Model training method, silicon wafer manufacturing method, device, storage medium and equipment

Abstract

The invention provides a model training method and device, a silicon wafer manufacturing method and device, a storage medium and equipment. The model training method comprises the steps that a first verification set and a first training set are determined from training samples; determining a second training set and a second verification set from the first training set, and performing model training on the initial model based on the first training set to obtain an alternative model of the current iteration round; verifying the alternative model through the first verification set and the second verification set to obtain a first verification result and a second verification result; and based on the second verification result, updating the first training set to obtain an updated first training set, and based on the updated first training set, returning to the step of performing model training on the initial model based on the first training set until the initial model converges, and determining the alternative model of which the first verification result meets a preset condition in the plurality of alternative models as a target model. According to the method, the model for predicting the quality of the silicon wafer can be accurately trained from the data with more noises.

Description

Technical Field

[0001] This invention relates to the field of semiconductor technology, specifically to a model training method, a silicon wafer manufacturing method, an apparatus, a storage medium, and a device. Background Technology

[0002] In the silicon wafer manufacturing process, the state of the production equipment will slowly drift. For example, in continuous production, due to the inherent errors of the equipment and the cumulative effects of internal and external disturbances (such as changes in heat load, fluctuations in abrasive concentration, etc.), the production equipment cannot maintain an absolutely stable set working point for a long time, which will cause the quality parameters of the produced silicon wafers to deviate.

[0003] With the development of artificial intelligence technology, using machine learning algorithms to build silicon wafer quality prediction models to optimize machine adjustment parameters is a good way to solve the above problems. However, the training data collected in industrial silicon wafer production sites is often of poor quality, containing a large number of outliers and noise. If these data are used directly to train the model, the model will learn incorrect patterns, which will not guarantee the accuracy of the silicon wafer quality prediction model and thus will prevent the effective optimization of machine adjustment parameters. Summary of the Invention

[0004] In view of this, embodiments of the present invention aim to provide a model training method, manufacturing method, apparatus, storage medium, and electronic device to solve the problem that the prediction accuracy of silicon wafer quality prediction models cannot be guaranteed in the prior art.

[0005] This invention provides a model training method, which includes: Acquire training samples, and determine a first validation set and a first training set from the training samples; the training samples include data recorded during the silicon wafer manufacturing process; the first validation set consists of data in the training samples whose noise meets a preset noise condition; A second training set and a second validation set are determined from the first training set, and the initial model is trained based on the first training set to obtain the candidate model corresponding to the current iteration round of the initial model; The candidate model is validated using the first validation set to obtain a first validation result, and the candidate model is validated using the second validation set to obtain a second validation result. Based on the second verification result, the first training set is updated to obtain an updated first training set. Based on the updated first training set, the step of training the initial model based on the first training set to obtain the candidate model corresponding to the current iteration round is returned to be executed until the initial model converges, and multiple candidate models are obtained. The candidate model whose first verification result meets the preset conditions among the multiple candidate models is determined as the target model. The target model is used to predict the quality parameters of the silicon wafer after production based on the initial parameters of the input silicon wafer and the machine adjustment parameters of the production machine.

[0006] In one embodiment, obtaining training samples includes: Obtain multiple data samples recorded during the silicon wafer manufacturing process; For each of the plurality of data samples, a score value is calculated for the data sample using a preset scoring algorithm. The score value represents the production quality of the silicon wafer by the production machine after the parameters of the production machine are adjusted based on the data sample. The weight of the data sample is determined based on the score value, and the score value is positively correlated with the weight. The plurality of data samples and the weight of each data sample are determined as the training samples.

[0007] In one embodiment, the second validation set includes a second input sample and a second target sample corresponding to the second input sample. The step of validating the candidate model using the second validation set to obtain a second validation result includes: The second input sample is predicted using the alternative model to obtain a second prediction result; The second prediction result is calculated using the preset scoring algorithm to obtain a second output score value; The second output score is compared with the score corresponding to the second target sample, and the comparison result is determined as the second verification result.

[0008] In one embodiment, updating the first training set based on the second verification result to obtain an updated first training set includes: If the second output score is greater than the score corresponding to the second target sample, then the second target sample in the second validation set of the first training set is replaced with the second prediction result to obtain the updated first training set.

[0009] In one embodiment, replacing the second target sample of the second validation set in the first training set with the second prediction result to obtain the updated first training set includes: The second target sample in the second validation set of the first training set is replaced with the second prediction result to obtain the updated second validation set. According to a specified ratio, a sample of target machine adjustment parameters is determined from the updated second verification set; The target machine adjustment parameter samples in the updated second verification set are enhanced by a specified multiplier to obtain the updated first training set.

[0010] In one embodiment, the target machine adjustment parameter samples include positive samples and negative samples. The positive samples are used for positive adjustment of the production machine, and the negative samples are used for negative adjustment of the production machine. The step of enhancing the target machine adjustment parameter samples in the updated second validation set by a specified factor to obtain the updated first training set includes: The negative samples in the target machine adjustment parameter sample are enhanced by a specified factor to obtain the updated first training set.

[0011] In one embodiment, updating the first training set based on the second verification result to obtain an updated first training set includes: If the second output score is greater than the score corresponding to the second target sample, then obtain the update probability; Based on the updated probability, the second target sample in the second validation set of the first training set is replaced with the second prediction result to obtain the updated first training set.

[0012] In one embodiment, the first verification set includes a first input sample and a first target sample corresponding to the first input sample, and the step of obtaining the update probability includes: For each of the candidate models, the first input sample is predicted using the candidate model to obtain a first prediction result; Calculate the mean absolute error between the first prediction result and the first target sample to obtain the mean absolute error corresponding to the alternative model; The first prediction result is calculated using the preset scoring algorithm to obtain the first output score value corresponding to the candidate model; The update probability is determined based on the first output score and mean absolute error corresponding to each of the candidate models.

[0013] In one embodiment, the step of calculating the second prediction result using the preset scoring algorithm to obtain the second output score value includes: Calculate the difference between the second target sample and the second input sample to obtain the first difference; Calculate the difference between the second prediction result and the second input sample to obtain the second difference; The second output score is determined based on the first difference and the second difference, wherein the absolute value of the difference between the first difference and the second difference is negatively correlated with the second output score.

[0014] In one embodiment, training the initial model based on the first training set includes: Based on the weights corresponding to each data sample in the first training set, the fitting priority corresponding to each data sample is determined, wherein the weights are positively correlated with the fitting priority; Based on the fitting priority and the first training set, the initial model is trained.

[0015] In one embodiment, the first validation set includes a first input sample and a first target sample corresponding to the first input sample. The step of validating the candidate model using the first validation set to obtain a first validation result includes: The first input sample is predicted using the alternative model to obtain a first prediction result; The average absolute error between the first prediction result and the first target sample is determined as the first verification result.

[0016] In one embodiment, determining the candidate model whose first verification result among the plurality of candidate models meets preset conditions as the target model includes: The candidate model with the smallest average absolute error among the multiple candidate models is determined as the target model.

[0017] Another aspect of the present invention provides a method for manufacturing a silicon wafer, the method comprising: Obtain the initial parameters of the silicon wafer to be manufactured and the machine adjustment parameters of the production equipment; The initial parameters and the machine adjustment parameters are input into the target model above for prediction to obtain the quality parameters of the silicon wafer after production. If the quality parameters do not meet the preset process requirements, then new machine adjustment parameters are determined based on the quality parameters. The production machine is adjusted according to the new machine adjustment parameters so as to manufacture the silicon wafer to be manufactured based on the adjusted production machine.

