Wafer surface trace particle identification method and system based on multispectral imaging

Abstract

The application discloses a wafer surface trace particle identification method and system based on multispectral imaging, relates to the field of wafer defect detection and metrology technology in a semiconductor manufacturing process, and has the following beneficial effects: through frequency domain transformation on multiple narrow-band images and introduction of a frequency domain background suppression model, selective attenuation is implemented on diffraction peak components corresponding to a periodic structure, then inverse transformation is performed to reconstruct a spatial domain representation, and the masking effect of a fixed mode background on particle detectability is reduced from the source; meanwhile, the frequency domain background suppression model is supervised or self-supervised by using a defect-free reference wafer spectrum and a layout simulation diffraction spectrum, so that the model learns the stable structure and the allowable fluctuation range of the background component, and the robustness and consistency under the conditions of pattern drift and micro-deformation are improved.

Description

Technical Field

[0001] This invention relates to the field of wafer defect detection and measurement technology in semiconductor manufacturing processes, and in particular to a method and system for identifying trace particulate matter on wafer surfaces based on multispectral imaging. Background Technology

[0002] As integrated circuit process nodes continue to shrink and enter advanced processes, the linewidth / spacing of wafer surface patterns is further reduced and the hierarchical structure is more complex. Trace particulate contaminants introduced during the manufacturing process, such as cleaning, handling, deposition and etching, may still cause failures such as open circuits, short circuits or leakage, thus adversely affecting device yield and reliability. Wafer defect detection systems usually need to achieve rapid location and identification of random defects such as particles under high throughput conditions.

[0003] Current wafer defect detection methods typically rely on optical imaging or scattering detection, such as bright-field / dark-field imaging, and perform candidate defect extraction and classification in the image or signal domain. In recent years, some solutions have introduced deep learning target detection networks to improve the automation of defect recognition. However, in high-density periodic pattern regions, the circuit pattern itself generates significant diffraction / scattering background. At the same time, minute process fluctuations such as line edge roughness, critical size fluctuations, and non-uniform film thickness can form so-called wafer noise / nuisance scattering events. This makes it difficult to obtain a stable signal-to-noise ratio and contrast for real defects (especially trace particles whose scattered signals are close to the noise level and whose size is smaller than the effective imaging scale of the system) under single imaging conditions, leading to an increase in missed detections or false alarms.

[0004] To address the aforementioned issues, existing methods for suppressing background, such as spatial filtering, background reconstruction, and differential methods, may still result in false suppression or over-detection under conditions of process drift or local pattern deformation. Therefore, there is an urgent need for a method and system for identifying trace particles on wafer surfaces that can improve defect-background separability in complex periodic pattern backgrounds while also ensuring online detection stability. Summary of the Invention

[0005] In view of the aforementioned existing problems, the present invention is proposed.

[0006] This invention provides a method and system for identifying trace particles on wafer surfaces based on multispectral imaging, which solves the problem of strong scattering of periodic patterns and the submersion of trace particle signals, leading to frequent false alarms and missed detections.

[0007] To solve the above-mentioned technical problems, the present invention provides the following technical solution: In a first aspect, embodiments of the present invention provide a method for identifying trace particulate matter on a wafer surface based on multispectral imaging, comprising: Step S1: Under the N narrow band illumination generated by the programmable illumination unit, the same surface area of ​​the wafer under test with a periodic pattern structure is imaged sequentially. Multiple images are acquired and registered during each band illumination period to obtain a multispectral image sequence. Step S2: Perform a two-dimensional Fourier transform on each of the multispectral image sequences to obtain a spectrum sequence; Step S3: Input the spectrum sequence into the frequency domain background suppression model to output the target spectrum after background suppression. The frequency domain background suppression model is trained under supervised or self-supervised conditions. The training data includes the spectrum obtained by the defect-free reference wafer under the narrow band illumination and / or the diffraction spectrum generated by computational optical simulation from the design layout of the periodic pattern structure. Step S4: Perform inverse Fourier transform on the target spectrum and fuse it by band to obtain a spatial domain feature map; input the spatial domain feature map into the defect detection neural network, and output the location coordinates and classification labels of trace particulate matter.

[0008] As a preferred embodiment of the method for identifying trace particulate matter on wafer surfaces based on multispectral imaging as described in this invention, wherein: N is 3 to 12, and the full width at half maximum (FWHM) of the narrow band is not greater than 20 nm.

[0009] As a preferred embodiment of the method for identifying trace particulate matter on wafer surfaces based on multispectral imaging as described in this invention, the center wavelength of the narrow band is selected from the range of 350nm to 1000nm, and different center wavelengths correspond to the absorption or scattering differences of silicon, silicon dioxide, metal materials and organic polymer materials.

[0010] As a preferred embodiment of the wafer surface trace particle identification method based on multispectral imaging described in this invention, the imaging is bright-field imaging, dark-field imaging, or a combination of both, and the corresponding exposure and illumination intensities are set for different bands so that the image grayscale is within the linear response range of the sensor.

[0011] As a preferred embodiment of the wafer surface trace particle identification method based on multispectral imaging described in this invention, the registration includes registration based on phase correlation, feature point matching or mutual information, used to obtain sub-pixel level alignment results between images of each band.

[0012] As a preferred embodiment of the wafer surface trace particle identification method based on multispectral imaging described in this invention, the frequency domain background suppression model outputs a frequency domain suppression weight map of the same size as the spectrum map, and attenuates the diffraction peak components corresponding to the periodic pattern structure to obtain the target spectrum map.

[0013] As a preferred embodiment of the wafer surface trace particulate matter identification method based on multispectral imaging described in this invention, the defect detection neural network has a dual-branch structure. The first branch uses the spatial domain feature map as input and output position coordinates, and the second branch uses the local frequency domain feature vector obtained from the target spectrum map as input and output classification labels.

