Crucible life detection method based on thermal microvariation microscopic substance evolution and device thereof

Abstract

The present application belongs to the field of quartz crucible crystal pulling production and service performance evaluation, discloses a kind of crucible life detection method based on thermal micro change micro matter evolution, this method takes transparent layer micro matter as detection object, specifically with micro bubble as the representative of micro matter to detect;Collect the focal plane image sequence of the same transparent layer micro area of quartz crucible sample to be measured before and after high temperature treatment;Each focal plane image is input into the improved SOLOv2 instance segmentation network, and a two-dimensional micro bubble instance mask is obtained;Three-dimensional micro bubble binary voxel mask is constructed by cross-focal plane micro bubble mask association, micro bubble wall breaking risk model and micro bubble fusion embrittlement model are constructed based on three-dimensional micro bubble parameters, critical breaking time and abnormal evolution evolution index are obtained respectively, and crucible life prediction is carried out;The present application also discloses a kind of crucible life detection device based on thermal micro change micro matter evolution.The present application provides data support for quartz crucible life evaluation and crystal pulling process optimization.

Description

Technical Field

[0001] This invention belongs to the field of quartz crucible crystal pulling production and service performance evaluation, specifically involving a crucible life detection method and device based on thermal micro-variation micro-material evolution. Background Technology

[0002] In the Czochralski process for growing single-crystal silicon, the transparent layer on the inner surface of the quartz crucible undergoes microscopic material and structural evolution under high temperatures. Microbubbles are a crucial manifestation of the micromaterials within the transparent layer; their size, quantity, morphology, and spatial distribution affect the crucible's resistance to crystallization, high-temperature structural strength, and service stability. The size of these microbubbles ranges from 10 to 100 micrometers. Under the influence of viscous flow, gas diffusion, and interfacial tension in the quartz material, these microbubbles dynamically change in size, morphology, and spatial distribution during the high-temperature treatment process. Continuous expansion, fusion, or microbubbles approaching the inner wall of the crucible can reduce local structural stability and affect the crucible's service life. Therefore, accurate collection and quantitative comparison of the micromaterial states, represented by microbubbles, within the same transparent layer microregion before and after high-temperature treatment are fundamental for studying the thermal evolution of transparent layer micromaterials, evaluating crucible quality, and optimizing single-crystal silicon pulling processes. With the photovoltaic and semiconductor industries continuously increasing their demands for single-crystal silicon quality, the industry is placing higher requirements on the imaging accuracy, analytical accuracy, and physical interpretability of micromaterial detection in the transparent layer of quartz crucibles.

[0003] Patent CN121304911A discloses a multi-view three-dimensional reconstruction and dynamic analysis system for microbubble swarms in quartz crucibles. This system collects microbubble data using a combination of optical and acoustic transmission imaging devices, constructs a three-dimensional model of the microbubbles, extracts feature parameters, and calculates microbubble growth rate, identifies aggregation and breakage behavior, and provides safety warnings based on time-series data. While this system achieves dynamic detection and early warning of microbubble swarms, it still has significant technical shortcomings. First, the system focuses on detecting the overall distribution of microbubble swarms in the crucible, lacking a mechanism for repeated localization and imaging resetting of the same micro-area, and therefore cannot obtain matching microbubble image pairs before and after high-temperature treatment. Second, it cannot effectively handle detection scenarios involving transparent layer microbubbles with large size spans, weak boundaries, and dense adhesion. Third, relying solely on geometric parameter statistics and threshold warnings, it fails to achieve a quantitative correlation between the detection results and the high-temperature physical evolution mechanism of quartz, limiting its detection accuracy and technical applicability.

[0004] In summary, existing micromaterial detection technologies for the transparent layer of quartz crucibles still suffer from problems such as insufficient micro-area positioning and retesting capabilities, poor precision segmentation of microbubbles in the transparent layer, and insufficient quantitative analysis capabilities of physical evolution, making it difficult to meet the actual needs of high-precision crucible quality evaluation and lifespan testing. Summary of the Invention

[0005] The technical problem to be solved by the present invention is to provide a crucible lifetime detection method based on thermal micro-variation micro-material evolution, which addresses the shortcomings of the prior art.

[0006] To solve the above-mentioned technical problems, the technical solution adopted by the present invention is: a crucible lifetime detection method based on thermal micro-variation micro-material evolution, comprising the following steps: S1. Sequentially acquire imaging focal planes along the optical axis of the industrial camera, continuously acquiring focal plane image sequences of the transparent layer micro-region of the quartz crucible sample before and after high-temperature treatment; high-temperature treatment refers to simulated crystal pulling high-temperature treatment. S2. Input the images of each focal plane before and after high-temperature processing into the improved SOLOv2 instance segmentation network to complete the identification and segmentation of two-dimensional microbubbles in the transparent layer micromaterials within each focal plane, output the mask of the two-dimensional microbubbles in each focal plane, and extract the quantitative feature parameters of each two-dimensional microbubble. S3. Based on the masks of two-dimensional microbubbles in each focal plane and quantitative feature parameters, according to the focal plane image acquisition sequence, with the current focal plane and the next focal plane as the matching unit, for each two-dimensional microbubble mask in the current focal plane, calculate its overlap coefficient, normalized centroid distance and normalized contour distance with each two-dimensional microbubble mask in the next focal plane. Mask pairs with an overlap coefficient greater than 0 and a normalized centroid distance less than or equal to 1 are identified as candidate association pairs. The association score of each candidate association pair is calculated. For each two-dimensional microbubble mask in the current focal plane, the mask with the highest association score is selected from its corresponding candidate association pairs to complete the pairing. A single two-dimensional microbubble mask is matched with only one corresponding mask. Two-dimensional microbubble masks that are matched and associated sequentially in consecutive focal planes are classified into the same three-dimensional microbubble, and a three-dimensional binary voxel mask corresponding to the three-dimensional microbubble is constructed.

[0007] S4. Based on the three-dimensional binary voxel mask of each three-dimensional microbubble, calculate the geometric feature parameters of the three-dimensional microbubble; the geometric feature parameters of the three-dimensional microbubble include the volume of the three-dimensional microbubble, the equivalent radius of the three-dimensional microbubble, the centroid of the three-dimensional microbubble, and the remaining distance from the outer boundary of the three-dimensional microbubble to the inner wall of the crucible; S5. Based on the geometric feature parameters of three-dimensional microbubbles, two types of quantitative models for integrating the physical mechanism of high-temperature quartz are constructed, namely the microbubble wall fracture risk model and the microbubble fusion embrittlement model. Based on the viscous flow of high-temperature quartz, the evolution of internal pressure of microbubbles, and the effect of interfacial tension, a microbubble wall-joining rupture risk model is constructed to calculate the effective driving pressure difference of each three-dimensional microbubble. When the effective driving pressure difference is >0, the critical rupture time of the corresponding three-dimensional microbubble is calculated; otherwise, it is marked as having no expansion wall-joining tendency, and its critical rupture time is not calculated. Based on the Ostwald ripening and vacancy diffusion mechanism, a microbubble fusion embrittlement model is constructed to calculate the abnormal evolution index of three-dimensional microbubbles. S6. When the minimum critical rupture time of the three-dimensional microbubbles in the whole field is less than or equal to the expected single crystal pulling time, the crucible life is determined to be insufficient and a rupture risk warning is generated; when the actual ripening rate is greater than the ripening rate constant, the microbubble evolution rate is determined to be abnormal and a matrix quality abnormality warning is generated.

[0008] This invention also discloses a crucible lifetime detection device based on thermal micro-variation micro-matter evolution, comprising: The image acquisition module is used to sequentially acquire imaging focal plane images along the optical axis of the industrial camera, continuously acquiring a sequence of focal plane images of the transparent layer micro-region of the quartz crucible sample before and after high-temperature treatment; high-temperature treatment refers to the simulated crystal pulling high-temperature treatment. The two-dimensional microbubble extraction module is used to input the images of each focal plane before and after high-temperature treatment into the improved SOLOv2 instance segmentation network to complete the identification and segmentation of two-dimensional microbubbles in the transparent layer micromaterials within each focal plane, output the mask of the two-dimensional microbubbles in each focal plane, and extract the quantitative feature parameters of each two-dimensional microbubble. The 3D microbubble construction module is used to construct a 3D binary voxel mask for each 2D microbubble in the next focal plane based on the masks and quantitative feature parameters of the 2D microbubble in each focal plane, according to the focal plane image acquisition sequence and using the current focal plane and the next focal plane as matching units. For each 2D microbubble mask in the current focal plane, the overlap coefficient, normalized centroid distance, and normalized contour distance with each 2D microbubble mask in the next focal plane are calculated. Mask pairs with an overlap coefficient greater than 0 and a normalized centroid distance less than or equal to 1 are identified as candidate association pairs, and the association score of each candidate association pair is calculated. For each 2D microbubble mask in the current focal plane, the mask with the highest association score is selected from its corresponding candidate association pairs to complete the pairing. A single 2D microbubble mask is matched with only one corresponding mask. The 2D microbubble masks that are matched and associated sequentially in consecutive focal planes are classified into the same 3D microbubble, and the corresponding 3D microbubble is constructed. The geometric feature calculation module is used to calculate the geometric feature parameters of three-dimensional microbubbles based on the three-dimensional binary voxel mask of each three-dimensional microbubble. The geometric feature parameters of three-dimensional microbubbles include the volume of the three-dimensional microbubble, the equivalent radius of the three-dimensional microbubble, the centroid of the three-dimensional microbubble, and the remaining distance from the outer boundary of the three-dimensional microbubble to the inner wall of the crucible. The physical model construction module is used to construct two types of quantitative models integrating the physical mechanisms of high-temperature quartz based on the geometric characteristic parameters of three-dimensional microbubbles: a microbubble wall-to-wall rupture risk model and a microbubble fusion embrittlement model. Based on the viscous flow of high-temperature quartz, the evolution of internal pressure of microbubbles, and the effect of interfacial tension, the microbubble wall-to-wall rupture risk model is constructed to calculate the effective driving pressure difference of each three-dimensional microbubble. When the effective driving pressure difference is >0, the critical rupture time of the corresponding three-dimensional microbubble is calculated; otherwise, it is marked as having no expansion wall-to-wall tendency, and its critical rupture time is not calculated. Based on the Ostwald ripening and vacancy diffusion mechanism, the microbubble fusion embrittlement model is constructed to calculate the abnormal evolution index of three-dimensional microbubbles. The early warning module is used to determine that the crucible life is insufficient and generate a risk warning when the minimum critical rupture time of the three-dimensional microbubbles in the whole field is less than or equal to the expected single crystal pulling time; and to determine that the microbubble evolution rate is abnormal and generate a matrix quality abnormality warning when the actual ripening rate is greater than the ripening rate constant.

