Process parameter regulation and control method and system for single crystal silicon carbide laser modification stripping

Abstract

The invention belongs to a laser modification stripping technology of single crystal silicon carbide, and relates to a method and a system for regulating and controlling technological parameters of laser modification stripping of single crystal silicon carbide. The method comprises the following steps: setting technological parameters of laser modification, and carrying out laser modification on single crystal silicon carbide; meanwhile, the three-dimensional shape of the crack nucleation of the laser modification area is monitored in real time through a sweep-frequency optical coherence tomography method; feature information is extracted based on the three-dimensional shape of the crack nucleation; the characteristic information comprises hole size, crack depth and distribution density characteristics; inputting the process parameters and the feature information into a pre-constructed convolutional neural network model, and outputting a prediction morphology of the crack nucleation; and based on the feature information, performing similarity comparison on the predicted morphology and the ideal morphology, if the similarity is lower than a preset similarity, dynamically adjusting the process parameters, performing iterative optimization until the similarity reaches the preset similarity, and determining that the adjustment of the process parameters is qualified. The problems of lack of real-time monitoring and feedback and uncontrollable crack nucleation and expansion are solved.

Description

TECHNICAL FIELD

[0001] The application belongs to the technical field of laser modification and exfoliation of single crystal SiC, and particularly relates to a process parameter regulation method and system for laser modification and exfoliation of single crystal silicon carbide. BACKGROUND

[0002] As an important wide-bandgap semiconductor material, single crystal SiC has great application potential in the fields of power devices, radio frequency devices, and optoelectronic devices due to its high hardness, high thermal conductivity, high breakdown electric field strength, and excellent high-temperature resistance and radiation resistance. However, the high brittleness and high hardness of single crystal SiC make it extremely difficult to process. Traditional mechanical cutting methods are prone to crack propagation and surface damage during processing, resulting in serious material loss, low processing precision, and high processing cost, which cannot meet the demand for high-quality wafers in modern semiconductor manufacturing.

[0003] As a new processing method, laser modification and exfoliation technology can induce the formation of a modification layer at a specific location in single crystal SiC, thereby achieving controlled fracture and providing a solution to the above problems. This technology uses the high energy density and precise control ability of laser to create a weakly bonded region at a specific depth in the crystal, which guides the crack to propagate along the predetermined direction, and finally realizes the efficient exfoliation of the wafer. Compared with traditional mechanical processing methods, laser modification and exfoliation technology can not only significantly reduce material loss, but also effectively improve processing efficiency and wafer quality, reduce subsequent grinding and polishing processes, and thus reduce overall processing cost.

[0004] However, existing laser modification and exfoliation technology still faces many challenges: Firstly, the crack nucleation process is affected by the complex coupling of various process parameters (such as laser power, scanning speed, pulse frequency, etc.), making it difficult to establish an accurate dynamic model, which makes it difficult to predict and control the nucleation and propagation behavior of the crack in real time during actual processing.

[0005] Secondly, traditional monitoring methods (such as scanning electron microscope SEM and white light interferometer) need to interrupt the processing flow, transfer the single crystal SiC sample to a special detection device for analysis, resulting in a long detection period and the inability to achieve real-time feedback, which cannot timely adjust the processing parameters to optimize the crack nucleation process, severely restricting the further improvement of processing efficiency and quality.

[0006] In addition, in the single crystal SiC laser modification exfoliation, the characterization of the crack morphology can be essentially regarded as the analysis of the specific image features formed in the processing. The crack is easy to deviate from the predetermined cleavage plane during the expansion process, resulting in the decline of the wafer surface quality after exfoliation, and affecting its application in high-end device manufacturing. In terms of crack morphology prediction, due to the complex coupling effect of various process parameters on the crack nucleation process, it is difficult to accurately describe the dynamic behavior of the traditional physical model.

[0007] In view of the above problems, it is particularly important to develop a method capable of real-time monitoring and dynamic control of the crack nucleation process. SUMMARY

[0008] The purpose of the present application is to provide a process parameter regulation method and system for single crystal silicon carbide laser modification exfoliation, which solves the problems of lack of real-time monitoring and feedback, and uncontrollable crack nucleation and expansion.

