Fusion control method, device and electronic equipment for shoulder placement growth process

Abstract

The application discloses a fusion control method and device for a shoulder growth process and electronic equipment. The method comprises the following steps: acquiring image sequence data and process sequence data; predicting a first pulling speed response value based on the image sequence data and the process sequence data through a first mapping model, wherein the first mapping model comprises a pre-trained residual neural network; predicting a second pulling speed response value based on the process sequence data through a second mapping model, wherein the second mapping model comprises a physical neural network; and outputting a pulling speed control response value according to the first pulling speed response value and the second pulling speed response value through a fusion control module, and controlling the silicon single crystal shoulder growth process according to the pulling speed control response value. The above technical solution uses the mapping model to predict the pulling speed response value according to the time sequence data of the image and the process respectively, and combines the two pulling speed response values to comprehensively determine the pulling speed control response value used for controlling the shoulder growth process, thereby improving the reliability of the automatic control of the shoulder growth process.

Description

Technical Field

[0001] This application relates to the field of single crystal process control technology, and in particular to a fusion control method, apparatus and electronic device for a shoulder growth process. Background Technology

[0002] Shoulder formation refers to the process in the early stages of single-crystal silicon growth where the diameter of the silicon crystal gradually increases from the seed crystal to the target diameter by controlling temperature and pulling speed. Existing methods for controlling the pulling speed during the shoulder formation process in single-crystal silicon growth include manual shoulder formation, proportional-integral-derivative (PID) speed control based on time-series data, fuzzy rule control based on image edge width, and model predictive control. Manual shoulder formation offers high flexibility and adaptability, but it suffers from high labor costs, poor operational consistency, and heavy reliance on the stable operation of individual experts, making it incompatible with large-scale long-term production. Furthermore, the discrepancies amplify over long-term operation, potentially leading to breakage. PID speed control based on time-series data only adjusts the pulling speed based on the deviation between the actual and set diameter values. This is unsuitable for the nonlinearity and large time delay issues in single-crystal silicon growth, leading to overshoot and undershoot, and ultimately, breakage during the shoulder formation process. Fuzzy rule control, due to the complexity of the silicon single crystal growth process and the issues of various product specifications and rapid technological iteration, requires extensive rule adjustments and optimizations under different application requirements. This necessitates numerous practical experiments, leading to high experimental costs and long experimental cycles. Model predictive control suffers from a problem: its control effect is highly dependent on model accuracy. However, silicon single crystal growth is a batch production process, and the model's accuracy struggles to adapt to equipment variability and product diversity. Furthermore, model building requires substantial data support, significantly limiting the application of model predictive control.

[0003] In summary, the current method for controlling the growth rate during the shoulder development process has limited applicability and is difficult to achieve reliable automated control. Summary of the Invention

[0004] This application provides a fusion control method, apparatus, and electronic device for the shoulder growth process, so as to achieve reliable automated control of the shoulder growth process.

[0005] In a first aspect, embodiments of this application provide a fusion control method for the shoulder growth process, comprising:

[0006] Acquire image sequence data and process sequence data of the shoulder formation process inside the single crystal furnace;

[0007] The first pulling speed response value is predicted based on image sequence data and process sequence data through the first mapping model, which includes a pre-trained residual neural network.

[0008] The second pulling speed response value is predicted based on the process sequence data through the second mapping model, which includes a physical neural network.

[0009] The integrated control module outputs a pulling speed control response value based on the first pulling speed response value and the second pulling speed response value, and controls the silicon single crystal shoulder growth process based on the pulling speed control response value.

[0010] Secondly, embodiments of this application also provide a fusion control device for the shoulder growth process, comprising:

[0011] The data acquisition module is used to acquire image sequence data inside the single crystal furnace and process sequence data of the shoulder formation process;

[0012] The first prediction module is used to predict the first pulling speed response value based on image sequence data and process sequence data through a first mapping model. The first mapping model includes a pre-trained residual neural network.

[0013] The second prediction module is used to predict the second pull speed response value based on the process sequence data through the second mapping model. The second mapping model includes a physical neural network.

[0014] The fusion control module is used to output a pulling speed control response value based on the first pulling speed response value and the second pulling speed response value, and to control the silicon single crystal shoulder growth process based on the pulling speed control response value.

[0015] Thirdly, embodiments of this application provide an electronic device, including:

[0016] One or more processors;

[0017] Storage device for storing one or more programs;

[0018] When one or more programs are executed by one or more processors, the one or more processors implement the fusion control method of the shoulder growth process as described in the first aspect.

[0019] Fourthly, embodiments of this application also provide a computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, implements the fusion control method for the shoulder growth process as described in the first aspect.

[0020] Fifthly, embodiments of this application also provide a computer program product, including a computer program and / or instructions, which, when executed by a processor, implement the fusion control method for the shoulder growth process as described in any of the above embodiments.

