Crystal bar resistivity control method and device, electronic equipment and storage medium

Abstract

The application discloses a crystal bar resistivity control method and device, electronic equipment and a storage medium. The crystal bar resistivity control method comprises the following steps: obtaining crystal bar production data of a current production period, wherein the crystal bar production data comprises a dopant feeding setting value, a process formula and production environment parameters; inputting the crystal bar production data into a resistivity prediction model to obtain a predicted resistivity of the current production period, wherein the resistivity prediction model is a deep learning model trained based on historical production data and a causal neural network, and the historical production data comprises crystal bar production data and corresponding crystal bar product resistivity in a preset period; obtaining a target resistivity of the current production period, determining a deviation value of the predicted resistivity and the target resistivity; and adjusting the dopant feeding setting value to a dopant feeding target value based on the deviation value. Through the crystal bar resistivity control method, the problem of low resistivity control precision in the crystal bar production process is solved.

Description

Technical Field

[0001] This application relates to the field of semiconductor processing equipment, and in particular to a method, apparatus, electronic device, and storage medium for controlling the resistivity of a crystal rod. Background Technology

[0002] Silicon single crystals are the core substrate material for solar cells and power semiconductor devices. Resistivity is a key electrical parameter of silicon single crystals, directly affecting device performance and yield. In the Czochralski (CZ) method of silicon single crystal growth, phosphorus (P) and antimony (Sb) co-doping techniques are commonly used to control resistivity distribution. Achieving satisfactory resistivity control with phosphorus and antimony co-doping is a challenging problem that urgently needs to be solved.

[0003] In related technologies, the doping amount is generally calculated based on empirical formulas using the previous feed amount and a fixed segregation coefficient. The segregation coefficient describes the proportion of material evenly distributed on both sides of the solid-liquid interface and is a parameter related to material solidification, crystal growth, and purification. However, in actual crystal growth, the segregation coefficient itself is affected by the equipment condition and crystal state, resulting in low accuracy of the doping amount calculated based on a fixed segregation coefficient, which in turn affects the resistivity stability of the produced crystal rod.

[0004] Due to the low resistivity control accuracy in related technologies, no effective solution has yet been proposed for crystal rod production. Summary of the Invention

[0005] To address the aforementioned technical problems, this application provides a method, apparatus, electronic device, and storage medium for controlling the resistivity of crystal rods, thereby solving the problem of low resistivity control accuracy in crystal rod production.

[0006] In a first aspect, a method for controlling the resistivity of a crystal rod is provided, the method comprising: Obtain the crystal rod production data for the current production cycle. The crystal rod production data includes dopant feeding settings, process formula, and production environment parameters. The production cycle is one or more furnaces of a single device, or one or more batches of multiple devices. The ingot production data is input into the resistivity prediction model to obtain the predicted resistivity of the current production cycle. The resistivity prediction model is a deep learning model trained based on historical production data and causal neural network. The historical production data includes ingot production data and corresponding ingot product resistivity within a preset time period. Obtain the target resistivity for the current production cycle, and determine the deviation between the predicted resistivity and the target resistivity; Based on the deviation value, the dopant feeding setting value is adjusted to the dopant feeding target value.

[0007] In one embodiment, the process formula includes raw material input data and process parameter data; the production environment parameters include thermal field status data and equipment status data.

[0008] In one embodiment, the step of inputting the crystal rod production data into the resistivity prediction model includes: acquiring historical production data and constructing a standard dataset based on the historical production data; determining a set of key feature variables and a causal topology graph between the set of key feature variables, the phosphorus-antimony segregation coefficient, and the crystal rod resistivity based on the standard dataset and a causal learning algorithm; and constructing the resistivity prediction model based on the set of key feature variables and the causal topology graph.

[0009] In one embodiment, acquiring historical production data and constructing a standard dataset based on the historical production data includes: grouping the historical production data according to the production cycle to obtain an original dataset; and performing missing value imputation, outlier removal, and data standardization on the original dataset to obtain the standard dataset.

[0010] In one embodiment, determining a set of key feature variables and a causal topology graph relating the set of key feature variables, the phosphorus-antimony segregation coefficient, and the resistivity of the crystal rod based on the standard dataset and the causal learning algorithm; constructing the resistivity prediction model based on the set of key feature variables and the causal topology graph includes: identifying a set of key feature variables related to the phosphorus-antimony segregation coefficient and the resistivity of the crystal rod in the standard dataset using a Bayesian Dirichlet equivalent uniform scoring algorithm and a greedy search algorithm, and establishing a causal topology graph; constructing an input layer based on the set of key feature variables; determining the connection relationships and connection weights between neurons in the hidden layer of the neural network model based on the constraints of the causal topology graph; constructing an output layer through an activation function; and using the predicted resistivity as the output result of the resistivity prediction model.

[0011] In one embodiment, after constructing the resistivity prediction model based on the set of key feature variables and the causal topology graph, the process includes: constructing a training set, a validation set, and a test set based on crystal growth data corresponding to the set of key feature variables in the standard dataset and the resistivity of crystal rod products; using the crystal growth data in the training set as input variables and the resistivity of the crystal rod products as labels, training the causal neural network model using a backpropagation algorithm to obtain a first prediction model; optimizing the hyperparameters of the target prediction model based on the validation set to obtain a second target model; verifying the prediction accuracy of the second prediction model based on the test set, and if the prediction accuracy meets a preset requirement, then using the second prediction model as the resistivity prediction model.

[0012] In one embodiment, after using the second prediction model as the resistivity prediction model, the method further includes: acquiring equipment maintenance information; determining equipment update nodes based on the equipment maintenance information; constructing an incremental dataset based on historical growth data after the equipment update nodes; and optimizing and training the resistivity prediction model based on the incremental dataset.

