Method, device and medium for predicting gallium oxide epitaxial layer

Abstract

The application relates to a gallium oxide epitaxial layer prediction method, device, equipment and medium, wherein the method comprises the following steps: collecting process parameters of MOCVD and epitaxial layer characterization data thereof, and performing data preprocessing to obtain preprocessed data; performing feature derivation and feature screening on the preprocessed data, performing data set division on the feature screened data to generate a training data set and a test data set; constructing an XGBoost regression model, performing hyperparameter search to obtain a target optimal parameter combination, training the XGBoost regression model, and generating a trained XGBoost regression model; testing the trained XGBoost regression model through the test data set to generate a target XGBoost regression model; and acquiring to-be-tested process parameters of the MOCVD, performing epitaxial layer prediction on the to-be-tested process parameters to obtain a target epitaxial layer thickness and a target full-width-at-half-maximum change amount. The application improves the prediction accuracy of the gallium oxide epitaxial layer.

Description

Technical Field

[0001] This application relates to the field of semiconductor material preparation technology, and in particular to a method, apparatus, device and medium for predicting gallium oxide epitaxial layers. Background Technology

[0002] Gallium oxide (β-Ga2O3), as an emerging ultra-wide bandgap semiconductor material, has shown great application potential in fields such as high-voltage power devices and deep-ultraviolet detectors. However, the growth of high-quality gallium oxide epitaxial layers using metal-organic chemical vapor deposition (MOCVD) still faces many technical challenges. First, in terms of process control, precursor materials (such as trimethylgallium with oxygen) are prone to gas-phase pre-reactions, leading to deterioration of the epitaxial layer surface morphology and making it difficult to control growth uniformity. Second, the epitaxial growth process involves complex multiphysics coupling effects, including dynamic processes such as gas-phase transport, surface adsorption, and pyrolysis reactions. These factors interact and exhibit strong nonlinear characteristics, making process parameter optimization extremely difficult.

[0003] In existing technologies, researchers have attempted to optimize process parameters through experimental trial and error, but this method is inefficient and costly. Although some studies have tried to introduce machine learning techniques for process optimization, existing models generally suffer from the following problems: First, the static modeling approach cannot accurately reflect the dynamic process characteristics of epitaxial growth, resulting in low prediction accuracy for epitaxial layers; second, the large amount of data required for model training, coupled with the high cost of actual MOCVD experiments leading to limited sample sizes, easily results in overfitting, further limiting prediction accuracy; third, the models have poor interpretability, making it difficult to deduce specific process improvement directions from the prediction results. Furthermore, traditional methods have shortcomings in feature engineering, failing to effectively handle the complex interactions between process parameters, resulting in prediction accuracy that is insufficient to meet actual production needs. Summary of the Invention

[0004] The purpose of this application is to provide a method, apparatus, device, and medium for predicting gallium oxide epitaxial layers, so as to improve the prediction accuracy of gallium oxide epitaxial layers.

[0005] To address the aforementioned technical problems, embodiments of this application provide a method for predicting gallium oxide epitaxial layers, comprising:

[0006] The process parameters of MOCVD and the corresponding epitaxial layer characterization data are collected, and the process parameters and epitaxial layer characterization data are preprocessed to obtain preprocessed data.

[0007] The preprocessed data is subjected to feature derivation and feature filtering to obtain feature-filtered data. The feature-filtered data is then divided into datasets to generate training datasets and test datasets.

[0008] An XGBoost regression model is constructed. The Grey Wolf optimization algorithm is used to search for hyperparameters based on the training dataset to obtain the target optimal parameter combination. The XGBoost regression model is then trained based on the target optimal parameter combination to generate the trained XGBoost regression model.

[0009] The trained XGBoost regression model was tested using a test dataset to generate the target XGBoost regression model.

[0010] The process parameters to be tested in the MOCVD are obtained, and the target is predicted based on the process parameters to obtain the target epitaxial layer thickness and the change in the target full width at half maximum (FWHM) using the target XGBoost regression model.

[0011] To address the aforementioned technical problems, embodiments of this application provide a prediction device for gallium oxide epitaxial layers, comprising:

[0012] The data acquisition module is used to acquire the process parameters of MOCVD and the corresponding epitaxial layer characterization data, and to preprocess the process parameters and epitaxial layer characterization data to obtain preprocessed data.

[0013] The data processing module is used to perform feature derivation and feature filtering on the preprocessed data to obtain feature-filtered data, and to divide the feature-filtered data into datasets to generate training datasets and test datasets.

[0014] The model training module is used to construct an XGBoost regression model. It uses the Grey Wolf optimization algorithm to perform hyperparameter search based on the training dataset to obtain the target optimal parameter combination, and trains the XGBoost regression model based on the target optimal parameter combination to generate the trained XGBoost regression model.

[0015] The model testing module is used to test the trained XGBoost regression model using a test dataset and generate the target XGBoost regression model.

[0016] The epitaxial layer prediction module is used to obtain the process parameters to be tested in the MOCVD, and to predict the epitaxial layer based on the process parameters to be tested using the target XGBoost regression model, thereby obtaining the target epitaxial layer thickness and the change in the target full width at half maximum (FWHM).

[0017] To solve the above-mentioned technical problems, one technical solution adopted by the present invention is to provide a computer device, including one or more processors; and a memory for storing one or more programs, such that the one or more processors implement the gallium oxide epitaxial layer prediction method described in any one of the above-mentioned methods.

[0018] To address the aforementioned technical problems, one technical solution adopted by the present invention is: a computer-readable storage medium storing a computer program, wherein the computer program, when executed by a processor, implements the gallium oxide epitaxial layer prediction method described in any one of the above embodiments. Embodiments of the present invention provide a method, apparatus, device, and medium for predicting gallium oxide epitaxial layers. The method includes: collecting MOCVD process parameters and corresponding epitaxial layer characterization data, and preprocessing the process parameters and epitaxial layer characterization data to obtain preprocessed data; performing feature derivation and feature filtering on the preprocessed data to obtain feature-filtered data, and dividing the feature-filtered data into a dataset to generate a training dataset and a test dataset; constructing an XGBoost regression model, using the Grey Wolf optimization algorithm to perform hyperparameter search based on the training dataset to obtain the target optimal parameter combination, and training the XGBoost regression model based on the target optimal parameter combination to generate a trained XGBoost regression model; testing the trained XGBoost regression model using the test dataset to generate a target XGBoost regression model; obtaining the MOCVD process parameters to be tested, and using the target XGBoost regression model to predict the epitaxial layer thickness and the target full width at half maximum (FWHM) variation. This invention constructs an XGBoost regression model combined with the Grey Wolf optimization algorithm for hyperparameter search and handles the complex relationships between process parameters. It can accurately predict the changes in epitaxial layer thickness and full width at half maximum (FWHM). It has the advantages of improving process optimization efficiency, enhancing model generalization ability, and realizing dynamic updates, thereby improving the prediction accuracy of gallium oxide epitaxial layers. Attached Figure Description