[0021] This application provides a fusion control method, apparatus, and electronic device for a shoulder growth process. The fusion control method includes: acquiring image sequence data within a single crystal furnace and process sequence data of the shoulder growth process; predicting a first pulling speed response value based on the image sequence data and process sequence data using a first mapping model, the first mapping model including a pre-trained residual neural network; predicting a second pulling speed response value based on the process sequence data using a second mapping model, the second mapping model including a physical neural network; and outputting a pulling speed control response value based on the first and second pulling speed response values ​​using a fusion control module, and controlling the silicon single crystal shoulder growth process based on the pulling speed control response value. The above technical solution utilizes mapping models to predict pulling speed response values ​​based on the time-series data of the image and process data respectively, and combines the two pulling speed response values ​​to comprehensively determine the final pulling speed control response value used to control the silicon single crystal shoulder growth process, thereby improving the reliability of automated control of the shoulder growth process. Attached Figure Description

[0022] The above and other features, advantages, and aspects of the embodiments of this disclosure will become more apparent from the accompanying drawings and the following detailed description. Throughout the drawings, the same or similar reference numerals denote the same or similar elements. It should be understood that the drawings are schematic, and the originals and elements are not necessarily drawn to scale.

[0023] Figure 1 A flowchart illustrating a fusion control method for a shoulder growth process provided in an embodiment of this application;

[0024] Figure 2 A schematic diagram of a fusion control process provided in an embodiment of this application.

[0025] Figure 3 A schematic diagram illustrating the output of a pulling speed control response value based on a first pulling speed response value and a second pulling speed response value, provided in an embodiment of this application.

[0026] Figure 4 A schematic diagram of a tension speed control curve for the shoulder-laying process under standard conditions, provided for an embodiment of this application;

[0027] Figure 5 A shoulder shape image at the end of shoulder placement is provided in an embodiment of this application;

[0028] Figure 6 A schematic diagram of the pulling speed control curve during the initial overheating process of shoulder placement, provided for an embodiment of this application;

[0029] Figure 7 Another shoulder shape image at the end of the shoulder placement is provided in the embodiments of this application;

[0030] Figure 8A schematic diagram of the structure of a fusion control device for a shoulder growth process provided in an embodiment of this application;

[0031] Figure 9 This is a schematic diagram of the structure of an electronic device provided in an embodiment of this application. Detailed Implementation

[0032] The present application will now be described in further detail with reference to the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are merely illustrative of the application and not intended to limit it. Furthermore, it should be noted that, for ease of description, the accompanying drawings show only the parts relevant to the present application, not the entire structure.

[0033] Before discussing the exemplary embodiments in more detail, it should be noted that some exemplary embodiments are described as processes or methods depicted as flowcharts. Although the flowcharts describe the steps as sequential processes, many of these steps can be performed in parallel, concurrently, or simultaneously. Furthermore, the order of the steps can be rearranged. A process can be terminated when its operation is complete, but it may also have additional steps not included in the figures. A process can correspond to a method, function, procedure, subroutine, subroutine, etc.

[0034] It should be noted that the concepts of "first" and "second" mentioned in the embodiments of this application are only used to distinguish different devices, modules, units or other objects, and are not used to limit the order or interdependence of the functions performed by these devices, modules, units or other objects.

[0035] Furthermore, the embodiments and features described in this application may be combined with each other, unless otherwise specified.

[0036] The acquisition, storage, use, and processing of data in this application all comply with the relevant provisions of national laws and regulations.

[0037] It should be noted that in the embodiments of this application, certain software, components, models and other existing solutions in the industry may be mentioned. These should be regarded as exemplary and are only intended to illustrate the feasibility of implementing the technical solution of this application. However, it does not mean that the applicant has used or necessarily used the relevant content of the solution.

[0038] Figure 1This is a flowchart illustrating a fusion control method for a shoulder-growing process provided in this application embodiment. This embodiment is applicable to situations involving real-time fusion control of video data during the shoulder-growing process. Specifically, this fusion control method for the shoulder-growing process can be executed by a fusion control device for the shoulder-growing process. This fusion control device can be implemented through software and / or hardware and integrated into an electronic device. The electronic device includes, but is not limited to, devices with computing capabilities such as computers, smartphones, or servers, and can also be a Central Processing Unit (CPU), System-on-Chips (SoC) computer, Field-Programmable Gate Array (FPGA), or Micro Controller Unit (MCU), etc.

[0039] like Figure 1 As shown, the method specifically includes the following steps:

[0040] S110. Acquire image sequence data and process sequence data of the single crystal furnace and the shoulder formation process.

[0041] In this embodiment, multiple image data (such as crystal shape images, liquid surface images, hot screen and / or reflection images, etc.) of the single crystal furnace over a period of time can be acquired using a charge-coupled device (CCD) based camera. These image data can be arranged into image sequence data according to time sequence. In addition, process parameters monitored during the shoulder growth process (such as power, crucible rotation, crystal rotation, pulling speed, crystal growth rate and / or pressure, etc.) can be acquired using sensors. These process parameters can be arranged into process sequence data according to time sequence. The image sequence data and process sequence data have a unified time label (or timestamp). The acquired time sequence data can cover various product specifications, process changes, and equipment changes. By acquiring image sequence data and process sequence data, multimodal sensing of the silicon single crystal shoulder growth process can be achieved, thereby providing a comprehensive and sufficient basis for pulling speed control.

