Wafer doping control method based on intelligent feedback regulation

Abstract

The invention relates to the technical field of semiconductor manufacturing, and discloses a wafer doping control method based on intelligent feedback adjustment, and the method comprises the steps: carrying out the atomic scale imaging of the surface of a wafer, so as to recognize a target doping site, and generating an atomic coordinate mapping graph, secondly, inputting the atomic coordinate mapping graph into a multi-scale simulation model to generate an initial process parameter set, and processing the initial process parameter set by combining a transformer model with a reinforcement learning algorithm to obtain a dynamic process data matrix; performing optimization processing on the dynamic process data matrix through a cross-material process migration graph neural network to obtain a process parameter migration mapping relation, comparing measured data with a multi-scale model prediction result through closed-loop feedback iteration, and outputting a corrected process parameter migration mapping relation and a multi-scale model updating instruction; according to the wafer doping method, the stability, the consistency and the yield of the wafer doping process are improved, and high accuracy and high uniformity of wafer doping are realized.

Description

Technical Field

[0001] The present application relates to the field of semiconductor manufacturing technology, and in particular to a wafer doping control method based on intelligent feedback regulation. Background Art

[0002] In the field of semiconductor manufacturing, wafer production is an extremely critical link. Wafer doping is the introduction of impurity atoms into specific areas during the semiconductor wafer manufacturing process to change the electrical properties of the semiconductor and form different device structures and functions, such as forming P-type or N-type semiconductor regions. Wafer doping is one of the key processes for manufacturing high-performance semiconductor devices. The accuracy and uniformity of doping have a crucial impact on the performance, reliability and yield of semiconductor devices.

[0003] Traditional wafer doping technologies, such as ion implantation and diffusion processes, often experience problems such as ion beam energy fluctuations, poor scanning uniformity, and space charge effects during ion implantation. The energy output by the ion source is difficult to maintain absolute stability, and slight energy deviations will cause differences in the depth of the implanted ions within the wafer, thus affecting the accuracy of doping. In addition, during the ion implantation process of large-area wafers, the scanning mechanism cannot ensure that the ion beam uniformly covers the entire wafer surface, resulting in inconsistent doping concentrations between the edge and center areas of the wafer, seriously affecting the consistency of chip performance. Especially during high current injection, the interference of space charge forces will cause the ion beam to distort, further worsening doping uniformity. Due to the limitations of traditional technologies, not only is the wafer doping accuracy low, but also the uniformity is poor. Summary of the Invention

[0004] The present application provides a wafer doping control method based on intelligent feedback regulation, which controls wafer doping through intelligent feedback regulation to achieve high requirements for wafer doping accuracy and improve the high accuracy and uniformity of wafer doping.

[0005] In a first aspect, the present application provides a wafer doping control method based on intelligent feedback regulation, the wafer doping control method based on intelligent feedback regulation comprising: performing atomic-scale imaging of the surface of the wafer to identify target doping sites and generate an atomic coordinate map; Inputting the atomic coordinate map into a multi-scale simulation model to generate an initial process parameter set; The initial process parameter set is processed by combining a transformer model with a reinforcement learning algorithm to obtain a dynamic process data matrix; Optimizing the dynamic process data matrix through a cross-material process migration graph neural network to obtain a process parameter migration mapping relationship; The measured data and the multi-scale model prediction results are compared through closed-loop feedback iteration, and a corrected process parameter migration mapping relationship and a multi-scale model update instruction are output.

[0006] In the technical scheme provided in the present application, the surface of the wafer is imaged at an atomic scale to identify a target doping site and generate an atomic coordinate mapping graph, the atomic coordinate mapping graph is input into a multi-scale simulation model to generate an initial process parameter set, the initial process parameter set is processed by combining a transformer model with a reinforcement learning algorithm to obtain a dynamic process data matrix, then the dynamic process data matrix is optimized by a cross-material process migration graph neural network to obtain a process parameter migration mapping relationship, and finally the measured data and the multi-scale model prediction results are compared through closed-loop feedback iteration, and a corrected process parameter migration mapping relationship and a multi-scale model update instruction are output. In the technical scheme, the target doping site on the surface of the wafer is accurately identified and the atomic coordinate mapping graph is generated through atomic scale imaging, the positioning accuracy is far higher than that of the traditional micron level, the process quality is ensured, the multi-scale simulation model generates the initial process parameter set based on the atomic coordinate mapping graph and comprehensive multi-scale information, which not only shortens the trial period and reduces experimental costs, but also is in line with the actual process. In addition, the initial process parameter set is processed by combining the transformer model with the reinforcement learning algorithm to generate the dynamic process data matrix that can reflect the dynamic changes of the parameters in real time, compared with static data, the flexibility and accuracy of process control are greatly improved. Furthermore, the dynamic process data matrix is optimized by the cross-material process migration graph neural network to establish the process parameter migration mapping relationship, which breaks the material limitation and improves the process stability and reliability. In addition, the measured and predicted data are continuously compared through closed-loop feedback iteration to correct the mapping relationship and update the model, which significantly improves the stability, consistency and yield of the wafer doping process. The present application realizes high precision and high uniformity of wafer doping. BRIEF DESCRIPTION OF DRAWINGS

[0007] In order to more clearly illustrate the technical solutions of the embodiments of the present application, the drawings needed in the embodiment description will be briefly introduced. Obviously, the drawings in the following description are some embodiments of the present application, and other drawings can be obtained by those skilled in the art without creative labor.

