A method, device, equipment and medium for controlling growth of indium phosphide crystal based on multi-modal deep learning

Abstract

The application discloses a kind of indium phosphide crystal growth control method, device, equipment and medium based on multi-modal deep learning, the method includes by X-ray acquisition solid-liquid interface in growth furnace interface image;The interface image is input machine vision model, and the global morphological characteristics of solid-liquid interface are captured by the machine vision model;By time series prediction module in growth control model, the future time length of phosphorus vapor pressure variation trend is determined based on historical observation data sequence;By reinforcement learning module in growth control model, control parameters are determined based on variation trend and global morphological characteristics;The growth of the indium phosphide crystal is controlled based on the control parameters.The application identifies global morphological characteristics by machine vision model and predicts the variation trend of phosphorus vapor pressure by time series prediction module, then determines control parameters in combination with global morphological characteristics and phosphorus vapor pressure variation trend, can realize the real-time dynamic regulation and control of indium phosphide crystal growth process, solve the response lag problem caused by traditional dependence on artificial experience fixed parameter control, effectively reduce the twin crystal rate and dislocation density of crystal, improve the growth yield and crystal quality of indium phosphide single crystal.

Description

Technical Field

[0001] This application relates to the field of semiconductor crystal growth technology, and in particular to a method, apparatus, equipment and medium for controlling the growth of indium phosphide crystals based on multimodal deep learning. Background Technology

[0002] Indium phosphide It is an important second-generation semiconductor material, widely used in optical communication modules, high-end radio frequency devices and sensor devices. The crystal has a melting point of 1070℃, and at this melting point, the dissociation pressure of phosphorus is as high as 2.75 MPa. Phosphorus is highly volatile, therefore... It is one of the most technically challenging materials in the field of semiconductor crystal growth.

[0003] Currently, the industry mainly uses two methods for growth. Single crystal: Liquid-sealed Czochralski method ( Liquid Encapsulated Czochralski) and vertical gradient condensation ( Vertical Gradient Freeze). When growing by method, it is usually with As a liquid sealant, the liquid seal layer thickness is approximately 30 mm. This inhibits phosphorus volatilization, while the growth process requires 2-3 days. It is carried out in a high-pressure nitrogen or argon atmosphere.

[0004] However, existing equipment mainly relies on the operator's experience to set fixed temperature profiles and pulling speeds, which is not only difficult to cope with the complex nonlinear coupling relationship between phosphorus vapor pressure and temperature during crystal growth, but also has a serious response lag problem when dealing with pressure changes caused by phosphorus volatilization, which in turn leads to the formation of twins and high dislocation density defects inside the crystal.

[0005] Therefore, the existing technology still needs to be improved and enhanced. Summary of the Invention

[0006] The technical problem to be solved by this application is to provide a method, apparatus, equipment and medium for controlling the growth of indium phosphide crystals based on multimodal deep learning, which addresses the shortcomings of the existing technology.

[0007] To address the aforementioned technical problems, the first aspect of this application provides a method for controlling the growth of indium phosphide crystals based on multimodal deep learning, wherein the method includes: Interface images of the solid-liquid interface in the growth furnace were acquired using X-rays. The interface image is input into the machine vision model, and the global morphological features of the solid-liquid interface are captured by the machine vision model. The trend of phosphorus vapor pressure variation in the future time period is determined by the time-series prediction module in the growth control model based on historical observation data sequences. The control parameters are determined based on the changing trend and global morphological features through the reinforcement learning module in the growth control model. The growth of the indium phosphide crystal is controlled based on the control parameters.

[0008] The aforementioned method for controlling the growth of indium phosphide crystals based on multimodal deep learning, wherein the acquisition of interface images of the solid-liquid interface in the growth furnace using X-rays specifically involves: X-rays are emitted into the growth furnace through an X-ray imaging unit installed on the side wall of the high-pressure chamber. The X-ray detector located on the opposite side of the X-ray imaging unit receives the X-rays passing through the side wall of the high-pressure chamber and... X-rays are taken after the liquid seal layer, and a grayscale projection image of the solid-liquid interface is formed based on the X-rays to obtain an interface image.

[0009] The aforementioned method for controlling the growth of indium phosphide crystals based on multimodal deep learning, wherein the step of inputting the interface image into a machine vision model and capturing the global morphological features of the solid-liquid interface through the machine vision model specifically includes: The interface image is input into the grayscale enhancement preprocessing module in the machine vision model. The grayscale enhancement preprocessing module enhances the contrast of the interface image to obtain an enhanced interface image. The enhanced interface image is input into the denoising module in the machine vision model, and the X-ray scattering noise is removed by the denoising module to obtain the denoised interface image. Based on the denoised interface image, it is filtered through a deblurred convolutional attention layer in the machine vision model. Image noise from the liquid sealing layer is used to extract the initial feature map; Based on the initial feature map, the target feature map is extracted through the self-attention layer in the machine vision model, and the global morphological features are determined based on the target feature map.

[0010] The aforementioned method for controlling the growth of indium phosphide crystals based on multimodal deep learning, wherein the global morphological features include the curvature of the solid-liquid interface, The thickness of the liquid seal layer and the state of seed crystal fusion.

[0011] The indium phosphide crystal growth control method based on multimodal deep learning includes control parameters such as temperature regulation, air pressure regulation, pulling speed regulation, and electromagnetic field strength.

[0012] The aforementioned method for controlling the growth of indium phosphide crystals based on multimodal deep learning includes a reinforcement learning module that uses the changing trend and global morphological features as the state space, the control parameters as the action space, and constructs a reward function based on the crystal dislocation density and twinning rate. The optimal control strategy is determined through a deep deterministic policy gradient approach.

