A gallium oxide preparation method and system based on deep learning and crucible lowering method

Abstract

The application relates to a gallium oxide preparation method and system based on deep learning and a crucible lowering method, and a prediction method comprises the following steps: acquiring preparation data of a gallium oxide single crystal; the preparation data comprises seed crystal data, environment data and control data; the preparation data is preprocessed to obtain preprocessed preparation data; and the preprocessed preparation data is input into a trained neural network model to obtain corresponding predicted property data of the gallium oxide single crystal. The preparation data is preprocessed to obtain preprocessed preparation data, the preprocessed preparation data is input into the trained neural network model to obtain corresponding predicted property data of the gallium oxide single crystal, the performance of the gallium oxide single crystal is predicted through the trained neural network model, the preparation data is adjusted to obtain the required performance of the gallium oxide single crystal, and the performance of the gallium oxide single crystal is optimized.

Description

Technical Field

[0001] This application relates to the field of gallium oxide preparation technology, and in particular to a gallium oxide preparation method and system based on deep learning and crucible lowering method. Background Technology

[0002] Gallium oxide (Ga₂O₃) single crystal is a transparent semiconductor oxide, belonging to the wide bandgap semiconductor materials. Generally, β-phase gallium oxide (β-Ga₂O₃) is more stable. β-Ga₂O₃ possesses many advantages, including a large bandgap, high saturated electron drift velocity, high thermal conductivity, high breakdown field strength, and chemical stability. Its high bandgap results in a high breakdown voltage, and combined with its high saturated electron drift velocity, high thermal conductivity, and chemical stability, β-Ga₂O₃ single crystals have broad application prospects in the field of electronic devices. The crucible lowering method is one method for preparing gallium oxide. However, when using this method, the preparation of gallium oxide is difficult to control, and too many factors influence the performance of the gallium oxide product, making it difficult to obtain gallium oxide with optimal properties.

[0003] Therefore, existing technologies need to be improved. Summary of the Invention

[0004] The technical problem to be solved by the present invention is to provide a gallium oxide preparation method and system based on deep learning and crucible lowering method, so as to optimize the performance of gallium oxide.

[0005] This invention provides a gallium oxide prediction method based on deep learning and crucible descent, comprising:

[0006] Acquire preparation data for gallium oxide single crystals; wherein, the preparation data includes: seed crystal data, environmental data, and control data;

[0007] The preparation data is preprocessed to obtain preprocessed preparation data;

[0008] The preprocessed preparation data is input into a trained neural network model, and the predicted property data corresponding to the gallium oxide single crystal is obtained through the trained neural network model.

[0009] The gallium oxide prediction method based on deep learning and crucible descent method, wherein the preprocessing of the preparation data to obtain preprocessed preparation data includes:

[0010] Based on the seed crystal data, the environmental data, and the control data, preprocessing preparation data is obtained; wherein, the preprocessing preparation data is a matrix formed by the seed crystal data, the environmental data, and the control data.

[0011] The gallium oxide prediction method based on deep learning and crucible descent method includes seed crystal data such as: seed crystal diffraction peak full width at half maximum (FWHM), seed crystal diffraction peak FWHM deviation value, and seed crystal diameter.

[0012] The environmental data includes: thermal resistance value of the insulation layer in the high temperature zone, thermal resistance value deviation value of the insulation layer in the high temperature zone, shape factor of the insulation layer in the high temperature zone, thermal resistance value of the insulation layer in the low temperature zone, thermal resistance value deviation value of the insulation layer in the low temperature zone, shape factor of the insulation layer in the low temperature zone, and shape factor of the insulation layer in the growth-driven zone.

[0013] The control data includes: input power in the high-temperature zone, cooling power in the high-temperature zone, input power in the low-temperature zone, cooling power in the low-temperature zone, and crucible descent speed.

[0014] The gallium oxide prediction method based on deep learning and crucible descent, wherein obtaining preprocessed preparation data based on the seed crystal data, the environmental data, and the control data includes:

[0015] Based on the seed crystal data, the environmental data, and the control data, a preparation vector is determined; wherein, the first element of the preparation vector is one of the seed crystal diffraction peak full width at half maximum (FWHM), the seed crystal diffraction peak FWHM deviation value, and the seed crystal diameter; the second element of the preparation vector is one of the thermal resistance value of the high-temperature zone insulation layer, the thermal resistance value deviation value of the high-temperature zone insulation layer, the shape factor of the high-temperature zone insulation layer, the thermal resistance value of the low-temperature zone insulation layer, the thermal resistance value deviation value of the low-temperature zone insulation layer, the shape factor of the low-temperature zone insulation layer, and the shape factor of the growth-driven zone insulation layer; the third element of the preparation vector is one of the high-temperature zone input power, the high-temperature zone cooling power, the low-temperature zone input power, the low-temperature zone cooling power, and the crucible descent speed;

[0016] Based on the preparation vector, the preparation data for the preprocessing is determined.

[0017] The gallium oxide prediction method based on deep learning and crucible descent includes the following predicted property data: predicted crack data, predicted impurity crystal data, predicted diffraction peak full width at half maximum (FWHM), predicted radial deviation of predicted FWHM, and predicted axial deviation of predicted FWHM.

