A silicon rod growth control method and system based on ultra-high-definition visual imaging

Abstract

The present application relates to a kind of based on ultra-high definition visual imaging's silicon rod growth control method and system.The method includes: obtaining the multiple spectrum and polarization dimension information of observation window transmission light;Based on the multiple spectrum and polarization dimension information, the influence of observation window on light propagation is analyzed, and light propagation influence parameter is generated;According to the light propagation influence parameter, the image of the silicon rod captured by main vision system is corrected to eliminate the artifact caused by observation window to image;Edge recognition and diameter calculation are carried out on the corrected silicon rod image, and the diameter information of the silicon rod is obtained;According to the diameter information of the silicon rod, control parameter in the process of silicon rod growth is adjusted.The present application uses multispectral polarization information to invert the optical distortion field of observation window and deduct its influence in measurement link, so that diameter data truly reflects the geometric size change of silicon rod, and the causal chain of control loop misoperation and crystal defect caused by observation window attachment is cut off from the root.

Description

Technical Field

[0001] This application relates to the field of silicon rod growth control technology, and in particular to a silicon rod growth control method and system based on ultra-high-definition visual imaging. Background Technology

[0002] High-purity, large-diameter silicon rods are the cornerstone of the semiconductor industry. The Czochralski method is a commonly used technique in silicon rod growth, typically employing a control system based on ultra-high-definition visual imaging. This system uses high-resolution cameras to capture real-time images of the silicon rod and dynamically adjusts growth parameters based on measurement data.

[0003] However, in actual production, the surface of the quartz observation window in the growth furnace develops complex, unevenly distributed silicate deposits due to the chemical reaction of silicon vapor under the extreme high temperatures within the furnace. These deposits cause minute and random deflections in the light emitted from the silicon rod's edge as it passes through the observation window, resulting in a high-frequency edge jitter artifact in the image captured by the camera. The image processing program misinterprets this optical deflection as rapid, minute fluctuations in the silicon rod's diameter, leading the diameter calculation unit to output incorrect measurement data.

[0004] After receiving diameter data filled with high-frequency noise, the control loop sends a large number of intensive fine-tuning commands to the control motor and heating components. These high-frequency fine-tuning actions cause continuous micro-amplitude vibrations that severely disrupt the ordered arrangement of silicon atoms in the crystal lattice, inducing a large number of dispersed micro-lattice dislocations and stress points. This leads to inconsistent electrical performance of semiconductor devices and a systemic, untraceable decline in product yield. To address this, we propose a silicon rod growth control method and system based on ultra-high-definition visual imaging. Summary of the Invention

[0005] This application provides a silicon rod growth control method and system based on ultra-high-definition visual imaging, which at least solves the problem in existing silicon rod growth control methods where image artifacts caused by deposits on the surface of the observation window lead to errors in diameter calculation and frequent fine-tuning of control parameters, ultimately resulting in lattice defects inside the silicon rod and a decrease in product yield.

[0006] In a first aspect, this application provides a method for controlling the growth of silicon rods based on ultra-high-definition visual imaging, comprising the following steps: Acquire multiple spectral and polarization dimensions of the light transmitted through the observation window; Based on the multiple spectral and polarization dimension information, the influence of the observation window on light propagation is analyzed, and light propagation influence parameters are generated. Based on the light propagation influence parameters, the silicon rod image captured by the main vision system is corrected to eliminate artifacts caused by the viewing window. Edge recognition and diameter calculation are performed on the corrected silicon rod image to obtain the diameter information of the silicon rod; Based on the diameter information of the silicon rod, the control parameters during the silicon rod growth process are adjusted.

[0007] Optionally, the step of analyzing the influence of the observation window on light propagation based on the multiple spectral and polarization dimension information and generating light propagation influence parameters includes: Obtain instruction information to adjust heating power and lifting speed; Based on the instruction information, predict the instantaneous change in the optical properties of the surface deposits on the observation window; The calibrated light propagation information is obtained by compensating for the instantaneous changes in the optical properties of the attached material from the multiple spectral and polarization dimensions of the transmitted light through the observation window. Based on the calibrated light propagation information, the influence of the observation window on light propagation is analyzed, and light propagation influence parameters are generated.

[0008] Optionally, the step of acquiring multiple spectral and polarization dimensions of the light transmitted through the observation window includes: During the growth of silicon rods, when there are transient micro-fluctuations on the surface of molten silicon or in the distribution of silicon vapor in the furnace, a high-speed imaging unit continuously acquires a sequence of original images containing information of different spectra and polarization dimensions at a sampling rate higher than the frequency of the transient micro-fluctuations. The original image sequence is timestamped, and spatial registration is performed on the image at each timestamp based on the timestamps to eliminate spatial misalignment caused by time asynchrony due to mechanical switching of filters during image acquisition. Based on the spatially registered image sequence, multiple spectral and polarization dimension information at each timestamp is extracted.

[0009] Optionally, the step of analyzing the influence of the observation window on light propagation based on the multiple spectral and polarization dimension information and generating light propagation influence parameters includes: Obtain furnace operating environment parameters, and predict the evolution trend of the surface deposits on the observation window based on the furnace operating environment parameters; The local light field characteristics of the transmitted light from the observation window are extracted, and the differences in the local light field characteristics are analyzed based on the predicted evolution trend of the attached structure. Based on the difference analysis results, the calculation weights and reference benchmarks of the optical distortion parameters are adaptively adjusted. Based on the adaptively adjusted optical distortion parameters and the reference benchmark, and using the multiple spectral and polarization dimension information, the influence of the observation window on light propagation is analyzed, and light propagation influence parameters are generated.

[0010] Optionally, the step of predicting the evolution trend of the surface deposits on the observation window based on the furnace operating environment parameters includes: Acquire local temperature and silicon vapor concentration information of multiple discrete regions on the surface of the observation window; Based on the local temperature information and silicon vapor concentration information of the multiple discrete regions, the growth rate and decomposition rate of the attachment in each discrete region are calculated, and the thickness change and morphological change of the attachment in each discrete region are predicted. The changes in the thickness and morphology of the attachments in each discrete region are combined to form a spatial distribution map of the evolution of the attachment structure on the observation window surface.

[0011] Optionally, the step of predicting the evolution trend of the surface deposits on the observation window based on the furnace operating environment parameters includes: Acquire real-time fluctuation data of furnace pressure, gas composition, and cooling water flow rate; Based on the real-time fluctuation data, identify the fluctuation type and fluctuation amplitude; For the identified fluctuation type and fluctuation amplitude, the corresponding attachment structure evolution adjustment factor is selected from the preset nonlinear response rule base; The attachment structure evolution adjustment factor is applied to the attachment structure evolution trend predicted based on conventional environmental parameters to obtain the corrected attachment structure evolution trend.

[0012] Optionally, the step of selecting the corresponding attachment structure evolution adjustment factor from a preset nonlinear response rule base includes: Obtain the process parameters for the current silicon rod growth stage, and based on the process parameters, assess the sensitivity of the current silicon rod growth stage to the adjustment factor of the attachment structure evolution. Multiple attachment structure evolution adjustment factors that match the identified fluctuation type and fluctuation amplitude are selected from the preset nonlinear response rule base; Based on the sensitivity, the selected multiple attachment structure evolution adjustment factors are weighted or ranked, and the attachment structure evolution adjustment factor with the highest weight or the highest ranking is selected.

