Silicon carbide epitaxial process formula generation method, device, equipment and storage medium
By using standardized processing and least squares regression algorithm to screen silicon carbide epitaxial process parameters, the problems of low efficiency and result deviation in traditional methods are solved, achieving efficient and reliable process parameter optimization, adapting to complex process scenarios, and reducing R&D costs.
Patent Information
- Application Number
- CN202511351035.4
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2025-09-22
- Publication Date
- 2025-10-28
- Estimated Expiration
- 2045-09-22
AI Technical Summary
Traditional methods for optimizing silicon carbide epitaxial process parameters are inefficient, costly, and fail to capture the complex coupling relationships between multiple parameters. Existing data-driven methods do not standardize input and output data, leading to biased results, and lack targeted screening mechanisms for process data characteristics.
By acquiring historical experimental process data, standardizing independent and dependent variables, using the principle of maximizing variance to screen core dependent variables, generating weight vectors and correlation coefficients of the second component, and combining the least squares regression algorithm to screen the optimal solution, a silicon carbide epitaxial process formula that meets the target indicators is generated.
It improves the efficiency of process parameter optimization, avoids the blindness of traditional trial and error, ensures the objectivity and reliability of analysis results, adapts to complex process scenarios, generates candidate formulations that meet target indicators, shortens the R&D cycle and reduces experimental costs.
Smart Images

Figure CN120851393A_ABST
Abstract
Description
Technical Field
[0001] This invention relates to the field of silicon carbide technology, specifically to a method, apparatus, equipment, and storage medium for generating silicon carbide epitaxial process formulations. Background Technology
[0002] In industrial production and experimental research, precise control of process parameters is crucial for product quality, production efficiency, and the reliability of experimental results. Traditional process parameter optimization methods often rely on trial and error based on human experience or single-variable experiments, resulting in low efficiency, high costs, and difficulty in capturing the complex coupling relationships between multiple parameters. While existing data-driven optimization methods incorporate statistical analysis or machine learning techniques, they often fail to standardize input and output data, leading to an imbalance in the weighting of parameters of different magnitudes on the results. Furthermore, in the feature extraction stage, the lack of targeted screening mechanisms based on the characteristics of process data makes them susceptible to interference from redundant information, hindering the accurate extraction of key influencing factors. In addition, the candidate solution screening stage often uses a single evaluation criterion, ignoring the multi-objective requirements of process indicators, resulting in deviations between optimization results and actual production goals, and failing to efficiently support precise decision-making regarding process parameters. Summary of the Invention In order to overcome the shortcomings of the prior art, the purpose of this invention is to provide a method, apparatus, equipment and storage medium for generating silicon carbide epitaxial process formulations.
[0003] To solve the above-mentioned technical problems, the technical solution adopted by the present invention is as follows: This invention provides a method for generating silicon carbide epitaxial process formulations, comprising: acquiring historical experimental process data; generating independent and dependent variables based on the historical experimental process data; standardizing the independent and dependent variables to obtain standardized independent and dependent variables; selecting a first-component correlation coefficient from the standardized dependent variables according to a preset maximum variance column condition; generating a weight vector based on the standardized independent variables and the first-component correlation coefficient; generating a second-component correlation coefficient based on the weight vector and the standardized independent variables; solving the historical experimental process data and preset target process indicators according to preset data selection conditions and the second-component correlation coefficient to obtain a set of candidate feasible solutions; screening the set of candidate feasible solutions according to a preset least squares regression algorithm and a preset expected dependent variable to obtain process parameters; and generating a silicon carbide epitaxial process formulation and evaluation report based on the process parameters.
[0004] Further, the step of solving the historical experimental process data and the preset target process indicators based on preset data selection conditions and the second component correlation coefficient to obtain a set of candidate feasible solutions includes: performing row number analysis on the historical experimental process data to obtain the number of historical data rows; calculating the historical experimental process data based on preset discard conditions, the number of historical data rows, and data selection conditions to obtain the mean squared error; performing data statistics on all mean squared errors to obtain the mean squared error count; calculating the standardized independent variable and the second component correlation coefficient to obtain the first loading; and determining whether the mean squared error count satisfies the requirements. The system sets a preset stopping condition for the inner loop iteration. If the mean squared error (MSE) does not meet the stopping condition, it returns to the previous iteration and calculates the historical experimental process data based on the discard criteria, the number of historical data rows, and the data selection criteria until the MSE meets the stopping condition. If the MSE meets the stopping condition, it analyzes the MSE and the correlation coefficient of the second component based on the preset stopping condition for the outer loop iteration to obtain the optimal number of components. It then acquires all the first loads to obtain the first load set and generates a set of candidate feasible solutions based on the optimal number of components, the first load set, and the target process indicators.
[0005] Further, the step of calculating the mean squared error based on preset discarding conditions, the number of historical data rows, and data selection conditions for historical experimental process data includes: filtering historical experimental process data according to discarding conditions and the number of historical data rows to obtain a training set and a discarded dataset; training a preset partial least squares regression algorithm on the training set to obtain a first partial least squares regression model; obtaining the input independent variable and the discarded dependent variable from the discarded dataset according to the data selection conditions; predicting the input independent variable according to the first partial least squares regression model to obtain the predicted value of the dependent variable; and calculating the mean squared error based on the predicted value of the dependent variable and the discarded dependent variable.
[0006] Further, the step of analyzing the mean squared error and the correlation coefficient of the second component based on the preset outer loop iteration stopping condition to obtain the optimal number of components includes: calculating the average of all mean squared errors to obtain a set of mean squared error averages; performing data statistics on the set of mean squared error averages to obtain the mean squared error average number; calculating the correlation coefficient of the standardized dependent variable and the second component to obtain the second loading; generating a residual matrix based on the second loading, the standardized dependent variable, and the correlation coefficient of the second component; determining whether the mean squared error average number or the residual matrix satisfies the outer loop iteration stopping condition; if the second mean squared error average number or the residual matrix satisfies the outer loop iteration stopping condition, then obtaining the minimum mean squared error average from the set of mean squared error averages; and obtaining the corresponding number of components based on the minimum mean squared error average to obtain the optimal number of components.
[0007] Furthermore, the step of generating a set of candidate feasible solutions based on the optimal number of components, the first load set, and the target process index includes: optimizing the first partial least squares regression model based on the optimal number of components to obtain an optimized least squares regression model; and solving the target process index based on the optimized least squares regression model and the first load set to obtain a set of candidate feasible solutions.
[0008] Further, the step of solving the target process index based on the optimized least squares regression model and the first loading set to obtain a set of candidate feasible solutions includes: obtaining all weight vectors to obtain a weight vector set; generating a first regression coefficient matrix and a generalized inverse matrix based on the weight vector set and the first loading set; determining whether the first regression coefficient matrix is a full-rank matrix; if the first regression coefficient matrix is a full-rank matrix, then solving the generalized inverse matrix, dependent variable, and first regression coefficient matrix according to a preset solution formula to obtain a first exact solution; solving the target process index based on the optimized least squares regression model and the first exact solution to obtain a set of candidate feasible solutions; otherwise, iteratively solving the generalized inverse matrix, dependent variable, and first regression coefficient matrix using a numerical method to obtain a second exact solution; and solving the target process index based on the optimized least squares regression model and the second exact solution to obtain a set of candidate feasible solutions.
