A model linkage-based gallium oxide epitaxial control method, device, equipment and medium

Abstract

The application discloses a gallium oxide epitaxy control method and device based on model linkage, equipment and medium. The method comprises the following steps: analyzing an epitaxy task intention in a natural language instruction interacted by an operator through a large model, and calling a target gallium oxide epitaxy model in a preset gallium oxide epitaxy model library based on the epitaxy task intention; acquiring an epitaxy parameter of gallium oxide epitaxy performed in an epitaxy device; determining a gallium oxide epitaxy control parameter based on the epitaxy parameter through the target gallium oxide epitaxy model; and controlling the epitaxy device based on a gallium source flow adjustment parameter to perform closed-loop control on the gallium oxide epitaxy. The application analyzes the epitaxy task intention in the natural language instruction interacted by the operator through the large model, selects the target gallium oxide epitaxy model established through subdivided working conditions according to the epitaxy task intention, predicts the epitaxy parameter through the target gallium oxide epitaxy model, and finally controls the epitaxy device based on the epitaxy parameter, so that the model dynamic matching and calling under the natural language interaction are realized, the gallium oxide epitaxy model does not need to be manually adapted by engineers, the dependence on artificial experience is reduced, and various epitaxy growth scenes, such as different substrate orientations and different doping types, can be adapted.

Description

Technical Field

[0001] This application relates to the field of semiconductor thin film epitaxial growth technology, and in particular to a gallium oxide epitaxial control method, apparatus, device and medium based on model linkage. Background Technology

[0002] Gallium oxide, as a fourth-generation ultrawide bandgap semiconductor material, has significant application value in fields such as ultra-high voltage and high-power power electronic devices and deep ultraviolet photodetectors. In order to manufacture high-performance Schottky barrier diodes (SBDs), metal-oxide-semiconductor field-effect transistors (MOSFETs), and other devices, it is necessary to epitaxially grow a gallium oxide epitaxial layer on a high-quality gallium oxide single crystal substrate with precise control over its thickness, doping concentration, and crystal quality.

[0003] Existing gallium oxide epitaxial growth methods (such as metal-organic chemical vapor deposition (MOCVD), halide vapor phase epitaxy (HVPE), molecular beam epitaxy (MBE), and mist chemical vapor deposition (Mist-CVD)) generally achieve multi-parameter control through a combination of physical models and PID control. However, the multiphase gas-phase reaction and thermodynamics are highly coupled during gallium oxide epitaxy, making it difficult to simultaneously achieve optimal control of all parameters using physical models and PID control methods. Furthermore, differences in parameters such as substrate orientation and doping type significantly alter the epitaxial growth window, making these control methods unsuitable for all growth scenarios and highly dependent on engineers' trial-and-error experience.

[0004] Therefore, the existing technology still needs to be improved and enhanced. Summary of the Invention

[0005] The technical problem to be solved by this application is to provide a gallium oxide epitaxial control method, device, equipment and medium based on model linkage, which addresses the shortcomings of the existing technology.

[0006] To address the aforementioned technical problems, the first aspect of this application provides a model-linked gallium oxide epitaxial control method, wherein the model-linked gallium oxide epitaxial control method specifically includes: The model parses the extended task intent in the natural language commands of the operator interaction, and calls the target gallium oxide epitaxial model in the preset gallium oxide epitaxial model library based on the extended task intent. Obtain the epitaxial parameters of gallium oxide epitaxy performed in the epitaxial device, wherein the epitaxial parameters include relative time, relative position, process relative deviation, and several historical gallium oxide epitaxial control parameters; Gallium oxide epitaxial control parameters are determined based on the epitaxial parameters using the target gallium oxide epitaxial model, wherein the gallium oxide epitaxial control parameters include at least gallium source flow rate adjustment parameters; The gallium source flow rate adjustment parameters are used to control the epitaxial device, thereby enabling closed-loop control of the gallium oxide epitaxy.

[0007] The model-linked gallium oxide epitaxial control method described herein includes at least one of the following epitaxial task objectives: epitaxial growth method, substrate surface type, epitaxial parameters, and doping type.

[0008] The aforementioned model-linked gallium oxide epitaxial control method, wherein calling the target gallium oxide epitaxial model from the preset gallium oxide epitaxial model library based on the epitaxial task intent specifically includes: Determine the matching degree between the epitaxial task intent and the epitaxial scenario of each gallium oxide epitaxial model in the preset gallium oxide epitaxial model library; Based on the matching degree, a target gallium oxide epitaxial model is selected from the preset gallium oxide epitaxial model library; The model configuration file of the target gallium oxide epitaxial model is read to dynamically invoke the target gallium oxide epitaxial model.

[0009] The model-linked gallium oxide epitaxial control method further includes, after obtaining the epitaxial parameters of the gallium oxide epitaxy performed in the epitaxial device, the method further includes: The process of changing the gallium oxide epitaxial model is detected. The changing operation is performed through a drop-down menu or by browsing a path. The new gallium oxide epitaxial model corresponding to the replacement operation is invoked, and the target gallium oxide epitaxial model is replaced with the new gallium oxide epitaxial model.

[0010] The gallium oxide epitaxial control method based on model linkage further includes real-time epitaxial growth rate and step-doped atom adsorption efficiency as gallium oxide epitaxial control parameters.

[0011] The aforementioned model-linked gallium oxide epitaxial control method, wherein the training process and / or offline tuning of each gallium oxide epitaxial model in the preset gallium oxide epitaxial model library includes: The training dataset is used to train the gallium oxide epitaxial model to be trained to obtain the intermediate gallium oxide epitaxial model, and the training log corresponding to the intermediate gallium oxide epitaxial model is obtained. The training log includes training loss, validation set determination coefficient, mean square error and feature parameters. The large model performs deep inference on the training logs to obtain the model state of the gallium oxide epitaxial model, and the gallium oxide epitaxial model is adjusted based on the model state. The process of retraining the untrained gallium oxide epitaxial model using the training dataset to obtain an intermediate gallium oxide epitaxial model continues until a preset termination condition is met. The intermediate gallium oxide epitaxial model that meets the preset termination condition is then used as the trained gallium oxide epitaxial model.

[0012] The aforementioned model-linked gallium oxide epitaxial control method, wherein the step of performing deep inference on the training logs using a large model to obtain the model state of the gallium oxide epitaxial model specifically includes: Obtain the epitaxial scene corresponding to the gallium oxide epitaxial model, and obtain the model evaluation index based on the epitaxial scene; Select the model index value corresponding to the model evaluation index from the training log, and determine the model state of the gallium oxide epitaxial model based on the model index value; The correspondence between the model evaluation metrics and the extended scenarios is as follows: When the epitaxial scenario is of the highly doped type, the model evaluation metric is the validation set determination coefficient; When the epitaxial scenario is unintentionally doped, the model evaluation metric is the feature parameters.

