A floating zone method crystal growth control method, system and floating zone furnace apparatus

Abstract

This application discloses a method, system, and floating zone furnace for controlling crystal growth using the floating zone method. The method includes: synchronously acquiring process parameter data and visual image data of the crystal growth interface characterizing the molten zone morphology within the floating zone furnace in real time; inputting the process parameter data and visual image data as multimodal inputs to a pre-trained crystal growth morphology prediction model, whereby the crystal growth morphology prediction model extracts and fuses features from the process parameter data and visual image data, outputting a quantitative prediction sequence of crystal growth morphology parameters in the future prediction time domain; based on the quantitative prediction sequence, using a real-time process control model employing a model predictive control framework, under the condition of satisfying preset operational constraints, continuously optimizing and solving for real-time control commands within the current control cycle that make the predicted output trajectory approximate the target growth trajectory; and sending the real-time control commands to the actuators of the floating zone furnace to adjust the process parameter data and control the crystal growth process.

Description

Technical Field

[0001] This application relates to the fields of advanced crystal material preparation and intelligent manufacturing technology, specifically to a floating zone crystal growth control method, system, and floating zone furnace equipment. Background Technology

[0002] The floating zone method, a crucible-free crystal growth technique, is a key method for preparing high-purity, high-melting-point functional crystals (such as semiconductor single crystals, laser crystals, and superconducting materials). Its basic principle is to form a molten zone in a polycrystalline raw material rod through localized heating, and to achieve directional solidification growth of the crystal by relying on the surface tension of the melt and the relative movement between the polycrystalline material and the already crystallized single crystal.

[0003] However, the traditional floating zone method faces severe challenges in process control. The core of the growth process—the solid-liquid interface morphology, melt zone stability, and crystal diameter—mainly relies on visual judgment and manual adjustment by operators through an observation window. This approach has inherent drawbacks, such as subjective differences in manual judgment and the fact that crystal defects often form by the time problems are discovered, ultimately resulting in low production efficiency and difficulty in meeting the requirements of modern high-end manufacturing in terms of crystal consistency and yield. Summary of the Invention

[0004] To address the aforementioned technical issues, this application discloses a floating zone crystal growth control method, system, and floating zone furnace equipment. By integrating real-time visual representation with process parameters, it utilizes an artificial intelligence model to quantitatively predict the growth morphology and achieves autonomous optimization and real-time adjustment of process parameters based on a model predictive control framework. Ultimately, it realizes adaptive, high-precision, and intelligent closed-loop control of the crystal growth process, improving crystal quality, yield, and process reproducibility.

[0005] Specifically, the technical solution of this application is as follows: In a first aspect, this application discloses a method for controlling crystal growth using the floating zone method, comprising the following steps: The process parameter data and crystal growth interface visual image data characterizing the molten zone morphology are collected in real time and synchronously. The process parameter data includes one or more of the following: heater power, crystal pulling speed, raw material feeding speed and crystal rotation speed. The process parameter data and the visual image data are used as multimodal inputs and input to a pre-trained crystal growth morphology prediction model. The crystal growth morphology prediction model extracts and fuses the features of the process parameter data and the visual image data, and outputs a quantitative prediction sequence of crystal growth morphology parameters in the future prediction time domain. Based on the quantized prediction sequence, the process real-time control model using the model predictive control framework is used to continuously optimize and solve the real-time control command within the current control cycle that makes the predicted output trajectory approximate the target growth trajectory, under the condition of satisfying the preset operation constraints. The real-time control command is sent to the actuator of the floating zone furnace to adjust the process parameter data and control the crystal growth process.

[0006] In some embodiments, the crystal growth morphology prediction model is a multimodal fusion neural network model, including: Feature extraction layer, including: The visual processing branch is used to extract image features reflecting the morphology of the melt zone and the characteristics of the solid-liquid interface from the visual image data. The parameter processing branch is used to extract timing features from the process parameter data; The feature fusion layer is used to deeply fuse the image features and the temporal features to obtain fused features; The output layer is used to output the quantized prediction sequence based on the fused features.

[0007] In some implementations, the vision processing branch employs a convolutional neural network, including at least one of a 3D convolutional network, a temporal convolutional network, or a Vision Transformer to extract image features. The parameter processing branch uses a recurrent neural network, including at least one of a long short-term memory network or a gated recurrent unit to extract temporal features; The feature fusion layer uses feature cascading or attention mechanisms for feature fusion.

[0008] In some implementations, the following is specifically included: Under the condition of satisfying preset operational constraints, continuously optimizing and solving for real-time control commands within the current control cycle to make the predicted output trajectory approximate the target growth trajectory; Within each control cycle, the crystal growth morphology prediction model is used as an internal dynamic model to predict the system output behavior under different candidate control command sequences in the future finite time domain. A cost function is constructed to quantify the deviation between the predicted output trajectory and the target growth trajectory, as well as the magnitude of the change in the control command; Under the preset operational constraints, the optimal control command that minimizes the cost function is solved by the real-time process control model.

[0009] Optionally, the real-time process control model is a first control model based on a reinforcement learning algorithm, wherein the reward function of the first control model is negatively correlated with the cost function; Alternatively, the real-time process control model may be a second control model trained based on an adaptive optimization algorithm.

