Edge-defined film-fed crystal growth device and method and construction method of quality optimization model

Abstract

The invention discloses an edge-defined film-fed crystal growth device and method and a construction method of a quality optimization model, and relates to the technical field of crystal growth. The construction method of the edge-defined film-fed crystal growth quality optimization model comprises the following steps: acquiring a multi-modal time sequence data set of edge-defined film-fed crystal growth, wherein the multi-modal time sequence data set comprises equipment data, visual data and operation record data; and preprocessing the multi-modal time series data set, and inputting the preprocessed multi-modal time series data set into the neural network model for training to obtain a trained neural network model, namely the edge-defined film-fed crystal growth quality optimization model. The edge-defined film-fed crystal growth quality optimization model replaces traditional logic engine control, can dynamically learn and adjust crystal growth parameters, does not depend on a preset fixed process any more, can reduce operation errors, improves the stability, accuracy and consistency of the crystal growth process, improves the crystal quality and production efficiency, and is suitable for large-scale production. And the stability and the consistency of the crystal quality are improved.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of crystal growth, in particular to a guided mode method crystal growth device and method, and a construction method of a quality optimization model. BACKGROUND

[0002] In the traditional crystal (for example, gallium oxide crystal) growth process, the growth operation relies on manual judgment and experience, and there is a large operation error and inconsistency. Due to the different judgment abilities of each operator and the sensitivity difference of the crystal state, the growth result is difficult to stabilize and the quality is not good. In addition, there is no complete precise correlation between the adjustment behavior of the operator each time and the result of the crystal growth, resulting in high uncertainty in the growth process and great influence of interference factors on the research and development and production process.

[0003] The growth control system on the market at present is usually based on traditional logic engine or programmed control, which needs manual presetting of detailed judgment process and countermeasures, and is difficult to cope with the complexity of various dynamic changes in the growth process. Especially in the absence of accurate judgment and feedback mechanism, the stability and consistency of the crystal quality are difficult to guarantee, the research and development cycle is long, and the production efficiency is low.

[0004] Therefore, the prior art still needs to be improved and developed. SUMMARY

[0005] Based on the above shortcomings of the prior art, the purpose of the present application is to provide a guided mode method crystal growth device and method, and a construction method of a quality optimization model, which aims to solve the problem that the existing crystal growth control system is usually based on traditional logic engine or programmed control, needs manual presetting of detailed judgment process and countermeasures, and is difficult to cope with the complexity of various dynamic changes in the growth process, resulting in difficulty in guaranteeing the stability and consistency of the crystal quality and low production efficiency.

[0006] The technical scheme of the present application is as follows:

[0007] In a first aspect of the present application, a construction method of a guided mode method crystal growth quality optimization model is provided, which comprises the following steps:

[0008] Obtain a multi-modal time series data set of guided mode method crystal growth, which includes device data, visual data and operation record data;

[0009] After the multi-modal time series data set is preprocessed, it is input into a neural network model for training to obtain a trained neural network model, that is, the guided mode method crystal growth quality optimization model.

[0010] Optionally, the equipment data comprises a crystal weight, a crystal weight change speed, a heating power, a pulling rod position, a pulling speed, an in-furnace temperature, an in-furnace pressure, a coil water flow, and a coil water temperature; and / or,

[0011] The visual data comprises a crystal image captured by a camera; and / or,

[0012] The operation record data comprises an adjusted parameter value.

[0013] Optionally, the operation record data comprises an adjusted parameter value, a parameter adjustment timing, a parameter adjustment direction, and a parameter adjustment amplitude.

[0014] The multi-modal time-series data set further comprises environmental data, and the environmental data comprises an environmental humidity and an environmental temperature.

[0015] Optionally, the step of inputting the preprocessed multi-modal time-series data set into the neural network model for training to obtain the trained neural network model comprises:

[0016] The preprocessed multi-modal time-series data set is inputted into the neural network model, and a predicted heating power and a predicted pulling speed are outputted;

[0017] A loss function value is obtained according to the predicted heating power and the predicted pulling speed and real heating power and real pulling speed corresponding to a good crystal quality;

[0018] Parameters in the neural network model are adjusted according to the loss function value until the loss function value converges, and a trained neural network model is obtained.

[0019] Optionally, the neural network model comprises a feature layer and a decision layer constructed based on a stacking generalization strategy and taking an output of the feature layer as a new feature for input;

[0020] The feature layer comprises four sub-models, which are a multi-layer perceptron residual network model, a bidirectional long short-term memory network model, a one-dimensional convolutional neural network model, and an extreme gradient boosting tree network model;

[0021] The decision layer comprises a meta-learner based on an attention mechanism.

