Wafer doping control method based on intelligent feedback adjustment

Abstract

The application relates to the technical field of semiconductor manufacturing, and discloses a wafer doping control method based on intelligent feedback adjustment, which comprises the following steps: atomic scale imaging is performed on the surface of a wafer to identify target doping sites and generate an atomic coordinate mapping diagram, the atomic coordinate mapping diagram is input into a multi-scale simulation model to generate an initial process parameter set, a dynamic process data matrix is obtained by combining a transformer model with a reinforcement learning algorithm to process the initial process parameter set, 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 measured data and multi-scale model prediction results are compared through closed-loop feedback iteration and output of a corrected process parameter migration mapping relationship and a multi-scale model update instruction. The application improves the stability, consistency and yield of wafer doping processes, and realizes high precision and high uniformity of wafer doping.

Description

Technical Field

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

[0002] In the semiconductor manufacturing field, wafer production is a crucial step. Wafer doping is the process of introducing impurity atoms into specific regions during semiconductor wafer manufacturing 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, and the precision and uniformity of doping have a vital impact on the performance, reliability, and yield of semiconductor devices.

[0003] Traditional wafer doping technologies, such as ion implantation and diffusion processes, frequently encounter problems during ion implantation, including energy fluctuations in the ion beam, poor scanning uniformity, and space charge effects. The energy output from the ion source is difficult to maintain absolute stability; even small energy deviations can lead to differences in the depth of implanted ions within the wafer, thus affecting doping accuracy. Furthermore, in ion implantation of large-area wafers, the scanning mechanism cannot guarantee uniform ion beam coverage across the entire wafer surface, resulting in inconsistent doping concentrations between the wafer edges and the center. This severely impacts chip performance consistency, especially during high-current implantation, where space charge forces distort the ion beam, further deteriorating doping uniformity. Due to the limitations of traditional technologies, wafer doping suffers from both low accuracy and poor uniformity. Summary of the Invention

[0004] This application provides a wafer doping control method based on intelligent feedback adjustment. By controlling wafer doping through intelligent feedback adjustment, the high precision requirements of wafer doping can be met, thereby improving the high accuracy and high uniformity of wafer doping.

[0005] In a first aspect, this application provides a wafer doping control method based on intelligent feedback regulation, the wafer doping control method based on intelligent feedback regulation comprising:

[0006] The surface of the wafer is imaged at the atomic scale to identify target doping sites and generate an atomic coordinate mapping map;

[0007] The atomic coordinate mapping diagram is input into the multi-scale simulation model to generate an initial set of process parameters;

[0008] A dynamic process data matrix is ​​obtained by processing the initial process parameter set by combining a transformer model with a reinforcement learning algorithm.

[0009] The process parameter migration mapping relationship is obtained by optimizing the dynamic process data matrix through a cross-material process migration graph neural network.

[0010] The measured data and the prediction results of the multi-scale model are compared through closed-loop feedback iteration, and the corrected process parameter migration mapping relationship and the multi-scale model update instruction are output.

[0011] In the technical solution provided in this application, atomic-scale imaging of the wafer surface is performed to identify target doping sites and generate an atomic coordinate mapping map. The atomic coordinate mapping map is then 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. Next, the dynamic process data matrix is ​​optimized by a cross-material process migration graph neural network to obtain a process parameter migration mapping relationship. Finally, 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 these features, atomic-scale imaging accurately identifies target doping sites on the wafer surface and generates an atomic coordinate mapping map. The atomic-level positioning accuracy far exceeds that of traditional micrometer-level positioning, 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. This shortens the exploration cycle, reduces experimental costs, and closely matches the actual process. In addition, the transformer model and reinforcement learning algorithm are combined 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, this greatly improves the flexibility and accuracy of process control. Furthermore, the cross-material process migration graph neural network optimizes the dynamic process data matrix and establishes a process parameter migration mapping relationship, breaking material limitations and improving process stability and reliability. Moreover, closed-loop feedback iteration continuously compares measured and predicted data to correct the mapping relationship and update the model, significantly improving the stability, consistency, and yield of the wafer doping process. This application achieves high precision and high uniformity in wafer doping. Attached Figure Description

[0012] To more clearly illustrate the technical solutions of the embodiments of the present invention, the drawings used in the description of the embodiments will be briefly introduced below. Obviously, the drawings described below are some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0013] Figure 1 This is a schematic flowchart of one embodiment of the wafer doping control method based on intelligent feedback adjustment in this application. Detailed Implementation

[0014] This application provides a wafer doping control method based on intelligent feedback regulation. The terms "first," "second," "third," "fourth," etc. (if present) in the specification, claims, and accompanying drawings of this application are used to distinguish similar objects and are not necessarily used to describe a specific order or sequence. It should be understood that such data can be interchanged where appropriate so that the embodiments described herein can be implemented in a sequence other than that illustrated or described herein. Furthermore, the terms "comprising" or "having" and any variations thereof are intended to cover a non-exclusive inclusion; for example, a process, method, system, product, or apparatus that comprises a series of steps or units is not necessarily limited to those steps or units explicitly listed, but may include other steps or units not explicitly listed or inherent to such processes, methods, products, or apparatus.