[0009] The present application is realized by the following technical solutions: The present application discloses a process parameter regulation method for single crystal silicon carbide laser modification exfoliation, comprising the following steps: S1, setting the process parameters of laser modification, and performing laser modification on single crystal silicon carbide; at the same time, using swept frequency optical coherence tomography to monitor the three-dimensional morphology of crack nucleation in the laser modification area in real time; S2, extracting feature information based on the three-dimensional morphology of crack nucleation; the feature information includes hole size, crack depth and distribution density characteristics; S3, inputting the process parameters of S1 and the feature information extracted in S2 into a pre-constructed convolutional neural network model, and outputting the predicted morphology of crack nucleation; S4, based on the feature information, comparing the predicted morphology with the ideal morphology in terms of similarity, if the similarity is lower than the preset similarity, dynamically adjusting the process parameters, repeating steps S1-S3, and iterating optimization until the similarity reaches the preset similarity, then the process parameter adjustment is qualified.

[0010] Further, in S1, the process parameters of laser modification include laser power, repetition frequency, scanning speed and pulse width.

[0011] Further, in S1, the three-dimensional morphology of crack nucleation in the laser modification area is monitored in real time by using swept frequency optical coherence tomography, specifically: Through coaxial light path design, the processing and monitoring are synchronized, the MHz level scanning rate of the swept frequency source is matched with the femtosecond laser processing frequency, and the three-dimensional morphology of crack nucleation in the laser modification area is obtained in real time.

[0012] Further, in S2, the feature information is extracted based on the three-dimensional morphology of crack nucleation, specifically: In the feature extraction stage of the swept-source optical coherence tomography method, the spectral features of the crack nucleation region are analyzed by using Fourier transform, the frequency domain signal is mapped to the spatial domain, and the hole size, crack depth and distribution density characteristics are used to reflect the crack nucleation region.

[0013] Further, in S3, the construction process of the pre-constructed convolutional neural network model is: Obtain historical data, the historical data including the process parameters of laser modification and the three-dimensional morphology of crack nucleation of the corresponding laser modification region; Feature extraction is performed on the three-dimensional morphology of crack nucleation to obtain feature information; The process parameters of laser modification and the feature information are input into the convolutional neural network model, and the convolutional neural network model is trained based on the cross-entropy loss function and the PyTorch deep learning framework; The output is a feature vector of the 4-dimensional process parameters corresponding to the three-dimensional morphology of crack nucleation.

[0014] Further, in S4, based on the feature information, the predicted morphology is compared with the ideal morphology in terms of similarity, specifically: The hole size, crack depth and distribution density are respectively standardized and normalized to the [0, 1] interval to obtain the actual value corresponding to each feature information ; For each feature information, the absolute value similarity of the actual value and the ideal value is calculated. Three features are assigned a weight ratio alpha, and the weighted combination of the similarity is calculated to obtain the similarity. If the similarity is greater than or equal to 80%, it is determined that the process parameter adjustment is qualified. If the similarity is less than 80%, trigger feedback and adjust the process parameters, and cycle comparison until the similarity meets the standard.

[0015] Further, the similarity calculation formula is as follows:

[0016] Wherein, S is the similarity, d is the hole size, h is the crack depth, p is the distribution density; is the actual value corresponding to the i-th feature information; is the ideal value corresponding to the i-th feature information; is the weight ratio corresponding to the i-th feature information.