[0022] Optionally, the step of inputting the preprocessed multi-modal time-series data set into the neural network model for training to obtain the trained neural network model comprises:

[0023] First, the preprocessed multi-modal time-series data set is inputted into the four sub-models respectively, the four sub-models are trained, and optimal parameters of the four sub-models are saved and fixed;

[0024] Then the meta-learner based on the attention mechanism is trained, the meta-learner automatically learns the weight distribution of the outputs of the four sub-models, the attention weight is normalized through a softmax function, and a trained neural network model is obtained.

[0025] Optionally, the preprocessed multi-modal time series data set is input into the four sub-models respectively, and the step of training the four sub-models specifically comprises:

[0026] After the multi-modal time series data set is cleaned, denoised and standardized, matrix processing is performed to obtain an original feature matrix;

[0027] Then, the original feature matrix is standardized or normalized and input into a multi-layer perceptron residual network model, and the multi-layer perceptron residual network model is trained;

[0028] After the original feature matrix is processed using a sliding window to generate sequence data, it is input into a bidirectional long short-term memory network model, and the bidirectional long short-term memory network model is trained;

[0029] After one-dimensional convolution processing is performed on the original feature matrix, it is input into a one-dimensional convolutional neural network model, and the one-dimensional convolutional neural network model is trained;

[0030] The original feature matrix is input into a limit gradient boosting tree network model, and the limit gradient boosting tree network model is trained.

[0031] In a second aspect of the present application, a guided mode crystal growth device is provided, wherein the guided mode crystal growth device comprises a guided mode crystal growth furnace and a control system, and the control system is embedded with a guided mode crystal growth quality optimization model constructed by the construction method of the present application.

[0032] In a third aspect of the present application, a crystal growth method is provided, wherein the crystal growth method is based on the guided mode crystal growth quality optimization model constructed by the construction method of the present application or based on the guided mode crystal growth device of the present application, and the crystal growth method comprises the following steps:

[0033] The heating power and the pulling speed required for crystal growth are output by the guided mode crystal growth quality optimization model, and the crystal is grown by using the heating power and the pulling speed; or,

[0034] The raw material required for the crystal is placed in the guided mode crystal growth furnace, and then the control system is used to obtain the heating power and the pulling speed required for crystal growth, and the crystal is grown by using the heating power and the pulling speed.

[0035] Optionally, the crystal growth method further includes the following steps:

[0036] After each crystal growth is completed, the guided crystal growth quality optimization model or control system updates and optimizes itself based on crystal quality, appearance, integrity, twinning status, crystal X-ray diffraction parameters, and defect status.

[0037] Beneficial Effects: In this invention, a neural network model is trained using a multimodal time-series dataset including equipment data, visual data, and operation record data. The resulting guided-mode crystal growth quality optimization model can automatically generate decision logic by learning the relationship between different parameters and crystal quality. This guided-mode crystal growth quality optimization model replaces traditional logic engine control, enabling dynamic learning and adjustment of crystal growth parameters (such as pulling speed and heating power). It no longer relies on a preset fixed process, reducing operational errors, improving the stability, accuracy, and consistency of the crystal growth process, enhancing crystal quality and production efficiency, and further improving the stability and consistency of crystal quality. Specifically, the guided-mode crystal growth quality optimization model can analyze data (such as equipment data and crystal images) in real time during the current growth process and automatically adjust relevant parameters (such as pulling speed and heating power) to optimize each step of the crystal growth process, ensuring the stability and consistency of crystal growth, improving the accuracy of the growth process and crystal quality. It achieves a high degree of automation and intelligence in the crystal growth process, can replicate the operating strategies of highly skilled operators, avoid human error, reduce judgment differences in manual operation, reduce human intervention and operational errors, avoid losses caused by improper human operation, ensure the standardization and consistency of the growth process, improve production efficiency, and reduce the production cycle.

[0038] In addition, the crystal growth quality optimization model provided by the present invention has a self-learning function, which can provide data feedback in each experiment and production process, avoid losses caused by improper human operation, continuously optimize its judgment logic, help R&D personnel quickly verify the impact of different process parameters on crystal quality, improve process optimization efficiency, and also improve the success rate and stability of crystal growth operation. Attached Figure Description

[0039] Figure 1 This is a schematic diagram illustrating the construction process of the crystal growth quality optimization model using the guided model method.

[0040] Figure 2 This is a partial data graph used for training the crystal growth quality optimization model using the guided model method.