[0014] As a preferred embodiment of the wafer surface trace particle identification method based on multispectral imaging described in this invention, the second branch includes a physical guidance module, which stores standard scattering spectrum templates of particles of different materials and sizes pre-calculated based on electromagnetic scattering theory. The second branch assists in classification decision by calculating the matching degree between the local frequency domain feature vector and each of the standard scattering spectrum templates. The second branch calculates the matching degree between the local frequency domain feature vector and the standard scattering spectrum template by: vectorizing the local frequency domain block and normalizing the intensity of the local frequency domain feature vector and the standard scattering spectrum template respectively; introducing scale adaptation and constructing a matching cost with multi-band weighted residuals, wherein the multi-band weights are obtained by statistically normalizing the energy of the signal region and the energy of the background region of the local frequency domain block; and mapping the matching cost to a matching probability through exponential normalization to assist classification decision.

[0015] As a preferred embodiment of the wafer surface trace particle identification method based on multispectral imaging described in this invention, wherein: the defect detection neural network is trained using a joint loss function that includes physical consistency constraints, so that the classification label is consistent with the material properties of the standard scattering spectrum template; The defect detection neural network is trained using a joint loss, which includes positional branch loss, classification cross-entropy loss, and physical consistency constraint term. The physical consistency constraint term constructs a template-induced distribution based on the distance between the sample's local frequency domain representation and the standard scattering spectrum template representation, and constrains the difference between the template-induced distribution and the network's predicted class distribution. The joint loss further includes a spectral consistency constraint term and a contrastive interval constraint term.

[0016] In a second aspect, the present invention provides a wafer surface trace particle identification system based on multispectral imaging, comprising: Programmable illumination unit for generating multiple narrow-band illumination lights; A stage for carrying and moving the wafer under test to align it with the same surface area; An imaging sensor is used to acquire the multispectral image sequence during each narrow-band illumination period; Processor and memory; The memory stores the model parameters of the frequency domain background suppression model and the defect detection neural network, as well as the computer program. The processor is configured to perform: registration and two-dimensional Fourier transform on the multispectral image sequence to obtain a spectrum sequence; perform background suppression on the spectrum sequence using the frequency domain background suppression model to output a target spectrum; perform inverse Fourier transform on the target spectrum and fuse them to obtain a spatial domain feature map; and output the location coordinates and classification labels of trace particulate matter through the defect detection neural network.

[0017] Through the above technical solution, the present invention can achieve at least the following beneficial effects: To address the problem that high-density periodic patterns lead to a dominant diffraction background and make it difficult for trace particle scattering signals to form an effective contrast in the spatial domain, this paper proposes to perform frequency domain transformation on multiple narrow-band images and introduce a frequency domain background suppression model. This model selectively attenuates the diffraction peak components corresponding to the periodic structure and then reconstructs the spatial domain representation through inverse transformation, thereby reducing the masking effect of fixed-mode background on particle detectability from the source.

[0018] To address the issue that micro-deformation of patterns caused by process fluctuations can lead to over-suppression or a large number of false alarms in background suppression algorithms, a supervised or self-supervised training method is used to train the frequency domain background suppression model by simultaneously utilizing the spectrum of a defect-free reference wafer and the diffraction spectrum of the layout simulation. This allows the model to learn the stable structure and allowable fluctuation range of the background components, thereby improving its robustness and consistency under pattern drift and micro-deformation conditions.

[0019] To address the issues of insufficient information in a single band and difficulty in reflecting differences in particle material, multi-narrow-band imaging and band fusion are used to introduce the spectral differences in absorption and scattering of different materials into a unified spatial domain feature expression. Furthermore, the differences in band signal-to-noise ratio can be combined for weighting, so that the more discriminative bands contribute more to the final identification.

[0020] To address the issue that detection models relying solely on spatial feature learning are unstable in strong noise environments and prone to misclassifying particles as graphic textures or process textures as defects, a dual-branch structure is used to decouple localization and classification. The localization branch regresses coordinates on the spatial domain feature map after background suppression, while the classification branch extracts local frequency domain representations on the target spectrum map, thereby improving discrimination stability from two complementary perspectives in the spatial and frequency domains.

[0021] To address the issues of deep learning classification lacking physical constraints and being sensitive to cross-batch illumination intensity drift and band signal-to-noise ratio changes, a standard scattering spectrum template is introduced through a physical guidance module. Local frequency domain features are normalized, scale-adapted, and multi-band weighted matching is performed on the template to form a template matching probability as an auxiliary basis, making classification decisions more dependent on spectral shape and relative energy distribution rather than absolute amplitude.

[0022] To address the issues of classification results being disconnected from physical mechanisms and unstable boundaries between similar material or size categories, a joint loss mechanism is constructed during the training phase, which includes physical consistency, spectral consistency, and contrastive margin. This mechanism aims to make the network's predicted distribution and template-induced distribution more consistent, and to strengthen the similarity of spectral shapes within the same category and the separability of dissimilar categories. This, in turn, improves the reliability and interpretability of classification under residual background and noise conditions. Attached Figure Description

[0023] To more clearly illustrate the technical solutions of the embodiments of this application, the accompanying drawings used in the embodiments will be briefly described below. It should be understood that the following drawings only show some embodiments of this application and should not be regarded as a limitation on the scope of this application.

[0024] Figure 1 This is a flowchart of the method for identifying trace particles on the wafer surface in the embodiments.

[0025] Figure 2 This is a framework diagram of the wafer surface trace particle identification system in the embodiment. Detailed Implementation

[0026] To make the objectives, technical solutions, and advantages of this application clearer, the following detailed description is provided in conjunction with the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are merely illustrative and not intended to limit the scope of this application.