[0013] The model-linked gallium oxide epitaxial control method described herein includes, during the training and / or offline tuning process of each gallium oxide epitaxial model in the preset gallium oxide epitaxial model library, data cleaning is performed on the training dataset used for training and / or offline tuning. The data cleaning includes one or more of the following: Dorajal outlier removal, Kalman filter smoothing, multiple imputation missing value processing, gas switching transient time alignment, and derived feature generation.

[0014] The aforementioned model-linked gallium oxide epitaxial control method, wherein controlling the epitaxial device based on the gallium source flow adjustment parameters to achieve closed-loop control of the gallium oxide epitaxy specifically includes: The gallium source flow rate adjustment parameters are used as prior knowledge, and the actual gallium source flow rate of several preceding sampling steps is obtained. Based on the prior knowledge and the actual gallium source flow rate of the preceding sampling steps, the target gallium source flow rate adjustment parameters are determined using the Bayesian fluctuation suppression method. The target gallium source flow rate adjustment parameters are sent to the controller corresponding to the epitaxial device, and the controller controls the epitaxial device to perform closed-loop control of the gallium oxide epitaxy.

[0015] A second aspect of this application provides a model-linked gallium oxide epitaxial control device, wherein the model-linked gallium oxide epitaxial control device specifically includes: The intent parsing module is used to parse the extended task intent in the natural language instructions of the operator interaction through a large model, and to call the target gallium oxide epitaxial model in the preset gallium oxide epitaxial model library based on the extended task intent. The acquisition module is used to acquire the epitaxial parameters of gallium oxide epitaxy performed in the epitaxial device, wherein the epitaxial parameters include relative time, relative position, process relative deviation, and several historical gallium oxide epitaxial control parameters; The prediction module is used to determine gallium oxide epitaxial control parameters based on the epitaxial parameters using the target gallium oxide epitaxial model, wherein the gallium oxide epitaxial control parameters include at least gallium source flow rate adjustment parameters; The control module is used to control the epitaxial device based on the gallium source flow adjustment parameters, so as to perform closed-loop control of the gallium oxide epitaxy.

[0016] A third aspect of this application provides a computer-readable storage medium storing one or more programs that can be executed by one or more processors to implement the steps in the model-linked gallium oxide epitaxial control method described above.

[0017] A fourth aspect of this application provides a terminal device, which includes: a processor and a memory; The memory stores a computer-readable program that can be executed by the processor; When the processor executes the computer-readable program, it implements the steps in any of the above-described model-linked gallium oxide epitaxial control methods.

[0018] Beneficial effects: 1. This application uses a large model to parse the epitaxial task intent in the natural language commands of operator interaction, and selects a target gallium oxide epitaxial model established through detailed operating conditions based on the epitaxial task intent. Then, it predicts epitaxial parameters using the target gallium oxide epitaxial model, and finally controls the epitaxial equipment based on these parameters. This achieves dynamic model matching and invocation under natural language interaction, eliminating the need for engineers to manually adapt gallium oxide epitaxial models, reducing reliance on human experience, and adapting to various epitaxial growth scenarios such as different substrate orientations and different doping types. Simultaneously, through high-precision forward prediction of the target gallium oxide epitaxial model, optimal control of multiple indicators such as epitaxial resistivity, thickness, and mobility can be achieved simultaneously.

[0019] 2. This application enables interaction between the large model and the operator, allowing the operator to issue complex extended tasks through natural language commands, which greatly reduces the threshold for process development.

[0020] 3. This application introduces a large model to perform deep inference on the training logs during model training and offline tuning. It matches corresponding evaluation indicators to determine the model state and adjusts the model parameters according to different epitaxial scenarios, thereby improving the prediction accuracy of the trained gallium oxide epitaxial model in the corresponding scenarios and providing more accurate control parameters for epitaxial control.

[0021] 4. This application uses the Bayesian fluctuation suppression method to determine the final target gallium source flow rate adjustment parameters, which can smooth the precursor flow rate and temperature field fluctuations, effectively eliminate the widening of the interface transition region caused by gas phase transport lag, realize the adaptive control of the atomic-level ultra-steep epitaxial interface, and ensure that the thickness and doping uniformity of the epitaxial layer meet the preparation requirements. Attached Figure Description

[0022] To more clearly illustrate the technical solutions in the embodiments of this application, the accompanying drawings used in the description of the embodiments will be briefly introduced below. Obviously, the accompanying drawings described below are only some embodiments of this application. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0023] Figure 1 This is a schematic diagram of the system used in the model-linked gallium oxide epitaxial control method provided in the embodiments of this application.

[0024] Figure 2 A flowchart of a model-linked gallium oxide epitaxial control method provided in an embodiment of this application.

[0025] Figure 3 This is a schematic diagram of a model-linked gallium oxide epitaxial control device provided in an embodiment of this application.

[0026] Figure 4 A schematic block diagram of the terminal device provided in the embodiments of this application. Detailed Implementation

[0027] This application provides a model-linked gallium oxide epitaxial control method, apparatus, device, and medium. To make the objectives, technical solutions, and effects of this application clearer and more explicit, the following detailed description is provided with reference to the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are only for explaining this application and are not intended to limit this application.

[0028] Those skilled in the art will understand that, unless specifically stated otherwise, the singular forms “a,” “an,” “the,” and “the” used herein may also include the plural forms. It should be further understood that the term “comprising” as used in this application means the presence of the stated features, integers, steps, operations, elements, and / or components, but does not exclude the presence or addition of one or more other features, integers, steps, operations, elements, components, and / or groups thereof. It should be understood that when we say an element is “connected” or “coupled” to another element, it can be directly connected or coupled to the other element, or there may be intermediate elements. Furthermore, “connected” or “coupled” as used herein can include wireless connections or wireless coupling. The term “and / or” as used herein includes all or any units and all combinations of one or more associated listed items.

[0029] It will be understood by those skilled in the art that, unless otherwise defined, all terms used herein (including technical and scientific terms) have the same meaning as commonly understood by one of ordinary skill in the art to which this application pertains. It should also be understood that terms such as those defined in general dictionaries should be understood to have the same meaning as in the context of the prior art, and should not be interpreted in an idealized or overly formal sense unless specifically defined as herein.

[0030] It should be understood that the sequence number and size of each step in this embodiment do not imply the order of execution. The execution order of each process is determined by its function and internal logic, and should not constitute any limitation on the implementation process of this application embodiment.

[0031] Research has revealed that gallium oxide, as a fourth-generation ultrawide bandgap semiconductor material, has significant application value in fields such as ultra-high voltage and high-power power electronic devices and deep ultraviolet photodetectors. To manufacture high-performance Schottky barrier diodes (SBDs), metal-oxide-semiconductor field-effect transistors (MOSFETs), a gallium oxide epitaxial layer with precisely controllable thickness, doping concentration, and crystal quality must be epitaxially grown on a high-quality gallium oxide single-crystal substrate.