[0041] Figure 3 This is a photograph of the crystal front during crystal growth using the guided model method.

[0042] Figure 4 This is a schematic diagram illustrating the training of the crystal growth quality optimization model using the guided model method. Detailed Implementation

[0043] This invention provides a crystal growth apparatus and method using the guided-mode method, and a method for constructing a quality optimization model. To make the objectives, technical solutions, and effects of this invention clearer and more explicit, the invention is further described in detail below. It should be understood that the specific embodiments described herein are merely illustrative of the invention and are not intended to limit the invention.

[0044] Unless otherwise defined, all technical and scientific terms used herein have the same meaning as commonly understood by one of ordinary skill in the art to which this invention pertains. The terminology used in this specification is for the purpose of describing particular embodiments only and is not intended to be limiting of the invention.

[0045] If the embodiments of the present invention involve descriptions such as "first" or "second", such descriptions are for descriptive purposes only and should not be construed as indicating or implying their relative importance or implicitly specifying the number of technical features indicated.

[0046] This invention provides a method for constructing a quality control model for crystal growth using the Edge-defined Film-fed Growth (EFG) method (i.e., a model for controlling or optimizing the quality of crystal growth using the EFG method, specifically a prediction model for heating power and pulling speed in high-quality crystal growth using the EFG method). The method includes methods such as... Figure 1 As shown, it includes the following steps:

[0047] S1. Obtain a multimodal time-series dataset for crystal growth using the guided mode method. The multimodal time-series dataset includes device data, visual data, and operation record data.

[0048] S2. The multimodal time series dataset is preprocessed and then input into the neural network model for training to obtain the trained neural network model, which is the crystal growth quality optimization model of the guided mode method.

[0049] In this embodiment, a neural network model is trained using a multimodal time-series dataset including equipment data, visual data, and operation record data. The resulting guided-mode crystal growth quality optimization model can automatically generate decision logic by learning the relationship between different parameters and crystal quality. This guided-mode crystal growth quality optimization model replaces traditional logic engine control, enabling it to dynamically learn and adjust crystal growth parameters (such as pulling speed and heating power). It no longer relies on a preset fixed process, reducing operational errors, improving the stability, accuracy, and consistency of the crystal growth process, and ultimately enhancing crystal quality and production efficiency. Specifically, the guided-mode crystal growth quality optimization model can analyze data (such as equipment data and crystal images) in real time during the current growth process and automatically adjust relevant parameters (such as pulling speed and heating power) to optimize each step of the crystal growth process, ensuring the stability and consistency of crystal growth, improving the accuracy of the growth process and crystal quality. It achieves a high degree of automation and intelligence in the crystal growth process, can replicate the operating strategies of highly skilled operators, avoid human error, reduce judgment differences in manual operation, reduce human intervention and operational errors, avoid losses caused by improper human operation, ensure the standardization and consistency of the growth process, improve production efficiency, and reduce the production cycle.

[0050] In addition, the crystal growth quality optimization model provided by the present invention has a self-learning function, which can provide data feedback in each experiment and production process, avoid losses caused by improper human operation, continuously optimize its judgment logic, help R&D personnel improve process optimization efficiency, and also improve the success rate and stability of crystal growth operation.

[0051] In this embodiment, by integrating equipment data, visual data, and operation records, multi-dimensional dynamic optimization of the crystal growth process is achieved, significantly improving the stability and success rate of growth.

[0052] In step S1, the multimodal time series dataset is randomly divided into a training dataset and a validation dataset. The neural network model is trained using the training dataset and validated using the validation dataset, thereby achieving complete training of the neural network model and obtaining a trained neural network model.

[0053] In some embodiments, the device data includes crystal weight, crystal weight change rate, heating power, lifting rod position, lifting speed, furnace temperature, furnace pressure, coil water flow rate, and coil water temperature.

[0054] In other words, during crystal growth, data is collected sequentially based on the following parameters: crystal weight, rate of change of crystal weight, heating power, lifting rod position, lifting speed, furnace temperature, furnace pressure, coil water flow rate, and coil water temperature. Specifically, this data can be collected using various sensors, with the collection interval determined according to specific needs, generally between 1 and 60 seconds, for example, 1 second, 5 seconds, 10 seconds, 15 seconds, 20 seconds, 25 seconds, 30 seconds, 40 seconds, 50 seconds, or 60 seconds. The data also includes the time taken to collect the crystal weight, rate of change of crystal weight, heating power, lifting rod position, lifting speed, furnace temperature, furnace pressure, coil water flow rate, and coil water temperature.

