Optimization method and device for improving algorithm robustness in monocrystalline silicon czochralski method, equipment and storage medium

Abstract

The invention provides an optimization method and device for improving algorithm robustness in a monocrystalline silicon czochralski method, equipment and a storage medium. The optimization method comprises the following steps: for each rod time of a monocrystalline silicon czochralski method, acquiring and processing related data of process parameters in a coil base temperature adjustment stage to obtain a process parameter data set in the temperature adjustment stage; constructing a sensing model, and fitting data of each rod-time technological parameter in the technological parameter data set in the temperature adjustment stage to obtain a key fitting coefficient result set; clustering the key fitting coefficient result set to obtain an overall boundary of the key fitting coefficient result set; constructing a mapping model between output variables of other process algorithms and the perception model; obtaining a combination of two key fitting coefficients by utilizing a sensing model after a new rod temperature adjustment stage is started and an earlier-stage set time period is finished, judging whether the combination is within an overall boundary, if not, adjusting process parameters, and if so, fusing a mapping model into other original process algorithms; and the algorithm robustness can be improved.

Description

Technical Field

[0001] The invention relates to the technical field of single crystal silicon Czochralski method, and in particular to an optimization method for improving algorithm robustness in single crystal silicon Czochralski method. Background Art

[0002] Single crystal silicon is an important material for manufacturing semiconductors and solar cells. Its quality and production efficiency have a crucial impact on the performance and cost of the final product. In the traditional single crystal silicon Czochralski method, process personnel need to control key parameters such as temperature, crystal pulling speed, crystal diameter growth rate, etc. through PID control and manual adjustment based on experience. This method is not only time-consuming and labor-intensive, but also prone to human errors, resulting in unstable crystal production quality. The specific difficulties are: (1) Difficulty in multi-parameter control: During the crystal pulling process, multiple parameters such as temperature, crystal pulling speed, crucible rotation, crucible lifting, etc. need to be precisely controlled. These parameters affect each other and are complex to control; (2) Difficulty in real-time adjustment: Any change in parameters during the crystal pulling process requires real-time adjustment. Under the traditional process, manual operation is relied on, making it difficult to achieve rapid response and precise control; (3) Difficulty in parameter optimization: Due to the large number of parameters and their interdependence, finding the best combination of parameters to achieve optimal quality and efficiency is a complex optimization problem, and traditional PID has great limitations in solving this problem.

[0003] Based on the above difficulties and the rapid development of AI technology, strong technical support has been provided for the realization of crystal pulling technology. A method combining artificial intelligence (AI) technology to optimize the single crystal silicon production process has been born. At this stage, the key steps to optimize the single crystal silicon production process through AI algorithms include temperature adjustment, crystal induction, shoulder release and equal diameter process stages, which are the most important links in the single crystal silicon direct pulling process. In the traditional PID control scenario, these process stages involve multiple control parameters, some of which can be predicted and controlled by AI algorithms. The following are the detailed process parameters that need to be controlled by the AI ​​algorithm at different process stages, as shown in Table 1:

[0004] Process stage Process parameters Temperature Control Main power, liquid port distance seeding Crystal pulling speed Shoulders Crystal pulling speed, main power Equal diameter Crystal pulling speed, main power, crucible rotation

[0005] Table 1

[0006] The main power in Table 1 refers to the main heating power. Usually, in addition to the main power, there is also a bottom power, that is, the bottom heating power; the liquid port distance is the distance from the bottom of the heat screen to the liquid surface.

[0007] The above process parameters are production parameters that can be directly controlled by the equipment. Other process parameters (such as crystal diameter growth rate) usually change according to the changes in these parameters, and are essentially controlled by the above process parameters. The application of AI algorithms aims to achieve the following goals: (1) Real-time prediction. Use AI algorithms for data analysis and prediction; (2) Parameter optimization and adjustment. Based on the on-site process environment, AI algorithms can automatically adjust process parameters to improve the robustness and generalization of AI algorithms; (3) Improve stability and consistency. Reduce human intervention and improve the stability of the production process and the consistency of product quality.

[0008] However, during the production of single crystal silicon, the on-site process environment often changes. These changes may come from multiple aspects, such as: (1) changes in the thermal field. As the equipment is used for a long time, the thermal field in the furnace may change, resulting in uneven temperature distribution; (2) equipment wear. As the equipment runs for a long time, the wear of various components may affect its accuracy and stability; (3) environmental factors. The external temperature, argon gas flow rate, water cooling screen flow rate, crucible rotation and other factors in the production environment will also affect the crystal pulling process. These changes in the process environment make it a difficult problem to ensure the robustness of the AI ​​algorithm model.

