Deep learning model-based shouldering control method and device and related equipment

Abstract

The invention provides a shouldering control method and device based on a deep learning model, and related equipment, and belongs to the technical field of semiconductor manufacturing. The method comprises the steps that under the condition that shouldering is abnormal in the Czochralski silicon shouldering process, first shouldering parameters in the Czochralski silicon shouldering process are obtained, and the first shouldering parameters are used for representing attribute data and environment data of the Czochralski silicon in the shouldering process; predicting the shouldering process of the czochralski silicon according to the first shouldering parameter through a pre-trained deep learning model to obtain a second shouldering parameter; and carrying out shouldering on the czochralski silicon based on the second shouldering parameter. According to the method, the shouldering effect of the czochralski silicon in the shouldering process can be effectively improved.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of semiconductor manufacturing, in particular to a shoulder growth control method and device based on a deep learning model and related equipment. BACKGROUND

[0002] The shoulder growth of the Czochralski process refers to a process adjustment in the process of growing single crystal silicon by the Czochralski process (CZ) to improve the quality and production efficiency of the crystal. The concept of shoulder growth is usually related to factors such as temperature, pulling speed and environmental pressure during the growth of the crystal. In the manufacture of single crystal silicon, the Czochralski process is to heat the silicon melt to a high temperature and grow a single crystal silicon rod by using a seed crystal. In the prior art, for the shoulder growth of the Czochralski process, an algorithm with a fixed pulling speed is generally used for control. However, this method cannot adapt to the complex and variable conditions of the crystal growth environment during the shoulder growth process, thereby resulting in the problem of poor shoulder growth effect of the Czochralski process in the prior art. SUMMARY

[0003] To solve the above technical problems, the present application provides a shoulder growth control method and device based on a deep learning model and related equipment, which solves the problem of poor shoulder growth effect of the Czochralski process in the prior art.

[0004] To achieve the above purpose, the technical solution adopted by the embodiments of the present application is as follows: In the case of abnormal shoulder growth in the process of the Czochralski process, a first shoulder growth parameter in the process of the Czochralski process is obtained, and the first shoulder growth parameter is used to represent attribute data and environmental data of the Czochralski process in the process of the Czochralski process; A pre-trained deep learning model is used to predict the shoulder growth process of the Czochralski process according to the first shoulder growth parameter, and a second shoulder growth parameter is obtained; The Czochralski process is performed based on the second shoulder growth parameter.

[0005] In some embodiments, before the step of predicting the shoulder growth process of the Czochralski process according to the first shoulder growth parameter by using the pre-trained deep learning model to obtain the second shoulder growth parameter, the method further comprises: A plurality of training data sets are obtained, each of which includes a sample shoulder growth parameter and a sample label, wherein the sample shoulder growth parameter is an actual shoulder growth parameter of a sample Czochralski process, and the sample label is a shoulder growth parameter in the normal shoulder growth process of the sample Czochralski process; An initial deep learning model is trained based on the plurality of training data sets to obtain the pre-trained deep learning model.

[0006] In some embodiments, the training data set includes high-frequency data, medium-frequency data, and low-frequency data. The high-frequency data includes at least one of the following: melt temperature, pulling speed, crucible lifting speed, and heating power. The medium-frequency data includes at least one of the following: crystal diameter morphology image and crystal edge line morphology image. The low-frequency data includes historical shoulder data.

[0007] In some embodiments, the training of the initial deep learning model based on the plurality of training data sets to obtain the pre-trained deep learning model includes: Preprocessing the plurality of training data sets to obtain preprocessed plurality of training data sets, the preprocessing being used to align the timestamps of the plurality of training data sets and normalize the corresponding dimensions of the plurality of training data sets to a preset dimension; Encoding the high-frequency data based on the control variable encoder of the initial deep learning model to obtain a first encoding result, encoding the medium-frequency data based on the visual feature encoder of the initial deep learning model to obtain a second encoding result, and encoding the low-frequency data based on the historical knowledge encoder of the initial deep learning model to obtain a third encoding result; Fusion processing the first encoding result, the second encoding result, and the third encoding result based on the fusion module of the initial deep learning model to generate fusion features; Identifying the fusion features based on the classification module of the initial deep learning model and outputting predicted shoulder parameters; Updating the initial deep learning model based on the predicted shoulder parameters to obtain the pre-trained deep learning model.