[0055] In some implementations, the visual data includes crystal images captured by a camera.

[0056] In other words, during crystal growth, crystal images (i.e., images of the crystal front or photographs) are acquired sequentially to capture the dynamic changes in the crystal growth state. Specifically, the time interval for acquiring crystal images can be determined according to specific needs, for example, it could be 1 minute. Furthermore, the crystal images acquired by the camera are displayed on a screen for real-time monitoring of crystal growth, and the real-time monitoring footage can be analyzed at a frequency of 5 seconds per frame. The visual data also includes the time taken for the camera to acquire the crystal images.

[0057] In some implementations, the operation log data includes adjusted parameter values.

[0058] In some implementations, the operation record data includes adjusted parameter values ​​(such as adjusted input heating power and lifting speed), the timing (or time) of parameter adjustment, the direction of parameter adjustment (e.g., increasing or decreasing heating power, pulling the lifting rod up or down, increasing or decreasing the lifting speed), and the magnitude of parameter adjustment.

[0059] In other words, during crystal growth, the adjusted parameter values, the timing of parameter adjustments, the direction of parameter adjustments, and the magnitude of parameter adjustments are collected in real time. In this embodiment, the correlation between the operator's actions and changes in the crystal growth state (or quality) can be established by recording the operator's actions. The recording interval of the operation data can be determined according to specific needs, generally between 1 and 60 seconds, for example, 1 second, 5 seconds, 10 seconds, 15 seconds, 20 seconds, 25 seconds, 30 seconds, 40 seconds, 50 seconds, or 60 seconds, etc.

[0060] In this embodiment of the invention, particular attention is paid to the decision-making behavior of highly skilled operators. The model can record and learn the specific timing, magnitude, and direction of parameter adjustments made by operators. This dynamic learning mechanism enables the model to simulate the operational behavior of highly skilled personnel, thereby effectively replicating excellent process capabilities.

[0061] In some embodiments, the multimodal time-series dataset further includes environmental data, including ambient humidity and ambient temperature (specifically, workshop humidity and workshop temperature). Because workshop humidity and temperature can also affect crystal growth, this embodiment collects data on workshop humidity and temperature at intervals that can be determined according to specific needs, generally between 1 and 60 seconds, for example, 1 second, 5 seconds, 10 seconds, 15 seconds, 20 seconds, 25 seconds, 30 seconds, 40 seconds, 50 seconds, or 60 seconds. The environmental data also includes the time taken to collect the ambient humidity and temperature data.

[0062] like Figure 2 The data shown includes some training parameters such as power, weight, growth rate, lifting rod position, and lifting speed. The AI ​​left-side recognition and AI right-side recognition data are detection data and are not used as training data. AI left-side recognition indicates whether the left-side crystal image during single-crystal growth is identified as a single crystal or a mixed crystal; GOOD indicates a single crystal, and the value after COOD indicates the probability of a single crystal. The same logic applies to AI right-side recognition.

[0063] Furthermore, in this invention, the multimodal time-series dataset may include not only device data, visual data, and operation record data, but also data derived from these data. Examples include relative time (i.e., the time interval between the acquisition time of each data point and the initial time), growth rate (i.e., the rate of change of crystal weight) / pulling speed, growth rate / crystal weight, crystal weight / pulling speed, growth rate × pulling speed, crystal weight × pulling speed, etc.

[0064] In step S2, in some embodiments, the step of preprocessing the multimodal time-series dataset and then inputting it into a neural network model for training to obtain a trained neural network model specifically includes:

[0065] The preprocessed multimodal time series dataset is input into the neural network model, which outputs the predicted heating power and lifting speed.

[0066] Based on the predicted heating power and pulling speed, and assuming good crystal quality (i.e., X-ray diffraction (XRD) test results of the crystal, the half-width at half maximum (FWHM) of the rocking curve is <150°, and the dislocation density is less than 10⁻⁶),... 3 cm -3 The loss function value is obtained by determining the actual heating power and lifting speed corresponding to the time.

[0067] The parameters in the neural network model are adjusted based on the loss function value until the loss function value converges, thus obtaining a trained neural network model.

[0068] Specifically, the loss function value is obtained based on the loss function and according to the predicted heating power and pulling speed, as well as the actual heating power and pulling speed corresponding to good crystal quality. The loss function includes a mean square error loss function, a smoothed average absolute error loss function, and a root mean square error loss function.

