A method for controlling the grain boundary morphology in the growth of indium phosphide crystals

Abstract

The application discloses a method for controlling grain boundary morphology in indium phosphide crystal growth, comprising: obtaining crystal surface original image data, performing scale remapping, gray scale standardization and multi-round noise suppression processing; extracting geometric description features, moment feature description features and contour sequence description features of connected domains; constructing a grain boundary morphology category recognition model to recognize the grain boundary morphology category to which the connected domain corresponding grain boundary region belongs; extracting grain boundary multi-scale texture features and grain boundary morphology features; constructing a grain boundary crack risk level evaluation model to evaluate the grain boundary crack risk level; predicting the grain boundary merging event in a future preset time period, judging the crack risk level after the grain boundary merging; and evaluating the effectiveness of the process adjustment scheme and updating the process adjustment scheme. The application effectively reduces the crack failure risk caused by grain boundary evolution in the crystal growth process, and significantly improves the stability, intelligent level and overall quality consistency of the finished crystal in the indium phosphide crystal growth process.

Description

Technical Field

[0001] This invention relates to the field of next-generation information technology, and in particular to a method for controlling grain boundary morphology during indium phosphide crystal growth. Background Technology

[0002] Indium phosphide (IPS), an important group III-V compound semiconductor material, is widely used in high-speed communication devices, optoelectronic devices, and high-frequency microwave devices due to its high electron mobility, direct bandgap, and excellent optoelectronic properties. During IPS crystal growth, the formation, evolution, and morphological stability of grain boundaries have a decisive impact on the crystal's mechanical properties, electrical properties, and yield. However, during IPS crystal growth, due to the combined effects of temperature gradients, compositional fluctuations, growth rate changes, and external disturbances, different types of grain boundary structures inevitably form within the crystal, such as small-angle grain boundaries, large-angle grain boundaries, twin boundaries, and pseudo-grain boundaries. These grain boundaries easily become stress concentration areas during subsequent cooling and use, inducing crack initiation, propagation, and even crystal failure, severely affecting the yield and application reliability. Therefore, accurate identification, risk assessment, and effective control of grain boundary morphology are key technical issues for improving the quality of IPS crystal growth. In existing technologies, grain boundary detection and analysis methods mostly rely on manual interpretation of microscopic images or automatic identification methods based on single image features. These methods are typically sensitive to imaging conditions, noise levels, and surface texture variations, making it difficult to maintain consistency and stability of identification results across different batches, devices, or growth stages. Furthermore, traditional methods often focus on static grain boundary detection, merely determining their presence without delving into the complexity of grain boundary morphology, texture inhomogeneity, and evolutionary trends, hindering quantitative assessment of potential grain boundary crack risks. Moreover, grain boundaries are not static during crystal growth; they migrate, evolve, and even merge depending on growth conditions. Existing process control methods rely primarily on empirical parameters or post-processing results for adjustments, failing to provide early warnings of crack risks arising from grain boundary migration and merging, thus making it difficult to prevent sudden crystal cracking or failure. Therefore, there is an urgent need for a grain boundary morphology control method capable of highly consistent, quantifiable identification of grain boundary morphology during indium phosphide crystal growth, enabling multi-scale fine characterization, crack risk prediction, and process parameter optimization. This would improve the intelligence level of the crystal growth process, reduce crack failure risk, and enhance the structural quality and application reliability of indium phosphide crystals. Summary of the Invention

[0003] To address the problems existing in the prior art, this invention provides a method for controlling grain boundary morphology during indium phosphide crystal growth, mainly comprising:

[0004] The original image data of the crystal surface is acquired by an optical camera, and scale remapping is completed by combining pixel size and spatial calibration parameters. Gray-level normalization and noise suppression are performed based on gray-level statistics and cumulative distribution mapping.

[0005] Based on the noise-suppressed crystal surface image, gradient feature images are obtained using gradient operators, and geometric description features, moment feature description features, and contour sequence description features of connected domains are extracted.

[0006] Based on the geometric description features, moment feature description features, and contour sequence description features of connected domains, a grain boundary morphology category identification model is constructed to identify the grain boundary morphology category to which the corresponding grain boundary region belongs.

[0007] A multi-scale sliding window is used to divide the local neighborhood of the grain boundary, and the gray-level co-occurrence matrix and gray-level run matrix are calculated in each scale sub-region of the grain boundary to extract multi-scale texture features and grain boundary morphology features.

[0008] Based on the multi-scale texture characteristics and morphological characteristics of grain boundaries, a grain boundary crack risk assessment model is constructed to assess the risk level of grain boundary cracks and predict the risk of grain boundary cracks within a preset time period.

[0009] By continuously monitoring and predicting grain boundary migration data using an optical camera, grain boundary merging events can be identified within a preset time period in the future, and the crack risk level after grain boundary merging can be determined.

[0010] Based on the comparison of the risk levels of grain boundary cracks before and after the process adjustment, the effectiveness of the process adjustment plan is evaluated, and the process adjustment plan is updated in conjunction with continuous monitoring feedback.

[0011] Furthermore, the process of acquiring raw image data of the crystal surface through an optical camera, performing scale remapping by combining pixel size and spatial calibration parameters, and performing grayscale normalization and noise suppression processing based on grayscale statistics and cumulative distribution mapping includes:

[0012] An optical camera installed inside the indium phosphide crystal growth apparatus acquires raw image data of the crystal surface in real time. Using a unified image encoding format, bit depth parameters, and color channel configuration, the raw image data is normalized. Scale remapping is then performed by reading the pixel size and spatial calibration parameters of the optical camera to obtain standardized crystal surface image data with consistent spatial scale. A grayscale statistical method is used to count the number of pixels at each gray level in the crystal surface image, constructing a grayscale probability distribution function. This probability distribution function is then accumulated to obtain a cumulative distribution function. Based on the cumulative distribution function, a mapping relationship between the raw grayscale values ​​and the target grayscale values ​​is established, and the grayscale value of each pixel in the image is replaced point-by-point to obtain a grayscale-normalized crystal surface image. Based on the grayscale-normalized crystal surface image, a multi-round median filtering algorithm is used to traverse the image matrix through a window. The change in the number of isolated noise pixels before and after each round of filtering is counted, and stability is assessed. After stability is determined, anisotropic diffusion filtering is performed. The diffusion coefficient is calculated based on the local gradient magnitude, and the pixel intensity is iteratively updated to obtain a noise-suppressed crystal surface image.

[0013] Furthermore, based on the noise-suppressed crystal surface image, a gradient feature image is obtained using a gradient operator, and geometric description features, moment feature description features, and contour sequence description features of the connected components are extracted, including:

[0014] Based on the noise-suppressed crystal surface image, a gradient operator is used to calculate the first-order partial derivatives in the horizontal and vertical directions respectively. Gradient feature images are obtained by synthesizing gradient magnitude and gradient direction matrices. A multi-directional filter kernel is used to perform convolution operations on the gradient feature images, recording the response intensity in each direction. Based on the angle between the principal gradient direction and the filter kernel direction, the response contribution is determined and the gradient response intensity is calculated. By comparing the gradient response intensity with a preset response intensity threshold, continuous high-response regions are identified, and an initial binary image of the grain boundary contour is obtained. A connected component labeling algorithm is used to group adjacent edge pixels into regions, determining the set of pixel coordinates for each connected component. The average value of all pixel coordinates in the connected component is calculated to obtain the centroid coordinates of each connected component. The extension length is determined by calculating the aspect ratio of the minimum bounding rectangle of the connected component, and the compactness formula is used. Determine the compactness of connected components. The ratio of the area of ​​the connected region to the area of ​​its smallest bounding rectangle is calculated to determine the rectangularity. The area and perimeter are then combined to form the geometric description feature of the connected region, where A is the area of ​​the connected region and P is the perimeter. The second and third central moments of the connected region are calculated and normalized. Based on the normalized moments, seven Hu invariant moments are obtained as moment feature description features. The approximate curvature of each point on the contour of the connected region is calculated, and its mean, variance, and maximum value are statistically analyzed. The contour point sequence is subjected to discrete Fourier transform, and the amplitude of the previously preset number of low-frequency coefficients is taken as the contour sequence description feature. The obtained geometric description feature, moment feature description feature, and contour sequence description feature of the connected region are stored in the indium phosphide crystal growth monitoring database.