[0008] In some embodiments, the classification module includes a regression submodule and a classification submodule, and identifying the fusion features based on the classification module of the initial deep learning model and outputting predicted shoulder parameters includes: Identifying the fusion features based on the regression submodule of the initial deep learning model and outputting predicted shoulder control parameters; Identifying the fusion features based on the classification submodule of the initial deep learning model and outputting a probability value, the predicted probability value being a probability value of the training data set being abnormal in the shoulder; The predicted shoulder parameters include the predicted shoulder control parameters and the predicted probability value.

[0009] In some embodiments, the updating of the initial deep learning model based on the predicted shoulder parameters to obtain the pre-trained deep learning model comprises: loss calculation of the predicted shoulder parameters and the sample label based on a preset loss function to obtain a loss value; updating of the initial deep learning model based on the loss value to obtain the pre-trained deep learning model.

[0010] In a second aspect, the present application provides a shoulder control device based on a deep learning model, which comprises: an acquisition module configured to acquire a first shoulder parameter in a case where an abnormal shoulder occurs in a Czochralski silicon shoulder process, the first shoulder parameter being used to represent attribute data and environmental data of the Czochralski silicon in the shoulder process; a prediction module configured to predict a shoulder process of the Czochralski silicon based on the first shoulder parameter by using a pre-trained deep learning model to obtain a second shoulder parameter; a shoulder module configured to perform a shoulder on the Czochralski silicon based on the second shoulder parameter.

[0011] In a third aspect, the present application further provides an electronic device comprising a processor, a memory, and a computer program stored in the memory and executable on the processor, wherein the computer program is executed by the processor to implement the steps in the method of the first aspect.

[0012] In a fourth aspect, the present application further provides a computer readable storage medium having a computer program stored thereon, wherein the computer program is executed by a processor to implement the steps in the method of the first aspect.

[0013] In a fifth aspect, the present application further provides a computer program product comprising computer instructions, wherein the computer instructions are executed by a processor to implement the steps in the method of the first aspect.

[0014] The present application has the following beneficial effects: Through the technical scheme of the present application, the shoulder state of the Czochralski silicon is detected in the Czochralski silicon shoulder process, the first shoulder parameter is acquired in a case where an abnormal shoulder occurs, the first shoulder parameter is predicted by using a pre-trained deep learning model to generate a second shoulder parameter in a normal shoulder process, and the Czochralski silicon is shouldered based on the second shoulder parameter, thereby avoiding the failure of the Czochralski silicon shoulder and effectively improving the shoulder effect of the Czochralski silicon. BRIEF DESCRIPTION OF DRAWINGS

[0015] In order to more clearly illustrate the technical solutions of the present application, the drawings required to be used in the description of the present application will be briefly introduced as follows. Obviously, the drawings in the following description are only some embodiments of the present application, and other drawings can be obtained by those skilled in the art without creative labor on the basis of these drawings.

[0016] Figure 1 is a flow chart of a shoulder placing control method based on a deep learning model provided in an embodiment of the present application; Figure 2 is a deep learning model structure schematic diagram in an embodiment of the present application; Figure 3 is a structure diagram of a shoulder placing control device based on a deep learning model provided in an embodiment of the present application; Figure 4 is a structure diagram of an electronic device provided in an embodiment of the present application. DETAILED DESCRIPTION

[0017] The technical solutions in the embodiments of the present application will be described clearly and completely in combination with the drawings in the embodiments of the present application. Obviously, the described embodiments are only some of the embodiments of the present application, not all the embodiments. Based on the embodiments in the present application, all other embodiments obtained by those skilled in the art without creative labor are within the protection scope of the present application.

[0018] The terms "first", "second", and the like in the embodiments of the present application are used to distinguish similar objects, and do not necessarily indicate a specific order or sequence. In addition, the terms "include" and "have" and any variations thereof are intended to cover non-exclusive inclusion, for example, a process, method, system, product or device including a series of steps or units does not necessarily limit to those clearly listed steps or units, but can include other steps or units not clearly listed or inherent to these processes, methods, products or devices. In addition, "and / or" is used in the present application to represent at least one of the connected objects, for example, A and / or B and / or C represents 7 cases including A alone, B alone, C alone, A and B both exist, B and C both exist, A and C both exist, and A, B and C all exist.