[0027] All terms used in this application (including technical and scientific terms) have the meanings commonly understood by those skilled in the art, unless otherwise defined. It should be noted that the terms used herein should be interpreted in a manner consistent with the context of this specification, and not in an idealized or overly rigid way. Example

[0028] like Figure 1 As shown, this application proposes a method for identifying trace particulate matter on wafer surfaces based on multispectral imaging, including the following steps: Step S1: Under the N narrow-band illumination generated by the programmable illumination unit, the same surface area of ​​the wafer under test with a periodic pattern structure is imaged sequentially. Multiple images are acquired and registered during each band illumination period to obtain a multispectral image sequence. In this embodiment, the multiple narrow-band illumination lights are a set of illumination beams with different center wavelengths and limited bandwidth. The number of narrow bands is subsequently limited to 3-12, and the bandwidth adopts an engineering aperture of no more than 20 nanometers at half maximum width and width. The center wavelength of each narrow band is output by the programmable illumination unit in a preset order, and the band number and acquisition timestamp are recorded at each exposure to maintain the time alignment and band correspondence of the multispectral image sequence in subsequent processing, avoiding mismatches introduced by sequential imaging. For example, the same surface area can be reproduced by the same coordinates of the stage and with the same field of view setting. The acquisition process is based on a single recipe execution to ensure that the illumination intensity, exposure time, and sensor gain are consistent under the same recipe, so that the frequency domain statistics are comparable between different batches. Specifically, multiple images can be understood as a set of images repeatedly acquired under the same narrow-band illumination conditions for the same field of view. The number of images can be set to 1-10 depending on throughput and signal-to-noise ratio requirements, with the principle of minimizing the time interval within the same narrow band to reduce alignment errors caused by mechanical and thermal drift. When acquiring multiple images within the same narrow band, alignment can be completed within that band first, followed by averaging or median fusion to suppress random noise, before proceeding to the cross-band registration process to improve the stability of subsequent spectrogram sequences. Furthermore, in engineering, registration can be achieved by selecting one narrow-band image as a reference image and estimating the two-dimensional geometric transformation of the remaining narrow-band images before resampling them to the reference coordinate system. Subpixel-level alignment results can be obtained through interpolation resampling, maintaining the linear relationship of pixel values ​​during resampling to avoid nonlinear distortion in subsequent two-dimensional Fourier transforms. Optionally, when feature points are insufficient in some narrow bands, phase correlation can be used to obtain the translation amount first, and feature point matching or mutual information strategies can be superimposed when available to compensate for small rotations and scale differences, thereby ensuring that the same physical location in different narrow bands has a consistent pixel mapping in the sequence.

[0029] Step S2 involves performing a two-dimensional Fourier transform on each image in the multispectral image sequence to obtain a spectrum sequence. Similarly, the two-dimensional Fourier transform can be implemented using a discrete fast Fourier transform, performing it independently on each registered narrow-band image to obtain the corresponding complex frequency domain representation. The amplitude is used to characterize the scattering intensity distribution, and the phase is used to preserve structural information. The frequency domain coordinates can use a unified origin and remain consistent throughout the sequence, so that the subsequent frequency domain background suppression model can process the spectrum images of different narrow bands with the same aperture. To ensure the comparability of spectral amplitudes between different narrow bands, each image can be linearly normalized in intensity or equivalently converted according to exposure and illumination intensity before entering the frequency domain. The conversion coefficients are recorded as metadata along with the image sequence to support reproduction and traceability.

[0030] Step S3: Input the spectrum sequence into the frequency domain background suppression model to output the target spectrum after background suppression. The frequency domain background suppression model is trained under supervised or self-supervised conditions. The training data includes the spectrum obtained from a defect-free reference wafer under narrow-band illumination and the diffraction spectrum generated by computational optical simulation of the design layout of the periodic pattern structure. In this embodiment, the frequency domain background suppression model takes the spectrum sequence as input and outputs a frequency domain suppression weight map of the same size as the spectrum or an equivalent frequency domain modulation result. Its engineering meaning is to attenuate the stable diffraction components generated by the periodic pattern structure, while preserving as much as possible the non-periodic components related to trace particle scattering. The attenuation intensity can be expressed by the weights output by the model, which can continuously take values ​​between 0 and 1, where close to zero indicates strong suppression and close to one indicates weak suppression. Furthermore, to take into account the micro-deformation of the pattern in different process batches, the same frequency domain coordinate convention can be maintained during the training and inference stages, and the diffraction peak positions are allowed to have a small range of shifts in the local neighborhood, so that the model learns the allowable drift range of the diffraction components, thereby reducing over-suppression or residual background caused by micro-deformation.

[0031] Furthermore, computational optical simulation based on the design layout can generate a priori diffraction spectra using simulation settings consistent with the actual narrow-band center wavelength and bandwidth, and ensure that its frequency domain sampling interval is consistent with the measured spectrum. This allows it to be directly used as a target representation of the frequency domain background or as a priori constraint for the position of diffraction peaks during training. When there is a systematic deviation between the measured reference spectrum and the simulated spectrum, both can be used together for supervised or self-supervised training, enabling the model to learn the differences introduced by the equipment transfer function and process fluctuations, thereby enhancing its adaptability to real production line data.

[0032] For example, a defect-free reference wafer can be a wafer with the same process and pattern, and confirmed by existing measurement or re-judgment procedures to be free of particle defects. Its spectrum acquisition should use the same narrow band sequence, exposure and illumination intensity settings as the wafer under test, and record the formula version and equipment status during acquisition to reduce system bias caused by formula differences. When the reference data comes from multiple batches, the spectrum amplitude can be statistically normalized by batch, so that the frequency domain background suppression model pays more attention to the spectrum shape and peak structure rather than the absolute amplitude, thereby improving the cross-batch generalization ability.