[0009] Further, the step of generating the first regression coefficient matrix and the generalized inverse matrix based on the weight vector set and the first load set includes: generating a weight matrix based on the weight vector set; obtaining all the second loads to obtain a second load set; generating a second load matrix based on the second load set; generating a first load matrix based on the first load set; generating a first regression coefficient matrix based on the weight matrix, the first load matrix, and the second load matrix; and transforming the first regression coefficient matrix to obtain the generalized inverse matrix.
[0010] Further, the step of filtering a set of candidate feasible solutions based on a preset least squares regression algorithm and a preset expected dependent variable to obtain process parameters includes: randomly sampling and replacing historical experimental process data according to a preset number of repetitions to obtain multiple bootstrap samples; constructing multiple second partial least squares regression models and multiple second regression coefficient matrices based on the least squares regression algorithm and the multiple bootstrap samples; performing inverse solving on the multiple second regression coefficient matrices based on the preset expected dependent variable and the multiple second partial least squares regression models to obtain multiple estimates; generating confidence intervals based on the multiple estimates; and filtering a set of candidate feasible solutions based on the confidence intervals to obtain process parameters.
[0011] Furthermore, the silicon carbide epitaxial process formulation generation device includes: a process data acquisition module for acquiring historical experimental process data; a variable generation module for generating independent and dependent variables based on the historical experimental process data; a standardization module for standardizing the independent and dependent variables to obtain standardized independent and dependent variables; a selection module for selecting the first component correlation coefficient from the standardized dependent variables according to a preset maximum variance column condition; a weight vector generation module for generating a weight vector based on the standardized independent variables and the first component correlation coefficient; a coefficient generation module for generating a second component correlation coefficient based on the weight vector and the standardized independent variables; a solution module for solving the historical experimental process data and preset target process indicators according to preset data selection conditions and the second component correlation coefficient to obtain a set of candidate feasible solutions; a screening module for screening the set of candidate feasible solutions according to a preset least squares regression algorithm and a preset expected dependent variable to obtain process parameters; and a formulation and report generation module for generating a silicon carbide epitaxial process formulation and evaluation report based on the process parameters.
[0012] Furthermore, the silicon carbide epitaxial process formulation generation apparatus includes: a memory and at least one processor, wherein the memory stores instructions; at least one processor invokes the instructions in the memory to cause the silicon carbide epitaxial process formulation generation apparatus to perform the various steps of the silicon carbide epitaxial process formulation generation method as described in any one of the above descriptions.
[0013] Furthermore, a computer-readable storage medium stores instructions that, when executed by a processor, implement the various steps of the silicon carbide epitaxial process formulation generation method as described above.
[0014] In the technical solution of this invention, the core dependent variable is screened by maximizing variance, and the influence intensity of the independent variable is quantified by combining weight vectors. Key control factors are located from multiple sets of process parameters (such as hydrogen carrier gas flow rate, process temperature, etc.), improving optimization efficiency and avoiding the blindness of traditional trial and error. By eliminating data interference and standardizing processing, the problem of differences in dimensions and orders of magnitude is solved. The component iteration mechanism dynamically optimizes the correlation between independent and dependent variables, removes noise and focuses on core laws, ensuring the objectivity and reliability of the analysis results. It is adaptable to complex process scenarios. The parameter synergy effect is captured by the correlation coefficient of the second component, and the optimal solution is screened by combining least squares regression. Candidate formulations that meet the target indicators (such as epitaxial layer thickness and uniformity) are generated, taking into account both flexibility and practicality. It provides an interpretable and reproducible optimization scheme for silicon carbide epitaxial process research and development, shortens the research and development cycle and reduces experimental costs. Attached Figure Description
[0015] The above and / or additional aspects and advantages of the present invention will become apparent and readily understood from the description of the embodiments taken in conjunction with the following drawings, in which: Figure 1 This is a first flowchart of a method for generating a silicon carbide epitaxial process formulation according to an embodiment of the present invention; Figure 2 This is a second flowchart of the method for generating silicon carbide epitaxial process formulations provided in an embodiment of the present invention; Figure 3 This is a third flowchart of the silicon carbide epitaxial process formulation generation method provided in the embodiments of the present invention; Figure 4 This is a fourth flowchart of the silicon carbide epitaxial process formulation generation method provided in this embodiment of the invention; Figure 5 This is a fifth flowchart of the silicon carbide epitaxial process formulation generation method provided in the embodiments of the present invention; Figure 6 This is a sixth flowchart of the silicon carbide epitaxial process formulation generation method provided in the embodiments of the present invention; Figure 7 This is a seventh flowchart of a method for generating a silicon carbide epitaxial process formulation according to an embodiment of the present invention; Figure 8 This is the eighth flowchart of the silicon carbide epitaxial process formulation generation method provided in the embodiments of the present invention; Figure 9 This is a schematic diagram of the silicon carbide epitaxial process formulation generation apparatus provided in an embodiment of the present invention; Figure 10 This is a schematic diagram of the structure of the silicon carbide epitaxial process formulation generation equipment provided in an embodiment of the present invention. Detailed Implementation
[0016] The terms "first," "second," "third," "fourth," etc. (if present) in the specification, claims, and accompanying drawings of this invention are used to distinguish similar objects and are not necessarily used to describe a specific order or sequence. It should be understood that such data can be interchanged where appropriate so that the embodiments described herein can be implemented in orders other than those illustrated or described herein. Furthermore, the terms "comprising" or "having" and any variations thereof are intended to cover a non-exclusive inclusion; for example, a process, method, system, product, or apparatus that comprises a series of steps or units is not necessarily limited to those steps or units explicitly listed, but may include other steps or units not explicitly listed or inherent to such processes, methods, products, or apparatus.
