AI-driven silicon carbide crystal growth process intelligent generation and dynamic optimization method and system

Abstract

The application provides an AI-driven silicon carbide crystal growth process intelligent generation and dynamic optimization method and system, relates to the field of silicon carbide crystal growth intelligent control, and the core innovation is that a full closed-loop mechanism of "process generation-equipment adaptation-scene matching-dynamic correction-defect tracing-model iteration" is constructed; through AI double model cooperation of multi-algorithm fusion, accurate equipment state perception and scene demand individualized adaptation, precise and self-adaptive generation of the silicon carbide crystal growth process and millisecond-level dynamic correction of the growth process are realized; the crystal quality stability and production efficiency are significantly improved; and the method is suitable for the silicon carbide crystal growth large-scale production of multiple processes such as PVT and liquid phase method and multiple scenes such as new energy vehicles, AI data centers and AR.

Description

Technical Field

[0001] This invention relates to the field of intelligent control of silicon carbide crystal growth, specifically to an AI-driven intelligent generation and dynamic optimization method and system for silicon carbide crystal growth process. Background Technology

[0002] Silicon carbide (SiC), with its advantages such as high breakdown electric field, high thermal conductivity, and low switching losses, has become a core supporting material for high-end equipment such as new energy vehicles, AI data centers, and photovoltaic inverters. With the accelerated development of new application areas such as AI, AR, and interposers, the market demand for silicon carbide crystals is experiencing a new surge, placing higher demands on crystal quality, production efficiency, and process adaptability.

[0003] The mainstream silicon carbide crystal growth processes are physical vapor transport (PVT) and liquid phase methods. These processes require extremely high levels of coordinated control over multiple parameters, including temperature, pressure, gas flow rate, and growth rate, which exhibit strong coupling and nonlinear correlations. Currently, process control in the industry is still largely based on manual experience. Operators manually adjust the process parameters for the next furnace based on the growth results from the previous furnace (such as crystal defect density, size, and resistivity). This approach suffers from low efficiency, large human error, and poor process consistency, making it difficult to adapt to the growth requirements of crystals of different specifications and for different applications.

[0004] In recent years, although some technologies have attempted to use machine learning to achieve intelligent control of silicon carbide crystal growth, they are limited to closed-loop adjustment of a single parameter and have not achieved automated generation of all process parameters. Furthermore, a linkage mechanism between historical data from the furnace, equipment status, and target parameters for the next furnace cycle has not been established, limiting process adaptability. In addition, existing AI applications focus only on boundary assessment of process parameters or post-growth data feedback, lacking a real-time dynamic correction mechanism during process execution. This makes it impossible to cope with real-time interference such as temperature fluctuations and equipment status changes during crystal growth, making it difficult to guarantee the stability of crystal growth quality. Simultaneously, a single model cannot adapt to different crystal growth processes such as PVT and liquid phase methods, as well as specific application scenarios. Moreover, the correlation between crystal defects and process parameters is only achieved through simple data statistics, resulting in low positioning accuracy and low process optimization efficiency.

[0005] In existing technology 1, only multi-system integration and multi-machine monitoring of silicon carbide crystal growth equipment are realized. AI is only used for basic data analysis and parameter sharing, without involving the intelligent generation and dynamic adaptation of all process parameters. In existing technology 2, the focus is on defining the boundary of process parameters based on experimental-simulation data. It lacks deep linkage with the real-time status of equipment and actual application scenarios, and there is no real-time optimization strategy during process execution. Summary of the Invention

[0006] The objective of this invention is achieved through the following technical solutions.

[0007] To address the shortcomings of the existing technologies, this invention proposes an intelligent system and method for the entire silicon carbide crystal growth process, which integrates precise equipment status perception, personalized adaptation to scenario requirements, dynamic defect correction during the growth process, and continuous iterative optimization of the model. Through the collaboration of AI dual models and the fusion of multiple algorithms, it achieves full-dimensional linkage between process, equipment, and scenario, breaking through the limitations of existing technologies in process generation, dynamic control, and defect tracing, and significantly improving the quality and production efficiency of silicon carbide crystal growth.

[0008] Specifically, this invention provides an AI-driven intelligent generation and dynamic optimization system for silicon carbide crystal growth, comprising:

[0009] The parameter input module is used to obtain the upper furnace crystal parameters, lower furnace target parameters, and scenario requirements for silicon carbide crystal growth.

[0010] The equipment status sensing module is used to collect the hardware status parameters of the growth furnace in real time and generate constraints for process adjustment through a hardware-process adaptation algorithm.

[0011] The AI ​​dual-model collaborative module includes a master generating model and a dynamic correction model;

[0012] The scenario adaptation module has a built-in multi-scenario parameter library, which is used to dynamically adjust the target weight generated by the process according to the scenario requirements.

[0013] The process execution and feedback module is used to convert process plans into equipment control commands for issuance, and to collect data and defect monitoring signals during the growth process in real time.

[0014] The defect tracing and optimization module is used to trace the defects of the grown crystal, generate targeted optimization strategies, and feed them back to the AI ​​dual-model collaborative module for iterative training.

[0015] The host computer interaction module provides a visual interactive interface; and

[0016] The data storage module is used to store all data in a hierarchical and categorized manner.

[0017] The system constructs a fully closed-loop control mechanism of "process generation - equipment adaptation - scenario matching - dynamic correction - defect tracing - model iteration" through the serial connection and closed-loop feedback between various modules.

[0018] This invention also provides an AI-driven intelligent generation and dynamic optimization method for silicon carbide crystal growth using the system described above, comprising the following steps:

[0019] S1: Multi-dimensional data input and standardized preprocessing: Multi-dimensional data is acquired through the parameter input module, equipment status perception module and process execution and feedback module, and standardized preprocessing is performed to generate a standardized multi-dimensional dataset;

[0020] S2: Scene adaptation and initial process scheme generation: The scene adaptation module adjusts the process weights according to the scene requirements, and the main generation model of the AI ​​dual-model collaborative module generates an initial process scheme based on a standardized multi-dimensional dataset.

[0021] S3: Multi-dimensional verification and iterative optimization of process constraints: Verify whether the initial process scheme meets the constraints of equipment hardware limits, thermodynamic balance and scenario performance requirements. If it does not meet the constraints, iterative optimization is performed until it meets the constraints, and the optimized initial process scheme is output.

[0022] S4: Dynamic process execution and millisecond-level real-time correction: The process execution and feedback module issues the process plan, and during the growth process, the dynamic correction model dynamically adjusts the process parameters at the millisecond level based on real-time data and defect signals;

[0023] S5: Defect tracing and AI dual-model iterative training: After growth is completed, the defect tracing and optimization module traces the crystal defects, generates targeted optimization strategies, and constructs an iterative dataset to iteratively train the AI ​​dual model.