[0032] Currently, existing gallium oxide epitaxial growth methods mainly include metal-organic chemical vapor deposition (MOCVD), halide vapor phase epitaxy (HVPE), molecular beam epitaxy (MBE), and misting chemical vapor deposition (Mist-CVD). However, existing gallium oxide epitaxial growth methods generally face the following problems: 1. High Coupling of Multiphase Gas-Phase Reactions and Thermodynamics: Taking the mainstream epitaxial growth processes MOCVD and HVPE as examples, the quality of the epitaxial layer (such as surface roughness RMS, mobility, doping concentration, etc.) is highly nonlinearly coupled with multiple physical quantities such as growth temperature, chamber pressure, VI / III ratio, carrier gas flow rate, spray nozzle spacing, and substrate rotation speed. The pre-reaction of the precursor in the gas phase, boundary layer transport, and surface adsorption-desorption kinetics are extremely complex. It is difficult to achieve simultaneous optimization of multiple indicators (such as thickness uniformity and high mobility) based on physical models and PID control methods.

[0033] 2. Significant differences in growth windows across different substrate orientations and doping types: Gallium oxide possesses a unique monoclinic crystal structure, resulting in strong anisotropy in its epitaxial growth. The temperature and pressure windows for its two-dimensional "step-flow" growth mode differ drastically across substrates with different orientations (e.g., (100), (010), (001), (-201), etc.) and different off-cut angles. Furthermore, the self-compensation effect and impurity adsorption efficiency during the growth process vary significantly among Si-doped, Sn-doped, and undoped (UID) epitaxial layers. This necessitates adapting a single physical model to all growth scenarios, heavily relying on engineers' trial-and-error experience.

[0034] 3. Lack of intelligent human-computer interaction and adaptive decision-making capabilities: Existing epitaxial control devices can only execute fixed recipes and cannot understand the operator's natural language intentions. They also cannot make adaptive adjustments when anomalies (such as surface roughening, three-dimensional island growth, and parasitic reactions) are detected in situ (such as spectral reflectance and RHEED oscillation).

[0035] 4. Insufficient model prediction accuracy and difficulty in adaptive parameter tuning: Although some studies have attempted to introduce traditional machine learning models for prediction of extension rate or mobility, the hyperparameters of these models (such as MLP and XGBoost) have a great impact on prediction accuracy. Manual parameter tuning is inefficient and it is difficult to ensure that the coefficient of determination of regression analysis reaches the optimal under variable extension conditions.

[0036] To address the aforementioned issues, this application embodiment parses the epitaxial task intent from the operator's natural language commands using a large model, and calls a target gallium oxide epitaxial model from a preset gallium oxide epitaxial model library based on the epitaxial task intent; obtains the epitaxial parameters of the gallium oxide epitaxy being performed in the epitaxial device; determines gallium oxide epitaxial control parameters based on the epitaxial parameters using the target gallium oxide epitaxial model; and controls the epitaxial device based on the gallium source flow adjustment parameters to achieve closed-loop control of the gallium oxide epitaxy. This application parses the epitaxial task intent from the operator's natural language commands using a large model, selects a target gallium oxide epitaxial model established through detailed operating conditions based on the epitaxial task intent, predicts the epitaxial parameters using the target gallium oxide epitaxial model, and finally controls the epitaxial device based on the epitaxial parameters. This achieves dynamic model matching and calling under natural language interaction, eliminating the need for engineers to manually adapt gallium oxide epitaxial models, reducing reliance on human experience, and adapting to various epitaxial growth scenarios such as different substrate orientations and different doping types. Meanwhile, through high-precision forward prediction of the target gallium oxide epitaxial model, optimal control of multiple indicators such as epitaxial resistivity, thickness and mobility can be achieved simultaneously.

[0037] The application content will be further explained below with reference to the accompanying drawings and the description of the embodiments.

[0038] This embodiment provides a model-linked gallium oxide epitaxial control method, applying a model-linked gallium oxide epitaxial control system, such as... Figure 1 As shown, the model-linked gallium oxide epitaxial control system is based on the PyQt framework and adopts a front-end and back-end separation architecture (retaining the .ui interface file to ensure that the software startup and page switching process is extremely smooth, without any flickering or brief display abnormalities).

[0039] The front-end includes a GUI (User-Guided Interface) layer, providing operators with an interactive interface to obtain natural language instructions and display the epitaxial growth process. The front-end may include a process recipe editing panel, a natural language dialog box, a real-time epitaxial parameter curve display panel (temperature, pressure, reflectivity, flow rate), and model training and automatic parameter tuning tabs.

[0040] The backend is used for bidirectional communication with the epitaxial device and to execute a model-based gallium oxide epitaxial control method. This involves parsing the operator's natural language commands to understand the epitaxial task intent using a large model, and then calling a target gallium oxide epitaxial model from a pre-defined gallium oxide epitaxial model library based on this intent. The backend also acquires the epitaxial parameters of the gallium oxide epitaxial process being performed on the epitaxial device; determines the gallium oxide epitaxial control parameters based on these parameters using the target gallium oxide epitaxial model; and controls the epitaxial device based on the gallium source flow adjustment parameters to achieve closed-loop control of the gallium oxide epitaxy. The backend can communicate bidirectionally with the PLC controller of the epitaxial device via the OPC UA industrial Ethernet protocol.

[0041] like Figure 2 As shown, the model-linked gallium oxide epitaxial control method specifically includes: S10. The extended task intent in the natural language instructions of the operator interaction is parsed through the large model, and the target gallium oxide epitaxial model is called in the preset gallium oxide epitaxial model library based on the extended task intent. Specifically, the large model is used to interact with operators to obtain their natural language instructions. These instructions can be completed in a single round of interaction or determined through multiple rounds of dialogue. The large model is deployed on an edge industrial control computer within the production line's local area network. It employs a lightweight, open-source large model that has been incrementally fine-tuned using LoRA and specializes in parsing epitaxial process corpora to meet the confidentiality requirements of semiconductor production lines.

[0042] Natural language instructions include epitaxial parameters required for gallium oxide epitaxy, such as epitaxial layer thickness, target resistivity, doping concentration range, and crystal orientation information. The large model, acting as the upper-level decision layer, performs semantic parsing of the natural language instructions, extracting the epitaxial parameters to form the epitaxial task intent. This epitaxial task intent can include epitaxial growth methods, substrate surface type, epitaxial parameters, and doping types.

[0043] For example: Suppose the operator's natural language instruction is "I need to grow 3 on a homogeneous substrate with a (010) plane using MOCVD". Thick, doping concentration is "Low-resistivity gallium oxide thin film". The large model performs semantic parsing on the natural language instruction, extracting the epitaxial method as MOCVD, the substrate orientation as (010), and the doping type as The target thickness is 3 The target concentration is Based on the extracted target extension parameters, an extension task intent is formed, which is as follows: Extensional method: MOCVD, Substrate orientation: (010), Doping type: , Target thickness: 3 , Target concentration: }

[0044] It should be noted that, in addition to homoepitaxial epitaxy, the epitaxial scenario in the embodiments of this application can also be sapphire, silicon, etc. ), silicon carbide ( Epitaxial growth on heterogeneous substrates such as Mutually, In the growth scenarios of gallium oxide and other ultra-wide bandgap semiconductor materials (such as AlN and diamond).

