Self - alignment process anomaly prediction method for silicon carbide high - voltage MOSFETs

Abstract

This application relates to the field of semiconductor technology, and provides a method for predicting self-alignment process anomalies for high-voltage silicon carbide MOSFETs. The method includes: obtaining gate oxide layer deposition parameters, generating a simulated gate oxide layer, performing a first anomaly prediction to obtain a first anomaly probability; obtaining gate metal deposition parameters, and performing a second anomaly prediction in combination with the simulated gate oxide layer to obtain a second anomaly probability; obtaining source-drain doping parameters, configuring a doping ratio according to the second anomaly probability, and predicting a doping anomaly probability; calculating a predicted anomaly probability based on the first, second, and doping anomaly probabilities as the self-alignment process prediction result. This application solves the technical problem that it is difficult to predict and control anomalies in high-voltage silicon carbide MOSFETs during the self-alignment process due to the complexity of multiple process steps, and realizes the effect of optimizing manufacturing quality and device performance by means of a multi-step combined anomaly prediction method, improving the accuracy and prediction ability of process anomaly identification.

Description

Technical Field

[0001] This application relates to the field of semiconductor technology, and particularly to a method for predicting self-alignment process anomalies for silicon carbide high-voltage MOSFETs. Background Art

[0002] Silicon carbide (SiC), as a wide-bandgap semiconductor material, is increasingly widely used in high-voltage, high-power, and high-frequency devices. Especially in high-voltage MOSFET devices, silicon carbide has become an important material for realizing high-performance and miniaturized power electronic systems due to its excellent high-temperature resistance, high breakdown field strength, and low on-resistance. However, during the manufacturing process of SiC high-voltage MOSFETs, due to their unique material properties and precise manufacturing requirements, there are many technical difficulties, especially in key process steps such as the formation of the gate oxide layer, the deposition of the gate metal, and the doping of the source and drain regions, where anomalies are likely to occur. These anomalies may lead to a decline in device performance and even affect its reliability.

[0003] As a high-precision manufacturing technology, the self-alignment process can accurately position the source and drain by using the already formed gate structure, effectively reducing the alignment error and improving the device integration and performance. However, the self-alignment process involves multiple complex manufacturing steps, and the parameter selection and control of each step directly affect the quality of the final device. For example, the thickness uniformity of the gate oxide layer, the integrity of the gate metal deposition, and the precise positioning of the doping region are all crucial for the performance of SiC high-voltage MOSFETs. Due to the interdependence between steps, any anomaly in a single link may trigger a chain reaction, resulting in the disqualification of the entire batch of devices.

[0004] Currently, the methods for predicting and controlling anomalies in the self-alignment process of SiC high-voltage MOSFETs are still relatively limited. Traditional quality control mainly relies on post-process inspection, which has the problems of lagging anomaly discovery and inability to prevent in advance. In addition, due to the high manufacturing costs of SiC materials and high-voltage MOSFET devices, the scrapping caused by anomalies not only causes economic losses but also affects production efficiency. Therefore, how to predict possible anomalies in real time during the process and provide a quantitative reference for the anomaly probability has become a key technical issue for improving production quality and process stability. Summary of the Invention

[0005] This application provides a method for predicting self-alignment process anomalies for silicon carbide high-voltage MOSFETs, aiming to solve the technical problem that it is difficult to predict and control anomalies in the self-alignment process of silicon carbide high-voltage MOSFETs due to the complexity of multiple process steps.

[0006] In view of the above problems, this application provides a method for predicting self-alignment process anomalies for silicon carbide high-voltage MOSFETs.

[0007] The present application provides a method for predicting self - alignment process anomalies for silicon carbide high - voltage MOSFETs. The method includes: during the manufacturing of silicon carbide high - voltage MOSFETs using the self - alignment process, obtaining the deposition parameters of the gate oxide layer, simulating the generation of the gate oxide layer to obtain a simulated gate oxide layer, performing a primary anomaly prediction to obtain a primary anomaly probability; obtaining the deposition parameters of the gate metal layer, combining with the simulated gate oxide layer, performing a secondary anomaly prediction to obtain a secondary anomaly probability; obtaining the doping parameters for the self - alignment doping of the source and drain, configuring a doping anomaly prediction ratio according to the secondary anomaly probability, and predicting the integrated doping anomaly probability according to the doping parameters to obtain a doping anomaly probability; calculating a predicted anomaly probability based on the primary anomaly probability, secondary anomaly probability, and doping anomaly probability as the result of the self - alignment process anomaly prediction.