[0015] For ease of understanding, the specific process of the embodiments of this application is described below. Please refer to [link / reference]. Figure 1 One embodiment of the wafer doping control method based on intelligent feedback adjustment in this application includes:

[0016] Step S101: Perform atomic-scale imaging on the surface of the wafer to identify target doping sites and generate an atomic coordinate mapping map;

[0017] Step S102: Input the atomic coordinate mapping diagram into the multi-scale simulation model to generate an initial process parameter set;

[0018] Step S103: The initial process parameter set is processed by combining the transformer model with a reinforcement learning algorithm to obtain a dynamic process data matrix;

[0019] Step S104: Optimize the dynamic process data matrix using a cross-material process migration graph neural network to obtain the process parameter migration mapping relationship;

[0020] Step S105: Compare the measured data with the multi-scale model prediction results through closed-loop feedback iteration, and output the corrected process parameter migration mapping relationship and the multi-scale model update instruction.

[0021] In one specific embodiment, atomic-scale imaging of the wafer surface and generation of an atomic coordinate mapping map are performed, which enables precise identification of target doping sites at the atomic level. This provides an extremely accurate positional information basis for subsequent processes, greatly improving the accuracy of wafer doping positioning. It can advance the positioning accuracy from the micrometer level to the atomic scale, ensuring that the foundation and guarantee are laid in the most critical initial step.

[0022] In addition, the multi-scale simulation model generates an initial process parameter set based on the atomic coordinate mapping diagram. The initial process parameter set generated by the multi-scale simulation model by integrating information from multiple scales is not only efficient, but also highly consistent with actual process requirements, which greatly shortens the process parameter exploration cycle, reduces the number of experiments and costs, and improves the fit between the initial parameters and the actual process.

[0023] Furthermore, by combining the transformer model with reinforcement learning algorithms to process the initial process parameter set, the complex correlations and changing trends of the initial process parameter set at different stages can be dynamically captured to generate a dynamic process data matrix. The dynamic process data matrix reflects the dynamic changes of process parameters in real time, providing rich and accurate data support for process optimization. Compared with static data processing, it can more accurately adapt to the complex and ever-changing wafer doping process, improving the flexibility and accuracy of process control.

[0024] Among them, by optimizing the dynamic process data matrix through a cross-material process migration graph neural network, the migration mapping relationship of process parameters is established, which realizes the effective transfer of process knowledge between different materials, breaks the material limitation, enables mature processes to be quickly adapted to newly developed materials, accelerates the development process of new material processes, reduces R&D costs and time, and improves the stability and reliability of new processes on different materials.

[0025] In one specific embodiment, the measured data is compared with the prediction results of the multi-scale model in a closed-loop feedback iterative process to continuously correct the migration mapping relationship of process parameters and update the multi-scale model. This continuous iterative mechanism promotes continuous process optimization, significantly improves the stability, consistency and yield of wafer doping process, ensures continuous improvement of wafer doping quality, and meets the stringent requirements for high-performance and high-reliability chip production.

[0026] In one specific embodiment, step S101 specifically includes:

[0027] (1) The surface of the wafer is scanned by a scanning probe microscope to obtain atomic-level surface morphology data;

[0028] (2) Input 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 list of target doping sites;

[0029] (3) Transform the local coordinates of the target doping site list to the wafer global coordinate system to generate the atomic coordinate mapping.

[0030] In one specific embodiment, a scanning probe microscope is used to perform a fine scan of the wafer surface. The scanning probe microscope can acquire extremely 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 extremely high precision, converting the microscopic features such as the arrangement and undulation of surface atoms into electrical signals, and further converting them into digital atomic-level surface morphology data. The data not only contains the precise position information of the atoms on the wafer surface, but also reflects details such as surface roughness, flatness, and the relative positional relationships between atoms. The acquired atomic-level surface morphology data provides the most original and accurate source of information for the subsequent accurate identification of target doping sites. Compared with traditional low-resolution imaging technology, it can capture more subtle surface features, greatly improving the accuracy of understanding the microstructure of the wafer surface.

