A method, apparatus, equipment, and medium for growing lithium niobate crystals based on machine vision and sequence prediction.

By combining machine vision with sequence prediction, the parameter control problem in the shoulder formation stage of lithium niobate crystal growth was solved, achieving precise control of lithium niobate crystal growth and improving crystal yield and uniformity.

CN123108722APending Publication Date: 2026-11-17HANGZHOU FUJIA GALLIUM TECH CO LTD
View PDF 0 Cites 0 Cited by

Patent Information

Application Number
CN202611205918.9
Authority / Receiving Office
CN · China
Patent Type
Applications(China)
Current Assignee / Owner
Filing Date
2026-08-10
Publication Date
2026-11-17

AI Technical Summary

Technical Problem

In existing lithium niobate crystal growth processes, parameter control during the shoulder formation stage relies on manual experience, which can easily lead to excessively large or distorted shoulder angles, resulting in thermal stress cracking and polycrystalline regions, affecting crystal yield. Furthermore, insufficient precision in composition control can lead to photorefractive defects.

Method used

A method combining machine vision and sequence prediction is adopted. The melt surface map is acquired by a vision acquisition device, the visual feature data is identified by a machine vision model, and the operating time series data is constructed by combining the data of the growth furnace sensor. The growth prediction model is used to capture long-distance parameter correlation features, determine the control parameters, and achieve precise control of lithium niobate crystal growth.

Benefits of technology

Direct visual closed-loop control of the shoulder formation process was achieved, which reduced the probability of shoulder deformation and thermal stress cracking, improved the growth yield and uniformity of lithium niobate crystals, and enhanced prediction accuracy and control stability.

✦ Generated by Eureka AI based on patent content.

Smart Images

  • Figure CN123108722A_ABST
    Figure CN123108722A_ABST
Patent Text Reader

Abstract

This application discloses a method, apparatus, equipment, and medium for growing lithium niobate crystals based on machine vision and sequence prediction. The method includes acquiring a melt surface image using a vision acquisition device installed on a growth furnace; identifying the melt surface image using a machine vision model to obtain visual feature data, the visual feature data including at least the shoulder expansion angle; determining current operating condition data based on the visual feature data and sensor data from the growth furnace, and constructing operating condition time-series data based on the current operating condition data and historical operating condition data of a preset duration; capturing long-distance parameter correlation features using a growth prediction model based on the operating condition time-series data, and determining control parameters based on the parameter correlation features; and controlling the growth of the lithium niobate crystal based on the control parameters. This application embodiment acquires a melt surface image using a vision acquisition device installed on a growth furnace, and then identifies the melt surface image to perform real-time visual monitoring of features such as the shoulder expansion angle. This achieves direct visual closed-loop control of the shoulder formation process, effectively avoiding deviations caused by human experience judgment, reducing the probability of abnormal situations such as shoulder deformities and thermal stress cracking, and improving the growth yield of lithium niobate crystals.
Need to check novelty before this filing date? Find Prior Art

Description

Technical Field

[0001] This application relates to the field of crystal preparation technology, and in particular to a method, apparatus, equipment and medium for growing lithium niobate crystals based on machine vision and sequence prediction. Background Technology

[0002] Lithium niobate ( Abbreviation Crystals, due to their excellent electro-optic, acousto-optic, nonlinear optical, and piezoelectric properties, are used in surface acoustic waves (SAW). Lithium niobate crystals are widely used in filters, electro-optic modulators, nonlinear optical devices, and integrated optical chips. It can be classified into lithium niobate with the same composition according to the proportion of its components ( , ≈48.6%) and near-stoichiometric lithium niobate ( , ≈50%). Currently, industrial production is based on... The crystal is the main component, and its growth process includes the following stages in sequence: heating and material preparation, preheating of seed crystal, inoculation, necking (neck diameter 8-12mm), shoulder formation (expansion angle about 45°), constant diameter growth (diameter 50-100mm), and cooling and annealing.

[0003] The expansion of lithium niobate crystals from the narrow neck (8-12 mm) to the shoulder formation stage (50-100 mm) places high demands on the cooling rate and the pulling speed. However, currently, these parameters are generally controlled manually based on experience, which can easily lead to excessively large shoulder angles (causing thermal stress cracking) or deformed shoulders (forming polycrystalline regions), thus affecting the crystal yield.

[0004] Therefore, existing technologies still need to be improved and enhanced. Summary of the Invention

[0005] The technical problem to be solved by this application is to provide a method, apparatus, equipment and medium for growing lithium niobate crystals based on machine vision and sequence prediction, in order to address the shortcomings of the prior art.

[0006] To address the aforementioned technical problems, the first aspect of this application provides a method for growing lithium niobate crystals based on machine vision and sequence prediction, wherein the method includes: The melt surface image is acquired using a vision acquisition device installed on the growth furnace; The melt surface diagram is identified by a machine vision model to obtain visual feature data, which includes at least one or more of the following: shoulder expansion angle, neck diameter, equal diameter deviation, and melt convection mode. The current operating condition data is determined based on the visual feature data and the sensor data of the growth furnace, and the operating condition time series data is constructed based on the current operating condition data and historical operating condition data of a preset duration. Based on the aforementioned operating condition time series data, a growth prediction model is used to capture long-distance parameter correlation characteristics, and control parameters are determined based on these parameter correlation characteristics. The growth of the lithium niobate crystal is controlled based on the control parameters.

[0007] The lithium niobate crystal growth method based on machine vision and sequence prediction is described in which the vision acquisition device is deployed outside the observation window on the side wall of the growth furnace, and its viewing angle covers the inner cavity of the growth furnace, so that the acquired melt surface image includes the melt surface and the crystal contact interface.