[0017] For ease of understanding, the specific process of the embodiment of the present invention is described below. Figure 1One embodiment of the silicon carbide epitaxial process formulation generation method of the present invention includes: 101. Obtain historical experimental process data; In this embodiment, historical experimental process data includes two types of core information: adjustable "process operation parameters" (such as temperature, flow rate, etc.) and "performance indicators" (such as product purity, thickness, etc.) that reflect the experimental results. 102. Generate independent and dependent variables based on historical experimental process data; In this embodiment, the independent variables include commonly used process parameters in silicon carbide epitaxy, such as hydrogen carrier gas flow rate, TCS flow rate, and ethylene flow rate. Flow rate, air flotation flow rate, and process temperature; dependent variables include silicon carbide epitaxial process parameters: epitaxial layer thickness, thickness uniformity, doping concentration, doping concentration uniformity, and surface roughness; historical experimental process data also refers to these physical quantities; the number of columns in the independent variable matrix represents the number of process parameters (i.e., the number of independent variables), and the number of columns in the dependent variable matrix represents the number of epitaxial process parameters (i.e., the number of dependent variables). The number of rows in the independent variable matrix and the dependent variable matrix are equal. One row represents one complete epitaxial process, and the number of rows represents how many epitaxial processes were performed in historical experiments. 103. Standardize the independent and dependent variables to obtain standardized independent and dependent variables; In this embodiment, the independent and dependent variables are standardized (e.g., Z-score standardization, so that the mean is 0 and the standard deviation is 1); the weight imbalance caused by the difference in the units (e.g., temperature unit °C, flow rate unit sccm) or the order of magnitude (e.g., thickness and concentration) of different variables is eliminated, so as to ensure that each variable "participates fairly" in subsequent calculations and avoid a certain variable from dominating the analysis results due to its excessively large value; 104. The first component correlation coefficient is obtained from the standardized dependent variables based on the preset condition of the column with the largest variance; In this embodiment, the column with the largest variance from the standardized dependent variables is selected as the initial component u (i.e., the first component correlation coefficient). 105. Generate a weight vector based on the standardized independent variables and the correlation coefficient of the first component; In this embodiment, the weight vector Where X is the standardized independent variable, The apparatus matrix for standardized independent variables; 106. Generate the second-component correlation coefficient based on the weight vector and standardized independent variables; In this embodiment, t = Xw, where t is the part of the standardized independent variable X that is most relevant to the current u (i.e., the first component correlation coefficient) (i.e., the second component correlation coefficient); the weight vector is calculated based on the standardized dependent variable Y and the second component correlation coefficient t. Then u = Yc, u is updated to reflect the u in Y that is most relevant to t; w is the weight vector of X and u. Its calculation goal is to maximize the covariance between t and u. Covariance is a statistic that measures the degree of correlation. Therefore, according to the definition, the calculated w multiplied by X gives t, which is the part that is most consistent with u in direction and changes most synchronously. 107. Based on the preset data selection conditions and the correlation coefficient of the second component, solve the historical experimental process data and the preset target process indicators to obtain a set of candidate feasible solutions. In this embodiment, historical experimental process data is known. A PLS model (least squares regression model) is generated based on the historical experimental process data. Then, the model is solved in reverse (equivalent to obtaining its inverse). Given the target process index, the inverse of the PLS model is substituted to obtain a set of candidate feasible solutions. 108. Based on the preset least squares regression algorithm and the preset expected dependent variable, a set of candidate feasible solutions are screened to obtain process parameters; In this embodiment, the process parameters include hydrogen carrier gas flow rate, TCS flow rate, ethylene flow rate, N2 flow rate, and process temperature. 109. Generate silicon carbide epitaxial process formulation and evaluation report based on process parameters; In this embodiment, one or more candidate recipes can be generated through the output module, and the output module also allows users to manually adjust the generated candidate recipes; In this embodiment, the core dependent variable is screened by maximizing variance, and the influence intensity of independent variables is quantified by weight vectors. Key control factors are located from multiple sets of process parameters (such as hydrogen carrier gas flow rate, process temperature, etc.), improving optimization efficiency and avoiding the blindness of traditional trial and error. By eliminating data interference and standardizing processing, the problem of differences in dimensions and orders of magnitude is solved. The component iteration mechanism dynamically optimizes the correlation between independent and dependent variables, removes noise to focus on core laws, and ensures the objectivity and reliability of the analysis results. It is adaptable to complex process scenarios. The parameter synergy effect is captured by the correlation coefficient of the second component, and the optimal solution is screened by least squares regression. Candidate formulations that meet the target indicators (such as epitaxial layer thickness and uniformity) are generated, taking into account both flexibility and practicality. This provides an interpretable and reproducible optimization scheme for silicon carbide epitaxial process research and development, shortens the research and development cycle, and reduces experimental costs.
[0018] Please see Figure 2 In the second embodiment of the silicon carbide epitaxial process formulation generation method of the present invention, step 107 includes: 201. Perform row count analysis on historical experimental process data to obtain the number of rows in the historical data; In this embodiment, the historical experimental process data contains a total of n rows of data entries. The number of historical data rows represents the scale of the historical experimental process data. The number of historical data rows directly reflects the total number (n) of the historical experimental process data entries. This step provides the basis for the sample division of subsequent cross-validation, ensuring data coverage and verifiability. 202. Calculate the mean square error of historical experimental process data based on preset discard conditions, number of historical data rows, and data selection conditions. In this embodiment, based on preset discarding conditions (such as "row-by-row discarding", discarding one row of data starting from the first row, using the remaining n-1 rows as the training set, and the discarded row as the validation set (including the substitution of independent variables and the discarding of dependent variables)), the mean squared error (MSE) is calculated through cross-validation to evaluate the model's predictive ability on data not used in the training; the inner loop is repeated n times to obtain n MSEs, completing one round of error evaluation in the inner loop; 203. Perform statistical analysis on all mean square errors to obtain the mean square error count; 204. Calculate the correlation coefficients of the standardized independent variables and the second component to obtain the first loading; In this embodiment, the load p of X (i.e., the first load) is calculated. ), 205. Determine whether the mean square error count meets the preset inner loop iteration stopping condition; 206. If the mean square error does not meet the inner loop iteration stop condition, return to the previous step to calculate the historical experimental process data based on the discard conditions, the number of historical data rows, and the data selection conditions, until the mean square error meets the inner loop iteration stop condition. In this embodiment, the stopping condition for the inner loop iteration is that h mean square errors can be calculated from all the mean square errors, and n mean square errors will be calculated in each round of the inner loop; by allowing each data entry in the historical experimental process data to participate in the evaluation as a validation set, and each round of the inner loop stops when "n mean square errors are calculated"; 207. If the mean squared error number satisfies the inner loop iteration stopping condition, then analyze the mean squared error and the correlation coefficient of the second component according to the preset outer loop iteration stopping condition to obtain the optimal number of components; 208. Obtain all the first loads to obtain the first load set; In this embodiment, the first load p reflects the contribution of the independent variable to the correlation coefficient t of the second component: the larger the absolute value of the load (such as the flow rate of a certain gas), the more significant the impact on the process results, providing a basis for parameter priority ranking; the first load set originates from the full sample cross-validation of n rows of data entries in the historical experimental process data in the inner loop. By discarding data row by row and dynamically adjusting the input parameters (X and t), the first load adapted to different training sets is repeatedly calculated and finally summarized into the first load set; the inner loop iteration stops when the outer loop iterates through different component numbers and repeats multiple rounds of the inner loop until the average value of h mean square errors can be calculated from all mean square errors, then the inner loop iteration stops. 209. Generate a set of candidate feasible solutions based on the optimal number of components, the first load set, and the target process parameters; In this embodiment, the average value of n MSEs is calculated as the prediction performance index of the current number of components; after the outer loop iterates for different numbers of components, the number of components corresponding to the smallest average value of all MSEs is selected as the optimal number of components A. This step avoids the problems of "underfitting due to too few components" or "overfitting due to too many components", ensuring the generalization ability of the model. In this embodiment, systematic cross-validation and load analysis provide support for process parameter optimization. First, the data scale is clarified through row count analysis, laying the foundation for cross-validation. The training set is divided using a "row-by-row truncation" method, and the mean squared error (MSE) is calculated through n inner loops to comprehensively evaluate the model's prediction accuracy, maximize the use of historical data, and reduce additional experimental costs. Simultaneously, the contribution of the first load analytical independent variable is calculated to clarify the priority of key parameters and improve the interpretability of optimization. When the MSE equals the number of data rows, the optimal number of components is screened through an outer loop, selecting the component with the smallest MSE to balance model complexity and generalization ability, avoiding overfitting or underfitting. Finally, candidate solutions are generated by combining the optimal number of components, the load set, and the target index, which not only meet process requirements but also are based on data patterns, providing a basis for efficiently screening reliable process parameters.