[0031] Furthermore, the acquired atomic-level surface morphology data is input into a specially trained convolutional neural network (CNN). Through a combination of multiple convolutional layers, pooling layers, and fully connected layers, the CNN automatically extracts key features from the atomic-level surface morphology data. Trained on a large amount of labeled atomic-level surface morphology data, the CNN learns the characteristic patterns of the target doping sites. When new atomic-level surface morphology data is input, the CNN can quickly analyze the data and identify the target doping sites. Once identified, the CNN further labels the three-dimensional coordinates of these sites in the original data coordinate system and compiles them into a list of target doping sites. This achieves automatic, efficient, and accurate identification and positioning of target doping sites, significantly improving accuracy and efficiency compared to manual identification, reducing human error, and providing precise target location information for subsequent process operations.

[0032] In one specific embodiment, the coordinates in the target doping site list are initially based on a local coordinate system. To enable the application of this site information uniformly throughout the entire wafer process, it is necessary to transform the local coordinates to the wafer global coordinate system. The transformation process is achieved through coordinate transformation algorithms and an understanding of the overall wafer structure. Through accurate coordinate transformation, each target doping site has a unique and definite position in the wafer global coordinate system. Integrating this target doping site information with global coordinates generates an atomic coordinate mapping map. The atomic coordinate mapping map intuitively shows the distribution of target doping sites on the wafer surface in the global coordinate system. This provides a unified, standardized, and easy-to-understand coordinate information foundation for subsequent multi-scale simulation model input and the entire wafer doping process planning, ensuring accurate grasp and effective utilization of doping sites throughout the entire wafer and improving the consistency and accuracy of process planning.

[0033] From acquiring atomic-level surface morphology data to accurately identifying and marking the target doping sites, and then generating atomic coordinate mapping maps, a high-precision and high-efficiency information acquisition and processing flow was constructed. This provided crucial basic information for subsequent wafer doping processes, and promoted the realization of high precision and high quality in wafer manufacturing processes.

[0034] In one specific embodiment, step S101 is followed by:

[0035] Noise in the atomic coordinate mapping graph is removed by Gaussian filtering, and missing data in the atomic coordinate mapping graph is filled by spline interpolation.

[0036] The data in the atomic coordinate mapping maps of different imaging regions are unified into the same global coordinate system and the data in the coordinate system are normalized.

[0037] In one specific embodiment, step S102 specifically includes:

[0038] (1) Convert the atomic coordinate mapping diagram into a database format to obtain wafer feature data, wherein the wafer feature data can be recognized by the multi-scale simulation model;

[0039] (2) Set the boundary conditions and initial conditions of the multi-scale simulation model, and use the multi-scale simulation model to perform simulation calculations on the wafer feature data to generate simulation results;

[0040] (3) Extract the initial process parameter set from the simulation results.

[0041] In one specific embodiment, Gaussian filtering is used to process the atomic coordinate mapping data. Due to various microscopic interferences during atomic-scale imaging, noise is present in the data. This noise can affect the accuracy of subsequent analyses of doping sites and related processes. Gaussian filtering can specifically attenuate noise signals while preserving effective feature information in the atomic coordinate mapping, making the data clearer and more accurate. Furthermore, spline interpolation is used to fill in missing data in the atomic coordinate mapping. During atomic-scale imaging, some data may be missing due to equipment malfunctions, signal obstruction, etc. Spline interpolation constructs a piecewise polynomial function and reasonably estimates the values ​​of missing data based on the distribution of known data points, thus ensuring the integrity of the atomic coordinate mapping data. This is crucial for the comprehensive analysis of the atomic distribution on the wafer surface and subsequent process simulation, avoiding analytical biases caused by missing data.

[0042] Ideally, the data in the atomic coordinate maps of different imaging regions should be unified into the same global coordinate system. Since different local coordinate systems are used when imaging different regions of the wafer, there is an inconsistency in the position of the data when merging and analyzing them. After unifying into the global coordinate system, all data have consistency in spatial position, which facilitates comprehensive processing. Then, the data in the global coordinate system should be normalized. Since the data magnitudes of different parameters in the atomic coordinate maps may vary greatly, normalization can unify the data to the same scale range, eliminate the impact of magnitude differences on subsequent calculations and model processing, and improve the efficiency and accuracy of data processing.