[0008] The lithium niobate crystal growth method based on machine vision and sequence prediction, wherein the step of identifying the melt surface diagram using a machine vision model to obtain lithium niobate crystal growth parameters specifically includes: The melt surface image is input into the machine vision model, and the crystal boundary and melt surface in the melt surface image are detected by the vision detection module in the machine vision model to determine the candidate boundary of the crystal profile. The candidate boundary is used to locate the region of interest in the melt surface map. The feature extraction module in the machine vision model is used to extract the sub-pixel level edge of the region of interest, and visual feature data is determined based on the sub-pixel level edge.

[0009] The lithium niobate crystal growth method based on machine vision and sequence prediction includes sensor data such as crystal weight, rotation speed, lifting speed, and temperature data. The crystal weight includes the weight collected by a weighing sensor deployed on the top of the lifting rod of the growth furnace, and the temperature data includes the temperature of each independent heating zone collected by a temperature controller deployed around the crucible of the growth furnace.

[0010] The lithium niobate crystal growth method based on machine vision and sequence prediction, wherein the step of capturing long-range parameter correlation features using the growth prediction model based on the operating condition time-series data, and determining control parameters based on the parameter correlation features specifically includes: The operating condition time series data is input into the Transformer encoder in the growth prediction model, and the Transformer encoder uses a multi-head self-attention mechanism to capture long-distance parameter correlation features. The parameter correlation features are input into the GRU decoder in the growth prediction model, and the control parameters for a future preset duration are predicted by the GRU decoder.

[0011] The lithium niobate crystal growth method based on machine vision and sequence prediction includes control parameters such as the speed adjustment of the pulling servo motor, the speed adjustment of the crucible rotation servo motor, and the power distribution matrix of the multi-temperature zone heater.

[0012] The lithium niobate crystal growth method based on machine vision and sequence prediction further includes: Historical data on the growth process of multiple lithium niobate crystals are aggregated through a cloud server. The Transformer encoder in the growth prediction model is fine-tuned using the historical data so that the side device used to perform the lithium niobate crystal growth method based on machine vision and sequence prediction can determine the control parameters using the fine-tuned growth prediction model. During the fine-tuning process, the attention weights of the Transformer encoder are optimized using a contrastive learning approach.

[0013] A second aspect of this application provides a lithium niobate crystal growth apparatus based on machine vision and sequence prediction, wherein the apparatus includes: The acquisition module is used to acquire melt surface images via a vision acquisition device installed on the growth furnace; The vision module is used to identify the melt surface pattern through a machine vision model to obtain visual feature data; The construction module is used to determine the current operating condition data based on the visual feature data and the sensor data of the growth furnace, and to construct the operating condition time series data based on the current operating condition data and historical operating condition data of a preset duration. The control parameter determination module is used to capture long-distance parameter correlation features based on the operating condition time series data using a growth prediction model, and determine control parameters based on the parameter correlation features. A control module is used to control the growth of the lithium niobate crystal based on the control parameters.

[0014] A third aspect of this application provides a computer-readable storage medium storing one or more programs that can be executed by one or more processors to implement the steps in the lithium niobate crystal growth method based on machine vision and sequence prediction as described above.

[0015] A fourth aspect of this application provides a terminal device, which includes: a processor and a memory; The memory stores a computer-readable program that can be executed by the processor; When the processor executes the computer-readable program, it implements the steps in the lithium niobate crystal growth method based on machine vision and sequence prediction as described above.

[0016] Beneficial effects: 1. This application acquires a melt surface image by installing a visual acquisition device on the growth furnace, and then identifies the melt surface image to perform real-time visual monitoring of the shoulder expansion angle, etc., thereby realizing direct visual closed-loop control of the shoulder formation process. This can effectively avoid the deviation caused by human experience judgment, reduce the probability of abnormal situations such as shoulder deformation and thermal stress cracking, and improve the growth yield of lithium niobate crystals.

[0017] 2. This application constructs working condition time series data by fusing real-time visual feature data and sensor data, and then uses the multi-head self-attention mechanism in the growth prediction model to capture long-distance parameter correlation features, thereby realizing the advanced prediction and active compensation of component drift in the later stage of growth. It can more accurately predict the control parameters required in future periods. Compared with the traditional control method that relies on single-point data feedback, the prediction accuracy and control stability are significantly improved.

[0018] 3. By predicting the power distribution matrix of the multi-temperature zone heater, this application can more accurately match and control the temperature of each heating zone, realize the active reshaping of the thermodynamic field in the melt, reduce the negative impact of melt convection inhomogeneity on crystal growth, and further improve the uniformity and quality stability of the grown lithium niobate crystal.

[0019] 4. This application supports fine-tuning the model by aggregating multiple batches of growth data in the cloud and optimizing attention weights by combining comparative learning. It can continuously iterate and optimize model performance in actual production, adapt to the working characteristics of different growth furnaces, and meet the application needs of large-scale industrial production. Attached Figure Description

[0020] To more clearly illustrate the technical solutions in the embodiments of this application, the accompanying drawings used in the description of the embodiments will be briefly introduced below. Obviously, the accompanying drawings described below are only some embodiments of this application. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0021] Figure 1 A flowchart of a lithium niobate crystal growth method based on machine vision and sequence prediction provided for embodiments of this application.

[0022] Figure 2 This is an architecture diagram of the lithium niobate crystal growth method based on machine vision and sequence prediction provided in the embodiments of this application.

[0023] Figure 3 This is a flowchart of the one-click crystal growth data closed loop.

[0024] Figure 4A schematic diagram of the lithium niobate crystal growth apparatus based on machine vision and sequence prediction provided in the embodiments of this application.