[0009] Based on the process parameters in Table 1 and the above process environment changes, it can be seen that all process parameters and changes in the process environment are closely related to the "liquid surface temperature" parameter, specifically: (1) the main power can directly control the liquid surface temperature; (2) the liquid mouth distance can affect the liquid surface temperature to a certain extent; (3) the mapping relationship between the crystal pulling speed and the crystal diameter growth rate is greatly affected by the liquid surface temperature; (4) the change of the thermal field has a great influence on the liquid surface temperature; (5) the wear of the equipment, such as the insulation performance of the crucible and the aging of the heater, has a great influence on the liquid surface temperature; (6) environmental factors have a great influence on the liquid surface temperature. Therefore, the liquid surface temperature can be used as a perception variable in the field of single crystal silicon Czochralski method to reflect the current process environment. However, the liquid surface temperature of the single crystal silicon Czochralski method is usually around 1450 degrees, and the temperature change is highly sensitive to the crystal pulling process. Current technical means cannot accurately measure such high temperatures, and changes in the process environment are difficult to accurately decouple and evaluate the impact on the liquid surface temperature in the field of AI algorithms. Therefore, the measurement of perception variables in AI crystal pulling is a difficult problem.

[0010] Chinese patent application CN113913923A proposes an automatic temperature control method for a single crystal furnace. The automatic temperature control method specifically performs cyclic temperature control on the single crystal furnace to slowly lower the current temperature of the liquid level and adjust the seeding power. The control system logic of the single crystal furnace automatic temperature control detects the current temperature of the liquid level in the single crystal furnace in real time during the cycle, and calculates the power reduction according to the starting seeding power and the set target temperature, cooling cycle, cycle cycle, and power adjustment coefficient. The display power after a temperature control cycle is calculated based on the power reduction and the set seeding power and the cycle is performed.

[0011] The above method assumes that the process environment does not change and the current temperature of the liquid surface can accurately reflect the actual temperature of the liquid surface, but there is no such ideal environment in the actual production process. This causes the method to have large differences under different process environments, resulting in significant lags and errors. In addition, this method relies on the control system to automatically record some process parameters under the same process conditions several times and perform average processing, but due to the differences in the on-site environment, it is difficult to find several temperature adjustment processes with exactly the same process conditions, and the average processing will average the differences in each temperature adjustment, and differentiated processing cannot be achieved.

[0012] Chinese patent application CN117107355A proposes a crystal pulling method and system for automatically adjusting crystal pulling process parameters based on artificial intelligence, including collecting equal-diameter state photos of crystal rods in a furnace, and data-marking the equal-diameter state photos to obtain marked equal-diameter state photos; performing deep learning on a computer based on the marked equal-diameter state photos to obtain the optimal state of crystal rods of different lengths, and establishing a variable function of power P and time T based on the optimal state, thereby obtaining a crystal pulling recognition model; when crystal pulling starts, AI automatically detects the state of the crystal rod and makes a judgment based on the crystal pulling recognition model; if the crystal rod meets the standard state, the current crystal pulling process parameters are maintained for crystal pulling, and the crystal pulling recognition model continues to make a judgment until the crystal pulling ends; if the crystal rod does not meet the standard state, the current crystal pulling process parameters are adjusted based on the variable function until the state of the crystal rod returns to the standard state, and then the crystal is pulled according to the standard crystal pulling process parameters, and the crystal pulling recognition model continues to make a judgment until the crystal pulling ends.

[0013] The variable function of power and time established by the above method is constructed based on all historical data. However, this method is difficult to accurately reflect the differences in the process environment during each crystal pulling process. For example, different crystal pulling batches may have differences in temperature, pressure, atmosphere composition, etc., and the changes in these variables are not fully considered in the variable function. Therefore, when dealing with different process environments, the model may not be able to effectively adjust the parameters, resulting in a decrease in the robustness and accuracy of the crystal pulling identification model.

[0014] Chinese patent application CN115125611A proposes an intelligent crystal pulling process method. The method includes obtaining the first optimal pulling speed of the crystal seeding process, including: collecting big data of the crystal seeding pulling speed when the crystal seeding is alive in the historical thermal field; obtaining the first set of big data according to the big data; obtaining the first optimal pulling speed of the crystal seeding under the thermal field according to the first set; obtaining the average pulling speed of the crystal seeding process, including: recording the instantaneous pulling speed per second in the crystal seeding process under the current thermal field; obtaining the second set according to the instantaneous pulling speed; in the crystal seeding process, when the instantaneous pulling speed oscillates in the second set within the first preset time, the instantaneous pulling speed at this time is the first instantaneous pulling speed, and the first instantaneous pulling speed feedbacks the change of temperature; all the first instantaneous pulling speeds obtain the instantaneous pulling speed set, and obtain the target average pulling speed according to the instantaneous pulling speed set; compare the first optimal pulling speed and the average pulling speed, and jump to the appropriate shoulder release process.

