filters to 1 to ensure that the output meets physical constraints, completes image reconstruction, and outputs a 1-channel grayscale image as wafer detection data at the target wavelength.

[0096] In this embodiment, the total loss function is constructed by transforming physical laws into trainable loss terms.

[0097] The total loss function based on physical law constraints includes:

[0098] Based on the wafer inspection data at the source wavelength, the ideal target wafer inspection data is calculated, and then, based on the scattering intensity wavelength dependence model, the scattering intensity loss between the wafer inspection data at the target wavelength and the ideal target wafer inspection data is calculated.

[0099] In this embodiment, the ideal target wafer detection data (i.e., the ideal target image) is obtained by multiplying the wafer detection data (i.e., the source image) at the source wavelength by a preset ratio, as shown below:

[0100] ;

[0101] in, Represents the coordinates in the ideal target image Pixel intensity at that location; Represents the coordinates in the source image The pixel intensity at that location.

[0102] To ensure that the intensity distribution of the target image matches that of the ideal target image, the formula for the scattering intensity loss is defined as follows:

[0103] ;

[0104] in, This indicates the loss of scattering intensity; Indicates the height of the image (i.e., the number of vertical pixels); Indicates the width of the image (i.e., the number of horizontal pixels); Coordinates in the target image The pixel intensity at that location.

[0105] In this embodiment, the scattering intensity loss is used to constrain the scattered light intensity of the wafer inspection data at the target wavelength to conform to wavelength dependence. In adversarial testing, this loss acts as a "physical corrector" for the generator, preventing the generator from outputting images that violate optical laws to deceive the discriminator.

[0106] Based on the selected contrast attenuation model, the defect region contrast at the source wavelength is calculated using wafer inspection data at the source wavelength. Then, the defect region contrast at the target wavelength is calculated using wafer inspection data at the target wavelength. Finally, the defect type contrast loss between the defect region contrast at the source wavelength and the defect region contrast at the target wavelength is calculated.

[0107] In this embodiment, the source images are input into a pre-trained U-Net network to obtain defect region masks. The pre-trained U-Net network is trained independently using an encoder-decoder architecture. The encoder progressively compresses the size of the input image and extracts multi-level features, from low-level texture to high-level semantic information, through four downsampling stages (the number of filters increases progressively from 64 to 512). The decoder progressively restores the image resolution through four upsampling stages (each stage includes deconvolution, and the number of filters decreases progressively from 512 to 64), and finally obtains the defect region mask through convolution.

[0108] Then, based on the defect area mask The contrast ratios of the defect regions in the source image and the target image are calculated separately and expressed as follows:

[0109] ;

[0110] ;

[0111] in, The contrast of a certain defect type in the source image (266nm); This represents the total number of pixels within the defective area. mask for defective regions Any pixel in; For pixels in the source region Pixel intensity; This represents the average pixel intensity of the source region; The contrast of the same defect type in the target image (355nm); For pixels in the target area Pixel intensity; This represents the average pixel intensity of the target region.

[0112] Based on the contrast ratios of the defect regions in the source image and the target image, the formula for defect type contrast loss is defined as follows:

[0113] ;

[0114] in, Indicates the type of defect, contrast loss; Indicates the defect type For scratch defects and particle defects A defect in; Represents the pixels in the target image Contrast; Represents the pixels in the source image Contrast; This refers to the contrast attenuation coefficient, which in this embodiment includes the contrast attenuation coefficient for scratch defects. Contrast attenuation coefficient of particle defects .

[0115] In adversarial scenarios, defect type contrast loss is used to force the generator to retain the physical properties of defects, thereby avoiding distortion of details.

[0116] Based on the improved generator, the scattering intensity scaling factor and defect type contrast attenuation coefficient currently learned by the improved generator are obtained, and the physical feature embedding loss is calculated by combining the physical parameter library.

[0117] In this embodiment, a physical feature embedding loss is designed to constrain the scattering intensity scaling factor learned by the physical feature embedding layer in the improved generator. Contrast attenuation coefficient with defect type The formula for defining the physical feature embedding loss is:

[0118] ;

[0119] in, Represents the physical feature embedding loss; The scattering intensity scaling factor is stored in the physical parameter library; These are configurable parameters.