[0025] Figure 5 A schematic block diagram of the terminal device provided in the embodiments of this application. Detailed Implementation

[0026] This application provides a method, apparatus, device, and medium for growing lithium niobate crystals based on machine vision and sequence prediction. To make the objectives, technical solutions, and effects of this application clearer and more explicit, the following detailed description, with reference to the accompanying drawings and embodiments, further illustrates this application. It should be understood that the specific embodiments described herein are merely illustrative and not intended to limit the scope of this application.

[0027] Those skilled in the art will understand that, unless specifically stated otherwise, the singular forms “a,” “an,” “the,” and “the” used herein may also include the plural forms. It should be further understood that the term “comprising” as used in this application means the presence of the stated features, integers, steps, operations, elements, and / or components, but does not exclude the presence or addition of one or more other features, integers, steps, operations, elements, components, and / or groups thereof. It should be understood that when we say an element is “connected” or “coupled” to another element, it can be directly connected or coupled to the other element, or there may be intermediate elements. Furthermore, “connected” or “coupled” as used herein can include wireless connections or wireless coupling. The term “and / or” as used herein includes all or any units and all combinations of one or more associated listed items.

[0028] It will be understood by those skilled in the art that, unless otherwise defined, all terms used herein (including technical and scientific terms) have the same meaning as commonly understood by one of ordinary skill in the art to which this application pertains. It should also be understood that terms such as those defined in general dictionaries should be understood to have the same meaning as in the context of the prior art, and should not be interpreted in an idealized or overly formal sense unless specifically defined as herein.

[0029] It should be understood that the sequence number and size of each step in this embodiment do not imply the order of execution. The execution order of each process is determined by its function and internal logic, and should not constitute any limitation on the implementation process of this application embodiment.

[0030] Research has found that lithium niobate ( Abbreviation Crystals, due to their excellent electro-optic, acousto-optic, nonlinear optical, and piezoelectric properties, are used in surface acoustic waves (SAW). Lithium niobate crystals are widely used in filters, electro-optic modulators, nonlinear optical devices, and integrated optical chips. It can be classified into lithium niobate with the same composition according to the proportion of its components ( , ≈48.6%) and near-stoichiometric lithium niobate ( , ≈50%). Currently, industrial production is based on... The crystal is the main component, and its growth process includes the following stages in sequence: heating and material preparation, preheating of seed crystal, inoculation, necking (neck diameter 8-12mm), shoulder formation (expansion angle about 45°), constant diameter growth (diameter 50-100mm), and cooling and annealing.

[0031] The expansion of lithium niobate crystals from the narrow neck (8-12 mm) to the shoulder formation stage (50-100 mm) places high demands on the cooling rate and the pulling speed. However, currently, these parameters are generally controlled manually based on experience, which can easily lead to excessively large shoulder angles (causing thermal stress cracking) or deformed shoulders (forming polycrystalline regions), thus affecting the crystal yield.

[0032] In addition, due to It belongs to a lithium-deficient structure, and minute changes in the solid-liquid interface morphology during its growth process can lead to changes in the axial and radial directions of the crystal. The drift in solid-liquid interface leads to photorefractive defects, severely impacting device performance. Existing methods rely solely on weighing sensors to predict solid-liquid interface morphology, which can result in insufficient precision in component control. Furthermore, The melt contains a complex coupling of thermal buoyancy flow and rotating centrifugal flow. Sudden changes in the convection mode can lead to drastic changes in the solid-liquid interface morphology, which in turn can cause fluctuations in crystal diameter and affect crystal yield.

[0033] To address the aforementioned issues, in this embodiment, a visual acquisition device installed on the growth furnace acquires a melt surface image; a machine vision model is used to identify the melt surface image to obtain visual feature data, which includes at least the shoulder expansion angle; based on the visual feature data and sensor data from the growth furnace, current operating condition data is determined, and operating condition time-series data is constructed based on the current operating condition data and historical operating condition data of a preset duration; based on the operating condition time-series data, a growth prediction model is used to capture long-distance parameter correlation features, and control parameters are determined based on the parameter correlation features; the growth of the lithium niobate crystal is controlled based on the control parameters. This embodiment acquires a melt surface image using a visual acquisition device installed on the growth furnace, and then identifies the melt surface image to perform real-time visual monitoring of features such as the shoulder expansion angle. This achieves direct visual closed-loop control of the shoulder formation process, effectively avoiding deviations caused by human experience judgment, reducing the probability of abnormal situations such as shoulder deformities and thermal stress cracking, and improving the growth yield of lithium niobate crystals. Meanwhile, by integrating real-time visual feature data and sensor data to construct operating condition time series data, and then using the multi-head self-attention mechanism in the growth prediction model to capture long-distance parameter correlation features, advanced prediction and active compensation for component drift in the later stage of growth are achieved. This enables more accurate prediction of the control parameters required in future periods. Compared with the traditional control method that relies solely on single-point data feedback, both prediction accuracy and control stability are significantly improved.

[0034] It should be noted that the execution entity of each embodiment of this application can be a computing service system with data processing, network communication, and program execution functions, such as an electronic system capable of realizing the above functions, a lithium niobate crystal growth system based on machine vision and sequence prediction, etc. The following uses a lithium niobate crystal growth system based on machine vision and sequence prediction (hereinafter referred to as "the system") as an example to describe the following embodiments.

[0035] The application content will be further explained below with reference to the accompanying drawings and through the description of the embodiments.

[0036] This embodiment provides a method for growing lithium niobate crystals based on machine vision and sequence prediction, such as... Figure 1 and Figure 2 As shown, the lithium niobate crystal growth method based on machine vision and sequence prediction specifically includes: S10. Acquire a melt surface image using a vision acquisition device installed on the growth furnace.