[0024] S6: Scenario-based process library update and reuse: Categorize and store the compliant process solutions and their associated data in the scenario-based process library, and update the scenario parameter library for reference in subsequent process generation.

[0025] The advantages of this invention are:

[0026] (1) High degree of automation: The entire process of “input parameters - one-click generation - process issuance” is fully automated, eliminating the need for manual process writing and reducing the process generation time from several hours to several seconds, thus greatly improving production efficiency;

[0027] (2) Strong process adaptability: Based on the linkage of upper and lower furnace parameters, equipment status constraints, and scenario requirements, the process is customized and generated. The feedback value of historical growth data is fully utilized, and the process scheme is more than 90% more in line with the actual growth requirements, effectively reducing the crystal defect rate.

[0028] (3) Lowering the operational threshold: No need for operators to have extensive experience in silicon carbide crystal growth. Ordinary employees can complete the process generation by inputting basic parameters through a visual interface, which facilitates the large-scale production and promotion of the technology.

[0029] (4) Iterable model: Through continuous feedback of real-time data during the growth process and defect data after growth, the AI ​​model can learn and optimize itself. The accuracy of the process generation gradually improves with the accumulation of data, and can adapt to the growth needs of silicon carbide crystals of different specifications, different processes and different scenarios.

[0030] (5) Reduced equipment wear: The process parameters are adapted to the real-time status of the equipment to avoid process execution that exceeds the hardware limits, reducing the overall equipment wear by more than 20% and extending the service life of the equipment. Attached Figure Description

[0031] Various other advantages and benefits will become apparent to those skilled in the art upon reading the following detailed description of preferred embodiments. The accompanying drawings are for illustrative purposes only and are not intended to limit the invention. Furthermore, the same reference numerals denote the same parts throughout the drawings. In the drawings:

[0032] Figure 1 A flowchart of an AI process generation method according to an embodiment of the present invention is shown. Detailed Implementation

[0033] Exemplary embodiments of the present disclosure will now be described in more detail with reference to the accompanying drawings. While exemplary embodiments of the present disclosure are shown in the drawings, it should be understood that the present disclosure may be implemented in various forms and should not be limited to the embodiments set forth herein. Rather, these embodiments are provided so that this disclosure will be thorough and complete, and will fully convey the scope of the disclosure to those skilled in the art.

[0034] This invention discloses an AI-driven intelligent generation and dynamic optimization system and method for silicon carbide crystal growth, relating to the field of intelligent control of silicon carbide crystal growth. The core innovation lies in constructing a closed-loop mechanism of "process generation - equipment adaptation - scenario matching - dynamic correction - defect tracing - model iteration". Through the collaborative AI dual-model fusion of multiple algorithms, precise perception of equipment status, and personalized adaptation to scenario requirements, the system achieves precise silicon carbide crystal growth, adaptive generation, and millisecond-level dynamic correction of the growth process. This significantly improves crystal quality stability and production efficiency, and is applicable to the large-scale production of silicon carbide crystals in multiple scenarios such as PVT, liquid phase method, new energy vehicles, AI data centers, and AR.

[0035] 1.1 Technical Solution: System Architecture (Core Components)

[0036] This system is an integrated intelligent control system combining hardware and software, consisting of eight core modules: parameter input module, equipment status perception module, AI dual-model collaboration module, scene adaptation module, host computer interaction module, process execution and feedback module, defect tracing and optimization module, and data storage module. Each module achieves data exchange and command transmission through industrial communication protocols. The modules are interconnected through a series + closed-loop feedback relationship. The specific structure and implementation details are as follows:

[0037] 1.1.1 Parameter Input Module

[0038] It enables standardized input of multi-dimensional parameters related to silicon carbide crystal growth, supporting four input methods: manual input, Excel template import, automatic device upload, and scene parameter library call. All input data from these four methods are uniformly converted into JSON format for subsequent processing. After data entry, field validation is automatically performed, and missing / erroneous fields are alerted in real time via pop-up windows.

[0039] Crystal parameter input for furnace loading: The fixed input dimensions are crystal size (diameter / thickness, in mm) and defect density (in cm). -2 ), resistivity (unit Ω·cm), growth time (unit h), growth process type (PVT / liquid phase method), crystal yield (%), etc.;

[0040] Input of target parameters for furnace operation: The fixed input dimensions are target size (diameter / thickness, in mm) and target defect level (Level I-V, corresponding to defect density ≤2000 / 3000 / 5000 / 8000 / 10000 cm³). -2 ), target resistivity (unit: Ω·cm), target growth rate (unit: mm / h), target yield (%), etc.;

[0041] Scenario requirement input: Supports drop-down selection of specific application scenarios, and allows manual addition of specific performance requirements for the scenario;

[0042] Automatic equipment upload: It interfaces with the crystal growth equipment control system via OPC UA protocol to automatically collect furnace process parameters and equipment operation data at a frequency of 1Hz;

[0043] Excel template import: Provides standardized templates with fixed fields and data formats. After import, the system automatically parses the field information and removes invalid rows.

[0044] 1.1.2 Equipment Status Sensing Module

[0045] It provides equipment hardware constraints for process generation, collects full-dimensional hardware status parameters of the growth furnace in real time, realizes the quantitative correlation between hardware status and process parameters through hardware-process adaptation algorithm, outputs specific constraints for process adjustment, and provides boundary input for process generation of AI dual-model collaborative module.

[0046] (1) Acquisition parameters and acquisition method: The following hardware status parameters are acquired by using dedicated sensors (temperature / wear / pressure sensors) deployed in the growth furnace at a sampling frequency of 50ms. All parameters are quantitatively scored in real time (0-100 points, the higher the score, the better the hardware status):

[0047] Heating mechanism: Heating mechanism loss rate (%), maximum achievable temperature (°C), temperature control accuracy (±°C);

[0048] Crucible: Remaining life (%), crucible inner wall wear (μm), crucible high temperature resistance threshold (°C);

[0049] Vacuum system: Vacuum control accuracy (Pa), sealing performance (leakage rate, Pa·L / s);

[0050] Airflow control system: airflow valve response accuracy (%), gas flow control error (sccm);

[0051] (2) Hardware-Process Adaptation Algorithm: A multivariate linear regression algorithm is used to construct a correlation model between hardware status and process parameters. The formula is as follows: ,in These are process parameter constraint values. Quantify and score the status of each hardware component. , where b is the weighting coefficient for each hardware state (calibrated according to the crystal growth process type; for PVT process, the heating mechanism has a weight of 0.3 and the crucible has a weight of 0.25; for liquid phase method, the crucible has a weight of 0.35 and the vacuum system has a weight of 0.2), and b is the correction coefficient.