[0018] Another aspect of the present invention provides a model training apparatus, the model training apparatus comprising: A first acquisition module is used to acquire training samples and determine a first validation set and a first training set from the training samples; the training samples include data recorded during the silicon wafer manufacturing process; the first validation set is data in the training samples whose noise meets a preset noise condition; The training module is used to determine a second training set and a second validation set from the first training set, and to train the initial model based on the first training set to obtain the candidate model corresponding to the current iteration round of the initial model. The verification module is used to verify the candidate model using the first verification set to obtain a first verification result, and to verify the candidate model using the second verification set to obtain a second verification result; The update module is used to update the first training set based on the second verification result to obtain the updated first training set, and based on the updated first training set, return to the step of training the initial model based on the first training set to obtain the candidate model corresponding to the current iteration round of the initial model, until the initial model converges and multiple candidate models are obtained. The determination module is used to determine the candidate model whose first verification result meets the preset conditions among the multiple candidate models as the target model. The target model is used to predict the quality parameters of the silicon wafer after production based on the initial parameters of the input silicon wafer and the machine adjustment parameters of the production machine.

[0019] In another aspect, the present invention provides a silicon wafer manufacturing apparatus, the silicon wafer manufacturing apparatus comprising: The second acquisition module is used to acquire the initial parameters of the silicon wafer to be manufactured and the machine adjustment parameters of the production machine. The prediction module is used to input the initial parameters and the machine adjustment parameters into the target model above for prediction, so as to obtain the quality parameters of the silicon wafer after production; The parameter update module is used to determine new machine adjustment parameters based on the quality parameters if the quality parameters do not meet the preset process requirements. An adjustment module is used to adjust the parameters of the production machine according to the new machine adjustment parameters, so as to manufacture the silicon wafer to be manufactured based on the adjusted production machine.

[0020] In another aspect, the present invention provides a computer-readable storage medium having stored thereon computer-executable instructions, which, when executed by a processor, implement the model training method or silicon wafer manufacturing method as described in any of the above embodiments.

[0021] In another aspect, the present invention provides an electronic device, the electronic device comprising: processor; Memory used to store the processor's executable instructions; The processor is used to execute the model training method or silicon wafer manufacturing method described in any of the above embodiments.

[0022] Compared with related technologies, the model training method provided by this invention has the following beneficial effects: The model training method provided by this invention involves acquiring training samples and determining a first validation set and a first training set from the training samples. The training samples include data recorded during the silicon wafer manufacturing process. The first validation set consists of training samples containing data whose noise meets preset noise conditions. Then, a second training set and a second validation set are determined from the first training set. Based on the first training set, the initial model is trained to obtain a candidate model corresponding to the current iteration round. The candidate model is validated using the first validation set to obtain a first validation result, and the candidate model is validated using the second validation set to obtain a second validation result. Based on the second validation result, the first training set is updated to obtain an updated first training set. Based on the updated first training set, the process of training the initial model using the first training set to obtain a candidate model corresponding to the current iteration round is repeated until the initial model converges, resulting in multiple candidate models. Finally, the candidate model whose first validation result meets the preset conditions is determined as the target model. The target model is used to predict the quality parameters of the silicon wafer after production based on the initial parameters of the input silicon wafer and the machine adjustment parameters of the production machine. In other words, during training, the first training set can be updated using the second validation results to continuously generate new and higher-quality first training sets. This achieves iterative data purification of the first training set, enabling the automatic and efficient training of a stable, reliable, and high-precision target model from noisy industrial data. This target model can then be used to effectively optimize machine adjustment parameters. Furthermore, by utilizing a lower-noise first validation set and selecting a model with higher prediction accuracy from multiple iterations as the target model, the production quality of silicon wafers can be effectively improved. Attached Figure Description

[0023] Figure 1 The diagram shown is a flowchart of a model training method provided in an embodiment of the present invention.

[0024] Figure 2 The diagram shown is a flowchart illustrating the specific implementation of steps 110 to 150 according to an embodiment of the present invention.

[0025] Figure 3 The diagram shown is a flowchart illustrating the specific implementation of the step "updating the second verification set" according to an embodiment of the present invention.

[0026] Figure 4 The diagram shown is a flowchart illustrating the specific implementation of the "data augmentation" step according to an embodiment of the present invention.

[0027] Figure 5 The diagram shown is a schematic flow chart of a silicon wafer manufacturing method according to an embodiment of the present invention.

[0028] Figure 6 The figure shown is a GBIR-SPC control diagram obtained by adjusting the target model according to an embodiment of the present invention.

[0029] Figure 7 The diagram shown is a GBIR-SPC control diagram obtained by an engineer through machine adjustment, according to an embodiment of the present invention.

[0030] Figure 8 The diagram shown is a schematic block diagram of a model training device provided in an embodiment of the present invention.

[0031] Figure 9 The diagram shown is a schematic block diagram of a silicon wafer manufacturing apparatus provided in an embodiment of the present invention.

[0032] Figure 10 The diagram shown is a block diagram of an electronic device provided in an embodiment of the present invention. Detailed Implementation

[0033] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0034] In the industrial production of silicon wafers, the equipment can experience state drift as operating time increases, leading to fluctuations in the quality parameters of the produced wafers and compromising production quality. Therefore, engineers often need to adjust the parameters of the production equipment in real time during silicon wafer production. However, in actual production, engineers' parameter adjustments are often not timely or accurate enough, resulting in inconsistent silicon wafer quality.

[0035] Some related technologies propose using machine learning algorithms to construct silicon wafer quality prediction models to optimize machine adjustment parameters. However, due to the influence of process experiments, production environment, and other factors, the training data collected on-site in silicon wafer production is often of poor quality, usually containing a large number of outliers and noise. If these data are used directly to train the model, the model will learn incorrect patterns, and the accuracy of the silicon wafer quality prediction model cannot be guaranteed.

[0036] Furthermore, silicon wafer production requires various processes, and a single machine adjustment parameter cannot meet the diverse process requirements. For example, the Global Backsurface-referenced Ideal plane / Range (GBIR), a flatness evaluation indicator, specifically focuses on the difference between the maximum and minimum thickness points on the silicon wafer. However, Double Side Polishing (DSP) removes the thickness across the entire silicon wafer, making it impossible to precisely adjust the maximum and minimum points. This results in many measured quality parameters often not having a good match with a single machine adjustment parameter, leading to poor quality of the training samples determined on-site.

[0037] In addition, on-site machine adjustments are often minor and can mislead the model into adjusting a certain parameter to guide the results to improve monotonically. This often results in the model only being able to provide a direction for adjustment (i.e., whether to increase or decrease a certain parameter) and not being able to provide specific parameter adjustment amounts. Therefore, the machine adjustment parameters determined on-site cannot serve as good training samples, resulting in poor quality training samples.

[0038] In view of the above problems, one embodiment of the present invention provides a model training method, which can be executed by a computer device (e.g., a server or a user terminal). Figure 1 As shown, the model training method may include: 110. Obtain training samples and determine the first validation set and the first training set from the training samples; the training samples include data recorded during the silicon wafer manufacturing process; the first validation set is the data in the training samples whose noise meets the preset noise conditions.

[0039] The training samples can be data obtained by recording the silicon wafer parameters and the machine parameters of the production equipment during the silicon wafer manufacturing process.

