Training method of electrical property prediction model of semiconductor tape-out and computer equipment

Abstract

The invention relates to the technical field of machine learning, and discloses a training method for an electrical performance prediction model of a semiconductor tape-out, and computer equipment, and the method comprises the steps: obtaining training data which comprises the technological parameters of a plurality of semiconductor tape-out pieces and the real electrical performance parameters of the semiconductor tape-out pieces; and inputting the training data into a preset electrical performance prediction model, and training the electrical performance prediction model by using a fusion loss function composed of a mean square error, a mean absolute error and a mean absolute percentage error to obtain the trained electrical performance prediction model. The method has the beneficial effect that the prediction accuracy and robustness of the electrical performance prediction model are improved.

Description

Technical Field

[0001] This application relates to the field of machine learning technology, and in particular to a training method and computer device for predicting the electrical performance of semiconductor wafers. Background Technology

[0002] In the semiconductor manufacturing field, accurately predicting the electrical properties of semiconductor wafers can effectively avoid anomalies and optimize production during the semiconductor wafer fabrication process.

[0003] Currently, the prediction of the electrical performance of semiconductor wafers is mainly achieved through machine learning methods such as linear neural networks, random forest regression, and support vector regression.

[0004] However, conventional machine learning methods typically employ a single loss function, which results in poor model training performance. Summary of the Invention

[0005] This application provides a training method and computer device for a prediction model of the electrical performance of semiconductor wafers, which solves the technical problem of poor model training effect and achieves the technical effect of improving the prediction accuracy and robustness of the prediction model of electrical performance.

[0006] To achieve the above objectives, the main technical solutions adopted in this application include: In a first aspect, embodiments of this application provide a method for training a prediction model of the electrical performance of semiconductor wafers, comprising: Acquire training data, which includes process parameters of multiple semiconductor wafers and actual electrical performance parameters of the semiconductor wafers; During the iterative training of the electrical performance prediction model, the process parameters are input into the electrical performance prediction model for processing to obtain predicted electrical performance parameters; and based on multiple preset loss functions, the model parameters of the electrical performance prediction model are corrected using the actual electrical performance parameters and the predicted electrical performance parameters to obtain the corrected model parameters of the electrical performance prediction model.

[0007] In this embodiment, by acquiring training data containing multiple semiconductor wafer fabrication process parameters and actual electrical performance parameters, and during iterative training, using multiple loss functions to correct the model parameters of the electrical performance prediction model based on the predicted electrical performance parameters output by the electrical performance prediction model and the actual electrical performance parameters contained in the training data, the accuracy and reliability of the prediction model for the electrical performance of semiconductor wafer fabrication are improved.

[0008] In one example, based on multiple preset loss functions, the model parameters of the electrical performance prediction model are corrected using the actual electrical performance parameters and the predicted electrical performance parameters to obtain the corrected model parameters of the electrical performance prediction model, including: Based on multiple preset loss functions, multiple loss feature values ​​are calculated between the predicted electrical performance and the actual electrical performance parameters; the multiple loss feature values ​​are fused to obtain the fused loss feature value for the current iteration. The model parameters of the electrical performance prediction model are corrected using the fusion loss feature values ​​to obtain the corrected model parameters of the electrical performance prediction model in the current iteration.

[0009] In this example, by using multiple loss functions, multiple loss feature values ​​are calculated between the predicted electrical performance parameters and the actual electrical performance parameters. These multiple loss feature values ​​are then fused to obtain a fused loss feature value. The model parameters are then corrected based on this fused loss feature value. This approach further improves the model training efficiency and enhances the prediction accuracy and robustness of the trained model.

[0010] In one example, the loss function includes a mean squared error loss function, a mean absolute error loss function, and a mean absolute percentage error loss function; based on multiple preset loss functions, multiple loss feature values ​​are calculated between the predicted electrical performance and the actual electrical performance parameters; the multiple loss feature values ​​are fused to obtain the fused loss feature value for the current iteration, including: The mean square error loss function, mean absolute error loss function, and mean absolute percentage error loss function are used to calculate the mean square error loss characteristic value, mean absolute error loss characteristic value, and mean absolute percentage error loss characteristic value based on the predicted electrical performance parameters and the actual electrical performance parameters, respectively. Based on preset weighting coefficients, the weighted sum of the mean squared error loss feature value, the mean absolute error loss feature value, and the mean absolute percentage error loss feature value is calculated to obtain the fusion loss feature value for the current iteration round.

[0011] In this example, the mean squared error loss feature value, mean absolute error loss feature value, and mean absolute percentage error loss feature value are calculated separately, and then fused according to the weighting coefficient to obtain the fused loss feature value. This method achieves accurate calculation of the fused loss feature value. Based on the weighting coefficient, the utilization rate of different error features is improved, the convergence speed and optimization direction accuracy of the electrical performance prediction model in the iteration process are enhanced, and the prediction effect of the electrical performance prediction model is optimized.

[0012] In one example, the method further includes: Based on the mean squared error loss feature value, the mean absolute error loss feature value, and the mean absolute percentage error loss feature value, the weighting coefficients of the mean squared error, the mean absolute error, and the mean absolute percentage error are optimized.

[0013] In this example, the weight coefficients are optimized by dynamically optimizing the corresponding weight coefficients based on the mean squared error loss feature value, the mean absolute error loss feature value, and the mean absolute percentage error loss feature value, thereby improving the training efficiency and training accuracy of the electrical performance prediction model.

[0014] In one example, obtaining training data includes: The energy and dose of ion implantation in multiple processes during the production of each semiconductor wafer are collected to generate the process parameters of the semiconductor wafer. The breakdown voltage and on-resistance of each semiconductor wafer are measured to obtain the actual electrical performance parameters of the semiconductor wafer. The process parameters and the actual electrical performance parameters are normalized to generate the dataset of the semiconductor wafer fabrication. The process parameters and actual electrical performance parameters of a portion of semiconductor wafer fabrication are selected from the dataset and used as the training data.

[0015] In this example, by collecting the energy and dose of ion implantation, the breakdown voltage and on-resistance of semiconductor wafers, and generating a dataset after normalization, and selecting a portion of it as training data, a high-quality training dataset is constructed, which improves the training effect of the electrical performance prediction model.

[0016] In one example, before normalizing the process parameters and the actual electrical performance parameters to generate the dataset for the semiconductor wafer fabrication, the method further includes at least one of the following: If the process parameters exceed the preset specifications, then delete the process parameters and the actual electrical performance parameters of the semiconductor wafer. If the energy curve of ion implantation in each process step of the process parameters changes abnormally, or if the dose curve of ion implantation in each process step changes abnormally, then the process parameters and the actual electrical performance parameters of the semiconductor wafer are deleted. If the actual electrical performance parameters exceed the preset range, then the process parameters and the actual electrical performance parameters of the semiconductor wafer are deleted.

[0017] In this example, by filtering data based on process parameter specifications, ion implantation curve trends, and actual electrical performance parameter ranges, the quality and reliability of semiconductor wafer fabrication datasets are improved.

[0018] In one example, process parameters and actual electrical performance parameters of a portion of semiconductor wafer fabrication are selected from the dataset as training data, including: Based on the process parameters and actual electrical performance parameters of each semiconductor wafer, the semiconductor wafers are clustered to obtain at least one classification cluster; each classification cluster includes multiple semiconductor wafers. The same number of semiconductor wafers are obtained from each of the aforementioned classification clusters, and the process parameters and actual electrical performance parameters of the semiconductor wafers are used as the training data.

[0019] In this example, by clustering semiconductor wafer fabrication based on process and electrical performance parameters and selecting an equal amount of data from each cluster as training data, the training data is balanced and comprehensive, thereby improving the model's generalization ability.

[0020] In one example, the method further includes: Obtain the process parameters of the semiconductor wafer under test; and input the process parameters of the semiconductor wafer under test into the trained electrical performance prediction model to obtain and output the electrical performance prediction results of the semiconductor wafer under test. The electrical performance prediction results include the predicted second electrical performance parameters of the semiconductor process wafer under test.