[0015] The above method relies on collecting a large amount of historical data under the same process environment and obtaining corresponding sets based on these data, which makes it difficult to flexibly apply in new process environments and cannot consider the impact of other process environment variables besides thermal field changes. In addition, since the drawing speed is controlled by the equipment PID, it cannot fully reflect the temperature changes. This means that even if the temperature changes, the adjustment of the drawing speed may not be timely or accurate enough, resulting in the actual temperature changes not being effectively fed back and controlled. This situation may cause deviations in the calculation results, reducing the accuracy and reliability of the model prediction. Summary of the invention

[0016] In order to solve at least one technical problem in the prior art, the embodiments of the present invention provide an optimization method, device, equipment and storage medium for improving the robustness of the algorithm in the single crystal silicon Czochralski method, which can improve the robustness of the process algorithm. To achieve the above technical objectives, the technical solution adopted by the embodiments of the present invention is:

[0017] In a first aspect, an embodiment of the present invention provides an optimization method for improving algorithm robustness in a single crystal silicon Czochralski method, comprising the following steps:

[0018] For each single crystal silicon Czochralski process, relevant data of process parameters in the furnace temperature adjustment stage are obtained and processed to obtain a data set of process parameters in the temperature adjustment stage;

[0019] Construct a perception model, and fit the data of each rod process parameter in the process parameter data set of the temperature adjustment stage to obtain a result set of key fitting coefficients;

[0020] Cluster the key fitting coefficient result set to obtain the overall boundary of the key fitting coefficient result set;

[0021] Construct a mapping model between the output variables of other process algorithms and the perception model, integrate the mapping model into the original other process algorithms and perform end-to-end training to obtain the regression coefficient result combination of the mapping model;

[0022] After the new temperature adjustment phase begins and the previous set period ends, the perception model is used to obtain the combination of the two key fitting coefficients, and then the overall boundary of the key fitting coefficient result set is used to determine whether the combination is within the overall boundary. If not, the process parameters are adjusted to make it within the overall boundary. If so, the mapping model is integrated into the original other process algorithms to obtain the output variables of the corrected predicted output.

[0023] Furthermore, for each single crystal silicon Czochralski process, obtaining and processing relevant data of process parameters in the furnace temperature adjustment stage specifically includes:

[0024] Obtain relevant data of process parameters in the early set period after the start of each rod temperature adjustment stage, wherein the process parameters include liquid surface temperature, main power, remaining material amount, crucible position and crucible speed;

[0025] Smoothing the data of liquid surface temperature; performing first-order difference processing on the data of main boost power and liquid surface temperature;

[0026] Then the data of the process parameters of the same rod batch are normalized to obtain a standardized data record; the relevant data of the process parameters of the temperature adjustment stage of all rod batches are processed to obtain a data set of process parameters of the temperature adjustment stage.

[0027] Furthermore, the perception model is constructed, and the data of each rod process parameter in the temperature adjustment stage process parameter data set is fitted to obtain the key fitting coefficient result set, which specifically includes:

[0028] constructing a perception model, wherein the perception model is represented by a fitting equation;

[0029] For the fitting coefficients in the fitting equation, two key fitting coefficients are determined, and the fitting equation is solved based on the data of each rod process parameter in the process parameter data set of the temperature adjustment stage to obtain a key fitting coefficient result set.

[0030] Furthermore, the fitting equation is shown in formula (1);

[0031]

[0032] Where ΔT is the difference in liquid surface temperature, t is the step time, Δp is the average of the difference in main power in the set period before time t, w is the remaining material, z is the crucible position, r is the crucible speed, A, B, D, θ and is the fitting coefficient, C is the constant term;

[0033] The two key fitting coefficients are θ and

[0034] The method of solving the fitting equation based on the data of each rod process parameter in the temperature adjustment stage process parameter data set includes: solving the fitting equation using a trust region reflection algorithm based on the data of each rod process parameter in the temperature adjustment stage process parameter data set, and limiting two key fitting coefficients θ and The value range of is [-1,1], and the key fitting coefficient result set about the two key fitting coefficients is obtained.

[0035] Furthermore, clustering the key fitting coefficient result set to obtain the overall boundary of the key fitting coefficient result set specifically includes:

[0036] The key fitting coefficient result set is processed by a density-based clustering algorithm, and the key fitting coefficient result set is divided into different clusters according to density distribution;

[0037] The boundaries of each cluster are calculated using a convex hull algorithm.

[0038] The boundaries of all clusters are integrated to form the overall boundary of the key fitting coefficient result set.

[0039] Furthermore, the mapping model between the output variables of other process algorithms and the perception model is constructed, the mapping model is integrated into the original other process algorithms and end-to-end training is performed, and the regression coefficient result combination of the mapping model is obtained specifically includes:

[0040] Constructing a mapping model between other process algorithm output variables and the perception model, the mapping model is shown in formula (2);

[0041]

[0042] in, is the output variable of other process algorithms, y is The true value of; k1, j1, k2 and j2 are the regression coefficients of the mapping model, and C is the constant term;

[0043] The mapping model is integrated into other existing process algorithms, and end-to-end training is performed according to relevant training methods of other existing process algorithms to obtain a combination of regression coefficient results of mapping models k1, j1, k2 and j2.