[0120] In this embodiment, the physical feature embedding loss is used to supervise the physical feature embedding layer in the improved generator, so that the learned parameters (such as scattering coefficients) approach the theoretical values. This introduces physical constraints at the feature level and improves the intrinsic rationality of the improved generator.

[0121] Based on the CycleGAN network with physical constraint loss, the cycle consistency loss and generative adversarial loss are calculated respectively.

[0122] In this embodiment, the cycle consistency loss function and generative adversarial loss function in the traditional CycleGAN network are used to calculate the cycle consistency loss and generative adversarial loss.

[0123] In adversarial training, a generative adversarial loss is used to ensure that the generated target is visually realistic. Specifically, by maximizing the generative adversarial loss, the discriminator outputs a high probability (close to 1) for the source image and a low probability (close to 0) for the target image; by minimizing the generative adversarial loss, the target image is misclassified as real by the discriminator. To prevent the game between the generator and the discriminator from getting stuck in trivial solutions (e.g., the generator always outputs the same result) during adversarial training, a cycle consistency loss is used to constrain the generator to maintain the image content structure, improving training stability. The scattering intensity loss, defect type contrast loss, and physical feature embedding loss are combined as a physical constraint loss in the total loss function, and this physical constraint loss is used as a regularization term to work in conjunction with the adversarial loss: the adversarial loss pursues visual realism, while the physical loss ensures physical correctness. During training, the weights are dynamically adjusted to balance the two.

[0124] The total loss function based on physical law constraints is obtained by weighted summation of scattering intensity loss, defect type contrast loss, physical feature embedding loss, cycle consistency loss, and generative adversarial loss.

[0125] ;

[0126] in, Total loss; This is due to the loss of circular consistency. To generate adversarial losses; , , , , All are weights.

[0127] We obtained wafer defect detection datasets at the source wavelength and at the target wavelength, respectively, and constructed a training set.

[0128] The CycleGAN network based on physical constraint loss is trained using the training set to obtain the trained CycleGAN network based on physical constraint loss.

[0129] In this embodiment, the peak signal-to-noise ratio (PSNR) and structural similarity (SSIM) are monitored in real time to quantify pixel fidelity and structural rationality. Specifically, during training, the weights of each loss term in the total loss function are continuously and dynamically adjusted based on the trends of PSNR and SSIM to achieve the optimal effect of wafer inspection data.

[0130] The specific content of the CycleGAN network trained using the training set based on physical constraint loss is as follows:

[0131] Set the initial weights for scattering intensity loss, defect type contrast loss, physical feature embedding loss, and generative adversarial loss in the total loss function, and set the target thresholds for peak signal-to-noise ratio and structural similarity.

[0132] In this embodiment, the initial weights of each loss term in the total loss function and the target threshold in the dynamic adjustment rule are set before training begins, i.e., the initial weights are set. , , , The target threshold for peak signal-to-noise ratio (PSNR) is: The target threshold for structural similarity is .

[0133] Set a training cycle and use the training set to train a CycleGAN network based on physical constraint loss.

[0134] During the training process, at the end of each training cycle, the current scattering intensity loss, defect type contrast loss, structural similarity index, and peak signal-to-noise ratio are calculated respectively.

[0135] For each of the following losses—scattering intensity loss, defect type contrast loss, physical feature embedding loss, and generative adversarial loss—the current weight of that loss is adjusted based on the calculated scattering intensity loss, defect type contrast loss, structural similarity index, and peak signal-to-noise ratio, combined with preset dynamic adjustment rules and target thresholds for peak signal-to-noise ratio and structural similarity.

[0136] In this embodiment, at the end of each training cycle, the quality of the wafer defect detection data at the target wavelength is evaluated, and based on the quality evaluation results, the current weights of each loss are adjusted according to a preset dynamic adjustment rule. The adjustment logic of the dynamic adjustment rule is as follows:

[0137] (1) Adjust the current weights of the scattering intensity loss:

[0138] The calculated scattering intensity loss Loss of scattering intensity relative to the preset target (In this embodiment, we take) ) compare, if Then Increase to This is used to enhance the generator's adherence to the laws governing scattering intensity; if Then Reduce to This is used to avoid excessively suppressing the diversity of image generation.