[0015] Furthermore, the step of constructing a grain boundary morphology category identification model based on the geometric description features of connected domains, the moment feature description features, and the contour sequence description features to identify the grain boundary morphology category to which the grain boundary region corresponding to each connected domain belongs includes:

[0016] Historical data on the geometric, moment, and contour sequence description features of connected domains were obtained from the indium phosphide crystal growth monitoring database. The grain boundary morphology categories of each connected domain were labeled, and a random forest algorithm was used for model training to construct a grain boundary morphology category recognition model. The grain boundary morphology categories include small-angle grain boundaries, large-angle grain boundaries, twin boundaries, pseudo-grain boundaries, and grain boundary network nodes. Pseudo-grain boundaries are connected domains that are caused by surface textures, processing marks, or noise and do not meet the criteria for true grain boundaries. Based on the geometric, moment, and contour sequence description features of connected domains extracted in real-time from crystal surface images, the grain boundary morphology category recognition model was used to identify the grain boundary morphology category of each connected domain. Connected domains identified as pseudo-grain boundaries were removed and no longer monitored.

[0017] Furthermore, the method employs a multi-scale sliding window to divide the local neighborhood of the grain boundary, and within each scale sub-region of the grain boundary, calculates the gray-level co-occurrence matrix and gray-level run-length matrix to extract multi-scale texture features and grain boundary morphology features, including:

[0018] Based on the set of pixel coordinates for each connected domain, the grain boundary positions are marked using pixel coordinate mapping, and the corresponding set of grain boundary region image blocks is extracted to obtain grain boundary region index data. Using the grain boundary line as the center, along its normal direction, a multi-scale sliding window is used to divide the local neighborhood of the grain boundary within a preset scale range, constructing a grain boundary sub-region dataset at each scale and recording the corresponding scale identification information. Within each scale grain boundary sub-region, the gray-level co-occurrence matrix and gray-level run length matrix are calculated, and texture features extracted from the same grain boundary in different scale sub-regions are extracted, including contrast, homogeneity, energy, entropy, correlation, directional consistency, and run length non-uniformity. The mean, standard deviation, maximum value, and quantile of the texture features obtained for the same grain boundary in different scale sub-regions are calculated respectively as multi-scale texture features of the grain boundary. By fitting the curvature of the grain boundary contour edge point sequence, the local curvature, rate of change of curvature, and inflection point distribution density are calculated as grain boundary morphological features. The multi-scale texture features and grain boundary morphological features are stored in the indium phosphide crystal growth monitoring database.

[0019] Furthermore, the step of constructing a grain boundary crack risk level assessment model based on multi-scale texture features and grain boundary morphology features, assessing the grain boundary crack risk level, and predicting the grain boundary crack risk within a preset time period includes:

[0020] Historical data on multi-scale texture features and morphological characteristics of grain boundaries, along with corresponding grain boundary morphology categories, were acquired using an indium phosphide crystal growth monitoring database. Crack risks for each grain boundary were then labeled. A random forest algorithm was used to train a model to construct a grain boundary crack risk assessment model, categorizing crack risks into high, medium, and low. Based on real-time acquired grain boundary morphology categories, multi-scale texture features, and morphological characteristics, the grain boundary crack risk assessment model was used to evaluate the grain boundary crack risk level. A tiered early warning system was implemented based on the assessment results. This included highlighting high-risk grain boundaries and generating process adjustment plans, marking and tracking medium-risk grain boundaries, and recording low-risk grain boundaries. The process adjustment plans included adjusting the local temperature within the growth apparatus. The system employs various methods to predict grain boundary characteristics. These include adjusting the gradient, regulating the rate of pressure change within the growth apparatus, adjusting the cooling gradient and rhythm inside and outside the furnace, setting monitoring markers for high-risk merging areas, and implementing a shutdown mechanism. Historical data on the geometric, moment, and contour sequence characteristics, multi-scale texture, and grain boundary morphology of medium- and low-risk grain boundaries are obtained from the indium phosphide crystal growth monitoring database. A long short-term memory network is used to train the model, constructing a grain boundary feature prediction model. This model predicts the geometric, moment, contour sequence, multi-scale texture, and grain boundary morphology characteristics of grain boundaries within a predetermined timeframe. Combined with a grain boundary morphology category identification model and a grain boundary crack risk level assessment model, the model predicts the time point for the grain boundary crack risk level transition and provides early warnings for the grain boundary.

[0021] Furthermore, the step of continuously monitoring and predicting grain boundary migration data using an optical camera, identifying grain boundary merging events within a preset time period, and determining the crack risk level after grain boundary merging includes:

[0022] Historical grain boundary migration data is continuously acquired in a time series using an optical camera. A long short-term memory network is used to train the model, constructing a grain boundary migration prediction model to predict grain boundary migration data within a preset future time period. The migration data includes the centroid coordinates of the grain boundary, the overall displacement direction of the grain boundary, the migration velocity, the spacing between adjacent grain boundaries, and the intersection angle. If the minimum spatial spacing, the opposite migration condition, and the intersection angle condition all meet preset judgment conditions within the predicted future time period, then a grain boundary merging event is determined to exist within the future preset time period. If a grain boundary merging event exists, the merging crack risk assessment formula is applied. Determine the crack risk index R after grain boundary merging, where, and These represent the grain boundary areas where grain boundary merging events occurred. and These are the grain boundary perimeters where the grain boundary merging event occurred. and Here, represents the entropy of the grain boundary where the grain boundary merging event occurred. and These represent the rate of change of grain boundary curvature during a grain boundary merging event. and These represent the grain boundary migration velocities at which grain boundary merging events occur. The geometric angle of grain boundary convergence at which a grain boundary merging event occurs. The grain boundary morphology merging coefficient corresponding to the combination of grain boundary morphology categories before merging is obtained from the grain boundary morphology merging coefficient table obtained by fitting historical crystal growth data; if the crack risk index after grain boundary merging is greater than the preset index threshold, the crack risk level after grain boundary merging is judged to be high, triggering a high-level warning, generating a process adjustment plan, and making process adjustments in advance.

[0023] Furthermore, the effectiveness of the process adjustment plan is evaluated based on the comparison of grain boundary crack risk levels before and after the process adjustment, and the process adjustment plan is updated in conjunction with continuous monitoring feedback, including:

[0024] Obtain the grain boundary crack risk level before and after process adjustment. By evaluating the change in crack risk level after process adjustment, determine the effectiveness of the process adjustment plan. If the grain boundary crack risk level decreases, the process adjustment measures are deemed effective, and the process adjustment plan is saved as a valid process operation template. If the grain boundary crack risk level does not change or increases, optimize the process adjustment plan, including adjusting the temperature gradient of the crystal growth device, optimizing the raw material supply rate, adjusting the crystal rotation speed, or modifying the grain boundary intersection angle. Continuously monitor the change in grain boundary crack risk level after implementing the optimized process adjustment plan, and continuously update the process adjustment plan based on real-time monitoring feedback until the grain boundary crack risk level is controlled within the preset safety range.

