A data-driven seeding power determination method and related apparatus

Abstract

The application discloses a data-driven seeding power determination method and related device, and belongs to the field of Czochralski single crystal in photovoltaics. The disclosed method comprises the following steps: determining whether a last crystal pulling process has a shoulder breaking according to last crystal pulling process data; if the last crystal pulling process has no shoulder breaking, obtaining process parameters of a seeding process and a shoulder process in the last crystal pulling process and on-site parameters of the current crystal pulling, and determining a seeding power of the current crystal pulling by using a first seeding power model according to the obtained parameters; and if the last crystal pulling process has a shoulder breaking and there is no record of entering an equal-diameter process in the current furnace, obtaining process parameters of the seeding process in the last crystal pulling process and on-site parameters of the current crystal pulling, and determining the seeding power of the current crystal pulling by using a second seeding power model according to the obtained parameters. The application realizes automatic and intelligent determination of the seeding power based on data driving, improves the determination accuracy of the seeding power, and improves the single crystal production efficiency and reduces the breaking rate.

Description

TECHNICAL FIELD

[0001] The application belongs to the technical field of Czochralski single crystal technology in photovoltaics, and particularly relates to a seed crystal power determination method based on data driving and related devices. BACKGROUND

[0002] In a single crystal growth control system, the current industry method for determining seed crystal power is still in the exploratory stage, mainly being an artificial determination method, specifically, a field process engineer determines the seed crystal power of this time according to experience.

[0003] However, this determination method is prone to cause non-ideal and inaccurate seed crystal power setting due to human subjective factors and insufficient field experience, and a large number of front-line workers need to pay attention to and operate at all times, which will cause certain impact on single crystal production cost, yield and production efficiency, and the seed crystal power determination method based on the known technology will also cause a high disconnection rate in the crystal pulling process. SUMMARY

[0004] Therefore, the application provides a seed crystal power determination method based on data driving and related devices to solve the problems of non-ideal and inaccurate seed crystal power setting and high disconnection rate in the known technology.

[0005] The specific technical solutions are as follows:

[0006] A seed crystal power determination method based on data driving, comprising:

[0007] obtaining last crystal pulling process data;

[0008] determining whether a shoulder breaking disconnection occurs in the last crystal pulling process according to the last crystal pulling process data;

[0009] if the last crystal pulling process does not occur shoulder breaking disconnection, obtaining process parameters of a seed crystal pulling process, process parameters of a shoulder process and field parameters of this time of crystal pulling, and determining the seed crystal power of this time of crystal pulling by using a first seed crystal power model according to the process parameters of the seed crystal pulling process, the process parameters of the shoulder process and the field parameters of this time of crystal pulling;

[0010] if the last crystal pulling process occurs shoulder breaking disconnection and there is no record of entering a constant diameter process in this furnace time, obtaining process parameters of a seed crystal pulling process and field parameters of this time of crystal pulling, and determining the seed crystal power of this time of crystal pulling by using a second seed crystal power model according to the process parameters of the seed crystal pulling process and the field parameters of this time of crystal pulling;

[0011] The second seed crystal power model and the first seed crystal power model are different agents trained by performing feature selection and feature standard interval range solving on historical process parameters in a crystal pulling process to construct a data set for different scenarios of whether a last crystal pulling process has a shoulder breaking.

[0012] Optionally, the process parameters of the seed crystal pulling process include at least part of the parameters of seed crystal pulling power, seed crystal average pulling speed, seed crystal late stage pulling speed, seed crystal pulling time, seed crystal pulling temperature, argon flow, crystal rotation speed, crucible rotation speed, re-feeding times, and seed crystal remaining material amount.

[0013] The process parameters of the shoulder process include at least part of the parameters of last shoulder time, shoulder height, shoulder early stage height, shoulder middle stage height, shoulder opening time, argon flow, crystal rotation speed, and crucible rotation speed.

[0014] The on-site parameters of the current crystal pulling include at least part of the parameters of current crystal pulling remaining material amount, liquid surface temperature, re-feeding times, thermal field size, and single crystal size.

[0015] Optionally, the construction process of the first seed crystal power model includes:

[0016] Obtaining source data of each process in the crystal pulling preparation process from historical data of the crystal pulling preparation process;

[0017] Performing feature selection processing matched with a first crystal pulling scenario on the source data of each process to obtain each first feature matched with the first crystal pulling scenario and satisfying a first correlation condition with seed crystal pulling power; the first crystal pulling scenario is a scenario in which a last crystal pulling process does not have shoulder breaking;

[0018] Solving a standard interval range of each first feature according to process data from seed crystal pulling to equal diameter by equal diameter performance screening; the process data includes data of each first feature;

[0019] According to the standard interval range of each first feature, performing data selection on each first feature, combining selected feature data and a first seed crystal pulling power corresponding to the feature data in historical data to obtain a first data set;

[0020] Performing machine learning model training based on the first data set to obtain the first seed crystal pulling power model.

[0021] Optionally, the performing feature selection processing matched with a first crystal pulling scenario on the source data of each process to obtain each first feature matched with the first crystal pulling scenario and satisfying a first correlation condition with seed crystal pulling power includes:

[0022] determine the correlation between different process parameters in the source data of the respective process;

[0023] eliminate part of the process parameters satisfying the preset correlation relationship to obtain each reserved parameter;

[0024] determine the importance of each reserved parameter relative to the seeding power using a gradient boosting decision tree model, and select each reserved parameter corresponding to the importance satisfying the importance condition and matching the first crystal pulling scene as the respective first feature.

[0025] Optionally, the construction process of the second seeding power model comprises:

[0026] obtain source data of each process in the crystal pulling preparation process from historical data of the crystal pulling preparation process;

[0027] perform feature selection processing matching the second crystal pulling scene on the source data of each process to obtain each second feature matching the second crystal pulling scene and satisfying a second correlation condition with the seeding power based on the feature selection processing; the second crystal pulling scene is a scene in which the last crystal pulling process has a record of shoulder breakage and the current furnace has no record of entering the constant diameter process;

[0028] solve the standard interval range of each second feature by filtering process data from seeding to constant diameter according to the constant diameter performance; the process data includes data of each second feature;

[0029] perform data selection on each second feature according to the standard interval range of each second feature, combine the selected feature data with the corresponding second seeding power of the feature data in the historical data to obtain a second data set;

[0030] train a machine learning model based on the second data set to obtain the second seeding power model.

[0031] Optionally, the performing feature selection processing matching the second crystal pulling scene on the source data of each process to obtain each second feature matching the second crystal pulling scene and satisfying a second correlation condition with the seeding power comprises:

[0032] determine the correlation between different process parameters in the source data of the respective process;

[0033] eliminate part of the process parameters satisfying the preset correlation relationship to obtain each reserved parameter;

[0034] The gradient boosting decision tree model is used to determine the importance of each reserved parameter relative to the seeding power, and each reserved parameter corresponding to the importance satisfying the importance condition and matching the second crystal pulling scene is selected as the respective second feature.