[0008] One or more technical solutions provided in the present application have at least the following technical effects or advantages:

[0009] For the above - mentioned method for predicting self - alignment process anomalies for silicon carbide high - voltage MOSFETs, the method obtains the deposition parameters of the gate oxide layer, generates a simulated gate oxide layer through these parameters, and performs a primary anomaly prediction on the process with the generated gate oxide layer to obtain the probability of possible problems in the gate oxide layer. This step can effectively identify the possible impacts caused by uneven oxide layer thickness or deposition condition fluctuations, laying a stable foundation for the subsequent process; Subsequently, by obtaining the deposition parameters of the gate metal layer and combining with the gate oxide layer generated in the previous step, a secondary anomaly prediction is performed to accurately analyze the quality problems of the gate metal deposition, identify potential problems in the metal layer thickness and coverage integrity, thereby reducing the electrical performance fluctuations caused by gate metal deposition defects and improving the process consistency; Then, for the self - alignment doping link of the source and drain, the corresponding doping parameters are obtained, the doping anomaly prediction ratio is dynamically configured according to the anomaly prediction results of the previous two steps, and using these configured ratios, the integrated doping anomaly probability prediction is further combined with the doping parameters to help evaluate in real - time whether the doping area is accurately positioned; Finally, by comprehensively considering the primary anomaly probability, secondary anomaly probability, and doping anomaly probability, the overall process anomaly prediction result is calculated. This multi - step joint prediction method can not only deeply identify potential problems in each link but also dynamically adjust during the entire process flow, improving the process reliability and device consistency, thereby optimizing production quality and efficiency.

[0010] The above description is only an overview of the technical solution of the present application. In order to be able to understand the technical means of the present application more clearly, it can be implemented according to the content of the specification. And in order to make the above and other purposes, features, and advantages of the present application more obvious and understandable, the following specifically exemplifies the specific implementation manners of the present application. BRIEF DESCRIPTION OF THE DRAWINGS

[0011] To more clearly illustrate the technical solutions in the embodiments of the present invention, the following will briefly introduce the accompanying drawings required for the description of the embodiments. Obviously, the accompanying drawings in the following description are only some embodiments of the present invention. For those of ordinary skill in the art, without creative efforts, other accompanying drawings can be obtained based on these drawings.

[0012] Figure 1 It is a schematic flowchart of a self - alignment process anomaly prediction method for a silicon carbide high - voltage MOSFET in an embodiment.

[0013] Figure 2 It is a schematic flowchart of obtaining a simulated gate oxide layer for a self - alignment process anomaly prediction method for a silicon carbide high - voltage MOSFET in an embodiment. Detailed implementation manners

[0014] By providing a self - alignment process anomaly prediction method for a silicon carbide high - voltage MOSFET in the embodiments of the present application, the technical problem that it is difficult to predict and control anomalies due to the complexity of multiple process steps in the self - alignment process of silicon carbide high - voltage MOSFETs is solved.

[0015] The following will clearly and completely describe the technical solutions in the embodiments of the present application with reference to the accompanying drawings in the embodiments of the present application. Obviously, the described embodiments are only a part of the embodiments of the present application, rather than all the embodiments. Based on the embodiments in the present application, all other embodiments obtained by those of ordinary skill in the art without creative efforts belong to the scope of protection of the present application.

[0016] It should be noted that the terms "including" and "having" and any variations thereof are intended to cover non - exclusive inclusion. For example, a process, method, system, product, or server that includes a series of steps or units does not necessarily have to be limited to those clearly listed steps or units, but may include other steps or modules that are not clearly listed or are inherent to these processes, methods, products, or devices.

[0017] Embodiment, as Figure 1 As shown, the present application provides a self - alignment process anomaly prediction method for a silicon carbide high - voltage MOSFET, and the method includes:

[0018] During the process of manufacturing a silicon carbide high - voltage MOSFET using a self - alignment process, obtain the deposition parameters of the gate oxide layer, perform simulation generation of the gate oxide layer, obtain a simulated gate oxide layer, perform a primary anomaly prediction, and obtain a primary anomaly probability.

[0019] In the embodiment of the present application, in the self-alignment process manufacturing of a silicon carbide high-voltage MOSFET, first, by obtaining the deposition parameters of the gate oxide layer, a generative adversarial network is used to simulate and generate the thickness distribution of the oxide layer, forming a simulated gate oxide layer containing the thickness information of multiple gate regions. This simulated gate oxide layer not only restores the thickness distribution characteristics after actual deposition but also can be used to detect and evaluate the uniformity of the oxide layer thickness in subsequent steps. The detection of uniformity is because if the thickness distribution of the oxide layer is uneven, it may have a serious impact on subsequent process steps. When the oxide layer is thinner in some regions, the subsequent metal layer deposition may not completely cover, and even the doping regions of the source and drain may invade into the gate, resulting in electrical isolation failure. This situation will directly affect the performance of the MOSFET, including increased leakage current and deteriorated switching characteristics. Therefore, through the simulation of the generative adversarial network and the thickness distribution analysis, an anomaly prediction can be carried out at an early stage. By combining a first-order variant classifier, possible oxide layer non-uniformity problems can be accurately identified, and a first-order anomaly probability is generated. The completion of this step not only helps to intervene in potential manufacturing defects in advance but also provides reliable data references for subsequent gate metal layer deposition and doping processes, ensuring the stability and consistency of device performance from the source.

[0020] Further, as Figure 2 shown, the present application provides obtaining the deposition parameters of the gate oxide layer, performing simulation generation of the gate oxide layer, and obtaining a simulated gate oxide layer, including:

[0021] Obtaining the deposition parameters of the gate oxide layer as the deposition parameters of the gate oxide layer; training an oxide layer simulation channel based on a generative adversarial network; inputting the deposition parameters of the gate oxide layer into the oxide layer simulation channel to simulate and generate a simulated gate oxide layer.