[0025] The technical solutions provided by the embodiments of the present invention may include the following beneficial effects:

[0026] This invention provides a method for controlling grain boundary morphology during indium phosphide crystal growth. By performing scale remapping, grayscale normalization, and multi-round noise suppression on crystal surface images, this invention significantly reduces the impact of imaging conditions and noise interference on grain boundary detection results, achieving stable and quantifiable identification of grain boundary regions. Combining gradient features and connected component analysis, this invention utilizes geometric description features, moment feature description features, and contour sequence description features to stably characterize grain boundary morphology and constructs a grain boundary morphology category recognition model, which can accurately distinguish multiple grain boundary types, significantly reducing the risk of false detection and misjudgment. Furthermore, this invention extracts multi-scale texture and morphological features of grain boundaries through a multi-scale sliding window, enabling precise characterization of local anomalies and morphological instabilities. Combined with a crack risk level assessment model, it achieves quantitative classification and trend prediction of grain boundary crack risk. Simultaneously, through continuous monitoring and prediction of grain boundary migration and merging behavior, high-risk grain boundary merging events can be identified in advance, reducing stress concentration and crack propagation probability. This invention further integrates risk assessment results before and after process adjustments to form a closed-loop feedback mechanism, achieving adaptive optimization control of process parameters. This significantly improves the stability, intelligence level, and overall quality consistency of the indium phosphide crystal growth process. The grain boundary morphology control method provided by this invention enables stable identification of grain boundary morphology, prediction and assessment of crack risk, and closed-loop adaptive control of process parameters. This effectively reduces the risk of crack failure caused by grain boundary evolution during crystal growth, significantly improving the stability, intelligence level, and overall quality consistency of the indium phosphide crystal growth process. Attached Figure Description

[0027] Fig. 1 This is a flowchart of a method for controlling grain boundary morphology during indium phosphide crystal growth according to the present invention;

[0028] Fig. 2 This is a schematic diagram of a method for controlling grain boundary morphology during indium phosphide crystal growth according to the present invention.

[0029] Fig. 3 This is another schematic diagram of a method for controlling grain boundary morphology during the growth of indium phosphide crystals according to the present invention. Detailed Implementation

[0030] To make the objectives, technical solutions, and advantages of this invention clearer, the invention will be described in detail below with reference to the accompanying drawings and specific embodiments.

[0031] This invention relates to CN201720133446.0, a device for growing indium phosphide single crystals, comprising a high-pressure furnace, a nitrogen system, a heating system, a cooling system, and a control system. The high-pressure furnace includes a support frame, a furnace body, and a reciprocating tilting mechanism. The nitrogen system includes a gas cylinder and a controllable electric valve. The heating system includes a heater and a crucible holder disposed within the furnace body cavity. The cooling system includes a water tank, a radiator, a water pump, and a hollow copper tube connected in sequence. The control system includes a pressure sensor, a thermocouple, a temperature sensor, and a programmable logic controller. This invention provides a highly automated indium phosphide single crystal growth device that effectively improves the growth quality and production efficiency of indium phosphide single crystals.

[0032] like Figs. 1-3 This embodiment of a method for controlling grain boundary morphology during indium phosphide crystal growth may specifically include:

[0033] Step S101: Obtain the original image data of the crystal surface through an optical camera, complete the scale remapping by combining the pixel size and spatial calibration parameters, and perform grayscale normalization and noise suppression processing based on grayscale statistics and cumulative distribution mapping.

[0034] An optical camera installed inside the indium phosphide crystal growth apparatus acquires raw image data of the crystal surface in real time. Using a unified image encoding format, bit depth parameters, and color channel configuration, the raw image data is normalized. Scale remapping is then performed by reading the pixel size and spatial calibration parameters of the optical camera to obtain standardized crystal surface image data with consistent spatial scale. Gray-level statistics are used to count the number of pixels at each gray level in the crystal surface image, constructing a gray-level probability distribution function. This probability distribution function is then accumulated to obtain a cumulative distribution function. Based on the cumulative distribution function, a mapping relationship between the raw gray values ​​and the target gray values ​​is established, and the gray value of each pixel in the image is replaced point-by-point to obtain a gray-level standardized crystal surface image. Based on the gray-level standardized crystal surface image, a multi-round median filtering algorithm is used to traverse the image matrix through a window. The change in the number of isolated noise pixels before and after each round of filtering is counted, and stability is assessed. After stability is determined, anisotropic diffusion filtering is performed. The diffusion coefficient is calculated based on the local gradient magnitude, and the pixel intensity is iteratively updated to obtain a noise-suppressed crystal surface image.

[0035] For example, during an indium phosphide crystal growth process, an optical camera installed inside the indium phosphide crystal growth apparatus continuously images the crystal surface at a sampling frequency of 5 frames per second, obtaining a raw crystal surface image with a resolution of 2048×2048 pixels, a bit depth of 12 bits, and an RGB three-channel format. Since the image encoding format may differ between different batches of cameras or under different imaging conditions, the raw images are first uniformly converted to a grayscale single-channel format, and the bit depth is uniformly mapped to an 8-bit grayscale range, i.e., a grayscale value range of 0–255. Simultaneously, the pixel size parameters of the optical camera are read (the actual physical size of a single pixel is 5μm × 5μm), along with the furnace space calibration results, to perform scale remapping on the image. This ensures that the distance between any two pixels in the image accurately corresponds to the actual physical distance on the crystal surface, thereby obtaining a standardized crystal surface image with consistent spatial scale. After obtaining the standardized crystal surface image, the number of pixels at each gray level in the entire image is statistically analyzed. The distribution of pixels with gray values ​​from 0 to 255 is obtained, with the gray values ​​concentrated in the range of 40 to 120, indicating that the overall brightness of the crystal surface is too dark and the contrast is insufficient. Based on this statistical result, a gray-level probability distribution function is constructed, and a cumulative distribution function is obtained by accumulating the values. For example, when the cumulative distribution function reaches 0.5 at a gray value of 80, it indicates that 50% of the pixels have a gray value no higher than 80. A mapping relationship between the original gray values ​​and the target gray values ​​is established according to the cumulative distribution function, stretching the originally concentrated gray range to a wider target gray range, such as mapping the original gray values ​​of 40 to 120 to the target gray values ​​of 30 to 220, thereby enhancing the contrast of the crystal surface. The gray value of each pixel in the image is then replaced point by point according to this mapping relationship to obtain a gray-normalized crystal surface image with a more uniform gray distribution. After grayscale normalization, to further suppress random noise and isolated bright spots generated during imaging, multiple rounds of median filtering are performed on the image. For example, a 3×3 filtering window is used for the first round of median filtering, and the change in the number of isolated pixels with abrupt grayscale changes before and after filtering is statistically analyzed. When the change rate of the number of isolated noise pixels after the second round of median filtering is less than 5%, the median filtering result is considered to be stable. Based on the stability determination, anisotropic diffusion filtering is performed on the image. During diffusion, the diffusion coefficient is dynamically calculated based on the local gradient magnitude. For example, diffusion is enhanced in smooth regions with a gradient magnitude of less than 10 to eliminate noise, while diffusion is suppressed in edge regions with a gradient magnitude greater than 30 to maintain the grain boundary structure. After multiple iterations to update the pixel intensity, a noise-suppressed crystal surface image is finally obtained, which effectively suppresses background noise while maintaining the clarity of the grain boundary edges.

