An adaptive beam shaping method for internal three-dimensional structures of single crystal silicon carbide
By employing adaptive beam shaping and multimodal real-time monitoring, and utilizing a physical information neural network model, the femtosecond laser processing process is controlled in real time. This solves the problems of poor processing quality and low consistency of the internal three-dimensional structure of single-crystal silicon carbide, achieving a high aspect ratio and low damage processing effect.
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- YUNNAN UNIV
- Filing Date
- 2026-03-11
- Publication Date
- 2026-06-09
Smart Images
Figure CN122165023A_ABST
Abstract
Description
Technical Field
[0001] This invention relates to the field of laser processing technology, and more specifically, to an adaptive beam shaping method for the internal three-dimensional structure of single-crystal silicon carbide. Background Technology
[0002] Femtosecond lasers, with their ultrashort pulse widths and extremely high peak power, have become a key tool for three-dimensional direct-write processing within transparent materials. By modulating the beam wavefront using a spatial light modulator (SLM), it can be shaped into a specific beam configuration (e.g., a Bessel beam to extend the depth of focus). However, this approach, aimed at achieving a predetermined, static light field distribution, has fundamental limitations when applied to brittle and difficult-to-machine materials like single-crystal silicon carbide, which have high thermal conductivity and hardness. The interaction between the femtosecond laser and single-crystal silicon carbide is a dynamic, nonlinear, transient process accompanied by complex phenomena such as plasma generation, phase transitions, and thermal accumulation. Current SLM-based processing often employs static or pre-programmed control methods, fixing the beam shaping parameters before processing and failing to respond to real-time changes in the internal state during the process. Furthermore, existing process monitoring methods (such as transmission imaging) primarily acquire contour information, lacking the ability to directly and in-situ detect key physicochemical states within the material (such as phase transitions and chemical composition).
[0003] This leads to a core contradiction: on the one hand, we are facing a highly dynamic and invisible processing "black box"; on the other hand, we can only operate it using a set of static "open-loop" instructions. The direct consequence is that when processing complex three-dimensional structures within challenging materials such as single-crystal silicon carbide, processing quality (such as aspect ratio, sidewall smoothness, and crack-free characteristics) is difficult to guarantee, the process window is narrow, and repeatability is poor. Therefore, how to achieve in-situ sensing of the inherent physicochemical changes during processing and, based on this, perform real-time, dynamic beam control is a bottleneck that urgently needs to be overcome in this field. Summary of the Invention
[0004] The technical problem to be solved by the present invention in view of the above-mentioned prior art is: how to overcome the problem of poor processing quality and low consistency caused by the imperceptible and uncontrollable static lag in the processing of the internal three-dimensional structure of difficult-to-process materials such as single-crystal silicon carbide when the existing femtosecond laser processing method is used.
[0005] To address the above problems, this invention provides an adaptive beam shaping method for the internal three-dimensional structure of single-crystal silicon carbide, comprising the following steps:
[0006] S1: Inject the femtosecond laser beam into the spatial light modulator;
[0007] S2: Use a spatial light modulator to perform wavefront modulation on a femtosecond laser beam and focus the modulated beam inside a single-crystal silicon carbide workpiece for processing.
[0008] S3: During the processing, at least two modes of process monitoring signals from the processing area are collected synchronously in real time. The process monitoring signals include plasma emission spectrum signals reflecting the transient physicochemical state of the material and interference morphology signals reflecting the changes in processing morphology.
[0009] S4: Based on real-time acquired process monitoring signals, the physical information neural network model dynamically calculates and generates control commands for the spatial light modulator to adjust the phase diagram loaded on the spatial light modulator in real time, thereby achieving adaptive closed-loop control of the three-dimensional spatial distribution of beam energy at the processing point; the physical information neural network model embeds the constraint equations describing the physical laws of the interaction between femtosecond laser and single-crystal silicon carbide into the learning process of the neural network.
[0010] As a further improvement of the present invention, the constraint equations include the two-temperature equations or a simplified model thereof.
[0011] As a further improvement of the present invention, the physical information neural network model has an online adaptive learning mechanism, which can use the deviation between the real-time monitoring signal and the model prediction value as a loss function to dynamically fine-tune the model parameters during the processing.