[0013] Optionally, the step of adaptively adjusting the calculation weights and reference benchmarks of the optical distortion parameters includes: The surface of the observation window is divided into multiple micro-regions; For each micro-region, the difference analysis results of its local light field characteristics are obtained, and the evolution characteristics of the attached structure in the micro-region are determined; Based on the evolution characteristics of the attached structure in each micro-region, the calculation weights and reference benchmarks of the optical distortion parameters of that micro-region are dynamically determined; The calculated weights and reference standards of the optical distortion parameters in each micro-region are combined to form a spatial distribution map of the calculated weights and reference standards of the optical distortion parameters on the observation window surface.

[0014] Optionally, the step of dynamically determining the calculation weights and reference benchmarks of the optical distortion parameters of each micro-region based on the evolution characteristics of the attached structure includes: Obtain growth stage information for each micro-region; Based on the growth stage information, a parameter adjustment strategy is selected from a preset parameter adjustment strategy library; Based on the parameter adjustment strategy, the calculation weights and reference benchmarks are determined.

[0015] Secondly, this application provides a silicon rod growth control system based on ultra-high-definition visual imaging, the system comprising: The information acquisition module is used to acquire multiple spectral and polarization dimensions of the light transmitted through the observation window; The parameter generation module is used to analyze the influence of the observation window on light propagation based on the multiple spectral and polarization dimension information, and generate light propagation influence parameters; The image correction module is used to correct the silicon rod image captured by the main vision system according to the light propagation influence parameters, so as to eliminate the artifacts caused by the viewing window to the image. The diameter calculation module is used to perform edge recognition and diameter calculation on the corrected silicon rod image to obtain the diameter information of the silicon rod; The control parameter adjustment module is used to adjust the control parameters during the silicon rod growth process based on the diameter information of the silicon rod.

[0016] Compared with related technologies, the silicon rod growth control method and system based on ultra-high-definition visual imaging provided in this application have at least the following technical advantages: By acquiring multiple spectral and polarization dimensions of the light transmitted through the observation window, the influence of the observation window on light propagation is analyzed, and light propagation influence parameters are generated. The silicon rod image captured by the main vision system is then corrected to eliminate artifacts caused by the observation window, thereby obtaining accurate silicon rod diameter information and avoiding diameter calculation errors due to artifacts. Based on this, the silicon rod growth control parameters are adjusted, eliminating the need for high-frequency fine-tuning of control parameters caused by erroneous diameter data. This fundamentally solves the problem of internal lattice defects and decreased yield in silicon rods caused by a chain reaction resulting from deposits on the observation window.

[0017] In summary, this application transforms the traditional silicon rod growth control based on direct image processing into a precise control mechanism based on multi-dimensional optical property analysis and artifact correction, providing a systematic solution to the optical interference problem in high-temperature visual measurement scenarios.

[0018] Details of one or more embodiments of this application are set forth in the following drawings and description to make other features, objects and advantages of this application more readily apparent. Attached Figure Description

[0019] The accompanying drawings, which are included to provide a further understanding of this application and form part of this application, illustrate exemplary embodiments and are used to explain this application, but do not constitute an undue limitation of this application. In the drawings: Figure 1 This is a flowchart illustrating a silicon rod growth control method based on ultra-high-definition visual imaging according to an exemplary embodiment.

[0020] Figure 2 This is a flowchart illustrating step S1 according to an exemplary embodiment.

[0021] Figure 3 This is one of the flowcharts illustrating step S2 according to an exemplary embodiment.

[0022] Figure 4 This is a second flowchart illustrating step S2 according to an exemplary embodiment.

[0023] Figure 5 This is a flowchart illustrating step S2A according to an exemplary embodiment.

[0024] Figure 6 This is a flowchart illustrating step S2C according to an exemplary embodiment.

[0025] Figure 7 This is a block diagram illustrating a silicon rod growth control system based on ultra-high-definition visual imaging, according to another exemplary embodiment. Detailed Implementation

[0026] To make the objectives, technical solutions, and advantages of this application clearer, the application is described and illustrated below with reference to the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are merely illustrative and not intended to limit the scope of this application. All other embodiments obtained by those skilled in the art based on the embodiments provided in this application without inventive effort are within the scope of protection of this application.

[0027] Obviously, the accompanying drawings described below are merely some examples or embodiments of this application. Those skilled in the art can apply this application to other similar scenarios based on these drawings without any inventive effort. Furthermore, it is understood that although the efforts made in this development process may be complex and lengthy, for those skilled in the art related to the content disclosed in this application, any changes to design, manufacturing, or production based on the technical content disclosed in this application are merely conventional technical means and should not be construed as insufficient disclosure of the content of this application.

[0028] In this application, the reference to "embodiment" means that a specific feature, structure, or characteristic described in connection with an embodiment may be included in at least one embodiment of this application. The appearance of this phrase in various places in the specification does not necessarily refer to the same embodiment, nor is it a separate or alternative embodiment that is mutually exclusive with other embodiments. It will be explicitly and implicitly understood by those skilled in the art that the embodiments described in this application may be combined with other embodiments without conflict.

[0029] Unless otherwise defined, the technical or scientific terms used in this application shall have the ordinary meaning understood by one of ordinary skill in the art to which this application pertains. The terms “a,” “an,” “an,” “the,” and similar words used in this application do not indicate quantity limitation and may indicate singular or plural. The terms “comprising,” “including,” “having,” and any variations thereof used in this application are intended to cover non-exclusive inclusion; for example, a process, method, system, product, or device that includes a series of steps or modules (units) is not limited to the listed steps or units, but may also include steps or units not listed, or may include other steps or units inherent to these processes, methods, products, or devices. The terms “connected,” “linked,” “coupled,” and similar words used in this application are not limited to physical or mechanical connections, but may include electrical connections, whether direct or indirect. “Multiple” used in this application refers to two or more. “And / or” describes the relationship between related objects, indicating that three relationships may exist; for example, “A and / or B” can represent: A alone, A and B simultaneously, and B alone. The character " / " generally indicates that the preceding and following objects are in an "or" relationship. The terms "first," "second," and "third" used in this application are merely to distinguish similar objects and do not represent a specific ordering of the objects.

[0030] Example 1 Embodiment 1 of this application provides a method for controlling the growth of silicon rods based on ultra-high-definition visual imaging. Figure 1 This is a flowchart illustrating a silicon rod growth control method based on ultra-high-definition visual imaging, according to an exemplary embodiment. Figure 1 As shown, the method includes the following steps: S1, acquire multiple spectral and polarization dimensions of the light transmitted through the observation window; In this step, unlike traditional vision control schemes where the main vision camera directly captures images of the silicon rod through the observation window for measurement, with the observation window assumed to be a transparent, ideal window, this step takes a different approach. It adds an independent optical characteristic detection channel between the observation window and the main vision system. Using a beam splitter and polarizer array, it simultaneously collects intensity distribution information of light passing through the same observation window area across multiple spectral bands, including the visible and near-infrared bands, and at different polarization directions, such as 0°, 45°, 90°, and 135°. This multi-dimensional data is necessary because the silicate film attached to the observation window surface exhibits differentiated absorption, refraction, and scattering effects on light of different wavelengths and polarization states. Information from a single spectrum or polarization state is insufficient to fully describe the optical distortion behavior of the observation window. This step essentially transforms the observation window, originally treated as a transparent medium, into an independently characterizable optical device.