[0036] Step S102: Based on the noise-suppressed crystal surface image, a gradient feature image is obtained using a gradient operator, and geometric description features, moment feature description features, and contour sequence description features of the connected domains are extracted.

[0037] Based on the noise-suppressed crystal surface image, a gradient operator is used to calculate the first-order partial derivatives in the horizontal and vertical directions respectively. Gradient feature images are obtained by synthesizing the gradient magnitude matrix and gradient direction matrix. A multi-directional filter kernel is used to perform convolution operations on the gradient feature images, recording the response intensity in each direction. Based on the angle between the principal gradient direction and the filter kernel direction, the response contribution is determined and the gradient response intensity is calculated. By comparing the gradient response intensity with a preset response intensity threshold, continuous high-response regions are identified, and an initial binary image of the grain boundary contour is obtained. A connected component labeling algorithm is used to group adjacent edge pixels into regions, determining the set of pixel coordinates for each connected component. The average value of all pixel coordinates in the connected component is calculated to obtain the centroid coordinates of each connected component. The extension length is determined by calculating the aspect ratio of the minimum bounding rectangle of the connected component, and the compactness formula is used. Determine the compactness of connected components. The ratio of the area of ​​a connected region to the area of ​​its smallest bounding rectangle is calculated to determine its rectangularity. This rectangularity, combined with the area and perimeter, serves as the geometric description feature of the connected region, where A is the area and P is the perimeter. The second and third central moments of the connected region are calculated, normalized, and seven Hu invariant moments are obtained based on these normalized moments, serving as moment feature description features. The approximate curvature of each point on the contour of the connected region is calculated, and its mean, variance, and maximum value are statistically analyzed. A discrete Fourier transform is performed on the contour point sequence, and the amplitude of a predetermined number of low-frequency coefficients is used as the contour sequence description feature. The obtained geometric description features, moment feature description features, and contour sequence description features of the connected region are stored in the indium phosphide crystal growth monitoring database.

[0038] For example, during an indium phosphide crystal growth process, an optical camera installed inside the indium phosphide crystal growth apparatus acquires a grayscale image of the crystal surface with a size of 2048×2048 pixels and a pixel size of 5μm×5μm. After image format normalization, grayscale standardization, and noise suppression processing, gradient analysis is performed on the crystal surface image. The Sobel gradient operator is used to calculate the first-order partial derivatives in the horizontal and vertical directions, respectively, and a response band with a gradient magnitude significantly higher than that of the background region is obtained in a certain local region. By synthesizing gradient magnitude and gradient direction matrices, and performing convolution operations on the gradient feature image using directional filter kernels in four directions (0°, 45°, 90°, 135°), the contribution of each directional gradient response to the grain boundary structure is determined based on the angle relationship between the principal gradient direction of the pixel and the direction of each directional filter kernel. Responses with higher directional consistency are assigned higher weights. The gradient response intensity of each direction is recorded, and weighted summaries are performed according to the contribution of different directions to the grain boundary structure to obtain a comprehensive gradient response intensity. Finally, a continuous high-response band with a gradient response intensity of 180 is obtained in this region, which is higher than the preset threshold of 120, thus forming an initial grain boundary contour after binarization. After performing connected component labeling on this binary grain boundary contour, an independent connected component is identified, which contains 312 pixels. By averaging the coordinates of these 312 pixels, the centroid coordinates of this connected component are calculated to be (1024.6, 987.2). In physical space, the corresponding position of this centroid is (5123μm, 4936μm). Further analysis of the bounding rectangle of this connected component reveals that its longest side is 78 pixels and its shortest side is 6 pixels, corresponding to physical dimensions of 390μm and 30μm respectively. Therefore, the elongation of this connected component is calculated to be 78÷6=13, indicating a distinctly elongated linear structure. Within this connected component, the actual area is 312 pixels, corresponding to a physical area of ​​312×25μm²=7800μm². Simultaneously, the area of ​​the smallest bounding rectangle is 78×6=468 pixels, resulting in a rectangularity of 312÷468≈0.667. After boundary tracing of the connected component's outline, its perimeter is found to be 176 pixels, corresponding to a physical perimeter of 880μm. Substituting the area and perimeter into the compactness formula... The compactness of the connected domain was calculated to be 0.126, which is significantly smaller than that of the compactness of a regular closed region, further indicating that the connected domain has typical grain boundary characteristics. Subsequently, the second and third central moments of the connected domain were calculated and normalized. Based on the normalized central moments, seven Hu invariant moments were calculated, with values ​​of (0.238, 0.041, 0.0021, 0.00036, −0.000012, 0.0000018, −0.0000004). These moment characteristics are used to describe the morphological properties of the grain boundary that remain unchanged under rotation, translation, and scale changes. In the contour analysis stage, the approximate curvature was calculated point-by-point for 176 contour points on the contour of the connected domain, and the average curvature was statistically obtained as 0.018 μm. -1 The curvature variance is 0.006 μm. -2 The maximum curvature is 0.052 μm. -1 This indicates that the grain boundary exhibits a certain degree of local bending. Further discrete Fourier transform was performed on the contour point sequence, and the amplitudes of the first 10 low-frequency coefficients were selected as the contour sequence description features. Their amplitudes are concentrated between [0.85, 0.12], reflecting that the overall morphology of the grain boundary is smooth but there are local morphological fluctuations. Finally, the geometric description features of this connected region (elongation 13, compactness 0.126, rectangularity 0.667, area 7800 μm², perimeter 880 μm, moment feature description features, 7 Hu invariant moments, contour sequence description features, and 10 low-frequency Fourier coefficients) were uniformly stored in the indium phosphide crystal growth monitoring database.

[0039] Step S103: Based on the geometric description features of connected domains, the moment feature description features, and the contour sequence description features, a grain boundary morphology category identification model is constructed to identify the grain boundary morphology category to which the grain boundary region corresponding to each connected domain belongs.

[0040] Historical data on the geometric, moment, and contour sequence description features of connected domains were obtained from the indium phosphide crystal growth monitoring database. The grain boundary morphology category of each connected domain was labeled, and a random forest algorithm was used to train the model, constructing a grain boundary morphology category recognition model. The grain boundary morphology categories include small-angle grain boundaries, large-angle grain boundaries, twin boundaries, pseudo-grain boundaries, and grain boundary network nodes. Pseudo-grain boundaries are connected domains that are caused by surface textures, processing marks, or noise and do not meet the criteria for true grain boundaries. Based on the geometric, moment, and contour sequence description features of connected domains extracted in real-time from crystal surface images, the grain boundary morphology category recognition model was used to identify the grain boundary morphology category of each connected domain. Connected domains identified as pseudo-grain boundaries were removed and no longer monitored.