[0018] A method for preparing gallium oxide based on deep learning and crucible lowering method, the method comprising:

[0019] Obtain target property data for the target gallium oxide single crystal;

[0020] Based on the target property data and the trained neural network model, the target preparation data corresponding to the target gallium oxide single crystal is determined; wherein, the target preparation data includes: seed crystal data, environmental data, and control data;

[0021] Based on the crucible descent method, the target gallium oxide single crystal was prepared according to the target preparation data.

[0022] The gallium oxide preparation method based on deep learning and crucible descent method determines the target preparation data corresponding to the target gallium oxide single crystal based on the target property data and the trained neural network model, including:

[0023] Obtain preset preparation data, and preprocess the preset preparation data to obtain preprocessed preset preparation data;

[0024] The pre-processed preset preparation data is input into a trained neural network model, and the predicted property data corresponding to the gallium oxide single crystal is obtained through the trained neural network model;

[0025] Based on the predicted property data and the target property data, the preset preparation data is corrected to obtain the target preparation data corresponding to the target gallium oxide single crystal.

[0026] The gallium oxide preparation method based on deep learning and crucible lowering method, wherein the trained neural network model is obtained by the following steps:

[0027] Acquire training data for gallium oxide single crystals and the corresponding actual property data; wherein, the training data includes: seed crystal training data, environmental training data, and control training data;

[0028] The preprocessed training data is further preprocessed to obtain preprocessed training data;

[0029] The preprocessed training data is input into a preset neural network model, and the predicted generation property data corresponding to the training data is obtained through the preset neural network model.

[0030] The model parameters of the preset neural network model are adjusted and corrected based on the predicted property data and the actual property data to obtain a trained neural network model.

[0031] The gallium oxide preparation method based on deep learning and crucible lowering method, wherein the preset neural network model includes: a feature extraction module and a fully connected module.

[0032] The step of inputting the preprocessed training data into a preset neural network model and obtaining the prediction generation property data corresponding to the preprocessed training data through the preset neural network model includes:

[0033] The preprocessed training data is input into the feature extraction module, and the feature extraction module obtains the feature vector corresponding to the preprocessed training data.

[0034] The feature vector is input into the fully connected module, and the prediction generation property data obtained from the preprocessed training data is obtained through the fully connected module.

[0035] A gallium oxide preparation system based on deep learning and crucible lowering method includes a memory and a processor, wherein the memory stores a computer program, characterized in that the processor executes the computer program to implement the steps of the prediction method described above, or the steps of the preparation method described above.

[0036] Compared with the prior art, the embodiments of the present invention have the following advantages:

[0037] First, the preparation data is preprocessed to obtain preprocessed preparation data. Then, the preprocessed preparation data is input into a trained neural network model. The trained neural network model is used to obtain the predicted property data corresponding to the gallium oxide single crystal. The trained neural network model can predict the performance of the gallium oxide single crystal. Therefore, the preparation data can be adjusted to obtain the desired performance of the gallium oxide single crystal, thereby optimizing the performance of the gallium oxide single crystal. Attached Figure Description

[0038] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are only some embodiments recorded in the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0039] Figure 1 This is a flowchart of the gallium oxide prediction method based on deep learning and crucible descent method in an embodiment of the present invention;

[0040] Figure 2 This is a schematic diagram of the crystal growth furnace in an embodiment of the present invention;

[0041] Figure 3 This is a schematic diagram showing the position and temperature of the crystal inside the furnace in an embodiment of the present invention;

[0042] Figure 4 This is an internal structure diagram of the gallium oxide preparation system based on deep learning and crucible lowering method in an embodiment of the present invention. Detailed Implementation

[0043] To enable those skilled in the art to better understand the present invention, the technical solutions of the present invention will be clearly and completely described below with reference to the accompanying drawings of the embodiments of the present invention. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0044] The crucible descent method, also known as the Bridgman crystal growth method, is a commonly used crystal growth method. The material for crystal growth is placed in a cylindrical crucible and slowly lowered through a furnace with a specific temperature gradient, the furnace temperature controlled slightly above the material's melting point. As it passes through the heated zone, the material in the crucible melts. As the crucible continues to descend, the temperature at the bottom first drops below the melting point, and crystallization begins. The crystal continues to grow as the crucible descends, such as... Figure 2 As shown. When growing gallium oxide crystals, corrosion of the crucible material must be avoided.

[0045] Various non-limiting embodiments of the present invention will now be described in detail with reference to the accompanying drawings.

[0046] See Figures 1-3 This paper illustrates a gallium oxide prediction method based on deep learning and the crucible descent method in an embodiment of the present invention. In this embodiment, the prediction method may include, for example, the following steps:

[0047] S100. Obtain preparation data for gallium oxide single crystals; wherein, the preparation data includes: seed crystal data, environmental data, and control data.

[0048] Specifically, preparation data refers to data from the preparation of gallium oxide single crystals. Obtaining preparation data for gallium oxide single crystals can be done by configuring the data as needed. For example, if it is necessary to predict the performance of a gallium oxide single crystal under certain preparation data, it is only necessary to determine the preparation data, preprocess the data to obtain preprocessed preparation data, and then input the preprocessed preparation data into a trained neural network model. The predicted property data is obtained through the trained neural network model. In other words, without conducting experiments, the property data of gallium oxide single crystals can be predicted once the preparation data is determined.

