Light source control method and system based on data analysis

Abstract

The invention provides a light source control method and system based on data analysis, and the method comprises the steps: carrying out the multi-channel spectral characteristic collection of an LED light source through a spectrum analyzer array and an optical fiber probe, obtaining a spectral response curve and an interference image, carrying out the multi-scale analysis processing, and generating a multi-dimensional data tensor; performing interference mode and defect signal decoupling analysis on the multi-dimensional data tensor to obtain an interference mode feature and an interference-defect discrimination result; performing spatial distribution analysis on the interference sensitivity of the wafer surface by adopting preset kernel density estimation and a graph neural network according to the interference mode characteristics and the interference-defect judgment result, and generating an interference sensitivity distribution matrix; according to the interference sensitivity distribution matrix, the current waveform and light beam phase distribution of the LED light source are accurately modulated, intelligent control over the LED light source is achieved, the influence of multi-mode interference on semiconductor detection is effectively eliminated, and the accuracy of nanoscale defect detection is improved.

Description

Technical Field

[0001] The present invention relates to the technical field of light source control, and particularly to a light source control method and system based on data analysis. Background Art

[0002] In the semiconductor manufacturing process, LED light sources are widely used in wafer inspection systems to identify nanoscale defects and structural anomalies. Traditional LED light source control methods mainly adopt fixed wavelength and intensity parameter settings, and are realized through simple switching and brightness adjustment. However, this control method fails to consider the multi-mode emission characteristics of LED light sources, resulting in microscopic interference fringes on the semiconductor surface after the light beam passes through a complex optical system.

[0003] These interference fringes are superimposed on the actual defect signals, making it difficult for the inspection system to distinguish real defects from optical artifacts. Especially in process nodes below 10nm, this interference effect seriously affects the accuracy and sensitivity of inspection, resulting in an increase in missed detections and false positive rates. There is a lack of an effective identification and compensation mechanism for the multi-mode interference effect of LED light sources in the prior art, and it is difficult to meet the requirements of high-precision inspection in semiconductor manufacturing. Summary of the Invention

[0004] The main purpose of the present invention is to solve the technical problem that the existing LED light source control method cannot effectively identify and compensate the multi-mode interference effect, resulting in difficulty in distinguishing real defects from optical artifacts during semiconductor inspection; The first aspect of the present invention provides a light source control method based on data analysis, and the light source control method based on data analysis includes: Performing multi-channel spectral characteristic acquisition on the LED light source through a spectral analyzer array and an optical fiber probe, and performing multi-scale analysis and processing on the acquired spectral response curve and interference image to obtain a multi-dimensional data tensor; Performing decoupling analysis of the interference mode and defect signal on the LED light source according to the multi-dimensional data tensor to obtain an interference mode feature and an interference-defect discrimination result; Performing spatial distribution analysis of the interference sensitivity on the wafer surface according to the interference mode feature and the interference-defect discrimination result through preset kernel density estimation and a graph neural network to obtain an interference sensitivity distribution matrix; Modulating the current waveform and beam phase distribution of the LED light source according to the interference sensitivity distribution matrix to realize the control of the LED light source.

[0005] Optionally, in the first implementation manner of the first aspect of the present invention, the performing multi-channel spectral characteristic acquisition on the LED light source through a spectral analyzer array and an optical fiber probe, and performing multi-scale analysis and processing on the acquired spectral response curve and interference image to obtain a multi-dimensional data tensor includes: Collect the spectral response curve of the LED light source through a spectral analyzer array, perform peak detection and spectral analysis processing on the spectral response curve to obtain the wavelength parameter, intensity parameter, and bandwidth parameter of the light source; Collect the spatial distribution data and interference image of the LED light source through an optical fiber probe, perform Fourier transform and wavelet transform on the spatial distribution data and interference image respectively to obtain the spatial propagation characteristic data and interference fringe parameter of the light source; Perform emission characteristic analysis on the LED light source according to the wavelength parameter, intensity parameter, and bandwidth parameter of the light source and the spatial propagation characteristic data, obtain the emission characteristic parameter of the light source, and perform multi-dimensional organization on the emission characteristic parameter and the interference fringe parameter to obtain a multi-dimensional data tensor.

[0006] Optionally, in the second implementation manner of the first aspect of the present invention, the decoupling analysis of the interference mode and the defect signal on the LED light source according to the multi-dimensional data tensor to obtain the interference mode feature and the interference-defect discrimination result includes: Apply the singular value decomposition algorithm to the multi-dimensional data tensor to decompose it into a light source characteristic mode, an optical path propagation mode, and a wafer surface response mode; Construct a three-mode coupling analysis matrix according to the light source characteristic mode, the optical path propagation mode, and the wafer surface response mode, apply a non-linear manifold learning algorithm to analyze the three-mode coupling analysis matrix, and determine the manifold structure boundary of the interference mode and the defect signal; Apply the persistent homology analysis method to the manifold structure boundary, extract topological structure features, and classify the topological structure features into topological invariant features that do not change with the detection conditions and topological variable features that change with the detection conditions; Analyze the interference image according to the topological invariant features and the topological variable features to generate an interference mode feature and an interference-defect discrimination result.

[0007] Optionally, in the third implementation manner of the first aspect of the present invention, the application of the non-linear manifold learning algorithm to analyze the three-mode coupling analysis matrix to determine the manifold structure boundary of the interference mode and the defect signal includes: Apply the locally linear embedding algorithm in the non-linear manifold learning algorithm to the three-mode coupling analysis matrix, map the high-dimensional data to a low-dimensional manifold space, and obtain the dimensionality-reduced representation data; Construct a k-nearest neighbor graph structure for the dimensionality-reduced representation data, calculate the similarity between nodes in the k-nearest neighbor graph structure according to the Euclidean distance, and obtain a similarity matrix; Calculate the Laplacian matrix eigenvalue spectrum according to the similarity matrix, analyze the eigenvalue distribution in the Laplacian matrix eigenvalue spectrum through the spectral clustering method, and obtain the intrinsic dimension of the data manifold; Construct a manifold diffusion graph based on the intrinsic dimension, calculate the diffusion distance between point pairs in the manifold diffusion graph through a heat kernel function, and obtain a diffusion distance matrix; Apply a contour tracing algorithm to the diffusion distance matrix, identify the regions where the diffusion distance gradient changes abruptly, segment the gradient mutation regions through a density change rate threshold, construct a boundary determination function, and apply the boundary determination function to the data points in the low-dimensional manifold space to obtain the manifold structure boundary of the interference pattern and the defect signal.

[0008] Optionally, in the fourth implementation manner of the first aspect of the present invention, the spatial distribution analysis of the interference sensitivity of the wafer surface based on the interference pattern features and the interference-defect discrimination result through preset kernel density estimation and a graph neural network to obtain an interference sensitivity distribution matrix includes: Perform adaptive kernel density estimation calculation on the interference pattern features, perform probability density modeling on the interference pattern distributions at different positions on the wafer surface, and generate preliminary spatial sensitivity data; Construct a graph network topology structure according to the pre-acquired microstructural features of the wafer surface, map the preliminary spatial sensitivity data onto the graph network topology structure, calculate the mutual influence between the nodes of the graph network topology structure through a pre-deployed graph neural network, and obtain an output result; Combine the interference-defect discrimination result with the output result through a recursive Bayesian estimation algorithm, and estimate and update the interference sensitivity of each node by combining physical prior knowledge to obtain an updated interference sensitivity estimate value; Reorganize and arrange the interference sensitivity estimate values according to the wafer surface coordinate system to generate an interference sensitivity distribution matrix with spatial resolution.