[0040] Understandably, there can be multiple training samples, each of which can be considered an (X, Y) pair. Here, X represents the input features during model training, describing the wafer's process state before setup and the adjustments performed by the production machine. For example, input features describing the wafer's process state before setup might include GBIR, Site Front Least-squares Range (SFQR), Edge Site Front Least-squares Range (ESFQR), Edge Site Flatness Quality Deviation (ESFQD), Nanotopography (Nano), global parameters, and edge parameters. Input features describing the adjustments performed by the production machine might include changes in target thickness (Thk), upper platen pressure, lower platen pressure, upper platen rotation speed, and lower platen rotation speed. Global parameters are composite indicators describing the overall macroscopic shape of the wafer. Its main purpose is to quantify the thickness difference between the center and edge regions of a silicon wafer, i.e., whether the wafer is "convex," "concave," or "flat." For example, a large positive global parameter usually indicates that the wafer is thicker at the center and thinner at the edges (like a small hill). A large negative global parameter usually indicates that the wafer is thinner at the center and thicker at the edges (like a small bowl). A global parameter close to zero indicates that the thickness is the same at the center and the edges, and the overall shape is flat. In practical applications, when the final quality parameters (such as GBIR) deteriorate, the location of the problem can be determined by checking the sign and magnitude of the global parameters. Among them, the edge parameter is a composite index specifically used to describe the shape accuracy of the outermost region of the silicon wafer.

[0041] Here, Y represents the output label during model training, which describes the change in the quality of the produced silicon wafer after performing the aforementioned adjustments. For example, the output label may include the predicted GBIR, predicted SFQR, predicted ESFQR, and predicted ESFQD after processing. Furthermore, the model can determine the changes in GBIR, SFQR, ESFQR, and ESFQD of the processed silicon wafer relative to the unprocessed silicon wafer based on the predicted GBIR, predicted SFQR, predicted ESFQR, predicted ESFQD, and the GBIR, SFQR, ESFQR, and ESFQD features in the input data.

[0042] In some implementations, the specific implementation of the step "determine the first validation set and the first training set from the training samples" may include: According to the first preset ratio, the first validation set and the first training set are determined from the training samples.

[0043] For example, if the first preset ratio represents a 1:9 ratio between the first validation set and the first training set, then 10% of the data from the training samples can be selected as the first validation set, and the remaining 90% of the data can be selected as the first training set. The first validation set can be a "golden validation set" selected from the training samples. This first validation set can be used as an objective standard to measure the model's generalization ability. In this embodiment, the first validation set will be used as an unmodified validation set to verify the model's performance. The training samples in the first validation set can be the most reliable and noise-free data points selected from historical data of silicon wafer manufacturing processes. These data can accurately reflect the relationship between process parameter adjustments and quality changes.

[0044] In some implementations, a first validation set can be determined from the training samples by determining whether the noise in the data in the training samples meets a preset noise condition. Optionally, if the data in the training samples is determined to meet the preset noise condition, it can be determined that the data meets the following criteria: data completeness (e.g., no missing values), reasonable values ​​(e.g., all values ​​are within the physically possible range), and accurate measurement (e.g., the measurement uncertainty of key parameters is below a preset threshold).

[0045] In some implementations, the specific implementation of obtaining training samples in step 110 may include: 111. Obtain multiple data samples recorded during the silicon wafer manufacturing process.

[0046] In this embodiment, the data sample can be equivalent to the (X, Y) pair described above.

[0047] 112. For each data sample among multiple data samples, calculate the score value of the data sample using a preset scoring algorithm. The score value represents the production quality of silicon wafers produced by the production machine after the parameters of the production machine are adjusted based on the data sample.

[0048] For example, taking a data sample that includes the GBIR of the silicon wafer before DSP, the actual GBIR of the silicon wafer after DSP, and the expected estimated GBIR of the silicon wafer after DSP, when the GBIR of the silicon wafer after DSP is too large (e.g., the actual GBIR of the silicon wafer after DSP is 109nm and the estimated GBIR is 80nm), the relationship between the expected quality parameter and the actual tuning effect can be calculated as: r = (actual GBIR - GBIR) / (estimated GBIR - GBIR).

[0049] Understandably, when parameter r is negative, it indicates that the actual optimization or degradation of GBIR is opposite to the expected amount, suggesting poor quality of the current data sample. When parameter r is greater than 0 and less than 1, it indicates that the machine adjustment parameters corresponding to the DSP are effective but have not achieved the expected results, suggesting high quality of the current data sample. When parameter r is 1, it indicates that the machine adjustment parameters corresponding to the DSP perfectly match the expected results, suggesting ideal quality of the current data sample. When r is greater than 1, it indicates that the effect of the machine adjustment parameters corresponding to the DSP is greater than the expected effect, suggesting poor quality of the current data sample.

[0050] Then, the score value corresponding to parameter r (i.e., the score value corresponding to the above data sample) can be calculated using the following preset scoring algorithm: S(r)=exp(-k(r-1)) 2 ).

[0051] Where k is a fixed parameter that can be customized according to actual needs; optionally, k can be 2. Where S(r) is the score value corresponding to parameter r. Where exp is the natural exponential function. It can be understood that a higher score value for a data sample indicates a higher quality data sample.

[0052] 113. The weight of the data sample is determined based on the score value, and the score value is positively correlated with the weight.

[0053] 114. Determine the training samples by taking multiple data samples and their weights.

[0054] As can be seen, in this embodiment, the score value of the data sample is calculated by a preset scoring algorithm, thereby accurately quantifying the quality of the data sample. The weight of the data sample is determined according to the score value. Multiple data samples and the weight of each data sample are determined as training samples, so that the training samples contain the weight of each data sample, so that the model can combine the weight of the data sample for training, thereby improving the efficiency and accuracy of model training.

[0055] 120. Determine the second training set and the second validation set from the first training set, and train the initial model based on the first training set to obtain the candidate model corresponding to the current iteration round of the initial model.

[0056] In some implementations, the initial model can be based on a gradient boosting tree or a deep neural network (DNN), and there is no limitation on this. Since both models are suitable for tabular data, they are also suitable for the training samples containing silicon wafer process data in this embodiment.

[0057] In some implementations, the specific implementation of training the initial model based on the first training set in step 120 may include: Based on the weights corresponding to each data sample in the first training set, the fitting priority for each data sample is determined, and the weights are positively correlated with the fitting priority.

[0058] Based on the fitting priority and the first training set, the initial model is trained.

[0059] For example, taking an initial model using a gradient boosting tree (XGBoost) as an example, the weights corresponding to each data sample can be directly substituted into the calculation formula for the split gain. The gradient (g) and second derivative (h) of each sample are multiplied by its weight during calculation. Thus, for samples with high weights, their gradient values ​​are amplified during each round of tree construction when calculating the split gain, thereby having a greater impact on the selection of decision tree split points. This results in these data samples having a higher fitting priority during model training, causing the model to prioritize fitting these samples with higher fitting priority. In other words, by introducing sample weights for model training, the model can focus more on high-quality sample data during training, effectively improving the accuracy of the trained model.

[0060] 130. The candidate model is validated using the first validation set to obtain the first validation result, and the candidate model is validated using the second validation set to obtain the second validation result.

[0061] In some implementations, the second validation set includes a second input sample and a second target sample corresponding to the second input sample. The second input sample may correspond to the input feature X in the (X, Y) pair in the above embodiments, and the second target sample may correspond to the output label Y in the (X, Y) pair. In step 130, the specific implementation of validating the candidate model using the second validation set to obtain the second validation result may include: 131. Using alternative models, predict the second input sample to obtain the second prediction result.