[0035] Optionally, the standard interval range of the corresponding feature is solved by screening the process data of seeding to the constant diameter according to the constant diameter performance, including:

[0036] The process data of seeding to the constant diameter each time is screened according to the constant diameter performance;

[0037] The standard interval of the corresponding feature is solved according to the statistical information of the data in the screened process data of seeding to the constant diameter;

[0038] The corresponding feature is a corresponding first feature or a second feature corresponding to the process data.

[0039] Optionally, the machine learning model is trained based on the corresponding data set, including:

[0040] The gradient boosting decision tree model is used to determine the importance of each reserved parameter relative to the seeding power, and each reserved parameter corresponding to the importance satisfying the importance condition and matching the second crystal pulling scene is selected as the respective second feature.

[0041] The corresponding data set is a first data set, and the corresponding seeding power model is a first seeding power model, or the corresponding data set is a second data set, and the corresponding seeding power model is a second seeding power model.

[0042] A data-driven seeding power determination device, comprising:

[0043] An acquisition module is configured to acquire last time crystal pulling process data.

[0044] A first determination module is configured to determine whether a shoulder breaking occurs in the last time crystal pulling process according to the last time crystal pulling process data.

[0045] A second determination module is configured to acquire process parameters of a seeding process, process parameters of a shoulder forming process and on-site parameters of the current crystal pulling in the last time crystal pulling process, and determine the seeding power of the current crystal pulling by using a first seeding power model according to the process parameters of the seeding process, the process parameters of the shoulder forming process and the on-site parameters of the current crystal pulling in the last time crystal pulling process.

[0046] The third determining module is used to obtain the process parameters of the crystal pulling process in the previous crystal pulling process and the field parameters of the current crystal pulling process when there is a record of a shoulder breakage during the previous crystal pulling process and no equal diameter process is entered in the current furnace. The second crystal pulling power model is used to determine the crystal pulling power of the current crystal pulling process based on the process parameters of the crystal pulling process in the previous crystal pulling process and the field parameters of the current crystal pulling process.

[0047] The second crystal pulling power model and the first crystal pulling power model are different intelligent agents trained by constructing datasets for different scenarios in the crystal pulling process, such as whether the previous crystal pulling process resulted in shoulder breakage. These models are constructed by selecting features that match the target scenario and solving the standard range of the features based on the process parameters in the historical crystal pulling process.

[0048] A computer-readable medium having a computer program stored thereon, which, when executed by a processor, can be used to implement the data-driven method for determining the lead-in power as described in any of the preceding claims.

[0049] As can be seen from the above scheme, the data-driven crystal pulling power determination method and related device provided in this application, for different scenarios in the crystal pulling process, whether the previous crystal pulling caused shoulder breakage, constructs a dataset by selecting features that match the target scenario and solving the standard range of the features based on the process parameters in the historical crystal pulling process, and then trains different agents to obtain the first crystal pulling power model and the second crystal pulling power model adapted to different scenarios. Based on this, by acquiring data from the previous crystal pulling process, it is determined whether a shoulder break occurred during the previous crystal pulling process. Then, based on the determination result, a crystal pulling power model adapted to the current actual scenario (whether a shoulder break actually occurred or not) is invoked. Using the invoked model and its corresponding input parameters, the crystal pulling power for the current crystal pulling is determined. This achieves automated and intelligent determination of crystal pulling power based on data-driven methods, adapting to different breakage situations / scenarios. It effectively overcomes the problems of inaccurate and unsatisfactory crystal pulling power settings caused by subjective human factors and insufficient on-site experience in manual determination methods. Furthermore, it eliminates the need for a large number of frontline workers to constantly monitor and operate the system, thus not affecting production costs, output, or efficiency. This improves the accuracy of crystal pulling power determination, increases single-crystal production efficiency, and reduces the breakage rate. Attached Figure Description

[0050] For ease of reference or clarity, the relevant abbreviations or terms used in the embodiments of this application are explained as follows:

[0051] Crystal pulling: This is a front-end process in the photovoltaic field, specifically the process of producing silicon rods from silicon material using the Czochralski method in a single crystal furnace.

[0052] Seed crystal: the original 16mm seed crystal is gradually reduced in size and stabilized at a certain uniformity of fine crystal.

[0053] Seed crystal: a single crystal with a certain crystal orientation, used as a replication sample to make the drawn silicon ingot have the same crystal orientation as the seed crystal.

[0054] GBDT model: GBDT stands for Gradient Boosting Decision Tree, which is a gradient boosting algorithm. GBDT optimizes the loss function using Taylor series expansion and negative gradient direction, and iteratively updates the model parameters to approximate the optimal solution of the loss function.

[0055] Constant diameter performance: the performance of whether the constant diameter length reaches the expected value.

[0056] In order to more clearly illustrate the technical solutions in the embodiments of the present application or prior art, the following will briefly introduce the drawings needed in the embodiment or prior art description. Obviously, the drawings in the following description are only embodiments of the present application, and those skilled in the art can obtain other drawings according to the provided drawings without creative labor.

[0057] Figure 1 is the construction process schematic diagram of the first seed crystal power model provided by the present application;

[0058] Figure 2 is the construction process schematic diagram of the second seed crystal power model provided by the present application;

[0059] Figure 3 is the flow chart of the data-driven seed crystal power determination method provided by the present application;

[0060] Figure 4 is the component structure diagram of the data-driven seed crystal power determination device provided by the present application;

[0061] Figure 5 is the component structure diagram of the electronic device provided by the present application. DETAILED DESCRIPTION

[0062] 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 part 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 scope of protection of the present application.

[0063] The embodiment of the present application provides a data-driven seed crystal power determination method and a related device, which is used for constructing a seed crystal power model of a seed crystal power process of a single crystal silicon prepared by a Czochralski method based on a machine learning method, to realize data-driven automatic and intelligent prediction of the seed crystal power, so as to solve the problems of non-ideal and inaccurate setting of the seed crystal power and high wire breakage rate in the prior art.

[0064] Further, the embodiment of the present application specifically uses a data-driven method to select parameters in a seed crystal process and a shoulder process and the like as model input features from historical data by using feature selection or even combining artificial experience, and performs data screening on the selected feature parameters by solving a standard interval range of each selected feature parameter, to finally construct a data set, and then trains a model by using a machine learning gradient boosting decision tree model, to obtain a seed crystal power model, and then inputs parameters in the last seed crystal process and shoulder process and the like into the trained seed crystal power model, to predict the seed crystal power in the current crystal pulling.

[0065] The data-driven seed crystal power determination method provided by the embodiment of the present application needs to be based on pre-construction of a seed crystal power model, and the construction process of the seed crystal power model is first described below.