[0042] S120. The first pulling speed response value is predicted based on image sequence data and process sequence data through the first mapping model. The first mapping model includes a pre-trained residual neural network.

[0043] For example, the pulling speed response value can be understood as the response value of the crystal growth rate during the crystal pulling process in a single crystal furnace. For image sequence data, a mapping model based on a network structure including a residual neural network can be constructed, namely the first mapping model. This first mapping model, after pre-training, can learn the mapping relationship between the input (mainly image sequence data) and the output (predicted pulling speed response value). The residual neural network can be used to extract features from both the image sequence data and the process sequence data. By inputting the acquired image sequence data and process sequence data into the first mapping model, the predicted first pulling speed response value can be obtained. It is understood that the temporal order of image sequence data and process sequence data is generally the same, but the dimension of image features is much higher than that of process features; therefore, the first mapping model is a pulling speed response mapping primarily based on image sequence data.

[0044] Understandably, due to the time lag in the silicon single crystal growth process, the optimal control response value at the next moment is determined by the multimodal data within past time steps. This process can be understood as a rolling calculation, where the input image of the first mapping model is a sequence of images in the time domain that is continuously updated over time. For example, the input image for a single step can refer to a continuous temporal image sequence within the time domain [NP, N], where P is the input time interval. During the rolling calculation, for time N+1, the input image for a single step refers to a continuous temporal image sequence within the time domain [N-P+1, N+1], discarding the images from the NP steps.

[0045] S130. The second pulling speed response value is predicted based on the process sequence data through the second mapping model, which includes a physical neural network.

[0046] For example, for process sequence data, a mapping model based on a neural network of physical control laws can be constructed, namely the second mapping model. The second mapping model can learn the mapping relationship between the input (process sequence data) and the output (predicted pulling speed response value) after pre-training.

[0047] It is understandable that the optimal pulling speed response of crystal growth depends not only on the input of visual image data, but also on the temporal changes of process parameters. In this embodiment, the relationship between the process timing at each time step and the pulling speed response is also considered during the rolling optimization calculation, thereby solving for the optimal pulling speed response value.

[0048] S140. Through the fusion control module, output the pulling speed control response value according to the first pulling speed response value and the second pulling speed response value, and control the silicon single crystal shoulder growth process according to the pulling speed control response value.

[0049] In this embodiment, a first pulling speed response value is obtained based on image sequence data, and a second pulling speed response value is obtained based on process sequence data. The fusion control module can be used to fuse these two predicted pulling speed response values ​​to comprehensively determine the final pulling speed control response value used to control the silicon single crystal shoulder growth process. For example, the reliability of the first and second pulling speed response values ​​can be evaluated, and the pulling speed response value with higher reliability can be selected as the final pulling speed control response value. Alternatively, the first and second pulling speed response values ​​can be fitted, for example, by calculating the mean, median, or weighted sum, to obtain the final pulling speed control response value. Other fitting or fusion algorithms can also be used. Based on this, the characteristics of both image sequence data and process sequence data can be taken into account to accurately and reliably predict the pulling speed control response value.

[0050] Figure 2 This is a schematic diagram of a fusion control process provided in an embodiment of this application. Figure 2 As shown, image sequence data inside the single crystal furnace acquired by the CCD camera and process sequence data monitored by the sensor during the shoulder-forming process are acquired respectively. For the image sequence data, a first mapping model based on a network structure containing a residual neural network is established, and for the process sequence data, a second mapping model based on a neural network of physical control law is established. These two mapping models can be combined and integrated into the control module to ultimately realize the rolling calculation of the next pulling speed control response value.

[0051] This application provides a fusion control method for the shoulder growth process. It utilizes a mapping model to predict the pulling speed response value based on both image and process timing data, and combines these two values ​​to determine the final pulling speed control response value used to control the shoulder growth process of silicon single crystal. This improves the reliability of automated control of the shoulder growth process. Specifically, if only process timing data is considered, such as adjusting the pulling speed based solely on the deviation between the actual diameter value and the set diameter value, it is not suitable for the nonlinearity and large time delay issues of the silicon single crystal growth process, leading to control overshoot and undershoot, and consequently, discontinuity in the shoulder growth process. The method in this application also incorporates image timing data, achieving silicon single crystal growth process control through multimodal sensing end-to-end. This provides better adaptability and robustness to various problems during the growth process, effectively avoiding control overshoot and undershoot, and preventing discontinuity in the shoulder growth process. This significantly improves control stability and automation rate, and allows for faster response to changes in process settings, operating conditions, and product types. Furthermore, even considering the complexity of silicon single crystal growth processes, the diversity of product specifications, and the rapid pace of technological iteration, the method in this application embodiment does not require extensive practical experiments involving numerous rules and adjustments for different application requirements, thus reducing experimental costs and shortening the experimental cycle. Moreover, since it integrates two mapping models, it reduces the dependence on the accuracy of any one time series data or any one model. Even for the batch production process of silicon single crystals, despite differences in equipment, product diversity, and anomalies in any one time series data or any one model, the method in this application embodiment can still achieve accurate prediction of the pulling speed control response value.