[0139] (2) Adjust the weights of the contrast loss for defect types:

[0140] Calculate the contrast loss of the defect type Contrast loss compared to the preset target defect type (In this embodiment, we take) ) compare, if Then Increase to This is used to ensure the physical consistency of defect characteristics; if Then Reduce to This is used to avoid over-correction that could lead to distortion of image details.

[0141] (3) Adjust the weights of the physical feature embedding loss:

[0142] Calculate the structural similarity index of wafer defect detection data at the source wavelength. Structural similarity index of wafer defect detection data at the target wavelength .

[0143] ;

[0144] in, Represents the structural similarity index; This represents the local mean of wafer defect detection data at the target wavelength (i.e., the average pixel intensity in the target image). This represents the local mean of wafer defect detection data at the source wavelength. This represents the covariance between the wafer defect detection data at the source wavelength and the wafer defect detection data at the target wavelength. This represents the local variance of wafer defect detection data at the target wavelength. This represents the local variance of wafer defect detection data at the source wavelength. and These are all constant terms, used to stabilize the denominator, and are usually taken as... , ,in Indicates the dynamic range of pixels. , .

[0145] like This indicates that the wafer defect detection data at the target wavelength deviates from the wafer defect detection data at the source wavelength, and... Increase to This is used to force the generator to learn features that better conform to physical laws; if This indicates that the structure meets the rationality standards and that the appropriate measures have been taken. Reduce to It is used to balance physical accuracy and the smoothness of image generation.

[0146] (4) Adjust the weights of the generative adversarial loss:

[0147] Calculate the peak signal-to-noise ratio of the wafer defect detection data at the source wavelength. Peak signal-to-noise ratio of wafer defect detection data at the target wavelength .

[0148] ;

[0149] in, This represents the peak signal-to-noise ratio of wafer defect detection data at the source wavelength and wafer defect detection data at the target wavelength; is the maximum value in the wafer defect detection data; MSE is the mean square error of the wafer defect detection data.

[0150] like This indicates a significant difference, and will Increase to This is used to strengthen the adversarial training between the generator and the discriminator, thereby improving image realism; if This indicates that the difference is small, and will Reduce to This is used to prevent the generator from excessively pursuing pixel-level consistency and ignoring physical laws.

[0151] Meanwhile, during the weight adjustment process, upper and lower limits are set for each weight. If the adjusted weight is higher than its upper limit, the adjusted weight is set to the upper limit; if the adjusted weight is lower than the lower limit, the weight is set to the lower limit. This is to prevent training anomalies caused by the weight of a certain loss term being too large or too small.

[0152] Based on the adjusted weights of each loss, the total loss function is adjusted until the adjusted total loss function meets the preset training completion conditions. Then the training ends, and the trained CycleGAN network based on physical constraint loss is obtained.

[0153] In this embodiment, the improved generator, through the concatenation of the physical feature embedding layer and the feature layer, enables the encoder-decoder network to directly learn wavelength-related physical parameters (such as scattering coefficients) during feature extraction, rather than relying solely on pixel-level mapping. This design allows the improved generator to adaptively adjust the scattering intensity and contrast of the output data, conforming to optical laws. Its optimization objective is to minimize adversarial loss and physical constraint loss, enabling the generated image to "deceive" the discriminator. The discriminator is used to evaluate the authenticity of the output data. Overall, the discriminator takes the complete output data as input, internally extracts global features, and determines whether the output data is a real sample. Locally, the discriminator randomly crops the output data to generate local outputs, and performs real / fake identification on each local output, forcing the generator to conform to physical laws even in detailed areas (such as defect edges), avoiding local distortion. The discriminator's goal is to maximize the adversarial loss and correctly classify the input and output data. Through this multi-scale discrimination, the discriminator provides more refined feedback, guiding the generator to improve quality.