[0066] The seed crystal power model constructed by the embodiment of the present application includes a first seed crystal power model and a second seed crystal power model, and the second seed crystal power model and the first seed crystal power model are different agents trained by performing feature selection and feature standard interval range solving on process parameters in a historical crystal pulling process to construct a data set, for different scenarios of whether shoulder wire breakage occurs in the last crystal pulling.

[0067] Specifically, in an actual production process, wire breakage may occur in each process of the crystal pulling preparation process, different scenarios of whether shoulder wire breakage occurs are regarded as different scenarios, and different seed crystal power models are designed and constructed for different scenarios, wherein if shoulder wire breakage does not occur in the last crystal pulling process (no shoulder wire breakage means entering a constant diameter process), the first seed crystal power model is used for intelligent prediction of the seed crystal power of this time, and if shoulder wire breakage occurs in the last crystal pulling process and there is no record of entering the constant diameter process in this furnace, the second seed crystal power model is used for intelligent prediction of the seed crystal power of this time.

[0068] Referring to Figure 1 The construction process of the first seed crystal power model can be implemented as the following steps 101 to 105, and the processing steps are described in detail below.

[0069] Step 101, obtaining source data of each process in the crystal pulling preparation process from historical data of the crystal pulling preparation process.

[0070] Optionally, each process in the crystal pulling preparation process can include seeding, shoulder opening, shoulder turning, and constant diameter.

[0071] The source data of the seeding process can include but is not limited to: crystal length, crystal diameter, growth pulling speed, heater power, remaining material weight, crucible lifting speed, crystal lifting speed, liquid surface temperature, argon flow rate, re-feeding times, absolute meniscus distance.

[0072] The source data of the shoulder opening process can include but is not limited to: crystal length, crystal diameter, growth pulling speed, heater power, remaining material weight, crucible lifting speed, crystal lifting speed, liquid surface temperature, argon flow rate, re-feeding times, absolute meniscus distance, shoulder opening time length, shoulder opening cooling amount, shoulder opening opening time length.

[0073] The source data of the shoulder turning process can include but is not limited to: crystal length, crystal diameter, growth pulling speed, heater power, remaining material weight, crucible lifting speed, crystal lifting speed, liquid surface temperature, argon flow rate, re-feeding times, absolute meniscus distance, shoulder turning cooling amount.

[0074] The source data of the constant diameter process can include but is not limited to: crystal length, crystal diameter, growth pulling speed, heater power, remaining material weight, crucible lifting speed, crystal lifting speed, liquid surface temperature, argon flow rate, re-feeding times, absolute meniscus distance, process running time length.

[0075] Step 102, performing feature selection processing matched with a first crystal pulling scenario on the source data of each process, to obtain each first feature matched with the first crystal pulling scenario and satisfying a first correlation condition with the seeding power based on the feature selection processing.

[0076] The first crystal pulling scenario is a scenario in which the last crystal pulling process does not occur shoulder breaking and enters the constant diameter process.

[0077] The feature selection processing of this step aims to select features from the obtained process parameter data (source data) that are representative and can improve the performance of the model (i.e., affect the seeding power) for the first crystal pulling scenario. Specifically, each parameter matched with the first crystal pulling scenario and satisfying a first correlation condition with the seeding power can be selected as each first feature by correlation between the obtained process parameters or statistical methods, to reduce the dimension of the parameter data.

[0078] Optionally, step 102 can be further implemented as steps 1-1 to 1-3 as follows, each of which is described in detail as follows:

[0079] 1-1, determine the correlation between different process parameters in the source data of each process.

[0080] Optionally, a correlation coefficient matrix between process parameters can be calculated, and it is determined through the correlation coefficient matrix between process parameters which process parameters have a higher correlation.

[0081] 1-2, remove part of the parameters in the plurality of process parameters satisfying the preset correlation relationship, and obtain each reserved parameter.

[0082] Optionally, the preset correlation relationship can be set as the correlation coefficient between process parameters being close to 1 or -1, i.e. the absolute value of the difference between the correlation coefficient between process parameters and 1 or -1 is less than a set threshold.

[0083] If the correlation between two or more process parameters is too high to satisfy the correlation relationship, for example, the correlation coefficient is close to 1 or -1, part of the process parameters can be removed, thereby obtaining each reserved parameter to reduce redundancy. For example, if the correlation coefficient between two process parameters is close to 1 or -1, one of the parameters can be removed.

[0084] 1-3, determine the importance of each reserved parameter relative to the seed power by using a gradient boosting decision tree model, and select each reserved parameter corresponding to the importance satisfying the importance condition and matching the first crystal pulling scene as each first feature.

[0085] On the basis of step 1-2, the importance of each reserved parameter relative to the seed power can be further evaluated by using a machine learning GBDT model, and an importance evaluation result of each reserved parameter relative to the seed power is output, and then each reserved parameter corresponding to the importance satisfying the importance condition and matching the first crystal pulling scene is selected as each first feature.

[0086] The importance condition can be but not limited to set as the importance evaluation value reaching a threshold value, or the importance evaluation value being top_k in a descending sequence of importance evaluation values corresponding to each parameter, so as to remove parameters with low importance relative to the seed power among each reserved parameter by setting the importance condition, and the feature selection can also be combined with the experience of on-site process. Wherein, k is an integer greater than 0.

[0087] The finally selected each first feature can include but is not limited to: seed power, average seed pulling speed, seed late stage pulling speed, seed time length, seed temperature, argon flow, crystal rotation speed, crucible rotation speed, re-throwing number, seed remaining amount, shoulder length, shoulder height, shoulder early stage height, shoulder middle stage height, shoulder opening time length argon flow, crystal rotation speed, crucible rotation speed, remaining amount of this time pulling, liquid level temperature, re-throwing number, heat field size, single crystal size.

[0088] Step 103, solving the standard interval range of each first feature by screening the process data of seeding to isometric according to isometric performance; the process data includes the data of each first feature.

[0089] Since the historical data is generated by on-site manual setting, on-site process personnel may cause data errors due to subjective factors, lack of experience, etc. For this situation, the embodiment of the application first processes the selected first features before making the data set to determine the standard interval range of the parameter value of each first feature.

[0090] Specifically, for each first feature, the process data of seeding to isometric corresponding to each first feature (i.e. the process data includes the data of each first feature) can be screened according to isometric performance first, and then the screened process data of seeding to isometric is statistically processed. According to the statistical information of the data in the screened process data of seeding to isometric, the standard interval of the first feature is solved, and specifically, the value interval range of the first feature that meets the preset statistical feature can be taken as the standard interval range of the parameter value of the first feature.

[0091] It is worth noting that the historical data and the solved standard data range under different bases and different equipment may not be consistent, based on which the embodiment of the application does not give the specific standard interval range of each first feature (i.e. the process parameter selected as the first feature). In actual application, as long as the above technical concept of the embodiment of the application is used to determine the standard interval range of the feature parameter value, it belongs to the protection scope of the embodiment of the application.

