A Machine Learning-Based Ion Implantation Optimization Method for Silicon Carbide Power Devices

Abstract

This invention belongs to the field of silicon carbide power device manufacturing technology, specifically a machine learning-based ion implantation optimization method for silicon carbide power devices. The method includes acquiring historical ion implantation process data for silicon carbide power devices to create an ion implantation process optimization sample dataset; dividing the sample dataset into training set A1, training set A2, and test set; training an SVM prediction model using training set A1 to obtain an initial SVM prediction model; training and validating the initial SVM prediction model using a network-based transfer learning method based on training set A2 to obtain a trained SVM prediction model; and testing the trained SVM prediction model on the test set. This invention can improve the speed and accuracy of process optimization while reducing costs and scrap rates; it can efficiently adapt to various dynamic changes in ion implantation process conditions and corresponding process results.

Description

Technical Field

[0001] This invention belongs to the field of silicon carbide power device manufacturing technology, and specifically relates to an ion implantation method for silicon carbide power devices. Background Technology

[0002] In the fabrication of silicon carbide (SiC) power devices, ion implantation is a process that uses high-energy ions to inject into selective regions of a semiconductor to achieve doping, modification, and isolation. Because the diffusion coefficient of impurities in SiC is very small below 1800°C, selective doping using the thermal diffusion method in traditional silicon processes is not feasible. Ion implantation, however, allows for controlled and designed doping at appropriate high temperatures. In the fabrication of SiC power devices, ion implantation is mainly applied to ohmic contacts, bipolar transistors, MOSFETs, Schottky diodes, and JFETs. Ion implantation plays a crucial role in the manufacturing of SiC power devices, directly affecting their performance, stability, and reliability. Therefore, research into optimizing ion implantation processes is critical. Only with further optimization of ion implantation processes can SiC power devices meet a wider range of application needs, driving technological progress and industrial development in fields such as power electronics and automotive electronics.

[0003] Commonly used methods for optimizing power device processes include: (1) Experimental testing: First, the designed device is fabricated using semiconductor process technology, and then tested using testing instruments to predict and characterize the device performance; (2) Numerical simulation: Some simple models are used to explain and predict the physical behavior of the device; (3) Machine learning-based methods. Experimental testing introduces human error and external influences during the testing process, such as environment, contact resistance, test circuits and instruments, etc. In particular, wafer testing is very sensitive to ambient temperature, humidity, probe conditions and exposure time in the air. Although simulation tools can partially replace time-consuming and expensive process experiments, effectively saving design costs and time, their implementation also has problems such as non-convergence and reasonable mesh settings. Neither can adapt to the dynamic changes of various ion implantation process conditions and corresponding process results. With the development of computer technology and the big data environment, machine learning methods have been applied to process optimization. Currently, by using a large amount of ion implantation process parameter data and combining it with the idea of ​​machine learning, a transfer learning prediction model suitable for the optimization of ion implantation process of silicon carbide power devices is being constructed. By learning from a large amount of sample data through the model, the optimization of ion implantation process of silicon carbide power devices based on machine learning has become a research hotspot.

[0004] In 2018, M. Shin and J. Lee first used a CNN model to detect lithographic hotspots with an accuracy of 95.5%. In the same year, N. Hari et al. established a prediction model for the structural parameters and switching characteristics of GaN power devices using deep neural networks. This GaN power device prediction model can accurately predict the on-state characteristics of devices without complex device mechanism analysis. X. Zeng et al. established a current prediction model for IGBTs using artificial neural networks (ANNs). This current prediction model can achieve global IGBT current prediction with an average error of less than 3% without physical current sensors. In 2019, multilayer deep neural networks (DNNs) were... A predictive model for the performance of silicon junctionless nanowire field-effect transistors (FETs), including the basis factor of merit (BFOM), turn-off current, threshold voltage (Vth), and subthreshold slope (SS), was established. Comparison with the TCAD numerical simulation tool revealed that this predictive model significantly reduces computational costs. In 2020, SBKutub et al. established a predictive model for the electrical characteristics of silicon-based gallium nitride power devices based on artificial neural networks. Summary of the Invention

[0005] The purpose of this invention is to provide a machine learning-based ion implantation optimization method for silicon carbide power devices, so as to improve the stability and reliability of silicon carbide power devices and meet a wider range of application needs.