[0049] The preparation data includes: seed crystal data, environmental data, and control data. Seed crystal data refers to the data of the seed crystal used in the preparation of gallium oxide single crystals; environmental data refers to the data of the environment in which the crystal is located during the preparation of gallium oxide single crystals; and control data refers to the data on controlling crystal growth during the preparation of gallium oxide single crystals.

[0050] S200. The preparation data is preprocessed to obtain preprocessed preparation data.

[0051] Specifically, after obtaining the preparation data, the preparation data is first preprocessed to obtain preprocessed preparation data, which can then be input into the trained neural network model so that the trained neural network model can process the preprocessed preparation data.

[0052] In one implementation of this application embodiment, step S200, preprocessing the preparation data to obtain preprocessed preparation data, includes:

[0053] S210. Based on the seed crystal data, the environmental data, and the control data, preprocessing preparation data is obtained; wherein, the preprocessing preparation data is a matrix formed by the seed crystal data, the environmental data, and the control data.

[0054] Specifically, after obtaining the preparation data, the preparation data is first preprocessed to obtain preprocessed preparation data. Since the various sub-data in the preparation data (such as seed crystal data, environmental data, and control data) will affect each other, but the extent of the mutual influence between the sub-data is currently unclear, it is necessary to preprocess the preparation data to rearrange and combine the various sub-data in the preparation data to form preprocessed preparation data.

[0055] In one implementation of this application, the seed crystal data includes: the full width at half maximum (FWHM) of the seed crystal diffraction peak, the FWHM deviation value of the seed crystal diffraction peak, and the seed crystal diameter; the environmental data includes: the thermal resistance value of the insulation layer in the high-temperature zone, the thermal resistance deviation value of the insulation layer in the high-temperature zone, the shape factor of the insulation layer in the high-temperature zone, the thermal resistance value of the insulation layer in the low-temperature zone, the thermal resistance deviation value of the insulation layer in the low-temperature zone, the shape factor of the insulation layer in the low-temperature zone, and the shape factor of the insulation layer in the growth-driven zone; the control data includes: the input power in the high-temperature zone, the cooling power in the high-temperature zone, the input power in the low-temperature zone, the cooling power in the low-temperature zone, and the crucible descent speed.

[0056] Specifically, the full width at half maximum (FWHM) of the seed crystal diffraction peak can be measured using an X-ray diffractometer. The FWHM deviation values ​​include the radial deviation and the axial deviation. Radial deviation refers to the direction on the horizontal plane, and axial deviation refers to the direction perpendicular to the horizontal plane, i.e., the vertical axis. The radial deviation value can be obtained by measuring the FWHM on both radial sides of the seed crystal and calculating the difference between the FWHM values ​​on both sides. Similarly, the axial deviation value can be obtained by measuring the FWHM on both axial sides of the seed crystal and calculating the difference between the FWHM values ​​on both sides.

[0057] In the crucible-lowering method for preparing gallium oxide single crystals, the region containing the already grown crystal is the low-temperature region, the region containing the melt that has not yet grown into a crystal is the high-temperature region, and the region where the melt grows into a crystal is the growth-driven region, which lies between the high-temperature and low-temperature regions. Typically, the high-temperature region is located above the low-temperature region, such as... Figure 2 As shown, when crucible 1 is in the high-temperature zone, the gallium oxide in crucible 1 is melt 3; when crucible 1 is in the low-temperature zone, the gallium oxide in crucible 1 is crystal 2, as shown. Figure 2 and Figure 3 As shown, when the crucible is in the growth-driven zone, the gallium oxide in the crucible grows from melt into crystal due to the temperature drop. During the crucible's descent from the high-temperature zone to the low-temperature zone, the gallium oxide gradually grows into a gallium oxide single crystal. The bottom of crucible 1 narrows to form a pointed tip, and the seed crystal is located at this tip. That is, during crystal growth, as the crucible descends and gradually enters the low-temperature zone, the melt 3 begins to grow from the seed crystal at the bottom of crucible 1, gradually growing into crystal 2. Of course, the seed crystal can be placed at the tip of crucible 1 after the gallium oxide in crucible 1 has completely melted.

[0058] like Figure 2 As shown, an insulation layer is provided outside the induction coil 4 to maintain temperature. The thermal resistance of the insulation layer refers to the temperature difference between its two ends when a unit of heat passes through it per unit time. A higher thermal resistance indicates a stronger ability to resist heat transfer and better insulation performance. The thermal resistance value in the high-temperature zone refers to the thermal resistance of the insulation layer located within the high-temperature zone, and the thermal resistance value in the low-temperature zone refers to the thermal resistance of the insulation layer located within the low-temperature zone.

[0059] The thermal resistance deviation of the insulation layer includes two values: radial thermal resistance deviation and axial thermal resistance deviation. The radial thermal resistance deviation can be obtained by measuring the thermal resistance on both radial sides of the insulation layer and calculating the difference between these values. Similarly, the axial thermal resistance deviation can be obtained by measuring the thermal resistance on both axial sides of the insulation layer and calculating the difference between these values.