[0008] Figure 1 An embodiment flowchart of the wafer doping control method based on intelligent feedback adjustment in the embodiments of the present application. DETAILED DESCRIPTION

[0009] The embodiment of the present application provides a wafer doping control method based on intelligent feedback regulation. The terms "first", "second", "third", "fourth" and the like (if any) in the specification and claims of the present application and the above-mentioned drawings are used to distinguish similar objects, and do not have to be used to describe a specific order or sequence. It should be understood that the data used in this way can be interchanged under appropriate circumstances, so that the embodiments described herein can be implemented in an order other than that illustrated or described herein. In addition, the term "includes" or "has" and any variations thereof are intended to cover non-exclusive inclusion, for example, a process, method, system, product or device including a series of steps or units does not have to be limited to those steps or units clearly listed, but can include other steps or units not clearly listed or inherent to these processes, methods, products or devices.

[0010] For ease of understanding, the specific process of the embodiment of the present application is described below. Please refer to Figure 1 One embodiment of the wafer doping control method based on intelligent feedback regulation in the present application comprises: Step S101, atom-scale imaging is performed on the surface of the wafer to identify a target doping site and generate an atomic coordinate mapping; Step S102, inputting the atomic coordinate mapping into a multi-scale simulation model to generate an initial process parameter set; Step S103, processing the initial process parameter set by combining a transformer model with a reinforcement learning algorithm to obtain a dynamic process data matrix; Step S104, optimizing the dynamic process data matrix by a cross-material process migration graph neural network to obtain a process parameter migration mapping relationship; Step S105, comparing the measured data and the multi-scale model prediction results through closed-loop feedback iteration and outputting the corrected process parameter migration mapping relationship and the multi-scale model update instruction.

[0011] In a specific embodiment, the surface of the wafer is atomically imaged and an atomic coordinate mapping is generated, which realizes accurate identification of the target doping site at the atomic level and provides extremely accurate position information basis for subsequent processes, greatly improving the accuracy of wafer doping positioning and enabling the positioning accuracy to be pushed from the micron level to the atomic scale, ensuring that the foundation and protection are laid in the most critical starting step.

[0012] In addition, the multi-scale simulation model generates an initial process parameter set according to the atomic coordinate mapping. The initial process parameter set generated by the multi-scale simulation model which integrates multiple scale information is not only efficient, but also highly consistent with actual process requirements, greatly shortening the process parameter exploration period, reducing the number of experiments and costs, and improving the degree of fit between the initial parameters and the actual process.

[0013] Moreover, the transformer model combined with the reinforcement learning algorithm can dynamically capture the complex correlation and changing trend of the initial process parameter set at different stages to generate a dynamic process data matrix, which reflects the dynamic changes of the process parameters in real time and provides rich and accurate data support for process optimization. Compared with static data processing, it can more accurately adapt to the complex and changeable wafer doping process, and improve the flexibility and accuracy of process control.

[0014] In the process parameter migration mapping relationship, the process knowledge between different materials is effectively migrated, the material restriction is broken, the mature process is quickly adapted to the new research and development material, the process development process of the new material is accelerated, the research and development cost and time are reduced, and the stability and reliability of the new process on different materials are improved.

[0015] In a specific embodiment, the measured data and the multi-scale model prediction results are iteratively compared and fed back in a closed loop, and the process parameter migration mapping relationship is continuously corrected and the multi-scale model is updated. This continuous iteration mechanism promotes continuous process optimization, significantly improves the stability, consistency and yield of the wafer doping process, ensures continuous improvement of wafer doping quality, and meets the stringent requirements for high-performance and high-reliability chip production.

[0016] In a specific embodiment, step S101 specifically comprises: (1) scanning the surface of the wafer by a scanning probe microscope to obtain atomic-scale surface topography data; (2) inputting the atomic-scale surface topography data into a convolutional neural network to identify the target doping sites and mark the three-dimensional coordinates of the target doping sites to obtain a target doping site list; (3) converting the local coordinates of the target doping site list to a wafer global coordinate system to generate the atomic coordinate mapping map.

[0017] In a specific embodiment, the surface of the wafer is finely scanned by a scanning probe microscope. The scanning probe microscope can obtain very detailed surface information through the interaction between the probe and the wafer surface. During the scanning process, the probe moves point by point on the wafer surface with very high precision, converting the microscopic features such as the arrangement and fluctuation of surface atoms into electrical signals, and further converting them into digitized atomic-scale surface topography data. The data not only contains the accurate position information of the atoms on the wafer surface, but also reflects the details such as the roughness, flatness and relative position relationship between atoms on the surface. Through the obtained atomic-scale surface topography data, the most original and accurate information source for subsequent accurate identification of target doping sites is provided. Compared with traditional low-resolution imaging technology, it can capture more subtle surface features and greatly improve the cognitive accuracy of the microscopic structure of the wafer surface.