[0154] In this embodiment, the iterative process of adversarial training of the CycleGAN network based on physical constraint loss is as follows: 1) Fix the generator and update the discriminator: Update the discriminator parameters by calculating the generative adversarial loss to improve the discrimination ability. 2) Fix the discriminator and update the generator: The generator optimizes its parameters based on the feedback from the discriminator to minimize the total loss and make the output results more realistic. 3) Iterative optimization: Repeat processes 1) and 2) until the generator and discriminator reach a balance, that is, the discriminator can no longer reliably distinguish between real and fake data, and the iteration is completed.

[0155] The wafer inspection data at the source wavelength is input into the trained CycleGAN network based on physical constraint loss, and the output is the wafer inspection data at the target wavelength.

[0156] Example 2:

[0157] This embodiment presents a wafer inspection data conversion system based on physical constraint loss, such as... Figure 4 As shown, the system includes:

[0158] The detection data acquisition module is used to acquire wafer detection data at the source wavelength and determine the target wavelength to be converted from the wafer detection data;

[0159] The mathematical model selection module is used to match a mathematical model from a pre-set physical law model library to describe the physical laws of the wafer inspection scenario based on the wafer inspection data at the source wavelength and the target wavelength.

[0160] The detection data conversion module is used to map the wafer detection data at the source wavelength to the target wavelength based on the mathematical model and using a CycleGAN network based on physical constraint loss, so as to obtain the wafer detection data at the target wavelength.

[0161] Example 3:

[0162] This embodiment proposes an electronic device, including: one or more processors, and a memory, wherein the memory is used to store instructions, and when the instructions are executed by the one or more processors, the one or more processors execute the aforementioned wafer inspection data conversion method based on physical constraint loss.

[0163] The electronic device may be a mobile phone, computer, or tablet computer, etc., and includes a memory and a processor. The memory stores a computer program, which, when executed by the processor, implements a wafer inspection data conversion method based on physical constraint loss as described in the embodiments. It is understood that the electronic device may also include input / output (I / O) interfaces and communication components.

[0164] The processor is used to execute all or part of the steps in the wafer inspection data conversion method based on physical constraint loss as described in the above embodiments. The memory is used to store various types of data, which may include, for example, instructions for any application or method in an electronic device, as well as application-related data.

[0165] The processor can be implemented as an Application Specific Integrated Circuit (ASIC), Digital Signal Processor (DSP), Programmable Logic Device (PLD), Field Programmable Gate Array (FPGA), controller, microcontroller, microprocessor, or other electronic components, and is used to execute the wafer inspection data conversion method based on physical constraint loss described in the above embodiments.

[0166] Example 4:

[0167] This embodiment proposes a computer-readable storage medium that stores executable instructions. When these instructions are executed, if they are implemented as software functional units and sold or used as independent products, they can be stored in a computer-readable storage medium.

[0168] The computer software product is stored in a storage medium and includes several instructions to cause a computer device (which may be a personal computer, a server, or a network device, etc.) to execute all or part of the steps of the wafer inspection data conversion method based on physical constraint loss described in various embodiments of this application.

[0169] The aforementioned storage media include: flash memory, hard disks, multimedia cards, card-type memory (e.g., SD (Secure Digital Memory Card) or DX (Memory Data Register, MDR) memory), random access memory (RAM), static random-access memory (SRAM), read-only memory (ROM), electrically erasable programmable read-only memory (EEPROM), programmable read-only memory (PROM), magnetic storage, disks, optical discs, servers, APP (Application) application stores, and other media capable of storing program verification codes. These media store computer programs, which, when executed by a processor, can implement the various steps of the aforementioned wafer inspection data conversion method based on physical constraint loss.

[0170] Example 5:

[0171] This embodiment proposes a computer program product, including a computer program or instructions, which, when executed by a processor, implements the aforementioned wafer inspection data conversion method based on physical constraint loss.

[0172] Based on this understanding, the technical solution of this application, in essence, or the part that contributes to the prior art, or part of the technical solution, can be embodied in the form of a computer program product.