[0060] Of course, when testing the high-temperature zone, the deviation value of the thermal resistance of the insulation layer in the high-temperature zone can be obtained; when testing the low-temperature zone, the deviation value of the thermal resistance of the insulation layer in the low-temperature zone can be obtained.

[0061] The shape factor of an insulation layer refers to the value of the shape and dimensions of the insulation area. For example, when using a cylindrical insulation layer, the shape factor includes the diameter and height of the insulation layer. When using a cubic insulation layer, the shape factor includes the length, width, and height of the insulation layer. Since crystal growth mainly occurs in the low-temperature region and the growth-driving region, the shape factor of the insulation layer in the low-temperature region and the shape factor in the growth-driving region will affect crystal growth.

[0062] Once the crucible and crystal growth furnace are determined, the shape factors of the insulation layer in the low-temperature zone and the growth-driven zone are also determined. As the crystal growth furnace is used, the thermal resistance values, deviation values, and shape factors of the insulation layer in the high-temperature zone, as well as the thermal resistance values ​​and deviation values ​​in the low-temperature zone, will change. However, these changes will not occur in the short term, and these environmental data can be retested after a certain number of crystal growth cycles.

[0063] The high-temperature zone input power refers to the input power of the induction coil in the high-temperature zone during crystal growth, while the low-temperature zone input power refers to the input power of the induction coil in the low-temperature zone during crystal growth. The high-temperature zone cooling power refers to the power required for cooling in the high-temperature zone, and the low-temperature zone cooling power refers to the power required for cooling in the low-temperature zone. Since the induction coil is hollow, a cooling medium is introduced into it during cooling, forming a cooling coil. Cooling is achieved by the continuous flow of the cooling medium within the cooling coil. The cooling power in the high-temperature and low-temperature zones can be determined based on the type and flow rate of the cooling medium. Cooling media types include water, oil, and gas, and the flow rate can be determined based on the flow velocity of the cooling medium and the diameter of the cooling coil. The crucible descent speed refers to the speed at which the crucible descends during crystal growth.

[0064] In one implementation of this application embodiment, step S210, obtaining preprocessed preparation data based on the seed crystal data, the environmental data, and the control data, includes:

[0065] S211. Based on the seed crystal data, the environmental data, and the control data, determine the preparation vector; wherein, the first element of the preparation vector is one of the seed crystal diffraction peak full width at half maximum (FWHM), the seed crystal diffraction peak FWHM deviation value, and the seed crystal diameter; the second element of the preparation vector is one of the thermal resistance value of the high-temperature zone insulation layer, the thermal resistance value deviation value of the high-temperature zone insulation layer, the shape factor of the high-temperature zone insulation layer, the thermal resistance value of the low-temperature zone insulation layer, the thermal resistance value deviation value of the low-temperature zone insulation layer, the shape factor of the low-temperature zone insulation layer, and the shape factor of the growth-driving zone insulation layer; the third element of the preparation vector is one of the high-temperature zone input power, the high-temperature zone cooling power, the low-temperature zone input power, the low-temperature zone cooling power, and the crucible descent speed.

[0066] S212. Determine the preparation data for the preprocessing based on the preparation vector.

[0067] Specifically, the preparation vector (A, B, C) is determined based on seed crystal data A, environmental data B, and control data C. Seed crystal data A is selected from: seed crystal diffraction peak full width at half maximum (FWHM) A1, seed crystal diffraction peak FWHM deviation value A2, and seed crystal diameter A3. Environmental data B is selected from: high-temperature zone insulation layer thermal resistance value B1, high-temperature zone insulation layer thermal resistance deviation value B2, high-temperature zone insulation layer shape factor B3, low-temperature zone insulation layer thermal resistance value B4, low-temperature zone insulation layer thermal resistance deviation value B5, low-temperature zone insulation layer shape factor B6, and growth-driven zone insulation layer shape factor B7. Control data C is selected from: high-temperature zone input power C1, high-temperature zone cooling power C2, low-temperature zone input power C3, low-temperature zone cooling power C4, and crucible descent speed C5. In other words, in the preparation vector (A, B, C), A can be one of A1, A2, A3, B can be one of B1, B2, B3, B4, B5, B6, B7, and C can be one of C1, C2, C3, C4, C5. This can form 105 preparation vectors.

[0068] Arrange all the preparation vectors in order of their numbers to form a matrix, and you will get the preprocessed preparation data.

[0069] Specifically, the preparation data for the pretreatment are as follows:

[0070]

[0071] Of course, other arrangement methods were also used to obtain preprocessed preparation data.

[0072] S300. Input the preprocessed preparation data into the trained neural network model, and obtain the predicted property data corresponding to the gallium oxide single crystal through the trained neural network model.

[0073] Specifically, the predicted property data includes: predicted crack data, predicted impurity crystal data, predicted diffraction peak full width at half maximum (FWHM), predicted diffraction peak FWHM radial deviation value, and predicted diffraction peak FWHM axial deviation value.

[0074] Crack data refers to crack grade data, while predicted crack data refers to predicted crack grade data. For example, cracks can be divided into multiple grades. For instance, if cracks are divided into 3 grades, the crack data would be 1, 2, and 3 respectively.