[0018] Further, the obtained atomic-level surface topography data is input into a specially trained convolutional neural network. The convolutional neural network can automatically extract key features in the atomic-level surface topography data through the combination of multiple convolutional layers, pooling layers, and fully connected layers. The convolutional neural network is trained with a large amount of labeled atomic-level surface topography data and learns the characteristic patterns of the target doping sites. When new atomic-level surface topography data is input, the convolutional neural network can quickly analyze the atomic-level surface topography data and identify the target doping sites. When the target doping sites are identified, the convolutional neural network further marks the three-dimensional coordinates of these sites in the original data coordinate system and organizes them into a target doping site list. This realizes automatic, efficient, and accurate identification and positioning of target doping sites. Compared with manual identification, the accuracy and efficiency of identification are greatly improved, and human errors are reduced. Precise target position information is provided for subsequent process operations.

[0019] In a specific embodiment, the coordinates in the target doping site list are initially based on a local coordinate system. In order to uniformly apply these site information in the entire wafer process operation, it is necessary to convert the local coordinates to the wafer global coordinate system. The conversion process is realized through coordinate transformation algorithms and understanding of the overall structure of the wafer. Through accurate coordinate conversion, each target doping site has a unique position in the wafer global coordinate system. Integrating these target doping site information with global coordinates generates an atomic coordinate mapping map. The atomic coordinate mapping map directly displays the distribution of target doping sites on the wafer surface in the global coordinate system. It provides a unified, standard, and easy-to-understand and use coordinate information basis for subsequent multi-scale simulation model input and entire wafer doping process planning. It ensures accurate grasp and effective use of doping sites in the entire wafer range, and improves the consistency and accuracy of process planning.

[0020] Among them, from the acquisition of atomic-level surface topography data to the accurate identification and coordinate marking of target doping sites, and then to the generation of atomic coordinate mapping map, a high-precision and high-efficiency information acquisition and processing flow is constructed, which provides important basic information for subsequent wafer doping process and promotes the realization of high precision and high quality of wafer manufacturing process.

[0021] In a specific embodiment, step S101 specifically includes: The noise of the data in the atomic coordinate mapping map is removed by Gaussian filtering, and the missing data in the atomic coordinate mapping map is filled by spline interpolation. The data in the atomic coordinate mapping maps of different imaging regions are unified to the same global coordinate system, and the data in the coordinate system are normalized.

[0022] In a specific embodiment, step S102 specifically comprises: (1) converting the atomic coordinate mapping into a database format to obtain wafer feature data, wherein the wafer feature data can be recognized by the multi-scale simulation model; (2) setting boundary conditions and initial conditions of the multi-scale simulation model, and simulating and calculating the wafer feature data by the multi-scale simulation model to generate a simulation result; (3) extracting the initial set of process parameters from the simulation result.

[0023] In a specific embodiment, the data of the atomic coordinate mapping is processed by Gaussian filtering. Since the atomic scale imaging is interfered by various microscopic factors, noise exists in the data, which will affect the accuracy of subsequent analysis of the doping site and related processes. Gaussian filtering can attenuate the noise signal and retain the effective feature information in the atomic coordinate mapping, so that the data in the atomic coordinate mapping is clearer and more accurate. In addition, the missing data in the atomic coordinate mapping is filled by spline interpolation. During the atomic scale imaging process, some data is missing due to equipment failure, signal shielding, etc. Spline interpolation estimates the value of the missing data by constructing a piecewise polynomial function according to the distribution of the known data points, thereby ensuring the integrity of the atomic coordinate mapping data, which is crucial for comprehensive analysis of the atomic distribution on the wafer surface and subsequent process simulation, and can avoid analysis deviation caused by data loss.

[0024] Best of all, the data in the atomic coordinate mapping of different imaging regions is unified to the same global coordinate system. Since different local coordinate systems are used when imaging different regions of the wafer, the data of each region has a position inconsistency when analyzed together. After being unified to the global coordinate system, all data has consistency in spatial position, which facilitates comprehensive processing. Then, the data in the global coordinate system is normalized. Since the data magnitudes of different parameters in the atomic coordinate mapping may differ greatly, normalization can unify the data to the same scale range, eliminate the influence of magnitude difference on subsequent calculation and model processing, and improve the efficiency and accuracy of data processing.

[0025] In a specific embodiment, after processing the data in the atomic coordinate mapping, the atomic coordinate mapping is converted into a database format to obtain wafer feature data. The multi-scale simulation model usually requires data input in a specific format. This format conversion can ensure that the generated wafer feature data can be smoothly recognized and read by the multi-scale simulation model, so that the data information contained in the atomic coordinate mapping can smoothly enter the processing flow of the multi-scale simulation model.