[0013] In one embodiment, adjusting the dopant feeding setting value to the dopant feeding target value based on the deviation value includes: obtaining the ingot production data of the previous production cycle, comparing the ingot production data of the current production cycle with the ingot production data of the previous production cycle; if they are different, then establishing a first optimization function based on the ingot production data, the predicted ingot resistivity data, and the target ingot resistivity data, with the target ingot resistivity data as the core constraint; if they are the same, then obtaining the dopant residue amount of the current furnace based on the dopant feeding amount, crystal length, and equipment running time of the current furnace, establishing a second optimization function based on the ingot production data, the predicted ingot resistivity data, the target ingot resistivity data, and the dopant residue amount, with the target ingot resistivity data as the core constraint; and solving for the dopant feeding target value that makes the deviation value between the predicted ingot resistivity data and the target ingot resistivity data within a preset range based on the first optimization function or the second optimization function, and adjusting the dopant feeding setting value to the dopant feeding target value.

[0014] Secondly, this application also provides a crystal rod resistivity control device, the device comprising: The acquisition module is used to acquire the crystal rod production data of the current production cycle. The crystal rod production data includes dopant feeding settings, process formula and production environment parameters. The production cycle is one or more furnaces of a single device, or one or more batches of multiple devices. The prediction module is used to input the crystal rod production data into the resistivity prediction model to obtain the predicted resistivity of the current production cycle. The resistivity prediction model is a deep learning model trained based on historical production data and causal neural network. The historical production data includes crystal rod production data within a preset time period and the corresponding crystal rod product resistivity. The calculation module is used to obtain the target resistivity of the current production cycle and determine the deviation between the predicted resistivity and the target resistivity. The control module is used to adjust the dopant feeding setting value to the dopant feeding target value based on the deviation value.

[0015] Thirdly, an electronic device is provided, including a memory and a processor, the memory storing a computer program, characterized in that the processor executes the computer program to implement the steps of the method as described in any embodiment of the first aspect above.

[0016] Fourthly, a computer-readable storage medium is provided, on which a computer program is stored, characterized in that the computer program, when executed by a processor, implements the steps of the method as described in any embodiment of the first aspect above.

[0017] The crystal rod resistivity control method provided in this application constructs a resistivity prediction model through historical production data of crystal rods and a causal neural network. Based on the crystal rod production data of the current production cycle, it predicts the resistivity of the crystal rod product, compares the predicted resistivity with the target resistivity, and adjusts the dopant feeding setting value of crystal rod production based on the deviation value between the predicted resistivity and the target resistivity. This method controls the resistivity of crystal rods from the source of production, which improves the control accuracy of crystal rod resistivity compared with the existing technology that calculates the dopant feeding amount based on a fixed segregation coefficient. Attached Figure Description

[0018] Figure 1 This is a schematic flowchart of a method for controlling the resistivity of a crystal rod according to an embodiment of this application; Figure 2 This is a schematic flowchart of a method for controlling the resistivity of a crystal rod according to another embodiment of this application; Figure 3 This is a schematic flowchart of a method for controlling the resistivity of a crystal rod according to another embodiment of this application; Figure 4 This is a schematic flowchart of a method for controlling the resistivity of a crystal rod according to another embodiment of this application; Figure 5 This is a schematic flowchart of a method for controlling the resistivity of a crystal rod according to another embodiment of this application; Figure 6 This is a schematic flowchart of a method for controlling the resistivity of a crystal rod according to another embodiment of this application; Figure 7 This is a schematic flowchart of a method for controlling the resistivity of a crystal rod according to another embodiment of this application; Figure 8 This is a schematic flowchart of a method for controlling the resistivity of a crystal rod according to another embodiment of this application; Figure 9 This is a comparison diagram of predicted resistivity and actual resistivity according to a specific embodiment of this application; Figure 10 This is a comparison diagram of doping resistivity distribution based on another specific embodiment and comparative example of this application; Figure 11 This is a schematic diagram of the structure of a crystal rod resistivity control device according to a specific embodiment of this application. Detailed Implementation

[0019] To make the above-mentioned objectives, features, and advantages of this application more apparent and understandable, the specific embodiments of this application are described in detail below with reference to the accompanying drawings. Many specific details are set forth in the following description to provide a thorough understanding of this application. However, this application can be implemented in many other ways different from those described herein, and those skilled in the art can make similar modifications without departing from the spirit of this application. Therefore, this application is not limited to the specific embodiments disclosed below.

[0020] Czochralski (CZ) method, a core process for producing high-purity monocrystalline silicon, plays a decisive role in controlling the electrical properties of silicon wafers through its doping stage. By adding dopants such as phosphorus and antimony, the resistivity of silicon rods can be effectively adjusted, thereby affecting the conversion efficiency and reliability of solar cells subsequently processed from silicon rod slices. Traditional doping processes mainly rely on manual experience to calculate the dopant dosage based on the amount of raw materials and a fixed segregation coefficient. However, many factors affect the resistivity of silicon rods during production, resulting in poor adaptability of the doping dosage calculation formula and poor control accuracy of the actual silicon rod resistivity based on the dopant dosage. Therefore, there is an urgent need for a dopant dosage calculation method that can improve the accuracy of silicon rod resistivity control.

[0021] To address the problem of low accuracy in controlling the resistivity of crystal rods in traditional technologies, this application provides a method for controlling the resistivity of crystal rods, such as... Figure 1 As shown, the method for controlling the resistivity of the crystal rod includes: Step S101: Obtain the crystal rod production data for the current production cycle. The crystal rod production data includes the dopant feeding setting value, process formula, and production environment parameters. The production cycle is one or more furnaces of a single device, or one or more batches of multiple devices.

