Dynamic adjustment method for large beam uniformity of ion implanter

Abstract

The invention relates to the technical field of ion implantation, and discloses a dynamic adjustment method for large beam uniformity of an ion implanter. According to the method, a beam spot distribution data sequence in a beam transmission path is collected in real time, uniformity drift prediction is carried out in combination with an injection process parameter set, and beam uniformity prediction parameters are generated. And capturing a current beam spot form image by using a beam spot imaging device, inputting the current beam spot form image and the prediction parameters into a plurality of beam spot classifiers containing independent classification paths in parallel, and identifying to obtain actual beam spot distortion parameters. And executing electromagnetic field parameter prediction based on the parameter, generating a predicted electromagnetic field adjustment parameter, collecting an adjusted beam spot image through a feedback mechanism, and inputting the beam spot image into a plurality of electromagnetic field parameter classifiers in parallel again to output an actual adjustment parameter. And finally, integrating the distortion parameter and the adjustment parameter to generate a dynamic adjustment instruction set, and realizing closed-loop control of beam uniformity. According to the method, the problem of uniformity drift in the large-beam ion implantation process is effectively solved, and the process stability and the implantation efficiency are improved.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of ion implantation, in particular to a dynamic adjustment method for large beam uniformity of an ion implanter. BACKGROUND

[0002] In the field of semiconductor manufacturing, as a key equipment, the performance of the ion implanter directly affects the electrical characteristics and yield of the chip. During the ion implantation process, beam uniformity is one of the core indicators for measuring the performance of the equipment, which determines the consistency of ion doping concentration on the wafer surface. With the development of semiconductor devices towards high precision and high integration, more stringent requirements are put forward for the beam uniformity of the ion implanter, especially in the large beam injection scenario, the control difficulty of beam uniformity is significantly increased.

[0003] The existing beam uniformity adjustment method of the ion implanter relies on static calibration or periodic detection, which is difficult to cope with the dynamic changes in the beam transmission process. The beam is easily disturbed by various factors in the transmission path, such as electromagnetic field fluctuation, vacuum environment change, ion source stability difference, and slight deformation of transmission components, etc. These factors will cause random distortion of the beam spot shape, and further destroy the beam uniformity. The traditional adjustment method usually adjusts in a fixed mode based on the preset process parameters, and lacks dynamic perception of real-time beam spot data. When the beam drifts suddenly, the adjustment response is lagged, and it is difficult to quickly correct the deviation.

[0004] In the prior art, the recognition of beam spot distortion mainly uses a single detection path or a single classification model, which is limited by the generalization ability of the model and is difficult to fully capture the beam spot shape changes under complex working conditions. For example, when the beam spot has both symmetric distortion and asymmetric shift, the single classifier is prone to incomplete feature extraction, resulting in inaccurate distortion parameter recognition. In the electromagnetic field adjustment link, the traditional method often generates adjustment parameters based on empirical formulas or simplified models, without fully combining real-time feedback data after beam action, so that the adjustment parameters deviate from the actual demand, affecting the uniformity correction effect.

[0005] With the widespread application of large beam injection process, the increase of beam intensity further amplifies the instability in the transmission process, the drift rate of beam uniformity is accelerated, and the time resolution and adjustment accuracy of the traditional adjustment method cannot meet the process requirements of modern semiconductor manufacturing, so there is an urgent need for a beam uniformity control method that can realize real-time perception, accurate prediction and dynamic adjustment. SUMMARY

[0006] The purpose of the present application is to provide a dynamic adjustment method for large beam uniformity of an ion implanter to solve the problems raised in the background art.

[0007] To achieve the above object, the application provides a dynamic regulation method for ion implanter large beam uniformity, which comprises the following steps:

[0008] Real-time acquisition of beam spot distribution data sequence in the beam transmission path of the ion implanter, synchronous acquisition of the injection process parameter set, uniformity drift prediction based on the beam spot distribution data sequence, and generation of beam uniformity prediction parameters;

[0009] Capture of a current beam spot morphology image by a beam spot imaging device, parallel input of the current beam spot morphology image into multiple beam spot classifiers in combination with the beam uniformity prediction parameters, and identification of actual beam spot distortion parameters, wherein each beam spot classifier comprises multiple independently running classification paths;

[0010] Execution of an electromagnetic field parameter prediction operation according to the actual beam spot distortion parameters, generation of predicted electromagnetic field regulation parameters, acquisition of a beam spot feedback image after electromagnetic field action by the beam spot imaging device, parallel input of the beam spot feedback image into multiple electromagnetic field parameter classifiers according to the predicted electromagnetic field regulation parameters, and output of actual electromagnetic field regulation parameters;

[0011] Integration of the actual beam spot distortion parameters and the actual electromagnetic field regulation parameters to generate a beam uniformity dynamic regulation instruction set.

[0012] Preferably, the real-time acquisition of the beam spot distribution data sequence in the beam transmission path of the ion implanter, the synchronous acquisition of the injection process parameter set, the uniformity drift prediction based on the beam spot distribution data sequence, and the generation of the beam uniformity prediction parameters comprise the following steps:

[0013] Continuous acquisition of the beam spot distribution data sequence by a beam probe array during continuous operation of the injection process;

[0014] Periodic recording of an injection process parameter set composed of injection energy, injection dose, and scanning frequency;

[0015] Extraction of statistical characteristic quantities of the beam spot distribution data sequence to generate a beam spot stability index;

[0016] Combination of the injection process parameter set and the beam spot stability index to execute a uniformity drift prediction operation to output beam uniformity prediction parameters.

[0017] Preferably, the combination of the injection process parameter set and the beam spot stability index to execute a uniformity drift prediction operation to output beam uniformity prediction parameters comprises the following steps:

[0018] Acquisition of a historical process database of the same type of ion implanter, extraction of a historical beam spot distribution data sequence set, a historical injection process parameter set, and labeling of beam spot uniformity deviation amounts at corresponding historical process times to form a beam spot uniformity deviation sample set.

[0019] constructing a beam uniformity prediction model based on time series decomposition;

[0020] supervising training and verification of the beam uniformity prediction model using the set of historical beam spot distribution data, the set of historical injection process parameters, and the set of beam spot uniformity deviation samples;

[0021] inputting the current set of injection process parameters and the current beam spot stability index into the trained beam uniformity prediction model, and outputting beam uniformity prediction parameters.