[0012] As a further improvement of the present invention, in step S3, the phase transition or stoichiometric change of the material is determined in real time by analyzing the intensity or peak shift of specific characteristic spectral lines of silicon atoms or carbon atoms in the plasma emission spectrum.
[0013] As a further improvement of the present invention, in step S3, the interference morphology signal is acquired by low coherence interferometry and used to obtain the depth information and surface roughness of the processing area in real time.
[0014] As a further improvement of the present invention, in step S4, the control commands calculated by the physical information neural network model based on the real-time signal include phase map corrections for dynamically adjusting the core size, depth of focus, or side lobe intensity of the Bessel beam.
[0015] As a further improvement of the present invention, in step S2, based on the digital model of the three-dimensional structure to be processed, an initial scanning path and a matching initial phase map sequence are pre-generated by a physical information neural network.
[0016] As a further improvement of the present invention, the specific process of step S4 is as follows: the real-time acquired spectral signal and morphology signal are fused into a high-dimensional vector, and input into the physical information neural network model along with the current processing position and the cumulative number of pulses. The model directly outputs the phase map data applied to the spatial light modulator at the next moment.
[0017] As a further improvement of the present invention, the method is used to process three-dimensional microchannels or microcavities with an aspect ratio greater than 10:1 inside single-crystal silicon carbide.
[0018] As a further improvement of the present invention, the sidewall roughness Ra of the three-dimensional microchannel or microcavity is less than 100 nm.
[0019] Compared with the prior art, the present invention has the following beneficial effects:
[0020] 1. The process has been made "transparent" and controllable. By introducing multimodal in-situ monitoring, especially the synchronous acquisition of plasma spectroscopy and interference morphology, this invention is the first to be able to acquire direct signals reflecting the evolution of the internal state and changes in the three-dimensional morphology of materials in real time, breaking the dilemma of "black box" processing.
[0021] 2. A dynamic adaptive control capability has been established. Based on real-time monitoring signals, this invention dynamically calculates and adjusts the phase diagram of the spatial light modulator through an intelligent decision model, enabling the beam shape and energy distribution to respond to transient changes in the processing, thus achieving a qualitative leap from "static preset" to "dynamic adaptation".
[0022] 3. Fundamentally improves processing limits and quality. This invention enables the processing of high aspect ratio, low damage, and high consistency three-dimensional microstructures inside single-crystal silicon carbide through a closed loop of "perception-decision-execution". Attached Figure Description
[0023] Figure 1 This is a flowchart of the method of the present invention;
[0024] Figure 2 This is a structural block diagram of the adaptive beam shaping system in this invention. Detailed Implementation
[0025] The present invention will now be described in further detail with reference to the accompanying drawings. It should be understood that the embodiments described herein are for illustrative purposes only and do not constitute a limitation on the scope of protection of the present invention.
[0026] Figure 1 An adaptive beam shaping method for the internal three-dimensional structure of single-crystal silicon carbide is shown. To make the technical solution of the present invention clearer, the following description will be combined with an adaptive beam shaping system for implementing the method.
[0027] Please see Figure 2 The adaptive beam shaping system mainly includes the following functional modules:
[0028] Femtosecond laser source: Used to generate femtosecond laser pulses with adjustable pulse width, repetition frequency, and single pulse energy.
[0029] The beam modulation and scanning module consists of a spatial light modulator (SLM), a beam expander, a two-dimensional scanning galvanometer, and a high numerical aperture objective lens. The spatial light modulator is the core actuator of this system, used to load a phase map according to control commands, thereby dynamically modulating the beam wavefront. The modulated beam is deflected by the scanning galvanometer and focused by the objective lens, then acts on a predetermined position inside the single-crystal silicon carbide workpiece.
[0030] Multimodal real-time monitoring unit: This is the system's "sensory organ," and includes at least:
[0031] Spectral diagnostic module: typically consists of a collecting lens, a fiber optic spectrometer, and a high-speed detector, used to acquire and analyze the plasma emission spectral signals generated at the processing point in real time.
[0032] Interference measurement module: preferably a white light interferometer or a confocal microscopy measurement system is used to acquire interference morphology signals of the processing area in real time.