[0010] In some implementations, the reward function is positively correlated with crystal growth quality evaluation indicators; The reinforcement learning algorithm includes one of deep deterministic policy gradient, proximal policy optimization, or soft actor-critic algorithm.

[0011] In some embodiments, the floating zone method for crystal growth control further includes: updating and optimizing the parameters of the crystal growth morphology prediction model and / or the real-time process control model based on the difference between the actual crystal growth result and the predicted result, or based on the quality evaluation result of the crystal after growth.

[0012] Secondly, this application also discloses a floating zone crystal growth control system, comprising: The data acquisition module is used to collect process parameter data and crystal growth interface visual image data characterizing the molten zone morphology in the floating zone furnace in real time and synchronously. The process parameter data includes one or more of the following: heater power, crystal pulling speed, raw material feeding speed, and crystal rotation speed. The data processing and prediction module has an embedded crystal growth morphology prediction model as described in any of the above embodiments, which is used to extract and fuse the features of the process parameter data and the visual image data, and output a quantitative prediction sequence of crystal growth morphology parameters in the future prediction time domain. The control decision module, which is embedded with the real-time process control model as described in any of the above embodiments, is used to perform rolling optimization to solve the real-time control command that makes the predicted output trajectory approximate the target growth trajectory within the current control cycle based on the quantized prediction sequence and under the condition of satisfying the preset operation constraints. The execution interface module is used to send the real-time control commands to the actuator of the floating zone furnace to adjust the process parameter data and control the crystal growth process.

[0013] Thirdly, this application also discloses an intelligent floating zone furnace device, which includes a furnace body, a heating device, a seed crystal rod, a raw material rod, and a motion mechanism; It also includes the floating zone crystal growth control system described in the above embodiments; And an actuator that is communicatively connected to the execution interface module of the control system, the actuator including at least one of a heater power controller, a lifting motor driver, a feeding motor driver, and a rotary motor driver.

[0014] In some embodiments, the furnace body of the intelligent floating zone furnace is provided with an optical observation window dedicated to the visual observation unit. The position and angle of the optical observation window are configured to clearly observe the crystal growth interface and the outline of the melting zone.

[0015] Compared with the prior art, this application has at least one of the following beneficial effects: 1. This application realizes closed-loop intelligent control through multi-source information fusion: For the first time, unstructured visual image information is deeply integrated with structured process parameters and introduced into the control closed loop as the core input, enabling the system to "see" and "understand" the growth state and make more accurate decisions.

[0016] 2. Possesses forward-looking control capabilities: Through morphological prediction models, the system can predict trends such as diameter drift and interface instability in advance, enabling the control behavior to change from passive response to active intervention, effectively preventing defects from occurring.

[0017] 3. Improved robustness and optimality of control: By adopting a model predictive control framework and solving for the optimal control sequence in the finite time domain through rolling optimization, it can systematically handle multivariate constraints (such as power upper limit and speed range) and continuously track the optimal growth trajectory in dynamic changes.

[0018] 4. It endows the system with self-learning and evolution capabilities: The model supports online learning and updates, and can continuously accumulate knowledge from actual growth results, adapt to different material properties or equipment differences, and achieve continuous performance improvement and automated accumulation of process knowledge.

[0019] 5. Promoted the standardization and automation of the process: reduced over-reliance on the personal experience of operators, and made the high-quality growth process of complex crystals more repeatable and scalable. Attached Figure Description

[0020] The preferred embodiments will now be described in a clear and easy-to-understand manner, in conjunction with the accompanying drawings, to further explain the above-mentioned characteristics, technical features, advantages, and implementation methods of this application.

[0021] Figure 1 This is a flowchart illustrating the steps of an embodiment of a floating zone crystal growth control method according to this application; Figure 2 This is an example of visual image data used to input the crystal growth morphology prediction model in this application embodiment; Figure 3 This is a schematic diagram of the visual branch extracting edge features from the input image data in this embodiment of the application, extracting the pixel edges of the crystal, and forming edge lines; Figure 4This is a schematic diagram of how the visual branch performs morphological calculations on the input image data and fits the crystal to obtain features such as the elliptical edge, center length, maximum width, and lateral deviation. Figure 5 This is a structural block diagram of an embodiment of a floating zone crystal growth control system according to this application. Detailed Implementation

[0022] In the following description, specific details such as particular system architectures and techniques are set forth for illustrative purposes and not for limitation, in order to provide a thorough understanding of the embodiments of this application. However, those skilled in the art will understand that this application can also be implemented in other embodiments without these specific details. In other instances, detailed descriptions of well-known systems, apparatuses, circuits, and methods have been omitted so as not to obscure the description of this application with unnecessary detail.

[0023] It should be understood that, when used in this specification and the appended claims, the term "comprising" indicates the presence of the described features, integrals, steps, operations, elements and / or components, but does not exclude the presence or addition of one or more other features, integrals, steps, operations, elements, components and / or sets.