[0037] Specifically, lithium niobate crystal growth does not require a liquid seal layer, and the melt surface is exposed, allowing for direct visual observation to obtain the melt surface image. To this end, a visual acquisition device can be installed on the side wall of the growth furnace. This device can acquire images of the melt surface through an observation window on the furnace side wall. The visual acquisition device can be an industrial camera (such as one with a resolution higher than 2048×2048 pixels and a frame rate higher than 10fps). This industrial camera is equipped with a high-temperature resistant sapphire protective window and a light-shielding filter to avoid interference from furnace heat radiation and stray light on image clarity, ensuring that the acquired melt surface image clearly shows the crystal outline and the boundary between the crystal and the melt surface.

[0038] Therefore, such as Figure 2 As shown, a visual acquisition device is deployed outside the observation window on the side wall of the growth furnace. The viewing angle of this device covers the inner cavity of the growth furnace (i.e., the viewing angle covers the melt surface and the crystal contact interface), so that the melt surface image acquired by the visual acquisition device includes both the melt surface and the crystal contact interface. The melt surface is the free surface of the lithium niobate melt, and the crystal contact interface is the boundary between the growing lithium niobate crystal and the melt surface. The acquired melt surface image clearly distinguishes the crystal region from the melt region, and can directly obtain the candidate boundary of the crystal outline at the current shoulder formation stage.

[0039] S20. The melt surface map is identified using a machine vision model to obtain visual feature data.

[0040] Specifically, the visual feature data is obtained by visually recognizing the melt surface diagram and is used to reflect the geometric morphological characteristics of crystal growth during the shoulder formation stage. The visual feature data can include geometric and thermodynamic apparent parameters affecting growth, such as the shoulder expansion angle, neck diameter, constant diameter deviation, and melt convection pattern. Furthermore, the visual feature data can be determined based on the growth stage of the lithium niobate crystal. In other words, the machine vision model can identify the shoulder expansion angle, neck diameter, constant diameter deviation, and melt convection pattern, and then determine the visual feature data to be output based on the growth stage of the lithium niobate crystal. For example, in the constant diameter growth stage, the visual feature data can include the constant diameter deviation and the melt convection pattern.

[0041] Among them, the measurement accuracy of the neck diameter is ±0.1mm, ensuring that the neck diameter is within the optimal dislocation range of 8-12mm; The shoulder expansion angle is used to calculate the expansion angle of the crystal shoulder in real time, with a target angle of approximately 45°. The equal diameter deviation is used to monitor diameter fluctuations during the equal diameter stage, triggering an alarm threshold of ±0.5mm; The melt convection mode is identified by analyzing the thermal striations pattern on the melt surface, which can be radial flow-dominated, circumferential flow-dominated, etc.

[0042] A machine vision model is used to identify visual feature data from acquired melt surface images. The input to this machine vision model is the melt surface image, and the output is the visual feature data. Specifically, the acquired melt surface image is input into the trained machine vision model, which then outputs visual feature data. This machine vision model can include a visual detection module and a feature extraction module. The visual detection module detects candidate boundaries of the crystal contour, and the feature extraction module extracts visual feature data based on the detected candidate boundaries.

[0043] Based on this, in one embodiment, the step of identifying the melt surface map using a machine vision model to obtain the lithium niobate crystal growth parameters specifically includes: The melt surface image is input into the machine vision model, and the crystal boundary and melt surface in the melt surface image are detected by the vision detection module in the machine vision model to determine the candidate boundary of the crystal profile. The candidate boundary is used to locate the region of interest in the melt surface map. The feature extraction module in the machine vision model is used to extract the sub-pixel level edge of the region of interest, and visual feature data is determined based on the sub-pixel level edge.

[0044] Specifically, the visual detection module is used to perform target detection to determine the crystal boundaries and melt surface positions in the melt surface image. It can be implemented using a pre-trained instance segmentation neural network model, which can accurately distinguish between crystal regions and melt regions to determine candidate boundaries for the crystal contour. In this embodiment, the visual detection module uses YOLOv8, which can quickly locate crystal boundaries (i.e., crystal region boundaries) and melt surfaces (i.e., solid-liquid interfaces) in the melt surface image to obtain bounding boxes for candidate crystal contours. The processing speed of this visual detection module is >30fps.

[0045] The feature extraction module is used to locate the region of interest (ROI) based on the candidate boundaries output by the visual detection module. Based on this ROI, it extracts the sub-pixel-level edges of the crystal contour. Then, based on the sub-pixel-level edges, it determines visual feature parameters such as the shoulder extension angle, neck diameter, and diameter deviation. Simultaneously, it identifies the melt convection pattern based on the thermal striations pattern of the melt surface region. In this embodiment, the feature extraction module uses a ResNet network. YOLOv8 and ResNet are jointly trained using a preset dataset to obtain a trained machine vision model. This allows YOLOv8 to extract candidate boundaries, and ResNet to extract the sub-pixel-level edges of the ROI. These sub-pixel-level edges are used to complete micrometer-level size measurements. Specifically, the ResNet model can identify the true boundary position between the crystal edge and the melt surface (with an accuracy higher than 1 pixel) to obtain a continuous, high-precision contour curve (i.e., sub-pixel-level edges).

[0046] In this embodiment, the pre-trained instance segmentation network YOLOv8 is first used to coarsely locate the crystal region and the melt region, and output the candidate boundary of the crystal contour. Then, the ResNet network is used to extract the region of interest within the candidate boundary in a refined manner to obtain the sub-pixel level high-precision edge curve, thereby obtaining accurate visual feature data. This not only takes into account the recognition efficiency and recognition accuracy, but also meets the frame rate and accuracy requirements of real-time monitoring during the crystal growth process.