[0022] Preferably, the current beam spot morphology image is captured by the beam spot imaging device, and the current beam spot morphology image is input into multiple beam spot classifiers in parallel based on the beam uniformity prediction parameters to identify and obtain actual beam spot distortion parameters, including:

[0023] acquiring a current beam spot morphology image of a target beam spot region by a beam spot imaging device;

[0024] According to the numerical range of the beam uniformity prediction parameters, the K beam spot distortion levels with the highest matching degree are selected, where K is a predetermined positive integer;

[0025] The deviation amplitudes of the K beam spot distortion levels and the beam uniformity prediction parameters are calculated, and the number of activated classification paths of each beam spot classifier is determined accordingly;

[0026] Selecting K beam spot classifiers corresponding to the K beam spot distortion levels, inputting the current beam spot morphology image into the activated classification paths of each beam spot classifier respectively, and generating K sets of beam spot classification result sets;

[0027] Statistically analyzing the proportion of valid judgment results in the K sets of beam spot classification result sets, and outputting K beam spot distortion level probability values;

[0028] Selecting the beam spot distortion level with the largest probability value as the actual beam spot distortion parameter.

[0029] Preferably, the deviation amplitudes of the K beam spot distortion levels and the beam uniformity prediction parameters are calculated, and the number of activated classification paths of each beam spot classifier is determined accordingly, including:

[0030] The deviation amplitude values of the K beam spot distortion levels and the beam uniformity prediction parameters are calculated respectively;

[0031] The deviation amplitude values are converted into the number of activated classification paths by using a preset coefficient conversion rule;

[0032] activating a pre-trained beam spot classifier set, wherein each beam spot classifier corresponds to a specific beam spot distortion level and includes multiple classification paths based on a residual network;

[0033] According to the number of activated classification paths, a corresponding number of classification paths in each beam spot classifier are randomly selected to perform image analysis.

[0034] Preferably, the electromagnetic field parameter prediction operation according to the actual beam spot distortion parameter generates a predicted electromagnetic field adjustment parameter, including:

[0035] Obtain electromagnetic field correction records corresponding to different beam spot distortion parameters in a historical process database;

[0036] Extract a historical electromagnetic field adjustment parameter sample set and a beam spot distortion parameter sample set;

[0037] Construct an electromagnetic field parameter predictor based on physical model correction;

[0038] Train the electromagnetic field parameter predictor using the beam spot distortion parameter sample set and the historical electromagnetic field adjustment parameter sample set;

[0039] Input the actual beam spot distortion parameter into the trained electromagnetic field parameter predictor to output a predicted electromagnetic field adjustment parameter.

[0040] Preferably, the beam spot feedback image after electromagnetic field action is collected by the beam spot imaging device, and the beam spot feedback image is input into multiple electromagnetic field parameter classifiers in parallel according to the predicted electromagnetic field adjustment parameter to output an actual electromagnetic field adjustment parameter, including:

[0041] After applying the predicted electromagnetic field adjustment parameter, a beam spot feedback image of the beam current transmission path is collected;

[0042] Select M electromagnetic field parameter levels adjacent to the predicted electromagnetic field adjustment parameter;

[0043] Call a group of electromagnetic field parameter classifiers corresponding to the M electromagnetic field parameter levels;

[0044] Input the beam spot feedback image into all classification paths of each electromagnetic field parameter classifier to generate M sets of electromagnetic field classification result sets;

[0045] Calculate the proportion of valid determination results of each set of electromagnetic field classification result sets and output M electromagnetic level probability distributions;

[0046] Select the electromagnetic field parameter level with the highest probability distribution as the actual electromagnetic field adjustment parameter.

[0047] Preferably, the calling of the group of electromagnetic field parameter classifiers corresponding to the M electromagnetic field parameter levels includes:

[0048] Load a pre-trained electromagnetic field parameter classifier library, wherein each classifier corresponds to a specific electromagnetic field parameter level and contains multiple classification paths based on attention mechanisms;

[0049] The training data of the classification path comprises sample beam spot images and their corresponding determination labels of electromagnetic field parameter levels.

[0050] Preferably, the integration of the actual beam spot distortion parameters and the actual electromagnetic field adjustment parameters generates a dynamic beam uniformity adjustment instruction set, comprising:

[0051] mapping the actual beam spot distortion parameters into beam current correction amounts;

[0052] converting the actual electromagnetic field adjustment parameters into magnetic field strength adjustment amounts and electric field angle adjustment amounts;

[0053] fusing the beam current correction amounts, the magnetic field strength adjustment amounts and the electric field angle adjustment amounts to generate a dynamic adjustment instruction set.

[0054] Preferably, the beam spot distribution data sequence is divided into a real-time acquisition unit and a historical analysis unit, comprising:

[0055] In a continuous injection process, the beam cross-sectional density distribution data is captured by a distributed sensor array to form a real-time beam spot distribution data sequence;

[0056] extracting beam fluctuation characteristic data under the same working condition from a historical process database to constitute a historical beam spot distribution data sequence;

[0057] inputting the real-time beam spot distribution data sequence and the historical beam spot distribution data sequence into a spatio-temporal feature fusion model to output an enhanced beam spot distribution feature set for subsequent uniformity drift prediction.

[0058] Compared with the prior art, the present application has the following beneficial effects:

[0059] The dynamic adjustment method for large beam uniformity of the ion implanter effectively solves the problem of uniformity control in the large beam injection process through the coordinated design of multiple links. In the data acquisition link, the beam spot distribution data sequence and the injection process parameter set are acquired in real time, the beam transmission state is comprehensively monitored, the subtle changes in beam uniformity can be captured in time, the limitations of the traditional method relying on static data are broken, and the adjustment process has a dynamic perception basis. The uniformity drift prediction based on the beam spot distribution data sequence can perceive the trend of the change in beam uniformity in advance, provide a forward-looking basis for subsequent adjustment, and avoid the lag problem caused by passive adjustment.

[0060] In the beam spot distortion identification link, the current beam spot shape image is captured by a beam spot imaging device, and the image is input into multiple beam spot classifiers in parallel in combination with the beam current uniformity prediction parameters. Each classifier includes multiple independently running classification paths. This multi-path parallel processing mode can extract beam spot features from different dimensions, cover more types of distortion modes, reduce the missed or misjudged of single classification path for complex distortion, and improve the comprehensiveness and reliability of the actual beam spot distortion parameter identification.

[0061] In the electromagnetic field adjustment parameter generation link, the preliminary adjustment parameters are generated by performing prediction operation according to the actual beam spot distortion parameters, and are not directly applied, but the feedback image of the beam spot after the action of the electromagnetic field is collected by the beam spot imaging device, and the feedback image is input into multiple electromagnetic field parameter classifiers in parallel again, and the actual adjustment parameters are output. This feedback-based secondary optimization mechanism can effectively correct the deviation between the preliminary prediction parameters and the actual effect, realize the dynamic calibration of the adjustment parameters, and make the electromagnetic field adjustment more suitable for the real-time state of the beam current.