[0045] Furthermore, after parsing the epitaxial task intent, the system retrieves the target GaO epitaxial model from a pre-defined GaO epitaxial model library based on the determined task intent. This library contains multiple GaO epitaxial models, each applicable to different epitaxial scenarios. Specifically, each GaO epitaxial model in the library is pre-labeled with an applicable scenario tag. These taggings can include information such as epitaxial method, substrate orientation range, and doping type range. After parsing the epitaxial task intent from the large model, the matching degree between the task intent and each scenario tag is calculated. The GaO epitaxial model corresponding to the scenario tag with the highest matching degree is selected as the target GaO epitaxial model, completing the dynamic matching and retrieval of the model without requiring manual selection of a suitable model.

[0046] Furthermore, the pre-defined gallium oxide epitaxial model library allows for the use of different model structures. For example, gallium oxide model A in the library uses a multilayer perceptron (MLP), model B uses a gradient boosting tree (XGBoost) model, and model C uses a long short-term memory (LSTM) network. These models with different structures can be trained using training data corresponding to different scene labels. Additionally, the library may contain some gallium oxide models that use the same model structure but are trained using training data corresponding to different scene labels. For instance, gallium oxide model D is a multilayer perceptron trained for gallium oxide scene a, and model E is a multilayer perceptron trained for gallium oxide scene b.

[0047] The pre-built gallium oxide epitaxial model library in this application stores multiple gallium oxide epitaxial models trained for different gallium oxide epitaxial scenarios. Each gallium oxide epitaxial model is trained and offline optimized for a specific gallium oxide epitaxial condition, and is adaptable to different epitaxial methods, substrate orientations, and doping types. Compared with generalized models, it can effectively improve the prediction accuracy for each gallium oxide epitaxial scenario and avoid prediction bias caused by differences in data distribution across different scenarios. Furthermore, this application allows for the configuration of gallium oxide epitaxial models with different model structures for different gallium oxide epitaxial scenarios, further improving the adaptability of gallium oxide epitaxial models to specific conditions. This eliminates the need for repeated manual adjustment of model parameters for new conditions, reducing reliance on engineers' experience.

[0048] In one embodiment, calling the target gallium oxide epitaxial model from the preset gallium oxide epitaxial model library based on the epitaxial task intent specifically includes: Determine the matching degree between the epitaxial task intent and the epitaxial scenario of each gallium oxide epitaxial model in the preset gallium oxide epitaxial model library; Based on the matching degree, a target gallium oxide epitaxial model is selected from the preset gallium oxide epitaxial model library; The model configuration file of the target gallium oxide epitaxial model is read to dynamically invoke the target gallium oxide epitaxial model.

[0049] Specifically, the matching degree reflects the degree of fit between the epitaxial task intent and the applicable scenario of the gallium oxide epitaxial model. It can be obtained by calculating the similarity (such as cosine similarity) between each epitaxial parameter index in the epitaxial task intent and the epitaxial parameters in the scenario label. In addition, in practical applications, a weight coefficient can be configured for each epitaxial parameter in the scenario label. Then, the similarity corresponding to each epitaxial parameter can be weighted using the weight coefficient to determine the matching degree. Moreover, the weight coefficient configured for the epitaxial parameter can be dynamically adjusted according to actual needs. For example, when the accuracy of the adaptation of growth parameters is required to be high, the weight of key parameters such as substrate orientation and doping type can be increased, thereby prioritizing the matching accuracy of core scenario parameters and avoiding the impact of core parameter adaptation errors on the final epitaxial quality.

[0050] Furthermore, after selecting the target gallium oxide epitaxial model, the model selection is dynamically achieved by reading the model configuration file of the target gallium oxide epitaxial model, thus avoiding hardcoding any specific model filenames in the code. The model configuration file dynamically records all trained gallium oxide epitaxial models and their physical applicability boundaries. The structure of this model configuration file can be as follows: { "epitaxy_models (list of epitaxy models)":[ { "model_id (model ID)":"sm_mocvd_010_si_v3", "algorithm":"XGBoost", "path (file path)":" / models / xgb_mocvd_010_si.pkl", "parameters":{ "max_depth (maximum tree depth)":6, "learning_rate": 0.05, "n_estimators (number of decision trees)": 200 }, "applicable_scope":{ "method (extensional method)": "MOCVD", "orientation":"010", "doping (doping elements)":" " }, "metrics (model evaluation metrics)":{ R2_growth_rate (growth rate determination coefficient): 0.988, "R2_carrier_conc (carrier concentration determination coefficient)": 0.982 } } ] }

[0051] Furthermore, after selecting the target gallium oxide epitaxial model, the model path of the target gallium oxide epitaxial model can be displayed on the front-end interface, allowing the operator to manually change the model used at any time before the task begins via a drop-down menu or path browsing, ensuring controllability. Based on this, after obtaining the epitaxial parameters of the gallium oxide epitaxial process performed in the epitaxial device, the method further includes: The process of changing the gallium oxide epitaxial model is detected. The changing operation is formed through a drop-down menu or path browsing. The new gallium oxide epitaxial model corresponding to the replacement operation is invoked, and the target gallium oxide epitaxial model is replaced with the new gallium oxide epitaxial model.

[0052] Specifically, when an operator judges, based on practical experience or on-site conditions, that the preset matching model does not meet the current requirements, they can directly select a model that meets the requirements on the front-end interactive interface. After the system detects the manual replacement operation, it will directly read the configuration file corresponding to the new model after the replacement, load the new model to replace the original target model, and update the prediction model used in subsequent calculations to the manually selected new model. This balances the efficiency of automatic matching with the flexibility of manual intervention, ensuring the controllability of the actual extension process.

[0053] S20. Obtain the epitaxial parameters of the gallium oxide epitaxial process performed in the epitaxial device.

[0054] Specifically, the epitaxial parameters are data collected online in real time from the epitaxial equipment. This equipment can be various gallium oxide epitaxial devices, such as MOCVD (Metal-Organic Chemical Vapor Deposition) furnaces, HVPE (Halide Vapor Phase Epitaxy) furnaces, MBE (Molecular Beam Epitaxy) furnaces, and Mist-CVD (Mist Chemical Vapor Deposition) furnaces. The epitaxial parameters serve as input data for the target gallium oxide epitaxial model, enabling the model to infer and determine the gallium oxide epitaxial control parameters based on these parameters.