[0019] The present application provides a shoulder placing control method, device and related equipment based on a deep learning model, which can effectively improve the shoulder placing effect of the Czochralski silicon in the shoulder placing process.

[0020] The embodiment of the present application provides a shoulder placing control method based on a deep learning model, as shown in Figure 1 , comprising: Step 101, in the case of abnormality in the straight pull single crystal silicon shoulder process, the first shoulder parameter in the straight pull single crystal silicon shoulder process is obtained, and the first shoulder parameter is used to represent the attribute data and environmental data of the straight pull single crystal silicon in the shoulder process.

[0021] In the embodiment, the straight pull single crystal silicon is a single crystal silicon material produced by the Czochralski method. The structure of the single crystal silicon is uniform and orderly, and it exhibits excellent performance in electronic and optoelectronic applications. In the embodiment, the straight pull single crystal silicon shoulder is formed by heating the silicon melt to a high temperature and using a seed crystal for crystal pulling growth to form a single crystal silicon rod.

[0022] In the shoulder process of the straight pull single crystal silicon, the first shoulder parameter in the straight pull single crystal silicon shoulder process is obtained by detecting the shoulder process of the straight pull single crystal silicon, and in the case of abnormality in the shoulder process of the straight pull single crystal silicon. The abnormality of the shoulder can include one of the following situations: irregular shape, decreased crystalline quality, temperature control problem, inconsistent stretching speed, and blank fracture or deviation.

[0023] It should be noted that in the embodiment, the first shoulder parameter is used to represent the attribute data and environmental data of the straight pull single crystal silicon in the shoulder process, wherein the attribute data is the data corresponding to the straight pull single crystal silicon, such as the photographed shoulder image, etc., and the environmental data is the data corresponding to the straight pull single crystal silicon shoulder environment, such as the melt temperature and the pulling speed, etc., which is not limited in the embodiment.

[0024] Step 102, by using the pre-trained deep learning model, the shoulder process of the straight pull single crystal silicon is predicted according to the first shoulder parameter, and the second shoulder parameter is obtained.

[0025] In the embodiment, the pre-trained deep learning model is a deep learning model trained in advance. In the embodiment, the deep learning model is taken as an example of CNN-LSTM hybrid network or spatio-temporal graph neural network. The above model can reflect the current state in real time and can accurately predict the crystal diameter change trend, the solid-liquid interface morphology evolution, the thermal stress distribution and the potential defect (such as dislocation initiation) risk in the next 5-10 seconds.

[0026] The first shoulder parameter is input into the pre-trained deep learning model for prediction to obtain the second shoulder parameter. Specifically, the second shoulder parameter can include continuous prediction control values and probability values. The continuous prediction control values are output, such as: recommended pulling speed adjustment amount (Δv), heating power adjustment amount (ΔP). The probability value represents the risk score of the current state deviating from the normal shoulder trajectory. When the risk score exceeds the threshold value, the system can trigger a high-level alarm or take more aggressive control strategies.

[0027] Step 103, based on the second shoulder parameters, the Czochralski silicon is shoulder.

[0028] In this embodiment, the Czochralski silicon is shoulder according to the determined second shoulder parameters, so as to ensure the normal shoulder process of the Czochralski silicon.

[0029] Through the technical scheme of the present application, the shoulder state of the Czochralski silicon is detected during the shoulder process of the Czochralski silicon, in the case of abnormal shoulder, the current first shoulder parameter is obtained, and the first shoulder parameter is predicted by the pre-trained deep learning model to predict the shoulder process of the Czochralski silicon, and the second shoulder parameter in the normal shoulder process is generated, so as to shoulder the Czochralski silicon based on the second shoulder parameter, thereby avoiding the failure of the Czochralski silicon shoulder, thereby effectively improving the shoulder effect of the Czochralski silicon.

[0030] In some embodiments, before the pre-trained deep learning model is used to predict the shoulder process of the Czochralski silicon according to the first shoulder parameter to obtain the second shoulder parameter, the method further comprises: Obtain a plurality of training data sets, each of which includes a sample shoulder parameter and a sample label, wherein the sample shoulder parameter is the actual shoulder parameter of the sample Czochralski silicon in the shoulder process, and the sample label is the shoulder parameter in the normal shoulder process of the sample Czochralski silicon. Based on the plurality of training data sets, an initial deep learning model is trained to obtain the pre-trained deep learning model.