[0044] Further, the method of using the perception model to obtain a combination of two key fitting coefficients and then judging whether the combination is within the overall boundary by the overall boundary of the key fitting coefficient result set includes: using the perception model to obtain two key fitting coefficients as θ and The combination is then judged whether the combination is within the overall boundary through the overall boundary of the key fitting coefficient result set;

[0045] The mapping model is integrated into the original other process algorithms to obtain the output variables of the modified predicted output, including: after obtaining the output variables of the predicted output of the other process algorithms, the two key fitting coefficients θ and The combination of and the output variables of other process algorithm prediction outputs are substituted into the mapping model to obtain the output variables of the corrected prediction output.

[0046] In a second aspect, an embodiment of the present invention provides an optimization device for improving algorithm robustness in a single crystal silicon Czochralski method, comprising:

[0047] The process parameter data acquisition and processing module is used to acquire and process the relevant data of the process parameters in the temperature adjustment stage of the furnace for each rod pulling process of single crystal silicon, so as to obtain a data set of process parameters in the temperature adjustment stage;

[0048] The perception model construction and data fitting module is used to construct the perception model and fit the data of each rod process parameter in the process parameter data set of the temperature adjustment stage to obtain the key fitting coefficient result set;

[0049] The clustering and boundary acquisition module is used to cluster the key fitting coefficient result set and obtain the overall boundary of the key fitting coefficient result set;

[0050] The mapping model construction and training module is used to construct a mapping model between the output variables of other process algorithms and the perception model, integrate the mapping model into the original other process algorithms and perform end-to-end training to obtain the regression coefficient result combination of the mapping model;

[0051] The judgment and execution module is used to obtain the combination of two key fitting coefficients using the perception model after the previous set period ends at the beginning of the new temperature adjustment stage, and then judge whether the combination is within the overall boundary through the overall boundary of the key fitting coefficient result set. If not, the process parameters are adjusted to make it within the overall boundary. If so, the mapping model is integrated into the original other process algorithms to obtain the output variable of the corrected predicted output.

[0052] In a third aspect, an embodiment of the present invention provides an electronic device, including:

[0053] a memory storing a computer program;

[0054] A processor is used to run the computer program, and when the computer program is running, the steps of the optimization method for improving the robustness of the algorithm in the single crystal silicon Czochralski method as described above are executed.

[0055] In a fourth aspect, an embodiment of the present invention provides a storage medium storing a computer program, wherein the computer program is configured to execute the steps of the optimization method for improving the robustness of the algorithm in the single crystal silicon Czochralski method as described above when running.

[0056] The technical solution provided by the embodiment of the present invention has the following beneficial effects:

[0057] 1) By constructing a perception model and introducing abstract perception variables, the anomalies of the process environment can be detected in time in the early stage of the temperature adjustment stage, so that the crystal pulling process can respond quickly at the beginning of the anomaly and adjust the process parameters in time to avoid the accumulation and expansion of potential problems.

[0058] 2) In the single crystal silicon Czochralski production process, process parameters such as liquid surface temperature and crucible rotation are strongly coupled in multiple process stages, and traditional methods are difficult to analyze and optimize separately; this application uses perception models and mapping models to effectively link different process stages, improves the model's generalization ability and robustness to different abnormal situations, and ensures that the results of the original prediction output of other process algorithms are more accurate and reliable. BRIEF DESCRIPTION OF THE DRAWINGS

[0059] Figure 1 The figure is a flow chart of an optimization method for improving algorithm robustness in an embodiment of the present invention.

[0060] Figure 2 Flow chart of the specific process of step S10 in the embodiment of the present invention.

[0061] Figure 3 Flow chart of the specific process of step S20 in the embodiment of the present invention.

[0062] Figure 4 Flow chart of the specific process of step S30 in the embodiment of the present invention.

[0063] Figure 5 Flow chart of the specific process of step S40 in the embodiment of the present invention.

[0064] Figure 6 Schematic diagram of an optimization device for improving algorithm robustness in an embodiment of the present invention.

[0065] Figure 7 Schematic diagram of the structure of an electronic device in an embodiment of the present invention. DETAILED DESCRIPTION

[0066] In order to make the purpose, technical solution and advantages of the present invention more clearly understood, the present invention is further described in detail below in conjunction with the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are only used to explain the present invention and are not used to limit the present invention.

[0067] In the description of the embodiments of the present invention, it should be noted that the terms "center", "upper", "lower", "left", "right", "vertical", "horizontal", "inner", "outer", etc., indicating the orientation or positional relationship, are based on the orientation or positional relationship shown in the drawings, and are only for the convenience of describing the present invention and simplifying the description, rather than indicating or implying that the device or element referred to must have a specific orientation, be constructed and operated in a specific orientation, and therefore cannot be understood as limiting the present invention. In addition, the terms "first", "second", and "third" are used for descriptive purposes only and cannot be understood as indicating or implying relative importance.