[0069] For example, such as Figure 4 As shown, the crystal weight, crystal weight change rate, heating power, lifting rod position, lifting speed, furnace pressure, coil water temperature, and crystal image (e.g.) are displayed. Figure 3 (As shown) After preprocessing, the input features are fed into the hidden layer of the neural network for training, resulting in a trained neural network model. The trained neural network model can then provide the corresponding heating power and pulling speed based on the input features, allowing for real-time adjustment of the heating power and pulling speed during crystal growth to obtain high-quality crystals.

[0070] In some implementations, the neural network model includes a feature layer and a decision layer constructed based on a stacked generalization strategy, which takes the output of the feature layer as input as new features.

[0071] The feature layer includes four sub-models: a multilayer perceptron (MLP) residual network model, a bidirectional long short-term memory (LSTM) network model, a one-dimensional convolutional neural network (1D CNN) model, and an extreme gradient boosting tree (XGBoost) network model.

[0072] The decision layer includes an attention-based meta-learner. Specifically, the attention-based meta-learner comprises a two-layer structure: the first layer contains 64 neurons and uses the Swish activation function, and the second layer is directly mapped to a two-dimensional output space.

[0073] The preprocessed multimodal time-series dataset is input into four sub-models for initial feature extraction. Specifically, an MLP residual network model learns nonlinear mappings, a bidirectional LSTM network model processes temporal dimension information, a 1D CNN model mines local correlations between features, and an XGBoost network model captures feature importance ranking. In this embodiment, a stacked generalization strategy is employed, using the outputs of the four sub-models as new features to construct a meta-learner. The meta-learner is designed as a two-layer neural network structure: the first layer contains 64 neurons and uses the Swish activation function, while the second layer directly maps to the two-dimensional output space. Furthermore, a feature attention layer is introduced into the meta-learner to automatically learn the weight allocation of each sub-model's output. The attention weights are normalized using a softmax function, enabling the model to dynamically adjust the importance of different sub-models based on the input.

[0074] In some implementations, the step of preprocessing the multimodal time-series dataset and then inputting it into a neural network model for training to obtain a trained neural network model specifically includes:

[0075] First, the multimodal time series dataset is preprocessed and then input into the four sub-models respectively. The four sub-models are trained, and their optimal parameters are saved and fixed.

[0076] Then, the attention-based meta-learner is trained. The meta-learner automatically learns the weight assignments of the four sub-model outputs. The attention weights are normalized by the softmax function (so that the model can dynamically adjust the importance of different sub-models according to the input) to obtain the trained neural network model.

[0077] In this embodiment, the step-by-step training strategy of first training four sub-models and then training the meta-learner avoids gradient conflicts and improves training efficiency. In this embodiment, a learnable fusion weight vector can also be designed to dynamically adjust the contribution of each model to the final prediction during training. In the early stages of training, the XGBoost network model is given higher weights to utilize its strong initialization capability; as training progresses, the weight ratio of the deep learning model is gradually increased. The final prediction result of this model is generated by combining the meta-learner output and a weighted average method. The weighting coefficients are dynamically adjusted based on the performance of the validation dataset to ensure optimal prediction results under different conditions. Furthermore, this invention further analyzes the prediction results of the guided-mode crystal growth quality optimization model using SHAP values. This not only provides the final prediction value but also explains the contribution of each sub-model and feature to the prediction result, enhancing the interpretability of the model.

[0078] In step S2, in some embodiments, the multimodal time-series dataset is preprocessed and then input into four sub-models respectively. The specific steps for training the four sub-models are as follows: steps S21 to S22:

[0079] S21. After cleaning, denoising and standardizing the multimodal time series dataset, perform matrix processing to obtain the original feature matrix;

[0080] In this step, preprocessing of the multimodal time-series dataset, including cleaning, denoising, and standardization, ensures the accuracy and consistency of the data input into the neural network model. For example, crystal growth data (such as crystal weight, crystal weight change rate, heating power, lifting rod position, lifting speed, furnace pressure, coil water temperature, and crystal images) are normalized; edge detection and feature extraction are performed on crystal images captured by cameras; zero-point correction and outlier removal are performed on data collected by sensors, such as equipment data; and time-series annotation is applied to operation record data.

[0081] S22. Then, the original feature matrix is ​​standardized or normalized (converted into vector form data suitable for MLP processing), and input into the MLP residual network model to train the MLP residual network model.

[0082] After the original feature matrix is ​​processed using a sliding window to generate sequence data, it is input into a bidirectional LSTM network model to train the LSTM network model.

[0083] After performing one-dimensional convolution on the original feature matrix, it is input into a 1D CNN model for training.

[0084] The original feature matrix is ​​input into the XGBoost network model to train the XGBoost network model.