[0024] To keep the drawings concise, each figure only schematically shows the parts relevant to the invention, and these do not represent the actual structure of the product. Furthermore, to facilitate understanding, in some figures, only one of components with the same structure or function is schematically depicted, or only one is labeled. In this document, "one" not only means "only one," but can also mean "more than one."

[0025] It should also be further understood that the term “and / or” as used in this application specification and the appended claims means any combination of one or more of the associated listed items and all possible combinations, and includes such combinations.

[0026] Furthermore, in the description of this application, the terms "first," "second," etc., are used only to distinguish descriptions and should not be construed as indicating or implying relative importance.

[0027] To more clearly illustrate the technical solutions in the embodiments of this application or the prior art, the specific implementation methods of this application will be described below with reference to the accompanying drawings. Obviously, the accompanying drawings described below are only some embodiments of this application. For those skilled in the art, other drawings and other implementation methods can be obtained based on these drawings without creative effort.

[0028] In the existing technology, the process control of the traditional floating zone method faces severe challenges. The core of its growth process, namely the solid-liquid interface morphology, melt zone stability and crystal diameter, mainly relies on the operator's visual judgment through the observation window and experience-based manual adjustment.

[0029] This model inevitably introduces a high degree of human subjectivity. Different operators, and even the same operator, may have inconsistent judgment standards under different conditions, leading to extremely poor process repeatability and huge batch-to-batch fluctuations in crystal quality. Furthermore, the experience of senior operators is difficult to quantify, solidify, and transfer, making process repeatability and stability highly susceptible to human factors.

[0030] Furthermore, the human eye has physiological limits in terms of observation ability and reaction speed. Operators often find it difficult to detect early defect signals such as subtle morphological changes in the molten zone and small temperature fluctuations in time. By the time the problem can be clearly observed with the naked eye, the growth process may have already deviated from the optimal state, or even lead to irreversible failures such as polycrystalline or cracking.

[0031] Prolonged periods of intense concentration for observation and manual adjustments not only increase the workload of operators but also make them prone to fatigue, further increasing the risk of misjudgment and misoperation. Ultimately, this leads to low production efficiency, making it difficult to meet the requirements of modern high-end material manufacturing for consistency, reliability, and yield.

[0032] In addition, existing automation solutions are mostly based on PID feedback control of single or a few process parameters such as temperature and speed, which cannot effectively integrate process parameters and key visual morphological information, resulting in limited control accuracy and anti-interference capabilities.

[0033] Although some studies have attempted to incorporate image processing techniques for diameter measurement or use machine learning for process parameter analysis, these methods are often isolated and fail to construct a real-time closed-loop control system of "multi-source sensing - intelligent prediction - optimization decision-making". Therefore, developing an intelligent floating zone system capable of interpreting growth visual information in real time, predicting growth dynamics, and automatically generating optimal control strategies is of great significance for overcoming the bottleneck of controllable preparation of high-end crystal materials.

[0034] Reference manual attached Figure 1 As shown, one embodiment of the floating zone method for controlling crystal growth according to this application specifically includes the following steps: S1, real-time synchronous acquisition of process parameter data and visual image data of crystal growth interface characterizing the molten zone morphology in the floating zone furnace.

[0035] Optionally, the process parameter data includes one or more of the following: heater power, crystal pulling speed, raw material feed speed, and crystal rotation speed.

[0036] In some implementations, heater power determines the heat input and temperature distribution within the molten zone. Crystal pulling speed controls the axial growth rate of the crystal. Feed rate is used to match the consumption of the polycrystalline ingot to maintain a constant molten zone volume. Crystal rotation speed influences the convection patterns within the melt and the flatness of the solid-liquid interface. These parameters collectively constitute structured data that quantitatively describes the growth process.

[0037] Visual image data characterizing the morphology of the molten zone are acquired by a high-speed camera with a narrow-band filter and a dedicated light source, which can clearly present key visual features such as the crystal's outline diameter, the degree of necking of the molten zone, and the curvature of the solid-liquid interface.

[0038] S2, the process parameter data and the visual image data are used as multimodal inputs and input to a pre-trained crystal growth morphology prediction model. The crystal growth morphology prediction model extracts and fuses the features of the process parameter data and the visual image data, and outputs a quantitative prediction sequence of crystal growth morphology parameters in the future prediction time domain.

[0039] Specifically, the core function of the crystal growth morphology prediction model is to extract high-dimensional features from multimodal data within a fixed time window before the current moment, and to mine the coupling relationship between process parameters and visual morphology through fusion learning, and finally output a quantitative prediction sequence of morphological parameters such as crystal diameter, melt zone width or interface curvature for a period of time in the future.

[0040] In an optional embodiment, the crystal growth morphology parameters include one or more of the following: crystal diameter, melt zone width, solid-liquid interface curvature, and melt zone stability index.

[0041] S3, based on the quantized prediction sequence, the process real-time control model using the model predictive control framework is used to continuously optimize and solve the real-time control command within the current control cycle that makes the predicted output trajectory approximate the target growth trajectory, under the condition of satisfying the preset operation constraints.

