A dg method crystal growth system and a crystal growth method

Abstract

The application discloses a DG method crystal growth system and a crystal growth method, and relates to the technical field of crystal growth. The system comprises a furnace body, an auxiliary heating mechanism, a rotating lifting mechanism and a liquid drop supply mechanism; a multi-modal perception unit for collecting image data of raw material drop formation, falling and growth interface morphology, acquiring two-dimensional temperature field distribution data of a growth interface region and crystal weight data; an intelligent control unit for generating instructions for regulating parameters of each actuator based on arbitration mechanism according to original data acquired by the multi-modal perception unit and issuing the instructions to each actuator. The system can detect the accurate three-dimensional morphology and volume of each raw material liquid drop in real time, and through the intelligent control unit, the priority and conflict of the regulation and control instructions can be judged in real time, and finally the rotating and lifting speed of the rotating lifting mechanism, the raw material liquid drop flow of the liquid drop supply mechanism and the heating parameters of the auxiliary heating mechanism are cooperatively regulated to respond to and compensate for the microscopic fluctuations caused by the falling of the raw material liquid drop.

Description

Technical Field

[0001] This invention relates to the field of crystal growth, and more particularly to a DG method crystal growth system and crystal growth method. Background Technology

[0002] The crucible lowering method is a traditional crystal growth method. However, during crystal growth, the crucible material (such as iridium or platinum) reacts with or dissolves in the melt at high temperatures, introducing impurities and reducing crystal purity. Furthermore, for high-melting-point crystals, crucible materials capable of withstanding their melting point are very expensive; for example, iridium crucibles are extremely costly, resulting in high equipment costs. Crucible-free growth technology fundamentally eliminates these problems, leading to the development of the DG (Drop-fed Growth) method, as detailed in invention patent CN 114000188 A. However, existing DG methods often employ fixed parameter settings or single-variable feedback (such as PID temperature control based on a single infrared source), resulting in coarse and lagging control that cannot detect and respond to microscopic disturbances caused by individual droplets. This leads to continuous fluctuations at the crystal growth interface and unstable crystal quality.

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

[0004] In view of the shortcomings of the prior art, the purpose of this invention is to provide a DG method crystal growth system and crystal growth method, which aims to solve the problem that the existing DG method growth device cannot sense and respond to the microscopic disturbances caused by a single droplet, resulting in continuous fluctuations in the crystal interface and unstable crystal quality.

[0005] The technical solution of the present invention is as follows: A first aspect of the present invention provides a DG method crystal growth system, comprising: Furnace body; The heating mechanism includes an auxiliary heating mechanism, which is disposed in the furnace body; A rotary lifting mechanism is located at the bottom of the furnace body and is used to place the seed crystal and enable the seed crystal to move up and down and rotate within the furnace body; A droplet supply mechanism, located at the top of the furnace body, is used to provide raw material droplets and allow the raw material droplets to fall onto the growth interface on the seed crystal surface. The multimodal sensing unit includes a visual sensing module, a thermal sensing module, and a weight sensing module. The visual sensing module is disposed outside the furnace body and is used to collect image data of the formation, falling, and growth interface morphology of the raw material droplets. The thermal sensing module is embedded in the auxiliary heating mechanism and is used to acquire two-dimensional temperature field distribution data of the growth interface region. The weight sensing module is integrated on the rotary lifting mechanism and is used to acquire weight data during the crystal growth process. The intelligent control unit is used to generate instructions for regulating the rotation and lifting speed of the rotary lifting mechanism, the raw material droplet flow rate of the droplet supply mechanism, and the heating parameters of the auxiliary heating mechanism based on the image data of raw material droplet formation, falling and growth interface morphology, two-dimensional temperature field distribution data of the growth interface region and weight data during crystal growth obtained by the multimodal sensing unit through an arbitration mechanism, and to send the instructions to the auxiliary heating mechanism, the rotary lifting mechanism and the droplet supply mechanism.

[0006] Optionally, the rotary lifting mechanism includes a seed crystal rod, a rotary motor for driving the seed crystal rod to rotate, and a lifting motor for driving the seed crystal rod to lift. The droplet supply mechanism includes an infusion tube, a flow valve located on the infusion tube, and a nozzle located at the end of the infusion tube; The heating mechanism also includes a main heating mechanism, which is disposed inside the furnace body; The auxiliary heating mechanism includes an induction heating coil, which is arranged around the outside of the growth interface.

[0007] Optionally, the visual perception module includes several cameras, and the fields of view of the several cameras are confocal on the path of the raw material droplets and the growth interface area.

[0008] Optionally, the thermal sensing module includes several thermocouples, with the temperature measuring ends of the thermocouples pointing towards the growth interface.

[0009] Optionally, the weight sensing module includes a weight sensor.

[0010] Optionally, the intelligent control unit includes: The data synchronization acquisition card is used to synchronously acquire raw data from the visual perception module, thermal perception module, and weight perception module. An industrial control computer is used to acquire and process the raw data collected by the data synchronization acquisition card. Based on the arbitration mechanism, it generates instructions for regulating the rotation and lifting speed of the rotary lifting mechanism, the raw material droplet flow rate of the droplet supply mechanism, and the heating parameters of the auxiliary heating mechanism. At the same time, it stores complete data for each crystal growth cycle. The programmable logic controller (PLC) is used to receive instructions generated by an industrial control computer and distribute them to the auxiliary heating mechanism, the rotary lifting mechanism, and the droplet supply mechanism.

[0011] Optionally, the industrial control computer includes: The computer vision processing platform is used to perform three-dimensional reconstruction of image data on the formation, falling and growth interface morphology of raw material droplets, and obtain the volume and morphology data of the raw material droplets. The multimodal data fusion center is used to acquire the volume and morphology data of the raw material droplets, the two-dimensional temperature field distribution data of the growth interface region, and the weight data during the crystal growth process, and aligns the timestamps of each data to generate a process state vector with confidence assessment. The process stability controller is used to acquire the process state vector generated from the multimodal data fusion center, and output fine-tuning instructions to compensate for single raw material droplet disturbances through feedforward-feedback composite calculation. The fine-tuning instructions include heating parameter fine-tuning instructions for the auxiliary heating mechanism, rotation speed fine-tuning instructions for the rotary elevator, and droplet flow rate fine-tuning instructions for the droplet supply mechanism. The mass balance controller is used to acquire the process state vector generated from the multimodal data fusion center and output the lifting speed adjustment command of the rotary lifting mechanism through an adaptive PID algorithm. The intelligent arbitrator receives instructions from the process stability controller and the quality balance controller, and, based on preset priority rules, combines the real-time confidence of the data used by each instruction to perform conflict resolution and instruction smoothing, outputting a safe collaborative control instruction set and sending it to the programmable controller.