[0021] In this example, by obtaining the process parameters of the semiconductor wafer under test and inputting them into a trained model for prediction, the effect of quickly and accurately obtaining the prediction results of the electrical performance of the semiconductor wafer under test is achieved.

[0022] Secondly, embodiments of this application provide a training apparatus for predicting the electrical performance of semiconductor wafers, comprising: The acquisition module is used to acquire training data, which includes process parameters of multiple semiconductor wafers and actual electrical performance parameters of the semiconductor wafers. The training module is used to input the process parameters into the electrical performance prediction model for processing during the iterative training process of the electrical performance prediction model to obtain predicted electrical performance parameters; and to correct the model parameters of the electrical performance prediction model based on multiple preset loss functions, using the actual electrical performance parameters and the predicted electrical performance parameters, to obtain the corrected model parameters of the electrical performance prediction model.

[0023] In this embodiment, by acquiring training data containing multiple semiconductor wafer fabrication process parameters and actual electrical performance parameters, and during iterative training, using multiple loss functions to correct the model parameters of the electrical performance prediction model based on the predicted electrical performance parameters output by the electrical performance prediction model and the actual electrical performance parameters contained in the training data, the accuracy and reliability of the prediction model for the electrical performance of semiconductor wafer fabrication are improved.

[0024] In one example, the training module is used for: Based on multiple preset loss functions, multiple loss feature values ​​are calculated between the predicted electrical performance and the actual electrical performance parameters; the multiple loss feature values ​​are fused to obtain the fused loss feature value for the current iteration. The model parameters of the electrical performance prediction model are corrected using the fusion loss feature values ​​to obtain the corrected model parameters of the electrical performance prediction model in the current iteration.

[0025] In this example, by using multiple loss functions, multiple loss feature values ​​are calculated between the predicted electrical performance parameters and the actual electrical performance parameters. These multiple loss feature values ​​are then fused to obtain a fused loss feature value. The model parameters are then corrected based on this fused loss feature value. This approach further improves the model training efficiency and enhances the prediction accuracy and robustness of the trained model.

[0026] In one example, the training module is used for: The mean square error loss function, mean absolute error loss function, and mean absolute percentage error loss function are used to calculate the mean square error loss characteristic value, mean absolute error loss characteristic value, and mean absolute percentage error loss characteristic value based on the predicted electrical performance parameters and the actual electrical performance parameters, respectively. Based on preset weighting coefficients, the weighted sum of the mean squared error loss feature value, the mean absolute error loss feature value, and the mean absolute percentage error loss feature value is calculated to obtain the fusion loss feature value for the current iteration round.

[0027] In this example, the mean squared error loss feature value, mean absolute error loss feature value, and mean absolute percentage error loss feature value are calculated separately, and then fused according to the weighting coefficient to obtain the fused loss feature value. This method achieves accurate calculation of the fused loss feature value. Based on the weighting coefficient, the utilization rate of different error features is improved, the convergence speed and optimization direction accuracy of the electrical performance prediction model in the iteration process are enhanced, and the prediction effect of the electrical performance prediction model is optimized.

[0028] In one example, the training module is used for: Based on the mean squared error loss feature value, the mean absolute error loss feature value, and the mean absolute percentage error loss feature value, the weighting coefficients of the mean squared error, the mean absolute error, and the mean absolute percentage error are optimized.

[0029] In this example, the weight coefficients are optimized by dynamically optimizing the corresponding weight coefficients based on the mean squared error loss feature value, the mean absolute error loss feature value, and the mean absolute percentage error loss feature value, thereby improving the training efficiency and training accuracy of the electrical performance prediction model.

[0030] In one example, the training module is used for: The energy and dose of ion implantation in multiple processes during the production of each semiconductor wafer are collected to generate the process parameters of the semiconductor wafer. The breakdown voltage and on-resistance of each semiconductor wafer are measured to obtain the actual electrical performance parameters of the semiconductor wafer. The process parameters and the actual electrical performance parameters are normalized to generate the dataset of the semiconductor wafer fabrication. The process parameters and actual electrical performance parameters of a portion of semiconductor wafer fabrication are selected from the dataset and used as the training data.

[0031] In this example, by collecting the energy and dose of ion implantation, the breakdown voltage and on-resistance of semiconductor wafers, and generating a dataset after normalization, and selecting a portion of it as training data, a high-quality training dataset is constructed, which improves the training effect of the electrical performance prediction model.

[0032] In one example, the training module is used for: If the process parameters exceed the preset specifications, then delete the process parameters and the actual electrical performance parameters of the semiconductor wafer. If the energy curve of ion implantation in each process step of the process parameters changes abnormally, or if the dose curve of ion implantation in each process step changes abnormally, then the process parameters and the actual electrical performance parameters of the semiconductor wafer are deleted. If the actual electrical performance parameters exceed the preset range, then the process parameters and the actual electrical performance parameters of the semiconductor wafer are deleted.

[0033] In this example, by filtering data based on process parameter specifications, ion implantation curve trends, and actual electrical performance parameter ranges, the quality and reliability of semiconductor wafer fabrication datasets are improved.

[0034] In one example, the training module is used for: Based on the process parameters and actual electrical performance parameters of each semiconductor wafer, the semiconductor wafers are clustered to obtain at least one classification cluster; each classification cluster includes multiple semiconductor wafers. The same number of semiconductor wafers are obtained from each of the aforementioned classification clusters, and the process parameters and actual electrical performance parameters of the semiconductor wafers are used as the training data.

[0035] In this example, by clustering semiconductor wafer fabrication based on process and electrical performance parameters and selecting an equal amount of data from each cluster as training data, the training data is balanced and comprehensive, thereby improving the model's generalization ability.

[0036] In one example, the prediction module is used for: Obtain the process parameters of the semiconductor wafer under test; and input the process parameters of the semiconductor wafer under test into the trained electrical performance prediction model to obtain and output the electrical performance prediction results of the semiconductor wafer under test. The electrical performance prediction results include the predicted second electrical performance parameters of the semiconductor process wafer under test.

[0037] In this example, by obtaining the process parameters of the semiconductor wafer under test and inputting them into a trained model for prediction, the effect of quickly and accurately obtaining the prediction results of the electrical performance of the semiconductor wafer under test is achieved.

[0038] Thirdly, embodiments of this application provide a computer device, including: a memory and a processor, wherein the memory and the processor are communicatively connected to each other, the memory stores computer instructions, and the processor executes the computer instructions to perform the method described in any of the above embodiments.

[0039] Fourthly, embodiments of this application provide a computer-readable storage medium storing computer instructions, which are used to cause a computer to perform the method described in any one of the above embodiments.

[0040] Fifthly, embodiments of this application provide a computer program product, including computer instructions, which are used to cause a computer to perform the method described in any of the above embodiments. Attached Figure Description

[0041] To more clearly illustrate the technical solutions in the specific embodiments of this application or the prior art, the drawings used in the description of the specific embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are some embodiments of this application. For those skilled in the art, other drawings can be obtained from these drawings without creative effort.

[0042] Figure 1 This is a flowchart illustrating a method for training a prediction model of the electrical performance of a semiconductor wafer according to an embodiment of this application. Figure 2 This is a flowchart illustrating a method for predicting the electrical performance of a semiconductor wafer according to an embodiment of this application. Figure 3 This is a structural diagram of a training device for a semiconductor wafer electrical performance prediction model according to an embodiment of this application; Figure 4 This is a structural diagram of a semiconductor wafer electrical performance prediction device according to an embodiment of this application; Figure 5 This is a schematic diagram of the structure of a computer device provided in an embodiment of this application. Detailed Implementation

[0043] To make the objectives, technical solutions, and advantages of the embodiments of this application clearer, the technical solutions of the embodiments of this application will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of this application, not all embodiments. Based on the embodiments of this application, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of this application.

[0044] In the field of semiconductor device performance prediction, machine learning methods such as linear neural networks, random forest regression, and support vector regression only achieve an accuracy of 95% when predicting electrical performance, which is unsatisfactory. Furthermore, deep neural networks (DNNs) trained with a single loss function have several problems. For example, the Mean Squared Error (MSE) loss function is sensitive to outliers, and a few large error samples can easily affect the results. Also, when the actual value is close to 0, training with the Mean Absolute Percentage Error (MAPE) loss function may fail to converge. Moreover, the MAE loss function measures the magnitude of the average error but ignores the direction; its derivative is constant, making it prone to oscillations at the minimum value of the loss function and difficult to converge to the optimal value.