[0019] To more clearly illustrate the solutions in this application, the accompanying drawings used in the description of the embodiments of this application will be briefly introduced below. Obviously, the accompanying drawings described below are some embodiments of this application. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0020] Figure 1 This is a flowchart illustrating the implementation of the method for predicting gallium oxide epitaxial layers provided in this application embodiment;

[0021] Figure 2 This is a flowchart illustrating the implementation of the first neutron flow of the prediction method for gallium oxide epitaxial layers provided in this application embodiment;

[0022] Figure 3 This is a flowchart illustrating the implementation of the second neutron flow in the prediction method for gallium oxide epitaxial layers provided in this application embodiment;

[0023] Figure 4 This is a flowchart illustrating the implementation of the third neutron flow in the prediction method for gallium oxide epitaxial layers provided in this application embodiment;

[0024] Figure 5 This is a flowchart illustrating the implementation of the fourth neutron flow in the prediction method for gallium oxide epitaxial layers provided in this application embodiment;

[0025] Figure 6 This is a flowchart illustrating the implementation of the fifth neutron process in the prediction method for gallium oxide epitaxial layers provided in this application embodiment;

[0026] Figure 7 This is a flowchart illustrating the implementation of the sixth neutron process in the prediction method for gallium oxide epitaxial layers provided in this application embodiment;

[0027] Figure 8 This is a schematic diagram showing the prediction and fitting effect of the model on the changes in thickness and full width at half maximum (FWHM) in the training dataset after using the XGBoost algorithm.

[0028] Figure 9 This is the prediction fit of the model on the test dataset after using the XGBoost algorithm for the changes in thickness and full width at half maximum (FWHM).

[0029] Figure 10 A schematic diagram of the high-quality epitaxial layer prepared under the guidance of the target XGBoost regression model;

[0030] Figure 11 This is a schematic diagram of the prediction device for gallium oxide epitaxial layers provided in an embodiment of this application;

[0031] Figure 12 This is a schematic diagram of the computer device provided in the embodiments of this application. Detailed Implementation

[0032] Unless otherwise defined, all technical and scientific terms used herein have the same meaning as commonly understood by one of ordinary skill in the art to which this application pertains; the terminology used herein in the specification of the application is for the purpose of describing particular embodiments only and is not intended to be limiting of the application; the terms "comprising" and "having," and any variations thereof, in the specification, claims, and foregoing drawings of this application, are intended to cover non-exclusive inclusion. The terms "first," "second," etc., in the specification, claims, or foregoing drawings of this application are used to distinguish different objects, not to describe a particular order.

[0033] In this document, the term "embodiment" means that a particular feature, structure, or characteristic described in connection with an embodiment may be included in at least one embodiment of this application. The appearance of this phrase in various places throughout the specification does not necessarily refer to the same embodiment, nor is it a separate or alternative embodiment mutually exclusive with other embodiments. It will be explicitly and implicitly understood by those skilled in the art that the embodiments described herein can be combined with other embodiments.

[0034] To enable those skilled in the art to better understand the present application, the technical solutions in the embodiments of the present application will be clearly and completely described below with reference to the accompanying drawings.

[0035] The present invention will now be described in detail with reference to the accompanying drawings and embodiments.

[0036] It should be noted that the method for predicting gallium oxide epitaxial layers provided in this application is generally executed by a server, and correspondingly, the device for predicting gallium oxide epitaxial layers is generally configured in the server.

[0037] In existing technologies, the growth of gallium oxide epitaxial layers suffers from problems such as pre-reaction control failure, narrow process windows, insufficient defect suppression, and low prediction accuracy of machine learning models. Traditional methods rely on trial and error to adjust process parameters, which cannot effectively handle the complexity of multi-physics coupling and high-dimensional parameter spaces, resulting in large fluctuations in epitaxial layer thickness and high defect density. Existing prediction models, due to neglecting dynamic process characteristics and overfitting to small samples, struggle to achieve accurate predictions, thus limiting the improvement of gallium oxide device performance.

[0038] To address the aforementioned issues, existing process parameter optimization methods fail to resolve the nonlinear relationships between parameters, and the characteristics of the dynamic growth process are not effectively extracted. By analyzing the coupling mechanism between gas-phase transport and surface reactions, it is recognized that the synergistic effect of temperature gradient and gas flow rate needs to be quantified through data modeling. To solve the problem of poor model generalization ability, this paper proposes embedding the physical mechanism into feature engineering and combining it with intelligent algorithms to optimize the model structure, thereby establishing a mapping relationship between process parameters and epitaxial layer quality. Therefore, this application proposes a method for predicting gallium oxide epitaxial layers, including: collecting MOCVD process parameters and corresponding epitaxial layer characterization data and preprocessing them; performing feature derivation and filtering on the preprocessed data to divide the dataset; constructing an XGBoost regression model and using the Grey Wolf optimization algorithm to search for hyperparameter combinations; validating the model using a test dataset, and then using the target model to predict epitaxial layer parameters.

[0039] Specifically, after collecting process parameters and epitaxial layer data, missing values ​​are processed through multiple interpolation, and Z-score standardization is used to eliminate dimensional differences. In the feature engineering stage, dynamic features such as the interaction term between temperature and flow rate, and the time derivative of the growth rate are introduced. Pearson correlation coefficients are used to select feature subsets strongly correlated with the target variable. When constructing the dual-task XGBoost model, the Grey Wolf algorithm searches for hyperparameters such as learning rate and tree depth within a preset range, and multi-fold cross-validation is used to evaluate the performance of parameter combinations. During model training, column sampling is used to randomly select feature subsets to enhance generalization ability, and second-order Taylor expansion approximates the influence of temperature on the diffusion coefficient, improving the fitting accuracy of the thermodynamic process. In the testing phase, extreme samples are used to verify the robustness of the model under process boundary conditions. Finally, the model can predict the epitaxial layer performance under different parameter combinations in real time.

[0040] Please see Figure 1 , Figure 1 A specific implementation of a method for predicting gallium oxide epitaxial layers is shown.

[0041] It should be noted that if substantially the same result is obtained, the method of this invention is not based on... Figure 1 Limited to the order of the processes shown, this method includes the following steps:

[0042] S1: Collect the process parameters of MOCVD and the corresponding epitaxial layer characterization data, and preprocess the process parameters and epitaxial layer characterization data to obtain preprocessed data.