[0033] Step S4: Perform inverse Fourier transform on the target spectrum and fuse it by band to obtain a spatial domain feature map; input the spatial domain feature map into the defect detection neural network, and output the location coordinates and classification labels of trace particulate matter; Optionally, band fusion can be achieved by pixel-wise weighted summation, channel-wise stitching followed by input to the network, or by statistical aggregation of features of each band. The weights of the weighted summation can be consistent with the multi-band weights described later, or an equal weight setting can be used to form a baseline before calculating the weights. When there are individual narrow band images with missing measurements or abnormal quality, the narrow band can be removed from the fusion and the remaining weights can be normalized to ensure the scale stability of the fusion result and avoid the transmission of missing measurement errors to the input of the defect detection neural network.

[0034] In this embodiment, N is 3 to 12, and the full width at half maximum (FWHM) of the narrow band is no greater than 20 nm; In this embodiment, the center wavelength of the narrow band is selected from the range of 350nm to 1000nm, and different center wavelengths correspond to the absorption or scattering differences of silicon, silicon dioxide, metal materials and organic polymer materials; In this embodiment, the imaging is bright field imaging, dark field imaging, or a combination of both, and the corresponding exposure and illumination intensities are set for different bands so that the image grayscale is within the linear response range of the sensor. In this embodiment, registration includes registration based on phase correlation, feature point matching, or mutual information, which is used to obtain sub-pixel level alignment results between images of different bands. In this embodiment, the frequency domain background suppression model outputs a frequency domain suppression weight map of the same size as the spectrum map, and attenuates the diffraction peak components corresponding to the periodic graphic structure to obtain the target spectrum map. Specifically, the diffraction peak components corresponding to the periodic pattern structure can be determined through two types of existing information: one type comes from the prior diffraction spectrum generated by the design layout simulation, which directly gives the position distribution of the main diffraction peak in the frequency domain coordinates; the other type comes from the measured spectrum of the defect-free reference wafer, which determines the actual peak position by searching for local maxima in the neighborhood of the prior peak position, in order to accommodate the slight shift of the peak position caused by process fluctuations. Furthermore, the size of the diffraction peak neighborhood can be set to the range of several frequency domain sampling points according to the spectral resolution and equipment stability, and the maximum response position in the neighborhood is used as the peak center for each peak position during inference, thereby improving the robustness of peak position positioning without introducing new processes; when the peak position is not obvious or the signal-to-noise ratio is low in a certain narrow band, it can degenerate to perform attenuation only based on the simulation prior, so as to ensure the continuous operation of the processing chain.

[0035] In this embodiment, the defect detection neural network has a dual-branch structure. The first branch uses the spatial domain feature map as the input and output position coordinates, and the second branch uses the local frequency domain feature vector obtained from the target spectrum map as the input and output classification label. In this embodiment, the second branch includes a physical guidance module, which stores standard scattering spectrum templates of particles of different materials and sizes that are pre-calculated based on electromagnetic scattering theory. The second branch assists in classification decision-making by calculating the matching degree between the local frequency domain feature vector and each standard scattering spectrum template. The second branch calculates the matching degree between the local frequency domain feature vector and the standard scattering spectrum template, including: vectorizing the local frequency domain block and normalizing the intensity of the local frequency domain feature vector and the standard scattering spectrum template respectively; introducing scale adaptation and constructing a matching cost with multi-band weighted residuals, wherein the multi-band weights are obtained by statistically normalizing the energy of the signal region and the background region of the local frequency domain block; and mapping the matching cost to a matching probability through exponential normalization, which is used to assist in classification decision-making. Furthermore, the temperature coefficient in exponential normalization is used to adjust the sharpness of the matching probability distribution. Its default value can be moderate to avoid the probability being too flat, resulting in insufficient discrimination, or too sharp, resulting in sensitivity to noise. The adjustable range of the temperature coefficient can follow the engineering range of 1-20 in the original paper, and can be selected based on the optimal point of consistency between the matching probability and the true class on the validation set. Similarly, when the multi-band weights satisfy the normalization constraint, it can avoid the result being dominated by single-band noise due to excessive weight of a single narrow band. Therefore, in actual settings, an empirical constraint can be imposed on the upper limit of the weights and the sum can be kept to one after normalization to improve numerical stability and cross-batch reproducibility.

[0036] In the second branch, the physical guidance module can calculate the matching degree between the local frequency domain feature vector and the standard scattering spectrum template by normalized similarity, multi-band weighting and probabilistic output, so as to reduce the impact of illumination intensity drift and signal-to-noise ratio differences in different bands. In one implementation, when the target spectrogram is cropped to obtain the first... A narrow band of local frequency domain blocks Then, its amplitude is vectorized to obtain the local frequency domain feature vector: (1) In equation (1), Indicates the narrow band number and , Indicates the number of narrowbands. Indicates the first A narrow-band local frequency domain block, Represents frequency domain coordinates, Indicates the first Local frequency domain eigenvectors of a narrow band This indicates that a two-dimensional matrix is ​​expanded into column vectors in a fixed order. Represents complex number amplitude operations; To eliminate the impact of illumination intensity drift on the overall amplitude scaling, intensity normalization is performed on the local vector and the template vector: (2) In equation (2), This represents the normalized local frequency domain eigenvector. Indicates the first in the physical boot module The template in the first The standard scattering spectrum template vector corresponding to each narrow band This represents the normalized template vector. Represents the L2 norm, Describes the normalized stable term and takes to To avoid the denominator being zero Indicates the template category index and , Indicates the total number of template categories; A scale-adaptive coefficient is introduced between the template and the local vector to compensate for the residual magnitude gain difference, with the residual used as the matching cost: (3) In equation (3), Indicates the first The template in the first Scale-adaptive coefficients on a narrow band Indicates transpose. Indicates the scalar solution for the stable term and takes to , Indicates the first The matching cost between a template and the sample to be tested under multi-band weighting Indicates the first Weighting coefficients for each narrow band and satisfying ; To correlate the weights with the signal-to-noise ratio, the band weights are constructed from the signal region energy / background region energy of the local frequency domain block: (4) In equation (4), Indicates the first Signal-to-noise ratio (SNR) for a narrow band Indicates the energy of the signal region. Indicates the energy of the background region. Indicates a stable term and takes to , Indicates a narrowband summation index; The energy in the signal region and the energy in the background region can be obtained by summing the squared magnitudes of the local frequency domain blocks: (5) In equation (5), This represents the set of frequency domain coordinates used to characterize the particle scattering response within a local frequency domain block. Indicates surrounding The set of neighborhood background coordinates is used to characterize the residual periodic structure and noise energy; In this embodiment, the engineering meanings of the signal region and the background region are respectively the subset of frequency domain coordinates within the local frequency domain block that are more relevant to the particle scattering response, and the subset of neighborhood coordinates surrounding the signal region and used to characterize the residual periodic structure and noise energy. The signal region can be determined based on the energy concentration position of the standard scattering spectrum template in each narrow band, and can maintain a fixed shape and fixed coordinate index within the same template category to ensure that the energy statistics between different samples have a consistent caliber. The background region can be a ring-shaped or strip-shaped neighborhood adjacent to the signal region, excluding the diffraction peak neighborhood, to reduce the pollution of the signal-to-noise ratio index by the residual periodic peaks. When the local frequency domain block is small and the sampling points in the background region are insufficient, degradation processing can be performed by shrinking the boundary of the signal region or expanding the boundary of the background region, and the non-zero setting of the numerical stability term can be maintained during weight normalization to avoid division by zero or weight collapse.