[0062] For example, taking the input sample as including the initial GBIR of the silicon wafer before DSP and the machine adjustment parameters corresponding to the DSP process, and the second target sample as the actual GBIR obtained after the silicon wafer is processed by the machine adjustment parameters corresponding to the DSP process, the above-mentioned initial GBIR and the above-mentioned machine adjustment parameters corresponding to the DSP process can be input into the candidate model, and the predicted GBIR of the silicon wafer after DSP output by the candidate model can be obtained, and the predicted GBIR can be determined as the second prediction result.

[0063] 132. Calculate the second prediction result using a preset scoring algorithm to obtain the second output score value.

[0064] In some implementations, the specific method for "calculating the second prediction result using a preset scoring algorithm to obtain the second output score value" in step 132 may include: Calculate the difference between the second target sample and the second input sample to obtain the first difference; Calculate the difference between the second prediction result and the second input sample to obtain the second difference; The second output score is determined based on the first difference and the second difference. The absolute value of the difference between the first difference and the second difference is negatively correlated with the second output score.

[0065] Following the example above, (actual GBIR - initial GBIR) can be determined as the first difference, and (predicted GBIR - initial GBIR) as the second difference. Then, using the preset scoring algorithm described above, a second output score value is calculated based on the first and second differences. Specifically, the parameter r can be calculated based on the first and second differences, where r = (actual GBIR - initial GBIR) / (predicted GBIR - initial GBIR), and then the result can be expressed by the formula S(r) = exp(-k(r-1)). 2 The algorithm calculates the score corresponding to parameter r and sets this score as the second output score. According to the formula above, the smaller the absolute value of the difference between the first and second differences, the closer parameter r is to 1. In this case, the higher the second output score, the higher the sample quality of the second validation set.

[0066] 133. Compare the second output score with the score corresponding to the second target sample, and determine the comparison result as the second verification result.

[0067] For example, the second output score value can be compared with the score value corresponding to the second target sample, and the comparison result indicates whether the second output score value is greater than the score value corresponding to the second target sample.

[0068] In some implementations, the first validation set includes a first input sample and a first target sample corresponding to the first input sample, wherein the first input sample may correspond to the input feature X in the (X, Y) pair in the above embodiments, and the first target sample may correspond to the output label Y in the (X, Y) pair. In step 130, the candidate model is validated using the first validation set to obtain a first validation result, including: The first prediction result is obtained by predicting the first input sample using alternative models.

[0069] The mean absolute error (MAE) between the first prediction result and the first target sample is determined as the first validation result.

[0070] The specific implementation of the step "predicting the first input sample using the alternative model to obtain the first prediction result" can be referred to step 131 above, and therefore will not be repeated here.

[0071] 140. Based on the second verification result, update the first training set to obtain the updated first training set. Based on the updated first training set, return to the step of training the initial model based on the first training set to obtain the candidate model corresponding to the current iteration round of the initial model, until the initial model converges and multiple candidate models are obtained.

[0072] In some implementations, the initial model can be determined to have converged when the number of iterations of the initial model reaches a specified number of iterations. For example, if the specified number of iterations is 100, the initial model can be determined to have converged when it has iterated 100 times.

[0073] In other implementations, the loss value of the initial model can be calculated using a loss function. If the loss value meets a preset loss condition, the initial model can be considered converged. Optionally, the loss function can be Huber Loss. Since there are many outliers in the samples, the error will be amplified when choosing a general MSE loss function. Therefore, using Huber Loss can combine the advantages of both the mean squared error loss function and the mean absolute error (MAE) loss function. It uses the square form for small errors and the absolute value form for large errors, reducing the sensitivity to outliers and systematically identifying and resisting the interference of outliers, thereby effectively achieving the goal of training a stable model from noisy data.

[0074] In some implementations, in step 140, the first training set is updated based on the second verification result to obtain the updated first training set. Specific implementations may include: If the second output score is greater than the score corresponding to the second target sample, then the second target sample in the second validation set of the first training set is replaced with the second prediction result to obtain the updated first training set.

[0075] Optionally, the method may further include: if the second output score is less than or equal to the score corresponding to the second target sample, the first training set can be kept unchanged, and the initial model can be trained in the next iteration based on the unchanged first training set.

[0076] In some implementations, the specific implementation of the step "replacing the second target sample of the second validation set in the first training set with the second prediction result to obtain the updated first training set" may include: A1. Replace the second target sample in the second validation set of the first training set with the second prediction result to obtain the updated second validation set.

[0077] For example, the second validation set includes a second input sample X1 and a second target sample Y1 corresponding to the second input sample X1. The candidate model in the current iteration outputs a second prediction result Y2 based on the second input sample X1. When the second output score corresponding to Y2 is greater than the score corresponding to Y1, Y1 in the second validation set can be replaced with Y2 to obtain an updated second validation set, i.e., this second validation set includes the sample (X1, Y2), thus obtaining the updated first training set. It is understood that samples other than (X1, Y1) in the second validation set can also be updated in the above manner, so this will not be elaborated upon here.

[0078] A2. Determine the target machine adjustment parameter sample from the updated second verification set according to the specified ratio.

[0079] For example, the updated second validation set includes updated samples (X1, Y2), (X2, Y3), (X3, Y4), and (X4, Y5). Since X1, X2, X3, and X4 are input samples, they all include machine adjustment parameter samples. Then, a specified proportion of machine adjustment parameter samples can be selected from the updated second validation set as target machine adjustment parameter samples. As an example, for instance, if the specified proportion is 50%, two updated samples can be randomly selected from the four updated samples. Then, the machine adjustment parameter samples from the input samples of these two updated samples are determined as target machine adjustment parameter samples. For example, the target machine adjustment parameter samples could be the machine adjustment parameter samples in X1 and X2, or they could be the machine adjustment parameter samples in X3 and X4.

[0080] Optionally, the specified ratio can be customized according to actual needs. For example, in some implementations, the specified ratio can be 10%.

[0081] A3. Enhance the target machine adjustment parameter samples in the updated second validation set according to the specified magnification to obtain the updated first training set.

[0082] For example, if the specified multiplier is 2, and the target machine's adjustment parameter sample includes an upper platen pressure change of 'd', then the adjustment range of the upper platen pressure change 'd' can be increased to twice its original value, resulting in an enhanced upper platen pressure change, i.e., (2*d). Then, the upper platen pressure change in the target machine's adjustment parameter sample is replaced with the enhanced upper platen pressure change. Based on the enhanced upper platen pressure change and a preset data relationship, new quality parameters for the silicon wafer can be calculated. Then, the original quality parameters corresponding to the upper platen pressure change in the first training set are replaced with the new quality parameters, thus obtaining an updated first training set. The preset data relationship can be obtained by fitting the historical platen pressure changes of the production machine and the historical quality parameters after silicon wafer processing.

[0083] In this embodiment, by enhancing the target machine adjustment parameter samples in the updated second validation set by a specified multiplier, an updated first training set is obtained. This further enhances the influence of high-quality samples on the model during model training, thereby improving the accuracy of the trained model.

[0084] In some implementations, the target machine adjustment parameter samples include positive and negative samples. Positive samples are used to positively adjust the production machine; for example, positive samples can be used to improve the quality parameters of the processed silicon wafers. Negative samples are used to negatively adjust the production machine; for example, negative samples can be used to decrease the quality parameters of the processed silicon wafers. In step A3, the target machine adjustment parameter samples in the updated second validation set are enhanced at a specified ratio to obtain the updated first training set. Specific implementations may include: The negative samples in the target machine adjustment parameter sample are enhanced by a specified magnification to obtain the updated first training set.