[0043] Specifically, in the MOCVD epitaxial growth process, parameters such as growth rate, pressure, O / Ga ratio, and temperature, through nonlinear coupling, regulate the diffusion coefficient Dm (lateral thermal diffusion rate) and deposition coefficient Fn (vertical deposition rate), thereby determining the epitaxial film growth mode: when Dm > Fn, atoms diffuse sufficiently on the substrate surface to form two-dimensional layered growth (2D mode), and positive temperature regulation significantly enhances diffusion capability; while when Dm < Fn, deposition dominates to form three-dimensional island-like growth (3D mode), and this process is closely related to the nonlinear relationship between pressure and O / Ga ratio. Therefore, in this embodiment, it is necessary to collect process parameters such as temperature, pressure, O / Ga ratio, carrier gas flow rate, and reaction chamber height from the actual epitaxial experiment, and simultaneously collect the corresponding measurement results of the epitaxial experiment, i.e., collect epitaxial layer characterization data. The epitaxial layer characterization data includes the thickness of the epitaxial layer, ΔFWHM (half-maximum width at half maximum), and surface roughness. To ensure data quality, the process parameters and epitaxial layer characterization data need to be preprocessed to obtain preprocessed data.

[0044] The full width at half maximum (FWHM) variation refers to the difference between the full width at half maximum (FWHM) of the epitaxial layer and the substrate. It directly quantifies crystal defects. The larger the value, the more defects such as dislocations and stacking faults introduced by epitaxial growth.

[0045] Please see Figure 2 , Figure 2 A specific implementation of step S1 is shown below:

[0046] S11: Collect the process parameters of the MOCVD and the corresponding epitaxial layer characterization data;

[0047] S12: The missing values ​​of the process parameters and the epitaxial layer characterization data are processed by multiple interpolation to generate data after missing value processing;

[0048] S13: Perform outlier cleaning and data normalization on the data after processing the missing values ​​to generate the preprocessed data.

[0049] Specifically, in the MOCVD process parameter preprocessing, firstly, missing parameters such as growth temperature and carrier gas flow rate are reconstructed using multiple interpolation, generating a complete dataset that conforms to reaction kinetics by utilizing the physical correlations between parameters. Subsequently, a local outlier detection method is employed to identify and remove abnormal deposition rate data caused by transient pressure sensor failures. Finally, Z-score normalization converts parameters such as substrate rotation speed and precursor partial pressure into a standard distribution with a mean of 0 and a variance of 1, eliminating interference from parameters of different dimensions on model training. These three processing steps form a progressive data optimization process, maintaining the physical integrity of the original data while establishing a high-quality data foundation for subsequent feature engineering. This application effectively solves the data quality problem in the process parameter preprocessing stage. By reconstructing missing data through multiple interpolation to maintain the physical correlations between parameters, removing interference noise through outlier cleaning, and eliminating dimensional differences through data normalization, it provides high-fidelity input data for subsequent model training. This method significantly improves the accuracy of epitaxial layer thickness and full width at half maximum (FWHM) prediction models, ensuring that the process parameter optimization process is based on reliable data.

[0050] Multiple imputation refers to a method that handles missing data by establishing multiple imputation models. Specifically, it can be implemented using chain equations or Bayesian regression models. Multiple complete datasets are generated iteratively and then merged for analysis, preserving the nonlinear relationships between variables. Compared to single imputation, this method effectively avoids information bias caused by missing data, ensuring that the input data for subsequent model training conforms to real physical laws. Outlier cleaning refers to a method of identifying and removing outlier data points based on statistical distribution characteristics. Specifically, it can be implemented using threshold judgment based on interquartile range or outlier detection algorithms based on density clustering, such as using the DBSCAN algorithm to identify isolated points deviating from the main data cluster. This step can eliminate noise interference introduced by momentary equipment failures or measurement errors, ensuring the reliability of training data. Data normalization refers to a method of mapping parameters with different dimensions to a unified numerical range. Specifically, it can be implemented using min-max normalization or Z-score normalization, for example, linearly transforming temperature parameters from 300-800℃ to the 0-1 range. This processing can eliminate the imbalance in model weight allocation caused by differences in parameter units, improving the convergence efficiency of the gradient descent algorithm.

[0051] S2: Perform feature derivation and feature filtering on the preprocessed data to obtain feature-filtered data, and divide the feature-filtered data into datasets to generate training datasets and test datasets.

[0052] Specifically, feature engineering is used to derive and filter features from the preprocessed data to obtain feature-filtered data. The feature-filtered data is then divided into training and testing datasets.

[0053] Feature derivation refers to the creation of physical interaction terms and dynamic process variables. Specifically, the product of temperature gradient and gas flow rate can be used as a new feature to characterize thermodynamic coupling effects.

[0054] Please see Figure 3 , Figure 3 A specific implementation of step S2 is shown below:

[0055] S21: Based on the preprocessed data, create physical interaction features and add dynamic process features to generate derived feature data;

[0056] S22: Calculate the Pearson correlation coefficient matrix of all parameters in the derived feature data, and perform feature filtering based on the Pearson correlation coefficient matrix to generate the feature-filtered data;

[0057] S23: Divide the data after feature filtering into datasets according to a preset division ratio to generate the training dataset and the test dataset.

[0058] Specifically, based on the preprocessed data, the system first constructs physical interaction features by nonlinearly combining independent parameters such as temperature and gas velocity to generate new features that reflect the synergistic effects of these parameters. For example, multiplying the square of the temperature by the reciprocal of the gas velocity simulates the interaction between thermodynamics and fluid dynamics. Then, dynamic process features are added, such as calculating the difference in growth rate between adjacent time points to characterize the acceleration of the epitaxial layer thickness change. Next, Pearson correlation coefficients are calculated for all derived features, and features with a correlation below 0.3 with the target variable are filtered out by setting a threshold to eliminate redundant information. Finally, a stratified random sampling method is used to divide the filtered data into training and test sets in a 7:3 ratio to ensure that the two datasets maintain consistency in the distribution of thickness and full width at half maximum (FWHM).

[0059] Among them, physical interaction features refer to composite features generated by mathematically combining different process parameters. Specifically, this can be achieved by multiplying temperature and gas flow rate or taking a logarithmic transformation, used to characterize the nonlinear coupling relationship between parameters. Dynamic process features refer to features reflecting the time-series changes during epitaxial growth. Specifically, this can be achieved by using the derivative or integral operation of the growth rate over time, used to capture dynamic growth patterns. The Pearson correlation coefficient matrix is ​​a statistical indicator used to quantify the degree of linear correlation between each feature and the target variable. Specifically, this can be achieved by calculating the ratio of covariance to standard deviation, used to screen features that are significantly correlated with the target variable. The preset split ratio refers to the proportion parameter for dividing the dataset into training and test sets. Specifically, this can be achieved by random stratified sampling, used to ensure the balance of data distribution.