[0062] By integrating the actual beam spot distortion parameters and the actual electromagnetic field adjustment parameters to generate a dynamic adjustment instruction set, the whole adjustment process forms a closed loop cooperation. From beam current state perception, distortion identification to adjustment parameter optimization, the data of each link supports each other, can quickly respond to the dynamic changes of the beam current uniformity, adapt to the differences in beam current characteristics under different injection process parameters, reduce the influence of external interference on the beam current uniformity, and improve the stability and adaptability of the uniformity control in the large beam current injection process. BRIEF DESCRIPTION OF DRAWINGS

[0063] Figure 1 The working principle diagram of the dynamic adjustment method of the ion implanter large beam current uniformity described in the application;

[0064] Figure 2 The flowchart for generating the beam current uniformity prediction parameters;

[0065] Figure 3 The flowchart for the uniformity drift prediction operation;

[0066] Figure 4 The flowchart for generating the predicted electromagnetic field adjustment parameters. DETAILED DESCRIPTION

[0067] The technical solutions in the embodiments of the application will be clearly and completely described below with reference to the drawings in the embodiments of the application. Obviously, the described embodiments are only part of the embodiments of the application, not all. Based on the embodiments in the application, all other embodiments obtained by those skilled in the art without creative labor are within the scope of protection of the application.

[0068] Please refer toFigure 1 The present application provides a method for dynamically adjusting the uniformity of a large beam in an ion implanter, the method comprising:

[0069] The method for dynamically adjusting the uniformity of a large beam in an ion implanter comprises the coordinated execution of multiple steps. During the operation of an ion implantation process, a sequence of beam spot distribution data is collected in real time in the beam transport path, while a set of implantation process parameters is simultaneously acquired. Based on the sequence of beam spot distribution data, a uniformity drift prediction operation is performed to generate beam uniformity prediction parameters. Subsequently, a current beam spot morphology image is captured by a beam imaging device, which is input into multiple beam spot classifiers in parallel along with the beam uniformity prediction parameters. Each beam spot classifier contains multiple independently running classification paths, through which actual beam spot distortion parameters are identified. Next, an electromagnetic field parameter prediction operation is performed based on the actual beam spot distortion parameters to generate predicted electromagnetic field adjustment parameters. After applying the predicted electromagnetic field adjustment parameters, a beam spot feedback image after the action of the electromagnetic field is collected using the beam imaging device. The beam spot feedback image is input into multiple electromagnetic field parameter classifiers in parallel along with the predicted electromagnetic field adjustment parameters, and actual electromagnetic field adjustment parameters are output. Finally, the actual beam spot distortion parameters and the actual electromagnetic field adjustment parameters are integrated to generate a set of beam uniformity dynamic adjustment instructions. This method ensures real-time monitoring and adaptive adjustment of beam uniformity, meeting the precision requirements in industrial applications.

[0070] Example 1: refer to Figure 2 In the method for dynamically adjusting the uniformity of a large beam in an ion implanter, Example 1 involves collecting a sequence of beam spot distribution data in the beam transport path in real time, simultaneously acquiring a set of implantation process parameters, and performing uniformity drift prediction based on these data to generate beam uniformity prediction parameters.

[0071] A beam probe array continuously collects a sequence of beam spot distribution data during the continuous operation of an implantation process. The array is composed of multiple high-precision sensors, including Faraday cups, electrostatic probes, or semiconductor detectors, distributed at key positions in the beam transport path, such as near focusing lenses, deflection magnets, and scanning units. The sensors record the density distribution of the beam cross-section in a high-frequency sampling manner, with a sampling frequency usually in the millisecond range, ensuring that transient changes in the beam can be captured. The data output by each sensor is processed by a signal conditioning circuit to eliminate noise interference and converted into digital signals, forming a time series of beam spot distribution data. These data contain information on the intensity distribution of the beam in the transverse and longitudinal directions, reflecting the shape, size, and uniformity of the beam spot.

[0072] The injection process parameter set is recorded synchronously during data acquisition, including injection energy, injection dose, and scanning frequency. The injection energy is provided by a high-voltage power module, which is monitored and recorded in real time by a voltage sensor; the injection dose is measured by a beam current integrator, which accumulates the number of ions per unit area; and the scanning frequency is fed back by the control system, reflecting the moving speed of the beam on the wafer surface. These parameters are updated at a fixed period, aligned with the timestamp of the beam spot distribution data sequence, to ensure data consistency. The record of process parameters not only contains the current value, but also stores the historical trend for process stability analysis.

[0073] The statistical characteristic quantities of the beam spot distribution data sequence are used to generate the beam spot stability index. The statistical characteristics include beam spot mean, variance, root mean square error, skewness, and kurtosis. The mean reflects the overall intensity level of the beam, the variance and root mean square error describe the fluctuation degree of the beam spot, and the skewness and kurtosis represent the asymmetry and sharpness of the beam spot distribution. These characteristics are calculated through a sliding window, and the window size is dynamically adjusted according to process requirements, usually covering a range of seconds to tens of seconds of data. The beam spot stability index is converted into a quantitative value through normalization, with a range of 0 to 1, and the higher the value, the more stable the beam uniformity.

[0074] The uniformity drift prediction operation combines the injection process parameter set and the beam spot stability index to output the beam uniformity prediction parameter. The prediction process relies on historical data analysis, comparing the current data with similar conditions in the historical process database to infer the trend of beam uniformity. The historical process database stores the running data of the same type of ion implanter under different process conditions, including beam spot distribution sequence, process parameters, and corresponding uniformity deviation. The data in the database is cleaned and labeled to ensure the quality of the samples used for prediction.

[0075] The beam spot distribution data sequence is divided into real-time acquisition units and historical analysis units. The real-time acquisition unit relies on the distributed sensor array to capture the current beam cross-sectional density distribution data, forming the real-time beam spot distribution data sequence. The historical analysis unit extracts the historical beam spot distribution data sequence under the same working condition from the process database, including the beam fluctuation characteristics under the same injection energy, dose, and scanning frequency. Real-time data and historical data are integrated through a spatiotemporal feature fusion model, which is based on a convolutional neural network structure and can extract spatial correlation and temporal dependence features simultaneously. The model input includes real-time beam spot distribution sequence and historical beam spot distribution sequence, and the output is an enhanced beam spot distribution feature set, which is used for subsequent uniformity drift prediction.