[0009] Optionally, in the fifth implementation manner of the first aspect of the present invention, the combining the interference-defect discrimination result with the output result through a recursive Bayesian estimation algorithm and estimating and updating the interference sensitivity of each node by combining physical prior knowledge to obtain an updated interference sensitivity estimate value includes: Construct a state space model for recursive Bayesian estimation, define the interference sensitivity as a hidden state variable, define the interference-defect discrimination result as an observation variable, and design a state transition equation and an observation equation according to physical optics theory to obtain a recursive Bayesian state space model; Perform particle filtering preprocessing on the output result of the graph neural network according to the recursive Bayesian state space model, generate a set of state particles that conform to the state transition equation through an importance sampling method, and assign an initial weight to each particle to obtain an initial weighted particle set; Substitute the initial weighted particle set into the state transition equation in the recursive Bayesian state space model for state prediction calculation, and combine physical prior knowledge to constrain and adjust the prediction result to obtain the prior state distribution; Substitute the prior state distribution and the interference-defect discrimination result into the observation equation, calculate the observation likelihood function value, and update the weights of each particle according to the observation likelihood function value; Perform a resampling operation according to the updated particle weights, eliminate particles with low weights and replicate particles with high weights, optimize the particle distribution while keeping the total number of particles unchanged, so as to obtain the posterior state distribution; Extract the interference sensitivity estimation values of each node from the posterior state distribution as the updated interference sensitivity estimation result.

[0010] Optionally, in the sixth implementation manner of the first aspect of the present invention, the controlling the LED light source by modulating the current waveform and the beam phase distribution according to the interference sensitivity distribution matrix includes: Divide the wafer surface into multiple control regions according to the sensitivity values in the interference sensitivity distribution matrix, and assign control weight coefficients to each control region; Calculate the multi-stage current waveform parameters of the LED light source according to the control weight coefficients, including rise time, peak current, duty cycle, and fall time, and generate a current control sequence; Map the interference sensitivity distribution matrix to the phase control plane of the spatial light modulator, calculate the compensation phase distribution, and generate a phase modulation signal; Send the current control sequence to the LED driver of the LED light source through a digital signal processing circuit, and at the same time send the phase modulation signal to the spatial light modulator to realize the control of the LED light source.

[0011] The second aspect of the present invention provides a light source control system based on data analysis, and the light source control system based on data analysis includes: A spectrum acquisition module, configured to collect multi-channel spectral characteristics of an LED light source through a spectrum analyzer array and an optical fiber probe, and perform multi-scale analysis and processing on the collected spectral response curve and interference image to obtain a multi-dimensional data tensor; A decoupling analysis module, configured to perform interference pattern and defect signal decoupling analysis on the LED light source according to the multi-dimensional data tensor to obtain interference pattern characteristics and an interference-defect discrimination result; A spatial analysis module, configured to perform spatial distribution analysis on the interference sensitivity of the wafer surface according to the interference pattern characteristics and the interference-defect discrimination result through preset kernel density estimation and a graph neural network to obtain an interference sensitivity distribution matrix; The light source control module is used to modulate the current waveform and beam phase distribution of the LED light source according to the interference sensitivity distribution matrix to achieve the control of the LED light source.

[0012] The above-mentioned light source control method and system based on data analysis collect the multi-channel spectral characteristics of the LED light source through a spectral analyzer array and an optical fiber probe, obtain the spectral response curve and interference image, and perform multi-scale analysis and processing to generate a multi-dimensional data tensor; perform decoupling analysis of the interference pattern and defect signal on the multi-dimensional data tensor to obtain the interference pattern characteristics and interference-defect discrimination results; use the preset kernel density estimation and graph neural network to perform spatial distribution analysis on the interference sensitivity of the wafer surface according to the interference pattern characteristics and interference-defect discrimination results to generate an interference sensitivity distribution matrix; according to the interference sensitivity distribution matrix, accurately modulate the current waveform and beam phase distribution of the LED light source to achieve intelligent control of the LED light source, effectively eliminate the influence of multi-mode interference on semiconductor detection, and improve the accuracy of nano-scale defect detection.

[0013] Other features and advantages of the present invention will be described in the following specification, and, in part, will be obvious from the specification, or will be understood by implementing the present invention. The objectives and other advantages of the present invention are realized and obtained by the structures specifically pointed out in the specification, claims, and drawings.

[0014] To make the above objectives, features, and advantages of the present invention more obvious and understandable, the following specific preferred embodiments are given in conjunction with the accompanying drawings and are described in detail as follows. Description of the Drawings

[0015] Figure 1 Schematic diagram of the first embodiment of the light source control method based on data analysis in the embodiments of the present invention; Figure 2 Schematic diagram of an embodiment of the light source control system based on data analysis in the embodiments of the present invention. Detailed Embodiments

[0016] To make the objectives, technical solutions, and advantages of the embodiments of the present invention clearer, the technical solutions of the present invention will be clearly and completely described below in conjunction with the accompanying drawings. Obviously, the described embodiments are some, but not all, of the embodiments of the present invention. All other embodiments obtained by those of ordinary skill in the art without creative efforts based on the embodiments of the present invention belong to the scope of protection of the present invention.

[0017] As used in the embodiments of the present invention, the terms "comprising" and "having" and any variations thereof are intended to cover non-exclusive inclusion. For example, a process, method, system, product, or device that includes a series of steps or units is not limited to the listed steps or units, but may optionally further include other steps or units not listed, or may optionally further include other steps or units inherent to these processes, methods, products, or devices.

[0018] For ease of understanding of this embodiment, first, a method for controlling a light source based on data analysis disclosed in the embodiments of the present invention will be introduced in detail. As Figure 1 shown, this method includes the following steps: 101. Collect multi-channel spectral characteristics of the LED light source through a spectral analyzer array and an optical fiber probe, and perform multi-scale analysis and processing on the collected spectral response curve and interference image to obtain a multi-dimensional data tensor; In one embodiment of the present invention, the step of collecting multi-channel spectral characteristics of the LED light source through a spectral analyzer array and an optical fiber probe, and performing multi-scale analysis and processing on the collected spectral response curve and interference image to obtain a multi-dimensional data tensor includes: collecting the spectral response curve of the LED light source through a spectral analyzer array, performing peak detection and spectral analysis and processing on the spectral response curve to obtain the wavelength parameter, intensity parameter, and bandwidth parameter of the light source; collecting the spatial distribution data and interference image of the LED light source through an optical fiber probe, performing Fourier transform and wavelet transform on the spatial distribution data and interference image respectively to obtain the spatial propagation characteristic data and interference fringe parameter of the light source; analyzing the emission characteristics of the LED light source according to the wavelength parameter, intensity parameter, and bandwidth parameter of the light source and the spatial propagation characteristic data to obtain the emission characteristic parameter of the light source, and organizing the emission characteristic parameter and the interference fringe parameter in multiple dimensions to obtain a multi-dimensional data tensor.