[0047] S30. Determine the current operating condition data based on the visual feature data and the sensor data of the growth furnace, and construct the operating condition time series data based on the current operating condition data and historical operating condition data of a preset duration.

[0048] Specifically, the operating condition time-series data includes a sequence of operating condition data from multiple consecutive acquisition times, arranged chronologically. A preset duration is a fixed time window, and the operating condition data corresponding to all acquisition times within this fixed duration window are considered historical operating condition data. All historical operating condition data corresponding to this preset duration and the current operating condition data, arranged chronologically by acquisition time, form the operating condition time-series data. This data reflects the current growth state of the lithium niobate crystal and its recent trends, providing a complete dynamic basis for subsequent adjustments to growth parameters.

[0049] The operating data at each acquisition moment includes visual feature data and growth furnace sensor data at that acquisition moment. The sensor data may include crystal weight, rotation speed, lifting and moving speed, and temperature data. The crystal weight includes the weight collected by the weighing sensor deployed on the top of the lifting rod of the growth furnace, and the temperature data includes the temperature of each independent heating zone collected by the temperature controller deployed around the crucible of the growth furnace.

[0050] To address this, in addition to installing a visual acquisition device (such as a high-resolution industrial camera) on the existing growth furnace, this embodiment also equips the growth furnace with an upper weighing sensor and a multi-zone temperature controller. The upper weighing sensor, mounted on the top of the lifting rod, is used to accurately measure the crystal weight in real time (accuracy ±0.1g) to calculate the instantaneous growth rate and crystal density change. The multi-zone temperature controller can independently control the heating power of heating zones at different axial positions around the crucible, thereby accurately acquiring temperature data for each heating zone. For example, if three independent heating zones (upper, middle, and lower) are configured around the crucible, controlling the axial temperature gradient of the melt (target gradient 3-5℃ / mm) respectively, the multi-zone temperature controller can independently control these three independent heating zones and acquire their temperature data. Furthermore, the rotation speed and lifting speed can be directly acquired by the encoder integrated into the lifting rod, allowing for real-time recording of the crystal's current rotation speed and lifting speed.

[0051] This embodiment integrates visual feature data collected at the same time with all sensor data to obtain the current operating condition data at that moment. Then, it arranges the current operating condition data and previously stored historical operating condition data within a preset time period in chronological order to construct operating condition time-series data that reflects the changing pattern of crystal growth state over time. Simultaneously, this embodiment sets multiple independent heating zones along the axial direction around the crucible, collects temperature data for each independent heating zone, and subsequently determines the heating power of each zone. This actively reshapes the thermohydrodynamic field within the melt, thereby precisely controlling the temperature gradient and solid-liquid interface shape during crystal growth, providing a control basis for obtaining high-quality crystals.

[0052] S40. Based on the operating condition time series data, use a growth prediction model to capture long-distance parameter correlation features, and determine control parameters based on the parameter correlation features.

[0053] Specifically, parameter correlation features are used to reflect the coupling influence between different growth parameters and the dynamic correlation between these parameters over time. Crystal growth is a strongly coupled, nonlinear, thermophysical process with large inertial time delays. Complex interactions exist between various control and state parameters. For example, the crystal quality at the current moment depends not only on the current temperature and pulling speed but also on the cumulative influence of process parameters from the past several hours or even tens of hours. Therefore, long-range parameter correlation features can reflect this causal relationship across time scales, thereby capturing the lagging effects of earlier parameter changes on the current and subsequent crystal growth states, and avoiding prediction biases caused by data fluctuations within short time windows.

[0054] The growth prediction model can adopt a Transformer structure to capture long-range parameter correlation features to determine control parameters, or it can adopt a Transformer and GRU structure, where the Transformer captures long-range parameter correlation features and the GRU decodes control parameters based on the parameter correlation features.

[0055] In one embodiment, the step of capturing long-range parameter correlation features using the growth prediction model based on the operating condition time-series data, and determining control parameters based on the parameter correlation features, specifically includes: The operating condition time series data is input into the Transformer encoder in the growth prediction model, and the Transformer encoder uses a multi-head self-attention mechanism to capture long-distance parameter correlation features. The parameter correlation features are input into the GRU decoder in the growth prediction model, and the control parameters for a future preset duration are predicted by the GRU decoder.

[0056] Specifically, the Transformer encoder is used to embed features into the operating condition data at each time step in the input operating condition time series data. Then, it calculates the self-attention weights between parameters at different positions through a multi-head self-attention mechanism, thereby highlighting the influence of key parameters at earlier times on the current growth state. This fully explores the parameter coupling relationship across time dimensions and extracts long-distance parameter correlation features that can reflect long-term change patterns.

[0057] Among these, long-range parameter correlation characteristics can include the correlation between the shoulder formation angle in the early stage of growth and the compositional uniformity in the later stage, the correlation between the abrupt change in pulling speed in the early stage of shoulder formation and the compositional drift in the later stage of constant diameter, and / or the correlation between thermal stripe fluctuations and small-angle grain boundary formation. The correlation between the abrupt change in pulling speed in the early stage of shoulder formation and the compositional drift in the later stage of constant diameter means that if the pulling speed is excessively increased in the early stage of shoulder formation (e.g., in the second hour) to accelerate shoulder formation, it will lead to... Excessive consumption of elements can cause a ratio imbalance, and this slight imbalance will appear in the later stages of constant diameter growth (such as at the 40th hour), manifesting as non-uniform refractive index of the crystal. The Transformer can capture the strong correlation between "abnormal pulling speed at the 2nd hour" and "decreased optical uniformity at the 40th hour", thus performing temperature compensation in advance.