[0022] Preferably, relevant parameters for depositing the gate oxide layer are collected from the actual production process as the gate oxide layer deposition parameters. These parameters include, but are not limited to, oxidation temperature, time, atmosphere composition (such as oxygen concentration), substrate surface cleanliness, and other influencing factors of the deposition environment. The accurate acquisition of these parameters is the basis for subsequent simulation generation. Subsequently, the collected gate oxide layer deposition parameters are input into a pre-constructed oxide layer simulation channel for simulating the generation of the gate oxide layer. This oxide layer simulation channel is constructed based on a generative adversarial network and includes a generator and a discriminator inside. The generator generates simulation data close to the real thickness distribution by inputting the gate oxide layer deposition parameters, and the discriminator judges the difference between the thickness distribution output by the generator and the real thickness distribution to form a supervisory feedback on the generator. By inputting the gate oxide layer deposition parameters obtained in the actual process into the trained oxide layer simulation channel, the generator outputs the thickness distribution of the simulated gate oxide layer covering the entire gate area. This distribution provides the thickness information of the oxide layer in each area, provides accurate basic data for subsequent process optimization and anomaly prediction, and reduces the time cost and material waste of process experiments.

[0023] Furthermore, the present application provides a method for training an oxide layer simulation channel based on a generative adversarial network, including:

[0024] Based on the gate oxide layer deposition data within a historical time, a set of sample gate oxide layer deposition parameters is collected, and the thickness distributions of the gate oxide layers deposited under different sample gate oxide layer deposition parameters are collected and labeled as a set of sample gate oxide layers. Based on the generative adversarial network, an oxide layer simulation channel is constructed, wherein the oxide layer simulation channel includes a generator and a discriminator inside. The set of sample gate oxide layer deposition parameters and the set of sample gate oxide layers are used to alternately supervise and train the generator and the discriminator until convergence, and a trained oxide layer simulation channel is obtained.

[0025] Optionally, a set of parameters for gate oxide layer deposition is extracted from the production data within the historical time as the sample gate oxide layer deposition parameter set, including oxidation temperature, time, atmosphere composition, etc. These parameters should cover different process conditions to ensure the diversity and comprehensiveness of the sample data. Under each set of deposition parameter conditions, the corresponding oxide layer thickness distribution data is collected, specifically including the thickness information of each gate region, and these thickness data are labeled as the sample gate oxide layer set. Subsequently, based on the generative adversarial network (GAN), an oxide layer simulation channel is constructed, which includes a generator and a discriminator. The design of the generator aims to generate an oxide layer thickness distribution close to the real one according to the input deposition parameters, while the discriminator is responsible for judging whether the input data is a real thickness distribution or a generated thickness distribution. During the training process, the parameters of the generator and the discriminator are randomly initialized, and relevant hyperparameters are set, such as the learning rate, batch size, etc. In each round of training, first, the sample gate oxide layer deposition parameter set is input into the generator to generate an initial simulated thickness distribution, and then the generated simulated thickness distribution and the real sample thickness distribution are input into the discriminator together to calculate the probability output of the discriminator for the authenticity of the data. The discriminator optimizes its parameters through the cross-entropy loss function so that it can accurately distinguish real data from generated data. At the same time, the parameters of the generator are updated through backpropagation, making the output of the generator gradually approach the real data distribution. The training process is carried out through the alternating optimization iteration of the generator and the discriminator. The generator constantly tries to deceive the discriminator, while the discriminator constantly improves its ability to distinguish real and generated data. As the training progresses, the output quality of the generator gradually improves, and the difference between the finally generated thickness distribution and the real thickness distribution gradually shrinks. The convergence conditions of the training include that the distribution difference between the output data of the generator and the real data reaches a preset threshold, or the discrimination accuracy of the discriminator on real data and generated data is close to random guessing (about 50%). When the convergence conditions are met, the trained oxide layer simulation channel can be used to input any gate oxide layer deposition parameters to generate a high-quality simulated gate oxide layer thickness distribution. This simulation channel provides a basis for subsequent process optimization and anomaly prediction and realizes the high digitization of the deposition process through deep learning technology.

[0026] Furthermore, this application provides a method for performing an anomaly prediction to obtain an anomaly probability, including:

[0027] According to the thickness distribution within the simulated gate oxide layer, the variance of the gate oxide layer thickness is calculated; based on the gate oxide layer deposition data within historical time, a set of sample gate oxide layer thickness variances is collected, and according to the proportion of abnormal occurrences of silicon carbide high-voltage MOSFETs under different sample gate oxide layer thickness variances, a set of sample primary anomaly probabilities is identified; using the set of sample gate oxide layer thickness variances and the set of sample primary anomaly probabilities, a primary variation classifier is constructed; the gate oxide layer thickness variance is input into the primary variation classifier, and a primary anomaly probability is obtained through classification prediction.