[0033] AI Control Unit: This is the system's "decision-making brain," with a physical information neural network model running at its core. This unit connects to the spectral diagnostic module, interferometry module, spatial light modulator, and scanning galvanometer signal, and is used to receive real-time sensor data, execute intelligent decision-making algorithms, and issue control commands.
[0034] Based on the above system, the specific steps of the method of the present invention are as follows:
[0035] S0, Processing Pre-planning:
[0036] Before physical processing begins, the physical information neural network model performs pre-calculation and optimization based on the digital model of the three-dimensional structure to be processed (such as an STL file), generating an initial scanning path and a matching initial beam phase map sequence as the starting strategy for subsequent adaptive closed-loop processing.
[0037] S1. Beam modulation and incidence:
[0038] A beam from a femtosecond laser source is incident on a spatial light modulator. The spatial light modulator is used to modulate the wavefront phase of the beam. In this invention, its core function is to dynamically load different phase diagrams according to the intelligent instructions of the AI control unit, thereby shaping the incident Gaussian beam into a non-diffraction beam or structured light field with a specific three-dimensional energy distribution, such as a Bessel beam, an Airy beam, or a vortex beam.
[0039] S2, Focusing and Processing:
[0040] The beam modulated by the spatial light modulator is focused onto predetermined three-dimensional coordinate points inside a single-crystal silicon carbide workpiece by a scanning galvanometer and a high numerical aperture objective lens system, and then subjected to point-by-point or layer-by-layer ablation or modification processing.
[0041] S3. Synchronous real-time acquisition of multimodal process signals:
[0042] This step is crucial for achieving "sensing" in this method. While the femtosecond laser pulse acts on the material, two types of process signals from the processing point are simultaneously acquired through two independent channels of the multimodal real-time monitoring unit:
[0043] Plasma emission spectral signals reflect the transient physicochemical state of materials and are acquired by a spectral diagnostic module. By analyzing the intensity, peak position, or broadening of specific elemental characteristic lines in the spectrum (e.g., the characteristic line SiI at 390.55 nm for silicon atoms and the characteristic line CI at 247.86 nm for carbon atoms) in real time, the instantaneous electron temperature and density at the processing point can be retrieved, and it can be determined whether the material has undergone an amorphous phase transition or a change in stoichiometry. As an example, a sustained abnormal increase in the intensity of the SiI 390.55 nm spectral line can serve as an effective early warning indicator that the material is undergoing an amorphous phase transition.
[0044] Interference topography signal: Reflects changes in the processing topography and is acquired through low-coherence interferometry (interferometric measurement module). By processing this signal, topography information such as the depth, width, and surface roughness of the pits or structures formed during processing can be obtained in real time.
[0045] S4. Adaptive closed-loop control based on physical information neural network:
[0046] This step is the core of this method for achieving "decision-making" and "execution". Based on the high-dimensional sensor data stream that integrates spectral and morphological information acquired in real time in step S3, the AI control unit performs real-time calculations through its internal physical information neural network model and generates control commands for the spatial light modulator, forming a precise closed loop with a microsecond-level response.
[0047] 4.1 Physical Information Neural Network Model
[0048] This model is the intelligent hub of this invention, and its key difference from the general data-driven model lies in:
[0049] Physical constraint embedding: During model training and inference, constraint equations describing the physical laws governing the interaction between femtosecond lasers and single-crystal silicon carbide (e.g., the two-temperature equations describing energy transfer between electrons and the lattice, or their simplified models) are embedded as physical loss terms. This ensures that the model's predicted output always follows fundamental physical conservation laws, improving its extrapolation accuracy and reliability in scenarios with scarce training data.
[0050] The model features an online adaptive learning mechanism: during processing, the system compares real-time monitoring signals with the model's predicted values, and the deviation is used to construct a loss function for dynamically fine-tuning the model's parameters. As an exemplary implementation, the system can be configured to trigger a learning cycle after processing a predetermined number of pulses (e.g., approximately 1000 pulses): the error between the predicted and measured values of key parameters such as morphology and depth over a past period (e.g., hundreds of pulse cycles) is used as the loss function, and optimization algorithms such as gradient descent are employed to fine-tune some parameters of the model (e.g., the weights of the last two layers in a fully connected layer). The learning rate can be set, for example, on the order of 0.001. This allows the model to adapt to minute batch variations of specific materials, environmental fluctuations, or long-term drift during processing.