[0012] Optionally, the industrial control computer also includes an autonomous optimization controller, which is used to acquire complete data for each crystal growth cycle, the control effects of the process stability controller and the quality balance controller, and output optimization strategy instructions based on a reward function mechanism; The intelligent arbitrator is used to receive instructions from the process stability controller, the quality balance controller, and the autonomous optimization controller, and, according to preset priority rules, combine the real-time confidence of the data used by each instruction to perform conflict resolution and instruction smoothing, output a safe collaborative control instruction set and send it to the programmable controller.

[0013] Optionally, the industrial control computer further includes: Digital twin models are used to predict disturbances caused by falling feed droplets based on complete data for each crystal growth cycle. The autonomous optimization controller is used to acquire complete data for each crystal growth cycle, the control effects of the process stability controller and the mass balance controller, and the prediction results of the digital twin model. Based on the reward function mechanism, it outputs optimization strategy instructions. The intelligent arbitrator is used to receive instructions from the process stability controller, the quality balance controller, and the autonomous optimization controller, and according to preset priority rules, combine the real-time confidence of the data used by each instruction to perform conflict resolution and instruction smoothing, output a safe collaborative control instruction set and send it to the programmable controller.

[0014] A second aspect of the present invention provides a crystal growth method, wherein, based on the DG method crystal growth system described above, the crystal growth method includes the following steps: The raw material is placed in the droplet supply mechanism, the seed crystal is placed on the rotary lifting mechanism, and the DG method crystal growth system is started to grow the crystal. The upper surface of the seed crystal melts to form a growth interface. The visual perception module collects image data of the formation, falling and growth interface morphology of the raw material droplet, the thermal perception module obtains two-dimensional temperature field distribution data of the growth interface area, and the weight perception module obtains weight data during the crystal growth process. Then, based on the arbitration mechanism, the intelligent control unit generates instructions to regulate the rotation and lifting speed of the rotary lifting mechanism, the raw material droplet flow rate of the droplet supply mechanism, and the heating parameters of the auxiliary heating mechanism, according to the image data of raw material droplet formation, falling and growth interface morphology obtained by the multimodal sensing unit, the two-dimensional temperature field distribution data of the growth interface region, and the weight data during the crystal growth process. The instructions are then sent to the auxiliary heating mechanism, the rotary lifting mechanism, and the droplet supply mechanism.

[0015] Beneficial Effects: The crystal growth system provided by this invention can detect the precise three-dimensional morphology and volume of each raw material droplet, the temperature field distribution at the growth interface, and the actual mass growth of the crystal in real time. Through the intelligent arbitration mechanism of the intelligent control unit, it can adjudicate the priority and conflict of control commands in real time, and finally coordinate the rotation and lifting speed of the rotating lifting mechanism, the raw material droplet flow rate of the droplet supply mechanism, and the heating parameters of the auxiliary heating mechanism. Therefore, the crystal growth system provided by this invention can sense, respond to, and compensate for the micro-fluctuations caused by the droplet falling, effectively solving the problem that existing crystal growth devices cannot sense and respond to the micro-disturbances caused by single droplets, resulting in continuous fluctuations at the crystal interface and unstable crystal quality. The crystal growth system provided by this invention can effectively improve the crystallization quality, electrical uniformity, and yield of single crystals. Attached Figure Description

[0016] Figure 1 This is a schematic diagram of the structure of the DG method crystal growth system.

[0017] Figure 2 This is a schematic diagram of the intelligent control principle of the DG method crystal growth system.

[0018] Figure 3 This is a flowchart of the logical decision-making process of an intelligent arbitrator.

[0019] Figure 4 This is a control hierarchy diagram of the DG method crystal growth system.

[0020] Appendix Figure 1 The labels in the text: 1. Furnace body; 11. Transparent viewing window; 2. Auxiliary heating mechanism; 3. Rotary lifting mechanism; 31. Seed crystal rod; 32. Rotary motor; 33. Lifting motor; 34. Multi-stage slip ring; 4. Seed crystal; 41. Growth interface; 5. Droplet supply mechanism; 51. Infusion pipe; 52. Flow valve; 53. Nozzle; 6. Multimodal sensing unit; 61. Visual sensing module; 611. High-speed industrial camera; 612. Narrow-band bandpass filter; 62. Thermal sensing module; 63. Weight sensing module. Detailed Implementation

[0021] This invention provides a DG method crystal growth system and method. To make the objectives, technical solutions, and effects of this invention clearer and more explicit, the invention is further described in detail below. It should be understood that the specific embodiments described herein are only for explaining the invention and are not intended to limit the invention.

[0022] Unless otherwise defined, all technical and scientific terms used herein have the same meaning as commonly understood by one of ordinary skill in the art to which this invention pertains. The terminology used in this specification is for the purpose of describing particular embodiments only and is not intended to be limiting of the invention.

[0023] The terms used in this document, such as “vertical,” “horizontal,” “up,” “down,” “left,” “right,” and similar expressions, are for illustrative purposes only and do not represent the only possible implementation.

[0024] In this invention, unless otherwise explicitly specified and limited, "above" or "below" the second feature can mean that the first feature is in direct contact with the second feature, or that the first feature is in indirect contact with the second feature through an intermediate medium. Furthermore, "above," "over," and "on top" of the second feature can mean that the first feature is directly above or diagonally above the second feature, or simply that the first feature is at a higher horizontal level than the second feature. "Below," "below," and "under" the second feature can mean that the first feature is directly below or diagonally below the second feature, or simply that the first feature is at a lower horizontal level than the second feature.

[0025] If the embodiments of the present invention involve descriptions such as "first" or "second", such descriptions are for descriptive purposes only and should not be construed as indicating or implying their relative importance or implicitly specifying the number of technical features indicated.

[0026] This invention provides a DG method crystal growth system (specifically applicable to the growth of crystals such as gallium oxide), wherein, as... Figure 1 As shown, the DG method crystal growth system includes: Furnace body 1; The heating mechanism includes an auxiliary heating mechanism 2, which is disposed in the furnace body 1 (the auxiliary heating mechanism surrounds the growth interface). A rotary lifting mechanism 3 is located at the bottom of the furnace body and is used to place the seed crystal 4 and enable the seed crystal to move up and down and rotate within the furnace body; The droplet supply mechanism 5 is located at the top of the furnace body and is used to provide raw material droplets and allow the raw material droplets to fall onto the growth interface 41 on the seed crystal surface. The multimodal sensing unit 6 includes a visual sensing module 61, a thermal sensing module 62, and a weight sensing module 63. The visual sensing module 61 is disposed outside the furnace body 1 and is used to collect image data of the formation, falling, and growth interface morphology of the raw material droplets. The thermal sensing module 62 is embedded in the auxiliary heating mechanism 2 and is used to acquire two-dimensional temperature field distribution data of the growth interface region. The weight sensing module 63 is integrated on the rotary lifting mechanism 3 and is used to acquire weight data (i.e., crystal weight data) during the crystal growth process. Intelligent control unit ( Figure 1 (Not shown in the diagram), installed outside the furnace body, is used to generate instructions for controlling the rotation and lifting speed of the rotary lifting mechanism, the raw material droplet flow rate of the droplet supply mechanism, and the heating parameters of the auxiliary heating mechanism based on the image data of raw material droplet formation, falling and growth interface morphology, two-dimensional temperature field distribution data of the growth interface region, and weight data during the crystal growth process obtained by the multimodal sensing unit through an arbitration mechanism. The instructions are then sent to the auxiliary heating mechanism, the rotary lifting mechanism, and the droplet supply mechanism.