[0092] Step 104, selecting data according to each first feature according to the standard interval range of each first feature, combining the selected feature data with the first seeding power corresponding to the feature data in the historical data to obtain a first data set.

[0093] On the basis of step 103, the present step can particularly perform data screening on each first feature according to the standard interval range of each first feature solved. First, assuming that each adjacent two times of seeding in the historical data are referred to as the last time of seeding and the present time of seeding, respectively, the record of the last time of seeding in the historical data which does not occur breakage into the equal-diameter process and the record of the equal-diameter performance in the present time of seeding which meets the expectation can be particularly screened, second, the parameters of the present time of seeding power, the present time of seeding average pulling speed, the present time of seeding late-stage pulling speed, the present time of seeding shoulder time length, the present time of seeding shoulder height and the like are screened according to the standard interval range of each first feature solved. Finally, the selected last time of seeding feature data, the present time of seeding feature data screened according to the standard interval range, and the seeding power of the present time of seeding are combined to form a data set for model training.

[0094] Here, the last time and the present time refer to the last time of seeding and the present time of seeding, respectively.

[0095] In the present step, the concept of feature data screening is that whether the parameters in the process of the last time of seeding and the shoulder process are in the standard range or not, as long as the data in the process of the present time of seeding and the shoulder process is in the standard range, the goal of model learning is to learn and predict the present time of seeding power according to different last time of seeding, so that the data of the seeding process and the subsequent process under the seeding power is in the standard range.

[0096] Step 105, performing machine learning model training based on the first data set to obtain the first seeding power model.

[0097] Specifically, the first data set can be used to train multiple decision trees in the gradient boosting decision tree model, each decision tree improves the model performance by reducing the residual error of the previous decision tree until the end condition is met, and the first seeding power model is obtained.

[0098] That is, in the present application, when the first data set is constructed to train the model to obtain the first seeding power model, the machine learning gradient boosting decision tree model (GBDT) is specifically used for model training. Through the applicant's research and experimental verification, compared with other machine learning models, the GBDT model can train multiple decision trees step by step, and each tree can reduce the residual error of the previous tree to better capture the complex relationship between a large number of different parameter data in different pulling processes and multiple processes in the straight pulling single crystal. Correspondingly, compared with other machine learning models, the present application can achieve better performance in the prediction of the seeding power by using the machine learning gradient boosting decision tree model (GBDT) for model training, and realize the accurate prediction of the seeding power.

[0099] The end condition can be, but is not limited to, being set as the model performance of the first seed crystal power model reaching a set performance, or the number of iterations of model training reaching a set number, or the model training duration reaching a set duration.

[0100] The embodiment can support subsequent data-driven automatic and intelligent prediction of the seed crystal power for the first crystal pulling scenario in which no shoulder breaking occurs in the last crystal pulling process, to overcome the problem of suboptimal and inaccurate seed crystal power setting caused by human subjective factors and insufficient field experience in the manual determination method, and to improve the crystal pulling production efficiency and reduce the breaking rate.

[0101] Referring to Figure 2 The construction process of the second seed crystal power model can be implemented as steps 201 to 205, which will be described in detail below.

[0102] Step 201, obtaining source data of each process in the crystal pulling preparation process from historical data of the crystal pulling preparation process.

[0103] Specifically, the source data of the seed crystal pulling, shoulder breaking, shoulder turning, and equal-diameter processes in the crystal pulling preparation process can be obtained from the historical data.

[0104] The obtained source data of each process can refer to the source data of each process obtained when constructing the first seed crystal power model, which will not be described here.

[0105] Step 202, performing feature selection processing matched with the second crystal pulling scenario on the source data of each process, to obtain each second feature matched with the second crystal pulling scenario and satisfying a second correlation condition with the seed crystal power based on the feature selection processing.

[0106] The second crystal pulling scenario is a scenario in which shoulder breaking occurs in the last crystal pulling process and there is no record of entering the equal-diameter process in the current furnace cycle.

[0107] The feature selection processing of this step aims to select features that are representative and can improve the model performance (i.e., affect the seed crystal power) from the obtained parameter data (source data) of each process for the second crystal pulling scenario. Specifically, each parameter matched with the second crystal pulling scenario and satisfying a second correlation condition with the seed crystal power can be selected as each second feature by correlation between the obtained process parameters or statistical methods, to reduce the dimension of the parameter data.

[0108] Similar to the feature selection step in the first seed power model construction process, this step can specifically determine the correlation between different process parameters in the source data of each process, eliminate part of the parameters in the multiple process parameters that meet the preset correlation relationship, obtain each reserved parameter, and determine the importance of each reserved parameter relative to the seed power using the gradient boosting decision tree model, select each reserved parameter corresponding to the importance that meets the importance condition and matches the second crystal pulling scene as the second feature. At the same time, the feature selection can also be combined with the experience of the field process.

[0109] The finally selected second features can include but are not limited to: seed power, average seed pulling speed, late seed pulling speed, seed time, seed temperature, argon flow, crystal rotation speed, crucible rotation speed, re-throwing number, seed remaining amount, remaining amount of this time, liquid surface temperature, re-throwing number, thermal field size, single crystal size.

[0110] Compared with each first feature, the second feature does not include the process parameters of the shoulder process such as shoulder time and shoulder height.

[0111] Step 203, solving the standard interval range of each second feature by screening the process data from seed pulling to constant diameter according to the constant diameter performance; the process data includes the data of each second feature.

[0112] This step can specifically screen the process data corresponding to each second feature from seed pulling to constant diameter according to the constant diameter performance for each second feature (i.e. the process data includes the data of each second feature), and statistically analyze the screened process data from seed pulling to constant diameter. According to the statistical information of the data in the screened process data from seed pulling to constant diameter, the standard interval of the second feature is solved, and specifically the value interval range of the second feature that meets the preset statistical feature can be taken as the standard interval range of the parameter value of the second feature.

[0113] Step 204, selecting data according to the standard interval range of each second feature, combining the selected feature data with the second seed power corresponding to the feature data in the historical data to obtain a second data set.

[0114] On the basis of step 203, this step can specifically perform data screening on each second feature according to the standard interval range of each second feature to be solved. First, assuming that the in-crystal pulling process before and after every two adjacent times in the historical data are respectively referred to as the last in-crystal pulling and the current in-crystal pulling in the historical data, the records of the last in-crystal pulling that occurs broken shoulder and does not enter the equal-diameter process and the records of the current in-crystal pulling that meets the expected equal-diameter performance can be specifically screened, second, the in-crystal pulling power, the average pulling speed of the current in-crystal pulling, the average pulling speed of the late stage of the current in-crystal pulling and other feature parameter data are screened according to the standard interval range of each second feature to be solved. Finally, the selected last feature data, the current feature data screened according to the standard interval range, and the in-crystal pulling power of the current in-crystal pulling are combined to form a data set for model training.