[0052] In one embodiment, the first mapping model includes a temporal convolutional network, a residual neural network, and a fully connected layer;

[0053] The first pulling speed response value is predicted based on image sequence data and process sequence data using the first mapping model, including:

[0054] S1210. Extract features of image sequence data and process sequence data within a time period using a residual neural network;

[0055] S1220. The features of the image sequence data and the features of the process sequence data are spliced ​​together according to the time step through the fully connected layer to obtain the spliced ​​features.

[0056] S1230. The first pull speed response value of the next moment of the time period is predicted by a temporal convolutional network based on the splicing features.

[0057] For example, for feature extraction from image sequence data, a first mapping model can first be trained based on the training set data. This first mapping model includes a residual neural network, a fully connected layer, and a temporal convolutional network. After training, it can learn the relationship between the image sequence data, process sequence data, and the pulling speed response value (or pulling speed change) at the next time step. Optionally, the loss function for the residual neural network and the fully connected layer is the Mean Square Error (MSE) function. The residual neural network can be used to extract features from the image sequence data and process sequence data, where the dimension of the image sequence data features within each time step is (m, P, 1), and the input dimension of the process time series data is (n, P, 1). The fully connected layer can be used to concatenate the features of the image sequence data and the process sequence data according to time steps. The concatenated feature dimension is (m+n, P, 1), and then the concatenated features are input into the temporal convolutional network. During the training of the first mapping model, parameters such as the loss function, input stride, and network structure of the temporal convolutional neural network can be optimized. The training of the first mapping model is completed when the preset loss function satisfies the error convergence condition. Based on this, the parameters of the residual neural network can be retained as a feature extraction layer for the image sequence data. In the actual inference process, after inputting the acquired image sequence data and process sequence data into the first mapping model, the optimal temporal convolutional neural network parameters can be obtained to predict the pull-speed response of the next time step.

[0058] It should be noted that this application uses neural networks for image feature extraction and judgment. When the process or equipment changes, there is no need to redesign the control rules; the model can be directly optimized based on the production images. Even considering the complexity of the silicon single crystal growth process, the diversity of product specifications, and the rapid technological iteration, the method in this application's embodiments does not require extensive actual experiments involving numerous rules and adjustments for different application requirements, effectively reducing experimental costs and shortening the experimental cycle.

[0059] Optionally, due to the large number of features in image sequence data, in practical applications, the first mapping model can be compressed using methods such as Dropout and / or model pruning. This can improve the response speed of the first mapping model while maintaining its computational accuracy. Then, the parameters of the residual neural network are retained as the feature extraction layer for the image sequence set.

[0060] In one embodiment, the loss function of the second mapping model is determined based on the loss function of the crystal dynamics model and the least squares loss function.

[0061] For example, the loss function of the second mapping model is designed as follows: Loss = μloss c +(1-μ)lossd Among them, loss c The loss function represents the loss function of the crystal dynamics model. d This is the data MSE loss function, where μ is the coefficient of the loss term, and 0 < μ < 1. Based on this, the generalization ability and industrial applicability of the process timing-based speed response control model (i.e., the second mapping model) can be enhanced.

[0062] In one embodiment, the fusion control module outputs a pulling speed control response value based on a first pulling speed response value and a second pulling speed response value, including: through the fusion control module,

[0063] When the process sequence data is abnormal but the image sequence data is normal, the first pulling speed response value will be used as the pulling speed control response value.

[0064] When the process sequence data is normal but the image sequence data is abnormal, the second pulling speed response value will be used as the pulling speed control response value.

[0065] When both the process sequence data and the image sequence data are normal, the first pull speed response value and the second pull speed response are fused based on the adaptive fusion function to obtain the pull speed control response value.

[0066] Figure 3 This is a schematic diagram illustrating an embodiment of the present application that outputs a pulling speed control response value based on a first pulling speed response value and a second pulling speed response value. Figure 3 As shown, the first pulling speed response value obtained primarily from image sequence data and the second pulling speed response value obtained from process sequence data are fused. If the image sequence data is abnormal, such as partial image loss or poor image quality, while the process sequence data is normal, the second pulling speed response value can be used as the final pulling speed control response value. If the process sequence data is abnormal, such as sensor failure, unstable measurement feedback output, incorrect process parameter format or numerical error, while the image sequence data is normal, the first pulling speed response value can be used as the final pulling speed control response value. If both the image sequence data and the process sequence data are normal, an adaptive fusion function is used to adaptively fuse the first and second pulling speed response values ​​to obtain the final pulling speed control response value.