[0052] (3) Constraint output: According to the above algorithm, the upper / lower limit thresholds of core process parameters such as temperature, growth rate and gas flow rate are output. For example, when the loss of the heating mechanism is ≥20%, the maximum heating temperature is reduced by 50-100℃; when the remaining life of the crucible is ≤50%, the upper limit of the growth rate is reduced by 0.05mm / h.

[0053] 1.1.3 AI Dual-Model Collaboration Module

[0054] The core computing module of this system consists of a main generation model and a dynamic correction model. The two models adopt a training mode of "pre-training + online iteration", with data exchange and collaborative work. The main generation model is responsible for the accurate generation of the initial process plan, while the dynamic correction model is responsible for the real-time dynamic adjustment of process parameters during the growth process, realizing a closed loop of process control of "static generation + dynamic optimization".

[0055] (1) Master Generator Model: Hybrid Model of CNN-LSTM and Genetic Algorithm

[0056] By integrating the feature extraction capabilities of Convolutional Neural Networks (CNN), the temporal data processing capabilities of Long Short-Term Memory Networks (LSTM), and the global optimization capabilities of Genetic Algorithms (GA), this method achieves feature fusion of multi-dimensional input data and the global optimal generation of process parameters. The input consists of standardized furnace loading and unloading parameters, scenario requirement parameters, and equipment state constraints. The output is a complete silicon carbide crystal growth process scheme containing 12 core parameters (temperature curve, pressure curve, Ar / N2 / H2 gas flow rate, seed crystal rotation speed, growth time, pulling speed, holding time, cooling rate, vacuum degree, raw material preheating temperature, crucible position, and gas switching time).

[0057] (1.1) CNN architecture design: used to extract spatial features of input data, consisting of 3 convolutional layers + 2 pooling layers + 1 fully connected layer:

[0058] Convolutional layers: Conv1 kernel size 3×3, stride 1, output channels 32, activation function ReLU; Conv2 kernel size 3×3, stride 1, output channels 64, activation function ReLU; Conv3 kernel size 2×2, stride 1, output channels 128, activation function ReLU.

[0059] Pooling layers: all are max pooling, with a 2×2 pooling kernel and a stride of 2, used for dimensionality reduction and preservation of key features;

[0060] Fully connected layer: 256 neurons, ReLU activation function, output 128-dimensional feature vector;

[0061] (1.2) LSTM structure design: used to process the temporal features of furnace growth data and equipment status data, and fused with the feature vector output by CNN. It contains 2 hidden layers, each with 128 neurons, 128-dimensional input, and 64-dimensional output feature vector. The forget gate threshold is set to 0.8 to avoid gradient vanishing.

[0062] (1.3) Genetic Algorithm Optimization: The 64-dimensional feature vector output by LSTM is used as the initial population. The fitness function is to maximize the crystal quality compliance rate, minimize equipment loss, and maximize the matching degree between process parameters and scenario requirements. The formula is as follows: ( To achieve the crystal quality compliance rate, This is the inverse value of equipment loss. For scene matching, the genetic algorithm parameters are set as follows: population size 100, crossover probability 0.85, mutation probability 0.05, number of iterations 50 generations, selection operator uses roulette wheel selection, crossover operator uses single-point crossover, mutation operator uses Gaussian mutation, and finally outputs the 12 globally optimal core process parameters.

[0063] (1.4) Model inference speed: The generation time of a single furnace process scheme is ≤3 seconds, which meets the needs of rapid industrial production.

[0064] (2) Dynamic correction model: PPO reinforcement learning model

[0065] The Proximal Policy Optimization (PPO) reinforcement learning algorithm is used to solve the problem of dynamic compensation for real-time interference during crystal growth. It takes real-time monitoring data during the growth process as state input, fine-tuning of process parameters as action output, and defect-free crystal growth and parameter stability as reward objectives, so as to achieve millisecond-level dynamic adjustment of process parameters with an adjustment delay of ≤50ms.

[0066] (2.1) Definition of intelligent agent and environment: The dynamic correction model is used as the intelligent agent, and the silicon carbide crystal growth furnace and growth process are used as the interactive environment. The state of the environment changes dynamically with real-time monitoring data.

[0067] (1.2) State space (S): 20-dimensional, including 20 real-time data such as real-time temperature (and fluctuation value), real-time pressure (and fluctuation value), actual gas flow rate, actual seed crystal rotation speed, actual growth rate, infrared spectral defect monitoring signal, stress monitoring signal, encapsulation monitoring signal, and real-time quantitative score of equipment status.

[0068] (2.3) Action Space (A): This is a continuous action space, including temperature, pressure, The fine-tuning range for the five core process parameters—gas flow rate, seed crystal rotation speed, and pulling speed—is ±0.5% ±5% (calibrated according to the process stage: ±0.5% ±1% in the early growth stage, ±1% ±3% in the middle growth stage, and ±3% ±5% in the later growth stage).

[0069] (2.4) Reward function (R): It adopts a combination of sparse reward and dense reward, and the formula is as follows: ,in A bonus is awarded for parameter stability (when the deviation between the actual and target values ​​of process parameters is ≤ ±1%). =5; when the deviation is ±1% to ±3%, =2; when the deviation is >3%, R1 = -5). For defect-free rewards (when no precursors to defects are detected), =10; When a slight precursor to a defect is detected. =0; When obvious signs of impending defects are detected. =-20), As a penalty for equipment loss (when process parameters are close to the equipment constraint threshold), =3; when the threshold is exceeded. =10);

[0070] (2.5) Training and iteration: The offline pre-training + online fine-tuning method is adopted. The offline pre-training uses 2000+ sets of historical data of long crystals and the training iterations are 1 million times. Both the policy network and the value network are 3 fully connected layers with 64 neurons in each layer. The online fine-tuning uses real-time data during the growth process as training samples. The model parameters are updated once after each batch of growth is completed to realize the self-optimization of the model.

[0071] 1.1.4 Scene Adaptation Module

[0072] It enables personalized matching of process solutions with specific application scenarios, and has a built-in standardized multi-scenario parameter library. For the performance priority differences of different application scenarios, it dynamically adjusts the fitness function weight of the main generation model through a weight allocation algorithm to achieve targeted optimization of key process parameters and ensure that the generated process solution fits the actual performance requirements of the scenario.

[0073] Scenario parameter library design: A structured database is used for storage, categorized into four main scenarios: new energy vehicle power devices, AI data center heat dissipation components, AR glasses optical components, and photovoltaic inverter devices. Each scenario predefines performance priorities, key process parameter weights, and core crystal quality indicators, as detailed in the table below:

[0074]

[0075] Weighting Allocation Algorithm: Based on the application scenario selected by the user, the weights of key process parameters in the scenario parameter library are retrieved, and the fitness function weights of the master generative model genetic algorithm are dynamically adjusted. The formula is as follows: ,in For context-specific fitness functions, Based on the fitness function, This is the weight matrix for key process parameters in the scenario.