[0022] Specifically, the Czochralski method is currently the mainstream method for producing monocrystalline silicon rods for semiconductors and photovoltaics. This method involves melting polycrystalline silicon in a single-crystal furnace and pulling the crystal rod by rotating the seed crystal. Rod production is often carried out in furnace runs or batches. A single furnace run refers to the growth of one rod in a single-crystal furnace. A batch refers to the production of rods using multiple pieces of equipment, such as single-crystal furnaces, with the same process and rod specifications. In mass production scenarios, rod production is often carried out in batches. The production cycle in this embodiment refers to the processing period for the same rod specifications; it can refer to a single furnace run or multiple furnace runs of a single piece of equipment, or a batch or multiple batches produced simultaneously by multiple pieces of equipment. Rod production data refers to various data involved in the rod production process. Among these, the dopant feeding setting value refers to the dosage determined based on the resistivity requirements and process formulation of the rod product. Furthermore, this dosage also includes dopant concentration and dopant weight. Dopant includes P-type dopant and N-type dopant. P-type dopant includes, but is not limited to, boron, aluminum, and gallium. N-type dopants include, but are not limited to, phosphorus, arsenic, antimony, and bismuth. Process formulation data refers to data related to the target ingot specifications within the current production cycle, such as raw material input data and corresponding processing parameters. Production environment parameters refer to equipment-related production data, such as thermal field status data and equipment status data. By acquiring ingot production data, multi-dimensional parameters related to this ingot production are obtained, providing data support for subsequent resistivity prediction.

[0023] Step S102: Input the crystal rod production data into the resistivity prediction model to obtain the predicted resistivity of the current production cycle. The resistivity prediction model is a deep learning model trained based on historical production data and causal neural network. The historical production data includes crystal rod production data and the corresponding crystal rod product resistivity within a preset time period.

[0024] Specifically, the resistivity of a crystal ingot product refers to the measured resistivity value corresponding to the crystal ingot product. This measured resistivity value includes the resistivity values ​​of one or more locations on the crystal ingot. For example, the measured resistivity value could be the head of a constant-diameter section of the crystal ingot, or it could include the resistivity test values ​​at locations such as the middle and tail of the crystal ingot. When the resistivity of the crystal ingot product includes measured resistivity values ​​from multiple locations on the crystal ingot, it can be used to evaluate the resistivity uniformity of the crystal ingot product, facilitating wafer grading after subsequent crystal ingot slicing. Historical production data includes crystal ingot production data from the crystal ingot production equipment prior to the current production cycle and the measured resistivity of the corresponding produced crystal ingot products. A deep learning model trained using historical production data and a causal neural network can deeply explore the causal relationship between various data variables in the historical production data and the resistivity of the crystal ingot product. The predicted resistivity obtained is determined based on the influence of multi-dimensional data variables. Compared to the prior art that only estimates resistivity based on the basic feed amount of raw materials and the dopant dosage, the predicted resistivity obtained in this embodiment is more accurate.

[0025] Step S103: Obtain the target resistivity of the current production cycle and determine the deviation between the predicted resistivity and the target resistivity.

[0026] Specifically, the target resistivity of the crystal ingots is determined based on the ingot production specifications within the production cycle. The target resistivity is the desired resistivity. The closer the predicted resistivity is to the target resistivity, the more the produced crystal ingot product meets expectations. By comparing the predicted resistivity and the target resistivity, the deviation between the predicted and target resistivity can be obtained.

[0027] Step S104: Adjust the dopant feeding setting value to the dopant feeding target value based on the deviation value.

[0028] Specifically, a deviation threshold can be set based on production quality requirements. The deviation value is compared with the deviation threshold. If the deviation value exceeds the threshold, it indicates that the quality of the crystal rod produced according to the current dopant feeding settings may not meet the standards, and adjustments are needed. The adjustment method can be based on matching the deviation value exceeding the deviation threshold with one of several pre-stored dopant feeding settings as the target dopant feeding value. The target dopant feeding value refers to the dopant feeding value that meets the expected resistivity of the crystal rod. Preferably, an optimization function can also be constructed based on a resistivity prediction model, and the target dopant feeding value can be determined through the optimization function.

[0029] The crystal rod resistivity control method of this application overcomes the limitations of fixed segregation coefficient by mining causal features and modeling neural networks with causal constraints, achieving high-precision prediction of resistivity. Based on the deviation between the predicted resistivity and the target resistivity, the dopant feeding setting value is adjusted, thus achieving precise control of crystal rod resistivity from the beginning of production.

[0030] In one embodiment, the process formulation includes raw material input data and process parameter data; the production environment parameters include thermal field status data and equipment status data.

[0031] Specifically, the raw material input data includes the amount of raw material input, the background resistivity of the raw material, the background doping concentration of the raw material, and the amount of raw material re-input. Process parameter data includes process setting data and real-time process data, specifically including crucible rotation speed, crystal rotation speed, crystal growth length, crystal pulling rate, argon pressure, argon flow rate, melt temperature, time from fusion to shoulder formation, and constant diameter growth rate. Thermal field status data includes the usage time of thermal field components, thermal field component replacement records, and thermal field attenuation coefficient. Equipment status data includes the cumulative runtime of the crystal pulling equipment, equipment hardware parameters, and equipment number. In this embodiment, the process formulation and production environment parameters are all process parameters highly correlated with the crystal rod growth process. By collecting the process formulation and production environment parameters, data analysis samples are provided for precise control of crystal rod resistivity, and data support is also provided for the prediction of crystal rod resistivity.

[0032] In one embodiment, such as Figure 2 As shown, before step S102 is executed, the following steps are also included: Step S201: Obtain historical production data and construct a standard dataset based on the historical production data.

[0033] Step S202: Based on the standard dataset and causal learning algorithm, determine the set of key feature variables and the causal topology diagram between the set of key feature variables, the phosphorus-antimony segregation coefficient and the resistivity of the crystal rod; construct a resistivity prediction model based on the set of key feature variables and the causal topology diagram.