[0075] Impurity data refers to impurity grade data, while predicted impurity data refers to predicted impurity grade data. For example, impurities can be divided into multiple grades. For instance, if impurities are divided into 3 grades, then the impurity data are 1, 2, and 3 respectively.

[0076] The predicted half-width at half-maximum (WHM) of a diffraction peak refers to the predicted WHM of the diffraction peak. The radial deviation value of the predicted WHM of the diffraction peak refers to the predicted difference between the WHM of the diffraction peak on both sides in the radial direction. The axial deviation value of the predicted WHM of the diffraction peak refers to the predicted difference between the WHM of the diffraction peak on both sides in the axial direction.

[0077] By inputting the preprocessed data into a trained neural network model, the model obtains predicted property data. It should be noted that the predicted property data can be one or more; for example, only crack prediction data may be needed.

[0078] In one implementation of this application, the trained neural network model is trained using the following steps:

[0079] A100, acquiring training data for gallium oxide single crystals and actual property data corresponding to the training data; wherein, the training data includes: seed crystal training data, environmental training data, and control training data.

[0080] Specifically, training data refers to the data used for preparing gallium oxide single crystals for training, while actual property data refers to the actual property data of the prepared gallium oxide single crystals. A training set is formed using the training data and actual property data. A pre-defined neural network model is trained based on this training set to obtain the trained neural network model. The seed crystal training data includes: seed crystal diffraction peak full width at half maximum (FWHM) training data, seed crystal diffraction peak FWHM deviation training data, and seed crystal diameter training data; the environmental training data includes: high-temperature zone insulation layer thermal resistance training data, high-temperature zone insulation layer thermal resistance deviation training data, high-temperature zone insulation layer shape factor training data, low-temperature zone insulation layer thermal resistance training data, low-temperature zone insulation layer thermal resistance deviation training data, low-temperature zone insulation layer shape factor training data, and growth-driven zone insulation layer shape factor training data; the control training data includes: high-temperature zone input power training data, high-temperature zone cooling power training data, low-temperature zone input power training data, low-temperature zone cooling power training data, and crucible descent speed training data. The actual property data includes: actual crack data, actual impurity crystal data, actual diffraction peak half-width at half-maximum (FWHM), actual diffraction peak FWHM radial deviation value, and actual diffraction peak FWHM axial deviation value.

[0081] When collecting data to obtain the training set, gallium oxide single crystals were prepared using the crucible lowering method, and the data from the preparation of the gallium oxide single crystals were recorded as training data. After obtaining the gallium oxide single crystals, their properties were analyzed to obtain actual property data. To facilitate the training of the neural network model, as much data as possible can be collected to form the training set.

[0082] A200. The training data is preprocessed to obtain preprocessed training data.

[0083] Specifically, after obtaining the training data, the training data is preprocessed to obtain preprocessed training data. The preprocessing process can be referred to step S200.

[0084] A300. Input the preprocessed training data into a preset neural network model, and obtain the prediction generation property data corresponding to the preprocessed training data through the preset neural network model.

[0085] Specifically, the preprocessed training data is input into a preset neural network model, and the predicted generated property data is obtained through the preset neural network model. The predicted generated property data includes: predicted generated crack data, predicted generated impurity crystal data, predicted generated diffraction peak full width at half maximum (FWHM), predicted generated diffraction peak FWHM radial deviation value, and predicted generated diffraction peak FWHM axial deviation value.

[0086] A400. Adjust the model parameters of the preset neural network model according to the predicted property data and the actual property data to obtain a trained neural network model.

[0087] Specifically, based on the predicted property data and the actual property data, the model parameters of the preset neural network model are corrected, and the process of inputting the preprocessed training data into the preset neural network model and obtaining the predicted property data corresponding to the preprocessed training data through the preset neural network model (i.e., step A300) continues until the preset training conditions are met, and a trained neural network model is obtained.

[0088] Specifically, based on the predicted property data and the actual property data, the model parameters of the preset neural network model are corrected, and the process continues to execute the step of inputting the preprocessed training data into the preset neural network model, and obtaining the predicted property data corresponding to the preprocessed training data through the preset neural network model, until the preset training conditions are met, resulting in a trained neural network model. That is, if the preset neural network model meets the preset training conditions, a trained neural network model is obtained. If the preset neural network model does not meet the preset training conditions, the process returns to step A300 until the preset neural network model meets the preset training conditions, resulting in a trained neural network model.

[0089] In one implementation of this invention, a loss function value for a preset neural network model is determined based on the predicted property data and the actual property data. The model parameters of the preset neural network model are then corrected based on the loss function value. Specifically, a gradient-based method is used to correct the parameters of the preset neural network model. After determining the loss function value of the preset neural network model, the corrected parameters of the preset neural network model are determined based on the gradient of the loss function value with respect to the parameters of the preset neural network model, the parameters of the preset neural network model, and a preset learning rate.

[0090] The preset training conditions include: the loss function value meets the first preset requirement and / or the preset number of training iterations of the neural network model reaches the first preset number of iterations.

