Map automatic calibration method based on semiconductor defect and yield

Abstract

The invention relates to the technical field of semiconductors, and discloses a semiconductor defect and yield Map-based automatic calibration method, which comprises the following steps of: acquiring wafer surface data in real time through a multi-modal sensor, analyzing station process parameters, and generating a time-space aligned wafer holographic data set; performing nonlinear deformation compensation based on the wafer holographic data set to generate a dynamic Die grid; utilizing a multi-task deep reinforcement learning model to predict initial Offset parameters of the defect Map and the yield Map; and performing global optimization on the initial Offset parameters through a quantum annealing algorithm, and outputting an optimal Offset. Data are collected through a multi-mode sensor, space-time alignment is carried out through Kalman filtering, a wafer holographic data set of space-time alignment is generated, the data space-time unification is achieved, the wafer state is accurately and completely reflected, and a high-quality data basis is provided for defect analysis and yield evaluation. The problem that data of the multi-modal sensor are inconsistent in time and space and are difficult to effectively fuse is solved.

Description

Technical Field

[0001] The present invention relates to the field of semiconductor manufacturing technology, and in particular to an automatic calibration method based on a semiconductor defect and yield map. Background Art

[0002] In today's semiconductor manufacturing industry, with the continuous advancement of processes and the increasing demand for product quality, accurate and comprehensive information on wafer status has become crucial to ensuring the yield of finished semiconductor products. To obtain multi-dimensional information about wafers, multi-modal sensors are often used. For example, optical sensors can capture microscopic topographic features of the wafer surface, LiDAR arrays can measure distance-related data, and RFID readers can obtain wafer identity and corresponding process parameters.

[0003] However, these sensors of different modalities each operate independently, with independent acquisition frequencies, acquisition times, and spatial coordinate reference systems. Due to the varying characteristics of these sensors, the collected data is often out of sync in time and lacks a unified spatial alignment standard. This leads to spatiotemporal inconsistencies in multimodal sensor data.

[0004] Traditional data processing methods and calibration means lack an effective integration mechanism when faced with such complex and heterogeneous data. It is difficult to deeply integrate these data from different sources and different time and space dimensions, and thus cannot fully tap the valuable information contained in the data to accurately describe the true state of the wafer. Ultimately, it affects the accurate judgment of wafer defects and the accurate assessment of yield, greatly restricting the improvement of semiconductor manufacturing efficiency and product quality. Summary of the Invention

[0005] In response to the shortcomings of the existing technology, the present invention provides an automatic calibration method based on semiconductor defect and yield map, which solves the problem that the temporal and spatial inconsistency of multimodal sensor data is difficult to effectively fuse.

[0006] To achieve the above objectives, the present invention is implemented through the following technical solutions: a method for automatic calibration based on semiconductor defect and yield map, comprising the following steps:

[0007] S1, real-time acquisition of wafer surface data through multimodal sensors, analysis of site process parameters, and generation of spatiotemporally aligned wafer holographic datasets;

[0008] S2, perform nonlinear deformation compensation based on the wafer holographic data set and generate a dynamic Die grid;

[0009] S3. Use a multi-task deep reinforcement learning model to predict the initial offset parameters of the defect map and yield map;

[0010] S4, globally optimize the initial Offset parameter by a quantum annealing algorithm, and output an optimal Offset;

[0011] S5, implement cross-site process knowledge migration based on a federated meta-learning framework, and update the calibration model parameter;

[0012] S6, iteratively optimize the model according to the calibration result, and complete dynamic self-adaptive calibration.

[0013] By adopting the above technical solutions, the multi-modal sensor is used to collect data to generate a holographic data set, a dynamic Die grid is generated by compensating for nonlinear deformation, the Offset parameter is predicted by a model and optimized by an algorithm, the model is updated by federated meta-learning, and dynamic self-adaptive calibration is realized by iterative optimization according to the calibration result, thereby solving problems such as multi-modal data fusion and assisting in semiconductor manufacturing.

[0014] Preferably, the multi-modal sensor in S1 comprises:

[0015] An optical sensor is used to collect wafer surface 3D point cloud data and perform denoising processing through anisotropic diffusion filtering;

[0016] A LiDAR array is used to construct a sparse depth matrix through time-of-flight measurement;

[0017] An RFID reader / writer is used to analyze the site ID and process parameters in the wafer label;

[0018] The space-time alignment is realized by Kalman filtering, and satisfies the formula Where Z k is a multi-sensor observation value, and K k is a dynamically adjusted Kalman gain.