[0028] Optionally, by simulating the thickness distribution within the gate oxide layer and calculating its thickness variance, the thickness distribution data within the gate oxide layer is sourced from the output of the oxide layer simulation channel, which contains thickness information for multiple gate regions. The calculation process uses statistical methods. By extracting the thickness values of all gate regions and calculating the square of their standard deviation, the thickness variance of the gate oxide layer is obtained. This variance value can quantify the uniformity of the oxide layer thickness, thereby reflecting the stability of the deposition process. Subsequently, based on the gate oxide layer deposition data collected over historical time, record the corresponding oxide layer thickness variance and construct a sample set of gate oxide layer thickness variances. These sample data cover various thickness distribution characteristics, including typical process results of uniform and non-uniform. At the same time, according to these thickness variance data, statistically calculate the proportion of abnormal occurrences of the corresponding silicon carbide high-voltage MOSFETs to label a set of sample one-time anomaly probabilities. This set is obtained by analyzing the anomaly occurrence rate under each variance value and can provide a clear mapping relationship for subsequent anomaly prediction. After that, use the sample set of gate oxide layer thickness variances and the sample set of one-time anomaly probabilities to construct a one-time variation classifier. The classifier is constructed using machine learning algorithms, such as support vector machine (SVM), random forest, or neural network, etc. Taking the fully connected neural network (FCNN) in the neural network as an example, use the fully connected neural network to construct the structure of the one-time variation classifier, including an input layer, a hidden layer, an output layer, etc. The input layer is used to receive the gate oxide layer thickness variance data, the output layer generates the predicted one-time anomaly probability value, and the hidden layer captures the complex relationship between the thickness variance and the anomaly probability through multiple layers of non-linear transformations. Initialize the weights and biases of the network using random numbers to make the network have a certain degree of randomness and avoid symmetry problems caused by initial weights. Use the collected set of gate oxide layer thickness variances as input features and the sample set of one-time anomaly probabilities as the target output and input them into the initialized one-time variation classifier. When performing forward propagation, the input data enters the neural network through the input layer and is transformed layer by layer through weighted operations and non-linear activation functions (such as ReLU) in the hidden layer to extract features. At the output layer, the result is converted into a probability distribution through the Softmax function, and the one-time anomaly probability corresponding to the current gate oxide layer thickness variance is output. Then use the binary cross-entropy loss function to calculate the loss between the predicted one-time anomaly probability and the sample one-time anomaly probability. Through the backpropagation algorithm, calculate the gradients of the loss function with respect to the weights and biases of each layer layer by layer, and adjust the network parameters to reduce the loss value. Use the Adam optimizer to optimize the network and adjust the weights and biases of the network using an adaptive learning rate to make the network converge to the global optimum faster;Repeat the above processes of forward propagation, loss calculation, backpropagation, and parameter optimization until the maximum number of iterations is reached or the loss value no longer decreases significantly. After the training is completed, use the data not used for training to test the performance of the current mutated classifier once, and evaluate the accuracy of the mutated classifier in predicting the probability of an anomaly once. If the performance of the classifier on the data not used for training reaches the preset accuracy target, output the currently trained mutated classifier once. If the performance does not meet the expectation, adjust the hyperparameters such as the learning rate, the number of hidden layers, and the number of neurons according to the feedback of the data not used for training, and retrain the model until the requirements are met. Finally, input the actually generated variance value of the gate oxide layer thickness into the trained mutated classifier once. Based on the mapping relationship learned from the training data, the classifier predicts the anomaly probability under the current thickness variance. This anomaly probability once serves as an early indicator of the process, providing a reference basis for parameter adjustment in subsequent processes, and can effectively reduce device performance problems caused by uneven oxide layer thickness.;

[0029] Obtain the deposition parameters of the gate metal layer deposition, and combine with the simulated gate oxide layer to perform secondary anomaly prediction to obtain the secondary anomaly probability.

[0030] In one embodiment, when performing secondary anomaly prediction of a silicon carbide high-voltage MOSFET, first obtain the deposition parameters of the gate metal layer, including the thickness of the metal deposition material, deposition rate, temperature, vacuum environment, etc. These parameters directly affect the uniformity and coverage quality of the gate metal layer. At the same time, use the previously generated simulated gate oxide layer as the basic data and input it into the secondary mutated classifier together with the gate metal layer deposition parameters to analyze potential anomalies during the metal layer deposition process. During the prediction process, the simulated gate oxide layer provides the basic information of the oxide layer thickness distribution, while the gate metal layer deposition parameters reflect the deposition conditions of the metal layer in different regions. If the oxide layer thickness distribution is uneven, it may cause insufficient deposition of the metal layer in the thinner region, resulting in uncovered or incomplete coverage. At the same time, if there are deviations in the deposition rate and uniformity of the metal layer, it may further exacerbate the overall deposition instability. These problems will directly affect the electrical performance of the MOSFET. For example, the leakage current increases or the switching characteristics deteriorate, resulting in performance anomalies. By combining the gate metal layer deposition parameters and the simulated gate oxide layer, the secondary mutated classifier can evaluate the uniformity and integrity of the metal layer deposition, identify potential defects during the deposition process, and output the corresponding secondary anomaly probability. This probability reflects the risk that the device performance may be abnormal under the current deposition conditions, providing a scientific basis for process optimization and parameter adjustment, and helping to prevent the occurrence of MOSFET performance anomalies at an early stage.