[0091] The first preset requirement is determined based on the accuracy and efficiency of the preset neural network model. For example, the loss function value of the preset neural network model reaches its minimum value or no longer changes. The first preset number of training iterations is the maximum number of training iterations for the preset neural network model, for example, 4000 times.

[0092] The loss functions of the preset neural network model include: mean squared error, root mean square error, and mean absolute error.

[0093] In one implementation of this application, the preset neural network model includes a feature extraction module and a fully connected module.

[0094] For example, the preset neural network model includes: a first convolutional unit, a second convolutional unit, a third convolutional unit, a fourth convolutional unit, and a fully connected unit. Specifically, the first convolutional unit includes: two convolutional layers and one pooling layer. The second, third, and fourth convolutional units each include three convolutional layers and one pooling layer. The fully connected unit includes three fully connected layers.

[0095] Convolutional and fully connected layers are responsible for mapping and transforming the input data. This process uses parameters such as weights and biases, and also requires activation functions. Pooling layers are fixed-function operations. Specifically, convolutional layers extract features; pooling layers perform pooling operations on the input features, changing their spatial dimensions; and fully connected layers fully connect all the data from the previous layer.

[0096] Step A300: Input the preprocessed training data into a preset neural network model, and obtain the prediction generation property data corresponding to the preprocessed training data through the preset neural network model, including:

[0097] A310. Input the preprocessed training data into the feature extraction module, and obtain the feature vector corresponding to the preprocessed training data through the feature extraction module;

[0098] A320. Input the feature vector into the fully connected module, and obtain the prediction generation property data obtained from the preprocessed training data through the fully connected module.

[0099] Specifically, the preprocessed training data is input into a preset neural network model, the feature extraction module in the preset neural network model outputs the feature vector corresponding to the preprocessed training data, and the feature vector is input into the fully connected module in the pre-trained model to obtain the prediction generation property data corresponding to the preprocessed training data output by the fully connected module.

[0100] Based on the above-mentioned gallium oxide prediction method based on deep learning and crucible lowering method, this embodiment provides a gallium oxide preparation method based on deep learning and crucible lowering method, the preparation method comprising:

[0101] B100: Obtain target property data for the target gallium oxide single crystal.

[0102] Specifically, if a target gallium oxide single crystal is required, the target property data of the target gallium oxide single crystal can be determined first, that is, the property data of the desired gallium oxide single crystal can be determined. The target property data includes: target crack data, target impurity crystal data, target diffraction peak full width at half maximum (FWHM), target diffraction peak radial deviation value, and target diffraction peak axial deviation value.

[0103] B200. Based on the target property data and the trained neural network model, determine the target preparation data corresponding to the target gallium oxide single crystal; wherein, the target preparation data includes: seed crystal data, environmental data, and control data.

[0104] Specifically, based on the target property data and the trained neural network model, the target preparation data corresponding to the target gallium oxide single crystal is determined. It should be noted that since different preparation data can yield the same property data, the target preparation data for the target gallium oxide single crystal is not unique. A single target preparation data point is determined based on the controllability of each data point among multiple target preparation data points, thereby facilitating the acquisition of the target gallium oxide single crystal.

[0105] In one implementation of this embodiment, B200 determines the target preparation data corresponding to the target gallium oxide single crystal based on the target property data and the trained neural network model, including:

[0106] B210. Obtain preset preparation data, and preprocess the preset preparation data to obtain preprocessed preset preparation data.

[0107] B220. Input the pre-processed preset preparation data into the trained neural network model, and obtain the predicted property data corresponding to the gallium oxide single crystal through the trained neural network model.

[0108] B230. Based on the predicted property data and the target property data, the preset preparation data is corrected to obtain the target preparation data corresponding to the target gallium oxide single crystal.

[0109] Specifically, pre-prepared data can be preset and preprocessed to obtain preprocessed pre-prepared data. The specific preprocessing process can be found in step S200. The preset pre-prepared data is input into a trained neural network model to obtain predicted property data. Then, based on the predicted property data and the target property data, the preset pre-prepared data is corrected. When the difference between the predicted property data and the target property data is less than a preset threshold, the preset pre-prepared data can be used as the target pre-prepared data. Correction of the preset pre-prepared data can be automatic or manual. Alternatively, a loss function value can be determined based on the predicted property data and the target property data, and the preset pre-prepared data can be corrected based on this loss function value. When the loss function value meets preset correction conditions, the preset pre-prepared data can be used as the target pre-prepared data. Preset correction conditions include: the loss function value meets a second preset requirement and / or the number of corrections to the preset pre-prepared data reaches a second preset number.

[0110] It should be noted that the preset preparation data includes: preset seed crystal data, preset environmental data, and preset control data; the preset seed crystal data includes: preset seed crystal diffraction peak full width at half maximum (FWHM), preset seed crystal diffraction peak FWHM deviation value, and preset seed crystal diameter; the preset environmental data includes: preset high-temperature zone insulation layer thermal resistance value, preset high-temperature zone insulation layer thermal resistance deviation value, preset high-temperature zone insulation layer shape factor, preset low-temperature zone insulation layer thermal resistance value, preset low-temperature zone insulation layer thermal resistance deviation value, preset low-temperature zone insulation layer shape factor, and preset growth-driving zone insulation layer shape factor; the preset control data includes: preset high-temperature zone input power, preset high-temperature zone cooling power, preset low-temperature zone input power, preset low-temperature zone cooling power, and preset crucible descent speed.