[0076] The uniformity drift prediction adopts a linear regression algorithm to calculate the correlation between the current beam spot stability index and historical data. The algorithm inputs include current process parameters, beam spot statistical features, and enhanced beam spot distribution features, and outputs a beam current uniformity prediction parameter. This parameter is expressed in a probabilistic form, reflecting the likelihood of the beam current uniformity deviating from the theoretical value in the future period. The prediction result is smoothed by a moving average filter to reduce the influence of random fluctuations and improve the reliability of the prediction.

[0077] The beam current uniformity prediction parameter is used as input for subsequent beam spot distortion recognition to guide the selection and activation of the beam spot classifier. The numerical range of the prediction parameter matches the preset beam spot distortion level, and the most likely distortion type is selected. The matching process uses the Euclidean distance method to calculate the closeness of the prediction parameter to each distortion level reference value, and selects the K levels with the smallest deviation as candidates. The number of activated paths of the beam spot classifier is determined by the deviation amplitude, and the smaller the deviation, the more paths are activated to improve recognition accuracy.

[0078] The current beam spot morphology image captured by the beam spot imaging device is combined with the prediction parameter and input into multiple beam spot classifiers for parallel processing. The classifiers are based on residual networks, each path independently analyzes the image features, and outputs the discrimination result of the beam spot distortion type. The effective determination proportion of each classifier is calculated to generate a beam spot distortion level probability distribution, and the level with the highest probability is finally selected as the actual beam spot distortion parameter. This parameter is used for subsequent prediction and optimization of electromagnetic field adjustment parameters to achieve closed-loop control of beam current uniformity.

[0079] Embodiment 2: see Figure 3 , which involves constructing and executing a uniformity drift prediction model based on historical process data training to predict the trend of beam current uniformity. The implementation process includes historical data acquisition, model construction, training and verification, and generation of final prediction parameters.

[0080] The historical process database stores the running data of the same type of ion implanter under different process conditions. The data in the database is collected from long-term process operation, including beam spot distribution data sequence, implantation process parameters, and corresponding beam spot uniformity deviation. The beam spot distribution data sequence is collected by a beam probe array, recording the intensity distribution of the beam in the transmission path; the implantation process parameters include implantation energy, implantation dose and scanning frequency, which are recorded in real time by the control system; the beam spot uniformity deviation is calculated by the difference between the actual measurement value and the theoretical uniformity value, and the labeling method includes manual labeling and automatic labeling. Manual labeling is performed by the operator according to the process specification, and automatic labeling is completed by a preset algorithm to ensure the accuracy and consistency of the data. The storage format of the historical data uses a time series database, which supports efficient query and analysis.

[0081] The beam spot uniformity deviation sample set is extracted from the historical process database for training the uniformity drift prediction model. The construction process of the sample set includes data cleaning, feature extraction, and label alignment. Data cleaning removes outliers and noise interference, such as invalid data caused by sensor failure; feature extraction calculates the statistical features of the beam spot distribution sequence, including mean, variance, skewness, and kurtosis; label alignment ensures that the beam spot distribution data, process parameters, and uniformity deviation of each sample match in time. The sample set is divided in chronological order to preserve the time continuity of the data, facilitating the model to learn time-dependent features.

[0082] The uniformity drift prediction model is constructed based on time series decomposition methods, integrating autoregressive integrated moving average algorithms and seasonal decomposition techniques. The structure of the model includes the extraction and combination of trend components, seasonal components, and residual components. The trend component reflects the long-term change direction of beam uniformity, calculated by the moving average algorithm; the seasonal component captures the periodic fluctuations of process parameters, such as regular changes in scan frequency; the residual component represents random fluctuations that cannot be explained by trends and seasonality. The input layer of the model receives the beam spot distribution data sequence and injection process parameters, which are input into the time series decomposition module after feature fusion. Each component is processed separately, and finally combined in the output layer to generate beam uniformity prediction parameters.

[0083] The training process of the model uses a set of historical beam spot distribution data sequences, a set of historical injection process parameters, and a set of beam spot uniformity deviation samples. The training data is divided into a training set and a validation set, with a ratio of 7:3. The training set is used to optimize model parameters, and the validation set is used to evaluate the generalization ability of the model. The gradient descent method is used for training algorithm, and the mean square error is selected as the loss function to measure the difference between the predicted value and the true deviation. The early stopping mechanism is applied during training to terminate training when the validation set loss no longer decreases, preventing overfitting. The model parameters are initialized using a random strategy, and the learning rate is dynamically adjusted to balance convergence speed and stability.

[0084] The trained model is evaluated for its prediction accuracy in the validation phase. The validation process uses an independent validation data set to calculate the correlation between the predicted parameters and the actual uniformity deviation. The evaluation indicators include root mean square error, mean absolute error, and determination coefficient, reflecting the prediction ability of the model under different process conditions. The validation results are used to adjust the model structure or hyperparameters, such as increasing the number of decomposition components or adjusting the length of the seasonal period. The final version of the model must meet the preset accuracy threshold before it can be put into practical application.

[0085] The current injection process parameter set and the current beam spot stability index are input into the trained model to perform a uniformity drift prediction operation. The preprocessing of the current data includes missing value filling and standardization. The missing value filling uses linear interpolation, and the standardization maps parameters of different dimensions to a unified range, such as the interval [0, 1]. The model performs time series decomposition on the preprocessed data to extract trend, seasonal, and residual features, and finally outputs beam uniformity prediction parameters. The prediction parameters are in the form of probabilistic values, indicating the likelihood of the beam uniformity deviating from the theoretical value in the future period of time.

[0086] The application of the prediction parameters includes guiding the screening of the beam spot classifier and optimizing the electromagnetic field adjustment parameters. The beam spot classifier selects the most likely beam spot distortion level for classification according to the numerical range of the prediction parameters; the prediction of the electromagnetic field adjustment parameters combines the prediction parameters with the actual beam spot distortion parameters to generate more accurate adjustment instructions. The dynamic updating mechanism of the prediction model is realized by periodic retraining, which optimizes the model parameters using new process data to adapt to changes in process conditions.

[0087] The implementation of embodiment 2 realizes accurate prediction of beam uniformity through a historical data-driven modeling method. The advantage of this method is that it fully utilizes historical process information to capture the complex patterns of beam behavior and provides a reliable basis for dynamic adjustment. The construction and training process of the model focuses on the representativeness of the data and the adaptability of the algorithm to ensure the effectiveness of the prediction results in practical applications.