[0019] Specifically, when collecting the spectral response curve of the LED light source using a spectral analyzer array, it is first necessary to arrange a high-precision spectral analyzer array at key positions between the LED light source and the wafer surface, including multiple spectrometers distributed at different positions along the optical path. The spectral analyzer is usually a CCD or CMOS spectrometer, and its working wavelength range covers the visible light band. During the collection process, different driving parameters are applied to modulate the LED light source, and at the same time, the spectral response curve of the LED is recorded. The Gaussian fitting and peak detection algorithms are applied to the obtained spectral response curve to identify the positions and intensities of the main peak and secondary peaks. Peak detection uses a local maximum search combined with a second derivative inflection point determination method to extract wavelength parameters (central wavelength, secondary mode wavelength), intensity parameters (peak intensity, integrated intensity), and bandwidth parameters (full width at half maximum, 1 / e² width). For example, when detecting a blue LED, the main peak and several secondary peaks can usually be identified, and these multi-mode characteristics have an important impact on the interference phenomenon in semiconductor detection. This step can accurately capture the spectral characteristics of the LED light source, especially the multi-mode characteristics, and provide basic spectral data for accurately characterizing the interference pattern.

[0020] Specifically, when collecting the spatial distribution data and interference images of the LED light source using an optical fiber probe, a high-precision optical fiber probe array is arranged in the optical path system. The optical fiber probe arranges measurement points at a certain interval along the light propagation direction, and forms a probe matrix in the plane perpendicular to the light propagation direction to realize the sampling of the three-dimensional distribution of the light field. The collected spatial distribution data is processed by two-dimensional Fourier transform to obtain the spatial frequency distribution characteristics, including spatial coherence, divergence angle, and wavefront aberration parameters. The continuous wavelet transform is applied to the interference image, and an appropriate mother wavelet function is used to decompose it at multiple scale parameters to obtain the multi-scale decomposition result. Parameters such as the direction, spacing, contrast, and curvature of the interference fringes are extracted from the wavelet coefficients. For example, in a semiconductor detection system, the characteristics of the interference fringes are highly correlated with the microscopic structure of the wafer surface. These spatial sampling and transformation processes can comprehensively characterize the propagation characteristics and interference phenomena of the LED light source in the spatial domain, and provide key data for accurately identifying the interference pattern.

[0021] Specifically, the spectral domain parameters and spatial domain parameters are combined to comprehensively analyze the emission characteristics of the LED light source. The covariance analysis method is used to calculate the correlation matrix between the wavelength parameters, intensity parameters, bandwidth parameters and spatial propagation characteristics. Based on the correlation matrix, the key emission characteristic parameters of the LED light source are extracted, including mode distribution characteristics, phase relationship and spatio-temporal coherence. Then, these emission characteristic parameters are combined with the interference fringe parameters to construct a multi-dimensional data tensor. This tensor has five dimensions: spatial coordinate dimension (x, y, z), spectral parameter dimension and time dimension, and each tensor element contains a complete description of the optical characteristics at this spatio-temporal point. The construction of this multi-dimensional data tensor breaks the separation between the spectral domain and the spatial domain in traditional light source analysis, realizes the unified characterization of the multi-mode interference characteristics of the LED light source, and provides a comprehensive data basis for the subsequent decoupling of interference-defect signals.

[0022] 102. Decouple and analyze the interference pattern and defect signal of the LED light source according to the multi-dimensional data tensor to obtain the interference pattern characteristics and the interference-defect discrimination result; In an embodiment of the present invention, the decoupling and analysis of the interference pattern and defect signal of the LED light source according to the multi-dimensional data tensor to obtain the interference pattern characteristics and the interference-defect discrimination result includes: applying the singular value decomposition algorithm to the multi-dimensional data tensor to decompose it into a light source characteristic mode, an optical path propagation mode and a wafer surface response mode; constructing a three-mode coupling analysis matrix according to the light source characteristic mode, the optical path propagation mode and the wafer surface response mode, and applying a non-linear manifold learning algorithm to analyze the three-mode coupling analysis matrix to determine the manifold structure boundary of the interference pattern and the defect signal; applying the persistent homology analysis method to the manifold structure boundary, extracting topological structure characteristics, and classifying the topological structure characteristics into topological invariant characteristics that do not change with the detection conditions and topological variable characteristics that change with the detection conditions; analyzing the interference image according to the topological invariant characteristics and the topological variable characteristics to generate the interference pattern characteristics and the interference-defect discrimination result.

[0023] Specifically, when applying the singular value decomposition algorithm to a multi-dimensional data tensor, the higher-order singular value decomposition (HOSVD) method is first used to decompose the five-dimensional data tensor into the product of a core tensor and modal matrices. HOSVD performs matrixization on each dimension and then applies the standard singular value decomposition to obtain three main modes: the light source characteristic mode, the optical path propagation mode, and the wafer surface response mode. The light source characteristic mode characterizes the wavelength, intensity, and bandwidth distribution characteristics of the LED; the optical path propagation mode includes the spatial coherence and diffraction characteristics of light on the propagation path; and the wafer surface response mode reflects how the microscopic structure of the wafer interacts with the incident light. In an actual semiconductor detection system, when detecting a 10nm process wafer, the multi-mode characteristics of the light source can be clearly separated through HOSVD. For example, the interference effect between the main peak and the secondary peak of a blue LED can be observed. This three-mode decomposition transforms the interference problem from the original high-dimensional complex representation into an analyzable low-dimensional structure, laying a foundation for subsequent analysis.

[0024] Specifically, when constructing a three-mode coupling analysis matrix based on the three modes, the tensor product operation is used to combine the light source characteristic mode, the optical path propagation mode, and the wafer surface response mode to form a third-order tensor. Each element of this tensor represents the coupling strength of the three modes under a specific combination. The locally linear embedding (LLE) algorithm is applied to this coupling analysis matrix. This is a non-linear manifold learning method that can map high-dimensional data to a low-dimensional space while preserving the intrinsic structure of the data. The LLE algorithm first finds k nearest neighbors for each data point, then reconstructs the point using a linear combination of these neighbors, and finally maintains this reconstruction relationship in the low-dimensional space. By calculating the eigenvectors of the reconstruction weight matrix, the natural interface between the interference pattern and the defect signal in the low-dimensional manifold space is determined. When detecting complex structures such as the wafer edge and contact holes, this interface can distinguish the interference fringes caused by the multi-mode characteristics of the LED from the actual defect signals, significantly improving the detection accuracy.

[0025] Specifically, when applying the persistent homology analysis method to the boundary of the manifold structure, the discrete Morse theory is used to construct the topological representation of the data. Persistent homology is a topological data analysis tool that extracts the topological structure of data by observing the emergence and disappearance of topological features (such as connected components, loops, cavities) at different scales. This method first constructs a filtered complex of data points and then calculates the persistent homology groups of different dimensions to generate a persistence diagram. The topological features extracted from the persistence diagram can be classified into two categories according to their stability: Topological invariant features are topological features that remain stable when detection conditions (such as light source parameters, temperature, etc.) change, representing the inherent characteristics of the interference pattern; topological variable features change with the detection conditions and are usually related to defect signals. For example, when detecting the patterns on a wafer after lithography, the fringe patterns generated by LED multimode interference are topologically represented as one-dimensional homology classes with low persistence, while real defects are represented as features with high persistence. This topology-based classification method can essentially distinguish interference artifacts from real defects and overcome the limitations of traditional image processing methods.