[0111] B300, based on the crucible lowering method, the target gallium oxide single crystal is prepared according to the target preparation data.

[0112] Specifically, after obtaining the target preparation data, the target gallium oxide single crystal can be prepared based on the crucible descent method according to the target preparation data.

[0113] Based on the above prediction method or the above preparation method, this invention provides a gallium oxide preparation system based on deep learning and crucible lowering method. This system can be a computer device, with an internal structure as follows: Figure 4 As shown, the system includes a processor, memory, network interface, display screen, and input devices connected via a system bus. The processor provides computational and control capabilities. The memory includes non-volatile storage media and internal memory. The non-volatile storage media stores the operating system and computer programs. The internal memory provides the environment for the operation of the operating system and computer programs in the non-volatile storage media. The network interface is used for communication with external terminals via a network connection. When the computer program is executed by the processor, it implements a gallium oxide prediction method based on deep learning and the crucible-descent method, or a gallium oxide preparation method based on deep learning and the crucible-descent method. The display screen can be a liquid crystal display (LCD) or an e-ink display. The input devices can be a touch layer covering the display screen, buttons, a trackball, or a touchpad mounted on the system casing, or an external keyboard, touchpad, or mouse.

[0114] Those skilled in the art will understand that Figure 4 The diagram shown is merely a partial structural diagram related to the present application and does not constitute a limitation on the system to which the present application is applied. A specific system may include more or fewer components than those shown in the diagram, or may combine certain components, or may have different component arrangements.

[0115] In one embodiment, a gallium oxide preparation system based on deep learning and crucible lowering method is provided, including a memory and a processor. The memory stores a computer program, and the processor executes the computer program to implement the steps of the prediction method or the steps of the preparation method.

[0116] The technical features of the above embodiments can be combined in any way. For the sake of brevity, not all possible combinations of the technical features in the above embodiments are described. However, as long as there is no contradiction in the combination of these technical features, they should be considered to be within the scope of this specification.

Claims

1. A gallium oxide prediction method based on deep learning and crucible lowering method, characterized in that, The prediction method includes: Acquire preparation data for gallium oxide single crystals; wherein, the preparation data includes: seed crystal data, environmental data, and control data; The preparation data is preprocessed to obtain preprocessed preparation data; The preprocessed preparation data is input into a trained neural network model, and the predicted property data corresponding to the gallium oxide single crystal is obtained through the trained neural network model. The preprocessing of the preparation data to obtain preprocessed preparation data includes: Based on the seed crystal data, the environmental data, and the control data, preprocessing preparation data is obtained; wherein, the preprocessing preparation data is a matrix formed by rearranging and combining the seed crystal data, the environmental data, and the control data; The seed crystal data includes: seed crystal diffraction peak full width at half maximum (FWHM) A1, seed crystal diffraction peak FWHM deviation value A2, and seed crystal diameter A3; the environmental data includes: thermal resistance value of the high-temperature insulation layer B1, thermal resistance value deviation value of the high-temperature insulation layer B2, shape factor of the high-temperature insulation layer B3, thermal resistance value of the low-temperature insulation layer B4, thermal resistance value deviation value of the low-temperature insulation layer B5, shape factor of the low-temperature insulation layer B6, and shape factor of the growth-driven insulation layer B7; the control data includes: high-temperature zone input power C1, high-temperature zone cooling power C2, low-temperature zone input power C3, low-temperature zone cooling power C4, and crucible descent speed C5; the pre-processing preparation data obtained based on the seed crystal data, the environmental data, and the control data includes: Based on the seed crystal data, the environmental data, and the control data, a preparation vector is determined; wherein, the first element of the preparation vector is one of the seed crystal diffraction peak full width at half maximum (FWHM), the seed crystal diffraction peak FWHM deviation value, and the seed crystal diameter; the second element of the preparation vector is one of the thermal resistance value of the high-temperature zone insulation layer, the thermal resistance value deviation value of the high-temperature zone insulation layer, the shape factor of the high-temperature zone insulation layer, the thermal resistance value of the low-temperature zone insulation layer, the thermal resistance value deviation value of the low-temperature zone insulation layer, the shape factor of the low-temperature zone insulation layer, and the shape factor of the growth-driven zone insulation layer; the third element of the preparation vector is one of the high-temperature zone input power, the high-temperature zone cooling power, the low-temperature zone input power, the low-temperature zone cooling power, and the crucible descent speed; Based on the preparation vector, determine the preparation data for the preprocessing; The preparation data for the pretreatment are as follows: The trained neural network model was obtained by training the following steps: Acquire training data for gallium oxide single crystals and corresponding actual property data; wherein, the training data includes: seed crystal training data, environmental training data, and control training data; The training data is preprocessed to obtain preprocessed training data; The preprocessed training data is input into a preset neural network model, and the predicted generation property data corresponding to the preprocessed training data is obtained through the preset neural network model. The model parameters of the preset neural network model are adjusted and corrected based on the predicted property data and the actual property data to obtain a trained neural network model. The preset neural network model includes: a feature extraction module and a fully connected module. The step of inputting the preprocessed training data into a preset neural network model and obtaining the prediction generation property data corresponding to the preprocessed training data through the preset neural network model includes: The preprocessed training data is input into the feature extraction module, and the feature extraction module obtains the feature vector corresponding to the preprocessed training data. The feature vector is input into the fully connected module, and the prediction generation property data corresponding to the preprocessed training data is obtained through the fully connected module.