[0017] The application discloses a process parameter regulation system for single crystal silicon carbide laser modification and stripping, which comprises a femtosecond laser processing module, an SS-OCT monitoring module, a CNN control module and a similarity comparison module. A femtosecond laser processing module is used to set process parameters of laser modification and perform laser modification on single crystal silicon carbide; An SS-OCT monitoring module is used to monitor the three-dimensional morphology of crack nucleation in the laser modification area in real time; A feature extraction module is used to extract feature information including hole size, crack depth and distribution density based on the three-dimensional morphology of crack nucleation; A CNN control module is used to input the process parameters and extracted planar features into a pre-constructed qualified convolutional neural network model, and output the predicted morphology of crack nucleation; A similarity comparison module is used to compare the key features of the predicted morphology and the ideal morphology to obtain a similarity; if the similarity is lower than a preset similarity, the process parameters are dynamically adjusted, and iterative optimization is performed until the similarity reaches the preset similarity, and then the process parameter adjustment is qualified.

[0018] Further, based on the feature information, the predicted morphology is compared with the ideal morphology, specifically: The hole size, crack depth and distribution density are standardized respectively, and after normalization, they are mapped to the [0, 1] interval to obtain the actual value corresponding to each feature information ; For each feature information, the absolute value similarity of the actual value and the ideal value is calculated; Three features are assigned a weight ratio α, and the weighted combination of the similarity is calculated to obtain the similarity; If the similarity is greater than or equal to 80%, it is determined that the process parameter adjustment is qualified; If the similarity is less than 80%, feedback is triggered and the process parameters are adjusted, and the comparison is repeated until the similarity meets the standard.

[0019] Further, the similarity calculation formula is as follows:

[0020] Wherein, S is the similarity, d is the hole size, h is the crack depth, p is the distribution density; is the actual value corresponding to the i-th feature information; is the ideal value corresponding to the i-th feature information; is the weight ratio corresponding to the i-th feature information.

[0021] Compared with the prior art, the present application has the following beneficial technical effects: The application discloses a process parameter regulation method for single crystal silicon carbide laser modification stripping, combines sweep frequency optical coherence tomography (SS-OCT) technology and laser modification technology, realizes synchronization of processing and monitoring through coaxial light path design, matches the MHz-level scanning rate of SS-OCT with the frequency of femtosecond laser processing, and realizes real-time acquisition of three-dimensional topographic data such as the depth of crack nucleation and hole distribution. The method solves the problem of crack nucleation dynamic modeling in the prior art, provides a high-resolution and real-time monitoring method for accurate characterization of crack nucleation, and proposes a crack topography prediction model based on a convolutional neural network (CNN). The process parameters and feature information extracted by SS-OCT are used as input, and the predicted topography of crack nucleation is output. The predicted topography is compared with an ideal topography (such as an ellipsoidal hole or a crack direction along a cleavage plane), and laser parameters are dynamically adjusted to realize accurate control of crack nucleation.

[0022] A closed-loop feedback control system is constructed, real-time feedback of SS-OCT monitoring data and CNN prediction results is realized, process parameters are adjusted, the spatial variation of the light beam is dynamically regulated, and the process parameters are real-time regulated by a computer. The stability and controllability of crack nucleation are significantly improved, crack deviation from the predetermined cleavage plane is avoided, and the surface quality of the wafer after stripping is ensured.

[0023] The application further discloses a process parameter regulation system for single crystal silicon carbide laser modification stripping, which comprises a femtosecond laser processing module, an SS-OCT monitoring module, a CNN control module and a similarity comparison module. The SS-OCT system and the laser modification platform are deeply integrated to ensure that the monitoring and processing positions are strictly synchronized and the interference of the additional detection station on the processing flow is avoided. The SS-OCT monitoring module realizes real-time capture of dynamic images of the modification area through a CMOS image sensor, realizes seamless connection of data acquisition, processing and control algorithms in combination with Matlab / Simulink, and provides efficient technical support for real-time regulation of crack nucleation. Specifically, the following aspects are included. 1. Real-time monitoring is realized in sub-second through the penetration imaging capability of SS-OCT, processing interruption is avoided, and the detection cycle is significantly shortened.

[0024] 2. The penetration imaging capability of SS-OCT is utilized to solve the problem that the internal topography cannot be acquired in real time by traditional monitoring methods (such as SEM), and the intelligent prediction and dynamic optimization of the crack nucleation process are realized in combination with the deep learning capability of CNN. Through multi-dimensional data fusion, comprehensive data support is provided for accurate modeling of crack nucleation.