[0085] In this step, customized preprocessing was performed on the input data for different sub-models. Specifically, for the MLP residual network model, the original feature matrix needs to be standardized or normalized to convert it into vector data suitable for MLP processing; for the bidirectional LSTM network model, a sliding window is used to generate sequence data; for the 1D CNN model, the original feature matrix is ​​processed by one-dimensional convolution; and for the XGBoost network model, the original feature matrix is ​​used.

[0086] The following is a detailed explanation of each model.

[0087] For MLP residual network models:

[0088] MLP residual network models can be built using PyTorch. They can standardize the input and output data, making features of different dimensions comparable and accelerating model convergence.

[0089] During the prediction phase, the MLP residual network model uses the same normalizer to transform the new data, and the core of the residual block is the skip connection (out += residual). This structure allows the network to learn the residual mapping between the input and output, solving the gradient vanishing and degradation problems in deep neural networks. BatchNorm and Dropout enhance the model's stability and generalization ability.

[0090] The MLP residual network model consists of an input layer, hidden layers, and an output layer, using fully connected layers for information propagation. The input layer maps multi-dimensional input features to the hidden layer dimension, and multiple residual blocks are cascaded to form a deep network. The output layer maps the hidden representations to a 2-dimensional output space and outputs a non-linear result through the ReLU activation function, enabling the model to learn complex patterns.

[0091] Specifically, the MLP residual network model is trained using the mean squared error (MSE) loss function until it converges (the MSE loss function is suitable for regression problems). The loss function value is obtained based on the MSE loss function and the predicted heating power and pulling speed, as well as the actual heating power and pulling speed corresponding to good crystal quality. The parameters in the MLP residual network model are adjusted according to this loss function value until it converges. Simultaneously, the Adam optimizer (combining the advantages of Adagrad and RMSProp) is used for optimization, and L2 regularization is implemented using the weight_decay parameter to prevent overfitting. When the validation loss stagnates, the learning rate is automatically reduced, which helps the model find a better solution in the later stages of training, improving training stability, especially suitable for deep networks and cases using the ReLU activation function.

[0092] For the bidirectional LSTM network model:

[0093] It can be built using PyTorch to process time-series crystal growth data (i.e., crystal weight, crystal weight change rate, heating power, lifting rod position, lifting speed, furnace pressure, coil water temperature, crystal images, etc.) using a bidirectional LSTM network model. The input data is first generated into sequence samples through a sliding window, and then fed into the bidirectional LSTM network model. The bidirectional LSTM network model contains three LSTM hidden layers, each with 64 neurons, capturing bidirectional dependencies in the time-series data through its bidirectional structure.

[0094] To address the vanishing gradient problem in long sequence training, Layer Normalization is used instead of Batch Normalization, with ReLU as the activation function. An attention mechanism is added before the output layer to automatically assign weights to features at different time steps, enhancing the ability to perceive key temporal features.

[0095] The bidirectional LSTM network model is trained using the smoothed mean absolute error (Huber) loss function. Based on the Huber loss function, and taking into account both the predicted heating power and pulling speed, as well as the actual heating power and pulling speed corresponding to good crystal quality, the loss function value is obtained. The parameters in the bidirectional LSTM network model are adjusted according to the loss function value until convergence, and then optimized using the AdamW optimizer. The model uses the Huber loss function to balance the influence of outliers, combined with the AdamW optimizer for parameter updates, and employs an early stopping strategy to prevent overfitting. During prediction, standardized time-series data is input into the model, outputting predicted sequences of power and pulling speed, which are then destandardized to restore the true values.

[0096] For 1D CNN models:

[0097] Considering the local correlations among crystal growth parameters, a 1D CNN-based model was designed. The input data (i.e., the preprocessed multimodal time-series dataset) is first dimension-mapped through a feature embedding layer, and then spatial correlation patterns between features are extracted through multiple convolutional blocks. Each convolutional block contains a 1D convolutional layer, a batch normalization layer, and a Leaky ReLU activation layer, with the kernel size gradually decreasing from 7 to 3 to capture feature dependencies at different scales.

[0098] The model introduces residual connections across convolutional blocks to alleviate gradient problems in deep networks. After feature extraction, a global average pooling layer compresses the feature map into a fixed-length vector, which is then mapped to the two-dimensional output space through a fully connected layer. To enhance the model's generalization ability, Dropout regularization is applied before the fully connected layer, and a cyclic learning rate strategy is used to dynamically adjust the learning rate. During training, the root mean square error (RMSE) loss function is used until it converges, and gradient accumulation techniques are combined to process large amounts of data to train the model, improving training efficiency. Specifically, the loss function is based on the RMSE loss function, and the loss function value is obtained according to the predicted heating power and pulling speed, as well as the actual heating power and pulling speed corresponding to good crystal quality. The parameters in the 1D CNN model are adjusted according to the loss function value until the loss function value converges.