[0034] Specifically, historical production data characterizes the growth of crystal rods within historical production cycles. Since historical production data includes multi-dimensional production data, data standardization is required before use to obtain a standard dataset. Based on the standard dataset and causal learning algorithms, key feature variable sets and causal topology graphs are mined. Causal learning algorithms are computational methods that automatically mine causal relationships between variables, quantify causal effects, and conduct interventions and counterfactual reasoning, starting from observational data, experimental data, or prior knowledge. Key feature variables refer to data features included in the standard dataset that have a causal relationship with crystal rod resistivity; the set of key feature variables is a collection of multiple key feature variables. The causal topology graph is a visualization model that integrates the system topology and causal relationships based on a directed acyclic graph (DAG). Processing the standard dataset using causal learning algorithms can mine variables in historical production data that have a causal relationship with the phosphorus-antimony segregation coefficient and crystal rod resistivity, and present them in the form of a causal topology graph. In traditional techniques, dopant dosage is calculated based on empirical formulas using the feed amount and a fixed segregation coefficient. For example, the formula is set by referring to the national standard "Conversion Procedure for Resistivity and Dopant Concentration of Boron-Doped, Phosphorus-Doped, and Arsenic-Doped Silicon Single Crystals" and combining it with actual production experience. However, the actual segregation coefficient fluctuates due to various factors, and these fluctuations are transmitted to the dopant calculation formula, leading to an overestimation or underestimation of the dopant content, ultimately resulting in a discrepancy between the ingot resistivity and the target resistivity. The ingot resistivity control method in this embodiment abandons the traditional empirical assumption of a fixed segregation coefficient. It uses a causal learning algorithm to analyze the impact of various production factors on the segregation coefficient during ingot production, identifying key characteristic variables that have a real causal influence on dopant segregation behavior and ingot resistivity, thus eliminating interference from spurious correlation features. Furthermore, a resistivity prediction model is constructed based on the set of key characteristic variables and a causal topology graph, enabling accurate prediction of the ingot resistivity for the current production cycle at the beginning of production. This provides real-time guidance for setting process parameters and dopant dosage, improving the control accuracy of ingot resistivity.

[0035] In one embodiment, such as Figure 3 As shown, step S201 above includes: Step S301: Group historical production data based on production cycle to obtain the original dataset.

[0036] Step S302: Perform missing value imputation, outlier removal, and data standardization on the original dataset to obtain a standard dataset.

[0037] In one specific embodiment, for the Czochralski silicon single crystal phosphorus-antimony co-doped production scenario, the process of constructing a standard dataset includes: acquiring full historical data of all single crystal furnaces in the plant area, which includes silicon material feeding data, dopant data, real-time process parameters, thermal field status data, equipment status data, historical batch data, and measured data of silicon single crystal resistivity of the corresponding produced crystal rods. The basic feed data includes polysilicon feed amount, polysilicon background resistivity, background doping concentration, and polysilicon refeed amount; dopant data includes: initial phosphorus dopant feed amount and antimony dopant feed amount per batch, and dopant purity; real-time process parameters include: crucible rotation speed, crystal rotation speed, real-time crystal growth length, crystal pulling rate, argon pressure, argon flow rate, melt temperature, time from melting to shoulder formation, and constant diameter growth rate; thermal field status data includes thermal field component usage time, thermal field component replacement record, and thermal field attenuation coefficient; equipment status data includes cumulative crystal pulling runtime, equipment hardware parameters, and equipment number; historical batch data includes: measured silicon single crystal resistivity values ​​of the previous batch / previous crystal on the same equipment, crucible retention amount of doped elements, and process parameter settings; measured silicon single crystal resistivity data includes the resistivity value of the head of the constant diameter section of the silicon single crystal. Data is grouped based on full historical data and production batches to construct an original dataset. This original dataset undergoes missing value imputation, outlier removal, and data standardization to create a preprocessed standard dataset. By constructing a standard dataset, multidimensional data can be integrated, improving the efficiency and accuracy of subsequent data analysis.

[0038] In one embodiment, such as Figure 4 As shown, step S202 further includes: Step S401: Identify the set of key feature variables related to the segregation coefficient of phosphorus and antimony and the resistivity of crystal rods in the standard dataset based on the Bayesian Dirichlet equivalent uniform scoring algorithm and the greedy search algorithm, and establish a causal topology graph.

[0039] Step S402: Construct an input layer based on the set of key feature variables, determine the connection relationships and connection weights between neurons in the hidden layer of the neural network model based on the constraints of the causal topology graph, construct an output layer through an activation function, and use the predicted resistivity as the output result of the resistivity prediction model.

[0040] Specifically, based on the preprocessed standard dataset, a causal structure learning algorithm is used to mine the causal relationship between each input variable and the phosphorus-antimony segregation coefficient and silicon single crystal resistivity, and to construct a causal topology graph between variables, thereby forming constraints on physical relationships.

[0041] Causal structure learning algorithms can be implemented using scoring functions, such as scoring optimization frameworks based on continuous optimization. This embodiment employs the Bayesian Dirichlet equivalent uniform scoring formula, which can accurately distinguish between direct causal relationships, indirect causal relationships, and spurious correlations between variables, thereby uncovering the correlations between crystal rod growth data and between crystal rod growth data and crystal rod resistivity.

[0042] The Bayesian Dirichlet equivalent uniform scoring formula with parent nodes is as follows:

[0043]

[0044] Where G is a directed acyclic graph (DAG), which is the causal structure of a Bayesian network; D is the observation dataset, containing N independent and identically distributed samples, each representing a parameter in a type of crystal rod growth data. ; parent node set Total number of configurations: When there is no parent node ; It is the Global Equivalent Sample Size (ESS), representing the strength of prior information; Discrete variables The number of value categories, ; For the first The node, the first Total number of samples under each parent configuration: ,satisfy ; For dataset D, and Take the sample count of the j-th configuration.

[0045] The Bayesian Dirichlet equivalent uniformity evaluation formula for nodes without parent nodes is as follows:

[0046] Whether to set a parent node can be chosen based on the data volume of the standard dataset and prior knowledge of crystal rod production.