[0031] S2, based on the multiple spectral and polarization dimension information, analyze the influence of the observation window on light propagation and generate light propagation influence parameters; In this step, after acquiring the multidimensional optical data of S1, the task is to convert these raw spectral and polarization information into quantization parameters that can be directly used for image correction. First, the optical properties of the silicate deposits are characterized by their complex refractive index: in, wavelength Complex refractive index at that point, It is the real refractive index (which describes the phase retardation of light). It is the extinction coefficient (describing the absorption and attenuation of light). The imaginary unit. At different wavelengths... and The differences form the physical basis of multispectral analysis. Comparative analysis of transmittance across different wavelength channels reveals that specific chemical components in the substrate produce characteristic absorptions at different wavelengths. By analyzing the morphology and intensity of the absorption spectra, the chemical composition and equivalent thickness distribution of the substrate can be inferred. Simultaneously, the degree of depolarization under different polarization states is quantitatively analyzed; the degree of depolarization is directly related to the scattering ability of the substrate's internal microstructure, with thicker or rougher substrates exhibiting a stronger depolarization effect. Finally, these analytical results are uniformly quantified into a set of light propagation influence parameters, including pixel-level refraction offset vectors, scattering point spread function radius, and transmission attenuation coefficients. This parameter set is essentially a virtual observation window distortion field derived from measured optical data, within which subsequent image correction is performed.

[0032] S3, Based on the light propagation influence parameters, the silicon rod image captured by the main vision system is corrected to eliminate artifacts caused by the viewing window to the image; In this step, physical distortion correction is performed on the raw silicon rod image captured by the main vision system based on the light propagation influence parameters output by S20, rather than performing blur filtering or statistical denoising at the image processing level. The specific correction operation includes obtaining the refraction offset vector according to the following formula: in, and These are the pixel offsets in the x and y directions, respectively. The local refractive index of the attachment at that location. The reference refractive index is the one without any adhering substances. and These represent the angular components of the incident light ray in the x and y directions. A reverse position offset is applied to each pixel to correct pixels that have been refracted and deviated from their true positions. Based on the scattering point spread function radius, deconvolution is used to compress the edge blur caused by scattering and restore edge sharpness. The transmission attenuation coefficient is obtained using the following formula to compensate for the grayscale values ​​of each pixel: in, To correct the grayscale value, The original image grayscale values. The pixels obtained by S2 calculation The transmission attenuation coefficient at the point is used to compensate for brightness distortion caused by absorption. Since the correction is based on measured optical transmission parameters rather than assumptions about image statistical properties, this correction process does not introduce artificial artifacts common in filtering algorithms. The corrected silicon rod image closely approximates the true optical imaging effect when the observation window is free of attachments, and high-frequency edge jitter artifacts are eliminated.

[0033] S4, perform edge recognition and diameter calculation on the corrected silicon rod image to obtain the diameter information of the silicon rod; In this step, the image has already eliminated the effects of optical distortion in the viewing window, so measurements can be performed as if processing an interference-free image. A sub-pixel precision edge detection algorithm is used, such as the improved Canny algorithm based on grayscale gradients or the Zernike moment method based on local region fitting, to extract the contour boundary positions on the left and right sides of the silicon rod. After obtaining the complete edge contour, multiple equally spaced cross-sections are selected along the silicon rod axis, and the diameter value at each cross-section is calculated. The average or weighted average of these values ​​is then used as the diameter information at the current moment. The key difference from traditional methods is that the edges extracted here are corrected true edges, free from high-frequency jitter components caused by objects attached to the viewing window. Therefore, false micrometer-level diameter jumps caused by artifacts no longer occur between adjacent frames, improving the temporal stability of the diameter data.

[0034] S5, adjust the control parameters during the silicon rod growth process according to the diameter information of the silicon rod; In this step, after obtaining diameter data that accurately reflects the geometric changes of the silicon rod, the operating mode of the control loop undergoes a fundamental change. In traditional schemes, noisy diameter data forces the PID controller to output tiny corrections almost every control cycle; however, in this step, because the diameter data provided by S40 is smooth and accurate, adjustment is only triggered immediately when a persistent deviation in the silicon rod diameter is detected, i.e., rather than single-frame noise fluctuations. Specific control parameters adjusted include: crystal pulling speed, used to control the axial growth rate of the silicon rod; crucible rotation speed, used to influence thermal convection and mass transport in the melt and heater power output; and temperature gradient at the solid-liquid interface. The reduced frequency and amplitude of adjustments decrease the micro-vibrations of the mechanical actuators, maintain higher stability of the thermal and flow fields at the growth interface, and allow silicon atoms to orderly enter their lattice positions without disturbance.

[0035] In the technical solution of the above embodiment, an independent optical characteristic analysis channel is introduced into the traditional visual imaging control process. This channel actively acquires the multispectral and polarization dimension information of the light transmitted through the observation window, converts it into light propagation influence parameters, and uses these parameters to perform physical distortion correction on the main visual image. This approach breaks the strong coupling between the optical distortion of the observation window and the measurement of the silicon rod image in traditional methods: traditional methods treat the observation window as a transparent window and directly equate the image captured through the observation window with the true morphology of the silicon rod; while this solution first characterizes the optical behavior of the observation window and subtracts the distortion effect during the measurement stage. The corrected diameter data accurately reflects the geometrical changes of the silicon rod, enabling the control loop to make accurate decisions under interference-free conditions. This fundamentally breaks the causal chain from observation window attachments to image artifacts to erroneous diameter data to PID high-frequency fine-tuning to mechanical micro-vibration to lattice dislocations to decreased yield.

[0036] In one possible design, Figure 2 This is a flowchart illustrating step S1 according to an exemplary embodiment. (Refer to the attached document.) Figure 2 In step S1, the step of obtaining multiple spectral and polarization dimension information includes: S11, During the growth of silicon rods, when there are transient micro-fluctuations in the molten silicon surface or the distribution of silicon vapor in the furnace, the high-speed imaging unit continuously acquires original image sequences containing different spectral and polarization dimension information at a sampling rate higher than the frequency of the transient micro-fluctuations. In this step, during silicon rod growth, natural convection at the molten silicon surface and localized fluctuations in silicon vapor concentration cause transient changes in the intensity of light transmitted through the observation window. These micro-fluctuations and the distortion caused by deposits on the observation window originate from different physical sources: the micro-fluctuations originate from the furnace flow field, while the distortion arises from the optical effects of the deposits. Therefore, they must be distinguished at the sampling level. To address this, a high-speed imaging unit is used, capturing raw image sequences at a rate significantly higher than the characteristic frequency of the transient micro-fluctuations. Typically, the sampling rate is set to tens to hundreds of hertz when the micro-fluctuations are on the order of a few hertz, ensuring sufficient sampling points are obtained within each micro-fluctuation cycle. Subsequent temporal filtering can then separate the two types of changes. During each acquisition, data from different spectral and polarization channels are alternately recorded in consecutive frames by rapidly switching between bandpass filters with different center wavelengths and polarizers with different orientations.