[0085] The specific implementation method for enhancing the negative samples in the target machine adjustment parameter sample by a specified magnification can be found in step A3, and therefore will not be repeated here.

[0086] In this embodiment, by performing enhancement processing on negative samples, negative samples with a larger adjustment range than normal are introduced for model training, which can effectively enhance the robustness of the model.

[0087] In other embodiments, in step 140, the first training set is updated based on the second verification result to obtain the updated first training set. Specific implementations may include: B1. If the second output score is greater than the score corresponding to the second target sample, then obtain the update probability.

[0088] In some implementations, the first verification set includes a first input sample and a first target sample corresponding to the first input sample. In step B1, the specific implementation of obtaining the update probability may include: B11. For each candidate model, predict the first input sample using the candidate model to obtain the first prediction result.

[0089] B12. Calculate the mean absolute error between the first prediction result and the first target sample to obtain the mean absolute error corresponding to the alternative model.

[0090] The specific implementation methods for steps B11 to B12 can be found in step 130, and therefore will not be repeated here.

[0091] B13. Calculate the first prediction result using a preset scoring algorithm to obtain the first output score value corresponding to the candidate model.

[0092] The specific implementation of step B13 can be found in step 132, so it will not be repeated here.

[0093] B14. Determine the update probability based on the first output score and mean absolute error corresponding to each of the candidate models.

[0094] In some implementations, the difference between the mean absolute errors (MAEs) of candidate models in every two adjacent iterations can be calculated first to obtain multiple error changes. Then, the quotient between each error change and the baseline MAE is calculated to obtain multiple error change rates. Finally, the error change rate with the largest value among these multiple error change rates is determined as the MAE improvement rate during model training.

[0095] For example, the formula for calculating the MAE improvement rate can be as follows: Improvement_MAE = max(0, (MAE_t-1 - MAE_t) / MAE_base) t>1 Where Improvement_MAE represents the MAE improvement rate, t represents the t-th iteration of the initial model, MAE_t represents the mean absolute error of the candidate model in the t-th iteration, MAE_t-1 represents the mean absolute error of the candidate model in the (t-1)-th iteration, and MAE_base represents the baseline mean absolute error.

[0096] Then, the difference between the first output score values ​​of the candidate models corresponding to each two adjacent iterations can be calculated to obtain multiple score changes. The quotient between each score change value and the baseline score value can be calculated to obtain multiple score change rates. The score improvement rate during the model training process is then determined from the multiple score change rates with the largest scores.

[0097] For example, the formula for calculating the rating improvement rate can be as follows: Improvement_Score = max(0, (Score_t - Score_t-1) / (100 - Score_base)) t>1 Where Improvement_Score represents the score improvement rate, Score_t represents the first output score value corresponding to the candidate model in the t-th iteration, Score_t-1 represents the first output score value corresponding to the candidate model in the (t-1)-th iteration, and Score_base represents the baseline score value.

[0098] Finally, the update probability can be determined based on the score improvement rate and MAE improvement rate.

[0099] For example, the update probability can be calculated using the following formula: P(t) = P_start + (P_end - P_start) * (w1 * Improvement_MAE + w2 *Improvement_Score) In this model training, w1 + w2 = 1, where w1 and w2 represent the weights allocated to the MAE improvement rate and the score improvement rate during training. This indicates whether the trained model is more data-driven or score-rule-driven. Optionally, w1 can be 0.3 and w2 can be 0.7. P_start and P_end represent the baseline acceptance rate and the maximum acceptance rate, respectively, indicating the basic and maximum level of trust in the scoring rule. Optionally, P_start and P_end can be 0.2 and 0.9, respectively.

[0100] B2. Based on the update probability, replace the second target sample in the second validation set of the first training set with the second prediction result to obtain the updated first training set.

[0101] For example, if the update probability is 50%, then when it is determined that the second output score is greater than the score corresponding to the second target sample, there is a 50% probability that the second target sample in the second validation set of the first training set will be replaced with the second prediction result. If the second target sample in the second validation set of the first training set is replaced with the second prediction result according to the 50% probability, then the updated first training set can be obtained. If the second target sample in the second validation set of the first training set is not replaced with the second prediction result according to the 50% probability, then the second validation set of the current iteration can remain unchanged and continue to be used for model training in the next iteration.

[0102] Considering that there may be factors in the training samples that the scoring algorithm has not taken into account, in this embodiment, the first training set is updated according to the update probability, instead of being updated directly. This can prevent the model from becoming cognitively fixed during training, maintain the model's evolutionary ability, and avoid getting stuck in local optima.

[0103] 150. The candidate model whose first verification result meets the preset conditions among multiple candidate models is determined as the target model. The target model is used to predict the quality parameters of the silicon wafer after production based on the initial parameters of the input silicon wafer and the machine adjustment parameters of the production machine.

[0104] In some implementations, the specific implementation of determining the candidate model whose first verification result meets the preset conditions among multiple candidate models as the target model in step 150 may include: The candidate model with the smallest mean absolute error among multiple candidate models is selected as the target model.

[0105] For example, the initial model can obtain a corresponding candidate model in each iteration. Each candidate model is validated through a first validation set, and a mean absolute error is obtained as the first validation result, thus obtaining multiple mean absolute errors. Then, the minimum mean absolute error can be determined from the multiple mean absolute errors, and the candidate model corresponding to the minimum mean absolute error is determined as the target model.

[0106] In practical applications, after obtaining the target model, the parameters of the production equipment in the silicon wafer production process can be adjusted according to the target model. For example, based on the initial parameters of the input silicon wafer and the equipment adjustment parameters, the target model predicts the quality parameters of the silicon wafer after production, and the equipment adjustment parameters can be adjusted according to the predicted quality parameters. As a more specific example, when the GBIR of the silicon wafer after DSP is too high (e.g., exceeding 80nm), the global parameter is checked to see if it is positive or negative. If it is positive, the upper mounting plate pressure is reduced; otherwise, the upper mounting plate pressure is increased. Furthermore, the quantitative relationship between the equipment adjustment parameters (e.g., upper mounting plate pressure) and the predicted quality parameters (global parameters) can be: global parameter change = 2 * upper mounting plate pressure change * 0.2 * global parameter, and the corresponding GBIR change = GBIR - (global parameter - global parameter change), and so on. The adjustment amount of other equipment adjustment parameters can also be determined in the same way, thereby enabling timely and accurate adjustment of the production equipment and effectively improving the quality of the produced silicon wafers.

[0107] For example, in practical applications, the specific implementation process of steps 110 to 150 can be as follows: Figure 2 As shown: First, training samples can be obtained by collecting data from the silicon wafer production site.

[0108] Then, 10% of the data from the training samples is selected as the first validation set, and the remaining 90% of the data is used as the first training set.

[0109] Next, the first training set can be divided into a second training set of 80% and a second validation set of 20%. After initializing the model parameters, the model can be trained using the first training set.

[0110] After the training of the model in the current round is completed, the candidate model obtained in the current iteration can be verified by the first validation set, and the first validation result corresponding to the candidate model can be recorded.

[0111] Next, the number of iteration rounds for the model is determined. If the model has iterated for N rounds, it indicates that the model has converged and training is complete. At this point, the candidate model with the best MAE can be selected as the target model based on the first validation results of the candidate models in each iteration round on the first validation set. Here, N is a positive integer.