[0068] In the description of the embodiments of the present invention, it should be noted that, unless otherwise clearly specified and limited, the terms "installed", "connected", and "connected" should be understood in a broad sense, for example, it can be a fixed connection, a detachable connection, or an integral connection; it can be a mechanical connection or an electrical connection; it can be a direct connection, or it can be indirectly connected through an intermediate medium, it can also be the internal connection of two components, it can be a wireless connection, or it can be a wired connection. For ordinary technicians in this field, the specific meanings of the above terms in the present invention can be understood according to specific circumstances.

[0069] In addition, the technical features involved in the different embodiments of the present invention described below can be combined with each other as long as they do not conflict with each other.

[0070] The embodiment of the present invention provides an optimization method for improving the robustness of the algorithm in the single crystal silicon Czochralski method, such as Figure 1 As shown, the following steps are included:

[0071] Step S10, for each single crystal silicon Czochralski process, obtain and process relevant data of process parameters in the temperature adjustment stage of the furnace to obtain a data set of process parameters in the temperature adjustment stage;

[0072] Step S20, constructing a perception model, and fitting the data of each rod process parameter in the temperature adjustment stage process parameter data set to obtain a key fitting coefficient result set;

[0073] Step S30, clustering the key fitting coefficient result set to obtain the overall boundary of the key fitting coefficient result set;

[0074] Step S40, constructing a mapping model between output variables of other process algorithms and the perception model, integrating the mapping model into the original other process algorithms and performing end-to-end training to obtain a regression coefficient result combination of the mapping model;

[0075] Step S50, after the previous set period ends after the new rod temperature adjustment stage begins, the perception model is used to obtain the combination of the two key fitting coefficients, and then the overall boundary of the key fitting coefficient result set is used to determine whether the combination is within the overall boundary. If not, the process parameters are adjusted to make it within the overall boundary. If so, the mapping model is integrated into the original other process algorithms to obtain the output variable of the corrected predicted output.

[0076] Liquid surface temperature is the most critical sensory variable in the crystal pulling process. However, whether it is measured directly or indirectly, it will cause large deviations due to measurement errors and interference from other physical factors. In addition, the change pattern of liquid surface temperature (fluctuation) is usually very complex, so it is necessary to indirectly characterize the liquid surface temperature through other variables. This characterization does not necessarily require interpretability. To optimize this indirect characterization, the selected data set should minimize the interference from process factors. To this end, the temperature adjustment stage best meets this requirement. In addition, the temperature adjustment stage is located at the initial stage of the same key process. By sensing the environmental status at the beginning of production, timely adjustments can be made to ensure the smooth progress of the crystal pulling process.

[0077] The embodiment of the present invention constructs an abstract perception method for representing the liquid surface temperature environment. The key fitting coefficients in the perception model can be abstractly used as perception variables, which can reflect the liquid surface temperature to a large extent. In the early stage of each crystal pulling process entering the temperature adjustment stage, the perception model can automatically detect the environmental state. If there is an abnormality, it can respond quickly at the beginning of the abnormality, adjust the process parameters in time, and avoid the accumulation and expansion of potential problems. Through the mapping model, different process stages can be effectively linked, which improves the generalization ability of the model and the robustness to different abnormal situations, ensuring that the results of the original prediction output of other process algorithms are more accurate and reliable.

[0078] like Figure 2 As shown, in step S10, for each single crystal silicon Czochralski process, obtaining and processing relevant data of process parameters in the furnace temperature adjustment stage specifically includes:

[0079] Step S101, obtaining relevant data of process parameters in the early set period after the start of each rod temperature adjustment stage, wherein the process parameters include liquid surface temperature, main power, remaining material amount, crucible position and crucible speed;

[0080] In a specific embodiment, relevant data of process parameters within 60 minutes after the start of the temperature adjustment stage of each rod can be obtained, and the data sampling granularity is 10s;

[0081] Step S102, smoothing the data of the liquid surface temperature; performing first-order difference processing on the data of the main power and the liquid surface temperature;

[0082] In a specific embodiment, the smoothing process and the first-order difference process are both performed with a time granularity of 10 seconds;

[0083] Step S103, normalizing the data of the process parameters of the same batch of bars to obtain a standardized data record; processing the relevant data of the process parameters of the temperature adjustment stage of all batches of bars to obtain a data set of process parameters of the temperature adjustment stage;

[0084] Step S103 is to further process based on step S102;

[0085] like Figure 3 As shown, the step S20 constructs a perception model and fits the data of each rod process parameter in the temperature adjustment stage process parameter data set to obtain a key fitting coefficient result set, which specifically includes:

[0086] Step S201, constructing a perception model, wherein the perception model is represented by a fitting equation;

[0087] Specifically, the fitting equation is shown in formula (1);

[0088]

[0089] Where ΔT is the difference in liquid surface temperature, t is the step time, Δp is the average of the difference in main power in the set period before time t, w is the remaining material, z is the crucible position, r is the crucible speed, A, B, D, θ and is the fitting coefficient, C is the constant term;

[0090] In this embodiment, Δp is the average of the main power differential values ​​within 20 minutes before time t; considering the delayed response effect of the liquid surface temperature on the main power, when analyzing the liquid surface temperature in the temperature adjustment stage, it is necessary to comprehensively consider the cumulative effect of the main power change at a certain moment on it;

[0091] Step S202, for the fitting coefficients in the fitting equation, two key fitting coefficients are determined, and the fitting equation is solved based on the data of the process parameters of each rod in the process parameter data set of the temperature adjustment stage to obtain a key fitting coefficient result set;

[0092] Specifically, the two key fitting coefficients are θ and The complexity of perceptual variables may not be represented by a single parameter and may not be represented by interpretable parameters; the two key fitting coefficients are θ and The combination of can be abstractly used as a perception variable to a certain extent. After eliminating the influence of the main power variable, these two key fitting coefficients completely determine the trend of the liquid surface temperature, so they can reflect the liquid surface temperature to a large extent.

[0093] Specifically, the method of solving the fitting equation based on the data of each rod process parameter in the temperature adjustment stage process parameter data set includes: solving the fitting equation using a trust region reflection algorithm based on the data of each rod process parameter in the temperature adjustment stage process parameter data set, and limiting two key fitting coefficients θ and The value range of is [-1,1], and the key fitting coefficient result set about the two key fitting coefficients is obtained;

[0094] Trust Region Reflective Algorithm is Trust Region Reflective Algorithm;

[0095] like Figure 4 As shown, the step S30, clustering the key fitting coefficient result set to obtain the overall boundary of the key fitting coefficient result set specifically includes:

[0096] Step S301, processing the key fitting coefficient result set by a density-based clustering algorithm, and dividing the key fitting coefficient result set into different clusters according to density distribution;

[0097] Specifically, a density-based clustering algorithm can use the DBSCAN algorithm;

[0098] Step S302, calculating the boundary of each cluster by a convex hull algorithm;

[0099] Specifically, the Convex Hull algorithm can be used; for each cluster point set, the Convex Hull algorithm is used to find the outermost points, and these points are connected in sequence to form a convex polygon, which represents the boundary of the cluster; the point here refers to the coordinate point represented by the combination of two key fitting coefficients;

[0100] Step S303, integrating the boundaries of all clusters to form the overall boundary of the key fitting coefficient result set;

[0101] Points that do not meet the boundary conditions of the overall boundary are considered abnormal points or outliers;

[0102] like Figure 5 As shown, the step S40, constructing a mapping model between the output variables of other process algorithms and the perception model, integrating the mapping model into the original other process algorithms and performing end-to-end training, and obtaining the regression coefficient result combination of the mapping model specifically includes:

[0103] Step S401, constructing a mapping model between other process algorithm output variables and the perception model, the mapping model is shown in formula (2);

[0104]

[0105] in, is the output variable of other process algorithms, y is The true value of; k1, j1, k2 and j2 are the regression coefficients of the mapping model, and C is the constant term;

[0106] Step S402, integrating the mapping model into other existing process algorithms, and performing end-to-end training according to relevant training methods of other existing process algorithms, to obtain a regression coefficient result combination of mapping models k1, j1, k2 and j2;

[0107] Through this integrated training process, the adaptability and consistency of other process algorithms can be ensured, the overall performance can be improved, and more accurate parameter output can be obtained;

[0108] Specifically, in step S50, the step of obtaining a combination of two key fitting coefficients by using the perception model, and then judging whether the combination is within the overall boundary by using the overall boundary of the key fitting coefficient result set includes: obtaining two key fitting coefficients θ and The combination is then judged whether the combination is within the overall boundary through the overall boundary of the key fitting coefficient result set;

[0109] The mapping model is integrated into the original other process algorithms to obtain the output variables of the modified predicted output, including: after obtaining the output variables of the predicted output of the other process algorithms, the two key fitting coefficients θ and The combination of and the output variables of other process algorithm prediction outputs are substituted into the mapping model to obtain the output variables of the corrected prediction output.