[0009] Compared with the prior art, the present invention has the following advantages: 1) By acquiring focal plane images before and after high-temperature treatment, spatial consistency of data before and after thermal evolution is ensured. Through preset positioning benchmarks and pose reproduction of a three-axis motion mechanism, the focal plane image sequence of the same transparent layer micro-region before and after high-temperature treatment is accurately acquired, providing a reliable data foundation for quantitative comparative analysis of the thermal evolution of transparent layer micro-materials represented by microbubbles.

[0010] 2) An improved SOLOv2 instance segmentation network is adopted. Spatial adaptive phase downsampling reduces the loss of weak boundary information in the early downsampling process. The spatial representation of microbubbles is enhanced by a high-resolution P2 layer feature routing strategy. Instance-level residual correction is performed on the uncertain boundary areas in the coarse mask through boundary-aware instance refinement. This effectively reduces microbubble missed detection, edge jaggedness, local offset and contour defects, and improves the accuracy of microbubble instance segmentation and subsequent morphological parameter calculation.

[0011] 3) By matching and associating two-dimensional microbubble masks across the focal plane, a three-dimensional microbubble voxel model is constructed and geometric parameters such as three-dimensional volume, equivalent radius, and spatial distance are directly calculated. This method abandons the traditional approach of approximating three-dimensional parameters from a single two-dimensional image, avoids the multiple solutions of direct two-dimensional conversion, and reduces the calculation error of three-dimensional parameters, thus providing accurate geometric input for subsequent physical quantization models.

[0012] 4) The microbubble wall rupture risk model takes the high-temperature quartz viscous flow law and gas internal pressure driving mechanism as the core, transforms the temperature dependence of quartz dynamic viscosity and the static geometric distance parameter into a dynamic critical rupture time; it realizes the quantitative assessment of microbubble distance, obtains the specific time of microbubble wall rupture, and realizes the forward risk prediction based on evolution trend.

[0013] 5) The microbubble fusion embrittlement model introduces the Ostwald ripening theory rate as a benchmark: the actual ripening rate is inverted by the cubic difference of the average radius of the three-dimensional microbubbles before and after high-temperature treatment, and then a ratioized abnormal evolution index is formed with the theoretical ripening rate constant at the same temperature; by comparing with the normal value of thermodynamic theory, the normal thermodynamic evolution of microbubbles and the accelerated ripening caused by abnormal factors of crucible can be distinguished, the intrinsic quality defects of crucible matrix can be located, and a quantifiable diagnostic basis is provided for the optimization of crucible raw material formulation and blanking process.

[0014] 6) By quantitatively analyzing the thermal evolution characteristics of transparent layer micromaterials represented by microbubbles, and setting judgment conditions from two aspects, namely the risk of microbubble wall rupture and the risk of abnormal evolution of quartz matrix, risk warning signals are generated accordingly. The detection results can be directly used for quartz crucible life evaluation, quality grading and crystal pulling process optimization, meeting the actual needs of high-precision crucible quality evaluation and life risk management.

[0015] The technical solution of the present invention will be further described in detail below with reference to the accompanying drawings and embodiments. Attached Figure Description

[0016] Figure 1 This is a flowchart of the method of the present invention; Figure 2 This is a schematic diagram of the architecture of the improved SOLOv2 network of the present invention; Figure 3 This is a schematic diagram of the spatial adaptive phase downsampling module structure of the method of the present invention; Figure 4 This is a schematic diagram of the detailed module structure for the boundary sensing instance of the method of the present invention. Figure 5 This is a flowchart illustrating steps S5 and S6 of the method of the present invention; Figure 6 This is a comparison diagram between the algorithm of the present invention and existing algorithms; Figure 7 This is a structural block diagram of the device of the present invention; Figure 8 This is a schematic diagram of the data acquisition platform of the device of the present invention; Figure 9 This is a physical diagram of the device of the present invention.

[0017] Explanation of reference numerals in the attached figures: 1. Clamping platform; 2. X-axis linear module; 3. Y-axis linear module; 4. Z-axis lifting module; 5. Industrial camera; 6. Upper light source; 7. Lower light source; 8. PLC controller; 9. Microscope head assembly. Detailed Implementation

[0018] Example of a crucible lifetime detection method based on thermal micro-variation micro-matter evolution: The microstructure refers to the microscopic features represented by microbubbles within the transparent layer of the quartz crucible. These microbubbles are the specific detection targets for instance segmentation, three-dimensional parameter calculation, and life risk analysis using this method.

[0019] like Figures 1-6 As shown, the crucible lifetime detection method based on thermal micro-variation micro-matter evolution includes the following steps: S1. Sequentially acquire focal plane images along the optical axis 5 of the industrial camera, continuously capturing a sequence of focal plane images of the transparent layer micro-region of the quartz crucible sample before and after high-temperature treatment; high-temperature treatment refers to simulated crystal pulling high-temperature treatment. The focal plane image sequence is acquired using an intelligent vision acquisition platform. For example... Figure 7 As shown, the intelligent vision acquisition platform includes: a clamping platform 1, used to fix the sample to be tested after the quartz crucible has been cut, and to ensure the stability of the sample during the acquisition process; an X-axis linear module 2, used to drive the industrial camera 5 to move along the axial direction of the quartz crucible to achieve lateral positioning of the transparent layer detection area; a Y-axis linear module 3, used to drive the clamping platform 1 to move in the horizontal direction to achieve switching and precise positioning of different detection areas of the sample to be tested; and a Z-axis lifting module 4, used to drive the industrial camera 5 to move along the optical axis, and to sequentially change the imaging focal plane according to the preset starting focal plane position, the number of focal planes, and the moving distance of adjacent focal planes. The system achieves automatic acquisition of focal plane image sequences. An industrial camera 5 and a microscope lens assembly 9 are used to acquire focal plane images of transparent layer microbubbles, obtaining images that meet the requirements for microbubble instance segmentation and 3D parameter calculation. An upper light source 6 and a lower light source 7 provide coordinated illumination to the transparent layer detection area, improving the imaging contrast between the transparent layer microbubbles and the background. A PLC controller 8 controls the movement of the X-axis linear module 2, Y-axis linear module 3, and Z-axis lifting module 4, and controls the Z-axis stepping movement according to preset focal plane acquisition parameters, triggering the industrial camera 5 to complete image acquisition at each focal plane position. The entire acquisition platform adopts a three-axis linkage control method, with each motion module using closed-loop positioning control, enabling the industrial camera 5 to automatically and repeatedly position any detection area of ​​the transparent layer, providing a hardware foundation for image acquisition of the same micro-area before and after crystal pulling.

[0020] The specific acquisition process is as follows: The test sample is cut from the straight wall area, arc transition area, or bottom area of ​​the quartz crucible at a predetermined axial height and circumferential position. The test sample is a plate-shaped or arc-shaped specimen retaining the complete wall thickness of the inner wall, transparent layer, and outer wall of the crucible. Its observation surface is a radial-axial cross-section encompassing both the crucible wall thickness direction and the axial direction. The observation surface is ground, polished, and cleaned, maintaining a predetermined distance between the target micro-area of ​​the transparent layer and the cut edge of the sample. The test sample is fixed to the intelligent vision acquisition platform according to a predetermined positioning reference. The position and imaging distance of the industrial camera 5 are adjusted using the XYZ three-axis motion mechanism, ensuring that the optical axis of the industrial camera 5 is substantially perpendicular to the observation surface and aligned with the target micro-area of ​​the transparent layer. The acquisition control module records the sample positioning reference, XYZ three-axis coordinates, lighting conditions, and imaging parameters of the industrial camera 5. It also controls the Z-axis lifting module 4 to move along the optical axis of the industrial camera 5, according to the initial focal plane position. The number of focal planes K and the distance of movement between adjacent focal planes The system sequentially acquires a sequence of focal plane images before and after high-temperature treatment. After the sample completes the simulated crystal pulling high-temperature treatment, it is re-fixed in the same position and orientation as before high-temperature treatment, based on the aforementioned positioning reference. The acquisition control module automatically calls up the recorded three-axis coordinates, focal plane acquisition parameters, and industrial camera 5 imaging parameters to repeatedly acquire images of the same transparent layer target micro-area, thereby obtaining a sequence of focal plane images before and after high-temperature treatment with spatial position and focal plane correspondence.

[0021] S2. Input the images of each focal plane before and after high-temperature treatment into the improved SOLOv2 instance segmentation network to identify and segment two-dimensional microbubbles in the transparent layer micromaterial within each focal plane. Output the masks of the two-dimensional microbubbles in each focal plane and extract the quantitative feature parameters of each two-dimensional microbubble. The extracted quantitative feature parameters of the two-dimensional microbubbles include the contour, area, equivalent radius, centroid coordinates, and planar distance from the inner wall of the crucible.