[0037] S12, timestamp the original image sequence, and spatially register the image under each timestamp according to the timestamp to eliminate spatial misalignment caused by time asynchrony due to mechanical switching of filters during image acquisition; In this step, multispectral and multipolarization channel images are acquired at different times by mechanically switching filters and polarizers. This physical implementation introduces two technical challenges. First, there is a time shift between the images of different channels; the silicon rod may undergo slight spatial displacement due to the lifting motion during acquisition. Second, the thermal expansion and slow rotation of the silicon rod under the high temperature environment inside the furnace also introduce additional spatial shifts. In this step, a high-precision timestamp is recorded for each frame of the original image, and the image of one of the channels is used as the spatial reference frame. A registration algorithm based on cross-correlation or feature point matching is used to transform the images of the remaining channels one by one into the spatial coordinate system of the reference frame. After spatial registration, the groups of different spectral and polarization images corresponding to the same timestamp are precisely aligned in space.

[0038] S13, based on the spatially registered image sequence, extract multiple spectral and polarization dimension information at each timestamp; In this step, after spatial registration, multidimensional optical features are extracted from the synchronized image group corresponding to each time stamp. For the spectral dimension, the average pixel radiance of the corresponding observation window region in each channel image is extracted to construct a spectral radiance vector. For the polarization dimension, parameters such as linear polarization degree and polarization angle are calculated based on intensity measurements in at least three different polarization directions. Furthermore, by performing differential calculations on the feature vectors of adjacent time stamps, the temporal change rate of information in each dimension can be obtained, providing a data foundation for subsequent transient disturbance separation schemes.

[0039] In the technical solution of the above embodiment, the multi-dimensional information acquisition process in S1 is further refined: high-speed oversampling solves the transient micro-fluctuation aliasing problem; timestamp marking and spatial registration solve the spatial deviation caused by the asynchrony of multi-channel images; and multi-dimensional feature extraction transforms the original pixel data into structured optical feature vectors. These three steps ensure the precise consistency of the data entering the S20 analysis stage in time and space, so that the subsequent analysis of light propagation influence parameters is based on reliable data.

[0040] In one example, during silicon rod growth, the molten silicon surface exhibits minute vibrations at a frequency of approximately 10 Hz. To accurately capture the impact of these vibrations on the transmitted light through the observation window, a high-speed CMOS camera is used as the high-speed imaging unit, with its frame rate set to 200 Hz, significantly higher than the 10 Hz fluctuation frequency. This camera continuously acquires images via a rapidly switching filter wheel containing red, green, and blue filters and polarizers. During acquisition, each time a filter switches and a frame is captured, a timestamp accurate to the millisecond level is added to that image. When the filter switches from red to green, due to the mechanical switching delay and minor jitter, there is a slight difference in the acquisition time between the red and green images, and the image center may shift by a few pixels. After acquiring the original image sequence, the timestamp of each frame is read first, and the first red frame in the sequence is selected as the reference image. For each subsequently acquired image frame, such as a green image, a blue image, or a polarized image, a spatial registration technique based on feature point matching and the RANSAC algorithm is used to calculate the geometric transformation matrix of that frame image relative to the reference image. This transformation matrix is ​​then applied for correction, ensuring precise spatial alignment with the reference image. In this way, even when mechanically switching filters causes time synchronization issues, it is possible to ensure that spatially aligned spectral and polarization dimension information is obtained at the same virtual time point.

[0041] In one possible design, Figure 3 This is one of the flowcharts illustrating step S2 according to an exemplary embodiment. (Refer to the attached diagram.) Figure 3 In step S2, the step of analyzing the effect of the observation window on light propagation includes: S21, Obtain instruction information for adjusting heating power and lifting speed; In this step, we start with the real-time behavior of the control loop and use feedforward information from the control commands to separate transient optical disturbances. We obtain the power and speed adjustment commands issued to the heating element and lifting motor in real time from the PID controller. The logic behind introducing command information is that any abrupt change in heating power and lifting speed will instantly alter the thermal flow field and silicon vapor concentration field near the observation window, causing instantaneous jumps in the optical properties of the deposits. If these are indiscriminately categorized as steady-state distortions, the resulting light propagation influence parameters will include error components introduced by the control actions themselves.

[0042] S22, based on the instruction information, predict the instantaneous change in the optical properties of the surface deposits on the observation window; In this step, a pre-calibrated thermo-optical coupling model is used to map the amplitude and direction of the command acquired in S21 into instantaneous changes in the refractive index and extinction coefficient of the attachment. This model is pre-established through offline experiments: under step excitation conditions of different heating powers and pulling speeds, the transmittance and reflectance curves of the attachment at multiple wavelengths are recorded on the observation window, and the transfer function between the command parameters and the changes in optical properties is fitted. For the heating power command, the refractive index of the attachment is mainly affected through the thermo-optical coefficient effect; for the pulling speed command, the equivalent extinction coefficient of the attachment surface layer is mainly affected by changing the deposition flux of silicon vapor on the observation window.

[0043] S23, from the multiple spectral and polarization dimension information of the transmitted light through the observation window, compensate for the instantaneous changes in the predicted optical properties of the attachment to obtain the calibrated light propagation information; In this step, the optical signal components corresponding to the instantaneous changes predicted in S22 are subtracted from the original spectral and polarization dimension information acquired in S1. For instantaneous changes in refractive index, the equivalent transmittance offset caused by the refractive index change is subtracted from the spectral transmittance data; for instantaneous changes in extinction coefficient, the additional depolarization component caused by the extinction change is subtracted from the polarization depolarization data. The remaining signal after subtraction is the calibrated light propagation information, representing the true optical transmission characteristics of the observation window under steady-state conditions.

[0044] S24, Based on the calibrated light propagation information, analyze the influence of the observation window on light propagation and generate light propagation influence parameters; In this step, based on the calibrated light propagation information, the spatial thickness distribution and equivalent refractive index distribution of the observation window attachments at the current moment are inverted using an optical forward model to generate light propagation influence parameters for subsequent image correction. Since the calibration process has removed transient optical disturbances caused by control commands, these parameters more accurately reflect the steady-state distortion caused by the observation window attachments themselves, avoiding the error of attributing the control system's own behavioral errors to changes in the optical properties of the observation window.

[0045] In the technical solution of the above embodiments, by combining control commands to predict and compensate for the instantaneous changes in the optical properties of the attachment, the transient optical disturbances caused by the adjustment of control parameters are separated from the steady-state distortion, thereby improving the generation accuracy of the light propagation influence parameters.

[0046] In one example, during silicon rod growth, a command is issued to increase the heating power by 10% from a preset value. Without the solution of this application, such a sudden increase in heating power might lead to a localized increase in silicon vapor concentration within the furnace, causing instantaneous deposition of silicon deposits on the observation window surface, resulting in a temporary decrease in the intensity of transmitted light or a slight change in polarization state. If the original spectral and polarization dimension information acquired at this time is used directly for analysis, it might be incorrectly assumed that the overall transmittance of the observation window has decreased or the polarization effect has increased, thus generating inaccurate light propagation influence parameters. However, according to the solution of this application, when the command information for increasing the heating power by 10% is received, the instantaneous change in the optical properties of the deposits on the observation window surface is immediately predicted based on a preset model, for example, predicting that the transmittance will instantaneously decrease by 0.5% and the polarization rotation angle will instantaneously increase by 0.1°. Subsequently, when extracting data from multiple spectral and polarization dimension information of the transmitted light from the observation window, these predicted instantaneous changes are used to compensate for the original data: increasing the original transmittance data by 0.5% and decreasing the original polarization rotation angle data by 0.1°, thereby obtaining calibrated light propagation information. Based on this calibrated information, the actual impact of the observation window on light propagation can be analyzed more accurately, generating light propagation influence parameters that are not affected by instantaneous heating power changes. This makes the correction of the silicon rod image more precise, ensuring the stability of silicon rod diameter measurement and the reliability of control.