[0031] Further, the present application provides obtaining deposition parameters for depositing a gate metal layer, combining the simulated gate oxide layer, performing secondary anomaly prediction, and obtaining a secondary anomaly probability, including:

[0032] Obtain the deposition parameters for depositing the gate metal layer as the gate metal layer deposition parameters; based on the production data of silicon carbide high-voltage MOSFETs within a historical time period, collect a set of sample gate oxide layer thickness variances and a set of sample gate metal layer deposition parameters, and identify a set of sample secondary anomaly probabilities according to the proportions of anomalies in silicon carbide high-voltage MOSFETs under different sample gate oxide layer thickness variances and sample gate metal layer deposition parameters; use the set of sample gate oxide layer thickness variances, the set of sample gate metal layer deposition parameters, and the set of sample secondary anomaly probabilities to construct a secondary variation classifier; input the gate metal layer deposition parameters and the gate oxide layer thickness variance into the secondary variation classifier, and classify and predict to obtain the secondary anomaly probability.

[0033] Optionally, obtaining the deposition parameters of the gate metal layer as the gate metal layer deposition parameters includes key process data such as thickness, deposition rate, temperature, and vacuum environment. These parameters directly affect the deposition uniformity and coverage performance of the metal layer. At the same time, based on the production data of silicon carbide high-voltage MOSFETs within a historical time period, collect sample data related to these parameters, including a set of sample gate oxide layer thickness variances and a set of sample gate metal layer deposition parameters; subsequently, according to the gate oxide layer thickness variances and gate metal layer deposition parameters of different samples, count the proportions of anomalies in silicon carbide high-voltage MOSFETs to label and obtain a set of sample secondary anomaly probabilities. This set is used to construct a secondary variation classifier. The training process of the classifier is similar to that of the aforementioned primary variation classifier, including steps such as model initialization, forward propagation, loss calculation, backward propagation, and parameter optimization. After the training of the primary variation classifier is completed, input the actually obtained gate metal layer deposition parameters and gate oxide layer thickness variance into the trained secondary variation classifier. According to these two sets of input data, the classifier combines the mapping relationship learned during the historical training process, predicts and outputs the secondary anomaly probability. This probability is used to quantify the possible anomaly risk to the performance of the MOSFET under the combined action of metal layer deposition and oxide layer deposition, providing an accurate reference basis for subsequent process adjustment and quality optimization.

[0034] Obtain the doping parameters for self-aligned doping of the source and drain, configure and obtain the doping anomaly prediction ratio according to the secondary anomaly probability, and predict the integrated doping anomaly probability according to the doping parameters to obtain the doping anomaly probability.

[0035] In one embodiment, when predicting anomalies in source-drain self-aligned doping, relevant parameters during the doping process are first obtained, including doping concentration, energy, dose, and scanning mode, etc. These doping parameters directly affect the formation quality of the source and drain. Subsequently, according to the previously calculated secondary anomaly probability, the number of doping anomaly prediction branches for predicting doping anomalies is dynamically configured. During the configuration process, the number of branches is adjusted based on the magnitude of the secondary anomaly probability. When the secondary anomaly probability is high, more doping anomaly prediction branches are allocated to improve the prediction accuracy by comprehensively analyzing potential anomalies during the doping process. When the secondary anomaly probability is low, the number of branches is reduced to lower the computational cost and save computing power. This dynamic branch configuration mechanism not only ensures the prediction accuracy but also optimizes the resource utilization efficiency in low-risk situations. In the integrated prediction stage, each branch separately receives the doping parameters as input and calculates the doping anomaly probability based on its respective analysis algorithm. The prediction results of all branches are integrated through weighted average or other integration methods, and finally, the overall doping anomaly probability is output. In this way, not only can various potential anomalies during the doping process be captured, but the computational intensity can also be dynamically adjusted according to the level of anomaly risk, thereby improving the computational efficiency while ensuring the prediction accuracy and providing accurate data support for subsequent process improvement and quality optimization.

[0036] Further, the present application provides obtaining doping parameters for source-drain self-aligned doping, configuring a doping anomaly prediction ratio according to the secondary anomaly probability, and predicting an integrated doping anomaly probability based on the doping parameters to obtain a doping anomaly probability, including:

[0037] Obtain doping parameters for source-drain self-aligned doping to obtain doping parameters; configure the secondary anomaly probability as the doping anomaly prediction ratio; based on ensemble learning, train a doping anomaly prediction channel of doping anomaly prediction branches including the number of branches; multiply the doping anomaly prediction ratio by the number of branches and round to obtain the number of doping analyses; input the doping parameters into randomly selected doping anomaly prediction branches of the number of doping analyses, predict the predicted doping anomaly probability of the number of doping analyses, and calculate the mean to obtain the doping anomaly probability.