[0047] Incremental frequency division calculation is performed based on the edge addition and deletion operations of the Greedy Search Algorithm (GES). Specifically, the Greedy Search Algorithm (GES) consists of a forward phase (FES, edge addition) and a backward phase (BES, edge deletion). Incremental score calculation means that each GES operation (adding / deleting an edge) only changes the parent node set of one node, therefore the change in score is minimal. Only the local score difference of this node needs to be recalculated; the global score does not need to be recalculated. Adding edges includes: giving Add a parent node ,

[0048] Edge deletion operations include: from Remove from parent node ,

[0049] The key feature variable with the optimal score is determined through a greedy iterative criterion. Specifically, GES selects the feature variable that yields the maximum positive score in each iteration. The operation continues until no operation can improve the score, and finally the equivalence class with the best score is obtained.

[0050] Based on the Bayesian Dirichlet equivalent uniformity scoring and greedy search algorithm, a causal topology graph is derived, ultimately determining the input feature set for the causal neural network model. This input feature set, i.e., the set of key feature variables, can be all parameters related to crystal growth contained in historical growth data, or one or more parameters selected from the crystal growth data based on the causal topology graph. Based on the key feature variable set selected in the steps and the causal topology graph, a neural network is constructed for each connection edge in the causal topology graph, thus building the causal neural network model. The model structure for each connection edge consists of an input layer, a hidden layer, and an output layer.

[0051] Each node in the input layer The input is equal to its parent node in the causal topology graph, i.e. Pa i It is the i-th parent node.

[0052] The hidden layer undergoes arbitrary nonlinear transformations, i.e.

[0053] in, yes The output; It is input; , It is a weight matrix. , It is a bias term; , It is an activation function. This involves the aggregation of parent node information. The output layer is built upon the activation function, i.e.,

[0054] in, Indicates that given input When, random variable Values The conditional probability; It is an activation function used to map the input to a probability distribution; Is with category The relevant weight matrix, It corresponds to the bias term. It is the prelude The output.

[0055] The design rule of the causal constraint hidden layer is as follows: the network connection weights strictly follow the constraints of the causal topology graph. Trainable connection weights exist between neurons only when two variables have a direct causal edge in the graph; neurons without a causal relationship are not connected. This design eliminates the interference of related features on the model, significantly improving its interpretability and generalization ability.

[0056] Model output settings: The model output is the predicted resistivity of the head of the target silicon single crystal.

[0057] In one embodiment, such as Figure 5 As shown, after step S202, the following steps are included: Step S501: Based on the crystal growth data corresponding to the key feature variable set in the standard dataset and the resistivity of the crystal rod product, construct the training set, validation set and test set.

[0058] Step S502: Based on the crystal growth data in the training set as input variables, and the resistivity of the crystal rod product as a label, the causal neural network model is trained using the backpropagation algorithm to obtain the first prediction model; the hyperparameters of the target prediction model are optimized based on the validation set to obtain the second target model; the prediction accuracy of the second prediction model is verified based on the test set. If the prediction accuracy meets the preset, the second prediction model is used as the resistivity prediction model.

[0059] Specifically, the standard dataset is divided into training, validation, and test sets in a ratio of 70%–80%: 10–15%: 10%–15%. Model training and optimization are then performed based on these sets. The training and optimization process includes: using the input feature set of the training set as input and the corresponding measured resistivity values ​​as labels, the causal neural network model is trained using the backpropagation algorithm to obtain the first prediction model; then, the hyperparameters of the first prediction model are optimized using the validation set to obtain the second prediction model; the prediction accuracy and generalization ability of the model are verified using the test set; when the prediction accuracy of the second prediction model on the test set meets a preset threshold, model training is complete, and the resistivity prediction model is obtained.

[0060] The resistivity prediction model of a causal neural network is evaluated using the mean decision error (MAE), calculated as follows:

[0061] in It is the target resistivity value. This is the predicted resistivity value. The number of samples.

[0062] In one embodiment, such as Figure 6 As shown, after step S502, the following steps are also included: Step S601: Obtain equipment maintenance information.

[0063] Step S602: Determine the equipment update node based on the equipment maintenance information.

[0064] Step S603: Construct an incremental dataset based on historical growth data after the device update node.

[0065] Step S604: Optimize and train the resistivity prediction model based on the incremental dataset.

[0066] Specifically, taking the silicon single crystal production scenario as an example, after the resistivity prediction model is trained and deployed, equipment maintenance information can be obtained after each batch or furnace of silicon rod growth. This equipment maintenance information includes equipment identification and maintenance details, such as equipment ID and maintenance logs. Based on this information, it can be determined whether the equipment has undergone an update. The equipment update node refers to the date or time of the equipment update, used to divide the production cycle before and after the update. If the current growth cycle is after the equipment update node, the full growth data of the current rods and the measured resistivity of the corresponding rod products are added to a new dataset. If the growth cycle of the corresponding data in the new dataset exceeds a preset number, this new dataset is used as an incremental dataset. The solidified resistivity prediction model is then fine-tuned based on the incremental dataset to quickly adapt to product changes or equipment modifications. This enables online self-updating of the model, replacing the original resistivity prediction model with the updated model for optimizing the doping amount in the next furnace. Based on the updated model, the causal verification of core features and the re-optimization of doping amount are automatically completed without the need for manual refitting of empirical formulas, realizing closed-loop intelligent optimization of resistivity control throughout the entire process.

[0067] In one embodiment, such as Figure 7 As shown, step S104 includes: Step S701: Obtain the ingot production data of the previous production cycle, and compare the ingot production data of the current production cycle with the ingot production data of the previous production cycle. If they are different, then take the ingot resistivity target data as the core constraint and establish a first optimization function based on the ingot production data, the ingot resistivity prediction data and the ingot resistivity target data.