[0019] Please see Figure 3 In the third embodiment of the silicon carbide epitaxial process formulation generation method of the present invention, step 202 includes: 301. Based on the selection criteria and the number of rows of historical data, the historical experimental process data is filtered to obtain the training set and the discarded dataset; In this embodiment, the historical data consists of n rows. The discarding condition is to discard rows one by one starting from the first row, and use the remaining n-1 rows to form the training set to achieve full sample coverage validation. This partitioning method ensures that each sample can be used as a validation set for evaluation, avoiding validation bias caused by fixed sample partitioning. The discarded dataset: the 1 row that is discarded is used as a validation sample, which includes the "independent variable" (the process parameter to be predicted) and the "discarded dependent variable" (the corresponding actual result index) for model accuracy verification. 302. Train the preset partial least squares regression algorithm based on the training set to obtain the first partial least squares regression model; In this embodiment, the advantage of partial least squares regression algorithm is that it can effectively handle the problem of multicollinearity of independent variables, and at the same time extract the core correlation between independent and dependent variables. It is especially suitable for multi-parameter coupled process scenarios (such as the interaction between multiple gas flow rates and temperatures in silicon carbide epitaxy). 303. Based on the data selection criteria, obtain the independent variables to be substituted and the dependent variables to be discarded from the discarded dataset; In this embodiment, the data selection criterion is to select the input independent variable and the discard dependent variable of a row that was just discarded in the inner loop. 304. Based on the first partial least squares regression model, predict the independent variables to obtain the predicted value of the dependent variable; 305. The mean squared error is calculated based on the predicted value of the dependent variable and the value of the discarded dependent variable. In this embodiment, the independent variable of a row that was just discarded in the inner loop is substituted into the temporary first partial least squares regression model to obtain the predicted value of the dependent variable for that row. The mean square error (MSE) between the predicted value of the dependent variable and the discarded dependent variable of the row is calculated to fully reflect the prediction accuracy of the model on different samples. In this embodiment, the cross-validation mechanism provides reliable support for the accuracy evaluation of the process model; the "row-by-row discarding" method is used to divide the samples, dynamically dividing n rows of historical data into a training set (n-1 rows) and a discarded dataset (1 row), ensuring that each sample participates in the validation, avoiding the bias of fixed division, and achieving an objective evaluation with full sample coverage; the partial least squares regression model trained on the training set can effectively handle the problem of multi-parameter collinearity, capture the core correlation between independent and dependent variables, and adapt to multi-parameter coupling scenarios such as silicon carbide epitaxy; by extracting the independent and dependent variables of the discarded rows through data selection conditions, the mean squared error (MSE) is calculated after substituting them into the model to obtain the predicted value, and the prediction accuracy of the model on different samples is fully reflected after n iterations; the value of historical data is maximized, providing a quantitative basis for subsequent model optimization and parameter selection, reducing R&D costs and improving process reliability.
[0020] Please see Figure 4In the fourth embodiment of the silicon carbide epitaxial process formulation generation method of the present invention, step 207 includes: 401. Calculate the average of all mean squared errors to obtain a set of mean squared error averages; In this embodiment, the mean squared error eliminates the influence of fluctuations in individual samples and more stably represents the overall prediction effect of the model under the current number of components (the higher the number of components, the higher the model complexity, and the mean needs to be used to evaluate whether it is overfitting). 402. Perform statistical analysis on the set of mean squared errors to obtain the mean squared error count; In this embodiment, n mean square errors are calculated in the inner loop, and the average of all n mean square errors is calculated to represent the prediction effect of that component. 403. Calculate the correlation coefficient between the standardized dependent variable and the second component to obtain the second loading; In this embodiment, the second load This quantifies the contribution of the dependent variable to the correlation coefficient t of the second component, supplementing the dependent variable association information not covered by the first loading; 404. Generate the residual matrix based on the second loading, the standardized dependent variable, and the correlation coefficient of the second component; In this embodiment, the residual matrix of the standardized dependent variable Y This reflects the dependent variable information not explained by the current second-component correlation coefficient t (the smaller the residual, the stronger the model's explanatory power for the dependent variable). 405. Determine whether the mean square error or the residual matrix satisfies the stopping condition of the outer loop iteration; In this embodiment, the outer loop iteration stops when the number of components selected by cross-validation is reached or the absolute value of the elements of the residual matrix F is close to 0 (i.e. the model has fully explained the dependent variable information). The elements of the residual matrix F being close to 0 means that mathematically continuing the iteration will no longer improve the accuracy of the model. The typical range is less than [1,5]. 406. When the average value of the second mean square error or the residual matrix satisfies the stopping condition of the outer loop iteration, obtain all the mean square error averages to obtain the set of mean square error averages. In this embodiment, conversely, a residual matrix is generated based on the first loading, the standardized independent variable, and the correlation coefficient of the second component. Update the dependent and independent variables based on the residual matrix F and residual matrix E respectively, return to execute the standardization of the independent and dependent variables, and extract the next set of t and u; 407. Obtain the minimum mean square error from the set of mean square error values; 408. Obtain the corresponding number of components based on the minimum mean square error average to obtain the optimal number of components; In this embodiment, until the outer loop calculation is completed, h mean squared errors are obtained. The component with the smallest mean squared error is selected from them, which is the optimal number of components A, ensuring that the model achieves the optimal balance between "lowest complexity" and "highest prediction accuracy". In this embodiment, a systematic iterative optimization mechanism is used to accurately select the optimal number of components, providing reliable parameter support for the process model. The mean square error of the n inner loops is averaged to eliminate the influence of single-sample fluctuations, stably reflecting the model's prediction effect based on the current number of components and avoiding overfitting risks. The contribution of the second load quantification dependent variable to the correlation coefficient of the second component is calculated to supplement variable correlation information. A residual matrix is generated to intuitively present information not explained by the model; the smaller the residual, the stronger the explanatory power. The outer loop dynamically determines the stopping condition; if the preset number of components is reached or the residual approaches 0, the iteration terminates; otherwise, the data is updated and new components are extracted. Finally, the number of components corresponding to the smallest mean square error from the h mean square errors is selected as the optimal value, achieving an optimal balance between model complexity and prediction accuracy. This process quantifies and standardizes evaluation, adapts to complex process scenarios, improves the effectiveness and reliability of process optimization, and reduces manual parameter tuning costs.