[0112] If the model has not iterated to N rounds, the second validation set can be updated, and the updated second validation set can be augmented and added to the first training set to obtain the updated first training set. Then, the model can be trained in the next round based on the updated first training set.

[0113] In some implementations, the specific implementation process of step "update the second verification set" can be as follows: Figure 3 As shown, when updating the second validation set, the second validation set can first be predicted using the candidate model in the current round. Then, it is checked whether the score of the prediction result of the candidate model is greater than the original score of the second validation set.

[0114] If the score of the prediction result of the candidate model is greater than the original score of the second validation set, the update probability is calculated, and the test result is accepted with the update probability. Then, the target output in the samples in the second validation set is updated based on the prediction result.

[0115] If the score of the prediction result of the candidate model is less than or equal to the original score of the second validation set, the original second validation set is retained, and the target output in the samples of the second validation set is updated based on the retained original second validation set. That is, the target output in the samples of the updated second validation set is consistent with the target output in the original second validation set.

[0116] In some implementations, the specific implementation process of the step "data augmentation of the updated second validation set" can be as follows: Figure 4As shown, firstly, 10% of the data is selected from the updated second validation set. Then, this 10% of data is augmented by a 2x tuning amplitude to obtain augmented data. Next, new quality parameters for the silicon wafer are calculated based on the augmented data. Finally, new samples are generated based on the new quality parameters and the augmented data. These new samples can be added to the first training set, thus obtaining the updated first training set.

[0117] As can be seen, in this implementation, training samples are acquired, and a first validation set and a first training set are determined from the training samples. The training samples include data recorded during the silicon wafer manufacturing process. The first validation set consists of data in the training samples whose noise meets preset noise conditions. Then, a second training set and a second validation set are determined from the first training set, and the initial model is trained based on the first training set to obtain a candidate model corresponding to the current iteration round. The candidate model is validated using the first validation set to obtain a first validation result, and the candidate model is validated using the second validation set to obtain a second validation result. Based on the second validation result, the first training set is updated to obtain an updated first training set. Based on the updated first training set, the process of training the initial model based on the first training set to obtain a candidate model corresponding to the current iteration round is repeated until the initial model converges, resulting in multiple candidate models. Finally, the candidate model whose first validation result meets the preset conditions among the multiple candidate models is determined as the target model. The target model is used to predict the quality parameters of the silicon wafer after production based on the initial parameters of the input silicon wafer and the machine adjustment parameters of the production machine. In other words, during training, the first training set can be updated using the second validation results to continuously generate new and higher-quality first training sets. This achieves iterative data purification of the first training set, enabling the automatic and efficient training of a stable, reliable, and high-precision target model from noisy industrial data. This target model can then be used to effectively optimize machine adjustment parameters. Furthermore, by utilizing a lower-noise first validation set and selecting a model with higher prediction accuracy from multiple iterations as the target model, the production quality of silicon wafers can be effectively improved.

[0118] One embodiment of the present invention provides a silicon wafer manufacturing method, which can be executed by computer equipment (e.g., a server or user terminal). Figure 5 As shown, the silicon wafer manufacturing method may include: 210. Obtain the initial parameters of the silicon wafer to be manufactured and the machine adjustment parameters of the production equipment.

[0119] 220. Input the initial parameters and machine adjustment parameters into the target model in the above embodiment for prediction to obtain the quality parameters of the silicon wafer after production.

[0120] 230. If the quality parameters do not meet the preset process requirements, new machine adjustment parameters shall be determined based on the quality parameters.

[0121] As an example, if a quality parameter is not within the specified range, it can be determined that the quality parameter does not meet the preset process requirements.

[0122] The specific implementation method for determining new machine adjustment parameters based on quality parameters in step 230 can refer to the specific implementation method of "adjusting the parameters of the production machine in the silicon wafer production process according to the target model" in the above embodiment, so it will not be repeated here.

[0123] 240. Adjust the parameters of the production machine according to the new machine adjustment parameters, so as to manufacture the silicon wafer to be manufactured based on the adjusted production machine.

[0124] For example, when adjusting the machine using the silicon wafer manufacturing method of this embodiment, the following can be obtained: Figure 6 The GBIR-SPC control diagram shown is shown. Figure 7 This is the GBIR-SPC control diagram obtained through engineer commissioning in related technologies. It can be seen that... Figure 7 The control chart obtained through engineer-led adjustments exhibited drastic fluctuations and frequent alarms. However, the control chart obtained through automatic adjustments based on the target model in this embodiment shows stable fluctuations and a well-controlled state. Therefore, automatic adjustments using the target model in this embodiment enable more timely adjustments and effectively improve silicon wafer quality and yield. SPC stands for Statistical Process Control. Figure 6 and Figure 7 The horizontal axis represents the batch of silicon wafers processed.

[0125] Figure 8 The diagram shown is a block diagram of a model training device provided in an embodiment of the present invention. Figure 8 As shown, the model training device 300 includes: The first acquisition module 310 is used to acquire training samples and determine a first validation set and a first training set from the training samples; the training samples include data recorded during the silicon wafer manufacturing process; the first validation set is the data in the training samples whose noise meets a preset noise condition; The training module 320 is used to determine the second training set and the second validation set from the first training set, and to train the initial model based on the first training set to obtain the candidate model corresponding to the current iteration round of the initial model. The verification module 330 is used to verify the candidate model using a first verification set to obtain a first verification result, and to verify the candidate model using a second verification set to obtain a second verification result. The update module 340 is used to update the first training set based on the second verification result to obtain the updated first training set, and based on the updated first training set, return to execute the step of training the initial model based on the first training set to obtain the candidate model corresponding to the current iteration round of the initial model, until the initial model converges and multiple candidate models are obtained. The determination module 350 is used to determine the candidate model whose first verification result meets the preset conditions among multiple candidate models as the target model. The target model is used to predict the quality parameters of the silicon wafer after production based on the initial parameters of the input silicon wafer and the machine adjustment parameters of the production machine.

[0126] In some implementations, the first acquisition module 310 is specifically used for: Obtain multiple data samples recorded during the silicon wafer manufacturing process; For each of the multiple data samples, a score value is calculated using a preset scoring algorithm. The score value represents the production quality of silicon wafers after the production machine adjusts its parameters based on the data sample. The weights of the data samples are determined based on the rating values, and the rating values ​​are positively correlated with the weights. Multiple data samples and the weights of each data sample are determined as training samples.

[0127] In some implementations, the second verification set includes a second input sample and a second target sample corresponding to the second input sample. Verification module 330 is specifically used for: The second input sample is predicted using the alternative model to obtain the second prediction result; The second prediction result is calculated using a preset scoring algorithm to obtain the second output score value; The second output score is compared with the score corresponding to the second target sample, and the comparison result is determined as the second verification result.

[0128] In some implementations, the update module 340 is specifically used for: If the second output score is greater than the score corresponding to the second target sample, then the second target sample in the second validation set of the first training set is replaced with the second prediction result to obtain the updated first training set.

[0129] In some implementations, the update module 340 is further used for: The second target sample in the second validation set of the first training set is replaced with the second prediction result to obtain the updated second validation set. According to the specified ratio, determine the target machine adjustment parameter sample from the updated second verification set; The target machine adjustment parameter samples in the updated second validation set are enhanced by a specified magnification to obtain the updated first training set.