[0076] Scene parameter library update: Supports manual addition / editing of scenes, and can automatically update the core crystal quality indicators and process parameter weights of each scene through feedback data from the defect tracing and optimization module, so as to achieve continuous optimization of scene adaptation.

[0077] 1.1.5 Host Computer Interaction Module

[0078] Provides operators with a visual, human-machine interactive interface, developed based on a B / S architecture, supporting access from multiple terminals such as computers and industrial tablets. The interface layout is clear, the operation is simple, and it includes seven functional areas. It also supports manual fine-tuning of process solutions, data visualization, and historical data backtracking.

[0079] Parameter input area: Integrates all the functions of the parameter input module, providing standardized input boxes, drop-down selection boxes, and file upload buttons;

[0080] Process Preview Area: Visualizes AI-generated process solutions in the form of curves and tables, including temperature curves, pressure curves, gas flow rate change curves, etc., and supports curve zooming, zooming out, and exporting;

[0081] One-click operation area: Includes four core buttons: "One-click process generation", "One-click process distribution", "One-click process pause", and "One-click process termination", enabling quick process operation;

[0082] Process Export / Save Area: Supports exporting process plans to formats such as Excel, PDF, and CSV. It also supports naming and categorizing process plans for saving, generating unique process numbers for easy retrieval.

[0083] Equipment status display area: Displays the hardware status parameters of the growth furnace in real time in the form of dashboard and bar chart, and issues a red warning when the parameters exceed the normal range;

[0084] Defect Origin Report Area: After growth is completed, a defect origin report is automatically generated, which includes defect type, defect location, process parameter causes, targeted optimization strategies, etc., and supports report export;

[0085] Dynamic Adjustment Trajectory Area: Displays the dynamic adjustment trajectory of process parameters during the growth process in the form of a time-series curve, records the adjustment time, adjustment parameters, adjustment amount, and adjustment reason, and supports trajectory backtracking and analysis;

[0086] Manual fine-tuning function: Operators can manually fine-tune the process plan generated by AI. The fine-tuning range is ±0.5% to ±10%. After fine-tuning, the system will automatically verify whether it meets the equipment constraints. If it does not meet the constraints, a pop-up window will remind the operator and provide correction suggestions.

[0087] 1.1.6 Process Execution and Feedback Module

[0088] It enables the conversion and issuance of process plans from "digital information" to "equipment control instructions," while simultaneously completing real-time data acquisition, defect monitoring signal reception, and process execution status feedback during the growth process. It serves as the core bridge connecting the AI ​​model and the crystal growth equipment, achieving a closed loop of "instruction issuance - data acquisition - status feedback."

[0089] Process instruction conversion and issuance: The process parameters (curves / values) output by the AI ​​dual-model collaborative module are converted into digital instructions that can be recognized by the crystal growth furnace control system according to the Modbus-RTU / TCP protocol. The instruction issuance frequency is matched with the process stage, with an issuance frequency of 10Hz in the early / late growth stage and 5Hz in the middle growth stage. It supports simultaneous issuance to multiple machines, and a single host computer can connect to ≥20 crystal growth furnaces.

[0090] Real-time data acquisition: Sensors deployed in the crystal growth furnace collect actual values ​​of process parameters such as temperature, pressure, gas flow rate, and seed crystal rotation speed during the growth process at a sampling frequency of 50ms. At the same time, real-time equipment status data is also collected and transmitted to the data storage module and dynamic correction model.

[0091] Defect monitoring interface: It connects with infrared spectroscopy detection equipment, stress detection equipment and package detection equipment via industrial Ethernet, supports millisecond-level defect monitoring signal reception, and after the detection signal is converted into a standardized digital signal (0-1, 1 represents the detection of a defect / defect precursor), it is transmitted in real time to the dynamic correction model and the host computer interaction module.

[0092] Process execution status feedback: Real-time acquisition of the process execution status (running / paused / faulted / completed) of the crystal growth furnace. When the status changes, an alert is immediately sent to the host computer interaction module, and the time and reason for the status change are recorded for subsequent analysis.

[0093] 1.1.7 Defect Origin Tracing and Optimization Module

[0094] The system enables precise correlation and traceability between silicon carbide crystal defects and process parameters, and generates targeted process optimization strategies. Simultaneously, it feeds back the traceability data and optimization strategies to the data storage module and the AI ​​dual-model collaboration module, enabling reverse iterative training of the dual models. This is the core support for continuous model optimization.

[0095] Defect type definition: The three main defects in silicon carbide crystal growth (dislocations, inclusions, and microcracks) are standardized and defined, and the characteristic indicators of each defect are clarified (determined by detection signals from optical / X-ray and stress testing equipment).

[0096] Defect-Process Correlation Algorithm: This algorithm combines the Apriori association rule mining algorithm with the XGBoost gradient boosting tree algorithm to accurately locate the causes of defects.

[0097] Apriori algorithm: Set minimum support of 0.2 and minimum confidence of 0.8 to mine the association rules between outliers in process parameters and defect types, such as "temperature fluctuation ≥ ±8℃ and Ar gas flow deviation ≥ ±5sccm → dislocation defect, confidence 0.92";

[0098] XGBoost algorithm: It takes process parameters, equipment status, and scenario requirements as feature inputs and defect type / defect density as label outputs to train a defect prediction model. After the model is trained, it locates the core process parameters that cause defects by ranking the feature importance. The feature importance threshold is set to 0.1.

[0099] Targeted optimization strategy generation: Based on the defect tracing results, quantitative fine-tuning strategies are generated for core process parameters. For example, if "temperature fluctuation ≥ ±8℃" is identified as the core cause of dislocation defects, an optimization strategy of "increasing the temperature control accuracy to ±3℃ and reducing the temperature gradient by 5℃ / cm during the growth phase" is generated. For different defect types, predefined optimization strategy templates are used, and quantitative parameters are automatically filled in based on the tracing results.

[0100] Model back-to-back iteration: Information such as defect type, core inducing process parameters, targeted optimization strategy, and growth data are integrated into a model iteration dataset, which is divided into training set, validation set, and test set in a 7:2:1 ratio. The dataset is fed back to the AI ​​dual-model collaborative module. The main generation model iterates once every 10 furnaces, and the dynamic correction model iterates once per furnace. After iteration, the model accuracy is verified. If the accuracy improves by ≥5%, the model parameters are updated.

[0101] 1.1.8 Data Storage Module

[0102] It provides multi-dimensional, structured, and highly reliable data storage and retrieval services for the entire system, adopting a three-tier storage architecture of "local server + edge computing node + cloud backup" to ensure the real-time performance, security, and integrity of data, while also providing data support for AI model training, defect tracing, and process library updates.