[0045] To address the aforementioned issues, this application proposes a novel training method for predicting the electrical performance of semiconductor wafers.

[0046] First, the computer equipment can acquire the process parameters of semiconductor wafer fabrication and, based on these parameters, analyze the impact of different process conditions on the electrical performance of the semiconductor wafer, thereby generating a dataset for the electrical performance prediction model. Specifically, the input process parameters can be the energy and dose of ion implantation at each stage of the process. These process parameters can be collected during the semiconductor wafer fabrication process. These process parameters can be output in a standard tabular format. The output electrical performance parameters mainly include breakdown voltage, on-resistance, and threshold voltage.

[0047] Next, the computer equipment can filter and delete failed semiconductor wafers and erroneous data, normalize the collected process parameters and electrical properties to the range of 0 to 1, and then divide them into training and test sets to achieve dataset preprocessing.

[0048] Then, the computer device can construct an electrical performance prediction model based on a deep neural network. This model can include a combined loss function. This combined loss function can be obtained by weighting MSE, MAPE, and Mean Absolute Error (MAE). These three loss functions can be calculated based on the predicted electrical performance parameters and the actual electrical performance parameters. The fusion of these three loss functions can be optimized during model training using gradient descent based on the initial weights to obtain the optimal value. The use of this combined loss function enhances the model's robustness to outliers during training and improves the convergence speed.

[0049] Finally, the computer device can input the training set into the electrical performance prediction model for training. During training, the computer device can use the loss function gradient to guide the updating of neuron weights and employ the ADAM optimizer to optimize the learning rate. At the end of training, the computer device can save the model weights and test the electrical performance prediction model using a test set. The computer device can determine the electrical performance prediction model as the final electrical performance prediction model when the test accuracy on the test set reaches a preset value.

[0050] According to an embodiment of this application, a training method for a prediction model of the electrical performance of semiconductor wafers is provided. It should be noted that the steps shown in the flowchart in the accompanying drawings can be executed in a computer device via a set of computer-executable instructions. Although a logical order is shown in the flowchart, in some cases, the steps shown or described can be executed in a different order. The computer device can be a terminal device, a personal computer, a server, or other device with a certain computational capacity.

[0051] Figure 1 This is a flowchart illustrating a training method for a semiconductor wafer electrical performance prediction model according to an embodiment of this application. Figure 1 As shown, with a computer device as the execution subject, the process includes the following steps: S101. Obtain training data, which includes process parameters of multiple semiconductor wafers and actual electrical performance parameters of the semiconductor wafers.

[0052] For example, the computer device can first acquire the process parameters of the semiconductor wafer during its manufacturing process. Furthermore, after the semiconductor wafer production is completed, the computer device can test the wafer to obtain its actual electrical performance parameters. The computer device can then use these process parameters and actual electrical performance parameters to form training data for an electrical performance prediction model. Based on this training data, the electrical performance prediction model can learn the intrinsic relationship between process parameters and electrical performance.

[0053] In one implementation, the process parameters may include the energy and dose of ion implantation. In the entire semiconductor manufacturing process, multiple different ion implantation processes are typically performed. Each ion implantation process has specific energy and dose conditions. Computer equipment can record this energy and dose to generate the process parameters.

[0054] In one implementation, the process parameters can be displayed in the form of a table or similar document.

[0055] In one implementation, the process parameters may also include parameters such as exposure time and development time in the photolithography process, and etching rate and etching time in the etching process.

[0056] In one implementation, the process parameters can be collected in real time by a data acquisition system during the actual production process.

[0057] In one implementation, the actual electrical performance parameters can reflect the performance of the semiconductor wafer in practical applications.

[0058] In one implementation, the actual electrical performance parameter may include a breakdown voltage. This breakdown voltage reflects the maximum voltage at which the semiconductor device can maintain its insulating state when subjected to voltage.

[0059] In one implementation, the actual electrical performance parameter may include on-resistance. This on-resistance reflects the degree to which the device impedes current flow when it is in the on-state.

[0060] In one implementation, the actual electrical performance parameter may include a threshold voltage. This threshold voltage is the minimum voltage required for the semiconductor device to begin conducting.

[0061] In one implementation, the actual electrical performance parameters may also include parameters such as leakage current and switching speed.

[0062] In one implementation, the actual electrical performance parameters can be collected in real time by a data acquisition system during actual production. Alternatively, the actual electrical performance parameters can be extracted from an existing experimental database. Or, the actual electrical performance parameters can be obtained through simulation.

[0063] S102. During the iterative training of the electrical performance prediction model, the process parameters are input into the electrical performance prediction model for processing to obtain the predicted electrical performance parameters. Based on multiple preset loss functions, the model parameters of the electrical performance prediction model are corrected using the actual electrical performance parameters and the predicted electrical performance parameters to obtain the corrected model parameters of the electrical performance prediction model.

[0064] For example, after acquiring training data, the computer device can input the training data into a preset electrical performance prediction model. This electrical performance prediction model can be trained using a fusion loss function to obtain a model capable of accurately predicting the electrical performance of semiconductor wafers.

[0065] In each iteration of the electrical performance prediction model training process, the computer first inputs the process parameters into the model for processing. Based on these process parameters, the model then outputs predicted electrical performance parameters.

[0066] The computer device can use multiple preset loss functions to determine the difference between the actual electrical performance parameters and the predicted electrical performance parameters. Then, based on this difference, the computer device can correct the model parameters of the electrical performance prediction model in the current iteration, obtaining the corrected model parameters.

[0067] The corrected model parameters will be used in the next iteration to output predicted electrical performance parameters based on the process parameters.

[0068] In one implementation, the computer device can construct an electrical performance prediction model based on machine learning or deep learning algorithms. For example, this electrical performance prediction model can be constructed based on models such as deep neural network models, support vector machine models, and decision tree models.

[0069] In one implementation, the electrical performance prediction model is used to predict the electrical performance parameters of a semiconductor wafer based on input process parameters.

[0070] In one implementation, the predicted electrical performance parameters are the electrical performance values ​​predicted by the electrical performance prediction model in this iteration based on the input semiconductor wafer fabrication process parameters.

[0071] In one implementation, the predicted electrical performance parameters may include breakdown voltage, on-resistance, threshold voltage, etc. These predicted electrical performance parameters are important indicators for evaluating the performance of semiconductor wafer fabrication.

[0072] In one implementation, the computer device can terminate the iteration when the preset number of iterations has been reached. Alternatively, the computer device can terminate the iteration when the loss function converges. The computer device can use the final output model parameters to construct the electrical performance prediction model.

[0073] In this embodiment, by acquiring training data containing multiple semiconductor wafer fabrication process parameters and actual electrical performance parameters, and during iterative training, using multiple loss functions to correct the model parameters of the electrical performance prediction model based on the predicted electrical performance parameters output by the electrical performance prediction model and the actual electrical performance parameters contained in the training data, the accuracy and reliability of the prediction model for the electrical performance of semiconductor wafer fabrication are improved.

[0074] In one example, in step S102 above, based on multiple preset loss functions, the model parameters of the electrical performance prediction model are corrected using real electrical performance parameters and predicted electrical performance parameters to obtain the corrected model parameters of the electrical performance prediction model, including: S1021. Based on multiple preset loss functions, calculate multiple loss feature values ​​between the predicted electrical performance and the actual electrical performance parameters. Fuse the multiple loss feature values ​​to obtain the fused loss feature value for the current iteration.

[0075] For example, the computer device can have multiple pre-defined loss functions. Based on the calculation formula of each loss function, the computer device can use the predicted electrical performance of each semiconductor wafer to participate in the calculation of the actual electrical performance parameters, thereby obtaining the loss feature value corresponding to each loss function. The computer device can then fuse these multiple loss feature values ​​to obtain the fused loss feature value for the current iteration.