[0026] The boundary conditions and initial conditions of the multi-scale simulation model need to be set. The boundary conditions simulate the actual environment boundary of the wafer doping process, such as the limitation of external conditions such as temperature and pressure. The initial conditions determine the initial values of various parameters at the beginning of the multi-scale simulation model calculation, for example: the initial state of the wafer, the initial distribution of the doping atoms. By setting these conditions, the multi-scale simulation model is close to the actual process situation. After setting, the wafer feature data is simulated and calculated by the multi-scale simulation model. The multi-scale simulation model comprehensively considers multi-scale information such as quantum mechanics, material properties, and macro process conditions. Through algorithm calculation, the physical changes of the wafer under different process conditions are simulated, and simulation results containing various process-related information are generated.

[0027] In a specific embodiment, the initial process parameter set is extracted from the simulation results. The simulation results include a large amount of information, but for actual wafer doping, key process parameters such as: dopant dose, implant energy, and annealing time need to be selected. These process parameters constitute the initial process parameter set, providing an important reference for subsequent actual process operation, providing reasonable planning and adjustment of the process, reducing the number of blind attempts in actual process exploration, improving the accuracy of process planning, and thus improving the efficiency and quality of the entire wafer doping process.

[0028] In a specific embodiment, step S103 specifically includes: (1) converting the initial process parameter set into a process parameter sequence; (2) inputting the process parameter sequence into the embedding layer of the transformer model to generate an embedding vector sequence; (3) analyzing and processing the embedding vector sequence through attention calculation and multi-head merging in the transformer model to obtain an attention weighted sequence; (4) inputting the attention weighted sequence into a feedforward neural network for feature transformation and nonlinear mapping to generate a high-order feature sequence; (5) processing the high-order feature sequence through the reinforcement learning algorithm to obtain the dynamic process data matrix.

[0029] In a specific embodiment, the processing of the high-order feature sequence through the reinforcement learning algorithm to obtain the dynamic process data matrix specifically includes: (1) performing multi-scale feature fusion on the high-order feature sequence and compressing the fused high-order feature sequence through an autoencoder algorithm to obtain a reinforcement learning state vector; (2) calculating the reinforcement learning state vector through a proximal policy optimization algorithm to obtain a corresponding action vector; (3) dynamically adjusting the action vector by a dynamic matrix and generating the dynamic process data matrix in time sequence according to the adjusted action vector.

[0030] In a specific embodiment, the initial process parameter set is converted into a process parameter sequence, and the originally unordered or less structured initial process parameters are arranged in a sequence according to a certain logical order, so that subsequent processing and analysis of the parameters can be performed in order, and the unified data format improves the processability of the data, laying a foundation for deep analysis using the transformer model. In addition, the process parameter sequence is input into the embedding layer of the transformer model, and the embedding layer converts each process parameter sequence from the original numerical space to a low-dimensional vector space through a specific encoding method to generate an embedding vector sequence. In this vector space, the relationship between the process parameters can be more easily understood and processed. The process parameters are converted into a vector representation suitable for processing, reducing the data dimension and reducing the computational load. At the same time, by capturing the correlation information between the process parameters through the relative position and distance between the vectors, the ability to extract data features is enhanced. In a specific embodiment, attention calculation is performed on the embedding vector sequence in the transformer model, so that the transformer model can automatically focus on the information at different positions in the embedding vector sequence, assign different weights to each position according to the importance, and then integrate the attention calculation results of multiple different "perspectives" through a multi-head merging operation to obtain an attention weighted sequence. The embedding vector model can focus on the key information in the process parameter sequence, effectively capture the long-distance dependence and complex interaction between the parameters, and improve the comprehensiveness and accuracy of the analysis of the process parameter sequence. Best of all, the attention weighted sequence is input into a feedforward neural network, which includes multiple hidden layers. Through a series of linear transformations and nonlinear activation functions, the input attention weighted sequence is transformed and nonlinearly mapped. In the processing process, simple features are combined and abstracted into more complex and representative high-order features to form a high-order feature sequence, further extracting and enhancing the complex features in the process parameters, and better fitting the complex relationships between the process parameters. In a specific embodiment, multi-scale feature fusion is performed on the high-order feature sequence to comprehensively consider the process parameter features at different scales to obtain more comprehensive information, and then the fused high-order feature sequence is compressed through a self-encoder algorithm. The self-encoder can reduce the data dimension while preserving key information to obtain a reinforcement learning state vector. The reinforcement learning state vector includes the core features of the process parameters after screening and compression. Multi-scale feature fusion ensures that important information is not missed, and self-encoder compression reduces data redundancy and improves computational efficiency while highlighting key features. Further, the reinforcement learning state vector is calculated by a proximal policy optimization algorithm. The proximal policy optimization algorithm continuously optimizes the policy network, so that in a given environment (i.e., a state space formed by process parameter characteristics), an action that can maximize the cumulative reward is taken. According to the reinforcement learning state vector, the proximal policy optimization algorithm calculates a corresponding action vector. The action vector makes a decision according to the current process parameter characteristics, for example, the adjustment direction and amplitude of some process parameters. Through the proximal policy optimization algorithm, a better decision strategy can be quickly searched in a complex process parameter state space, and a reasonable adjustment action can be dynamically generated according to the change of the process parameters.