[0103] Storage architecture design:

[0104] Edge computing nodes: Deployed in the crystal growth workshop, responsible for the temporary storage of real-time data (equipment status, growth process parameters, defect monitoring signals), with a storage period of 7 days and a sampling frequency of 50ms, to ensure the low-latency data requirements of the dynamic correction model.

[0105] Local server: Deployed in the enterprise's data center, responsible for the long-term storage of all data, with a storage period of ≥5 years, including all data such as furnace loading and unloading parameters, process plans, equipment status data, defect data, model iteration data, and scenario parameter library;

[0106] Cloud backup: Public cloud object storage is used to perform off-site backup of the core data of the local server. The backup frequency is daily incremental backup and weekly full backup to prevent data loss.

[0107] Data is categorized and stored as follows: Data is divided into 8 categories, each with a unique index (process number / furnace number / timestamp), supporting multi-dimensional retrieval by furnace number, time, scenario, defect type, etc., with a retrieval response time of ≤1 second.

[0108] Crystal parameters for upper and lower furnaces;

[0109] Device status parameter class;

[0110] Scene parameter library class;

[0111] Process solutions;

[0112] Real-time data on the growth process;

[0113] Defect data and source tracing reports;

[0114] Model training and iterative data classes;

[0115] Process execution status class;

[0116] Data format standardization: All stored data adopts a standardized format. Numerical data retains 4 decimal places, time-series data is indexed by timestamps (millisecond level), and text data uses UTF-8 encoding to ensure data interoperability and parsing efficiency.

[0117] 1.2 Technical Solution: Core Methods and Steps

[0118] The intelligent generation and dynamic optimization method for silicon carbide crystal growth process of this invention is based on the above eight core modules, constructing a fully closed-loop process of "input-generation-verification-execution-correction-traceability-iteration-update", which includes six core steps, each step is interconnected, realizing intelligent generation, dynamic control and continuous optimization of silicon carbide crystal growth process. The specific steps are as follows (corresponding to...). Figure 1 AI process generation method flowchart):

[0119] S1: Multi-dimensional data input and standardized preprocessing

[0120] Multi-dimensional data input: Operators can input furnace crystal parameters and furnace target parameters through the host computer interaction module and select specific application scenarios; the equipment status perception module collects the growth furnace hardware status parameters in real time through sensors; the process execution and feedback module automatically collects furnace process execution data from the crystal growth equipment control system; the scenario adaptation module retrieves the corresponding performance priority and process parameter weight from the scenario parameter library according to the scenario selected by the user.

[0121] Data standardization preprocessing: The system performs unified preprocessing on all input data to ensure data validity and consistency. The preprocessing steps are as follows:

[0122] Outlier removal: Outlier data is removed using the 3σ principle, that is, when a data value exceeds μ±3σ (μ is the mean and σ is the standard deviation), it is identified as an outlier, automatically removed, and supplemented using linear interpolation.

[0123] Unit unification: Convert all physical quantities to predefined standard units (such as dimensions mm, resistivity Ω·cm, temperature ℃), and automatically complete the unit conversion;

[0124] Normalization: Map all numerical data to the interval [0,1], using the following formula: Eliminate the influence of dimensions;

[0125] Scenario requirement standardization: The specific performance requirements of the scenario input by the user are converted into feature vectors that can be recognized by the AI ​​model, with a dimension of 16.

[0126] Output: Generate a standardized multi-dimensional dataset, including feature vectors of upper and lower furnace parameters, feature vectors of equipment status, feature vectors of scenario requirements, and equipment process constraints, which will serve as input for subsequent steps.

[0127] S2: Scene Adaptation and Initial Process Scheme Generation

[0128] Scene adaptation weight adjustment: The scene adaptation module retrieves the weight matrix from the scene parameter library based on the scene requirement feature vector in the standardized multi-dimensional dataset, and dynamically adjusts the fitness function weights of the master generator model (CNN-LSTM+genetic algorithm) through the weight allocation algorithm to complete scene adaptation.

[0129] Initial process scheme generation: The standardized multi-dimensional dataset is input into the main generator model of the AI ​​dual-model collaborative module. First, spatial features are extracted through CNN, then temporal features are processed through LSTM, and finally global optimization is performed through genetic algorithm to generate an initial process scheme containing 12 core parameters. The process scheme is transmitted to the host computer interaction module in the form of curves and tables for visualization.

[0130] Output: Initial process plan (JSON format + visual format).

[0131] S3: Multi-dimensional verification and iterative optimization of process constraints

[0132] Constraint settings: The system predefines three main categories of process constraints, all of which are hard constraints and none can be omitted:

[0133] Equipment hardware limit constraints: upper and lower threshold values ​​of process parameters output by the equipment status sensing module;

[0134] Thermodynamic equilibrium constraint: Based on the thermodynamic principle of silicon carbide crystal growth, a synergistic constraint relationship between temperature, pressure, and gas flow rate is predefined. For example, in the PVT process, when the temperature is ≥2200℃, the vacuum degree must be ≤10Pa.

[0135] Scene performance requirements constraints: The core crystal quality indicators retrieved by the scene adaptation module must be met by the process parameters of these indicators.

[0136] Multi-dimensional verification: The system compares the initial process plan with the above three categories of constraints one by one, and verifies whether each process parameter meets the constraint requirements. The verification results are displayed in real time on the host computer interaction module.

[0137] Iterative optimization: If the initial process scheme meets all constraints, proceed directly to the next step; if not, the verification deviation data is fed back to the master generator model, which adjusts the initial population value and fitness function of the genetic algorithm and re-infers and generates the scheme. The number of iterations is ≤3 times until the process scheme meets all constraints.

[0138] Output: Optimized initial process scheme.

[0139] S4: Dynamic process execution and millisecond-level real-time correction

[0140] This step is the actual execution phase of the process, realizing a closed loop of "process issuance - real-time monitoring - dynamic correction - status feedback," and is the core step to ensure crystal growth quality.

[0141] Process instruction issuance: The process execution and feedback module converts the optimized initial process scheme into Modbus protocol instructions and issues them to the crystal growth furnace control system to start the silicon carbide crystal growth process;

[0142] Real-time data and defect signal acquisition: The process execution and feedback module collects the actual values ​​of process parameters and real-time equipment status data during the growth process at a sampling frequency of 50ms. At the same time, it receives the millisecond-level detection signals from the defect monitoring equipment. All data is transmitted in real time to the dynamic correction model, the host computer interaction module and the data storage module.

[0143] Real-time dynamic correction: The dynamic correction model (PPO reinforcement learning) takes the collected real-time data as the state input and outputs the fine-tuning amount of process parameters through model inference. If a defect precursor or process parameter deviation exceeds the threshold is detected, the fine-tuning command is immediately sent to the crystal growth furnace control system to realize millisecond-level dynamic correction of process parameters. The adjustment record is transmitted in real time to the dynamic adjustment trajectory area of ​​the upper computer interaction module.