[0067] In one embodiment, the adaptive fusion function includes a Kalman-like fusion function; the pull speed control response value is obtained by weighting the first pull speed response and the second pull speed response, and the corresponding weights are updated based on the Kalman-like fusion function.

[0068] For example, the formula for updating the prediction error covariance of the Kalman filter-like fusion function over time is as follows: Among them, F i P is the state transition matrix of model i. i,n-1|n-1Q is the covariance of the estimation error of model i at time n-1. i It is the process noise covariance of model i. Model i can be understood as the speed control response value prediction model or multimodal sensing control model, that is, the model composed of the first mapping model, the second mapping model and the fusion control module, which is used to output the final speed control response value after fusion.

[0069] The gain calculation formula is as follows: Among them, H i R is the observation matrix of model i, and R is the observation noise covariance.

[0070]

[0071] By updating the weighting formula, the final pull speed control response value is:

[0072] O i,final =w i O i,Picture +(1-w i )O i,Tecnology , of which O i,final For the final pull-speed response output of adaptive fusion control, O i,Picture O is the first pulling speed response value corresponding to image sequence data. i,Tecnology This refers to the pulling speed response quantity corresponding to the process sequence.

[0073] In one embodiment, the method further includes:

[0074] S1010. Determine the coefficient of determination using the test dataset, which includes the predicted tension control response value and the actual tension control response value.

[0075] S1020. Evaluate the performance of the first mapping model, the second mapping model, and the fusion control module based on the coefficient of determination.

[0076] For example, the coefficient of determination between the predicted and actual speed control response values ​​is used as a performance index to evaluate the multimodal sensing control method, thereby ensuring the accuracy of the speed control response prediction model. For the acquired image sequence data and process sequence data, these sequence data can be divided into training and test sets, for example, the ratio of data volume in the training and test sets can be 7:3, and then the evaluation can be performed on the training and test sets respectively.

[0077] Coefficient of determination (R) 2 The calculation formula for ) is as follows: in, It is the predicted tension control response value, u i This is the actual pulling speed control response value. This represents the average value of the actual pulling speed control response, where N is the sample size, i.e., the number of samples. Based on this, the fluctuation of the pulling speed output response can be evaluated by assessing the coefficient of determination, while ensuring that the predetermined process and control objectives are met. According to the requirements of the shoulder forming process, while satisfying the crystal formation requirements, the smaller the pulling speed fluctuation, the fewer the invalid outputs of the controller, and the stronger the controller performance.

[0078] The following specific examples illustrate the fusion control method for the shoulder growth process.

[0079] Image data and process data are read from the CCD camera and furnace process sensors in real time and synchronously, and gradually accumulated to form image sequence data and process sequence data. The data is then input into the optimized and trained multimodal perception fusion control model (including the first mapping model, the second mapping model and the fusion control module), and the pull speed response value is output in a rolling optimization manner.

[0080] The optimized and trained multimodal sensing fusion control model can be evaluated. Table 1 shows a comparison of the performance of single-modal and multimodal sensing fusion control. The accuracy (coefficient of determination) Ri of the predicted and actual acceleration control response values ​​of the multimodal sensing fusion control model on the test set is also shown. 2 =0.8, indicating that multimodal perception pre-training modeling can accurately characterize the relationship between image sequences, process sequences and pulling speed response.

[0081] As shown in Table 1, compared with image modeling and process timing modeling, multimodal sensing pre-trained modeling has the advantages of low training cost, strong adaptability, and high prediction accuracy, which is more in line with the requirements of expert operation. Specifically, regarding training cost, the training cost of each model iteration of multimodal sensing pre-training is 1 / 5000 of the cost of pure image sequence modeling, which significantly reduces the training cost of multimodal sensing models; for the trained model, the inference time of the multimodal sensing pre-trained model is 0.1s, which meets the needs of industrial control. Compared with the existing patented methods that detect images or process parameters and rely on expert-established rule tables to output control response values, the fusion control method of the shoulder growth process in this application embodiment is an end-to-end control method. There is no expert experience in the fusion process from modeling to control, so multimodal sensing control is more adaptable and suitable for large-scale silicon single crystal production processes.

[0082] Table 1 Comparison of control effects between single-modal and multi-modal sensing fusion

[0083] Modeling pure image sequence data ￥50000 middle 0.75 / Pure process sequence data modeling ￥1 weak 0.51 0.01s Multimodal perception pre-training modeling ￥10 powerful 0.80 0.1s

[0084] In addition, the multimodal sensing fusion control method was tested on 10-inch and 10.5-inch drawing specifications, and under conditions of overheating, overcooling, and normal initial shoulder setting. The drawing speed control curve during the shoulder setting process under standard conditions is shown below. Figure 4 As shown, the corresponding shoulder shape image at the end of the shoulder relaxation is as follows: Figure 5 As shown; the drawing speed control curve during the initial overheating process of shoulder release is as follows. Figure 6 As shown, the corresponding shoulder shape image at the end of the shoulder relaxation is as follows: Figure 7 As shown, for the initial shoulder-forming states of undercooling, overheating, normal, and two drawing specifications, the shoulder-forming process was successful in all embodiments. The synchronized shoulder-forming completion image shows a smooth shoulder-forming process without steps or bright lines, meeting the requirements of process experts. Drawing speed data shows that the drawing speed during the shoulder-forming process was stable; compared to model predictive control, the drawing speed during shoulder-forming was more stable, and the diameter growth was stable, conforming to the process setting. Therefore, the multimodal sensing intelligent control method has strong anti-interference capabilities for measuring process noise data and exhibits high success rate and high adaptability.