[0031] In this embodiment, before obtaining the pre-trained deep learning model, the initial deep learning model needs to be trained by a plurality of training data sets until the initial deep learning model is trained to obtain the pre-trained deep learning model. Specifically, each training data set includes a sample shoulder parameter and a sample label, the sample shoulder parameter is the actual shoulder parameter of the sample Czochralski silicon in the shoulder process, and the sample label is the shoulder parameter in the normal shoulder process of the sample Czochralski silicon.

[0032] Therefore, the initial deep learning model is trained by a plurality of training data sets, and the pre-trained deep learning model is obtained after the model training is completed.

[0033] It should be noted that the sample shoulder parameter includes positive samples and negative samples, wherein a large amount of successful shoulder process data (including all the above modal data) is extracted from the historical database as the positive sample. Extract the shoulder process data of the failure, the occurrence of the edge, the dislocation and other abnormalities as the negative sample. Each sample is a time sequence segment, for example, containing all multi-modal data within t minutes after the shoulder starts.

[0034] In some embodiments, the training data set comprises high-frequency data, medium-frequency data and low-frequency data. The high-frequency data comprises at least one of the following: melt temperature, pulling speed, crucible lifting speed and heating power. The medium-frequency data comprises at least one of the following: crystal diameter morphology image and crystal edge line morphology image. The low-frequency data comprises historical shoulder setting data.

[0035] In the present embodiment, the plurality of training data sets are composed of high-frequency data, medium-frequency data and low-frequency data, and the optimal proportion of the three types of data is 1:1:1. Specifically, the high-frequency data is high-frequency time series data (in seconds), including real-time melt temperature (multiple temperature measurement points), pulling speed, crucible lifting speed and heating power, which constitute a real-time control variable flow. The medium-frequency data is medium-frequency visual / measurement data (in seconds), which is the crystal diameter and edge line morphology image captured by a camera or an infrared sensor. It should be noted that the morphology features such as diameter, edge line width and shoulder angle are extracted from the image by image processing algorithms (such as edge detection and semantic segmentation). The low-frequency data is low-frequency historical knowledge data (in furnace times): historical furnace time data (success and failure cases). Each data is a structured record, including but not limited to: final crystal bar quality label (success / failure and defect type), average process parameters of the furnace time, key event points (such as shoulder setting completion time and abnormal occurrence time), etc., which constitute a historical knowledge base.

[0036] In some embodiments, the training of the initial deep learning model based on the plurality of training data sets to obtain the pre-trained deep learning model comprises: preprocessing the plurality of training data sets to obtain preprocessed plurality of training data sets, the preprocessing being used to align the timestamps of the plurality of training data sets and normalize the corresponding dimensions of the plurality of training data sets to a preset dimension; encoding the high-frequency data based on a control variable encoder of the initial deep learning model to obtain a first encoding result, encoding the medium-frequency data based on a visual feature encoder of the initial deep learning model to obtain a second encoding result, and encoding the low-frequency data based on a historical knowledge encoder of the initial deep learning model to obtain a third encoding result; performing fusion processing on the first encoding result, the second encoding result and the third encoding result based on a fusion module of the initial deep learning model to generate fusion features; performing identification on the fusion features based on a classification module of the initial deep learning model to output predicted shoulder setting parameters; update the initial deep learning model based on the predicted shoulder parameters to obtain the pre-trained deep learning model.

[0037] In the embodiment, the plurality of training data sets need to be preprocessed first to obtain preprocessed plurality of training data sets, specifically, the preprocessing includes standardization processing and timestamp alignment.

[0038] Among them, the standardization processing refers to normalizing different dimensional data (such as temperature (℃), speed (mm / min), power (kW)) to the same scale (such as [-1, 1] interval). The timestamp alignment refers to using interpolation or sliding window method to unify all data streams to the same time sequence index to form a synchronous multi-modal data matrix, and the timestamp alignment is the premise of effective fusion.

[0039] As shown in FIG. 1, Figure 2 Figure 2 is a deep learning model structure diagram in the embodiment, after obtaining the preprocessed plurality of training data sets, the high-frequency data, the medium-frequency data and the low-frequency data are encoded by the control variable encoder, the visual feature encoder and the historical knowledge encoder respectively to generate the corresponding first encoding result, the second encoding result and the third encoding result.