[0068] Step S702: Based on the first optimization function, solve for the dopant feeding target value when the deviation between the predicted crystal resistivity data and the target crystal resistivity data is within a preset range, and adjust the dopant feeding setting value to the dopant feeding target value.

[0069] Specifically, it determines whether the current production cycle is a new production cycle, that is, whether the current processing furnace is a new furnace or whether the current processing batch is a new production batch, and constructs a first optimization function for the doping amount based on the determination result. The process of establishing the first optimization function includes: based on the resistivity prediction model, when the furnace is a new furnace, using the target resistivity as the core constraint, constructing the first optimization function:

[0070]

[0071]

[0072] in, The resistivity value of silicon single crystal is predicted by a causal neural network model. This refers to the amount of phosphorus dopant added. Let X be the amount of antimony dopant added, and let X be the set of other input features besides the dopant amount. The target resistivity is preset. , These represent the upper and lower limits of the phosphorus dopant dosage. , The upper and lower limits of the antimony dopant dosage are defined, and then iterative optimization is performed within the constraints of both phosphorus and antimony doping. Based on a global optimization algorithm, the first optimization function is solved iteratively to obtain the globally optimal initial phosphorus dopant dosage that satisfies the target resistivity requirement. With the optimal initial amount of antimony dopant The optimization algorithm can be one of the following: gradient descent iterative algorithm, particle swarm optimization algorithm, genetic algorithm, or least squares method.

[0073] In one embodiment, such as Figure 8 As shown, step S104 further includes: Step S801: Obtain the ingot production data of the previous production cycle, and compare the ingot production data of the current production cycle with the ingot production data of the previous production cycle. If they are the same, obtain the dopant residue of the current furnace based on the dopant feed amount, crystal length and equipment running time. With the ingot resistivity target data as the core constraint, establish a second optimization function based on the ingot production data, ingot resistivity prediction data, ingot resistivity target data and dopant residue.

[0074] Step S802: Based on the second optimization function, solve for the dopant feeding target value when the deviation between the predicted crystal resistivity data and the target crystal resistivity data is within a preset range, and adjust the dopant feeding setting value to the dopant feeding target value.

[0075] Specifically, taking silicon single crystal growth as an example, for continuous furnace / particle pulling production using the same equipment, the amount of phosphorus and antimony dopants remaining in the quartz crucible after the previous furnace on the same equipment is included in the calculation of the doping amount for a new furnace. , The dopant content is incorporated into the optimization calculation for this batch. The amount of dopant retained can be calculated based on the raw material input amount of the previous batch and the phosphorus-antimony segregation material balance formula. The revised second optimization function is:

[0076] in This is an input feature set that includes changes in process parameters, thermal field status updates, and equipment status updates for this furnace cycle. , ,in, The values ​​are the total phosphorus doping amount, total antimony doping amount, crystal length, and running time of the previous batch, respectively. The amounts of phosphorus and antimony dopants remaining in the quartz crucible are calculated as nonlinear functions of the total phosphorus doping amount, crystal length, and running time of the previous batch, respectively. The method for controlling the resistivity of the crystal rod in this embodiment further improves the accuracy of calculating the target dopant feeding value by incorporating the dopant retention amount from the previous batch / furnace.

[0077] In one specific embodiment, an original dataset is constructed based on all historical production data from 280 single-crystal furnaces over a three-month period. This historical production data includes: polysilicon feed rate (900-1000 kg / furnace), silicon resistivity, phosphorus / antimony dopant feed rate, crucible rotation (8-12 rpm), crystal rotation (5-10 rpm), crystal growth length, argon pressure (6-9 torr), argon flow rate (100-150 slm), hot zone graphite component usage time, cumulative equipment runtime, and the measured resistivity value of the previous furnace run. A resistivity prediction model is trained based on this original dataset after data standardization. Then, based on this resistivity prediction model and the ingot resistivity prediction method of this application, ingot resistivity prediction is performed, such as... Figure 9As shown in the comparison graph of predicted and actual resistivity, the model can accurately predict the resistivity of silicon rods, with a mean absolute error (MAE) of 0.05 Ω·cm on the test set. The training and test sets were obtained by regrouping data from the original dataset after data standardization. A perfect prediction line indicates that, under ideal conditions, the predicted resistivity is the same as the actual resistivity.

[0078] In another specific embodiment, resistivity tests were conducted over 1400 times in 70 single-crystal furnaces over a period of two months. The implementation conditions for the embodiment and the comparative example were the same: 70 single-crystal furnaces of the same model, and the quality of the polycrystalline silicon raw materials, the batches of quartz crucibles, the experimental environment, the personnel responsible for feeding and cleaning, the process standards followed, and the production specifications were all identical. Based on this, the embodiment used the ingot resistivity control method of this application to produce 11-inch single-crystal silicon rods, while the comparative example used a method based on manually determining the phosphorus and antimony doping setpoints to produce 11-inch single-crystal silicon rods. The comparative analysis results of the embodiment and the comparative example are shown in Table 1. Table 1. Comparison of test data between the examples and comparative examples.

[0079] For newly started furnace runs, the resistivity compliance rate of the embodiment within ±0.1 Ω·cm was 77%, 9% higher than the comparative example. For consecutive batches, the compliance rate within ±0.1 Ω·cm was 84%, 12% higher than the comparative example. The compliance rate within ±0.2 Ω·cm for consecutive batches reached as high as 98%, demonstrating extremely high resistivity compliance. A comparison of the doping resistivity distribution of the embodiment and the comparative example is shown in the figure below. Figure 10 As shown, the upper and lower bounds refer to the maximum and minimum values ​​within the allowable range of the target resistivity, respectively. Density refers to the data density per unit width within each interval of the histogram. Figure 10 It can be seen that the resistivity distribution of the embodiment is more concentrated and has a smaller proportion of outliers compared to the resistivity distribution of the comparative example, indicating that the resistivity results of the embodiment are more stable. In summary, the target resistivity achievement rate of the embodiment is significantly higher than that of the comparative example, corresponding to an increase in effective output.