[0025] 3. The CNN model is combined with feedback control technology to solve the problem of multi-parameter coupled modeling, and the prediction accuracy is above 90 %.

[0026] 4. By replacing the traditional Gaussian light beam with a linear light beam and combining the SLM to dynamically regulate the light intensity distribution, a single scanning large-area uniform processing is realized, and the processing time is significantly reduced; by optimizing the crack nucleation morphology, the warping degree, bending degree and roughness of the wafer surface are maximally reduced, and the overall processing precision is improved.

[0027] 5. After actual processing by the process parameters obtained by the regulation system, the peeling surface roughness of the single crystal silicon carbide is reduced by 30%, the warping degree is less than or equal to 70 microns, and it is suitable for 6-inch and above SiC wafers. BRIEF DESCRIPTION OF DRAWINGS

[0028] Figure 1 A monitoring flowchart of a single crystal SiC laser modification and peeling process based on SS-OCT technology real-time monitoring and CNN online control is provided for the embodiments of the present application. Figure 2 A SS-OCT technology real-time monitoring device for laser modification of SiC is provided for the embodiments of the present application. Figure 3 A polarization converter in the SS-OCT divides the interference principle diagram of the target light beam and the reference light beam. Figure 4 A crack nucleation controller block diagram is provided for the embodiments of the present application, including a CNN prediction module, a parameter adjustment logic and a feedback execution process.

[0029] In the figure, 1 is a femtosecond laser; 2 is a half-wave plate; 3 is a polarization beam splitter; 4 is a concave lens; 5 is a convex lens; 6 is a mirror; 7 is a spatial light modulator; 8 is a galvanometer; 9 is a white LED light source; 10 is a three-axis moving platform; 11 is a sample; 12 is an objective lens; 13 is a dichroic mirror; 14 is an infrared cutoff filter; 15 is a CMOS image sensor; 16 is a polarization converter a; 17 is a single-mode fiber ring; 18 is a swept light source; 19 is a polarization converter b; 20 is a single-mode fiber coupler; 21 is a photodetector; 22 is a data acquisition device; 23 is a computer; 24 is a laser beam; 25 is a digital micromirror device; and 26 is a convolutional neural network model. DETAILED DESCRIPTION

[0030] In order to make the purpose, technical scheme and advantages of the present application clearer and more clear, the following will be further described in detail in combination with the drawings and examples. It should be understood that the specific examples described herein are only used to explain the present application and are not used to limit the present application, that is, the described examples are only a part of the examples of the present application, but not all examples.

[0031] The components described and illustrated in the accompanying drawings and examples of the present application can be arranged and designed in a wide variety of different configurations, therefore, the following detailed description of the embodiments of the present application provided in the accompanying drawings is not intended to limit the scope of the application, but merely represents one selected embodiment of the present application. Based on the drawings and examples of the present application, all other embodiments obtained by those skilled in the art without creative work fall within the scope of the present application.

[0032] SS-OCT, namely Swept Source Optical Coherence Tomography, is a kind of Optical Coherence Tomography technology, which has the advantages of fast scanning speed, high sensitivity and strong penetration, and is widely used in the medical field.

[0033] Principle: SS-OCT uses a wavelength-swept light source to temporarily sweep the wavelength, irradiates the output light onto the sample 11, the scattered and reflected light of the sample 11 combines with the reference light to produce interference light, and then the interference light is detected by a photodetector 21. Finally, the depth direction image of the sample 11 is obtained by performing inverse Fourier transform on the obtained wavelength information.

[0034] As shown in Figure 2 , the SS-OCT technology real-time monitoring device for laser modification of SiC specifically includes a femtosecond laser processing module, an SS-OCT monitoring module, and a CNN control module.

[0035] The femtosecond laser processing module includes a femtosecond laser 1, a digital micromirror device 25, and an optical path assembly.