[0060] S3: Construct an XGBoost regression model, use the Grey Wolf optimization algorithm to search for hyperparameters based on the training dataset to obtain the target optimal parameter combination, and train the XGBoost regression model based on the target optimal parameter combination to generate the trained XGBoost regression model.

[0061] Specifically, an XGBoost regression model is constructed using the XGBoost algorithm. This XGBoost regression model is a dual-task regression tree structure, used to predict the thickness and full width at half maximum (FWHM) variation of the epitaxial layer, respectively. The gray wolf optimization algorithm is a metaheuristic algorithm that simulates the hunting behavior of gray wolf packs. Specifically, it can improve search efficiency by iteratively updating the positions of α, β, and δ wolves and searching for the optimal solution in the hyperparameter space.

[0062] Please see Figure 4 , Figure 4 A specific implementation of step S3 is shown below:

[0063] S31: Construct the XGBoost regression model, wherein the XGBoost regression model is a dual-task regression tree structure. The first task regression tree branch of the XGBoost regression model predicts thickness using the minimum mean square error function, and the second task regression tree branch of the XGBoost regression model predicts the full width at half maximum (FWHM) change using the mean absolute error function. S32: Use the Gray Wolf optimization algorithm to perform hyperparameter search based on the training dataset to obtain the target optimal parameter combination. S33: Iteratively train the XGBoost regression model based on the target optimal parameter combination. During the iterative training process, column sampling is used to randomly select a subset of features for each tree in the XGBoost regression model, and a second-order Taylor expansion is used to fit the exponential relationship between temperature and diffusion coefficient. S34: When the number of iterations reaches a preset number or the loss of the model generated in the current iteration is less than a preset threshold, stop the iterative training and generate the trained XGBoost regression model.

[0064] Specifically, in the model building stage, considering the physical differences in the variation of epitaxial layer thickness and full width at half maximum (FWHM), minimum mean square error (MMS) and mean absolute error (MAE) are designed as loss functions, respectively, and independent optimization of different error criteria is achieved through a dual-branch structure. In the hyperparameter optimization stage, the Grey Wolf algorithm simulates the wolf pack hierarchy and cooperative hunting mechanism, performing a global search within a preset parameter range. Its dynamic adjustment of the search step size effectively balances the exploration and development process. During model training, the column sampling strategy forces each tree to learn different feature combinations by randomly masking some features, thereby improving the model's robustness to noise features. Simultaneously, during model training, regularization terms (L1 / L2) are used to directly constrain the complexity of the tree structure, preventing overfitting to local process windows. For example, when an abnormal pressure value in an experiment causes a sudden change in Fn, the penalty mechanism can suppress the model's excessive response to that noise data. To address the nonlinear influence of temperature parameters on epitaxial layer quality, a higher-order derivative term of the temperature gradient is introduced into the loss function through a second-order Taylor expansion, enabling the model to accurately capture the exponential effect of temperature changes on the material diffusion rate. This application addresses the problems of accuracy imbalance, low efficiency of hyperparameter optimization, and insufficient dynamic process modeling in existing models during multi-task prediction.

[0065] The dual-task regression tree structure treats thickness prediction and full width at half maximum (FWHM) change prediction as independent regression tasks, achieving multi-objective joint learning through shared feature layers. Specifically, it can be implemented using a parallel tree structure with two loss function branches. This structure can address the issue of differing error distributions for different prediction objectives. The gray wolf optimization algorithm is a swarm intelligence algorithm simulating the hunting behavior of gray wolves. Specifically, it can use an encirclement-tracking mechanism guided by α, β, and δ wolves for hyperparameter space search. This algorithm overcomes the local optima trap of traditional grid search in high-dimensional parameter spaces. Column sampling involves randomly selecting some features for node splitting during the generation of each tree. This can be achieved using sampling without replacement, which reduces collinearity interference between features. Second-order Taylor expansion approximates the exponential function relationship between temperature and diffusion coefficient using a quadratic polynomial. This can be achieved by introducing a second derivative term in the loss function calculation. This method can accurately fit the nonlinear thermodynamic effects during epitaxial growth.

[0066] Please see Figure 5 , Figure 5 A specific implementation of step S32 is shown below:

[0067] S321: Randomly generate a preset number of hyperparameter combinations based on the training dataset;

[0068] S322: Perform multi-fold cross-validation on each set of hyperparameter combinations to generate mean squared error values;

[0069] S323: Update the hyperparameter combination based on the sorting of the mean squared error values, and re-perform multi-fold cross-validation based on the updated hyperparameter combination to iteratively generate new mean squared error values;

[0070] S324: Determine the optimal combination of parameters for the target based on the new mean square error value.

[0071] Specifically, within the Grey Wolf optimization algorithm framework, an initial population covering different regions of the parameter space is first generated through random sampling. Each hyperparameter combination undergoes a multi-fold cross-validation process, and the average mean squared error of each validation set is calculated as the fitness evaluation index. Individuals in the population are ranked according to their error values, retaining individuals with high fitness as leader wolves, while random mutation is introduced to generate new individuals to supplement the population. Through an iterative update mechanism, the search process gradually focuses on high-potential parameter regions while maintaining population diversity to prevent getting trapped in local optima. In each iteration, multi-fold cross-validation effectively reduces the evaluation bias caused by a single data partition, ensuring the stability of parameter optimization. Finally, after a preset number of iterations, the hyperparameter combination with the smallest validation error is selected as the global optimum.

[0072] The preset number of hyperparameter combinations refers to the initial set of candidate solutions randomly selected from the parameter space. This can be achieved using uniform distribution or Latin hypercube sampling methods, providing a diverse basis for subsequent optimization. Multi-fold cross-validation involves dividing the training data into multiple mutually exclusive subsets for iterative validation. This can be achieved using stratified sampling to maintain data distribution consistency and is used to evaluate the generalization performance of hyperparameter combinations. Mean squared error ranking update involves ranking candidate solutions based on the model validation results. This can be achieved using an elite retention strategy combined with a random perturbation mechanism, preserving superior solutions while avoiding premature convergence.

[0073] S4: Test the trained XGBoost regression model using the test dataset to generate the target XGBoost regression model.

[0074] Specifically, during the model validation phase, the accuracy of the predicted thickness and FWHM is verified by reserving a test set, ensuring that high generalization performance is maintained under different temperature ranges and extreme O / Ga ratio process windows.

[0075] Please see Figure 6 , Figure 6 A specific implementation of step 4 is shown below:

[0076] S41: The trained XGBoost regression model is used to predict the epitaxial layer based on the test dataset to obtain the test epitaxial layer thickness and the test full width at half maximum (FWHM) change.