2. The gallium oxide prediction method based on deep learning and crucible lowering method according to claim 1, characterized in that, The predicted property data includes: predicted crack data, predicted impurity crystal data, predicted diffraction peak full width at half maximum (FWHM), predicted diffraction peak FWHM radial deviation value, and predicted diffraction peak FWHM axial deviation value.

3. A method for preparing gallium oxide based on deep learning and crucible lowering method, characterized in that, The preparation method includes: Obtain target property data for the target gallium oxide single crystal; Based on the target property data and the trained neural network model, the target preparation data corresponding to the target gallium oxide single crystal is determined; wherein, the target preparation data includes: seed crystal data, environmental data, and control data; Based on the crucible lowering method, the target gallium oxide single crystal was prepared according to the target preparation data. The trained neural network model was obtained by training the following steps: Acquire training data for gallium oxide single crystals and corresponding actual property data; wherein, the training data includes: seed crystal training data, environmental training data, and control training data; The training data is preprocessed to obtain preprocessed training data; The preprocessed training data is input into a preset neural network model, and the predicted generation property data corresponding to the preprocessed training data is obtained through the preset neural network model. The model parameters of the preset neural network model are adjusted and corrected based on the predicted property data and the actual property data to obtain a trained neural network model. The preprocessing of the training data to obtain preprocessed training data includes: Based on the seed crystal training data, the environment training data, and the control training data, preprocessed training data is obtained; wherein, the preprocessed training data is a matrix formed by rearranging and combining the seed crystal training data, the environment training data, and the control training data; The seed crystal training data includes: seed crystal diffraction peak full width at half maximum (FWHM) training data A1, seed crystal diffraction peak FWHM deviation value training data A2, and seed crystal diameter training data A3; the environmental training data includes: high-temperature zone insulation layer thermal resistance value training data B1, high-temperature zone insulation layer thermal resistance value deviation value training data B2, high-temperature zone insulation layer shape factor training data B3, low-temperature zone insulation layer thermal resistance value training data B4, low-temperature zone insulation layer thermal resistance value deviation value training data B5, low-temperature zone insulation layer shape factor training data B6, and growth-driven zone insulation layer shape factor training data B7; the control training data includes: high-temperature zone input power training data C1, high-temperature zone cooling power training data C2, low-temperature zone input power training data C3, low-temperature zone cooling power training data C4, and crucible descent speed training data C5; the preprocessed training data obtained based on the seed crystal training data, the environmental training data, and the control training data includes: Based on the seed crystal training data, the environmental training data, and the control training data, a preparation vector is determined; wherein, the first element of the preparation vector is one of the seed crystal diffraction peak full width at half maximum (FWHM) training data, the seed crystal diffraction peak FWHM deviation value training data, and the seed crystal diameter training data; the second element of the preparation vector is one of the high-temperature zone insulation layer thermal resistance value training data, the high-temperature zone insulation layer thermal resistance value deviation value training data, the high-temperature zone insulation layer shape factor training data, the low-temperature zone insulation layer thermal resistance value training data, the low-temperature zone insulation layer thermal resistance value deviation value training data, the low-temperature zone insulation layer shape factor training data, and the growth-driven zone insulation layer shape factor training data; the third element of the preparation vector is one of the high-temperature zone input power training data, the high-temperature zone cooling power training data, the low-temperature zone input power training data, the low-temperature zone cooling power training data, and the crucible descent speed training data; Based on the prepared vector, the preprocessed training data is determined; The preprocessed training data is as follows: The step of determining the target preparation data corresponding to the target gallium oxide single crystal based on the target property data and the trained neural network model includes: Obtain preset preparation data, and preprocess the preset preparation data to obtain preprocessed preset preparation data; The pre-processed preset preparation data is input into a trained neural network model, and the predicted property data corresponding to the gallium oxide single crystal is obtained through the trained neural network model; Based on the predicted property data and the target property data, the preset preparation data is corrected to obtain the target preparation data corresponding to the target gallium oxide single crystal; The preset neural network model includes: a feature extraction module and a fully connected module. The step of inputting the preprocessed training data into a preset neural network model and obtaining the prediction generation property data corresponding to the preprocessed training data through the preset neural network model includes: The preprocessed training data is input into the feature extraction module, and the feature extraction module obtains the feature vector corresponding to the preprocessed training data. The feature vector is input into the fully connected module, and the prediction generation property data corresponding to the preprocessed training data is obtained through the fully connected module.

4. A gallium oxide preparation system based on deep learning and crucible lowering method, characterized in that, The device includes a memory and a processor, the memory storing a computer program, characterized in that the processor executes the computer program to implement the steps of the prediction method according to any one of claims 1 to 2, or the steps of the preparation method according to claim 3.

Images