[0173] The various embodiments in this application are described in a progressive manner. The same or similar parts between the various embodiments can be referred to each other. Each embodiment focuses on describing the differences from other embodiments.

[0174] The scope of protection of this application is not limited to the embodiments described above. Obviously, those skilled in the art can make various modifications and variations to this disclosure without departing from the scope and spirit of this disclosure. If such modifications and variations fall within the scope of this disclosure and its equivalents, then the intent of this disclosure also includes these modifications and variations.

Claims

1. A wafer inspection data conversion method based on physical constraint loss, characterized in that, This method includes the following steps: Acquire wafer inspection data at the source wavelength and determine the target wavelength to be converted from the wafer inspection data; Based on the wafer inspection data at the source wavelength and the target wavelength, a mathematical model for describing the physical laws of the wafer inspection scenario is matched from a pre-set physical law model library. Based on the mathematical model, the wafer inspection data at the source wavelength is mapped to the target wavelength using the CycleGAN network based on physical constraint loss, thus obtaining the wafer inspection data at the target wavelength.

2. The wafer inspection data conversion method based on physical constraint loss according to claim 1, characterized in that, The specific method for matching a mathematical model describing the physical laws of the wafer inspection scenario from a pre-set physical law model library based on the wafer inspection data at the source wavelength and the target wavelength is as follows: Construct a mathematical model to describe the physical laws of wafer inspection scenarios; The mathematical models mentioned include: a wavelength-dependent scattering intensity model and a contrast attenuation model based on defect type and scattering mechanism; Define a physical parameter library for storing predefined parameters; wherein the predefined parameters include: the source wavelength of wafer inspection data conversion, the target wavelength of wafer inspection data conversion, the scattering intensity scaling factor, and the contrast attenuation coefficient for different defect types; Based on the wafer inspection data at the source wavelength, determine the defect type corresponding to the wafer inspection data; Based on the wafer inspection data at the source wavelength, the target wavelength, and the defect type, and in conjunction with the physical parameter library, the scattering intensity wavelength dependence model and contrast attenuation model corresponding to the wafer inspection data are selected.

3. The wafer inspection data conversion method based on physical constraint loss according to claim 1, characterized in that, The method for constructing the mathematical model used to describe the physical laws of the wafer inspection scenario is as follows: Based on the universal law of Rayleigh scattering, a wavelength-dependent model of scattering intensity is constructed; Contrast attenuation models based on defect type and scattering mechanism are constructed, including: contrast attenuation model based on Mie scattering and contrast attenuation model based on Rayleigh scattering; The contrast attenuation model based on Mie scattering is applicable to defects with an aspect ratio greater than a preset threshold; the contrast attenuation model based on Rayleigh scattering is applicable to defects whose size characteristics satisfy the Rayleigh scattering condition.

4. The wafer inspection data conversion method based on physical constraint loss according to claim 1, characterized in that, Based on the mathematical model, the CycleGAN network based on physical constraint loss is used to map the wafer inspection data at the source wavelength to the target wavelength, resulting in the following specific content of the wafer inspection data at the target wavelength: Based on the mathematical model, a CycleGAN network based on physical constraint loss is constructed; wherein the CycleGAN network based on physical constraint loss is: replacing the generator in the traditional CycleGAN network with an improved generator, and constructing a total loss function based on physical constraint according to the mathematical model used to describe the physical laws of the wafer inspection scenario. We obtained wafer defect detection datasets at the source wavelength and at the target wavelength, respectively, and constructed a training set. The training set is used to train a CycleGAN network based on physical constraint loss, resulting in a trained CycleGAN network based on physical constraint loss. The wafer inspection data at the source wavelength is input into the trained CycleGAN network based on physical constraint loss, and the output is the wafer inspection data at the target wavelength.