[0037] When mapping the matching cost to the matching degree and forming the category probability, the matching score can be set as follows: The template matching probability is obtained by Softmax: (6) In equation (6), Indicates the first The matching probability of a template. This represents the temperature coefficient, ranging from 1 to 20, used to adjust the sharpness of the probability distribution. Indicates a category summation index. Indicates the first The matching score of each template; thus, the matching degree can be used. This indicates that it is used for internal comparison, and can also be used... This indicates that the output is used for classification decision; In the classification process of the second branch, local frequency domain blocks of each narrow band are first obtained by cropping the target spectrum map according to the candidate region of the particles, and the amplitude of the local frequency domain blocks is expanded into local frequency domain feature vectors in a fixed order. The vectorization process corresponds to Equation (1). Subsequently, in order to suppress the overall amplitude scaling caused by the illumination intensity drift, the local frequency domain feature vectors and the standard scattering spectrum template are normalized respectively. The normalization process corresponds to Equation (2). In the matching stage, in order to compensate for the residual gain difference, scale adaptation is introduced and template matching cost is constructed with weighted residuals. The cost calculation corresponds to Equation (3). The band weights in the multi-band weighting are obtained in the matching stage based on the energy statistics of the signal area and the background area of ​​the local frequency domain block, and are normalized. The weight construction and energy statistics correspond to Equations (4) and (5) respectively. The template matching cost is mapped to the matching score and exponentially normalized to obtain the matching probability of each template. The probabilistic output corresponds to Equation (6). The temperature adjustment factor of exponential normalization is preset by the technicians in the method and system configuration stage in combination with the performance of historical samples to control the dispersion of the probability distribution. This matching probability can serve as an auxiliary basis for the classification decision in the second branch, and form a consistent confidence expression with the network classification output.

[0038] Specifically, the above implementation provides a path for calculating the matching degree between local frequency domain features and the standard scattering spectrum template. The matching process is completed in the second branch without changing the physical guidance module's template storage and network branch structure. By vectorizing the local frequency domain blocks, the responses of different frequency domain coordinates are incorporated into a unified representation. Then, normalization is used to weaken the overall amplitude change caused by illumination intensity drift, making the matching focus on the spectral shape and relative energy distribution. Scale adaptation is introduced in the matching stage to compensate for residual gain differences, and weighted residuals are used as a cost to make the differences between the template and the sample in multiple bands quantifiable and comparable. The weights are constructed from the energy of the signal region and the background region, so that bands with higher signal-to-noise ratios or those that better reflect material differences contribute more to the results, thereby reducing the pull of low signal-to-noise bands on misjudgments. After converting the matching cost into a probabilistic form, the classification output has a comparable confidence expression, which is convenient for integration with the network's classification decision and provides a unified measurement basis for the subsequent introduction of physical consistency constraints. In this embodiment, the training of the defect detection neural network adopts a joint loss function that includes physical consistency constraints, so that the classification label is consistent with the material properties of the standard scattering spectrum template. The defect detection neural network is trained using a joint loss, which includes positional branch loss, classification cross-entropy loss, and physical consistency constraint term. The physical consistency constraint term constructs a template-induced distribution based on the distance between the sample's local frequency domain representation and the standard scattering spectrum template representation, and constrains the difference between the template-induced distribution and the network's predicted class distribution. The joint loss further includes a spectral consistency constraint term and a contrastive margin constraint term to enhance the spectral similarity of the same class and improve the discrimination of different classes. During the training phase, the training objective of the defect detection neural network is specifically defined as the joint loss of position branch constraint, classification branch constraint, and physical consistency constraint. The joint loss is used to simultaneously compress position coordinate error, reduce the probability of misclassification, and ensure that the classification result is consistent with the standard scattering spectrum template in the physical guidance module at the frequency domain feature level. In one implementation: when there are a total of [number] items in the same batch When dealing with labeled particle samples, the joint loss can be constructed as follows: (7) In equation (7), Denotes the joint loss function. This represents the position coordinate branch loss. Represents the classification cross-entropy loss. Represents physical consistency constraints. This represents the spectral consistency constraint term. This indicates a comparison constraint term that indicates closer categories and farther categories. This represents the position loss weighting coefficient. This represents the weighting coefficients of the cross-entropy loss. This represents the physical consistency weighting coefficient. Indicates the spectral constraint weight coefficients. Indicates the comparison constraint weight coefficient; Furthermore, the loss weight coefficients are used to balance the gradient scale between location error, classification error, and physical consistency constraints. Their default settings can adopt engineering experience, using location and classification as equal magnitudes, physical consistency and spectral consistency as smaller magnitudes, and contrastive constraints as medium magnitudes, to avoid the physical constraint term suppressing the supervision signal in the early stages of training. The adjustable range of each weight coefficient can be coarsely adjusted in the order of 0.1-10, and then finely adjusted gradually based on the optimal combination of localization error and classification accuracy on the validation set. Optionally, in the early stages of training, the weights of physical consistency and contrastive constraints can be reduced to stabilize convergence, and then gradually increased in the later stages of training to enhance inter-class separation. This adjustment only changes the order and combination of the existing weight coefficients and does not introduce new training procedures.