[0006] The ion implantation optimization method for silicon carbide power devices based on machine learning provided in this invention uses Support Vector Machine (SVM) as the predictive model for machine learning optimization. The specific steps are as follows:

[0007] Step 1, Sample Collection and Preprocessing: Collect historical silicon carbide power device ion implantation process data, preprocess it, and create an ion implantation process optimization sample dataset. The sample dataset includes an implantation process dataset and a label set.

[0008] (1) Production of the implantation process dataset: The information in the implantation process dataset includes the implanted ion type, implantation energy and dose, implantation angle and implantation temperature, which serve as the input source for the support vector machine prediction model. The data is normalized and scaled to between [0,1] to prevent certain features (especially those with a large numerical range) from dominating the model training, while also ensuring that all features have the same dimensions and proportions.

[0009] (2) Tag set creation: The tag set is the process result corresponding to the input injection process parameters, including resistivity, activation rate, channel depth, electron mobility and sheet resistance;

[0010] (3) Sample set division: The ion implantation process optimization sample dataset is divided into training set A1, training set A2 and test set; training set A1 accounts for 20% of the sample dataset and is used for model pre-training and validation, training set A2 accounts for 70% of the sample dataset and is used for further model training and validation, and the remaining 10% is used as the test set for the final test of the model.

[0011] Step two: Following the sample preprocessing in step one, a labeled training set A1 is obtained. This labeled training set A1 is then input into the SVM prediction model for pre-training and validation, yielding the initial training configuration parameters, i.e., the initial SVM prediction model. Configuration parameters refer to various hyperparameters that affect the performance and behavior of the SVM model; these parameters include:

[0012] Kernel function: The algorithm uses kernel function to map data from the original space to a higher-dimensional feature space, thereby finding a linearly separable hyperplane in this higher-dimensional space;

[0013] The C-value, a hyperparameter of the error rate, controls the maximum error rate the model can tolerate. A larger C-value means the model is less tolerant of error rates, but may lead to overfitting; a smaller C-value means the model is more tolerant of error rates, but may lead to underfitting.

[0014] The kernel function parameters affect the model's performance and generalization ability, and need to be determined based on the specific data and task.

[0015] Regularization parameters control model complexity and the accuracy of fitting data. A larger regularization parameter value results in a simpler model and lower accuracy in fitting data. Conversely, a smaller regularization parameter value results in a more complex model and higher accuracy in fitting data.

[0016] Step 3: Using the initial training configuration parameters of the model obtained in Step 2, the model's optimal performance is obtained. Then, using a network-based transfer learning method, the SVM model is trained and validated using the training set A2. Based on the continuous learning of the characteristics of ion implantation process parameters and process results by the SVM model, the final training configuration parameter file is generated, which is the trained SVM model.

[0017] Step four: Use the final training configuration parameter file generated in step three to perform final model testing on the test set, and measure the results using RMSE (root mean square error) and R-squared. 2 The coefficient of determination (RMSE) is used as an evaluation metric. RMSE is the square root of the average of the sum of squares of the differences between predicted and actual values. The smaller the RMSE value, the smaller the difference between the prediction model and the actual data, and the more accurate the prediction model. 2 R is the square of the correlation coefficient between the actual and predicted values. 2 The closer the value is to 1, the better the model's performance.