[0051] Input and Output: The model's input includes at least real-time multimodal sensing data, the three-dimensional coordinates of the current processing point, and the number of pulses applied, among other state information. The model's output is the phase map data to be loaded onto the spatial light modulator at the next moment, which is essentially a correction to the beam wavefront.
[0052] 4.2 Closed-loop control process
[0053] The specific process of adaptive closed-loop control is as follows:
[0054] Information Fusion and Input: The AI control unit fuses the synchronously acquired plasma spectral signals (extracted features such as specific spectral line intensities) and interference topography signals (extracted features such as depth and roughness) into a multi-dimensional feature vector in real time. This feature vector, along with the current processing point's three-dimensional coordinates, the accumulated number of laser pulses, and other processing status information, is input into the physical information neural network model. Based on this, the model directly maps and outputs the specific phase map data to be applied to the spatial light modulator at the next moment.
[0055] Intelligent Decision-Making: Based on its embedded physical laws and learned mapping relationships, the model calculates the optimal three-dimensional energy distribution of the beam required to achieve the ideal modification effect (such as uniform ablation, avoidance of thermal damage) for the current processed voxel point. This decision is directly reflected in the adjustment of the phase diagram. Those skilled in the art will understand that by designing a specific phase diagram using computational holographic algorithms, it is possible to precisely control physical parameters such as the core size, diffraction-free length (depth of focus), and sidelobe intensity distribution of the generated Bessel beam. Therefore, this invention, by adjusting the phase diagram in real time, can achieve operations such as dynamically optimizing the core size of the Bessel beam, thereby precisely controlling the energy deposition in the axial and radial directions.
[0056] Precise execution: The phase map data output by the model is sent to the spatial light modulator in real time by the AI control unit. The spatial light modulator changes the beam wavefront accordingly, so that the shape of the beam acting on the next processing voxel is corrected in real time, thereby achieving "dynamic shaping" of the spatiotemporal distribution of processing energy.
[0057] This invention is particularly suitable for processing complex three-dimensional microstructures with an aspect ratio greater than 10:1 inside single-crystal silicon carbide, such as vertical microchannels, microcavities, or three-dimensional photonic crystals. Through the closed-loop control described above, the processing quality can be significantly improved, achieving a highly smooth surface with a sidewall roughness Ra of less than 100 nm, and effectively suppressing microcracks and amorphization.
[0058] To more intuitively demonstrate the effects of the method of the present invention, the following description uses a specific processing task based on the above system and method as an example.
[0059] Processing objective: To fabricate a vertical cylindrical microchannel with a diameter of 30μm and a depth of 400μm inside a 4H-SiC single wafer with a thickness of 500μm.
[0060] Implementation process:
[0061] The workpiece is fixed, and the target 3D model is input. Based on S0, the AI control unit pre-generates the initial scanning path (top-down spiral scan) and the initial Bessel beam phase map.
[0062] The system initiates closed-loop processing from S1 to S4. When processing to a depth of approximately 250 μm, the interferometry module detects that the material removal rate is lower than expected, while the spectral diagnostic module detects an abnormal increase in the intensity of silicon characteristic spectral lines.
[0063] The physical information neural network model in the AI control unit comprehensively judges that this phenomenon is caused by nonlinear absorption accumulation leading to beam waist energy saturation and potential thermal damage risk. The model instantly calculates the solution: generating a phase correction map with a specific ring-shaped occlusion.
[0064] The correction plot was immediately loaded into the spatial light modulator, which partially suppressed the sidelobe energy of the subsequent processed beam and optimized the core energy distribution.
[0065] Processing continues until completion.
[0066] Comparison of effects:
[0067] Compared with open-loop machining using a fixed Bessel beam, the machining results obtained by this method show significant improvements in aspect ratio (reaching 13:1), channel diameter consistency (deviation < ±1 μm), and sidewall roughness (Ra < 80 nm). Raman spectroscopy confirms that the edges of the structures machined by this method are essentially free of amorphous phases, achieving "cold machining".