[0039] When outputting the particle position coordinates in the first branch, a smooth coordinate residual function can be used to suppress the impact of outliers on the gradient. The position branch loss can be written as: (8) In equation (8), Represents the sample index and , Indicates the first Predicted location coordinates vector for each sample Indicates the first The location coordinate label vector of each sample The segmented threshold representing the coordinate residual. Represents the smoothing residual function; The smoothing residual function can be written in terms of components as follows: (9) For example, the segmentation threshold for coordinate residuals can be set to one to five pixels based on engineering experience regarding annotation error and pixel resolution. This provides a smoother gradient in small residual intervals and suppresses the adverse effects of outliers on training in large residual intervals. The threshold should not be zero to avoid instability of the residual function near zero, nor should it be too large to avoid losing effective penalty for moderate errors. Furthermore, when the image resolution or scaling factor differs between different process layers, the threshold can be adjusted proportionally to the pixel scale of the spatial domain feature map to maintain a consistent physical length scale, thereby enhancing cross-layer applicability.

[0040] In equation (9), Represents the coordinate residual vector. express In the component The value on, Indicates the coordinate component index and , The meaning is the same as the definition above; When the second branch outputs the classification label, let the network... The class probability of each sample output is Its supervision label is a one-hot vector. The cross-entropy term can be written as: (10) In equation (10), Indicates the total number of categories. Indicates category index and , Indicates the first Each sample in category The labeled components, Indicates the first Each sample in category The predicted probability components on, Represents a numerically stable term and takes to Specifically, the numerical stability term is used to prevent infinity or non-numerical values ​​from occurring when the probability is logarithmic. Its default value can be on the order of 10 to the power of -6, and can be adjusted within the range of 10 to the power of -8 to 10 to the power of -4. This value should not be too small to avoid numerical underflow on low-probability samples, nor should it be too large to avoid significant bias to the true probability distribution. Similarly, the settings of the normalization stability term, the scaling stability term, and the signal-to-noise ratio stability term all follow the same principle, that is, to maintain strictly positive values ​​and be in a range much smaller than the typical denominator, so as to balance numerical stability and controllable bias.

[0041] The construction of physical consistency constraints is consistent with the template matching logic in the aforementioned implementation: let the local frequency domain representation vector of the second branch before Softmax be... The first physical guidance module The characterization vector corresponding to the standard scattering spectrum template is The degree of deviation between the two can be described by the squared distance: (11) In equation (11), Indicates the first The local frequency domain representation vector of each sample. Indicates the first The frequency domain representation vector of the template class. Indicates the first The sample and the first Distance cost between class templates Represents the L2 norm; Transforming the distance cost into a template-induced distribution and requiring it to be consistent with the network classification distribution, the physical consistency term can be written as a constraint based on KL divergence: (12) In equation (12), Indicates the first Each sample has a class probability component induced by template distance. The temperature coefficient representing the distance to probability is set to 0.5-20. Indicates the index of the Softmax summation and , The physical consistency constraint term is represented by the symbol 'physical consistency', with the remaining symbols retaining their meanings from the previous definitions. This constraint ensures that the network's frequency domain representation and the corresponding category template are statistically consistent when providing classification labels, thus incorporating material-related scattering spectrum differences into the classification decision. Furthermore, the temperature coefficient from distance to probability controls the sensitivity of the template-induced distribution to distance differences. Its default value can be a medium range of 3-10 to ensure higher probabilities for nearby templates while avoiding excessive sharpness that could lead to sensitivity to noise and residual background. When the distance between categories is small and easily confused, this coefficient can be appropriately increased to enhance discriminative power; when distance noise is large, it can be appropriately decreased to improve robustness. Optionally, this coefficient can be fixed during the training phase for stable convergence, and maintained at the same value during the inference phase to ensure comparability between the template-induced distribution and the network output distribution, thereby facilitating the interpretation and reproduction of the physical consistency constraint term.

[0042] Under certain process conditions, similar amplitudes do not necessarily equate to similar spectral shapes. A spectral shape consistency constraint can be added within the same framework, aligning the first and second differences of the representation vectors with the template difference, as follows: (13) In equation (13), Describes the first-order difference operator matrix. Represents a second-order difference operator matrix. Describes the norm, Indicates the first The true class index of each sample This represents the weight coefficient of the second-order difference term, taking values ​​between 0.1 and 2. This represents the spectral consistency constraint term. Similarly, the weight coefficient of the second-order difference term is used to balance the contributions of the first-order and second-order differences to spectral consistency. Its default value can be in the range of 0.5-1 to emphasize the difference in spectral curvature without excessively amplifying high-frequency noise. When the noise in the local frequency domain representation is large, this coefficient can be reduced to avoid the second-order difference amplifying noise. When the spectral bending feature is more sensitive to material differentiation, this coefficient can be increased to enhance the discriminative power. Furthermore, the adjustment of this coefficient should be coordinated with the temperature coefficient and the physical consistency weight coefficient to avoid the simultaneous enhancement of multiple constraints leading to training instability.