[0021] Please see Figure 5 In the fifth embodiment of the silicon carbide epitaxial process formulation generation method of the present invention, step 209 includes: 501. Optimize the first partial least squares regression model based on the optimal number of components to obtain the optimized least squares regression model; In this embodiment, the optimized least squares regression model represents the mapping relationship between the independent variable matrix and the dependent variable matrix. The optimized model clearly defines the mapping relationship between the independent variable matrix (process parameters, such as hydrogen carrier gas flow rate, process temperature, etc.) and the dependent variable matrix (outcome indicators, such as epitaxial layer thickness). This relationship has been calibrated with the optimal component number, which not only retains the core correlation information, but also avoids excessive model complexity caused by redundant parameters. 502. Solve the target process parameters based on the optimized least squares regression model and the first load set to obtain a set of candidate feasible solutions; In this embodiment, in the silicon carbide epitaxial process, the commonly used target process indicators include epitaxial layer thickness, thickness uniformity, doping concentration, doping concentration uniformity, and surface roughness. Finally, a set of process parameter combinations that meet the constraints of the target indicators are generated, namely, candidate feasible solutions, which lay the foundation for subsequent accurate selection of optimal process parameters. In this embodiment, after optimizing the first partial least squares regression model with the optimal component number, the resulting optimized model clearly defines the mapping relationship between process parameters (such as H2 hydrogen carrier gas flow rate and process temperature) and result indicators (such as epitaxial layer thickness). After calibration with the optimal component number, the core correlation information is retained while avoiding redundant parameters that lead to excessive model complexity, achieving a balance between prediction accuracy and generalization ability. Combining the optimized model with the first load set, the core target indicators (thickness, uniformity, doping concentration, etc.) of silicon carbide epitaxy are solved to generate process parameter combinations that meet the indicator constraints, i.e., candidate feasible solutions. This process focuses on key quality requirements, the candidate solutions fit the optimization objectives, reduce the cost of ineffective screening, and the parameter correlation is traceable, which is convenient for engineers to understand and adjust, efficiently connects the preceding and following processes, and accelerates the iteration of process development.
[0022] Please see Figure 6 In the sixth embodiment of the silicon carbide epitaxial process formulation generation method of the present invention, step 502 includes: 601. Obtain all weight vectors to obtain a weight vector set; In this embodiment, these vectors quantify the influence of independent variables on the correlation coefficient of the second component, and are the key basis for constructing the regression relationship; 602. Generate the first regression coefficient matrix and the generalized inverse matrix based on the weight vector set and the first loading set; In this embodiment, the first regression coefficient matrix is essentially a quantitative expression of the mapping relationship between independent and dependent variables, directly related to process parameters and result indicators; the generalized inverse matrix solves the problem that the inverse matrix of non-square or singular matrices does not exist, providing mathematical tools for subsequent solutions; 603. Determine whether the first regression coefficient matrix is a full-rank matrix. In this embodiment, it is determined whether the first regression coefficient matrix is a full-rank matrix (the column vectors are linearly independent). This determination determines the choice of solution method and ensures the validity of the solution. 604. When the first regression coefficient matrix is a full-rank column matrix, the generalized inverse matrix, dependent variable and first regression coefficient matrix are solved according to the preset solution formula to obtain the first exact solution; In this embodiment, the first exact solution is analytical and can accurately satisfy the target index constraints; In this embodiment, ,Right now This is the first exact solution, where I is the identity matrix, with 1s on the main diagonal and 0s on all other elements. Let B be the generalized inverse matrix of the first regression coefficient matrix. The dependent variable matrix; 605. Solve the target process parameters based on the optimized least squares regression model and the first exact solution to obtain a set of candidate feasible solutions; 606. Conversely, the generalized inverse matrix, dependent variable, and first regression coefficient matrix are iteratively solved using numerical methods to obtain the second exact solution. In this embodiment, B is mostly a non-square matrix and does not have a full rank, so a numerical iterative solution is needed, with the goal of making B such that... Minimize, and gradually approach the optimal solution through iterative optimization, adapting to complex data structures; 607. Solve for the target process indicators based on the optimized least squares regression model and the second exact solution to obtain a set of candidate feasible solutions; In this embodiment, the target process indicators include the target range of epitaxial indicators (such as a thickness of 10±0.5um), the priority of epitaxial indicators, and whether to use certain indicators based on the priority of the epitaxial targets. Some constraints of the process are defined (such as the longest process time, a specific gas flow range, etc.). The solution process takes into account the priority of indicators to ensure that key quality parameters are met first. In this embodiment, a systematic matrix operation and flexible solution mechanism are used to support the optimization of process parameters. The influence of independent variables on the correlation coefficient of the second component is quantified by collecting a weight vector set. Combined with the first load set, a first regression coefficient matrix is generated, clearly presenting the mapping relationship between process parameters and result indicators. A generalized inverse matrix solves the problem of solving non-square or singular matrices. An appropriate solution method is selected by determining whether the matrix is full rank: when full rank, an analytical method is used to find the first exact solution, satisfying the indicator constraints; when not full rank, a numerical method is used iteratively to find the second exact solution, approximating the optimal solution by minimizing the error, adapting to complex data structures. The solution process incorporates the target process indicator range, priority, and process constraints (such as flow limits) to ensure that key quality parameters are met first. This solution balances theoretical rigor with practical applicability; the solution generation logic is traceable, efficiently connecting the model and application, and accelerating the implementation of process formulations.