[0085] The fusion control method for the shoulder growth process provided in this application embodiment is designed for multi-modal sensing and intelligent control under the non-steady-state Czochralski single-crystal silicon conditions in mass production. It achieves high automation and intelligence in the shoulder control during mass production, which can improve shoulder stability, automation rate, intelligence rate and survival rate, reduce ineffective working hours and lower production and operating costs.

[0086] Figure 8 This is a schematic diagram of a fusion control device for a shoulder-growing process provided in an embodiment of this application. The fusion control device for the shoulder-growing process provided in this embodiment includes:

[0087] The data acquisition module 210 is used to acquire image sequence data and process sequence data of the shoulder formation process inside the single crystal furnace;

[0088] The first prediction module 220 is used to predict the first pulling speed response value based on image sequence data and process sequence data through a first mapping model. The first mapping model includes a pre-trained residual neural network.

[0089] The second prediction module 230 is used to predict the second pull speed response value based on the process sequence data through the second mapping model. The second mapping model includes a physical neural network.

[0090] The fusion control module 240 is used to output a pulling speed control response value based on the first pulling speed response value and the second pulling speed response value, and to control the silicon single crystal shoulder growth process based on the pulling speed control response value.

[0091] The device uses a mapping model to predict the pulling speed response value based on the timing data of the image and the process, and combines the two pulling speed response values ​​to determine the final pulling speed control response value used to control the shoulder growth process of silicon single crystal, thereby improving the reliability of automated control of the shoulder growth process.

[0092] Based on any of the above embodiments, the first mapping model includes a temporal convolutional network, a residual neural network, and a fully connected layer;

[0093] The first prediction module 220 includes:

[0094] The extraction unit is used to extract features of image sequence data and process sequence data within a time period using a residual neural network;

[0095] The stitching unit is used to stitch together the features of image sequence data and process sequence data according to time steps through a fully connected layer to obtain stitched features;

[0096] The prediction unit is used to predict the first pull speed response value of the next moment of the time period based on the splicing features through a temporal convolutional network.

[0097] Based on any of the above embodiments, the loss function of the second mapping model is determined according to the loss function of the crystal dynamics model and the least squares loss function.

[0098] Based on any of the above embodiments, the fusion control module 240 is specifically used for:

[0099] When the process sequence data is abnormal but the image sequence data is normal, the first pulling speed response value will be used as the pulling speed control response value.

[0100] When the process sequence data is normal but the image sequence data is abnormal, the second pulling speed response value will be used as the pulling speed control response value.

[0101] When both the process sequence data and the image sequence data are normal, the first pull speed response value and the second pull speed response are fused based on the adaptive fusion function to obtain the pull speed control response value.

[0102] Based on any of the above embodiments, the adaptive fusion function includes a Kalman filter-like fusion function; the pull speed control response value is obtained by weighting the first pull speed response and the second pull speed response, and the corresponding weights are updated based on the Kalman filter-like fusion function.

[0103] Based on any of the above embodiments, the device further includes:

[0104] The testing module is used to determine the coefficient of determination using a test dataset, which includes the predicted and actual pull speed control response values.

[0105] The evaluation module is used to evaluate the performance of the first mapping model, the second mapping model, and the fusion control module based on the coefficient of determination.

[0106] The fusion control device for the shoulder growth process provided in this application embodiment can be used to execute the fusion control method for the shoulder growth process provided in any of the above embodiments, and has corresponding functions and beneficial effects.

[0107] Figure 9 A schematic diagram of an electronic device 10, which can be used to implement embodiments of this application, is shown. The electronic device 10 is intended to represent various forms of digital computers, such as laptop computers, desktop computers, workstations, personal digital assistants, servers, blade servers, mainframe computers, and other suitable computers. The electronic device 10 may also represent various forms of mobile devices, such as personal digital processors, cellular phones, smartphones, user equipment, and other similar computing devices. The components shown herein, their connections and relationships, and their functions are merely illustrative and are not intended to limit the implementation of the application described and / or claimed herein.