[0043] When it's necessary to increase the distance between templates of the same and different classes, a contrastive distance constraint can be added, making the distance between the truth class templates less than the distance between the nearest different class template, written as: (14) In equation (14), Indicates the truncation operator and , The distance interval threshold is defined as 0.5-1, with the remaining symbols retaining their previous meanings. This constraint creates a structure in the template space where similar templates are closer and dissimilar templates are further apart, making templates of similar materials or sizes but different categories less likely to be confused. For example, the distance interval threshold specifies the minimum interval between the ground truth template and the closest dissimilar template. Its default value can be in the middle range of 0.7 to 0.9, to enhance inter-class separation while avoiding excessive penalties for hard samples. The threshold should not be too small to render the interval constraint ineffective, nor should it be too large to prevent training from converging poorly on similar material categories. Furthermore, when the total number of categories is large and the templates are relatively close, a smaller threshold can be used in conjunction with increasing the weight coefficient of the comparative constraint to obtain a smoother optimization process, changing only the combination of existing parameter values ​​without altering the overall process.

[0044] During the training phase, for each batch of labeled particle samples, the joint loss is composed of positional branch constraints, classification branch supervision, and physical consistency constraints. The weighted combination form of the joint loss corresponds to equation (7). The weight factors of each loss term are pre-set by technicians during the training configuration phase and can be adjusted according to the validation set effect. The positional branch measures the deviation between the predicted coordinates and the labeled coordinates, and uses a smooth residual form to suppress gradient fluctuations of abnormal samples. The positional loss and residual function correspond to equations (8) and (9), respectively. The classification branch uses the cross-entropy form to constrain the predicted category distribution to be consistent with the one-hot encoded label. The classification supervision corresponds to equation (10). The physical consistency constraint part first calculates the distance between the local frequency domain representation of the sample and the template representation of each category. The distance calculation corresponds to equation (11). Then, the distance is mapped to the category distribution induced by the template, and the distribution difference is used to measure it to make it consistent with the network's predicted distribution. The physical consistency term corresponds to equation (12). The temperature adjustment factor from distance to distribution is also preset during the training configuration phase. To highlight spectral consistency rather than just similar amplitudes, consistency constraints of first-order and second-order differences are introduced within the same framework, with the spectral consistency term corresponding to equation (13). When it is necessary to enhance the stability of class boundaries, an interval constraint of closer similar classes and farther dissimilar classes is introduced, with the contrastive constraint corresponding to equation (14), where the interval threshold is preset during the training configuration phase. During training, the network parameters are updated using the joint loss mentioned above, so that the classification results and template material properties are consistent at the frequency domain representation level. In this embodiment, the number of samples in a batch is used to statistically average the contribution of each loss term. If some class samples are missing in a batch, the contrastive interval constraint can be made to only take effect on samples with comparable true value classes and dissimilar candidate classes, and the uncalculated terms are treated as zero contribution to avoid introducing non-numerical values. When the energy statistics are unavailable due to out-of-bounds pruning or missing measurements of a local frequency domain block, the corresponding sample can be marked as missing and the physical consistency and spectral consistency related terms can be skipped in the batch, retaining only the position and classification supervision terms, so as to ensure that the training process is continuous and will not be interrupted by individual abnormal samples. Furthermore, when encountering missing measurements in individual narrow-band images during the inference stage, the missing bands can be removed and the multi-band weights can be renormalized using the aforementioned fusion degradation caliber, thereby maintaining output stability without changing the network structure.

[0045] Specifically, the above implementation method revolves around the construction of a joint loss function, combining position coordinate constraints, classification supervision, and physical consistency constraints in the same training objective. The position branch compresses the difference between predicted coordinates and labeled coordinates through smoothing residuals, limiting the impact of individual outliers on the training process. The classification branch uses cross-entropy constraints to reduce the probability of misclassification, aligning the network output with the labeled category. The physical consistency term utilizes the standard scattering spectrum template in the physical guidance module to map the distance between the local frequency domain representation and the template into a template-induced distribution, and constrains this distribution to be consistent with the network classification distribution, ensuring that the classification decision is consistent with the material-related scattering spectrum differences. The spectral shape consistency term emphasizes the spectral shape features through differential alignment, avoiding confusion caused by relying solely on the similarity of overall amplitudes. The contrastive constraint widens the separation between templates of the same and different classes through distance intervals, making the class boundaries more stable and improving reliability under noise and residual background conditions.

[0046] Example 2: Based on Example 1, this application also proposes a wafer surface trace particulate matter identification system based on multispectral imaging, comprising: A programmable illumination unit for generating N narrow-band illumination lights; A stage is used to hold and move the wafer under test to align it with the same surface area. An imaging sensor for acquiring multispectral image sequences during each narrow-band illumination period; Processor and memory; The memory stores the model parameters of the frequency domain background suppression model and the defect detection neural network, as well as the computer program. The processor is configured to perform the following: register and perform two-dimensional Fourier transform on the multispectral image sequence to obtain a spectrum sequence; perform background suppression on the spectrum sequence using the frequency domain background suppression model to output the target spectrum; perform inverse Fourier transform on the target spectrum and fuse them to obtain a spatial domain feature map; and output the location coordinates and classification labels of trace particulate matter through the defect detection neural network. Furthermore, model parameters can be stored in memory as read-only data bound to the recipe version. A computer program sequentially executes registration, two-dimensional Fourier transform, frequency domain background suppression, inverse transform fusion, and neural network inference operations on the processor. To ensure traceability, the program can write existing parameters such as narrowband number, center wavelength configuration, exposure and illumination intensity settings, temperature coefficient, and values ​​of various stability terms into the runtime log, allowing the same input data to reproduce the same output results under the same configuration. Optionally, when communication between the processor and the imaging sensor is briefly interrupted, the acquired multispectral image sequence can be temporarily stored in memory and subsequent steps can continue after communication is restored, avoiding the entire sequence being invalidated due to a single interruption. The minimum duration of the buffer can be set according to the time required to complete the acquisition of a single multispectral sequence, thus avoiding the introduction of additional processes.