[0058] The correlation between thermal stripe fluctuations and small-angle grain boundary formation is manifested in the following hours when high-frequency thermal stripe fluctuations (characterizing strong thermal convection) continuously appear on the melt surface for more than 30 minutes, making it highly likely that small-angle grain boundaries will form in the crystal. Therefore, by capturing long-range parameter correlation characteristics, convection can be suppressed and the defect formation chain can be interrupted immediately by adjusting the crucible rotation speed when thermal stripe anomalies first appear.

[0059] Furthermore, after capturing the parameter correlation features, these features are input into the GRU decoder. The GRU decoder then outputs control parameters for a future preset time period (e.g., 5 minutes) based on these features. These control parameters may include the speed adjustment of the lifting servo motor (target lifting speed 0.5-2.0 mm / h), the speed adjustment of the crucible rotation servo motor (target rotation speed 8-20 rpm), and the power distribution matrix of the multi-temperature zone heaters (the power ratio of each axially deployed independent heating zone).

[0060] This application adopts a growth prediction model that combines Transformer and GRU. By leveraging the advantages of Transformer in capturing long-distance temporal dependencies and the efficient decoding capabilities of GRU, it can fully explore the hysteretic influence of historical process parameters on the current crystal growth state. This breaks through the limitations of traditional control methods that rely solely on data at the current moment for adjustment, and achieves accurate parameter prediction based on global growth process information, effectively addressing the large inertial time delay characteristics of the crystal growth process.

[0061] S50. Control the growth of the lithium niobate crystal based on the control parameters.

[0062] Specifically, based on the control parameters output by the growth prediction model, adjustment commands are issued to each actuator of the growth furnace: a speed adjustment command is issued to the lifting servo motor to adjust the lifting speed to the target speed; a speed adjustment command is issued to the crucible rotation servo motor to adjust the crucible speed to the target speed; and a power adjustment command is issued to the multi-temperature zone temperature controller to adjust the heating power of each independent heating zone according to the power distribution matrix. This completes the dynamic adjustment of the melt temperature gradient and the solid-liquid interface shape, offsets the hysteresis effect caused by historical parameter fluctuations, suppresses the generation of crystal defects, and ensures that the lithium niobate crystal is always in a stable growth state.

[0063] In this embodiment, after controlling the growth of the lithium niobate crystal based on the aforementioned control parameters, the step of acquiring the melt surface image through a vision acquisition device installed on the growth furnace is re-executed to form a closed-loop control cycle. This continuously acquires the latest growth status data and updates the operating timeline data, iteratively adjusting the control parameters to achieve dynamic and precise control over the entire lithium niobate crystal growth process. Compared to traditional control methods that rely on manual experience to periodically adjust parameters, this closed-loop control method offers more timely response and higher control precision, effectively reducing errors caused by manual operation and improving the overall quality and yield of the entire crystal.

[0064] Specifically, this closed-loop control process can perform a round of state acquisition, visual feature data acquisition, control parameter prediction, and growth control at fixed acquisition intervals (such as 1 minute), continuously and dynamically correcting the crystal growth process. This ensures that the growth parameters at each stage always match the needs of the current actual growth state, and that parameter adjustments do not lag behind actual state changes due to large inertial time delay characteristics. This fundamentally reduces problems such as crystal defects and uneven composition caused by untimely parameter control.

[0065] For example, let's illustrate a control process using temperature compensation during the constant diameter growth stage. In this example, the control process includes: Status acquisition (T=0ms): The industrial camera acquires the current solid-liquid interface image, the upper weighing sensor acquires the current weight signal, and the temperature controller transmits the temperature data of each independent heating zone.

[0066] Visual feature data acquisition (T=5ms): The visual feature data identified by the machine vision model on the edge device is that the diameter deviation is 0.2mm and the melt convection mode is weakened circumferential flow.

[0067] Control parameter prediction (T=35ms): The industrial controller predicts the parameter correlation characteristics within the next 10 minutes based on visual feature data and sensor data, and the Transformer encoder combines historical operating data from the past 120 minutes. The GRU decoder determines the required control parameters based on the parameter correlation characteristics: the pulling speed needs to be reduced by 0.1mm / h, the heating power in the middle zone needs to be reduced by 0.5kW, and the crucible rotation speed needs to be increased by 2rpm to enhance circumferential flow.

[0068] Instruction issuance in growth control (T=50ms): The industrial control computer converts the above control parameters into low-level control instructions and sends them directly to the PLC via the EtherCAT bus.

[0069] Hardware execution in growth control (T=51ms): The PLC drives the servo motor and thyristor. Based on the existing PID closed-loop control, the control parameters are superimposed on the PID setpoint as feedforward compensation signals to achieve smooth adjustment and avoid overshoot and oscillation of traditional simple PID control.

[0070] In one embodiment, to improve computational efficiency and meet the real-time control requirements of the crystal growth process, the method of this application is implemented based on a cloud-edge collaborative architecture. For example... Figure 3 As shown, specifically: The edge computing node is deployed with machine vision models; specifically, YOLOv8 and ResNet models are deployed on the edge computing node after being quantized by TensorRT (FP16 precision). This edge computing node is directly connected to an industrial camera via the GigE Vision protocol, realizing the integration of image acquisition and vision model deployment and inference, with an end-to-end latency of <30ms.