[0130] In some implementations, the target machine adjustment parameter sample includes positive and negative samples. Positive samples are used for positive adjustment of the production machine, and negative samples are used for negative adjustment of the production machine. The update module 340 is further used for: The negative samples in the target machine adjustment parameter sample are enhanced by a specified magnification to obtain the updated first training set.

[0131] In some implementations, the update module 340 is specifically used for: If the second output score is greater than the score corresponding to the second target sample, then obtain the update probability; Based on the update probability, the second target sample in the second validation set of the first training set is replaced with the second prediction result to obtain the updated first training set.

[0132] In some implementations, the first verification set includes a first input sample and a first target sample corresponding to the first input sample. The update module 340 is further used for: For each candidate model, the first input sample is predicted using the candidate model to obtain the first prediction result; Calculate the mean absolute error between the first prediction result and the first target sample to obtain the mean absolute error corresponding to the candidate model; The first prediction result is calculated using a preset scoring algorithm to obtain the first output score value corresponding to the candidate model; The update probability is determined based on the first output score and mean absolute error of each candidate model.

[0133] In some implementations, the verification module 330 is further used for: Calculate the difference between the second target sample and the second input sample to obtain the first difference; Calculate the difference between the second prediction result and the second input sample to obtain the second difference; The second output score is determined based on the first difference and the second difference. The absolute value of the difference between the first difference and the second difference is negatively correlated with the second output score.

[0134] In some implementations, the training module 320 is specifically used for: Based on the weights corresponding to each data sample in the first training set, the fitting priority for each data sample is determined, and the weights are positively correlated with the fitting priority. Based on the fitting priority and the first training set, the initial model is trained.

[0135] In some implementations, the first verification set includes a first input sample and a first target sample corresponding to the first input sample. The verification module 330 is further configured to: The first prediction result is obtained by predicting the first input sample using the alternative model; The mean absolute error between the first prediction result and the first target sample is determined as the first verification result.

[0136] In some implementations, the determining module 350 is specifically used for: The candidate model with the smallest mean absolute error among multiple candidate models is selected as the target model.

[0137] Figure 9 The diagram shown is a block diagram of a silicon wafer manufacturing apparatus according to an embodiment of the present invention. Figure 9 As shown, the silicon wafer manufacturing apparatus 400 may include: The second acquisition module 410 is used to acquire the initial parameters of the silicon wafer to be manufactured and the machine adjustment parameters of the production machine. The prediction module 420 is used to input the initial parameters and machine adjustment parameters into the target model mentioned above for prediction, so as to obtain the quality parameters of the silicon wafer after production.

[0138] The parameter update module 430 is used to determine new machine adjustment parameters based on the quality parameters if the quality parameters do not meet the preset process requirements.

[0139] The adjustment module 440 is used to adjust the parameters of the production machine according to the new machine adjustment parameters, so as to manufacture the silicon wafer to be manufactured based on the adjusted production machine.

[0140] The specific implementation process of the functions and roles of each module in the above device can be found in the implementation process of the corresponding steps of the model training method in the above embodiments, and will not be repeated here.

[0141] Figure 10 The diagram shown is a block diagram of an electronic device 500 provided in an embodiment of the present invention.

[0142] Reference Figure 10 The electronic device 500 includes a processing component 510, which further includes one or more processors, and memory resources represented by memory 520 for storing instructions, such as application programs, that can be executed by the processing component 510. The application programs stored in memory 520 may include one or more modules, each corresponding to a set of instructions. Furthermore, the processing component 510 is configured to execute instructions to perform the aforementioned model training method.

[0143] Electronic device 500 may also include a power supply component configured to perform power management of electronic device 500, a wired or wireless network interface configured to connect electronic device 500 to a network, and an input / output (I / O) interface. Electronic device 500 may operate on an operating system stored in memory 520, such as Windows Server™, Mac OSX™, Unix™, Linux™, FreeBSD™, or similar.

[0144] A non-transitory computer-readable storage medium, wherein when the instructions in the storage medium are executed by the processor of the aforementioned electronic device 500, the aforementioned electronic device 500 is able to execute the aforementioned model training method or silicon wafer manufacturing method.

[0145] Those skilled in the art will recognize that the units and algorithm steps of the various examples described in conjunction with the embodiments disclosed herein can be implemented in electronic hardware, or a combination of computer software and electronic hardware. Whether these functions are implemented in hardware or software depends on the specific application and design constraints of the technical solution. Those skilled in the art can use different methods to implement the described functions for each specific application, but such implementations should not be considered beyond the scope of this invention.

[0146] Those skilled in the art will clearly understand that, for the sake of convenience and brevity, the specific working processes of the systems, devices, and units described above can be referred to the corresponding processes in the foregoing method embodiments, and will not be repeated here.

[0147] In the several embodiments provided in this application, it should be understood that the disclosed systems, apparatuses, and methods can be implemented in other ways. For example, the apparatus embodiments described above are merely illustrative; for instance, the division of units is only a logical functional division, and in actual implementation, there may be other division methods. For example, multiple units or components may be combined or integrated into another system, or some features may be ignored or not executed. Furthermore, the coupling or direct coupling or communication connection shown or discussed may be through some interfaces; the indirect coupling or communication connection between apparatuses or units may be electrical, mechanical, or other forms.

[0148] The units described as separate components may or may not be physically separate. The components shown as units may or may not be physical units; that is, they may be located in one place or distributed across multiple network units. Some or all of the units can be selected to achieve the purpose of this embodiment according to actual needs.

[0149] In addition, the functional units in the various embodiments of the present invention can be integrated into one processing unit, or each unit can exist physically separately, or two or more units can be integrated into one unit.

[0150] If the aforementioned functions are implemented as software functional units and sold or used as independent products, they can be stored in a computer-readable storage medium. Based on this understanding, the technical solution of this invention, or the part that contributes to the prior art, or a portion of the technical solution, can be embodied in the form of a software product. This computer software product is stored in a storage medium and includes several instructions to cause a computer device (which may be a personal computer, server, or network device, etc.) to execute all or part of the steps of the methods described in the various embodiments of this invention. The aforementioned storage medium includes various media capable of storing program verification codes, such as USB flash drives, portable hard drives, read-only memory (ROM), random access memory (RAM), magnetic disks, or optical disks.

[0151] Furthermore, it should be noted that the combination of the various technical features in this case is not limited to the combination methods described in the claims of this case or the combination methods described in the specific embodiments. All technical features described in this case can be freely combined or combined in any way, unless they contradict each other.

[0152] It should be noted that the above examples are merely specific embodiments of the present invention, and the present invention is obviously not limited to the above embodiments, with many similar variations. All modifications that can be directly derived or conceived by those skilled in the art from the content disclosed in this invention should fall within the protection scope of this invention.

[0153] It should be understood that the terms "first," "second," etc., mentioned in the embodiments of the present invention are merely for the purpose of more clearly describing the use of the technical solutions in the embodiments of the present invention, and are not intended to limit the scope of protection of the present invention.

[0154] The above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent substitutions, etc., made within the spirit and principles of the present invention should be included within the protection scope of the present invention.