[0026] Specifically, when analyzing the interference image according to the topological invariant features and topological variable features, a pixel-level classification method based on a discriminant function is adopted. First, the topological features are mapped back to the original image space to construct a feature vector field. For each pixel point, the similarity between its feature vector and the topological invariant / variable features is calculated to generate an interference pattern probability map and a defect signal probability map. Through threshold segmentation and edge enhancement processing, the final interference pattern feature representation and the interference-defect discrimination result are obtained. This topology-based image analysis method performs particularly well when dealing with complex wafer surfaces (such as 3D NAND structures with deep aspect ratio trenches), and can effectively distinguish interference fringes caused by the multimode characteristics of LEDs from nanoscale surface defects such as particle contamination, micro-scratches or pattern deformation. Compared with traditional intensity- or contrast-based analysis methods, this topology feature-driven analysis method is more robust and less affected by noise and light intensity fluctuations, providing more reliable defect recognition results for semiconductor inspection.

[0027] Further, applying the non-linear manifold learning algorithm to analyze the three-modal coupling analysis matrix to determine the manifold structure boundary between the interference pattern and the defect signal includes: applying the locally linear embedding algorithm in the non-linear manifold learning algorithm to the three-modal coupling analysis matrix to map the high-dimensional data to the low-dimensional manifold space and obtain the reduced-dimensional representation data; constructing a k-nearest neighbor graph structure for the reduced-dimensional representation data, calculating the similarity between nodes in the k-nearest neighbor graph structure according to the Euclidean distance, and obtaining the similarity matrix; calculating the eigenvalue spectrum of the Laplacian matrix according to the similarity matrix, analyzing the eigenvalue distribution in the eigenvalue spectrum of the Laplacian matrix through spectral clustering method, and obtaining the intrinsic dimension of the data manifold; constructing a manifold diffusion graph based on the intrinsic dimension, calculating the diffusion distance between point pairs in the manifold diffusion graph through the heat kernel function, and obtaining the diffusion distance matrix; applying the contour tracking algorithm to the diffusion distance matrix, identifying the region of sudden change in the diffusion distance gradient, segmenting the gradient mutation region through the density change rate threshold, constructing a boundary decision function, and applying the boundary decision function to the data points in the low-dimensional manifold space to obtain the manifold structure boundary between the interference pattern and the defect signal.

[0028] Specifically, when applying the locally linear embedding (LLE) algorithm to the three-modal coupling analysis matrix, first input this high-dimensional matrix into the LLE processing flow. The LLE algorithm is a non-linear dimensionality reduction method, which is particularly suitable for processing data distributed along a low-dimensional manifold in a high-dimensional space. In the semiconductor detection scenario, the specific operation is to run the LLE algorithm program on a high-performance computing cluster to map the original three-modal coupling analysis matrix to a low-dimensional space (usually 2-5 dimensions). In the mapping parameter setting, it is necessary to select an appropriate number of neighborhoods according to the data scale. For the multi-mode interference problem of LED light sources, the number of neighborhoods is usually set to 0.5%-1% of the total number of data points. For example, when processing the detection data of 7nm process node wafers, the LLE algorithm can compress the high-dimensional data including the light source mode, propagation mode, and surface response mode into a three-dimensional manifold space, while retaining the essential differences between the interference pattern and the defect signal. This dimensionality reduction processing not only greatly reduces the computational complexity, but more importantly, reveals the inherent low-dimensional structure of the data, making the distribution patterns of the interference pattern and the defect signal more obvious in the low-dimensional space.

[0029] Specifically, when constructing a k-nearest neighbor graph structure for the dimensionality-reduced representation data, the spatial relationship between data points needs to be calculated according to the Euclidean distance metric. First, for each data point after dimensionality reduction, calculate its Euclidean distance from all other points, and select k nearest neighbors for each point (the value of k is usually selected between 10 and 30 and adjusted according to the data density). Based on these neighbor relationships, construct a k-nearest neighbor graph, where the nodes in the graph represent data points and the edges represent neighbor relationships. Then calculate the similarity between nodes. The similarity is defined as the reciprocal of the Euclidean distance between point pairs or the value after being processed by the Gaussian kernel function to obtain a complete similarity matrix. In the actual semiconductor detection application, this graph structure representation can effectively capture the essential differences between the specific patterns formed by LED multi-mode interference and the random defect distribution. For example, the data points generated by the interference pattern often show a regular connection structure in the k-nearest neighbor graph, while the defect signals show a discrete and irregular sub-graph structure.

[0030] Specifically, when calculating the Laplacian matrix according to the similarity matrix, first construct the degree matrix of the graph (a diagonal matrix, where the diagonal elements are the sum of the number of edges of each node), and then subtract the similarity matrix from the degree matrix to obtain the Laplacian matrix. Perform eigenvalue decomposition on this matrix to obtain the eigenvalue spectrum. By analyzing the "spectral gap" (the significant gap between adjacent eigenvalues) of the eigenvalue distribution, use the spectral clustering method to determine the intrinsic dimension of the data manifold. In the LED multi-mode interference analysis, the distribution pattern of the eigenvalue spectrum directly reflects the separation degree of the interference pattern and the defect signal in the manifold space. For example, when processing the detection data of wafers with FinFET structures, there is usually an obvious spectral gap after the 3rd - 4th eigenvalues in the eigenvalue spectrum, indicating that the intrinsic dimension of the data is 3 or 4. At this time, the interference pattern and the defect signal can be effectively separated in this dimensional space.

[0031] Specifically, when constructing a manifold diffusion graph based on the determined intrinsic dimension, use the heat kernel function as the diffusion kernel to calculate the diffusion distance between any two points on the manifold. The parameter settings of the heat kernel function are adjusted according to the local curvature characteristics of the manifold, which reflects the rate difference of the heat diffusion process in different regions. By solving the diffusion equation, obtain the diffusion distance matrix representing the similarity between point pairs. Apply the contour tracing algorithm to this matrix to identify the regions where the diffusion distance gradient changes abruptly. These regions usually correspond to the natural interface between the interference pattern and the defect signal. Use an adaptive density change rate threshold (usually set to 1.5 - 2 times the local average gradient) to segment the gradient mutation regions and construct a boundary determination function. Apply this function to each data point in the low-dimensional manifold space to accurately identify whether it belongs to the interference pattern or the defect signal. In actual applications, such as detecting minute defects in DRAM capacitor trenches, this boundary determination method based on diffusion distance can accurately distinguish the regular interference fringes caused by the LED multi-mode characteristics from the real structural defects, significantly reducing the false positive detection results and improving the accuracy and reliability of the detection.

[0032] 103. Perform spatial distribution analysis on the interference sensitivity of the wafer surface based on the interference pattern features and the interference-defect discrimination results through preset kernel density estimation and a graph neural network, and obtain an interference sensitivity distribution matrix. In one embodiment of the present invention, the performing spatial distribution analysis on the interference sensitivity of the wafer surface based on the interference pattern features and the interference-defect discrimination results through preset kernel density estimation and a graph neural network to obtain an interference sensitivity distribution matrix includes: performing adaptive kernel density estimation calculation on the interference pattern features, performing probability density modeling on the interference pattern distributions at different positions on the wafer surface, and generating preliminary spatial sensitivity data; constructing a graph network topology according to the microstructural features of the wafer surface obtained in advance, mapping the preliminary spatial sensitivity data onto the graph network topology, calculating the mutual influence between the nodes of the graph network topology through a pre-deployed graph neural network, and obtaining an output result; combining the interference-defect discrimination result with the output result through a recursive Bayesian estimation algorithm, and estimating and updating the interference sensitivity of each node in combination with physical prior knowledge to obtain an updated interference sensitivity estimation value; reorganizing and arranging the interference sensitivity estimation values according to the wafer surface coordinate system to generate an interference sensitivity distribution matrix with spatial resolution.