[0022] In this embodiment, as Figure 2 As shown, the improved SOLOv2 instance segmentation network includes a ResNet-50 backbone network with an introduced spatial adaptive phase downsampling module, a feature pyramid network (FPN), a dynamic prediction head, and a boundary-aware instance refinement module; the aforementioned dynamic prediction head includes a category branch, a dynamic convolutional kernel branch, and a mask feature branch.

[0023] like Figure 2 As shown, Figure 2 The diagram in (a) shows the original SOLOv2 instance segmentation network structure. Figure 2 Block diagram (b) shows the backbone network structure of this embodiment. Figure 2 Block diagram (c) shows the feature pyramid network structure of this embodiment. Figure 2The block diagram in section (d) shows the structure of the dynamic prediction head and the boundary awareness instance refinement module in this embodiment.

[0024] in, Figure 2 In block diagram (b), the input image first undergoes 7×7 convolution with a stride of 2, batch normalization (BN), and rectified linear unit (RELU) activation. Then, a spatial adaptive phase downsampling module with a stride of 2 further reduces the spatial resolution of the feature map to 1 / 4 of the input image. The subsequent residual stage adopts the standard bottleneck residual structure of ResNet-50, sequentially outputting multi-level features C2, C3, C4, and C5. The spatial resolutions of C2, C3, C4, and C5 relative to the input image are 1 / 4, 1 / 8, 1 / 16, and 1 / 32 of the original image, respectively.

[0025] 1) The initial max pooling layer of the ResNet-50 backbone network is replaced by a spatial adaptive phase downsampling module to extract multi-level features of the input image and reduce the loss of weak boundary and local morphological information of small bubbles during early downsampling.

[0026] Specifically, the spatial adaptive phase downsampling module includes a standard pooling path and a phase-preserving path. The standard pooling path performs max pooling with a step size of 2 on the input features to obtain standard downsampling features. The phase-preserving path divides the input features into four spatial phase components through pixel rearrangement, calculates the selection weight of each spatial phase after grouping the feature channels, and selects the corresponding spatial phase feature from each group according to the selection weight. Boundary gating weights are generated based on the local response differences between different spatial phases, and the selected spatial phase features are injected into the standard downsampling features through controlled residual connections to output the phase-preserving downsampling features.

[0027] like Figure 3 As shown, the spatial adaptive phase downsampling module includes a standard pooling path and a phase-preserving path. For the input feature map... The standard pooling path uses a 3×3 max pooling operation with a stride of 2 to obtain standard downsampled features:

[0028] The phase-preserving path employs a doubling pixel rearrangement operation, which rearranges each 2×2 neighborhood in the input feature map into four spatial phase components. The spatial phase mentioned above represents the sampling position of different sub-pixels within a 2×2 neighborhood, and does not represent the optical phase. The input feature channels are divided into multiple channel groups, allowing different channel groups to select different spatial phases.

[0029] A phase selection descriptor is constructed based on the centered response, absolute deviation, and characteristic descriptors of the standard pooling path of the four spatial phase components. The phase weights are calculated using a lightweight phase selection network:

[0030] in, For phase selection networks, The temperature coefficient is Softmax. These are the soft selection weights corresponding to the four spatial phases.

[0031] To select a single spatial phase during forward propagation and preserve gradients during backward propagation, a pass-through estimation method is used to obtain hard phase selection weights:

[0032] in, Indicates one-hot encoding. This indicates that gradient calculation is stopped during backpropagation. Spatial phase numbering.

[0033] The selected phase feature is obtained based on the hard phase selection weight:

[0034] The spatial adaptive phase downsampling module further calculates the local phase contrast based on the difference between the maximum and minimum responses among the four spatial phase descriptors, and normalizes it according to the local feature amplitude. The normalized local phase contrast is then input into the boundary gating network, and the boundary gating weights are obtained after Sigmoid activation. This allows phase compensation to primarily act on weak boundary regions with significant spatial variations.

[0035] The difference between the selected phase feature and the standard downsampling feature is mapped through grouped 1×1 convolutions and injected into the standard downsampling path via residual connections under boundary gating weights. The output of the spatial adaptive phase downsampling module is:

[0036] in, For grouped 1×1 convolution mapping, For learnable residual scaling parameters, This module is used to limit the phase compensation amplitude. By preserving the standard max-pooling path and employing a controlled residual injection method, the network maintains the stability of pre-trained ResNet-50 features in the early stages of training and gradually learns the spatial phase response related to the weak boundaries of microbubbles during training. This module achieves feature dimensionality reduction while selectively preserving high-frequency details of the weak boundaries of microbubbles, effectively reducing edge information loss caused by early downsampling in the backbone network and laying a solid feature foundation for subsequent accurate segmentation. Figure 3 In multiphase grouping, G=8 indicates that the number of channel groups is 8; This represents the four spatial phase components.

[0037] 2) The aforementioned feature pyramid network is used to perform multi-scale fusion of the multi-level features output by the backbone network, outputting five different resolution levels of predicted feature maps from P2 to P6. A high-resolution P2 layer feature routing strategy is adopted, using the high-resolution P2 layer features as the main prediction level for small-scale two-dimensional microbubbles, and simultaneously accessing the category branch and the dynamic convolutional kernel branch. This achieves the preservation of high-resolution spatial information of the microbubbles.

[0038] Specifically, in the original SOLOv2 feature routing, the P2 layer features need to be downsampled by 0.5 times before entering the category branch and the dynamic convolution kernel branch, which increases the effective stride of the first prediction layer from 4 to 8, resulting in the loss of spatial information related to the position, outline and weak boundary of the microbubbles in the P2 layer features.

[0039] In this embodiment, the high-resolution P2 layer feature routing strategy cancels the 0.5x downsampling operation before the P2 layer features enter the dynamic prediction head, and directly uses the original spatial resolution P2 layer features with an effective stride of 4 as the input of the first prediction layer, and simultaneously inputs the category branch and the dynamic convolution kernel branch; the P3 layer features, P4 layer features, P5 layer features, and P6 layer features after double bilinear upsampling are used as the input of the second to fifth prediction layers, respectively, and the effective strides of the five prediction layers are 4, 8, 16, 32 and 32, respectively.

[0040] In this system, P2 layer features are directly input to the category branch and the dynamic convolution kernel branch to enhance the category prediction, center localization, and instance dynamic convolution kernel generation capabilities of microbubbles. P2 and P3 features are simultaneously input to the boundary-aware instance refinement module to construct high-resolution boundary detail features for microbubbles. The spatial adaptive phase downsampling module reduces boundary information loss caused by early downsampling in the backbone network, while the high-resolution P2 layer feature routing strategy transmits the preserved high-resolution spatial information to the microbubble instance prediction and boundary refinement stages. This setup directly strengthens the spatial representation capability of the first prediction layer, significantly improving the localization accuracy and contour restoration of ten-micrometer-level microbubbles, and reducing the probability of missed detection of small targets.

[0041] 3) The dynamic prediction head includes a category branch, a dynamic convolution kernel branch, and a mask feature branch. The category branch outputs the microbubble category results; the dynamic convolution kernel branch generates a corresponding instance dynamic convolution kernel for each candidate microbubble instance; the mask feature branch fuses features P2 to P5 to generate shared high-resolution mask features.

[0042] The aforementioned category branch performs category prediction on the features of each prediction level from P2 to P6, and outputs the microbubble category confidence score corresponding to each grid position. The aforementioned dynamic convolution kernel branch works in conjunction with the mask feature branch to generate a coarse two-dimensional microbubble instance mask; specifically, the dynamic convolution kernel branch generates a corresponding instance dynamic convolution kernel for each candidate microbubble instance; the mask feature branch fuses features from P2 to P5 to generate shared high-resolution mask features.

[0043] For the current focal plane, the th Let there be a two-dimensional microbubble instance, and let its corresponding instance dynamic convolution kernel be... Shared mask features are The Logit representation of the coarse-to-micro bubble instance mask is:

[0044] In the formula, This indicates a dynamic convolution operation. Indicates the first Logit for a coarse mask of a two-dimensional microbubble instance.

[0045] Because the coarse microbubble instance mask is generated from a shared mask feature with a fixed resolution, small-sized, weakly boundaryed, and locally adhered microbubbles may still exhibit problems such as jagged edges, local offsets, incomplete contours, or inaccurate separation of adjacent instances. Therefore, as... Figure 4 As shown, the present invention sets up a boundary-aware instance refinement module after the dynamic prediction head; this module includes three stages: boundary detail decoding, boundary-aware point selection, and instance condition residual refinement.

[0046] In the boundary detail decoding stage, convolutional modules containing 3×3 convolutions, group normalization, and corrected linear unit activation are used to process P2 and P3 features respectively. The processed P3 features are then upsampled twice to the spatial resolution of the P2 layer features and concatenated with the processed P2 layer features. The concatenated features are then fused through two 3×3 convolutional modules to obtain the boundary detail features. Boundary detail features After further 1×1 convolution, the boundary response map is obtained. Among them, boundary detail features Used to provide local texture and contour information, boundary response map Used to characterize the location where microbubble boundaries may exist.

[0047] For the For a two-dimensional microbubble instance, its initial refinement input is set as follows:

[0048] in, Logit represents the coarse-to-fine bubble instance mask generated by dynamic convolution. and These represent the mask Logit after the first and second boundary refinements, respectively.

[0049] In the In the next refinement process, the spatial positions are calculated by combining the current mask's uncertainty and boundary response. Priority of refinement:

[0050] in, For the first A two-dimensional microbubble instance in spatial location Prioritize the refinement of details; This is a Sigmoid function; the first term increases the priority of locations with a current prediction probability close to 0.5, and the second term increases the priority of locations with microbubble boundary responses. A preset number of spatial locations are selected according to their refinement priority from high to low to form the set of points to be refined. Superscript The first instance refinement module of the boundary awareness module The refinement stage does not represent exponentiation; in this embodiment... The value can be 0, 1, or 2.