[0013] The aforementioned method for controlling the growth of indium phosphide crystals based on multimodal deep learning further includes: The complete growth process data of multiple indium phosphide crystals are aggregated through a cloud server; Federated learning is performed on the machine vision model and the growth control model using complete growth process data to jointly fine-tune the machine vision model and the growth control model, so that the side device executing the indium phosphide crystal growth control method based on multimodal deep learning can determine the control parameters using the jointly fine-tuned machine vision model and the growth control model.

[0014] A second aspect of this application provides an indium phosphide crystal growth control device based on multimodal deep learning, wherein the device includes: The acquisition module is used to acquire interface images of the solid-liquid interface in the growth furnace using X-rays. The capture module is used to input the interface image into the machine vision model and capture the global morphological features of the solid-liquid interface through the machine vision model. The time-series prediction module is used to determine the trend of phosphorus vapor pressure changes over a future time period based on historical observation data sequences, using the time-series prediction module in the growth control model. The determination module is used to determine control parameters based on the changing trend and global morphological features through the reinforcement learning module in the growth control model. A control module is used to control the growth of the indium phosphide crystal based on the control parameters.

[0015] A third aspect of this application provides a computer-readable storage medium storing one or more programs that can be executed by one or more processors to implement the steps in the indium phosphide crystal growth control method based on multimodal deep learning as described above.

[0016] A fourth aspect of this application provides a terminal device, which includes: a processor and a memory; The memory stores a computer-readable program that can be executed by the processor; When the processor executes the computer-readable program, it implements the steps in the indium phosphide crystal growth control method based on multimodal deep learning as described above.

[0017] Beneficial effects: 1. This application identifies global morphological features through a machine vision model and predicts the trend of phosphorus vapor pressure change through a time-series prediction module. Then, it combines the global morphological features and the trend of phosphorus vapor pressure change to determine control parameters, which can realize real-time dynamic control of the indium phosphide crystal growth process. This solves the response lag problem caused by the traditional reliance on fixed parameter control based on human experience, effectively reduces the twinning rate and dislocation density of the crystal, and improves the growth yield and crystal quality of indium phosphide single crystals.

[0018] 2. This application uses the trend of phosphorus vapor pressure change as the basis for determining control parameters, and makes advance predictions and adjustments for possible pressure changes and operating condition fluctuations, thereby achieving effective feedforward compensation, further avoiding crystal defects caused by lag control, and improving the stability of the crystal growth process.

[0019] 3. This application achieves this by allowing X-rays to pass through the sidewalls of the high-pressure chamber and The liquid seal layer, which captures images of the solid-liquid interface inside the growth furnace, solves the technical problem that traditional optical methods cannot directly observe the shape of the solid-liquid interface and the seed fusion. It enables the visualization of the solid-liquid interface state and provides intuitive and reliable data support for the control of the crystal growth process.

[0020] 4. This application effectively filters image interference and X-ray scattering noise caused by the liquid seal layer by adding grayscale enhancement preprocessing, convolutional attention denoising and self-attention feature extraction to the visual framework, ensuring the accuracy and reliability of global morphological features and providing accurate state basis for subsequent control parameter determination.

[0021] 5. This application adopts a reinforcement learning method based on deep deterministic policy gradient to determine the optimal control parameters. The reward function is constructed based on the crystal dislocation density and twinning rate to guide the optimization direction of the model. It can quickly adapt to complex nonlinear growth conditions, output the optimal control parameters, and realize the adaptive regulation of the crystal growth process.

[0022] 6. This application uses a federated learning approach to jointly fine-tune the model using historical data from multiple growth scenarios. While ensuring the data privacy of each device, it continuously optimizes the model performance, enabling the model to adapt to the growth control requirements of indium phosphide crystals under different working conditions, thereby improving the generalization ability and control accuracy of this method. Attached Figure Description

[0023] To more clearly illustrate the technical solutions in the embodiments of this application, the accompanying drawings used in the description of the embodiments will be briefly introduced below. Obviously, the accompanying drawings described below are only some embodiments of this application. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0024] Figure 1 A flowchart of the indium phosphide crystal growth control method based on multimodal deep learning provided in the embodiments of this application.

[0025] Figure 2 This is an architecture diagram of the indium phosphide crystal growth control method based on multimodal deep learning provided in the embodiments of this application.

[0026] Figure 3 This is a flowchart of the one-click crystal growth data closed loop.

[0027] Figure 4 This is a schematic diagram of the indium phosphide crystal growth control device based on multimodal deep learning provided in an embodiment of this application.

[0028] Figure 5 A schematic block diagram of the terminal device provided in the embodiments of this application. Detailed Implementation

[0029] This application provides a method, apparatus, device, and medium for controlling the growth of indium phosphide crystals based on multimodal deep learning. To make the objectives, technical solutions, and effects of this application clearer and more explicit, the following detailed description, with reference to the accompanying drawings and embodiments, further illustrates this application. It should be understood that the specific embodiments described herein are merely illustrative and not intended to limit the scope of this application.

[0030] Those skilled in the art will understand that, unless specifically stated otherwise, the singular forms “a,” “an,” “the,” and “the” used herein may also include the plural forms. It should be further understood that the term “comprising” as used in this application means the presence of the stated features, integers, steps, operations, elements, and / or components, but does not exclude the presence or addition of one or more other features, integers, steps, operations, elements, components, and / or groups thereof. It should be understood that when we say an element is “connected” or “coupled” to another element, it can be directly connected or coupled to the other element, or there may be intermediate elements. Furthermore, “connected” or “coupled” as used herein can include wireless connections or wireless coupling. The term “and / or” as used herein includes all or any units and all combinations of one or more associated listed items.