[0041] For example, using the indium phosphide crystal growth monitoring database, historical data of 2000 connected component geometric description features, moment description features, and contour sequence description features identified by the connected component labeling algorithm were obtained. For one connected component numbered C3, its geometric description features, moment description features, and contour sequence description features were extracted. This connected component has a pixel area of ​​842 pixels and a corresponding perimeter of 196 pixels. Its minimum bounding rectangle has an aspect ratio of 7.8, a compactness calculation result of 0.23, and a rectangularity of 0.18, indicating that this connected component exhibits a distinct elongated linear structure. Further calculation and normalization of its second and third central moments yielded seven Hu invariant moments, with the first Hu moment being 0.0021 and the second Hu moment being 1.7 × 10⁻⁶. -4 The remaining Hu moments are all at 10. -5 The magnitude range indicates that the connected region morphology possesses strong rotational invariance. Furthermore, performing a Discrete Fourier Transform on the contour point sequence of this connected region and extracting the amplitudes of the first 10 low-frequency coefficients reveals a maximum amplitude of 31.6, with a monotonic amplitude decay trend, indicating overall contour continuity and the absence of significant high-frequency noise disturbances. Combining the obtained geometric, moment, and contour sequence description features of the connected region with growth process records, the grain boundary morphology category of this connected region was manually labeled as a large-angle grain boundary. This label was used as a feature vector input to a random forest algorithm for model training, constructing a grain boundary morphology category recognition model. The grain boundary morphology categories include small-angle grain boundaries, large-angle grain boundaries, twin boundaries, pseudo-grain boundaries, and grain boundary network nodes. After the model training was completed, the system entered the real-time monitoring stage. Another connected component, numbered C8, had a pixel area of ​​only 96 pixels, a minimum bounding rectangle aspect ratio of 1.3, a compactness of 0.91, and a rectangularity of 0.76. The Hu invariant moment distribution showed a high degree of concentration, and the low-frequency coefficient amplitude of the contour Fourier had no obvious dominant component. Using the grain boundary morphology category recognition model, the grain boundary morphology category was identified as a pseudo grain boundary. Its formation was explained as being caused by the superposition of local reflection texture and noise on the crystal surface. Therefore, this connected component was automatically removed in the recognition stage and did not enter the grain boundary texture analysis and crack risk assessment process.

[0042] Step S104: A multi-scale sliding window is used to divide the local neighborhood of the grain boundary, and the gray-level co-occurrence matrix and gray-level run matrix are calculated in each scale grain boundary sub-region to extract multi-scale texture features and grain boundary morphology features.

[0043] Based on the set of pixel coordinates for each connected component, grain boundary locations are marked using pixel coordinate mapping, and corresponding set of grain boundary region image blocks is extracted to obtain grain boundary region index data. Centered on the grain boundary line, a multi-scale sliding window is used to divide the local neighborhood of the grain boundary within a preset scale range along its normal direction, constructing grain boundary sub-region datasets at each scale and recording the corresponding scale identification information. Within each scale grain boundary sub-region, the gray-level co-occurrence matrix and gray-level run-length matrix are calculated, and texture features extracted from the same grain boundary in different scale sub-regions are extracted, including contrast, homogeneity, energy, entropy, correlation, directional consistency, and run-length non-uniformity. The mean, standard deviation, maximum value, and quantile of the texture features obtained for the same grain boundary in different scale sub-regions are calculated as multi-scale texture features of the grain boundary. By fitting the curvature of the grain boundary contour edge point sequence, the local curvature, rate of change of curvature, and inflection point distribution density are calculated as grain boundary morphological features. The multi-scale texture features and grain boundary morphological features are stored in the indium phosphide crystal growth monitoring database.

[0044] For example, a continuous grain boundary is detected from a currently acquired standardized crystal surface image. Its connected region consists of 860 edge pixels, with x=420–615 and y=280–305 in the image coordinate system. The grain boundary location is marked based on this set of pixel coordinates. Using this coordinate range as a basis, a 220×60 pixel grain boundary region image block is cropped from the original image. A unique region index number is assigned to this grain boundary for subsequent feature association and historical tracing. Using the grain boundary centerline as a reference, the normal direction of the grain boundary is calculated. Three scale sliding windows are constructed along the normal direction on both sides of the grain boundary: a small-scale window of 9×9 pixels, a medium-scale window of 21×21 pixels, and a large-scale window of 41×41 pixels. This forms a multi-scale grain boundary sub-region dataset within the local neighborhood of the grain boundary, and the scale identifier corresponding to each sub-region is recorded. Taking a mesoscale window as an example, 18 sub-regions were obtained along the grain boundary at this scale. The gray-level co-occurrence matrix and gray-level run-length matrix were calculated for each sub-region. The gray-level co-occurrence matrix of one sub-region showed a contrast of 4.6, homogeneity of 0.71, energy of 0.18, entropy of 3.2, and correlation of 0.83. The run-length non-uniformity corresponding to the gray-level run-length matrix was 0.64. After statistically analyzing the corresponding features of all 18 sub-regions of the same grain boundary at this scale, the mean contrast was 4.2, the standard deviation was 0.8, the maximum value was 5.9, and the 90th percentile was 5.1. The mean entropy was 3.0, the standard deviation was 0.5, the maximum value was 4.1, and the 90th percentile was 3.7. Using the same method to process the small-scale and large-scale windows, the statistical results obtained at the three scales were combined to form a multi-scale texture feature description of the grain boundary. Simultaneously, morphological analysis was performed on the contour edge point sequence of the grain boundary. After smoothing and fitting the contour, the local curvature distribution was calculated. The average local curvature was 0.012 / pixel, the maximum curvature was 0.046 / pixel, and the average rate of curvature change was 0.009. Six distinct inflection points were identified along the grain boundary length, with a corresponding inflection point density of approximately 0.028 / pixel. Finally, the multi-scale texture features and grain boundary morphology features corresponding to this grain boundary were uniformly stored in the indium phosphide crystal growth monitoring database.

[0045] Step S105: Based on the multi-scale texture features and morphological features of grain boundaries, construct a grain boundary crack risk level assessment model, assess the grain boundary crack risk level, and predict the grain boundary crack risk within a preset time period.

[0046] Historical data on multi-scale texture features and morphological characteristics of grain boundaries, along with corresponding grain boundary morphology categories, were acquired using an indium phosphide crystal growth monitoring database. Crack risks for each grain boundary were then labeled. A random forest algorithm was used to train the model, constructing a grain boundary crack risk level assessment model, categorized as high, medium, and low. Based on the real-time acquired grain boundary morphology categories, multi-scale texture features, and morphological characteristics, the grain boundary crack risk level assessment model was used to evaluate the grain boundary crack risk level. A tiered early warning system was implemented based on the assessment results. This included highlighting high-risk grain boundaries and generating process adjustment plans; marking and tracking medium-risk grain boundaries; and recording low-risk grain boundaries. The process adjustment plans included adjusting the local temperature gradient within the growth unit, regulating the pressure change rate within the growth unit, adjusting the cooling gradient and cooling rhythm inside and outside the furnace, establishing monitoring markers for high-risk merging areas, and triggering a shutdown mechanism. By using the indium phosphide crystal growth monitoring database, historical data on the geometric, moment, and contour sequence characteristics, multi-scale texture, and morphological characteristics of medium- and low-risk grain boundaries are obtained. A long short-term memory network is used to train the model and construct a grain boundary feature prediction model. This model predicts the geometric, moment, contour sequence, multi-scale texture, and morphological characteristics of grain boundaries within a preset time period. Combined with a grain boundary morphology category identification model and a grain boundary crack risk level assessment model, the model predicts the time point when the risk level of grain boundary cracks changes and provides corresponding early warning for the grain boundary.