[0023] Please see Figure 7 In the seventh embodiment of the silicon carbide epitaxial process formulation generation method of the present invention, step 602 includes: 701. Generate a weight matrix based on the weight vector set; In this embodiment, the weight vector quantifies the influence strength of the independent variable on the correlation coefficient of the second component, while the weight matrix systematically integrates multiple sets of weight information to fully reflect the influence pattern of the independent variable under different components. 702. Obtain all the second loads to obtain the second load set; In this embodiment, the second load set originates from the full-range test of multiple sets of second component correlation coefficients in the outer loop. By iterating over different second component correlation coefficients, the second loads that adapt to different second component correlation coefficients are repeatedly calculated and finally summarized into the second load set. 703. Generate the second load matrix based on the second load set; In this embodiment, the second loading matrix reflects the correlation between the dependent variable and the second component correlation coefficient, and the two together constitute a key quantitative indicator of variable correlation. 704. Generate the first load matrix based on the first load set; 705. Generate the first regression coefficient matrix based on the weight matrix, the first loading matrix, and the second loading matrix; In this embodiment, the formula for solving the first regression coefficient matrix is: In the formula, This is the first regression coefficient matrix. This is the weight matrix. This is the first load matrix. This is the transpose of the first load matrix. The first loading matrix is the transpose of the second loading moment. The first loading matrix reflects the contribution of the independent variable to the correlation coefficient of the second component, and the second loading matrix reflects the relationship between the dependent variable and the correlation coefficient of the second component. Together, they constitute the key quantitative indicators of variable relationship. The first regression coefficient matrix is a direct quantitative expression of the mapping relationship between the independent variable and the dependent variable. For example, the element values in the matrix intuitively reflect the influence coefficient of a certain process parameter (such as H2 flow rate) on the result indicator (such as epitaxial layer thickness), providing a mathematical basis for parameter optimization. 706. Transform the first regression coefficient matrix to obtain the generalized inverse matrix; In this embodiment, ( Let be the generalized inverse matrix of the first regression coefficient matrix B. Let be the transpose of the first regression coefficient matrix B. In process data, the regression coefficient matrix is often a non-square matrix or a singular matrix (not full rank). The traditional inverse matrix does not exist. However, the generalized inverse matrix solves this problem through mathematical transformation, ensuring that it can still be effectively solved in complex data scenarios and providing the necessary mathematical tools for subsequent accurate solution calculation. In this embodiment, a systematic matrix construction and transformation provides mathematical support for process parameter optimization. A weight matrix is generated from a weight vector set, systematically integrating multiple sets of weight information to fully reflect the influence of independent variables under different components. The load matrices generated by combining the first and second load sets demonstrate the contribution of independent variables to the correlation coefficient of the second component and reflect the relationship between the dependent variable and the correlation coefficient of the second component, collectively constituting key quantitative indicators. The first regression coefficient matrix, generated through a specific formula, directly quantifies the mapping relationship between independent and dependent variables, clarifying the influence coefficients of process parameters on the result indicators. The generalized inverse matrix obtained by transforming the regression coefficient matrix solves the problem of solving non-square or singular matrices, adapting to complex data scenarios. This solution systematically integrates dispersed information, improving model interpretability and solution rigor, laying the foundation for subsequent accurate solution calculation, and efficiently promoting process parameter optimization.
[0024] Please see Figure 8 In the eighth embodiment of the silicon carbide epitaxial process formulation generation method of the present invention, step 108 includes: 801. Randomly extract and replace historical experimental process data according to the preset number of repetitions to obtain multiple Bootstrap samples; In this embodiment, N samples are randomly selected with replacement from historical experimental process data, and repeated M times. Typically, the value of M ranges from 1000 to 10000, resulting in multiple Bootstrap samples. By simulating data fluctuations through sampling, the generated sample set retains the statistical characteristics of the original data and reflects the uncertainties in the data (such as experimental errors and sample bias), providing a basis for subsequent model robustness analysis. 802. Based on the least squares regression algorithm and multiple bootstrap samples, multiple second partial least squares regression models and multiple second regression coefficient matrices are constructed; In this embodiment, a partial least squares regression model is constructed for each Bootstrap sample, resulting in a set of second partial least squares regression models and a set of second regression coefficient matrices. This group It can reflect the uncertainty of B, and is especially suitable for the common problem of "unstable parameter correlation" in process data (such as the fluctuation of the synergistic influence of multiple gas flow rates), and can cover possible correlation patterns through multiple models; 803. Based on the preset expected dependent variable and multiple second partial least squares regression models, the multiple second regression coefficient matrices are solved in reverse to obtain multiple estimated values; In this embodiment, for the desired dependent variable, for each Perform a reverse calculation on each one to obtain M estimated values; 804. Generate confidence intervals based on multiple estimates; In this embodiment, a confidence interval is used: the 2.5% and 97.5% quantiles of the estimated value are used as the 95% confidence interval. This interval reflects the possible range of values of the process parameter with a 95% probability, and quantifies the uncertainty of the parameter. 805. Filter a set of candidate feasible solutions based on the confidence interval to obtain the process parameters; In this embodiment, the candidate feasible solutions generated in the early stage are screened according to the confidence interval, and the parameter combinations that fall within the interval are retained to finally obtain the process parameters; 806. Evaluate the rule base to obtain process formulations and evaluation reports; In this embodiment, the selected parameters are evaluated in conjunction with a rule base (such as process safety constraints and equipment limitations) to generate a process formula and an evaluation report that includes parameter uncertainty and compliance. In this embodiment, a process formula and evaluation report can be generated based on process parameters. The specific steps are as follows: The core entity is obtained by defining the target process indicators, preset process steps, and process parameters. The core relationship is defined based on the pre-defined subordinate relationships, pre-defined constraint relationships, and mapping relationships; Acquire knowledge in the field of silicon carbide processes, and extract this knowledge to obtain structured data, semi-structured data, and unstructured data. A knowledge graph is constructed based on core entities, core relationships, structured data, semi-structured data, and unstructured data. (The main function of knowledge graphs is to classify process formulation steps and instructions, and to verify their safety and rationality. The main steps in their construction are:) 1. Define core entities (process objectives, process steps, process parameters, equipment, materials, etc.); 2. Define core relationships: subordinate relationships (e.g., process steps include temperature parameters), constraint relationships (e.g., the value of parameter A is affected by parameter B), and mapping relationships (e.g., process objective G corresponds to process step sequence S); 3. Data Acquisition and Preprocessing: Knowledge of silicon carbide processes is collected from multiple dimensions, including but not limited to literature, industry manuals, safety regulations, and expert experience. Data processing includes removing duplicate information and standardizing terminology. 4. Knowledge Extraction: Structured data (tables) can be directly mapped to entities (e.g., "Reactor A: Maximum temperature 1800℃" in the equipment manual can be extracted as (Reactor A, Maximum temperature, 1800℃), semi-structured data (process documents) can be matched or parsed using templates, and unstructured data (such as papers and expert interviews) can be extracted using NLP techniques and domain dictionaries. 5. Knowledge storage: Choose a graph database for storage; Rules are extracted from the knowledge graph; Entity association analysis is performed on the rules and knowledge graph to obtain entity associations; A rule base is constructed based on preset rule types and entity relationships; (The specific steps for building the rule base are as follows:) 1. Rule Types: Process Step Division Rules, Safety Verification Rules, Reasonableness Verification Rules 2. Rule Extraction: Extract rules from the knowledge graph (e.g., extract safety rules from "Reactor A - Maximum Temperature - 1800℃" in the knowledge graph; extract rationality rules from expert experience "When the epitaxial growth time exceeds 60 min, the temperature needs to be reduced by 50℃ to avoid substrate damage"). 3. Relationship between rules and knowledge graph: Entities in the rules are associated with entities in the knowledge graph. During reasoning, parameters are obtained through the knowledge graph, and process constraints are transformed into executable logical rules based on the information in the knowledge graph. In this embodiment, Bootstrap sampling and multi-model analysis provide reliable assurance for process parameter screening. N samples are extracted from historical data with replacement, and repeated 1000-10000 times to generate multiple Bootstrap samples. This preserves the statistical characteristics of the original data while reflecting uncertainties such as experimental errors. A partial least squares regression model and regression coefficient matrix are constructed based on each sample. Multiple models cover the fluctuation patterns of parameter correlations, adapting to unstable scenarios in complex processes. Multiple estimated values are obtained by inversely solving the expected dependent variable. A 95% confidence interval is constructed using the 2.5% and 97.5% quantiles to quantify the possible value range of the parameters. Candidate solutions are screened based on the confidence intervals, retaining parameter combinations within the intervals. Process formulations and evaluation reports are generated by combining rule base verification. This scheme quantifies uncertainty, enhances parameter robustness, reduces invalid experiments, and the output report is both compliant and interpretable, providing support for process optimization with high stability requirements.