[0031] It will be understood by those skilled in the art that, unless otherwise defined, all terms used herein (including technical and scientific terms) have the same meaning as commonly understood by one of ordinary skill in the art to which this application pertains. It should also be understood that terms such as those defined in general dictionaries should be understood to have the same meaning as in the context of the prior art, and should not be interpreted in an idealized or overly formal sense unless specifically defined as herein.

[0032] It should be understood that the sequence number and size of each step in this embodiment do not imply the order of execution. The execution order of each process is determined by its function and internal logic, and should not constitute any limitation on the implementation process of this application embodiment.

[0033] Research has revealed that indium phosphide It is an important second-generation semiconductor material, widely used in optical communication modules, high-end radio frequency devices and sensor devices. The crystal has a melting point of 1070℃, and at this melting point, the dissociation pressure of phosphorus is as high as 2.75 MPa. Phosphorus is highly volatile, therefore... It is one of the most technically challenging materials in the field of semiconductor crystal growth.

[0034] Currently, the industry mainly uses two methods for growth. Single crystal: Liquid-sealed Czochralski method ( Liquid Encapsulated Czochralski) and vertical gradient condensation ( Vertical Gradient Freeze). When growing by method, it is usually with As a liquid sealant, the liquid seal layer thickness is approximately 30 mm. This inhibits phosphorus volatilization, while the growth process requires 2-3 days. It is carried out in a high-pressure nitrogen or argon atmosphere.

[0035] However, existing equipment relies primarily on operator experience to set fixed temperature profiles and pulling speeds. This makes it difficult to handle the complex nonlinear coupling between phosphorus vapor pressure and temperature during crystal growth, and it also suffers from severe response hysteresis when dealing with pressure surges caused by phosphorus volatilization, leading to the formation of twins and high dislocation density defects within the crystal. Furthermore, due to… Due to the obstruction of the liquid seal layer, traditional optical methods cannot directly observe the solid-liquid interface morphology and seed crystal fusion state. Instead, offline defect detection of the crystal must be performed after growth, which further increases the probability of producing defective crystals.

[0036] To address the aforementioned issues, in this embodiment, an interface image of the solid-liquid interface in the growth furnace is acquired using X-rays; this interface image is input into a machine vision model, which captures the global morphological features of the solid-liquid interface; the time-series prediction module in the growth control model determines the future trend of phosphorus vapor pressure based on historical observation data sequences; the reinforcement learning module in the growth control model determines control parameters based on the trend and global morphological features; and the growth of the indium phosphide crystal is controlled based on these control parameters. This application, by identifying global morphological features using a machine vision model and predicting phosphorus vapor pressure trends using a time-series prediction module, and then combining these features with the determination of control parameters, enables real-time dynamic control of the indium phosphide crystal growth process. This solves the response lag problem caused by traditional fixed-parameter control based on manual experience, effectively reducing the twinning rate and dislocation density of the crystal, and improving the growth yield and crystal quality of indium phosphide single crystals. Meanwhile, this application uses the changing trend of phosphorus vapor pressure as the basis for determining control parameters, making advance predictions and adjustments for possible pressure changes and operating condition fluctuations, achieving effective feedforward compensation, further avoiding crystal defects caused by lag control, and improving the stability of the crystal growth process.

[0037] It should be noted that the execution entity of each embodiment of this application can be a computing service system with data processing, network communication, and program execution functions, such as an electronic system capable of realizing the above functions, or an indium phosphide crystal growth control system based on multimodal deep learning. The following description uses an indium phosphide crystal growth control system based on multimodal deep learning (hereinafter referred to as "the system") as an example to illustrate the following embodiments.

[0038] The application content will be further explained below with reference to the accompanying drawings and through the description of the embodiments.

[0039] This embodiment provides a method for controlling the growth of indium phosphide crystals based on multimodal deep learning, such as... Figure 1 and Figure 2 As shown, the indium phosphide crystal growth control method based on multimodal deep learning specifically includes: S10. Acquire interface images of the solid-liquid interface in the growth furnace using X-rays.

[0040] Specifically, the growth of indium phosphide crystals requires As a liquid sealant to suppress phosphorus volatilization, interface images can be directly obtained through visual observation. Therefore, embodiments of this application utilize X-rays to acquire interface images of the solid-liquid interface in the growth furnace. Specifically, an X-ray-based visual acquisition device is mounted on the side wall of the high-pressure chamber of the growth furnace. This device transmits X-rays into the growth furnace; the X-rays enter from the side wall of the high-pressure chamber and pass through… The liquid seal layer reaches the X-ray detector on the other side, forming an interface image on the X-ray detector.

[0041] Therefore, the high-pressure chamber of the growth furnace is equipped with an X-ray imaging unit and an X-ray detector, located on opposite sides of the high-pressure chamber. X-rays are emitted from the X-ray emitting end and pass sequentially through the side walls of the high-pressure chamber and... The liquid seal layer is ultimately received by the X-ray detector, generating a grayscale projection image of the solid-liquid interface. This image can intuitively present the outline shape of the solid-liquid interface and the fusion of the seed crystals, solving the problem that traditional optical methods cannot penetrate the liquid seal layer for observation.

[0042] Based on this, the specific steps of acquiring the interface image of the solid-liquid interface in the growth furnace using X-rays are as follows: X-rays are emitted into the growth furnace through an X-ray imaging unit installed on the side wall of the high-pressure chamber. The X-ray detector located on the opposite side of the X-ray imaging unit receives the X-rays passing through the side wall of the high-pressure chamber and... X-rays are taken after the liquid seal layer, and a grayscale projection image of the solid-liquid interface is formed based on the X-rays to obtain an interface image.