[0047] For example, historical data covering 1000 different growth batches of grain boundary multi-scale texture features and grain boundary morphology features were extracted from the indium phosphide crystal growth monitoring database. This historical data, combined with the corresponding grain boundary morphology categories, was used to train a model using a random forest algorithm to construct a grain boundary crack risk assessment model, categorized as high, medium, and low. During the real-time monitoring phase, a new grain boundary signal was captured. This grain boundary, within a length of approximately 1.8 mm, was divided into 96 multi-scale sub-regions. The texture features of this grain boundary within a small-scale analysis window with a side length of 32 pixels included a mean contrast of 385, a standard deviation of 112, a maximum value of 910, a 90th percentile of 760, and a mean entropy of 3.62 with a maximum value of 4.05. This indicates significant texture inhomogeneity within a local area of ​​the grain boundary. The medium-scale and large-scale texture features of this grain boundary were then acquired. Meanwhile, the morphological characteristics of this grain boundary show that its mean curvature is 0.41, the standard deviation of the rate of change of curvature is 0.19, and the density of inflection points is 2.7 per millimeter. Using the grain boundary crack risk level assessment model, its grain boundary crack risk level is determined to be high. Based on the grain boundary crack risk level assessment results, a graded early warning process is implemented, including highlighting high-risk grain boundaries and generating process adjustment plans, marking and tracking medium-risk grain boundaries, and recording low-risk grain boundaries. The process adjustment plans include adjusting the local temperature gradient within the growth apparatus, regulating the pressure change rate within the growth apparatus, adjusting the cooling gradient and cooling rhythm inside and outside the furnace, changing the growth direction or the direction of gravity through a reciprocating flipping mechanism to improve the thermal stress distribution and thus change the stress state of the melt and crystal in the gravitational field, promoting convection uniformity within the melt, reducing the local temperature gradient, effectively alleviating the thermal and structural stress generated during crystal growth, setting monitoring marks for high-risk merging areas, and triggering a shutdown mechanism. The current risk level of grain boundary cracks is high. The specific process adjustment plan implemented is as follows: By adjusting the heating power and heating rate of the heating system, the overall temperature change process in the furnace is dynamically optimized, reducing the temperature gradient characterization value corresponding to the grain boundary region from approximately 25℃ / cm to approximately 22℃ / cm to alleviate the thermal stress concentration near the grain boundary. At the same time, by controlling the opening and closing status of the controllable electric valve in the nitrogen system, the pressure change rate in the furnace is reduced from 0.15MPa / min to 0.05MPa / min to suppress abnormal migration and unstable evolution of grain boundaries. In addition, the cooling system is coordinated to reduce the external cooling water flow rate from 12L / min to 6L / min, extending the slow cooling stage time, thereby reducing the risk of grain boundary merging and crack formation.Meanwhile, for grain boundaries classified as medium-risk or low-risk, their historical feature sequences from the past 60 minutes are retrieved for time-series modeling. If it is observed that their entropy value has been continuously increasing at a rate of 0.15 per 10 minutes in the past hour, and the rate of curvature change shows a non-linear upward trend, a long short-term memory network is used to train the model and construct a grain boundary feature prediction model. This model predicts the geometric description features, moment feature description features, contour sequence description features, multi-scale texture features, and grain boundary morphology features of the grain boundary within a preset time period. Combined with the grain boundary morphology category recognition model and the grain boundary crack risk level assessment model, it is predicted that the grain boundary will change from a medium-risk state to a high-risk state in the next 45 minutes, thus issuing an early warning signal in advance. This allows process engineers to take preventive measures such as temporarily reducing the raw material supply rate by 20% before the crack actually occurs.

[0048] Step S106: Continuously monitor and predict grain boundary migration data using an optical camera, identify grain boundary merging events within a preset time period, and determine the crack risk level after grain boundary merging.

[0049] Historical grain boundary migration data is continuously acquired in a time series using an optical camera. A long short-term memory (LSTM) network is used to train a model to construct a grain boundary migration prediction model. This model predicts grain boundary migration data within a predetermined time period. The migration data includes the centroid coordinates of the grain boundary, the overall displacement direction of the grain boundary, the migration velocity, the spacing between adjacent grain boundaries, and the intersection angle. If the predicted minimum spatial spacing, the opposite migration condition, and the intersection angle condition all meet the predetermined criteria within the predetermined time period, then a grain boundary merging event is determined to occur within that time period. If a grain boundary merging event occurs, the merging crack risk assessment formula is applied. Determine the crack risk index R after grain boundary merging, where, and These represent the grain boundary areas where grain boundary merging events occurred. and These are the grain boundary perimeters where the grain boundary merging event occurred. and Here, represents the entropy of the grain boundary where the grain boundary merging event occurred. and These represent the rate of change of grain boundary curvature during a grain boundary merging event. and These represent the grain boundary migration velocities at which grain boundary merging events occur. The geometric angle of grain boundary convergence at which a grain boundary merging event occurs. The grain boundary morphology merging coefficients, corresponding to the combinations of grain boundary morphology categories before merging, are obtained from a table of grain boundary morphology merging coefficients obtained by fitting historical crystal growth data. If the crack risk index after grain boundary merging is greater than the preset threshold, the crack risk level after grain boundary merging is judged to be high, triggering a high-level warning, generating a process adjustment plan, and making process adjustments in advance.

[0050] For example, an optical camera captured the continuous time series of grain boundary migration history data of two adjacent grain boundaries, grain boundary A and grain boundary B, which are approaching each other. The model was trained on the grain boundary migration history data through a long short-term memory network to build a grain boundary migration prediction model. The model predicts that in the next 30 minutes, the centroid coordinates of grain boundary A will move from the current (420, 310) to (435, 325), and the migration speed will evolve uniformly from 0.5 micrometers per minute. Meanwhile, grain boundary B will move in the opposite direction, causing the predicted minimum spatial distance to converge rapidly from 35 micrometers to 2 micrometers. The predicted geometric intersection angle between the two will decrease from 55° to 38°. When these predicted values ​​simultaneously meet the criteria of a distance of less than 5 micrometers, a positive cosine value of the directional angle, and angle convergence, it is determined that a grain boundary merging event will occur in the future. The predicted areas of grain boundaries A and B are 1250 and 1080 square pixels, respectively, with perimeters of 480 and 425 pixels. The entropy values ​​representing texture complexity are taken as the average of their sub-region aggregations: 4.2 and 3.9, respectively. The curvature change rates are 0.25 and 0.31, respectively, and the migration speeds are 0.62 and 0.58 micrometers per minute, respectively. The intersection angle is the predicted 38°. Furthermore, based on the grain boundary morphology categories of grain boundaries A and B before merging, the grain boundary morphology merging coefficient corresponding to the morphology category combination obtained from the grain boundary morphology merging coefficient table obtained by fitting historical crystal growth data is 1. According to the merging crack risk assessment formula... The calculated crack risk index R after grain boundary merging is 82.5, where, and These represent the grain boundary areas where grain boundary merging events occurred. and These are the grain boundary perimeters where the grain boundary merging event occurred. and Here, represents the entropy of the grain boundary where the grain boundary merging event occurred. and These represent the rate of change of grain boundary curvature during a grain boundary merging event. and These represent the grain boundary migration velocities at which grain boundary merging events occur. The geometric angle of grain boundary convergence at which a grain boundary merging event occurs. The grain boundary morphology merging coefficients, corresponding to the combinations of grain boundary morphology categories before merging, are obtained from a table of grain boundary morphology merging coefficients obtained by fitting historical crystal growth data. The crack risk index after grain boundary merging is 82.5, exceeding the preset threshold of 70. Therefore, the crack risk level of this merging event is determined to be high, immediately triggering a high-level red alert. The process parameters are automatically adjusted to reduce the local temperature gradient by 4°C to alleviate thermal stress. By controlling the opening and closing status of the controllable electric valve in the nitrogen system, the pressure change rate inside the furnace is reduced from 0.15MPa / min to 0.05MPa / min. The cooling system is also coordinated to reduce the external cooling water flow rate from 12L / min to 6L / min. These early intervention measures effectively avoid the crack risk caused by grain boundary merging.