[0055] The epitaxial parameters include relative time, relative position, process relative deviation, and several historical gallium oxide epitaxial control parameters; wherein: Relative time is the merging of relative time in the extension phase. , This refers to the merging time during the extension phase. The target time for the extension phase; The relative position is the current cumulative thickness of the epitaxial layer. ; Relative process deviations include temperature deviations. Dimensionless pressure ratio Carrier gas flow ratio and relative doping ratio ,in, This indicates the measured substrate temperature. Indicates the melting point of the material. This indicates the actual pressure in the reaction chamber. Indicates standard atmospheric pressure. Indicates carrier gas flow rate, Indicates the total gas flow rate. Indicates the flow rate of the doped source gas. Indicates the gallium source gas flow rate; Several historical gallium oxide epitaxial control parameters are a sequence of historical source flux difference values ​​from a preset number of sampling steps prior to the current sampling step, which can be expressed as: , , in, This represents a sequence of historical source flow differences. This indicates the number of historical source flow differences in the historical source flow difference sequence. Indicates the current sampling step. This represents the difference in historical source traffic. This indicates the real-time measured flow rate of the gallium source. This represents the initial steady-state flow rate of the gallium source.

[0056] This application uses relative time, relative position, and process relative deviation as input data, which can eliminate the influence of different equipment ranges and units, and adapt to various gallium oxide epitaxial models. Simultaneously, this application incorporates several historical gallium oxide epitaxial control parameters into the input data. By using historical data to reflect the process drift trend of the epitaxial process, it helps the model more accurately capture process change patterns and improves the accuracy of predictive control parameters.

[0057] It should be noted that during the inference process of the gallium oxide epitaxial model, the epitaxial values ​​for each acquisition step (such as a 1Hz sampling period) are actually acquired online in real time by sensors (such as the flow rate feedback from the mass flow meter MFC, the measured temperature from the pyrometer, etc.); while the target setpoint for each acquisition step (such as the target flow rate for the next step of the PLC) is calculated from the inference output of the gallium oxide epitaxial model, thereby achieving true adaptive closed-loop control. The epitaxial parameters can also include other sensor data, such as the crystal plane angle acquired by the in-situ infrared CCD, the multi-point temperature distribution on the outer wall of the epitaxial equipment, and the residual concentration of precursors at the waste gas treatment end, to further enrich the feature information.

[0058] For example, the relationship between the actual data collected online in real time by the sensor and the epitaxial parameters of the input target gallium oxide epitaxial model can be shown in Table 1.

[0059] Table 1

[0060] S30. Determine gallium oxide epitaxial control parameters based on the epitaxial parameters using the target gallium oxide epitaxial model.

[0061] Specifically, the input of the target gallium oxide epitaxial model is the epitaxial parameters, and the output is the gallium oxide epitaxial control parameters. In other words, the target gallium oxide epitaxial model can directly infer and output the gallium oxide epitaxial control parameters that best meet the current operating conditions based on the input epitaxial parameters, without the need for repeated manual adjustments to the set values ​​of the control parameters.

[0062] For example, the target gallium oxide epitaxial model adopts the XGBoost model. The epitaxial parameters are input into the XGBoost model, and forward propagation calculation is performed by accumulating the values ​​of the leaf nodes of multiple decision trees (such as 200 trees) to obtain the gallium oxide epitaxial control parameters. These gallium oxide epitaxial control parameters are expressed as follows:

[0063] in, Indicates sampling step Gallium oxide epitaxial control parameters, Indicates the number of decision trees. Indicates the first A decision tree, Indicates sampling step The extensional parameters.

[0064] The gallium oxide epitaxial control parameters may include gallium source flow rate adjustment parameters, as well as real-time epitaxial growth rate and step-doping atom adsorption efficiency. The gallium source flow rate adjustment parameter is the core predictive output, used to send data to the epitaxial equipment. The real-time epitaxial growth rate and step-doping atom adsorption efficiency are auxiliary predictive outputs, used to help operators intuitively grasp the current growth status of the epitaxial layer, assist in judging whether the process is stable and whether the model prediction is reasonable, and provide intuitive reference for manual judgment.

[0065] In this embodiment, the gallium source flow rate adjustment parameter can be the gallium source flow rate adjustment difference. This ensures the model's cross-device generalization performance. The gallium source flow adjustment parameter is sent to the flow control system of the epitaxial device. The flow control system automatically determines the target absolute flow rate of the gallium source to complete the flow adjustment. The target absolute flow rate of the gallium source is expressed as: , in, This represents the absolute flow rate of the gallium source target. This indicates the difference in gallium source flow rate adjustment. This represents the initial steady-state flow rate of the gallium source.

[0066] S40. Control the epitaxial device based on the gallium source flow adjustment parameters to perform closed-loop control of the gallium oxide epitaxy.

[0067] Specifically, after obtaining the gallium source flow adjustment parameters, these parameters are sent to the flow control system (such as a PLC controller) of the epitaxial equipment. The flow control system calculates the target absolute flow rate of the gallium source based on the parameters and directly adjusts the valve opening in the pipeline to stably adjust the gallium source flow rate to the target absolute flow rate, completing the closed-loop adjustment of this sampling step. In the next sampling step, the epitaxial parameters under the current operating conditions are re-acquired, and the target gallium oxide epitaxial model is input again to calculate new gallium source flow adjustment parameters. This iterative adjustment is repeated to achieve dynamic closed-loop control of the entire gallium oxide epitaxial process, thereby offsetting the fluctuations in growth rate and carrier concentration caused by process drift in real time, ensuring the doping uniformity and thickness uniformity of the gallium oxide epitaxial layer, and stabilizing the growth quality of the gallium oxide epitaxial layer.

[0068] In one embodiment, during gallium oxide epitaxial growth (especially the growth of superlattice, heterojunction, or modulated doped structures), even small fluctuations in precursor flow rate or growth temperature can lead to a widening of the epitaxial layer interface transition region, thereby severely degrading device performance. Therefore, before controlling the epitaxial equipment based on the gallium source flow rate adjustment parameters, these parameters can be used as prior knowledge, and then smoothed by combining them with the stability constraints of historical multi-sampling step control to avoid frequent and large fluctuations in flow rate within a short period of time, thus reducing the impact on the stability of the growth interface.

[0069] For example, the gallium source flow adjustment parameter controls the epitaxial device to perform closed-loop control of the gallium oxide epitaxy, specifically including: The gallium source flow rate adjustment parameters are used as prior knowledge, and the actual gallium source flow rate of several preceding sampling steps is obtained. Based on the prior knowledge and the actual gallium source flow rate of the preceding sampling steps, the target gallium source flow rate adjustment parameters are determined using the Bayesian fluctuation suppression method. The target gallium source flow rate adjustment parameters are sent to the controller corresponding to the epitaxial device, and the controller controls the epitaxial device to perform closed-loop control of the gallium oxide epitaxy.

[0070] Specifically, the Bayesian fluctuation suppression method establishes a fluctuation probability model based on the prior gallium source flow rate adjustment parameters and the actual gallium source flow rate of several preceding sampling steps. By constraining the total fluctuation amplitude of the flow rate adjustment, the target gallium source flow rate adjustment parameters that meet the stability requirements are selected. This ensures that process drift can be offset while avoiding large fluctuations in flow rate in a short period of time, thus ensuring that the epitaxial interface reaches atomic-level ultra-steepness.