[0108] like Figure 9 As shown, the electronic device 10 includes at least one processor 11 and a memory, such as a read-only memory (ROM) 12 or a random access memory (RAM) 13, communicatively connected to the at least one processor 11. The memory stores computer programs executable by the at least one processor. The processor 11 can perform various appropriate actions and processes based on the computer program stored in the ROM 12 or loaded from storage unit 18 into the RAM 13. The RAM 13 can also store various programs and data required for the operation of the electronic device 10. The processor 11, ROM 12, and RAM 13 are interconnected via a bus 14. An input / output (I / O) interface 15 is also connected to the bus 14.

[0109] Multiple components in electronic device 10 are connected to I / O interface 15, including: input unit 16, such as keyboard, mouse, etc.; output unit 17, such as various types of displays, speakers, etc.; storage unit 18, such as disk, optical disk, etc.; and communication unit 19, such as network card, modem, wireless transceiver, etc. Communication unit 19 allows electronic device 10 to exchange information / data with other devices through computer networks such as the Internet and / or various telecommunications networks and wireless networks.

[0110] Processor 11 can be a variety of general-purpose and / or special-purpose processing components with processing and computing capabilities. Some examples of processor 11 include, but are not limited to, a central processing unit (CPU), a graphics processing unit (GPU), various special-purpose artificial intelligence (AI) computing chips, various processors running machine learning model algorithms, a digital signal processor (DSP), and any suitable processor, controller, microcontroller, etc. Processor 11 performs the various methods and processes described above.

[0111] In some embodiments, the methods described above can be implemented as a computer program tangibly contained in a computer-readable storage medium, such as storage unit 18. In some embodiments, part or all of the computer program can be loaded and / or installed on electronic device 10 via ROM 12 and / or communication unit 19. When the computer program is loaded into RAM 13 and executed by processor 11, one or more steps of the methods described above can be performed. Alternatively, in other embodiments, processor 11 can be configured to perform the methods of any of the embodiments described above by any other suitable means (e.g., by means of firmware).

[0112] Various embodiments of the systems and techniques described above herein can be implemented in digital electronic circuit systems, integrated circuit systems, field-programmable gate arrays (FPGAs), application-specific integrated circuits (ASICs), application-specific standard products (ASSPs), systems-on-a-chip (SoCs), payload-programmable logic devices (CPLDs), computer hardware, firmware, software, and / or combinations thereof. These various embodiments may include implementations in one or more computer programs that can be executed and / or interpreted on a programmable system including at least one programmable processor, which may be a dedicated or general-purpose programmable processor, capable of receiving data and instructions from a storage system, at least one input device, and at least one output device, and transmitting data and instructions to the storage system, the at least one input device, and the at least one output device.

[0113] Computer programs used to implement the methods of this application may be written in any combination of one or more programming languages. These computer programs may be provided to a processor of a general-purpose computer, a special-purpose computer, or other programmable data processing device, such that when executed by the processor, the computer programs cause the functions / operations specified in the flowcharts and / or block diagrams to be performed. The computer programs may be executed entirely on a machine, partially on a machine, or as a standalone software package, partially on a machine and partially on a remote machine, or entirely on a remote machine or server.

[0114] In the context of this application, a computer-readable storage medium can be a tangible medium that may contain or store a computer program for use by or in conjunction with an instruction execution system, apparatus, or device. A computer-readable storage medium can be, but is not limited to, electronic, magnetic, optical, electromagnetic, infrared, or semiconductor systems, apparatus, or devices, or any suitable combination of the foregoing. Alternatively, a computer-readable storage medium can be a machine-readable signal medium. More specific examples of machine-readable storage media include electrical connections based on one or more wires, portable computer disks, hard disks, random access memory (RAM), read-only memory (ROM), erasable programmable read-only memory (EPROM or flash memory), optical fiber, portable compact disk read-only memory (CD-ROM), optical storage devices, magnetic storage devices, or any suitable combination of the foregoing.

[0115] To provide interaction with a user, the systems and techniques described herein can be implemented on an electronic device 10, which includes: a display device (e.g., a CRT (cathode ray tube) or LCD (liquid crystal display) monitor) for displaying information to the user; and a keyboard and pointing device (e.g., a mouse or trackball) through which the user provides input to the electronic device 10. Other types of devices can also be used to provide interaction with the user; for example, feedback provided to the user can be any form of sensory feedback (e.g., visual feedback, auditory feedback, or tactile feedback); and input from the user can be received in any form (including sound input, voice input, or tactile input).

[0116] The systems and technologies described herein can be implemented in computing systems that include backend components (e.g., as data servers), or computing systems that include middleware components (e.g., application servers), or computing systems that include frontend components (e.g., user computers with graphical user interfaces or web browsers through which users can interact with implementations of the systems and technologies described herein), or any combination of such backend, middleware, or frontend components. The components of the system can be interconnected via digital data communication of any form or medium (e.g., communication networks). Examples of communication networks include local area networks (LANs), wide area networks (WANs), blockchain networks, and the Internet.