[0051] For each location to be refined Point features are extracted from boundary detail features, the current mask, the current mask gradient, and the coordinates relative to the instance center:

[0052] in, For the first The center coordinates of a two-dimensional microbubble instance Indicates position Boundary detail features at the location, This represents the mask gradient at the current position. This represents the coordinate offset of the position to be refined relative to the center of the microbubble instance. For the first The first two-dimensional microbubble instance in the... In the next refinement stage, spatial location The feature vector at that point, For gradient operators, For the first The first two-dimensional microbubble instance in the... In the next refinement stage, position The mask logic value at that location.

[0053] The first The dynamic convolution kernel corresponding to each microbubble instance Input kernel embedding network to generate instance conditional modulation parameters:

[0054] Applying the above example conditional modulation parameters to the point feature encoding result:

[0055] in, For point feature encoding networks, For core embedded network, and They are respectively with the first The modulation scaling parameters and modulation bias parameters corresponding to each microbubble instance. This represents element-wise multiplication. By using instance-dynamic convolution kernels to conditionally modulate point features, adjacent microbubbles with similar local grayscale and boundary features can obtain different boundary correction results based on their respective instance identities.

[0056] The instance conditional point refinement network predicts the mask Logit residual based on the modulated point features and updates the position to be refined:

[0057] in, This is a mask Logit residual prediction network. Unselected locations retain the current mask Logit value, and residual correction is only applied to locations with uncertain boundaries but supported by boundary responses. This improves the local boundary quality of microbubble instances while controlling computational cost. This fixed-point refinement mechanism balances inference efficiency and segmentation accuracy, achieving differentiated boundary correction for adjacent adhered bubbles through instance conditional modulation, effectively solving the problems of jagged edges and incomplete contours in coarse masks.

[0058] During training, Focal loss is used for the category branch, Dice loss is used for the coarse and fine bubble instance masks, point-level binary cross-entropy loss is used for the points to be refined, and binary cross-entropy loss and Dice loss are used for the boundary response map. The overall loss function is:

[0059] in, For category loss, For coarse mask loss, For details regarding the loss, and These are the binary cross-entropy loss and Dice loss for the boundary response map, respectively. , , These are the weighting coefficients for the corresponding loss terms.

[0060] 4) The boundary-aware instance refinement module described above uses the features of layer P2, layer P3, coarse two-dimensional microbubble instance mask and corresponding dynamic convolution kernel to correct the uncertain regions of the mask boundary and output the refined two-dimensional microbubble instance mask.

[0061] The boundary-aware instance refinement module comprises three stages: boundary detail decoding, boundary-aware point selection, and instance condition residual refinement. In the boundary detail decoding stage, convolution processing, upsampling and feature fusion are performed on the features of P2 layer and P3 layer to obtain detail feature map and boundary response map; In the boundary sensing point selection stage, based on the uncertainty and boundary response map of the coarse two-dimensional microbubble instance mask, the boundary positions to be refined are screened from the coarse mask; In the instance conditional residual refinement stage, the detailed features, mask prediction values, mask gradients, and coordinate information relative to the center of the microbubble instance are extracted from the location to be refined. The instance conditional modulation parameters are generated by combining the corresponding dynamic convolution kernels, the mask residuals are predicted, and the mask values ​​at the location to be refined are updated.

[0062] The boundary-aware instance refinement module performs two progressive boundary corrections on the coarse mask. The first correction outputs the first refined mask logic value, and the second correction outputs the second refined mask logic value for the j-th instance. The final formula for calculating the mask probability of microbubble instances is:

[0063] in, It is the Sigmoid activation function. For the first The final mask probability map of each two-dimensional microbubble instance is shown, with the superscript (2) indicating the refinement stage number. The refined microbubble instance mask is then processed by mask quality scoring and matrix nonmaximum suppression to remove low-confidence and duplicate microbubble instances, resulting in the final two-dimensional microbubble instance segmentation results in each focal plane. This improved network has good adaptability to microbubbles with large size spans, low boundary contrast, and local dense adhesion. The output two-dimensional mask is more in line with the real bubble morphology, effectively improving the accuracy of subsequent morphological parameter extraction.

[0064] Figure 4 In this context, Logit represents the current mask; Top-K coordinates to be refined represent the coordinates of a preset number of points to be refined, selected according to priority.

[0065] The improved SOLOv2 instance segmentation network proposed in this invention reduces the loss of weak boundary information of microbubbles caused by early downsampling of the backbone network through a spatial adaptive phase downsampling module, enhances the spatial localization of microbubbles and the ability to generate instance dynamic convolutional kernels through a high-resolution P2 layer feature routing strategy, and performs instance-level residual correction on the boundary uncertainty region in the coarse microbubble mask through a boundary-aware instance refinement module. This improves the network's ability to identify and segment microbubbles with large size spans, weak boundary contrast, local dense adhesion, and irregular shape.

[0066] After automatically identifying and segmenting microbubble instances in each focal plane, corresponding two-dimensional microbubble instance masks are obtained, and parameters such as the contour, area, two-dimensional equivalent radius, and centroid coordinates of each microbubble are extracted. Simultaneously, the focal plane number and Z-axis position corresponding to each instance mask are recorded. The instance segmentation results of each focal plane before and after high-temperature processing are input into the focal plane 3D reconstruction module, providing data for subsequent cross-focal plane microbubble association and 3D geometric feature parameter calculation.

[0067] It should be noted that the symbols in the formula of step S2 are only applicable to the relevant calculations of the improved SOLOv2 instance segmentation network.

[0068] Figure 2 In this context, ResNet represents the residual network; P2 to P6 represent the multi-level features output by the feature pyramid network; and NMS represents matrix nonmaximum suppression.

[0069] To verify the performance of the improved SOLOv2 instance segmentation network proposed in this invention, the improved SOLOv2 instance segmentation network was compared with other instance segmentation models under the same training data, test data, and evaluation criteria. The comparison results are as follows: Figure 6 As shown, Figure 6 In the diagram, the icons Miss Instance represent missed instances, Lacked EdgeDetail represents missing edge details, Incomplete Instance represents incomplete instance segmentation, and False Instance represents false positives. 0h, 100h, 300h, and 500h represent different high-temperature processing durations. Original represents the original image, and Ours represents the improved SOLOv2 instance segmentation result of this invention. Other instance segmentation models include Mask R-CNN (region-based convolutional neural network masking model), Cascade Mask R-CNN, YOLACT (YouOnly Look At Coefficients), Conditional Convolutional Instance Segmentation, SOLO, and SOLOv2 models. Figure 6As can be observed, the images processed by the improved SOLOv2 instance segmentation network in this embodiment showed no missed detections, missing edge details, incomplete instance segmentation, or false detections during each high-temperature processing stage. The results demonstrate a significant advantage under conditions of small-sized microbubbles and high cross-section over union (COB) ratio, indicating that it can more accurately identify tiny microbubbles in the transparent layer and obtain instance masks that better fit the real boundaries and have higher integrity. Therefore, the algorithm of this invention has stronger adaptability to complex imaging situations such as weak boundaries, large size spans, and local dense adhesion of microbubbles in the transparent layer. Its comprehensive detection and segmentation performance is superior to other comparative models, providing more reliable basic data for subsequent calculation of microbubble 3D geometric parameters and crucible life risk assessment.

[0070] S3. Because the instance segmentation network outputs two-dimensional instance masks of microbubbles on different focal planes, but the subsequent model judgment uses relevant data of three-dimensional microbubbles, it is necessary to perform cross-focal plane association of two-dimensional microbubble masks on different focal planes and obtain the three-dimensional geometric parameters of microbubbles through three-dimensional voxel reconstruction.

[0071] Specifically, based on the masks of two-dimensional microbubbles in each focal plane and quantitative feature parameters, according to the focal plane image acquisition sequence, with the current focal plane and the next focal plane as the matching unit, for each two-dimensional microbubble mask in the current focal plane, the overlap coefficient, normalized centroid distance and normalized contour distance with each two-dimensional microbubble mask in the next focal plane are calculated respectively.

[0072] In this embodiment, the focal plane image sequence is assumed to contain K focal planes, and the first... The Z-axis position of each focal plane is:

[0073] in, The initial focal plane position, The distance traveled between two consecutive Z-axis acquisitions is directly obtained from the position encoder of the Z-axis lifting module.

[0074] When the optical axis of the industrial camera 5 is substantially perpendicular to the sample observation surface, the axial position of the focal plane is corrected based on the refractive index of the quartz material. Under paraxial imaging conditions, the actual distance between adjacent focal planes inside the sample is:

[0075] in, For quartz materials at imaging wavelengths The refractive index below, denoted as , where is the refractive index of the medium between the camera and the sample.

[0076] To ensure that the two-dimensional microbubble masks at different focal planes can characterize the effective cross-section of the microbubble at the corresponding depth, the movement distance between adjacent focal planes should not exceed half of the axial resolution of the imaging system. Sharpness calculations are performed on the images at each focal plane, retaining the microbubble mask region within the effective depth of field of the current focal plane while suppressing out-of-focus areas.

[0077] Specifically, the formula for calculating the overlap coefficient is:

[0078] in, For the first focal plane center The two-dimensional microbubble and the first In the focal plane, the th The mask overlap factor of a two-dimensional microbubble. For the first focal plane center A collection of two-dimensional microbubble mask regions For the first focal plane center A collection of two-dimensional microbubble mask regions. To perform the minimum value operation, This is a function for calculating the area of ​​a region, used to determine the pixel area of ​​the corresponding mask region. This is the intersection operation of sets.

[0079] The formula for calculating the normalized centroid distance is:

[0080] in, For the first focal plane center The two-dimensional microbubble and the first In the focal plane, the th The normalized centroid distance of a two-dimensional microbubble For the first focal plane center The centroid of a two-dimensional microbubble mask For the first focal plane center The centroid of a two-dimensional microbubble mask For the first focal plane center The equivalent radius of a two-dimensional microbubble mask, For the first focal plane center The equivalent radius of a two-dimensional microbubble mask, It is a two-dimensional Euclidean norm.