[0027] The crystal growth system provided by this invention can detect the precise three-dimensional morphology and volume of each raw material droplet, the temperature field distribution at the growth interface, and the actual weight gain of the crystal in real time. Through the intelligent arbitration mechanism of the intelligent control unit, it can adjudicate the priority and conflict of control commands in real time, and finally coordinate the rotation and lifting speed of the rotating lifting mechanism, the raw material droplet flow rate of the droplet supply mechanism, and the heating parameters of the auxiliary heating mechanism. Therefore, the crystal growth system provided by this invention can sense, respond to, and compensate for the micro-fluctuations caused by the droplet falling, effectively solving the problem that existing crystal growth devices cannot sense and respond to the micro-disturbances caused by single droplets, resulting in continuous fluctuations at the crystal interface and unstable crystal quality. The crystal growth system provided by this invention can effectively improve the crystallization quality, electrical uniformity, and yield of single crystals.

[0028] In some implementations, such as Figure 1As shown, the rotary lifting mechanism 3 includes a seed crystal rod 31, a rotary motor 32 for driving the seed crystal rod to rotate, a lifting motor 33 for driving the seed crystal rod to move up and down, and servo drivers for the rotary motor and the lifting motor. The seed crystal rod is used to place and support the seed crystal. The rotary lifting mechanism 3 may also include multi-stage slip rings 34 to prevent wire breakage or signal interruption caused by wire entanglement during rotation.

[0029] In some implementations, such as Figure 1 As shown, the droplet supply mechanism 5 includes a delivery tube 51, a flow valve 52 located on the delivery tube, and a nozzle 53 located at the end of the delivery tube. The flow valve is used to regulate the flow rate of the raw material droplets. It is understood that the droplet supply mechanism may also include a storage mechanism for containing the raw material liquid.

[0030] In some embodiments, the heating mechanism further includes a main heating mechanism, wherein the main heating mechanism is in Figure 1 Not shown in the diagram, it is located in the furnace body and may be located on the periphery of the growth chamber, and is used to provide the main heating for the furnace body and crystal growth. Specifically, it may be a resistance heater.

[0031] The auxiliary heating mechanism includes an induction heating coil, which is arranged around the growth interface to provide local temperature compensation and uniform heating for the growth interface and the area near the nozzle. This prevents the raw material droplets from cooling down and solidifying before dripping, thus preventing them from clogging the nozzle. At the same time, it stabilizes the local thermal field and reduces the interface temperature disturbance caused by the raw material droplets.

[0032] In some embodiments, the visual perception module includes a plurality of cameras, the fields of view of which are confocal on the path of the raw material droplet and the growth interface region (the plurality of cameras are used to acquire images of droplet formation, droplet falling and growth interface morphology).

[0033] In some specific implementations, the plurality of cameras includes a plurality of high-speed industrial cameras, a plurality of structured light 3D scanning cameras, or a plurality of time-of-flight (ToF) cameras.

[0034] like Figure 1 As shown, the visual perception module 61 includes three high-speed industrial cameras 611, each located at a certain angle (e.g., at a trisection) along the circumference of the furnace body. The visual perception module 61 also includes narrow-band bandpass filters 612 positioned in front of the lenses of the three high-speed industrial cameras. A transparent viewing window 11 is provided on the side wall of the furnace body. The high-speed industrial cameras can acquire images of droplet formation, falling, and growth interface morphology through the transparent viewing window (which can be made of quartz), with a acquisition speed of up to 10,000 FPS. Alternatively, the transparent viewing window can be replaced with a positive-pressure clean gas curtain, a high-temperature resistant protective lens, and a water-cooling jacket.

[0035] In some embodiments, the thermal sensing module includes a plurality of thermocouples, the temperature measuring ends of which point towards the growth interface.

[0036] In some specific embodiments, the thermal sensing module includes three platinum-rhodium thermocouples (high temperature resistant, with an accuracy of ±0.1℃), which are located at three equal divisions in the circumferential direction of the induction heating coil, and are used to measure the two-dimensional temperature field distribution data of the crystal growth interface region.

[0037] In some embodiments, the weight sensing module includes a weight sensor (with an accuracy of ±0.01g) for real-time monitoring of quality data during crystal growth. Specifically, the weight sensor is integrated onto the seed crystal rod.

[0038] In some embodiments, the intelligent control unit includes: The data synchronization acquisition card is used to synchronously acquire raw data from the visual perception module, thermal perception module, and weight perception module. An industrial control computer is used to acquire and process the raw data collected by the data synchronization acquisition card. Based on the arbitration mechanism, it generates instructions for regulating the rotation and lifting speed of the rotary lifting mechanism, the raw material droplet flow rate of the droplet supply mechanism, and the heating parameters of the auxiliary heating mechanism. At the same time, it stores complete data for each crystal growth cycle (including but not limited to status data, instruction data, and result data). The programmable controller is used to receive instructions generated by an industrial control computer (and convert them into drive signals), and then send them to the auxiliary heating mechanism, the rotary lifting mechanism, and the droplet supply mechanism.

[0039] In this embodiment, the data synchronization acquisition card is triggered by a unified clock (IEEE 1588 protocol) source to achieve a unified and accurate time base. It synchronously acquires the raw data (i.e., image data of raw material droplet formation, falling and growth interface morphology, two-dimensional temperature field distribution data of the growth interface region, and weight data during crystal growth) from the visual perception module, thermal perception module, and weight perception module, and transmits them to the industrial control computer. The industrial control computer processes the raw data and generates control instructions based on the arbitration mechanism, which are then sent to the programmable controller. The programmable controller receives the control instructions generated by the industrial control computer (and converts them into drive signals), and then distributes them to the auxiliary heating mechanism, the rotary lifting mechanism, and the droplet supply mechanism.

[0040] The industrial control calculator provided in this invention can have three structures.