[0018] Technical features and performance advantages of this invention:

[0019] This invention utilizes machine learning to optimize the ion implantation process for silicon carbide power devices, achieving precise control and prediction of key ion implantation processes. It also allows for continuous updating and retraining of the model as new ion implantation data accumulates, adapting to process changes and optimization needs. Through continuous experimentation and data analysis, the model and process parameters are further optimized. Compared to traditional ion implantation methods for silicon carbide power devices, this invention improves the speed and accuracy of process optimization while reducing costs and scrap rates. It can efficiently adapt to various dynamic changes in ion implantation process conditions and corresponding process results. Attached Figure Description

[0020] Figure 1 This is a flowchart of a machine learning-based ion implantation optimization method for silicon carbide power devices. Detailed Implementation

[0021] The ion implantation optimization method for silicon carbide power devices based on machine learning provided by this invention uses support vector machine (SVM) as the predictive model for machine learning optimization. The specific steps are as follows:

[0022] Step 1, Sample Collection and Preprocessing: Collect historical silicon carbide power device ion implantation process data, preprocess it, and create an ion implantation process optimization sample dataset. The sample dataset includes an implantation process dataset and a label set.

[0023] (1) Production of the implantation process dataset: The information in the implantation process dataset includes the implanted ion type, implantation energy and dose, implantation angle and implantation temperature, which serve as the input source for the support vector machine prediction model. The data is normalized and scaled to between [0,1] to prevent certain features (especially those with a large numerical range) from dominating the model training, while also ensuring that all features have the same dimensions and proportions.

[0024] The formula for data normalization is:

[0025]

[0026] Where x′ is the normalized value, x is the original value, min(x) is the minimum value of the original value, and man(x) is the maximum value of the original value.

[0027] (2) Tag set creation: The tag set is the process result corresponding to the input injection process parameters, including resistivity, activation rate, channel depth, electron mobility and sheet resistance.

[0028] (3) Sample set division: The ion implantation process optimization sample dataset is divided into training set A1, training set A2 and test set; training set A1 accounts for 20% of the sample dataset and is used for model training and validation, training set A2 accounts for 70% of the sample dataset and is used for model training and validation, and the remaining 10% is used as the test set for the final test of the model.

[0029] Step two: Following the sample preprocessing in step one, a labeled training set A1 is obtained. This labeled training set A1 is then input into the SVM prediction model for pre-training and validation, yielding the initial training configuration parameters. These configuration parameters refer to various hyperparameters that affect the performance and behavior of the SVM model; these parameters include:

[0030] Kernel function: The algorithm uses kernel function to map data from the original space to a higher-dimensional feature space, thereby finding a linearly separable hyperplane in this higher-dimensional space;

[0031] The kernel function can be a linear kernel, a polynomial kernel, a radial basis function (RBF) kernel, or a sigmoid kernel, etc.

[0032] The C-value, a hyperparameter of the error rate, controls the maximum error rate the model can tolerate. A larger C-value means the model is less tolerant of error rates, but may lead to overfitting; a smaller C-value means the model is more tolerant of error rates, but may lead to underfitting.

[0033] The kernel function parameters affect the model's performance and generalization ability, and need to be determined based on the specific data and task.

[0034] Regularization parameters control model complexity and the accuracy of fitting data. A larger regularization parameter value results in a simpler model and lower accuracy in fitting data. Conversely, a smaller regularization parameter value results in a more complex model and higher accuracy in fitting data.

[0035] Step 3: Using the initial training configuration parameters of the model obtained in Step 2, the model's optimal performance is obtained. Then, the training set A2 is trained and validated using a network-based transfer learning method. Based on the continuous learning of the characteristics of ion implantation process parameters and process results by the SVM model, the final training configuration parameter file is generated.

[0036] Step four: Use the final training configuration parameter file generated in step three to perform final model testing on the test set, and measure the results using RMSE (root mean square error) and R-squared. 2 (Determination coefficient) is used as an evaluation index.