[0036] The optical path assembly includes a half-wave plate 2, a polarization beam splitter 3, a concave lens 4, a convex lens 5, a mirror 6, a spatial light modulator 7, a galvanometer 8, and a convex lens assembly.

[0037] The convex lens assembly includes two convex lenses arranged in a plane opposite to each other.

[0038] As shown in Figure 4 , the femtosecond laser 1 generates a laser beam 24, which passes through the digital micromirror device 25 and reaches the dichroic mirror 13, is reflected into the objective lens 12, and then enters the sample 11 (single crystal silicon carbide) placed on the three-axis moving platform 10. The digital micromirror device 25 is connected in parallel to one side of the optical path assembly.

[0039] Processing and monitoring principle of the femtosecond laser processing module: the femtosecond laser processing module realizes non-thermal melting modification of the single crystal silicon carbide through high-precision optical modulation and focusing. Figure 2As shown, the specific process of light path transmission and regulation is as follows: the femtosecond laser 1 outputs ultrashort pulse laser, which first passes through the half-wave plate 2, and the polarization direction of the laser is adjusted by rotating the wave plate angle to optimize the energy distribution efficiency of the subsequent polarization beam splitter 3. The laser is split by the polarization beam splitter 3 according to the preset ratio, part of the energy is used for real-time power monitoring, and the main beam enters the beam expansion system. The beam expansion is carried out by the combination of concave lens 4 and convex lens 5. The expanded laser enters the spatial light modulator 7 through the mirror 6, and the laser wavefront phase is dynamically modulated based on the computer-generated hologram, which converts the traditional Gaussian beam into a linear beam to realize uniform energy deposition. The modulated laser is turned by the mirror, then enters the galvanometer 8 system through the convex lens assembly and the mirror, and the laser scanning path is controlled by the two-dimensional deflection of the high-speed galvanometer to ensure the coverage accuracy of the processing area. A white LED light source 9 is arranged at the bottom of the workbench for transmission observation, and a convex lens is arranged above the white LED light source 9. The scanned laser is focused to the inside of the single crystal silicon carbide sample 11 on the three-axis moving platform 10 through the convex lens, and the focusing depth is accurately controlled by the Z-axis displacement. The three-axis moving platform 10 and the galvanometer 8 move synchronously to realize high-speed scanning of a large area, ensuring the uniformity and processing efficiency of the modified layer.

[0040] As shown in Figure 2 , it is the interference principle diagram of the polarization converter in the SS-OCT, which includes: polarization converter a16, swept source 18, polarization converter b19, photodetector 21; the swept source 18 emits output light to the polarization converter a16 and the polarization converter b19, one light source directly reaches the photodetector 21 through the polarization converter b19, and the polarization converter a16 refracts another part of the light source to the mirror, and this part of the light source is transmitted into the sample 11. At the same time, the sample 11 also reflects the light to the polarization converter a16 again, which is reflected to the photodetector 21, and the photodetector 21 detects the interference light.

[0041] As shown in Figure 2 and Figure 3 , the SS-OCT monitoring module includes a CMOS image sensor 15, a polarization converter a16, a single-mode fiber circulator 17, a swept source 18, a polarization converter b19, a single-mode fiber coupler 20, a photodetector 21, and a data acquisition device 22.

[0042] Principle: The machining beam output by the femtosecond laser 1 is split by the dichroic mirror 13, and one beam is reflected to the monitoring light path as the signal beam of the swept source optical coherence tomography (SS-OCT) system. The beam passes through the convex lens and the mirror in turn, passes through the infrared cutoff filter 14, suppresses the residual infrared noise of the femtosecond laser, and ensures that the CMOS image sensor 15 only receives the detection light signal of the SS-OCT. The CMOS image sensor 15 captures the scattering light signal of the modified area in real time at a frame rate of MHz level for auxiliary positioning and surface topography monitoring. Another part of the laser is used as a reference beam, which is reflected by the dichroic mirror 13 and enters the core light path of the SS-OCT: the polarization converter a 16 dynamically adjusts the polarization state of the reference beam to maximize the interference efficiency with the backscattered light of the sample 11. The single-mode fiber circulator 17 isolates the reflected noise and ensures one-way transmission of the light signal to the interference module. The swept source 18 outputs to the polarization converter a 16 and the polarization converter b 19, and the single-mode fiber coupler 20 couples the backscattered light of the reference beam and the machining beam of the sample 11 to generate an interference signal. The photodetector 21 detects the time-domain changes of the interference signal and eliminates common-mode noise. The data acquisition device 22 converts the frequency domain signal into depth-resolved spatial domain data through inverse Fourier transform, and finally reconstructs the three-dimensional morphology of crack nucleation on the computer 23 through algorithm.