[0071] For example, the actual gallium source flow rate of several preceding sampling steps includes the actual gallium source flow rate of the previous sampling step, and the target gallium source flow rate adjustment parameter can be expressed as: , in, Indicates sampling step Gallium source flow adjustment parameters This represents the absolute flow rate of the gallium source target. This indicates the actual gallium source flow rate in the previous sampling step. Indicates dynamic weighting coefficients. Represents the Bayesian suppression coefficient. This represents the variance of the predicted flow rate.

[0072] In one embodiment, this application further configures a reverse prediction solver. The reverse prediction solver (such as a reverse solving network based on Gaussian process regression or MLP) solves for the optimal combination of process parameters based on the core target parameters (such as target thickness, growth rate, target carrier concentration, etc.) input by the operator within the actual physical feasible domain (allowing negative values ​​and other parameters that conform to physical laws, with input limits entirely based on actual physical feasibility). This provides a reference for setting initial steady-state parameters for forward dynamic control, reduces the difficulty of initial process parameter debugging, and further ensures that gallium oxide epitaxial growth meets design requirements from the initial stage.

[0073] In one embodiment, to improve the adaptability of the gallium oxide epitaxial model to actual process conditions, the training process of the gallium oxide epitaxial model can also adopt a model linkage architecture. The gallium oxide epitaxial model derived based on physical mechanisms is combined with a data-driven large model. The gallium oxide epitaxial model provides training index data for the large model, and the large model analyzes the training index data for the process drift characteristics of specific equipment and provides adjustment strategies for the gallium oxide epitaxial model. This ensures that the model output conforms to the basic physical laws and can adapt to the individual differences of different equipment, further improving control accuracy.

[0074] For example, the training process and / or offline tuning of each gallium oxide epitaxial model in the preset gallium oxide epitaxial model library includes: The training dataset is used to train the gallium oxide epitaxial model to be trained to obtain the intermediate gallium oxide epitaxial model, and the training logs corresponding to the intermediate gallium oxide epitaxial model are obtained. The large model performs deep inference on the training logs to obtain the model state of the gallium oxide epitaxial model, and the gallium oxide epitaxial model is adjusted based on the model state. The process of retraining the untrained gallium oxide epitaxial model using the training dataset to obtain an intermediate gallium oxide epitaxial model continues until a preset termination condition is met. The intermediate gallium oxide epitaxial model that meets the preset termination condition is then used as the trained gallium oxide epitaxial model.

[0075] Specifically, to ensure the standardization of the training dataset, it needs to include device metadata, initial recipe values, and time-series sampling data. For example, an example of training data for MOCVD homoepitaxy (010) Si-doped gallium oxide is as follows: #Metadata Header Device_ID (Device Number): MOCVD-01-

[0076] Epiwafer ID: EP-20260603-01 Substrate Orientation: 010 Dopant_Type (Doping Type): Si Initial_TMGa_Flow_sccm (Initial trimethylgallium flow rate): 25.0 Initial_ _Flow_sccm (Initial Oxygen Flow Rate): 1500.0 #Time-Series Data Time_s (time), Stage_ID (process stage number), Substrate_Temp_C (substrate temperature), Chamber_Press_Torr (chamber pressure), TMGa_Flow_sccm (trimethylgallium flow rate). _Flow_sccm (oxygen flow rate) _Flow_sccm (silane flow rate), Reflectance_au (reflectivity), Epi_Thickness_ (Thickness of epitaxial layer) 0,1,850.0,50.0,25.0,1500.0,0.0,0.620,0.000 10,1,850.2,50.1,25.0,1500.0,0.0,0.621,0.005 120,2,850.0,50.0,25.0,1500.0,5.0,0.615,0.052 600,3,849.8,49.9,25.2,1500.0,5.1,0.550,0.480 1800,3,850.1,50.0,25.1,1500.0,5.0,0.420,1.520 3600,4,850.0,50.0,0.0,1500.0,0.0,0.620,3.020.

[0077] In this example, the system automatically merges the data from stage 3 (steady-state growth) and stage 4 (cooling and purging) and uses them as "stage 34" for unified feature extraction and model training.

[0078] Furthermore, the gallium oxide epitaxial model was trained using the training dataset, and the training loss and training set determination coefficients were recorded in log form after training. Coefficient of determination of the validation set The training process involves analyzing the mean squared error and feature parameters, then inputting the training logs into a large model. This allows the large model to leverage its language understanding and reasoning capabilities to analyze the trends in prediction errors across different parameter ranges, pinpointing the dimensions where the current model parameters do not match the equipment's process characteristics, and providing specific parameter adjustment directions. This assists the gallium oxide crystal growth model to be trained in quickly converging to the optimal parameter range suitable for the current equipment. This linked training mode, where the mechanistic model determines the basic laws and the large model assists in parameter tuning, avoids the abnormal outputs that easily occur in purely data-driven models, which violate physical laws, and solves the problem that purely mechanistic models are difficult to adapt to the individual differences of different equipment. It effectively improves the prediction accuracy and cross-device adaptability of the gallium oxide epitaxial model.

[0079] In one embodiment, the step of performing deep inference on the training logs using a large model to obtain the model state of the gallium oxide epitaxial model specifically includes: Obtain the epitaxial scene corresponding to the gallium oxide epitaxial model, and obtain the model evaluation index based on the epitaxial scene; Select the model index value corresponding to the model evaluation index from the training log, and determine the model state of the gallium oxide epitaxial model based on the model index value.

[0080] Specifically, different epitaxial scenarios have varying accuracy requirements for gallium oxide epitaxial models, and the corresponding model evaluation metrics also differ. For example, in conductive layer growth scenarios where carrier concentration control is crucial, the mean square error of carrier concentration prediction is prioritized as the core evaluation metric. Conversely, in power device substrate epitaxial scenarios where thickness uniformity is critical, the coefficient of determination for thickness prediction is prioritized. The large model matches the corresponding model evaluation metric to the current epitaxial scenario, extracts the numerical changes of the corresponding metric from the training logs across different parameter ranges, analyzes the deviation range of the current model parameters, and finally outputs a determined model state, providing an accurate basis for subsequent parameter adjustments.

[0081] The correspondence between the model evaluation metrics and the extended scenarios is as follows: When the epitaxial scenario is of the highly doped type, the model evaluation metric is the validation set determination coefficient; When the epitaxial scenario is unintentionally doped, the model evaluation metric is the feature parameters.

[0082] Specifically, highly doped epitaxial layers inherently have high carrier concentrations, and small-scale concentration fluctuations have a relatively limited impact on device performance. However, thickness uniformity is a core requirement for ensuring consistency between subsequent substrate processing and device performance. Therefore, selecting the validation set determination coefficient corresponding to the thickness prediction accuracy as the core evaluation index better aligns with the actual production needs of highly doped gallium oxide epitaxy. This allows model optimization to better fit the control objectives in this scenario, further improving the growth quality stability of highly doped gallium oxide epitaxy. Undoped gallium oxide epitaxial layers, on the other hand, have lower requirements for impurity concentration control but higher requirements for intrinsic carrier concentration stability. Therefore, extracting characteristic parameters at the physical feature level as evaluation indices can more accurately reflect the model's fit to the basic growth law. Larger models can directly locate deviations at the mechanism adaptation level through changes in characteristic parameters, enabling more efficient correction of model parameters.