[0051] Step S107: Based on the comparison results of the risk level of grain boundary cracks before and after the process adjustment, evaluate the effectiveness of the process adjustment plan, and update the process adjustment plan in combination with continuous monitoring feedback.

[0052] Obtain the grain boundary crack risk level before and after process adjustment. Evaluate the change in crack risk level after process adjustment to determine the effectiveness of the process adjustment plan. If the grain boundary crack risk level decreases, the process adjustment measures are deemed effective, and the process adjustment plan is saved as a valid process operation template. If the grain boundary crack risk level does not change or increases, optimize the process adjustment plan, including adjusting the temperature gradient of the crystal growth apparatus, optimizing the raw material supply rate, adjusting the crystal rotation speed, or modifying the grain boundary intersection angle. Continuously monitor the change in grain boundary crack risk level after implementing the optimized process adjustment plan, and continuously update the process adjustment plan based on real-time monitoring feedback until the grain boundary crack risk level is controlled within a preset safety range.

[0053] For example, in the closed-loop control of indium phosphide crystal growth, the effectiveness of the process parameter adjustment scheme is verified by comparing data before and after the adjustment. For instance, for a grain boundary predicted to have a high risk of grain boundary cracking, an initial adjustment scheme was implemented, reducing the local temperature gradient from 28℃ / cm to 22℃ / cm. Subsequently, feedback data obtained through real-time monitoring showed that the crack risk level of the grain boundary had decreased to medium. Based on this, the cooling measure was deemed effective, and this parameter combination was automatically saved as an effective process operation template for this type of large-angle grain boundary merging risk. For another grain boundary region where the risk level remained high after adjustments, the original solution was deemed ineffective, and an optimization program was immediately initiated. This involved dynamically optimizing the overall temperature change process within the furnace by adjusting the heating power and rate of the heating system. The temperature gradient characteristic value corresponding to the grain boundary region was reduced from approximately 22℃ / cm to approximately 18℃ / cm. Furthermore, by controlling the opening and closing of the controllable electric valve in the nitrogen system, the pressure change rate within the furnace was reduced from 0.05MPa / min to 0.04MPa / min, suppressing abnormal grain boundary migration and unstable evolution, and attempting to release local stress by slowing growth kinetics. Within 20 minutes of implementing the optimized process adjustment plan, continuous monitoring showed a significant decrease in key characteristics such as the average contrast and curvature abrupt change rate at this location. Combining the grain boundary morphology identification model and the grain boundary crack risk level assessment model, the real-time crack risk level was identified as low risk, successfully entering the preset medium / low risk safety range. If the real-time crack risk level is still identified as high risk, the process adjustment plan is continuously updated based on real-time monitoring feedback until the grain boundary crack risk level is controlled within the preset safety range.

[0054] The above description is merely a preferred embodiment of this application and an explanation of the technical principles employed. Those skilled in the art should understand that the scope of the invention involved in this application is not limited to technical solutions formed by specific combinations of the above-described technical features, but should also cover other technical solutions formed by arbitrary combinations of the above-described technical features or their equivalents without departing from the concept of this application. For example, technical solutions formed by substituting the above features with (but not limited to) technical features with similar functions disclosed in this application.

Claims

1. A method for controlling grain boundary morphology during indium phosphide crystal growth, characterized in that, The method includes: The original image data of the crystal surface is acquired by an optical camera, and scale remapping is completed by combining pixel size and spatial calibration parameters. Gray-level normalization and noise suppression are performed based on gray-level statistics and cumulative distribution mapping. Based on the noise-suppressed crystal surface image, gradient feature images are obtained using gradient operators, and geometric description features, moment feature description features, and contour sequence description features of connected domains are extracted. Based on the geometric description features, moment feature description features, and contour sequence description features of connected domains, a grain boundary morphology category identification model is constructed to identify the grain boundary morphology category to which the corresponding grain boundary region belongs. A multi-scale sliding window is used to divide the local neighborhood of the grain boundary, and the gray-level co-occurrence matrix and gray-level run matrix are calculated in each scale sub-region of the grain boundary to extract multi-scale texture features and grain boundary morphology features. Based on the multi-scale texture characteristics and morphological characteristics of grain boundaries, a grain boundary crack risk assessment model is constructed to assess the risk level of grain boundary cracks and predict the risk of grain boundary cracks within a preset time period. By continuously monitoring and predicting grain boundary migration data using an optical camera, grain boundary merging events can be identified within a preset time period in the future, and the crack risk level after grain boundary merging can be determined. Based on the comparison of the risk levels of grain boundary cracks before and after the process adjustment, the effectiveness of the process adjustment plan is evaluated, and the process adjustment plan is updated in conjunction with continuous monitoring feedback.

2. The method according to claim 1, wherein, The process of acquiring raw image data of the crystal surface through an optical camera, performing scale remapping by combining pixel size and spatial calibration parameters, and performing grayscale normalization and noise suppression processing based on grayscale statistics and cumulative distribution mapping includes: An optical camera installed inside the indium phosphide crystal growth apparatus acquires raw image data of the crystal surface in real time. Using a unified image encoding format, bit depth parameters, and color channel configuration, the raw image data is normalized. Scale remapping is then performed by reading the pixel dimensions and spatial calibration parameters of the optical camera to obtain standardized crystal surface image data with consistent spatial scale. Gray-scale statistical methods are used to perform gray-scale statistics, cumulative distribution calculation, and mapping transformation on each image to obtain a gray-scale standardized crystal surface image. Based on the gray-scale standardized crystal surface image, multiple rounds of median filtering are performed, recording the changes in noise points before and after each round of filtering. After stabilization, anisotropic diffusion filtering is performed, calculating the diffusion coefficient based on the gradient magnitude and iteratively updating the pixel values ​​to obtain noise-suppressed image data.

3. The method according to claim 1, wherein, The crystal surface image after noise suppression is used to obtain gradient feature images using gradient operators, and geometric description features, moment feature description features, and contour sequence description features of connected components are extracted, including: Based on the noise-suppressed crystal surface image, the first-order partial derivatives in the horizontal and vertical directions are calculated using the gradient operator. The gradient magnitude matrix and gradient direction matrix are synthesized to obtain the gradient feature image. A multi-directional filter kernel is used to perform convolution operations on the gradient feature image, recording the response intensity in each direction. Based on the angle between the principal gradient direction and the filter kernel direction, the response contribution is determined and the gradient response intensity is calculated. By comparing the gradient response intensity with a preset response intensity threshold, continuous high-response regions are identified, and an initial binary image of the grain boundary contour is obtained. A connected component labeling algorithm is used to group adjacent edge pixels into regions, determining the set of pixel coordinates for each connected component, and calculating the average of all pixel coordinates in the connected component to obtain the centroid coordinates of each connected component. The extension length is determined by calculating the aspect ratio of the minimum bounding rectangle of the connected component, and the compactness formula is used. Determine the compactness of connected components. The ratio of the area of ​​the connected region to the area of ​​its smallest bounding rectangle is calculated to determine the rectangularity. The area and perimeter are then combined to form the geometric description feature of the connected region, where A is the area of ​​the connected region and P is the perimeter of the connected region. The second and third central moments of the connected region are calculated, the central moments are normalized, and seven Hu invariant moments are obtained based on the normalized moments as moment feature description features. By calculating the approximate curvature of each point on the contour of the connected domain, calculating its average, variance and maximum values, and performing a discrete Fourier transform on the contour point sequence, the amplitude of the previously preset number of low-frequency coefficients is taken as the contour sequence description features, and the obtained geometric description features, moment feature description features and contour sequence description features of the connected domain are stored in the indium phosphide crystal growth monitoring database.