[0041] In some embodiments, the first industrial control computer includes: The computer vision processing platform is used to perform three-dimensional reconstruction of image data on the formation, fall, and growth interface morphology of raw material droplets, and obtain the volume and morphology data of the raw material droplets (the combination of the visual perception module and three-dimensional reconstruction can achieve sub-micro level precision measurement of the volume of irregular raw material droplets, overcome the error of assuming that the raw material droplets are regular spheres, and provide accurate data for the feedforward compensation of the L1 controller and the mass balance calculation of the L2 controller, thereby reducing control errors from the source). The multimodal data fusion center is used to acquire volume and morphology data of raw material droplets, two-dimensional temperature field distribution data of the growth interface region, and weight data during crystal growth. It also aligns the timestamps of each data point to generate a unified process state vector with confidence assessment. The process stability controller (L1 controller) is used to acquire the process state vector generated from the multimodal data fusion center, and output fine-tuning instructions for compensating for single raw material droplet disturbances through feedforward-feedback composite calculation. The fine-tuning instructions include fine-tuning instructions for the heating parameters of the auxiliary heating mechanism (e.g., the heating power of the induction coil), fine-tuning instructions for the rotation speed of the rotary elevator, and fine-tuning instructions for the droplet flow rate of the droplet supply mechanism (e.g., the opening degree of the flow valve of the droplet supply mechanism). The mass balance controller (L2 controller) is used to acquire the process state vector generated from the multimodal data fusion center and output the lifting speed adjustment command of the rotary lifting mechanism through the adaptive PID algorithm (which is a control algorithm that combines proportional, integral and derivative elements). The intelligent arbitrator receives instructions from the process stability controller and the quality balance controller, and performs conflict resolution and instruction smoothing based on preset priority rules (e.g., safety > quality > optimization), combined with the real-time confidence of the data used in generating each instruction (such as data acquired by the multimodal sensing unit, specifically the volume and morphology data of the raw material droplets, the two-dimensional temperature field distribution data of the growth interface region, and the weight data during the crystal growth process). It outputs a safe collaborative control instruction set and sends it to the programmable controller.

[0042] Because the volume of each droplet of raw material may vary, the fluctuations at the growth interface will also differ. Therefore, the process parameters related to crystal growth need to be adaptively fine-tuned according to the different interface fluctuations to ensure crystal quality. In this embodiment, the L1 controller, based on the process state vector generated by the multi-modal data fusion center (the L1 controller automatically acquires the droplet volume and morphology, and growth interface temperature data from the process state vector during calculation), performs feedforward-feedback composite calculations and outputs fine-tuning commands for the heating parameters (induction coil heating power) of the auxiliary heating mechanism, the rotation speed of the rotary elevator, and the droplet flow rate of the droplet supply mechanism to compensate for single raw material droplet disturbances. This effectively solves the impact of single raw material droplet disturbances.

[0043] In addition, the L2 controller can provide phased feedback based on the crystal growth rate, triggering pull-up speed correction through the material balance principle to achieve growth rate adaptation. Specifically, the L2 controller acquires the process state vector generated by the multimodal data fusion center, automatically obtains the weight data of the crystal growth process in the process state vector during calculation, compares the theoretical crystal weight calculated from the cumulative drop volume with the actual crystal weight measured by the weight sensor, and outputs the lifting speed adjustment command of the rotary lifting mechanism (i.e., pull-up speed) through an adaptive PID algorithm.

[0044] In some implementations, the second type of industrial control computer includes: The computer vision processing platform is used to perform three-dimensional reconstruction of image data on the formation, falling and growth interface morphology of raw material droplets, and obtain the volume and morphology data of the raw material droplets. The multimodal data fusion center is used to acquire volume and morphology data of raw material droplets, two-dimensional temperature field distribution data of the growth interface region, and weight data during crystal growth. It also aligns the timestamps of each data point to generate a unified process state vector with confidence assessment. The process stability controller (L1 controller) is used to acquire the process state vector generated from the multimodal data fusion center, and output fine-tuning instructions for compensating for single raw material droplet disturbances through feedforward-feedback composite calculation. The fine-tuning instructions include fine-tuning instructions for the heating parameters of the auxiliary heating mechanism (e.g., the heating power of the induction coil), fine-tuning instructions for the rotation speed of the rotary elevator, and fine-tuning instructions for the droplet flow rate of the droplet supply mechanism (e.g., the opening degree of the flow valve of the droplet supply mechanism). The mass balance controller (L2 controller) is used to acquire the process state vector generated from the multimodal data fusion center and output the lifting speed adjustment command of the rotary elevator through an adaptive PID algorithm; The autonomous optimization controller (L3 controller) is used to acquire complete data (process status history, stored in the industrial control calculator) for each crystal growth cycle, as well as the control effects of the process stability controller (L1 controller) and the quality balance controller (L2 controller). Based on the reward function mechanism, it outputs optimization strategy instructions. The intelligent arbitrator receives instructions from the process stability controller, quality balance controller, and autonomous optimization controller. Based on preset priority rules (e.g., safety > quality > optimization), it combines the real-time confidence of the data used in generating each instruction (such as data acquired by the multimodal sensing unit, specifically the volume and morphology data of the raw material droplets, the two-dimensional temperature field distribution data of the growth interface region, and the weight data during the crystal growth process) to perform conflict resolution and instruction smoothing. It then outputs a safe collaborative control instruction set and sends it to the programmable controller.

[0045] In this embodiment, the L1 controller, based on the process state vector generated by the multimodal data fusion center (the L1 controller automatically acquires the droplet volume and morphology, and growth interface temperature data in the process state vector during calculation), outputs fine-tuning commands for the heating parameters (induction coil heating power) of the auxiliary heating mechanism, the rotation speed of the rotary elevator, and the droplet flow rate of the droplet supply mechanism to compensate for single raw material droplet disturbances through feedforward-feedback composite calculations, effectively solving the impact of single raw material droplet disturbances.

[0046] In addition, the L2 controller can provide phased feedback based on the crystal growth rate, triggering pull-up speed correction through the material balance principle to achieve growth rate adaptation. Specifically, the L2 controller acquires the process state vector generated by the multimodal data fusion center, automatically obtains the weight data of the crystal growth process in the process state vector during calculation, compares the theoretical crystal weight calculated from the cumulative drop volume with the actual crystal weight measured by the weight sensor, and outputs the lifting speed adjustment command of the rotary lifting mechanism (i.e., pull-up speed) through an adaptive PID algorithm.

[0047] It is used to obtain complete data (process status history), control effect of process stability controller (L1 controller) and quality balance controller (L2 controller) for each crystal growth cycle, and outputs optimization strategy instructions based on reward function mechanism; Based on long-cycle process data and the control effects of the L1 and L2 controllers, the L3 controller achieves autonomous process optimization through a reward function.