[0047] In one possible design, Figure 4 This is a second flowchart illustrating step S2 according to an exemplary embodiment. (Refer to the attached diagram.) Figure 4 In step S2, the step of analyzing the effect of the observation window on light propagation includes: S2A: Obtain furnace operating environment parameters, and predict the evolution trend of the surface deposits structure based on the furnace operating environment parameters; The core logic of this step is that the root cause of optical distortion in the observation window lies in the structure of the deposit itself, including its thickness distribution and micromorphology. The evolution of these structures is directly driven by the furnace operating environment. Therefore, predicting structural evolution based on environmental parameters is fundamental to achieving accurate optical modeling. The collected furnace operating environment parameters include furnace temperature distribution, gas pressure, protective gas component ratio, and cooling water flow rate. These parameters are input into a pre-established physicochemical model of deposit deposition and decomposition. This model encompasses the chemical vapor deposition kinetics of silicon vapor on the observation window surface, the thermal decomposition reaction rate of the deposit at high temperatures, and the scouring effect of airflow on the deposit surface. It predicts the evolution trend of deposit thickness and micromorphology at different locations on the observation window over a future period, providing a reference benchmark for subsequent analysis of optical field differences.

[0048] S2B: Extract the local light field characteristics of the transmitted light from the observation window, and perform a difference analysis on the local light field characteristics based on the predicted evolution trend of the attached structure. In this step, local optical field features are extracted from the multispectral and polarization data obtained from S1, according to different regions of the observation window. These features include the average transmittance at each wavelength, the polarization degree variation pattern with direction, and the optical field gradient information between adjacent regions. Then, the extracted local optical field features are compared region by region with the expected optical field features simulated based on the attached structure predicted by S2A. Regions with significant differences usually indicate that the actual evolution path of the attached organism deviates from the prediction based on conventional environmental parameters, requiring more attention and weight adjustment in downstream processing.

[0049] S2C adaptively adjusts the calculation weights and reference benchmarks of optical distortion parameters based on the difference analysis results. In this step, based on the S2B difference analysis results, the calculation strategies for optical distortion parameters, including refraction offset vector, scattering diffusion radius, and transmission attenuation coefficient, are dynamically adjusted for different observation window regions. For regions with large differences, i.e., regions where the evolution of the attachment deviates significantly from the prediction, the calculation weight of the optical distortion parameters in these regions is increased, and the latest frame's measured data is used as a reference benchmark to ensure the timeliness of the correction. For stable regions with smaller differences, the calculation weight is appropriately reduced, and the moving average of multiple frames of historical data is used as a reference benchmark to suppress random noise. This adaptive weight allocation strategy allows the optical distortion parameters to both track rapid changes in the attachment and maintain a low noise level in stable regions.

[0050] S2D, based on the adaptively adjusted optical distortion parameters and the reference benchmark, analyzes the influence of the observation window on light propagation based on the multiple spectral and polarization dimension information, and generates light propagation influence parameters. In this step, the adaptively adjusted weights and baseline are applied to the optical forward model to calculate the equivalent optical path delay and transmission attenuation coefficient for each pixel location. This results in a set of light propagation influence parameters with unevenly distributed weights that dynamically update with changes in the attached material. Compared to using globally uniform weights, the adaptively adjusted parameters have higher spatial resolution and temporal response speed in regions with drastic changes in the attached material, achieving spatially adaptive and accurate modeling of the optical influence of the viewing window as a whole.

[0051] In the technical solution of the above embodiments, by predicting the evolution trend of the attached structure and adaptively adjusting the optical distortion parameters in combination with the differences in local light field characteristics, spatial adaptive accurate modeling of the optical influence of the observation window is achieved.

[0052] In one possible design, Figure 5 This is a flowchart illustrating step S2A according to an exemplary embodiment. (See attached diagram.) Figure 5 In step S2A, the step of predicting the evolution trend of the attached structure includes: S2A1 acquires local temperature and silicon vapor concentration information for multiple discrete regions on the surface of the observation window; In this step, building upon S2A, this scheme provides a more specific implementation method for predicting the evolution of deposit structures using furnace operating environment parameters. First, by placing temperature-sensing elements, such as thermocouple arrays and gas sensors, at different locations around the observation window, local temperature and silicon vapor concentration data for multiple discrete regions on the observation window surface are acquired. The selection of discrete regions is based on the thermal and flow field distribution characteristics of the observation window; that is, denser sampling points are typically placed at locations with larger temperature and airflow velocity gradients. These local environmental data directly determine the competitive relationship between the deposition and decomposition rates of deposits in each region.

[0053] S2A2, based on the local temperature information and silicon vapor concentration information of the multiple discrete regions, calculate the growth rate and decomposition rate of the attachment in each discrete region, and predict the change in the thickness and morphology of the attachment in each discrete region. In this step, the silicon vapor chemical vapor deposition rate Calculated according to the Arrhenius formula: in, For the deposition rate, Pre-exponential factor, It is the activation energy of a chemical reaction. The gas constant is... To observe the local thermodynamic temperature of the window surface, Local silicon vapor concentration; decomposition rate Then calculate according to the following formula: in, To decompose pre-exponential factors, The decomposition activation energy is used. The net deposition rate is: Based on the above rate equations, the silicon vapor chemical vapor deposition rate and the thermal decomposition rate of the deposits at high temperatures are calculated for each discrete region. The net deposition rate equals the deposition rate minus the decomposition rate; a positive value indicates that the deposits thicken during that time period, while a negative value indicates thinning. Multiplying the net rate by the predicted time span yields the change in deposit thickness for that region. Simultaneously, the morphological change direction of the deposits is determined based on the ratio of the deposition rate to the decomposition rate: high deposition / low decomposition tends towards dense, layered growth, while low deposition / high decomposition tends towards a porous, loose structure.

[0054] S2A3 combines the changes in the thickness and morphology of the attachments in each discrete region to form a spatial distribution map of the evolution of the attachment structure on the surface of the observation window. In this step, the predicted results for each discrete region, including thickness variation (a scalar quantity) and morphological variation (indicating densification or loosening trends), are used to generate a continuous two-dimensional spatial distribution map through spatial interpolation algorithms, such as radial basis function-based interpolation or Kriging interpolation. This distribution map visually demonstrates the expected structural and morphological evolution of the attachments at different locations on the observation window surface over a future period, serving as a spatial reference for subsequent optical field characteristic difference analysis and adaptive adjustment of optical distortion parameters.

[0055] In one possible design, another approach to predicting the evolution trend of the surface deposits on the observation window based on furnace operating environment parameters in step S2A includes: S2Aa acquires real-time fluctuation data of furnace pressure, gas composition, and cooling water flow rate; In this step, the above S2A1-S2A3 schemes are based on temperature and concentration field data under stable environmental conditions for prediction. However, in actual growth furnace operation, the furnace pressure, protective gas composition, and cooling water flow rate are not constant; that is, they fluctuate in real time due to the regulation of the external gas supply system and the circulating cooling system. In this step, real-time time-series data of these environmental parameters are simultaneously acquired from the growth furnace pressure sensor, online mass spectrometer, and cooling water flow meter. The impact of such real-time fluctuations on the evolution of the deposit structure is non-linear: small pressure fluctuations may have a negligible impact on the deposition rate, but exceeding a certain threshold will cause a qualitative change in the deposition mode.