[0081] The formula for calculating the normalized profile distance is:

[0082] in, For the first focal plane center The two-dimensional microbubble and the first In the focal plane, the th The normalized profile distance of a two-dimensional microbubble. This refers to the Hausdorff distance operator. For the first In the focal plane, the th The outline of a two-dimensional microbubble For the first In the focal plane, the th The outline of a two-dimensional microbubble.

[0083] In this embodiment, the formula for calculating the association score is:

[0084] in, For the first focal plane center The two-dimensional microbubble and the first In the focal plane, the th The correlation score of a two-dimensional microbubble. It is a natural exponential function.

[0085] In this embodiment, mask pairs with an overlap coefficient greater than 0 and a normalized centroid distance less than or equal to 1 are identified as candidate association pairs, and an association score is calculated for each candidate association pair. The association score is then used to determine the association score. A one-to-one matching process is employed, starting from the largest to the smallest. Each 2D microbubble mask is associated with only one 2D microbubble mask in an adjacent focal plane. Specifically, for each 2D microbubble mask within the current focal plane, the mask with the highest association score is selected from its corresponding candidate association pairs to complete the pairing. The remaining candidate masks are then used as new microbubbles for subsequent focal plane associations. Each 2D microbubble mask is matched with only one corresponding mask. 2D microbubble masks that are sequentially matched and associated in consecutive focal planes are grouped into the same 3D microbubble, constructing a 3D binary voxel mask for the corresponding 3D microbubble. This multi-dimensional matching association strategy effectively reduces the probability of mismatches and omissions across focal planes, ensuring the topological correctness of the 3D voxel model and providing a reliable structural foundation for subsequent 3D parameter calculations.

[0086] S4. Based on the three-dimensional binary voxel mask of each three-dimensional microbubble, calculate the geometric feature parameters of the three-dimensional microbubble; the geometric feature parameters of the three-dimensional microbubble include the volume of the three-dimensional microbubble, the equivalent radius of the three-dimensional microbubble, the centroid of the three-dimensional microbubble, and the remaining distance from the outer boundary of the three-dimensional microbubble to the inner wall of the crucible.

[0087] In this embodiment, for the first A three-dimensional microbubble, whose three-dimensional binary voxel mask is defined as:

[0088] in, These are the pixel coordinates within the focal plane. Number the focal plane. Therefore, only when the pixel Located within an effective two-dimensional microbubble mask, and this mask has been associated with the first... When a three-dimensional microbubble is formed, .

[0089] The horizontal pixel size of industrial camera 5 on the sample plane is:

[0090] in, , The pixel size of the sensor for the industrial camera 5. , This represents the magnification of the imaging system in both directions.

[0091] No. The three-dimensional microbubble in the first The cross-sectional area on each focal plane is:

[0092] in, These represent the image width and height, respectively.

[0093] The calculation steps for the geometric characteristic parameters of three-dimensional microbubbles are as follows: 1) The volume of the three-dimensional microbubble is calculated using the trapezoidal integral of the cross-sectional areas of adjacent focal planes:

[0094] in For the first The volume of a three-dimensional microbubble, This represents the actual horizontal dimension of a single pixel in the imaging system within the sample plane. This represents the actual vertical dimension of a single pixel in the sample plane of the imaging system. This represents the actual distance between adjacent focal planes within the sample. The total number of focal planes, The width in pixels of a single focal plane image. The height in pixels of a single focal plane image. For the first The three-dimensional binary voxel mask of the three-dimensional microbubble is in the first... Focal plane coordinates The value at that location, For the first The three-dimensional binary voxel mask of the three-dimensional microbubble is in the first... Focal plane coordinates The value at position; when the first Coordinates in a focal plane The pixel at that location belongs to the one already associated with the first... When masking the effective cross-section of a three-dimensional microbubble... ,otherwise , For the first The three-dimensional binary voxel mask of the three-dimensional microbubble is in the first... Focal plane coordinates The value at point [0, 1] is calculated using the trapezoidal integral of the cross-sectional areas of adjacent focal planes. The volume of each three-dimensional microbubble is determined to reduce the volume calculation error caused by the discrete focal plane.

[0095] 2) The formula for calculating the equivalent radius of a three-dimensional microbubble is:

[0096] in, For the first The equivalent radius of a three-dimensional microbubble; 3) Three-dimensional microbubble centroid The expression is:

[0097] in, For the first The total number of effective voxels in three-dimensional microbubbles These are the actual spatial coordinates corresponding to the origin of the three-dimensional coordinate system.

[0098] Total effective voxels The calculation formula is:

[0099] 4) The formula for calculating the remaining distance from the outer boundary of the three-dimensional microbubble to the inner wall of the crucible is:

[0100] in, For the first The remaining distance from the outer boundary of each three-dimensional microbubble to the inner wall of the crucible. For the first The distance from the centroid of a three-dimensional microbubble to the inner wall of the crucible. For the three-dimensional point set of the inner wall of the crucible, For any spatial point in the three-dimensional point set on the inner wall of the crucible. Based on the direct calculation method of continuous focal plane voxels, the approximate error of traditional two-dimensional projection to infer three-dimensional parameters is avoided, and the obtained geometric parameters such as volume and spacing are more consistent with the real spatial shape of the bubble.

[0101] Using the above method, the volume, equivalent radius, centroid coordinates, and shortest distance from the centroid of the three-dimensional microbubble to the inner wall of the crucible are directly obtained. These parameters are then used as input parameters for the subsequent microbubble wall rupture risk model and microbubble fusion embrittlement model.

[0102] S5. Based on the geometric feature parameters of three-dimensional microbubbles, two types of quantitative models for the fusion of high-temperature quartz physical mechanisms are constructed, namely the microbubble wall fracture risk model and the microbubble fusion embrittlement model.

[0103] Based on the viscous flow of high-temperature quartz, the evolution of internal pressure of microbubbles, and the effect of interfacial tension, a microbubble wall-joining rupture risk model is constructed to calculate the effective driving pressure difference of each three-dimensional microbubble. When the effective driving pressure difference is >0, the critical rupture time of the corresponding three-dimensional microbubble is calculated; otherwise, it is marked as having no expansion wall-joining tendency, and its critical rupture time is not calculated. Based on the Ostwald ripening and vacancy diffusion mechanism, a microbubble fusion embrittlement model is constructed to calculate the abnormal evolution index of three-dimensional microbubbles.

[0104] 1) The expression for the microbubble wall rupture risk model is determined as follows. Specifically, assuming the microbubble is approximately spherical, the quartz matrix is ​​an incompressible, high-viscosity continuous medium, and the inertial effect during the microbubble expansion process is negligible compared to the viscous resistance, the change in the microbubble radius over time is expressed as:

[0105] in, The radius of the microbubble; For quartz materials at temperature The dynamic viscosity at the following levels; This refers to the internal pressure of the microbubbles. External pressure of the microbubbles; This represents the interfacial tension between the microbubbles and the quartz matrix. Dynamic viscosity can also be obtained from high-temperature viscosity measurement data tables or other calibrated viscosity models.

[0106] Specifically, the dynamic viscosity of quartz materials can be calculated using the experimentally calibrated viscosity-temperature relationship. When using the Vogel-Fulcher-Tammann equation, its expression is:

[0107] in, , and The fitting parameters are based on the high-temperature viscosity test of quartz material. Absolute temperature is used.

[0108] Under the condition that the amount of gaseous substance inside the microbubble remains approximately constant within the predicted time, according to the ideal gas law, the internal pressure of the microbubble is expressed as:

[0109] in, This refers to the amount of substance of the gas inside the microbubble. It is the gas constant; The volume is the equivalent of a spherical microbubble.

[0110] Step S4 obtains the current three-dimensional equivalent radius of the microbubble. Three-dimensional volume and the three-dimensional shortest distance from the centroid of the microbubble to the inner wall of the crucible. Input the microbubble wall rupture risk model. Based on the above critical geometric conditions for rupture risk, the current remaining distance between the outer boundary of the microbubble and the inner wall of the crucible is... .when When the microbubble has come into contact with or crossed the boundary of the inner wall of the crucible, the system directly identifies the microbubble as a high-risk microbubble.

[0111] To facilitate engineering calculations, under quasi-steady-state conditions where temperature, quartz dynamic viscosity, and effective driving pressure difference of the microbubbles remain approximately constant within the prediction time, the formula for the change of the microbubble radius over time is approximated using the current state of the microbubbles. Wherein... The formula for calculating the effective driving pressure difference of a three-dimensional microbubble is:

[0112] in, This refers to the amount of substance of the gas within the microbubbles. The gas constant is The crystal pulling temperature, For temperature Interfacial tension between microbubbles and the quartz matrix.

[0113] Ultimately, the critical rupture time required for the microbubble to expand from its current state to the inner wall of the crucible, i.e., the expression for the microbubble wall-to-wall rupture risk model, can be obtained as follows:

[0114] in, For the first The critical rupture time of a three-dimensional microbubble. For temperature The dynamic viscosity of quartz below For the first The effective driving pressure difference of a three-dimensional microbubble The external environmental pressure of the microbubbles is the working pressure inside the furnace cavity during the single crystal pulling process. Its value is determined based on the argon atmosphere pressure parameter set in the single crystal pulling process.

[0115] when When this occurs, it indicates that the microbubbles have a tendency to continue expanding under the current temperature and pressure conditions. The theoretical critical contact time is calculated using the model expression. When the current effective driving pressure difference is insufficient to cause the microbubble to continue expanding outward, the finite critical contact time is not calculated, and the microbubble is marked as having no obvious expansion and contact tendency under the current conditions. This model transforms the static geometric spacing into a dynamic critical rupture time, realizing a quantitative prediction of the microbubble contact risk and providing a time-dimensional quantitative indicator for the assessment of the crucible's remaining service life.