[0048] In some implementations, the third type of industrial control computer includes: The computer vision processing platform is used to perform three-dimensional reconstruction of image data on the formation, falling and growth interface morphology of raw material droplets, and obtain the volume and morphology data of the raw material droplets. The multimodal data fusion center is used to acquire volume and morphology data of raw material droplets, two-dimensional temperature field distribution data of the growth interface region, and weight data during crystal growth. It also aligns the timestamps of each data point to generate a unified process state vector with confidence assessment. Storage unit, used to store complete data for each crystal growth cycle; The digital twin model is used to predict disturbances caused by falling raw material droplets based on complete data for each crystal growth cycle (such as disturbances to growth interface temperature, growth interface liquid level, internal crystal defects (such as dislocations, stress, inclusions and striations), crystal defect risk index, crystal growth rate, and crystal diameter). The digital twin model sends the prediction results to the L3 controller for feedforward control, making the system more stable and forward-looking. The process stability controller (L1 controller) is used to acquire the process state vector generated from the multimodal data fusion center, and output fine-tuning instructions for compensating for single raw material droplet disturbances through feedforward-feedback composite calculation. The fine-tuning instructions include fine-tuning instructions for the heating parameters of the auxiliary heating mechanism (e.g., the heating power of the induction coil), fine-tuning instructions for the rotation speed of the rotary elevator, and fine-tuning instructions for the droplet flow rate of the droplet supply mechanism (e.g., the opening degree of the flow valve of the droplet supply mechanism). The mass balance controller (L2 controller) is used to acquire the process state vector generated from the multimodal data fusion center and output the lifting speed adjustment command of the rotary elevator through an adaptive PID algorithm; The autonomous optimization controller (L3 controller) is used to acquire complete data for each crystal growth cycle (process status history), the control effects of the process stability controller (L1 controller) and the quality balance controller (L2 controller), as well as the prediction results of the digital twin model. Based on the reward function mechanism, it outputs optimization strategy instructions. The intelligent arbitrator receives instructions from the process stability controller, quality balance controller, and autonomous optimization controller. Based on a preset priority rule (priority rule: safety > quality > optimization), and combined with the real-time confidence level of the data used in generating each instruction (such as data acquired by the multimodal sensing unit, specifically the volume and morphology data of the raw material droplets, the two-dimensional temperature field distribution data of the growth interface region, and the weight data during the crystal growth process), it performs conflict resolution and instruction smoothing, outputs a safe collaborative control instruction set, and sends it to the programmable controller.

[0049] In this embodiment, the digital twin model can be constructed using the following approach: COMSOL / ANSYS (mechanism core) + Python / PyTorch (AI algorithm) + Siemens platform (system integration).

[0050] In this embodiment, the L1 controller, based on the process state vector generated by the multimodal data fusion center (the L1 controller automatically acquires the droplet volume and morphology, and growth interface temperature data in the process state vector during calculation), outputs fine-tuning commands for the heating parameters (induction coil heating power) of the auxiliary heating mechanism, the rotation speed of the rotary elevator, and the droplet flow rate of the droplet supply mechanism to compensate for single raw material droplet disturbances through feedforward-feedback composite calculations, effectively solving the impact of single raw material droplet disturbances.

[0051] The L2 controller provides phased feedback based on the crystal growth rate, triggering pull-up speed correction through material balance principles to achieve growth rate adaptation. The L2 controller acquires a process state vector generated from a multimodal data fusion center, compares the theoretical crystal growth weight calculated from the cumulative drop volume with the actual crystal weight measured by a weight sensor, obtains the mass balance error, and outputs an adjustment command for the lifting speed of the rotary elevator (i.e., the crystal pull-up speed) through an adaptive PID algorithm, ensuring a dynamic balance between mass input and growth output. Additionally, a laser rangefinder can be added to the DG method crystal growth system to measure the crystal diameter in real time, serving as a parallel or verification input for the L2 controller.

[0052] The theoretical crystal growth weight (or mass) flow rate is: _theory(t) = ρ_melt × (dV_droplet / dt) in: _theory(t): The theoretical crystal growth gravimetric flow rate at time t (g / s or g / h); ρ_melt: Melt density of seed crystal (e.g., gallium oxide) (g / cm³) 3 Material constants; dV_droplet / dt: The derivative of the feed droplet volume with respect to time, i.e., the real-time droplet volumetric flow rate (cm). 3 / s); The actual crystal weight (or mass) growth rate is: _actual(t) = ρ_crystal × A_crystal × V_pull(t) in: _actual(t): The actual weight growth rate of the crystal at time t (g / s or g / h); ρ_crystal: Density of the crystal (e.g., gallium oxide crystal) (g / cm³) 3 (), slightly higher than the melt density; A_crystal: Current cross-sectional area of ​​the crystal (cm²) 2 The diameter is determined by the target diameter. V_pull(t): The crystal pulling speed (cm / s or mm / h) at time t, which is the direct output of the L2 controller; The mass balance error (data used in the L2 controller PID calculation) is: e(t) =

[0053] Where e(t): the accumulated mass error (g) at time t; τ: Integral time variable.

[0054] The L3 controller is specifically a deep reinforcement learning (DRL) agent that makes master process optimization decisions based on a reward function. Its goal is to maximize the long-term cumulative reward score by adjusting various parameters (actions) throughout the crystal growth process. The L3 controller receives complete data for each crystal growth cycle (process state history), the control effects of the L1 / L2 controllers, and prediction results from a digital twin model. Its internal value network evaluates the long-term benefits of the current state, while the policy network generates optimization policy instructions aimed at maximizing the reward function. Additionally, the aforementioned DG crystal growth system can be equipped with acoustic emission sensors or high-temperature cameras to monitor potential defects (such as cracks and inclusions) during crystal growth, generating signals that serve as strong negative feedback in the reward function or as an emergency interruption signal for the intelligent arbitrator. Specifically, the reward function is designed as the sum of positive rewards for factors beneficial to crystal quality (such as temperature stability and quality balance) and negative penalties for unfavorable factors (such as predicting defect risk), as follows: R t = (w T σ T ) (w M |e M ∣) (w D D t )+(w S S t ) (w P P t ) in, R t The immediate reward obtained at time t (or within a control period); σ T : A measure of the fluctuation in the temperature field at the growth interface (from filtered thermocouple data); e M : Mass balance error; D t Defect risk index predicted by digital twin model; S t Rewards for stable process operation; P t Energy consumption penalty; w T w M w D w S w P : These are the weighting coefficients for each item.

[0055] Reward function R t It is constructed as a weighted sum of multiple process quality indicators, including: a penalty term for interface temperature fluctuations. σ T Penalty for mass balance error |e M | and the penalty term for the defect risk index predicted by the digital twin model. D t The goal of the L3 controller is to maximize the cumulative reward ∑R throughout the growth process through policy optimization. t This enables global autonomous optimization of crystal quality, stability, and efficiency.

[0056] Taking a camera as the visual perception module, a thermocouple as the thermal perception module, and a weight sensor as the weight perception module as an example, the intelligent control principle in this embodiment is as follows: Figure 2As shown. First, the data acquisition card synchronously acquires raw data from the camera, thermocouple, and weight sensor (image data of raw material droplet formation, falling, and growth interface morphology; two-dimensional temperature field distribution data of the crystal growth interface region; and weight data during crystal growth, respectively). The computer vision processing platform performs three-dimensional reconstruction on the image data of raw material droplet formation, falling, and interface morphology to obtain the volume and morphology data of the raw material droplets. The multimodal data fusion center acquires the volume and morphology data of the raw material droplets, the two-dimensional temperature field distribution data of the growth interface region, and the weight data during crystal growth, and aligns the timestamps of each data point to generate a unified process state vector with confidence assessment (i.e., IEEE). The 1588 protocol ensures a unified hardware clock source and its confidence level, packaging them into unified time frames to generate a unified process state vector with confidence assessment. Then, the L1, L2, and L3 controllers acquire the process state vector. The L1 controller performs feedforward-feedback composite calculations based on droplet volume and morphology, and interface temperature, quickly outputting fine-tuning commands for heating parameters (such as induction coil heating power) of the auxiliary heating mechanism, the rotation speed (crystal lifting speed) of the rotary elevator, and the droplet flow rate (such as valve opening) of the droplet supply mechanism to compensate for single droplet disturbances. The L2 controller compares the theoretical crystal clock weight calculated from the cumulative droplet volume with the actual crystal weight measured by the weight sensor to obtain the mass balance error. It then outputs adjustment commands for the lifting speed of the rotary elevator (i.e., the crystal lifting speed) through an adaptive PID algorithm, ensuring a dynamic balance between mass input and growth output. The L3 controller, at its core a DRL agent, receives complete data for each crystal growth cycle (process state history), the control effects of the L1 / L2 controllers, and prediction results from the digital twin model. Its internal value network evaluates the long-term benefits of the current state, while the policy network generates optimized policy instructions aimed at maximizing the reward function (see the formula above).