[0056] S2Ab identifies the fluctuation type and fluctuation amplitude based on the real-time fluctuation data; In this step, spectral analysis is performed on the acquired real-time time series, such as using Fast Fourier Transform and statistical feature extraction, to identify the type of fluctuation: periodic fluctuations, such as the pulsation frequency of a cooling water pump; occasional impacts, such as pressure spikes or random noise caused by gas cylinder switching; and the amplitude and duration of each type of fluctuation. Different types of fluctuations affect the attached structure through different mechanisms: periodic pressure fluctuations mainly cause periodic changes in the density of the attached material, while occasional impacts may form interlayer interface defects in the attached material.

[0057] S2Ac selects the corresponding attachment structure evolution adjustment factor from the preset nonlinear response rule base for the identified fluctuation type and fluctuation amplitude. In this step, based on the fluctuation type and amplitude identified by S2Ab, a matching adjustment factor for the attachment structure evolution is selected from a pre-built nonlinear response rule base. This rule base is established through extensive experimental calibration: under different types and amplitudes of environmental fluctuations, the difference between the actual measured deviation of the attachment structure evolution within the observation window and the predicted value based on steady-state assumptions is quantified into an adjustment factor. The role of the adjustment factor is to correct the basic evolution trend predicted based on conventional steady-state parameters.

[0058] S2Ad applies the attachment structure evolution adjustment factor to the attachment structure evolution trend predicted based on conventional environmental parameters to obtain the corrected attachment structure evolution trend. In this step, the adjustment factor selected in S2Ac is superimposed or multiplied with the evolution trend of the basic attachment structure predicted by the aforementioned schemes S2A1-S2A3 based on steady-state environmental parameters, to obtain the corrected evolution trend. The corrected trend reflects the actual evolution path of the attachment structure under the influence of real-time environmental fluctuations, and is no longer limited by the deviation of the steady-state assumption. Therefore, it can still give prediction results that are consistent with reality when there are large dynamic fluctuations in the furnace environment.

[0059] In the technical solution of the above embodiments, by identifying environmental fluctuation characteristics and using a nonlinear response rule base for correction, the evolution trend of the attached structure can still be accurately predicted even when there are dynamic fluctuations in the furnace environment.

[0060] In one example, during silicon rod growth, a sudden, short-duration pulse of pressure surged within the furnace, accompanied by slight fluctuations in cooling water flow. First, real-time data on furnace pressure, gas composition, and cooling water flow were acquired using high-precision sensors. Next, this data was processed to identify the pressure fluctuation as a short-duration high-pressure pulse with an amplitude of X Pa, and the cooling water flow fluctuation as a slight decrease with an amplitude of Y liters per minute. Then, based on the identified fluctuation type and amplitude, a query was performed in a pre-defined nonlinear response rule base. This base contains a rule: when a short-duration high-pressure pulse and a slight decrease in cooling water flow occur, the attachment growth rate will instantaneously increase by Z%. A corresponding attachment structure evolution adjustment factor was selected, such as a multiplicative factor of 1+Z%, and this factor was applied to the attachment structure evolution trend predicted based on conventional environmental parameters. For example, if the conventional prediction is that the attachment thickness will increase by 0.1 micrometers in the next minute, the corrected attachment structure evolution trend is obtained: the corrected prediction is that the attachment thickness will increase by 0.1 multiplied in the next minute, with the factor being 1+Z% micrometers. In this way, even when transient fluctuations occur in the furnace environment, the structural evolution of the deposits on the observation window surface can be accurately predicted.

[0061] In one possible design, step S2Ac, which involves selecting the corresponding attachment structure evolution adjustment factor from a preset nonlinear response rule base, includes: S2Ac1: Obtain the process parameters of the current silicon rod growth stage, and based on the process parameters, evaluate the sensitivity of the current silicon rod growth stage to the adjustment factor of the attachment structure evolution. In this step, significant differences exist in the furnace thermal field distribution and silicon vapor generation at different stages of silicon rod growth—namely, the seeding stage, the constant-diameter growth stage, and the final stage. These differences lead to varying deposition rates and chemical compositions of the deposits at different stages. By acquiring information about the current silicon rod growth stage and its key process parameters, such as the target diameter, crystal pulling speed range, and heater power range, the tolerance for deviations in the deposit structure evolution prediction at this stage can be assessed. For example, the constant-diameter growth stage has much higher requirements for diameter accuracy than the seeding stage; therefore, this stage is more sensitive to deposit evolution prediction errors and requires the selection of a more responsive adjustment factor.

[0062] S2Ac2, selects multiple attachment structure evolution adjustment factors from the preset nonlinear response rule base that match the identified fluctuation type and fluctuation amplitude; In this step, based on the volatility type identified by S2Aa (including periodic, occasional shock, and random volatility) and volatility amplitude, the rule base is initially filtered to identify all candidate adjustment factors that meet the matching conditions. Each adjustment factor in the rule base is associated with a set of applicable condition tags, including volatility type, amplitude range, and applicable growth stage tags, allowing for rapid identification of the candidate set through multi-condition combination queries.

[0063] S2Ac3, based on the sensitivity, weight or sort the multiple attachment structure evolution adjustment factors selected, and select the attachment structure evolution adjustment factor with the highest weight or the highest ranking. In this step, the sensitivity of the current stage assessed by S2Ac1 is used as a secondary screening criterion to weight and rank the candidate adjustment factors. Stages with high sensitivity tend to select adjustment factors with fast response speeds but potentially introducing small overshoots, while stages with low sensitivity prioritize adjustment factors with smooth responses but slight lags. Finally, the adjustment factor with the highest overall score is selected as the correction factor for the current moment, ensuring that the factor selection results match the control requirements of this stage.

[0064] In the technical solution of the above embodiments, by combining the sensitivity assessment of the silicon rod growth stage to weight the selection of adjustment factors, adaptive optimization of the prediction of the evolution of attachments at different growth stages is achieved.

[0065] In one possible design, Figure 6 This is a flowchart illustrating step S2C according to an exemplary embodiment. (Refer to the attached diagram.) Figure 6 In step S2C, the step of adaptively adjusting the optical distortion parameters includes: S2C1 divides the surface of the observation window into multiple micro-regions; This step, building upon S2C, presents a spatially refined implementation method for adaptively adjusting optical distortion parameters. First, the observation window surface is divided into multiple micro-regions according to a preset grid, such as a regular grid of 10×10 or 20×20, or a variable-density grid based on thermal field distribution. The choice of the number of micro-regions is a trade-off: too few micro-regions fail to reflect the spatial non-uniformity of the attached structure, while too many increase computational load and noise sensitivity. Typically, a finer grid is used in the central region of the observation window, which is also the key area for primary visual imaging.