[0076] In one implementation, the fusion loss function combines the advantages of multiple loss functions, improving the model's sensitivity to different types of errors, thereby enhancing the model's prediction accuracy and robustness.

[0077] In one implementation, the computer device can perform fusion processing through methods such as weighted fusion, optimal selection, and calculation of the average.

[0078] S1022. By fusing loss eigenvalues, the model parameters of the electrical performance prediction model are corrected to obtain the corrected model parameters of the electrical performance prediction model in the current iteration round.

[0079] For example, the computer device optimizes the model parameters of the electrical performance prediction model based on the loss value and the preset learning rate, and completes the training of the electrical performance prediction model in the current iteration.

[0080] The computer device determines the direction of parameter adjustment for the electrical performance prediction model based on the fusion loss feature values. It can also determine the step size for parameter adjustment based on a preset learning rate. Furthermore, the computer device can correct the model parameters of the electrical performance prediction model in the current iteration and obtain the corrected model parameters.

[0081] Computer equipment can improve model performance by continuously adjusting model parameters to make the model's predictions closer to the actual electrical performance parameters.

[0082] In one implementation, the loss value reflects the magnitude of the model's prediction error under the current parameters.

[0083] In one implementation, the learning rate controls the step size for updating model parameters. An excessively large learning rate may cause the model parameters to update too quickly, missing the optimal solution. An excessively small learning rate will slow down the model training process and result in excessively long convergence times.

[0084] In one implementation, common optimization algorithms include gradient descent, stochastic gradient descent, mini-batch gradient descent, and Adam optimization algorithm.

[0085] Gradient descent updates the parameters by calculating the gradient of the loss function with respect to the model parameters and then moving in the opposite direction of the gradient. Adam optimization, on the other hand, combines the concepts of momentum and adaptive learning rate, enabling more efficient parameter updates.

[0086] For example, by using multiple loss functions to calculate multiple loss feature values ​​between predicted electrical performance parameters and actual electrical performance parameters, and fusing these multiple loss feature values ​​to obtain a fused loss feature value, and then correcting the model parameters based on the fused loss feature value, the model training efficiency is further improved, and the prediction accuracy and robustness of the trained model are enhanced.

[0087] In one example, the loss function may include the mean squared error loss function, the mean absolute error loss function, and the mean absolute percentage error loss function.

[0088] The mean squared error (MSE) measures the model's prediction error by calculating the average of the squares of the errors between the predicted and actual electrical performance parameters. This MSE can penalize larger errors more severely and is suitable for scenarios where the square of the error is sensitive.

[0089] The mean absolute error (MAE) measures the error by calculating the average of the absolute differences between the predicted and actual electrical performance parameters. This MAE directly reflects the average deviation between the predicted and actual electrical performance parameters and is relatively insensitive to outliers.

[0090] The mean absolute percentage error (MASE) is calculated by taking the average of the ratios of the absolute values ​​of the errors between the predicted and actual electrical performance parameters to the actual values. This MASE can be expressed as a percentage and reflects the relative accuracy of the prediction, making it suitable for situations where the dimensions of the target variables differ significantly.

[0091] In step S1021 above, based on multiple preset loss functions, multiple loss feature values ​​are calculated between the predicted electrical performance and the actual electrical performance parameters; these multiple loss feature values ​​are then fused to obtain the fused loss feature value for the current iteration, including: Step 211: Using the mean square error loss function, the mean absolute error loss function, and the mean absolute percentage error loss function, calculate the mean square error loss characteristic value, the mean absolute error loss characteristic value, and the mean absolute percentage error loss characteristic value based on the predicted electrical performance parameters and the actual electrical performance parameters, respectively.

[0092] For example, after obtaining the predicted electrical performance parameters, the computer device calculates the mean square error, mean absolute error, and mean absolute percentage error based on the predicted electrical performance parameters and the actual electrical performance parameters, respectively, to obtain the mean square error loss characteristic value, the mean absolute error loss characteristic value, and the mean absolute percentage error loss characteristic value.

[0093] In one implementation, the mean square error (MSE) is an indicator that measures the magnitude of the error by calculating the average of the squares of the errors between the predicted and actual electrical performance parameters. The formula for calculating the MSE can be: in, This represents the mean square error. These are actual electrical performance parameters. To predict electrical performance parameters. The number of training data samples used in the current iteration.

[0094] In one implementation, the mean absolute error (MAE) is the average of the absolute differences between the predicted and actual electrical performance parameters. The formula for calculating the MAE can be: in, This represents the mean absolute error. These are actual electrical performance parameters. To predict electrical performance parameters. The number of training data samples used in the current iteration.

[0095] In one implementation, the mean absolute percentage error (MASE) is the average of the ratios of the absolute values ​​of the errors between the predicted and actual electrical performance parameters to the actual values. This MASE is expressed as a percentage. The formula for calculating the MASE is: in, The mean absolute percentage error. These are actual electrical performance parameters. To predict electrical performance parameters. The number of training data samples used in the current iteration.

[0096] Step 212: Based on the preset weight coefficients, calculate the weighted sum of the mean squared error loss feature value, the mean absolute error loss feature value, and the mean absolute percentage error loss feature value to obtain the fusion loss feature value for the current iteration round.

[0097] For example, the computer device calculates the weighted sum of the mean squared error loss feature value, the mean absolute error loss feature value, and the mean absolute percentage error loss feature value according to preset weight coefficients, and obtains the loss value for this iteration.

[0098] In one implementation, the weighting coefficients can be set empirically. The computer device can then use these set weighting coefficients to train the electrical performance prediction model during a single training session.

[0099] In one implementation, the weight coefficients can be adjusted during training based on actual conditions. The computer can optimize the weight coefficients simultaneously with the model parameters in each iteration.

[0100] In one implementation, the computer device can employ a grid search method, pre-setting a series of combinations of weight coefficients, and then iterating through these combinations to calculate the model's performance metrics under different combinations. The computer device can then select the weight coefficient combination that optimizes the model's performance.

[0101] In one implementation, the formula for calculating the loss value can be: in, This is the loss value. This represents the mean square error. This represents the mean absolute error. The mean absolute percentage error. These are the weighting coefficients for the mean squared error loss characteristic value, the mean absolute error loss characteristic value, and the mean absolute percentage error loss characteristic value.

[0102] In this example, the electrical performance prediction model is iteratively optimized by calculating the fused loss value based on the predicted electrical performance parameters and the actual electrical performance parameters during the iterative training of the electrical performance prediction model, and by optimizing the model parameters based on the loss value and the preset learning rate. This improves the convergence speed and the accuracy of the optimization direction of the electrical performance prediction model during the iteration process, thereby optimizing the prediction effect of the electrical performance prediction model.

[0103] In one example, during each iteration, the computer device can also optimize the weighting coefficients of each loss based on the error loss feature values ​​of each error. This process may include: Step 213: Based on the mean square error loss feature value, the mean absolute error loss feature value, and the mean absolute percentage error loss feature value, optimize the weighting coefficients of mean square error, mean absolute error, and mean absolute percentage error.

[0104] For example, during the iteration process, the computer device can optimize the weighting coefficients of the mean squared error, mean absolute error, and mean absolute percentage error based on the mean squared error loss feature value, mean absolute error loss feature value, and mean absolute percentage error loss feature value for that iteration.

[0105] In one implementation, the computer device can redetermine the weighting coefficients of each loss based on the magnitude of each error loss characteristic value. For example, after normalizing the mean squared error loss characteristic value, the mean absolute error loss characteristic value, and the mean absolute percentage error loss characteristic value to a range of 0-1, the computer device can calculate the ratio of the normalized mean squared error loss characteristic value, the mean absolute error loss characteristic value, and the mean absolute percentage error loss characteristic value. The computer device can then determine the weighting coefficients of these characteristic values ​​based on this ratio. For example, when the ratio of the mean squared error loss characteristic value, the mean absolute error loss characteristic value, and the mean absolute percentage error loss characteristic value is 1:2:5, the weighting coefficient of the mean squared error loss characteristic value can be 0.125, the weighting coefficient of the mean absolute error loss characteristic value can be 0.25, and the weighting coefficient of the mean absolute percentage error loss characteristic value can be 0.625.