[0043] The data acquisition device 22 is connected with the host computer 23, and the host computer 23 is installed with a feature extraction module and a CNN control module.

[0044] The feature extraction module is used for mapping the frequency domain signal collected by the data acquisition device 22 to the spatial domain through inverse Fourier transform, reflecting the hole size, crack depth and distribution density characteristics.

[0045] The photodetector 21 detects the gray scale change of the internal defects of the sample 11, and converts the frequency domain to the spatial domain through inverse Fourier transform, so as to obtain a real-time image.

[0046] As Figure 4As shown, the three-axis mobile platform 10 is integrated with a high-brightness white LED light source 9 at the bottom, which forms a uniform illumination beam through a collimating lens, and then penetrates through the three-axis mobile platform 10 and the modified region of the single crystal silicon carbide sample 11, and finally is captured by the CMOS image sensor 15 in real time to obtain the surface topography. The laser beam 24 output by the femtosecond laser 1 is incident to the digital micromirror device 25, and the intensity and direction of the light beam are modulated by the binary hologram. The modulated laser is split by the dichroic mirror, and the main light path is focused by the focusing objective lens to accurately focus the laser inside the single crystal SiC sample 11 to induce nonlinear absorption to form a modified layer, and the monitoring light path is reflected to the CMOS image sensor 15. The image data obtained from the CMOS image sensor 15 is preprocessed to extract the feature factors related to the crack topography, which is transmitted to the convolutional neural network model 26 in real time for analysis and processing to evaluate the similarity between the current crack topography and the ideal transverse type II crack. According to the prediction result of the convolutional neural network model 26, the feedback control mechanism dynamically adjusts the output parameters of the femtosecond laser to realize accurate control of the femtosecond laser modification of the single crystal silicon carbide wafer.

[0047] In summary, the functions of the femtosecond laser processing module are as follows: the femtosecond laser 1 emits femtosecond pulses with specific wavelength, pulse width and repetition frequency, generates a linear light beam through a spatial light modulator 7, and expands and focuses the light beam to the single crystal SiC sample 11 to achieve uniform processing on a large area. The functions of the SS-OCT monitoring module are as follows: using a swept source 18 and a double-beam interference structure, the three-dimensional morphology of crack nucleation is monitored in real time, the interference of the processing light on the monitoring signal is avoided, the dynamic images of the modified region are captured in real time by the CMOS image sensor 15, and the integration of data acquisition, processing and control algorithm is realized by Matlab / Simulink; by detecting the backscattering light signals at different depths inside the sample 11, the three-dimensional morphology of crack nucleation is reconstructed by inverse Fourier transform. During the processing, the SS-OCT updates the three-dimensional morphology data every 0.5 s; The CNN control module is trained based on the PyTorch framework, the input layer includes 4 process parameters (laser power, repetition frequency, scanning speed, pulse width) and 3 SS-OCT features (hole size, crack depth, distribution density). The output layer is a feature vector of the 4-dimensional process parameters corresponding to the obtained crack topography. The model training uses the cross-entropy loss function, the optimizer is Adam, and the training accuracy reaches 92%.