[0080] The crystal rod resistivity control method of this application can significantly improve the accuracy of crystal rod resistivity prediction and the adaptability of doping amount, thereby significantly increasing the yield of crystal rods meeting the target. It can significantly improve the stability of crystal rod resistivity targeting in continuous production and eliminate the lag in optimization. Furthermore, the resistivity prediction model of this application has strong interpretability and good industrial applicability.

[0081] This application also provides a crystal rod resistivity control device, such as... Figure 11 As shown, the crystal rod resistivity control device includes: The acquisition module 10 is used to acquire the crystal rod production data of the current production cycle. The crystal rod production data includes the dopant feeding setting value, process formula and production environment parameters. The production cycle is one or more furnaces of a single device, or one or more batches of multiple devices. Prediction module 20 is used to input crystal rod production data into resistivity prediction model to obtain the predicted resistivity of the current production cycle. The resistivity prediction model is a deep learning model trained based on historical production data and causal neural network. The historical production data includes crystal rod production data within a preset time period and the corresponding crystal rod product resistivity. The calculation module 30 is used to obtain the target resistivity of the current production cycle and determine the deviation between the predicted resistivity and the target resistivity. The control module 40 is used to adjust the dopant feeding setpoint to the dopant feeding target value based on the deviation value.

[0082] In one embodiment, the process formulation includes raw material input data and process parameter data; the production environment parameters include thermal field status data and equipment status data.

[0083] In one embodiment, before inputting the crystal rod production data into the resistivity prediction model, the process includes: acquiring historical production data and constructing a standard dataset based on the historical production data; determining a set of key feature variables and a causal topology graph between the set of key feature variables, the phosphorus-antimony segregation coefficient, and the crystal rod resistivity based on the standard dataset and a causal learning algorithm; and constructing a resistivity prediction model based on the set of key feature variables and the causal topology graph.

[0084] In one embodiment, acquiring historical production data and constructing a standard dataset based on the historical production data includes: grouping the historical production data according to the production cycle to obtain the original dataset; and performing missing value imputation, outlier removal, and data standardization on the original dataset to obtain the standard dataset.

[0085] In one embodiment, based on a standard dataset and a causal learning algorithm, a set of key feature variables and a causal topology graph relating the set of key feature variables, the phosphorus-antimony segregation coefficient, and the resistivity of the crystal rod are determined. Constructing a resistivity prediction model based on the set of key feature variables and the causal topology graph includes: identifying the set of key feature variables related to the phosphorus-antimony segregation coefficient and the resistivity of the crystal rod in the standard dataset using a Bayesian Dirichlet equivalent uniform scoring algorithm and a greedy search algorithm, and establishing a causal topology graph; constructing an input layer based on the set of key feature variables; determining the connection relationships and connection weights between neurons in the hidden layer of the neural network model based on the constraints of the causal topology graph; constructing an output layer through an activation function; and using the predicted resistivity as the output of the resistivity prediction model.

[0086] In one embodiment, after constructing a resistivity prediction model based on a set of key feature variables and a causal topology graph, the process includes: constructing a training set, a validation set, and a test set based on crystal growth data corresponding to the set of key feature variables in a standard dataset and the resistivity of crystal rod products; using the crystal growth data in the training set as input variables and the resistivity of crystal rod products as labels, training the causal neural network model using a backpropagation algorithm to obtain a first prediction model; optimizing the hyperparameters of the target prediction model based on the validation set to obtain a second target model; and verifying the prediction accuracy of the second prediction model based on the test set. If the prediction accuracy meets the preset requirements, the second prediction model is used as the resistivity prediction model.

[0087] In one embodiment, after using the second prediction model as the resistivity prediction model, the method further includes: acquiring equipment maintenance information; determining equipment update nodes based on the equipment maintenance information; constructing an incremental dataset based on historical growth data after the equipment update nodes; and optimizing and training the resistivity prediction model based on the incremental dataset.

[0088] In one embodiment, the control module 40 is further configured to acquire the ingot production data of the previous production cycle, compare the ingot production data of the current production cycle with the ingot production data of the previous production cycle, and if they are different, then with the ingot resistivity target data as the core constraint, a first optimization function is established based on the ingot production data, the ingot resistivity prediction data, and the ingot resistivity target data; if they are the same, then with the dopant feed amount, crystal length, and equipment running time of the current furnace, the dopant residue amount of the current furnace is acquired, and with the ingot resistivity target data as the core constraint, a second optimization function is established based on the ingot production data, the ingot resistivity prediction data, the ingot resistivity target data, and the dopant residue amount; based on the first optimization function or the second optimization function, a dopant feed target value is calculated that makes the deviation between the ingot resistivity prediction data and the ingot resistivity target data within a preset range, and the dopant feed setting value is adjusted to the dopant feed target value.

[0089] This application provides an electronic device, including a memory and a processor, wherein the memory stores a computer program, characterized in that the processor executes the computer program to implement the steps of the ingot resistivity control method as described in any of the above embodiments.

[0090] This application also provides a computer-readable storage medium having a computer program stored thereon, characterized in that the computer program, when executed by a processor, implements the steps of the ingot resistivity control method as described in any of the above embodiments.

[0091] The technical features of the above embodiments can be combined in any way. For the sake of brevity, not all possible combinations of the technical features in the above embodiments are described. However, as long as there is no contradiction in the combination of these technical features, they should be considered to be within the scope of this specification.

[0092] The embodiments described above are merely illustrative of several implementation methods of this application, and while the descriptions are relatively specific and detailed, they should not be construed as limiting the scope of the invention patent. It should be noted that those skilled in the art can make various modifications and improvements without departing from the concept of this application, and these all fall within the protection scope of this application. Therefore, the protection scope of this invention patent should be determined by the appended claims.