[0025] The method for generating silicon carbide epitaxial process formulations in the embodiments of the present invention has been described above. The apparatus for generating silicon carbide epitaxial process formulations in the embodiments of the present invention is described below. Please refer to [link / reference]. Figure 9 One embodiment of the silicon carbide epitaxial process formulation generation apparatus of the present invention includes: Process data acquisition module 1 is used to acquire historical experimental process data; Variable generation module 2 is used to generate independent and dependent variables based on historical experimental process data; Standardization module 3 is used to standardize the independent and dependent variables to obtain standardized independent and dependent variables; Select module 4 to select the first component correlation coefficient from the standardized dependent variable based on the preset maximum variance column condition; Weight vector generation module 5 is used to generate weight vectors based on standardized independent variables and the first component correlation coefficient; Coefficient generation module 6 is used to generate the second-component correlation coefficient based on the weight vector and standardized independent variables; The solution module 7 is used to solve the historical experimental process data and the preset target process index based on the preset data selection conditions and the correlation coefficient of the second component, so as to obtain a set of candidate feasible solutions. The screening module 8 is used to screen a set of candidate feasible solutions based on a preset least squares regression algorithm and a preset expected dependent variable to obtain process parameters. Formula and report generation module 9 is used to generate silicon carbide epitaxial process formulas and evaluation reports based on process parameters; In this embodiment, the core dependent variable is screened by maximizing variance, and the influence intensity of independent variables is quantified by weight vectors. Key control factors are located from multiple sets of process parameters (such as hydrogen carrier gas flow rate, process temperature, etc.), improving optimization efficiency and avoiding the blindness of traditional trial and error. By eliminating data interference and standardizing processing, the problem of differences in dimensions and orders of magnitude is solved. The component iteration mechanism dynamically optimizes the correlation between independent and dependent variables, removes noise to focus on core laws, and ensures the objectivity and reliability of the analysis results. It is adaptable to complex process scenarios. The parameter synergy effect is captured by the correlation coefficient of the second component, and the optimal solution is screened by least squares regression. Candidate formulations that meet the target indicators (such as epitaxial layer thickness and uniformity) are generated, taking into account both flexibility and practicality. This provides an interpretable and reproducible optimization scheme for silicon carbide epitaxial process research and development, shortens the research and development cycle, and reduces experimental costs.
[0026] Figure 10 This is a schematic diagram of the structure of a silicon carbide epitaxial process formulation generation device 900 provided in an embodiment of the present invention. The silicon carbide epitaxial process formulation generation device 900 can vary significantly due to different configurations or performance. It may include one or more central processing units (CPUs) 910 (e.g., one or more processors) and a memory 920, and one or more storage media 930 (e.g., one or more mass storage devices) storing application programs 933 or data 932. The memory 920 and storage media 930 can be temporary or persistent storage. The program stored in the storage media 930 may include one or more modules (not shown in the diagram), each module may include a series of instruction operations on the silicon carbide epitaxial process formulation generation device 900. Furthermore, the processor 910 may be configured to communicate with the storage media 930 and execute a series of instruction operations in the storage media 930 on the silicon carbide epitaxial process formulation generation device 900 to implement the steps of the silicon carbide epitaxial process formulation generation method provided in the above-described method embodiments.
[0027] The silicon carbide epitaxial process formulation generation equipment 900 may also include one or more power supplies 940, one or more wired or wireless network interfaces 950, one or more input / output interfaces 960, and / or one or more operating systems 931, such as Windows Server, MacOSX, Unix, Linux, FreeBSD, etc. Those skilled in the art will understand that... Figure 10 The illustrated structure of the silicon carbide epitaxial process formulation generation equipment does not constitute a limitation on the silicon carbide epitaxial process formulation generation equipment. It may include more or fewer components than illustrated, or combine certain components, or have different component arrangements.
[0028] The present invention also provides a computer-readable storage medium, which can be a non-volatile computer-readable storage medium or a volatile computer-readable storage medium, wherein the computer-readable storage medium stores instructions that, when executed on a computer, cause the computer to perform the steps of the silicon carbide epitaxial process formulation generation method.
[0029] Those skilled in the art will clearly understand that, for the convenience and brevity of description, the specific working processes of the above-described systems, devices, and units can refer to the corresponding processes in the aforementioned method embodiments and will not be repeated here.
[0030] If the integrated unit is implemented as a software functional unit and sold or used as an independent product, it can be stored in a computer-readable storage medium. Based on this understanding, the technical solution of the present invention, in essence, or the part that contributes to the prior art, or all or part of the technical solution, can be embodied in the form of a software product. This computer software product is stored in a storage medium and includes several instructions to cause a computer device (which may be a personal computer, server, or network device, etc.) to execute all or part of the steps of the methods described in the various embodiments of the present invention. The aforementioned storage medium includes various media capable of storing program code, such as USB flash drives, portable hard drives, read-only memory (ROM), random access memory (RAM), magnetic disks, or optical disks.
[0031] Finally, it should be noted that the above descriptions are merely preferred embodiments of the present invention and are not intended to limit the present invention. Although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art can still modify the technical solutions described in the foregoing embodiments or make equivalent substitutions for some of the technical features. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the protection scope of the present invention.
Claims
1. A method for generating a silicon carbide epitaxial process formulation, characterized in that, include: Obtain historical experimental process data; Independent and dependent variables are generated based on historical experimental process data; The independent and dependent variables are standardized to obtain standardized independent and standardized dependent variables; The first component correlation coefficient is obtained from the standardized dependent variables based on the preset maximum variance column condition; A weight vector is generated based on the standardized independent variables and the correlation coefficient of the first component. The second-component correlation coefficient is generated based on the weight vector and standardized independent variables; Based on the preset data selection conditions and the correlation coefficient of the second component, the historical experimental process data and the preset target process indicators are solved to obtain a set of candidate feasible solutions. Based on the preset least squares regression algorithm and the preset expected dependent variable, a set of candidate feasible solutions are screened to obtain process parameters; Generate silicon carbide epitaxial process formulations and evaluation reports based on process parameters.