[0115] The concept of the above feature data screening is that, regardless of whether the feature parameters in the last in-crystal pulling and the shoulder process are within the standard range, as long as the data in the current in-crystal pulling and the shoulder process is within the standard range, the goal of model learning is to learn and predict the current in-crystal pulling power according to different last conditions, so that the in-crystal pulling process and subsequent process data under the in-crystal pulling power are within the standard range.

[0116] Step 205, performing machine learning model training based on the second data set to obtain the second in-crystal pulling power model.

[0117] Similarly, this step can specifically train multiple decision trees in the gradient boosting decision tree model using the second data set, each decision tree improves the model performance by reducing the residual error of the previous decision tree until the end condition is met, and the second in-crystal pulling power model is obtained.

[0118] That is, when performing model training based on the constructed second data set to obtain the second in-crystal pulling power model, the present application also adopts the machine learning gradient boosting decision tree model (GBDT) to perform model training, to gradually train multiple decision trees based on the GBDT model, and to make each tree reduce the residual error of the previous tree, so as to better capture the complex relationship between a large number of various parameter data involved in different crystal pulling processes and multiple processes in the direct single crystal pulling, and to achieve better performance in in-crystal pulling power prediction, and to realize accurate prediction of the in-crystal pulling power.

[0119] The end condition here can be but is not limited to being set as: the model performance of the second in-crystal pulling power model reaches a set performance, or the number of iterations of model training reaches a set number, or the duration of model training reaches a set duration.

[0120] The second seed crystal pulling power model is constructed, so that the second seed crystal pulling scenario of the last seed crystal pulling process with shoulder breaking and no entering the equal diameter process in the current furnace is supported, data-driven seed crystal pulling power automation and intelligent determination is performed, the problems of the non-ideal and inaccurate seed crystal pulling power setting caused by the subjective factors and insufficient field experience in the manual determination are overcome, and the seed crystal pulling production efficiency is improved and the breaking rate is reduced.

[0121] Based on the constructed first seed crystal pulling power model and the second seed crystal pulling power model, referring to the data-driven seed crystal pulling power determination method flowchart shown in FIG. 6, the data-driven seed crystal pulling power determination method provided by the embodiment of the present application can include the following processing steps: Figure 3

[0122] Step 301, obtaining last seed crystal pulling process data.

[0123] Specifically, the process parameter data of each process of the seed crystal pulling process can be obtained.

[0124] Step 302, determining whether the last seed crystal pulling process has shoulder breaking according to the last seed crystal pulling process data.

[0125] Specifically, the scenario to which the seed crystal pulling power prediction of the current seed crystal pulling belongs can be determined by determining whether the last seed crystal pulling process has shoulder breaking according to the obtained last seed crystal pulling process data. If it is determined that the last seed crystal pulling process has not shoulder breaking and enters the equal diameter process, the scenario to which the seed crystal pulling power prediction of the current seed crystal pulling belongs is the first seed crystal pulling scenario. If it is determined that the last seed crystal pulling process has shoulder breaking and there is no record of entering the equal diameter process in the current furnace, the scenario to which the seed crystal pulling power prediction of the current seed crystal pulling belongs is the second seed crystal pulling scenario.

[0126] Step 303, if the last seed crystal pulling process has not shoulder breaking, obtaining the process parameters of the seed crystal pulling process, the process parameters of the shoulder breaking process, and the field parameters of the current seed crystal pulling, and determining the seed crystal pulling power of the current seed crystal pulling by using the first seed crystal pulling power model according to the process parameters of the seed crystal pulling process, the process parameters of the shoulder breaking process, and the field parameters of the current seed crystal pulling.

[0127] If the last seed crystal pulling process has not shoulder breaking, that is, the scenario to which the seed crystal pulling power prediction of the current seed crystal pulling belongs is the first seed crystal pulling scenario, for this case, the model input data matched with the first seed crystal pulling scenario can be specifically obtained. The obtained model input data includes the process parameters of the seed crystal pulling process, the process parameters of the shoulder breaking process, and the field parameters of the current seed crystal pulling.

[0128] ​The process parameters of the seed crystal introduction process in the last crystal pulling process include, but are not limited to, at least part of parameters such as a seed crystal introduction power, an average seed crystal introduction speed, a seed crystal introduction late stage speed, a seed crystal introduction time length, a seed crystal introduction temperature, an argon flow rate, a crystal rotation speed, a crucible rotation speed, a re-feeding number and a seed crystal introduction residual amount. The process parameters of the shoulder forming process in the last crystal pulling process include, but are not limited to, at least part of parameters such as a last shoulder forming time length, a shoulder forming height, a shoulder forming early stage height, a shoulder forming middle stage height, a shoulder forming opening time length, an argon flow rate, a crystal rotation speed and a crucible rotation speed. The on-site parameters of the current crystal pulling include, but are not limited to, at least part of parameters such as a current crystal pulling residual amount, a liquid surface temperature, a re-feeding number, a thermal field size and a single crystal size.

[0129] Then, the obtained parameters, i.e., the process parameters of the seed crystal introduction process in the last crystal pulling process and the on-site parameters of the current crystal pulling, are input into the first seed crystal introduction power model, so that the first seed crystal introduction power model predicts and outputs the seed crystal introduction power of the current crystal pulling based on the input parameters.

[0130] In step 304, if the last crystal pulling process has the shoulder forming broken wire and there is no record of entering the constant diameter process in the current furnace cycle, the process parameters of the seed crystal introduction process in the last crystal pulling process and the on-site parameters of the current crystal pulling are obtained, and the second seed crystal introduction power model is used to determine the seed crystal introduction power of the current crystal pulling according to the process parameters of the seed crystal introduction process in the last crystal pulling process and the on-site parameters of the current crystal pulling.

[0131] The second seed crystal introduction power model and the first seed crystal introduction power model are different intelligent agents trained by performing feature selection and feature standard interval range solving on the process parameters in the historical crystal pulling process to construct a data set for different scenarios of whether the last crystal pulling has the shoulder forming broken wire in the crystal pulling preparation process.

[0132] If the last crystal pulling process has the shoulder forming broken wire and there is no record of entering the constant diameter process in the current furnace cycle, i.e., the scenario to which the seed crystal introduction power prediction of the current crystal pulling belongs is the second crystal pulling scenario, for this case, the model input data matching the second crystal pulling scenario can be obtained. The obtained model input data includes the process parameters of the seed crystal introduction process in the last crystal pulling process and the on-site parameters of the current crystal pulling.

[0133] The process parameters of the seed crystal introduction process in the last crystal pulling process and the on-site parameters of the current crystal pulling can be specifically referred to the above description, which will not be repeated here.

[0134] Then, the obtained parameters, i.e., the process parameters of the seed crystal introduction process in the last crystal pulling process and the on-site parameters of the current crystal pulling, are input into the second seed crystal introduction power model, so that the second seed crystal introduction power model predicts and outputs the seed crystal introduction power of the current crystal pulling based on the input parameters.