[0038] Preferably, the doping parameters for obtaining self-aligned doping of the source and drain are obtained, including key process parameters such as doping concentration, energy, and dose. These parameters directly affect the accuracy and uniformity of the doped region. Record these parameters and use them as input features for subsequent analysis. Subsequently, according to the quadratic anomaly probability calculated in the previous step, configure it as the doping anomaly prediction ratio. This ratio is used to dynamically adjust the number of doping anomaly prediction branches used in the integrated prediction process, so as to achieve reasonable allocation of resources at different anomaly risk levels. The doping anomaly prediction branches are built into the doping anomaly prediction channel. These doping anomaly prediction branches are trained using historical doping data through the method of integrated training and can independently predict the doping anomaly probability. During the training process, a set of sample doping parameters and the corresponding anomaly probability are collected as input and target, and multiple prediction branches are constructed using random forest, gradient boosting tree, deep learning, neural network, etc., to ensure that each can finally accurately predict the anomaly probability. In the prediction stage, multiply the doping anomaly prediction ratio by the total number of branches in the prediction channel and round up to obtain the actual number of doping analyses participating in the analysis. The dynamic adjustment of the number of branches ensures that more branches are allocated in the case of high anomaly probability to improve the analysis accuracy, and fewer branches are reduced in the case of low anomaly probability to optimize the calculation efficiency. Then, input the actually obtained doping parameters into the prediction branches of the randomly selected number of doping analyses. Each branch independently outputs a predicted doping anomaly probability, and these prediction results are combined through the method of mean calculation to obtain the final doping anomaly probability. This integrated prediction process can make full use of the flexibility and reliability of the multi-branch channel, improve the anomaly detection accuracy while effectively managing the computing resources. The finally obtained doping anomaly probability provides key data support for the quality evaluation and process optimization of the self-aligned doping process, and helps to further improve the production quality and device performance of the silicon carbide high-voltage MOSFET.

[0039] Furthermore, the present application provides a doping anomaly prediction channel based on ensemble learning to train doping anomaly prediction branches including the number of branches, including:

[0040] According to the production data of silicon carbide high-voltage MOSFETs within a historical time period, collect a set of sample doping parameters, and obtain the proportion of silicon carbide high-voltage MOSFETs with anomalies under different sample doping parameters, and mark to obtain a set of sample doping anomaly probabilities; randomly draw multiple copies of doping supervised training data with the number of branches from the set of sample doping parameters and the set of sample doping anomaly probabilities; respectively use the multiple copies of doping supervised training data, based on ensemble learning, train to obtain doping anomaly prediction branches with the number of branches, and integrate to obtain a doping anomaly prediction channel.

[0041] Optionally, based on the production data within a historical time period, collect a set of doping parameters related to silicon carbide high-voltage MOSFETs. These parameters include variables such as doping concentration, energy, and dose, covering different combinations of doping processes during historical production. At the same time, statistically calculate the proportion of MOSFETs showing anomalies under each set of doping parameter conditions. For example, the occurrence frequencies of phenomena such as electrical performance degradation and increased leakage current are counted, and these proportions are labeled as the sample doping anomaly probability set. To construct a multi-branch supervised training dataset, randomly sample from the sample doping parameter set and the sample doping anomaly probability set using the sampling method with replacement. According to the required number of branches, draw multiple copies of data. Each dataset contains doping parameters and the corresponding anomaly probability annotations. In this way, multiple sets of independent supervised training data are generated to ensure that the sample characteristics of each dataset have a certain degree of randomness while maintaining the consistency of the overall data distribution. Subsequently, use each supervised training dataset to train multiple doping anomaly prediction branches based on the ensemble learning method. These branches can be based on common ensemble learning models, such as Random Forest, Gradient Boosting Decision Trees (GBDT), or Deep Neural Network (DNN). Each branch is trained independently. By learning the mapping relationship between doping parameters and anomaly probabilities, a doping anomaly prediction branch capable of predicting doping anomaly probabilities is generated. Exemplarily, when using a deep neural network, the training process can be completed through steps such as forward propagation, loss calculation, backpropagation, and parameter optimization in the same way as constructing a single-variant classifier described above. When using a random forest, input a supervised training dataset into the random forest model. The random forest generates multiple sub-datasets from this supervised training dataset through the random sampling method with replacement. These sub-datasets are used to train different decision trees in the random forest. For the training of each decision tree, randomly select a subset of all input features as candidate splitting features and make splitting decisions at each node. The splitting criterion can be based on information gain, Gini coefficient, or mean squared error (MSE) to maximize the purity of the data or minimize the prediction error. Recursive splitting continues until the stopping condition is met, such as reaching the maximum depth, insufficient number of samples at the node, or insignificant reduction in error caused by splitting. Then, the splitting stops and leaf nodes are generated. Each leaf node records the anomaly probability of this node as the prediction result. After all the decision trees in the random forest are independently trained, model integration is performed. For new input doping parameters, the random forest makes individual predictions through each decision tree to obtain multiple anomaly probability results. By taking the average of the prediction results of all decision trees, the final doping anomaly probability prediction value is output. The random forest utilizes the ensemble of multiple decision trees to improve the robustness and anti-overfitting ability of the model;Finally, all trained doping anomaly prediction branches are integrated in parallel to construct a complete doping anomaly prediction channel, which can efficiently process a variety of doping parameter combinations, provide reliable anomaly probability prediction, and provide strong support for subsequent process optimization and quality control. ;

[0042] According to the primary abnormality probability, the secondary abnormality probability and the doping abnormality probability, a predicted abnormality probability is calculated and obtained as a self-alignment process abnormality prediction result.