[0043] Specifically, the X-rays emitted by the X-ray imaging unit, after passing through the side wall of the high-pressure chamber, will continue to penetrate the growth furnace. The liquid seal layer reaches the solid-liquid interface region within the growth furnace, then exits through the side wall of the high-pressure chamber on the other side and is received by an X-ray detector (such as a flat panel detector) on the opposite side of the X-ray imaging unit. Due to the differences in density between the seed crystal, melt, liquid seal layer, and crystal, the attenuation of X-rays varies. Therefore, the X-ray intensities received by the X-ray detector differ, resulting in the output of a grayscale projection image with differences in brightness and darkness, thus obtaining the interface image.

[0044] S20. Input the interface image into the machine vision model and capture the global morphological features of the solid-liquid interface through the machine vision model.

[0045] Specifically, the global morphological features are obtained through visual recognition of the interface image using a machine vision model. These features reflect the contour shape of the solid-liquid interface, the seed crystal fusion depth, and the overall unevenness of the solid-liquid interface, providing a basis for subsequent control parameter adjustments. These global morphological features include the solid-liquid interface curvature, The thickness of the liquid seal layer and the fusion state of the seed crystals. Among them, the curvature of the solid-liquid interface is used to determine the overall concavity and convexity of the solid-liquid interface, which can be convex (stable), flat (critical), or concave (dangerous), etc. The thickness of the liquid seal layer can be determined based on the grayscale gradient, which is used to reflect... The current evaporation level of the liquid seal layer. The seed crystal fusion status is used to reflect the degree of fusion between the seed crystal and the masterbatch melt, to determine whether the fusion process meets expectations, and to avoid subsequent crystal growth defects caused by over-fusion or under-fusion.

[0046] In one embodiment, the interface image is obtained by X-ray penetration through the sidewalls of the high-pressure chamber and When X-rays are captured through a liquid seal layer, interference from the chamber's metal structure and the liquid seal layer can cause noise and blurred edges in the acquired interface images. Therefore, when using a machine vision model to capture the global morphological features of the solid-liquid interface, denoising can be performed on the interface. This can be done through methods such as Gaussian filtering to remove background noise, or by adding a grayscale enhancement preprocessing operation to the machine vision model to improve the contrast of the interface image and reduce background noise.

[0047] Based on this, the step of inputting the interface image into the machine vision model and capturing the global morphological features of the solid-liquid interface through the machine vision model specifically includes: The interface image is input into the grayscale enhancement preprocessing module in the machine vision model. The grayscale enhancement preprocessing module enhances the contrast of the interface image to obtain an enhanced interface image. The enhanced interface image is input into the denoising module in the machine vision model, and the X-ray scattering noise is removed by the denoising module to obtain the denoised interface image. Based on the denoised interface image, it is filtered through a deblurred convolutional attention layer in the machine vision model. Image noise from the liquid sealing layer is used to extract the initial feature map; Based on the initial feature map, the target feature map is extracted through the self-attention layer in the machine vision model, and the global morphological features are determined based on the target feature map.

[0048] Specifically, the machine vision model is based on the Vision Transformer (ViT) architecture, and adds a grayscale enhancement preprocessing module, a denoising module, and a deblurring convolutional attention layer to ViT. The grayscale enhancement preprocessing module is connected to the input layer and is used to improve the contrast of the interface image to obtain an enhanced interface image. For example, the grayscale enhancement preprocessing module uses a linear layer to expand the low-contrast interface image to a preset grayscale range. The grayscale difference between the solid-liquid interface and the surrounding area is amplified to obtain an enhanced interface map. A denoising module connected to this grayscale enhancement preprocessing module is used to remove X-ray scattering noise. For example, this denoising module can use a Gaussian filter convolutional layer to smooth the enhanced interface map and filter out random scattering noise generated during X-ray imaging to obtain a denoised interface map. A deblurring convolutional attention layer connected to the denoising module divides the denoised interface map into several image blocks and assigns different weights to features in different regions through a multi-head attention mechanism, emphasizing the feature response of the solid-liquid interface region and suppressing background noise in the uniform region of the liquid seal layer, outputting an initial feature map. Finally, a self-attention layer models the global dependency of the initial feature map, capturing the global contour information of the solid-liquid interface, outputting a target feature map, and then regressing global morphological features such as the solid-liquid interface curvature, liquid seal layer thickness, and seed crystal fusion state from the target feature map.

[0049] S30. Using the time-series prediction module in the growth control model, determine the trend of phosphorus vapor pressure changes over future time periods based on historical observation data sequences.

[0050] Specifically, the historical observation data sequence includes multiple sets of observation data collected chronologically during the indium phosphide crystal growth process. Each set of observation data includes furnace temperature distribution, crystal weight, crystal pulling speed, chamber pressure, and heating power at the corresponding acquisition time. The chamber pressure can include the pressure within the chamber. Partial pressure and / or total pressure within the chamber. Data sampling intervals are set from 30 to 120 minutes depending on the crystal growth stage; for example, observation data from the past 60 minutes are collected to form a historical observation data sequence. The future trend of phosphorus vapor pressure variation is predicted by the time-series prediction module in the growth control model based on the historical observation data sequence. This reflects the overall direction and fluctuation amplitude of phosphorus vapor pressure variation in the furnace over a certain period, enabling early identification of pressure abrupt changes caused by phosphorus volatilization, thus facilitating effective feedforward compensation.