[0106] In one implementation, the computer device can determine the adjustment value of the weight coefficient of each loss based on the magnitude of each error loss characteristic value. For example, after normalizing the mean squared error loss characteristic value, the mean absolute error loss characteristic value, and the mean absolute percentage error loss characteristic value to a range of 0-1, the computer device can calculate the mean of the normalized mean squared error loss characteristic value, the mean absolute error loss characteristic value, and the mean absolute percentage error loss characteristic value. The computer device can then calculate the difference between the normalized mean squared error loss characteristic value, the mean absolute error loss characteristic value, and the mean absolute percentage error loss characteristic value and the mean. The computer device can then multiply a preset compensation base value by each of these three differences to obtain three adjustment values. The computer device can then adjust the three weight coefficients according to these three adjustment values.

[0107] In one implementation, the computer device can determine the range for generating new weight coefficients for each loss based on the magnitude of each error loss feature value. Within this range, the computer device can use a random search method to generate new error loss feature values.

[0108] In one implementation, the computer device can use a gradient descent optimization method to update the weight coefficients by calculating the gradient of the loss function with respect to the weight coefficients, and then updating the weight coefficients in the reverse direction of the gradient. This gradually approximates the optimal solution, and this method is highly efficient when dealing with large-scale data and complex models.

[0109] In this example, the weight coefficients are optimized by dynamically optimizing the corresponding weight coefficients based on the mean squared error loss feature value, the mean absolute error loss feature value, and the mean absolute percentage error loss feature value, thereby improving the training efficiency and training accuracy of the electrical performance prediction model.

[0110] In one example, step S101 above, obtaining training data, includes: S1011. Collect the energy and dose of ion implantation in multiple processes during the production of each semiconductor wafer to generate process parameters for the semiconductor wafer.

[0111] For example, computer equipment performs data acquisition in the semiconductor wafer fabrication process, accurately collecting ion implantation energy and dosage information for each semiconductor wafer in multiple processes of ion implantation, and generating process parameters for semiconductor wafer fabrication based on this information.

[0112] In one implementation, ion implantation is a key process in semiconductor manufacturing. It involves implanting ions of specific energies into a semiconductor material to alter its electrical properties. The energy of the ion implantation determines the depth to which the ions can penetrate the semiconductor material, while the dose represents the number of ions implanted per unit area. Different combinations of energy and dose can significantly impact the performance of semiconductor wafer fabrication.

[0113] In one implementation, when the computer device collects this data, it can utilize a data acquisition module equipped with a dedicated ion implantation device, which can record the energy and dose parameters of each ion implantation in real time. Alternatively, independent sensors and data recording instruments can be used to monitor and store relevant parameters during the ion implantation process.

[0114] S1012. Measure the breakdown voltage and on-resistance of each semiconductor wafer to obtain the actual electrical performance parameters of the semiconductor wafer.

[0115] For example, the computer device can also acquire the breakdown voltage and on-resistance obtained by measuring the electrical performance of each semiconductor wafer, and then use them as the true electrical performance parameters of the semiconductor wafer.

[0116] In one implementation, the breakdown voltage refers to the critical voltage value at which a semiconductor device transitions from an insulating state to a conductive state when subjected to voltage. When the voltage exceeds the breakdown voltage, the semiconductor device may be damaged and malfunction.

[0117] In one implementation, the on-resistance is the degree to which a semiconductor device impedes the flow of current when it is in the on state. The smaller the on-resistance, the lower the energy loss of the device when it is in the on state, and the better its performance.

[0118] In one implementation, when measuring these parameters, the computer equipment can utilize specialized electrical performance testing instruments, such as high-precision voltmeters, ammeters, and resistance testers. These instruments can accurately measure the breakdown voltage and on-resistance of the semiconductor wafer under different conditions. Alternatively, an automated testing system can be used, where the semiconductor wafer is placed in a test fixture, and the computer-controlled testing instruments automatically complete the measurement process and record the results.

[0119] S1016. Normalize the process parameters and actual electrical performance parameters to generate a dataset for semiconductor wafer fabrication.

[0120] For example, the computer device normalizes the collected process parameters and the measured actual electrical performance parameters to generate a dataset of semiconductor wafer fabrication.

[0121] In one implementation, normalization can eliminate the differences in dimensions and orders of magnitude between different parameters, enabling the parameters to be compared and analyzed within the same scale range, thereby improving the training efficiency and prediction accuracy of the model.

[0122] In one implementation, normalization can map the process parameters and actual electrical performance parameters to a range of 0-1. The calculation formula is as follows: in, The value obtained from the mapping. This is the original value. It is the minimum value. This is the maximum value.

[0123] It is important to note that this normalization process treats each value individually.

[0124] For example, when normalizing the energy in process parameters, It is the minimum value among all energies. It is the maximum value among all energies. This represents the original value of the energy currently being processed. This represents the new energy value.

[0125] In one implementation, normalization can convert the process parameters and actual electrical performance parameters into a distribution with a mean of 0 and a standard deviation of 1. The calculation formula is as follows: in, The value obtained from the mapping. This is the original value. This is the mean. The standard deviation is denoted as .

[0126] It is important to note that this normalization process treats each value individually.

[0127] For example, when normalizing the energy in process parameters, This represents the original value of the energy currently being processed. This represents the new energy value. This is the average of all energy. The standard deviation is calculated based on the total energy values ​​and the mean.

[0128] S1017. Select some semiconductor wafer fabrication process parameters and actual electrical performance parameters from the dataset as training data.

[0129] For example, a computer device selects a portion of the semiconductor wafer process parameters and actual electrical performance parameters from the generated dataset as training data to form a training dataset. This training data is used to train the model.

[0130] In one implementation, the computer device can use the remaining data in the dataset as test data to form a test dataset.

[0131] In one implementation, when selecting training data, the computer device can use a random sampling method to randomly select a certain proportion of samples from the dataset as training data.

[0132] In one implementation, the computer device can use a stratified sampling method to divide the dataset into several layers based on certain characteristics of the semiconductor wafer fabrication process, and then extract samples from each layer according to a certain proportion to form training data. For example, the layers can be stratified based on different process parameters.

[0133] In one implementation, when selecting training data, the computer device can also incorporate data augmentation techniques to transform and expand the selected training data, increasing its diversity and further improving the model's generalization ability. For example, this could involve applying small random perturbations to process parameters or reasonably scaling the actual electrical performance parameters.

[0134] In this example, by collecting the energy and dose of ion implantation, the breakdown voltage and on-resistance of semiconductor wafers, and generating a dataset after normalization, and selecting a portion of it as training data, a high-quality training dataset is constructed, which improves the training effect of the electrical performance prediction model.

[0135] In one example, in step S1017 above, selecting process parameters and actual electrical performance parameters of a portion of semiconductor wafer fabrication data from the dataset as training data includes: Step 171: Based on the process parameters and actual electrical performance parameters of each semiconductor wafer, cluster the semiconductor wafers to obtain at least one classification cluster. Each classification cluster includes multiple semiconductor wafers.

[0136] For example, after obtaining the dataset, the computer device can treat the process parameters and actual electrical performance parameters corresponding to each semiconductor wafer as a single data set.

[0137] Computer equipment can cluster semiconductor wafers based on data from each individual wafer. The computer equipment can then create at least one cluster, with each cluster containing multiple semiconductor wafers. This process aims to group semiconductor wafers with similar process characteristics and electrical properties into one category, helping to discover the inherent patterns and differences between different types of semiconductor wafers, thus facilitating more accurate model training and performance analysis later on.

[0138] In one implementation, the clustering used in this application is an unsupervised learning algorithm. This clustering algorithm does not require pre-labeling of data categories, but rather divides the data into different clusters based on the characteristics of the data itself.

[0139] In one implementation, the clustering algorithm can be an existing clustering algorithm or an improved clustering algorithm.

[0140] Step 172: Obtain the same number of semiconductor wafers from each cluster, and use the process parameters and actual electrical performance parameters of the semiconductor wafers as training data.