[0088] Embodiment 3: involves capturing the current beam spot morphology image through a beam spot imaging device, inputting multiple beam spot classifiers in parallel combining the beam uniformity prediction parameters to identify the actual beam spot distortion parameters. This process includes image acquisition, distortion level screening, classification path activation, parallel classification processing, and final distortion parameter determination, forming a complete beam spot distortion identification process.

[0089] The beam spot imaging device uses a high-resolution CCD or CMOS sensor and is configured at the end of the beam transmission path of the ion implanter. The device captures beam spot morphology images of the target area at a fixed frequency, with an image resolution of no less than 1024x1024 pixels and a gray level of 12 bits. Optical filtering technology is used during imaging to eliminate environmental light interference, and an electronic shutter is used to control the exposure time to ensure accurate representation of the beam intensity within the dynamic range. The captured current beam spot morphology image is preprocessed, including non-uniformity correction, bad pixel repair, and flat field correction, to eliminate the response differences of the sensor itself and the distortions introduced by the optical system.

[0090] The beam uniformity prediction parameters serve as prior information to guide the subsequent screening of the beam spot distortion levels. The numerical range of the prediction parameters is matched with the pre-set K beam spot distortion levels, and the matching process uses the following distance measurement method:

[0091]

[0092] wherein, represents the deviation distance of the i-th distortion level from the prediction parameter, P is the value of the beam uniformity prediction parameter, is the reference value of the i-th distortion level. The value of K is usually set to 5, covering typical distortion levels from slight to severe. The reference value is obtained by historical data analysis, representing the characteristic central value of each type of distortion. After calculating the deviation distance of each level, the distortion levels corresponding to the K smallest distances in ascending order are selected as the candidate set.

[0093] The determination of the number of activated paths is directly related to the deviation distance. The number of activated paths is calculated according to the following rule: when is less than the threshold value T, the maximum value is taken; when is greater than or equal to T, it is reduced in linear proportion. Wherein T is the preset distance threshold, is the maximum number of paths for a single classifier. This dynamic activation mechanism can not only ensure sufficient recognition of high matching degree levels, but also avoid excessive calculation of low matching degree levels. Each beam spot classifier contains multiple independent classification paths based on residual network, with a configurable number of paths between 10-20. The residual network adopts a 50-layer deep structure, containing multiple residual blocks and skip connections, which can effectively extract multi-scale features of beam spot images.

[0094] The training data of the beam spot classifier comes from the labeled beam spot image samples in historical processes. The sample collection covers various process conditions and beam states to ensure the generalization ability of the classifier. During training, data augmentation techniques are used, including random rotation, translation, and brightness adjustment, to improve the robustness of the model to image changes. The loss function uses cross-entropy loss, and the optimization algorithm uses stochastic gradient descent with momentum. The trained classifier is evaluated on an independent validation set, and the classification accuracy must meet the preset standard before it can be put into use.

[0095] The current beam spot morphology image is input into the activated paths of the K selected classifiers in parallel for analysis. Each activated path runs independently and outputs the judgment result of the image belonging to the distortion level of that path. The classification process first normalizes the input image, then extracts features through the residual network, and finally obtains the class probability through the fully connected layer and the softmax function. Due to the different parameter initialization of each path and the differential enhancement of training data, even different paths of the same classifier may produce different outputs. This design increases the diversity of classification.

[0096] The statistical analysis of the K-group sclereogram classification results uses a majority voting mechanism. For each distortion level, the proportion of classes correctly classified across different classification paths is calculated. Path weights are considered in the proportion calculation, and these weights are positively correlated with the accuracy of the path on the validation set. Levels with a valid classification proportion exceeding 50% are considered candidate results. When multiple levels meet the criteria, the level with the highest proportion is selected; when no level meets the criteria, a review mechanism is triggered, increasing the number of activated paths for reclassification.

[0097] The final determination of actual beam spot distortion parameters also requires time-series consistency testing. The system maintains a distortion level cache queue of length L, recording the intermediate results of the most recent L identifications. When the current identification result is inconsistent with most historical results in the queue, time-domain correlation analysis is initiated to examine whether the evolution trend of the beam spot morphology supports the current judgment. This spatiotemporal joint verification mechanism can effectively suppress random errors in single identifications and improve the reliability of the parameters. The determined actual beam spot distortion parameters not only include the level category but also include a confidence score, reflecting the degree of certainty of the judgment.

[0098] The online update mechanism for the classifier group is implemented through incremental learning. New speckle image samples are collected periodically, and the model is incrementally trained to adapt to changes in process conditions while maintaining the original classification capabilities. The update process employs an elastic weight solidification algorithm to balance the retention of old and new knowledge. Simultaneously, a classification performance monitoring mechanism is established. When the recognition accuracy consistently falls below a threshold, a model reconstruction process is triggered, retraining the entire classifier group.

[0099] Accurate identification of beamsaw distortion parameters is achieved through parallel processing of multiple classifiers and a dynamic path activation mechanism. Image acquisition and preprocessing ensure the quality of input data, level-based filtering based on predicted parameters improves recognition efficiency, residual network classification paths provide powerful feature extraction capabilities, and majority voting and spatiotemporal verification ensure the reliability of results. The entire process design fully considers various influencing factors in the actual process environment and can operate stably under complex conditions. The online update mechanism of the classifier ensures long-term adaptability, enabling the system to continuously optimize recognition performance by tracking changes in process conditions.

[0100] Example 4: Reference Figure 4 Based on the actual beam distortion parameters, predicted electromagnetic field adjustment parameters are generated, and these parameters are optimized through a feedback mechanism. This process comprises four main stages: electromagnetic field parameter prediction, adjustment implementation, feedback acquisition, and parameter optimization, forming a closed-loop control system. The implementation details of each stage are illustrated below with specific examples.

[0101] The electromagnetic field parameter prediction operation is based on the correction records in the historical process database. The database stores successful electromagnetic field adjustment schemes corresponding to different beam spot distortion parameters, which are organized in a structured manner for quick retrieval and analysis. When the system receives the actual beam spot distortion parameters, it first queries similar cases in the database. The similarity judgment considers factors such as distortion type, degree, and current process conditions. The query results are sorted by matching degree, and the top few most similar cases are selected as references. The electromagnetic field parameter predictor analyzes the adjustment parameters of these reference cases, combines the current beam state, and generates predicted electromagnetic field adjustment parameters. The prediction process considers the physical characteristics of the electromagnetic field, including magnetic field strength, electric field direction, and their combined effects. The predicted parameters are represented in vector form, containing suggested values for each adjustment dimension.