[0135] According to the above scheme, the data-driven seed power determination method provided by the application can be known, for different scenarios of whether the last crystal pulling process occurs shoulder breaking, the data set is constructed by selecting the characteristics of the process parameters in the historical crystal pulling process and solving the standard interval range of the characteristics to train different agents, and the first seed power model and the second seed power model respectively adapted to different scenarios are obtained. On this basis, by obtaining the last crystal pulling process data, it is determined whether the last crystal pulling process occurs shoulder breaking according to the obtained data, and the seed power model adapted to the current actual scenario (whether the actual shoulder breaking) is called according to the determination result, and the model called is used to determine the seed power of the current crystal pulling based on the adaptive model input parameters, realizing the automatic and intelligent determination of the seed power based on data driving adapted to different breaking situations / scenarios. It can effectively overcome the problem of inaccurate seed power setting caused by human subjective factors, insufficient field experience and other reasons in manual determination method, and does not need a large number of front-line workers to pay attention to and operate at all times, which will not affect the production cost, yield and production efficiency, thereby improving the determination accuracy of the seed power, and improving the single crystal production efficiency and reducing the breaking rate.

[0136] Corresponding to the above method, the embodiment of the application also provides a data-driven seed power determination device, see Figure 4 the component structure diagram, the device comprises:

[0137] The acquisition module 401 is used for acquiring the last crystal pulling process data;

[0138] The first determination module 402 is used for determining whether the last crystal pulling process occurs shoulder breaking according to the last crystal pulling process data;

[0139] The second determination module 403 is used for acquiring the process parameters of the seed pulling process, the process parameters of the shoulder process and the field parameters of the current crystal pulling in the last crystal pulling process under the condition that the last crystal pulling process does not occur shoulder breaking, and determining the seed power of the current crystal pulling by the first seed power model according to the process parameters of the seed pulling process, the process parameters of the shoulder process and the field parameters of the current crystal pulling in the last crystal pulling process;

[0140] The third determination module 404 is used for acquiring the process parameters of the seed pulling process and the field parameters of the current crystal pulling in the last crystal pulling process under the condition that the last crystal pulling process occurs shoulder breaking and there is no record of entering the constant diameter process in the current furnace, and determining the seed power of the current crystal pulling by the second seed power model according to the process parameters of the seed pulling process and the field parameters of the current crystal pulling in the last crystal pulling process;

[0141] The second seed crystal power model and the first seed crystal power model are different agents trained by performing feature selection and feature standard interval range solving on historical process parameters in a crystal pulling process to construct a data set for different scenarios of whether the last crystal pulling process occurred shoulder breakage.

[0142] In an optional embodiment, the process parameters of the seed crystal pulling process include at least part of the parameters of the seed crystal pulling power, the average seed crystal pulling speed, the late seed crystal pulling speed, the seed crystal pulling time, the seed crystal pulling temperature, the argon flow, the crystal rotation speed, the crucible rotation speed, the re-feeding number, and the remaining material amount in the last crystal pulling process.

[0143] The process parameters of the shoulder process include at least part of the parameters of the last shoulder time, the shoulder height, the pre-shoulder height, the mid-shoulder height, the shoulder opening time, the argon flow, the crystal rotation speed, and the crucible rotation speed.

[0144] The on-site parameters of the current crystal pulling include at least part of the parameters of the remaining material amount, the liquid surface temperature, the re-feeding number, the thermal field size, and the single crystal size in the current crystal pulling.

[0145] In an optional embodiment, the above device further includes a model construction module for constructing a seed crystal power model; when constructing the first seed crystal power model, the model construction module is specifically configured to:

[0146] Obtain source data of each process in the crystal pulling preparation process from historical data of the crystal pulling preparation process;

[0147] Perform feature selection processing matched with the first crystal pulling scenario on the source data of each process, to obtain each first feature matched with the first crystal pulling scenario and satisfying a first correlation condition with the seed crystal power; the first crystal pulling scenario is a scenario in which the last crystal pulling process does not occur shoulder breakage;

[0148] Solve the standard interval range of each first feature according to the equal-diameter performance screening of the process data from seed crystal pulling to equal diameter; the process data includes data of each first feature;

[0149] According to the standard interval range of each first feature, perform data selection on each first feature, combine the selected feature data with the first seed crystal power corresponding to the feature data in the historical data to obtain a first data set;

[0150] Perform machine learning model training based on the first data set to obtain the first seed crystal power model.

[0151] In an optional implementation, the model construction module, when performing the feature selection processing on the source data of the respective processes to obtain the respective first features that match the first crystal pulling scenario and satisfy the first correlation condition with the seed power, is specifically configured to:

[0152] determine the correlation between different process parameters in the source data of the respective processes;

[0153] eliminate part of the process parameters that satisfy the preset correlation relationship to obtain respective remaining parameters;

[0154] determine the importance of each remaining parameter with respect to the seed power by using a gradient boosting decision tree model, and select respective remaining parameters that satisfy the importance condition and match the first crystal pulling scenario as the respective first features.

[0155] In an optional implementation, the model construction module, when constructing the second seed power model, is specifically configured to:

[0156] obtain source data of respective processes in a crystal pulling preparation process from historical data of the crystal pulling preparation process;

[0157] perform feature selection processing on the source data of the respective processes to obtain respective second features that match a second crystal pulling scenario and satisfy a second correlation condition with the seed power, based on the feature selection processing; the second crystal pulling scenario is a scenario in which a last crystal pulling process has a record of a shoulder break and the current furnace has no record of entering an equal-diameter process;

[0158] solve a standard interval range of each of the second features by filtering process data from seed power to equal-diameter according to equal-diameter performance; the process data includes data of each of the second features;

[0159] perform data selection on each of the second features according to the standard interval range of each of the second features, combine the selected feature data with the corresponding second seed power of the feature data in the historical data to obtain a second data set;

[0160] train a machine learning model based on the second data set to obtain the second seed power model.

[0161] In an optional implementation, the model construction module, when performing the feature selection processing on the source data of the respective processes to obtain the respective second features that match the second crystal pulling scenario and satisfy the second correlation condition with the seed power, is specifically configured to:

[0162] determine a correlation between different process parameters in the source data of the respective process;

[0163] prune part of the process parameters that meet the preset correlation relationship, to obtain respective reserved parameters;

[0164] determine the importance of each reserved parameter with respect to the seed power by using a gradient boosting decision tree model, and select respective reserved parameters that meet an importance condition and match the second crystal pulling scene as the respective second features.

[0165] In an optional implementation, the model construction module is specifically configured to:

[0166] filter the process data of each seed pulling to the constant diameter according to the constant diameter performance;

[0167] solve the standard interval range of the corresponding feature according to the statistical information of the data in the filtered process data of the seed pulling to the constant diameter;

[0168] The corresponding feature is a corresponding first feature or a second feature corresponding to the process data.