[0141] For example, a computer device can select the same number of semiconductor wafers from each category cluster. The computer device can then use the corresponding process parameters and actual electrical performance parameters of these semiconductor wafers as training data to form a training dataset. This selection process ensures that the training data covers the characteristics of different types of semiconductor wafers, avoiding bias in model training due to an excessive or insufficient amount of data of certain types, thereby improving the model's generalization ability and accuracy.

[0142] In one implementation, the computer device can randomly select a specified number of semiconductor wafers from each cluster.

[0143] In one implementation, if there are different subcategories or hierarchical structures within a classification cluster, the computer device can select semiconductor wafers from each subcategory according to a certain proportion to ensure that the training data can more comprehensively reflect the characteristics of the classification cluster.

[0144] In one implementation, the computer device can also prioritize semiconductor wafers with higher importance based on their significance within the classification cluster, thereby improving the quality and representativeness of the training data. For example, this importance can be determined based on the distance between the semiconductor wafer and the cluster center within the classification cluster. Alternatively, the importance can be determined based on the contribution of the semiconductor wafer to the classification features within the classification cluster.

[0145] In one implementation, if the number of semiconductor wafers in a cluster is less than the required number, the computer device can generate simulated process parameters based on the manufacturing process of the semiconductor wafers in the cluster. Based on these simulated process parameters, a model is used to perform simulations to obtain the simulated actual electrical performance parameters. The computer device can then add the data generated from the simulated process parameters and the simulated actual electrical performance parameters to the dataset to improve the dataset's balance.

[0146] In this example, by clustering semiconductor wafer fabrication based on process and electrical performance parameters and selecting an equal amount of data from each cluster as training data, the training data is balanced and comprehensive, thereby improving the model's generalization ability.

[0147] In one example, before normalizing the process parameters and actual electrical performance parameters in step S1016 above to generate the semiconductor wafer fabrication dataset, the computer device can also preprocess the data, which may include at least one of the following steps: S1013. If the process parameters exceed the preset specifications, delete the process parameters and actual electrical performance parameters of the semiconductor wafer fabrication.

[0148] For example, the computer equipment checks the collected semiconductor wafer fabrication process parameters to determine if they exceed preset specifications. If any process parameter is found to exceed the preset specifications, the computer equipment immediately deletes the corresponding process parameters and actual electrical performance parameters for that semiconductor wafer. This operation aims to ensure that the data retained in the dataset conforms to production specifications and requirements, preventing abnormal data from interfering with subsequent model training and performance analysis, and improving data quality and reliability.

[0149] In one implementation, the preset specification is a pre-defined range of reasonable values ​​for process parameters. This range is determined based on the process requirements and quality control standards of semiconductor manufacturing.

[0150] In one implementation, when the computer equipment determines whether the process parameters exceed the preset specifications, it can use a step-by-step comparison method to compare each process parameter with the upper and lower limits in the preset specifications.

[0151] S1014. If the energy curve of ion implantation in each process step has an abnormal trend, or the dose curve of ion implantation in each process step has an abnormal trend, then delete the process parameters and actual electrical performance parameters of the semiconductor wafer fabrication.

[0152] For example, the computer device analyzes the trends of energy and dose curves generated during ion implantation at each stage of the semiconductor wafer fabrication process. If an abnormal trend is found in the energy curve or the dose curve, the computer device deletes the corresponding process parameters and actual electrical performance parameters for that semiconductor wafer. This step is to exclude data showing abnormal process fluctuations during production, ensuring the stability and consistency of the data in the dataset, preventing data corresponding to abnormal process curves from misleading model training, and improving the model's ability to fit normal production processes.

[0153] In one implementation, the energy curve is plotted with the process step as the x-axis and the ion implantation energy as the y-axis. It visually demonstrates the changes in ion implantation energy at different processes.

[0154] In one implementation, the dose curve is a curve plotted with the process as the horizontal axis and the ion implantation dose as the vertical axis, reflecting the changing trend of the ion implantation dose in different processes.

[0155] In one implementation, abnormal trends may manifest as sudden jumps in the curve, excessive fluctuations, or deviations from normal process variation patterns.

[0156] In one implementation, when analyzing the trend of curve changes, the computer device can use numerical analysis methods to calculate characteristic parameters such as the slope and curvature of the curve, and determine whether the curve is abnormal by setting a reasonable threshold.

[0157] In one implementation, the computer device can also utilize anomaly detection algorithms from machine learning. For example, the Isolation Forest algorithm, a type of Support Vector Machine algorithm, etc., can be used to automatically identify and classify curves.

[0158] In one implementation, the computer device can also use a preset standard curve and calculate the matching degree between the energy curve or dose curve and the preset standard curve to determine whether the energy curve or dose curve is abnormal.

[0159] Optionally, standard curves can be set separately for the energy curve and the dose curve.

[0160] Optionally, the computer device can determine that an anomaly exists when the matching degree is determined to be lower than the anomaly threshold by setting an anomaly threshold.

[0161] S1015. If the actual electrical performance parameters exceed the preset range, delete the process parameters and actual electrical performance parameters of the semiconductor wafer fabrication.

[0162] For example, the computer device performs a range check on the measured actual electrical performance parameters of the semiconductor wafer to determine if they exceed a preset range. When the actual electrical performance parameters exceed the preset range, the computer device deletes the corresponding process parameters and actual electrical performance parameters for that semiconductor wafer. This measure ensures that the electrical performance parameters included in the dataset meet product performance requirements, preventing data with abnormal performance from entering the dataset, thereby improving the quality of the dataset and enabling the model trained on this dataset to more accurately predict the normal electrical performance of the semiconductor wafer.

[0163] In one implementation, the preset range is a reasonable range of electrical performance parameters pre-defined according to the design requirements and usage standards of the semiconductor product.

[0164] In one implementation, when the computer device determines whether the actual electrical performance parameters exceed the preset range, it can use a simple conditional judgment statement to compare each parameter with the upper and lower limits of the preset range.

[0165] In this example, by filtering data based on process parameter specifications, ion implantation curve trends, and actual electrical performance parameter ranges, the quality and reliability of semiconductor wafer fabrication datasets are improved.

[0166] Figure 2 This is a flowchart illustrating a method for predicting the electrical performance of a semiconductor wafer according to an embodiment of this application, as shown below. Figure 2 As shown, based on the above steps, an electrical performance prediction model is trained. The computer device can also utilize this electrical performance prediction model through the following steps: S201. Obtain the process parameters of the semiconductor wafer to be tested.

[0167] For example, the computer device first initiates a data acquisition program to obtain the process parameters of the semiconductor wafer under test. The acquisition of these process parameters is the same as in step S1011 described above. Further details are omitted here.

[0168] S202. Input the process parameters of the semiconductor wafer under test into the trained electrical performance prediction model, obtain and output the electrical performance prediction results of the semiconductor wafer under test. The electrical performance prediction results include the predicted second electrical performance parameters of the semiconductor wafer under test.

[0169] For example, a computer device can input the process parameters of the semiconductor wafer under test into a trained electrical performance prediction model. This electrical performance prediction model is used to predict the electrical performance of a semiconductor wafer under test. Figure 1 The illustrated embodiment shows an electrical performance prediction model trained to obtain prediction results based on the process parameters. These prediction results include a second predicted electrical performance parameter. A computer device can output this second predicted electrical performance parameter.

[0170] In this example, by obtaining the process parameters of the semiconductor wafer under test and inputting them into a trained model for prediction, the effect of quickly and accurately obtaining the prediction results of the electrical performance of the semiconductor wafer under test is achieved.

[0171] Figure 3 This is a structural diagram of a training device for a semiconductor wafer electrical performance prediction model according to an embodiment of this application, as shown below. Figure 3 As shown, the training device 300 includes: The acquisition module 301 is used to acquire training data, which includes process parameters of multiple semiconductor wafers and actual electrical performance parameters of the semiconductor wafers.

[0172] The training module 302 is used to input process parameters into the electrical performance prediction model for processing during the iterative training process of the electrical performance prediction model to obtain predicted electrical performance parameters; and based on multiple preset loss functions, to correct the model parameters of the electrical performance prediction model using the real electrical performance parameters and the predicted electrical performance parameters to obtain the corrected model parameters of the electrical performance prediction model.