[0042] Specifically, in each control cycle, the model predictive control framework uses the aforementioned morphological prediction model as an internal dynamic model to predict the future output behavior of the system under different candidate control commands. Then, by solving a constrained optimization problem online, it finds a set of control commands that make the predicted output trajectory as close as possible to the preset constant diameter or variable diameter growth trajectory. The first command in the command sequence is then issued and executed, and this rolling optimization process is repeated in the next cycle.

[0043] S4, the real-time control command is sent to the actuator of the floating zone furnace to adjust the process parameter data and control the crystal growth process.

[0044] Specifically, the actuators include a programmable power supply for powering the heater, and precision servo motors for driving the lifting and feeding shafts. The actuators receive real-time control commands and convert them into physical actions, ultimately achieving closed-loop regulation of the heater power, lifting speed, and feeding speed, thus completing closed-loop control of the entire crystal growth process.

[0045] In some optional implementations, the process parameter data is adjusted to control the crystal growth process, specifically including: adjusting the heating power to change the temperature distribution and thermal stability of the molten zone; adjusting the pulling speed to control the axial growth rate of the crystal; adjusting the feed speed to match the consumption rate of the polycrystalline raw material rod to maintain a constant molten zone volume; adjusting the crystal rotation speed to affect the melt convection mode and the flatness of the solid-liquid interface; adjusting the raw material rod rotation speed to improve the temperature uniformity and composition mixing in the molten zone; adjusting the cooling water flow rate or temperature to fine-tune the thermal environment inside the furnace; and adjusting the atmosphere pressure or airflow speed in an atmosphere-controlled floating zone furnace to suppress component volatilization or regulate the surface tension of the melt. Through the comprehensive adjustment of the above single or multiple process parameters, precise control of the crystal diameter, molten zone morphology, and solid-liquid interface stability can be achieved.

[0046] Based on the above embodiments, this application discloses another embodiment of a floating zone crystal growth control method, wherein the crystal growth morphology prediction model is a multimodal fusion neural network model, including: Feature extraction layer, including: The visual processing branch is used to extract image features reflecting the morphology of the melt zone and the characteristics of the solid-liquid interface from the visual image data.

[0047] The parameter processing branch is used to extract timing features from the process parameter data.

[0048] The feature fusion layer is used to deeply fuse the image features and the temporal features to obtain fused features.

[0049] The output layer is used to output the quantized prediction sequence based on the fused features.

[0050] Specifically, in some embodiments, the crystal growth morphology prediction model is a multimodal fusion neural network model, whose inputs are time-series data of process parameters within a time window and a synchronized sequence of visual images, as detailed in the appendix to the specification. Figure 2 As shown, Figure 2 This is an example of visual image data used to input the crystal growth morphology prediction model in this embodiment.

[0051] The model architecture employs a dual-branch fusion network. This is to fully utilize the two types of heterogeneous data generated during the floating zone crystal growth process, including time-series data of process parameters reflecting thermodynamic states and visual image data that intuitively presents the dynamics of the molten zone. By jointly learning, the deep coupling relationship between the two is explored, thereby achieving high-precision quantitative prediction of crystal growth morphology.

[0052] In this model's architecture, the vision branch is used to extract high-dimensional semantic information from the original images. Specifically, the vision branch first uses a convolutional neural network to encode each frame of the crystal growth interface image, extracting features that characterize the melt zone contour, the degree of solid-liquid interface curvature, and the crystal diameter. (See attached instruction manual.) Figure 3 , 4 As shown, Figure 3 This embodiment illustrates how the visual branch extracts edge features from the input image data, extracting the pixel edges of the crystal to form an edge line. Figure 4 In this embodiment, the visual branch performs morphological calculations on the input image data and fits the crystal to obtain a schematic diagram showing the characteristics of the crystal's elliptical edge, center length, maximum width, and lateral deviation.

[0053] Optionally, the visual processing branch employs a convolutional neural network (ResNet, EfficientNet), including at least one of a 3D convolutional network, a temporal convolutional network, or a Vision Transformer to extract image features. This application is not specifically limited. Subsequently, to capture the evolution of these visual features over time, the model further introduces a 3D convolutional or temporal convolutional network to learn the dynamic changing trend of the melt zone morphology by temporally modeling the features of multiple consecutive frames of images.

[0054] Meanwhile, the parameter branch focuses on processing time-series data composed of various physical quantities closely related to the crystal growth process, including heating power, pulling speed, feed speed, rotation speed of the crystal and the raw material rod, and melting zone temperature.

[0055] Optionally, the parameter processing branch employs a recurrent neural network, including at least one of a long short-term memory network or a gated recurrent unit to extract temporal features. These network structures are particularly adept at capturing long-range dependencies in temporal data due to their built-in memory units and gating mechanisms, and can effectively learn how historical changes in process parameters affect the current growth state.

[0056] After visual features and temporal features are extracted separately, the model deeply fuses the two through a feature fusion layer, which uses feature cascading or attention mechanisms for feature fusion.

[0057] Optionally, the feature fusion layer employs a feature concatenation approach, concatenating two types of feature vectors into a higher-dimensional comprehensive representation. Alternatively, the feature fusion layer uses an attention mechanism, allowing the model to automatically learn whether to emphasize visual information or process parameter information at different growth stages, and assigning appropriate weights to features from different sources, thereby achieving adaptive dynamic fusion.