Claims

1. A method for controlling the resistivity of a crystal rod, characterized in that, The method includes: Obtain the crystal rod production data for the current production cycle. The crystal rod production data includes dopant feeding settings, process formula, and production environment parameters. The production cycle is one or more furnaces of a single device, or one or more batches of multiple devices. The ingot production data is input into the resistivity prediction model to obtain the predicted resistivity of the current production cycle. The resistivity prediction model is a deep learning model trained based on historical production data and causal neural network. The historical production data includes ingot production data and corresponding ingot product resistivity within a preset time period. Obtain the target resistivity for the current production cycle, and determine the deviation between the predicted resistivity and the target resistivity; Based on the deviation value, the dopant feeding setting value is adjusted to the dopant feeding target value.

2. The method for controlling the resistivity of a crystal rod according to claim 1, characterized in that, The process formula includes raw material input data and process parameter data; the production environment parameters include thermal field status data and equipment status data.

3. The method for controlling the resistivity of a crystal rod according to claim 1, characterized in that, Before inputting the crystal rod production data into the resistivity prediction model, the following steps are included: Acquire historical production data and construct a standard dataset based on the historical production data; Based on the standard dataset and causal learning algorithm, a set of key feature variables and a causal topology graph relating the set of key feature variables, the phosphorus-antimony segregation coefficient, and the resistivity of the crystal rod are determined; the resistivity prediction model is constructed based on the set of key feature variables and the causal topology graph.

4. The method for controlling the resistivity of a crystal rod according to claim 3, characterized in that, The process of acquiring historical production data and constructing a standard dataset based on the historical production data includes: The historical production data is grouped based on the production cycle to obtain the original dataset; The original dataset is subjected to missing value imputation, outlier removal, and data standardization to obtain the standard dataset.

5. The method for controlling the resistivity of a crystal rod according to claim 3, characterized in that, Based on the standard dataset and causal learning algorithm, a set of key feature variables and a causal topology diagram between the set of key feature variables, the phosphorus-antimony segregation coefficient and the crystal rod resistivity are determined. The resistivity prediction model is constructed based on the set of key feature variables and the causal topology graph, including: The set of key feature variables related to the segregation coefficient of phosphorus and antimony and the resistivity of crystal rods in the standard dataset were identified using the Bayesian Dirichlet equivalent uniform scoring algorithm and the greedy search algorithm, and a causal topology graph was established. An input layer is constructed based on the set of key feature variables. The connection relationships and connection weights between neurons in the hidden layer of the neural network model are determined based on the constraints of the causal topology graph. An output layer is constructed through an activation function, and the predicted resistivity is used as the output result of the resistivity prediction model.

6. The method for controlling the resistivity of a crystal rod according to claim 3, characterized in that, After constructing the resistivity prediction model based on the set of key feature variables and the causal topology graph, the process includes: Training, validation and test sets are constructed based on crystal growth data and resistivity of crystal rod products corresponding to the set of key feature variables in the standard dataset. Using the crystal growth data in the training set as input variables and the resistivity of the crystal rod product as a label, the causal neural network model is trained using the backpropagation algorithm to obtain the first prediction model. The hyperparameters of the target prediction model are optimized based on the validation set to obtain a second target model; The prediction accuracy of the second prediction model is verified based on the test set. If the prediction accuracy meets the preset requirement, the second prediction model is used as the resistivity prediction model.

7. The method for controlling the resistivity of a crystal rod according to claim 6, characterized in that, The step of using the second prediction model as the resistivity prediction model further includes: Obtain equipment maintenance information; The equipment update node is determined based on the equipment maintenance information; An incremental dataset is constructed based on historical growth data after the device update node. The resistivity prediction model is optimized and trained based on the incremental dataset.

8. The method for controlling the resistivity of a crystal rod according to claim 1, characterized in that, The step of adjusting the dopant doping setpoint to the dopant doping target value based on the deviation value includes: Obtain the ingot production data from the previous production cycle and compare the ingot production data of the current production cycle with the ingot production data of the previous production cycle to see if they are the same. If they are different, then the target resistivity data of the crystal rod is used as the core constraint, and a first optimization function is established based on the crystal rod production data, the predicted resistivity data of the crystal rod, and the target resistivity data of the crystal rod. If they are the same, the residual amount of dopant in the current furnace is obtained based on the dopant feed amount, crystal length and equipment running time of the current furnace. The target data of the crystal rod resistivity is used as the core constraint. A second optimization function is established based on the crystal rod production data, the predicted data of the crystal rod resistivity, the target data of the crystal rod resistivity and the residual amount of dopant. Based on the first optimization function or the second optimization function, the target value for dopant feeding is calculated so that the deviation between the predicted data of the crystal rod resistivity and the target data of the crystal rod resistivity is within a preset range. The dopant feeding setting value is then adjusted to the target value for dopant feeding.

9. A crystal rod resistivity control device, characterized in that, The device includes: The acquisition module is used to acquire the crystal rod production data of the current production cycle. The crystal rod production data includes dopant feeding settings, process formula and production environment parameters. The production cycle is one or more furnaces of a single device, or one or more batches of multiple devices. The prediction module is used to input the crystal rod production data into the resistivity prediction model to obtain the predicted resistivity of the current production cycle. The resistivity prediction model is a deep learning model trained based on historical production data and causal neural network. The historical production data includes crystal rod production data within a preset time period and the corresponding crystal rod product resistivity. The calculation module is used to obtain the target resistivity of the current production cycle and determine the deviation between the predicted resistivity and the target resistivity. The control module is used to adjust the dopant feeding setting value to the dopant feeding target value based on the deviation value.

10. An electronic device comprising a memory and a processor, wherein the memory stores a computer program, characterized in that, When the processor executes the computer program, it implements the steps of the method according to any one of claims 1 to 8.

11. A computer-readable storage medium having a computer program stored thereon, characterized in that... When the computer program is executed by a processor, it implements the steps of the method according to any one of claims 1 to 8.

Images