2. The method for generating the silicon carbide epitaxial process formulation as described in claim 1, characterized in that, The process involves solving historical experimental process data and preset target process indicators based on preset data selection conditions and the correlation coefficient of the second component to obtain a set of candidate feasible solutions, including: Perform row count analysis on historical experimental process data to obtain the number of rows in the historical data; The mean square error is calculated based on the preset discard criteria, the number of rows of historical data, and the data selection criteria. Statistical analysis was performed on all mean squared errors to obtain the mean squared error count; The correlation coefficients of the standardized independent variables and the second component are calculated to obtain the first loading; Determine whether the mean square error count meets the preset inner loop iteration stopping condition; If the mean square error does not meet the inner loop iteration stop condition, return to the previous step to calculate the historical experimental process data based on the discard condition, the number of historical data rows, and the data selection condition, until the mean square error meets the inner loop iteration stop condition. If the mean square error number satisfies the inner loop iteration stopping condition, then the mean square error and the correlation coefficient of the second component are analyzed according to the preset outer loop iteration stopping condition to obtain the optimal number of components. Obtain all the first loads to obtain the first load set; A set of candidate feasible solutions is generated based on the optimal number of components, the first load set, and the target process parameters.
3. The method for generating the silicon carbide epitaxial process formulation as described in claim 2, characterized in that, The step of calculating the mean square error of historical experimental process data based on preset discarding criteria, the number of historical data rows, and data selection criteria includes: Historical experimental process data are filtered based on the selection criteria and the number of rows of historical data to obtain the training set and the discarded dataset. The pre-defined partial least squares regression algorithm is trained based on the training set to obtain the first partial least squares regression model. Based on the data selection criteria, obtain the independent variables and dependent variables to be substituted and discarded from the discarded dataset; The dependent variable is predicted by substituting the independent variables into the first partial least squares regression model. The mean squared error is calculated based on the predicted value of the dependent variable and the value of the discarded dependent variable.
4. The method for generating the silicon carbide epitaxial process formulation as described in claim 3, characterized in that, The step of analyzing the mean square error and the correlation coefficient of the second component based on the preset outer loop iteration stopping condition to obtain the optimal number of components includes: Calculate the average of all mean squared errors to obtain a set of mean squared error averages; Perform statistical analysis on the set of mean squared errors to obtain the mean squared error count; The correlation coefficient between the standardized dependent variable and the second component is calculated to obtain the second loading; The residual matrix is generated based on the second load, the standardized dependent variable, and the correlation coefficient of the second component. Determine whether the mean squared error or the residual matrix satisfies the stopping condition of the outer loop iteration; If the average value of the second mean square error or the residual matrix satisfies the stopping condition of the outer loop iteration, then the minimum mean square error is obtained from the set of average mean square errors. The number of components is obtained by taking the average value of the minimum mean square error, so as to obtain the optimal number of components.
5. The method for generating the silicon carbide epitaxial process formulation as described in claim 4, characterized in that, The step of generating a set of candidate feasible solutions based on the optimal number of components, the first load set, and the target process parameters includes: The first partial least squares regression model is optimized based on the optimal number of components to obtain the optimized least squares regression model. The target process parameters are solved using the optimized least squares regression model and the first load set to obtain a set of candidate feasible solutions.
6. The method for generating the silicon carbide epitaxial process formulation as described in claim 5, characterized in that, The step involves solving for the target process parameters based on the optimized least squares regression model and the first load set to obtain a set of candidate feasible solutions, including: Obtain all weight vectors to obtain a weight vector set; Generate the first regression coefficient matrix and the generalized inverse matrix based on the weight vector set and the first loading set; Determine whether the first regression coefficient matrix is a full-rank column matrix; When the first regression coefficient matrix is a full-rank column matrix, the generalized inverse matrix, dependent variable and first regression coefficient matrix are solved according to the preset solution formula to obtain the first exact solution; The target process parameters are solved using the optimized least squares regression model and the first exact solution to obtain a set of candidate feasible solutions; Conversely, the generalized inverse matrix, dependent variable, and first regression coefficient matrix are iteratively solved using numerical methods to obtain the second exact solution; The target process parameters are solved using the optimized least squares regression model and the second exact solution to obtain a set of candidate feasible solutions.
7. The method for generating the silicon carbide epitaxial process formulation as described in claim 6, characterized in that, The step of generating the first regression coefficient matrix and the generalized inverse matrix based on the weight vector set and the first loading set includes: Generate a weight matrix based on the weight vector set; Obtain all the second loads to obtain the second load set; Generate a second load matrix based on the second load set; Generate a first load matrix based on the first load set; The first regression coefficient matrix is generated based on the weight matrix, the first loading matrix, and the second loading matrix; The first regression coefficient matrix is transformed to obtain the generalized inverse matrix.
8. The method for generating the silicon carbide epitaxial process formulation as described in claim 1, characterized in that, The step of filtering a set of candidate feasible solutions based on a preset least squares regression algorithm and a preset expected dependent variable to obtain process parameters includes: Based on a preset number of repetitions, historical experimental process data are randomly sampled and replaced to obtain multiple Bootstrap samples. Based on the least squares regression algorithm and multiple bootstrap samples, multiple second partial least squares regression models and multiple second regression coefficient matrices were constructed. Based on the pre-defined expected dependent variable and multiple second partial least squares regression models, multiple second regression coefficient matrices are solved in reverse to obtain multiple estimated values; Generate confidence intervals based on multiple estimates; A set of candidate feasible solutions is screened based on the confidence interval to obtain the process parameters.
9. A silicon carbide epitaxial process formulation generation apparatus, characterized in that, include: The process data acquisition module is used to acquire historical experimental process data; The variable generation module is used to generate independent and dependent variables based on historical experimental process data. The standardization module is used to standardize independent and dependent variables to obtain standardized independent and dependent variables. The selection module is used to select the first component correlation coefficient from the standardized dependent variables based on the preset maximum variance column conditions; The weight vector generation module is used to generate a weight vector based on the standardized independent variables and the first-component correlation coefficient. The coefficient generation module is used to generate the second-component correlation coefficient based on the weight vector and standardized independent variables. The solution module is used to solve historical experimental process data and preset target process indicators based on preset data selection conditions and the correlation coefficient of the second component, so as to obtain a set of candidate feasible solutions. The filtering module is used to filter a set of candidate feasible solutions based on a preset least squares regression algorithm and a preset expected dependent variable to obtain process parameters. The formulation and report generation module is used to generate silicon carbide epitaxial process formulations and evaluation reports based on process parameters.
10. Equipment for generating silicon carbide epitaxial process formulations, characterized in that, The silicon carbide epitaxial process formula generation device includes: a memory and at least one processor, wherein the memory stores instructions; At least one of the processors invokes the instructions in the memory to cause the silicon carbide epitaxial process formulation generation apparatus to perform the steps of the silicon carbide epitaxial process formulation generation method as described in any one of claims 1-8.
Citation Information
Patent Citations
Industrial process technology optimization method based on partial least squares
CN111476405A
Cited By
Ion source beam current self-adaption method, device and equipment and storage medium
CN121748249A