[0043] In one specific embodiment, after processing the data in the atomic coordinate mapping diagram, the atomic coordinate mapping diagram is converted into a database format to obtain wafer feature data. Multi-scale simulation models usually require data input in a specific format. This format conversion can ensure that the generated wafer feature data can be successfully recognized and read by the multi-scale simulation model, so that the data information contained in the atomic coordinate mapping diagram can be successfully entered into the processing flow of the multi-scale simulation model.

[0044] This involves setting boundary and initial conditions for the multi-scale simulation model. Boundary conditions simulate the actual environmental boundaries of the wafer doping process, such as external constraints like temperature and pressure. Initial conditions determine the initial values ​​of various parameters at the start of the multi-scale simulation model calculation, such as the initial state of the wafer and the initial distribution of doped atoms. By setting these conditions, the multi-scale simulation model closely approximates the actual process conditions. After setting these conditions, the multi-scale simulation model is used to simulate and calculate the wafer's characteristic data. The multi-scale simulation model comprehensively considers multi-scale information such as quantum mechanics, material properties, and macroscopic process conditions. Through algorithm calculations, it simulates the physical changes of the wafer under different process conditions and generates simulation results containing various process-related information.

[0045] In one specific embodiment, an initial set of process parameters is extracted from the simulation results. The simulation results contain a large amount of information, but for actual wafer doping, it is necessary to select key process parameters, such as dopant dosage, implantation energy, and annealing time. These process parameters constitute the initial set of process parameters, which provides an important reference for subsequent actual process operations. This enables reasonable planning and adjustment of the process, reduces the number of blind attempts in actual process exploration, improves the accuracy of process planning, and thus improves the efficiency and quality of the entire wafer doping process.

[0046] In one specific embodiment, step S103 specifically includes:

[0047] (1) Convert the initial set of process parameters into a sequence of process parameters;

[0048] (2) Input the process parameter sequence into the embedding layer of the transformer model for encoding to generate an embedding vector sequence;

[0049] (3) The embedded vector sequence is analyzed and processed by attention calculation and multi-head merging in the transformer model to obtain the attention-weighted sequence;

[0050] (4) Input the attention-weighted sequence into a feedforward neural network to perform feature transformation and nonlinear mapping to generate a higher-order feature sequence;

[0051] (5) The dynamic process data matrix is ​​obtained by processing the higher-order feature sequence through the reinforcement learning algorithm.

[0052] In one specific embodiment, the step of processing the high-order feature sequence using the reinforcement learning algorithm to obtain the dynamic process data matrix specifically includes:

[0053] (1) Perform multi-scale feature fusion on the high-order feature sequence and compress the fused high-order feature sequence using an autoencoder algorithm to obtain a reinforcement learning state vector;

[0054] (2) The corresponding action vector is obtained by calculating the reinforcement learning state vector through the proximal policy optimization algorithm;

[0055] (3) The motion vector is dynamically adjusted by a dynamic matrix and the dynamic process data matrix is ​​generated in chronological order based on the adjusted motion vector.

[0056] In one specific embodiment, the initial process parameter set is converted into a process parameter sequence. The original disordered or poorly structured initial process parameters are arranged into a sequence according to a certain logical order, so that the parameters can be processed and analyzed in sequence. The unified data format improves the data processability and lays the foundation for in-depth analysis using the transformer model.

[0057] Furthermore, the process parameter sequence is input into the embedding layer of the transformer model. The embedding layer transforms each process parameter sequence from the original numerical space to a low-dimensional vector space through a specific encoding method, generating an embedded vector sequence. In this vector space, the relationship between process parameters can be more easily understood and processed. The process parameters are converted into vector representations suitable for processing, reducing the data dimensionality and computational load. At the same time, the correlation information between process parameters is captured through the relative position and distance between vectors, enhancing the ability to extract data features.

[0058] In one specific embodiment, attention calculation is performed on the embedded vector sequence in the transformer model. This allows the transformer model to automatically focus on information at different positions within the embedded vector sequence, assigning different weights to each position based on its importance. Then, through a multi-head merging operation, the attention calculation results from multiple different "perspectives" are integrated to obtain an attention-weighted sequence. The embedded vector model can focus on key information in the process parameter sequence, effectively capturing long-distance dependencies and complex interactions between parameters, thus improving the comprehensiveness and accuracy of the process parameter sequence analysis.