[0173] In one example, module 302 is used for: Based on multiple preset loss functions, multiple loss feature values ​​between the predicted electrical performance and the actual electrical performance parameters are calculated; the multiple loss feature values ​​are fused to obtain the fused loss feature value for the current iteration. By fusing loss eigenvalues, the model parameters of the electrical performance prediction model are corrected, resulting in the corrected model parameters for the current iteration.

[0174] In one example, the loss function includes the mean squared error loss function, the mean absolute error loss function, and the mean absolute percentage error loss function; training module 302 is used for: The mean square error loss function, mean absolute error loss function, and mean absolute percentage error loss function are used to calculate the mean square error loss characteristic value, mean absolute error loss characteristic value, and mean absolute percentage error loss characteristic value based on the predicted electrical performance parameters and the actual electrical performance parameters, respectively. Based on the preset weighting coefficients, the weighted sum of the mean squared error loss feature value, the mean absolute error loss feature value, and the mean absolute percentage error loss feature value is calculated to obtain the fusion loss feature value for the current iteration round.

[0175] In one example, training module 302 is used for: Based on the mean squared error loss feature value, the mean absolute error loss feature value, and the mean absolute percentage error loss feature value, the weight coefficients of mean squared error, mean absolute error, and mean absolute percentage error are optimized.

[0176] In one example, training module 302 is used for: The energy and dose of ion implantation are collected in multiple steps of the semiconductor wafer fabrication process to generate the process parameters of the semiconductor wafer fabrication.

[0177] The breakdown voltage and on-resistance of each semiconductor wafer are measured to obtain the true electrical performance parameters of the semiconductor wafer.

[0178] The process parameters and actual electrical performance parameters are normalized to generate a dataset for semiconductor wafer fabrication.

[0179] Select process parameters and actual electrical performance parameters of a portion of semiconductor wafer fabrication data from the dataset as training data.

[0180] In one example, training module 302 is used for: If the process parameters exceed the preset specifications, then delete the process parameters and actual electrical performance parameters of the semiconductor wafer fabrication.

[0181] If the energy curve of ion implantation in each process step has an abnormal trend, or if the dose curve of ion implantation in each process step has an abnormal trend, then delete the process parameters and actual electrical performance parameters of the semiconductor wafer fabrication.

[0182] If the actual electrical performance parameters exceed the preset range, then delete the process parameters and actual electrical performance parameters of the semiconductor wafer fabrication.

[0183] In one example, training module 302 is used for: Based on the process parameters and actual electrical performance parameters of each semiconductor wafer, the semiconductor wafers are clustered to obtain at least one classification cluster. Each classification cluster includes multiple semiconductor wafers.

[0184] The same number of semiconductor wafers were obtained from each cluster, and the process parameters and actual electrical performance parameters of the semiconductor wafers were used as training data.

[0185] Further functional descriptions of the above modules and units are the same as those in the corresponding embodiments described above, and will not be repeated here.

[0186] Figure 4 This is a structural diagram of a semiconductor wafer electrical performance prediction device according to an embodiment of this application, as shown below. Figure 4As shown, the prediction device 400 includes: The acquisition module 401 is used to acquire the process parameters of the semiconductor wafer under test.

[0187] The prediction module 402 is used to input the process parameters of the semiconductor wafer under test into the trained electrical performance prediction model, and obtain and output the electrical performance prediction results of the semiconductor wafer under test. The electrical performance prediction results include the predicted second electrical performance parameters of the semiconductor wafer under test.

[0188] Further functional descriptions of the above modules and units are the same as those in the corresponding embodiments described above, and will not be repeated here.

[0189] In this embodiment, the training device and the prediction device are presented in the form of functional units. Here, a unit refers to an application-specific integrated circuit (ASIC), a processor and memory that execute one or more software or fixed programs, and / or other devices that can provide the above functions.

[0190] Figure 5 This is a schematic diagram of the structure of a computer device according to an embodiment of this application, such as... Figure 5 As shown, the computer device 500 includes one or more processors 501, memory 502, and interfaces for connecting the components, including high-speed interfaces and low-speed interfaces. The components are interconnected via different buses and can be mounted on a common motherboard or otherwise installed as needed. The processors can process instructions executed within the computer device, including instructions stored in or on memory to display graphical information of a GUI on external input / output devices (such as display devices coupled to the interface). In some alternative implementations, multiple processors and / or multiple buses can be used with multiple memories and multiple memory modules, if desired. Similarly, multiple computer devices can be connected, each providing some of the necessary operations (e.g., as a server array, a group of blade servers, or a multiprocessor system). Figure 5 Take a processor 501 as an example.

[0191] Processor 501 may be a central processing unit, a network processor, or a combination thereof. Processor 501 may further include a hardware chip. The hardware chip may be an application-specific integrated circuit (ASIC), a programmable logic device (PLD), or a combination thereof. The programmable logic device may be a complex programmable logic device (CAMP), a field-programmable gate array (FPGA), a general-purpose array logic (GDA), or any combination thereof.

[0192] The memory 502 stores instructions executable by at least one processor 501 to cause at least one processor 501 to perform the method shown in the above embodiments.

[0193] Memory 502 may include a program storage area and a data storage area. The program storage area may store the operating system and applications required for at least one function. The data storage area may store data created based on the use of the computer device. Furthermore, memory 502 may include high-speed random access memory and may also include non-transitory memory, such as at least one disk storage device, flash memory device, or other non-transitory solid-state storage device. In some alternative embodiments, memory 502 may optionally include memory remotely located relative to processor 501, which can be connected to the computer device via a network. Examples of such networks include, but are not limited to, the Internet, intranets, local area networks, mobile communication networks, and combinations thereof.

[0194] Memory 502 may include volatile memory, such as random access memory. Memory may also include non-volatile memory, such as flash memory, hard disk, or solid-state drive. Memory 502 may also include combinations of the above types of memory.

[0195] The computer device also includes a communication interface 503 for communicating with other devices or communication networks.

[0196] This application also provides a computer-readable storage medium. The methods described in this application can be implemented in hardware or firmware, or implemented as recordable on a storage medium, or implemented as computer code downloaded over a network and originally stored on a remote storage medium or a non-transitory machine-readable storage medium and subsequently stored on a local storage medium. Thus, the methods described herein can be processed by software stored on a storage medium using a general-purpose computer, a dedicated processor, or programmable or dedicated hardware. The storage medium can be a magnetic disk, optical disk, read-only memory, random access memory, flash memory, hard disk, or solid-state drive, etc. Further, the storage medium may also include combinations of the above types of memory. It is understood that computers, processors, microprocessor controllers, or programmable hardware include storage components capable of storing or receiving software or computer code, which, when accessed and executed by the computer, processor, or hardware, implements the methods shown in the above embodiments.

[0197] This application provides a computer program product including computer instructions stored in a computer-readable storage medium. A processor of a computer device reads the computer instructions from the computer-readable storage medium and executes the computer instructions, causing the computer device to perform the method of any embodiment of this application.

[0198] Although embodiments of this application have been described in conjunction with the accompanying drawings, those skilled in the art can make various modifications and variations without departing from the spirit and scope of this application, and all such modifications and variations fall within the scope defined by the appended claims.

[0199] It is understood that the data involved in this technical solution (including but not limited to the data itself, the acquisition or use of the data) shall comply with the requirements of relevant laws, regulations and related provisions.

[0200] It is understood that in the specific implementation of this application, data such as user information, location information, and navigation data are involved. When the above embodiments of this application are applied to specific products or technologies, user permission or consent is required, and the collection, use and processing of related data must comply with the relevant laws, regulations and standards of the relevant countries and regions.

[0201] The devices and modules described in the above embodiments can be implemented by computer chips or physical entities, or by products with certain functions. A typical implementation device is a computer. Specifically, a computer can be, for example, a personal computer, a laptop computer, a cellular phone, a camera phone, a smartphone, a personal digital assistant, a media player, a navigation device, an email device, a game console, a tablet computer, a wearable device, or any combination of these devices.

[0202] For ease of description, the above devices are described separately by function as various units. Of course, in implementing this application, the functions of each unit can be implemented in one or more software and / or hardware.