[0019] Preferably, the nonlinear deformation compensation in S2 adopts a thin plate spline function modeling, and satisfies Where φ(r)=r 2 lnr is a radial basis function, and the weight w i is solved by a least squares method.

[0020] Preferably, the dynamic Die grid generation method in S2 comprises:

[0021] The coordinates after deformation compensation are mapped to a hexagonal adaptive grid;

[0022] The grid distortion is optimized by a Laplace smoothing algorithm, and satisfies the iterative formula Where N(i) is a set of adjacent nodes.

[0023] Preferably, the quantum annealing algorithm in 4 comprises: constructing a Hamiltonian of a QUBO model Where Jij For the inter-die space correlation weight,

[0024] Preferably, the federal meta-learning framework in S5 comprises:

[0025] Based on the process similarity, a knowledge graph is constructed, and the edge weight calculation satisfies wij=Softmax(MLP([v i ;v j ]));The model parameter aggregation rule is

[0026] Preferably, the multi-task deep reinforcement learning model in S3 comprises: the reward function of the policy network is Where is the point correlation operator.

[0027] Preferably, the knowledge graph node in S5 comprises: site ID, equipment model, Die layout template, historical Offset mean

[0028] Preferably, the S5 further comprises a model distillation process, and the loss function is L kd =KL(f edge (x)||f cloud (x));Where f edge is an edge lightweight model, and f cloud is a cloud complex model.

[0029] Preferably, the calibration verification condition in S6 comprises: the correlation coefficient ρ of the overlapping area of the defect Map and the yield Map is greater than or equal to 0.95.

[0030] The application provides a kind of based on semiconductor defect and yield Map automatic calibration method. With the following

[0031] Beneficial effects:

[0032] 1、The application generates a spatio-temporally aligned wafer holographic dataset by collecting data through multi-modal sensors and using Kalman filtering spatio-temporal alignment, unifies data spatio-temporally, accurately and completely reflects wafer state, and provides high-quality data basis for defect analysis and yield evaluation. The problem of inconsistent spatio-temporal multi-modal sensor data fusion is solved.

[0033] 2. This invention uses thin plate spline function modeling to compensate for nonlinear wafer deformation. Combining coordinate mapping, hexagonal adaptive grids, and Laplace smoothing algorithms, it generates a dynamic Die grid, resulting in uniform grid node distribution and controlled distortion. This provides accurate wafer shape and appropriate spatial structure for defect detection and yield analysis, improving analysis accuracy. This solves the problem of feature position offset caused by nonlinear wafer deformation in semiconductor manufacturing, and the lack of accuracy and reliance on manual adjustment in traditional methods for defect detection and yield assessment.

[0034] 3. This invention, through a multi-task deep reinforcement learning model and corresponding reward function, can fully exploit the complex characteristics and inherent connections between the defect map and yield map data. Leveraging the advantages of multi-task learning, it can more accurately predict the initial offset parameters. Simultaneously, the reward function guides and constrains the agent's behavior, improving the reliability and stability of the prediction results and providing strong data support for subsequent calibration and other tasks. This solves the problem that traditional methods have difficulty accurately predicting the initial offset parameters of the defect map and yield map by leveraging the complex relationships between data.

[0035] 4. This invention uses a federated meta-learning framework to build a knowledge graph to clarify site process knowledge associations. It then updates calibration model parameters based on model parameter aggregation rules to integrate the advantages of multiple sites, improving calibration accuracy and robustness. It also uses model distillation to achieve knowledge transfer, allowing the edge to utilize limited resources to assist in calibration, optimizing calibration results and improving overall efficiency. This solves the problems of traditional methods that make it difficult to integrate cross-site process knowledge, the difficulty of developing high-quality models at the edge due to resource constraints, and the difficulty of knowledge transfer and coordination between models at different levels, resulting in a lack of comprehensiveness and adaptability in model calibration. BRIEF DESCRIPTION OF THE DRAWINGS

[0036] Figure 1 This is a flow chart of an automatic calibration method based on semiconductor defect and yield map proposed by the present invention;

[0037] Figure 2 This is a timing diagram of the automatic calibration method based on the semiconductor defect and yield map proposed by the present invention. DETAILED DESCRIPTION

[0038] The following will clearly and completely describe the technical solution of the present invention in conjunction with the accompanying drawings. Obviously, the embodiments described are only some embodiments of the present invention, not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without making creative efforts are within the scope of protection of the present invention.