[0169] In an optional implementation, the model construction module is specifically configured to:

[0170] train multiple decision trees in the gradient boosting decision tree model by using the corresponding data set, each decision tree improves the model performance by reducing the residual error of the previous decision tree until a termination condition is met, to obtain a corresponding seed power model;

[0171] The corresponding data set is a first data set, and the corresponding seed power model is a first seed power model, or the corresponding data set is a second data set, and the corresponding seed power model is a second seed power model.

[0172] For the data-driven seed power determination apparatus disclosed in the embodiments of the present application, since it corresponds to the data-driven seed power determination method disclosed in the above method embodiments, the description is relatively simple, and the relevant similarities can be found in the above method embodiments, which will not be described in detail here.

[0173] The embodiments of the present application also disclose an electronic device, and the component structure of the electronic device is shown in Figure 5 at least includes:

[0174] a memory 10 for storing a computer instruction set;

[0175] The computer instruction set can be implemented in the form of a computer program.

[0176] The processor 20 is configured to implement the data-driven seeding power determination method according to any one of the method embodiments by executing the computer instruction set.

[0177] The processor 20 can be a central processing unit (CPU), an application-specific integrated circuit (ASIC), a digital signal processor (DSP), a field programmable gate array (FPGA), a neural network processor (NPU), a deep learning processor (DPU), or other programmable logic devices, etc.

[0178] The electronic device can be provided with a display device and / or a display interface, and can be externally connected to a display device.

[0179] In addition, the electronic device can further include a communication interface, a communication bus, and the like. The memory, the processor, and the communication interface can communicate with each other through the communication bus.

[0180] The communication interface is configured to communicate between the electronic device and other devices. The communication bus can be a peripheral component interconnect (PCI) bus or an extended industry standard architecture (EISA) bus, etc. The communication bus can be divided into an address bus, a data bus, a control bus, etc.

[0181] In addition, the present application also provides a computer readable medium having a computer program stored thereon, and the computer program includes program codes for executing the data-driven seeding power determination method according to any one of the method embodiments.

[0182] In the context of this application, a computer readable medium (machine readable medium) can be a tangible medium that can contain, or store a program for use by or in connection with an instruction execution system, apparatus, or device. The machine readable medium can be a machine readable signal medium or a machine readable storage medium. The machine readable medium can include, but is not limited to, electronic, magnetic, optical, electromagnetic, infrared, or semiconductor system, apparatus, or device, or any suitable combination of the foregoing. More specific examples of the machine readable storage medium will include one or more lines of electrical connections, portable computer disks, hard disks, random access memory (RAM), read-only memory (ROM), erasable programmable read-only memory (EPROM or Flash memory), optical fibers, portable compact disc read-only memories (CD-ROMs), optical storage devices, magnetic storage devices, or any suitable combination of the foregoing.

[0183] It should be noted that each of the embodiments in the present specification is described in a progressive manner, and each embodiment focuses on the differences from other embodiments. The same or similar parts of each embodiment can be understood by referring to each other.

[0184] For the convenience of description, the above system or device is described by dividing into various modules or units respectively. Of course, the functions of each unit can be implemented in the same or multiple software and / or hardware in the implementation of the present application.

[0185] From the above description of the embodiments, those skilled in the art can clearly understand that the present application can be implemented by means of software and the necessary general hardware platform. Based on such understanding, the technical solutions of the present application can be embodied in the form of a software product, which can be stored in a storage medium, such as a ROM / RAM, a magnetic disk, an optical disk, etc., and includes a number of instructions for making a computer device (which can be a personal computer, a server, or a network device, etc.) execute the methods described in each embodiment or some parts of the embodiments of the present application.

[0186] Finally, it needs to be pointed out that in this document, relational terms such as first, second, third, and the like, are used solely to distinguish one entity or action from another, without necessarily requiring or implying any actual relationship or order between such entities or actions. Moreover, the terms "comprises", "comprising", or any other variations thereof, are intended to cover a non-exclusive inclusion, such that a process, method, article, or apparatus that comprises a list of elements does not include only those elements but can include other elements not expressly listed or inherent to such process, method, article, or apparatus. An element proceeded by "comprises... a" does not, without more constraints, exclude the existence of additional identical elements in the process, method, article, or apparatus that comprises the element.

[0187] The above description is merely preferred embodiments of the present application, and it is to be pointed out that those skilled in the art can make a number of improvements and refinements without departing from the principles of the present application, and these improvements and refinements shall also be considered as falling within the scope of the present application.

Claims

1. A data-driven based seeding power determination method, characterized in that, The method comprises: obtaining last crystal pulling process data; determining whether the last crystal pulling process has occurred shoulder breakage according to the last crystal pulling process data; if the last crystal pulling process has not occurred shoulder breakage, obtaining process parameters of a seed pulling process, process parameters of a shoulder forming process and on-site parameters of the current crystal pulling, and determining seed pulling power of the current crystal pulling by using a first seed pulling power model according to the process parameters of the seed pulling process, the process parameters of the shoulder forming process and the on-site parameters of the current crystal pulling; if the last crystal pulling process has occurred shoulder breakage and there is no record of entering the equal diameter process in the current furnace cycle, obtaining process parameters of a seed pulling process and on-site parameters of the current crystal pulling, and determining seed pulling power of the current crystal pulling by using a second seed pulling power model according to the process parameters of the seed pulling process and the on-site parameters of the current crystal pulling; wherein the second seed pulling power model and the first seed pulling power model are different intelligent agents trained by constructing data sets through feature selection and standard interval range solving of historical crystal pulling process parameters for different scenarios of whether the last crystal pulling has occurred shoulder breakage in the crystal pulling preparation process; wherein the construction process of the first seed pulling power model comprises: obtaining source data of each process in the crystal pulling preparation process from historical data of the crystal pulling preparation process; performing feature selection processing matched with the first crystal pulling scenario on the source data of each process to obtain each first feature matched with the first crystal pulling scenario and satisfying a first correlation condition with seed pulling power; the first crystal pulling scenario is a scenario in which the last crystal pulling process has not occurred shoulder breakage; solving the standard interval range of each first feature by filtering process data from seed pulling to equal diameter according to equal diameter performance; the process data includes data of each first feature; selecting data of each first feature according to the standard interval range of each first feature, combining the selected feature data with the first seed pulling power corresponding to the feature data in the historical data to obtain a first data set; training a machine learning model based on the first data set to obtain the first seed pulling power model; the construction process of the second seed pulling power model comprises: obtaining source data of each process in the crystal pulling preparation process from historical data of the crystal pulling preparation process; performing feature selection processing matched with the second crystal pulling scenario on the source data of each process to obtain each second feature matched with the second crystal pulling scenario and satisfying a second correlation condition with seed pulling power; the second crystal pulling scenario is a scenario in which the last crystal pulling process has occurred shoulder breakage and there is no record of entering the equal diameter process in the current furnace cycle; solving the standard interval range of each second feature by filtering process data from seed pulling to equal diameter according to equal diameter performance; the process data includes data of each second feature; According to the standard interval range of each second feature, data selection is performed on each second feature, and the selected feature data is combined with the corresponding second seeding power in the historical data to obtain a second data set; wherein each adjacent seeding in the historical data is referred to as the last seeding and the current seeding, respectively, and the second data set is obtained by screening the records of the last seeding in the historical data that did not enter the equal-diameter process after the shoulder breakage and the records of the current seeding that meet the expected equal-diameter performance; Based on the second data set, a machine learning model is trained to obtain the second seeding power model.