[0116] This model, as a risk analysis branch of crucible life testing methods targeting near-wall microbubble expansion failure modes, calculates the time required for microbubbles to evolve from their current state to the critical state of near-wall rupture risk, providing time-dimensional characteristic parameters for crucible remaining life testing. Under the high-temperature conditions of single-crystal pulling, quartz glass exhibits high-viscosity rheological properties. Microbubbles near the inner wall of the crucible expand under the combined effects of internal microbubble pressure, external pressure, interfacial tension, and the viscous resistance of the quartz matrix. As the microbubbles continue to expand and approach the inner wall of the crucible, the thickness of the remaining quartz layer between the microbubbles and the inner wall gradually decreases, potentially causing a decrease in local strength of the inner wall and increasing the risk of rupture and melt contamination.

[0117] 2) The process of determining the expression for the microbubble fusion embrittlement model is as follows: First, the fusion and growth phenomenon exhibited by microbubbles within the transparent layer at high temperatures is essentially an Ostwald ripening process, following the Lifshitz-Slyozov-Wagner (LSW) theory. At the microscopic level, small microbubbles have high internal pressure and chemical potential, while large microbubbles have low internal pressure and chemical potential. Driven by the chemical potential gradient, gas atoms or vacancies diffuse from small microbubbles to large microbubbles through the quartz matrix, leading to the dissolution of small microbubbles and the growth of large microbubbles.

[0118] The expression for the microbubble fusion embrittlement model is as follows:

[0119] in, This is an index of abnormal evolution. For temperature The ripening rate constant under the condition, This is the actual ripening rate constant.

[0120] According to LSW theory, in a diffusion-controlled matrix, the cube of the average microbubble radius increases linearly with time. Based on this scaling relationship, the equivalent average microbubble radius obtained through 3D reconstruction before and after high-temperature treatment is calculated. The formula for calculating the actual ripening rate constant obtained by inversion is:

[0121] in, The average equivalent radius of the three-dimensional microbubbles in the entire field before high-temperature treatment. The average equivalent radius of the three-dimensional microbubbles in the entire field after high-temperature treatment. This refers to the duration of high-temperature processing.

[0122] temperature The formula for calculating the aging rate constant is:

[0123] in, This is the diffusion coefficient of the gaseous solute in the quartz matrix. The equilibrium solubility of the gas in quartz. This represents the molar volume of the gas. Using the standard thermodynamic theoretical rate as a reference, this index can effectively distinguish between the normal thermal evolution and abnormally accelerated growth of microbubbles. It can also reversely locate purity defects and crystallization risks in the crucible matrix, providing a diagnostic basis for optimizing raw materials and billet preparation processes.

[0124] S6, such as Figure 5 As shown, when the minimum critical rupture time of the full-field three-dimensional microbubbles is less than or equal to the expected single crystal pulling time, the crucible life is deemed insufficient, and a rupture risk warning is generated. Minimum critical rupture time of the full-field three-dimensional microbubbles The calculation formula is:

[0125] Specifically, the minimum critical rupture time of a three-dimensional microbubble represents the shortest time required for the most dangerous microbubble in the entire process to reach the critical state of wall rupture risk under current process conditions. ≤Expected single crystal pulling time When this occurs, it indicates that the remaining life under the microbubble wall rupture failure mode cannot cover the expected remaining drawing time, therefore an early warning should be issued. When this occurs, it indicates that the predicted lifetime under this failure mode can cover the expected remaining holding time, and no warning is needed.

[0126] When the actual maturation rate exceeds the maturation rate constant, an abnormal microbubble evolution rate and an abnormal matrix quality are detected, triggering a warning. Alternatively, an abnormal microbubble evolution index can be defined. Its expression is:

[0127] like A value greater than 1 indicates that the actual microbubble growth rate far exceeds the theoretical diffusion rate, strongly suggesting the presence of numerous microscopic defects within the quartz matrix, the transformation from amorphous to crystalline state (crystallization), or impurities such as alkali metals accelerating the diffusion process. Therefore, In reality, it's a quantitative evaluation index that uses macroscopic microbubble images to infer the microscopic purity and defect level of a quartz crucible. In risk assessment, when... When the threshold value exceeds 1, the system determines that the quartz matrix has abnormal microscopic quality degradation and triggers an alarm for abnormal matrix quality.

[0128] The dual-dimensional risk assessment system covers two failure modes: sudden fracture failure and long-term matrix performance degradation. The test results can directly support the quality classification, life prediction and process optimization of quartz crucibles.

[0129] This invention performs cross-focal-plane correlation on microbubble instance masks at each focal plane to construct a three-dimensional binary voxel region for the microbubble, and calculates the three-dimensional volume, three-dimensional equivalent radius, three-dimensional centroid coordinates, and three-dimensional distance from the crucible inner wall of the microbubble. Firstly, a microbubble wall-joint rupture risk model is constructed by combining high-temperature viscous flow of quartz, microbubble internal pressure evolution, and interfacial tension. The critical rupture time is calculated using the current three-dimensional equivalent radius of the microbubble and the remaining distance from the crucible inner wall, and the minimum critical rupture time across the entire field is used as a lifetime constraint parameter. Secondly, combining Ostwald ripening and vacancy diffusion mechanisms, the actual ripening rate is inverted based on the change in the three-dimensional equivalent average radius before and after high-temperature treatment, and the abnormal evolution exponent of the microbubble is calculated. Based on this, the minimum critical rupture time across the entire field is compared with the expected remaining single-crystal pulling time to achieve crucible lifetime status evaluation and automatic risk warning.

[0130] Based on existing technologies for macroscopic overall detection and geometric statistics, this invention achieves a leap in transparent layer microbubble detection from "random single qualitative observation" to "fixed-point comparative quantitative analysis", from "phenomenon statistics" to "mechanism correlation", and from "manual interpretation" to "automatic early warning" through the systematic integration of intelligent visual fixed-point acquisition of the same micro-area focal plane, a dedicated improved instance segmentation network, focal plane three-dimensional reconstruction, and material physical quantification model. It has made substantial improvements in acquisition reliability, segmentation accuracy, small target adaptability, and physical interpretability, and is more suitable for the actual needs of the monocrystalline silicon industry for fine quality assessment and service risk management of quartz crucible transparent layers.

[0131] Example of a crucible lifetime detection device based on thermal micro-variation micro-matter evolution: like Figure 7 , Figure 8 , Figure 9As shown, the crucible lifetime detection device based on thermal micro-variation micro-matter evolution includes: The image acquisition module is used to sequentially acquire imaging focal planes along the optical axis 5 of the industrial camera, continuously acquiring focal plane image sequences of the transparent layer micro-region of the quartz crucible sample before and after high-temperature treatment; high-temperature treatment refers to simulated crystal pulling high-temperature treatment. The two-dimensional microbubble extraction module is used to input the images of each focal plane before and after high-temperature treatment into the improved SOLOv2 instance segmentation network to complete the identification and segmentation of two-dimensional microbubbles in the transparent layer micromaterials within each focal plane, output the mask of the two-dimensional microbubbles in each focal plane, and extract the quantitative feature parameters of each two-dimensional microbubble. The 3D microbubble construction module is used to construct a 3D binary voxel mask for each 2D microbubble in the next focal plane based on the masks and quantitative feature parameters of the 2D microbubble in each focal plane, according to the focal plane image acquisition sequence and using the current focal plane and the next focal plane as matching units. For each 2D microbubble mask in the current focal plane, the overlap coefficient, normalized centroid distance, and normalized contour distance with each 2D microbubble mask in the next focal plane are calculated. Mask pairs with an overlap coefficient greater than 0 and a normalized centroid distance less than or equal to 1 are identified as candidate association pairs, and the association score of each candidate association pair is calculated. For each 2D microbubble mask in the current focal plane, the mask with the highest association score is selected from its corresponding candidate association pairs to complete the pairing. A single 2D microbubble mask is matched with only one corresponding mask. The 2D microbubble masks that are matched and associated sequentially in consecutive focal planes are classified into the same 3D microbubble, and the corresponding 3D microbubble is constructed. The geometric feature calculation module is used to calculate the geometric feature parameters of three-dimensional microbubbles based on the three-dimensional binary voxel mask of each three-dimensional microbubble. The geometric feature parameters of three-dimensional microbubbles include the volume of the three-dimensional microbubble, the equivalent radius of the three-dimensional microbubble, the centroid of the three-dimensional microbubble, and the remaining distance from the outer boundary of the three-dimensional microbubble to the inner wall of the crucible. The physical model construction module is used to construct two types of quantitative models integrating the physical mechanisms of high-temperature quartz based on the geometric characteristic parameters of three-dimensional microbubbles: a microbubble wall-to-wall rupture risk model and a microbubble fusion embrittlement model. Based on the viscous flow of high-temperature quartz, the evolution of internal pressure of microbubbles, and the effect of interfacial tension, the microbubble wall-to-wall rupture risk model is constructed to calculate the effective driving pressure difference of each three-dimensional microbubble. When the effective driving pressure difference is >0, the critical rupture time of the corresponding three-dimensional microbubble is calculated; otherwise, it is marked as having no expansion wall-to-wall tendency, and its critical rupture time is not calculated. Based on the Ostwald ripening and vacancy diffusion mechanism, the microbubble fusion embrittlement model is constructed to calculate the abnormal evolution index of three-dimensional microbubbles. The early warning module is used to determine that the crucible life is insufficient and generate a risk warning when the minimum critical rupture time of the three-dimensional microbubbles in the whole field is less than or equal to the expected single crystal pulling time; and to determine that the microbubble evolution rate is abnormal and generate a matrix quality abnormality warning when the actual ripening rate is greater than the ripening rate constant.