[0057] Then, the intelligent arbitrator receives all instructions from the L1, L2, and L3 controllers, and according to the preset priority rules (i.e., safety > quality > optimization), it combines the real-time confidence of the data used in generating each instruction (i.e., process parameters, such as data obtained by the multimodal sensing unit, specifically the volume and morphology data of the raw material droplets, the two-dimensional temperature field distribution data of the growth interface region, and the weight data during the crystal growth process) to perform conflict resolution and instruction smoothing, and outputs the final, safe collaborative control instruction set. This set includes instructions for regulating the valve opening of the droplet supply mechanism (controlling the flow rate of the raw material droplets), the induction coil heating power supply (controlling the heating power of the induction coil), and the motor of the rotary lifting mechanism (controlling the lifting speed and rotation speed of the rotary lifting mechanism), thus completing the crystal growth process. Complete data of the crystal growth process (including state data, instruction data, and result data) is stored in the industrial control computer. Based on the complete data of the crystal growth process, the digital twin model predicts the disturbances (disturbances on the growth interface temperature and crystal stress distribution) caused by the falling raw material droplets. It can calculate in advance the disturbances of the raw material droplets on the interface temperature and crystal stress distribution, providing a basis for the feedforward control of the L3 controller.

[0058] like Figure 3 As shown, the intelligent arbitrator includes a security detection module and a confidence module. The logical decision-making process of the intelligent arbitrator is as follows: S11. Receive instructions from the L1 controller, L2 controller and L3 controller, denoted as U_L1, U_L2 and U_L3 respectively.

[0059] S12. The safety detection module detects whether there are any anomalies, that is, whether U_L1, U_L2 and U_L3 exceed the safety threshold (i.e. the safety boundary).

[0060] S13. If yes, then reject the over-limit instruction and enable the safety preset value; if no, then use the confidence assessment module to evaluate the credibility of each instruction source data (i.e., the data used to generate each instruction). This step is the confidence assessment node. The confidence assessment module performs validity verification and confidence scoring on the volume and morphology data of the raw material droplets, the two-dimensional temperature field distribution data of the growth interface region, and the weight data during the crystal growth process. If the data confidence is lower than the threshold, then the "data failure" flag is triggered.

[0061] S14. Then, perform conflict detection on the U_L1, U_L2 and U_L3 instructions, and perform rule arbitration (make a decision according to the preset rules).

[0062] This step involves command conflict nodes and arbitration decision nodes. For example, when checking command conflicts, if command U_L1 requests "increase heating power" to maintain temperature, while command U_L3 requests "decrease heating power" to optimize thermal field uniformity, it is marked as "conflict".

[0063] When making arbitration decisions, a ruling is made according to preset rules. For example, if a "data failure" flag is triggered, a conservative control strategy based on high-confidence sensors is prioritized, and instructions that rely on failed data are ignored. If there are no data issues but instruction conflicts exist, the corresponding instructions are executed according to the default priority of "L1 (maintain instantaneous stability) > L2 (maintain macroscopic quality) > L3 (long-term optimization)," or a more complex weighted voting mechanism is initiated. If there are neither data issues nor instruction conflicts, the instruction set is output directly.

[0064] S15, Output the final coordinated control instruction set (sent to each actuator).

[0065] In summary, such as Figure 4 As shown, the system provided by this invention is mainly divided into three layers. The top layer is the data processing and intelligent decision-making layer, which mainly includes an industrial control computer (running the core intelligent control algorithm) and a data synchronization acquisition card, and the two communicate with each other via Ethernet. The middle layer is the real-time control and security logic layer, which mainly includes a programmable controller. The bottom layer is the output and interface interaction layer, which mainly includes output interfaces.

[0066] At the top layer, the data synchronization acquisition card synchronously acquires the raw data obtained by the multimodal sensing unit. Then, the industrial control computer acquires and processes the raw data, and outputs a secure collaborative control instruction set through the L1, L2 and L3 controllers, the digital twin model and the intelligent arbitrator, and sends it to the programmable controller in the middle layer.

[0067] In the middle layer, the programmable controller receives the cooperative control instruction set, converts it into drive signals, executes the underlying safety interlock logic, handles high-speed sequential control, and communicates directly with the execution structure.

[0068] At the bottom layer, the output interface layer outputs drive signals to each actuator (such as auxiliary heating mechanism, rotary lifting mechanism and droplet supply mechanism).

[0069] This invention also provides a crystal growth method, wherein, based on the DG method crystal growth system described above, the crystal growth method includes the following steps: The raw material is placed in the droplet supply mechanism, the seed crystal is placed on the rotary lifting mechanism, the DG method crystal growth system is turned on to grow the crystal, the auxiliary heating mechanism melts the upper surface of the seed crystal to form a growth interface, the visual perception module collects image data of the formation, falling of the raw material droplet and the morphology of the growth interface; the thermal perception module obtains two-dimensional temperature field distribution data of the growth interface area; the weight perception module obtains weight data during the crystal growth process. Then, based on the arbitration mechanism, the intelligent control unit generates instructions to regulate the rotation and lifting speed of the rotary lifting mechanism, the raw material droplet flow rate of the droplet supply mechanism, and the heating parameters (e.g., the heating power of the induction heating coil) of the auxiliary heating mechanism, according to the image data of raw material droplet formation, falling and growth interface morphology obtained by the multimodal sensing unit, the two-dimensional temperature field distribution data of the growth interface region, and the weight data during the crystal growth process. The instructions are then sent to the auxiliary heating mechanism, the rotary lifting mechanism, and the droplet supply mechanism.