[0031] In an embodiment, the action vector is dynamically adjusted by a dynamic matrix. The dynamic matrix further corrects and optimizes the action vector according to real-time conditions in the process or pre-set rules. Then, the adjusted action vectors are sequentially combined to generate a dynamic process data matrix in time sequence. The dynamic process data matrix reflects the dynamic decisions made according to the change of process parameter characteristics at different time points, i.e., the dynamic change of process parameters. The dynamic matrix adjustment makes the decision more flexible and in line with actual process requirements. The generated dynamic process data matrix can accurately reflect the dynamic evolution of process parameters in the time dimension, provide accurate data support for real-time control and optimization of wafer doping processes, and help improve the stability of the process and product quality. In an embodiment, step S104 specifically includes: (1) performing data processing on the dynamic process data matrix to construct a process graph of node and edge relationships; (2) mapping the process characteristics of source materials and target materials in the process graph to the same feature space through feature mapping to realize feature alignment and construct a graph neural network model; (3) training the graph neural network model to obtain the cross-material process migration graph neural network; (4) calculating the corresponding features of the updated process parameter nodes in the dynamic process data matrix through the cross-material process migration graph neural network, and calculating and analyzing the feature relationships between the corresponding features to obtain the process parameter migration mapping relationship.

[0032] In an embodiment, the calculation of the corresponding features of the updated process parameter nodes in the dynamic process data matrix through the cross-material process migration graph neural network includes: (1) taking the process parameters in the dynamic process data matrix as nodes in the process graph and setting initial feature vectors for the nodes; (2) constructing a graph structure according to the correlations between the process parameters; (3) processing the initial feature vector of the node and the graph structure through an embedding layer of the graph neural network and mapping the initial feature vector of the node to a low-dimensional vector space to obtain an embedding representation of the node; (4) exchanging and updating information between adjacent nodes through the graph neural network and the embedding representation of the node to generate an updated corresponding feature of the process parameter node.

[0033] In a specific embodiment, data processing is performed on the dynamic process data matrix, each process parameter is regarded as a node, and a node and edge relationship is constructed according to the mutual correlation between process parameters, thereby forming a process graph. For example, if two process parameters have a direct influence relationship in physics, such as a dopant dose and the electrical performance of a wafer, a connection (edge) is established between the nodes representing the two parameters. The weight of the edge can be set according to the closeness of the parameter correlation. The complex relationship between process parameters is presented through an intuitive process graph, and the originally abstract data matrix is converted into a form that is easier to understand and analyze.

[0034] Optimally, the process characteristics of source materials and target materials in the process graph are mapped to the same feature space through a feature mapping method. The process characteristics of different materials can have different representations and scales. Feature alignment is to eliminate these differences so that they can be compared and fused in the same space. For example, the doping temperature and time characteristics of source materials and target materials are transformed into feature vectors with the same dimension and comparable scale through a specific linear or nonlinear transformation. Then, a graph neural network model is constructed, the structure (such as the number of layers, the number of neurons in each layer, etc.) and the connection mode of the network are determined, the unified representation of process characteristics of different materials is realized, the graph neural network can learn and analyze the commonality and differences of process parameters between different materials in the same framework, which lays a foundation for cross-material process transfer and widens the application scope of process optimization methods.

[0035] In a specific embodiment, the constructed graph neural network model is trained through a large number of samples containing process data of different materials. During the training process, the graph neural network model continuously adjusts its parameters (such as weights and biases) to minimize the error (such as mean square error) between the predicted results and the actual results. After multiple rounds of training, the graph neural network model gradually learns the complex relationships and patterns between process parameters of different materials, thereby obtaining a cross-material process transfer graph neural network. The graph neural network model is given the ability to transfer process knowledge across materials, which can use process data of known materials to predict and optimize process parameters of target materials, reducing the cost and time of process research and development for new target materials.

[0036] In a specific embodiment, the process parameters in the dynamic process data matrix are taken as nodes in the process graph, and an initial feature vector is set for each node. The initial feature vector can include the original numerical value of the process parameter, or a pre-processed (such as normalized, standardized) numerical value, or it can also combine some process-related additional information, such as the importance level of the parameter in the entire process flow, to provide an initial feature representation for the node as the starting point for subsequent feature updating and analysis, ensuring that each process parameter can be processed by the graph neural network in an appropriate manner.

[0037] Optimally, the graph structure is further refined and improved according to the internal correlation between process parameters. In addition to the basic connection relationship established earlier, the direction (directed or undirected edges) and weight of the edges can also be adjusted based on new information such as the order of process steps, causal relationships, etc. For example, if the change of a certain process parameter is directly caused by another parameter, the corresponding edge can be set as a directed edge, and the weight is adjusted according to the strength of the causal influence. This builds a more accurate and more reflective graph structure of the actual process, enabling the graph neural network model to better capture the interactions between process parameters during information transmission and feature updating.