4. The method according to claim 1, wherein, The process involves constructing a grain boundary morphology category recognition model based on the geometric description features of connected domains, the moment feature description features, and the contour sequence description features. This model identifies the grain boundary morphology category to which the grain boundary region corresponding to each connected domain belongs, including: Historical data on the geometric, moment, and contour sequence description features of connected domains were obtained from the indium phosphide crystal growth monitoring database. The grain boundary morphology category of each connected domain was labeled, and a random forest algorithm was used to train the model to construct a grain boundary morphology category recognition model. The grain boundary morphology categories include small-angle grain boundaries, large-angle grain boundaries, twin boundaries, pseudo-grain boundaries, and grain boundary network nodes. Among them, pseudo-grain boundaries are connected domains that are caused by surface texture, processing marks, or noise and do not meet the criteria for true grain boundaries. Based on the geometric, moment, and contour sequence description features of connected domains extracted in real time from crystal surface images, the grain boundary morphology category recognition model was used to identify the grain boundary morphology category of the corresponding grain boundary region of each connected domain, and connected domains identified as pseudo-grain boundaries were removed.

5. The method according to claim 1, wherein, The process employs a multi-scale sliding window to divide the local neighborhood of the grain boundary, and within each scale sub-region of the grain boundary, calculates the gray-level co-occurrence matrix and gray-level run-length matrix to extract multi-scale texture features and grain boundary morphology features, including: Based on the set of pixel coordinates for each connected domain, the grain boundary positions are marked using pixel coordinate mapping, and the corresponding set of grain boundary region image blocks is extracted to obtain grain boundary region index data. Using the grain boundary line as the center, along its normal direction, a multi-scale sliding window is used to divide the local neighborhood of the grain boundary within a preset scale range, constructing a grain boundary sub-region dataset at each scale and recording the corresponding scale identification information. Within each scale grain boundary sub-region, the gray-level co-occurrence matrix and gray-level run length matrix are calculated, and texture features extracted from the same grain boundary in different scale sub-regions are extracted, including contrast, homogeneity, energy, entropy, correlation, directional consistency, and run length non-uniformity. The mean, standard deviation, maximum value, and quantile of the texture features obtained for the same grain boundary in different scale sub-regions are calculated respectively as multi-scale texture features of the grain boundary. By fitting the curvature of the grain boundary contour edge point sequence, the local curvature, rate of change of curvature, and inflection point distribution density are calculated as grain boundary morphological features. The multi-scale texture features and grain boundary morphological features are stored in the indium phosphide crystal growth monitoring database.

6. The method according to claim 1, wherein, The process involves constructing a grain boundary crack risk level assessment model based on multi-scale texture features and grain boundary morphology characteristics, assessing the grain boundary crack risk level, and predicting the grain boundary crack risk within a preset time period, including: Historical data on multi-scale texture features and morphological features of grain boundaries, as well as corresponding grain boundary morphology categories, were obtained from the indium phosphide crystal growth monitoring database. The crack risk of each grain boundary was labeled, and a grain boundary crack risk level assessment model was constructed using the random forest algorithm. The crack risk levels include high, medium, and low. Based on the real-time acquired grain boundary morphology categories, multi-scale texture features, and morphological features, the grain boundary crack risk level assessment model was used to evaluate the grain boundary crack risk level. Based on the risk level assessment results of grain boundary cracks, a graded early warning process is implemented. This includes highlighting high-risk grain boundaries and generating process adjustment plans, marking and tracking medium-risk grain boundaries, and recording low-risk grain boundaries. The process adjustment plans include adjusting the local temperature gradient within the growth unit, regulating the pressure change rate within the growth unit, adjusting the cooling gradient and cooling rhythm inside and outside the furnace, setting monitoring markers for high-risk merging areas, and triggering a shutdown mechanism. Through the indium phosphide crystal growth monitoring database, historical data on the geometric, moment, and contour sequence characteristics, multi-scale texture, and grain boundary morphology of medium- and low-risk grain boundaries are obtained. A long short-term memory network is used to train the model, constructing a grain boundary feature prediction model to predict the geometric, moment, contour sequence, multi-scale texture, and grain boundary morphology characteristics of grain boundaries within a preset time period. Combined with the grain boundary morphology category recognition model and the grain boundary crack risk level assessment model, the time point for the grain boundary crack risk level transition is predicted, and corresponding early warning measures are taken for the grain boundary in advance.

7. The method according to claim 1, wherein, The method of continuously monitoring and predicting grain boundary migration data using an optical camera, identifying grain boundary merging events within a preset time period, and determining the crack risk level after grain boundary merging includes: By continuously acquiring historical data of grain boundary migration over a continuous time series using an optical camera, a grain boundary migration prediction model is constructed using a long short-term memory network to predict the migration data of grain boundaries within a preset time period in the future. The migration data includes the centroid coordinates of the grain boundary, the overall displacement direction of the grain boundary, the migration velocity, the spacing between adjacent grain boundaries, and the intersection angle. If the minimum spatial spacing, the opposite migration condition, and the intersection angle condition of the predicted grain boundaries all meet the preset judgment conditions within the preset time period in the future, then it is determined that a grain boundary merging event exists within the preset time period in the future. If grain boundary merging events occur, then the merging crack risk assessment formula should be used. Determine the crack risk index R after grain boundary merging, where, and These represent the grain boundary areas where grain boundary merging events occurred. and These are the grain boundary perimeters where the grain boundary merging event occurred. and Here, represents the entropy of the grain boundary where the grain boundary merging event occurred. and These represent the rate of change of grain boundary curvature during a grain boundary merging event. and These represent the grain boundary migration velocities at which grain boundary merging events occur. The geometric angle of grain boundary convergence at which a grain boundary merging event occurs. The grain boundary morphology merging coefficient corresponding to the combination of grain boundary morphology categories before merging is obtained from the grain boundary morphology merging coefficient table obtained by fitting historical crystal growth data; if the crack risk index after grain boundary merging is greater than the preset index threshold, the crack risk level after grain boundary merging is judged to be high, triggering a high-level warning, generating a process adjustment plan, and making process adjustments in advance.

8. The method according to claim 1, wherein, The effectiveness of the process adjustment plan is evaluated based on the comparison of grain boundary crack risk levels before and after the process adjustment, and the plan is updated in conjunction with continuous monitoring feedback, including: Obtain the risk level of grain boundary cracks before and after process adjustment, and determine the effectiveness of the process adjustment plan by evaluating the change in crack risk level after process adjustment; If the risk level of grain boundary cracks is reduced, the process adjustment measures are deemed effective, and the process adjustment plan is saved as a valid process operation template. If the risk level of grain boundary cracks does not change or increases, the process adjustment plan is optimized, including adjusting the temperature gradient of the crystal growth device, optimizing the raw material supply rate, adjusting the crystal rotation speed, or modifying the grain boundary intersection angle. The changes in the risk level of grain boundary cracks after implementing the optimized process adjustment plan are continuously monitored, and the process adjustment plan is continuously updated based on real-time monitoring feedback until the risk level of grain boundary cracks is controlled within the preset safe range.

Images