[0117] A computing system can include clients and servers. Clients and servers are generally located far apart and typically interact through communication networks. The client-server relationship is created by computer programs running on the respective computers and having a client-server relationship with each other. The server can be a cloud server, also known as a cloud computing server or cloud host, which is a hosting product within the cloud computing service system to address the shortcomings of traditional physical hosts and VPS services, such as high management difficulty and weak business scalability.

[0118] This application also provides a computer program product, including a computer program and / or instructions, which, when executed by a processor, implement the fusion control method for the shoulder growth process as described in any of the above embodiments.

[0119] It should be understood that the various forms of processes shown above can be used to rearrange, add, or delete steps. For example, the steps described in this application can be executed in parallel, sequentially, or in different orders, as long as the desired result of the technical solution of this application can be achieved, and this is not limited herein.

[0120] The specific embodiments described above do not constitute a limitation on the scope of protection of this application. Those skilled in the art should understand that various modifications, combinations, sub-combinations, and substitutions can be made according to design requirements and other factors. Any modifications, equivalent substitutions, and improvements made within the spirit and principles of this application should be included within the scope of protection of this application.

Claims

1. A method for controlling the fusion process of shoulder growth, characterized in that, include: Acquire image sequence data and process sequence data of the shoulder formation process inside the single crystal furnace; The first pulling speed response value is predicted based on image sequence data and process sequence data through the first mapping model, which includes a pre-trained residual neural network. The second pulling speed response value is predicted based on the process sequence data through the second mapping model, which includes a physical neural network. The fusion control module outputs a pulling speed control response value based on the first pulling speed response value and the second pulling speed response value, and controls the silicon single crystal shoulder growth process based on the pulling speed control response value. The first mapping model includes a temporal convolutional network, the residual neural network, and a fully connected layer; The first pulling speed response value is predicted based on image sequence data and process sequence data using the first mapping model, including: Features of image sequence data and process sequence data within a time period are extracted using a residual neural network. By using a fully connected layer, the features of image sequence data and process sequence data are spliced ​​together according to time steps to obtain spliced ​​features; The first pull-speed response value of the next moment in the time period is predicted by a temporal convolutional network based on splicing features; The loss function of the second mapping model is determined based on the loss function of the crystal dynamics model and the least squares loss function.

2. The method according to claim 1, characterized in that, The fusion control module outputs a pulling speed control response value based on the first pulling speed response value and the second pulling speed response value, including: Through the fusion control module When the process sequence data is abnormal but the image sequence data is normal, the first pulling speed response value will be used as the pulling speed control response value. When the process sequence data is normal but the image sequence data is abnormal, the second pulling speed response value will be used as the pulling speed control response value. When both the process sequence data and the image sequence data are normal, the first and second pull speed response values ​​are fused based on the adaptive fusion function to obtain the pull speed control response value.

3. The method according to claim 2, characterized in that, Adaptive fusion functions include Kalman filter-like fusion functions; The pulling speed control response value is obtained by weighting the first pulling speed response value and the second pulling speed response value, and the corresponding weights are updated based on the Kalman filter-like fusion function.

4. The method according to claim 1, characterized in that, Also includes: The coefficient of determination was determined using a test dataset, which included both the predicted and actual pull speed control response values. The performance of the first mapping model, the second mapping model, and the fusion control module is evaluated based on the coefficient of determination.

5. A fusion control device for the shoulder growth process, characterized in that, include: The data acquisition module is used to acquire image sequence data inside the single crystal furnace and process sequence data of the shoulder formation process; The first prediction module is used to predict the first pulling speed response value based on image sequence data and process sequence data through a first mapping model. The first mapping model includes a pre-trained residual neural network. The second prediction module is used to predict the second pull speed response value based on the process sequence data through the second mapping model. The second mapping model includes a physical neural network. The fusion control module is used to output a pulling speed control response value based on the first pulling speed response value and the second pulling speed response value, and to control the silicon single crystal shoulder growth process based on the pulling speed control response value. The first mapping model includes a temporal convolutional network, a residual neural network, and a fully connected layer; The first prediction module is specifically used for: Features of image sequence data and process sequence data within a time period are extracted using a residual neural network. By using a fully connected layer, the features of image sequence data and process sequence data are spliced ​​together according to time steps to obtain spliced ​​features; The first pull-speed response value of the next moment in the time period is predicted by a temporal convolutional network based on splicing features; The loss function of the second mapping model is determined based on the loss function of the crystal dynamics model and the least squares loss function.

6. An electronic device, characterized in that, include: At least one processor; A memory that is communicatively connected to at least one processor; wherein, The memory stores a computer program that can be executed by at least one processor, the computer program being executed by at least one processor to enable at least one processor to perform the fusion control method for the shoulder growth process as described in any one of claims 1-4.

7. A computer-readable storage medium having a computer program stored thereon, characterized in that, When executed by the processor, the program implements the fusion control method for the shoulder growth process as described in any one of claims 1-4.

8. A computer program product, comprising a computer program and / or instructions, characterized in that, When a computer program and / or instructions are executed by a processor, they implement a fusion control method for the shoulder growth process as described in any one of claims 1-4.

Images