[0043] In one embodiment, during the self-alignment process abnormality prediction process, the final predicted abnormality probability is calculated by comprehensively considering the primary abnormality probability, the secondary abnormality probability and the doping abnormality probability. The primary abnormality probability mainly reflects the problems that may exist in the gate oxide deposition stage, such as uneven thickness distribution. The secondary abnormality probability combines the parameters of the gate oxide layer and the gate metal layer deposition to evaluate the abnormalities that may be caused during the metal layer deposition process. The doping abnormality probability is based on the parameters of the source and drain self-aligned doping to predict the accuracy and uniformity of the doping area. By comprehensively calculating these probability values, a predicted abnormality probability can be obtained as a self-alignment process abnormality prediction result. This predicted abnormality probability fully considers the impact of each key step in the process flow on the final device performance, reflects the independence of each link, and captures the linkage effect between them, and provides a quantitative assessment of the abnormality risk of the entire self-alignment process. It can be used as an important basis for production process monitoring and process optimization, and provides strong support for improving the manufacturing quality of silicon carbide high-voltage MOSFETs.

[0044] Further, the present application provides a method for calculating and obtaining a predicted abnormality probability according to the primary abnormality probability, the secondary abnormality probability and the doping abnormality probability as a self-alignment process abnormality prediction result, including:

[0045] The average of the primary abnormality probability and the secondary abnormality probability is calculated to obtain the gate abnormality probability; the gate abnormality probability and the doping abnormality probability are added to obtain the predicted abnormality probability as the self-alignment process abnormality prediction result.

[0046] Preferably, the primary anomaly probability and the secondary anomaly probability are used as inputs, which are respectively derived from the anomaly prediction results in the gate oxide layer deposition stage and the gate metal layer deposition stage. The primary anomaly probability and the secondary anomaly probability are averaged to obtain the gate anomaly probability. The purpose of this step is to comprehensively evaluate the overall anomaly risk of the gate-related processes and reflect the stability of the gate oxide layer and the metal layer deposition links. Subsequently, the gate anomaly probability is added to the doping anomaly probability to calculate the final predicted anomaly probability. The doping anomaly probability is obtained by analyzing the parameters and anomaly risks in the source and drain doping processes. By directly adding these two probability values, a comprehensive predicted anomaly probability is obtained. This predicted anomaly probability value comprehensively reflects the anomaly risk levels of the three key stages (gate oxide layer deposition, gate metal layer deposition, and doping) in the self-alignment process. The entire calculation process integrates the multi-stage prediction results into a unified quantitative indicator, which can provide a clear reference basis for the monitoring and anomaly management of the production process. If the predicted anomaly probability exceeds the preset threshold, it can indicate that there is a high anomaly risk in the production process, and further process optimization or parameter adjustment is required to ensure that the final product quality meets the expected standards.

[0047] In summary, the embodiments of the present application at least have the following technical effects:

[0048] The embodiments of the present application cover the complete process flow from the gate oxide layer deposition, the gate metal layer deposition to the self-alignment doping of the source and drain. By obtaining the deposition parameters of the gate oxide layer, using the generative adversarial network to simulate and generate the oxide layer thickness distribution for primary anomaly prediction, combining the gate metal layer deposition parameters and the simulated gate oxide layer data for secondary anomaly prediction, then configuring the doping anomaly prediction ratio according to the doping parameters and the secondary anomaly probability, and predicting the doping anomaly probability through the ensemble learning model. After completing the anomaly probability prediction for each stage, the primary anomaly probability, the secondary anomaly probability, and the doping anomaly probability are comprehensively calculated to generate the final self-alignment process anomaly prediction result; these technical effects jointly solve the technical problem that it is difficult to predict and control anomalies due to the complexity of multiple process steps in the self-alignment process of silicon carbide high-voltage MOSFETs, and achieve the effect of improving the accuracy and prediction ability of process anomaly recognition through a multi-step joint anomaly prediction method, thereby optimizing the manufacturing quality and device performance.

[0049] It should be noted that the above sequence of the embodiments of the present application is only for description and does not represent the superiority or inferiority of the embodiments. And the above description of specific embodiments of this specification has been made. The processes depicted in the drawings do not necessarily require the specific order and continuous order shown to achieve the desired results. In some embodiments, multitasking and parallel processing are also possible or may be advantageous.

[0050] The foregoing are only the preferred embodiments of the present application and are not intended to limit the present application. Any modifications, equivalent substitutions, improvements, etc. made within the spirit and principle of the present application shall be included within the protection scope of the present application.

[0051] This specification and the drawings are merely exemplary descriptions of the present application and are considered to have covered any and all modifications, variations, combinations, or equivalents within the scope of the present application. Obviously, those skilled in the art can make various changes and modifications to the present application without departing from the scope of the present application. Thus, if these modifications and variations of the present application fall within the scope of the present application and its equivalent technologies, the present application is intended to include these changes and modifications therein.