[0040] It needs to be noted that the control variable encoder refers to using one-dimensional convolutional neural network (1D-CNN) to extract local time sequence pattern for time sequence data such as pulling speed and temperature, followed by long short-term memory network (LSTM) to capture long-term dependency. The visual feature encoder refers to using convolutional neural network (CNN) (such as ResNet backbone network) to extract rich spatial morphological features for crystal diameter and ridge image. The historical knowledge encoder refers to using fully connected layer (FC) to encode the historical furnace data to generate a “knowledge vector” representing historical experience.

[0041] ​After obtaining the first encoding result, the second encoding result and the third encoding result, the fusion module of the initial deep learning model is used to fuse the first encoding result, the second encoding result and the third encoding result to generate a fusion feature. Specifically, the fusion module can be a cross-modal attention fusion module, which allows the model to learn to "focus" on the importance of different data sources at a specific moment. The working principle is to use the control variable sequence as Query (Q). The visual feature sequence and the historical knowledge vector are used as Key (K) and Value (V). By calculating the attention weight, the model can dynamically determine whether to pay more attention to the visual form change of the crystal or to refer to the experience of a successful case in history at the current moment; for example, when the edge line appears a slight fluctuation, the model will "pay attention" to the change of the visual feature and refer to the case of successfully adjusting the power under similar conditions in the historical knowledge to generate a control instruction. Through the dynamic and nonlinear fusion mode, the understanding and decision-making ability of the model for complex working conditions are greatly improved.

[0042] After the fusion feature is generated, a classification module including several fully connected networks is used to output a predicted shoulder parameter, i.e., a prediction result, and the initial deep learning model is updated based on the prediction result to obtain the pre-trained deep learning model.

[0043] In some embodiments, the classification module includes a regression submodule and a classification submodule, and the classification module based on the initial deep learning model identifies the fusion feature and outputs a predicted shoulder parameter, including: The regression submodule based on the initial deep learning model identifies the fusion feature and outputs a predicted shoulder control parameter; The classification submodule based on the initial deep learning model identifies the fusion feature and outputs a probability value, and the predicted probability value is a probability value of the training data set being in an abnormal shoulder condition; The predicted shoulder parameter includes the predicted shoulder control parameter and the predicted probability value.

[0044] In this embodiment, the classification module includes a regression submodule and a classification submodule, wherein the regression submodule is used to output continuous predicted control values, such as a recommended pulling speed adjustment amount (Δv) and a heating power adjustment amount (ΔP). The classification submodule is used to output a probability value representing a risk score of the current state deviating from the normal shoulder trajectory. When the risk score exceeds a threshold value, the system can trigger a high-level alarm or take a more aggressive control strategy.

[0045] In some embodiments, updating the initial deep learning model based on the predicted shoulder parameter to obtain the pre-trained deep learning model includes: Based on a preset loss function, the loss is calculated using the predicted shoulder placement parameters and the sample labels to obtain the loss value; The initial deep learning model is updated based on the loss value to obtain the pre-trained deep learning model.

[0046] In this embodiment, the preset loss function is a composite loss function: total loss = α * mean squared error loss (MSELoss) + β * cross-entropy loss (Cross-Entropy Loss).

[0047] Among them, MSE Loss measures the error between predicted process parameters (such as the ideal pulling speed in the next few seconds) and the actual value. Cross-Entropy Loss determines the probability that the current state is closer to a "success" mode or a "failure" mode.

[0048] The predicted bearing parameters and sample labels are calculated using a preset loss function to obtain a loss value. This loss value is then used to update the initial deep learning model, ultimately resulting in a trained deep learning model. Specifically, the update process includes inputting preprocessed and aligned multimodal time series data into AMSTF-Net. The composite loss function is optimized using the backpropagation algorithm, iteratively updating all network weight parameters. Training objective: To enable the model to learn optimal control strategies from successful historical data and to learn how to mitigate risks promptly from failed data.