[0058] The fused feature vector is ultimately passed to a regression output layer composed of fully connected layers, and finally the predicted value is regressed through the fully connected layers. Optionally, the number of neurons in the regression output layer corresponds to the number of predicted targets, which in this application means outputting quantified values ​​of crystal growth morphology parameters at multiple time points in the future prediction time domain.

[0059] In other embodiments of this example, the crystal growth morphology prediction model requires a large amount of historical experimental data covering different growth conditions during the model training phase. Each training sample uses a sequence of process parameters and a synchronized visual image sequence within a fixed time window before a certain moment as input features, and uses the actual crystal morphology parameter sequence after that moment as supervision labels, forming paired data of "parameter-image-subsequent real morphology". Supervised learning is used during training, employing a mean squared error loss function or a smoothed L1 loss function to measure the deviation between the model's predicted values ​​and the actual values. The backpropagation algorithm is used to continuously optimize hundreds of millions of learnable parameters in the network until the model can accurately predict future crystal growth trends from the input multimodal data.

[0060] This application provides another embodiment of a floating zone crystal growth control method. Based on any embodiment of the above method, step S3 involves: based on the quantization prediction sequence, using a model predictive control framework-based real-time process control model, and under the condition of satisfying preset operational constraints, continuously optimizing and solving for real-time control commands within the current control cycle that make the predicted output trajectory approximate the target growth trajectory. Specifically, this includes: S31, within each control cycle, the crystal growth morphology prediction model is used as an internal dynamic model to predict the system output behavior under different candidate control command sequences in the future finite time domain.

[0061] S32, Construct a cost function, which is used to quantify the deviation between the predicted output trajectory and the target growth trajectory, as well as the magnitude of the change in the control command.

[0062] S33, Under the preset operational constraints, the optimal control command that minimizes the cost function is solved by the real-time process control model.

[0063] Specifically, this embodiment utilizes a real-time process control model based on Model Predictive Control (MPC). MPC is a model-based rolling optimization control strategy. In each control cycle, the controller uses an internal dynamic model to predict the system's output behavior under different control input sequences within a finite future time domain.

[0064] The controller solves an optimization problem to find a sequence of control commands that makes the predicted output trajectory as close as possible to the desired reference trajectory, while satisfying various operational constraints (such as power limits). Mathematically, this can be expressed as minimizing a cost function that includes tracking error and control variable changes. Specifically, the cost function is a mathematical expression defined in this application that describes the desired optimal state of the system. The task of the optimization solver is to find a set of control commands that minimizes the value of this cost function.

[0065] Optionally, the cost function is a weighted sum of the deviation between the predicted crystal diameter and the target diameter, as well as the variation in the control command, to prevent drastic fluctuations in power and speed.

[0066] Solving this optimization problem can be instantiated as a reinforcement learning agent (e.g., trained using Deep Deterministic Policy Gradient Algorithm (DDPG) or Proximal Policy Optimization (PPO) algorithms), whose reward function is the negative of the aforementioned cost function. Alternatively, it can be solved online using traditional numerical optimization methods (such as gradient descent or sequential quadratic programming). Reinforcement learning is more suitable for complex nonlinear systems and facilitates online integration.

[0067] In some optional implementations, the real-time process control model is a first control model based on a reinforcement learning algorithm. The reward function is positively correlated with the crystal growth quality evaluation index and negatively correlated with the cost function.

[0068] Specifically, reinforcement learning algorithms are essentially neural networks trained through extensive trial and error. During the training phase, the agent interacts with a simulated environment or historical data. Its learning objective is guided by a reward function designed to be positively correlated with crystal growth quality evaluation indicators. For example, positive rewards are given when crystal diameter tracking errors are small, melting zone stability is high, and control actions are smooth; conversely, negative penalties are imposed. Simultaneously, this reward function is negatively correlated with an implicit cost function. This means that in pursuing the maximization of cumulative rewards, the agent is actually learning how to minimize the cost function value, where the cost function is typically used to quantify the degree of diameter deviation, control fluctuations, and constraint violations. After training, the agent can directly output optimal control commands based on the current input quantized prediction sequence and system state.

[0069] Optionally, the reinforcement learning algorithm includes one of the following specific algorithms: for example, a deep deterministic policy gradient algorithm, which is suitable for control problems in continuous action spaces and can directly output continuous values ​​such as heating power adjustment or lifting speed adjustment. Alternatively, a proximal policy optimization algorithm, which improves training stability by limiting the magnitude of policy updates, and is particularly suitable for scenarios involving pre-training in interaction with a simulation environment. Or, a soft actor-critic (SAC) algorithm, which encourages exploration by maximizing policy entropy, effectively avoiding getting trapped in local optima. Reinforcement learning is more suitable for complex nonlinear systems and is easy to integrate into online learning.

[0070] In some alternative implementations, the real-time process control model is a second control model trained based on an adaptive optimization algorithm.

[0071] Specifically, the adaptive optimization algorithm does not rely on pre-training. Instead, it constructs a cost function online in each control cycle, which includes a diameter tracking error term and a control stability term. Under the condition of satisfying operational constraints such as the upper limit of heating power and the speed adjustment range, it uses a numerical optimization algorithm to solve in real time the optimal control command sequence that minimizes the cost function.