Claims

1. A self-aligned process anomaly prediction method for silicon carbide high-voltage MOSFET, characterized in that: The method comprises: In the process of manufacturing silicon carbide high-voltage MOSFET by using a self-aligned process, the deposition parameters of the gate oxide layer are obtained, the gate oxide layer is simulated and generated, the simulated gate oxide layer is obtained, an abnormality prediction is performed, and an abnormality probability is obtained; Obtaining deposition parameters of the gate metal layer deposition, combining the simulated gate oxide layer, performing secondary anomaly prediction, and obtaining secondary anomaly probability; Acquire doping parameters for self-aligned doping of the source and drain, configure and obtain a doping anomaly prediction ratio according to the secondary anomaly probability, perform integrated doping anomaly probability prediction according to the doping parameters, and obtain a doping anomaly probability; According to the primary abnormality probability, the secondary abnormality probability and the doping abnormality probability, a predicted abnormality probability is calculated as a self-alignment process abnormality prediction result; Among them, making an abnormal prediction and obtaining an abnormal probability include: According to the simulated gate oxide layer thickness distribution, a gate oxide layer thickness variance is calculated; Based on the gate oxide deposition data in the historical time, the sample gate oxide thickness variance set is collected, and according to the abnormal proportion of silicon carbide high-voltage MOSFET under different sample gate oxide thickness variances, the sample one-time abnormal probability set is marked; Using the sample gate oxide thickness variance set and the sample primary anomaly probability set, constructing a primary variation classifier; Inputting the gate oxide layer thickness variance into the primary variation classifier, classifying and predicting to obtain a primary abnormality probability; Wherein, obtaining doping parameters for self-aligned doping of the source and drain, configuring to obtain a doping anomaly prediction ratio according to the secondary anomaly probability, performing integrated doping anomaly probability prediction according to the doping parameters, and obtaining a doping anomaly probability, includes: Obtaining doping parameters of the source and drain for self-aligned doping to obtain doping parameters; configuring the secondary anomaly probability as a doping anomaly prediction ratio; Based on ensemble learning, a doping anomaly prediction channel including doping anomaly prediction branches of a number of branches is trained; The doping anomaly prediction ratio is multiplied by the number of branches and the result is rounded to obtain the doping analysis number; The doping parameters are respectively input into the doping anomaly prediction branches of the randomly selected doping analysis quantity, the predicted doping anomaly probability of the doping analysis quantity is predicted, and the mean is calculated to obtain the doping anomaly probability.

2. The self-aligned process anomaly prediction method for silicon carbide high-voltage MOSFET according to claim 1, characterized in that: Obtaining deposition parameters of the gate oxide layer, performing simulation generation of the gate oxide layer, and obtaining a simulated gate oxide layer, including: Obtaining a deposition parameter of a gate oxide layer as a gate oxide layer deposition parameter; Based on the generative adversarial network, the oxide layer simulation channel is trained; The gate oxide layer deposition parameters are input into the oxide layer simulation channel, and a simulated gate oxide layer is generated by simulation.

3. The self-aligned process anomaly prediction method for silicon carbide high-voltage MOSFET according to claim 2, characterized in that: Based on the generative adversarial network, the oxide layer simulation channel is trained, including: Based on the gate oxide deposition data in the historical time, a sample gate oxide deposition parameter set is collected, and the thickness distribution of the gate oxide layer deposited under different sample gate oxide deposition parameters is collected and marked as a sample gate oxide set; Based on a generative adversarial network, an oxide layer simulation channel is constructed, wherein the oxide layer simulation channel includes a generator and an adversary; The sample gate oxide layer deposition parameter set and the sample gate oxide layer set are used to perform alternate supervised training on the generator and the adversary until convergence, thereby obtaining a trained oxide layer simulation channel.

4. The self-aligned process anomaly prediction method for silicon carbide high-voltage MOSFET according to claim 1, characterized in that: Obtaining deposition parameters of the gate metal layer deposition, combining the simulated gate oxide layer, performing secondary anomaly prediction, and obtaining secondary anomaly probability, including: Obtaining deposition parameters of the gate metal layer as gate metal layer deposition parameters; Based on the historical production data of silicon carbide high-voltage MOSFET, a set of sample gate oxide thickness variances and a set of sample gate metal layer deposition parameters are collected, and according to the proportion of abnormalities of silicon carbide high-voltage MOSFET under different sample gate oxide thickness variances and sample gate metal layer deposition parameters, a set of sample secondary abnormality probabilities is obtained by identification; A secondary variation classifier is constructed using the sample gate oxide layer thickness variance set, the sample gate metal layer deposition parameter set and the sample secondary anomaly probability set; The gate metal layer deposition parameters and the gate oxide layer thickness variance are input into the secondary variation classifier, and the secondary anomaly probability is obtained through classification prediction.

5. The self-aligned process anomaly prediction method for silicon carbide high-voltage MOSFET according to claim 1, characterized in that: Based on ensemble learning, the doping anomaly prediction channel of the doping anomaly prediction branch including the number of branches is trained, including: According to the production data of silicon carbide high-voltage MOSFET in the historical period, a set of sample doping parameters is collected, and the proportion of abnormal silicon carbide high-voltage MOSFET under different sample doping parameters is obtained, and a set of sample doping abnormality probability is obtained by identification; There are multiple copies of doping supervision training data of randomly selected branches with replacement in the sample doping parameter set and the sample doping abnormal probability set; The plurality of doping supervision training data are respectively used to train a number of doping anomaly prediction branches based on ensemble learning, and the doping anomaly prediction channels are obtained through integration.

6. The self-aligned process anomaly prediction method for silicon carbide high-voltage MOSFET according to claim 1, characterized in that: According to the primary abnormality probability, the secondary abnormality probability and the doping abnormality probability, a predicted abnormality probability is calculated as a self-alignment process abnormality prediction result, including: Calculate the average of the primary abnormality probability and the secondary abnormality probability to obtain the gate abnormality probability; The gate abnormality probability and the doping abnormality probability are added to obtain a predicted abnormality probability as a self-alignment process abnormality prediction result.

Images