[0049] The technical solution of this invention detects the shoulder formation state of Czochralski single-crystal silicon during the shoulder formation process. In the event of abnormal shoulder formation, the current first shoulder formation parameter is obtained, and a pre-trained deep learning model is used to predict the shoulder formation process of Czochralski single-crystal silicon based on the first shoulder formation parameter. A second shoulder formation parameter that conforms to the normal shoulder formation process is generated, and the shoulder formation of Czochralski single-crystal silicon is performed based on the second shoulder formation parameter. This avoids the failure of shoulder formation of Czochralski single-crystal silicon and effectively improves the shoulder formation effect of Czochralski single-crystal silicon.

[0050] This invention also provides a shoulder-releasing control device 300 based on a deep learning model, such as... Figure 3 As shown, it includes: The acquisition module 310 is used to acquire the first shoulder formation parameter during the shoulder formation process of the Czochralski single crystal silicon when the shoulder formation is abnormal. The first shoulder formation parameter is used to represent the attribute data and environmental data of the Czochralski single crystal silicon during the shoulder formation process. The prediction module 320 is used to predict the shoulder formation process of the Czochralski single crystal silicon based on the first shoulder formation parameter using a pre-trained deep learning model, and obtain the second shoulder formation parameter. The shoulder-forming module 330 is used to perform shoulder-forming on the Czochralski single-crystal silicon based on the second shoulder-forming parameters.

[0051] In some embodiments, it also includes: The sample acquisition module is used to acquire multiple training datasets. Each training dataset includes a sample shoulder placement parameter and a sample label. The sample shoulder placement parameter is the actual shoulder placement parameter of the sample Czochralski single crystal silicon during the shoulder placement process, and the sample label is the shoulder placement parameter of the sample Czochralski single crystal silicon during the normal shoulder placement process. The training module is used to train the initial deep learning model based on the multiple training datasets to obtain the pre-trained deep learning model.

[0052] In some embodiments, the training dataset includes high-frequency data, mid-frequency data, and low-frequency data; The high-frequency data includes at least one of the following: melt temperature, lifting speed, crucible lifting speed, and heating power; The intermediate frequency data includes at least one of the following: crystal diameter morphology image and crystal edge morphology image; The low-frequency data includes historical shoulder data.

[0053] In some embodiments, the training module includes: The preprocessing submodule is used to preprocess the multiple training datasets to obtain multiple preprocessed training datasets. The preprocessing is used to align the timestamps of the multiple training datasets and normalize the units corresponding to the multiple training datasets to preset units. The encoding submodule is used to encode the high-frequency data based on the control variable encoder of the initial deep learning model to obtain a first encoding result, to encode the mid-frequency data based on the visual feature encoder of the initial deep learning model to obtain a second encoding result, and to encode the low-frequency data based on the historical knowledge encoder of the initial deep learning model to obtain a third encoding result. The fusion submodule is used to perform fusion processing on the first encoding result, the second encoding result, and the third encoding result based on the fusion module of the initial deep learning model to generate fused features; The identification submodule is used by the classification module based on the initial deep learning model to identify the fused features and output the predicted parameters. The update submodule is used to update the initial deep learning model based on the predicted shoulder parameters to obtain the pre-trained deep learning model.

[0054] In some embodiments, the classification module includes a regression submodule and a classification submodule, and the identification submodule includes: The first identification unit is used to identify the fused features based on the regression submodule of the initial deep learning model and output the predicted shoulder control parameters. The second identification unit is used to identify the fused features based on the classification submodule of the initial deep learning model and output a probability value. The predicted probability value is the probability value that the training dataset is a case of abnormal shoulder placement. The predicted shoulder placement parameters include the predicted shoulder placement control parameters and the predicted probability value.

[0055] In some embodiments, the update submodule includes: The calculation unit is used to perform loss calculation on the predicted shoulder parameters and the sample labels based on a preset loss function to obtain a loss value; An update unit is used to update the initial deep learning model based on the loss value to obtain the pre-trained deep learning model.

[0056] The technical solution of this invention detects the shoulder formation state of Czochralski single-crystal silicon during the shoulder formation process. In the event of abnormal shoulder formation, the current first shoulder formation parameter is obtained, and a pre-trained deep learning model is used to predict the shoulder formation process of Czochralski single-crystal silicon based on the first shoulder formation parameter. A second shoulder formation parameter that conforms to the normal shoulder formation process is generated, and the shoulder formation of Czochralski single-crystal silicon is performed based on the second shoulder formation parameter. This avoids the failure of shoulder formation of Czochralski single-crystal silicon and effectively improves the shoulder formation effect of Czochralski single-crystal silicon.