Claims

1. A model training method, characterized in that, include: Obtain training samples, and determine a first validation set and a first training set from the training samples; The training samples include data recorded during the silicon wafer manufacturing process; The first validation set consists of training samples whose noise meets a preset noise condition. A second training set and a second validation set are determined from the first training set, and the initial model is trained based on the first training set to obtain the candidate model corresponding to the current iteration round of the initial model; The candidate model is validated using the first validation set to obtain a first validation result, and the candidate model is validated using the second validation set to obtain a second validation result. Based on the second verification result, the first training set is updated to obtain an updated first training set. Based on the updated first training set, the step of training the initial model based on the first training set to obtain the candidate model corresponding to the current iteration round is returned to be executed until the initial model converges, and multiple candidate models are obtained. The candidate model whose first verification result meets the preset conditions among the multiple candidate models is determined as the target model. The target model is used to predict the quality parameters of the silicon wafer after production based on the initial parameters of the input silicon wafer and the machine adjustment parameters of the production machine.

2. The method according to claim 1, characterized in that, The acquisition of training samples includes: Obtain multiple data samples recorded during the silicon wafer manufacturing process; For each of the plurality of data samples, a score value is calculated for the data sample using a preset scoring algorithm. The score value represents the production quality of the silicon wafer by the production machine after the parameters of the production machine are adjusted based on the data sample. The weight of the data sample is determined based on the score value, and the score value is positively correlated with the weight. The plurality of data samples and the weight of each data sample are determined as the training samples.

3. The method according to claim 2, characterized in that, The second validation set includes a second input sample and a second target sample corresponding to the second input sample. The step of validating the candidate model using the second validation set to obtain a second validation result includes: The second input sample is predicted using the alternative model to obtain a second prediction result; The second prediction result is calculated using the preset scoring algorithm to obtain a second output score value; The second output score is compared with the score corresponding to the second target sample, and the comparison result is determined as the second verification result.

4. The method according to claim 3, characterized in that, The step of updating the first training set based on the second verification result to obtain the updated first training set includes: If the second output score is greater than the score corresponding to the second target sample, then the second target sample in the second validation set of the first training set is replaced with the second prediction result to obtain the updated first training set.

5. The method according to claim 3, characterized in that, The step of replacing the second target sample in the second validation set of the first training set with the second prediction result to obtain the updated first training set includes: The second target sample in the second validation set of the first training set is replaced with the second prediction result to obtain the updated second validation set. According to a specified ratio, a sample of target machine adjustment parameters is determined from the updated second verification set; The target machine adjustment parameter samples in the updated second verification set are enhanced by a specified multiplier to obtain the updated first training set.

6. The method according to claim 5, characterized in that, The target machine adjustment parameter samples include positive samples and negative samples. The positive samples are used for positive adjustment of the production machine, and the negative samples are used for negative adjustment of the production machine. The updated first training set is obtained by enhancing the target machine adjustment parameter samples in the updated second validation set according to a specified factor, including: The negative samples in the target machine adjustment parameter sample are enhanced by a specified factor to obtain the updated first training set.

7. The method according to claim 3, characterized in that, The step of updating the first training set based on the second verification result to obtain an updated first training set includes: If the second output score is greater than the score corresponding to the second target sample, then obtain the update probability; Based on the updated probability, the second target sample in the second validation set of the first training set is replaced with the second prediction result to obtain the updated first training set.

8. The method according to claim 7, characterized in that, The first verification set includes a first input sample and a first target sample corresponding to the first input sample. The step of obtaining the update probability includes: For each of the candidate models, the first input sample is predicted using the candidate model to obtain a first prediction result; Calculate the mean absolute error between the first prediction result and the first target sample to obtain the mean absolute error corresponding to the alternative model; The first prediction result is calculated using the preset scoring algorithm to obtain the first output score value corresponding to the candidate model; The update probability is determined based on the first output score and mean absolute error corresponding to each of the candidate models.

9. The method according to claim 3, characterized in that, The step of calculating the second prediction result using the preset scoring algorithm to obtain the second output score value includes: Calculate the difference between the second target sample and the second input sample to obtain the first difference; Calculate the difference between the second prediction result and the second input sample to obtain the second difference; The second output score is determined based on the first difference and the second difference, wherein the absolute value of the difference between the first difference and the second difference is negatively correlated with the second output score.

10. The method according to any one of claims 2 to 9, characterized in that, The step of training the initial model based on the first training set includes: Based on the weights corresponding to each data sample in the first training set, the fitting priority corresponding to each data sample is determined, wherein the weights are positively correlated with the fitting priority; Based on the fitting priority and the first training set, the initial model is trained.

11. The method according to any one of claims 1 to 9, characterized in that, The first validation set includes a first input sample and a first target sample corresponding to the first input sample. The step of validating the candidate model using the first validation set to obtain a first validation result includes: The first input sample is predicted using the alternative model to obtain a first prediction result; The average absolute error between the first prediction result and the first target sample is determined as the first verification result.

12. The method according to claim 11, characterized in that, The step of determining the candidate model whose first verification result among the plurality of candidate models meets the preset conditions as the target model includes: The candidate model with the smallest average absolute error among the multiple candidate models is determined as the target model.

13. A method for manufacturing a silicon wafer, characterized in that, include: Obtain the initial parameters of the silicon wafer to be manufactured and the machine adjustment parameters of the production equipment; The initial parameters and the machine adjustment parameters are input into the target model as described in any one of claims 1 to 12 for prediction to obtain the quality parameters of the silicon wafer after production. If the quality parameters do not meet the preset process requirements, then new machine adjustment parameters are determined based on the quality parameters. The production machine is adjusted according to the new machine adjustment parameters so as to manufacture the silicon wafer to be manufactured based on the adjusted production machine.

14. A model training device, characterized in that, include: The first acquisition module is used to acquire training samples and determine a first validation set and a first training set from the training samples; The training samples include data recorded during the silicon wafer manufacturing process; The first validation set consists of training samples whose noise meets a preset noise condition. The training module is used to determine a second training set and a second validation set from the first training set, and to train the initial model based on the first training set to obtain the candidate model corresponding to the current iteration round of the initial model. The verification module is used to verify the candidate model using the first verification set to obtain a first verification result, and to verify the candidate model using the second verification set to obtain a second verification result; The update module is used to update the first training set based on the second verification result to obtain the updated first training set, and based on the updated first training set, return to the step of training the initial model based on the first training set to obtain the candidate model corresponding to the current iteration round of the initial model, until the initial model converges and multiple candidate models are obtained. The determination module is used to determine the candidate model whose first verification result meets the preset conditions among the multiple candidate models as the target model. The target model is used to predict the quality parameters of the silicon wafer after production based on the initial parameters of the input silicon wafer and the machine adjustment parameters of the production machine.

15. A silicon wafer manufacturing apparatus, characterized in that, include: The second acquisition module is used to acquire the initial parameters of the silicon wafer to be manufactured and the machine adjustment parameters of the production machine. The prediction module is used to input the initial parameters and the machine adjustment parameters into the target model as described in any one of claims 1 to 12 for prediction, so as to obtain the quality parameters of the silicon wafer after production; The parameter update module is used to determine new machine adjustment parameters based on the quality parameters if the quality parameters do not meet the preset process requirements. An adjustment module is used to adjust the parameters of the production machine according to the new machine adjustment parameters, so as to manufacture the silicon wafer to be manufactured based on the adjusted production machine.

16. A computer-readable storage medium having computer-executable instructions stored thereon, characterized in that, When the executable instructions are executed by the processor, they implement the model training method as described in any one of claims 1 to 12 or the silicon wafer manufacturing method as described in claim 13.

17. An electronic device, characterized in that, The electronic device includes: processor; Memory used to store the processor's executable instructions; The processor is used to execute the model training method according to any one of claims 1 to 12 or the silicon wafer manufacturing method according to claim 13.

Images