[0059] Ideally, the attention-weighted sequence is input into a feedforward neural network, which contains multiple hidden layers. Through a series of linear transformations and nonlinear activation functions, the input attention-weighted sequence undergoes feature transformation and nonlinear mapping. During processing, simple features are combined and abstracted into more complex and representative higher-order features, forming a higher-order feature sequence. This further extracts and enhances the complex features in the process parameters, enabling a better fit to the complex relationships between process parameters.

[0060] In one specific embodiment, multi-scale feature fusion is performed on the high-order feature sequence to comprehensively consider the process parameter features at different scales in order to obtain more comprehensive information. Then, the fused high-order feature sequence is compressed by an autoencoder algorithm. By constructing an encoder and decoder structure, the autoencoder can reduce the data dimensionality while retaining 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 no important information is missed, and autoencoder compression reduces data redundancy, improves computational efficiency, and highlights key features.

[0061] Furthermore, the reinforcement learning state vector is calculated using a proximal policy optimization algorithm. By continuously optimizing the policy network, the proximal policy optimization algorithm enables actions that maximize cumulative rewards in a given environment (i.e., the state space constituted by process parameter features). Based on the reinforcement learning state vector, the proximal policy optimization algorithm calculates the corresponding action vector. The action vector represents the decision made based on the current process parameter features, such as the direction and magnitude of adjustment for certain process parameters. Through the proximal policy optimization algorithm, a better decision strategy can be quickly searched in the complex process parameter state space, and reasonable adjustment actions can be dynamically generated based on changes in process parameters.

[0062] In one specific embodiment, the action vectors are dynamically adjusted using a dynamic matrix. The dynamic matrix further corrects and optimizes the action vectors based on real-time conditions during the process or pre-set rules. Then, the adjusted action vectors are sequentially combined in chronological order to generate a dynamic process data matrix. The dynamic process data matrix reflects the dynamic decisions made based on changes in process parameter characteristics at different time points, i.e., the dynamic changes in process parameters. The dynamic matrix adjustment makes the decision-making 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, providing precise data support for real-time control and optimization of wafer doping processes, and helping to improve process stability and product quality.

[0063] In one specific embodiment, step S104 specifically includes:

[0064] (1) Perform data processing on the dynamic process data matrix to construct a process diagram of node and edge relationships;

[0065] (2) By mapping the process features of the source material and the target material in the process diagram to the same feature space through feature mapping, feature alignment is achieved to construct a graph neural network model;

[0066] (3) The graph neural network model is trained to obtain the cross-material process transfer graph neural network;

[0067] (4) The process parameter migration mapping relationship is obtained by 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 the feature relationship between the corresponding features is calculated and analyzed.

[0068] In one specific embodiment, the step of calculating the updated corresponding features of process parameter nodes in the dynamic process data matrix using the cross-material process migration graph neural network includes:

[0069] (1) Use the process parameters in the dynamic process data matrix as nodes in the process diagram and set an initial feature vector for the nodes;

[0070] (2) Construct a graph structure based on the relationship between the process parameters;

[0071] (3) The initial feature vector of the node and the graph structure are processed by the embedding layer of the graph neural network, and the initial feature vector of the node is mapped to a low-dimensional vector space to obtain the embedding representation of the node;

[0072] (4) Adjacent nodes exchange and update information through the graph neural network and the embedded representation of the nodes to generate the corresponding features of the updated process parameter nodes.

[0073] In one specific embodiment, the dynamic process data matrix is ​​processed, and each process parameter is regarded as a node. The relationship between nodes and edges is constructed based on the interrelationship between process parameters, thereby forming a process graph. For example, if two process parameters have a direct physical influence relationship, such as the doping dose being closely related to the electrical performance of the wafer, a connection (edge) is established between the nodes representing these two parameters. The weight of the edge can be set according to the tightness of the parameter relationship. The complex relationship between process parameters is presented through an intuitive process graph, transforming the originally abstract data matrix into a form that is easier to understand and analyze.

[0074] Ideally, feature mapping methods are used to map the process features of the source and target materials in the process diagram to the same feature space. Process features of different materials may have different representations and scales. Feature alignment aims to eliminate these differences and enable them to be compared and integrated in the same space. For example, the doping temperature, time, and other features of the source and target materials can be transformed into feature vectors with the same dimension and comparable scale through specific linear or nonlinear transformations. Then, a graph neural network model is constructed to determine the network structure (such as the number of layers, the number of neurons per layer, etc.) and connection method, thereby realizing a unified representation of the process features of different materials. This allows the graph neural network to learn and analyze the commonalities and differences of process parameters between different materials within the same framework, laying the foundation for cross-material process transfer and broadening the applicability of process optimization methods.