[0048] The CNN model is used to predict the crack nucleation morphology, and the CNN prediction result is compared with the ideal morphology (ellipsoidal hole, crack direction deviation <5°). If the similarity is <80%, the SLM hologram is adjusted to optimize the linear beam intensity distribution, and the process parameters such as laser power and repetition frequency are adjusted. Through continuous iterative adjustment, the similarity of the crack nucleation morphology and the ideal morphology is ensured to be more than 80%, the crack nucleation morphology is ensured to be stable and controllable, and the parameters are locked to complete the modification layer processing.

[0049] Based on the feature information, the predicted morphology is compared with the ideal morphology, specifically: The hole size, crack depth and distribution density are standardized respectively, and after normalization, the actual value of each feature is obtained ; For each feature information, the absolute value similarity of the actual value and the ideal value is calculated. Three features are assigned a weight ratio α, and the weighted combination of the similarity is calculated to obtain the comprehensive similarity S ; The comprehensive similarity calculation formula is as follows:

[0050] Wherein, S is the similarity, d is the hole size, h is the crack depth, p is the distribution density; is the actual value corresponding to the i-th feature information; is the ideal value corresponding to the i-th feature information; is the weight ratio corresponding to the i-th feature information.

[0051] If the similarity is ≥80%, it is determined that the process parameter adjustment is qualified, and subsequent processing can be carried out; If the similarity is <80%, trigger feedback and adjust the process parameters, and compare in cycles until the similarity meets the standard.

[0052] The following is a verification example of the present application: The method is implemented on a 6-inch 4H-SiC wafer sample, and the ideal morphology process parameters are controlled. Then, the 6-inch 4H-SiC wafer is processed, and the surface roughness Ra is reduced from 1.2 μm of the traditional method to 0.5 μm, the warpage is optimized from 200 μm to 70 μm, and the processing efficiency is improved by 40%.

[0053] It should be pointed out finally that the above embodiments are only used to illustrate the technical solutions of the present application but not to limit it. Although the present application has been described in detail with reference to the above embodiments, it should be understood by those skilled in the art that the specific embodiments of the present application can be modified or replaced equivalently without departing from the spirit and scope of the present application, and any modification or equivalent replacement should be covered in the protection scope of the claims of the present application.

Claims

1. A method for controlling process parameters in laser-modified exfoliation of single-crystal silicon carbide, characterized in that, Includes the following steps: S1. Set the process parameters for laser modification and perform laser modification on single-crystal silicon carbide; at the same time, use swept-frequency optical coherence tomography to monitor the three-dimensional morphology of crack nucleation in the laser-modified region in real time. S2. Extract feature information based on the three-dimensional morphology of crack nucleation; the feature information includes pore size, crack depth, and distribution density. S3. Input the process parameters of S1 and the feature information extracted in S2 into the pre-constructed convolutional neural network model, and output the predicted morphology of the crack nucleus. S4. Based on the feature information, compare the predicted morphology with the ideal morphology. If the similarity is lower than the preset similarity, dynamically adjust the process parameters and repeat steps S1-S3. Iterate and optimize until the similarity reaches the preset similarity, then the process parameter adjustment is qualified.

2. The method for controlling process parameters of laser-modified exfoliation of single-crystal silicon carbide according to claim 1, characterized in that, In S1, the process parameters for laser-based heat treatment include laser power, repetition frequency, scanning speed, and pulse width.

3. The method for controlling process parameters of laser-modified exfoliation of single-crystal silicon carbide according to claim 1, characterized in that, In S1, the three-dimensional morphology of crack nucleation in the laser-modified region is monitored in real time using swept-frequency optical coherence tomography. Specifically: By using a coaxial optical path design to achieve synchronous processing and monitoring, and by matching the MHz-level scanning rate of the frequency-sweeping light source with the femtosecond laser processing frequency, the three-dimensional morphology of crack nucleation in the laser-modified region can be obtained in real time.

4. The method for controlling process parameters of laser-modified exfoliation of single-crystal silicon carbide according to claim 1, characterized in that, In S2, the extraction of feature information based on the three-dimensional morphology of crack nucleation specifically includes: In the feature extraction stage, the swept-frequency optical coherence tomography method uses Fourier transform to analyze the spectral characteristics of the crack nucleation region, mapping the frequency domain signal to the spatial domain to reflect the characteristics of hole size, crack depth, and distribution density.