[0051] In order to collect observation data, such as Figure 2 As shown, in addition to an X-ray imaging unit and an X-ray detector on the growth furnace, this embodiment also equips the growth furnace with an infrared thermometer array and a phosphorus vapor pressure sensor. The infrared thermometer array is arranged at a downward angle on the outer wall of the high-pressure chamber, allowing for multi-point non-contact temperature measurement of the liquid seal surface and crystal neck through the furnace's observation window to obtain the furnace's internal temperature distribution. Its temperature measurement accuracy is [insert accuracy here]. The phosphorus vapor pressure sensor is installed on the top exhaust pipe of the high-pressure chamber and can continuously collect the current phosphorus vapor pressure value in the chamber (including the pressure inside the chamber). (partial pressure and total pressure within the chamber), its response time is less than 100. .

[0052] The time series prediction module is constructed using a Long Short-Term Memory (LSTM) network, which can contain... Hidden layers, each containing The LSTM network contains 3 hidden layers, each with 256 hidden units, and uses Dropout regularization to prevent overfitting. After inputting the time-series observation data sequence into the time-series prediction module, the module models the time dependencies of data at different times in the historical observation data sequence, learns the variation law of phosphorus vapor pressure with parameters such as furnace temperature and pulling speed, and finally outputs the predicted values ​​of phosphorus vapor pressure at each time point within a preset future time period, and summarizes the trend of phosphorus vapor pressure change over the preset future time period. Of course, in practical applications, other time series prediction models can also be used, such as the Transformer model or the Prophet model, which can all achieve trend prediction based on historical data; no specific limitation is made here.

[0053] This application embodiment predicts the phosphorus vapor pressure change trend over a preset time period using a timing prediction module, and then uses this preset time period phosphorus vapor pressure change trend as the basis for determining control parameters. This can capture the evolution trend of phosphorus vapor pressure abrupt changes in advance, and complete the pre-adjustment of control parameters before the pressure abrupt change occurs to perform effective feedforward compensation. This avoids the problem that traditional feedback control can only be passively adjusted after the pressure abrupt change occurs, which is prone to oscillation, and provides a stable pressure environment for the crystal growth process.

[0054] S40. The control parameters are determined based on the change trend and global morphological features through the reinforcement learning module in the growth control model.

[0055] Specifically, the control parameters are used to regulate the indium phosphide crystal growth process, and may include at least one of temperature regulation, gas pressure regulation, pulling speed regulation, and electromagnetic field strength. In the embodiments of this application, the control parameters include temperature regulation, gas pressure regulation, pulling speed regulation, and electromagnetic field strength. The temperature regulation is used to adjust the heating power of different heating zones, change the axial and radial temperature distribution in the furnace, and thus adjust the morphology and position of the solid-liquid interface. The gas pressure regulation is used to adjust the amount of inert gas filling in the high-pressure chamber, maintain the stability of the total pressure in the chamber, and suppress excessive volatilization of phosphorus. The pulling speed regulation is used to adjust the pulling speed of the crystal pulling device to match the crystal growth rate and the solid-liquid interface movement speed. The electromagnetic field strength is used to regulate melt convection, uniformly distribute the temperature and composition inside the melt, and avoid crystal defects caused by supercooling of the composition.

[0056] The reinforcement learning module is built upon a deep deterministic policy gradient algorithm, which is suitable for decision-making problems with continuous action spaces and can directly output multi-dimensional control parameters with continuously varying values. Specifically, this reinforcement learning module uses the aforementioned changing trends and global morphological features as the state space, the aforementioned control parameters as the action space, and constructs a reward function based on crystal dislocation density and twinning rate. Then, the optimal control strategy is determined using the deep deterministic policy gradient approach. For example, the reward function uses a larger final defect-free crystal length, a lower probability of twinning, and a lower dislocation density as positive reward terms, and system oscillations caused by excessive control parameter adjustments as penalty terms, allowing the model to gradually learn the optimal control strategy corresponding to different solid-liquid interface morphologies and phosphorus vapor pressure changing trends.

[0057] It is understandable that when using this reinforcement learning module for online control decision-making, the global morphological features of the solid-liquid interface and the trend of phosphorus vapor pressure change over a preset time period are used as state inputs and input into the reinforcement learning module that has been trained. The reinforcement learning module directly outputs the optimal control parameter adjustment scheme based on the current state, thereby completing the dynamic decision-making for the growth process.

[0058] S50. Control the growth of the indium phosphide crystal based on the control parameters.

[0059] Specifically, after obtaining the control parameters, the control parameters are sent to the corresponding actuators, which then perform the corresponding adjustment actions: the temperature adjustment is sent to the induction heater of the growth furnace, which adjusts the heating power of the corresponding heating zone; the gas pressure adjustment is sent to the pressure controller of the high-pressure chamber, which adjusts the intake or exhaust volume to adjust the total pressure and phosphorus partial pressure in the chamber; the pulling speed adjustment is sent to the lifting motor, which adjusts the linear speed of crystal pulling; and the electromagnetic field strength parameter is sent to the electromagnetic stirring coil outside the furnace body, which adjusts the intensity of melt convection control.

[0060] This application embodiment uses a combination of machine vision model and growth control model to determine control parameters, and controls the growth of indium phosphide crystals based on these control parameters. This breaks through the limitations of traditional manual setting of fixed parameters and can achieve real-time dynamic control of the growth process without relying on the experience and judgment of operators. It not only solves the control problem caused by the nonlinear coupling of phosphorus vapor pressure and temperature, but also responds to pressure changes caused by phosphorus volatilization in advance through feedforward compensation, effectively reducing the generation of crystal defects such as twins and dislocations, and improving the overall growth yield and crystal quality of indium phosphide single crystals.