[0075] In one specific embodiment, the constructed graph neural network model is trained using a large number of samples containing process data of different materials. During the training process, the graph neural network model continuously adjusts its own 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. This endows the graph neural network model with the ability to transfer process knowledge across materials, enabling it to use process data of known materials to predict and optimize process parameters of target materials, thereby reducing the process development costs and time for new target materials.

[0076] In one specific embodiment, the process parameters in the dynamic process data matrix are used as nodes in the process graph, and an initial feature vector is set for each node. The initial feature vector may contain the original value of the process parameter, or it may be the value after preprocessing (such as normalization and standardization). It may also combine some process-related additional information, such as the importance level of the parameter in the entire process flow, to provide the node with an initial feature representation as the starting point for subsequent feature updates and analysis, ensuring that each process parameter can be processed by the graph neural network in an appropriate way.

[0077] Ideally, the graph structure can be further refined and improved based on the inherent relationships between process parameters. In addition to the basic connections established earlier, the direction (directed or undirected) and weight of edges can be adjusted according to new information, such as the sequence of process steps and causal relationships. 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 can be adjusted according to the strength of the causal influence. This constructs a more accurate graph structure that better reflects the actual process situation, enabling the graph neural network model to better capture the interaction between process parameters during information transmission and feature updates.

[0078] In one specific embodiment, the initial feature vectors of nodes 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) of the graph to perform a linear transformation on the initial feature vector of each node, and performs a weighted summation by combining the information of neighboring nodes, mapping the initial feature vector of the node to a low-dimensional vector space to obtain the embedded representation of the node. The dimension of this low-dimensional vector space is usually determined through experiments and model performance evaluation. It is necessary to ensure that the node features can be fully extracted without making the computation too large. The high-dimensional and complex initial features are transformed into a low-dimensional, compact embedding representation containing more structural information, reducing the data dimension, improving the computational efficiency, and enhancing the distinguishability of features and the ability to reflect the graph structure.

[0079] In this system, adjacent nodes exchange information through a graph neural network. Under the message passing mechanism of the graph neural network, each node updates its own embedding representation based on the embedding representation of its neighbors and the weights of the edges. For example, in a graph convolutional network (GCN), a node aggregates information from its neighbors, processes it through weighted averaging and nonlinear activation functions to obtain updated features. Through iterative calculations of multiple layers of the network, the node continuously absorbs information from its neighbors, enriches and improves its own feature representation, and finally generates the corresponding updated features of the process parameter nodes. This allows the nodes to capture information from surrounding nodes and the local and global relationships between process parameters. The generated updated features more comprehensively and accurately reflect the role and status of process parameters in the entire process system.

[0080] In one specific embodiment, the corresponding features after the process parameter node is updated are calculated and analyzed. The Pearson correlation coefficient between features is calculated using a correlation analysis algorithm to determine the degree of correlation between different process parameter features. Causal analysis methods (such as Granger causality test) are used to determine the causal relationship between parameters. Based on the analysis results, the process parameter migration mapping relationship is extracted, that is, the rules of how process parameters between different materials are related and transformed. This realizes the intrinsic connection and change law of process parameters between different materials, providing clear guidance for the migration and optimization of wafer doping processes between different materials. It can quickly determine the appropriate process parameters for new target materials, improving the efficiency and accuracy of process development.

[0081] In one specific embodiment, step S105 specifically includes:

[0082] (1) Compare the measured data with the prediction results of the multi-scale model point by point and calculate the error value, and construct an error matrix from the error value;

[0083] (2) Analyze the error matrix and generate an error cause analysis report;

[0084] (3) The error cause analysis report and the current process parameter migration mapping relationship are used to correct the process parameter migration mapping relationship and generate a corrected process parameter migration mapping relationship;

[0085] (4) Generate the multi-scale model update instruction based on the error cause analysis results in the error cause analysis report;

[0086] (5) Determine whether the error value is less than the set threshold. If the error value is greater than the set threshold, compare and analyze the corrected process parameter migration mapping relationship and the updated multi-scale model again, and update the process parameter migration mapping relationship and the multi-scale model.

[0087] In one specific embodiment, the step of comparing the measured data with the multi-scale model prediction results point by point and calculating the error value, and constructing the error value into an error matrix, includes:

[0088] The measured data and the multi-scale model prediction results are cleaned and normalized to generate preprocessed measured data and multi-scale model prediction results.