[0039] Please see the attached Figure 1 -Attached Figure 2, an embodiment of the present invention provides an automatic calibration method based on a semiconductor defect and yield map, comprising the following steps:

[0040] S1. Real-time wafer surface data is collected through multimodal sensors, site process parameters are analyzed, and a spatiotemporally aligned wafer holographic dataset is generated.

[0041] The multimodal sensors in S1 include:

[0042] Optical sensor, used to collect 3D point cloud data of the wafer surface and perform denoising through anisotropic diffusion filtering;

[0043] LiDAR array, which constructs a sparse depth matrix through time-of-flight measurements;

[0044] RFID reader / writer, which parses the site ID and process parameters in the wafer tag;

[0045] Spatiotemporal alignment is achieved through Kalman filtering, satisfying the formula where Z k is the multi-sensor observation value, K k is the dynamically adjusted Kalman gain.

[0046] Specifically, it is based on the principle of optical imaging, uses a specific light source to illuminate the wafer surface, and collects 3D point cloud data of the wafer surface by receiving reflected light. Due to the influence of factors such as the acquisition environment and the accuracy of the sensor itself, the original collected data will have noise interference, so it is further denoised through anisotropic diffusion filtering. Anisotropic diffusion filtering is based on the gradient information of different areas in the image. While smoothing the noise, it retains the image's edge and other detailed features as much as possible. Specifically, it constructs partial differential equations and sets appropriate diffusion coefficients and other parameters so that the diffusion along the edge is weaker and the diffusion perpendicular to the edge is stronger. In this way, noise is removed to obtain relatively clear and accurate 3D point cloud data related to the microscopic morphology of the wafer surface.

[0047] Using the time-of-flight (TOF) measurement principle, laser pulses are emitted to the wafer surface, and the time it takes for the laser to return is recorded. Based on the speed of light in air (approximately the speed of light), the distance from each measurement point to the LiDAR sensor is calculated using the formula "distance = speed × time / 2" (divided by 2 because the laser goes through a round-trip process), and then a sparse depth matrix is ​​constructed to provide supplementary data for describing the wafer status from the distance dimension.

[0048] Depend on the radio frequency identification technology, through the emission of a specific frequency of radio frequency signal, when the wafer with RFID tag into its working area, the reader and the tag between electromagnetic induction or electromagnetic coupling, so that the tag is activated and the information stored in it (such as site ID and process parameters, etc.) in the form of wireless communication transmission back to the reader, the analysis of the wafer tag related information.

[0049] Kalman filter is a kind of optimal estimation algorithm based on linear system state space model, used for real-time estimation of dynamic system state and data fusion. In this application scenario, the data collected by multi-modal sensor is regarded as the observation value of a dynamic system.

[0050] First, the system has a state transition equation, which describes the evolution law of the system state with time, such as the change of the position, attitude and other state information of the wafer at different time (although in this scenario, the state alignment of data fusion layer may be more focused, but from the perspective of the whole Kalman filter, there is such a basic model setting), its general form can be expressed as Where the system state vector is F k is the state transition matrix, is the state vector at the last time, B k is the control input vector, and uk corresponds to the coefficient matrix.

[0051] At the same time, there is an observation equation to relate the system state and the observation value of the sensor, that is, the relationship between the data collected by the multi-modal sensor and the actual state of the system, which is Where H k is the observation matrix, and uk is the observation noise vector.

[0052] In specific application, Kalman filter will be initialized first, and the initial system state estimation value and the corresponding covariance matrix (representing the uncertainty degree of the estimation value) are given. Then, at each sampling time (corresponding to the time of sensor data collection), the next time state is predicted based on the current state estimation value, and the prediction process is carried out according to the state transition equation, to obtain the predicted state value And the corresponding prediction covariance

[0053] Then, when the new sensor observation value Z k is obtained, the Kalman gain K k is calculated, and its calculation formula is Where R k is the observation noise covariance matrix, H kThe Kalman gain plays a role in balancing the weight of the predicted value and the observed value, and dynamically adjusts whether to rely more on the prediction or the observation according to the reliability of the observed value and the uncertainty of the current prediction.