[0047] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of this application, and are not intended to limit them. Although this application has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some or all of the technical features therein. Such modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the scope of the technical solutions of the embodiments of this application.

[0048] Furthermore, those skilled in the art will understand that although some embodiments herein include certain features included in other embodiments but not others, combinations of features from different embodiments are meant to be within the scope of this application and form different embodiments. For example, all the embodiments above can be used in any combination. The information disclosed in this background section is intended only to enhance the understanding of the general background of this application and should not be construed as an admission or in any way implying that such information constitutes prior art known to those skilled in the art.

Claims

1. A method for identifying trace particulate matter on wafer surfaces based on multispectral imaging, characterized in that, include: Step S1: Under the N narrow band illumination generated by the programmable illumination unit, the same surface area of ​​the wafer under test with a periodic pattern structure is imaged sequentially. Multiple images are acquired and registered during each band illumination period to obtain a multispectral image sequence. Step S2: Perform a two-dimensional Fourier transform on each of the multispectral image sequences to obtain a spectrum sequence; Step S3: Input the spectrum sequence into the frequency domain background suppression model to output the target spectrum after background suppression. The frequency domain background suppression model is trained under supervised or self-supervised conditions. The training data includes the spectrum obtained by the defect-free reference wafer under the narrow band illumination and / or the diffraction spectrum generated by computational optical simulation from the design layout of the periodic pattern structure. Step S4: Perform inverse Fourier transform on the target spectrum and fuse it by band to obtain a spatial domain feature map; The spatial domain feature map is input into the defect detection neural network, which outputs the location coordinates and classification labels of trace particulate matter. The defect detection neural network includes a classification branch, which takes a local frequency domain feature vector obtained from the target spectrogram as input and outputs the classification label. The classification branch includes a physical guidance module, which stores standard scattering spectrum templates of particles of different materials and sizes that are pre-calculated based on electromagnetic scattering theory. The classification branch assists in classification decisions by calculating the matching degree between the local frequency domain feature vector and each of the standard scattering spectrum templates.

2. The method for identifying trace particulate matter on a wafer surface based on multispectral imaging according to claim 1, characterized in that, The value of N is 3 to 12, and the full width at half maximum (FWHM) of the narrow band is no greater than 20 nm.

3. The method for identifying trace particulate matter on a wafer surface based on multispectral imaging according to claim 2, characterized in that, The center wavelength of the narrow band is selected from the range of 350nm to 1000nm, and different center wavelengths correspond to the absorption or scattering differences of silicon, silicon dioxide, metal materials and organic polymer materials.

4. The method for identifying trace particulate matter on a wafer surface based on multispectral imaging according to claim 3, characterized in that, The imaging is a bright-field imaging, a dark-field imaging, or a combination of both, and the corresponding exposure and illumination intensities are set for different bands so that the image grayscale is within the linear response range of the sensor.

5. The method for identifying trace particulate matter on a wafer surface based on multispectral imaging according to claim 4, characterized in that, The registration includes registration based on phase correlation, feature point matching, or mutual information, used to obtain sub-pixel level alignment results between images of different bands.

6. The method for identifying trace particulate matter on a wafer surface based on multispectral imaging according to claim 5, characterized in that, The frequency domain background suppression model outputs a frequency domain suppression weight map of the same size as the spectrum, and attenuates the diffraction peak components corresponding to the periodic graphic structure to obtain the target spectrum.

7. The method for identifying trace particulate matter on a wafer surface based on multispectral imaging according to claim 6, characterized in that, The defect detection neural network has a dual-branch structure. The first branch uses the spatial domain feature map as input and output position coordinates, and the second branch uses the local frequency domain feature vector obtained from the target spectrum map as input and output classification labels.

8. The method for identifying trace particulate matter on a wafer surface based on multispectral imaging according to claim 7, characterized in that, The second branch calculates the matching degree between the local frequency domain feature vector and the standard scattering spectrum template by: vectorizing the local frequency domain block and normalizing the intensity of the local frequency domain feature vector and the standard scattering spectrum template respectively; introducing scale adaptation and constructing a matching cost with multi-band weighted residuals, wherein the weight of each band in the multi-band weighted residuals is obtained by statistically normalizing the energy of the signal region and the energy of the background region of the local frequency domain block; and mapping the matching cost to a matching probability through exponential normalization, which is used to assist in classification decision-making.

9. The method for identifying trace particulate matter on a wafer surface based on multispectral imaging according to claim 8, characterized in that, The defect detection neural network is trained using a joint loss function that includes physical consistency constraints, so that the classification label is consistent with the material properties of the standard scattering spectrum template. The defect detection neural network is trained using a joint loss, which includes positional branch loss, classification cross-entropy loss, and physical consistency constraint term. The physical consistency constraint term constructs a template-induced distribution based on the distance between the sample's local frequency domain representation and the standard scattering spectrum template representation, and constrains the difference between the template-induced distribution and the network's predicted class distribution. The joint loss further includes a spectral consistency constraint term and a contrastive interval constraint term.

10. A wafer surface trace particulate matter identification system based on multispectral imaging, employing the wafer surface trace particulate matter identification method based on multispectral imaging as described in any one of claims 1 to 9, characterized in that, include: Programmable illumination unit for generating multiple narrow-band illumination lights; A stage for carrying and moving the wafer under test to align it with the same surface area; An imaging sensor is used to acquire the multispectral image sequence during each narrow-band illumination period; Processor and memory; The memory stores the model parameters of the frequency domain background suppression model and the defect detection neural network, as well as the computer program. The processor is configured to perform: registration and two-dimensional Fourier transform on the multispectral image sequence to obtain a spectrum sequence; perform background suppression on the spectrum sequence using the frequency domain background suppression model to output a target spectrum; perform inverse Fourier transform on the target spectrum and fuse them to obtain a spatial domain feature map; and output the location coordinates and classification labels of trace particulate matter through the defect detection neural network.

Images