[0083] Furthermore, after determining the model evaluation metrics, the model state of the gallium oxide epitaxial model can be determined based on the metric values. This model state reflects the adaptability of the current model to the epitaxial scenario and the current production equipment. If the model metric values ​​do not reach the preset threshold for the scenario, the large model will directly output the adjustment step size and direction for the corresponding parameter dimensions, automatically triggering the next round of training and parameter tuning until the model metric meets the preset requirements, ultimately obtaining the optimal gallium oxide epitaxial model adapted to the current equipment and doping scenario. The model state can include generalization error and volatility, training convergence and response latency, and accumulated timing deviation.

[0084] Furthermore, when adjusting the gallium oxide epitaxial model based on its state, the model structure can be read, and then a model adjustment strategy can be selected based on this structure and state. The model structure, state, and adjustment strategy can then be adjusted accordingly. The correspondence between model structure, state, and adjustment strategy can be pre-defined. For example, the correspondence between these relationships can be shown in Table 2.

[0085] Table 2

[0086] Furthermore, the large model writes the modified hyperparameters into the configuration file, automatically triggering the retraining of the gallium oxide epitaxial model. This closed-loop iteration continues until the determination coefficient of the gallium oxide epitaxial model is reached. Once a preset threshold is reached (e.g., >0.98), further increases are deferred. After training, only the best-performing model training file is retained to avoid redundancy. Simultaneously, the large model decision layer automatically generates a training summary report, providing the operator with a detailed explanation of the physical and mathematical logic behind hyperparameter tuning.

[0087] In one embodiment, during the training and / or offline tuning process of each gallium oxide epitaxial model in the preset gallium oxide epitaxial model library, the training dataset used for training and / or offline tuning is cleaned, wherein the data cleaning includes at least one of the following: Dorajal outlier removal: A sliding window is used to address electromagnetic spikes generated by thermocouples, flow meters, and pressure gauges during gas switching. The principle is to automatically eliminate; Kalman filtering smoothing: In-situ smoothing and noise reduction of high-frequency noise signals such as spectral reflectance and pyrometer temperature, while preserving the true physical interference trend of thin film growth; Multiple imputation for missing values: For non-aligned missing data caused by inconsistent sensor sampling rates, multiple imputation is performed using the Random Forest algorithm to reconstruct the complete time series; Transient time alignment during gas switching: In epitaxial growth, there is a gas phase transport delay of several seconds from the valve opening to the substrate surface of the precursor. For example, by calculating the cross-correlation function, the flow control signal and the in-situ reflectivity change signal can be automatically aligned to eliminate the feature misalignment caused by the time lag. Derivative feature generation: Perform pairwise multiplication, division, and inverse division operations on core features such as "current temperature", "VI / III ratio", and "growth rate" (e.g., generate "temperature-VI / III ratio product", "growth rate-temperature ratio", etc.) to enrich the feature space and significantly improve the fitting ability of single or dual multilayer perceptron (MLP) models.

[0088] In summary, this embodiment provides a model-linked gallium oxide epitaxial control method. The method includes parsing the epitaxial task intent in the operator's natural language commands using a large model, and calling a target gallium oxide epitaxial model from a preset gallium oxide epitaxial model library based on the epitaxial task intent; obtaining the epitaxial parameters of the gallium oxide epitaxial process being performed in the epitaxial device; determining gallium oxide epitaxial control parameters based on the epitaxial parameters using the target gallium oxide epitaxial model; and controlling the epitaxial device based on the gallium source flow adjustment parameters to achieve closed-loop control of the gallium oxide epitaxy. This application parses the epitaxial task intent in the operator's natural language commands using a large model, selects a target gallium oxide epitaxial model established through detailed operating conditions based on the epitaxial task intent, predicts the epitaxial parameters using the target gallium oxide epitaxial model, and finally controls the epitaxial device based on the epitaxial parameters. This achieves dynamic model matching and calling under natural language interaction, eliminating the need for engineers to manually adapt gallium oxide epitaxial models, reducing reliance on human experience, and adapting to various epitaxial growth scenarios such as different substrate orientations and different doping types. Meanwhile, through high-precision forward prediction of the target gallium oxide epitaxial model, optimal control of multiple indicators such as epitaxial resistivity, thickness and mobility can be achieved simultaneously.

[0089] Based on the above-described model-linked gallium oxide epitaxial control method, this embodiment provides a model-linked gallium oxide epitaxial control device, such as... Figure 3 As shown, the model-linked gallium oxide epitaxial control device specifically includes: The intent parsing module 100 is used to parse the extended task intent in the natural language instructions of the operator interaction through a large model, and to call the target gallium oxide epitaxial model in the preset gallium oxide epitaxial model library based on the extended task intent. The acquisition module 200 is used to acquire the epitaxial parameters of gallium oxide epitaxy performed in the epitaxial device, wherein the epitaxial parameters include relative time, relative position, process relative deviation and several historical gallium oxide epitaxial control parameters; The prediction module 300 is used to determine gallium oxide epitaxial control parameters based on the epitaxial parameters using the target gallium oxide epitaxial model, wherein the gallium oxide epitaxial control parameters include at least gallium source flow rate adjustment parameters; The control module 400 is used to control the epitaxial device based on the gallium source flow adjustment parameters, so as to perform closed-loop control of the gallium oxide epitaxy.

[0090] Based on the above-described model-linked gallium oxide epitaxial control method, this embodiment provides a computer-readable storage medium storing one or more programs that can be executed by one or more processors to implement the steps in the model-linked gallium oxide epitaxial control method described in the above embodiment.

[0091] Based on the above-described model-linked gallium oxide epitaxial control method, this application also provides a terminal device, such as... Figure 4 As shown, it includes at least one processor 20; a display screen 21; and a memory 22, and may also include a communications interface 23 and a bus 24. The processor 20, display screen 21, memory 22, and communications interface 23 can communicate with each other via the bus 24. The display screen 21 is configured to display a preset user guide interface in the initial setup mode. The communications interface 23 can transmit information. The processor 20 can invoke logical instructions in the memory 22 to execute the methods described in the above embodiments.

[0092] Furthermore, the logical instructions in the aforementioned memory 22 can be implemented as software functional units and, when sold or used as independent products, can be stored in a computer-readable storage medium.

[0093] The memory 22, as a computer-readable storage medium, can be configured to store software programs, computer-executable programs, such as program instructions or modules corresponding to the methods in the embodiments of this disclosure. The processor 20 executes functional applications and data processing by running the software programs, instructions, or modules stored in the memory 22, thereby implementing the methods in the above embodiments.