[0054] Finally, the updated system state estimation value is obtained by fusing the predicted value and the observed value The calculation formula is The covariance matrix is updated at the same time, which is used for the next round of iterative calculation. Through such a process of continuous iteration, the data collected by different sensors at different times is used to realize the alignment of data in different space-time according to the respective observation equation and the common state transition framework, so that the finally generated data set has consistency in time and space.

[0055] Through the working principle of collecting data by multi-modal sensors and using Kalman filtering for space-time alignment, the effect of generating a wafer holographic data set that is space-time aligned is finally realized. This data set comprehensively covers the wafer surface micro-topography information (3D point cloud data after denoising) obtained from optical sensors, the distance-related sparse depth matrix provided by LiDAR array, and the station ID and process parameters analyzed by RFID reader, etc. Information, and each data is synchronized in time and unified in space, which can accurately and completely reflect the overall state of the wafer under the corresponding manufacturing link, and provide a high-quality data basis for further analyzing the defect condition of the wafer and evaluating the yield, etc. Operation. Solving the problem of inconsistent space-time of multi-modal sensor data difficult to effectively fuse.

[0056] S2, based on the wafer holographic data set, nonlinear deformation compensation is performed to generate a dynamic Die grid.

[0057] The nonlinear deformation compensation in S2 adopts thin plate spline function modeling, which satisfies where φ(r) = r 2 lnr is a radial basis function, and the weight w i Solved by least squares method.

[0058] The dynamic Die grid generation method in S2 includes:

[0059] Mapping the coordinates after deformation compensation to a hexagonal adaptive grid;

[0060] Optimizing the grid distortion by Laplace smoothing algorithm, satisfying the iterative formula Where N(i) is the set of adjacent nodes.

[0061] Specifically, in the semiconductor wafer manufacturing process, the wafer often produces nonlinear deformation due to various process factors (such as thermal stress, mechanical stress, etc.). In order to accurately compensate for such nonlinear deformation, a thin plate spline function (TPS) is used for modeling.

[0062] The thin plate spline function is a mathematical function based on a physical model, commonly used to handle deformation fitting problems of object surfaces with elastic deformation. Its expression is where φ(r) = r 2 lnr is a radial basis function that describes the relationship between the distance from a point in space to a control point and the influence of deformation. Here, r represents the Euclidean distance from the point (x, y) to the control point (x i ,y i ). w i is the weight coefficient, which determines the degree of influence of each control point on the deformation of the target point. a x and a y are the coefficients of the linear term, used to handle global linear transformations.

[0063] In order to determine the weight w i , the least squares method is used. The goal of the least squares method is to find a set of weights w i such that the sum of the squared errors between the predicted values of the thin plate spline function at the control points and the actual observed values is minimized. Specifically, assuming we have N control points (x i , y i ) and their corresponding deformation observations d i , we can construct an error function By taking the partial derivative of E with respect to w i and setting it to zero, we get a system of linear equations, which can be solved to obtain the weights w i .

[0064] After completing the nonlinear deformation compensation, the coordinates after deformation compensation are mapped to a hexagonal adaptive grid. The hexagonal grid has better symmetry and space-filling properties, and can more effectively represent the information of the wafer surface compared to the traditional rectangular grid. By matching the coordinates after deformation compensation with the nodes of the hexagonal grid, each point is assigned to the appropriate hexagonal grid cell, thereby completing the mapping of the coordinates.

[0065] The Laplace smoothing algorithm is a commonly used method for improving grid quality. Its iterative formula is where represents the coordinates of node i at the kth iteration. N(i) is the set of neighboring nodes of node i. is the coordinates of the neighboring node j of node i at the kth iteration.

[0066] The basic idea of the algorithm is that the new position of each node is a weighted average of its current position and the positions of neighboring nodes. Through multiple iterations, the distribution of the grid nodes is made more uniform, and the distortion of the grid is reduced. At each iteration, node i moves a certain distance in the direction of the center of gravity of its neighboring nodes, and this distance is controlled by the coefficient 0.2.

[0067] By using thin plate spline function modeling for nonlinear deformation compensation, the nonlinear deformation of the wafer is accurately corrected. In combination with mapping the coordinates to the adaptive hexagonal grid and using the Laplace smoothing algorithm to optimize the grid distortion to generate a dynamic Die grid, the grid node distribution is uniform, the distortion is effectively controlled, and more accurate wafer shape and more suitable spatial structure are provided for subsequent defect detection and yield analysis, which helps to improve the accuracy of analysis. The problem of feature position offset caused by nonlinear deformation of the wafer due to various factors in semiconductor manufacturing is solved, and the problem of insufficient accuracy and dependence on manual adjustment in traditional analysis methods in defect detection and yield evaluation is solved.