[0070] This invention also provides a crystal growth method, wherein, based on the DG method crystal growth system described above, the crystal growth method includes the following steps: The raw material is placed in the droplet supply mechanism, the seed crystal is placed on the seed crystal rod of the rotary lifting mechanism, and the DG method crystal growth system is turned on to grow the crystal. The upper surface of the seed crystal melts to form a growth interface. The camera collects image data of the formation, falling and growth interface morphology of the raw material droplet; the thermocouple acquires (or measures) the two-dimensional temperature field distribution data of the growth interface region; and the weight sensor acquires (or measures) the weight data during the crystal growth process. Then, the data synchronization acquisition card is triggered by a unified clock (IEEE 1588 protocol) source to simultaneously acquire image data of raw material droplet formation, falling and growth interface morphology, two-dimensional temperature field distribution data of the growth interface region, and weight data during crystal growth. The computer vision processing platform performs three-dimensional reconstruction on the image data of raw material droplet formation, falling and interface morphology to obtain the volume and morphology data of the raw material droplets; The multimodal data fusion center acquires the volume and morphology data of the raw material droplets, the two-dimensional temperature field distribution data of the growth interface region, and the weight data during the crystal growth process, and aligns the timestamps of each data to generate a process state vector with confidence assessment. The L1 controller acquires the process state vector generated from the multimodal data fusion center and outputs fine-tuning instructions to compensate for single raw material droplet disturbances through feedforward-feedback composite calculation. The fine-tuning instructions include fine-tuning instructions for the heating power of the induction coil, fine-tuning instructions for the rotation speed of the rotary elevator, and fine-tuning instructions for the valve opening of the droplet supply mechanism. The L2 controller acquires the process state vector generated from the multimodal data fusion center and outputs the lifting speed adjustment command of the rotary elevator (i.e., crystal pulling speed) through the adaptive PID algorithm; The L3 controller acquires complete data for each crystal growth cycle and the control effects of the L1 and L2 controllers, and outputs optimization strategy instructions based on the reward function mechanism. The intelligent arbitrator receives instructions from the L1, L2, and L3 controllers and, based on preset priority rules ("safety > quality > optimization"), combines the real-time confidence level of the data used to generate each instruction to perform conflict resolution and instruction smoothing. It outputs a safe set of collaborative control instructions (such as valve opening fine-tuning instructions, heating power fine-tuning instructions, rotation speed fine-tuning instructions, and lifting speed adjustment instructions) and sends them to the programmable controller. The programmable controller receives control instructions generated by the industrial control computer (and converts them into drive signals) and sends them to the induction heating coil, the rotary lifting mechanism, and the droplet supply mechanism.

[0071] This invention also provides a crystal growth method, wherein, based on the DG method crystal growth system described above, the crystal growth method includes the following steps: The raw material is placed in the droplet supply mechanism, the seed crystal is placed on the seed crystal rod of the rotary lifting mechanism, the DG method crystal growth system is turned on to grow the crystal, the induction heating coil melts the upper surface of the seed crystal to form a growth interface, the camera collects image data of the formation, falling and growth interface morphology of the raw material droplet; the thermocouple acquires (or measures) the two-dimensional temperature field distribution data of the growth interface region; the weight sensor acquires (or measures) the weight data during the crystal growth process. Then, the data synchronization acquisition card is triggered by a unified clock (IEEE 1588 protocol) source to synchronously acquire image data of raw material droplet formation, falling and growth interface morphology, two-dimensional temperature field distribution data of the growth interface region, and weight data during crystal growth. The computer vision processing platform performs three-dimensional reconstruction on the image data of raw material droplet formation, falling and interface morphology to obtain the volume and morphology data of the raw material droplets; The multimodal data fusion center acquires the volume and morphology data of the raw material droplets, the two-dimensional temperature field distribution data of the growth interface region, and the weight data during the crystal growth process, and aligns the timestamps of each data to generate a process state vector with confidence assessment. The digital twin model predicts the disturbances (disturbances on growth interface temperature and crystal stress distribution) caused by falling raw material droplets based on complete data of each crystal growth cycle (stored in a storage unit). The L1 controller acquires the process state vector generated from the multimodal data fusion center and outputs fine-tuning instructions to compensate for single raw material droplet disturbances through feedforward-feedback composite calculation. The fine-tuning instructions include fine-tuning instructions for the heating power of the induction coil, fine-tuning instructions for the rotation speed of the rotary elevator, and fine-tuning instructions for the valve opening of the droplet supply mechanism. The L2 controller acquires the process state vector generated from the multimodal data fusion center and outputs the lifting speed adjustment command of the rotary elevator (i.e., crystal pulling speed) through the adaptive PID algorithm; The L3 controller acquires complete data for each crystal growth cycle, the control effects of the L1 and L2 controllers, and the prediction results of the digital twin model. Based on the reward function mechanism, it outputs optimization strategy instructions. The intelligent arbitrator receives instructions from the L1, L2, and L3 controllers and, based on preset priority rules (priority rule: safety > quality > optimization), combines the real-time confidence of the data used to generate each instruction to perform conflict resolution and instruction smoothing, outputting a safe set of collaborative control instructions (such as valve opening fine-tuning instructions, heating power fine-tuning instructions, rotation speed fine-tuning instructions, and lifting speed adjustment instructions) and sending them to the programmable controller; the programmable controller receives control instructions generated by the industrial control computer (and converts them into drive signals), and sends them to the induction heating coil, the rotary lifting mechanism, and the droplet supply mechanism.

[0072] In summary, this invention constructs an intelligent closed loop of "perception-decision-execution-evolution," realizing full-process adaptive closed-loop control driven by a single drop event. This allows for the systematic and proactive suppression of growth disturbances, and the system can adaptively smooth the dynamic disturbances brought by each drop of raw material, achieving stability of the growth interface. Furthermore, through autonomous learning and safety arbitration, the process is continuously optimized during long-cycle growth. Ultimately, while significantly reducing the cost of precious metals (an inherent advantage of the DG method), it simultaneously achieves a significant improvement in crystal quality (low dislocation density, high uniformity) and growth yield, resolving the core contradiction of the DG method's difficulty in simultaneously achieving "cost reduction" and "quality improvement." Specifically, the system uses a multimodal sensing unit to detect the precise three-dimensional morphology and volume of each raw material droplet, the temperature field at the growth interface, and the actual mass growth of the crystal in real time. Based on this, a multi-level control architecture consisting of L1 / L2 / L3 controllers and an intelligent arbitrator enables the system to achieve a balance between control robustness and intelligence. A forward-looking reward function drives a deep reinforcement learning agent to make autonomous process optimization decisions, and a top-level intelligent arbitration mechanism is designed to adjudicate the priority and conflicts of multi-level control commands in real time, ultimately coordinating the control of actuators such as dropleting, heating, lifting, and rotation. The intelligent arbitration mechanism endows the system with human-like decision-making capabilities, prioritizing stability when in doubt. When the data acquired by the multimodal sensing unit is abnormal, the system automatically degrades to a conservative mode, avoiding exacerbating the problem with erroneous data, greatly improving the reliability and safety of the system during long-term, uninterrupted crystal growth, and reducing the risk of ingot scrap due to control misjudgments. Ultimately, this allows the system to cope with high-frequency disturbances, perform long-term optimization, and never lose control.

[0073] It should be understood that the application of the present invention is not limited to the examples above. Those skilled in the art can make improvements or modifications based on the above description, and all such improvements and modifications should fall within the protection scope of the appended claims.