[0033] Specifically, when performing adaptive kernel density estimation calculation on the interference pattern features, a variable bandwidth kernel density estimation method is adopted, which can automatically adjust the kernel bandwidth according to the data distribution characteristics. In specific implementation, first convert the interference pattern features into spatial distribution data on the wafer surface, with the wafer coordinates (x, y) as the index. Then apply a multivariate Gaussian kernel function at each position, and the kernel bandwidth is automatically adjusted according to the local data density, using a wider bandwidth in sparse data regions and a narrower bandwidth in dense data regions. The calculation process uses a numerical integration method with an adaptive step size, divides the entire wafer surface into a high-resolution grid (such as 50nm×50nm), and calculates the probability density of the interference pattern at each grid point. For example, when processing the SRAM region with a regular array structure, the kernel density estimation can accurately reflect the change in interference sensitivity caused by the microstructure edges, while showing a relatively uniform distribution in the flat region. This adaptive kernel density estimation method can accurately reflect the interference pattern distribution characteristics at different positions on the wafer surface, generate preliminary spatial sensitivity data, and provide a basis for subsequent accurate analysis.

[0034] Specifically, when constructing a graph network topology based on the pre-acquired microscopic structure characteristics of the wafer surface, it is first necessary to extract the microscopic structure information of the wafer surface from the wafer design file (GDSII). The wafer surface is represented as a complex graph network, where the nodes represent specific regions on the surface (such as key structures like lines, contact holes, gates, etc.), and the edges represent the physical adjacency relationships and optical coupling relationships between regions. In this graph network, attribute features such as structure type, height, and width are assigned to each node, and relationship features such as distance and direction are assigned to each edge. Then, the preliminary spatial sensitivity data is mapped as the initial feature of the nodes onto this graph network. Next, the pre-deployed graph convolutional neural network (GCN) algorithm is run to perform multi-layer graph convolutional operations. Each layer of operation aggregates the information of adjacent nodes and updates the representation of the current node. This message-passing mechanism can simulate the spatial propagation characteristics of the interference effect on the wafer surface, calculate the mutual influence between nodes, and finally output the interference sensitivity considering the microscopic structure characteristics. For example, when processing FinFET devices with complex three-dimensional structures, the graph neural network can accurately capture the complex interference pattern interactions between fin structures of different heights.

[0035] Specifically, when combining the interference-defect discrimination result with the output result through the recursive Bayesian estimation algorithm, a state-observation framework is constructed. In this framework, the interference sensitivity is defined as the hidden state variable, and the interference-defect discrimination result is defined as the observation variable. Recursive Bayesian estimation realizes data fusion by alternately executing the prediction step and the update step. In the prediction step, the output result of the graph neural network is used as prior information, and combined with the interference pattern propagation law in physical optics theory (such as Fresnel diffraction and the principle of light wave coherence), the interference sensitivity distribution of each node is predicted. In the update step, the interference-defect discrimination result is used as the observation input, and the posterior probability distribution is calculated through Bayes' formula to update the interference sensitivity estimation of each node. In practical applications, such as detecting 3D NAND memories with deep trench structures, recursive Bayesian estimation can organically combine the interference pattern analysis result with the actual defect distribution pattern, greatly improving the recognition accuracy of interference-sensitive regions in high aspect ratio structures. This data fusion method based on the Bayesian framework not only considers the algorithm prediction result but also integrates physical prior knowledge, making the interference sensitivity estimation more in line with the actual optical phenomenon.

[0036] Specifically, when reorganizing and arranging the updated interference sensitivity estimates according to the wafer surface coordinate system, a high-precision interpolation algorithm is used to convert the estimates at discrete nodes into a continuous spatial distribution. In the specific implementation, first, a two-dimensional grid matching the physical size of the wafer is defined. The grid resolution is set according to the optical resolution of the detection system, usually 1 / 4 to 1 / 3 of the optical resolution. Then, the interference sensitivity estimates of each node in the graph network are mapped to the closest grid points. For grid points without a directly corresponding node, the radial basis function interpolation is used to calculate their values. Finally, a smoothing filter is applied to the interpolated spatial distribution to reduce the influence of noise, obtaining the final interference sensitivity distribution matrix. This matrix intuitively shows the sensitivity of different positions on the wafer surface to the multimode interference of the LED light source, presented as a two-dimensional array in the form of a heat map. The high-value areas indicate positions with high interference sensitivity, and the low-value areas indicate positions less susceptible to interference. In practical applications, such as detecting logic chips with multiple metal interconnections, this distribution matrix can accurately indicate which areas need to specifically adjust the LED light source parameters to avoid interference artifacts, thereby improving the overall detection accuracy.

[0037] Furthermore, combining the interference-defect discrimination result with the output result through the recursive Bayesian estimation algorithm and estimating and updating the interference sensitivity of each node by combining physical prior knowledge to obtain the updated interference sensitivity estimate values includes: constructing a state space model for recursive Bayesian estimation, defining the interference sensitivity as the hidden state variable, defining the interference-defect discrimination result as the observation variable, designing the state transition equation and the observation equation according to the physical optics theory to obtain the recursive Bayesian state space model; performing particle filter preprocessing on the output result of the graph neural network according to the recursive Bayesian state space model, generating a set of state particles that conform to the state transition equation through the importance sampling method, and assigning an initial weight to each particle to obtain the initial weighted particle set; substituting the initial weighted particle set into the state transition equation in the recursive Bayesian state space model for state prediction calculation, and constraining and adjusting the prediction result by combining physical prior knowledge to obtain the prior state distribution; substituting the prior state distribution and the interference-defect discrimination result into the observation equation, calculating the observation likelihood function value, and updating the weights of each particle according to the observation likelihood function value; performing a resampling operation according to the updated particle weights, eliminating particles with low weights and replicating particles with high weights, optimizing the particle distribution while keeping the total number of particles unchanged, so as to obtain the posterior state distribution; extracting the interference sensitivity estimate values of each node from the posterior state distribution as the updated interference sensitivity estimation result.

[0038] Specifically, when constructing the state space model of recursive Bayesian estimation, the state variables and observation variables are first clearly defined. In the semiconductor detection scenario, the interference sensitivity of each node on the wafer surface is defined as the hidden state variables, which cannot be directly measured but need to be estimated; the interference-defect discrimination results are defined as the observation variables, which can be obtained through the interference-defect discrimination process in the aforementioned steps. The state transition equation describes the law of the change of interference sensitivity over time or space and is designed based on physical optics theory, especially the Fresnel diffraction principle and the optical wave coherence theory. The observation equation describes the relationship between the interference-defect discrimination results and the actual interference sensitivity. In actual implementation, the state transition equation takes into account the multi-mode characteristics of the LED light source and the propagation characteristics of the interference mode on the wafer surface, while the observation equation takes into account the characteristics of the discrimination algorithm and the influence of measurement noise. For example, when detecting a 3D NAND wafer with a deep trench structure, the state transition equation needs to specifically consider the waveguide effect and multiple reflection effect of the aspect ratio structure on the interference mode. The constructed recursive Bayesian state space model lays a theoretical framework for subsequent accurate estimation, making the estimation process conform to physical laws and adapt to complex detection environments.