[0077] S42: Calculate the verification index based on the thickness of the tested epitaxial layer and the change in the full width at half maximum (FWHM) of the tested epitaxial layer, and determine whether the test passes based on the verification index to obtain the judgment result. The verification index includes the coefficient of determination, the minimum root mean square error, and the mean absolute error.

[0078] S43: If the judgment result is that the test is passed, then an extreme special sample is constructed, and the trained XGBoost regression model is tested based on the extreme special sample. If the test is passed, the target XGBoost regression model is obtained.

[0079] Specifically, in the model testing phase, preliminary validation is first performed using a standard test set. The coefficient of determination is calculated to determine the model's ability to capture the trends of thickness and full width at half maximum (FWHM). Simultaneously, the root mean square error (RMSE) and mean absolute error (MAE) are used to analyze the dispersion of predicted values ​​and the distribution of error amplitude. When all three indicators reach preset thresholds, a specialized sample set containing extreme process parameters is further constructed. For example, the reaction chamber pressure is set to 1.5 times the standard operating condition, or the precursor flow rate reaches the critical value of the theoretical deposition rate. This sample set is used to test the model's predictive stability in the edge regions of the parameter space. If the model's validation indicators on extreme samples still meet the requirements, it is determined to possess the robustness required for industrial applications.

[0080] The test dataset refers to a subset of data selected from the feature-filtered data for model performance evaluation. This can be achieved by randomly splitting the feature-filtered data into training and test sets according to a preset ratio. Its purpose is to provide samples independent of the training process to verify the model's generalization ability. Validation metrics are mathematical evaluation standards used to quantify the model's predictive performance. Specifically, the coefficient of determination reflects the trend of agreement between predicted and true values; the root mean square error measures the overall level of prediction deviation; and the mean absolute error characterizes the absolute magnitude of the error. Their combined application can comprehensively evaluate the model's prediction accuracy under normal operating conditions. Extreme-specific samples refer to test data simulating extreme combinations of process parameters. These can be generated through parameter space boundary sampling or physical field simulation. Their purpose is to verify the model's stability under critical conditions. For example, samples can be constructed in a high-temperature, high-pressure region close to the gas-phase pre-reaction threshold to test the model's response to abnormal operating conditions.

[0081] The XGBoost algorithm used in this application for predictive intelligent optimization of epitaxial layers is reflected in the following aspects: (1) Key control factors are identified by ranking the importance of features. For example, it is found that temperature contributes 62% to Dm, while the weight of O / Ga ratio to Fn exceeds 45%, guiding the process engineer to prioritize the adjustment of core parameters. (2) Redundant parameter interference is eliminated by column subsampling. For example, when there is collinearity between carrier gas flow rate and pressure, random feature subset selection can enhance the robustness of the model. Combined with intelligent algorithms such as Grey Wolf Optimization (GWO), hyperparameter search is performed. Within the preset space of n_estimators (100-1000), max_depth (3-10), and learning_rate (0.01-0.3), the global optimal solution is found through 5-fold cross-validation, which improves the efficiency of process parameter recommendation by more than 3 times. This data-driven method has achieved the accuracy of 2D / 3D growth mode switching within ±5%, while reducing the optimization cycle of crystal quality from 20-30 experiments in the traditional method to 5-8 experiments.

[0082] S5: Obtain the process parameters to be tested in the MOCVD, and use the target XGBoost regression model to predict the epitaxial layer based on the process parameters to obtain the target epitaxial layer thickness and the target full width at half maximum (FWHM) variation.

[0083] Specifically, the above steps involve training and testing the XGBoost regression model to generate a target XGBoost regression model, which is then deployed. When prediction is required in practice, the process parameters to be measured in MOCVD are obtained, and the epitaxial layer is predicted based on these parameters using the target XGBoost regression model to obtain the target epitaxial layer thickness and the change in the target full width at half maximum (FWHM).

[0084] Please see Figure 7 , Figure 7 A specific implementation method following step S5 is shown below:

[0085] S61: Store the process parameters to be tested, the target epitaxial layer thickness, and the target full width at half maximum (FWHM) variation in a historical database;

[0086] S62: When the number of predictions of the target XGBoost regression model reaches the preset number of predictions, the target XGBoost regression model is dynamically and incrementally updated using data from the historical database to obtain the updated XGBoost regression model.

[0087] Specifically, after each epitaxial layer prediction, process parameters, predicted thickness, and full width at half maximum (FWHM) data are automatically written to the historical database, forming a timestamped process record. When the cumulative number of predictions reaches a preset threshold, the system extracts new data from the database and uses an incremental learning algorithm to iteratively update the weights of the leaf nodes of the XGBoost regression model. During this process, the model retains its original tree structure, adjusting only the node splitting threshold and weight allocation, thus avoiding overfitting by limiting the magnitude of parameter adjustments in each update. Simultaneously, importance sampling techniques are used to select historical data highly correlated with the current process state for training, reducing redundant computation.

[0088] The historical database refers to a structured data storage system used to store process parameters, epitaxial layer thickness, and full width at half maximum (FWHM) variations. It can be implemented using a time-series database or a relational database, accumulating process data that changes over time to provide a data foundation for model updates. Dynamic incremental updates refer to adjusting model parameters based solely on newly added data without retraining all historical data. This can be achieved using online gradient descent algorithms or sliding window sampling methods, reducing computational resource consumption through local parameter optimization while maintaining the model's adaptability to process drift. The preset prediction count refers to the threshold condition for triggering model updates. This can be set based on process stability or equipment maintenance cycles; for example, initiating the update process after every 10 predictions to ensure the model update frequency matches the rate of process change.

[0089] Please see Figures 8 to 10 , Figure 8 This is a schematic diagram showing the prediction and fitting effect of the model on the changes in thickness and full width at half maximum (FWHM) in the training dataset after using the XGBoost algorithm. Figure 9 This is the prediction fit of the model on the test dataset after using the XGBoost algorithm for the changes in thickness and full width at half maximum (FWHM). Figure 10 A schematic diagram of the high-quality epitaxial layer prepared under the guidance of the target XGBoost regression model.