[0110] like Figure 6 As shown, the embodiment of the present invention further provides an optimization device for improving algorithm robustness in a single crystal silicon Czochralski method, comprising:

[0111] The process parameter data acquisition and processing module is used to acquire and process the relevant data of the process parameters in the temperature adjustment stage of the furnace for each rod pulling process of single crystal silicon, so as to obtain a data set of process parameters in the temperature adjustment stage;

[0112] The perception model construction and data fitting module is used to construct the perception model and fit the data of each rod process parameter in the process parameter data set of the temperature adjustment stage to obtain the key fitting coefficient result set;

[0113] The clustering and boundary acquisition module is used to cluster the key fitting coefficient result set and obtain the overall boundary of the key fitting coefficient result set;

[0114] The mapping model construction and training module is used to construct a mapping model between the output variables of other process algorithms and the perception model, integrate the mapping model into the original other process algorithms and perform end-to-end training to obtain the regression coefficient result combination of the mapping model;

[0115] The judgment and execution module is used to obtain the combination of two key fitting coefficients using the perception model after the previous set period ends at the beginning of the new temperature adjustment stage, and then judge whether the combination is within the overall boundary through the overall boundary of the key fitting coefficient result set. If not, the process parameters are adjusted to make it within the overall boundary. If so, the mapping model is integrated into the original other process algorithms to obtain the output variable of the corrected predicted output.

[0116] like Figure 7 As shown, an embodiment of the present invention further proposes an electronic device, comprising: a processor and a memory; the processor and the memory communicate with each other, for example, are connected and communicate with each other through a bus; a computer program is stored in the memory; the processor is used to run the computer program, and when the computer program is running, the steps of the optimization method for improving the robustness of the algorithm in the single crystal silicon direct pulling method as described above are executed; the processor can be a CPU, or other general-purpose processors, digital signal processors (Digital Signal Processor, DSP), application specific integrated circuits (Application Specific Integrated Circuit, ASIC), field programmable gate arrays (Field-Programmable Gate Array, FPGA) or other programmable logic devices, discrete gates or transistor logic devices, discrete hardware components and other devices; the memory can include volatile memory, such as random access memory; the memory can also include non-volatile memory, such as flash memory, hard disk or solid state drive; the memory can also include a combination of the above-mentioned types of memory.

[0117] An embodiment of the present invention further proposes a storage medium, in which a computer program is stored, and the computer program is configured to execute the steps of the optimization method for improving the robustness of the algorithm in the single crystal silicon direct pulling method as described above when running; the storage medium includes a magnetic disk, an optical disk, a read-only memory (ROM), a random access memory (RAM), a flash memory (Flash Memory), a hard disk (Hard Disk Drive, abbreviated as: HDD) or a solid-state drive (SSD), etc. and any combination thereof.

[0118] Finally, it should be noted that the above specific implementation methods are only used to illustrate the technical solution of the present invention rather than to limit it. Although the present invention has been described in detail with reference to the embodiments, those skilled in the art should understand that the technical solution of the present invention can be modified or replaced by equivalents without departing from the spirit and scope of the technical solution of the present invention, which should be included in the scope of the claims of the present invention.

Claims

1. An optimization method for improving algorithm robustness in a single crystal silicon Czochralski method, characterized in that: The following steps are involved: For each single crystal silicon Czochralski process, relevant data of process parameters in the furnace temperature adjustment stage are obtained and processed to obtain a data set of process parameters in the temperature adjustment stage; Construct a perception model, and fit the data of each rod process parameter in the process parameter data set of the temperature adjustment stage to obtain the key fitting coefficient result set; Cluster the key fitting coefficient result set to obtain the overall boundary of the key fitting coefficient result set; Construct a mapping model between the output variables of other process algorithms and the perception model, integrate the mapping model into the original other process algorithms and perform end-to-end training to obtain the regression coefficient result combination of the mapping model; After the new temperature adjustment phase begins and the previous set period ends, the perception model is used to obtain the combination of the two key fitting coefficients, and then the overall boundary of the key fitting coefficient result set is used to determine whether the combination is within the overall boundary. If not, the process parameters are adjusted to make it within the overall boundary. If so, the mapping model is integrated into the original other process algorithms to obtain the output variables of the corrected predicted output.

2. The optimization method for improving algorithm robustness in the single crystal silicon Czochralski method according to claim 1, characterized in that: The method of obtaining and processing the relevant data of the process parameters in the furnace temperature adjustment stage for each single crystal silicon Czochralski process specifically includes: Obtain relevant data of process parameters in the early set period after the start of each rod temperature adjustment stage, wherein the process parameters include liquid surface temperature, main power, remaining material amount, crucible position and crucible speed; Smoothing the data of liquid surface temperature; performing first-order difference processing on the data of main boost power and liquid surface temperature; Then the data of the process parameters of the same rod batch are normalized to obtain a standardized data record; the relevant data of the process parameters of the temperature adjustment stage of all rod batches are processed to obtain a data set of process parameters of the temperature adjustment stage.