[0066] S2C2: For each micro-region, obtain the difference analysis results of its local light field characteristics, and determine the evolution characteristics of the attached structure in the micro-region; In this step, for each micro-region, the local optical field characteristic deviation corresponding to that region is extracted from the aforementioned global difference analysis results; that is, the direction and magnitude of the deviation between the measured optical field and the predicted optical field. Simultaneously, considering the evolution characteristics of the attached structure in that micro-region, including whether the attached structure is currently in a rapid thickening phase or a slow, stable phase, and whether the evolution trend is linear or accelerating, a comprehensive judgment is made on the required adjustment space and speed for the optical distortion parameters of that micro-region.

[0067] S2C3 dynamically determines the calculation weights and reference benchmarks of the optical distortion parameters of each micro-region based on the evolution characteristics of the attached structure in each micro-region. In this step, based on the output of S2C2, a calculation strategy for optical distortion parameters is independently determined for each micro-region. For micro-regions in a rapid evolution phase, such as those where the deposits are rapidly thickening or undergoing morphological changes, a higher calculation weight is assigned, and the measured data from the most recent frames is used as a reference to ensure the timeliness of the correction parameters. For micro-regions in a stable evolution phase, the weight is appropriately reduced, and a moving average of a longer historical window is used as a reference to improve the signal-to-noise ratio. This region-by-region differentiated strategy avoids the trade-off between insufficient response in active regions and excessive noise in stable regions, which is a problem with a globally uniform strategy.

[0068] S2C4 combines the calculated weights and reference benchmarks of the optical distortion parameters of each micro-region to form a spatial distribution map of the calculated weights and reference benchmarks of the optical distortion parameters on the observation window surface. In this step, the system combines the weights and baseline values ​​of all micro-regions into a two-dimensional distribution map according to their spatial location.

[0069] In the technical solution of the above embodiments, spatial adaptive optimization of optical distortion parameters is achieved by dividing micro-regions and dynamically adjusting weights based on the evolution characteristics of local attachments.

[0070] In one example, during silicon rod growth, the non-uniformity of the temperature field and silicon vapor flow field within the furnace on the observation window surface leads to differences in the growth rate and optical properties of the deposits in different regions. The lower region of the observation window, being closer to the molten silicon surface, experiences higher silicon vapor concentration and a larger temperature gradient, potentially forming thicker amorphous silicon deposits with complex optical properties; while the upper region may form thinner crystalline silicon deposits with relatively simpler optical properties. In this case, the observation window surface is first divided into a 10×10 micro-grid. For each micro-region, multiple spectral and polarization dimensions of its transmitted light are continuously acquired, and combined with furnace operating environment parameters such as local temperature and silicon vapor concentration, differences in local light field characteristics are analyzed. By analyzing the polarization state changes and spectral absorption characteristics of the transmitted light, it can be identified that the deposits in the lower micro-region have higher scattering and absorption effects and a faster growth rate, while the deposits in the upper micro-region have lower scattering and absorption effects and a relatively slower growth rate. Based on these determined characteristics of the attachment structure evolution, the computational weights and reference benchmarks for the optical distortion parameters of each micro-region are dynamically determined: for micro-regions with thicker lower attachments and more complex optical properties, higher computational weights are assigned to distortion parameters related to scattering and absorption, and the reference benchmark is adjusted to more accurately reflect the refractive index and extinction coefficient of amorphous silicon; for micro-regions with thinner upper attachments, lower computational weights are assigned, and a reference benchmark closer to crystalline silicon is used. Finally, these dynamically determined computational weights and reference benchmarks for each micro-region are combined into a spatial distribution map. When the main vision system captures an image of the silicon rod, differentiated correction parameters are applied to different pixel regions according to this distribution map.

[0071] In one possible design, step S2C3, which involves dynamically determining the calculation weights and reference benchmarks for the optical distortion parameters of each micro-region based on the evolution characteristics of the attached structure, includes: S2C31, to obtain growth stage information for each micro-region; In this step, building upon the aforementioned micro-region analysis, the dimension of silicon rod growth stages is further introduced to guide the refined selection of parameter adjustment strategies. The projection position of each micro-region on the observation window corresponds to a specific axial position of the silicon rod. During the growth process, different axial positions of the silicon rod are at different growth stages; that is, the lower region near the solid-liquid interface is in the latest crystallization stage, the middle region has undergone a period of cooling and diameter stabilization, and the upper region is nearing completion. Therefore, the growth stage of each micro-region actually refers to the process state of the silicon rod position directly opposite that micro-region, rather than a globally consistent single stage.

[0072] S2C32, based on the growth stage information, selects a parameter adjustment strategy from a preset parameter adjustment strategy library; In this step, the preset parameter adjustment strategy library contains adjustment strategy templates customized for different growth stages. For example, for micro-regions corresponding to the solid-liquid interface, the strategy tends to be high-frequency updates and small adjustments to ensure a rapid response to diameter changes; for micro-regions in the stabilized segment, the strategy tends to be low-frequency updates and mean filtering to reduce the interference of random noise. Based on the growth stage information of each micro-region obtained above, the most suitable strategy template is selected for each micro-region from the strategy library.

[0073] S2C33, based on the parameter adjustment strategy, determines the calculation weights and reference benchmarks; In this step, based on the adjustment strategy selected for each micro-region, the calculation weights of the optical distortion parameters for that micro-region are specifically determined. These weights include the time decay coefficient and the contribution ratio of the current frame, the update cycle of the reference baseline (in frames), and the outlier removal threshold. Finally, the strategy parameters for each micro-region are aggregated into a unified set of execution parameters to drive the subsequent optical distortion parameter calculation process.

[0074] In the technical solution of the above embodiments, by introducing growth stage information to guide the selection of parameter adjustment strategies, the determination of calculation weights and reference benchmarks is deeply coupled with the silicon rod growth process, thereby improving the accuracy of optical distortion correction.

[0075] In summary, the silicon rod growth control method based on ultra-high-definition visual imaging provided in Embodiment 1 of this application, through multispectral polarization imaging analysis of the optical influence of the observation window, generation of light propagation influence parameters through two complementary schemes, and physical-level image distortion correction and accurate diameter calculation, fundamentally severs the causal chain from observation window attachments to image artifacts to erroneous diameter data to PID high-frequency fine-tuning to mechanical micro-vibration to lattice dislocations to yield reduction. Step S2 provides two parameter generation paths: one based on control command feedforward and the other based on environmental parameter modeling, addressing the optical disturbance separation problem from different angles. Each subordinate scheme further elaborates on its corresponding sub-steps, covering a complete technical hierarchy from discrete region prediction and nonlinear fluctuation correction to micro-region spatial adaptive adjustment.

[0076] Example 2 Embodiment 2 of this application also provides a silicon rod growth control system based on ultra-high-definition visual imaging. Figure 7 This is a block diagram illustrating a silicon rod growth control system based on ultra-high-definition visual imaging, according to another exemplary embodiment. (See attached diagram.) Figure 7 The system includes: Information acquisition module 01 is used to acquire multiple spectral and polarization dimensions of the light transmitted through the observation window.

[0077] The parameter generation module 02 is used to analyze the influence of the observation window on light propagation based on the multiple spectral and polarization dimension information and generate light propagation influence parameters.

[0078] Image correction module 03 is used to correct the silicon rod image captured by the main vision system according to the light propagation influence parameters to eliminate artifacts.

[0079] The diameter calculation module 04 is used to perform edge recognition and diameter calculation on the corrected silicon rod image.

[0080] The control parameter adjustment module 05 is used to adjust the control parameters during the silicon rod growth process according to the diameter information of the silicon rod.