[0068] S3, predicting the initial Offset parameters of the defect Map and the yield Map by using a multi-task deep reinforcement learning model.

[0069] The multi-task deep reinforcement learning model in S3 includes: the reward function of the policy network is Where is the point-wise correlation operator.

[0070] Specifically, deep reinforcement learning is a technology that combines the powerful feature representation capability of deep learning and the decision optimization capability based on environmental feedback of reinforcement learning. Multi-task deep reinforcement learning aims to handle multiple related tasks simultaneously, and more efficiently learns the common and characteristic knowledge between different tasks through sharing part of the network structure, so as to improve the overall learning effect and prediction performance.

[0071] The model mainly includes key elements such as agents (Agents), environments (Environments), and reward mechanisms. The agent takes a series of actions (Actions) in the environment, and the environment feeds back the corresponding state (States) and reward (Rewards) according to the actions of the agent, and the agent adjusts its strategy according to these feedback information to maximize the long-term cumulative reward.

[0072] In this specific application scenario, the agent needs to decide what action to take according to the wafer-related data it has obtained (such as the data set after the previous step processing, etc.), which may involve the selection of data features, the decision of different processing methods, etc., and then affect the prediction of the initial Offset parameters of the defect Map and the yield Map.

[0073] The reward function of the policy network is where y defect represents the relevant data vector corresponding to the defect Map, y yield represents the relevant data vector corresponding to the yield Map, and is a point-by-point correlation operator, through which the correlation degree of the defect Map and the yield Map at each corresponding position can be calculated, is a certain norm operation (here, the square of the Frobenius norm) on this correlation degree, which reflects the overall correlation between the defect Map and the yield Map. The role of this part is that when the correlation between the defect Map and the yield Map is higher, that is, the two are more matched and more in line with the actual situation, the value will be larger, which means giving the agent a relatively higher positive reward, guiding the agent to take actions in the direction that makes the correlation between the two better.a t represents the action taken by the agent at time t, and ||a t || 2 is a measure of the action, multiplied by the coefficient 0.1 and then subtracted from the reward function, which is to constrain the action taken by the agent to a certain extent, avoiding the agent taking too extreme or unreasonable actions, so that the agent can consider the rationality and appropriateness of the action while pursuing the high correlation between the defect Map and the yield Map, ensuring the stability and effectiveness of the entire learning and prediction process.

[0074] Through continuous interaction in the environment, the agent uses the deep neural network to learn and update the parameters of the policy network according to the state of each step and the reward obtained, so that the policy network can output appropriate actions according to the input wafer data, and then gradually optimize the prediction of the initial Offset parameter of the defect Map and the yield Map.

[0075] Through the multi-task deep reinforcement learning model and the corresponding reward function, the complex features and internal relations of the defect Map and the yield Map related data can be fully tapped, and the initial Offset parameter can be more accurately predicted with the help of the multi-task learning advantage, while the reward function guides and constrains the agent behavior, improves the reliability and stability of the prediction result, and provides strong data support for subsequent calibration and other work. It solves the problem that traditional methods cannot accurately predict the initial Offset parameter of the defect Map and the yield Map by utilizing the complex association between the data.

[0076] S4, globally optimize the initial Offset parameter through a quantum annealing algorithm, and output the optimal Offset. The quantum annealing algorithm in S4 includes constructing the Hamiltonian of the QUBO model where J ij is the spatial correlation weight between dies,

[0077] Specifically, the quantum annealing algorithm is an effective method based on quantum mechanics principles for solving combinatorial optimization problems. In the context of semiconductor defect and yield map calibration, we need to globally optimize the initial offset parameters obtained through the previous steps to obtain the optimal offset, at which time the quantum annealing algorithm plays an important role.

[0078] The core is to construct the Hamiltonian of the QUBO model, whose expression is Where, J ij represents the spatial correlation weight between dies, which reflects the degree of correlation between different dies on the wafer in spatial position. For example, if two dies are adjacent in space or in similar process affected areas, the correlation between them may be strong, and the corresponding J ij value will be relatively large; conversely, when the correlation is weak, the J ij value is small. This weight is crucial for depicting the relationship between different positions of dies on the entire wafer and their influence on the final optimization result, and is a key factor in incorporating spatial structure information when constructing the optimization model.