[0039] Specifically, when performing particle filter preprocessing on the output result of the graph neural network according to the recursive Bayesian state space model, the sequential importance sampling method is used to generate a particle set. Particle filter is a Monte Carlo method that approximately represents the probability distribution of the state through a large number of random samples (particles). In specific implementation, first, according to the interference sensitivity distribution output by the graph neural network, a certain number (usually 500 - 1000) of state particles are generated for each node, and each particle represents a possible interference sensitivity value. The generation of particles follows the constraints of the state transition equation to ensure that the generated particle distribution conforms to the expectations of physical optics theory. Then, an initial weight is assigned to each particle, and the initial weight is usually set to an equal value (i.e., 1 / number of particles). For different types of wafer microstructures, the number of particles needs to be adjusted accordingly. For example, for a complex FinFET structure, due to its more complex interference mode, more particles (such as 1500 - 2000) are required to accurately represent the state distribution. This particle-based representation method can effectively handle the non-linear and multi-modal characteristics of the interference sensitivity distribution, avoiding the limitations of traditional Gaussian approximation methods when dealing with complex distributions.

[0040] Specifically, when substituting the initial weighted particle set into the state transition equation for state prediction calculation, the state transition equation is applied to each particle to predict the next state based on the current state. The state transition equation embeds the propagation law of interference patterns in physical optics, especially considering the wavelength characteristics of the light source, coherence, and the influence of the micro-structure on the wafer surface on the interference pattern. During the prediction process, physical prior knowledge is combined to constrain and adjust the prediction results, including the physical upper limit of interference sensitivity, gradient continuity constraint, and correlation constraint with the wafer structure, etc. For example, when processing a DRAM wafer with a periodic array structure, the prior knowledge includes that the interference sensitivity should have similar values on similar structures, and the periodic distribution characteristics of the interference pattern on the periodic structure. These physical constraints ensure the physical rationality of the prediction results and avoid physically unreasonable solutions that may be caused by pure data-driven methods. After prediction and constraint processing, all particles jointly form a prior state distribution, which represents the best estimate of the interference sensitivity before the update of the observed data.

[0041] Specifically, when substituting the prior state distribution and the interference-defect discrimination result into the observation equation, the value of the observation likelihood function corresponding to each particle is calculated. The observation likelihood function quantifies the probability of observing the current interference-defect discrimination result given a specific interference sensitivity state. During the calculation process, the state represented by each particle is substituted into the observation equation, compared with the actually observed interference-defect discrimination result, and the difference between the two is calculated and converted into a likelihood value. Then, the weights of each particle are updated according to the likelihood value, and the particles with higher likelihood values obtain larger updated weights. In practical applications, such as when detecting through-silicon via (TSV) structures, the interference-defect discrimination result often shows high uncertainty around the vertical structure. Therefore, the observation equation needs to specifically consider the influence of this structural characteristic on the measurement noise. In this way, the prior knowledge and the observed data are organically combined to achieve the core mechanism of Bayesian update. The weight update makes those particles that are more consistent with the observed data have a greater influence in subsequent processing, thus guiding the estimation result closer to the true state.

[0042] Specifically, when performing resampling according to the updated particle weights, the systematic resampling algorithm is used to implement the screening and replication of particles. The purpose of resampling is to solve the particle degeneracy problem, that is, after multiple updates, the weights of a few particles approach 1 while the weights of most particles approach 0. In the specific implementation, the cumulative distribution function is constructed based on the normalized particle weights, and then sampling points are evenly drawn in the interval [0,1]. The corresponding particles are selected according to the positions of the sampling points on the cumulative distribution function. This process eliminates the particles with low weights and replicates the particles with high weights multiple times while keeping the total number of particles unchanged. The trigger condition for resampling is usually based on the calculation of the effective number of particles (Neff). When Neff is lower than a certain proportion (such as 50%) of the total number of particles, resampling is performed. After resampling, the weights of all particles are reset to be equal, forming a new particle set, that is, the posterior state distribution. Finally, the interference sensitivity estimation values of each node are extracted from the posterior state distribution, usually using the weighted average method, that is, the weighted sum of all particle state values. When processing high-precision semiconductor detection data, this particle-based state estimation method can accurately capture the complex distribution characteristics of the interference sensitivity and provide more accurate estimation results than the traditional Kalman filter, especially showing significant advantages when dealing with non-linear observation relationships and non-Gaussian noise.

[0043] 104. Modulate the current waveform and beam phase distribution of the LED light source according to the interference sensitivity distribution matrix to achieve the control of the LED light source.

[0044] In an embodiment of the present invention, the modulating the current waveform and beam phase distribution of the LED light source according to the interference sensitivity distribution matrix to achieve the control of the LED light source includes: dividing the wafer surface into multiple control regions according to the sensitivity values in the interference sensitivity distribution matrix, and assigning control weight coefficients to each control region; calculating the multi-stage current waveform parameters of the LED light source according to the control weight coefficients, including the rise time, peak current, duty cycle, and fall time, and generating a current control sequence; mapping the interference sensitivity distribution matrix to the phase control plane of the spatial light modulator, calculating the compensation phase distribution, and generating a phase modulation signal; sending the current control sequence to the LED driver of the LED light source through a digital signal processing circuit, and at the same time sending the phase modulation signal to the spatial light modulator to achieve the control of the LED light source.

[0045] Specifically, when dividing the wafer surface into control regions according to the interference sensitivity distribution matrix, an adaptive segmentation algorithm is used to perform clustering analysis on the sensitivity values. First, the threshold segmentation method is applied to the interference sensitivity distribution matrix to preliminarily classify the sensitivity values into different levels (such as high, medium, and low). Then, combined with the physical structure characteristics of the wafer (such as device boundaries, trench regions, planar regions, etc.), the region growing algorithm is used to merge regions with similar sensitivity characteristics and physical adjacency to form several control regions. A control weight coefficient is assigned to each control region, and the weight coefficient is proportional to the average sensitivity value of the region, while considering the size and importance of the region. For example, when detecting a FinFET wafer, the sensitivity of the fin region is usually high and the structure is critical, so a larger control weight is assigned; while the sensitivity of the large flat region is low, a smaller control weight is assigned. This sensitivity-based region segmentation and weight assignment method can achieve a reasonable allocation of detection resources, finely control key regions, and roughly control non-key regions, improving the system efficiency while ensuring the detection accuracy.

[0046] Specifically, when calculating the multi-stage current waveform parameters of the LED light source according to the control weight coefficient, a non-linear optimization algorithm is used to solve the optimal parameter combination. The current waveform parameters include the rise time (controlling the LED startup characteristics), the peak current (controlling the maximum light intensity), the duty cycle (controlling the average light intensity), and the fall time (controlling the turn-off characteristics). The optimization goal is to minimize the interference effect, and the constraint conditions include the physical limitations of the LED and the requirements of the detection system. In specific implementation, first, a mapping relationship between the control weight and the waveform parameters is established, and the high-weight region corresponds to more precise timing control and more stable light intensity output. Then, a series of waveform parameters are calculated through the iterative optimization algorithm to generate a complete current control sequence. In practical applications, such as detecting the deep trench structure of 3D NAND memories, by precisely controlling the pulse characteristics of the LED, differential illumination of structures at different depths can be achieved, reducing the interference effect caused by multiple reflections of the bottom structure. This multi-stage current waveform control method breaks through the limitation of traditional LED control that only adjusts a constant current and realizes dynamic and precise regulation of the LED light source characteristics in the time domain.