[0132] In this embodiment, the image acquisition module uses an intelligent vision acquisition platform to acquire focal plane image sequences of the transparent layer micro-area of ​​the quartz crucible sample before and after high-temperature treatment; the intelligent vision acquisition platform includes a clamping platform 1, an X-axis linear module 2, a Y-axis linear module 3, a Z-axis lifting module 4, an industrial camera 5, a microscope head assembly 9, an upper light source 6, a lower light source 7, and a PLC controller 8.

[0133] The clamping platform 1 is used to fix the test sample after the quartz crucible is cut, and to maintain the stability of the sample position during the acquisition process; the X-axis linear module 2 is used to drive the industrial camera 5 to move along the sample axis to achieve axial positioning of the transparent layer detection area; the Y-axis linear module 3 is used to drive the clamping platform 1 to move in the horizontal direction to achieve switching and precise positioning of different circumferential detection areas of the test sample; the Z-axis lifting module 4 is used to drive the industrial camera 5 to move along the imaging optical axis, and to change the imaging focal plane sequentially according to the preset starting focal plane position, the number of focal planes and the moving distance of adjacent focal planes to achieve automatic acquisition of focal plane image sequences; the industrial camera 5 and the microscope head assembly 9 are used to acquire microscopic microbubble images of the transparent layer to obtain focal plane images that meet the spatial resolution requirements; the upper light source 6 and the lower light source 7 are used to provide coordinated illumination of the transparent layer detection area; the PLC controller 8 is connected to the X-axis linear module, the Y-axis linear module and the Z-axis lifting module respectively.

[0134] This invention provides a micro-material intelligent vision acquisition scheme based on sample positioning reference, XYZ three-axis coordinate positioning, automatic recall of focal plane acquisition parameters, and automatic reset of industrial camera 5. The cut quartz crucible sample is fixed on the intelligent vision acquisition platform according to a preset positioning reference. The XYZ three-axis coordinates of the target micro-region in the transparent layer, the sample positioning reference, and the initial focal plane position are set and recorded. Number of focal planes Distance of movement between adjacent focal planes The imaging parameters of industrial camera 5 are used; the acquisition control module controls the Z-axis lifting module 4 to sequentially change the focal plane position along the optical axis of industrial camera 5, acquiring a sequence of focal plane images before high-temperature treatment. After the sample completes the simulated crystal pulling high-temperature treatment, the sample is re-fixed according to the above positioning reference, and the recorded positioning and focal plane acquisition parameters are automatically called to repeatedly acquire the same transparent layer target micro-region, obtaining a sequence of focal plane images before and after high-temperature treatment with spatial position and focal plane correspondence. This method improves the consistency of sample repetitive positioning and focal plane acquisition conditions, and provides reliable image data for cross-focal plane correlation of transparent layer microbubbles, calculation of three-dimensional equivalent parameters, and detection of feature evolution before and after high-temperature treatment.

[0135] This embodiment is implemented with reference to the above method embodiment, and will not be repeated here.

[0136] The above description is merely a preferred embodiment of the present invention and does not constitute any limitation on the present invention. Any simple modifications, alterations, or equivalent structural changes made to the above embodiments based on the technical essence of the present invention shall still fall within the protection scope of the present invention.

Claims

1. A crucible lifetime detection method based on thermal micro-variation micro-matter evolution, characterized in that: The steps include the following: S1. Acquire the imaging focal plane sequentially along the optical axis of the industrial camera (5), and continuously acquire the focal plane image sequence of the transparent layer micro-region of the quartz crucible sample before and after high temperature treatment; high temperature treatment refers to simulated crystal pulling high temperature treatment; S2. Input the images of each focal plane before and after high-temperature processing into the improved SOLOv2 instance segmentation network to complete the identification and segmentation of two-dimensional microbubbles in the transparent layer micromaterials within each focal plane, output the mask of the two-dimensional microbubbles in each focal plane, and extract the quantitative feature parameters of each two-dimensional microbubble. S3. Based on the masks of two-dimensional microbubbles in each focal plane and quantitative feature parameters, according to the focal plane image acquisition sequence, with the current focal plane and the next focal plane as the matching unit, for each two-dimensional microbubble mask in the current focal plane, calculate its overlap coefficient, normalized centroid distance and normalized contour distance with each two-dimensional microbubble mask in the next focal plane. Mask pairs with an overlap coefficient greater than 0 and a normalized centroid distance less than or equal to 1 are identified as candidate association pairs, and the association score of each candidate association pair is calculated. For each two-dimensional microbubble mask in the current focal plane, the mask with the highest association score is selected from its corresponding candidate association pairs to complete the pairing. A single two-dimensional microbubble mask is matched with only one corresponding mask. The two-dimensional microbubble masks that are matched and associated in successive focal planes are classified into the same three-dimensional microbubble, and a three-dimensional binary voxel mask of the corresponding three-dimensional microbubble is constructed. S4. Based on the three-dimensional binary voxel mask of each three-dimensional microbubble, calculate the geometric feature parameters of the three-dimensional microbubble; the geometric feature parameters of the three-dimensional microbubble include the volume of the three-dimensional microbubble, the equivalent radius of the three-dimensional microbubble, the centroid of the three-dimensional microbubble, and the remaining distance from the outer boundary of the three-dimensional microbubble to the inner wall of the crucible; S5. Based on the geometric feature parameters of three-dimensional microbubbles, two types of quantitative models for integrating the physical mechanism of high-temperature quartz are constructed, namely the microbubble wall fracture risk model and the microbubble fusion embrittlement model. Based on the viscous flow of high-temperature quartz, the evolution of internal pressure of microbubbles, and the effect of interfacial tension, a microbubble wall-joining rupture risk model is constructed to calculate the effective driving pressure difference of each three-dimensional microbubble. When the effective driving pressure difference is >0, the critical rupture time of the corresponding three-dimensional microbubble is calculated; otherwise, it is marked as having no expansion wall-joining tendency, and its critical rupture time is not calculated. Based on the Ostwald ripening and vacancy diffusion mechanism, a microbubble fusion embrittlement model is constructed to calculate the abnormal evolution index of three-dimensional microbubbles. S6. When the minimum critical rupture time of the three-dimensional microbubbles in the whole field is less than or equal to the expected single crystal pulling time, the crucible life is determined to be insufficient and a rupture risk warning is generated. When the actual maturation rate is greater than the maturation rate constant, an abnormal microbubble evolution rate and abnormal matrix quality are detected, triggering an early warning.

2. The crucible lifetime detection method based on thermal micro-variation micro-matter evolution according to claim 1, characterized in that: The quantitative characteristic parameters of the two-dimensional microbubble include the outline, area, equivalent radius, centroid coordinates, and planar distance from the inner wall of the crucible. The formula for calculating the overlap coefficient is: in, For the first focal plane center The two-dimensional microbubble and the first In the focal plane, the th The mask overlap factor of a two-dimensional microbubble. For the first focal plane center A collection of two-dimensional microbubble mask regions For the first focal plane center A collection of two-dimensional microbubble mask regions To perform the minimum value operation, This is a function for calculating the area of ​​a region, used to determine the pixel area of ​​the corresponding mask region. This refers to the intersection operation of sets. The formula for calculating the normalized centroid distance is: in, For the first focal plane center The two-dimensional microbubble and the first In the focal plane, the th The normalized centroid distance of a two-dimensional microbubble For the first focal plane center The centroid of a two-dimensional microbubble mask For the first focal plane center The centroid of a two-dimensional microbubble mask For the first focal plane center The equivalent radius of a two-dimensional microbubble mask, For the first focal plane center The equivalent radius of a two-dimensional microbubble mask, It is a two-dimensional Euclidean norm; The formula for calculating the normalized profile distance is: in, For the first focal plane center The two-dimensional microbubble and the first In the focal plane, the th The normalized profile distance of a two-dimensional microbubble. This is the Hausdorff distance operator. For the first In the focal plane, the th The outline of a two-dimensional microbubble For the first In the focal plane, the th The outline of a two-dimensional microbubble.

3. The crucible lifetime detection method based on thermal micro-variation micro-matter evolution according to claim 2, characterized in that: The formula for calculating the correlation score is as follows: in, For the first focal plane center The two-dimensional microbubble and the first In the focal plane, the th The correlation score of a two-dimensional microbubble. It is a natural exponential function.

4. The crucible lifetime detection method based on thermal micro-variation micro-matter evolution according to claim 1, characterized in that: The calculation steps for the geometric feature parameters of the three-dimensional microbubble are as follows: 1) The volume of the three-dimensional microbubble is calculated using the trapezoidal integral of the cross-sectional areas of adjacent focal planes: in, For the first The volume of a three-dimensional microbubble, This represents the actual horizontal dimension of a single pixel in the imaging system within the sample plane. This represents the actual vertical dimension of a single pixel in the sample plane of the imaging system. This represents the actual distance between adjacent focal planes within the sample. The total number of focal planes, The width in pixels of a single focal plane image. The height in pixels of a single focal plane image. For the first The three-dimensional binary voxel mask of the three-dimensional microbubble is in the first... Focal plane coordinates The value at that location, For the first The three-dimensional binary voxel mask of the three-dimensional microbubble is in the first... Focal plane coordinates The value at; 2) The formula for calculating the equivalent radius of a three-dimensional microbubble is: in, For the first The equivalent radius of a three-dimensional microbubble; 3) Three-dimensional microbubble centroid The expression is: in, For the first The total number of effective voxels in three-dimensional microbubbles These are the actual spatial coordinates corresponding to the origin of the three-dimensional coordinate system. 4) The formula for calculating the remaining distance from the outer boundary of the three-dimensional microbubble to the inner wall of the crucible is: in, For the first The remaining distance from the outer boundary of each three-dimensional microbubble to the inner wall of the crucible. For the first The distance from the centroid of a three-dimensional microbubble to the inner wall of the crucible. For the three-dimensional point set of the inner wall of the crucible, Let be any spatial point in the three-dimensional point set on the inner wall of the crucible.