[0038] In a specific embodiment, the initial feature vector of the node and the graph structure are processed by the embedding layer of the graph neural network. The embedding layer uses the adjacency matrix (describing the graph structure) and the node feature matrix (initial feature vector) to perform linear transformation on the initial feature vector of each node, and performs weighted summation with the information of the neighbor nodes, mapping the initial feature vector of the node to a low-dimensional vector space to obtain the embedding representation of the node. The dimension of this low-dimensional vector space is usually determined through experiments and model performance evaluation, ensuring that the node features can be fully extracted without excessive computational complexity. The high-dimensional complex initial features are converted into low-dimensional compact embedding representations that contain more structural information, reducing the data dimension and improving the computational efficiency, while enhancing the distinguishability of the features and the ability to reflect the graph structure.

[0039] In which, the adjacent nodes exchange information through the graph neural network. Under the message passing mechanism of the graph neural network, each node updates its embedding representation based on the embedding representation of its neighbor nodes and the weight of the edge. For example, in a graph convolutional network (GCN), the node updates its feature by aggregating the information of the neighbor nodes, performing weighted averaging and nonlinear activation function processing, and iteratively calculating through multiple layers of network. The node continuously absorbs the information of the neighbor nodes, enriches and perfects its own feature representation, and finally generates the updated feature corresponding to the process parameter node, enabling the node to capture the information of the surrounding nodes and the local and global mutual relationships between the process parameters, and the generated updated feature more comprehensively and accurately reflects the role and position of the process parameters in the entire process system.

[0040] In a specific embodiment, the corresponding features of the process parameter node are updated, the Pearson correlation coefficient index between the features is calculated through a correlation analysis algorithm, the correlation degree between different process parameter features is determined, the causal relationship between the parameters is determined through a causal analysis method (such as Granger causality test), based on the analysis result, the process parameter migration mapping relationship is extracted, that is, the rules of how the process parameters of different materials are related and converted, the internal relationship and change rule of the process parameters of different materials are realized, clear guidance is provided for migrating and optimizing the wafer doping process between different materials, the appropriate process parameters for the new target material can be quickly determined, and the efficiency and accuracy of process development are improved.

[0041] In a specific embodiment, step S105 specifically includes: (1) The measured data and the multi-scale model prediction results are compared point by point and the error value is calculated, and the error matrix is constructed from the error value; (2) The error matrix is analyzed and an error cause analysis report is generated; (3) The error cause analysis report and the current process parameter migration mapping relationship correct the process parameter migration mapping relationship to generate a corrected process parameter migration mapping relationship; (4) The error cause analysis report generates the multi-scale model update instruction according to the error cause analysis result of the error cause analysis report; (5) Determine whether the error value is less than a set threshold value, when the error value is greater than the set threshold value, the corrected process parameter migration mapping relationship and the updated multi-scale model are re-compared and analyzed, and the process parameter migration mapping relationship and the multi-scale model are updated.

[0042] In a specific embodiment, before the measured data and the multi-scale model prediction results are compared point by point and the error value is calculated, the error matrix is constructed from the error value, it includes: The measured data and the multi-scale model prediction results are data cleaned and normalized to generate preprocessed measured data and multi-scale model prediction results.

[0043] In a specific embodiment, the measured data has missing values and outliers, and the prediction results of the multi-scale model also have unreasonable data due to errors. Data cleaning is to identify and handle these problems, such as using interpolation method to fill in missing values, detecting and correcting or rejecting outliers based on statistical methods (such as Z-score), so that the data is more complete and reliable. The measured data and the prediction results may have different dimensions and value ranges, and normalization processing unifies them to the same scale, such as scaling the data to the [0, 1] interval, which improves the data quality, eliminates the differences in dimensions and ranges between data, lays a foundation for subsequent accurate comparison and error calculation, and avoids analysis deviation caused by data problems. The corresponding data points of the preprocessed measured data and the prediction results of the multi-scale model are compared one by one, the error values of each corresponding point are calculated, the calculation method of the error values can be selected according to the specific circumstances, and then the calculated error values are arranged into a matrix form according to certain rules, which is convenient for subsequent analysis and processing, and directly shows the differences between the measured data and the prediction results of the multi-scale model at each data point.

[0044] Among them, the error matrix is analyzed in multiple dimensions, such as the distribution of errors (mean, variance, etc.), the distribution characteristics of the error matrix at different process parameters or different process stages, and the possible causes of errors are inferred in combination with the process and actual production conditions, such as unreasonable model assumptions, measurement equipment precision problems, process condition fluctuations, etc., and the analysis results are arranged into an error cause analysis report to comprehensively find out the root cause of the error, and provide a clear method for correcting the process parameter migration mapping relationship and updating the multi-scale model.