[0090] This application's embodiments utilize the XGBoost algorithm to achieve intelligent analysis and prediction of MOCVD process parameters, and its technical advantages have been fully verified in the prediction of film thickness and crystal quality. For example... Figure 8 and Figure 9 As shown, in one specific embodiment, the training set accuracy for film thickness prediction reaches R0. 2 (Coefficient of determination) = 0.995 (RMSE (Root Mean Square Error) = 22.903, MAE (Mean Absolute Error) = 17.059), test set hold R 2 =0.978 (RMSE=54.309, MAE=36.517), proving that the model accurately captures the physical laws of growth rate; ΔFWHM prediction achieves R on the training set. 2 =0.922 (RMSE=9.410, MAE=6.918), test set R 2 =0.859 and RMSE=13.539, MAE=9.887, significantly better than the traditional empirical model under the same working conditions. 2 =0.282, RMSE=30.547, MAE=20.557. Particularly in tests within extreme process windows (temperature >1050℃, O / Ga ratio <0.2), the film thickness prediction error was reduced by 82.7% compared to the manual trial-and-error method (traditional method test set R). 2=0.842, RMSE=144.681, MAE=94.231), while ΔFWHM's MAE=7.816 is 63.4% higher than the traditional method. This is due to the model's precise analysis of the exponential control law of temperature on Dm (Dm ∝ exp(-Ea / kT)) and the interaction effect of pressure-O / Ga ratio through second-order Taylor expansion, which enhances process stability and provides an intelligent solution for semiconductor epitaxial manufacturing that combines theoretical depth and engineering reliability.

[0091] Through continuous model iteration and database expansion, the model's prediction accuracy for extrapolated results has become increasingly higher. Ultimately, guided by the model and based on historical data, the model achieved the following results in just five iterations: Figure 10 The high-quality homoepitaxial β-Ga2O3 layer shown has a rocking curve full width at half maximum (FWHM) of only 75.6 arcsec and a surface roughness of only 0.619 nm, and exhibits a clear two-dimensional growth morphology, which greatly reduces the cost of process development.

[0092] In this embodiment, MOCVD process parameters and corresponding epitaxial layer characterization data are collected, and the process parameters and epitaxial layer characterization data are preprocessed to obtain preprocessed data. Feature derivation and feature filtering are performed on the preprocessed data to obtain feature-filtered data. The feature-filtered data is then divided into training and testing datasets. An XGBoost regression model is constructed, and the Grey Wolf optimization algorithm is used to search for hyperparameters based on the training dataset to obtain the target optimal parameter combination. The XGBoost regression model is trained based on the target optimal parameter combination to generate a trained XGBoost regression model. The trained XGBoost regression model is tested using the testing dataset to generate a target XGBoost regression model. The MOCVD process parameters to be tested are obtained, and the target XGBoost regression model is used to predict the epitaxial layer thickness and the target full width at half maximum (FWHM) variation based on the process parameters to be tested, thus obtaining the target epitaxial layer thickness and the target FWHM variation. This invention constructs an XGBoost regression model combined with the Grey Wolf optimization algorithm for hyperparameter search and handles the complex relationships between process parameters. It can accurately predict the changes in epitaxial layer thickness and full width at half maximum (FWHM). It has the advantages of improving process optimization efficiency, enhancing model generalization ability, and realizing dynamic updates, thereby improving the prediction accuracy of gallium oxide epitaxial layers.

[0093] This application achieves accurate prediction and process parameter optimization of the epitaxial layer growth process. Data preprocessing and feature engineering effectively extract key influencing factors, solving the challenge of dimensionality reduction in high-dimensional parameter spaces. The optimized XGBoost model simultaneously predicts changes in thickness and full width at half maximum (FWHM), providing dual constraints for process adjustments. The dynamic incremental update mechanism enables the model to continuously adapt to changes in equipment status, ensuring long-term prediction reliability. This method significantly improves the uniformity and defect suppression capability of epitaxial layers, laying the foundation for the fabrication of high-performance gallium oxide devices.

[0094] Please refer to Figure 11 As a response to the above Figure 1 The implementation of the method shown in this application provides an embodiment of a prediction device for gallium oxide epitaxial layers, which is similar to... Figure 1 Corresponding to the method embodiments shown, this device can be specifically applied to various electronic devices.

[0095] like Figure 11 As shown, the prediction device for gallium oxide epitaxial layers in this embodiment includes: a data acquisition module 71, a data processing module 72, a model training module 73, a model testing module 74, and an epitaxial layer prediction module 75, wherein:

[0096] The data acquisition module 71 is used to acquire the process parameters of MOCVD and the corresponding epitaxial layer characterization data, and to preprocess the process parameters and epitaxial layer characterization data to obtain preprocessed data.

[0097] The data processing module 72 is used to perform feature derivation and feature filtering on the preprocessed data to obtain feature-filtered data, and to divide the feature-filtered data into datasets to generate training datasets and test datasets.

[0098] The model training module 73 is used to construct an XGBoost regression model. It uses the Grey Wolf optimization algorithm to perform hyperparameter search based on the training dataset to obtain the target optimal parameter combination, and trains the XGBoost regression model based on the target optimal parameter combination to generate a trained XGBoost regression model.

[0099] Model testing module 74 is used to test the trained XGBoost regression model using a test dataset and generate a target XGBoost regression model.

[0100] The epitaxial layer prediction module 75 is used to obtain the process parameters to be tested in the MOCVD, and to predict the epitaxial layer based on the process parameters to be tested using the target XGBoost regression model, thereby obtaining the target epitaxial layer thickness and the change in the target full width at half maximum (FWHM).

[0101] Furthermore, the model training module 73 includes:

[0102] The model building unit is used to build the XGBoost regression model, wherein the XGBoost regression model is a dual-task regression tree structure, the first task regression tree branch of the XGBoost regression model predicts the thickness using the minimum mean square error function, and the second task regression tree branch of the XGBoost regression model predicts the full width at half maximum (FWHM) change using the mean absolute error function.

[0103] The hyperparameter search unit is used to perform hyperparameter search based on the training dataset using the Grey Wolf optimization algorithm to obtain the optimal combination of parameters for the target.

[0104] The training unit is used to iteratively train the XGBoost regression model based on the target optimal parameter combination. During the iterative training process, column sampling is used to randomly select a subset of features for each tree in the XGBoost regression model, and a second-order Taylor expansion is used to fit the exponential relationship between temperature and diffusion coefficient.

[0105] The model generation unit is used to stop iterative training and generate the trained XGBoost regression model when the number of iterative training iterations reaches a preset number or the loss of the model generated in the current iteration is less than a preset threshold.

[0106] Furthermore, the hyperparameter search unit includes:

[0107] The hyperparameter combination generation unit is used to randomly generate a preset number of hyperparameter combinations based on the training dataset.

[0108] The mean squared error value generation unit is used to perform multi-fold cross-validation on each set of hyperparameter combinations to generate a mean squared error value.

[0109] The hyperparameter combination update unit is used to update the hyperparameter combination based on the sorting of the mean square error values, and to re-perform multi-fold cross-validation based on the updated hyperparameter combination to iteratively generate new mean square error values.

[0110] The optimal combination of parameters for the target is determined based on the new mean square error value.

[0111] Furthermore, the model testing module 74 includes:

[0112] The test metric generation unit is used to predict the epitaxial layer based on the test dataset using the trained XGBoost regression model to obtain the test epitaxial layer thickness and the test full width at half maximum (FWHM) change.