[0037] RMSE is the square root of the average of the sum of squares of the differences between predicted and actual values. The smaller the RMSE value, the smaller the difference between the prediction model and the actual data, and the more accurate the prediction model. 2 R is the square of the correlation coefficient between the actual and predicted values. 2 The closer the value is to 1, the better the model's performance. RMSE and R 2 The formula for calculation is:

[0038]

[0039] Among them, Q obs,i Q refers to the measured value. mod,i This refers to simulated values. This represents the average of the measured values. This represents the average value of the simulated values, where n is the number of samples.

Claims

1. A machine learning-based ion implantation optimization method for silicon carbide power devices, characterized in that, The Support Vector Machine (SVM) is used as the prediction model for machine learning optimization. The specific steps are as follows: Step 1, Sample Collection and Preprocessing: Collect historical silicon carbide power device ion implantation process data, preprocess it, and create an ion implantation process optimization sample dataset. The sample dataset includes an implantation process dataset and a label set. (1) Production of implantation process dataset: The information in the implantation process dataset includes implanted ion type, implantation energy and dose, implantation angle and implantation temperature, which serve as the input source for the support vector machine prediction model; The data is scaled to between [0,1] through normalization to prevent certain features from dominating model training, while ensuring that all features have the same dimensions and scale. (2) Tag set creation: The tag set is the process result corresponding to the input injection process parameters, including resistivity, activation rate, channel depth, electron mobility and sheet resistance; (3) Sample set division: The ion implantation process optimization sample dataset is divided into training set A1, training set A2 and test set; training set A1 accounts for 20% of the sample dataset and is used for model pre-training and validation, training set A2 accounts for 70% of the sample dataset and is used for further model training and validation, and the remaining 10% is used as the test set for the final test of the model. Step two involves inputting the training set A1 into the SVM prediction model for pre-training and validation, and obtaining the initial training configuration parameters. These configuration parameters refer to various hyperparameters that affect the performance and behavior of the SVM model; the parameters include: Kernel functions map data from the original space to a higher-dimensional feature space, thereby finding a linearly separable hyperplane in this higher-dimensional space. The C-value, a hyperparameter of the error rate, controls the maximum error rate allowed by the model; The kernel function parameters affect the model's performance and generalization ability, and need to be determined based on the specific data and task. Regularization parameters are used to control model complexity and the accuracy of fitting data. The larger the value of the regularization parameter, the simpler the model and the lower the accuracy of fitting data; the smaller the value of the regularization parameter, the more complex the model and the higher the accuracy of fitting data. Step 3: Using the initial training configuration parameters of the model obtained in Step 2, the model's optimal performance is obtained. Then, using a network-based transfer learning method, the SVM model is trained and validated using the training set A2. The SVM model continuously learns the characteristics of ion implantation process parameters and process results, generating the final training configuration parameter file, which is the trained SVM model. Step four: Use the final training configuration parameter file generated in step three to perform final testing of the model on the test set, and use the root mean square error (RMSE) and coefficient of determination (R²) as the test parameters. 2 As an evaluation indicator.

2. The ion implantation optimization method for silicon carbide power devices based on machine learning according to claim 1, characterized in that, The data is normalized, and the calculation formula is: Where x′ is the normalized value, x is the original value, min(x) is the minimum value of the original value, and man(x) is the maximum value of the original value.

3. The ion implantation optimization method for silicon carbide power devices based on machine learning according to claim 1, characterized in that, The kernel function is a linear kernel, a polynomial kernel, a radial basis function (RBF) kernel, or a sigmoid kernel.

4. The ion implantation optimization method for silicon carbide power devices based on machine learning according to claim 1, characterized in that, The root mean square error (RMSE) is the square root of the average of the sum of squares of the differences between the predicted and actual values. A smaller RMSE value indicates a smaller discrepancy between the prediction model and the actual data, and a more accurate prediction model. The coefficient of determination (R²) 2 R is the square of the correlation coefficient between the actual and predicted values. 2 The closer the value is to 1, the better the model's performance; its calculation formula is: Among them, Q obs,i Q refers to the measured value. mod,i This refers to simulated values. This represents the average of the measured values. This represents the average value of the simulated values, where n is the number of samples.

Images