[0061] Furthermore, to ensure the real-time performance of control parameter inference and growth furnace control, embodiments of this application employ a collaborative framework combining end-side and edge-side approaches. Specifically, The machine vision model is deployed in the processing unit supporting the X-ray imaging system (that is, composed of an X-ray imaging unit and an X-ray detector). After the interface image is collected by the X-ray imaging system, the interface image is directly subjected to image recognition through the machine vision model deployed in its processing unit to obtain global morphological features. The X-ray imaging system communicates with the industrial control computer via Ethernet, and the inference latency is less than 50ms per frame.

[0062] The growth furnace control is deployed in the form of Docker containers in the industrial control computer (such as Windows / Linux industrial control computers) of the growth furnace. The industrial control computer performs two-way data exchange with the growth controller (such as a PLC controller) of the growth furnace through the OPC UA (IEC 62541) industrial communication protocol. The industrial control computer reads real-time observation data (with a sampling period of 1s) from the growth controller, and writes the determined control parameters back to the register of the growth controller in the form of control instructions, so as to drive the corresponding actuators through the growth controller.

[0063] Wherein, the cooperative control process of the end-side and edge-side cooperative framework does not directly adopt the existing single PID control process, but adopts a cooperative control process of upper-layer decision-making and bottom-layer PID execution. The cooperative control process is specifically as follows: Upper-layer decision-making: The reinforcement learning model outputs the control parameter items of each actuator (that is, temperature adjustment amount, air pressure adjustment amount, pulling speed adjustment amount, and electromagnetic field intensity) according to the change trend of phosphorus vapor pressure in the future period predicted by the time series prediction module and the global morphological features identified by the machine vision model.

[0064] Bottom-layer PID: Each actuator (that is, the induction heater, the pressure controller, the pulling motor, and the electromagnetic stirring coil) is equipped with an independent PID controller, which is responsible for accurately tracking the control parameter items issued by the upper-layer decision-making. The bottom-layer PID utilizes the existing PID control loop, but its set value is no longer given by manual experience, but dynamically given in real time by the growth control model.

[0065] Therefore, the embodiments of the present application do not abandon the existing PID control, but superimpose an upper-layer decision-making layer on the existing PID execution layer, so that the target set value of PID is no longer a fixed curve, but a target value dynamically optimized by the upper-layer decision-making through the generation control model according to real-time working conditions. The priority and timing rules followed by the cooperative control of each actuator are shown in Table 1.

[0066] Table 1 The following priority and timing rules are followed for the cooperative control of each actuator

[0067] Furthermore, within each control cycle (e.g., 5 seconds), the growth control model simultaneously determines the temperature adjustment, air pressure adjustment, pulling speed adjustment, and electromagnetic field strength. However, the PLC control executes these parameters sequentially according to priority order from priority 1 to priority 4. During execution, if the air pressure deviation exceeds the safety threshold... If the pressure deviation exceeds 10 MPa, the pressure controller will be executed first and other actuators will be suspended; if the pressure deviation does not exceed the safety threshold, the temperature regulation, pressure regulation, pull speed regulation and electromagnetic field strength will be issued in parallel.

[0068] To further illustrate the embodiments of this application, the control cycle of the constant diameter growth stage is taken as an example. The specific control cycle is as follows: Initial state (t=0): Current temperature: 1065℃ (set value 1065℃, deviation 0); Chamber pressure: 2.65MPa (set value 2.60MPa, deviation +0.05MPa); Pulling speed: 8.0 mm / h (set value 8.0 mm / h, deviation 0); Electromagnetic field strength: 50mT; Machine vision model output: Global morphological characteristics: The curvature of the solid-liquid interface changes from convex to flat (the radius of curvature increases from 120mm to 200mm). Time series prediction module (future 5 minutes): The trend of phosphorus vapor pressure change: The phosphorus vapor pressure will continue to rise to 2.80 MPa (exceeding the safe threshold of 2.70 MPa). Predicted cause: The drop in melt level leads to a thinning of the B2O3 liquid seal layer, which accelerates the phosphorus escape rate; Reinforcement learning module output: Temperature regulation amount: (Cooling down to reduce the rate of phosphorus volatilization); Air pressure regulation amount: (Slightly open the exhaust valve to release excess phosphorus vapor); Speed ​​adjustment amount: (Slightly accelerate the pulling process to allow the crystal to escape the high-temperature zone as quickly as possible); Electromagnetic field modulation amount: (Enhanced stirring improves temperature uniformity below the liquid surface).

[0069] Collaborative control: Step 1 (0-500ms): Determine the air pressure deviation. It did not exceed the safety threshold, and the four paths executed in parallel; Step 2 (500ms-1s): The pressure controller adjusts the exhaust proportional valve opening from 12% to 18%, and the target air pressure is set to... ; Step 3 (1-3s): The induction heater transfers power from... Down to The target temperature is set to ; Step 4 (Immediately): The lifting motor will increase the pulling speed from... Adjust to ; Step 5 (3-10s): The current in the electromagnetic stirring coil changes from... Rise to The magnetic field strength from Rise to .

[0070] Execution result (at the start of the next control cycle): Chamber pressure: (Approaching the target value) temperature: (Currently descending) Speed ​​of pulling: (In place) Machine vision model feedback: The trend of interface curvature decrease has slowed down.