5. The wafer inspection data conversion method based on physical constraint loss according to claim 4, characterized in that, The improved generator includes: a physical feature embedding layer, a splicing layer, and an encoder-decoder network; The physical feature embedding layer is used to encode the source wavelength and the target wavelength respectively, and to use convolution to map the obtained wavelength encoding into a physical modulation vector; The splicing layer is used to splice the physical modulation vector with the wafer detection data at the source wavelength in the channel dimension to obtain a fused feature map; The encoder-decoder network is used to extract and reconstruct features from the fused feature map based on the physical modulation vector to obtain wafer detection data at the target wavelength.

6. The wafer inspection data conversion method based on physical constraint loss according to claim 5, characterized in that, The encoder-decoder network includes, in sequence, a first encoder, a second encoder, a third encoder, a first decoder, a second decoder, and a third decoder; wherein the second encoder is skipped by the first decoder; and the first encoder and the second decoder are skipped by each other. The first encoder, second encoder, third encoder, first decoder, second decoder, and third decoder are each implemented using a different number of filters.

7. The wafer inspection data conversion method based on physical constraint loss according to claim 4, characterized in that, The total loss function based on physical law constraints includes: Based on the wafer inspection data at the source wavelength, the ideal target wafer inspection data is calculated. Then, based on the preset constraint of consistent scattered light intensity distribution, the scattering intensity loss between the wafer inspection data at the target wavelength and the ideal target wafer inspection data is calculated. Based on the selected contrast attenuation model, the defect region contrast at the source wavelength is calculated according to the wafer inspection data at the source wavelength, and the defect region contrast at the target wavelength is calculated according to the wafer inspection data at the target wavelength. Then, the defect type contrast loss between the defect region contrast at the source wavelength and the defect region contrast at the target wavelength is calculated. Based on the improved generator, the scattering intensity scaling factor and defect type contrast attenuation coefficient currently learned by the improved generator are obtained, and the physical feature embedding loss is calculated by combining the physical parameter library. Based on the CycleGAN network with physical constraint loss, calculate the cycle consistency loss and generative adversarial loss respectively; The total loss function based on physical law constraints is obtained by weighted summation of scattering intensity loss, defect type contrast loss, physical feature embedding loss, cycle consistency loss, and generative adversarial loss.

8. The wafer inspection data conversion method based on physical constraint loss according to claim 7, characterized in that, The specific content of the CycleGAN network trained using the training set based on physical constraint loss is as follows: Set the initial weights for scattering intensity loss, defect type contrast loss, physical feature embedding loss, and generative adversarial loss in the total loss function, and set the target thresholds for peak signal-to-noise ratio and structural similarity. Set the training cycle and use the training set to train the CycleGAN network based on physical constraint loss; During the training process, at the end of each training cycle, the current scattering intensity loss, defect type contrast loss, structural similarity index, and peak signal-to-noise ratio are calculated respectively. For each of the following losses: scattering intensity loss, defect type contrast loss, physical feature embedding loss, and generative adversarial loss, the current weight of the loss is adjusted based on the calculated scattering intensity loss, defect type contrast loss, structural similarity index, and peak signal-to-noise ratio, combined with the preset dynamic adjustment rules and the target thresholds for peak signal-to-noise ratio and structural similarity. Based on the adjusted weights of each loss, the total loss function is adjusted until the adjusted total loss function meets the preset training completion conditions. Then the training ends, and the trained CycleGAN network based on physical constraint loss is obtained.

9. A wafer inspection data conversion system based on physical constraint loss, used to implement the wafer inspection data conversion method based on physical constraint loss as described in any one of claims 1-8, characterized in that, The system includes: The detection data acquisition module is used to acquire wafer detection data at the source wavelength and determine the target wavelength to be converted from the wafer detection data; The mathematical model selection module is used to match a mathematical model from a pre-set physical law model library to describe the physical laws of the wafer inspection scenario based on the wafer inspection data at the source wavelength and the target wavelength. The detection data conversion module is used to map the wafer detection data at the source wavelength to the target wavelength based on the mathematical model and using a CycleGAN network based on physical constraint loss, so as to obtain the wafer detection data at the target wavelength.

10. A computer program product, characterized in that, Includes a computer program or instructions that, when executed by a processor, implement the wafer inspection data conversion method based on physical constraint loss as described in any one of claims 1-8.

Images