[0089] In one specific embodiment, the measured data contains missing values ​​and outliers, and the multi-scale model prediction results also contain unreasonable data due to errors. Data cleaning is to identify and handle these problems, such as using interpolation to fill in missing values, and using statistical methods (such as Z-score) to detect, correct, or remove outliers, making the data more complete and reliable. The measured data and prediction results may have different dimensions and value ranges. Normalization processing unifies them to the same scale, such as scaling the data to the [0, 1] interval, which improves data quality, eliminates differences in dimensions and ranges between data, lays the foundation for accurate comparison and error calculation in the future, and avoids analytical bias caused by data problems. The preprocessed measured data and the corresponding data points of the multi-scale model prediction results are compared one by one, and the error value of each corresponding point is calculated. The method for calculating the error value can be selected according to the specific situation. Then, the calculated error values ​​are arranged into a matrix form according to certain rules, which facilitates subsequent analysis and processing and intuitively shows the differences between the measured data and the multi-scale model prediction results at each data point.

[0090] This involves multi-dimensional analysis of the error matrix, such as the distribution of statistical errors (mean, variance, etc.), analyzing the distribution characteristics of the error matrix under different process parameters or different process stages, and speculating on possible causes of errors based on process and actual production conditions, such as unreasonable model assumptions, measurement equipment accuracy issues, and fluctuations in process conditions. The analysis results are then compiled into an error cause analysis report to comprehensively identify the root causes of errors and provide a clear method for correcting the migration mapping relationship of process parameters and updating the multi-scale model.

[0091] In one specific embodiment, the process parameter migration mapping relationship is corrected based on the error cause analysis report and the current process parameter migration mapping relationship. For example, if the analysis finds that the mapping relationship of a certain process parameter causes a large 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, improving the accuracy of process parameter prediction, and thus optimizing the wafer doping process.

[0092] Ideally, based on the results of the error cause analysis report, the parts and methods that need to be updated in the multi-scale model are determined, and multi-scale model update instructions are generated. For example, if the error is caused by inaccurate settings of certain physical parameters in the multi-scale model, these parameters are updated; if the multi-scale model structure is unreasonable, the multi-scale model structure is adjusted. This not only enables the multi-scale model to continuously adapt to changes in actual processes, but also improves the prediction accuracy and reliability of the multi-scale model.

[0093] The process involves setting an error threshold and comparing the calculated error value with this threshold. If the error value exceeds the threshold, it indicates that the current correction and update have not achieved the desired effect. In this case, the corrected process parameter migration mapping relationship and the updated multi-scale model need to be re-compared and analyzed, and the process parameter migration mapping relationship and multi-scale model need to be corrected and updated again. This process is iterated repeatedly until the error value is less than the threshold. Through continuous feedback and iteration, the process parameter migration mapping relationship and multi-scale model are continuously optimized, making the predicted results increasingly closer to the measured data. This improves the stability and consistency of the wafer doping process and enhances product quality.

[0094] In the technical solution provided in this application, atomic-scale imaging of the wafer surface is performed to identify target doping sites and generate an atomic coordinate mapping map. The atomic coordinate mapping map is then 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. Next, the dynamic process data matrix is ​​optimized by a cross-material process migration graph neural network to obtain a process parameter migration mapping relationship. Finally, 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 these features, atomic-scale imaging accurately identifies target doping sites on the wafer surface and generates an atomic coordinate mapping map. The atomic-level positioning accuracy far exceeds that of traditional micrometer-level positioning, 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. This shortens the exploration cycle, reduces experimental costs, and closely matches the actual process. In addition, the transformer model and reinforcement learning algorithm are combined 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, this greatly improves the flexibility and accuracy of process control. Furthermore, the cross-material process migration graph neural network optimizes the dynamic process data matrix and establishes a process parameter migration mapping relationship, breaking material limitations and improving process stability and reliability. Moreover, closed-loop feedback iteration continuously compares measured and predicted data to correct the mapping relationship and update the model, significantly improving the stability, consistency, and yield of the wafer doping process. This application achieves high precision and high uniformity in wafer doping.

[0095] Those skilled in the art will understand that, for the sake of convenience and brevity, the specific working process of the system and units described above can be referred to the corresponding process in the foregoing method embodiments, and will not be repeated here.