[0068] It should be noted that the core principle of the adaptive closed-loop control method based on multimodal real-time sensing and physical information neural network provided by this invention lies in achieving dynamic response and precise control to the inherent physical changes during the processing. This principle has universal guiding significance for solving the internal three-dimensional structure processing problems of other hard and brittle materials (such as diamond and sapphire) that face similar dynamic nonlinear effects and thermal management challenges in femtosecond laser processing. Those skilled in the art, based on their understanding of the inventive concept, can apply the method of this invention to the processing of related materials by adjusting the specific material physical parameters (such as thermal conductivity and nonlinear absorption coefficient) embedded in the physical information neural network. This invention uses single-crystal silicon carbide, a typical and highly challenging material, as an example for detailed explanation, and should not be construed as limiting the scope of protection of this invention.
Claims
1. An adaptive beam shaping method for the internal three-dimensional structure of single-crystal silicon carbide, characterized in that, Includes the following steps: S1: Inject the femtosecond laser beam into the spatial light modulator; S2: The femtosecond laser beam is wavefront modulated using the spatial light modulator, and the modulated beam is focused into the interior of a single-crystal silicon carbide workpiece for processing. S3: During the processing, at least two modes of process monitoring signals from the processing area are collected synchronously and in real time. The process monitoring signals include plasma emission spectrum signals reflecting the transient physicochemical state of the material and interference morphology signals reflecting the changes in processing morphology. S4: Based on the real-time acquired process monitoring signals, the physical information neural network model dynamically calculates and generates control commands for the spatial light modulator to adjust the phase diagram loaded on the spatial light modulator in real time, thereby realizing adaptive closed-loop control of the three-dimensional spatial distribution of beam energy at the processing point; the physical information neural network model embeds the constraint equations describing the physical laws of the interaction between femtosecond laser and single-crystal silicon carbide into the learning process of the neural network.
2. The adaptive beam shaping method for the internal three-dimensional structure of single-crystal silicon carbide according to claim 1, characterized in that, The constraint equations include the two-temperature equations or their simplified models.
3. The adaptive beam shaping method for the internal three-dimensional structure of single-crystal silicon carbide according to claim 1, characterized in that, The physical information neural network model has an online adaptive learning mechanism, which can use the deviation between the real-time monitoring signal and the model prediction as a loss function to dynamically fine-tune the model parameters during the processing.
4. The adaptive beam shaping method for the internal three-dimensional structure of single-crystal silicon carbide according to claim 1, characterized in that, In step S3, the material phase transition or stoichiometric changes are determined in real time by analyzing the intensity or peak shift of specific characteristic spectral lines of silicon or carbon atoms in the plasma emission spectrum.
5. The adaptive beam shaping method for the internal three-dimensional structure of single-crystal silicon carbide according to claim 1, characterized in that, In step S3, the interference morphology signal is acquired by low coherence interferometry and is used to obtain the depth information and surface roughness of the processing area in real time.
6. The adaptive beam shaping method for the internal three-dimensional structure of single-crystal silicon carbide according to claim 1, characterized in that, In step S4, the physical information neural network model calculates control commands based on real-time signals, including phase diagram corrections for dynamically adjusting the core size, depth of focus, or side lobe intensity of the Bessel beam.
7. The adaptive beam shaping method for the internal three-dimensional structure of single-crystal silicon carbide according to claim 1, characterized in that, In step S2, based on the digital model of the three-dimensional structure to be processed, the physical information neural network pre-generates an initial scanning path and a matching initial phase map sequence.
8. The adaptive beam shaping method for the internal three-dimensional structure of single-crystal silicon carbide according to claim 1, characterized in that, The specific process of step S4 is as follows: the real-time acquired spectral signal and morphology signal are fused into a high-dimensional vector, and the vector is input into the physical information neural network model along with the current processing position and the cumulative number of pulses. The model then directly outputs the phase map data applied to the spatial light modulator at the next moment.
9. The adaptive beam shaping method for the internal three-dimensional structure of single-crystal silicon carbide according to claim 1, characterized in that, Used for fabricating three-dimensional microchannels or microcavities with an aspect ratio greater than 10:1 inside single-crystal silicon carbide.
10. The adaptive beam shaping method for the internal three-dimensional structure of single-crystal silicon carbide according to claim 9, characterized in that, The sidewall roughness Ra of the three-dimensional microchannel or microcavity is less than 100 nm.