[0099] For the XGBoost network model:

[0100] As a representative algorithm of ensemble learning, the XGBoost network performs excellently in handling nonlinear regression problems. In this invention, for a multimodal time-series dataset of crystal growth, feature importance analysis is first performed to select feature combinations that significantly influence the output. The model adopts a multi-objective regression mode, simultaneously predicting two objective variables: heating power and pulling speed.

[0101] For parameter optimization, Bayesian optimization is used to search for optimal hyperparameter combinations, including maximum tree depth (8-12), learning rate (0.01-0.1), and subsampling rate (0.7-0.9). To handle complex interactions between features, the XGBoost feature cross option is enabled, and the gamma parameter is set to control the tree complexity. An early stopping mechanism is introduced to monitor validation set error and prevent overfitting.

[0102] During training, k-fold cross-validation was used to evaluate model stability, and the prediction results of multiple base models were finally integrated. In the prediction phase (i.e., when using the trained neural network model to predict the heating power and pulling speed required for crystal growth), the input data underwent the same standardization process as in the training phase. After outputting the predicted values ​​through the XGBoost model, inverse standardization was performed to obtain the final prediction result. This model demonstrates strong robustness and interpretability when handling small-sample, high-dimensional crystal growth data.

[0103] This invention also provides a crystal growth apparatus for the guided-mode method, which includes a guided-mode crystal growth furnace and a control system. The control system is embedded with a guided-mode crystal growth quality optimization model constructed using the construction method described above.

[0104] During crystal growth, the guided-mode crystal growth quality optimization model analyzes growth status data provided by sensors and cameras in real time, as well as personnel operation data, to comprehensively determine the current crystal growth status. If the guided-mode crystal growth quality optimization model detects that operation adjustments or growth parameters deviate from the optimal range, it will automatically adjust the parameters (such as the heating power and pulling speed of crystal growth, and determine whether remelting has occurred and to what extent, based on real-time camera images), or prompt the operator to make corrections.

[0105] The "one-click crystal growth" device provided by this invention collects a large amount of operation data and screen information, and applies this data to the crystal growth process, resulting in a crystal growth success rate that is significantly higher than that of manual operation.

[0106] This invention also provides a crystal growth method, wherein the crystal growth method comprises the following steps, based on the model-guided crystal growth quality optimization model constructed by the construction method described above or the model-guided crystal growth apparatus described above:

[0107] The heating power and pulling speed required for crystal growth are input into the crystal growth quality optimization model using the guided model method, and the crystal is grown using the heating power and pulling speed.

[0108] Alternatively, the raw materials required for crystal growth can be placed in a mold-guided crystal growth furnace, and then the heating power and pulling speed required for crystal growth can be obtained using a control system. The crystal can then be grown using the heating power and pulling speed.

[0109] In some embodiments, the crystal growth method further includes the following steps:

[0110] After each crystal growth is completed, the crystal growth quality optimization model of the guided model method updates and optimizes itself based on crystal quality, appearance, integrity, twinning status, crystal X-ray diffraction (XRD) parameters, and defect status.

[0111] In this implementation, the model's judgment accuracy and optimization ability are gradually improved by continuously accumulating and learning new experimental data.

[0112] In a specific implementation, the crystal quality test data, the entire growth process, and the thermal field information of the furnaces with better crystal quality are matched one by one. Then, the matched data are organized into training data, which includes crystal images (or photos), crystal XRD parameters, defect test data, crystal growth process, and thermal field information as input values. The model is then trained a second time to obtain a more accurate model that can grow high-quality crystals.

[0113] This invention also provides a computer-readable storage medium storing a computer program that, when executed by a processor, implements the construction method described above or the crystal growth method described above.

[0114] This invention also provides an electronic device, which includes a memory and a processor. The memory stores a computer program that can run on the processor. When the computer program is executed by the processor, it implements the construction method of this invention as described above or the crystal growth method of this invention.

[0115] In some embodiments, the crystal is gallium oxide crystal. Gallium oxide (GaO) crystals, or single crystals, are widely used in high-voltage electronic devices, ultraviolet optoelectronic devices, and other fields.

[0116] It should be understood that the application of the present invention is not limited to the examples above. Those skilled in the art can make improvements or modifications based on the above description, and all such improvements and modifications should fall within the protection scope of the appended claims.