[0081] The above five modules work in series according to the data flow direction: the information acquisition module 01 transmits multi-dimensional optical data to the parameter generation module 02, the parameter generation module 02 transmits the generated light propagation influence parameters to the image correction module 03, the image correction module 03 transmits the image after artifact elimination to the diameter calculation module 04, and the diameter calculation module 04 transmits the accurate diameter information to the control parameter adjustment module 05.

[0082] The technical features of the above embodiments can be combined in any way. For the sake of brevity, not all possible combinations of the technical features in the above embodiments are described. However, as long as there is no contradiction in the combination of these technical features, they should be considered to be within the scope of this specification.

[0083] The embodiments described above are merely illustrative of several implementation methods of this application, and while the descriptions are relatively specific and detailed, they should not be construed as limiting the scope of the invention patent. It should be noted that those skilled in the art can make various modifications and improvements without departing from the concept of this application, and these all fall within the protection scope of this application. Therefore, the protection scope of this patent application should be determined by the appended claims.

Claims

1. A method for controlling silicon rod growth based on ultra-high-definition visual imaging, characterized in that, Includes the following steps: Acquire multiple spectral and polarization dimensions of the light transmitted through the observation window; Based on the multiple spectral and polarization dimension information, the influence of the observation window on light propagation is analyzed, and light propagation influence parameters are generated. Based on the light propagation influence parameters, the silicon rod image captured by the main vision system is corrected to eliminate artifacts caused by the viewing window. Edge recognition and diameter calculation are performed on the corrected silicon rod image to obtain the diameter information of the silicon rod; Based on the diameter information of the silicon rod, the control parameters during the silicon rod growth process are adjusted.

2. The method according to claim 1, characterized in that, The step of analyzing the influence of the observation window on light propagation and generating light propagation influence parameters based on the multiple spectral and polarization dimension information includes: Obtain instruction information to adjust heating power and lifting speed; Based on the instruction information, predict the instantaneous change in the optical properties of the surface deposits on the observation window; The calibrated light propagation information is obtained by compensating for the instantaneous changes in the optical properties of the attached material from the multiple spectral and polarization dimensions of the transmitted light through the observation window. Based on the calibrated light propagation information, the influence of the observation window on light propagation is analyzed, and light propagation influence parameters are generated.

3. The method according to claim 1, characterized in that, The steps for obtaining multiple spectral and polarization dimension information of the transmitted light through the observation window include: During the growth of silicon rods, when there are transient micro-fluctuations on the surface of molten silicon or in the distribution of silicon vapor in the furnace, a high-speed imaging unit continuously acquires a sequence of original images containing information of different spectra and polarization dimensions at a sampling rate higher than the frequency of the transient micro-fluctuations. The original image sequence is timestamped, and spatial registration is performed on the image at each timestamp based on the timestamps to eliminate spatial misalignment caused by time asynchrony due to mechanical switching of filters during image acquisition. Based on the spatially registered image sequence, multiple spectral and polarization dimension information at each timestamp is extracted.

4. The method according to claim 1, characterized in that, The step of analyzing the influence of the observation window on light propagation and generating light propagation influence parameters based on the multiple spectral and polarization dimension information includes: Obtain furnace operating environment parameters, and predict the evolution trend of the surface deposits on the observation window based on the furnace operating environment parameters; The local light field characteristics of the transmitted light from the observation window are extracted, and the differences in the local light field characteristics are analyzed based on the predicted evolution trend of the attached structure. Based on the difference analysis results, the calculation weights and reference benchmarks of the optical distortion parameters are adaptively adjusted. Based on the adaptively adjusted optical distortion parameters and the reference benchmark, and using the multiple spectral and polarization dimension information, the influence of the observation window on light propagation is analyzed, and light propagation influence parameters are generated.

5. The method according to claim 4, characterized in that, The step of predicting the evolution trend of the surface deposits on the observation window based on the furnace operating environment parameters includes: Acquire local temperature and silicon vapor concentration information of multiple discrete regions on the surface of the observation window; Based on the local temperature information and silicon vapor concentration information of the multiple discrete regions, the growth rate and decomposition rate of the attachment in each discrete region are calculated, and the thickness change and morphological change of the attachment in each discrete region are predicted. The changes in the thickness and morphology of the attachments in each discrete region are combined to form a spatial distribution map of the evolution of the attachment structure on the observation window surface.

6. The method according to claim 4, characterized in that, The step of predicting the evolution trend of the surface deposits on the observation window based on the furnace operating environment parameters includes: Acquire real-time fluctuation data of furnace pressure, gas composition, and cooling water flow rate; Based on the real-time fluctuation data, identify the fluctuation type and fluctuation amplitude; For the identified fluctuation type and fluctuation amplitude, the corresponding attachment structure evolution adjustment factor is selected from the preset nonlinear response rule base; The attachment structure evolution adjustment factor is applied to the attachment structure evolution trend predicted based on conventional environmental parameters to obtain the corrected attachment structure evolution trend.

7. The method according to claim 6, characterized in that, The step of selecting the corresponding attachment structure evolution adjustment factor from the preset nonlinear response rule base includes: Obtain the process parameters for the current silicon rod growth stage, and based on the process parameters, assess the sensitivity of the current silicon rod growth stage to the adjustment factor of the attachment structure evolution. Multiple attachment structure evolution adjustment factors that match the identified fluctuation type and fluctuation amplitude are selected from the preset nonlinear response rule base; Based on the sensitivity, the selected multiple attachment structure evolution adjustment factors are weighted or ranked, and the attachment structure evolution adjustment factor with the highest weight or the highest ranking is selected.

8. The method according to claim 4, characterized in that, The steps for adaptively adjusting the calculation weights and reference benchmarks of the optical distortion parameters include: The surface of the observation window is divided into multiple micro-regions; For each micro-region, the difference analysis results of its local light field characteristics are obtained, and the evolution characteristics of the attached structure in the micro-region are determined; Based on the evolution characteristics of the attached structure in each micro-region, the calculation weights and reference benchmarks of the optical distortion parameters of that micro-region are dynamically determined; The calculated weights and reference standards of the optical distortion parameters in each micro-region are combined to form a spatial distribution map of the calculated weights and reference standards of the optical distortion parameters on the observation window surface.

9. The method according to claim 8, characterized in that, The step of dynamically determining the calculation weights and reference benchmarks of the optical distortion parameters of each micro-region based on the evolution characteristics of the attached structure includes: Obtain growth stage information for each micro-region; Based on the growth stage information, a parameter adjustment strategy is selected from a preset parameter adjustment strategy library; Based on the parameter adjustment strategy, the calculation weights and reference benchmarks are determined.

10. A silicon rod growth control system based on ultra-high-definition visual imaging, characterized in that, The system includes: The information acquisition module is used to acquire multiple spectral and polarization dimensions of the light transmitted through the observation window; The parameter generation module is used to analyze the influence of the observation window on light propagation based on the multiple spectral and polarization dimension information, and generate light propagation influence parameters; The image correction module is used to correct the silicon rod image captured by the main vision system according to the light propagation influence parameters, so as to eliminate the artifacts caused by the viewing window to the image. The diameter calculation module is used to perform edge recognition and diameter calculation on the corrected silicon rod image to obtain the diameter information of the silicon rod; The control parameter adjustment module is used to adjust the control parameters during the silicon rod growth process based on the diameter information of the silicon rod.

Images