[0102] Table 1: Shows some beam spot distortion cases in the historical database and their corresponding electromagnetic field adjustment parameters.

[0103] Case Number Type of Distortion Degree of Distortion Magnetic Field Strength (T) Electric Field Angle (°) Adjustment Effect Rating C-1024 Edge Spread Moderate 0.85 12.5 A C-1025 Central Depression Mild 0.72 8.2 B+ C-1026 Asymmetry Severe 1.05 15.8 A- C-1027 Multi-Pole Distortion Moderate 0.91 10.3 B C-1028 Edge Spread Mild 0.68 7.5 A

[0104] In a specific example, the currently detected actual beam spot distortion parameter is "edge spread - moderate", which has similar characteristics with cases C-1024 and C-1028. The predictor analyzes the adjustment parameters of these two cases, considers that the current beam energy is higher than case C-1024 but lower than C-1028, and finally generates the predicted electromagnetic field adjustment parameters as magnetic field strength 0.82T and electric field angle 11.2°. This predicted value is between the two reference cases, while considering the differences in current process conditions.

[0105] After the predicted electromagnetic field adjustment parameters are generated, the control system converts them into specific device instructions. Magnetic field strength adjustment is achieved by changing the electromagnet excitation current, and the control system calculates the corresponding current set value according to the pre-set current-magnetic field strength conversion curve. Electric field angle adjustment is completed by adjusting the voltage distribution of the electrostatic deflection plate, and the voltage distribution scheme is based on the electric field simulation calculation results. These instructions are sent to the respective actuators through a digital communication interface, and the execution process uses closed-loop control to monitor the actual achieved electromagnetic field parameters in real time, ensuring consistency with the set values.

[0106] The beam spot imaging device collects feedback images after the electromagnetic field adjustment is completed. The collection timing is selected after the adjustment is stable, usually waiting for 3-5 beam scanning periods. The collection conditions of the feedback images remain consistent with the original images, including exposure time, gain, and other parameter settings. The collected feedback images go through the same preprocessing process as the original images, including noise removal, non-uniformity correction, etc. The preprocessed images are input into the beam spot analysis module to calculate various characteristic parameters of the current beam spot, including shape factor, symmetry index, and uniformity coefficient, etc. These characteristic parameters are compared with the baseline values before adjustment to evaluate the adjustment effect.

[0107] The electromagnetic field parameter classifier group is used to analyze the feedback images and optimize the adjustment parameters. The classifier group contains multiple specially trained neural network models, each of which is optimized for a specific range of electromagnetic field parameters. In the example, the predicted parameters are 0.82T and 11.2°, and the system selects the classifiers corresponding to the adjacent 3 parameter levels for analysis. The classifier inputs the feedback image and the current electromagnetic field parameters, and outputs the adjustment effect score and the recommended adjustment direction. The score is based on the closeness of the image features to the ideal beam spot, and the adjustment direction indicates the trend of increasing or decreasing each parameter. The output results of multiple classifiers are weighted and fused, and the weights depend on the accuracy of each classifier in the verification phase.

[0108] The parameter optimization process adopts an iterative adjustment strategy. According to the comprehensive recommendations of the classifier group, the system generates new electromagnetic field parameter candidate values. The generation of candidate values takes into account the adjustment step limit to avoid excessive single adjustment. The new candidate parameters are implemented again and feedback is collected, and this cycle continues until the termination condition is met. The termination conditions include reaching the preset beam spot quality indicators, exceeding the maximum number of iterations, or the adjustment effect no longer significantly improving, etc. In the example, after two iterations of adjustment, the final actual electromagnetic field adjustment parameters are determined as magnetic field strength 0.83T and electric field angle 10.8°. This result, compared with the initial predicted value, shows a slight increase in magnetic field strength and a slight decrease in electric field angle, reflecting the system's fine adjustment to the actual beam response.

[0109] The training data of the electromagnetic field parameter predictor comes from historical successful adjustment cases. The training samples contain beam spot distortion features, process conditions and corresponding optimal electromagnetic field adjustment parameters. The model training uses a supervised learning method, with the input being the beam spot features and process parameters, and the output being the electromagnetic field adjustment parameter recommendation. Regularization techniques are applied during training to prevent overfitting, and cross-validation is used to evaluate model performance. The model is regularly incrementally trained using new successful cases to maintain the timeliness of the prediction ability.

[0110] The design of the feedback adjustment mechanism takes into account the response characteristics of the beam system. Different electromagnetic field parameters have different degrees and speeds of influence on the beam spot shape, and the system establishes a response characteristic knowledge base. In parameter adjustment, parameters with fast response are adjusted first, and parameters with slow response are adjusted subsequently. This phased adjustment strategy improves the optimization efficiency and avoids the mutual interference between parameters. At the same time, the system records the parameter combination and its effect of each adjustment, enriching the content of the historical database and providing more references for subsequent prediction.

[0111] The dynamic determination of electromagnetic field adjustment parameters is achieved through a closed-loop process of prediction-execution-feedback-optimization. The effective use of historical data improves the accuracy of prediction, and the parallel analysis of multiple classifiers ensures the reliability of parameter optimization. The iterative adjustment strategy enables the final parameters to accurately match the actual requirements. The entire process fully considers the complexity of beam control in ion implantation processes and can generate appropriate adjustment schemes under various working conditions. The self-learning ability of the system enables it to continuously improve and maintain excellent regulation performance in the long term.

[0112] Example 5: Invocation of the electromagnetic field parameter classifier group and generation of the beam uniformity dynamic adjustment instruction set. This process is based on the actual beam spot distortion parameters and actual electromagnetic field adjustment parameters obtained in the early stage. Through systematic conversion and fusion operations, executable device control instructions are finally formed.

[0113] The invocation process of the electromagnetic field parameter classifier group begins with the system loading the pre-trained classifier library. This classifier library is stored in a dedicated non-volatile memory and contains classification models trained for different electromagnetic field parameter levels. Each classification model corresponds to a specific electromagnetic field parameter interval, which covers various possible combinations within the normal working range of the device. The training data for the classifier comes from beam spot image samples collected during historical processes. These samples have been strictly screened and labeled to ensure coverage of various typical beam states. Sample labeling not only includes electromagnetic field parameter level information but also records corresponding process conditions and device state parameters. The network structure of the classifier uses a deep convolutional neural network, which includes multiple convolutional and pooling layers, and can effectively extract local and global features of beam spot images. Data augmentation techniques, including random rotation, scaling, and brightness adjustment, are used during the training process to improve the model's generalization ability.