[0096] The above-described embodiments are only used to illustrate the technical solutions of this application, and are not intended to limit them. Although this application has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some of the technical features. Such modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the spirit and scope of the technical solutions of the embodiments of this 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 adjustment includes: The surface of the wafer is imaged at the atomic scale to identify target doping sites and generate an atomic coordinate mapping map; The atomic coordinate mapping diagram is input into the multi-scale simulation model to generate an initial set of process parameters; A dynamic process data matrix is ​​obtained by processing the initial process parameter set by combining a transformer model with a reinforcement learning algorithm. The process includes: 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; analyzing and processing the embedding vector sequence through 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 higher-order feature sequence; performing multi-scale feature fusion on the higher-order feature sequence and compressing the fused higher-order feature sequence using an autoencoder algorithm to obtain a reinforcement learning state vector; calculating the corresponding action vector using a proximal policy optimization algorithm on the reinforcement learning state vector; dynamically adjusting the action vector using a dynamic matrix and generating the dynamic process data matrix in chronological order based on the adjusted action vector. The process parameter migration mapping relationship is obtained by optimizing the dynamic process data matrix through a cross-material process migration graph neural network. The measured data and the prediction results of the multi-scale model are compared iteratively through closed-loop feedback, and the corrected process parameter migration mapping relationship and the multi-scale model update instruction are output.

2. The wafer doping control method based on intelligent feedback adjustment according to claim 1, characterized in that, Atomic-scale imaging of the wafer surface to identify target doping sites and generate an atomic coordinate mapping map includes: Atomic-level surface morphology data were obtained by scanning the surface of the wafer using a scanning probe microscope. The atomic-level surface morphology data is input 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 list of target doping sites; The local coordinates of the target doping site list are transformed to the wafer global coordinate system to generate the atomic coordinate mapping.

3. The wafer doping control method based on intelligent feedback adjustment according to claim 2, characterized in that, The method further includes performing atomic-scale imaging of the wafer surface to identify target doping sites and generate an atomic coordinate mapping map, and also includes: Noise in the atomic coordinate mapping graph is removed by Gaussian filtering, and missing data in the atomic coordinate mapping graph is filled by spline interpolation. The data in the atomic coordinate mapping maps of different imaging regions 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, The atomic coordinate mapping is input into the multi-scale simulation model to generate an initial set of process parameters, including: The atomic coordinate mapping is converted into a database format to obtain wafer feature data, wherein the wafer feature data can be recognized by the multi-scale simulation model; The boundary conditions and initial conditions of the multi-scale simulation model are set, and the simulation results are generated by simulating the wafer feature data through the multi-scale simulation model. The initial process parameter set is extracted from the simulation results.

5. The wafer doping control method based on intelligent feedback adjustment according to claim 1, characterized in that, The dynamic process data matrix is ​​optimized using a cross-material process migration graph neural network to obtain the process parameter migration mapping relationship, including: The dynamic process data matrix is ​​processed to construct a process graph showing the relationships between nodes and edges; Feature mapping is used to map the process features of the source material and the target material in the process diagram to the same feature space to achieve feature alignment and construct a graph neural network model. The graph neural network model is trained to obtain the cross-material process transfer graph neural network; The process parameter migration mapping relationship is obtained by 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 by calculating and analyzing the feature relationship between the corresponding features.

6. The wafer doping control method based on intelligent feedback regulation according to claim 5, characterized in that, The updated features of process parameter nodes in the dynamic process data matrix are obtained by calculating the dynamic process data matrix using the cross-material process migration graph neural network, including: The process parameters in the dynamic process data matrix are used as nodes in the process diagram, and an initial feature vector is set for the nodes. A graph structure is constructed based on the relationships between the process parameters; The initial feature vector of the node and the graph structure are processed by the embedding layer of the graph neural network, and the initial feature vector of the node is mapped to a low-dimensional vector space to obtain the 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 the corresponding features of the updated process parameter nodes.

7. The wafer doping control method based on intelligent feedback regulation according to claim 1, characterized in that, 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, including: The measured data is compared point by point with the prediction results of the multi-scale model, and the error value is calculated. The error values ​​are then used to construct an error matrix. The error matrix is ​​analyzed and an error cause analysis report is generated; Based on the error cause analysis report and the current process parameter migration mapping relationship, the process parameter migration mapping relationship is corrected and a corrected process parameter migration mapping relationship is generated. The multi-scale model update instruction is generated based on the error cause analysis results in the error cause analysis report; Determine whether the error value is less than a set threshold. If the error value is greater than the set threshold, then 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.

8. The wafer doping control method based on intelligent feedback regulation according to claim 7, characterized in that, The measured data is compared point-by-point with the prediction results of the multi-scale model, and the error values ​​are calculated. The error values ​​are then used to construct an error matrix, which includes: The measured data and the multi-scale model prediction results are cleaned and normalized to generate preprocessed measured data and multi-scale model prediction results.

Images