[0079] h i is defined as Here, h can be regarded as a weight coefficient related to the offset of the i-th die, which is related to specific process characteristics, statistical characteristics of data and other factors, and is used to measure the relative importance of the offset of different dies in the optimization process. is a measure of the difference between the defect map and the yield map corresponding to the i-th die, which can be a certain distance measure or a numerical representation of the difference degree. By multiplying it with h , it comprehensively reflects the influence of the difference between the defect and yield of the i-th die on the overall optimization.

[0080] x i is a binary variable whose value is 0 or 1, representing a certain decision or state. In the QUBO model, the solution that minimizes the Hamiltonian H is found by determining the combination of these binary variables, and this solution corresponds to the optimal offset parameter after global optimization.

[0081] Quantum annealing algorithm utilizes the characteristics of quantum bits (Qubits). During the algorithm running process, Qubits can be in the superposition state of 0 and 1, and can overcome the local optimal solution trap of traditional classical algorithm through quantum tunneling effect and other quantum mechanical phenomena, and conduct a more extensive search in the entire solution space, so as to find the global optimal solution. With the evolution of the algorithm, the system gradually anneals from the quantum state to the classical state, and finally converges to a certain binary solution, that is, the solution corresponding to the optimal Offset we expect.

[0082] By using quantum annealing algorithm to construct QUBO model for calculation, the spatial correlation between different Dies on the wafer and the difference in defects and yield of each Die can be fully considered, and complex factors can be integrated for global optimization. In addition, it can also take advantage of the advantage of exceeding classical algorithm to avoid falling into local optimal solution, and output more accurate optimal Offset parameter, provide high-quality parameter basis for calibration, and improve calibration accuracy and reliability. It solves the problem that traditional optimization method is easy to fall into local optimal solution, difficult to consider the spatial correlation of different Dies on the wafer, the difference in defects and yield, and cannot obtain the best Offset parameter to realize the optimal calibration effect.

[0083] S5, based on the federal meta-learning framework, realizes cross-site process knowledge migration, and updates the calibration model parameters.

[0084] The federal meta-learning framework in S5 includes:

[0085] Based on the process similarity, a knowledge graph is constructed, and the edge weight calculation satisfies wij=Softmax(MLP([v i ;v j ]));The model parameter aggregation rule is

[0086] S5 also includes a model distillation process, and its loss function is L kd =KL(f edge (x)||f cloud (x));Where f edge is an edge lightweight model, and f cloud is a cloud complex model.

[0087] Specifically, in semiconductor manufacturing, different sites often have their own process characteristics and data accumulation. The federal meta-learning framework first constructs a knowledge graph based on process similarity, aiming to integrate these scattered knowledge information. By analyzing the process parameters, equipment types, production processes and other factors of each site, the similarity of the processes between different sites is measured. For example, comparing site A and site B, if they use the same type of key production equipment, and the parameter settings of some core process links are close, then the process similarity between the two sites is relatively high.

[0088] The edge weight calculation satisfies This formula. Where v i and v j represent the process feature vectors of site i and site j respectively, which are the results of vectorizing the site process-related information, containing information such as equipment model, process parameter range, etc. MLP (Multi-Layer Perceptron) is a commonly used neural network structure, which is used here to process the spliced feature vectors and map them to a suitable space to mine more complex process similarity features. The Softmax function further converts the output of the MLP into a probability distribution, so that the edge weight wij is between 0 and 1, and the sum of all connected edge weights is 1, so as to accurately quantify the degree of association between site i and site j. In the knowledge graph, the weight of the edge is higher, indicating that the two sites are more closely related in terms of process knowledge, which is more conducive to subsequent knowledge transfer.

[0089] The model parameter aggregation rule is Here N represents the total number of sites participating in federated learning, θ i is the local model parameter of site i, and wio is the weight between site i and a reference site (which can be the center site or a site set as a reference). Through such an aggregation rule, the model parameters of different sites are weighted and summed according to the weight relationship between each site and the reference site, so that the final meta-model parameter θ meta integrates the knowledge advantages of each site, taking into account the characteristics and experience of different sites in the process, realizing cross-site knowledge aggregation and integration, updating and calibrating model parameters, so that the calibration model can absorb the beneficial process knowledge of multiple sites and improve its performance and adaptability.