[0045] In a specific embodiment, according to the error cause analysis report and the current process parameter migration mapping relationship, the process parameter migration mapping relationship is corrected, for example: if it is found through analysis that the mapping relationship of a certain process parameter leads to a larger error, the coefficient or function form of the parameter in the mapping relationship is adjusted to generate a corrected process parameter migration mapping relationship, so that the process parameter migration mapping relationship is more in line with the actual process situation, and the accuracy of process parameter prediction is improved, and the wafer doping process is optimized.

[0046] Best, according to the results of the error cause analysis report to determine the part and method of the multi-scale model that needs to be updated and generate multi-scale model update instructions, for example: if the error is caused by inaccurate setting of some physical parameters in the multi-scale model, update these parameters, if the multi-scale model structure is unreasonable, adjust the multi-scale model structure, not only to make the multi-scale model continuously adapt to the actual process changes, but also to improve the prediction accuracy and reliability of the multi-scale model.

[0047] An error threshold is set and the calculated error value is compared with the error threshold. If the error value exceeds the set threshold, it indicates that the current corrections and updates have not yet achieved the desired effect. The corrected process parameter migration mapping relationship and the updated multi-scale model need to be re-compared and analyzed. The process parameter migration mapping relationship and the multi-scale model must be corrected again and updated again. This process is repeated until the error value is less than the set threshold. Through continuous feedback and iteration, the process parameter migration mapping relationship and the multi-scale model are continuously optimized, making the predicted results increasingly close to the measured data, thereby improving the stability and consistency of the wafer doping process and enhancing product quality.

[0048] In the technical solution provided in the present application, atomic scale imaging is performed on the surface of the wafer to identify the target doping sites and generate an atomic coordinate mapping map, and then the atomic coordinate mapping map is input into a multi-scale simulation model to generate an initial process parameter set. The initial process parameter set is processed by combining a transformer model with a reinforcement learning algorithm to obtain a dynamic process data matrix. Subsequently, the dynamic process data matrix is ​​optimized by a cross-material process migration graph neural network to obtain a process parameter migration mapping relationship. Then, the measured data and the multi-scale model prediction results are compared through closed-loop feedback iteration and the corrected process parameter migration mapping relationship and the multi-scale model update instruction are output. Among them, atomic-scale imaging accurately identifies the target doping sites on the wafer surface and generates an atomic coordinate mapping map. The atomic-level positioning accuracy far exceeds the traditional micron level, ensuring process quality. The multi-scale simulation model generates an initial process parameter set based on the atomic coordinate mapping map and integrates multi-scale information, which not only shortens the trial cycle and reduces experimental costs, but also fits the actual process. In addition, the transformer model is combined with the reinforcement learning algorithm to process the initial process parameter set and generate a dynamic process data matrix that can reflect the dynamic changes of parameters in real time. Compared with static data, the flexibility and accuracy of process control are greatly improved. Not only that, the cross-material process migration map neural network optimizes the dynamic process data matrix and establishes a process parameter migration mapping relationship, breaking the material limitations and improving process stability and reliability. In addition, the closed-loop feedback iteration continuously compares the measured and predicted data to correct the mapping relationship and update the model, which significantly improves the stability, consistency and yield of the wafer doping process. This application achieves high precision and high uniformity of wafer doping.

[0049] Those skilled in the art will clearly understand that, for the convenience and brevity of description, the specific working processes of the above-described systems, systems and units can refer to the corresponding processes in the aforementioned method embodiments and will not be repeated here.

[0050] The above-described embodiments are only used to illustrate the technical solutions of the present application, and not to limit them; although the present application has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that they can still modify the technical solutions recorded in the foregoing embodiments, or make equivalent replacements for part of the technical features; and these modifications or replacements do not make the essence of the corresponding technical solutions deviate from the spirit and scope of the technical solutions of the embodiments of the present application.

Claims

1. A wafer doping control method based on intelligent feedback regulation, characterized in that: The wafer doping control method based on intelligent feedback regulation includes: performing atomic-scale imaging of the surface of the wafer to identify target doping sites and generate an atomic coordinate map; Inputting the atomic coordinate map into a multi-scale simulation model to generate an initial process parameter set; The initial process parameter set is processed by combining a transformer model with a reinforcement learning algorithm to obtain a dynamic process data matrix; Optimizing the dynamic process data matrix through a cross-material process migration graph neural network to obtain a process parameter migration mapping relationship; The measured data is compared with the prediction result of the multi-scale model through closed-loop feedback iteration, and the revised process parameter migration mapping relationship and the multi-scale model update instruction are output.

2. The wafer doping control method based on intelligent feedback regulation according to claim 1, characterized in that: Performing atomic-scale imaging of the wafer surface to identify target doping sites and generate an atomic coordinate map, including: Scanning the surface of the wafer using a scanning probe microscope to obtain atomic-level surface morphology data; Inputting the atomic-level surface morphology data into a convolutional neural network to identify the target doping sites and mark the three-dimensional coordinates of the target doping sites to obtain a target doping site list; The local coordinates of the target doping site list are converted to a wafer global coordinate system to generate the atomic coordinate map.