[0072] This application provides another embodiment of a floating zone crystal growth control method, which, based on any of the above embodiments, further includes step S5, updating and optimizing the parameters of the crystal growth morphology prediction model and / or the real-time process control model based on the difference between the actual crystal growth result and the predicted result, or based on the quality evaluation result of the crystal after growth.

[0073] Specifically, after each crystal growth cycle, the quality of the finished crystal, including indicators such as diameter consistency, interface flatness, and defect density, can be evaluated to ultimately verify the effectiveness of the entire control strategy. Based on these quality evaluation results, the real-time process control model can be fine-tuned offline or the reinforcement learning reward signal can be recalibrated.

[0074] Through a dual optimization mechanism that combines prediction deviation-driven and final quality feedback, the model can continuously accumulate process knowledge under different batches, materials, and operating conditions, learn from each growth and continuously evolve, ultimately enabling the control system to have adaptability and process optimization capabilities that exceed fixed programs.

[0075] Based on the same concept, this application also discloses a floating zone crystal growth control system. (See attached specification.) Figure 5 An embodiment of the floating zone crystal growth control system described in this application specifically includes: The data acquisition module is used to synchronously acquire process parameter data and visual image data of the crystal growth interface characterizing the molten zone morphology in real time within the floating zone furnace. The process parameter data includes one or more of the following: heater power, crystal pulling speed, raw material feed speed, and crystal rotation speed.

[0076] Specifically, the data acquisition module consists of a process parameter sensor group and a vision observation unit, which is responsible for high-frequency and synchronous capture of physical signals and image signals reflecting the growth status.

[0077] The data processing and prediction module is embedded with the crystal growth morphology prediction model described in any of the above embodiments, which is used to extract and fuse the features of the process parameter data and the visual image data, and output a quantitative prediction sequence of crystal growth morphology parameters in the future prediction time domain.

[0078] Specifically, the data processing and prediction module receives raw data and performs preprocessing (signal filtering, image enhancement, and data alignment). The core is a crystal growth morphology prediction model. This model learns complex mapping relationships from historical data, extracts features from current multimodal data, and outputs a quantitative prediction sequence of key morphological parameters for a future period. In an optional implementation, the crystal growth morphology parameters include one or more of the following: crystal diameter, melt zone width, solid-liquid interface curvature, and melt zone stability index.

[0079] The control decision module, which embeds the real-time process control model described in any of the above embodiments, is used to perform rolling optimization to solve the real-time control command that makes the predicted output trajectory approximate the target growth trajectory within the current control cycle, based on the quantized prediction sequence and under the condition of satisfying preset operation constraints.

[0080] Specifically, the core of the control decision module is a real-time process control model using a model predictive control framework. This module takes the morphological prediction sequence, the current state, and the target growth trajectory (such as a constant diameter) as inputs. Within a forward prediction time domain, it repeatedly solves the optimization problem to calculate the optimal control command sequence that makes the predicted output closest to the target trajectory, and outputs the first command in the sequence.

[0081] The execution interface module is used to send the real-time control commands to the actuator of the floating zone furnace to adjust the process parameter data and control the crystal growth process.

[0082] Specifically, the execution interface module converts control commands into drive signals, which control the actuators such as the heater power supply and precision motor, thereby changing the actual growth conditions.

[0083] This application discloses another embodiment of a floating zone crystal growth control system. Based on the above embodiment, the data acquisition module includes: a process parameter sensor group and a visual observation unit.

[0084] The process parameter sensor group includes thermocouples, photoelectric encoders, flow meters, etc., and is used to monitor at least one of the following: heater power, temperature in each zone, crystal pulling speed, raw material feed speed, crystal rotation speed, furnace atmosphere pressure and composition.

[0085] The visual observation unit includes at least one high-speed industrial camera and a matching lighting source, used to acquire clear images of the molten zone and crystal growth interface.

[0086] This application provides another embodiment of a floating zone crystal growth control system. Based on any of the above embodiments, the system further includes a model training and update module for performing offline training and online learning and updating of the model. Specifically, the model training and update module updates and optimizes the parameters of the crystal growth morphology prediction model and / or the real-time process control model based on the difference between the actual crystal growth results and the predicted results, or based on the quality evaluation results of the crystal after growth.

[0087] Based on the same technical concept, this application also discloses: an intelligent floating zone furnace device, the intelligent floating zone furnace device including a furnace body, a heating device, a seed crystal rod, a raw material rod, and a motion mechanism.

[0088] It also includes the floating zone crystal growth control system described in the above embodiments.

[0089] And an actuator that is communicatively connected to the execution interface module of the control system, the actuator including at least one of a heater power controller, a lifting motor driver, a feeding motor driver, and a rotary motor driver.

[0090] In some embodiments, the furnace body of the intelligent floating zone furnace is provided with an optical observation window dedicated to the visual observation unit. The position and angle of the optical observation window are configured to clearly observe the crystal growth interface and the outline of the melting zone.