[0071] The industrial control unit deploys a growth prediction model, specifically a Transformer+GRU structured growth prediction model running as a Python service. This industrial control unit communicates directly with the lifting servo drive and heater controller via a high-speed EtherCAT (IEC 61158) real-time industrial bus, with a control cycle of 1ms, ensuring real-time control commands. The industrial control unit connects to edge computing nodes via a local area network (LAN), employing ZeroMQ message queues to achieve low-latency data transmission (latency <5ms) between the visual model output and the growth model input.

[0072] The cloud platform (acting as a training server) is responsible for collecting historical data on the lithium niobate crystal growth process. Based on this historical data, the Transformer encoder in the growth prediction model is fine-tuned so that the side-side equipment executing the machine vision and sequence prediction-based lithium niobate crystal growth method can use the fine-tuned growth prediction model to determine control parameters. The cloud platform pushes model update packages to the industrial control computer via the MQTT protocol, and the system automatically performs a hot-swap of the growth prediction model before the next batch of growth begins.

[0073] Based on this, in one embodiment, the method further includes: Historical data on the growth process of multiple lithium niobate crystals are aggregated through a cloud server. The Transformer encoder in the growth prediction model is fine-tuned using the historical data so that the side device used to perform the lithium niobate crystal growth method based on machine vision and sequence prediction can determine the control parameters using the fine-tuned growth prediction model. During the fine-tuning process, the attention weights of the Transformer encoder are optimized using a contrastive learning approach.

[0074] Specifically, after crystal growth is completed, the processed crystal is subjected to Curie temperature testing, optical uniformity testing, and birefringence testing in sequence. Among these, the Curie temperature test can reflect the crystal's properties. The axial distribution of the refractive index was measured using an interferometer to determine its uniformity. The system automatically selected growth batches with uniform composition based on the axial standard deviation of the Curie temperature. These batches were designated as positive samples, while those with substandard composition uniformity were designated as negative samples. The system also acquired complete growth process data for each batch, including melt surface diagram sequences, operating time sequence data, and control command sequences.

[0075] Then, a contrastive learning training pair is constructed based on positive and negative samples. Through contrastive learning, the Transformer encoder narrows the distance between the same parameter correlation features in the feature space in positive samples and widens the distance between different parameter correlation features in negative samples. This optimizes the attention weight allocation of the multi-head self-attention module, enabling the model to more accurately identify long-distance parameter correlations that have a key impact on the final crystal quality and improve the model's prediction accuracy of the impact of abnormal parameters.

[0076] Finally, after fine-tuning, a model update package is generated and sent to each side industrial control unit via the MQTT protocol. The model is automatically updated before the start of the next crystal growth cycle, realizing the continuous accumulation of growth data from multiple cycles and the iterative optimization of model performance, thereby continuously improving control accuracy and crystal yield.

[0077] In summary, this embodiment provides a lithium niobate crystal growth method based on machine vision and sequence prediction. The method includes: acquiring a melt surface image using a vision acquisition device installed on a growth furnace; identifying the melt surface image using a machine vision model to obtain visual feature data, which includes at least the shoulder expansion angle; determining current operating condition data based on the visual feature data and sensor data from the growth furnace, and constructing operating condition time-series data based on the current operating condition data and historical operating condition data of a preset duration; capturing long-distance parameter correlation features using a growth prediction model based on the operating condition time-series data, and determining control parameters based on the parameter correlation features; and controlling the growth of the lithium niobate crystal based on the control parameters. This embodiment acquires a melt surface image using a vision acquisition device installed on the growth furnace, and then identifies the melt surface image to perform real-time visual monitoring of features such as the shoulder expansion angle. This achieves direct visual closed-loop control of the shoulder formation process, effectively avoiding deviations caused by human experience judgment, reducing the probability of abnormal situations such as shoulder deformities and thermal stress cracking, and improving the growth yield of lithium niobate crystals.

[0078] Based on the aforementioned lithium niobate crystal growth method based on machine vision and sequence prediction, this embodiment provides a lithium niobate crystal growth apparatus based on machine vision and sequence prediction, such as... Figure 4 As shown, the device includes: The acquisition module 100 is used to acquire a melt surface image through a vision acquisition device installed on the growth furnace. The vision module 200 is used to identify the melt surface map through a machine vision model to obtain visual feature data; The construction module 300 is used to determine the current operating condition data based on the visual feature data and the sensor data of the growth furnace, and to construct the operating condition time series data based on the current operating condition data and historical operating condition data of a preset duration. The control parameter determination module 400 is used to capture long-distance parameter correlation features based on the operating condition time series data using a growth prediction model, and determine control parameters based on the parameter correlation features. The control module 500 is used to control the growth of the lithium niobate crystal based on the control parameters.

[0079] Based on the above-described lithium niobate crystal growth method based on machine vision and sequence prediction, this embodiment provides a computer-readable storage medium storing one or more programs that can be executed by one or more processors to implement the steps in the lithium niobate crystal growth method based on machine vision and sequence prediction as described in the above embodiment.

[0080] Based on the aforementioned lithium niobate crystal growth method based on machine vision and sequence prediction, this application also provides a terminal device, such as... Figure 5 As shown, it includes at least one processor 20; a display screen 21; and a memory 22, and may also include a communications interface 23 and a bus 24. The processor 20, display screen 21, memory 22, and communications interface 23 can communicate with each other via the bus 24. The display screen 21 is configured to display a preset user guide interface in the initial setup mode. The communications interface 23 can transmit information. The processor 20 can invoke logical instructions in the memory 22 to execute the methods described in the above embodiments.

[0081] Furthermore, the logical instructions in the aforementioned memory 22 can be implemented as software functional units and, when sold or used as independent products, can be stored in a computer-readable storage medium.