3. The wafer doping control method based on intelligent feedback regulation according to claim 2, characterized in that: Performing atomic-scale imaging of the surface of the wafer to identify target doping sites and generate an atomic coordinate map, further comprising: removing noise from the data in the atomic coordinate map by Gaussian filtering and filling missing data in the atomic coordinate map by spline interpolation; The data in the atomic coordinate maps of different imaging areas are unified into the same global coordinate system and the data in the coordinate system are normalized.

4. The wafer doping control method based on intelligent feedback regulation according to claim 1, characterized in that: Inputting the atomic coordinate map into a multi-scale simulation model to generate an initial process parameter set includes: Converting the atomic coordinate map into a database format to obtain wafer feature data, wherein the wafer feature data can be recognized by the multi-scale simulation model; Setting boundary conditions and initial conditions of the multi-scale simulation model, and performing simulation calculations on the wafer feature data using the multi-scale simulation model to generate simulation results; The initial process parameter set is extracted from the simulation results.

5. The wafer doping control method based on intelligent feedback regulation according to claim 1, characterized in that: The initial process parameter set is processed by combining the transformer model with the reinforcement learning algorithm to obtain a dynamic process data matrix, including: Converting the initial process parameter set into a process parameter sequence; Inputting the process parameter sequence into the embedding layer of the transformer model for encoding to generate an embedding vector sequence; The embedding vector sequence is analyzed and processed by attention calculation and multi-head merging in the transformer model to obtain an attention weighted sequence; Inputting the attention weighted sequence into a feedforward neural network for feature transformation and nonlinear mapping to generate a high-order feature sequence; The high-order feature sequence is processed by the reinforcement learning algorithm to obtain the dynamic process data matrix.

6. The wafer doping control method based on intelligent feedback regulation according to claim 5, characterized in that: Processing the high-order feature sequence by the reinforcement learning algorithm to obtain the dynamic process data matrix includes: Performing multi-scale feature fusion on the high-order feature sequence and compressing the fused high-order feature sequence through an autoencoder algorithm to obtain a reinforcement learning state vector; Calculating the reinforcement learning state vector using a proximal strategy optimization algorithm to obtain a corresponding action vector; The motion vector is dynamically adjusted through a dynamic matrix and the dynamic process data matrix is ​​generated in chronological order according to the adjusted motion vector.

7. The wafer doping control method based on intelligent feedback regulation according to claim 1, characterized in that: The dynamic process data matrix is ​​optimized by a cross-material process migration graph neural network to obtain a process parameter migration mapping relationship, including: Performing data processing on the dynamic process data matrix to construct a process diagram of node and edge relationships; Mapping the process features of the source material and the target material in the process graph to the same feature space through feature mapping to achieve feature alignment to construct a graph neural network model; Training the graph neural network model to obtain the cross-material process migration graph neural network; The dynamic process data matrix is ​​calculated by the cross-material process migration graph neural network to obtain the corresponding features of the updated process parameter nodes in the dynamic process data matrix, and the feature relationship between the corresponding features is calculated and analyzed to obtain the process parameter migration mapping relationship.

8. The wafer doping control method based on intelligent feedback regulation according to claim 7, characterized in that: Calculating the dynamic process data matrix through the cross-material process migration graph neural network to obtain updated corresponding features of the process parameter nodes in the dynamic process data matrix, including: Using the process parameters in the dynamic process data matrix as nodes in the process graph and setting initial feature vectors for the nodes; Constructing a graph structure based on the association between the process parameters; Processing the initial feature vector of the node and the graph structure through the embedding layer of the graph neural network and mapping the initial feature vector of the node to a low-dimensional vector space to obtain an embedded representation of the node; Adjacent nodes exchange and update information through the graph neural network and the embedded representation of the nodes to generate corresponding features of the updated process parameter nodes.

9. The wafer doping control method based on intelligent feedback regulation according to claim 1, characterized in that: Comparing the measured data with the prediction results of the multi-scale model through closed-loop feedback iteration and outputting the revised process parameter migration mapping relationship and the multi-scale model update instruction, including: Comparing the measured data with the prediction results of the multi-scale model point by point and calculating error values, and constructing the error values ​​into an error matrix; Analyzing the error matrix and generating an error cause analysis report; The error cause analysis report and the current process parameter migration mapping relationship are used to correct the process parameter migration mapping relationship to generate a corrected process parameter migration mapping relationship; generating the multi-scale model update instruction according to the error cause analysis result of the error cause analysis report; Determine whether the error value is less than a set threshold. When the error value is greater than the set threshold, re-compare and analyze the corrected process parameter migration mapping relationship and the updated multi-scale model, and update the process parameter migration mapping relationship and the multi-scale model.

10. The wafer doping control method based on intelligent feedback regulation according to claim 9, characterized in that: The measured data is compared point by point with the prediction result of the multi-scale model and error values ​​are calculated. Before the error values ​​are constructed into an error matrix, the following steps are included: Data cleaning and normalization are performed on the measured data and the multi-scale model prediction results to generate the pre-processed measured data and the multi-scale model prediction results.

Images