5. The method for controlling process parameters of laser-modified exfoliation of single-crystal silicon carbide according to claim 1, characterized in that, In S3, the construction process of the pre-built convolutional neural network model is as follows: Acquire historical data, including the process parameters of laser refining and the corresponding three-dimensional morphology of crack nucleation in the laser refining region; Feature extraction is performed on the three-dimensional morphology of crack nucleation to obtain feature information; The process parameters and feature information of laser-modified heat treatment are input into the convolutional neural network model, and the convolutional neural network model is trained based on the cross-entropy loss function and the PyTorch deep learning framework. The output is a feature vector of the 4D process parameters corresponding to the three-dimensional morphology of the crack nucleation.

6. The method for controlling process parameters of laser-modified exfoliation of single-crystal silicon carbide according to claim 1, characterized in that, In S4, based on feature information, the predicted shape is compared with the ideal shape in terms of similarity, specifically as follows: The pore size, crack depth, and distribution density are standardized and normalized to the [0,1] interval to obtain the actual value corresponding to each feature. ; For each feature, calculate the actual value. Compared with ideal value The absolute value similarity; Assign a weight ratio α to the three features, and calculate the weighted sum of similarities to obtain the similarity score; If the similarity is ≥80%, the process parameter adjustment is deemed qualified. If the similarity is less than 80%, feedback is triggered and process parameters are adjusted. The comparison is repeated until the similarity meets the standard.

7. The method for controlling process parameters of laser-modified exfoliation of single-crystal silicon carbide according to claim 6, characterized in that, The similarity calculation formula is as follows: in, S For similarity, d For the hole size, h The crack depth. ρ The distribution density; This represents the actual value corresponding to the i-th feature. This represents the ideal value corresponding to the i-th feature. This represents the weight ratio corresponding to the i-th feature information.

8. A process parameter control system for laser-modified exfoliation of single-crystal silicon carbide, characterized in that, It includes a femtosecond laser processing module, an SS-OCT monitoring module, a CNN control module, and a similarity comparison module; The femtosecond laser processing module is used to set the process parameters for laser modification and to perform laser modification on single-crystal silicon carbide. The SS-OCT monitoring module is used to monitor the three-dimensional morphology of crack nucleation in the laser-modified region in real time. The feature extraction module is used to extract feature information based on the three-dimensional morphology of crack nuclei, including pore size, crack depth, and distribution density features; The CNN control module is used to input process parameters and extracted planar features into a pre-constructed qualified convolutional neural network model and output the predicted morphology of the crack kernel. The similarity comparison module is used to compare the predicted morphology with the ideal morphology in terms of key feature similarity to obtain the similarity. If the similarity is lower than the preset similarity, the process parameters are dynamically adjusted and iteratively optimized until the similarity reaches the preset similarity, then the process parameter adjustment is qualified.

9. The process parameter control system for laser-modified exfoliation of single-crystal silicon carbide according to claim 8, characterized in that, Based on feature information, the predicted shape is compared with the ideal shape in terms of similarity, specifically as follows: The pore size, crack depth, and distribution density are standardized and normalized to the [0,1] interval to obtain the actual value corresponding to each feature. ; For each feature, calculate the actual value. Compared with ideal value The absolute value similarity; Assign a weight ratio α to the three features, and calculate the weighted sum of similarities to obtain the similarity score; If the similarity is ≥80%, the process parameter adjustment is deemed qualified. If the similarity is less than 80%, feedback is triggered and process parameters are adjusted. The comparison is repeated until the similarity meets the standard.

10. The process parameter control system for laser-modified exfoliation of single-crystal silicon carbide according to claim 9, characterized in that, The similarity calculation formula is as follows: in, S For similarity, d For the hole size, h The crack depth. ρ The distribution density; This represents the actual value corresponding to the i-th feature. This represents the ideal value corresponding to the i-th feature. This represents the weight ratio corresponding to the i-th feature information.

Images