[0071] The key difference in the above process is that traditional PID can only operate when the air pressure has already risen to a certain level. The previous approach only responded passively, while this application uses a time-series prediction module to predict in advance and a reinforcement learning module to make feedforward decisions, before the air pressure gets out of control (only the deviation is significant). It proactively adopts a multi-actuator coordinated adjustment. This "predictive-feedforward-coordinated" control paradigm is something that existing single-loop PID control cannot achieve.

[0072] In one embodiment, to ensure that the growth control model and machine vision model can continuously adapt to the indium phosphide crystal growth process, a training and update strategy of offline pre-training + online fine-tuning is adopted. Specifically, based on the training data generated from multiple indium phosphide crystal growth cycles, the growth control model and machine vision model are initially pre-trained in the cloud, resulting in trained growth control and machine vision models. Then, during actual production operation, after each successful crystal growth cycle, all interface images, observation data, and control parameters for that cycle are transmitted back to the cloud server. The cloud server fine-tunes the growth control model and machine vision model based on the new data, and then sends the updated model parameters back to the end-side and edge-side deployment, enabling the model to gradually adapt to the individual characteristics of a specific growth furnace and continuously improve control accuracy as production progresses.

[0073] Based on this, in one embodiment, such as Figure 3 As shown, the method further includes: The complete growth process data of multiple indium phosphide crystals are aggregated through a cloud server; Federated learning is performed on the machine vision model and the growth control model using complete growth process data to jointly fine-tune the machine vision model and the growth control model, so that the side device executing the indium phosphide crystal growth control method based on multimodal deep learning can determine the control parameters using the jointly fine-tuned machine vision model and the growth control model.

[0074] Specifically, after crystal growth is completed, the processed crystal is subjected to XRD diffraction rocking curve test (to obtain full width at half maximum (FWHM)), etch pit density (EPD) test and optical uniformity test to obtain crystal quality. Then, crystal batches with crystal quality higher than the preset quality threshold are selected from the crystal growth data, and the complete growth process data of each crystal in the crystal batch is obtained and uploaded to the cloud database. The complete growth process data includes interface image sequence, observation data sequence and control parameter sequence.

[0075] Then, the cloud server periodically triggers federated learning training tasks, using complete growth process data from the cloud database to jointly fine-tune the machine vision model and the growth control model. After the joint fine-tuning is completed, a model update package is generated and sent to each side industrial control unit via the MQTT protocol. The model is automatically updated before the start of the next crystal growth cycle, realizing the continuous accumulation of growth data from multiple cycles and the iterative optimization of model performance, thereby continuously improving control accuracy and crystal yield.

[0076] In summary, this embodiment provides a method for controlling the growth of indium phosphide crystals based on multimodal deep learning. This method includes: acquiring an interface image of the solid-liquid interface in a growth furnace using X-rays; inputting the interface image into a machine vision model to capture the global morphological features of the solid-liquid interface; determining the phosphorus vapor pressure variation trend over a future time period based on historical observation data sequences using a time-series prediction module in the growth control model; determining control parameters based on the variation trend and global morphological features using a reinforcement learning module in the growth control model; and controlling the growth of the indium phosphide crystal based on the control parameters. This application, by identifying global morphological features through a machine vision model and predicting the phosphorus vapor pressure variation trend through a time-series prediction module, and then combining the global morphological features and the phosphorus vapor pressure variation trend to determine control parameters, enables real-time dynamic control of the indium phosphide crystal growth process. This solves the response lag problem caused by traditional fixed parameter control relying on manual experience, effectively reducing the twinning rate and dislocation density of the crystal, and improving the growth yield and crystal quality of indium phosphide single crystals. Meanwhile, this application uses the changing trend of phosphorus vapor pressure as the basis for determining control parameters, making advance predictions and adjustments for possible pressure changes and operating condition fluctuations, achieving effective feedforward compensation, further avoiding crystal defects caused by lag control, and improving the stability of the crystal growth process.

[0077] Based on the above-mentioned indium phosphide crystal growth control method based on multimodal deep learning, this embodiment provides an indium phosphide crystal growth control device based on multimodal deep learning, such as... Figure 4 As shown, the device includes: The acquisition module 100 is used to acquire interface images of the solid-liquid interface in the growth furnace using X-rays. The capture module 200 is used to input the interface image into the machine vision model and capture the global morphological features of the solid-liquid interface through the machine vision model. The time series prediction module 300 is used to determine the trend of phosphorus vapor pressure change over a future time period based on historical observation data sequences through the time series prediction module in the growth control model. Module 400 is used to determine control parameters based on the changing trend and global morphological features through the reinforcement learning module in the growth control model. The control module 500 is used to control the growth of the indium phosphide crystal based on the control parameters.

[0078] Based on the above-described method for controlling the growth of indium phosphide crystals based on multimodal deep learning, this embodiment provides a computer-readable storage medium storing one or more programs that can be executed by one or more processors to implement the steps in the method for controlling the growth of indium phosphide crystals based on multimodal deep learning as described in the above embodiment.

[0079] Based on the above-mentioned method for controlling the growth of indium phosphide crystals using multimodal deep learning, this application also provides a terminal device, such as... Figure 5 As shown, it includes at least one processor 20; a display screen 21; and a memory 22, and may also include a communications interface 23 and a bus 24. The processor 20, display screen 21, memory 22, and communications interface 23 can communicate with each other via the bus 24. The display screen 21 is configured to display a preset user guide interface in the initial setup mode. The communications interface 23 can transmit information. The processor 20 can invoke logical instructions in the memory 22 to execute the methods described in the above embodiments.

[0080] Furthermore, the logical instructions in the aforementioned memory 22 can be implemented as software functional units and, when sold or used as independent products, can be stored in a computer-readable storage medium.