[0113] The judgment result generation unit is used to calculate the verification index based on the test epitaxial layer thickness and the test full width at half maximum variation, and to judge whether the test passes based on the verification index to obtain the judgment result. The verification index includes the coefficient of determination, the minimum root mean square error, and the mean absolute error.

[0114] The target model generation unit is used to construct extreme specialized samples if the judgment result is that the test is passed, and to perform model testing on the trained XGBoost regression model based on the extreme specialized samples. If the test is passed, the target XGBoost regression model is obtained.

[0115] Furthermore, the data acquisition module 71 includes:

[0116] Epitaxial layer characterization data acquisition unit, used to acquire the process parameters of the MOCVD and the corresponding epitaxial layer characterization data;

[0117] The missing value processing unit is used to process the missing values ​​of the process parameters and the epitaxial layer characterization data using a multiple interpolation method, and generate data after missing value processing.

[0118] The data normalization unit is used to perform outlier cleaning and data normalization on the data after missing value processing to generate the preprocessed data.

[0119] Furthermore, the data processing module 72 includes:

[0120] A derived feature data generation unit is used to create physical interaction features and add dynamic process features based on the preprocessed data to generate derived feature data.

[0121] The feature filtering unit is used to calculate the Pearson correlation coefficient matrix of all parameters in the derived feature data, and perform feature filtering based on the Pearson correlation coefficient matrix to generate the feature-filtered data.

[0122] The data partitioning unit is used to partition the data after feature filtering according to a preset partitioning ratio to generate the training dataset and the test dataset.

[0123] Furthermore, the epitaxial layer prediction module 75 also includes:

[0124] A data storage unit is used to store the process parameters to be measured, the target epitaxial layer thickness, and the target full width at half maximum (FWHM) variation in a historical database.

[0125] The dynamic incremental update unit is used to dynamically and incrementally update the target XGBoost regression model using data from the historical database when the number of predictions of the target XGBoost regression model reaches a preset number of predictions, so as to obtain the updated XGBoost regression model.

[0126] To address the aforementioned technical problems, embodiments of this application also provide a computer device. Please refer to [link / reference needed]. Figure 12 , Figure 12 This is a basic structural block diagram of the computer device in this embodiment.

[0127] Computer device 8 includes a memory 81, a processor 82, and a network interface 83 that are interconnected via a system bus. It should be noted that... Figure 12 Only a computer device 8 with three components—memory 81, processor 82, and network interface 83—is shown. It should be understood that implementing all shown components is not required; more or fewer components may be implemented alternatively. Those skilled in the art will understand that this computer device is a device capable of automatically performing numerical calculations and / or information processing according to pre-set or stored instructions. Its hardware includes, but is not limited to, microprocessors, application-specific integrated circuits (ASICs), field-programmable gate arrays (FPGAs), digital signal processors (DSPs), and embedded devices.

[0128] Computer devices can include desktop computers, laptops, handheld computers, and cloud servers. These devices allow for human-computer interaction with users through keyboards, mice, remote controls, touchpads, or voice-activated devices.

[0129] The memory 81 includes at least one type of readable storage medium, including flash memory, hard disk, multimedia card, card-type memory (e.g., SD or DX memory), random access memory (RAM), static random access memory (SRAM), read-only memory (ROM), electrically erasable programmable read-only memory (EEPROM), programmable read-only memory (PROM), magnetic memory, magnetic disk, optical disk, etc. In some embodiments, the memory 81 may be an internal storage unit of the computer device 8, such as the hard disk or memory of the computer device 8. In other embodiments, the memory 81 may also be an external storage device of the computer device 8, such as a plug-in hard disk, smart media card (SMC), secure digital (SD) card, flash card, etc., equipped on the computer device 8. Of course, the memory 81 may include both internal storage units and external storage devices of the computer device 8. In this embodiment, the memory 81 is typically used to store the operating system and various application software installed on the computer device 8, such as program code for a method for predicting gallium oxide epitaxial layers. In addition, the memory 81 may also be used to temporarily store various types of data that have been output or will be output.

[0130] In some embodiments, processor 82 may be a central processing unit (CPU), controller, microcontroller, microprocessor, or other data processing chip. This processor 82 is typically used to control the overall operation of the computer device 8. In this embodiment, processor 82 is used to run program code stored in memory 81 or process data, for example, to run the program code of the gallium oxide epitaxial layer prediction method described above, to implement various embodiments of the gallium oxide epitaxial layer prediction method.

[0131] The network interface 83 may include a wireless network interface or a wired network interface, which is typically used to establish a communication connection between the computer device 8 and other electronic devices.

[0132] This application also provides another embodiment, namely, a computer-readable storage medium storing a computer program that can be executed by at least one processor to cause the at least one processor to perform the steps of the gallium oxide epitaxial layer prediction method described above.

[0133] Through the above description of the embodiments, those skilled in the art can clearly understand that the methods of the above embodiments can be implemented by means of software plus necessary general-purpose hardware platforms. Of course, they can also be implemented by hardware, but in many cases the former is a better implementation method. Based on this understanding, the technical solution of this application, in essence, or the part that contributes to the prior art, can be embodied in the form of a software product. This computer software product is stored in a storage medium (such as ROM / RAM, magnetic disk, optical disk) and includes several instructions to cause a terminal device (which may be a mobile phone, computer, server, air conditioner, or network device, etc.) to execute the methods of the various embodiments of this application.

[0134] Obviously, the embodiments described above are merely some embodiments of this application, not all embodiments. The accompanying drawings show preferred embodiments of this application, but do not limit the scope of this application. This application can be implemented in many different forms; rather, these embodiments are provided to provide a more thorough and comprehensive understanding of the disclosure of this application. Although this application 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 specific embodiments, or make equivalent substitutions for some of the technical features. Any equivalent structures made using the content of this application's specification and drawings, directly or indirectly applied to other related technical fields, are similarly within the scope of protection of this application.