[0144] Process execution status feedback: The system monitors the process execution status of the crystal growth furnace in real time. If equipment failure, process interruption or other abnormal situation occurs, the process issuance will be suspended immediately, an alarm will be sent to the host computer, and abnormal information will be recorded at the same time.

[0145] Output: Execution record of optimized process scheme, dynamic correction record (including adjustment time, adjustment parameters, adjustment amount, and adjustment reason), and real-time dataset of growth process.

[0146] S5: Defect Origin Tracing and AI Dual-Model Iterative Training

[0147] This step is performed after the single-furnace silicon carbide crystal growth process is completed, enabling precise defect tracing and continuous model optimization, providing more accurate model support for the generation of processes in subsequent furnaces:

[0148] Crystal quality inspection and defect data acquisition: Perform full-dimensional quality inspection on the grown silicon carbide crystal, collect quality data such as defect type, defect density, defect location, crystal size, and resistivity, and transmit them to the defect tracing and optimization module;

[0149] Precise Defect Source Tracing: The defect source tracing and optimization module adopts the Apriori+XGBoost fusion algorithm to correlate and analyze the real-time dataset of the growth process, process plan, equipment status data and defect data, locate the core process parameter causes of defects, and generate a defect source tracing report.

[0150] Targeted optimization strategy generation: Based on the defect tracing results, quantitative targeted process optimization strategies are generated for the core process parameter causes, clarifying the direction and amount of process parameter adjustment;

[0151] Model iteration dataset construction: Defect data, defect tracing results, targeted optimization strategies, real-time growth process datasets, and process execution records are integrated into a model iteration dataset, which is divided into training set, validation set, and test set in a 7:2:1 ratio.

[0152] AI Dual-Model Iterative Training: The model iteration dataset is fed back to the AI ​​dual-model collaborative module. The main generating model is trained iteratively using batch gradient descent. The model parameters are updated when the validation set accuracy is ≥95%. The dynamic correction model is trained iteratively using online fine-tuning. The model parameters are updated when the real-time adjustment response latency is ≤50ms after training.

[0153] Outputs: Defect source tracing report, targeted process optimization strategy, model iteration parameters, and deviation dataset.

[0154] S6: Scenario-based process library updates and reuse

[0155] This step is the final step in the closed-loop process, enabling the process solution to be consolidated and reused, thereby improving the efficiency and accuracy of process generation for subsequent furnaces.

[0156] Process scheme selection: The data storage module selects qualified process schemes based on the quality inspection results of single-furnace crystals (crystal quality meets the core indicators required by the scenario, and yield rate is ≥ target value).

[0157] Process scheme classification and storage: The qualified process scheme is associated with the corresponding furnace parameters, equipment status data, scenario requirements, dynamic correction records, and defect traceability results. It is classified and stored in the scenario-based process library according to three dimensions: application scenario, crystal growth process type, and crystal specifications, generating a unique process number and supporting multi-dimensional retrieval.

[0158] Scene parameter library update: The defect tracing and optimization module feeds back the defect-process association rules and targeted optimization strategies in the tracing results to the scene adaptation module, automatically updating the process parameter weights and crystal quality core indicators in the scene parameter library to achieve continuous optimization of scene adaptation;

[0159] Process library reuse: When generating processes for subsequent furnaces, the system can retrieve compliant process solutions for similar scenarios / specifications from the scenario-based process library as initial reference values ​​for the main generation model, greatly improving the efficiency and accuracy of process generation;

[0160] Feedback: The updated scenario-based process library and scenario parameter library serve as input data for step S1, enabling closed-loop iteration throughout the entire process.

[0161] The AI-driven intelligent generation and dynamic optimization system and method for silicon carbide crystal growth process of the present invention has completed software and hardware deployment and real-world scenario verification, realizing intelligent generation, dynamic correction and defect tracing of silicon carbide crystal growth process. The following is the best implementation method, including three parts: system hardware deployment, AI dual-model training process and practical application case. Those skilled in the art can adapt and implement this implementation scheme under different production scales and crystal growth equipment models.

[0162] 2.1 System Hardware Deployment

[0163] The hardware deployment of this system adopts a four-layer architecture of "perception layer - transmission layer - computing layer - application layer". The hardware selection and deployment methods of each layer are as follows to ensure the real-time performance, stability and scalability of the system and adapt it to industrial-scale production:

[0164] Sensing Layer: This is the foundational layer for data acquisition, deployed next to the silicon carbide crystal growth furnace, and includes:

[0165] Specialized sensors: heating coil temperature sensor (accuracy ±1℃), crucible wear sensor (accuracy ±1μm), vacuum sensor (accuracy ±0.1Pa), gas flow sensor (accuracy ±1sccm), stress sensor (accuracy ±1MPa), etc.

[0166] Defect monitoring equipment: Infrared spectroscopy detection equipment (detection frequency 10Hz, defect identification accuracy ≥98%), package detection equipment (detection accuracy ±5μm);

[0167] Deployment method: All sensors are embedded and seamlessly connected to the crystal growth furnace. The defect monitoring equipment is installed at the observation port of the crystal growth furnace through an industrial bracket to ensure the detection field of view.

[0168] Transport layer: Enables high-speed, low-latency data transmission, including:

[0169] 5G Industrial Gateway: Supports 5G / wired dual links, with a transmission rate of ≥1Gbps and a latency of ≤20ms, enabling the transmission of data from the perception layer to the computing layer and application layer;

[0170] Industrial Ethernet switches: Gigabit speed, PoE power supply supported, enabling networking communication among multiple devices;

[0171] Communication Protocol: The perception layer and the transmission layer adopt the Modbus protocol, and the transmission layer and the computing layer adopt the OPC UA protocol to ensure the standardization and compatibility of data transmission.

[0172] Computational Layer: This is the core computational layer of the system, responsible for AI model inference, data processing, and storage, including:

[0173] Edge computing nodes: Industrial-grade edge computing servers (such as CPU, GPU, memory: 64GB, hard disk: 1TBSSD) are deployed in the crystal growth workshop and are responsible for temporary storage of real-time data and dynamic correction of model inference.

[0174] Local server: Enterprise-grade rack server (such as CPU, GPU, memory: 256GB, hard disk: 10TBSSD) is deployed in the enterprise data center and is responsible for the inference of the main generator model, long-term storage of the full data and iterative training of the model.

[0175] Cloud backup server: Utilizes public cloud object storage with a storage capacity of ≥100TB to achieve off-site backup of core data.

[0176] Application layer: This is the interaction layer for operators, including industrial tablets (10.1 inches, Android system, supporting touch operation) and computers (such as desktop computers, CPU, memory: 32GB, graphics card, deployed in the control room of the crystal growth workshop and the enterprise technology department, supporting simultaneous access and operation by multiple terminals).