[0094] The memory 22 may include a program storage area and a data storage area. The program storage area may store the operating system and application programs required for at least one function; the data storage area may store data created based on the use of the terminal device. Furthermore, the memory 22 may include high-speed random access memory (RAM) and non-volatile memory. Examples include 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, as well as transient storage media.

[0095] Furthermore, the specific process of loading and executing multiple instruction processors in the aforementioned storage medium and terminal device has been described in detail in the above method, and will not be repeated here.

[0096] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of this application, and are not intended to limit them. Although this application has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some of the technical features. Such modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the spirit and scope of the technical solutions of the embodiments of this application.

Claims

1. A model-linked control method for gallium oxide epitaxy, characterized in that, The model-linked gallium oxide epitaxial control method specifically includes: The model parses the extended task intent in the natural language commands of the operator interaction, and calls the target gallium oxide epitaxial model in the preset gallium oxide epitaxial model library based on the extended task intent. Obtain the epitaxial parameters of gallium oxide epitaxy performed in the epitaxial device, wherein the epitaxial parameters include relative time, relative position, process relative deviation, and several historical gallium oxide epitaxial control parameters; Gallium oxide epitaxial control parameters are determined based on the epitaxial parameters using the target gallium oxide epitaxial model, wherein the gallium oxide epitaxial control parameters include at least gallium source flow rate adjustment parameters; The gallium source flow rate adjustment parameters are used to control the epitaxial device, thereby enabling closed-loop control of the gallium oxide epitaxy.

2. The gallium oxide epitaxial control method based on model linkage according to claim 1, characterized in that, The epitaxial task intent includes at least one of the following: epitaxial growth method, substrate surface type, epitaxial parameters, and doping type.

3. The gallium oxide epitaxial control method based on model linkage according to claim 1, characterized in that, The step of calling the target gallium oxide epitaxial model from the preset gallium oxide epitaxial model library based on the epitaxial task intent specifically includes: Determine the matching degree between the epitaxial task intent and the epitaxial scenario of each gallium oxide epitaxial model in the preset gallium oxide epitaxial model library; Based on the matching degree, a target gallium oxide epitaxial model is selected from the preset gallium oxide epitaxial model library; The model configuration file of the target gallium oxide epitaxial model is read to dynamically invoke the target gallium oxide epitaxial model.

4. The gallium oxide epitaxial control method based on model linkage according to claim 1, characterized in that, After obtaining the epitaxial parameters of the gallium oxide epitaxy performed in the epitaxial device, the method further includes: The process of changing the gallium oxide epitaxial model is detected. The changing operation is performed through a drop-down menu or by browsing a path. The new gallium oxide epitaxial model corresponding to the replacement operation is invoked, and the target gallium oxide epitaxial model is replaced with the new gallium oxide epitaxial model.

5. The gallium oxide epitaxial control method based on model linkage according to claim 1, characterized in that, The gallium oxide epitaxial control parameters also include real-time epitaxial growth rate and step-doped atom adsorption efficiency.

6. The gallium oxide epitaxial control method based on model linkage according to claim 1, characterized in that, The training process and / or offline tuning of each gallium oxide epitaxial model in the preset gallium oxide epitaxial model library includes: The training dataset is used to train the gallium oxide epitaxial model to be trained to obtain the intermediate gallium oxide epitaxial model, and the training log corresponding to the intermediate gallium oxide epitaxial model is obtained. The training log includes training loss, validation set determination coefficient, mean square error and feature parameters. The large model performs deep inference on the training logs to obtain the model state of the gallium oxide epitaxial model, and the gallium oxide epitaxial model is adjusted based on the model state. The process of retraining the untrained gallium oxide epitaxial model using the training dataset to obtain an intermediate gallium oxide epitaxial model continues until a preset termination condition is met. The intermediate gallium oxide epitaxial model that meets the preset termination condition is then used as the trained gallium oxide epitaxial model.

7. The gallium oxide epitaxial control method based on model linkage according to claim 6, characterized in that, The step of performing deep inference on the training logs using a large model to obtain the model state of the gallium oxide epitaxial model specifically includes: Obtain the epitaxial scene corresponding to the gallium oxide epitaxial model, and obtain the model evaluation index based on the epitaxial scene; Select the model index value corresponding to the model evaluation index from the training log, and determine the model state of the gallium oxide epitaxial model based on the model index value; The correspondence between the model evaluation metrics and the extended scenarios is as follows: When the epitaxial scenario is of the highly doped type, the model evaluation metric is the validation set determination coefficient; When the epitaxial scenario is unintentionally doped, the model evaluation metric is the feature parameters.

8. The gallium oxide epitaxial control method based on model linkage according to claim 1, characterized in that, During the training and / or offline tuning process of each gallium oxide epitaxial model in the preset gallium oxide epitaxial model library, the training dataset used for training and / or offline tuning is cleaned. The data cleaning includes one or more of the following: Dorajal outlier removal, Kalman filter smoothing, multiple imputation missing value processing, gas switching transient time alignment, and derived feature generation.

9. The gallium oxide epitaxial control method based on model linkage according to claim 1, characterized in that, The control of the epitaxial device based on the gallium source flow adjustment parameters to achieve closed-loop control of the gallium oxide epitaxy specifically includes: The gallium source flow rate adjustment parameters are used as prior knowledge, and the actual gallium source flow rate of several preceding sampling steps is obtained. Based on the prior knowledge and the actual gallium source flow rate of the preceding sampling steps, the target gallium source flow rate adjustment parameters are determined using the Bayesian fluctuation suppression method. The target gallium source flow rate adjustment parameters are sent to the controller corresponding to the epitaxial device, and the controller controls the epitaxial device to perform closed-loop control of the gallium oxide epitaxy.

10. A gallium oxide epitaxial control device based on model linkage, characterized in that, The model-linked gallium oxide epitaxial control device specifically includes: The intent parsing module is used to parse the extended task intent in the natural language instructions of the operator interaction through a large model, and to call the target gallium oxide epitaxial model in the preset gallium oxide epitaxial model library based on the extended task intent. The acquisition module is used to acquire the epitaxial parameters of gallium oxide epitaxy performed in the epitaxial device, wherein the epitaxial parameters include relative time, relative position, process relative deviation, and several historical gallium oxide epitaxial control parameters; The prediction module is used to determine gallium oxide epitaxial control parameters based on the epitaxial parameters using the target gallium oxide epitaxial model, wherein the gallium oxide epitaxial control parameters include at least gallium source flow rate adjustment parameters; The control module is used to control the epitaxial device based on the gallium source flow adjustment parameters, so as to perform closed-loop control of the gallium oxide epitaxy.

11. A computer-readable storage medium, characterized in that, The computer-readable storage medium stores one or more programs, which can be executed by one or more processors to implement the steps in the model-linked gallium oxide epitaxial control method as described in any one of claims 1-9.

12. A terminal device, characterized in that, include: Processor and memory; The memory stores a computer-readable program that can be executed by the processor; When the processor executes the computer-readable program, it implements the steps in the model-linked gallium oxide epitaxial control method as described in any one of claims 1-9.

Images