2. The data-driven seeding power determination method of claim 1, wherein, The process parameters of the seeding process include at least part of the parameters of the seeding power, the average seeding speed, the late seeding speed, the seeding time, the seeding temperature, the argon flow, the crystal rotation speed, the crucible rotation speed, the number of re-feeding, and the remaining material amount in the last crystal pulling process; The process parameters of the shoulder process include at least part of the parameters of the last shoulder time, the shoulder height, the pre-shoulder height, the mid-shoulder height, the shoulder opening time, the argon flow, the crystal rotation speed, and the crucible rotation speed; The on-site parameters of the current crystal pulling include at least part of the parameters of the remaining material amount, the liquid surface temperature, the number of re-feeding, the thermal field size, and the single crystal size.

3. The data-driven seeding power determination method of claim 1, wherein, The feature selection processing matched with the first crystal pulling scene is performed on the source data of each process to obtain each first feature matched with the first crystal pulling scene and meeting the first correlation condition with the seeding power, including: determining the correlation between different process parameters in the source data of each process; eliminating part of the parameters in the multiple process parameters that meet the preset correlation relationship to obtain each reserved parameter; using a gradient boosting decision tree model to determine the importance of each reserved parameter relative to the seeding power, and selecting each reserved parameter corresponding to the importance meeting the importance condition and matched with the first crystal pulling scene as the each first feature.

4. The data-driven seeding power determination method of claim 1, wherein, The feature selection processing matched with the second crystal pulling scene is performed on the source data of each process to obtain each second feature matched with the second crystal pulling scene and meeting the second correlation condition with the seeding power, including: determining the correlation between different process parameters in the source data of each process; eliminating part of the parameters in the multiple process parameters that meet the preset correlation relationship to obtain each reserved parameter; using a gradient boosting decision tree model to determine the importance of each reserved parameter relative to the seeding power, and selecting each reserved parameter corresponding to the importance meeting the importance condition and matched with the second crystal pulling scene as the each second feature.

5. The data-driven seeding power determination method of claim 1, wherein, The standard interval range of the corresponding feature is solved by screening the process data from seeding to equal-diameter according to the equal-diameter performance, including: screening the process data from seeding to equal-diameter for each time according to the equal-diameter performance; solving the standard interval of the corresponding feature according to the statistical information of the data in the screened process data from seeding to equal-diameter; wherein the corresponding feature is the corresponding first feature or second feature corresponding to the process data.

6. The data-driven seeding power determination method of claim 1, wherein, The machine learning model training is performed based on a corresponding data set, including: training multiple decision trees in a gradient boosting decision tree model using the corresponding data set, each decision tree improving model performance by reducing the residual of the previous decision tree until a termination condition is met, to obtain a corresponding seeding power model; wherein the corresponding data set is a first data set and the corresponding seeding power model is a first seeding power model, or the corresponding data set is a second data set and the corresponding seeding power model is a second seeding power model.

7. A data-driven based seeding power determination apparatus, characterized in that, including: an acquisition module configured to acquire last crystal pulling process data; a first determination module configured to determine whether a shoulder breaking occurred in a last crystal pulling process according to the last crystal pulling process data; a second determination module configured to, in a case where the shoulder breaking did not occur in the last crystal pulling process, acquire process parameters of a seeding process and a shoulder breaking process in the last crystal pulling process and on-site parameters of the current crystal pulling, and determine a seeding power of the current crystal pulling by using a first seeding power model according to the process parameters of the seeding process and the shoulder breaking process in the last crystal pulling process and the on-site parameters of the current crystal pulling; a third determination module configured to, in a case where the shoulder breaking occurred in the last crystal pulling process and there is no record of entering an equal diameter process in the current furnace cycle, acquire process parameters of a seeding process in the last crystal pulling process and on-site parameters of the current crystal pulling, and determine a seeding power of the current crystal pulling by using a second seeding power model according to the process parameters of the seeding process in the last crystal pulling process and the on-site parameters of the current crystal pulling; wherein the second seeding power model and the first seeding power model are different intelligent agents trained by performing feature selection and standard interval range solving of features matched with different scenarios according to historical process parameters; wherein the construction process of the first seeding power model includes: acquiring source data of each process in the crystal pulling preparation process from historical data of the crystal pulling preparation process; performing feature selection processing matched with a first crystal pulling scenario on the source data of each process to obtain each first feature matched with the first crystal pulling scenario and satisfying a first correlation condition with the seeding power; the first crystal pulling scenario is a scenario where the shoulder breaking did not occur in the last crystal pulling process; solving a standard interval range of each first feature by filtering process data from seeding to equal diameter according to equal diameter performance; the process data includes data of each first feature; performing data selection on each first feature according to the standard interval range of each first feature, combining the selected feature data with the first seeding power corresponding to the feature data in the historical data to obtain a first data set; performing machine learning model training based on the first data set to obtain the first seeding power model; the construction process of the second seeding power model includes: acquiring source data of each process in the crystal pulling preparation process from historical data of the crystal pulling preparation process; The source data of each process is subjected to a feature selection process matched with a second crystal pulling scenario to obtain each second feature matched with the second crystal pulling scenario and satisfying a second correlation condition with a seed power based on the feature selection process; the second crystal pulling scenario is a scenario in which a last crystal pulling process has a record of a break during a shoulder-to-equal-diameter transition and the current furnace cycle does not have a record of entering the equal-diameter process; A standard interval range of each second feature is solved by filtering process data of seed pulling to equal-diameter according to equal-diameter performance; the process data includes data of each second feature; Data of each second feature is selected according to the standard interval range of each second feature, and selected feature data is combined with a second seed power corresponding to the feature data in historical data to obtain a second data set; wherein each adjacent seed pulling in historical data is referred to as a last seed pulling and a current seed pulling, respectively, and the second data set is obtained by filtering a record of a last seed pulling in historical data having a break during a shoulder-to-equal-diameter transition and not entering the equal-diameter process and a record of equal-diameter performance in a current seed pulling record satisfying an expectation; A machine learning model is trained based on the second data set to obtain the second seed power model.

8. A computer readable medium having stored thereon a computer program, characterized in that, The computer program, when executed by a processor, can be used to implement the data-driven seed power determination method according to any one of claims 1-6.

Images