Claims

1. A DG method crystal growth system, characterized in that, include: Furnace body; The heating mechanism includes an auxiliary heating mechanism, which is disposed in the furnace body; A rotary lifting mechanism is located at the bottom of the furnace body and is used to place the seed crystal and enable the seed crystal to move up and down and rotate within the furnace body; A droplet supply mechanism, located at the top of the furnace body, provides raw material droplets and allows them to fall onto the growth interface of the seed crystal. A multimodal sensing unit includes a visual sensing module, a thermal sensing module, and a weight sensing module. The visual sensing module is located outside the furnace body and collects image data on the formation, fall, and morphology of the raw material droplets and the growth interface. The thermal sensing module is embedded in the auxiliary heating mechanism and acquires two-dimensional temperature field distribution data of the growth interface region. The weight sensing module is integrated into the rotary lifting mechanism and acquires weight data during crystal growth. The intelligent control unit is used to generate instructions for regulating the rotation and lifting speed of the rotary lifting mechanism, the raw material droplet flow rate of the droplet supply mechanism, and the heating parameters of the auxiliary heating mechanism based on the image data of raw material droplet formation, falling and growth interface morphology, two-dimensional temperature field distribution data of the growth interface region and weight data during crystal growth obtained by the multimodal sensing unit through an arbitration mechanism, and to send the instructions to the auxiliary heating mechanism, the rotary lifting mechanism and the droplet supply mechanism.

2. The DG method crystal growth system according to claim 1, characterized in that, The rotary lifting mechanism includes a seed crystal rod, a rotary motor for driving the seed crystal rod to rotate, and a lifting motor for driving the seed crystal rod to lift. The droplet supply mechanism includes an infusion tube, a flow valve located on the infusion tube, and a nozzle located at the end of the infusion tube; The heating mechanism also includes a main heating mechanism, which is disposed inside the furnace body; The auxiliary heating mechanism includes an induction heating coil, which is arranged around the growth interface.

3. The DG method crystal growth system according to claim 1, characterized in that, The visual perception module includes several cameras, and the fields of view of the several cameras are confocal on the path of the raw material droplets falling and the growth interface area.

4. The DG method crystal growth system according to claim 1, characterized in that, The thermal sensing module includes several thermocouples, with the temperature measuring ends of the thermocouples pointing towards the growth interface.

5. The DG method crystal growth system according to claim 1, characterized in that, The weight sensing module includes a weight sensor.

6. The DG method crystal growth system according to claim 1, characterized in that, The intelligent control unit includes: The data synchronization acquisition card is used to synchronously acquire raw data from the visual perception module, thermal perception module, and weight perception module. An industrial control computer is used to acquire and process the raw data collected by the data synchronization acquisition card. Based on the arbitration mechanism, it generates instructions for regulating the rotation and lifting speed of the rotary lifting mechanism, the raw material droplet flow rate of the droplet supply mechanism, and the heating parameters of the auxiliary heating mechanism. At the same time, it stores complete data for each crystal growth cycle. The programmable logic controller (PLC) is used to receive instructions generated by an industrial control computer and distribute them to the auxiliary heating mechanism, the rotary lifting mechanism, and the droplet supply mechanism.

7. The DG method crystal growth system according to claim 6, characterized in that, The industrial control computer includes: The computer vision processing platform is used to perform three-dimensional reconstruction of image data on the formation, falling and growth interface morphology of raw material droplets, and obtain the volume and morphology data of the raw material droplets. The multimodal data fusion center is used to acquire the volume and morphology data of the raw material droplets, the two-dimensional temperature field distribution data of the growth interface region, and the weight data during the crystal growth process, and aligns the timestamps of each data to generate a process state vector with confidence assessment. The process stability controller is used to acquire the process state vector generated from the multimodal data fusion center, and output fine-tuning instructions to compensate for single raw material droplet disturbances through feedforward-feedback composite calculation. The fine-tuning instructions include heating parameter fine-tuning instructions for the auxiliary heating mechanism, rotation speed fine-tuning instructions for the rotary elevator, and droplet flow rate fine-tuning instructions for the droplet supply mechanism. The mass balance controller is used to acquire the process state vector generated from the multimodal data fusion center and output the lifting speed adjustment command of the rotary lifting mechanism through an adaptive PID algorithm. The intelligent arbitrator receives instructions from the process stability controller and the quality balance controller, and, based on preset priority rules, combines the real-time confidence of the data used by each instruction to perform conflict resolution and instruction smoothing, outputting a safe collaborative control instruction set and sending it to the programmable controller.

8. The DG method crystal growth system according to claim 7, characterized in that, The industrial control computer also includes an autonomous optimization controller, which is used to acquire complete data for each crystal growth cycle, the control effects of the process stability controller and the quality balance controller, and output optimization strategy instructions based on the reward function mechanism. The intelligent arbitrator is used to receive instructions from the process stability controller, the quality balance controller, and the autonomous optimization controller, and, according to preset priority rules, combine the real-time confidence of the data used by each instruction to perform conflict resolution and instruction smoothing, output a safe collaborative control instruction set and send it to the programmable controller.

9. The DG method crystal growth system according to claim 7, characterized in that, The industrial control computer also includes: Digital twin models are used to predict disturbances caused by falling feed droplets based on complete data for each crystal growth cycle. The autonomous optimization controller is used to acquire complete data for each crystal growth cycle, the control effects of the process stability controller and the mass balance controller, and the prediction results of the digital twin model. Based on the reward function mechanism, it outputs optimization strategy instructions. The intelligent arbitrator is used to receive instructions from the process stability controller, the quality balance controller, and the autonomous optimization controller, and according to preset priority rules, combine the real-time confidence of the data used by each instruction to perform conflict resolution and instruction smoothing, output a safe collaborative control instruction set and send it to the programmable controller.

10. A crystal growth method, characterized in that, Based on the DG method crystal growth system according to any one of claims 1-9, the crystal growth method includes the following steps: The raw material is placed in the droplet supply mechanism, the seed crystal is placed on the rotary lifting mechanism, and the DG method crystal growth system is started to grow the crystal. The upper surface of the seed crystal melts to form a growth interface. The visual perception module collects image data of the formation, falling and growth interface morphology of the raw material droplet, the thermal perception module obtains two-dimensional temperature field distribution data of the growth interface area, and the weight perception module obtains weight data during the crystal growth process. Then, based on the arbitration mechanism, the intelligent control unit generates instructions to regulate the rotation and lifting speed of the rotary lifting mechanism, the raw material droplet flow rate of the droplet supply mechanism, and the heating parameters of the auxiliary heating mechanism, according to the image data of raw material droplet formation, falling and growth interface morphology obtained by the multimodal sensing unit, the two-dimensional temperature field distribution data of the growth interface region, and the weight data during the crystal growth process. The instructions are then sent to the auxiliary heating mechanism, the rotary lifting mechanism, and the droplet supply mechanism.

Images