[0057] This invention also provides an electronic device. Please refer to [link to relevant documentation]. Figure 4 The electronic device may include a processor 401, a memory 402, and a program 4021 stored in the memory 402 and executable on the processor 401.

[0058] When program 4021 is executed by processor 401, it can achieve the following: Figure 1 Any step in the corresponding method embodiment can achieve the same technical effect, and will not be repeated here to avoid repetition.

[0059] This invention also provides a computer-readable storage medium storing a computer program. When executed by a processor, this computer program implements the various processes of the above-described embodiments of the deep learning model-based load control method and achieves the same technical effects. To avoid repetition, it will not be described again here. The computer-readable storage medium may be a read-only memory (ROM), a random access memory (RAM), a magnetic disk, or an optical disk.

[0060] This invention also provides a computer program product, which is stored in a storage medium and executed by at least one processor to implement the various processes of the above-described embodiments of the shoulder control method based on a deep learning model, and can achieve the same technical effect. To avoid repetition, it will not be described again here.

[0061] It should be noted that, in this document, the terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or apparatus that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such a process, method, article, or apparatus. Unless otherwise specified, an element defined by the phrase "comprising one..." does not exclude the presence of other identical elements in the process, method, article, or apparatus that includes that element.

[0062] Through the above description of the embodiments, those skilled in the art can clearly understand that the methods of the above embodiments can be implemented by means of software plus necessary general-purpose hardware platforms. Of course, they can also be implemented by hardware, but in many cases the former is a better implementation method. Based on this understanding, the technical solution of the present invention, or the part that contributes to the prior art, can be embodied in the form of a software product. This computer software product is stored in a storage medium (such as ROM / RAM, magnetic disk, optical disk) and includes several instructions to cause a terminal (which may be a mobile phone, computer, server, air conditioner, or network device, etc.) to execute the methods described in the various embodiments of the present invention.

[0063] The embodiments of the present invention have been described above with reference to the accompanying drawings. However, the present invention is not limited to the specific embodiments described above. The specific embodiments described above are merely illustrative and not restrictive. Those skilled in the art can make many other forms under the guidance of the present invention without departing from the spirit and scope of the claims, and all of these forms are within the protection scope of the present invention.

[0064] Through the above description of the embodiments, those skilled in the art can clearly understand that the methods of the above embodiments can be implemented by means of software plus necessary general-purpose hardware platforms. Of course, they can also be implemented by hardware, but in many cases the former is a better implementation method. Based on this understanding, the technical solution of the present invention, or the part that contributes to the prior art, can be embodied in the form of a software product. This computer software product is stored in a storage medium (such as ROM / RAM, magnetic disk, optical disk) and includes several instructions to cause a terminal (which may be a mobile phone, computer, server, air conditioner, or network device, etc.) to execute the methods described in the various embodiments of the present invention.

[0065] In the various method embodiments of this disclosure, the sequence numbers of each step are not intended to limit the order of the steps. For those skilled in the art, any changes in the order of the steps are within the scope of protection of this disclosure without any creative effort.

[0066] It should be noted that the various embodiments in this specification are described in a progressive manner, and the same or similar parts between the various embodiments can be referred to mutually. Each embodiment focuses on describing the differences from other embodiments. In particular, since the embodiments are basically similar to the product embodiments, the descriptions are relatively simple, and the relevant parts can be referred to the descriptions of the product embodiments.

[0067] The above description represents the preferred embodiments of this disclosure. It should be noted that those skilled in the art can make various improvements and modifications without departing from the principles described herein, and these improvements and modifications should also be considered within the scope of protection of this disclosure.

Claims

1. A shoulder-releasing control method based on a deep learning model, characterized in that, The method includes: In the event of abnormal shoulder formation during the Czochralski single-crystal silicon shoulder formation process, a first shoulder formation parameter is obtained during the shoulder formation process of the Czochralski single-crystal silicon. The first shoulder formation parameter is used to represent the attribute data and environmental data of the Czochralski single-crystal silicon during the shoulder formation process. The second shoulder formation parameter is obtained by predicting the shoulder formation process of the Czochralski single crystal silicon based on the first shoulder formation parameter using a pre-trained deep learning model. The Czochralski single-crystal silicon is subjected to shoulder formation based on the second shoulder formation parameter.