[0047] Specifically, when mapping the interference sensitivity distribution matrix to the phase control plane of the spatial light modulator, coordinate transformation and phase calculation are required. A spatial light modulator is an optical device capable of adjusting the phase of light waves and is usually composed of a large number of independently controlled liquid crystal cells. First, through the coordinate mapping relationship of the optical path system, the interference sensitivity distribution on the wafer plane is mapped to the spatial light modulator plane. Then, based on the interference principle and wavefront modulation theory, the phase compensation value required for each position is calculated to generate a complete phase distribution map. Finally, the phase distribution map is converted into a control signal for the spatial light modulator. When detecting wafers with high-reflectivity materials, such as logic chips with metal interconnect layers, the incident light wavefront can be precisely modulated by the spatial light modulator, actively changing the position and intensity of the interference fringes, so that the interference fringes avoid key detection areas or align with known regular patterns for easy filtering. This phase modulation-based method realizes the spatial domain control of the interference pattern, complementing the current waveform control in the time domain.

[0048] Specifically, when implementing the control signal transmission through a digital signal processing circuit, a high-precision timing control system is adopted to ensure the precise synchronization of the current control sequence and the phase modulation signal. The digital signal processing circuit includes a high-speed processor, a D / A converter, and a signal driving module. The current control sequence is converted into a driving signal through the D / A converter and a power amplifier and sent to the LED driver; the phase modulation signal is sent to the spatial light modulator through a dedicated interface. In this process, the timing control of the signal is crucial, and it is necessary to ensure the precise coordination between the change of the current waveform and the phase modulation. For example, when detecting complex wafer structures that require multi-angle illumination, the system may need to dynamically adjust the current waveform and phase distribution of the LED at different illumination angles, which requires the control signal to have a nanosecond-level timing accuracy. Through this coordinated control of the current waveform and phase distribution, multi-parameter and high-precision regulation of the LED light source is achieved, fundamentally changing the formation conditions of LED multi-mode interference, greatly reducing the impact of the interference pattern on the detection result, and improving the accuracy and reliability of semiconductor detection.

[0049] In this embodiment, a multi-channel spectral characteristic acquisition of the LED light source is performed through a spectral analyzer array and an optical fiber probe, a spectral response curve and an interference image are obtained and subjected to multi-scale analysis and processing to generate a multi-dimensional data tensor; a decoupling analysis of the interference pattern and the defect signal is performed on the multi-dimensional data tensor to obtain the interference pattern characteristics and the interference-defect discrimination result; a preset kernel density estimation and a graph neural network are used to perform a spatial distribution analysis of the interference sensitivity on the surface of the wafer according to the interference pattern characteristics and the interference-defect discrimination result to generate an interference sensitivity distribution matrix; according to the interference sensitivity distribution matrix, the current waveform and the beam phase distribution of the LED light source are accurately modulated to achieve intelligent control of the LED light source, effectively eliminating the influence of multi-mode interference on semiconductor detection and improving the accuracy of nano-scale defect detection.

[0050] The method for controlling a light source based on data analysis in the embodiment of the present invention is described above. Next, the light source control system based on data analysis in the embodiment of the present invention will be described. Please refer to Figure 2 , an embodiment of the light source control system based on data analysis in the embodiment of the present invention includes: A data acquisition module 201, configured to acquire real-time operation data of production equipment in multiple regions of a factory, perform standardized fusion processing on the real-time operation data, and obtain a standardized data stream with a unified data structure; An association analysis module 202, configured to analyze the association relationship between production factors according to the standardized data stream in combination with a preset digital factory knowledge graph, and obtain key influencing factors in the current production state; A task optimization module 203, configured to perform optimized allocation of production tasks based on the key influencing factors in combination with a preset multi-dimensional cost analysis model, and obtain a production control instruction for regional collaboration; An instruction execution module 204, configured to send the production control instruction for regional collaboration to the production equipment in the corresponding region, monitor the change of production parameters of each region's production equipment after executing the control instruction, and dynamically adjust the production control instruction according to the change of the production parameters.

[0051] In the embodiments of the present invention, the light source control system based on data analysis operates the above-mentioned light source control method based on data analysis. The light source control system based on data analysis collects the multi-channel spectral characteristics of the LED light source through a spectral analyzer array and an optical fiber probe, obtains the spectral response curve and the interference image, and performs multi-scale analysis and processing to generate a multi-dimensional data tensor; performs decoupling analysis on the interference pattern and the defect signal of the multi-dimensional data tensor to obtain the interference pattern characteristics and the interference-defect discrimination result; uses the preset kernel density estimation and graph neural network to perform spatial distribution analysis on the interference sensitivity of the wafer surface according to the interference pattern characteristics and the interference-defect discrimination result to generate an interference sensitivity distribution matrix; according to the interference sensitivity distribution matrix, accurately modulate the current waveform and the beam phase distribution of the LED light source to realize the intelligent control of the LED light source, effectively eliminate the influence of multi-mode interference on semiconductor detection, and improve the accuracy of nano-scale defect detection.

[0052] Those skilled in the art can clearly understand that for the convenience and simplicity of description, the specific working processes of the above-described system, device, or unit can refer to the corresponding processes in the foregoing method embodiments and will not be repeated here.

[0053] If the integrated unit is implemented in the form of a software functional unit and sold or used as an independent product, it can be stored in a computer-readable storage medium. Based on such an understanding, the technical solution of the present invention, in essence, or the part that contributes to the prior art, or all or part of the technical solution, can be embodied in the form of a software product. The computer software product is stored in a storage medium and includes several instructions for causing 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 methods described in the various embodiments of the present invention. The foregoing storage medium includes: various media such as a USB flash drive, a mobile hard disk, a read-only memory (ROM), a random access memory (RAM), a magnetic disk, or an optical disc that can store program codes.

[0054] The above is the case. The above embodiments are only used to illustrate the technical solutions of the present invention and are not intended to limit them. Although the present invention has been described in detail with reference to the foregoing embodiments, those of ordinary skill in the art should understand that they can still modify the technical solutions described in the foregoing embodiments or perform equivalent replacements for some of the technical features. These modifications or replacements do not cause the essence of the corresponding technical solutions to deviate from the spirit and scope of the technical solutions of the various embodiments of the present invention.

Claims

1. A method for controlling a light source based on data analysis, characterized in that, The light source control method based on data analysis includes: Collecting the multi-channel spectral characteristics of the LED light source through a spectral analyzer array and an optical fiber probe, and performing multi-scale analysis and processing on the collected spectral response curves and interference images to obtain a multi-dimensional data tensor; Performing decoupling analysis of the interference mode and defect signal on the LED light source according to the multi-dimensional data tensor to obtain interference mode characteristics and interference-defect discrimination results; Performing spatial distribution analysis of the interference sensitivity on the wafer surface according to the interference mode characteristics and interference-defect discrimination results through preset kernel density estimation and graph neural network to obtain an interference sensitivity distribution matrix; According to the interference sensitivity distribution matrix, modulating the current waveform and beam phase distribution of the LED light source to achieve the control of the LED light source.