[0090] The model distillation process is also included in the federated meta-learning framework, and its loss function is L kd = KL(f edge (x) || f cloud (x)); where f edge represents the edge lightweight model, which is usually deployed near the data source (such as the local device end of each site), has relatively simple structure, consumes less computing resources, etc., and is convenient for local rapid data processing and feature extraction. f cloud is a cloud complex model, which often has a deeper and more complex network structure, can learn more comprehensive and fine-grained feature representations, but has relatively high requirements for computing resources and network environment.

[0091] KL refers to relative entropy, also known as KL divergence, which is used to measure the difference between two probability distributions. In the model distillation process, the KL divergence between the output distribution of the edge lightweight model f edge (x) and the output distribution of the cloud complex model f cloud (x) is calculated as the loss function L kd During training, the goal is to minimize this loss function, so that the edge lightweight model can learn the key knowledge contained in the cloud complex model, distill the "knowledge" of the cloud complex model into the edge lightweight model, achieve effective transmission and simplification of knowledge, and enable the edge to have good prediction ability in the case of limited resources, while further assisting the update of model parameter calibration, so that the model can better adapt to different environments and complete the calibration task of semiconductor defect and yield map.

[0092] Through the federal meta-learning framework, the site process knowledge association is clearly defined by constructing a knowledge graph, and the calibration model parameters are updated by integrating the advantages of multiple sites according to the model parameter aggregation rules, improving the calibration accuracy and robustness. In addition, knowledge transfer is achieved through model distillation, allowing the edge to use limited resources to assist calibration, optimize calibration results, and improve overall efficiency. The traditional method cannot integrate cross-site process knowledge, the edge is limited by resources and cannot have a high-quality model, and there are difficulties in knowledge transfer and coordination between different levels of models, resulting in a lack of comprehensiveness and adaptability in model calibration.

[0093] S6, according to the calibration result feedback iteration optimization model, complete dynamic self-adaptive calibration.

[0094] The calibration verification condition in S6 includes: the correlation coefficient p of the overlap area of the defect map and the yield map is greater than or equal to 0.95.

[0095] Specifically, in the semiconductor manufacturing process, in order to achieve accurate defect and yield map calibration, the method of iterating and optimizing the model according to the calibration result feedback is adopted. After completing a calibration operation, relevant result information after calibration is obtained, which includes the specific situation of the defect map and the yield map after calibration, such as their spatial distribution, feature matching, etc.

[0096] Then, these calibration results are input as feedback data into the model, and the model will analyze these feedback data according to the pre-set evaluation mechanism and algorithm to determine whether the current calibration result meets the expected standard or whether there is further optimization space. For example, it will check whether the correlation between the defect position marked in the calibrated defect map and the corresponding area in the yield map conforms to the actual process logic, and the matching degree of each feature area, etc.

[0097] Based on this analysis and judgment, the model will make targeted adjustments to its internal parameters and related parts of the algorithm structure. For example, if it is found that the correspondence between defects and yield in a certain area is still unsatisfactory, the weight parameters involved in the calibration calculation in that area may be adjusted, or the algorithm module that processes the corresponding data features may be optimized. The data will then be processed again based on the new parameters and algorithms to carry out the next round of calibration. This cycle will be repeated, and the model will be iteratively optimized to enable it to better adapt to the ever-changing conditions in actual semiconductor manufacturing and the differences in characteristics of different batches of wafers, thereby achieving dynamic adaptive calibration.

[0098] During the calibration process, a key calibration verification condition is established, focusing on the correlation coefficient ρ of the overlap area between the defect map and the yield map. The overlap area here refers to the spatially shared area between the defect map and the yield map, and the correlation coefficient ρ is a measure of the degree of correlation between the data in the overlap area between the two maps.

[0099] The specific method for calculating the correlation coefficient ρ is usually based on statistical principles. For example, the Pearson correlation coefficient can be used (of course, other appropriate correlation measurement methods may also be used according to actual conditions). By analyzing the data points at corresponding positions in the overlapping area, the covariance between them and their respective standard deviations are calculated, and the value of the correlation coefficient ρ is obtained.

[0100] When ρ ≥ 0.95, the data correlation between the defect map and the yield map in the overlapping area is very high, indicating a good match between the two. The current calibration result meets the requirements and can be considered to accurately reflect the actual defect and yield situation of the semiconductor wafer. Conversely, if the value of ρ is lower than 0.95, the calibration effect needs to be improved. According to the feedback iterative optimization mechanism described above, the model needs to be adjusted and recalibrated until the verification condition of ρ ≥ 0.95 is met.