Claims

1. A method of predicting an epitaxial layer of gallium oxide, characterized by, The method comprises the following steps: Collecting process parameters of MOCVD and corresponding epitaxial layer characterization data, and preprocessing the process parameters and the epitaxial layer characterization data to obtain preprocessed data; Deriving and screening features from the preprocessed data to obtain screened feature data, and dividing the screened feature data into a training data set and a test data set; Building an XGBoost regression model, searching for hyperparameters based on the training data set using a grey wolf optimization algorithm to obtain a target optimal parameter combination, training the XGBoost regression model based on the target optimal parameter combination to generate a trained XGBoost regression model; Model testing of the trained XGBoost regression model through the test data set to generate a target XGBoost regression model; Obtaining the process parameters to be measured of the MOCVD, and predicting the epitaxial layer based on the process parameters to be measured through the target XGBoost regression model to obtain a target epitaxial layer thickness and a target full width at half maximum change; The method of building an XGBoost regression model, searching for hyperparameters based on the training data set using a grey wolf optimization algorithm to obtain a target optimal parameter combination, and training the XGBoost regression model based on the target optimal parameter combination to generate a trained XGBoost regression model comprises: Building the XGBoost regression model, wherein the XGBoost regression model is a double-task regression tree structure, the first task regression tree branch of the XGBoost regression model performs thickness prediction through a least mean square error function, and the second task regression tree branch of the XGBoost regression model performs full width at half maximum change prediction through an average absolute error function; Searching for hyperparameters based on the training data set using the grey wolf optimization algorithm to obtain the target optimal parameter combination; Iteratively training the XGBoost regression model based on the target optimal parameter combination, and in the iterative training process, randomly selecting a feature subset for each tree in the XGBoost regression model using column sampling, and fitting the exponential relationship between temperature and diffusion coefficient through a second-order Taylor expansion; When the number of iterations reaches a preset number or the loss of the model generated in the current iteration is less than a preset threshold, stopping the iterative training to generate the trained XGBoost regression model.

2. The method of predicting a gallium oxide epitaxial layer according to claim 1, characterized by, The method of searching for hyperparameters based on the training data set using the grey wolf optimization algorithm to obtain the target optimal parameter combination comprises: Randomly generating a preset number of hyperparameter combinations based on the training data set; Performing multi-fold cross-validation on each of the hyperparameter combinations to generate a mean square error value; Updating the hyperparameter combinations based on the ranking of the mean square error values, and re-performing multi-fold cross-validation based on the updated hyperparameter combinations to iteratively generate a new mean square error value; Determining the target optimal parameter combination based on the new mean square error value.

3. The method of claim 1, wherein the method is a method of predicting a gallium oxide epitaxial layer, characterized by: The model testing on the trained XGBoost regression model through the test data set generates a target XGBoost regression model, including: Performing epitaxial layer prediction based on the test data set through the trained XGBoost regression model to obtain test epitaxial layer thickness and test full width at half maximum change; Calculating a verification index based on the test epitaxial layer thickness and the test full width at half maximum change, and determining whether the test passes based on the verification index to obtain a determination result, wherein the verification index includes a determination coefficient, a minimum root mean square error, and an average absolute error; If the determination result is that the test passes, an extreme special sample is constructed, and the trained XGBoost regression model is tested based on the extreme special sample, and if the test passes, the target XGBoost regression model is obtained.

4. The method of claim 1, wherein the method is a method of predicting a gallium oxide epitaxial layer, characterized by: The process parameters of the MOCVD and the corresponding epitaxial layer characterization data are collected, and the process parameters and the epitaxial layer characterization data are preprocessed to obtain preprocessed data, including: Collecting the process parameters of the MOCVD and the corresponding epitaxial layer characterization data; Performing missing value processing on the process parameters and the epitaxial layer characterization data using a multiple imputation method to generate missing value processed data; Performing outlier cleaning and data normalization on the missing value processed data to generate the preprocessed data.

5. The method of claim 1, wherein the method is a method of predicting a gallium oxide epitaxial layer, characterized by: The preprocessed data is subjected to feature derivation and feature screening to obtain feature screened data, and the feature screened data is subjected to data set division to generate a training data set and a test data set, including: Creating physical interaction features and adding dynamic process features based on the preprocessed data to generate derived feature data; Calculating a Pearson correlation coefficient matrix of all parameters in the derived feature data, and performing feature filtering based on the Pearson correlation coefficient matrix to generate the feature screened data; The feature screened data is divided into data sets according to a preset division ratio to generate the training data set and the test data set.

6. The method of predicting a gallium oxide epitaxial layer according to any one of claims 1 to 5, characterized by, After obtaining the to-be-tested process parameters of the MOCVD and performing epitaxial layer prediction based on the to-be-tested process parameters through the target XGBoost regression model to obtain target epitaxial layer thickness and target full width at half maximum change, the method further includes: Storing the to-be-tested process parameters, the target epitaxial layer thickness, and the target full width at half maximum change in a historical database; When the number of predictions of the target XGBoost regression model reaches a preset number of predictions, dynamically updating the target XGBoost regression model based on the data in the historical database to obtain an updated XGBoost regression model.

7. A prediction device of a gallium oxide epitaxial layer, characterized by comprising: including: A data collection module for collecting process parameters of a MOCVD and corresponding epitaxial layer characterization data, and preprocessing the process parameters and the epitaxial layer characterization data to obtain preprocessed data; The data processing module is configured to derive features from the preprocessed data, screen the features, obtain screened feature data, and divide the screened feature data into a training data set and a test data set; The model training module is configured to construct an XGBoost regression model, search for hyperparameters based on the training data set using a grey wolf optimization algorithm, obtain a target optimal parameter combination, train the XGBoost regression model based on the target optimal parameter combination, and generate a trained XGBoost regression model; The model testing module is configured to test the trained XGBoost regression model based on the test data set, and generate a target XGBoost regression model; The epitaxial layer prediction module is configured to obtain process parameters to be tested of the MOCVD, and predict an epitaxial layer based on the process parameters to be tested using the target XGBoost regression model, to obtain a target epitaxial layer thickness and a target full width at half maximum change; The model training module comprises: The model construction unit is configured to construct the XGBoost regression model, wherein the XGBoost regression model is a double-task regression tree structure, a first task regression tree branch of the XGBoost regression model is used to predict a thickness by a least mean square error function, and a second task regression tree branch of the XGBoost regression model is used to predict a full width at half maximum change by an average absolute error function; The hyperparameter searching unit is configured to search for hyperparameters based on the training data set using the grey wolf optimization algorithm, to obtain the target optimal parameter combination; The training unit is configured to iteratively train the XGBoost regression model based on the target optimal parameter combination, and in the iterative training process, randomly select a feature subset for each tree in the XGBoost regression model by column sampling, and fit an exponential relationship between a diffusion coefficient and a temperature by a second-order Taylor expansion; The model generation unit is configured to stop the iterative training when a number of iterations reaches a preset number or a model loss generated in a current iteration is less than a preset threshold, and generate the trained XGBoost regression model.

8. A computer device, comprising: A memory and a processor are included, the memory stores a computer program, and the processor implements the prediction method of the gallium oxide epitaxial layer in any one of claims 1 to 6 when executing the computer program.

9. A computer-readable storage medium, characterized in that, A computer program is stored on the computer readable storage medium, and the computer program is executed by the processor to implement the prediction method of the gallium oxide epitaxial layer in any one of claims 1 to 6.

Images