[0091] Specifically, the intelligent floating zone furnace equipment of this invention integrates intelligent technology into a traditional structure. The furnace body is equipped with an optimized optical observation window to ensure that the vision unit obtains distortion-free, high-contrast interface images. All actuators (heaters, lifting / rotating / feeding motors) are controlled by intelligent control system software within a central industrial control computer. The entire system achieves a seamless connection from physical growth to digital decision-making.

[0092] To make the objectives, technical solutions, and advantages of this invention clearer, the intelligent floating zone furnace equipment in this embodiment is described in a non-limiting manner below with reference to a preferred embodiment using lithium niobate crystal growth as an example. Those skilled in the art should understand that the embodiments described herein are for illustrative purposes only and are not intended to limit the invention.

[0093] In the intelligent growth control scenario of lithium niobate single crystals, the system hardware configuration includes: Floating zone furnace: employing radio frequency induction heating and equipped with a water cooling system. Vision unit: a 5-megapixel CMOS camera with a frame rate of 60fps, equipped with a narrow-band filter (center wavelength 532nm to reduce thermal radiation interference) and a telephoto lens. Coaxial arrangement of high-brightness LED cold light source. Sensors: an RF generator with a built-in power meter, a high-precision encoder monitoring the position and speed of three-axis (lifting, feeding, and rotating) servo motors, and an infrared thermometer monitoring the temperature around the molten zone.

[0094] The system synchronously executes the following steps: Data acquisition and synchronization: Develop host computer software to read process parameters through the OPC UA protocol, capture images through the Camera Link interface, and use hardware triggering to ensure data timestamp synchronization.

[0095] Morphology prediction model training: Data preparation: Collect data from 50 successful lithium niobate growth experiments. For the image sequence, use image algorithms to semi-automatically annotate the crystal diameter (pixels) and interface contour of each frame.

[0096] Model construction: The image branch uses a lightweight MobileNetV3, and the parameter branch uses a two-layer GRU. After fusion, it is fed into a two-layer fully connected network to output a diameter prediction sequence of one point per second for the next 30 seconds.

[0097] Training: The PyTorch framework was used, with MSE as the loss function and the Adam optimizer for training.

[0098] Implementation of Real-Time Process Control Model (MPC Based on Reinforcement Learning): Environment Definition: The "floating zone furnace + morphology prediction model" is used as the reinforcement learning environment. The state (st) consists of the historical process parameters for the past minute, image features from the past 10 frames (extracted by the visual encoder of the prediction model), and the current diameter prediction sequence. The action (at) is a continuous value of [power change ΔP, lifting speed change ΔV]. The reward (rt) = -w1*|D_pred - D_target| - w2*|ΔP| - w3*|ΔV|, where D_pred is the mean of the predicted diameter, and weights w1, w2, and w3 are used to balance tracking accuracy and control stability.

[0099] Training: First, the agent is trained using the SAC algorithm in a simplified simulation of lithium niobate growth based on physical rules until it can stably maintain the target diameter in the simulation. Then, the trained policy network is transferred to the actual system as the initial controller.

[0100] In practical implementation, the online control process is as follows: The operator sets the target diameter (e.g., 10mm) and starts growth. The system executes the following cyclical steps once per second: Data acquisition - Prediction model outputs future diameter sequence - Control model (SAC policy network) outputs ΔP and ΔV based on the current state and prediction sequence → Commands are sent to the actuator after being limited.

[0101] The control model performs model prediction based on the most recent data every 15 minutes (i.e., rolling optimization of MPC), fine-tuning its strategy to adapt to the specific circumstances of the current furnace.

[0102] Performance Verification: A comparative experiment was conducted with traditional PID control. Under slight disturbances in the density of the raw material, PID control resulted in significant diameter fluctuations (±0.5mm). In contrast, this intelligent control system, by anticipating the thinning trend, proactively increased the heating power, suppressing the diameter fluctuations to within ±0.1mm and achieving a flatter solid-liquid interface, significantly improving the optical uniformity of the crystal.

[0103] It should be noted that the above embodiments are merely preferred embodiments of the present invention and are not intended to limit the present invention. Those skilled in the art can make various modifications and variations to the system architecture, the specific network structure of the AI ​​model, the training algorithm, the implementation method of the control framework, etc., without departing from the spirit and scope of the present invention. For example, the morphological prediction model can use Vision Transformer instead of CNN, and the control model can also be solved online using a purely optimized MPC algorithm. All such modifications and variations should be covered within the protection scope of the present invention.

[0104] The floating zone crystal growth control method, system, and floating zone furnace equipment of this application have the same technical concept, and the technical details of the embodiments of the two are applicable to each other. In order to reduce repetition, they will not be repeated here.

[0105] Those skilled in the art will clearly understand that, for the sake of convenience and brevity, the above-described division of program modules is merely an example. In practical applications, the above functions can be assigned to different program modules as needed, that is, the internal structure of the device can be divided into different program units or modules to complete all or part of the functions described above. The program modules in the embodiments can be integrated into one processing unit, or each unit can exist physically separately, or two or more units can be integrated into one processing unit. The integrated unit can be implemented in hardware or as a software program unit. Furthermore, the specific names of the program modules are only for easy differentiation and are not intended to limit the scope of protection of this application.