[0114] The attention mechanism design of the classifier is its core feature. During the image processing stage, the network automatically learns to focus on key areas of the beam spot, such as edge distortion areas or intensity anomaly areas. This attention allocation is dynamic and automatically adjusts the focus according to the characteristics of the input image. The design of multiple classification paths increases the robustness of the system. Different paths may focus on different features of the image, and the final judgment results of each path are integrated through ensemble learning. The classifier undergoes strict verification tests before being put into use, using independent test data sets to evaluate its classification accuracy and generalization performance. The verification process simulates various scenarios in the actual working environment, including different beam intensities, energy levels, and scanning modes.

[0115] The mapping of actual beam spot distortion parameters to beam current correction quantities is achieved through preset conversion rules. These rules are based on a deep understanding of the physical properties of the beam and take into account the influence mechanisms of different distortion types on beam transmission. The conversion process not only considers the distortion level but also combines the current process parameters and device state. For example, for edge spreading type distortion, the correction quantity calculation considers the relationship between beam energy and beam spot size; for center depression type distortion, it combines the current settings of beam intensity and focusing system. The conversion result is represented in digital form and can be directly input into the beam adjustment unit for execution.

[0116] The conversion of actual electromagnetic field adjustment parameters to magnetic field intensity adjustment quantities and electric field angle adjustment quantities follows the device control protocol. The conversion process takes into account the physical coupling effects of electromagnetic field parameters, i.e., the mutual influence relationship between magnetic field and electric field. The system maintains a parameter coupling matrix that describes the comprehensive effects under different parameter combinations. The conversion algorithm decouples the original parameters based on this matrix to obtain independent magnetic field and electric field adjustment quantities. These adjustment quantities are further converted into device-specific control instructions, such as electromagnet current set values or deflection plate voltage distribution schemes. Range checks are implemented during the conversion process to ensure that the generated instructions are within the safe working range of the device.

[0117] The fusion of beam current correction quantities, magnetic field intensity adjustment quantities, and electric field angle adjustment quantities uses a weighted synthesis strategy. The weight distribution is based on the expected effect of each adjustment method under the current process conditions, which comes from the analysis of historical adjustment records. The fusion algorithm considers the timing characteristics of the adjustment and reasonably arranges the execution order and time interval of different adjustment actions. For example, electric field adjustment usually responds faster than magnetic field adjustment, so it may be executed first. The fusion process also includes consistency checks to ensure that there is no conflict or contradiction between the adjustment instructions. The final dynamic adjustment instruction set is encoded in a standardized format, including instruction type, parameter value, execution timing, etc.

[0118] The execution of the dynamic adjustment instruction set uses a phased verification mechanism. The system first verifies the logical reasonableness of the instructions in simulation mode, checking whether the parameters of each instruction are within the allowed range. Then it actually executes some instructions in a small range test to observe the response of the beam. Finally, it fully implements all adjustment instructions, and the entire process is monitored in real time, and safety protection programs are started immediately upon detection of abnormalities. During instruction execution, the system continuously collects beam feedback data and compares it with the expected effect, making dynamic adjustments as necessary.

[0119] The adaptive learning function of the system enables continuous optimization of the regulation strategy. Each successful regulation case is recorded and analyzed, extracting valuable experience for improving the subsequent regulation process. The learning algorithm focuses on the difference between the regulation effect and the initial prediction, gradually correcting the parameter conversion rules and fusion strategy. This learning is carried out progressively under the premise of ensuring system stability, avoiding sudden changes affecting process consistency. The system also implements regular performance evaluations to check whether the learning process brings the expected improvement effect.

[0120] Through the carefully designed classifier invocation and instruction generation process, efficient conversion from diagnostic parameters to execution instructions is achieved. The attention mechanism and multi-path design of the electromagnetic field parameter classifier provide reliable parameter recognition capability, the parameter conversion rules are based on physical principles to ensure the rationality of regulation, and the instruction fusion strategy considers the response characteristics of actual equipment, and the final generated dynamic regulation instruction set can accurately correct the beam uniformity. The design of the entire process fully considers various constraint conditions in the actual industrial environment, ensuring process stability and equipment safety while ensuring regulation effect. The learning ability of the system enables it to adapt to various changes in long-term use, continuously maintaining high-level regulation performance.

[0121] It should be noted that, in the present document, relational terms such as first and second and the like can be used solely to distinguish one entity or action from another entity or action without necessarily requiring or implying any actual such relationship or order between such entities or actions. Moreover, the terms "comprises", "comprising", or any other variation thereof, are intended to cover a non-exclusive inclusion, such that a process, method, article, or apparatus that comprises a list of elements does not include only those elements but can include other elements not expressly listed or inherent to such process, method, article, or apparatus.

[0122] While embodiments of the present application have been shown and described with reference to particular embodiments thereof, it will be understood by those skilled in the art that various changes in form and details can be made therein without departing from the spirit and scope of the application. The scope of the application is defined by the appended claims and their equivalents.

Claims

1. A method for dynamically adjusting the uniformity of a large beam current in an ion implanter, characterized in that, The method includes: Real-time acquisition of beam spot distribution data sequence in the beam transmission path of the ion implanter, simultaneous acquisition of implantation process parameter set, and prediction of uniformity drift based on beam spot distribution data sequence to generate beam uniformity prediction parameters; The beam spot morphology image is captured by the beam spot imaging device, and combined with the beam uniformity prediction parameters, the current beam spot morphology image is input into multiple beam spot classifiers in parallel to identify and obtain the actual beam spot distortion parameters. Each beam spot classifier contains multiple independently running classification paths. Based on the actual beam spot distortion parameters, an electromagnetic field parameter prediction operation is performed to generate predicted electromagnetic field adjustment parameters. A beam spot feedback image after the electromagnetic field is applied is acquired using a beam spot imaging device. Based on the predicted electromagnetic field adjustment parameters, the beam spot feedback image is input in parallel into multiple electromagnetic field parameter classifiers, and the actual electromagnetic field adjustment parameters are output. By integrating the actual beam spot distortion parameters with the actual electromagnetic field adjustment parameters, a dynamic adjustment command set for beam uniformity is generated.

2. The method for dynamically adjusting the uniformity of the large beam current in an ion implanter according to claim 1, characterized in that, The real-time acquisition of beam spot distribution data sequences in the beam transmission path of the ion implanter, simultaneous acquisition of a set of implantation process parameters, and prediction of uniformity drift based on the beam spot distribution data sequences generate beam uniformity prediction parameters, including: During the continuous operation of the injection process, beam spot distribution data sequences are continuously acquired through a beam detector array; A set of injection process parameters consisting of injection energy, injection dose, and scanning frequency is periodically recorded; Statistical features of the beam spot distribution data sequence are extracted to generate a beam spot stability index; By combining the injection process parameter set with the beam spot stability index, a uniformity drift prediction operation is performed to output beam uniformity prediction parameters.