[0081] The memory 22, as a computer-readable storage medium, can be configured to store software programs, computer-executable programs, such as program instructions or modules corresponding to the methods in the embodiments of this disclosure. The processor 20 executes functional applications and data processing by running the software programs, instructions, or modules stored in the memory 22, thereby implementing the methods in the above embodiments.

[0082] The memory 22 may include a program storage area and a data storage area. The program storage area may store the operating system and application programs required for at least one function; the data storage area may store data created based on the use of the terminal device. Furthermore, the memory 22 may include high-speed random access memory (RAM) and non-volatile memory. Examples include various media capable of storing program code, such as USB flash drives, portable hard drives, read-only memory (ROM), random access memory (RAM), magnetic disks, or optical disks, as well as transient storage media.

[0083] Furthermore, the specific process of loading and executing multiple instruction processors in the aforementioned storage medium and terminal device has been described in detail in the above method, and will not be repeated here.

[0084] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of this application, and are not intended to limit them. Although this application has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some of the technical features. Such modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the spirit and scope of the technical solutions of the embodiments of this application.

Claims

1. A method for controlling the growth of indium phosphide crystals based on multimodal deep learning, characterized in that, The method includes: Interface images of the solid-liquid interface in the growth furnace were acquired using X-rays. The interface image is input into the machine vision model, and the global morphological features of the solid-liquid interface are captured by the machine vision model. The trend of phosphorus vapor pressure variation in the future time period is determined by the time-series prediction module in the growth control model based on historical observation data sequences. The control parameters are determined based on the changing trend and global morphological features through the reinforcement learning module in the growth control model. The growth of the indium phosphide crystal is controlled based on the aforementioned control parameters.

2. The indium phosphide crystal growth control method based on multimodal deep learning according to claim 1, characterized in that, The specific steps of acquiring the interface image of the solid-liquid interface in the growth furnace using X-rays are as follows: X-rays are emitted into the growth furnace through an X-ray imaging unit installed on the side wall of the high-pressure chamber. The X-ray detector located on the opposite side of the X-ray imaging unit receives light transmitted through the side wall of the high-pressure chamber and... X-rays are taken after the liquid seal layer, and a grayscale projection image of the solid-liquid interface is formed based on the X-rays to obtain an interface image.

3. The indium phosphide crystal growth control method based on multimodal deep learning according to claim 1, characterized in that, The step of inputting the interface image into a machine vision model and capturing the global morphological features of the solid-liquid interface through the machine vision model specifically includes: The interface image is input into the grayscale enhancement preprocessing module in the machine vision model. The grayscale enhancement preprocessing module enhances the contrast of the interface image to obtain an enhanced interface image. The enhanced interface image is input into the denoising module in the machine vision model, and the X-ray scattering noise is removed by the denoising module to obtain the denoised interface image. Based on the denoised interface image, it is filtered through a deblurred convolutional attention layer in the machine vision model. Image noise from the liquid sealing layer is used to extract the initial feature map; Based on the initial feature map, the target feature map is extracted through the self-attention layer in the machine vision model, and the global morphological features are determined based on the target feature map.

4. The method for controlling the growth of indium phosphide crystals based on multimodal deep learning according to claim 1 or 3, characterized in that, The global morphological features include the curvature of the solid-liquid interface, The thickness of the liquid seal layer and the state of seed crystal fusion.

5. The method for controlling the growth of indium phosphide crystals based on multimodal deep learning according to claim 1, characterized in that, The control parameters include temperature regulation, air pressure regulation, pulling speed regulation, and electromagnetic field strength.

6. The method for controlling the growth of indium phosphide crystals based on multimodal deep learning according to claim 1 or 5, characterized in that, The reinforcement learning module uses the changing trend and global morphological features as the state space, the control parameters as the action space, and constructs a reward function based on the crystal dislocation density and twinning rate. It then determines the optimal control policy through a deep deterministic policy gradient approach.

7. The method for controlling the growth of indium phosphide crystals based on multimodal deep learning according to claim 1, characterized in that, The method further includes: The complete growth process data of multiple indium phosphide crystals are aggregated through a cloud server; Federated learning is performed on the machine vision model and the growth control model using complete growth process data to jointly fine-tune the machine vision model and the growth control model, so that the side device executing the indium phosphide crystal growth control method based on multimodal deep learning can determine the control parameters using the jointly fine-tuned machine vision model and the growth control model.

8. A device for controlling the growth of indium phosphide crystals based on multimodal deep learning, characterized in that, The device includes: The acquisition module is used to acquire interface images of the solid-liquid interface in the growth furnace using X-rays. The capture module is used to input the interface image into the machine vision model and capture the global morphological features of the solid-liquid interface through the machine vision model. The time-series prediction module is used to determine the trend of phosphorus vapor pressure changes over a future time period based on historical observation data sequences, using the time-series prediction module in the growth control model. The determination module is used to determine control parameters based on the changing trend and global morphological features through the reinforcement learning module in the growth control model. A control module is used to control the growth of the indium phosphide crystal based on the control parameters.

9. A computer-readable storage medium, characterized in that, The computer-readable storage medium stores one or more programs, which can be executed by one or more processors to implement the steps in the indium phosphide crystal growth control method based on multimodal deep learning as described in any one of claims 1-7.

10. A terminal device, characterized in that, include: Processor and memory; The memory stores a computer-readable program that can be executed by the processor; When the processor executes the computer-readable program, it implements the steps in the indium phosphide crystal growth control method based on multimodal deep learning as described in any one of claims 1-7.

Images