[0177] 2.2 AI Dual-Model Training Process

[0178] The AI ​​dual-model (master generator model and dynamic correction model) of this invention adopts a training mode of "offline pre-training + online iteration". After pre-training, it is deployed to the system and iterated online using actual production data. The model accuracy continues to improve with the accumulation of data. The specific training process and parameters are as follows:

[0179] Training data acquisition and preprocessing:

[0180] Scope of data collection: Collect over 2000 sets of historical data on PVT / liquid phase crystal growth from a silicon carbide production company, including crystal parameters of upper and lower furnaces, equipment status data, scenario requirements, defect data, process schemes, and real-time data of the growth process;

[0181] Data preprocessing: Outlier removal, unit unification, and normalization are performed according to the method in step S1. The processed data is divided into training and validation sets in a 7:3 ratio, and a test set (containing 300 sets of actual production data that were not used in training) is constructed.

[0182] Training of the master generative model (CNN-LSTM + genetic algorithm):

[0183] Training steps: First, train the CNN-LSTM model using the Adam optimizer, with a learning rate of 0.001, a batch size of 32, and 100 training iterations. After the validation set loss converges, use the output of the CNN-LSTM as the initial population for the genetic algorithm optimization.

[0184] Training metrics: The matching degree between the validation set process scheme and the actual optimal process scheme is ≥95%, and the process generation time per furnace is ≤3 seconds;

[0185] Test set verification: The crystal quality compliance rate of the test set is ≥92%, and the defect density is reduced by more than 30% on average.

[0186] Training of the Dynamically Corrected Model (PPO Reinforcement Learning):

[0187] Training steps: First, perform offline pre-training using 2000+ sets of historical data as training samples, with 1 million training iterations, a learning rate of 0.0003, a batch size of 64, and after the reward function converges, deploy to edge computing nodes for online fine-tuning.

[0188] Training metrics: Real-time adjustment response delay ≤50ms, process parameter deviation compensation rate ≥98%, and defect precursor identification and correction accuracy ≥95%;

[0189] Test set verification: During the growth process of the test set, the fluctuation range of process parameters was ≤±3%, and the crystal defect rate was reduced by more than 40% on average.

[0190] Online iterative training: After the model is deployed, after each batch of growth is completed, the actual production data is used as new training samples to fine-tune the dual models online. The main generation model iterates once every 10 batches, and the dynamic correction model iterates once per batch. After each iteration, the model indicators are verified, and if the indicators improve, the model parameters are updated.

[0191] 2.3 Practical Application Cases

[0192] To verify the effectiveness and practicality of this invention, a practical application verification was conducted on the PVT crystal growth production line of a silicon carbide production company in China. The following is a production case of SiC crystals for power devices in new energy vehicles, with specific parameters and execution results as follows:

[0193] Scenario requirement: SiC crystals for power devices in new energy vehicles, using PVT process, with key specifications of 205mm diameter and defect density ≤2000cm³. -2 Resistivity 0.015~0.025Ω·cm, thermal stability ≥400℃, yield ≥90%;

[0194] Input data:

[0195] Crystal parameters for furnace loading: diameter 205mm, defect density 2700cm³ -2 Resistivity 0.025 Ω·cm, growth time 72 h, PVT process, yield 85%;

[0196] Target parameters for furnace loading: diameter 205mm, target defect level II (≤3000cm) -2 The target resistivity is 0.020 Ω·cm, the target growth rate is 0.2 mm / h, and the target yield is ≥90%.

[0197] Equipment status parameters: heating mechanism loss rate 15% (quantitative score 85 points), crucible remaining life 80% (quantitative score 80 points), vacuum system leakage rate 0.01 Pa·L / s (quantitative score 90 points), airflow valve response accuracy 98% (quantitative score 98 points).

[0198] Equipment process constraints: maximum heating temperature 2300℃, maximum growth rate 0.22mm / h, Ar gas flow rate range 50~200sccm.

[0199] System execution process:

[0200] S1-S3: The system completes multi-dimensional data preprocessing and scenario adaptation in 3 seconds, generates an initial process plan and passes constraint verification. The core parameters of the optimized initial process plan are: temperature curve (heating to 2250℃ / 10h, holding at 2250℃ / 100h, cooling to room temperature / 20h), pressure curve (10Pa / entire process), Ar gas flow rate 80sccm, seed crystal rotation speed 0.5r / min, and pulling speed 0.05mm / h.

[0201] S4: After the process was issued, a temperature fluctuation of ±5℃ occurred during the growth process. The dynamic correction model responded within 50ms, fine-tuned the heating power by +3%, and quickly compensated the temperature fluctuation to within ±2℃. No defect precursors were detected throughout the process.

[0202] S5: After growth is completed, the crystal quality inspection data is fed back to the system. After analysis by the defect tracing and optimization module, it is determined that there are no defects. The model iteration dataset is updated, and the two models are fine-tuned online.

[0203] S6: This process solution has been determined to be a compliant process solution and will be stored in the scenario-based process library according to the category of "New Energy Vehicle Power Devices / PVT Process / 205mm".

[0204] Execution result:

[0205] Crystal quality: Final crystal defect density 1900 cm⁻¹ -2 With a resistivity of 0.019Ω·cm, thermal stability of 420℃, and diameter deviation of ±0.03mm, it meets the requirements of power devices for new energy vehicles.

[0206] Production efficiency: The process generation time is 3 seconds, which is 4800 times faster than manually writing processes (4 hours). The growth process requires no manual intervention and is 100% automated.

[0207] Equipment wear and tear: With precise matching of process parameters and equipment status, the wear rate of heating mechanism and crucible is reduced by 20% compared with traditional process, and the overall equipment operating cost is reduced by 15%;

[0208] Yield: The yield of this production was 95%, an increase of 10 percentage points compared to the previous batch, meeting the target yield requirement.

[0209] Multi-batch validation: Fifty consecutive batches of SiC crystal production for power devices in new energy vehicles were conducted using the system and method of this invention, achieving an average crystal defect density of 1850 cm⁻¹. -2 The average resistivity is 0.020 Ω·cm, the average yield is 93%, and the process consistency is ≥95%, which verifies the stability and effectiveness of the present invention.

[0210] The above description is merely a preferred embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any variations or substitutions that can be easily conceived by those skilled in the art within the technical scope disclosed in the present invention should be included within the scope of protection of the present invention. Therefore, the scope of protection of the present invention should be determined by the scope of the claims.