2. The method according to claim 1, characterized in that, Before obtaining the second shoulder formation parameter by predicting the shoulder formation process of the Czochralski-grown single-crystal silicon using a pre-trained deep learning model based on the first shoulder formation parameter, the method further includes: Multiple training datasets are obtained, each training dataset including a sample shoulder setting parameter and a sample label, wherein the sample shoulder setting parameter is the actual shoulder setting parameter of the sample Czochralski single crystal silicon during the shoulder setting process, and the sample label is the shoulder setting parameter of the sample Czochralski single crystal silicon during the normal shoulder setting process; The initial deep learning model is trained based on the multiple training datasets to obtain the pre-trained deep learning model.

3. The method according to claim 2, characterized in that, The training dataset includes high-frequency data, mid-frequency data, and low-frequency data; The high-frequency data includes at least one of the following: melt temperature, lifting speed, crucible lifting speed, and heating power; The intermediate frequency data includes at least one of the following: crystal diameter morphology image and crystal edge morphology image; The low-frequency data includes historical shoulder data.

4. The method according to claim 3, characterized in that, The step of training the initial deep learning model based on the multiple training datasets to obtain the pre-trained deep learning model includes: The multiple training datasets are preprocessed to obtain multiple preprocessed training datasets. The preprocessing is used to align the timestamps of the multiple training datasets and normalize the units corresponding to the multiple training datasets to preset units. The control variable encoder based on the initial deep learning model encodes the high-frequency data to obtain a first encoding result; the visual feature encoder based on the initial deep learning model encodes the mid-frequency data to obtain a second encoding result; and the historical knowledge encoder based on the initial deep learning model encodes the low-frequency data to obtain a third encoding result. The fusion module based on the initial deep learning model performs fusion processing on the first encoding result, the second encoding result, and the third encoding result to generate fused features; The classification module based on the initial deep learning model identifies the fused features and outputs predicted shoulder parameters. The initial deep learning model is updated based on the predicted shoulder parameters to obtain the pre-trained deep learning model.

5. The method according to claim 4, characterized in that, The classification module includes a regression submodule and a classification submodule. The classification module based on the initial deep learning model identifies the fused features and outputs predicted parameters, including: Based on the regression submodule of the initial deep learning model, the fused features are identified, and the predicted shoulder control parameters are output. Based on the classification submodule of the initial deep learning model, the fused features are identified and a probability value is output. The predicted probability value is the probability that the training dataset is a case of abnormal shoulder placement. The predicted shoulder placement parameters include the predicted shoulder placement control parameters and the predicted probability value.

6. The method according to claim 4, characterized in that, The step of updating the initial deep learning model based on the predicted shoulder parameters to obtain the pre-trained deep learning model includes: Based on a preset loss function, the loss is calculated using the predicted shoulder placement parameters and the sample labels to obtain the loss value; The initial deep learning model is updated based on the loss value to obtain the pre-trained deep learning model.

7. A shoulder-releasing control device based on a deep learning model, characterized in that, The device includes: The acquisition module is used to acquire the first shoulder formation parameter during the shoulder formation process of the Czochralski single crystal silicon when abnormal shoulder formation occurs. The first shoulder formation parameter is used to represent the attribute data and environmental data of the Czochralski single crystal silicon during the shoulder formation process. The prediction module is used to predict the shoulder formation process of the Czochralski single crystal silicon based on the first shoulder formation parameter using a pre-trained deep learning model, and obtain the second shoulder formation parameter. The shoulder-forming module is used to form a shoulder on the Czochralski single-crystal silicon based on the second shoulder-forming parameters.

8. An electronic device, characterized in that, include: A processor, a memory, and a program stored in the memory and executable on the processor, wherein the program, when executed by the processor, implements the steps of the method as described in any one of claims 1 to 6.

9. A computer-readable storage medium, characterized in that, The computer-readable storage medium stores a computer program that, when executed by a processor, implements the steps of the method as described in any one of claims 1 to 6.

10. A computer program product, characterized in that, Includes computer instructions that, when executed by a processor, implement the steps of the method as described in any one of claims 1 to 6.

Images