3. The optimization method for improving algorithm robustness in the single crystal silicon Czochralski method according to claim 1 or 2, characterized in that: The constructing perception model and fitting the data of each rod process parameter in the process parameter data set of the temperature adjustment stage to obtain the key fitting coefficient result set specifically includes: constructing a perception model, wherein the perception model is represented by a fitting equation; For the fitting coefficients in the fitting equation, two key fitting coefficients are determined, and the fitting equation is solved based on the data of each rod process parameter in the process parameter data set of the temperature adjustment stage to obtain a key fitting coefficient result set.

4. The optimization method for improving algorithm robustness in the single crystal silicon Czochralski method according to claim 3, characterized in that: The fitting equation is shown in formula (1); Where ΔT is the difference in liquid surface temperature, t is the step time, Δp is the average of the difference in main power in the set period before time t, w is the remaining material, z is the crucible position, r is the crucible speed, A, B, D, θ and is the fitting coefficient, C is the constant term; The two key fitting coefficients are θ and The method of solving the fitting equation based on the data of each rod process parameter in the temperature adjustment stage process parameter data set includes: solving the fitting equation using a trust region reflection algorithm based on the data of each rod process parameter in the temperature adjustment stage process parameter data set, and limiting two key fitting coefficients θ and The value range of is [-1,1], and the key fitting coefficient result set about the two key fitting coefficients is obtained.

5. The optimization method for improving algorithm robustness in the single crystal silicon Czochralski method according to any one of claims 4, characterized in that: The clustering of the key fitting coefficient result set to obtain the overall boundary of the key fitting coefficient result set specifically includes: The key fitting coefficient result set is processed by a density-based clustering algorithm, and the key fitting coefficient result set is divided into different clusters according to density distribution; The boundaries of each cluster are calculated using a convex hull algorithm. The boundaries of all clusters are integrated to form the overall boundary of the key fitting coefficient result set.

6. The optimization method for improving algorithm robustness in the single crystal silicon Czochralski method according to claim 5, characterized in that: The mapping model between the output variables of other process algorithms and the perception model is constructed, the mapping model is integrated into the original other process algorithms and end-to-end training is performed, and the regression coefficient result combination of the mapping model is obtained specifically includes: Constructing a mapping model between other process algorithm output variables and the perception model, the mapping model is shown in formula (2); in, is the output variable of other process algorithms, y is The true value of; k1, j1, k2 and j2 are the regression coefficients of the mapping model, and C is the constant term; The mapping model is integrated into other existing process algorithms, and end-to-end training is performed according to relevant training methods of other existing process algorithms to obtain a combination of regression coefficient results of mapping models k1, j1, k2 and j2.

7. The optimization method for improving algorithm robustness in the single crystal silicon Czochralski method according to claim 6, characterized in that: The method of using the perception model to obtain a combination of two key fitting coefficients and then judging whether the combination is within the overall boundary by using the overall boundary of the key fitting coefficient result set includes: using the perception model to obtain two key fitting coefficients θ and The combination is then judged whether the combination is within the overall boundary through the overall boundary of the key fitting coefficient result set; The mapping model is integrated into the original other process algorithms to obtain the output variables of the modified predicted output, including: after obtaining the output variables of the predicted output of the other process algorithms, the two key fitting coefficients θ and The combination of and the output variables of other process algorithm prediction outputs are substituted into the mapping model to obtain the output variables of the corrected prediction output.

8. An optimization device for improving algorithm robustness in a single crystal silicon Czochralski method, characterized in that: include: The process parameter data acquisition and processing module is used to acquire and process the relevant data of the process parameters in the temperature adjustment stage of the furnace for each rod pulling process of single crystal silicon, so as to obtain a data set of process parameters in the temperature adjustment stage; The perception model construction and data fitting module is used to construct the perception model and fit the data of each rod process parameter in the process parameter data set of the temperature adjustment stage to obtain the key fitting coefficient result set; The clustering and boundary acquisition module is used to cluster the key fitting coefficient result set and obtain the overall boundary of the key fitting coefficient result set; The mapping model construction and training module is used to construct a mapping model between the output variables of other process algorithms and the perception model, integrate the mapping model into the original other process algorithms and perform end-to-end training to obtain the regression coefficient result combination of the mapping model; The judgment and execution module is used to obtain the combination of two key fitting coefficients using the perception model after the previous set period ends at the beginning of the new temperature adjustment stage, and then judge whether the combination is within the overall boundary through the overall boundary of the key fitting coefficient result set. If not, the process parameters are adjusted to make it within the overall boundary. If so, the mapping model is integrated into the original other process algorithms to obtain the output variable of the corrected predicted output.

9. An electronic device, characterized in that: include: a memory storing a computer program; A processor is used to run the computer program, and when the computer program is running, the steps of the optimization method for improving the robustness of the algorithm in the single crystal silicon Czochralski method according to any one of claims 1 to 7 are executed.

10. A storage medium, characterized in that: The storage medium stores a computer program, and the computer program is configured to execute the steps of the optimization method for improving algorithm robustness in the single crystal silicon Czochralski method according to any one of claims 1 to 7 when running.

Images