Claims

1. An AI-driven intelligent generation and dynamic optimization system for silicon carbide crystal growth process, characterized in that, The system includes: The parameter input module is used to obtain the upper furnace crystal parameters, lower furnace target parameters, and scenario requirements for silicon carbide crystal growth. The equipment status sensing module is used to collect the hardware status parameters of the growth furnace in real time and generate constraints for process adjustment through a hardware-process adaptation algorithm. The AI ​​dual-model collaborative module includes a master generating model and a dynamic correction model; The scenario adaptation module has a built-in multi-scenario parameter library, which is used to dynamically adjust the target weight generated by the process according to the scenario requirements. The process execution and feedback module is used to convert process plans into equipment control commands for issuance, and to collect data and defect monitoring signals during the growth process in real time. The defect tracing and optimization module is used to trace the defects of the grown crystal, generate targeted optimization strategies, and feed them back to the AI ​​dual-model collaborative module for iterative training. The host computer interaction module provides a visual interactive interface; and The data storage module is used to store all data in a hierarchical and categorized manner. The system constructs a fully closed-loop control mechanism through the serial connection and closed-loop feedback between various modules, which includes process generation, equipment adaptation, scenario matching, dynamic correction, defect tracing, and model iteration.

2. The AI-driven intelligent generation and dynamic optimization system for silicon carbide crystal growth process according to claim 1, characterized in that, The device status sensing module specifically includes: Multiple sensors are used to collect hardware status parameters of the heating mechanism, crucible, vacuum system, and airflow control system at a fixed frequency, and to perform real-time quantitative scoring of each parameter; The hardware-process adaptation algorithm unit uses a multiple linear regression algorithm to construct a correlation model between hardware status and process parameters. Based on the quantitative score and the preset process type weight coefficient, it outputs the upper / lower limit threshold of the core process parameters as the constraint condition.

3. The AI-driven intelligent generation and dynamic optimization system for silicon carbide crystal growth process according to claim 1, characterized in that, In the AI ​​dual-model collaborative module: The main generation model is a hybrid model of CNN-LSTM and genetic algorithm. It is used to generate an initial process scheme containing multiple core parameters based on the data from the parameter input module, the equipment status perception module and the scene adaptation module. The model extracts spatial features through CNN, processes temporal features through LSTM, and performs global optimization through genetic algorithm. The dynamic correction model is a model based on the PPO reinforcement learning algorithm. It is used to achieve online dynamic adjustment and closed-loop control of process parameters during crystal growth by taking real-time monitoring data as state input, fine-tuning of process parameters as action output, and defect-free crystal growth and parameter stability as reward objectives.

4. The AI-driven intelligent generation and dynamic optimization system for silicon carbide crystal growth process according to claim 3, characterized in that, The scene adaptation module is further used for: The storage contains a multi-scenario parameter library containing different application scenarios and their corresponding performance priorities, key process parameter weights, and core crystal quality indicators; Based on the scenario selected by the user, the corresponding key process parameter weights are retrieved, and the fitness function of the genetic algorithm in the master generation model is dynamically corrected through a weight allocation algorithm to achieve personalized matching between the process scheme and the subdivided application scenario.

5. The AI-driven intelligent generation and dynamic optimization system for silicon carbide crystal growth process according to claim 1, characterized in that, The defect tracing and optimization module specifically includes: The defect-process correlation algorithm unit integrates the Apriori association rule mining algorithm and the XGBoost gradient boosting tree algorithm to locate the core process parameter causes of defects; The targeted optimization strategy generation unit is used to automatically generate a quantitative process fine-tuning strategy based on the core process parameter causes. The model reverse iteration unit is used to construct an iterative dataset from defect data and source tracing results, and feed it back to the AI ​​dual-model collaborative module to iteratively train the main generating model and the dynamic correction model.

6. A method for intelligent generation and dynamic optimization of silicon carbide crystal growth process driven by AI using the system described in any one of claims 1 to 5, characterized in that, Includes the following steps: S1: Multi-dimensional data input and standardized preprocessing: Multi-dimensional data is acquired through the parameter input module, equipment status perception module and process execution and feedback module, and standardized preprocessing is performed to generate a standardized multi-dimensional dataset; S2: Scene adaptation and initial process scheme generation: The scene adaptation module adjusts the process weights according to the scene requirements, and the main generation model of the AI ​​dual-model collaborative module generates an initial process scheme based on a standardized multi-dimensional dataset. S3: Multi-dimensional verification and iterative optimization of process constraints: Verify whether the initial process scheme meets the constraints of equipment hardware limits, thermodynamic balance and scenario performance requirements. If it does not meet the constraints, iterative optimization is performed until it meets the constraints, and the optimized initial process scheme is output. S4: Dynamic process execution and real-time correction: The process execution and feedback module issues the process plan, and during the growth process, the dynamic correction model dynamically adjusts the process parameters based on real-time data and defect signals; S5: Defect tracing and AI dual-model iterative training: After growth is completed, the defect tracing and optimization module traces the crystal defects, generates targeted optimization strategies, and constructs an iterative dataset to iteratively train the AI ​​dual model. S6: Scenario-based process library update and reuse: Categorize and store the compliant process solutions and their associated data in the scenario-based process library, and update the scenario parameter library for reference in subsequent process generation. 7.The AI-driven SiC crystal growth process intelligent generation and dynamic optimization method of claim 6, wherein, In step S1, the data standardization preprocessing includes: removing outliers using the 3σ principle, unifying physical quantity units, normalizing numerical data to the [0,1] interval, and converting scenario requirements into feature vectors that the model can recognize. 8.The AI-driven SiC crystal growth process intelligent generation and dynamic optimization method of claim 6, wherein, The constraints in step S3 include: Equipment hardware limit constraints, namely the upper and lower limit thresholds of process parameters output by the equipment status sensing module; Thermodynamic equilibrium constraint refers to the synergistic constraint relationship of multiple process parameters predefined based on the silicon carbide crystal growth principle; The scenario performance requirement constraint refers to the process parameter requirements corresponding to the core crystal quality indicators retrieved by the scenario adaptation module. 9.The AI-driven SiC crystal growth process intelligent generation and dynamic optimization method of claim 6, wherein, In step S4, the dynamic adjustment of the dynamic correction model further includes: The actual values ​​of process parameters, equipment status, and defect monitoring signals during the growth process are collected at a frequency of 50ms. Real-time data is input into the PPO reinforcement learning model, and the model infers and outputs one or more of the following fine-tuning parameters: temperature, pressure, gas flow rate, seed crystal rotation speed, and pulling speed. Fine-tuning commands are sent to the crystal growth furnace control system to achieve dynamic compensation for adjustment delays of ≤50ms. 10.The AI-driven SiC crystal growth process intelligent generation and dynamic optimization method of claim 6, wherein, The defect tracing in step S5 specifically includes: The Apriori algorithm was used to mine association rules between outliers in process parameters and defect types. The XGBoost algorithm was used to train a defect prediction model, and the core process parameters that cause defects were located by ranking the importance of features. Based on the source tracing results, a targeted optimization strategy containing quantitative adjustment parameters is generated from a predefined optimization strategy template.

Images