Claims

1. A method for controlling crystal growth using the floating zone method, characterized in that, Includes the following steps: The process parameter data and crystal growth interface visual image data characterizing the molten zone morphology are collected in real time and synchronously. The process parameter data includes one or more of the following: heater power, crystal pulling speed, raw material feeding speed and crystal rotation speed. The process parameter data and the visual image data are used as multimodal inputs and input to a pre-trained crystal growth morphology prediction model. The crystal growth morphology prediction model extracts and fuses the features of the process parameter data and the visual image data, and outputs a quantitative prediction sequence of crystal growth morphology parameters in the future prediction time domain. Based on the quantized prediction sequence, the process real-time control model using the model predictive control framework is used to continuously optimize and solve the real-time control command within the current control cycle that makes the predicted output trajectory approximate the target growth trajectory, under the condition of satisfying the preset operation constraints. The real-time control command is sent to the actuator of the floating zone furnace to adjust the process parameter data and control the crystal growth process.

2. The floating zone method for controlling crystal growth as described in claim 1, characterized in that, The crystal growth morphology prediction model is a multimodal fusion neural network model, including: Feature extraction layer, including: The visual processing branch is used to extract image features reflecting the morphology of the melt zone and the characteristics of the solid-liquid interface from the visual image data. The parameter processing branch is used to extract timing features from the process parameter data; The feature fusion layer is used to deeply fuse the image features and the temporal features to obtain fused features; The output layer is used to output the quantized prediction sequence based on the fused features.

3. The floating zone method for controlling crystal growth as described in claim 2, characterized in that: The visual processing branch employs a convolutional neural network, including at least one of a 3D convolutional network, a temporal convolutional network, or a VisionTransformer to extract image features. The parameter processing branch uses a recurrent neural network, including at least one of a long short-term memory network or a gated recurrent unit to extract temporal features; The feature fusion layer uses feature cascading or attention mechanisms for feature fusion.

4. The floating zone method for controlling crystal growth as described in claim 1, characterized in that, The aforementioned method involves, under the condition of satisfying preset operational constraints, continuously optimizing and solving for real-time control commands within the current control cycle to make the predicted output trajectory approximate the target growth trajectory; specifically, it includes: Within each control cycle, the crystal growth morphology prediction model is used as an internal dynamic model to predict the system output behavior under different candidate control command sequences in the future finite time domain. A cost function is constructed to quantify the deviation between the predicted output trajectory and the target growth trajectory, as well as the magnitude of the change in the control command; Under the preset operational constraints, the optimal control command that minimizes the cost function is solved by the real-time process control model.

5. The floating zone method for controlling crystal growth as described in claim 4, characterized in that, The real-time process control model is a first control model based on a reinforcement learning algorithm, and the reward function of the first control model is negatively correlated with the cost function. Alternatively, the real-time process control model may be a second control model trained based on an adaptive optimization algorithm.

6. The floating zone method for controlling crystal growth as described in claim 5, characterized in that, The reward function is positively correlated with the crystal growth quality evaluation index; The reinforcement learning algorithm includes one of deep deterministic policy gradient, proximal policy optimization, or soft actor-critic algorithm.

7. A method for controlling crystal growth using the floating zone method as described in any one of claims 1-6, characterized in that, It also includes updating and optimizing the parameters of the crystal growth morphology prediction model and / or the real-time process control model based on the difference between the actual crystal growth results and the predicted results, or based on the quality evaluation results of the crystal after growth.

8. A floating zone crystal growth control system, characterized in that, include: The data acquisition module is used to collect process parameter data and visual image data of crystal growth interface characterizing the molten zone morphology in real time and synchronously. The process parameter data includes one or more of the following: heater power, crystal pulling speed, raw material feed speed, and crystal rotation speed; The data processing and prediction module is embedded with the crystal growth morphology prediction model as described in any one of claims 1-7, which is used to extract and fuse the features of the process parameter data and the visual image data, and output a quantitative prediction sequence of crystal growth morphology parameters in the future prediction time domain. The control decision module, which is embedded with the real-time process control model as described in any one of claims 1-7, is used to perform rolling optimization to solve the real-time control command that makes the predicted output trajectory approximate the target growth trajectory within the current control cycle based on the quantized prediction sequence and under the condition of satisfying preset operation constraints. The execution interface module is used to send the real-time control commands to the actuator of the floating zone furnace to adjust the process parameter data and control the crystal growth process.

9. An intelligent floating zone furnace device, characterized in that, The intelligent floating zone furnace equipment includes a furnace body, a heating device, a seed crystal rod, a raw material rod, and a motion mechanism; It also includes the floating zone crystal growth control system as described in claim 8; And an actuator that is communicatively connected to the execution interface module of the control system, the actuator including at least one of a heater power controller, a lifting motor driver, a feeding motor driver, and a rotary motor driver.

10. The intelligent floating zone furnace equipment as described in claim 9, characterized in that, The intelligent floating zone furnace equipment has an optical observation window on its furnace body, which is dedicated to the visual observation unit. The position and angle of the optical observation window are configured to clearly observe the crystal growth interface and the outline of the melting zone.

Images