[0203] Those skilled in the art will understand that the methods provided by the embodiments of this application can be implemented by a computer program product. Therefore, this application can take the form of a completely hardware embodiment, a completely software embodiment, or an embodiment combining software and hardware aspects. Furthermore, this application can take the form of a computer program product embodied on one or more computer-usable storage media (including but not limited to disk storage, CD-ROM, optical storage, etc.) containing computer-usable program code.

[0204] This application is described with reference to flowchart illustrations and / or block diagrams of methods, apparatus, and computer program products according to embodiments of this application. It will be understood that each block of the flowchart illustrations and / or block diagrams, and combinations of blocks in the flowchart illustrations and / or block diagrams, can be implemented by computer program instructions. These computer program instructions can be provided to a processor of a general-purpose computer, special-purpose computer, embedded processor, or other programmable data processing apparatus to produce a machine, such that the instructions, which execute via the processor of the computer or other programmable data processing apparatus, generate instructions for implementing the flowchart... Figure 1 One or more processes and / or boxes Figure 1 A device that provides the functions specified in one or more boxes.

[0205] These computer program instructions may also be stored in a computer-readable storage medium that can direct a computer or other programmable data processing device to function in a particular manner, such that the instructions stored in the computer-readable storage medium produce an article of manufacture including instruction means, which are implemented in a process Figure 1 One or more processes and / or boxes Figure 1 The function specified in one or more boxes.

[0206] These computer program instructions may also be loaded onto a computer or other programmable data processing equipment to cause a series of operational steps to be performed on the computer or other programmable equipment to produce a computer-implemented process, thereby providing instructions that execute on the computer or other programmable equipment for implementing the process. Figure 1 One or more processes and / or boxes Figure 1 The steps of the function specified in one or more boxes.

[0207] It should also be noted that the terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or apparatus that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such process, method, article, or apparatus. Unless otherwise specified, an element defined by the phrase "comprising one..." does not exclude the presence of other identical elements in the process, method, article, or apparatus that includes that element.

[0208] The various embodiments in this specification are described in a progressive manner. Similar or identical parts between embodiments can be referred to mutually. Each embodiment focuses on describing the differences from other embodiments. In particular, the apparatus embodiments are basically similar to the method embodiments, so the description is relatively simple; relevant parts can be referred to the descriptions of the method embodiments.

[0209] The above are merely embodiments of this application and are not intended to limit the scope of this application. Various modifications and variations can be made to this application by those skilled in the art. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of this application should be included within the scope of the claims of this application.

[0210] Although embodiments of this application have been described in conjunction with the accompanying drawings, those skilled in the art can make various modifications and variations without departing from the spirit and scope of this application, and such modifications and variations all fall within the scope defined by the appended claims.

Claims

1. A training method for a predictive model of the electrical performance of semiconductor wafers, characterized in that, The method includes: Acquire training data, which includes process parameters of multiple semiconductor wafers and actual electrical performance parameters of the semiconductor wafers; During the iterative training of the electrical performance prediction model, the process parameters are input into the electrical performance prediction model for processing to obtain predicted electrical performance parameters; and based on multiple preset loss functions, the model parameters of the electrical performance prediction model are corrected using the actual electrical performance parameters and the predicted electrical performance parameters to obtain the corrected model parameters of the electrical performance prediction model.

2. The method according to claim 1, characterized in that, Based on multiple preset loss functions, the model parameters of the electrical performance prediction model are corrected using the actual electrical performance parameters and the predicted electrical performance parameters, resulting in the corrected model parameters of the electrical performance prediction model, including: Based on multiple preset loss functions, multiple loss feature values ​​are calculated between the predicted electrical performance and the actual electrical performance parameters; the multiple loss feature values ​​are fused to obtain the fused loss feature value for the current iteration. The model parameters of the electrical performance prediction model are corrected using the fusion loss feature values ​​to obtain the corrected model parameters of the electrical performance prediction model in the current iteration.

3. The method according to claim 2, characterized in that, The loss function includes the mean squared error loss function, the mean absolute error loss function, and the mean absolute percentage error loss function; based on the preset multiple loss functions, multiple loss feature values ​​are calculated between the predicted electrical performance and the actual electrical performance parameters. By fusing the multiple loss feature values, the fused loss feature value for the current iteration is obtained, including: The mean square error loss function, mean absolute error loss function, and mean absolute percentage error loss function are used to calculate the mean square error loss characteristic value, mean absolute error loss characteristic value, and mean absolute percentage error loss characteristic value based on the predicted electrical performance parameters and the actual electrical performance parameters, respectively. Based on preset weighting coefficients, the weighted sum of the mean squared error loss feature value, the mean absolute error loss feature value, and the mean absolute percentage error loss feature value is calculated to obtain the fusion loss feature value for the current iteration round.

4. The method according to claim 3, characterized in that, The method further includes: Based on the mean squared error loss feature value, the mean absolute error loss feature value, and the mean absolute percentage error loss feature value, the weighting coefficients of the mean squared error, the mean absolute error, and the mean absolute percentage error are optimized.

5. The method according to any one of claims 1-4, characterized in that, Obtain training data, including: The energy and dose of ion implantation in multiple processes during the production of each semiconductor wafer are collected to generate the process parameters of the semiconductor wafer. The breakdown voltage and on-resistance of each semiconductor wafer are measured to obtain the actual electrical performance parameters of the semiconductor wafer. The process parameters and the actual electrical performance parameters are normalized to generate the dataset of the semiconductor wafer fabrication. The process parameters and actual electrical performance parameters of a portion of semiconductor wafer fabrication are selected from the dataset and used as the training data.

6. The method according to claim 5, characterized in that, Before normalizing the process parameters and the actual electrical performance parameters to generate the dataset for the semiconductor wafer fabrication, the method further includes at least one of the following: If the process parameters exceed the preset specifications, then delete the process parameters and the actual electrical performance parameters of the semiconductor wafer. If the energy curve of ion implantation in each process step of the process parameters changes abnormally, or if the dose curve of ion implantation in each process step changes abnormally, then the process parameters and the actual electrical performance parameters of the semiconductor wafer are deleted. If the actual electrical performance parameters exceed the preset range, then the process parameters and the actual electrical performance parameters of the semiconductor wafer are deleted.

7. The method according to claim 5, characterized in that, The process parameters and actual electrical performance parameters of a portion of semiconductor wafer fabrication data are selected from the dataset and used as the training data, including: Based on the process parameters and actual electrical performance parameters of each semiconductor wafer, the semiconductor wafers are clustered to obtain at least one classification cluster; each classification cluster includes multiple semiconductor wafers. The same number of semiconductor wafers are obtained from each of the aforementioned classification clusters, and the process parameters and actual electrical performance parameters of the semiconductor wafers are used as the training data.

8. The method according to any one of claims 1-4, characterized in that, The method further includes: Obtain the process parameters of the semiconductor wafer under test; and input the process parameters of the semiconductor wafer under test into the trained electrical performance prediction model to obtain and output the electrical performance prediction results of the semiconductor wafer under test. The electrical performance prediction results include the predicted second electrical performance parameters of the semiconductor wafer under test.

9. A training device for predicting the electrical performance of semiconductor wafers, characterized in that, The device includes: The acquisition module is used to acquire training data, which includes process parameters of multiple semiconductor wafers and actual electrical performance parameters of the semiconductor wafers. The training module is used to input the process parameters into the electrical performance prediction model for processing during the iterative training process of the electrical performance prediction model to obtain predicted electrical performance parameters; and to correct the model parameters of the electrical performance prediction model based on multiple preset loss functions, using the actual electrical performance parameters and the predicted electrical performance parameters, to obtain the corrected model parameters of the electrical performance prediction model.

10. A computer device, characterized in that, include: A memory and a processor, the memory and the processor being communicatively connected to each other, the memory storing computer instructions, the processor executing the computer instructions to perform the method of any one of claims 1 to 8.

11. A computer-readable storage medium, characterized in that, The computer-readable storage medium stores computer instructions for causing the computer to perform the method of any one of claims 1 to 7.

12. A computer program product, characterized in that, Includes computer instructions for causing a computer to perform the method of any one of claims 1 to 8.

Images