3. The method for dynamically adjusting the uniformity of the large beam current in an ion implanter according to claim 2, characterized in that, The step of combining the injection process parameter set with the beam spot stability index to perform a uniformity drift prediction operation to output beam uniformity prediction parameters includes: Obtain the historical process database of the same model of ion implanter, extract the historical beam spot distribution data sequence set and the historical implantation process parameter set, and mark the beam spot uniformity deviation at the corresponding historical process time to form a beam spot uniformity deviation sample set. Construct a beam uniformity prediction model based on time series decomposition; The historical beam spot distribution data sequence set, historical injection process parameter set, and beam spot uniformity deviation sample set are used to supervise the training and verification of the beam uniformity prediction model. Input the current set of injection process parameters and the current beam spot stability index into the trained beam uniformity prediction model, and output the beam uniformity prediction parameters.

4. The method for dynamically adjusting the uniformity of the large beam current in an ion implanter according to claim 1, characterized in that, The process involves capturing a current beam spot morphology image using a beam spot imaging device, combining it with the beam uniformity prediction parameters, and inputting the current beam spot morphology image in parallel into multiple beam spot classifiers to identify and obtain actual beam spot distortion parameters, including: Acquire the current beam spot morphology image of the target beam spot region using a beam spot imaging device; Based on the numerical range of the beam uniformity prediction parameters, the K beam spot distortion levels with the highest matching degree are selected, where K is a preset positive integer; Calculate the deviation magnitude between the K beam spot distortion levels and the beam uniformity prediction parameters, and determine the number of classification path activations for each beam spot classifier accordingly; Select the K fascicle classifiers corresponding to the K fascicle distortion levels, input the current fascicle morphology image into the activation classification path of each fascicle classifier, and generate K sets of fascicle classification results. Calculate the proportion of valid judgment results in the K sets of beam spot classification results, and output the probability values ​​of K beam spot distortion levels; The beam spot distortion level with the highest probability value is selected as the actual beam spot distortion parameter.

5. The method for dynamically adjusting the uniformity of the large beam current in an ion implanter according to claim 4, characterized in that, The calculation of the deviation magnitude between the K beam spot distortion levels and the beam uniformity prediction parameters, and the determination of the number of classification path activations for each beam spot classifier based on this deviation, includes: Calculate the deviation magnitude between the K beam spot distortion levels and the beam uniformity prediction parameters respectively; The deviation magnitude value is converted into the number of classification path activations using a preset coefficient conversion rule; Activate a pre-trained set of fascicle classifiers, where each fascicle classifier corresponds to a specific fascicle distortion level and contains multiple classification paths constructed based on residual networks; Based on the number of activated classification paths, a corresponding number of classification paths are randomly selected from each speckle classifier to perform image analysis.

6. The method for dynamically adjusting the uniformity of large beam current in an ion implanter according to claim 1, characterized in that, The step of performing electromagnetic field parameter prediction based on the actual beam spot distortion parameters to generate predicted electromagnetic field adjustment parameters includes: Obtain electromagnetic field correction records corresponding to different beam spot distortion parameters from the historical process database; Extract historical electromagnetic field modulation parameter sample sets and beam spot distortion parameter sample sets; Construct an electromagnetic field parameter predictor based on a physical model correction; The electromagnetic field parameter predictor is trained using the beam spot distortion parameter sample set and the historical electromagnetic field adjustment parameter sample set. The actual beam distortion parameters are input into the trained electromagnetic field parameter predictor, which outputs the predicted electromagnetic field adjustment parameters.

7. The method for dynamically adjusting the uniformity of large beam current in an ion implanter according to claim 1, characterized in that, The beam spot imaging device acquires beam spot feedback images after the electromagnetic field is applied. Based on the predicted electromagnetic field adjustment parameters, the beam spot feedback images are input in parallel into multiple electromagnetic field parameter classifiers to output the actual electromagnetic field adjustment parameters, including: After applying the predicted electromagnetic field conditioning parameters, beam spot feedback images of the beam transmission path are acquired; Select M electromagnetic field parameter levels adjacent to the predicted electromagnetic field adjustment parameters; Call the electromagnetic field parameter classifier group corresponding to M electromagnetic field parameter levels; The beam spot feedback image is input into all the classification paths of each electromagnetic field parameter classifier to generate M sets of electromagnetic field classification results. Calculate the percentage of valid judgment results for each electromagnetic field classification result set, and output the probability distribution of M electromagnetic field levels. The electromagnetic field parameter level with the highest probability distribution is selected as the actual electromagnetic field adjustment parameter.

8. The method for dynamically adjusting the uniformity of large beam current in an ion implanter according to claim 7, characterized in that, The invocation of the electromagnetic field parameter classifier group corresponding to the M electromagnetic field parameter levels includes: Load a pre-trained electromagnetic field parameter classifier library, where each classifier corresponds to a specific electromagnetic field parameter level and contains multiple classification paths based on attention mechanisms; The training data for the classification path includes sample beam pattern images and their corresponding electromagnetic field parameter level determination labels.

9. The method for dynamically adjusting the large beam current uniformity of an ion implanter according to claim 1, characterized in that, The process of integrating the actual beam spot distortion parameters and the actual electromagnetic field adjustment parameters to generate a dynamic beam uniformity adjustment command set includes: Map the actual beam spot distortion parameters to beam correction values; The actual electromagnetic field adjustment parameters are converted into magnetic field strength adjustment and electric field angle adjustment. The beam correction amount, magnetic field strength adjustment amount, and electric field angle adjustment amount are combined to generate a dynamic adjustment command set.

10. The method for dynamically adjusting the large beam current uniformity of an ion implanter according to claim 1, characterized in that, The beam spot distribution data sequence is divided into a real-time acquisition unit and a historical analysis unit, including: During continuous injection, beam cross-sectional density distribution data is captured by a distributed sensor array to form a real-time beam spot distribution data sequence. Extract beam fluctuation characteristic data under the same operating conditions from the historical process database to form a historical beam spot distribution data sequence; The real-time beam spot distribution data sequence and the historical beam spot distribution data sequence are input into the spatiotemporal feature fusion model, and the output enhanced beam spot distribution feature set is used for subsequent uniformity drift prediction.

Images