[0082] The memory 22, as a computer-readable storage medium, can be configured to store software programs, computer-executable programs, such as program instructions or modules corresponding to the methods in the embodiments of this disclosure. The processor 20 executes functional applications and data processing by running the software programs, instructions, or modules stored in the memory 22, thereby implementing the methods in the above embodiments.

[0083] The memory 22 may include a program storage area and a data storage area. The program storage area may store the operating system and application programs required for at least one function; the data storage area may store data created based on the use of the terminal device. Furthermore, the memory 22 may include high-speed random access memory (RAM) and non-volatile memory. Examples include various media capable of storing program code, such as USB flash drives, portable hard drives, read-only memory (ROM), random access memory (RAM), magnetic disks, or optical disks, as well as transient storage media.

[0084] Furthermore, the specific process of loading and executing multiple instruction processors in the aforementioned storage medium and terminal device has been described in detail in the above method, and will not be repeated here.

[0085] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of this application, and are not intended to limit them. Although this application has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some of the technical features. Such modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the spirit and scope of the technical solutions of the embodiments of this application.

Claims

1. A method for growing lithium niobate crystals based on machine vision and sequence prediction, characterized in that, The method includes: The melt surface image is acquired using a vision acquisition device installed on the growth furnace; The melt surface diagram is identified by a machine vision model to obtain visual feature data, which includes at least one or more of the following: shoulder expansion angle, neck diameter, equal diameter deviation, and melt convection mode. The current operating condition data is determined based on the visual feature data and the sensor data of the growth furnace, and the operating condition time series data is constructed based on the current operating condition data and historical operating condition data of a preset duration. Based on the aforementioned operating condition time series data, a growth prediction model is used to capture long-distance parameter correlation characteristics, and control parameters are determined based on these parameter correlation characteristics. The growth of the lithium niobate crystal is controlled based on the control parameters.

2. The lithium niobate crystal growth method based on machine vision and sequence prediction according to claim 1, characterized in that, The visual acquisition device is deployed outside the observation window on the side wall of the growth furnace, and its viewing angle covers the inner cavity of the growth furnace so that the acquired melt surface image includes the melt surface and the crystal contact interface.

3. The lithium niobate crystal growth method based on machine vision and sequence prediction according to claim 1, characterized in that, The step of identifying the melt surface pattern using a machine vision model to obtain the lithium niobate crystal growth parameters specifically includes: The melt surface image is input into the machine vision model, and the crystal boundary and melt surface in the melt surface image are detected by the vision detection module in the machine vision model to determine the candidate boundary of the crystal profile. The candidate boundary is used to locate the region of interest in the melt surface map. The feature extraction module in the machine vision model is used to extract the sub-pixel level edge of the region of interest, and visual feature data is determined based on the sub-pixel level edge.

4. The lithium niobate crystal growth method based on machine vision and sequence prediction according to claim 1, characterized in that, The sensor data includes crystal weight, rotation speed, lifting speed, and temperature data. The crystal weight includes the weight collected by a weighing sensor deployed on the top of the lifting rod of the growth furnace, and the temperature data includes the temperature of each independent heating zone collected by a temperature controller deployed around the crucible of the growth furnace.

5. The lithium niobate crystal growth method based on machine vision and sequence prediction according to claim 1, characterized in that, The step of capturing long-range parameter correlation features using the growth prediction model based on the operating condition time series data, and determining control parameters based on the parameter correlation features, specifically includes: The operating condition time series data is input into the Transformer encoder in the growth prediction model, and the Transformer encoder uses a multi-head self-attention mechanism to capture long-distance parameter correlation features. The parameter correlation features are input into the GRU decoder in the growth prediction model, and the control parameters for a future preset duration are predicted by the GRU decoder.

6. The lithium niobate crystal growth method based on machine vision and sequence prediction according to claim 1, characterized in that, The control parameters include the speed adjustment of the lifting servo motor, the speed adjustment of the crucible rotation servo motor, and the power distribution matrix of the multi-temperature zone heater.

7. The lithium niobate crystal growth method based on machine vision and sequence prediction according to claim 1, characterized in that, The method further includes: Historical data on the growth process of multiple lithium niobate crystals are aggregated through a cloud server. The Transformer encoder in the growth prediction model is fine-tuned using the historical data so that the side device used to perform the lithium niobate crystal growth method based on machine vision and sequence prediction can determine the control parameters using the fine-tuned growth prediction model. During the fine-tuning process, the attention weights of the Transformer encoder are optimized using a contrastive learning approach.

8. A lithium niobate crystal growth apparatus based on machine vision and sequence prediction, characterized in that, The device includes: The acquisition module is used to acquire melt surface images via a vision acquisition device installed on the growth furnace; The vision module is used to identify the melt surface pattern through a machine vision model to obtain visual feature data; The construction module is used to determine the current operating condition data based on the visual feature data and the sensor data of the growth furnace, and to construct the operating condition time series data based on the current operating condition data and historical operating condition data of a preset duration. The control parameter determination module is used to capture long-distance parameter correlation features based on the operating condition time series data using a growth prediction model, and determine control parameters based on the parameter correlation features. A control module is used to control the growth of the lithium niobate crystal based on the control parameters.

9. A computer-readable storage medium, characterized in that, The computer-readable storage medium stores one or more programs, which can be executed by one or more processors to implement the steps in the lithium niobate crystal growth method based on machine vision and sequence prediction as described in any one of claims 1-7.

10. A terminal device, characterized in that, include: Processor and memory; The memory stores a computer-readable program that can be executed by the processor; When the processor executes the computer-readable program, it implements the steps in the lithium niobate crystal growth method based on machine vision and sequence prediction as described in any one of claims 1-7.