5. The crucible lifetime detection method based on thermal micro-variation micro-matter evolution according to claim 4, characterized in that: The expression for the microbubble joint rupture risk model is as follows: in, For the first The critical rupture time of a three-dimensional microbubble. For temperature The dynamic viscosity of quartz below For the first The effective driving pressure difference of a three-dimensional microbubble; No. The formula for calculating the effective driving pressure difference of a three-dimensional microbubble is: in, This refers to the amount of substance of the gas within the microbubbles. The gas constant is The crystal pulling temperature, For temperature Interfacial tension between microbubbles and the quartz matrix. The environmental pressure outside the microbubbles; The expression for the microbubble fusion embrittlement model is as follows: in, This is an index of abnormal evolution. For temperature The ripening rate constant under the condition, This is the actual ripening rate constant; The formula for calculating the actual ripening rate constant is: in, The average equivalent radius of the three-dimensional microbubbles in the entire field before high-temperature treatment. The average equivalent radius of the three-dimensional microbubbles in the entire field after high-temperature treatment. This refers to the duration of high-temperature processing; temperature The formula for calculating the aging rate constant is: in, This is the diffusion coefficient of the gaseous solute in the quartz matrix. The equilibrium solubility of the gas in quartz. Let be the molar volume of the gas.

6. The crucible lifetime detection method based on thermal micro-variation micro-matter evolution according to claim 1, characterized in that: In step S2, the improved SOLOv2 instance segmentation network includes a ResNet-50 backbone network, a feature pyramid network, a dynamic prediction head, and a boundary-aware instance refinement module; the dynamic prediction head includes a category branch, a dynamic convolutional kernel branch, and a mask feature branch. The initial max pooling layer of the ResNet-50 backbone network is replaced by a spatially adaptive phase downsampling module to extract multi-level features from the input image; The feature pyramid network is used to perform multi-scale fusion of multi-level features and output five different resolution levels of prediction feature maps from P2 to P6. The high-resolution P2 layer feature routing strategy is adopted, and the high-resolution P2 layer features are used as the main prediction level of small-scale two-dimensional microbubbles, and the category branch and dynamic convolution kernel branch are accessed simultaneously. The category branch performs category prediction on the input features and outputs the microbubble category confidence score corresponding to each grid position; the dynamic convolution kernel branch works with the mask feature branch to generate a coarse two-dimensional microbubble instance mask. The boundary-aware instance refinement module uses P2 layer features, P3 layer features, a coarse two-dimensional microbubble instance mask, and a corresponding dynamic convolution kernel to correct uncertain regions at the mask boundary and output a refined two-dimensional microbubble instance mask.

7. The crucible lifetime detection method based on thermal micro-variation micro-matter evolution according to claim 6, characterized in that: The spatial adaptive phase downsampling module includes a standard pooling path and a phase-preserving path. The standard pooling path performs max pooling with a step size of 2 on the input features to obtain standard downsampling features. The phase-preserving path divides the input features into four spatial phase components through pixel rearrangement, calculates the selection weight of each spatial phase after grouping the feature channels, and selects the corresponding spatial phase feature from each group according to the selection weight. Boundary gating weights are generated based on the local response differences between different spatial phases, and the selected spatial phase features are injected into the standard downsampling features through controlled residual connections to output the phase-preserving downsampling features. The high-resolution P2 layer feature routing strategy cancels the 0.5x downsampling operation before the P2 layer features enter the dynamic prediction head, and directly uses the original spatial resolution P2 layer features with an effective stride of 4 as the input of the first prediction level, and simultaneously inputs the class branch and the dynamic convolution kernel branch; the P3 layer features, P4 layer features, P5 layer features, and P6 layer features after double bilinear upsampling are used as the input of the second to fifth prediction levels, respectively, and the effective strides of the five prediction levels are 4, 8, 16, 32 and 32, respectively.

8. The crucible lifetime detection method based on thermal micro-variation micro-matter evolution according to claim 6, characterized in that: The boundary-aware instance refinement module includes three stages: boundary detail decoding, boundary-aware point selection, and instance condition residual refinement. In the boundary detail decoding stage, convolution processing, upsampling and feature fusion are performed on the features of P2 layer and P3 layer to obtain detail feature map and boundary response map; In the boundary sensing point selection stage, based on the uncertainty and boundary response map of the coarse two-dimensional microbubble instance mask, the boundary positions to be refined are screened from the coarse mask; In the instance conditional residual refinement stage, the detailed features, mask prediction values, mask gradients, and coordinate information relative to the center of the microbubble instance are extracted from the location to be refined. The instance conditional modulation parameters are generated by combining the corresponding dynamic convolution kernels, the mask residuals are predicted, and the mask values ​​at the location to be refined are updated.

9. A crucible lifetime detection device based on thermal micro-variation micro-matter evolution, characterized in that: include: The image acquisition module is used to sequentially acquire the imaging focal plane along the optical axis of the industrial camera (5), and continuously acquire the focal plane image sequence of the transparent layer micro-area of ​​the quartz crucible sample before and after high temperature treatment. High-temperature treatment refers to simulating the high-temperature treatment of crystal pulling; The two-dimensional microbubble extraction module is used to input the images of each focal plane before and after high-temperature treatment into the improved SOLOv2 instance segmentation network to complete the identification and segmentation of two-dimensional microbubbles in the transparent layer micromaterials within each focal plane, output the mask of the two-dimensional microbubbles in each focal plane, and extract the quantitative feature parameters of each two-dimensional microbubble. The 3D microbubble construction module is used to construct a 3D binary voxel mask for each 2D microbubble in the next focal plane based on the masks and quantitative feature parameters of the 2D microbubble in each focal plane, according to the focal plane image acquisition sequence and using the current focal plane and the next focal plane as matching units. For each 2D microbubble mask in the current focal plane, the overlap coefficient, normalized centroid distance, and normalized contour distance with each 2D microbubble mask in the next focal plane are calculated. Mask pairs with an overlap coefficient greater than 0 and a normalized centroid distance less than or equal to 1 are identified as candidate association pairs, and the association score of each candidate association pair is calculated. For each 2D microbubble mask in the current focal plane, the mask with the highest association score is selected from its corresponding candidate association pairs to complete the pairing. A single 2D microbubble mask is matched with only one corresponding mask. The 2D microbubble masks that are matched and associated sequentially in consecutive focal planes are classified into the same 3D microbubble, and the corresponding 3D microbubble is constructed. The geometric feature calculation module is used to calculate the geometric feature parameters of three-dimensional microbubbles based on the three-dimensional binary voxel mask of each three-dimensional microbubble. The geometric feature parameters of three-dimensional microbubbles include the volume of the three-dimensional microbubble, the equivalent radius of the three-dimensional microbubble, the centroid of the three-dimensional microbubble, and the remaining distance from the outer boundary of the three-dimensional microbubble to the inner wall of the crucible. The physical model construction module is used to construct two types of quantitative models integrating the physical mechanisms of high-temperature quartz based on the geometric characteristic parameters of three-dimensional microbubbles: a microbubble wall-to-wall rupture risk model and a microbubble fusion embrittlement model. Based on the viscous flow of high-temperature quartz, the evolution of internal pressure of microbubbles, and the effect of interfacial tension, the microbubble wall-to-wall rupture risk model is constructed to calculate the effective driving pressure difference of each three-dimensional microbubble. When the effective driving pressure difference is >0, the critical rupture time of the corresponding three-dimensional microbubble is calculated; otherwise, it is marked as having no expansion wall-to-wall tendency, and its critical rupture time is not calculated. Based on the Ostwald ripening and vacancy diffusion mechanism, the microbubble fusion embrittlement model is constructed to calculate the abnormal evolution index of three-dimensional microbubbles. The early warning module is used to determine that the crucible life is insufficient and generate a rupture risk warning when the minimum critical rupture time of the three-dimensional microbubbles in the whole field is less than or equal to the expected single crystal pulling time. When the actual maturation rate is greater than the maturation rate constant, an abnormal microbubble evolution rate and abnormal matrix quality are detected, triggering an early warning.

10. The crucible lifetime detection device based on thermal micro-variation micro-matter evolution according to claim 9, characterized in that: The image acquisition module uses an intelligent vision acquisition platform to acquire focal plane image sequences of the transparent layer micro-area of ​​the quartz crucible sample before and after high-temperature treatment; the intelligent vision acquisition platform includes a clamping platform (1), an X-axis linear module (2), a Y-axis linear module (3), a Z-axis lifting module (4), an industrial camera (5), a microscope head assembly (9), an upper light source (6), a lower light source (7), and a PLC controller (8). The clamping platform (1) is used to fix the test sample after the quartz crucible is cut, and to maintain the stability of the sample position during the acquisition process; the X-axis linear module (2) is used to drive the industrial camera (5) to move along the sample axis to realize the axial positioning of the transparent layer detection area; the Y-axis linear module (3) is used to drive the clamping platform (1) to move in the horizontal direction to realize the switching and precise positioning of different circumferential detection areas of the test sample; the Z-axis lifting module (4) is used to drive the industrial camera (5) to move along the imaging optical axis, and to change the imaging focal plane in sequence according to the preset starting focal plane position, the number of focal planes and the moving distance of adjacent focal planes to realize the automatic acquisition of the focal plane image sequence; the industrial camera (5) and the microscope head assembly (9) are used to acquire the microscopic microbubble image of the transparent layer and obtain the focal plane image that meets the spatial resolution requirements; the upper light source (6) and the lower light source (7) are used to provide coordinated illumination to the transparent layer detection area; the PLC controller (8) is connected to the X-axis linear module (2), the Y-axis linear module (3) and the Z-axis lifting module (4) respectively.

Images