2. The method for controlling a light source based on data analysis according to claim 1, characterized in that, The collecting the multi-channel spectral characteristics of the LED light source through a spectral analyzer array and an optical fiber probe, and performing multi-scale analysis and processing on the collected spectral response curves and interference images to obtain a multi-dimensional data tensor includes: Collecting the spectral response curve of the LED light source through a spectral analyzer array, and performing peak detection and spectral analysis and processing on the spectral response curve to obtain the wavelength parameter, intensity parameter, and bandwidth parameter of the light source; Collecting the spatial distribution data and interference image of the LED light source through an optical fiber probe, and performing Fourier transform and wavelet transform on the spatial distribution data and interference image respectively to obtain the spatial propagation characteristic data and interference fringe parameter of the light source; Performing emission characteristic analysis on the LED light source according to the wavelength parameter, intensity parameter, and bandwidth parameter of the light source and the spatial propagation characteristic data to obtain the emission characteristic parameters of the light source, and organizing the emission characteristic parameters and the interference fringe parameters in multiple dimensions to obtain a multi-dimensional data tensor.

3. The method for controlling a light source based on data analysis according to claim 1, characterized in that, The performing decoupling analysis of the interference mode and defect signal on the LED light source according to the multi-dimensional data tensor to obtain interference mode characteristics and interference-defect discrimination results includes: Applying the singular value decomposition algorithm to the multi-dimensional data tensor to decompose it into a light source characteristic mode, an optical path propagation mode, and a wafer surface response mode; Constructing a three-mode coupling analysis matrix according to the light source characteristic mode, the optical path propagation mode, and the wafer surface response mode, and applying a non-linear manifold learning algorithm to analyze the three-mode coupling analysis matrix to determine the manifold structure boundary of the interference mode and the defect signal; Applying the persistent homology analysis method to the manifold structure boundary, extracting topological structure features, and classifying the topological structure features into topological invariant features that do not change with the detection conditions and topological variable features that change with the detection conditions; Analyzing the interference image according to the topological invariant features and the topological variable features to generate interference mode characteristics and interference-defect discrimination results.

4. The method for controlling a light source based on data analysis according to claim 3, characterized in that, The applying a non-linear manifold learning algorithm to analyze the three-mode coupling analysis matrix to determine the manifold structure boundary of the interference mode and the defect signal includes: Applying the locally linear embedding algorithm in the non-linear manifold learning algorithm to the three-mode coupling analysis matrix to map the high-dimensional data to a low-dimensional manifold space to obtain the reduced-dimensional representation data; Construct a k-nearest neighbor graph structure for the dimensionality-reduced representation data, calculate the similarity between nodes in the k-nearest neighbor graph structure according to the Euclidean distance, and obtain a similarity matrix; Calculate the eigenvalue spectrum of the Laplacian matrix according to the similarity matrix, analyze the eigenvalue distribution in the eigenvalue spectrum of the Laplacian matrix by the spectral clustering method, and obtain the intrinsic dimension of the data manifold; Construct a manifold diffusion graph based on the intrinsic dimension, calculate the diffusion distance between point pairs in the manifold diffusion graph through the heat kernel function, and obtain a diffusion distance matrix; Apply the contour tracing algorithm to the diffusion distance matrix, identify the region of sudden change in the diffusion distance gradient, segment the gradient mutation region through the density change rate threshold, construct a boundary decision function, and apply the boundary decision function to the data points in the low-dimensional manifold space to obtain the manifold structure boundary of the interference pattern and the defect signal.

5. The method for controlling a light source based on data analysis according to claim 1, characterized in that, The spatial distribution analysis of the interference sensitivity on the wafer surface by the preset kernel density estimation and graph neural network according to the interference pattern features and the interference-defect discrimination results to obtain the interference sensitivity distribution matrix includes: Perform adaptive kernel density estimation calculation on the interference pattern features, perform probability density modeling on the interference pattern distributions at different positions on the wafer surface, and generate preliminary spatial sensitivity data; Construct a graph network topology structure according to the microstructural features of the wafer surface obtained in advance, map the preliminary spatial sensitivity data onto the graph network topology structure, and calculate the mutual influence between nodes in the graph network topology structure through the pre-deployed graph neural network to obtain an output result; Combine the interference-defect discrimination result with the output result through the recursive Bayesian estimation algorithm, and combine physical prior knowledge to estimate and update the interference sensitivity of each node to obtain an updated interference sensitivity estimate; Reorganize and arrange the interference sensitivity estimates according to the wafer surface coordinate system to generate an interference sensitivity distribution matrix with spatial resolution.

6. The method for controlling a light source based on data analysis according to claim 5, characterized in that, The combination of the interference-defect discrimination result and the output result through the recursive Bayesian estimation algorithm, and the estimation and update of the interference sensitivity of each node by combining physical prior knowledge to obtain an updated interference sensitivity estimate includes: Construct a state space model for recursive Bayesian estimation, define the interference sensitivity as a hidden state variable, define the interference-defect discrimination result as an observation variable, and design a state transition equation and an observation equation according to the physical optics theory to obtain a recursive Bayesian state space model; Perform particle filter preprocessing on the output result of the graph neural network according to the recursive Bayesian state space model, generate a set of state particles that conform to the state transition equation through the importance sampling method, and assign an initial weight to each particle to obtain an initial weighted particle set; Substitute the initial weighted particle set into the state transition equation in the recursive Bayesian state space model for state prediction calculation, and combine physical prior knowledge to constrain and adjust the prediction result to obtain a prior state distribution; Substitute the prior state distribution and the interference-defect discrimination result into the observation equation, calculate the observation likelihood function value, and update the weights of each particle according to the observation likelihood function value; Perform resampling operations according to the updated particle weights, eliminate particles with low weights and replicate particles with high weights, optimize the particle distribution while keeping the total number of particles unchanged, so as to obtain the posterior state distribution; Extract the interference sensitivity estimation values of each node from the posterior state distribution as the updated interference sensitivity estimation result.

7. The light source control method based on data analysis according to claim 1, wherein, The control of the LED light source by modulating the current waveform and beam phase distribution of the LED light source according to the interference sensitivity distribution matrix includes: Divide the wafer surface into multiple control regions according to the sensitivity values in the interference sensitivity distribution matrix, and assign control weight coefficients to each control region; Calculate the multi-stage current waveform parameters of the LED light source according to the control weight coefficients, including rise time, peak current, duty cycle and fall time, and generate a current control sequence; Map the interference sensitivity distribution matrix to the phase control plane of the spatial light modulator, calculate the compensation phase distribution, and generate a phase modulation signal; Send the current control sequence to the LED driver of the LED light source through the digital signal processing circuit, and at the same time send the phase modulation signal to the spatial light modulator to achieve the control of the LED light source.

8. A light source control system based on data analysis, wherein, The light source control system based on data analysis includes: A spectrum acquisition module for multi-channel spectrum characteristic acquisition of the LED light source through a spectrum analyzer array and an optical fiber probe, and performing multi-scale analysis and processing on the acquired spectrum response curve and interference image to obtain a multi-dimensional data tensor; A decoupling analysis module for decoupling analysis of the interference mode and defect signal of the LED light source according to the multi-dimensional data tensor to obtain the interference mode characteristics and the interference-defect discrimination result; A spatial analysis module for spatially analyzing the interference sensitivity on the wafer surface according to the interference mode characteristics and the interference-defect discrimination result through preset kernel density estimation and graph neural network to obtain an interference sensitivity distribution matrix; A light source regulation module for modulating the current waveform and beam phase distribution of the LED light source according to the interference sensitivity distribution matrix to achieve the control of the LED light source.

Images