[0101] By leveraging an iterative optimization mechanism with feedback from calibration results, the model can continuously improve itself based on actual calibration conditions, enhancing calibration accuracy and precision to align with actual production conditions. Clear calibration verification conditions (ρ ≥ 0.95) provide quantitative evaluation criteria, avoiding calibration blindness, ensuring high-quality final calibration results, and enhancing the reliability of defect detection and yield assessment in semiconductor manufacturing. This addresses the challenges of traditional calibration methods, which lack feedback mechanisms and clear quality evaluation criteria, resulting in difficulty in flexible adjustment, low calibration accuracy, inability to adapt to changing conditions, and inability to determine whether calibration results meet standards.

[0102] While embodiments of the application have been shown and described, it is to be understood that the embodiments described are merely exemplary of the principles and application of the present application. Numerous modifications and adaptions can be effected without departing from the spirit and scope of the present application, which is not limited to the exact construction and arrangement described. It is intended, therefore, to cover all modifications and adaptions that fall within the scope of the claims and their equivalents.

Claims

1. An automatic calibration method based on semiconductor defect and yield map, characterized in that: The following steps are involved: S1, real-time acquisition of wafer surface data through multimodal sensors, analysis of site process parameters, and generation of spatiotemporally aligned wafer holographic datasets; S2, perform nonlinear deformation compensation based on the wafer holographic data set and generate a dynamic Die grid; S3. Use a multi-task deep reinforcement learning model to predict the initial offset parameters of the defect map and yield map; S4. Globally optimize the initial Offset parameters using the quantum annealing algorithm and output the optimal Offset. S5. Implement cross-site process knowledge transfer based on the federated meta-learning framework and update calibration model parameters; S6. Iteratively optimize the model based on the calibration result feedback to complete dynamic adaptive calibration.

2. The method according to claim 1, wherein: The multimodal sensor in S1 includes: Optical sensor, used to collect 3D point cloud data of the wafer surface and perform denoising through anisotropic diffusion filtering; LiDAR array, which constructs a sparse depth matrix through time-of-flight measurements; RFID reader / writer, which parses the site ID and process parameters in the wafer tag; The spatiotemporal alignment is achieved by Kalman filtering, satisfying the formula where Z k is the multi-sensor observation value, K k is the dynamically adjusted Kalman gain.

3. The method for automatic calibration based on semiconductor defect and yield map according to claim 1, characterized in that: The nonlinear deformation compensation in S2 is modeled using thin plate spline function to meet where φ(r) = r 2 lnr is the radial basis function, weight w i Solved by the least squares method.

4. The method for automatic calibration based on semiconductor defect and yield map according to claim 1, characterized in that: The dynamic Die grid generation method in S2 includes: Mapping the deformation-compensated coordinates to a hexagonal adaptive grid; The Laplace smoothing algorithm is used to optimize the mesh distortion and satisfy the iterative formula Where N(i) is the set of adjacent nodes.

5. The method for automatic calibration based on semiconductor defect and yield map according to claim 1, characterized in that: The quantum annealing algorithm in S4 includes: constructing the Hamiltonian of the QUBO model Among them J ij is the spatial correlation weight between Dies, 6. The method for automatic calibration based on semiconductor defect and yield map according to claim 1, characterized in that: The federated meta-learning framework in S5 includes: The knowledge graph is constructed based on process similarity, and the edge weight calculation satisfies wij=Softmax(MLP([v i ;v j ])); The model parameter aggregation rule is 7. The method for automatic calibration based on semiconductor defect and yield map according to claim 1, characterized in that: The multi-task deep reinforcement learning model in S3 includes: the reward function of the policy network is where ⊙ is the point-wise correlation operator.

8. The method for automatic calibration based on semiconductor defect and yield map according to claim 1, characterized in that: The knowledge graph nodes in S5 include: site ID, device model, Die layout template, historical offset average 9. The method for automatic calibration based on semiconductor defect and yield map according to claim 1, characterized in that: The S5 also includes a model distillation process, whose loss function is L kd =KL(f edge (x)||f cloud (x)); where f edge is the edge lightweight model, f cloud For complex models in the cloud.

10. The method for automatic calibration based on semiconductor defect and yield map according to claim 1, characterized in that: The calibration verification condition in S6 includes: the correlation coefficient ρ of the overlapping area between the defect map and the yield map is ≥ 0.95.

Images