Claims

1. A method for constructing a crystal growth quality optimization model using the guided model method, characterized in that, The method comprises the following steps: obtaining a multi-modal time series data set of a crystal growth by a guiding mode method, the multi-modal time series data set comprising equipment data, visual data and operation record data; inputting the pre-processed multi-modal time series data set into a neural network model for training to obtain a trained neural network model, i.e., the crystal growth quality optimization model by the guiding mode method.

2. The construction method of claim 1, wherein, the equipment data comprises crystal weight, crystal weight change rate, heating power, pulling rod position, pulling speed, furnace temperature, furnace pressure, coil water flow and coil water temperature; and / or, the visual data comprises crystal images captured by a camera; and / or, the operation record data comprises adjusted parameter values.

3. The construction method of claim 2, wherein, the operation record data comprises adjusted parameter values, parameter adjustment timing, parameter adjustment direction and parameter adjustment amplitude; the multi-modal time series data set further comprises environmental data, the environmental data comprising environmental humidity and environmental temperature.

4. The construction method according to any one of claims 1 to 3, characterized in that, The step of inputting the pre-processed multi-modal time series data set into a neural network model for training to obtain a trained neural network model comprises: inputting the pre-processed multi-modal time series data set into the neural network model to output predicted heating power and pulling speed; obtaining a loss function value according to the predicted heating power and pulling speed and the real heating power and pulling speed corresponding to good crystal quality; adjusting parameters in the neural network model according to the loss function value until the loss function value converges to obtain the trained neural network model.

5. The construction method according to claim 4, characterized in that, The neural network model comprises a feature layer and a decision layer constructed based on a stacking generalization strategy and taking the output of the feature layer as new features for input; the feature layer comprises four sub-models, i.e., a multi-layer perceptron residual network model, a bidirectional long short-term memory network model, a one-dimensional convolutional neural network model and an extreme gradient boosting tree network model; the decision layer comprises a meta-learner based on an attention mechanism.

6. The construction method of claim 5, wherein, The step of inputting the pre-processed multi-modal time series data set into a neural network model for training to obtain a trained neural network model comprises: first, inputting the pre-processed multi-modal time series data set into the four sub-models respectively, training the four sub-models, saving and fixing the optimal parameters of each sub-model; then, training the meta-learner based on the attention mechanism, the meta-learner automatically learning the weight distribution output by the four sub-models, the attention weight being normalized by a softmax function to obtain the trained neural network model.

7. The construction method of claim 6, wherein, The step of inputting the pre-processed multi-modal time series data set into the four sub-models respectively and training the four sub-models comprises: after cleaning, denoising and standardizing the multi-modal time series data set, performing matrix processing to obtain an original feature matrix; then, after standardizing or normalizing the original feature matrix, inputting the original feature matrix into the multi-layer perceptron residual network model to train the multi-layer perceptron residual network model; After the original feature matrix is processed using a sliding window to generate sequence data, the sequence data is input into a bidirectional long short-term memory network model, and the bidirectional long short-term memory network model is trained; After the original feature matrix is processed using one-dimensional convolution, the one-dimensional convolutional neural network model is input into a one-dimensional convolutional neural network model, and the one-dimensional convolutional neural network model is trained; The original feature matrix is input into a limit gradient boosting tree network model, and the limit gradient boosting tree network model is trained.

8. A method of growing a crystal by the edge defined film growth method, characterized by, The guided mode method crystal growth device comprises a guided mode method crystal growth furnace and a control system, and the control system is embedded with the guided mode method crystal growth quality optimization model constructed by using the construction method according to any one of claims 1-7.

9. A method of growing a crystal, characterized by, The crystal growth method is based on the guided mode method crystal growth quality optimization model constructed by using the construction method according to any one of claims 1-7 or the guided mode method crystal growth device according to claim 8, and the crystal growth method comprises the following steps: The heating power and the pulling speed required for crystal growth are output by using the guided mode method crystal growth quality optimization model, and the crystal is grown by using the heating power and the pulling speed; or, The raw material required for the crystal is placed in the guided mode method crystal growth furnace, and then the heating power and the pulling speed required for crystal growth are obtained by using the control system, and the crystal is grown by using the heating power and the pulling speed.

10. The growth method of claim 9, wherein, The crystal growth method further comprises the following steps: After each crystal growth is completed, the guided mode method crystal growth quality optimization model or the control system is updated and optimized according to the crystal quality, appearance, integrity, twinning condition, crystal X-ray diffraction parameters, and defect condition.

Images